Lecture Notes in Electrical Engineering 758

Rosdiazli Ibrahim ·
K. Porkumaran · Ramani Kannan ·
Nursyarizal Mohd Nor ·
S. Prabakar Editors

International
Conference on Artificial
Intelligence for Smart
Community
AISC 2020, 17–18 December, Universiti
Teknologi Petronas, Malaysia
Lecture Notes in Electrical Engineering

Volume 758

Series Editors

Leopoldo Angrisani, Department of Electrical and Information Technologies Engineering, University of Napoli
Federico II, Naples, Italy
Marco Arteaga, Departament de Control y Robótica, Universidad Nacional Autónoma de México, Coyoacán,
Mexico
Bijaya Ketan Panigrahi, Electrical Engineering, Indian Institute of Technology Delhi, New Delhi, Delhi, India
Samarjit Chakraborty, Fakultät für Elektrotechnik und Informationstechnik, TU München, Munich, Germany
Jiming Chen, Zhejiang University, Hangzhou, Zhejiang, China
Shanben Chen, Materials Science and Engineering, Shanghai Jiao Tong University, Shanghai, China
Tan Kay Chen, Department of Electrical and Computer Engineering, National University of Singapore,
Singapore, Singapore
Rüdiger Dillmann, Humanoids and Intelligent Systems Laboratory, Karlsruhe Institute for Technology,
Karlsruhe, Germany
Haibin Duan, Beijing University of Aeronautics and Astronautics, Beijing, China
Gianluigi Ferrari, Università di Parma, Parma, Italy
Manuel Ferre, Centre for Automation and Robotics CAR (UPM-CSIC), Universidad Politécnica de Madrid,
Madrid, Spain
Sandra Hirche, Department of Electrical Engineering and Information Science, Technische Universität
München, Munich, Germany
Faryar Jabbari, Department of Mechanical and Aerospace Engineering, University of California, Irvine, CA,
USA
Limin Jia, State Key Laboratory of Rail Traffic Control and Safety, Beijing Jiaotong University, Beijing, China
Janusz Kacprzyk, Systems Research Institute, Polish Academy of Sciences, Warsaw, Poland
Alaa Khamis, German University in Egypt El Tagamoa El Khames, New Cairo City, Egypt
Torsten Kroeger, Stanford University, Stanford, CA, USA
Yong Li, Hunan University, Changsha, Hunan, China
Qilian Liang, Department of Electrical Engineering, University of Texas at Arlington, Arlington, TX, USA
Ferran Martín, Departament d’Enginyeria Electrònica, Universitat Autònoma de Barcelona, Bellaterra,
Barcelona, Spain
Tan Cher Ming, College of Engineering, Nanyang Technological University, Singapore, Singapore
Wolfgang Minker, Institute of Information Technology, University of Ulm, Ulm, Germany
Pradeep Misra, Department of Electrical Engineering, Wright State University, Dayton, OH, USA
Sebastian Möller, Quality and Usability Laboratory, TU Berlin, Berlin, Germany
Subhas Mukhopadhyay, School of Engineering & Advanced Technology, Massey University,
Palmerston North, Manawatu-Wanganui, New Zealand
Cun-Zheng Ning, Electrical Engineering, Arizona State University, Tempe, AZ, USA
Toyoaki Nishida, Graduate School of Informatics, Kyoto University, Kyoto, Japan
Federica Pascucci, Dipartimento di Ingegneria, Università degli Studi “Roma Tre”, Rome, Italy
Yong Qin, State Key Laboratory of Rail Traffic Control and Safety, Beijing Jiaotong University, Beijing, China
Gan Woon Seng, School of Electrical & Electronic Engineering, Nanyang Technological University,
Singapore, Singapore
Joachim Speidel, Institute of Telecommunications, Universität Stuttgart, Stuttgart, Germany
Germano Veiga, Campus da FEUP, INESC Porto, Porto, Portugal
Haitao Wu, Academy of Opto-electronics, Chinese Academy of Sciences, Beijing, China
Walter Zamboni, DIEM - Università degli studi di Salerno, Fisciano, Salerno, Italy
Junjie James Zhang, Charlotte, NC, USA
The book series Lecture Notes in Electrical Engineering (LNEE) publishes the
latest developments in Electrical Engineering—quickly, informally and in high
quality. While original research reported in proceedings and monographs has
traditionally formed the core of LNEE, we also encourage authors to submit books
devoted to supporting student education and professional training in the various
fields and applications areas of electrical engineering. The series cover classical and
emerging topics concerning:
• Communication Engineering, Information Theory and Networks
• Electronics Engineering and Microelectronics
• Signal, Image and Speech Processing
• Wireless and Mobile Communication
• Circuits and Systems
• Energy Systems, Power Electronics and Electrical Machines
• Electro-optical Engineering
• Instrumentation Engineering
• Avionics Engineering
• Control Systems
• Internet-of-Things and Cybersecurity
• Biomedical Devices, MEMS and NEMS

For general information about this book series, comments or suggestions, please
contact [email protected].
To submit a proposal or request further information, please contact the Publishing
Editor in your country:
China
Jasmine Dou, Editor ([email protected])
India, Japan, Rest of Asia
Swati Meherishi, Editorial Director ([email protected])
Southeast Asia, Australia, New Zealand
Ramesh Nath Premnath, Editor ([email protected])
USA, Canada
Michael Luby, Senior Editor ([email protected])
All other Countries
Leontina Di Cecco, Senior Editor ([email protected])
** This series is indexed by EI Compendex and Scopus databases. **
Rosdiazli Ibrahim · K. Porkumaran ·
Ramani Kannan · Nursyarizal Mohd Nor ·
S. Prabakar
Editors

International Conference
on Artificial Intelligence
for Smart Community
AISC 2020, 17–18 December, Universiti
Teknologi Petronas, Malaysia
Editors
Rosdiazli Ibrahim K. Porkumaran
Department of Electrical and Electronics Department of Electrical and Electronics
Engineering Engineering
Universiti Teknologi PETRONAS Sri Sairam Engineering College
Seri Iskandar, Perak, Malaysia Chennai, India

Ramani Kannan Nursyarizal Mohd Nor

Department of Electrical and Electronics Department of Electrical and Electronics
Engineering Engineering
Universiti Teknologi PETRONAS Universiti Teknologi PETRONAS
Seri Iskandar, Perak, Malaysia Seri Iskandar, Perak, Malaysia

S. Prabakar
Department of Biomedical Engineering
Sona College of Technology
Salem, India

ISSN 1876-1100 ISSN 1876-1119 (electronic)

Lecture Notes in Electrical Engineering
ISBN 978-981-16-2182-6 ISBN 978-981-16-2183-3 (eBook)
https://doi.org/10.1007/978-981-16-2183-3

© The Editor(s) (if applicable) and The Author(s), under exclusive license to Springer Nature
Singapore Pte Ltd. 2022
This work is subject to copyright. All rights are solely and exclusively licensed by the Publisher, whether
the whole or part of the material is concerned, specifically the rights of translation, reprinting, reuse
of illustrations, recitation, broadcasting, reproduction on microfilms or in any other physical way, and
transmission or information storage and retrieval, electronic adaptation, computer software, or by similar
or dissimilar methodology now known or hereafter developed.
The use of general descriptive names, registered names, trademarks, service marks, etc. in this publication
does not imply, even in the absence of a specific statement, that such names are exempt from the relevant
protective laws and regulations and therefore free for general use.
The publisher, the authors, and the editors are safe to assume that the advice and information in this book
are believed to be true and accurate at the date of publication. Neither the publisher nor the authors or
the editors give a warranty, expressed or implied, with respect to the material contained herein or for any
errors or omissions that may have been made. The publisher remains neutral with regard to jurisdictional
claims in published maps and institutional affiliations.

This Springer imprint is published by the registered company Springer Nature Singapore Pte Ltd.
The registered company address is: 152 Beach Road, #21-01/04 Gateway East, Singapore 189721,
Singapore
Contents

Design of PID Controller for Integrating Processes with Inverse

Response . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
A. Parvathi, M. Rathaiah, R. Kiranmayi, and K. Nagabhushanam
Environmental Feasibility Survey of Solar Photovoltaic Cells . . . . . . . . . 13
Sai Goutm Golive, B. Vijaya Krishna, B. Parama Sivam,
and K. Ramash Kumar
Intelligent Control Techniques for Parameter Tuning of PID
Controller for LFC with Emphasis on Genetic Algorithm . . . . . . . . . . . . 25
B. Venkata Prasanth, Raja Sathish Kumar, Vijaya Krishna Boyina,
and K. Ramash Kumar
Simplified Decoupler Based Fractional Order PID Controller
for Two Variable Fractional Order Process . . . . . . . . . . . . . . . . . . . . . . . . . 35
A. Divya Teja, R. Kiranmayi, K. Nagabhushanam, and N. Swathi
Improved Centralized PID Controller with Disturbance
Rejection for LTI Multivariable Processes . . . . . . . . . . . . . . . . . . . . . . . . . . 49
S. Anitha, R. Kiranmayi, K. Nagabhushanam,
and K. Firdose Kowser Ahamadia
Tuning of PID Controller Using SIMC Method for Systems
with Time Delay and RHP Poles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
P. Keerthi, R. Kiranmayi, K. Nagabhushanam,
and K. Firdose Kowser Ahamadia
Independent Controller Design for Non-minimum Phase Two
Variable Process with Time Delay . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81
C. Venkatasuresh, M. Rathaiah, R. Kiranmayi, and K. Nagabhushanam

v
vi Contents

Control of DC Link Voltage and Load Voltage Variations

in a Pitch Angle Controlled PMSG Based Wind Energy
Conversion System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
A. Nagarajan, K. Ramash Kumar, and G. Sureshkumaar
A Novel Approach of Wind MPPT Using Fuzzy Logic . . . . . . . . . . . . . . . . 107
Nesar Uddin, Md Saiful Islam, Jewel Sikder Joy,
and Ishaque Mozumder
Shale Gas Productive Volume Optimization . . . . . . . . . . . . . . . . . . . . . . . . . 117
Adamu Umar Ibrahim and Berihun Mamo Negash
A Comparative Study and Validation of Kinematic Analysis
of a Crank Rocker Engine Prototype Using MATLAB
and ADAMS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125
A. M. Albaghdadi, M. B. Baharom, and S. A. Sulaiman
Application of Machine Learning Models in Gas Hydrate
Mitigation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 135
Sachin Dev Suresh, Bhajan Lal, Ali Qasim, Khor Siak Foo,
and Jega Divan Sundramoorthy
Analysis of Signal Sensing with Adaptive Threshold for Energy
Detector in Cognitive Radio Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 145
Biji Rose and B. Aruna Devi
Collaborative Design in Concurrent Engineering of Industry 4.0
Through Virtual Reality Simulation in Achieving Accelerated
Time-To-Market During COVID-19 (Coronavirus) Pandemic
Outbreak . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 153
A. F. Fudzin, M. Amin, and A. Mokhtar
Stability Analysis of Semi-Markovian Discrete-Time Neural
Networks with Time-Varying Leakage Delays . . . . . . . . . . . . . . . . . . . . . . . 165
K. Maheswari, S. N. Shivapriya, and C. Ramkumar
Forecasting PM10 Concentration Based on a Hybrid Fuzzy
Time Series Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 177
Yousif Alyousifi and Mahmod Othman
Carbonated Water Injection for EOR and CO2 Storage:
A Simulation Study . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 185
A. Abdulrahman and J. Foroozesh
Study of Efficient FIR Filter Architecture for Fixed
and Reconfigurable Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 197
B. Aruna Devi, V. Prasanna Devi, and S. Preethi
Contents vii

Improved Deep Learning Based Prediction of Crop Yield Using

Bidirectional Long Short Term Memory . . . . . . . . . . . . . . . . . . . . . . . . . . . . 201
V. Saveetha, R. Kingsy Grace, S. Sophia, and S. Birundha
Instrument Cluster IoT Enclosure Design and Production
Implementation in Self Driven Vehicles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 211
R. Thirumalai, M. Seenivasan, A. Sivakumar, M. Vivekraj,
and V. Balaji
General Adversarial Networks: A Tool to Detect the Novel
Coronavirus from CT Scans . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 217
R. Shriram, T. R. Kaushek Kumar, V. Samuktha, and R. Karthika
VLSI Implementation of Multipliers for Artificial Intelligence
Applications: A Survey . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 231
R. Karthick, S. Shanmuga Raju, G. Abinaya, G. Ashwanth,
and G. Kaviya
Automated Boneage Analysis Using Machine Learning . . . . . . . . . . . . . . . 241
J. K. Krithika, Yayha Norashikin, K. Porkumaran, and S. Prabakar
Multi-Class SVM Prediction Model for Lung Cancer Diagnosis . . . . . . . 253
D. Lakshmi, J. Sivakumar, and S. Ramani
Survey on Fire Safety Robot & Implementation of Android
Application . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 265
K. Sakthisudhan, S. Rubika, S. Sadhasivam, V. Suguna,
and Sri Muruga Rajan
Detection of Emergency Vehicles Using Radio Frequency
Identification (RFID) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 271
Usha Mittal and Priyanka Chawla
Big Data-Enabled Solutions for Covid-19 . . . . . . . . . . . . . . . . . . . . . . . . . . . 279
Dirisala Saikrishna, Bavisetti Sai Dhanush, Madhusudhan Rao,
Priyanka Chawla, and Usha Mittal
Smart Traffic Monitoring System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 287
V. Radhika, V. Madhan Babu, and S. Jayaprakash
Harvesting Electrical Energy from Body Heat to Power Bio
Medical Wearables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 297
M. Preethika, Mohamad Radzi Bin Ahmad, K. Porkumaran,
S. Prabakar, and R. Sundar
The Significant Relationship Between E-Governance System
and Government Operation Excellence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 311
Almahdy Alhaj Saleh, Ahmed Rashid Alkhuwaylidee,
Murugan Thangiah, and Arif Razzaq
viii Contents

Trajectory Tracking Control of Industrial Robot Manipulator

Using VS Control System with STHWS Method . . . . . . . . . . . . . . . . . . . . . 325
S. Nandhakumar, B. Shreeram, R. Thirumalai, K. Sakthisudhan,
and P. Sivaprakash
Medical Fitness Device for Driving Assessment . . . . . . . . . . . . . . . . . . . . . . 335
K. Chamundeswari, R. Saranya, M. Smitha, S. Prabakar,
and K. Porkumaran
Application of Artificial Intelligence for Reservoir Storage
Prediction: A Case Study . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 343
Abdus Samad Azad, Pandian M. Vasant, José A. Gámez Vintaned,
and Junzo Watada
Fabrication of Parallel Mechanism Actuated by Pneumatic
Artificial Muscle for Rehabilitation Therapy . . . . . . . . . . . . . . . . . . . . . . . . 355
S. Krishnan, A. M. A. Rani, Laxamanan G. Kurappa,
and Sivajothi Paramasivam
A Review of Artificial Intelligence-Based Techniques to Estimate
Atmospheric Parameters Influencing the Performance
of Concentrating Photovoltaic/Thermal Systems . . . . . . . . . . . . . . . . . . . . . 365
F. Masood, P. Nallagownden, I. Elamvazuthi, J. Akhter,
and M. A. Alam
On the Performance of MIMO-UVLC System
over Turbulence-induced Fading Channels . . . . . . . . . . . . . . . . . . . . . . . . . . 373
Sajid Mumtaz, Azrina Abd Aziz, and Komal Masroor
Fabrication of One Degree of Freedom Knee Exoskeleton
for Nervous Disorder . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 385
J. Duwaraka, N. Sudharsana, S. Prabakar, K. Porkumaran,
Iraivan Ilamvaluthi, and Ramani Kannan
Semi-Automatic Detection and Measurement of Fetal
Parameters from Ultrasound Images and the Scope Automatic
System Using LabVIEW . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 393
S. Prabakar, K. Porkumaran, J. Samson Isaac, R. Karthikeyan,
G. Gopu, and Ramani Kannan
A Framework for Enhancing Network Lifetime in Internet
of Things Environment Using Clustering Formation . . . . . . . . . . . . . . . . . 401
Abdulrahman Aminu Ghali, Rohiza Ahmad, and Hitham Alhussian
An Empirical Analysis on Big Analytics for e-Healthcare
and Agriculture . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 409
N. Purandhar and S. Ayyasamy
Contents ix

Enhanced Approach in VANETs for Avoidance of Collision

with Reinforcement Learning Strategy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 419
A. Ganesh and S. Ayyasamy
VBlock—Blockchain-Based Traceability in Medical Products
Supply Chain Management: A Case Study in VietNam . . . . . . . . . . . . . . . 429
Tuan Khoi Nguyen Huynh, Tuan Anh Dao, Trong Nhan Van Pham,
Khang Hy Vuong Nguyen, Nghi Cong Tran, and Hoang Huong Luong
Development of Real-Time Internet of Things (IoT) Based
Water Quality Monitoring System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 443
Huzein Fahmi bin Hawari, Mohamad Nor Syahid bin Mokhtar,
and Sohail Sarang
Edge Computing Technology: State-of-the-Art and Challenges
in an Internet of Things . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 455
Deepa Raghunathan and M. Krishnamoorthi
Comparison of Rectangular Tunnel with Shield Jacking Support
and Pipe Roof Support in Finite Element Method . . . . . . . . . . . . . . . . . . . 469
M. Y. M. Nasir, H. Mohamad, and H. Alarifi
The Effect of Skill Type on Skill-Gap in the Nigerian
Construction Industry . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 479
M. S. Aminu, S. U. Kunya, I. Y. Mohammed, and S. A. Bustani
Propagation Model of Molecular Communication Based
Targeted Drug Delivery for Atherosclerosis Disease Therapy . . . . . . . . . 499
Pradeep Murugesan, S. Prabakar, K. Porkumaran, and R. Karthikeyan
Adaptation of Machine Learning and Blockchain Technology
in Cyber-Physical System Applications: A Concept Paper . . . . . . . . . . . . 517
Mujaheed Abdullahi, Hitham Alhussian, and Norshakirah Aziz
Application of Artificial Neural Network (ANN) and Adaptive
Neuro Fuzzy (ANFIS) Techniques for the Modelling
and Optimization of COD Adsorption Process . . . . . . . . . . . . . . . . . . . . . . . 525
Hifsa Khurshid, Muhammad Raza Ul Mustafa, and Yeek-Chia Ho
An Evolutionary Stream Clustering Technique Outlier Detection
in Medical Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 539
Nadilah Ayu Supardi, Said Jadid Abdulkadir, and Norshakirah Aziz
Analysis of Intelligent Agent and Its Components for the SLA
Negotiation Process in Cloud Computing . . . . . . . . . . . . . . . . . . . . . . . . . . . 547
Rishi Kumar, Mohd Fadzil Hassan, and Muhamad Hariz M. Adnan
x Contents

Early Detection of Myocardial Infarction Using Machine

Learning with Maximum Accuracy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 553
S. Abirami Manisa, B. J. Abarna, V. Geethanjali, G. V. Hari Venkat,
and R. Karthikeyan
Design and Development of a Real Time Mouse for Human
Computer Interaction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 565
U. Vinothkumar, D. Kavya, M. Abinesh, M. Kavin, and P. Aravind
Design and Implementation of Low Cost Energy Meter Using
MIT App Inventor . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 573
P. Gajendran, S. N. Deepa, and N. Rajasingam
Detection of Sybil Attack in Internet of Things . . . . . . . . . . . . . . . . . . . . . . 581
C. Vinothini, J. Priya, P. Anitha, and T. S. Reshmi
Diagnosis of Brain Tumor Using ANN with Spatial Fuzzy
Clustering and Active Level Set Contour . . . . . . . . . . . . . . . . . . . . . . . . . . . 589
Hafiza Akter Munira and Md Saiful Islam
Current Overview of Machine Learning Application
for Predicting Steam Huff and Puff Injection Production
Performance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 597
Mohammad Galang Merdeka, Syahrir Ridha,
Berihun Mamo Negash, and Suhaib Umer Ilyas
Enhanced Cluster Head Based Data Gathering (ECHGS)
Technique in IoT Based Smart Irrigation System . . . . . . . . . . . . . . . . . . . . 605
K. Padmanaban, A. M. SenthilKumar, A. K. Velmurugan,
and E. S. Madhan
Estimation of State of Charge Using EKF and SVR in Li-Ion
Battery . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 615
P. Sannihith Reddy and R. Shanmughasundaram
Evolving Spiking Neural Network for Prediction Problems
in Healthcare System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 625
Tasbiha Ibad, Said Jadid Abdulkadir, and Norshakirah Binti Ab Aziz
Prototyping Pro-Active Wearable Gadget for the Surveillance
of Coal Miners in Pakistan . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 633
Rahul Kumar, Ramani Kannan, Jayasankari Ganasan,
Ghulam E Mustafa Abro, Abdul Sattar, Vipin Kumar,
and Nirbhay Mathur
Design of a Self-Tuning PID Controller for a Temperature
Control System Using Fuzzy Logic . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 643
Md. Tauhidul Islam, Ariful Islam, Rahul Kumar,
Ghulam E Mustafa Abro, Sourav Majumdar, and Vipin Kumar Oad
Contents xi

Design of a Low-Cost High-Speed and Large-Capacity Data

Logger Using MicroSD Card . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 651
Muhammad Sajjad, Mohd Zuki Yusoff, and Junaid Ahmad
Machine Learning in Healthcare: Current Trends and the Future . . . . . 659
Usman Ahmad Usmani and Jafreezal Jaafar
Predictive Maintenance for a Turbofan Engine Using Data
Mining . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 677
Ismaila Mahmud, Idris Ismail, and Zuhairi Baharudin
Cyber Security Breaches and the Long-Run Effect on Firms’
Market Value: A Conceptual Framework . . . . . . . . . . . . . . . . . . . . . . . . . . . 689
Syed Emad Azhar Ali and Fong-Woon Lai
Towards Early Distribution of Container-Based Microservices
in Cloud Computing Environment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 699
Abdul Saboor, Ahmad Kamil Mahmood, Mohd Fadzil Hassan,
Syed Nasir Mehmood Shah, Muhammad Aadil Siddiqui,
Saeed Ahmed Magsi, and Muhammad Junaid
An AI-Based Chicken Disease Management System . . . . . . . . . . . . . . . . . . 707
Luyl-Da Quach, Nghi Pham Quoc, Nhien Huynh Thi,
Nhan Le Thanh, Xuan Truong Thi Thanh, Duc Chung Tran,
and Mohd Fadzil Hassan
Automatic Polyp Segmentation in Colonoscopy Images Using
Single Network Model: SegNet . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 717
Chin Yii Eu, Tong Boon Tang, and Cheng-Kai Lu
An Agile System for FiO2 Regulation in Ventilators . . . . . . . . . . . . . . . . . . 725
S. Prabakar, J. Samson Isaac, R. Karthikeyan, K. Porkumaran,
and Ramani Kannan
Computer-Aided Diagnostic Tool for Classification of Colonic
Polyp Assessment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 735
Win Sheng Liew, Tong Boon Tang, and Cheng-Kai Lu
Live Obstacle Detection Audio System for Visually Impaired
People . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 745
P. Anitha, C. Vinothini, J. Priya, and S. Guhan
Reducing the Aleatoric Uncertainties of Failure Prediction
Using Singular Value Decomposition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 755
Ahmad Kamal Mohd Nor, Srinivasa Rao Pedapati,
and Masdi Muhammad
xii Contents

Reducing Uncertainty in Failure Prediction Using Singular

Value Decomposition Feature Selection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 775
Ahmad Kamal Mohd Nor, Srinivasa Rao Pedapati,
and Masdi Muhammad
Unsupervised Classification of Acoustic Emission Signal
to Discriminate Composite Failure at Low Frequency . . . . . . . . . . . . . . . . 797
Noor A’in A. Rahman, Zazilah May, and Muhamad Shazwan Mahmud
Finding the EEG Footprint of Stress Resilience . . . . . . . . . . . . . . . . . . . . . . 807
Rumaisa Abu Hasan, Syed Saad Azhar Ali, Tong Boon Tang,
and Muhamad Saiful Bahri Yusoff
Modeling and Tuning of PID Controller for Continuous Stirred
Tank Reactor . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 817
A. Suguna, S. N. Deepa, and N. Rajasingam
Adaptive Security for Cognitive Robotic Process Automation
in Enterprise Computing Using AI-Powered Analytical Engine . . . . . . . 825
Mohamed Ibrahim Beer Mohamed and Mohd Fadzil Hassan
Multidirection Features Based Image Inpainting
with Color-Direction Patch-Sparsity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 837
B. Vidhya and M. Nikhil Madhav
Classification of Sub-frequency Bands Based Two-Class Motor
Imagery Using CNN . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 851
Muhammad Ahsan Awais, Mohd Zuki Yusoff, and Norashikin Yahya
Prediction of Methane Hydrate Formation Rate in Multiphase
System using Artificial Neural Network . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 859
Wan Adli Nuaim Bin Wan Ahmad Nadzri, Omar Nashed,
Bhajan Lal, Khor Siak Foo, and Khalik Mohd Sabil
Chemometrics Analysis and Wavelength Biomarker
Identification Using Fourier Transform Infrared Spectroscopy
for Lard Adulteration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 867
Muhammad Aadil Siddiqui, Mohd Haris Md Khir,
Gunawan Witjaksono, Muhammad Junaid, Saeed Ahmed Magsi,
and Abdul Saboor
Deep Learning Approach for Divergence Behavior Detection
at High Density Crowd . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 875
M. U. Farooq, M. N. Mohamad Saad, Y. Saleh, and S. Daud Khan
Anomaly Localization at High-Density Crowd Using Motion
Shape Image (MSI) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 889
M. U. Farooq, M. N. Mohamad Saad, S. Daud Khan, and Y. Saleh
Contents xiii

EEG Based Brain Controlled RC Car with Attention Level . . . . . . . . . . . 907

Muhammad Talha Ejaz, Ammara Zahid,
and Muhammad Mudassir Ejaz
Automated Pill and Syringe Dispenser . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 919
R. Paviya, S. Prabakar, K. Porkumaran, and Abu Bakar Sayuti Saman
Multi-classification of Brain Tumor Images Using Transfer
Learning Based Deep Neural Network . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 927
Pramit Dutta, Khaleda Akhter Sathi, and Md. Saiful Islam
A Meta Model Based Particle Swarm Optimization for Enhanced
Global Search . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 935
Rasel Ahmed, Shuhaimi Mahadzir, and Nor Erniza Mohammad Rozali
HCShare: Blockchain Technology in Healthcare System . . . . . . . . . . . . . 945
R. Sangeetha and M. Krishnamoorthi
SAZZ Converter Fed Fuzzy Logic Speed Controlled BLDC
Motor Drive . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 953
A. Senthilnathan, P. Palanivel, and R. Balakrishnan
Pedestrian Re-identification in Video Surveillance System
with Improved Feature Extraction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 961
Sina Salehian, Patrick Sebastian, and Abu Bakar Sayuti
Smart Children Management Using Data Analytics, Machine
Learning and IoT . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 977
Md Jobair Hossain Faruk and Muhamad Hariz Muhamad Adnan
Optical Wireless Communication Based Wireless Body Area
Network for Remote Patient Monitoring . . . . . . . . . . . . . . . . . . . . . . . . . . . . 985
Komal Masroor, Sajid Mumtaz, Micheal Drieberg, and Varun Jeoti
Development of an IoT-Enabled Stroke Rehabilitation System . . . . . . . . 993
Huzein Fahmi Bin Hawari and Safuan Bin Abu
MRI Brain Tumor Classification Technique Using Fuzzy
C-Means Clustering and Artificial Neural Network . . . . . . . . . . . . . . . . . . 1005
Angona Biswas and Md Saiful Islam
An IoT Based Automatic Vehicle Accident Detection and Rescue
System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1013
K. Gayathri Devi, K. Yasoda, B. Rajesh, R. Sowmiya,
and S. S. Vishalidevi
Automatic Firefighting System Using Unmanned Aerial Vehicle . . . . . . . 1023
K. Gayathri Devi, K. Yasoda, and Maria Nithin Roy
xiv Contents

State-of-the Art: Short Text Semantic Similarity (STSS)

Techniques in Question Answering Systems (QAS) . . . . . . . . . . . . . . . . . . 1033
Zaira Hassan Amur, Yewkwang Hooi, Irum Naz Sodhar,
Hina Bhanbhro, and Kamran Dahri
Modeling of Wireless Power Transmission for Pacemaker
Battery Charging in MATLAB Simulink Environment . . . . . . . . . . . . . . . 1045
G. P. Kiruthiga, Mohdfakhizan B. Romlie, K. Porkumaran,
and S. Prabakar
Recent t-way Test Generation Strategies Based on Optimization
Algorithms: An Orchestrated Survey . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1055
Ammar K. Alazzawi, Helmi Md Rais, Shuib Basri,
Yazan A. Alsariera, Abdullahi Abubakar Imam, Saad Adnan Abed,
Abdullateef Oluwagbemiga Balogun, and Ganesh Kumar
Research and Accomplishments in Applications of Non-negative
Matrix Factorization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1061
Phong Cao Nguyen, Cao Hong Nga, and Huong Hoang Luong
Feature Selection Using Correlation Matrix on Metagenomic
Data with Pearson Enhancing Inflammatory Bowel Disease
Prediction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1073
Huong Hoang Luong, Trong Thanh Tran, Ngoc Van Nguyen,
An Duc Le, Huyen Thi Thanh Nguyen, Khoi Dinh Nguyen,
Nghi Cong Tran, and Hai Thanh Nguyen
Design of PID Controller for Integrating
Processes with Inverse Response

A. Parvathi, M. Rathaiah, R. Kiranmayi, and K. Nagabhushanam

Abstract In chemical process control designing, forms showing an underlying

inverse reaction alters the final state can as often as possible in industries. Procedures
displaying such attributes are called inverse reaction procedures. Moreover, inverse
reaction procedures can incorporate an integrator in their corresponding model. There
are just a couple of studies tending to control of integrating/capacitive procedures
with inverse reaction are in literature. Existing techniques, generally, propose the
utilization of surely PID-controllers for the control of converse reaction forms with
an integrator. However, it has been demonstrated by concentrates in the literature
that PID controllers bring about unsuitable closed loop exhibitions for these forms.
Expansion of inverse reaction makes the control of an integrating procedure progres-
sively hard to control. Then, a filter is used as shown in proposed structure, performs
in exceptionally acceptable closed loop reactions for a class of these procedures. This
paper describes the utilization of PID controllers with filters to improve the response
of integrating/capacitive forms with reverse reaction. Simulation models are given
to represent the prevalence of the proposed technique compared to existing methods.

Keywords Integrating process · Inverse reaction · Servo operation · PID filter ·

Set point filter · Stability

1 Introduction

The standard PID (Proportional-Integral-Derivative) controller is having three terms

in its implementation. This control is well known in process automation and chemical
industries due to its easier implementation and improved performances. Its tuning
parameters can be discovered by utilizing one of the hypothetical methodologies
accepting that the procedure model and by utilizing an ordinary procedure model
and one of a few tuning rules, which can be found in the writing. In general, an error

A. Parvathi (B) · M. Rathaiah · R. Kiranmayi · K. Nagabhushanam

Department of EEE, JNTUA College of Engineering (Autonomous), Ananthapuramu, Andhra
Pradesh, India

© The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 1
R. Ibrahim et al. (eds.), International Conference on Artificial Intelligence
for Smart Community, Lecture Notes in Electrical Engineering 758,
https://doi.org/10.1007/978-981-16-2183-3_1
2 A. Parvathi et al.

signal which is the difference between the desired and actual output of the processes
is an input for the PID controller.
It offers numerous options to the researchers for tuning. The earlier tuning princi-
ples presented by Ziegler and Nichols [1]. They described many tuning procedures
for different types of systems. This tuning approach becomes complicated when the
process having the integrating terms and zeros on the right half of the s-plane. The
intricacy becomes more significant when the time delay (dead time) is related to
the procedure. If the process has large time delay in their model then this type of
design procedure may not give the accurate desired performance of the processes
considered. Therefore a new developments or modification in the basic structure of
PID controller/tuning plan is required.
Procedures with inverse reaction attributes are regularly experienced in industries,
for example, the degree of the drum kettle in a refining segment, the left temperature
of a cylindrical exothermic reactor. Integrating procedures with inverse reaction can
likewise be knowledgeable about mechanical applications, for example, level control
of a kettle steam drum. The model of this type of process includes the zero of transfer
function on the RHS.
It is noted that PID-type controllers bring about unacceptable performance exhi-
bitions in controlling the processes like integrating type. Numerous researchers have
projected different approaches for tuning this type of process models as: Jin and Liu
[2] proposed an upgraded inside the model controller and determined scientific PID
tuning rules for incorporating forms. Anil and Sree [3] planned a PID controller for
different types of industrial processes with time deferring based on directsynthesis
approach. Ajmeri and Ali [4] utilized a direct way to combine and deal with tuning
PID controllers for complete integrating, and twofold request integrating procedures
with huge time delays.
Then again, there are very few works tending to the control of integrating proce-
dures with reverse/inverse reaction and dead time. Luyben [5] proposed a PI and
additionally a PID controller in the traditional single-input single-yield (SISO) frame-
work to control incorporating forms with converse reaction. Pai et al. [6] have used
PID for inverse operating integrating process models.
In recent years, Kaya [7] has projected a PI–PD form of controller to improve
the performance of such (integrating with inverse operating) processes using Smith
indicator plan.
The target of this work is to structure an effective control methodology for the
stable-integrating processes of first order with delay in operating time. The goal is
cultivated by utilizing a PID controller along filter is designed. The controller param-
eters are inferred as far as procedure parameters with the assistance of polynomial
strategy.
The remaining portion of paper is structured as follows: The projected control
scheme is explained in part II, control system design in part III, and simulation
studies in part IV. Finally conclusion is presented.
Design of PID Controller for Integrating Processes … 3

2 Structure of Proposed Control System

The projected control loop structure for (capacity type) integrating processes is
given in Fig. 1. This control loop consist of controller—Gc (s), disturbance signal—
d, process with inverse response—Gp (s) and set point filter—F(s).In this structure
r—denotes the reference and Y —denotes the output response.
In this study the process considered is (capacity type) integrating type which
has one pole (type-1) at origin of s-plane. The form of integrating (capacity type)
processes with inverse response is given as:

K p (−αs + 1) −θs
G p (s) = e α > 0; τ, θ > 0 (1)
s(τ s + 1)

where, Kp —is the steady-state gain, τ—is time constant, and θ—is the time delay.
The servo and regulatory responses are derived as presented in Eqs. (2) and (3)
respectively.

Y (s) F(s)G c (s)G p (s)

= (2)
r (s) 1 + G c (s)G p (s)
Y (s) G p (s)
= (3)
d(s) 1 + G c (s)G p (s)

For deriving the controller parameters, the expression of process is rearranged as

shown in Eq. (4).

k(−αs + 1) −θs
G p (s) = e (4)
s(τ s + 1)
k(αs + 1)(−αs + 1) −θs
G p (s) = e
s(αs + 1)(τ s + 1)

The controller used in projected approach is of the form of Eq. (5).

Fig. 1 Structure of proposed closed loop PI control system

4 A. Parvathi et al.
 
ki
G c (s) = k p + + kd s (5)
s

These sections explains the steps to intend the overall controller which includes
set point filter F(s), tuning rules, selecting λ value and performance measurement.

2.1 Set Point Filter

It is clear from the controller presents zeroes in the servo reaction. The zeroes are the
main reason for undesired over shoot in the servo reaction prompting motions and
the settling time is very high. The presented strategy utilized a set point channel as
appeared in Eq. (6) so as to defeat the impact of zeroes presented by the controller
in servo reaction [8].

1
F(s) =   (6)
kd τi s 2 + k p τi s + 1

2.2 Tuning of PID Controller

The form of controller is,

  
ki α2 s 2 + α1 s + 1
G c (s) = k p + + kd s (7)
s β2 s 2 + β1 s + 1

To rearrange the inference, the controller parameters α1 α2 and β1 , β2 are accepted

as appeared below.
 
θ + 2s
α1 =
2
 
θ + 2s 2
α2 =
4

8λτ + 4λτ − 6θ τ + 20τ s + 2θ s − θ 2 + 8s 2

β1 =
32τ − 8λ + 6θ + 12s

s(−6θ τ + 4λτ + 8λs + 8λτ − 12τ s − 4θ s − θ 2 − 4s 2 )

β2 =
32τ − 8λ + 6θ + 12s
Design of PID Controller for Integrating Processes … 5

2.3 Selection of Tuning Parameter (Λ)

The suggested scope of MS is 1.2–2, which is a trade off between the speed of reaction
and powerful stability. Be that as it may, numerous multiple times, researchers think
about qualities past the most extreme point on account of precarious and incorporating
forms in order to accomplish desired speed of reaction. For the projected technique,
the loop transfer function (L) is,

L = G c (s)G p (s)

 −1
Ms = max 1 + G p (s)G c (s) (8)

2.4 Performance Measurement

The performance indexes considered in the current study are integral of absolute
error (IAE), integral of the square error (ISE). Lessening the IAE and ISE is the goal
of present paper.
These are expressed as:

∞
IAE = (|E(t)|)dt (9)
0

∞
 2 
ISE = E (t) dt (10)
0

The total variation (TV) of the input u(t),

∞

TV = |u i+1 − u i | (11)
i=1

3 Simulation Results and Discussions

Example 1 The model is given to describe the controller technique.

0.547(−0.418s + 1)
G p (s) = (12)
s(1.06s + 1)
6 A. Parvathi et al.

The PID filter is computed as given in Eq. (7)

0.0548s 2 + 0.468s + 1
f P I D (s) = (13)
0.0433s 2 + 0.5215s + 1

The set point filter of the process is calculated as given in Eq. (6).

1
F(s) = (14)
2.9216s 2 + 3.815s + 1

This procedure TF (transfer function) has a place with a mechanical contex-

tual investigation of heater steam drum and the level is constrained by controlling
evaporator. The PI-PD controller parameter utilized by Kaya [7] is used for compar-
ison purpose. Controller parameters for the current strategy recommended are Kc =
2.3174, ki = 0.6075 and kd = 1.7748. In this it is used that λ = 1.500.
Figure 2 shows reactions for all plan techniques to both a change in reference and
a stage upsets with size of −0.5, at time instant t = 25 s. The corresponding control
signal has been shown in Fig. 3.
Execution particulars for all structure techniques are condensed in Table 1. In all
design methods, for both nominal and distributed system the method that present in
this work has low values of ISE, IAE and TV.

Example 2 The model is given to describe the controller technique.

Fig. 2 Closed loop response of example 1

Design of PID Controller for Integrating Processes … 7

Fig. 3 Control signal of example 1

Table 1 Performance indices

Control scheme IAE ISE TV
of example 1
Proposed method 5.52e−05 5.45e−08 0.003
PID based on standard forms 1.7319 1.1135 0.0055

0.5(−0.5s + 1)
G p (s) = e−0.7s (15)
s(0.5s + 1)(0.4s + 1)(0.1s + 1)

In the design, the model is used as:

0.5183(−0.4699s + 1) −0.81s
G p (s) = e (16)
s(1.1609s + 1)

The PID filter is computed as given in Eq. (7)

0.1914s 2 + 0.875s + 1
f P I D (s) = (17)
0.0974s 2 + 0.677s + 1

The set point filter of the process is calculated as given in Eq. (6).

1
F(s) = (18)
6.1703s 2 + 6.475s + 1

This procedure move capacity has a place with a mechanical contextual investi-
gation of heater steam drum and the level is constrained by controlling evaporator.
8 A. Parvathi et al.

Fig. 4 Closed loop response of example 2

The PI-PD controller parameter utilized by Kaya [7] is used for comparison
purpose. Controller parameters for the current strategy recommended are Kc =
1.1292, ki = 0.1744 and kd = 1.0761. In this it is used that λ = 1.500.
Figure 4 shows reactions for all plan techniques to both a change in reference and
a stage upsets with size of −0.5 at time instant(t) = 25 s. The corresponding control
signal has been shown in Fig. 5.
Execution particulars for all structure techniques are condensed in Table 2. In all
methods, for both nominal and distributed system the method that present in this
work has low values of ISE, IAE and TV.

Example 3 The model is given to describe the controller technique.

(−s + 1) −7s
G p (s) = e (19)
s(10s + 1)

The PID filter is computed as given in Eq. (7)

0.218s 2 + 0.112s + 1
f P I D (s) = (20)
0.084s 2 + 0.571s + 1

The set point filter of the process is calculated as given in Eq. (6).

1
F(s) = (21)
1.4285s 2 + 3.2075s + 1
Design of PID Controller for Integrating Processes … 9

Fig. 5 Control signal of example 2

Table 2 Performance indices

Control scheme IAE ISE TV
of example 2
Proposed method 1.918 0.2358 0.0028
PID based on standard forms 2.5006 1.6787 0.0032

The PI-PD controller parameter utilized by Kaya [7] is used for comparison
purpose. Controller parameters for the current strategy recommended are Kc = 0.119,
ki = 0.0371 and kd = 0.053. In this it is used that λ = 1.500.
Figure 6 shows reactions for all plan techniques to both one unit reference change
and a stage upsets with size of −0.5, at time instant t = 250 s. Execution particulars
for all structure techniques are condensed in Table 3. In all design methods, for both
nominal and distributed system the method that present in this work has low values
of ISE, IAE and TV (Fig. 7).

4 Conclusions

This paper portrayed the controller plan for integrating procedure with backwards
reaction and pole at RHS. The PID controller has been planned in an adjusted structure
with arrangement of forward path. The set point channel/filter additionally used to
improve the reaction of process because of step changes in reference signal. The
proposed technique is connected to the two distinctive relative investigations and their
reactions have been appeared in Figs. 2 and 3. The reaction obviously clears that the
10 A. Parvathi et al.

Fig. 6 Closed loop response of example 3

Table 3 Performance indices

Control scheme IAE ISE TV
of example 3
Proposed method 1.545 0.358 0.127
PID based on standard forms 2.899 1.223 0.872

Fig. 7 Control signal of example 3

Design of PID Controller for Integrating Processes … 11

proposed strategy has improved reaction than the strategies utilized for comparison.
The proposed design plan is additionally simpler than other related methodologies
for integrating/capacitive type processes.

References

1. Ziegler JG, Nichols NB (1942) Optimum settings for automatic controllers. Am Soc Mech Eng
64:759–768
2. Jin QB, Liu Q (2014) Analytical IMC-PID design in term of performance/robustness tradeoff
for integrating processes: from 2-Dof to1-Dof. J Process Control 24:22–32
3. Ch, Anil, Sree RP (2015) Tuning of PID controllers for integrating processes using direct
synthesis method. ISA Trans 57:211–219
4. Ajmeri M, Ali A (2015) Simple tuning rules for integrating processes with large time delay.
Asian J Control 17:2033–2040
5. Uma S, Chidambaram M, Rao AS (2010) Set point weighted modified Smith predictor with
filter controllers for non-minimum-phase (NMP) integrating processes. Chem Eng Res Des
88:592–601
6. Pai NS, Chang SC, Huang CT (2010) Tuning PI/PID controllers for integrating processes with
dead-time and inverse response by simple calculations. J Process Control 20:726–733
7. Kaya I (2018) Controller design for integrating processes with inverse response and dead time
based on standard forms. Electr Eng 100(3):2011–2022
8. Kumar, Praveen M, Manimozhi M (2018) A new control scheme for integrating processes with
inverse response and time delay. Chem Product Process Modeling 13(4)
Environmental Feasibility Survey
of Solar Photovoltaic Cells
Sai Goutm Golive, B. Vijaya Krishna, B. Parama Sivam,
and K. Ramash Kumar

Abstract Sustainable power source advancements presents an outflow free of

methods for the energy collecting in the current power energy situation requesting
practical necessities. Be that as it may, the creation cycle of RETs particularly
sunlight based photovoltaic (SPV) is material what’s more, energy costly cycle.
In this paper such data and the energy concentrated creation cycles of Solar photo-
voltaics frameworks are examined. The frameworks is broke down up to a creation
stage with center given to the energy and material separation between various phases
of creation. The assessment of an outflows through creation step is completed for a
1 kWp Solar photovoltaics framework. A correlation is made through various kinds
of advancements i.e., mono Si, poly Si, slim film CdTe and strip Silicon. Likewise,
correlation of maintainability parts of housetop and ground layed establishments
was completed. In strip Silicon, energy process represented higher discharges as
opposed to material in instance of different innovations. In the layed construct, in
rooftop construct brought about least discharges among material investment funds
while, layed mounting constructs are ecologically costly because of high material
utilization. Additionally, the emanation portion of materials utilized underway stage
is decreased directly from mono crystalline Si to strip Si when in rooftop attaching
is utilized. Consequently, slight film CdTe, lace Si consists a naturally less expensive
alternatives of outfitting sun based energy and creation cycles of advances like mono
crystalline Si, poly crystalline Si should be enhanced to make them ecologically
reasonable.

S. G. Golive (B) · B. Vijaya Krishna

Department of Electrical and Electronics Engineering, Bapatla Engineering College, Bapatla,
Andhra Pradesh, India
e-mail: [email protected]
B. Parama Sivam
Department of Electrical and Electronics Engineering, Annamalai University, Chennai, India
K. Ramash Kumar
Department of Electrical and Electronics Engineering, Dr. N.G.P. Institute of Technology,
Coimbatore 641048, India

© The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 13
R. Ibrahim et al. (eds.), International Conference on Artificial Intelligence
for Smart Community, Lecture Notes in Electrical Engineering 758,
https://doi.org/10.1007/978-981-16-2183-3_2
14 S. G. Golive et al.

Keywords Solar Photovoltaic Cells · Life cycle inventory (LCI) · Life cycle
assessment (LCA) · ECO-indicator · Renewable energy Systems (RES)

1 Introduction

An Earth-wide temperature boost and environmental evolve have shifted worldwide

energy index to move from outflow escalated fossil powered assets to sustainable
power sources [1]. Flourishing with the maintainability and green advancement to
moderate atmosphere change issues non-customary and sustainable power source
innovations are given a help via financial approaches around the globe [2]. All through
the gross RESs are encouraged to chop downwards discharges through liberated and
changed approaches, for example, power buy agreements, premium feed in duties
because of their emanation easy nature of energy reaping [3]. In any case, to choose
maintainability parts of any RES life cycle emanations need to be review as opposed
to counting emanations that are stayed away from in operated stage.
Life cycle investigation presents a valuable device in breaking down emanations
over the whole life cycle by means of utilizing products and power sources i.e.,
life cycle stock in various periods of creation, utilize and removal of any item
[4]. The assembly phase of photovoltaics boards incorporate three arranges to be
specific wafer creation, cell manufacture and board fabricating. The SPV board is
the most costly piece of absolute PV framework monetarily as naturally examine
high measure of energy and input sources of info. Numerous examinations have
assessed the ecological effects of the PhotoVoltaics boards in the entire existence
with outflows per KWh source created as a mark [5, 6]. Notwithstanding, the sepa-
ration of emanations among various creation stages is the way to distinguish the
territory in which betterment is needed to extend PV more monetarily and natu-
rally manageable innovation. The existence cycle investigation apparatus is related
with four phases (Fig. 1) [4]. 1. Objective and degree definition consists illustrating
the target, framework limits concerning LCI and measures, characterizing the prac-
tical section like natural or financial grade 2. Life cycle stock examination counting
information assortment, isolation and standardization and so on. 3. The existence
cycle examination stage incorporate assessment of either financial or natural parts
of the item and 4. Translation of results to upgrade manageability lists of the item
in various periods of life assessment. Over last decades numerous investigations
helped out life assessment of SPV along records similar to energy restitution time
frame and outflows per KWh energy bridled. The main mono Silicon boards brought
about a normal 2–4 years of EPBT for various degrees of illumination and outflow
level of 160 to 300 g—CO2 eq./KWh were accounted for [5–8] The variety upon the
mark determined is because of various regardless of whether conditions and vari-
eties underway cycles of SPV. Too exemplified energy differs as per sort of mounting
utilized. Generally the housetop mount and open ground surrounding frameworks
had higher cycle and exemplified energies contrasted with veneer surrounded frame-
works [9]. Be that as it may, the veneer surrounded constructs may bring about higher
Environmental Feasibility Survey of Solar Photovoltaic Cells 15

Fig. 1 Life cycle

assessment framework [4]

compensation periods because of restricted sun based radiation [10]. Counting the
current examinations on life assessment investigation of PV, this examination thinks
about the cycle energy and typified energy of ascend constructs in assessing SPV
framework maintainability angles. Dissimilar to the vast majority of the investiga-
tions an incorporated way to deal with derive stage savvy discharges at all steps and
parts of SPV framework is concocted with this examination. The extent of its investi-
gation was limited to creation phase of PV frameworks barring the capacity compo-
nents similar to batteries and power modules. To measure outflows underway period
of SPV ECO-indicator instrument was utilized and useful unit is huge amounts of
CO2 per 1 kWp SPV framework creation. Rest of the paper is sorted out as includes:
Methodology received to assess life cycle emanations is portrayed in segment II.
Segment III diagrams the portrayal of 1 kWp framework for various kinds of PV
boards, body of construct and ascend constructs. The segment IV the outcomes are
introduced and examined as for generally speaking framework just as every segment.

2 Methodology Adopted

The supportability parts of Photo framework are assessed utilizing ECO-indicator

generally known as Eco Indicator 99 eco it configuration instrument [11]. The ECO-it
is a straight forward programming to adaptable information base with utilize. There
are two pointers together effection climate specifically ReCiPe (RCP) strategy and
16 S. G. Golive et al.

Fig. 2 Analysis stages of ECO-it eco design tool [11]

CO2 counterparts. The ReCiPe (RCP) esteems are unit results which mean the seri-
ousness to outflows with choosen it emor movement. Large data, high the rapid
articulated on effect on climate. The ECO-it device comprises off our areas/phases
of assessment. The initial stage gives data about the framework and any definitions
with respect to the frame work, practical element can be characterized to assist an
ce the professional in evaluation. The subsequent stage was the creation period of
the item/administration whose step and cycles for the frame work can be compre-
hended individually to assess coming about emanations. The ECO-it was encouraged
to incorporate various components like mass, zone and energy indifferent decimal
standards. Any product, energy and cycle in flux into the frame work or item can
be determined via five classifications of LCI gave i.e., product, energy, relocate,
handling and administration/framework (Fig. 2). When the information from LCI is
began coming about emanations can be seen as far as CO2 counter parts or ECO-it
scores. The score of each cycle or part can be utilized to assess the effect of specific
segment/measure in the entire life (Fig. 3).

3 Model Representation

In this complete scenario, life rotation discharges for a 1 kWp Solar framework is
dissected as four kinds of boards, three sorts of ascend cohesive and regularly utilized
BOS. The virtual and specialized attributes of a wide range of Solar boards utilized
Environmental Feasibility Survey of Solar Photovoltaic Cells 17

Fig. 3 LCI inventory inclusion in ECO-it eco design tool [11]

Table 1 SPV panel characteristics

Cell type Manufacturer Cells per Module Weight Modules Module
Module Rating module dimensions (kg) per 1 kWp area per
(Wp) (mm × mm) 1 kWp
(m2 )
1 M ono Si SunWorld [12] 60 1675 × 951 46.7 7 6.7
250
2 poly Si SunWorld [13] 60 1675 × 961 39.7 7 6.7
250
3 CdTe FirstSolar [14] 216 1200 × 600 26.45 11 7.78
92.5
4 Ribbon Evergreensolar 114 1650.5 × 41 5 7.65
[15] 205 951.3

in this examination are specified in Table 1. The average LCI of SPV boards per 1m2
zone of board is embraced [9] in discussion with zone needed for 1 kWp for different
sorts of SPV boards dissected in this investigation. For the ascend framework, three
kinds of ascend courses of action are examined in particular, in-rooftop, on-rooftop
and ground surrounded constructs. As the products and zone necessity differs amid
the diverse ascend surrounded considered, the in general emanations may likewise
shift. The boundaries of the ascend constructs are specified in Table 2. Since the
existence cycle stock of three ascend construction it tends to be seen that, ground
ascend constructs are material serious and in rooftop framework can spare material
contribution surrounded rooftop tile sparing. Be that as it may, the energy repay of in
rooftop frameworks should be lower contrasted with that of on rooftop frameworks
because of skewed environmental and air refrigeration on lower edge of PV boards
[12]. To examine outflows owed to BOS of SPV framework, the framework is larger
than usual by half [11]. Thusly, outflow investigation for 1 kWp will in the long run
outcome in discharges for 1.5 KVA of inverter limit. In the association part, LCI
of Direct wiring is additionally remembered for BOS itself instead of take up as
18 S. G. Golive et al.

Table 2 Mounting constructs LCI [9]

Mounting type On roof In roof Ground
Product maker PhonixSonnenstrom AG¨ Schletter Springerville
Steel (kg) 0 0 4.01
Stainless steel (kg) 0.49 0.28 0
Aluminium (kg) 0.54 1.21 0
Concrete (kg) 0 0 8.03

a different substance take up as a immaterial effect contrasted with different BOS

components, for example, inverter [16].

4 Results and Discussion

The 1 kWp SPV framework is planned from support to entryway regarding emana-
tions during assembling/creation phase of cell, ascending and body system of the
framework. The framework limit for module is found from wafer manufacturing
leaving Si raw material product and energy inputs.

4.1 Emanation Investigation for SPV Modules

The vitality and material concentrated assembling measure of SPV module applies
impressive test in prepare SPV a practical energy reaping innovation. The mono
translucent innovation is product just as energy serious there by bringing about most
noteworthy discharges of 1257 kg of CO2 for every 1 kWp of creation. While, strip
Si with low product and energy utilization fueled by present day creation innovation
brought about best natural alternative with 311 kg of CO2 per 1 kWp of creation
(Fig. 4).

1400
1200
EMISSIONS (Kg of

1000
800
CO2)

600
400
200
0
Mono Si Poly Si Thin film Ribbon cell

Fig. 4 Emission equivalents of different SPV technologies

Environmental Feasibility Survey of Solar Photovoltaic Cells 19

(1) LCA of mono-Si:

By a wide margin, the mono Si innovation is effectively the most earth costly because
of huge product and energy inputs. The wafer creation stage comprises over 61%
of complete discharges and the cell phase represents simple 2.9% of the discharges
and left emanations are from cell bundling into module. All things considered the
discharge because of the PV module creation in mono Si innovation can be classified
into five sorts in particular, gases, solids, fluids, power and direct emanations to air.
While, a portion of these are restricted to specific phases of creation. For instance, the
fluid and gases utilization is practically immaterial in the event of module readiness.
It tends to be seen that, aside from direct discharges what’s more, discharges from
solids utilized in PV board creation, wafer creation stage is answerable for larger part
of the discharges. The wafer phase of PV board has represented 35%, 71%, 94%,
85% of emanations from solids, fluid, gases and power respectively. Producing of
silicon material in wafer production stage is related with over 65% of the emissions
proceeded in wafer phase. Therefore, powerful wafer fabricating/creation innovation
is the need of great importance to reduce outflows through material contributions to
wafer stage. It is likewise saw that, over 60% of the outflows from solids were in
module phage. This can be ascribed to the fact that, module planning related with
board bundling is material escalated with front and back metelled, BSF coating on
back end, aluminum utilized for bundling and so forth. In the cell manufacture stage,
back metelled was the key cycle which caused lion’s share of the outflows in this
stage. In this manner, the emanation investigation of mono Si PV board surmisings
the urge of abridgement in energy just as material contributions to wafer creation
stage (Fig. 5).
(2) LCA of Poly-Si:
Lately, poly Si innovation developed as impressively proficient option for exor-
bitant and product concentrated mono Si innovation. When contrasted with the

Fig. 5 Emissions of mono

Si panel at different
production stages
20 S. G. Golive et al.

creation outflows of mono Si, emanations in wafer creation phase are simple 46.7%
and module bundling had brought about most noteworthy discharges of over 48%
of complete poly Si creation discharges. The principle distinction in the creation
measure is in to wafer phase of creation. As an aggregate decrease of 530 kg of CO2
for each 1 kWp wafer creation is seen which was proportionate to 46.7% contrasted
with 69.2% of mono Si. In this manner, poly Si wafer creation measure is naturally
22.4% suggested over mono Si wafer creation. Additionally, when contrasted with
mono Si the emanations from wafer creation measure are decreased by 7.3%, 58.7%,
7.27% and 22.88% in the event of solids, fluids, gases and power individually (Fig. 6).
(3) LCA of Thin Film CdTe:
Additionally, the slight film modules had came up long way from their presentation
in 1975. The record effectiveness of slight film CdTe modules is accounted for as
21.5% [18]. The primary bit of leeway of the flimsy film modules is the product
adequacy and wide scope of activity in both diffused and direct retention range
[19]. Notwithstanding, the effectiveness is lower contrasted with mono and poly Si
partners, there with heightening geological necessity to tackle same sum of energy
contrasted with previous PV advancements. Creation of 1 kWp of flimsy film PV
module brought about 475 kg-eq of CO2 which is 62%, 34.6% lower contrasted with
mono and poly Si advances separately (Fig. 1). The discharges from creation phase
of Cd-Te innovation are partitioned into three classes to be specific module measure,
plant compound utilization and plant equipment use. The module creation measure
involve 98% of absolute discharges while plant compound utilization and equipment
use represented 1.4%, 0.5% individually (Fig. 7). In the module cycle, material and
energy inputs involve emanation portions of 60, 40% individually.

Fig. 6 Emissions of mono

Si panel at different
production stages
Environmental Feasibility Survey of Solar Photovoltaic Cells 21

500 465

400

300

200

100
6.6 2.8
0
CdTe PV module Plant chemical usage Plant hardware usage

Fig. 7 LCA of thin film technology

300
250
Kg CO2 Eq

200
150
100
50
0
Material Electricity Emission to air

Fig. 8 LCA of ribbon Si wafer technology

(4) LCA of Ribbon Silicon Wafer:

The strip silicon wafer is main of the most recent PV advances empowering tremen-
dous product decrease alongside quick fabricating measure. Nonetheless, the effec-
tiveness of lace silicon wafer is exceptionally low contrasted with slight film too. This
is because of the way that, strip Si wafer have extremely dainty silicon construct with
lower likelihood of photon assimilation there by settling on productivity [15]. For
1 kWp creation of lace Si wafer the outflows are assessed from LCI accessible. The
lace Si wafer creation stage brought about 311 kg of CO2 −eq which was 75%, 57%
and 34.5% lower contrasted with mono, poly and flimsy film innovations individu-
ally. In contrast to mono, poly and tin film innovations, energy source is liable for
greater part of emanations instead of material contributions as if there should arise
an occurrence of another PV advances. Energy input oneself is answerable for over
87% of absolute discharges during wafer creation stage. While product and direct
emanations to air settled at 9% and 3.8% separately (Fig. 8).

4.2 LCA of Mounting Constructs

When the PV producing was finished, establishment measure is the key territory
which devours impressive measure of materials through ascend constructs [9]. The
ordinary LCI of three ascend constructs used in this examination is represented in
22 S. G. Golive et al.

160
140
120
Kg CO2 Eq

100
80
60
40
20
0
On roof Inroof Ground mounted

Fig. 9 LCA of mounting constructs

Table 2. In light of the undertaking type mounting constructs are separated into three
sorts specifically on-rooftop constructs which are sent in average rooftop top PV
frameworks, in rooftop constructs in which boards are coordinated to rooftop top
supplanting the tiles there by offering material reserve funds in building develop-
ment and the third sort is ground ascend constructs which require extra material as
concrete utilized for readiness of establishments. The outflows during establishment
measure are thought to be exclusively from exemplified energy of product utilized
for ascending constructs. The discharges three kinds of discharges are represented in
Fig. 9. Ground layed constructs brought about most elevated emanations of 135 kg
of CO2 –eq. thought about to 61 and 13 kg for on rooftop and in rooftop partners. The
contrast in emanations if there should be an occurrence of in-rooftop construct can
be credited to the way that, 1 m2 of PV boards conceivably can supplant 40 rooftop
tile there by decreasing epitomized energy of in general framework (Fig. 10).
LCA of generally framework To comprehend the portion of various framework
parts in emanations of entire framework consisting casing and BOS, LCA of entire
framework is to be performed. Here, LCA of entire framework is assessed for a
1 kWp framework including ascending and BOS prerequisites. The emanations of
1 kWp framework are investigated with in rooftop framework ascending for all PV
advances as appeared in Fig. 11. The emanation portion of PV module with the
entire framework is diminished from 96.3% for mono Si to 86.8% for lace Si wafer
innovation. While emanations with poly and slight film advances remained at 94%,
90% separately.

Chart Title
150
100
50
Kg CO2 Eq

0
Steel Sell convert Aluminium Concrete Roof Ɵle
-50
-100
-150

Fig. 10 LCA of in-roof mounting constructs

Environmental Feasibility Survey of Solar Photovoltaic Cells 23

5 Conclusion

Energy collecting through sustainable power source innovations is moderately

discharge free contrasted with non-renewable energy source advancements like coal,
flammable gas, atomic frameworks and so on. Notwithstanding, material further-
more, energy costly cycles utilized underway periods of RET advancements like SPV
presents maintainability challenges on their arrangement. In this manner, manage-
ability examination of SPV assistance in distinguishing energy and product costly
stages upon by congratulating fabricating/creation advances to improve the related
cycles so as to decrease natural weight. This work represents LCA of SPV creation
and establishment stages for various SPV innovations and ascend constructs. Key
discoveries of the problem close the strip Si wafer innovation as best natural alter-
native for saddling sun based energy. What’s more, for strip Si innovation energy
utilization is mindful for greater part of discharges though, its material’s exemplified
energy that mutual greater part of outflows if there should be an occurrence of mono,
poly what’s more, dainty film advancements. At long last, it is reasoned that, the
blend of lace Si innovation with in rooftop ascending might lessen by and large SPV
framework emanations by 73% contrasted with mono Silicon innovation.

References

1. The Emissions Gap Report 2012. United Nations Environment Programme (UNEP), Nairobi.
Advancing State Clean Energy Funds Options for Administration and Funding, U.S. Environ-
mental Protection Agency’s Climate Protection Partnerships Division, May 2008
2. Available [online] http://www.epa.gov/greenpower/documents/purchasingguideforweb.pdf
3. Guide to Purchasing Green Power Renewable Electricity, Renewable Energy Certificates, and
On-Site Renewable Generation
4. Environmental management—Life Cycle Assessment—Principles and framework, ANSI/ISO
14040, American National Standard (1997)
5. Kreith F, Norton P, Brown D (1990) A comparison of CO2 emissions from fossil and solar
power plants in the United States. Energy 15:1181–1198
6. Kannan R, Leong KC, Osman R, Ho HK, Tso (2006) CPO Life Cycle assessment study of
solar PV systems: an example of a 27 kWp distributed solar PV system in Singapore. Solar
Energy 80:55–63
7. Nawaz I, Tiwari GN (2006) Embodied energy analysis of photovoltaic (PV) system based on
macro- and micro-level. Energy Policy 34:3144–3152
8. Raghu N, Kshatri SS, Boyina VK (2020) Impact of temperature and irradiance on life time of
Pv inverters. Int J Res 2:283–291
9. Knapp K, Jester T (2001) Empirical investigation of the energy payback time for photovoltaic
modules. Sol Energy 71:165–172
10. Fthenakis V, Kim HC, Frischknecht R, Raugei M, Sinha P, Stucki M (2011) Life cycle inven-
tories and life cycle assessments of photovoltaic systems. In: International energy agency
photovoltaic power systems programme, October 2011
11. Eco-indicator 99 manual for designers, A damage oriented method for life cycle impact assess-
ment, Ministry of Housing, Spatial planning and environment, The Netherlands, October 2000
(Online): www.pre-sustainability.com/download/EI99Manual.pdf
24 S. G. Golive et al.

12. Vijaya Krishna B, Venkata Prasanth B, Sujatha P (2018) Design and implementation of DPFC
for multi-bus Power system. Int J Eng Technol 7(2.8):18–29
13. Sunmodule SW 250 POLY (33mm frame). (Online): http://www.solarworld-usa.com//media/
www/files/datasheets/sunmodule-pro-series/33mm-frame/sunmodule-solar-panel-pro-series-
poly-33mm-frame-ds.pdf
14. Vijaya Krishna B, Venkata Prasanth B, Sujatha P (2018) Enhancement of power quality in
distribution system using hybrid seven level H-bridge inverter based Dpfc. J Electrical Eng
2:61–71
15. Vijaya Krishna B, Venkata Prasanth B, Sujatha P (2018) MA TLAB/Simulink study of of multi-
level inverter topologies using minimized quantity of switches. Int J Eng Technol 7(1.5):209–
216
16. Evergreensolar, ES-A Series Photovoltaic panels. [Online] http://www.evergreensolar.com/upl
oad2/InstallManuals/2011-JANES-AIMWETLSM-0029-USRevE
Intelligent Control Techniques
for Parameter Tuning of PID Controller
for LFC with Emphasis on Genetic
Algorithm

B. Venkata Prasanth, Raja Sathish Kumar, Vijaya Krishna Boyina,

and K. Ramash Kumar

Abstract The theme of this work is carried out to reduce the deviations involved in
LFC of a developed system by via GA technique (intelligent controller) deployed PID
controller. In MATLAB environment, the system is modelled and as well as simu-
lated for conventional controller, Fuzzy, PSO and GA based PID tuning parameters.
Simulation results shows that there is an improved dynamic response of the system
when subjected to different step perturbations with proposed GA-PID controller.

Keywords Load Frequency Control · Integral Time Absolute Error · Fuzzy Logic
Controller · Particle Swarm Optimization · Genetic Algorithm · PID controller

1 Introduction

The goal of LFC during transient conditions is to maintain constant frequency, power
exchanges and tackling system complex model as well as variations involved [1–3].
In general case, the habitual apparent power produced by generators must be met in
accordance with the load power variations. The principal task before a power system
engineers is to provide good quality of power supply generated by a mixture of
renewable and non-renewable sources of energy to the utility customers without any
distortions. Load frequency control to be referred as by maintaining frequency within
permissible limits by regulating both wattage input power generation and demand

B. Venkata Prasanth
QIS College of Engineering & Technology, Ongole, India
R. Sathish Kumar
Keshav Memorial Institute of Technology, Hyderabad, India
V. K. Boyina (B)
Bapatla Engineering College, Bapatla, India
e-mail: [email protected]
K. Ramash Kumar
Department of Electrical and Electronics Engineering, Dr.N.G.P. Institute of Technology,
Coimbatore 641048, India

© The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 25
R. Ibrahim et al. (eds.), International Conference on Artificial Intelligence
for Smart Community, Lecture Notes in Electrical Engineering 758,
https://doi.org/10.1007/978-981-16-2183-3_3
26 B. Venkata Prasanth et al.

[4]. The variables (f and Ptie ) effects during power transient load power condi-
tions. Automatic control is more efficient method of load frequency control because
manual controls are sluggish and involve inherent human time lags. The problem of
ALFC resolves into not only measuring f but also analysing the measured change
from a reference measurement value. Keep view on this correction is initiated by
using control strategy to keep the system original measured value. The risk of the
interconnected system increases because of system design deals with more time
varying parameters. Valid assumptions are to be incorporated in design of a system
controller to enhance overall performance. Tyreus-Luyben, Cohen-Coon, Fertik, Z-
N and Integral Control Methods, are considered for controlling the LFC of isolated
single area system [3, 5, 6]. Integral gain of conventional controller restricts the
dynamic performance of the system. An optimal control scheme based PSO which
enhances the gain of controller for addressing LFC in power systems of single area
or multi area type power systems. The authors [7, 8] proposed the AGC by using
a optimal controller for two area power system as well as compared the results.
Intelligent control technique needs to be deployed to achieve further improvement
in power system dynamics [9–12]. Among them GA has been used to deal complex
optimization problems which uses H2/H∞ Controller [6–8].

2 Model Investigation and Controllers

The power plant model shown in Fig. 1 consists of governor, turbine, a generator and
load with speed regulation as feedback etc., are characterized by transfer functions.
The values of gain and time constants chosen for the model are listed out in Table 1.
The state variables are listed below for the model shown above in Fig. 1.

Fig. 1 Single area thermal power generation model

Table 1 Power system

Name of the component Gain Time constant (s)
parameters
Governor KG = 1 TG = 0.08
Turbine KT = 1 TT = 0.3
Power System/Generator KP = 120 TP = 20
Speed Regulation R = 2.4
Intelligent Control Techniques for Parameter Tuning of PID … 27
   
−1 KP
ẋ1 = x1 + x3 −K P PD
Tg TP

ẋ2 = x6
   
1 −1
ẋ3 = x2 + x3
Tt Tt

ẋ4 = x5
       
−K P KP −1 KP
ẋ5 = x2 + x3 + x5 + x6
TP Tt TP Tt2 Tg TP Tt
       
ki k p R−1 kd −1
ẋ6 = x1 + x4 + x5 + x6
Tg Tg R Tg Tg

In Power system, invariably the nature of the elements is non-linear. At a particular

working point linearization to be done to get a linear model from a nonlinear model.
The actual conditions and practical conditions are differing in general scenario. Clas-
sical algorithms are used to design a load frequency with PID controller in terms of
state space variables for the modelled system [13, 14]. The system dynamics are tabu-
lated for uncontrolled, conventional, Fuzzy rule based, PSO based and GA based PID
Controller in Sect. 3. The intelligent controllers implemented for the model to get
hold of better tuning of gains are listed below.
PID Controller:
Almost in all control applications, PID controllers are extensively used because of its
robustness and system performance. The proto type of PID controller has parameters
namely KP , Ki and Kd which are shown in the Fig. 2. To achieve better system
performance within desirable limits; it is necessary to tune the KP , Ki and Kd terms

Fig. 2 Proto type PID controller

28 B. Venkata Prasanth et al.

in the controller by appropriate optimization technique. Improper tuning leads to

adverse effects on system performance.
Fuzzy Logic Controller:
The general Fuzzy based logic controller consists of fuzzification, rule based control
and defuzzification shown in the Fig. 3. The logic is much closer to the human logics
and is better as compared with classical systems. We cannot achieve satisfactory
solutions from regular PID controllers due to the involvement of time varying vari-
ables while designing the system. The linguistic variables from [3] used for input e(t)
and [ė(t)] for the chosen model as well as for output also under fuzzification and rule
based fuzzy table framed and shown in Table 2 for the model shown in Fig. 1 based
on If and Then rules. Centroid method used for defuzzification process to transform
the output linguistic variable to a real value signal.
PSO Algorithm:
PSO is also one of the soft computing techniques to acquire better tuning values
for parameters involved in PID controller to solve more complex engineering prob-
lems. The procedure for minimization of performance index which is integral time
absolute error (ITAE) i.e., ITAE = ∫ t e(t)dt and frequency deviation is an error func-
tion is obtained by uniformly distributed particles, particle’s position and particles
velocities & weights. Figure 4 shows the flow chart of PSO algorithm.

Fig. 3 Fuzzy logic controller

Table 2 Fuzzy rule based table

ė e
NL NS ZZ PS PL
NL S S M M L
NS S M M L VL
ZZ M M L VL VL
PS M L VL VL VVL
PL L VL VL VVL VVL
Intelligent Control Techniques for Parameter Tuning of PID … 29

Fig. 4 PSO algorithm flow

chart

Genetic Algorithm:
To solve more complex problems GA isused and the process involved represented
by flow chart which is shown in Fig. 5. The procedure for minimization of perfor-
mance index which is integral time absolute error (ITAE) = ∫ t e(t)dt and frequency
deviation is an error function is obtained by reproduction, crossover and mutation.

3 Simulation Results

The dynamic response of the system is studied by applying step change in load
perturbation (PD = 0.01, 0.02, 0.03, 0.04 and 0.05) to the generator. The dynamics
are studied in terms of time domain specifications with GA, PSO and Fuzzy Logic
PID Controllers as well as concluded from the responses shown in Figs. 6, 7, 8,
9 and 10 that GA-PID controller perks up better transient behavior among all the
mentioned controllers (Table 3).
30 B. Venkata Prasanth et al.

Fig. 5 Genetic algorithm flow chart

Fig. 6 Dynamic response of isolated thermal system with a disturbance of PD = 0.01
Intelligent Control Techniques for Parameter Tuning of PID … 31

Fig. 7 Dynamic response of isolated thermal system with a disturbance of PD = 0.02

Fig. 8 Dynamic response of isolated thermal system with a disturbance of PD = 0.03

Fig. 9 Dynamic response of isolated thermal system with a disturbance of PD = 0.04
32 B. Venkata Prasanth et al.

Fig. 10 Dynamic response of isolated thermal system with a disturbance of PD = 0.05

Table 3 Comparison of different controllers

Disturbance with PD = 0.01
Type of Controller Undershoot Static Error Settling Time
Uncontrolled −0.03069 −0.02352 4.255
PID controlled −0.02942 0 23.780
Fuzzy-PID controlled −0.02716 0 6.755
PSO-PID controlled −0.01329 0 3.195
GA-PID controlled −0.01325 0 2.744
Disturbance with PD = 0.02
Type of Controller Undershoot Static Error Settling Time
Uncontrolled −0.06139 −0.04697 4.439
PID controlled −0.05855 0 24.030
Fuzzy-PID controlled −0.05355 0 7.641
PSO-PID controlled −0.02658 0 3.459
GA-PID controlled −0.02638 0 3.226
Disturbance with PD = 0.03
Type of Controller Undershoot Static Error Settling Time
Uncontrolled −0.09076 −0.07044 4.817
PID controlled −0.08821 0 24.860
Fuzzy-PID controlled −0.08035 0 8.032
PSO-PID controlled −0.04003 0 3.765
GA-PID controlled −0.03999 0 3.507
(continued)
Intelligent Control Techniques for Parameter Tuning of PID … 33

Table 3 (continued)
Disturbance with PD = 0.01
Type of Controller Undershoot Static Error Settling Time
Disturbance with PD = 0.04
Type of Controller Undershoot Static Error Settling Time
Uncontrolled −0.12201 −0.09405 5.045
PID controlled −0.11560 0 25.310
Fuzzy-PID controlled −0.10250 0 11.49
PSO-PID controlled −0.05332 0 4.566
GA-PID controlled −0.05228 0 4.255
Disturbance with PD = 0.05
Type of Controller Undershoot Static Error Settling Time
Uncontrolled −0.15340 −0.11770 5.685
PID controlled −0.14590 0 26.451
Fuzzy-PID controlled −0.12780 0 12.760
PSO-PID controlled −0.06599 0 5.349
GA-PID controlled −0.06579 0 5.021

4 Conclusion

A single area Load frequency control problem with PID controller is modelled and
state variables are derived. Conventional controller, Fuzzy based, PSO based and GA
based PID controllers are applied to the generator to get dead beat response from
the system. These controllers are used to minimize the system transient behaviour
pertaining to step load disturbances i.e., PD (p.u.) = 0.01 to 0.05 in steps of 0.01. GA
based PID tuned controller improves the system transient behaviour in contrast with
other controllers implemented. The proposed GA controller is more robust in nature
to minimize the system transient behaviour. There is a scope for further research
to where the robustness of power system can be improved by new soft computing
techniques which deals optimization of complex problems.

References

1. Sathish Kumar R (2018) LFC problem by using improved genetic algorithm tuning PID
controller. Int J Pure Appl Math 120:7899–7908
2. Prasanth BV (2018) Tuning of PID controller by improved GA technique for an Isolated Nuclear
Power System. Int J Pure Appl Math 120:8011–8020
3. Fagna R (2017) Load frequency control of single area thermal power plant using type 1 fuzzy
logic controller. Sci J Circuits Syst Signal Process 6: 50–56
4. Vijaya Krishna B (2018) Enhancement of power quality in distribution system using hybrid
seven level H-bridge inverter based DPFC. J Electr Eng 2:61–71
34 B. Venkata Prasanth et al.

5. Prasanth BV (2015) PSO-PID tuning technique for load frequency control problem. Int J Global
Innov 3:204–209
6. Rajani C (2015) Multi-objective design of load frequency problem using genetic algorithm
based PID controller. Int J Adv Eng Global Technol 3:307–318
7. Kumar RS (2014) Load frequency control of two area power system with fuzzy logic and ANN
control. Int J Adv Eng Global Technol 2:143–147
8. Prasanth BV (2013) New control strategy for improvement of transient behavior in multi area
interconnected power system with emphasis on robust genetic algorithm. Int J Eng ResTechnol
(IJERT) 2:3989–3993
9. Subramanyam B (2009) Comparative study of different control strategies for a load frequency
problem with emphasis on new fuzzy logic controller. IJEEE Spring Edn 1:25–33
10. Vijaya Krishna B (2018) MATLAB/simulink study of multi-level inverter topologies using
minimized quantity of switches. Int J Eng Technol 7:201–216
11. Vijaya Krishna B (2018) Design and implementation of DPFC for multi-bus power system. Int
J Eng Technol 7:18–29
12. Vijaya Krishna B (2016) Designing of multilevel DPFC to improve power quality. In: Inter-
national Conference on Electrical, Electronics and Optimization Techniques (ICEEOT) @
IEEE
13. Wood AJ, Wollenberg BF (1996) Power generation operation and control. Wiley, New Jersey
14. Kennedy J (1995) Particle swarm optimization. In: Proc IEEE Int Conf Neural Netw 4:1942–
1948
Simplified Decoupler Based Fractional
Order PID Controller for Two Variable
Fractional Order Process

A. Divya Teja, R. Kiranmayi, K. Nagabhushanam, and N. Swathi

Abstract In recent years, more research work has been proved the advantages of
using fractional order modeling and control techniques. This paper describes the
design of fractional order proportional integral derivative (FO-PID) controller along
with simplified decoupling method of two variable fractional order processes. The
structure of two variable processes is different from single variable process. The
interaction effects occur for two variable fractional order processes. So to reduce the
interaction effect, the process is decoupled by simplified decoupling method. The
maximum sensitivity based frequency domain strategy is suitable for parameters
tuning of FO-PID controller. The described controller yield better execution for
set point-tracking. Disturbances like white noise are included in the system so as
to exhibit the FO-PID method shows better result in rejection of disturbance. The
parameter uncertainties are added to the process. The course of action is more robust
to such dissimilarities. Simulation results will exhibit the better achievement of the
proposed method.

Keywords Fractional order control · Simplified decoupling · Fractional order

proportional integral derivative time delay · Two-input-two-output process ·
Inverted decoupling · Fractional calculus

1 Introduction

One of the oldest control method proposed in 1920s is PID control method. Mostly it
is utilized in several industrial fields because of good performance, simple design and
small settling time [1, 2]. Many of the PID control methods are modelled as integer
order system. In some cases the systems should be modelled as fractional order,
for the systems the fractional PID controller is employed. Many scientists employ
fractional order controllers for many applications to obtain desired performance of the
systems [3–5]. In 1994 I. Podlubny [6] introduced a PID controller for a fractional

A. Divya Teja (B) · R. Kiranmayi · K. Nagabhushanam · N. Swathi

Department of EEE, JNTUA College of Engineering (Autonomous), Ananthapuramu, Andhra
Pradesh, India

© The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 35
R. Ibrahim et al. (eds.), International Conference on Artificial Intelligence
for Smart Community, Lecture Notes in Electrical Engineering 758,
https://doi.org/10.1007/978-981-16-2183-3_4
36 A. Divya Teja et al.

order processes. He described the PID controller in generalized form introducing

fractional calculus and compare with the classical PID controller to exhibit good
response of the controller. The PID controllers contains three parameters (K p , Ki,
and K d ) to tune but in FOPID controllers contains other two parameters λ and μ. λ
is fractional integration order and that of μ is fractional derivative order. Therefore,
the FOPID contains five parameters (K p , K i , K d , λ and μ) for tuning purpose. This
makes the design of controller more flexible. Many literatures will mainly focus on
tuning method of FO-PID. For the tuning of FOPID literatures [7–9] demonstrates
different tuning methods.
In this paper, a simplified decoupler based FOPID controller is introduced for two
variable processes. For two variable process the interactions between the variables
must exist. So to eliminate the interactions for two control loops, simplified decou-
pling is used. The frequency domain tuning method is used to design the FOPID
controller and it was proposed by Monje et al. [8].
The left over portion of this paper is arranged as follows: In unit 2, introducing
the fractional order control, the fractional order PID controller and the structure of
the fractional order of two variable processes. In unit 3, the structure of simplified
decoupling, the frequency tuning strategy and the design of proposed controller are
given. For simulation study the numerical examples are considered with and without
time delay processes and comparisons of the presented method and FOIMC method
are structured in unit 4. Conclusion and future scope are present in unit 5.

2 Fractional Order Control

2.1 Fractional Calculus

The concept of fractional calculus is obtained from mathematical branch. It describes

the integrals and derivative theory of non-integer order. Fractional calculus has devel-
oped for a long time. Researchers have done their work in several areas of engineering
and technology by using fractional calculus during the last few decades. There are
many other definitions of fractional derivatives and integrals. Riemann and Liouville
had described fractional order integral definition using Taylor series

t
−v 1
a Dt f(t) = (t − τ )v−1 f(τ )dτ (1)
(γ )
a

The fractional calculus operator a Dt−v represents fractional derivative and f(t) is
an function of integer. G(γ) defines gamma function.
Simplified Decoupler Based Fractional Order PID Controller … 37

Fig. 1 The FO-PID standard structure

2.2 Fractional Order PID Controller

Podulbuny has first described the FOPID controller. It was an expansion of conven-
tional PID controller using fractional calculus. The fractional order PID based control
method is presented in Fig. 1. Gp (s) is the process, Gc (s) is the FOPID controller.
By adding three control actions, the expression for the control method is obtained
and it is described in time domain below:

de(t)
u(t) = Kp e(t) + Ki e(t)dt + (2)
dt

The expression of PID controller in s-domain is

u(S) 1
Gc (s) = Kp + Ki + Kp s (3)
e(s) s

2.3 Fractional Order TITO Process

A two variable fractional order linear system is represented by the following fractional
order differential equations:
γ
0 Dt x(t) = Ax(t) + Bu(t), (4)

Y(t) = Cx(t) (5)

Assuming initial conditions are zero and by applying Laplace transform, the
representation of state space equations can be transformed as

s γ Y(s) = (s) U(s) (6)

By substituting (4) and (5) in (6), the following equation can be obtained as:
38 A. Divya Teja et al.
 
Gp11 (s) G p12 (s)
−1
G p (s) = C(I−A) B = (7)
Gp21 (s) G p22 (s)

where each Gi j (s) is a fractional order transfer function and defined as follows:

bij1 sβij1 + bij0

Gi j (s) = , i, j = 1, 2 (8)
aij2 sαij2 + aij1 sαij1 + aij0

where βij and αij are the fractional order, aij and bij are the coefficients. The
structure of fractional order two variable process is introduced in Fig. 1.

3 Simplified Decoupling Based FO-TITO-PID Controller

Design

To design the FO-TITO-PID controller, two steps are required.

Step1: FO-TITO process is decoupled with simplified one.
Step2: FO-PID controller is designed for FO-TITO process.
The interactions are present between two control loops of the TITO process can be
controlled by decoupling method. According to [10] the decoupling methods are ideal
decoupling, simplified decoupling and inverted decoupling. Ideal decoupling is most
complex in design and the inverted decoupling has some limitations to the process i.e.,
the time delay for the diagonal element matrix must be minimum value, the outputs of
the control method exhibit large variations. To overcome the limitations, simplified
decoupling method is chosen. The structure of simplified decoupling represent in
Fig. 2.

Fig. 2 The structure of the simplified decoupling FO-PID method

Simplified Decoupler Based Fractional Order PID Controller … 39

The new decoupler can be obtained as:

G12 (s)
Gl1 (s) = − (9)
G11 (s)
G21 (s)
Gl2 (s) = − (10)
G22 (s)

The output of c1 , c2 and the input u1 , u2 of G(s) have the relationship as:

G12 (s)
u1 = c1 + c2 (11)
G11 (s)
G21 (s)
u2 = c2 + c1 (12)
G22 (s)

The relationship between the inputs u1, u2 of G(s) and the output is:

u1 G11 (s) + u2 G12 (s) = y1 (13)

u2 G22 (s) + u1 G21 (s) = y2 (14)

Substitute (10) into (11), the result will be:

G12 (s) G21 (s)

c1 (G11 (s) − ) = y1 (15)
G22 (s)
G12 (s) G21 (s)
c2 (G22 (s) − ) = y2 (16)
G11 (s)

The generalized loop transfer function is

 
G12 (s) G21 (s)
(G11 (s) − G22 (s)
) 0
G(s) = G12 (s) G21 (s) (17)
0 (G22 (s) − G11 (s)
)

Substituting (9) into (10), the decoupler can be rewrite as

 
b121 sβ121 + b120 (a112 sα112 + a111 sα111 + a110 )
Gl1 (s) = −   (18)
b111 sβ111 + b110 (a122 sα112 + a121 sα121 + a120 )
 
b211 sβ211 + b210 (a222 sα222 + a221 sα221 + a220 )
Gl1 (s) = −   (19)
b221 sβ221 + b220 (a212 sα112 + a211 sα211 + a210 )
40 A. Divya Teja et al.

3.1 Controller Design

A fractional order two variable processes is considered. After applying the simplified
decoupling to the two variable process, the interactions are present between two
control loops has been eliminated. The FOPID control method discussed in Sect. 2
obtained the following expressions:

1
C11 (s) = Kp1 + Ki1 + Kd1 s μ1 (20)
s λ1
1
C22 (s) = Kp2 + Ki2 Kd2 s μ2 (21)
s λ2
In controller design, parameter tuning is of great importance because of changes in
operating conditions. To solve this problem tuning the parameters based on maximum
sensitivity approach, the controller values are obtained. It is defined as

1
Ms = max (22)
ω∈[0,+∞] 1 + C(jω)G(jω)

4 Simulation Study

In this section, the TITO-FOIMC method will compare with the proposed method.
The model of simulink diagram is shown in Fig. 2.

4.1 Example 1

A FO-TITO process without time delay is considered as

 
1.2 0.6
G(s) = 2s0.5 +1 3s0.7. +1 (23)
0.5 1.5
s0.8 +1 3s0.6 +1

The new decoupler can be obtained as

2s0.5 + 1
Gl12 (s) = −   (24)
2 1 + 3s0.7
3s0.6 + 1
Gl21 (s) = −   (25)
3 1 + s0.8
Simplified Decoupler Based Fractional Order PID Controller … 41

The TITO-FOIMC method is compare with the proposed method in this part. A
maximum sensitivity based parameter tuning method is used to design the FOIMC
controller. The maximum sensitivity value of loop1 is equal to 1, the value of Ms2 for
loop 2 is equal to 1. The transfer functions of TITO-IMC controller are given below:

2s0.5
CIMC−1 (s) = (26)
1.2(1 + 0.667s)

3s0.6 + 1
CIMC−2 (s) = (27)
1.5(1 + 0.5843s)

The FO-PID controller is designed based on frequency domain tuning approach.

By tuning the five parameters represented above.
The following FOPID controllers can be obtained as

33.307
C11 (s) = 10.7613+ (28)
s0.8143
30.723
C22 (s) = 10.1953+ (29)
s0.7573
The output behaviour of the FO-PID method is obtained in Fig. 3. At two instants
at t = 0 s and at t = 4 s, a unit step input is applied and also for the two inputs, a
disturbance of the magnitude 0.1 are added at t = 10 s. From Fig. 3 the simplified

2.5

1.5
Output

0.5
y1 of the proposed method
y1 of the TITO FO-IMC method
0 y2 of the proposed method
y2 of the TITO FO-IMC method

-0.5
0 2 4 6 8 10 12 14 16 18 20
Time (s)

Fig. 3 Step response of FOIMC and the proposed method

42 A. Divya Teja et al.

Table 1 Time domain specifications of proposed and FO-IMC method

S. No Control method Rise time Peak time Settling time Overshoot (Mp )
(t r ) (t p ) (t s )
1 FO-IMC Y1 1.4716 0.9985 4.1943 11.0936
method Y2 1.2911 0.002 20 10.5483
2 Proposed Y1 1.0648 0.5544 4.0333 10.1183
method Y2 1.1695 0 9.9982 10.0673

decoupling of the FO-PID controller exhibit less overshoot when compared to FO-
IMC method. The output response of the proposed method reaches the steady state
from the disturbance faster than FO-IMC method.
The time domain specifications are listed in Table 1. The values of rise time,
settling time and peak overshoot have decreased when compare to the FO-IMC
method. The presented method having quicker settling time and shows better results.
The parameter uncertainties (50%) are added to the process to test the robustness
of the system. The following results are obtained in Figs. 4 and 5. In the case of
model mismatch with parameter uncertainties, the presented method is more robust
than FO-IMC method. The output response of the proposed method will not undergo
any change for parameter uncertainties.
Sometimes external noises are included in the system while operation of process.
In order to reject noise in the system and to exhibit the better results of the presented
design, a Gaussian white noise is included to the step input. The result is obtained in
Fig. 6. From the figure, the proposed method performs well in noise rejection.

1.5
Output

y1
y1 with 50% parameter perturbation
0.5
y2
y2 with 50% parameter perturbation

0
0 2 4 6 8 10 12 14 16 18 20
Time (s)

Fig. 4 Step response of the proposed method with 50% parameter perturbation
Simplified Decoupler Based Fractional Order PID Controller … 43

2.5

1.5
Output

0.5 y1
y1 with 50% parameter perturbation
0 y2
y2 with 50% paramter perturbation

-0.5
0 2 4 6 8 10 12 14 16 18 20
Time (s)

Fig. 5 Step response of FOIMC method with 50% parameter perturbation

2.5

1.5
Output

0.5
y1 of the proposed method
y1 of the TITO-FOIMC method
0 y2 of the proposed method
y2 of the TITO-FOiMC method

-0.5
0 2 4 6 8 10 12 14 16 18 20
Time (s)

Fig. 6 Tracking and disturbance rejection performance of the proposed method with white noisy

4.2 Example 2

A FO-TITO process with time delay is considered:

 
1.2e−0.2s 0.6e−0.3s
G(s) = 2s0.5 +1 3s0.7. +1 (30)
0.5e−0.4s 1.5e−0.3s
s0.8 +1 3s0.6 +1
44 A. Divya Teja et al.

The new decoupler can be obtained as

2s0.5 + 1 −0.1s
Gl12 (s) =  e (31)
2 1 + 3s0.7
3s0.6 + 1 −0.1s
Gl21 (s) = −  e (32)
3 1 + s0.8

The TITO-FOIMC method is compare with the proposed method in this part. A
maximum sensitivity based parameter tuning method is used to design the FOIMC
controller. The Maximum sensitivity value of loop1 is equal to 1.2785. The value of
Ms2 for loop 2 is equal to 1.2557.
The transfer functions of TITO-IMC controller are given below:

2s0.5 + 1
CIMC−1 (s) = (33)
1.2(1 + 0.6885s)

3s0.6 + 1
CIMC−2 (s) = (34)
1.5(1 + 0.6521s)

The following FOPID controllers can be obtained as

7.35
C11 (s) = 1.2 + (35)
s0.93
5.12
C22 (s) = 1 + (36)
s0.81
The output behaviour of the FO-PID method is obtained in Fig. 7. At two instants
at t = 0 s and at t = 4 s a unit step input is applied and also for the two inputs,
a disturbance of the maitude 0.1 are added at t = 10 s. From Fig. 7 the simplified
decoupling of the FO-PID controller exhibit less overshoot when compared to FO-
IMC method. The output response of the presented method has attained steady state
from the disturbance faster than FO-IMC method. The time domain specifications
are listed in Table 2. The values of rise time, settling time and peak overshoot have
decreased when compare to the FO-IMC method. The presented method having
quicker settling time and shows better results.
The parameter uncertainties (50%) are added to the process to test the robustness
of the system. The following results are obtained in Figs. 8 and 9. In the case of
model mismatch with parameter uncertainties, the presented method is more robust
than FO-IMC method. The output response of the proposed method will not undergo
any change for parameter uncertainties.
Sometimes external noises are included in the system while operation of process.
In order to reject noise in the system and to exhibit the better results of the presented
Simplified Decoupler Based Fractional Order PID Controller … 45

2.5

1.5
Output

0.5
y1 of the proposed method
y1 of the TITO-FOIMC method
0 y2 of the TITO-FOIMC method
y2 of the proposed method

-0.5
0 2 4 6 8 10 12 14 16 18 20
Time (s)

Fig. 7 Step response of TITO-FOIMC and the proposed method

Table 2 Time domain specifications of proposed and FO-IMC method for time delay process
S. No Control method Rise time Peak time Settling time Overshoot (Mp )
(t r ) (t p ) (t s )
1 FO-IMC Y1 1.1056 6 15.9399 9.6176
method Y2 0.7162 16.5009 15.6956 4.6633
2 Proposed Y1 0.4491 2.2389 15.6025 3.5137
method Y2 0.5921 8.0387 9.9982 0.3095

design, a Gaussian white noise is included to the step input. The result is obtained in
Fig. 10. From the figure, the proposed method performs well in noise rejection.

5 Conclusion

A simplified decoupling FO-PID controller is projected for FO-TITO process.

Decoupling interconnection must introduce between controller and process so as
to avoid the interactions within the two primary loops. The Fractional order PID
is decoupled by simplified decoupler and determine the two controllers separately.
Simulations results were proved that the proposed method does not undergo any
change to parameter uncertainties. Whenever a disturbance is added to two inputs,
the output response of the proposed method reaches steady state faster than FO-
IMC method. The time domain specification of presented design method is much
46 A. Divya Teja et al.

2.5

1.5
Output

0.5 y1
y1 with 50% parameter perturbation

0 y2
y2 with 50% parameter perturbation

-0.5
0 2 4 6 8 10 12 14 16 18 20
Time (s)

Fig. 8 Step response of the proposed method with 50% parameter perturbation with time delay

2.5

1.5
output

0.5
y1
y1 with 50% parameter perturbation
0 y2
y2 with 50% parameter perturbation

-0.5
0 2 4 6 8 10 12 14 16 18 20
Time (s)

Fig. 9 Step response of the TITO-FOIMC method with 50% parameter perturbation with time
delay

better than FO-IMC method. In few words, the FO-PID controller presented in this
paper has produced robust performance for both set point-tracking and rejection in
disturbance. The proposed method will extend to three variable FO-TITO process
for future work.
Simplified Decoupler Based Fractional Order PID Controller … 47

2.5

1.5
Output

0.5
y1 of the proposed method
y1 of the TITO-FOIMC method
0 y2 of the TITO-FOIMC method
y2 of the proposed method

-0.5
0 2 4 6 8 10 12 14 16 18 20
Time (s)

Fig. 10 Tracking and disturbance rejection performance of the proposed method with white noisy

References

1. Xue Y, Zhao H, Yang Q (2008) Self-tuning of PID parameters based on the modified particle
swarm optimization, IEEE
2. Maiti D, Acharya A, Chakraborty M, Konar A (2005) Tuning PID and PIDδ Controllers using
the integral time absolute error criterion, IEEE, Aug 2005
3. Monje CA, Vinagre BM, Feliu V, Chen Y (2008) Tuning and auto-tuning of fractional order
controllers for industry applications. Control Eng Prac 16(7):798–812
4. Chen Y, Bhaskaran T, Xue D (2008) Practical tuning rule development for fractional order
proportional and integral controllers. J Computat Nonlinear Dyn 3(2):021403
5. Zhao C, Xue D, Chen Y (2005) A fractional order PID tuning algorithm for a class of fractional
order plants. In: 2005 IEEE international conference mechatronics and automation, vol 1. IEEE,
pp 216–221
6. Podlubny I (1994) Fractional-order systems and fractional-order controllers. Inst Exp Phys
Slovak Acad Sci, Kosice
7. Polubny (1999) Fractional-order systems and PIλ Dμ controller. IEEE Trans Automatic Control
44:208–214
8. Monje CA, Vinagre BM, Feliu V (2008) Tuning andauto-tuning of fractional order controllers
for industry applications. Control Eng Pract 16(7):798–812
9. Samko SG, Kilbas AA, Marichev OI (1993) Fractional integrals and derivatives. Theory and
applications, Gordon and Breach, Yverdon
10. Li Z, Chen YQ (2014) Ideal, simplified and inverted decoupling of fractional order TITO
processes. In: IFAC Proceedings Volumes 2014, vol 47, no 3, pp 2897–2902
Improved Centralized PID Controller
with Disturbance Rejection for LTI
Multivariable Processes
S. Anitha, R. Kiranmayi, K. Nagabhushanam,
and K. Firdose Kowser Ahamadia

Abstract Decoupling control is a notable strategy for MIMO processes to dispose

of the impacts of undesirable process interactions. It is a compelling instrument for
MIMO process; outer disturbances actually have erase severe impacts on control
execution of the closed loop structures. The extended 2-DOF control calculation for
the MIMO measures in which the greater part of the controllers to decouple the
frameworks first and afterward the controller is intended for the decoupled loops.
This eventually builds the intricacy of the system design. This method requires the
decoupled PID controller i.e., ‘n’ diagonal controllers for ‘n’ variable processes
in addition to two off diagonal lead/lag filters for MIMO processes to achieve the
desired performance. But the decoupler requires the exact mathematical model of the
processes. If there is any errors occur in the modeling of physical processes, it may
not give the satisfactory performance. To overcome these limitations of decoupled
control system, a centralized PID controller is proposed. A centralized control system
requires, ‘n2 ’ PID controllers for ‘n’ variable processes will give better response
irrespective of model errors. Among these, main diagonal controllers improve the
performance and off diagonal reduce the combinability effects of variables in the
MIMO processes. The PID controller will be tuned by frequency response approaches
to reduce the steady state errors like IAE. The proposed method with disturbance
approach in the loop will give improved performance than existing approaches.

Keywords Interaction · MIMO process variables · Decoupler · Centralized

control · Diagonal loop · Off diagonal loop

1 Introduction

In present day design a controller in industry, vast number modules of mechanical

plants were displayed LTI processes. As indicated by the quantity of variables present
in the system, the LTI processes can be separated into SISO processes and Multi input

S. Anitha (B) · R. Kiranmayi · K. Nagabhushanam · K. Firdose Kowser Ahamadia

Department of EEE, JNTUA College of Engineering (Autonomous), Ananthapuramu, Andhra
Pradesh, India

© The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 49
R. Ibrahim et al. (eds.), International Conference on Artificial Intelligence
for Smart Community, Lecture Notes in Electrical Engineering 758,
https://doi.org/10.1007/978-981-16-2183-3_5
50 S. Anitha et al.

and Multi Output (MIMO) processes. Because of the presence of associations and
coupling in MIMO processes, there are a few problems to increase the difficulty of
designing use of SISO structures into design of MIMO structures directly [1].
In multiple loop control approach, the MIMO techniques are treated as combi-
nations of multiple-single controlling loops, and a main controller is organized and
designed on each controlling loop by in view of loop variable associations [2]. Due
to their reasonable sensitivity, and structure straight and direct forwardness, multiple
loop controlling has been a lot of acknowledge by controlling loop control industry
and broad attempt has been determined to get better the introduction of multiple-loop
PI/PID controllers. Different control structure, such as, detuning factor procedures,
chronological loop closing systems, self-governing control methods and decentral-
ized control methodologies, etc., have been proposed consistently. Exactly when
the associations along with different variables are nominal a multiple-loop PID/PI
controller is conventionally enough, in such case, the off diagonal- decentralized
control approach procedure may not succeed to give acceptable reaction if there be
present genuine circle interchanges [3].
Decoupling (Off-diagonal) control is a popular method for MIMO processes to
take out the impacts of process variable collaborations. This at last builds the diffi-
culty in designing a controller. In decoupling (Off-diagonal) PID-like controller have
intended for the MIMO systems directly with straightforwardness procedure as in
case of SISO systems. The methodology describes how to decouple (off-diagonal
elements) and control approach of the systems in a single step. Despite the fact that
decoupling control is a viable approach in favour of MIMO type processes, outer
disturbances actually have consequences for control execution of the structure. To
acquire a superior in the interruption limination, disturbance observers (DOB) be
presented [1]. DOBs have been recognized for their capacities in minimizing the
effect of disturbance.
Accordingly, various logical decentralized control schemes are used to tune and to
make sure system steadiness, which basis design procedure more simple and reduces
costs.
For the MIMO structures with extraordinary loop variable associations, the decou-
pling (off-diagonal) control format shave been regularly utilized [4]. The decoupling
(off diagonal) control as a rule necessitates two phases: (1) arrangement of decoupler
off diagonal elements to restrict the interrelation among control loops, (2) preparation
of the standard loop regulators for as a rule for carrying out control algorithm.
The important purposes of the decoupling (Off-diagonal) control approach are,
(1) it allows the usage of SISO control strategies, (2) if there ought to be an event
of sensor issues, it is decently effortless to adjust loop physically, adjust individual
loop is reasonably exaggerated by the breakdown. Regardless, such an arrangement
may outcome especially stable control composition, especially, when the structure
estimation is high. Thus, the large numbers of investigators are generally focused on
TITO structures.
For more number of control variables methods, Wang et al. [5]projected certain
procedures for fully coupled PID or other than PID regulators with deliberate arrange-
ment strategy and for the most part system structures [6]. Regardless, owing to
Improved Centralized PID Controller with Disturbance … 51

the mathematic assessment of the decouplers (off diagonal) for large number vari-
able system, the vitality of like a control structure can’t be definite. Subsequently,
stationary decoupling come close to which can in a general sense improve control
shows with their robustness like so as to of multiple-loop control structures are
obtained beside the production to control loop relationship for large variable MIMO
structures.
Decoupling systems only may not give the guaranteed stability in a control
strategy. In spite of the way that static off diagonal element is simple but difficult
to design and complete, they may possibly not commonly give sensible robustness
of loop but there subsist significant interactions. However, a combination of decou-
pler with diagonal PID will give a centralized control scheme and will improve the
response of the system.
In this projection, another off diagonal control procedure [7] is along with
frequency response based PID controller is anticipated. The primary objective of
this direct methodology is to get better the trouble dismissal when arrangement has
serious outside disturbance influences, and to simplify the control algorithm.
The organization of the proposal is structured as: Section 2 discusses the illus-
tration of IMC based centralized control system. Section 3 depicts the centralized
control arrangement design using frequency response approach, Section 4 presents
the simulation of case studies and projected algorithm justification, and summary
are depicted in Sect. 5.

2 Representation of IMC Based Centralized Control

System

A graphic description of broad and projected closed loop IMC control arrangement of
multiple control variable strategies are showed up in Figs. 1 and 2 independently. In

Fig. 1 General IMC structure with disturbance ‘d’

52 S. Anitha et al.

Fig. 2 Centralized PID Control system for TITO system

projected approach, the diagonal regulator is used for excusal of loop internal variable
combinability and desirable set value/reference feed is for next the set value. The
parts exist in projected control system are decentralized- main loop controller Gc(s),
Gm(s) is inside model and process Gp(s).
For ease, it is acknowledged that the number of restricted variables in multiple
variable techniques is comparable to numbered controlled components. This allows
mixing of single (main) variable control with a single variable in the course of an
investigation of controller.
In common the basic exhibition of the transfer function of the 2 × 2 (TITO)
process is:
 
Gp11 (s) Gp12 (s)
Gp (s) = (1)
Gp21 (s) Gp22 (s)

Where, the process G pi j (s) is embedded as first order process with delay time
(FOPDT), i.e.,

K
Gpij (s) = e - θs (2)
τs + 1
The RGA of the Eq. (1) can be written as:

RG A = G(0) ⊗ G(0)−T (3)

Improved Centralized PID Controller with Disturbance … 53

The arrangement of centralized (fully coupled) controller is of the outline as:

 
Gc11 (s) Gc12 (s)
Gc (s) = (4)
Gc21 (s) Gc22 (s)

The relation amid input and output can be written as [9]:

Y (s) = G p (s)U (s) (5)

   
y1 (s) u 1 (s)
Y (s) = , U (s) = (6)
y2 (s) u 2 (s)

Here, Y(s), U(s) are specifies an output, input vectors of the process correspond-
ingly. The large variable processing out variable association for the process with two
variables is able to be given as:

Y1 (s) = G p,11 (s)u 1 (s) + G p,12 (s)u 2 (s) (7)

Y2 (s) = G p,21 (s)u 1 (s) + G p,22 (s)u 2 (s) (8)

In the TITO (two variable) structure, while the resulting loop is off (closed), the
commitment from ui to yi has two different ways of sign transmission is feasible.
The grouping of two conduction ways is measured as productive open control loop
components. In case the later analysis controller would be in modified configuration,
with ys2 = 0, by then the as a rule closed loop model everywhere in the scope of y1
and u1 is known by

y1 (s) G p12 (s)G p21 (s)G c2 (s)

= G p11 (s) − (9)
u 1 (s) 1 + G c2 (s)G p22 (s)

These analytical expressions can be produced as:

y1 (s) G p12 (s)G p21 (s)(G c2 (s)G p22 (s))

= G p11 (s) − (10)
u 1 (s) G p22 (1 + G c2 (s)G p22 (s))

Likewise, for additional loops be as:

y2 (s) G p21 (s)G p12 (s)(G c1 (s)G p11 (s))

= G p22 (s) − (11)
u 2 (s) G p11 (1 + G c1 (s)G p11 (s))

In this projected two variable structures, open-loop components amid output

controlled variable (yi ) and controlled variable (ui) not simply depend upon the
looking at move limit model (gii ) yet what’s more depend upon various methods and
regulators in each and every former controlling loop. It deduces the regulation of
54 S. Anitha et al.

single controller is incomprehensible independently and this depends upon various

controllers. An abnormal relations of Eqs. (10) and (11) be able to modified by toler-
ating two notions: First, the perfect controller assume for the former loop (the output
achieves steady and fast state with no transients) was used to smooth out the Eqs. (8
and 9), that is,

y1 (s) G p12 (s)G p21 (s)

Ĝ 11 (s) = = G p11 (s) − (12)
u 1 (s) G p22 (s)

Similarly, the other ETFs are,

y2 (s) G p12 (s)G p21 (s)

Ĝ 22 (s) = = G p22 (s) − (13)
u 2 (s) G p11 (s)

Here, G11 and G22 are the incredible open-loop effective transfer function (EOTF).
These EOTFs are confounded models, and it is difficult to clearly use them for the
controller structure. The decouplers are resolved for the off diagonal positions as:

G p,12 (s)
D12 (s) = − (14)
G p,11 (s)
G p,21 (s)
D21 (s) = − (15)
G p,22 (s)

The diagonal controller and off diagonal decouplers together forms a centralized
control system.

3 Centralized Control System Design

The most of the industrial processes are represented by second order plus delay model
as,

b0,i j
G i j (s) = e−θi j s (16)
a2 i j s2 + a1 i j s + 1

Similarly, ETF is represented as,

G i j (0)
Ĝ i j (s) = e−θ̂i j s (17)
a2 i j s 2 + a1 i j s + 1

The Multiloop decentralized controller would then be able to be easily planned

by solitary loop advances dependent on relating ETFs. At this moment, we utilize
the increase and stage edges approach. It is essential in light of the fact that the
Improved Centralized PID Controller with Disturbance … 55

recurrence reaction technique gives great execution despite indistinctness in together

plant model and aggravations. The PID-type controller of every circle is assumed of
accompanying standard structure:

ki,i
G c,i (s) = k p,i + + kd,i s (18)
s
The controller preserved as,

kd,i s 2 + k p,i s + ki,i

G c,i (s) = (19)
s
Then the open loop system model becomes,

Ĝ i j (0) −θ̂i j s
G c,i (s)Ĝ i j (s) = k e (20)
s
Abbreviating the gain (GM) and phase margin (PM) provisions as Am,i and ψm,i ,
and their intersect frequencies as ωg,i and ω p,i , correspondingly.

arg[G c,i ( jωg,i )Ĝ i j ( jωg,i )] = −π (21)

 
 
Am,i G c,i ( jωg,i )Ĝ i j ( jωg,i ) = 1 (22)

 
gc,i ( jω p,i )ĝii ( jω p,i ) = 1 (23)

ψm,i = π + arg[G c,i ( jω p,i )Ĝ i j ( jω p,i )] (24)

By simplification we obtain,
π ωg,i
ωg,i θ̂i j = Am,i =
2 k Ĝ i j (0)
π
k Ĝ ii (0) = ω p,i ψm,i = − ω p,i θ̂i j
2
Which results,

π 1 π
ψm,i = (1 − ), k=
2 Am,i 2 Am,i θ̂i j Ĝ i j (0)
56 S. Anitha et al.

Table 1 Typical gain (GM) and Phase margins (PM)

   
ψm,i π 4 π 3 3π 8 2π 5
Am,i 2 3 4 5

As a result of this plan, the increase and phase (PM) and gain margins (GM)
are consistent to one another, some conceivable phase and gain margin choices are
provided in Table 1.
The controller considerations are obtained by,
⎡ ⎤ ⎡ ⎤
k p,i a1,i j
⎣ ki,i ⎦ = π ⎣ 1 ⎦ (25)
kd,i 2 Am,i θ̂i j Ĝ i j (0) a
2,i j

The complete design steps of proposed centralized control system are given below.
Step 1: Choose the minimum phase multivariable process and get the mathemat-
ical model of the process Gp (s).
Step 2: Determine the control loops by using relative gains of the variables in the
process as in Eq. (3).
Step 3: Approximate the mathematical model of equivalent processes and
determine the decoupler elements in off diagonal positions as in Eq. (16).
Step 4: Design the PID controllers of main diagonal using the frequency response
tuning method as given in Eq. (25).
Step 5: An exhibition of the projected unified control is assessed regarding indis-
pensable of integral over time of working time, i.e., necessary of IAE (Integral Abso-
lute Error), ISE (Integral Square Error) as given in Eqs. (25) and (26). The lesser the
estimations of these IAE and ISE shows better exhibition of control algorithm.
The expressions for two variable-TITO process are:
∞

IAE = (|E1 (t)| + |E2 (t)|)dt (26)

0
∞

ISE = (E21 (t) + E22 (t))dt (27)

0

4 Simulation Results

Example 1: A Wood–Berry twofold refining area methodology is utilized as the two

variable (TITO) systems that have(been broadly thought about by most of the experts.
This technique can be made out of the trade work structure as given by Eq. (28).
Improved Centralized PID Controller with Disturbance … 57

12.8e−3s −18.9e−3s
G p (s) = 16.7s+1 21s+1
6.6e−7s −19.4e−3s (28)
10.9s+1 14.4s+1

Since it is 2 × 2 process, there two loops should be controlled. The control loops
are determined by RGA method. Based on the outcome of RGA the pairing of control
loops have selected. The RGA of the given process is,
 
2.0094 −1.0094
RG A = (29)
−1.0094 2.0094

From the RGA [8, 9], it is concluded that the diagonal loops are need to have
controller for improving the performance of the process. The effective models of the
corresponding processes are:

ef f y1 (s) 12.8e−3s 18.9 × 6.6(14.4s + 1)e−7s

G 11 (s) = = − (30)
u 1 (s) 16.7s + 1 19.4(21s + 1)(10.9s + 1)

ef f y2 (s) −19.4e−3s 18.9 × 6.6(16.7s + 1)e−9s

G 22 (s) = = − (31)
u 2 (s) 14.4s + 1 12.8(21s + 1)(10.9s + 1)

The parameters for controller are obtained as:

1.3046s+0.0781
0
G c, pr oposed (s) = 10s+1
−0.7422s−0.0515 (32)
0 6s+1

An inverted decoupler of the process is designed and arranged in off diagonal to

reduce the inter combinability among the variable of the process. This is given in
equation below.

(189s+18.9)e−2s
1 (21s+1)
D(s)inver ted = (−39.66s−6.6)e−6s (33)
(10.9s+1)
1

For comparison, a decoupled control system with disturbance observer is used

in this paper. Also the method proposed by Shen et al. [10] has been used. The
simulation results of all these control method have shown. The first output of the
proposed method including the comparison methods is shown in Fig. 3 and second
output is shown in Fig. 4. Also the control signals are depicted in Figs. 5 and 6
correspondingly which compares the smoothness of the controller output.
For an intention of MATLAB simulation, a unit value change to the set or reference
input assistance for the control pair loop 1 at time t = 0 s and for later (second) control
loop 2 at time t = 100 s is incorporated. A step value change in disturbance signal
is also further incorporated to the simulation at t = 250 s. The index of presentation
of two control loops i.e., loop 1 and loop 2 with projected method and comparison
approaches is provided in Table 2. This proves the projected control algorithm is
58 S. Anitha et al.

Fig. 3 Closed loop response of first output (Y1 ) of WB column

Fig. 4 Closed loop response of second output (Y2 ) of WB column

Improved Centralized PID Controller with Disturbance … 59

Fig. 5 Control signal of first loop (Ys1 − U1 ) of WB column

Fig. 6 Control response of secondary loop (Ys2 − U2 ) of WB process

60 S. Anitha et al.

Table 2 Performance indices

Controller Control loop Set Point IAE Disturbance
of WB
method IAE
Proposed Y1 -U1 4.267 10.528
Method Y2 -U2 5.892 46.251
Decoupling Y1 -U1 6.576 14.768
Controller Y2 -U2 8.355 54.737
Shen et al Y1 -U1 12.789 27.137
Method Y2 -U2 13.962 97.860

enhanced the whole performance of the two variable process by attenuating influence
of disturbance. The proposed method shows IAE for set point is 4.267 and 5.892
respectively for two channels. Similarly IAE for disturbance is 10.528 and 46.251
for the two channels respectively. These values show improvement as compared to
existing method used for comparisons.
Example 2: The transfer function of the Shell Heavy Oil process (HOP) by [1] is
known as:
4.05e−27s 1.77e−28s
G p (s) = 27s+1 21s+1
5.39e−18s 5.72e−14s (34)
50s+1 60s+1

Since it is 2 × 2 process, there two loops should be controlled. The control loops
are determined by RGA method. Based on the outcome of RGA the pairing of control
loops have selected. The RGA [9] of the given process is,
 
1.7002 −0.7002
RG A = (35)
−0.7002 1.7002

From the RGA, it is concluded that the diagonal loops are need to have
controller for improving the performance of the process. The effective models of
the corresponding processes are:

ef f y1 (s) 4.05e−27s 1.77 × 5.39(60s + 1)e−32s

G 11 (s) = = − (36)
u 1 (s) 27s + 1 5.72(21s + 1)(50s + 1)

ef f y2 (s) 5.72e−14s 1.77 × 5.39(27s + 1)e−32s

G 22 (s) = = − (37)
u 2 (s) 60s + 1 4.05(21s + 1)(50s + 1)

The parameters for controller are obtained as:

0.6663s+0.2469
0
G c, pr oposed (s) = 19s+1
10.488s+0.1748 (38)
0 26s+1
Improved Centralized PID Controller with Disturbance … 61

An inverted decoupler of the process is designed and arranged in off diagonal to

reduce the inter combinability among the variable of the process. This is given in
equation below.

(−33.63s−1.77)e−2s
1 (60s+1)
D(s)inver ted = (−140.14s−5.39)e−4s (39)
(50s+1)
1

For comparison, a decoupled control system with disturbance observer is used

in this paper. Also the method ID IMC has been used. The simulation results of all
these control method have shown. The first output of the proposed method including
the comparison methods is shown in Fig. 7 and second output is shown in Fig. 8.
Also the control signals are shown in Figs. 9 and 10 respectively which compares
the smoothness of the controller output.
For an intention of MATLAB simulation, a unit value change to the set or reference
input assistance for the control pair loop 1 at time t = 0 s and for later (second) control
loop 2 at time t = 200 s is incorporated. A step value change in disturbance signal is
also further incorporated to the simulation at t = 800 s.
The index of presentation of two control loops i.e., loop 1 and loop 2 with projected
method and comparison approaches is provided in Table 3.
This proves the projected control algorithm is enhanced the whole performance
of the two variable process by attenuating influence of disturbance.

Fig. 7 Closed loop response of first output (Y1 ) of HOP

62 S. Anitha et al.

Fig. 8 Closed loop response of first output (Y2 ) of HOP

Fig. 9 Control signal of first loop (Ys1 − U1 ) of HOP

Improved Centralized PID Controller with Disturbance … 63

Fig. 10 Control signal of second loop (Ys2 − U2 ) of HOP

Table 3 Performance indices

Controller Control loop Set Point IAE Disturbance
of HOP
method IAE
Proposed Y1 − U1 31.159 135.226
Method Y2 − U2 29.357 198.498
Decoupling Y1 − U1 40.741 203.237
Controller Y2 − U2 31.148 215.139
ID IMC Y1 − U1 47.280 267.481
Method Y2 − U2 41.659 443.77

The proposed method shows IAE for set point is 31.159 and 29.357 respectively
for two channels. Similarly IAE for disturbance is 135.226 and 198.498 for the two
channels respectively.

5 Conclusions

The present paper discusses the design of controller for multi variable process. The
proposed method improves the set point response as well as disturbance response of
the processes. The pairing of multivariable systems has been decided by relative gains
of corresponding transfer function models of the given multivariable systems. The
64 S. Anitha et al.

performance of the system has improved by diagonal controllers and combinability

among the different variables of the processes has been reduced by off diagonal
controllers. The proposed method consists of diagonal and off diagonal controllers.
The controllers of main diagonal are designed by using frequency response approach.
Off diagonal are designed by inverted decoupled concept. The simulation results of
proposed method have been compared with two similar approaches namely decoupler
with disturbance observer and Shen et al. methods. The results are shows that the
present approach improves the performance for both set point and disturbance effects.
The proposed method is straight forward approach and it is simple to use in practical
environment.

References

1. Hao Y et al (2020) Decoupling controller with disturbance observer for LTI MIMO systems.
Asian J Control 22(2): 831–840
2. Wood RK, Berry MW (1973) Terminal composition control of a binary distillation column.
Chem Eng Sci 28(9):1707–1717
3. Zhang W (2011) Quantitative process control theory, vol 45. CRC Press, Boca Raton
4. Xiong Q, Cai W-J, He M-J (2005) A practical loop pairing criterion for multivariable processes.
J Process Control 15:741–747
5. Wang Q-G, Huang B, Guo X (2000) Auto-tuning of TITO decoupling controllers from step
tests. ISA Trans 39(4): 407–418
6. Vu TNL, Lee M (2009) Independent design of Multiloop PI/PID controllers for Multi-delay
processes, World Academy Science, Engineering and Technology (60):703–708
7. Tavakoli S, Griffin I, Fleming PJ (2006) Tuning of decentralised PI (PID) controllers for TITO
processes. Control Eng Pract 14(9):1069–1080
8. Seborg DE et al (2010) Process dynamics and control. Wiley
9. Hanuma NR, Ashok Kumar DV, Anjaneyulu KSR (2014) Control configuration selection and
controller design for multivariable processes using normalized gain. World Acad Sci Eng
Technol Int J Electr Comput Electron Commun Eng 8(10)
10. Shen Y, Sun Y, Li S (2012) Adjoint transfer matrix based decoupling control for multivariable
processes. Ind Eng Chem Res 51(50):16419–16426
Tuning of PID Controller Using SIMC
Method for Systems with Time Delay
and RHP Poles
P. Keerthi, R. Kiranmayi, K. Nagabhushanam,
and K. Firdose Kowser Ahamadia

Abstract To control a class of system with RHP poles and time delay, a revised
cascade control model is proposed in this work. This revised cascade control model
aims to enhance both regulatory and servo performance of the system. Structurally,
this proposed revised cascade control scheme has three controllers. To minimize the
set point variations, an auxiliary controller is also used. An analytical method based
on IMC design principle is used in designing the controller for set point tracing.
It is developed using H2 optimization and Internal Model Control theory. Distur-
bance rejecter is developed as P + I + D controller. To make the set point tracking
controller physically realizable, a low pass filter is cascaded with it. Simplified
Internal Modal Control (SIMC) with only one adjustable parameter ‘Tc ’ is used
for tuning the controllers. The simulation studies for proposed approach are done
using MATLAB/SIMULINK to demonstrate its effectiveness. This proposed revised
cascade control scheme depicts 3-Degree of Freedom Control Structure.

Keywords Time delay · Unstable processes · Cascade control · 3-DOF structure ·

P + I + D controller

1 Introduction

Processes with RHP poles are unstable in nature and are very difficult to control
when compared to stable systems. This RHP pole may lead to large setting time
and increased overshoot. In many systems, this unstable pole may also create an
imbalance between the input and the output due to some external disturbance [1].
Therefore the study of the processes with RHP poles and time delay are carried out
by some authors in past. In the initial stage, studies were focused mainly on tuning
methods [2]. Chidambaram along with his co-workers proposed several techniques
for systems having delay and RHP poles. They developed P + I controller using
Ziegler-Nichols tuning formulae [3]. A formula for P and P + I controllers using

P. Keerthi (B) · R. Kiranmayi · K. Nagabhushanam · K. Firdose Kowser Ahamadia

Department of EEE, JNTUA College of Engineering (Autonomous), Ananthapuramu, Andhra
Pradesh, India

© The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 65
R. Ibrahim et al. (eds.), International Conference on Artificial Intelligence
for Smart Community, Lecture Notes in Electrical Engineering 758,
https://doi.org/10.1007/978-981-16-2183-3_6
66 P. Keerthi et al.

IMC principle is explained in [4]. To control a class of linear and non-linear systems,
an Internal Model Controller cascaded with first order filter was suggested [5]. Anvil
[6] developed a P + I + D controller using Internal Model Control theory and
H2 optimization technique for process with RHP poles. In [7, 8] revised form of
SP structure based on 2-DOF control scheme is shown. But in general, one loop
control method offers a satisfied output for systems with delay in few cases. The
cascade control technique has an ability to reject the load disturbances very fast
which enhances the closed loop performance of the system. Generally, there are two
control loops in a cascade structure, one is secondary and the other is primary loop.
Reducing the disturbances in the inner loop itself and not letting them enter into the
primary loop is the main concept behind the cascade control technique [9]. A general
cascade control scheme which was derived from Internal Model Control P + I + D
method was proposed by Lee and Oh [10].
Cascade control structure alone might not give an effective regulatory control for
systems with RHP poles whose control quality should be held at a peek level. Smith
predictor (SP) can compensate the adverse effects of time delay. So combinations of
cascade control scheme and SP have the ability to achieve a better control perfor-
mance [11]. A cascade control scheme is derived from optimal Internal Model Control
technique for control of systems with RHP poles with combination of controllers
and filter was proposed by Dasari et al. [12]. In [13] the disturbance rejecter and set
point tracking were developed using Internal Model Control. The main motive of
the present study is to enhance both disturbance rejection and set point tracking. To
achieve this we are using SIMC tuning method with only one adjustable parameter
‘Tc ’. This SIMC tuning method reduces recovery time and enhances the regulatory
performance of the system.

2 Revised Cascade Control Model

The proposed revised cascade control scheme for systems consisting RHP poles is
displayed in Fig. 1. In Fig. 1, P2m and P1m are the process models of the primary
unstable systems and intermediate stable system, P1 and P2 respectively. P1mo and
P2mo are the perfect transfer function models of P1m and P2m respectively without
any time delay; i.e., P1m =P1mo e−θ1 mS and P2m =P2mo e−θ2 mS
Proposed revised cascade control model has three controllers namely, the set point
tracker C, inner loop load disturbance rejector F1 , and the outer loop load disturbance
rejector F2 . To minimize the set point output variations an auxiliary controller set Pc
is installed. F1 rejects the disturbances that are introduced in the intermediate system
and F2 rejects the disturbances that are introduced in the primary systems. Note that,
the auxiliary controller Pc has no effect on the controllers C, F1 and F2 . This revised
cascade control technique depicts the 3-DOF control scheme. This method offers a
better control in the model parameter error case.
The intermediate system is stable in nature,
Tuning of PID Controller Using SIMC Method … 67

Fig. 1 Revised cascade control model

P1(s) = K 1 e−θ1s /(T1s +1) (1)

The primary system is unstable in nature,

P2(s) = K 2 e−θ2s /(T2s −1) (2)

3 Procedure for Controller Design

3.1 Stabilizing Controller P c

The set point transfer function is derived from Fig. 1,

 
C P1 P2 1 + P1mo P2mo F2 e−(θ1m +θ2m )S (1 + P1m F1 )
Hr = ∗ (3)
1 + PC P1 moP2mo 1 + P1 F1 + P1 P2 F2 + P1 P2 P1m F1 F2

Considering the ideal case, i.e.,P1m and P2m are exact process models without any
time delay. Therefore, set point transfer function is simplified

C P1 P2
Hr =
1 + Pc P1mo P2mo

K 1 K 2 Ce−(θ1 +θ2 )S
= (4)
T1 T2 S2 + (T2 − T1 )S + K 1 K 2 PC − 1

If auxiliary controller (Pc ) wasn’t introduced i.e., Pc = 0 then Eq. 4 is stable.

Therefore, by using RH criterion for stability of its C.E, a realistic form of Pc is
obtained. The selection of Pc can be done in two ways.
68 P. Keerthi et al.

Case 1: T1 < T2 , consider Pc = K c and (K c > 1/K 1 K 2 ) Hence, C.E of set point
transfer function becomes T1 T2 S 2 + (T2 − T1 )S + K 1 K 2 K c = 0, is clearly stable as
per RH Criterion for stability.
Case 2: T1 >= T2 , consider Pc = K c + K d S, (K c > 1/K 1 K 2 ) K d >
(T1 − T2 )/(K 1 K 2 ). Hence, C.E of set point transfer function becomes T1 T2 S 2 +
(K 1 K 2 K d + T2 − T1 )S + K 1 K 2 K c − 1 = 0 which is stable as per RH criterion for
stability.
Note that,Pc can designed as a regular P + I + D controller but it may result in
complex tuning procedure. So, it is not suggestible.

3.2 Set Point Tracker C

If the set point tracker C is developed as stable and rational then there exist zero dead
time element in the denominator of Eq. 4,which
 may yield a better set point response.
We use H2 performance measure min e22 i.e., the integral square error (ISE) in
designing optimal set point tracker C, which should meet the min W (1 − Hr )22
condition.
Here W represents set point input weight function.
Using Pade’s approximation for delay term in Eq. 4 gives

e−(θ1 +θ2 )S = Q nn [−(θ1 + θ2 )S]/Q nn [(θ1 + θ2 )S]

where,


n
(2n − j)!n!
Q nn [(θ1 + θ2 )S] = ∗ [(θ1 + θ2 )S] j
j=0
(2n)! j!(n − j)!

There are 2 cases to perform the controller design procedure:

Case 1: T1 < T2 consider Pc = K c mentioned in above. Therefore, from Eq. 4
  2
1 K 1 K 2 C(S)Q nn [−(θ1 + θ2 )S] 
 
W (1 − Hr )22 =  1−   
S T1 T2 S 2 + (T2 − T1 )S + K 1 K 2 Pc − 1 Q nn [(θ1 + θ2 )S] 
2
 2
 Q [(θ + θ )S] 
 nn 1 2 K 1 K 2 C(S) 
= −  
 S Q nn [−(θ1 + θ2 )S] S T1 T2 S 2 + (T2 − T1 )S + K 1 K 2 Pc − 1 
2
Tuning of PID Controller Using SIMC Method … 69

As Q nn (0) = 1 and all zeroes in Q nn [−(θ1 + θ2 )S] term

are in complex RHP, using orthogonality property in H2 norm
 
 Q nn [(θ1 + θ2 )S] − Q nn [−(θ1 + θ2 )S] 2
W (1 − Hr )2 = 
2  
S Q nn [−(θ1 + θ2 )S] 
2
 
 T T S 2 + (T − T )S + K K K − 1 − K K C(S) 2 .
 1 2 2 1 1 2 C 1 2 
+   
 S T1 T2 S 2 + (T2 − T1 )S + K 1 K 2 K C − 1 
2
In order to design the ideal optimal controller, let its second term be zero.

T1 T2 S 2 + (T2 − T1 )S + K 1 K 2 K c − 1
C1m(s) =
K1 K2

Since, C1m(s) is improper function. To make it physically realizable, low-pass

filter Ic (S) = (T 1+1)2 is cascaded with it. Therefore, a practical optimal controller is
cS
derived as

T1 T2 S 2 + (T2 − T1 )S + K 1 K 2 K C − 1
C(S) = (5)
K 1 K 2 (TC S + 1)2

where Tc —adjustable term.

C attains the optimality only when Tc is adjusted to zero.
Case 2: T1 ≥ T2 follows a similar deriving method gives,

T1 T2 S 2 + (K 1 K 2 K d + T2 − T1 )S + K 1 K 2 K c − 1
C(S) = (6)
K 1 K 2 (TcS + 1)2

Remark 1: Substitute Eq. 5 or 6, (if T1 >= T2 ) in Eq. 4, then practical form of set
point transfer function is derived as,

1
Hr (S) = e−(θ1 +θ2 )S (7)
(Tcs + 1)2

Applying Inverse Laplace Transform,

0 , t ≤ θ1 + θ2
Yr (t) = −(t−θ1 −θ2 )/Tc (8)
1− 1+ t
Tc
e , t > θ1 + θ2

There exist zero overshoot in ideal set point output and by tuning Tc quantitative
time domain specifications are achieved.
From Eq. 8, tuning formula can be figured as tr = 3.8897Tc + θ1 + θ2 where tr
is the rising time.
From Eqs. 7 and 8, it is clear and obvious that Pc does not affect the set point tracker
C. In ideal case, i.e., when Tc is adjusted to zero, the set point response Hr (S) =
70 P. Keerthi et al.

e−(θ1 +θ2 )S i.e., system output reaches the desired value for set point immediately after
the overall systems’ delay time. This single adjustable parameter Tc gives the trade off
among the ideal performance of set point output and the output capacity of set point
tracker C. Therefore, when Tc is adjusted to small value, the setpoint tracing becomes
faster and the output of C is larger resulting in more aggressive dynamic behaviour.
When Tc is adjusted to a higher value, the set point tracing becomes slow, and the
output of C is small resulting in less aggressive dynamic behaviour. Therefore, it is
suggested to adjust ‘T C ’ to a value which approximately equals the system delay
initially. If tuning is unsatisfactory, then decrease/increase Tc to achieve required set
point response.

3.3 Disturbance Rejector F 1

In order to make the design procedure simpler, leave the primary outer loop and
consider the inner loop load disturbance output independently. This results in degra-
dation of actual system performance, due to load disturbance response oscillation
may seep into the inner loop.
From Fig. 1, the ideal load disturbance process models of inner loop are given as

y1 P1 (S)
Hd1 (S) = = (9)
d1 1 + F1 (S)P1 (S)
y1 1
Hd2 (S) = = (10)
d2 1 + F1 (S)P1 (S)

The inner loop complementary sensitivity function is termed as,

f1 F1 (S)P1 (S)
Td−inner = = (11)
d1 1 + F1 (S)P1 (S)

The required ideal complementary sensitivity function is Td−inner = e(−θ1s ) .

It means when a load disturbance d1 is continously introduced in P1 , the load
disturbance rejector F1 must observe the resultant P1 response error immediately
after system time delay θ1 and produce equal and opposite signal f 1 to oppose the
injected load disturbance.
The practical required complementary sensitivity function relied on combined H2
optimal performance measure of Internal Model Control theory is given

1 (−θ1s )
Td−inner (S) = e (12)
T f1 s +1

where, T f1 is as adjustable parameter.

From Eqs. 11 and 12,
Tuning of PID Controller Using SIMC Method … 71

Td−inner (S) 1
F1 (S) = (13)
1 − Td−inner (S) P1 (S)

From Eq. 13,

If T f1 is tuned to be small, the load disturbance rejection in the inner loop becomes
faster, the output of F is larger and vice-versa. Therefore, T f1 aims the trade-off among
load disturbance rejection capability and output capacity of F.

3.4 Disturbance Rejector F 2

From Fig. 1, the load disturbance model of outer loop is given as

y2 1
Hd3 (S) = = (14)
d3 1 + F2 (S)P1 (S)P2 (S)

The outer loop complementary sensitivity function is derived as,

f2 F2 (S)P1 (S)P2 (S)

Td−outer (S) = = (15)
d1 1 + F2 (S)P1 (S)P2 (S)

Similar analysis as in previous subsection is carried out to find required closed-

loop complementary sensitivity function in practical form.

aS + 1 −(θ1 +θ2 )S
Td− outer (S) =  3 e (16)
T f2 S + 1

where T f2 is adjustable parameter and a is obtained by an asymptotic constraint.

lim Hd3 (S) = 0

S→ T1
2

To reject the step load disturbances, the following asymptotic constraint is required

lim (1 − Td− outer (S)) = 0 (17)

S→ T1
2

Substitute Eq. 16 in Eq. 17,


aS + 1 −(θ1 +θ2 )S
lim 1−  3 e =0
S→ T1
2 T f2 S + 1

On solving above equation,

72 P. Keerthi et al.

 3 
T f2 (θ1 +θ2 )/T2
a = T2 +1 e −1 (18)
T2

from Eqs. 15, 16 and 18, the required load disturbance is obtained as,

1 (T1 S + 1)(T2 S − 1)(aS + 1)

F2 (S) =   (19)
K 1 K 2 T f2 + 1 3 − (aS + 1)e−(θ1 +θ2 )S

At S = 1/T2 , Eq. 19 has a zero-pole cancellation on Right Hand Plane, this may
result in unreliable working of F2 .
So, Maclaurin expansion formulae are used to regenerate the ideal disturbance
rejector.
Say F2 (S) = M(S)/S,
 
1 M (0) 2 M i (0) i
F2 (S) = M(0) + M (0)S + S +.+ S (20)
m 2! i!

Considering the first three terms of Eq. 20 resemble a standard P + I + D controller

1
F2−P I D (S) = K F + + TD S (21)
T1 S

where K F = M (0), T1 = 1/M(0), TD = M (0)/2.

First order low pass filter is cascaded with F2 in Eq. 21 to implement it physically.
In [14], third order approximation controller using Pade series expansion is given
as,

d3 S 2 + d2 S + d1 d0
F2−3/3 (S) = +   (22)
C2 S + C1 S + 1
2 S C2 S + C1 S + 1
2

The 1st part is a lead-lag controller of second order, whereas the 2nd part represents
an integrator cascaded with second order low pass filter.

4 Robust Stability Analysis

Considering the formulae developed in above section, system stability is determined.

Practically, the frequently occurred process uncertainties are systems parameters
perturbation, the response uncertainties of actuator and measurement errors of the
sensors.
According to small gain theorem, a system is robust stable only if,

m(S)Td (S)∞ < 1 (23)

Tuning of PID Controller Using SIMC Method … 73

m—Multiplicative system uncertainty bound,

Td —Complementary sensitivity function.
To find the robust stability constraint of inner loop, substitution Eq. 12 in Eq. 23
yield the robust stability constraint.
 
 m 1 (S) 
 
T S + 1 < 1 (24)
f1 ∞

m 1 (S) multiplicative uncertainty bound of P1 .

Say m 1 (S) = K 1 /K 1 , the robust stability constraint of T f1 is
 |K 1 |
T f21 w 2 + 1 > , ∀w > 0 (25)
K1

m 1 (S) = e−(θ1 )S − 1

θ1 Time delay uncertainty of P1 To be robust stability, tune T f1 as below

  
T f21 w 2 + 1 > e−( jQw) − 1, ∀w > 0 (26)

Substitute Eqs. 16 and 18 in Eq. 23 which gives the robust stability constraint.
  
 3 
 T2 T2
+1 e (θ1 +θ2 )/(T2 −1)S
+ 1
 T2  1
  3  < (27)
  m 2 (S)∞
 T f2 S + 1 
 
∞

m 2 —Multiplicative uncertainty bound of P1 ,P2 .

Increasing T f2 strengthens the robust stability but reduces performance of distur-
bance rejection and vice-versa. It is advised to initially adjust T f2 is almost equal to
overall process time delay and later decrease/increase T f2 until the desired response
and robust stability are achieved.

5 Simulation Examples

Three unstable cascade processes are simulated. For numerical comparison, perfor-
mance measure Total Variation (TV), Integral Absolute Error (IAE) and Peak Value
(PV) of the system output are considered.

∞
IAE = |e(t)|dt
0
74 P. Keerthi et al.


N
TV = |u n+1 − u n |
n=0

Example1: Consider an example studied by Cheng-Qiang Yin in [13]. The

e−3S −2S
primary process is P2 = 10S−1 and the intermediate system is P1 = 2eS+1 Method
suggested by Cheng-Qiang Yin [13] is considered for comparison. For the Proposed
method, F1 = S + 1/2S + 2, F2 = 0.8228 + (0.8228/46.522S) + 0.8228 ∗
100S 2 +4S+1
1.4649S, PC = 0.6 and C = 160S 2 +80S+10 .

Considering the above controller values, the performance of the two techniques
are simulated by enabling unit step change at t = 1 s in the set point and applying
negative step change at t = 150 s in disturbance d1 and negative step change at t =
350 s in disturbance d2 respectively. System output and control signals are depicted
in Figs. 2 and 3 for perfect model case.
To analyse the system robustness, assume a 20% raise in intermediate process
gain along with 20% raise in intermediate delay time. The system output is given in
Fig. 4 and control action response is given in Fig. 5. On observing Figs. 4 and 5, the

Fig. 2 System output for Example 1 (perfect)

Fig. 3 Control signals of Example 1 (perfect)

Tuning of PID Controller Using SIMC Method … 75

Fig. 4 System output for Example 1 (error).

Fig. 5 Primary control signals of Example 1 (error)

proposed technique provides gives better closed loop response and the control action
undergoes several oscillations at first which later becomes smoother obviously. The
performance measures for proposed method and other method are tabulated below
for both perfect and perturbed cases. From Table 1 the proposed method has lower
IAE and TV than Cheng-Qiang technique. Therefore, the proposed technique gives
an enhanced performance than the Cheng-Qiang technique.

Table 1 Performance
Method IAE PV TV
measures of Example 1
Perfect model parameter
Cheng-Qiang 47.4591 1.01 39.65
Proposed 25.1821 1.26 31.45
Error model parameter
Cheng-Qiang 55.81 1.52 62.74
Proposed 27.6376 1.0 59.28
76 P. Keerthi et al.

Example 2: Another example studied by Cheng-Qiang Yin in [13] is consid-

e−4S
ered whose primary process is P2 = 20S−1 and the intermediate process is
2e−2S 35.1847S 2 +24.8678S+1.5707
P1 = 20S+1
In proposed technique, F1 = 20S+1 F
m S +2 2
= 2.56S 2 +0.8655S+1
+
400S 2 +2S+1
0.0208
S (2.56S 2 +0.8655S+1)
, PC = 1, and C = 72S 2 +24S+2 .
Considering above controller parameters, the performance of two techniques are
simulated by enabling a unit step change at t = 1 s in the set point and applying a
negative unit step change at t = 150 s in disturbance d1 and negative unit step change
at t = 350 s in disturbance d2 respectively. The system output and control action
response are given in Figs. 6 and 7 for perfect model case. Therefore proposed method
gives a better performance than Cheng-Qiang method for disturbance rejection.
To analyse robust stability of the system, assume a +10% error in intermediate
process gain and delay and +20% error in primary process gain and delay. Its outputs
are shown in Figs. 8 and 9. On observing Figs. 8 and 9, oscillations exists in both
system responses and control actions.
The performance characteristics for two methods in both perfect and perturbed
case are tabulated in Table 2.

Fig. 6 System output for Example 2 (perfect)

Fig. 7 Control signals of Example 2 (perfect)

Tuning of PID Controller Using SIMC Method … 77

Fig. 8 System output for Example 2 (error)

Fig. 9 Control signals of Example 2 (error)

Table 2 Performance
Technique IAE PV TV
characteristics of Example 2
Perfect model parameter
Cheng-Qiang 28.3552 1.05 23.91
Proposed 24.2949 1.06 18.26
Error model parameter
Cheng-Qiang 31.10 1.16 33.53
Proposed 28.3028 1.0 27.83

Example 3 Consider another example studied by Cheng-Qiang in [13] whose

−0.399S
e−0.6S
primary process is P2 = e5S−1 and the intermediate process is P1 = 2.07S+1 . For
the proposed method, F1 = 2.07S+1
0.6S+2
, F2 = 4.4091 + 1
1.9636S
+ 7.1874S, PC = 2, and
10.35S 2 +2.938S+1
C= S 2 +2S+1
.
78 P. Keerthi et al.

Considering above controller parameters, the performance of these methods are

simulated by enabling a unit step change at t = 1 s at the set point and applying a
negative unit step change at t = 30 s in disturbance d1 and negative step change at t
= 60 s in disturbance d2 respectively. The resulting system outputs are displayed in
Fig. 10. The proposed method offers enhanced performance than other method.
To analyse robust stability, suppose a +20% error in intermediate process and
−20% error in process delay and also +20% error in primary process gain along
with +30% error in process delay. The corresponding outputs are displayed in
Fig. 11. Initially proposed method has oscillations, but it results in better disturbance
rejection.
The performance measures for the proposed and other method under perfect and
perturbed cases are tabulated in Table 3.

Fig. 10 Control signals of Example 3 (perfect)

Fig. 11 System output for Example 3 (error)

Tuning of PID Controller Using SIMC Method … 79

Table 3 Performance
Technique IAE PV TV
characteristics of Example 3
Perfect model parameter
Cheng-Qiang 4.1129 1.05 50.9
Proposed 3.4976 1.0 43.73
Error model parameter
Cheng-Qiang 4.14 1.05 76.23
Proposed 3.7929 1.0 64.154

6 Conclusion

This paper work proposes a revised cascade control method for systems with RHP
poles and time delay. The set point tracker C is developed derived from H2 optimiza-
tion and Internal Model Control theory. The disturbance rejector F2 is developed as a
P + I + D controller. A filter is cascaded in series with controller C to make it phys-
ically realizable. SIMC tuning method is used to tune the controllers. SIMC has ‘Tc ’
as variable measure which gives the trade-off among robustness and performance.
MATLAB simulated results depicts that suggested technique gives enhanced outputs
for both rejection of disturbance and also set point tracing. Thus, both regulatory and
servo performances are enhanced. This method also yields lower IAE, PV and TV
values.

References

1. Liu T, Gao FR (2011) Enhanced IMC design of load disturbance rejection for integrating and
unstable processes with slow dynamics. ISA Trans 50(2):239–248
2. Chidambaram M (1997) Control of unstable systems: a review. J Energy Heat Mass Transf
19:49–56
3. Kavdia M, Chidambaram M (1996) On-line controller tuning for unstable systems. Comput
Chem Eng 20(3):301–305
4. Jacob EF, Chidambaram M (1996) Design of controllers for unstable first-order plus time delay
systems. Comput Chem Eng 20(5):579–584
5. Saxena S, Hote YV (2017) Internal model control based P+I+D tuning using first-order filter.
Int J Control Autom Syst 15(1):149–159
6. Vanavil B, Anusha AVNL, Perumalsay M, RaoAS (2014) Enhanced IMC-P+I+D controller
design with lead lag filter for unstable and integrating processes with time delay. Chem Eng
Commun 201(11):1468–1496
7. Zhang WD, Gu DY, Wang W, Xu XM (2004) Quantitative performance design of a revised
Smith predictor for unstable processes with time delay. Ind Eng Chem Res 43(43):56–62
8. Rao AS, Rao VSR, Chidambaram M (2007) Simple analytical design of revised Smith predictor
with enhanced performance for unstable first-order plus time delay (FOPTD) processes. Ind
Eng Chem Res 46(13):4561–4571
9. Vijaya Krishna B, Venkata Prasanth B, Sujatha P (2018) Design and implementation of DPFC
for multi-bus Power system. Int J Eng Technol 7(2.8):18–29
10. Lee YH, Oh SG (2002) Enhanced control with a general cascade control structure. Ind Eng
Chem Res 41(11):2679–2688
80 P. Keerthi et al.

11. Vijaya Krishna B, Venkata Prasanth B, Sujatha P (2018) Enhancement of power quality in
distribution system using hybrid seven Level H-bridge inverter based Dpfc. J Electrical Eng
2:61–71
12. Vijaya Krishna B, Venkata Prasanth B, Sujatha P (2018) MATLAB/Simulink study of multi-
level inverter topologies using minimized quantity of switches. Int J Eng Technol 7(1.5):209–
216
13. Yin C-Q, Wang H-T, Sun Q, Zhao L. Enhanced cascade control system for a class of unstable
processes with time delay
14. Liu T, Zhang W, Gu D (2005) IMC-based control strategy for open-loop unstable cascade
processes
Independent Controller Design
for Non-minimum Phase Two Variable
Process with Time Delay

C. Venkatasuresh, M. Rathaiah, R. Kiranmayi, and K. Nagabhushanam

Abstract Non-minimum phase (NMP) second order with time delay Process
(SOPTDP) are stabilized by tuning process. Certain restrictions on the attainable
performance when system have right hand side zeros of s-plane. The PID controller
had been designed by Enhanced IMC method. The proposed Proportional Integral
and Derivative controller (PID) are implemented to non-minimum phase systems like
Fermenter, Isothermal CSTR. The performance was analysed under model uncer-
tainty and considering perturbations in one model parameter. The performance of
the system under recommend method compare with the existing method in the liter-
ature. From the performance detected as present Approach effectively stabilize and
advances the performance of process with disturbances. From simulated result of
three examples demonstrates present work offers better-quality presentation of the
set-point track and Hload rejection than existing method through better time domain
specifications.

Keywords Enhanced IMC · NMP · Second Order plus Time Delay systems
(SOPTDS) · Fermenter and Isothermal CSTR

1 Introduction

PID controller was a naturally uses controller amongst industries. Valuation of PID
controllers justifies conversation now. Normal PID controllers were exchanged by
electronic controller. These offers various improvements than earlier controller.
Thereafter, electronic controller is substituted with computer programs. presently
PID controller are software programs in place of physical controllers. Dealing
unstable processes were at all times difficulty with comparison to the stable processes.
Regrettably, various chemical processes are unstable processes. Some samples were
bottom level control of distillation column, chemical reactors, bioreactors, etc. The
difficulty controlling becomes higher though the processes have time delay. Time

C. Venkatasuresh (B) · M. Rathaiah · R. Kiranmayi · K. Nagabhushanam

JNTUA Collage of Engineering (Autonomous), Ananthapuramu, Andhra Pradesh, India

© The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 81
R. Ibrahim et al. (eds.), International Conference on Artificial Intelligence
for Smart Community, Lecture Notes in Electrical Engineering 758,
https://doi.org/10.1007/978-981-16-2183-3_7
82 C. Venkatasuresh et al.

delay may be due to process lag. The goal has to develop a simple controller for
stable/unstable SOPTD systems [1–5]. A tuning parameter selection depends on
maximum sensitivity (Ms) function. Set point filter was used to decrease overshoots
in servo problem. “Many Approaches are used to construction of PI/PID controller
for stable NMP Processes are stated in literature as Internal Model Control approach
(IMC) [6, 7], “PM and GM” [8], “optimization approach” [9, 10], control structure
approach [11–13] etc.
Here numerous approaches existing in literature to construct PI/PID controllers
to NMP FOPTD systems. But a very few approaches were offered to NMP SOPTDS
processes.

2 Proposed Method

Therefore, in this study, a SOPDS process model G p (s) given by Eq. 1 is considered
for the construct of PID controller G c (s)

k p (u − ts) −ds
G p (s) = e (1)
ps 2 + qs + r

Closed loop transfer function obtained as

C(s) G p (s)G c (s)

= (2)
R(s) 1 + G p (s)G c (s)

Wanted closed loop gain of the system is as follows

C(s) e−ds
= Z (s) = (3)
R(s) (1 + αs)2

From the closed loop transfer function controller transfer function derived as

1 Z (s)
G c (s) = (4)
G p (s) 1 − Z (s)

By substituting above functions G c (s) is obtained as

ps 2 + qs + r e−ds
G c (s) = (5)
k p (1 − ts) (1 + αs)2 − e−ds

By using Taylor series expansion G c (s) Approximated as follows

Independent Controller Design for Non-minimum … 83
⎡ ⎤
1⎣ ps 2 + qs + r e−ds
G c (s) =    ⎦
s k t (α 2 − d2 2
)s + u(α 2 − d 2 + t(2α + d))s + u(2α + d)
p 2 2
(6)
⎡ ⎤
ps 2 + qs + r e−ds
A(s) = ⎣     ⎦ (7)
d2 2 2 − d 2 + t(2α + d))s + u(2α + d)
k p t (α 2 − 2
)s + u(α 2

Therefore, controller transfer function is written as

A(s)
G c (s) = (8)
s
For the ideal form of PID apply Maclaurin series expansion theorem then

1 A(0) 2
G c (s) = A(0) + A(0) s + s + ... (9)
s 2

The ideal form of PID as given below

1
G c (s) = kc 1 + + τD s (10)
τI s

The controller parameters are assessed by comparing the constants of ‘s’ of. The
following terms for controller parameters are gotten and written as

kc A(0)
kc = A(0) , τ I = , τD = (11)
A(0) 2kc

3 Selection of α

For the controller design the tuning parameter α is required. A suitable tuning param-
eter α is chosen by the use of maximum sensitivity function. The chosen tuning
parameter for attained controller should offer better performance and robustness.
For smaller α value of gives a faster response and displays improved results in the
case of load disturbance for the processes. Ms is defined as

1
Ms = (12)
1 + G p (s)G c (s)
84 C. Venkatasuresh et al.

Ms value uses to determine a range for the Gain Margin (GM) and Phase Margin
(PM) and relation between Ms and GM and Ms and PM is given below.

Ms 1
GM ≥ ; P M ≥ 2 sin−1 (13)
Ms − 1 2Ms

If the Ms value Increases then GM and PM decreases. The smaller values of

GM is 1.6° and 34° have been suggested PM for stable processes. There are certain
boundaries of the controller for controlling process and to attain the wanted robust
closed loop performance. In this work, the α selected to reach Ms value in the range
of 1.1–1.7.

4 Simulation Studies

Three cases of NMP SOPDT Process were taken to the Valuation of present approach.
Simulation approach has been made to achieve the best value of α. For finding
optimum PID values, the original PID values were essentially varied frequently using
α through computer simulation till the lowest MS index and the required closed loop
performance reached. Effectiveness of the controllers stands measure in the form of
errors, total variation (TV) of operated variables and time domain specification.
Case study-1 (fermenter).
Fermenter process exact model has given below. In this the microbial growth was
supposed to following Monod kinetics has specified by the following equations.

dZ
= (μ−K d )Z−Zu (14)
dt
dS μ
= ( − m)Z − (S−St )u (15)
dt Y
μmax S
μ= (16)
km + S

From solving above equations. Resulted as follows

X (S) 0.2803(1 − 3.4063s) −0.3s

= e (17)
U (s) 3.1638s 2 + 5.7382s + 1

In this present approach, PID values gotten are k c = 3.0725, τ i = 5.7382, τ d =

0.5514 through choosing α = 0.1514, β = − 5.475α, γ = 1.925α correspondingly.
The performance has assessed for closed loop processes by unit step change in the
set point as exposed in Fig. 1 also unit step change in load as exposed in Fig. 2.
From figures analysed that the characteristics of a system like rise time (tr), settling
Independent Controller Design for Non-minimum … 85

Fig. 1 Servo performance of process with the controller parameters obtained from the enhanced
IMC, proposed method

Fig. 2 Regulatory performance to fermenter process with controller parameters attained through
enhanced IMC, proposed Approach
86 C. Venkatasuresh et al.

time (ts) were less than the available approach and these are recorded in Table 1.
From both responses of current and existing method observes the current method
accomplishes a smoother undershoot than the existing method. The current method
expresses important development in the controller through reduce of ITAE 11% than
enhanced IMC method. The controller actions of both servo and regulatory problems
specified in Figs. 3 and 4 correspondingly.
Case study 2 (isothermal CSTR).
Van De Vesse isothermal CSTR [13] is taken now.

P1 P2
X−
→ 3Y −
→Z

P3
2X −
→W

The mathematical model of the system is given below

dC X F 
= C X, f − C X − P1 C X − P3 C X2 (18)
dt V
dC X F
= − C B + P1 C X − P2 CY (19)
dt V
From solving above equations resulted function is as below

X (S) −1.117s + 3.1472 −0.1s

= 2 e (20)
U (s) s + 4.6429s + 5.3821

In this present approach, PID values gotten are k c = 1.4588, τ i = 0.8627, τ d =

0.2154 through choosing α = 0.1577, β = − 1.009α, γ = 2.586α correspondingly.
The performance has assessed for closed loop processes by unit step change in the
set point as exposed in Fig. 5 also unit step change in load as exposed in Fig. 6.
From figures analysed that the characteristics of a system like tr, ts were lower than
the available approach and these are recorded in Table 1. From both responses of
current and existing method observes the current method accomplishes a smoother
undershoot than the existing method. The current method expresses important devel-
opment in the controller through reduce of ITAE 25% than enhanced IMC method.
The controller actions of both servo and regulatory problems specified in Figs. 7 and
8 correspondingly.
Case study-3 (fermenter).
The linearized form of case study 1 transfer function is as follows

X (S) 0.0291(1 − 28.7121s ) −0.3s

= e (21)
U (s) 2.2024s 2 + 9.0155s + 1
Table 1 Comparisons of Performance indices of all case studies
Case study Design method Servo problem Regulatory problem TV Rise time (s) Settling time (s)
IAE ITAE IAE ITAE Servo Regulatory
1 Enhanced IMC 6.372 16.306 2.398 22.384 131.23 2.987 0.839 1.63
Proposed 6.042 15.019 2.064 21.31 109.58 2.489 0.802 1.108
Independent Controller Design for Non-minimum …

2 Enhanced IMC 1.068 0.533 0.651 1.103 26.654 1.748 3.58 8.24
Proposed 0.911 0.496 0.626 1.001 23.256 1.018 3.46 6.289
3 Enhanced IMC 36.423 267.67 3.575 57.261 148.81 6.347 11.3 26.7
Proposed 35.7 225.36 2.92 50.215 139 5.983 10.86 25.9
87
88 C. Venkatasuresh et al.

Fig. 3 Control action attained through servo response with PID parameters for enhanced IMC,
proposed Approach

Fig. 4 Control action attained through regulatory response with PID parameters for enhanced IMC,
proposed Approach
Independent Controller Design for Non-minimum … 89

Fig. 5 Servo performance of isothermal process with the controller parameters attained from the
enhanced IMC, proposed approach

Fig. 6 Regulatory performance to isothermal process with controller parameters attained through
enhanced IMC, proposed Approach
90 C. Venkatasuresh et al.

Fig. 7 Control action attained through servo response with PID parameters for enhanced IMC,
proposed Approach

Fig. 8 Control action attained through regulatory response with PID parameters for enhanced IMC,
proposed Approach
Independent Controller Design for Non-minimum … 91

Fig. 9 Servo performance of fermenter process with the controller parameters attained from the
enhanced IMC, proposed approach

For the present method, controller parameters attained are k c = 8.2075, τ i =

9.0155, τ d = 0.2443 by choosingα = − 0.0362, β = 87.78α, γ = 0.0362α corre-
spondingly. The controller parameters for enhanced IMC method are k c = 1.4832, τ i
= 0.8627 and τ d = 0. 1723. The performance has assessed for closed loop processes
by unit step change in the set point as exposed in Fig. 9 also unit step change in load
as exposed in Fig. 10. From figures analysed that the characteristics of a system like
rise time tr, ts were lower than the available approach and these are recorded in Table
1. From both responses of current and existing method observes the current method
accomplishes a smoother undershoot than the existing method. The current method
expresses important development in the controller through reduce of ITAE 15% than
enhanced IMC method. The controller actions of both servo and regulatory problems
specified in Figs. 11 and 12 correspondingly.

5 Conclusion

A tuning process was used to develop the PID for the NMP SOPDT systems. The
proposed method was associated with the method existing method in literature such
as Enhanced IMC. Analysed that from simulation results, proposed method gives
important development in the closed loop performance has detected in three cases
related to the current method in literature. The current method expresses improved
performance indices IAE and the ITAE. smother control action was attained by the
92 C. Venkatasuresh et al.

Fig. 10 Regulatory performance to fermenter process with controller parameters attained through
enhanced IMC, proposed Approach

Fig. 11 Control action attained through servo response with PID parameters for enhanced IMC,
proposed Approach
Independent Controller Design for Non-minimum … 93

Fig. 12 Control action attained through regulatory response with PID parameters for enhanced
IMC, proposed Approach

present Approach. The performance was enhanced in present approach than Existing
approach. This could be exposed that the present PID is able to control NMP SOPTD
systems.

References

1. Sankar Rao C (2019) Enhanced PID controller for non-minimum phase second order plus time
delay system. In: Chemical product and process modeling, 20180059
2. Normey-Rico JE, Camacho EF (2008) Simple robust dead-time compensator for first-order
plus dead-time unstable processes. Ind Eng Chem Res 47:4784–4790
3. Uma S, Rao AS (2016) Enhanced modified Smith predictor for second-order non-minimum
phase unstable processes. Int J Syst Sci 47:966–981
4. Begum KG, Rao AS, Radhakrishnan TK (2017) Enhanced IMC based PID controller design
for non-minimum phase (NMP) integrating processes with time delays. ISA T 68:223–234
5. Hauksdottir A, Sigurosson S (2011) The matching coefficients PID controller. In: American
control conference, San Francisco, CA, USA
6. Kuhlmann A, Bogle ID (2001) Controllability evaluation for non-minimum phase-processes
with multiplicity. AIChE J 47:2627–2632
7. Seaborg DE, Edgar TF, Mellichamp DA (2004) Process dynamics and control. Wiley, USA
8. Skogestad S (2003) Simple analytic rules for model reduction and PID controller tuning. J
Process Contr 13:291–309
9. Simi S, Chidambaram M (2013) A simple method of tuning series cascade controllers for
unstable systems. J Control Theory Appl 11:661
10. Sri RP, Chidambaram M (2003) A simple method of tuning PID controllers for integrator/dead-
time processes. Comput Chem Eng 27:211–215
94 C. Venkatasuresh et al.

11. Vijaya Krishna B (2018) MATLAB/simulink study of multi-level inverter topologies using
minimized quantity of switches. Int J Eng Technol 7:201–216
12. Vijaya Krishna B (2018) Design and implementation of DPFC for multi-bus power system. Int
J Eng Technol 7:18–29
13. Vijaya Krishna B (2018) Enhancement of power quality in distribution system using hybrid
seven level H-bridge inverter based DPFC. J Electr Eng 2:61–71
Control of DC Link Voltage and Load
Voltage Variations in a Pitch Angle
Controlled PMSG Based Wind Energy
Conversion System

A. Nagarajan, K. Ramash Kumar, and G. Sureshkumaar

Abstract The world is looking for a better alternative energy source to generate
electricity. Wind Energy Conversion Systems stands first in the choice of researchers
in terms of renewable energy. Permanent Magnet Synchronous Generator (PMSG)
plays an important role in wind energy generation system design. They exhibit a
character of generating electrical energy even at a low speed of operation. Also, the
control strategies of PMSG based WECS are quite easy compared to other generators
used in wind energy generators. A pitch angle control technique along with load
voltage variation control is proposed in this article. A MATLAB/Simulink based
simulation work has been proposed here and the test results are quite satisfactory.

Keywords PMSG · DC link · Wind Energy Conversion System (WECS) · Pitch

angle control · PIC controller

1 Introduction

Wind power conversion system is one of the cleanest sources of electrical power
generation systems. The power produced per unit area is quite impressive compared
to all other power generation techniques. Variable speed WECS are most popular in
wind power technologies. They can deliver maximum power point tracking as well.
Their efficiency is also compared to fixed speed and limited variable speed WECS.
The first choice of electrical generator now a days is doubly fed induction generator
(DFIG). In this article PMSG based WECS have been proposed. This generator
improves the reliability of the entire system. The features of PMSG are self-excitation

A. Nagarajan · G. Sureshkumaar
Karpagam College of Engineering, Coimbatore, India
e-mail: [email protected]
G. Sureshkumaar
e-mail: [email protected]
K. Ramash Kumar (B)
Dr. N.G.P Institute of Technology, Coimbatore, India

© The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 95
R. Ibrahim et al. (eds.), International Conference on Artificial Intelligence
for Smart Community, Lecture Notes in Electrical Engineering 758,
https://doi.org/10.1007/978-981-16-2183-3_8
96 A. Nagarajan et al.

Fig. 1 PMSG based wind turbine

capability, leading to a high power factor and high efficient operation. A standalone
PMSG based WECS can be used as a prime source of energy to the places where grid
integration is impossible. The reason is PMSG based WECS can generate electricity
even at a low speed of wind flow.
The main challenge of standalone power system is the feeding the end users with
a balanced power supply with proper control of power converters. In wind energy
conversion systems some power quality issues like voltage imbalalnce, generation of
harmonics, flickering and load variation will occur very often. The change in voltage
occurs often due to variation in the load. The oscillating voltage of WECS is mainly
due to change in their wind speed. Also, the harmonics are produced due the high
frequency operating power converters of WECS. Those power quality problem could
not tolerated by the consumers and so, require mitigation methods [1]. Controllers
for power converters were detailed in [2–9].
Figure 1 shows the schematic diagram of permanent magnet synchronous gener-
ator based wind energy conversion system. When load varies with in a short period,
that causes a huge variation in voltage at common point of coupling. This will result
in torque ripple sin the shaft which may affect the life of the machine,
A standalone small scale PMSG based WECS has been explained in this article.
The major considerations are,
1 Implementation of pitch angle control under heavy wind turbulences.
2 Maintaining the constant voltage in the DC link to minimize ripple in the DC
input to the generator side converter.

2 Mathematical Equations Analyis of Wind Turbine

The energy extracted from the wind is directly proportional to velocity of the wind.
The power traced from the wind energy is expressed as below,

P = 0.5ρ Avw3 (1)

where
ρ—air density (kg/m3 );
Control of DC Link Voltage and Load Voltage … 97

A—the area swept by blades (m2 ); and.

Vw —wind speed (m/s).
Betz limit explains the fraction at which the wind turbine can extract power from
the wind flow. It is limited to 59% as per the Betz theory. Power coefficient Cp is
related to pitch angle of the blade and TSR (Tip Speed Ratio). This power coefficient
describes the betz limit.
The power extracted from the wind by the turbine is expressed as,

Pw = 0.5C p (β, λ)ρ Avw3 (2)

where
Cp—power coefficient,
β—pitch angle in degrees.
λ—tip speed ratio.
The tip speed ratio is defined as the ratio of the speed at which the tip of blade
rotates to the speed of the wind (wind velocity) Vw,

R
λ= (3)
vw

where
—speed of the turbine rotor.
R—radius of turbine rotor tip swept area.
Figure 2 shows the relationship between speed of the turbine rotor and the mechan-
ical power developed in the turbine. It is clear from the picture that is the mechanical
power developed in the rotor is directly proportional to the speed of rotor also the
developed power reaches maximum when at a particular point for various wind
speeds.
The typical speed vs power extraction plot is shown in Fig. 3. It shows that all the
WECS starts to produce electricity at a particular speed of wind. Till that the wind
generator will be acting as a motor that is receiving power from the system. Once the

Fig. 2 Turbine rotor speed

versus mechanical power
98 A. Nagarajan et al.

Fig. 3 Simulated responses of wind speed versus power extracted

Fig. 4 MATLAB/Simulink model of control of pitch angle

cut-in velocity is attained the machine starts to generate electricity that is feeding the
power to the power system. One major advantage of using PMSG base WECS is that
it can produce electricity from a minimum wind velocity compared to all other types
of wind energy conversion systems. From the Fig. 4 it is evident that that extracted
power is maximum from the speed 12–25 m/s. if speed exceeds beyond this limit it
can be controlled either mechanically by using pitch angle control technique or yaw
mechanism or electrically using the power converters.

3 PAC of Wind Turbine

Pitch Angle Control (PAC) is defined as the angle at which the wind strikes the
turbine blades with respect to the angle of rotation. When the speed of wind exceeds
the wind blades experiences more torque as shown in the above figure. The figure
says that the power generated beyond the cut off velocity is zero. That is when the
speed of wind exceeding the maximum value of rated velocity the mechanical control
comes into action.
Control of DC Link Voltage and Load Voltage … 99

There are two ways to control the wind turbine mechanically which are
1. Pitch angle control
2. Yaw mechanism
Since the power extracted from the wind is depending mainly on pitch angle and
tip speed ratio the adjustment of pitch angle should be optimum below the rated
speed of the wind. The need of pitch angle control is as below;
(1) Maximizing the obtained power of the wind turbine, below the cut off speed
the angle must be set to derive maximum power from the system.
(2) Regulating the power generation within a safe value. Ensuring the machine
operates with in the SOA (safe operating area).

4 Simulink Model of PAC

Following is the simulation model of the pitch angle control of wind energy conver-
sion system. Here the actual wind speed is measured and compared with the reference
value of wind flow. The error is amplified with help of a gain. After the gain point a
rate limiter is introduced which limits the speed of the rotor within a safe value. In
practical case a servo mechanism is employed to adjust the pitch angle. When the
wind velocity exceeds a particular value the servo motor turns the wind blade to a
particular angle. The feedback is given via the sensors which collect the speed of the
rotor of the turbine.

5 Influence on Voltage at PCC and Its Compensation

It is not necessary that the load is to be constant always. Due to variation in load
connected to the inverter the current in all the phases will not be uniform. In distribu-
tion network this is a major issue. This sudden change in load may lead to unbalance
voltage at the point of common coupling. There for compensation of the load voltage
variation is more important in the design of converters. A voltage control technique
using the PWM inverter is employed in this proposal. PI controllers have been used
to control the pulse generation. Here the actual value is compared with the reference
value, based on the variation between actual and reference current magnitudes the
PWM pulses are generated with different modulation index values (Fig. 5).
100 A. Nagarajan et al.

Fig. 5 MATLAB/Simulink model of PWM inverter controller for unbalanced load compensation

6 Simulation Responses and Its Discussions

The following figure shows the simulation model of the power converter used in the
proposed system. The results are compared with and without the pitch angle control
strategy (Fig. 6).
The speed of wind is set to 12 m per second for first 3 s and it is set to 25 m/s up to
7 s the again 12 m/s up to 10 s. in this the base speed is set to 12 m/s. From Fig. 7 the
generator speed crosses more than 1.5 p.u which is 2550 rpm. In Fig. 8 it is shown
that the generator speed is controlled to 1.2p.u at the excess wind speed (25 m/s).
And parameters like generator torque and output voltage are also controlled.
Figure 9 shows the DC link voltage control. The disturbance occurs at 0.25 s of
the simulation period. But the system is providing a constant voltage over the entire
simulation period because of the control of the rectifier as explained in Sect. 5.
The load is varied at 0.25 s by using breaker in the Simulink model. In the above
Fig. 10, the voltage output is kept constant throughout the run time. The disturbance
is tolerated by the pulse generation. The magnitude of the modulating signal is shown
in below plot.
Control of DC Link Voltage and Load Voltage … 101

Fig. 6 MATLAB/Simulink model of power converters

102 A. Nagarajan et al.

Fig. 7 Simulated responses of WECS without pitch control

Fig. 8 Simulated responses of WECS with pitch control

Control of DC Link Voltage and Load Voltage … 103

Fig. 9 Simulated responses of dc link voltage and magnitude of reference signal for PWM
generation of rectifier

Fig. 10 Simulated responses of output voltage of inverter and magnitude of reference signal for
pwm generation of inverter
104 A. Nagarajan et al.

Table 1 Generator ratings

Stator phase resistance 0.085 
Rotor speed 1700 rpm
Line voltage 2500 V
Torque constant 16.54 Nm
No of pole pairs 4

7 Conclusion

It is evident that under mechanical discrepancy conditions like higher wind speed
and multi directional wind power mechanical control of the systems works better
compared to electrical control instantly. i.e. pitch angle control result is quick as
compared to the electrical control. Also controlling the DC link voltage with in a
safe limit saves the generator as well as the load. PI controller based control is giving
better results as shown in the result. Modern controllers may give better result.

Appendix

See Table 1.

References

1. Bhende CN, Mishra S, Malla SG (2011) Permanent magnet synchronous generator-based

standalone wind energy supply system. IEEE Trans Sustain Energy 2(4)
2. Rolan A, Luna A, Vazquez G, Aguilar D, Azevedo G (2009) Modeling of a variable speed wind
turbine with a permanent magnet synchronous generator. In: IEEE International Symposium on
Industrial Electronics Seoul Olympic Parktel, Seoul, Korea July pp 5–8
3. Gupta J, Ashwani Kumar. Fixed pitch wind turbine based permanent magnet synchronous
machine model for wind energy conversion, pdf. (Available www.onlinejet.net)
4. Zhang J, Cheng M, Chen Z, Fu X (2008) Pitch angle control for variable speed wind. In:
Third international conference on electric utility deregulation and restructuring and power
technologies, Nanjing, China, pp 6–9
5. Ramash Kumar K (2019) FLC and PWM SMC for KY boost converter. J Circuits Syst Comput.
https://doi.org/10.1142/S0218126619501846
6. Ramash Kumar K (2016) Implementation of sliding mode controller plus proportional inte-
gral controller for negative output elementary boost converter. Alexandria Eng J (Elsevier)
55(2):1429–1445
7. Ramash Kumar K, Jeevananthan S (2012) Analysis, design and implementation of hysteresis
modulation sliding mode controller for negative output elementary boost converter. J Electric
Power Components Syst 40(3):292–311
Control of DC Link Voltage and Load Voltage … 105

8. Gayathri Monicka J, Guna Sekhar NO, Ramash Kumar K (2011) Performance evaluation of
membership functions on fuzzy logic controlled AC voltage controller for speed control of
induction motor drive. Int J Comput Appl 13(5)
9. Ramash Kumar K, Arunkumar N, Sivakumaran TS (2019) Implementation of non-linear
controller for contemporary DC-DC converters. Int J Electrical Eng Informatics 11(4):622–637
A Novel Approach of Wind MPPT Using
Fuzzy Logic

Nesar Uddin, Md Saiful Islam, Jewel Sikder Joy, and Ishaque Mozumder

Abstract The key motive of this article is to carry out the optimum power as well
as supreme power from wind turbine using fuzzy control technique. The wind power
system largely relies on wind velocity and pitch angle, so Fuzzy control technique is
used to extract supreme power in an efficient method which minimizes the dislocation
in the system reaction. For developing the fuzzy-based wind MPPT system, a DC/DC
step-up power converter is constituted to trigger the gate of its MOS-FET switch by
applying Fuzzy rules for regulating rotor acceleration of wind generator. The fuzzy
controller provides a signal as well as duty cycle as output through pulse width
modulation (PWM) that switches the gate of DC/DC converter. Simulation results
show that the developed fuzzy-based wind MPPT structure can perform nonlinear
functions of random difficulty and cover an extensive range of working conditions.

Keywords Fuzzy Logic Controller (FLC) · Maximum power point tracking

(MPPT) · Wind Energy Conversion System (WECS)

1 Introduction

In the mint condition, electricity is fundamental imperious of humankind, socioeco-

nomic, artificial development of several countries [1]. Yet again, most of the regions
did not entree to the light of electricity. The significant source of energy generation
in the creation is fossil fuel that raises some gases that influence the human body [2].

N. Uddin (B)
Department of EEE, Chittagong University of Engineering and Technology, Chattogram 4349,
Bangladesh
M. Saiful Islam
Department of ETE, Chittagong University of Engineering and Technology, Chattogram 4349,
Bangladesh
J. Sikder Joy · I. Mozumder
Department of IET, Chittagong University of Engineering and Technology, Chattogram 4349,
Bangladesh

© The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 107
R. Ibrahim et al. (eds.), International Conference on Artificial Intelligence
for Smart Community, Lecture Notes in Electrical Engineering 758,
https://doi.org/10.1007/978-981-16-2183-3_9
108 N. Uddin et al.

In 2019, be close to 19 TW fossil fuels have been used [3]. It also increases air pollu-
tion and damages the environment. To surpass the problem, the knowledge of new
technology of sustainable energy from solar irradiation, hydro source, wind resource,
geothermal store and biogas supply, are a more effective and reliable method in the
creation [4].
Several regions in Bangladesh are also probable for wind resource as well as wind
power generation via Patenga, Cox’s Bazar, around Teknaf area, Kuakhata and Kutub-
diya beach [5]. Generally, a wind turbine cannot generate extreme power by tradi-
tional control strategy [6]. To surpass these difficulties, MPPT methods are planned
to extract supreme power by applying FLC. Howsoever, Fuzzy control technique is
an expert intelligence thinking method that has beaten fruition in numerous cases to
take effective decisions like human beings by adjusting themselves for present and
future situations automatically [7]. Therefore, to extract supreme power from the
wind turbine, the fuzzy control technique has been used in this paper. The recom-
mended Fuzzy control technique can run an input/output planning by depending on
the human sense which is a more effectual and optimum way.

2 Methodology

2.1 Block Diagram for Wind MPPT Structure

Maximum power point tracking (MPPT) is a method that is used to bring out supreme
power from renewable foundations. The suggested wind MPPT is designed for
extracting efficient power that stands on wind turbine, DC/DC step-up converter,
fuzzy based control method and AC load as well as the electric appliance. The gener-
ated electricity from wind turbine relies on the atmospheric situation as well as wind
speed and direction of pitch that is distributed through DC/DC step-up converter.
The step-up converter is coupled with the fuzzy controller as well as the control unit.
The control division acts according to fuzzy rules which give duty cycle as output.
Subsequently, this output signal switches the gate of MOS-FET device through PWM
incessantly to get the extreme power. Control units have double inputs viz; error signal
and change in the error signal. The error signal is denoted as ‘e’ and change in the
error signal is denoted as ‘de’. Again ‘e’ and ‘de’ relies on a ratio of change in power
and change in voltage. Change in power is denoted as ‘P’ and change in voltage is
denoted as ‘V’. Overall block diagram has been presented in Fig. 1.

2.2 Fuzzy Based Efficient Wind MMPT

Generally, wind power conversion method, the output power of the wind turbine is
not linear because of unstable weather conditions. The fuzzy-logic-based controller
A Novel Approach of Wind MPPT Using Fuzzy Logic 109

Fig. 1 Proposed Fuzzy based wind MPPT structure

helps to get better performance in such a situation. This arrangement is made up of

two inputs and a single output that is illustrated below (Fig. 2). Error signal (e) and
change in error signal (de) are the inputs. Five triangular memberships have been
introduced for error signal. Those are Very Negative, Negative, Zero’ side, Positive
and Very Positive (denoted by VN, N, Z, P, VP), given below in Fig. 3. Three
triangular memberships have been considered for changing the error signal. Those
are Negative, Zero’s side and Positive (denoted by N, Z, P), exposed in Fig. 4. Duty
cycle represents the result as output membership function which has three triangular
memberships. These functions are Very Negative, Zero and Very Positive (denoted
by VN, Z, VP). They are illustrated in Fig. 5.
In support of MPPT method, we want e = (P ÷ V) is equal zero. Here, P
indicates power variation and V indicates voltage variation. In condition, e = (P
÷ V) greater than zero, we supposed to raise the voltage to tap the MPPT. In
condition, e = (P ÷ V) less than zero, it requires to lesser the voltage to tap the
maximum power [8].

Fig. 2 Overall block diagram of wind MPPT

110 N. Uddin et al.

Fig. 3 Membership function of error signal (e)

Fig. 4 Membership function of change in error signal (de)

Fig. 5 Membership function of duty cycle

A Novel Approach of Wind MPPT Using Fuzzy Logic 111

Table 1 Proposed fuzzy

IF THEN
rules for MPPT system
Number of Error signal, e Change in error Output signal
Fuzzy rules signal, de
1–3 VN/VN/VN N/Z/P VP/VP/VP
4–6 N/N/N N/Z/P VP/Z/Z
7–9 Z/Z/Z N/Z/P Z/Z/Z
10–12 P/P/P N/Z/P Z/Z/VN
13–15 VP/VP/VP N/Z/P VN/VN/VN

Fig. 6 Control unit of MPPT structure

The rotor velocity is the main dependence of the voltage of the wind turbine. As
a result, the FLC controller is introduced here to be in command of rotor speed for
higher power generation. The fuzzy formulas are précised in Table 1. The maximum
output is picked up using fuzzy control technique. DC/DC step-up converter is applied
here to position the duty ratio in the course of PWM. The error signal is a ratio of
power deviation and voltage deviation with respect to time. A mathematical model
of the control unit of MPPT system has been given in Fig. 6.

P P(t) − P(t − 1)
e= = (1)
V V (t) − V (t − 1)

The FLC is associated with double input (e and de) and single output (Duty Cycle)
that rely on IF–THEN affirmations which is named as fuzzy rules where IF indicates
the input and THEN indicates the outputs parts. Overall, 15 rules have been set
to produce supreme power from the scheme where various working conditions are
described that is given in Table 1.

2.3 Simulink Structural Design of Suggested Wind MPPT

The ultimate schematic illustration of MPPT system is exposed in Fig. 7 where the
arrangement is constructed with a wind turbine, DC/DC step-up converter, electric
112 N. Uddin et al.

Fig. 7 Overall Simulink of suggested wind MPPT system

load and other system’s controlling unit. The outputs have been culled by varying
the wind speed from the range of (3–8) meter per second.

3 Evaluation of Wind MPPT

The 3-D plane inspection of FLC is represented in Fig. 8 where X-axis is representing
the error signal (e) and Y-axis is indicating a change in error signal (de). The value
of ‘e’ is measured from negative 10 to positive 10 that rely on the fluctuation of
power and voltage. The value of ‘de’ is calculated as -3 to 3 which is précised for
the interruption period to bring the utmost power. The outcome signal of the FLC is
the duty cycle that varies with reference to ‘e’ and ‘de’.

Fig. 8 Surface view of duty

cycle of wind MPPT model
A Novel Approach of Wind MPPT Using Fuzzy Logic 113

The value of the error signal has been considered negative ten to positive ten
because of standard power as well as voltage of wind turbine. The standard power
and voltage of WES5 tulipo wind turbine are 2.5 kW and 400 V. The standard voltage
is converted to 220 V by 1-phase rectifier that is shown in Fig. 7. Therefore, we can
get a ratio between power and voltage which is approximately 10. This value can vary
negative ten to positive ten by depending on wind velocity. But the range of change
in the error signal does not change gradually. So, this value has been considered
negative three to positive three which is lower than the value of the error signal.
Case study—A
According to Fig. 9, when the value of the error signal is very negative and the
value of change in the error signal is negative, then the duty cycle is very positive
to get supreme power. Generally, it is not possible to get supreme power when the
value of the error signal is very negative and the value of change in the error signal is
negative. At any time, this situation may occur because of unavailable wind velocity.
Therefore, it is impossible to get maximum power from the turbine. So, the duty
cycle has been increased to bring out supreme power.
Case study—B
In this situations, when error signal is negative and change in error signal is
considered as the value of zero’ side, then duty cycle is assumed zero’s side. That
means the variation of change in error signal is not high by following error signal.
So, it is not necessary to change the duty cycle (Fig. 10).
Case study—C
In this case, the value of the error signal and change in error signal are considered
as positive and negative. So, the value of the duty cycle is assumed zero’s side because
of its operating condition. Error signal positive means the high operating condition

Fig. 9 When e-VN, de-N then Duty cycle-VP

Fig. 10 When e-N, de-Z then Duty cycle- Z

114 N. Uddin et al.

Fig. 11 When e-P, de-N then Duty cycle- Z

Table 2 Power comparison of proposed MPPT system with conventional system

Speed of wind in Direction of Pitch in Power-without MPPT in Power-Fuzzy Based
meter per second degree Watt Wind MPPT in Watt
3 0 28.22 30.95
3 1 28.22 30.95
4 0 81.62 89.79
4 1 81.62 89.79
5 0 233 252
5 1 233 252.7
6 0 438.8 482.9
6 1 438.8 482.9
7 0 1008 1086
7 1 1008 1086
8 0 1799 1988
8 1 1799 1988

of wind turbine and change in error signal negative means it does not influence on
the operating system. So, the duty cycle has remained beside zero (Fig. 11).
Power difference of Fuzzy based wind MPPT system and without MPPT system
is given in Table 2. From the Table, it is clear that Fuzzy based wind MPPT system
is more efficient than a conventional system.

4 Conclusion

The suggested fuzzy based wind MPPT system that has diverse benefits over previ-
ously accessible systems. Fuzzy system is used which depends on nonlinear func-
tions which is very modern and reliable. The petition of Fuzzy technique is swiftly
mounting due to its strong performance regardless of complex situations and elas-
ticity. It is possible to increase the efficiency by adapting supplementary intellectual
fuzzy rules. A problem of the proposed structure is their remittent environment of
A Novel Approach of Wind MPPT Using Fuzzy Logic 115

wind speed that is dependent on weather variations. Fuzzy technique has been applied
to extract extreme power by two input parameters via Wind velocity & pitch angle.
The results of the Fuzzy based wind MPPT model has been compared with the
conventional system under a similar climate situation. Simulation result provides a
grand performance of Fuzzy based wind MPPT model and less error. The applica-
tion of renewable sources minimizes various gas discharges in the system. Besides,
the system is environment kindly, favorable. Developed Fuzzy based wind MPPT
structure more efficient than any other operable system as well as a grid-connected
system that must reduce power loss.

References

1. Soedibyo T, Hadi A, Putra HP, Suryoatmojo H, Ashari M (2016) Integration of hybrid PV/wind
generation system using fuzzy MPPT in grid connected system for remote area. In: MATEC web
of conferences, International Conference on Frontiers of Sensors Technologies (ICFST 2016),
vol 59, pp 1–2
2. Gowri M, Vinothkumar V, Punitha K (2015) ANFIS based MPPT and droop controller for power
sharing in interline power system. Int J Sci Res Dev 3(04)
3. Khalil GA, Lee DC, Seok jk (2014) Variable speed, wind power generation system based on
fuzzy logic control for maximum output power tracking. In: 35th annual IEEE power electronics
specialists conference Aachen, Germany, pp 5–6
4. Uddin N, Islam MS (2018) Optimal fuzzy logic based smart energy management system for real
time application integrating RES, grid and battery. In: 4th international conference on Electrical
Engineering and Information & Communication Technology, IEEE
5. Islam MR, Rahman MN, Mannan MA (2016) Study of wind power in different parts of
Bangladesh. Int Res J Eng Technol (IRJET) 03(09): 1290. ISSN: 2395-0056
6. Uddin N, Islam MS (2018) Optimization of PV energy generation based on ANFIS. In:
International conference on Innovations in Science, Engineering & Technology, IEEE
7. Hailemariam MM, Mekonnen T, Sudheendra H (2015) Novel approach to fuzzy logic controller
based hybrid solar/micro hydro/bio-mass generation, A real time analysis (Barsoma Village,
Ethiopia). Int J Innov Res Eng Multidisciplinary Phys Sci (IJIRMPS) 3(1). ISSN: 2349-7300
ISO 9001
8. Minh HQ, Cuong NC, Chau TN (2014) A fuzzy-logic based MPPT method for stand-alone wind
turbine system. Am J Eng Res (AJER) 3(9): 177–184. e-ISSN: 2320-0847 p-ISSN: 2320-0936
Shale Gas Productive Volume
Optimization

Adamu Umar Ibrahim and Berihun Mamo Negash

Abstract Recently, a modified mathematical model was presented in order to deter-

mine the productivity of multi-fractured shale gas wells. However, the considerable
challenge of the model was assuming an average fracture width for the determination
of the practical fracture geometry of the stimulated reservoir volume. This challenge
led to over prediction of the gas well. Therefore, this paper presents a simple and
accurate method using a pseudo-3D model to determine the fracture width in shale
gas reservoirs. The method utilized the hydraulic fracturing propagation capacity of
the MATLAB software to visualize the fracture geometry, which includes the fracture
height, fracture length, and the fracture width. So, the obtained fracture width can be
incorporated with analytical models to predict long term productivity for multistage
fractured shale gas wells. An accurate result for the productivity will assist in deter-
mining the ultimate gas recovery, propped volume, optimal fracture length, fracture
spacing and predict the future performance of the well.

Keywords Shale gas · Hydraulic fracturing · Fracture geometry

1 Introduction

The combination of horizontal drilling and hydraulic fracturing technologies has

already boosted the exploitation of shale reservoirs by many countries around the
globe [1]. This unconventional resource compliment the hydrocarbon demand all
over the world. Therefore, these technologies are revolutionary and will remain vital
in shale gas production [2].
One of the significant challenges in shale gas production optimization is the fore-
casting of long-term productivity of shale gas wells. Currently, three types of models

A. Umar Ibrahim (B)

Petroleum Engineering Department, Universiti Teknologi PETRONAS, Perak, Malaysia
e-mail: [email protected]
B. Mamo Negash
Shale Gas Research Center (SGRC), Universiti Teknologi PETRONAS, Perak, Malaysia
e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 117
R. Ibrahim et al. (eds.), International Conference on Artificial Intelligence
for Smart Community, Lecture Notes in Electrical Engineering 758,
https://doi.org/10.1007/978-981-16-2183-3_10
118 A. Umar Ibrahim and B. Mamo Negash

are applied to predict productivity. These models are basically: analytical transient,
numerical and empirical models [3, 4]. Recently, a simple model was presented,
which successfully predict productivity with less than 3% over prediction. The analyt-
ical model is applied in pseudo-steady state flow conditions. However, the fracture
width is assumed and computed as an average in the model [4]. Therefore, there
is a need to find a more reliable way of obtaining a fracture width that will close
the overprediction gap of the model. An accurate fracture width is vital to proper
field performance prediction, fracture completion design and adequate placement of
proppants in shale gas wells [5]. Also, this will give rise to designing a fracturing
treatment that will propagate in a more optimal direction and avoid unnecessary
damage to the subsurface environment.
This paper presents a simple and more accurate method using a pseudo-3D model
to determine the fracture width in shale gas reservoirs. The model gives the visual
result of the fracture geometry using MATLAB software. So, the obtained fracture
width can be incorporated with the latest analytical model to forecast the long-term
productivity of multistage fractured shale gas wells.

2 Methods and Procedures

Multistage hydraulic fracture in horizontal wells initiates many fractures from each
perforation cluster. These change the formation stress field and affect the propagation
of hydraulic fracture. Also, the low-viscosity fracturing fluid will leak off from the
surface of hydraulic fracture into the reservoir, therefore, increasing its pressure.
Consequently, the natural fractures will be stimulated by the induced stress. If the
shear stress is greater than the shear strength, shear failure occurs [6]. Likewise, if
the natural fractures inner pressure is greater than the tensile strength, tensile failure
happens [7, 8]. Therefore, these increase the shale reservoir permeability. The fracture
geometry, which has the fracture height, length, and width, directly control most of
the productivity of the well [9, 10]. Proper design of these parameters aid in gas
production optimization during fracturing treatment design.
Furthermore, it is good to visualize the fracture geometry using the pseudo-3D
model in MATLAB software to determine the fracture width. The P3D model is
compared with the well-known KGD and PKN models (Table 1).
Where, Qg = production rate of the gas (MScf/D); nf = number of fractures; μ =
viscosity of the fluid (cp); T = temperature of the formation (°R); km = permeability
of matrix (md); h = reservoir thickness (ft); p = average reservoir pressure (psia); pw
= wellbore pressure (psia); e = exponential function; xf = fracture half-length (ft);
kf = permeability of fracture (md); w = average fracture width (in); Sf = fracture
spacing (ft).
Shale Gas Productive Volume Optimization 119

Table 1 KGD, PKN & P3D models comparison

MODEL: KGD/2D Assumptions
8G Q 3 1/6 2/3
L= 0.48( (1−v)μ ) .t • Height is constant
• A plane strain fracture in the horizontal plane
8(1−v) Q 3 μ
W 0 = 1.32( G )1/6 .t 1/3 • Viscous and Newtonian fracturing fluid
3 • Elasticity theory used
2G Qμ 1/4
Pw = σ mi n + 0.96( (1−v) 3 L2 ) • No leak off into the formation
MODEL: PKN/2D
 
L = π cQ1 h t 1/2 • Plane strain fracture
• The fracture toughness does not influence the
Q μ 1/4 1/8
2
W 0 = 4( 2(1−v)
(1−v)μ ) .t fracture geometry
• No leak off into the formation
G 4 Q 2 μ 1/5 1/5
Pw = 2.5( ) .t
(1−v)4 h6
MODEL: PSEUDO-3D
P enet r at i ondept h(R) = • The vertical stress distribution is uniform
3 • The shape of fracture is elliptical
0.548( GGQ )1/9 .t 1/9 • No leak off into the formation [11]
3μ
W 0 = 1.32( 8(1−v)
G
Q
)1/6 .t 1/3
5 GW 0 (r w )
Pw = σ mi n − 4π R ln R
G = shear modulus (psi); Q = pumping rate
(bbl/min); W 0 = maximum width (ft); v =
poisson ratio; L = fracture length (ft); r w =
wellbore radius (ft); σmin = minimum stress
(psi); μ = fluid viscosity (cp)
 
5.87×10−5 n f km h Ṗ 2 −Pw2
MULTIFRACTURED SHALE GAS WELLS ⎡ ⎤
PRODUCTIVITY √
µT S f c⎣  −1√cx − 1
√ ⎦
1−e f (3x f c)
96km
c= k f wS f

3 Results and Conclusions

The single hydraulic fracture propagation is first shown to compare a P3D model with
the conventional KGD and PKN model. The visualized fracture width provides the
quantitative result of the width, which makes it easy to be applied in the productivity
model.

4 Discussion

From Figs. 1, 2, 3, 4, 5, 6, 7 and 8, it is shown that the length of the fracture

improves with greater rate than the fracture width of the fracture, which increases
120 A. Umar Ibrahim and B. Mamo Negash

Fig. 1 Fracture length

against time (KGD)

Fig. 2 Max width against

time (KGD)

Fig. 3 Max width against

length (KGD)

Fig. 4 KGD in 3D
Shale Gas Productive Volume Optimization 121

Fig. 5 Fracture width

against length (PKN)

Fig. 6 Fracture width

against height (PKN)

Fig. 7 Fracture height

against width (P3D)

Fig. 8 Pseudo-3D model

122 A. Umar Ibrahim and B. Mamo Negash

in proportional to t1/ 3. These models show a curvy tip which indicates that in the
KGD model, the effect of the tip of the fracture is not large. Both KGD and PKN
showed an approximation the width of the fracture and length. The pseudo-3D model
forecasts a complete ellipsoid in 3D space at a time which can be seen in Figs. 9,
10 and 11. Therefore, the P3D model is good for fracture propagation modelling.
The model can display the variation of the fracture height with length as the fracture
propagates.

Fig. 9 Fracture length

against time (PKN)

Fig. 10 Max. width against

time (PKN)

Fig. 11 P3D predicted

shape
Shale Gas Productive Volume Optimization 123

The fracture width is directly taken from the visualized fracture geometry and
subsequently inputted in the multi fractured shale gas wells. This method is more
dependable to determine the fracture width of a fractured shale gas reservoir to
determine the productivity of gas wells. Also, it can be applied in order to optimize
production by adequate fracture spacing and proppant placement.

References

1. Abdelaziz A et al (2019) Unconventional shale hydraulic fracturing under true triaxial labora-
tory conditions, the value of understanding your reservoir. In: SPE annual technical conference
and exhibition, Society of Petroleum Engineers
2. Quainoo AK, Negash BM, Bavoh CB, Ganat TO, Tackie-Otoo BN (2020) A perspective on
the potential application of bio-inhibitors for shale stabilization during drilling and hydraulic
fracturing processes (in English). J Natural Gas Sci Eng 79:103380
3. Mahmoodi S, Abbasi M, Sharifi M (2019) New fluid flow model for hydraulic fractured wells
with non-uniform fracture geometry and permeability. J Natural Gas Sci Eng, Article 68, Art
no 102914
4. Li G, Guo B, Li J, Wang MJSD, and Completion (2019) A mathematical model for predicting
long-term productivity of modern multifractured shale-gas/oil wells, vol 34, no 02, pp 114–127
5. Belyadi H, Fathi E, Belyadi F (2019) Hydraulic fracturing in unconventional reservoirs:
theories, operations, and economic analysis. Gulf Professional Publishing
6. Wu Y-S (2018) Hydraulic fracture modeling. Gulf Professional Publishing
7. Ren L, Lin R, Zhao JZ (2018) Stimulated reservoir volume estimation and analysis of hydraulic
fracturing in shale gas reservoir, (in English). Arab J Sci Eng 43(11):6429–6444
8. Ren L, Lin R, Zhao JZ, Rasouli V, Zhao JY, Yang H (2018) Stimulated reservoir volume
estimation for shale gas fracturing: mechanism and modeling approach, (in English). J Petrol
Sci Eng 166:290–304
9. Ahn CH, Dilmore R, Wang JY (2017) Modeling of hydraulic fracture propagation in
shale gas reservoirs: a three-dimensional, two-phase model. J Energy Resources Technology
139(1):012903
10. Berawala DS, Andersen PO, Ursin JR (2019) Controlling parameters during continuum flow
in shale-gas production: a fracture/matrix-modeling approach, (in English). SPE J Conf Paper
24(3):1378–1394
11. Yew CH, Weng X (2104) Mechanics of hydraulic fracturing. Gulf Professional Publishing
A Comparative Study and Validation
of Kinematic Analysis of a Crank Rocker
Engine Prototype Using MATLAB
and ADAMS

A. M. Albaghdadi , M. B. Baharom, and S. A. Sulaiman

Abstract This work presents an analytical study of a suggested mechanism as

part of a design process of a Crank-Rocker (CR) engine. The basic geometry of
this engine is based on the principal of a four-bar crank rocker mechanism, which
requires optimization for effective operation. Kinematic study of this mechanism is
achieved using two analysis approaches, analytical and graphical model simulation
approach. To achieve the objective, two softwares, namely MATLAB and ADAMS,
were used to conduct this analysis and to validate mechanism functionality. Under
certain constraints and operational conditions, the simulation results by both applica-
tions were compared and found to be almost identical, this gives an indication about
mechanism operational behaviour. Moreover, this study provides clear idea about
using different approaches and applications for system kinematic study purposes.

Keywords Crank-Rocker engine · Kinematic study · Mechanism synthesis ·

Model simulation

1 Introduction

A new engine called Crank-Rocker engine (CR) was patented in 2018 by engineers
of Universiti Teknologi Petronas, Malaysia. Despite showing promising results, the
engine is still considered belonging to early development stage. Of these studies
which has been performed on crank-rocker engine and used as base reference for
this paper and any further improvements on this engine, work in researches [1–4].

A. M. Albaghdadi (B) · M. B. Baharom · S. A. Sulaiman

Department of Mechanical Engineering, Universiti Teknologi Petronas (UTP), 32610 Bandar Seri
Iskandar, Perak, Malaysia
e-mail: [email protected]
M. B. Baharom
e-mail: [email protected]
S. A. Sulaiman
e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 125
R. Ibrahim et al. (eds.), International Conference on Artificial Intelligence
for Smart Community, Lecture Notes in Electrical Engineering 758,
https://doi.org/10.1007/978-981-16-2183-3_11
126 A. M. Albaghdadi et al.

It is desired to enhance the crank rocker (CR) engine performance by upgrading

engine components through geometry design optimization. An initial mechanism
configuration of the crank rocker engine is suggested using algorithm optimization
method. The next task is to perform kinematic and dynamic study of the configured
as part of the design process.
The study of system kinematics can be done using analytical method, graph-
ical method or iterative numerical methods [5]. Nowadays, it is significant to use
rapid and effective method that is capable of solving real life problems, rather than
using conventional methods which might consume designer time and effort. Hence,
many applications and softwares are being used to accomplish design objectives
effectively, with more reliable outcome in considerable time. Of these software,
the ones used to perform kinematic analysis and simulation are namely MATLAB,
MSC ADAMS, SAM, GIM, Universal Mechanisms, MMtool, and Linkage [5]. On
one hand, for coded softwares such as MATLAB, the analyses are performed by
constructing models and solving systems of mathematical equations. The imple-
mentation of such software is considered more efficient than solving these equations
manually. However, the previous process might be very tedious due to the listing of
all equations in a logical sequence that must be followed and understood by soft-
ware compiler. On the other hand, application software such as ADAMS, is a virtual
modeling tool, meaning that it only requires mechanism geometry and initial oper-
ation condition to generate full body analysis and animate corresponding actions
[6].
In this paper, a kinematic analysis study of a particular crank rocker engine
mechanism analyzed using two softwares, namely MATLAB and ADAMS. Posi-
tion, velocity and acceleration of this mechanism are studied under specified driving
motion and system constraints. Moreover, to validate mechanism kinematics, a
comparison between both simulation methods is conducted using the outcome results
of both applications.

2 Analytical Analysis of Crank Rocker Mechanism

in MATLAB

Since a MATLAB code is utilized to perform analytical study on this system [7–
9], it is required to include related mathematical equations that represent system
physics. Hence this section introduces kinematic fundamentals which are involved
in analyzing the system. Once the mechanism geometry and other conditions such
as Grashof’s law and initial conditions are identified, mathematical equations are
constructed. This study adopted the usage of vector-loop method for a crank-rocker
mechanism kinematic analysis. Figure 1, shows the crank rocker basic geometry
which consists of four main linkages namely ground R1 , Crank R2 , Coupler R3 and
Rocker R4 . Also, each respected link angular position with reference to the ground
is set to θ1 , θ2 , θ3 and θ4 , which are (ground linkage angle, crank angle, coupler
A Comparative Study and Validation of Kinematic … 127

Fig. 1 Crank-Rocker basic

geometry

angle and rocker angle) respectively. The angle between the coupler and rocker is
called the transmission angle γ. Another angle of concern in this design process is
the throw angle ϕ, which represents maximum rocker displacement, in other words
it is obtained by subtracting maximum and minimum crank rocker angle (θ4 ).
In this case, the vector loop equation is represented in Eq. (1):

R2 + R3 = R1 + R4 (1)

By arranging this relation, we have

R2 + R3 − R1 − R4 = 0 (2)

Arbitrary link position relative to the X–Y axis, for each components of Eq. (2)
can be rewritten as:

R2 cos θ2 + R3 cos θ3 − R4 cos θ4 − R1 cos θ1 = 0 (3)

R2 sin θ2 + R3 sin θ3 − R4 sin θ4 − R1 sin θ1 = 0 (4)

Equation (3) represents system component position in the X-axis direction, while
Eq. (4) represents system component position in the Y-axis direction. When applying
first time derivative on Eqs. (3) and (4), also since is ground link is fixed (ω1 = 0),
we get angular velocity Eqs. (5) and (6):

−R2 sin θ2 ω2 − R3 sin θ3 ω3 + R4 sin θ4 ω4 = 0 (5)

R2 cos θ2 ω2 + R3 cos θ3 ω3 − R4 cos θ4 ω4 = 0 (6)

In the above Equations ω2 , ω3 and ω4 are angular velocities of links R2 , R3 , and

R4 respectively. Similarly, when applying second time derivative on both Eqs. (3)
and (4), they yield angular acceleration Eqs. (7) and (8):
128 A. M. Albaghdadi et al.

− R2 sin θ2 α2 − R2 cos θ2 ω22 − R3 sin θ3 α3 − R3 cos θ3 ω32

+ R4 sin θ4 α4 + R4 cos θ4 ω42 = 0 (7)

R2 cos θ2 α2 − R2 sin θ2 ω22 + R3 cos θ3 α3 − R3 sin θ3 ω32

− R4 cos θ4 α4 + R4 sin θ4 ω42 = 0 (8)

Also, α2 , α3 , and α4 above represent the angular accelerations of links R2, R3,
and R4 respectively. Linear velocity for joints between linkages, in this case point A
and B as illustrated in Fig. 1, basically can be found by Eq. (9):

υ =ω×R (9)

where, R is linkage length and ω is the angular speed. Likewise, linear acceleration
for both points can be found by Eq. (10):

a = at + an (10)

where, at is the tangential acceleration of this point and an is the normal acceleration
of the same point, and both can be found by:

at = α × R (11)

v2
an = (12)
R
Here, α is the angular acceleration of link R, and v is the linear velocity at the
measured point. Using all of the above equations with known crank position (0° ≤ θ2
≤ 360°) and constant rotational speed ω2 = C, MATLAB codes are programmed for
system analysis. All remaining parameters are calculated and translated into system
simulation analysis.

3 Modelling and Simulation in ADAMS

Unlike programming with softwares such as MATLAB [6, 10, 11], ADAMS has the
advantage of less need to write mathematical equations to identify system physics.
The required task is only to represent the geometry of this system into the application
interface, and then ADAMS has the ability to translate this system into a model for
simulation and analysis. ADAMS also allows users to perform static, kinematic and
dynamic study on different system conditions, in addition to the ability to simulate
and predict the behavior of each case. However, other procedures are required to
A Comparative Study and Validation of Kinematic … 129

Fig. 2 Crank-rocker model in ADAMS

accomplish valid simulation such as identifying all links and joints dimensions.
Also, it is desired to create proper measurements and variables of each object to
be included and calculated during the simulation process. After model set-up is
finished, the simulation process can be performed, and a full evaluation of system
characteristics is created. Figure 2, shows ADAMS workspace environment, the
crank rocker mechanism is drawn with corresponding system geometry. Also, joints
between linkages are identified with crank joint having a motion function for system
initiation.

4 Results and Discussion

Selection of mechanism dimensions was previously conducted using algorithm opti-

mization method [12]. Table 1, shows the geometry specifications of the crank rocker
mechanism.
The above data were then supplied to MATLAB and ADAMS as a system config-
uration. Also, a constant crank rotation speed of 200 rad/sec was applied at the crank
link. The ground linkage angle θ1 is assumed to be zero so that there is no offset
between the crank and rocker linkages. The path motion of the mechanism is illus-
trated in Fig. 3, by MATLAB. In ADAMS, an animated simulation of the system is
firstly performed to validate this mechanism then carry on system analysis.

Table 1 Crank-rocker mechanism dimensions

R1 R2 /mm R3 R4 θ2 γ
/mm /mm /mm /degree /degree
33.23 14.03 35 21.07 180 84.34
130 A. M. Albaghdadi et al.

Fig. 3 Crank-Rocker mechanism path motion in MATLAB

Figure 3, shows the complete rotation of the crank link (green), and the corre-
sponding rocker path motion (red). After analyzing this system path behavior, a
simulation was conducted in both softwares for studying the kinematic performance
of this mechanism. The outcome results were plotted in overlayed graphs for better
illustration and verification purpose. The plots of angular position, velocity and accel-
eration versus crank position (θ2 ), obtained from MATLAB and ADAMS simulations
are compared and shown in Figs. 4, 5, 6, 7 and 8.
Figures 4, 5, 6, 7 and 8, shows overlaying graphs of the different kinematic anal-
ysis for the crank-rocker mechanism by MATLAB and ADAMS. These data are
representing mechanism various positions, angular and linear velocities, angular

Fig. 4 Kinematic simulation comparison results between MATLAB and ADAMS, Mechanism
angular position
A Comparative Study and Validation of Kinematic … 131

Fig. 5 Kinematic simulation

comparison results between
MATLAB and ADAMS,
Mechanism linear velocity

Fig. 6 Kinematic simulation

comparison results between
MATLAB and ADAMS,
Mechanism angular velocity

Fig. 7 Kinematic simulation

comparison results between
MATLAB and ADAMS,
Mechanism linear
acceleration
132 A. M. Albaghdadi et al.

Fig. 8 Kinematic simulation

comparison results between
MATLAB and ADAMS,
Mechanism angular
acceleration

and linear accelerations with respect to the crank position. When examining and
comparing results by both softwares, it can be noticed that both models are showing
almost same readings, which gives an indication that both results are verifying each
other. Hence, these results can be considered and applied for further system analysis
and enhancement.

5 Conclusion

This paper studied the kinematic characteristics of a possible mechanism design

of the crank rocker engine. Two different approaches were implemented for this
goal, analytical coding using MATLAB and animated model using ADAMS. Both
methods simulated this system for determined crank rotation speed and results show
this mechanism kinematic behavior in term of position, velocity and acceleration. A
comparison between outcome results by both softwares validated each other, where
all plots were almost identical. These two methods considered practical and useful
for kinematic analysis of such system.

References

1. Mohammed SE, Baharom MB, Rashid A, Aziz A (2016) Thermodynamic analysis of a single
cylider crank-rocker engine. ARPN J Eng Appl Sci 11:12239–12246
2. Mohammed SE, Baharom MB, Aziz ARA (2017) Performance and combustion characteristics
of a novel crank-rocker engine. J Mech Sci Technol 31:3563–3571
3. Hassan SEM (2018) Development, performance and emission assessment of a single-cylinder
air-cooled gasoline Crank-Rocker engine. Universiti Teknologi Petronas
4. Mohammed SE, Baharom MB (2019) Modelling of combustion characteristics of a single
curved-cylinder spark-ignition crank-rocker engine. Energies 1–14
A Comparative Study and Validation of Kinematic … 133

5. Verma SK, Kumar R, Chittawadigi RG, Saha SK (2016) Kinematic analysis of mechanisms
using velocity and acceleration diagrams (VAD) module in mechanalyzer software. In: The
proceedings of the Asian conference on multibody dynamics, p. 01_771082
6. Hroncová D, Delyová I, Frankovský P (2014) Kinematic analysis of mechanisms using MSC
Adams. Appl Mech Mater 611:83–89
7. Norton RL (2012) Design of machinery: an introduction to the synthesis and analysis of
mechanisms and machines, McGraw-Hill, New York
8. Mohammadzadeh A (2007) Analytical synthesis and analysis of mechanisms using MATLAB
and SIMULINK. Am Soc Eng Educ 12(242):1
9. Myszka DH, Celik A (2012) Machines and Mechanism. Prentice Hall
10. Hroncová D, Binda M, Šarga P, Kičák F (2012) Kinematical analysis of crank slider mechanism
using MSC ADAMS/View. Procedia Eng 48:213–22
11. Hroncová D, Frankovský P, Virgala I, Delyová I (2014) Kinematic analysis of the press
mechanism using MSC Adams. Am J Mech Eng 2:312–315
12. Albaghdadi AM, Baharom MB, bin Sulaiman SA (2021) Parameter design optimization of
the crank-rocker engine using the FMINCON function in MATLAB. IOP Conference Series:
Materials Science and Engineering 1088(1):012072. https://doi.org/10.1088/1757-899X/1088/
1/012072
Application of Machine Learning Models
in Gas Hydrate Mitigation

Sachin Dev Suresh, Bhajan Lal , Ali Qasim , Khor Siak Foo,
and Jega Divan Sundramoorthy

Abstract The production of oil and natural gas contributes to a significant amount of
revenue generation in Malaysia thereby strengthen the country’s economy. The flow
assurance industry is faced with impediments during smooth operation of the trans-
mission pipeline in which gas hydrate formation is the most important. It affects the
normal operation of the pipeline by plugging it. Gas hydrate is a crystalline struc-
ture composed of a network of hydrogen bonds between host molecules of water
and guest molecules of the incoming gases under high pressure and low temper-
ature conditions. Industry uses different types of chemical inhibitors in pipeline
to suppress hydrate formation. To overcome this problem, machine learning algo-
rithm has been introduced as part of risk management strategies. The objective of
this research is to evaluate the various types of machine learning models used to
predict the gas hydrate formation where the input parameters are gas composition,
pressure and concentration of inhibitor and the output parameter is hydrate deposi-
tion/formation temperature (HDFT). Three machine learning models are compared:
Artificial Neural Network (ANN), Least Square version of Support Vector Machine
(LSSVM), and Extremely Randomized Trees (Extra Trees). Comparison of the three
different machine learning models is based on the correlation coefficient, R2 . The
best choice of machine learning model that has highest R2 is obtained by Extra Trees

S. D. Suresh · B. Lal (B) · A. Qasim

Chemical Engineering Department, Universiti Teknologi PETRONAS, 32610 Bandar Seri
Iskandar Perak, Malaysia
e-mail: [email protected]
B. Lal · A. Qasim
CO2 Research Centre (CO2RES), Universiti Teknologi PETRONAS, 32610 Bandar Seri Iskandar
Perak, Malaysia
K. S. Foo
PTTEP-Malaysia, Petronas Twin Towers Kuala Lumpur City Centre, Level 26-30, Tower 2, 50088
Kuala Lumpur, Malaysia
J. D. Sundramoorthy
Federal Territory of Kuala Lumpur, Baker Hughes (M) Sdn. Bhd, 207 Jalan Tun Razak, 50400
Kuala Lumpur, Malaysia

© The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 135
R. Ibrahim et al. (eds.), International Conference on Artificial Intelligence
for Smart Community, Lecture Notes in Electrical Engineering 758,
https://doi.org/10.1007/978-981-16-2183-3_12
136 S. D. Suresh et al.

model of 0.9991 compared to other two machine learning models which predicted
R2 value greater than 0.96.

Keywords Gas hydrate · Machine learning · Hydrate formation temperature

1 Introduction

In the 1930s, gas hydrate deposition was discovered in pipelines where several
attempts were made to understand the clathrate hydrate in terms of compound,
composition, equilibrium state and manner of handling them [1]. Gas hydrate, mostly
known as clathrate hydrates exist when sufficient amount of water molecules in oil
and hydrocarbon vapour form crystallize during transient flow condition [2, 3]. Gas
hydrates are developed in the crude oil and natural gas industries’ output, refining and
transmission facilities. [4–8]. Depending on the type and size of guest molecules,
the gas hydrates are produced in 3 structures, which are I, II and H [7, 9]. The
unit cells of Structure I had eight cavities (two smalls and six large ones) and were
formed by hydrogen-bound 46 water molecules. Structure II unit cells, by contrast,
contained 136 molecules of water and were enclosed in 24 cavities, including 16
small and 8 wide cavities. Whereas structure H is a mixture of methane, propane,
isobutene, CO2 , H2 S, N2 , and stabilises with one of the two hydrate structures forming
micro cavities[10, 11]. According to Qin et al. [12] and Olajire [13], at high pres-
sure and low temperatures as well as long distances, gas hydrates are produced,
often contained in subsea oil and gas pipelines. Although gas hydrate production
contributes to critical flow assurance problem, it is therefore important to address
this issue by implementing effective preventive and predictive strategies [4].
There are several risk management strategies that has been used to prevent the
growth of hydrates which are dehydration of natural gas, usage of inhibitors such as
thermodynamic hydrate inhibitors (THIs) and low dosage hydrate inhibitors (LDHIs)
[12–16]. Methanol (MeOH) and monoethylene glycol (MEG) are two commercially
used Thermodynamic hydrate inhibition compounds [17]. The goal of these THIs is to
interrupt the process in which hydrates form, often like water–ice antifreeze, to ensure
safety under the most severe hydrate forming conditions. However, the downside of
these THIs is that the hydrate condition gets extreme as the subcooling process is of
better consistency, leading to more inhibitor requirements. This leads to a rate limit
of methanol supply due to availability, storage and injection restrictions, leading to
an adverse effect on capital costs, operating costs, non-optimum performance, as
well as an increased risk of hydrate plug formation [1].
In contrast with THIs, LDHIs play a crucial position as an alternative to THIs in
terms of lower doses of inhibitors used and dosage levels that cause lower loss of
inhibitors due to evaporation. Other than that, LDHIs would decrease the chemical
storage and injection needs, which will have a beneficial impact by minimising
CAPEX and OPEX [1]. Two kinds of LDHIs, such as kinetic hydrate inhibitors (KHI),
are used to slow hydrate crystallisation and anti-agglomerants (AAs), which facilitate
Application of Machine Learning Models in Gas Hydrate Mitigation 137

hydrate molecule distortion [12]. For simulation purposes, computer intelligence can
be used, such as machine learning (ML) models, to test the accuracy of the inhibitor
used, composition of gas molecules, gas hydrate temperature and pressure regulation
[4].
Usually, ML is called a universal toolbox, ready to be used for classification issues,
defining an appropriate group for a new collection of findings, and regression tasks
that approximate the relationship between groups of data provided. ML is fundamen-
tally a numerical representation of a phenomenon, provided a certain value and based
on a certain setting, aimed at performing a task. ML techniques may be classified
based on learning activity priorities that define patterns for detection and estima-
tion, action learning, or inductive methods of learning. The algorithms can be further
grouped into three distinct families of learning, known as supervised learning, unsu-
pervised learning and reinforcement learning [18]. The main objective of this paper is
to compare numerous forms of ML models that will be used, such as Aritifical Neural
Network (ANN), Least Square Support Vector Machine (LSSVM) and Extremely
Randomized Trees (Extra Trees), which are used to help create advanced gas hydrate
control techniques [12]. In this article, theory and method employed for different ML
models have been discussed. Afterwards, a comparison among the models has been
performed and the their effectiveness has been assesed and identified. Finally, on
these basis, some important conclusion have been drawn and discussed.

2 Methods and Theory

Authors like Rebai et al. [19], stated that machine learning techniques such as ANN,
LSSVM and Extra Trees are considered eligible to reproduce any form a function and
have been successful in predicting the required output of the gas hydrate formation
which is hydrate formation/deposition temperature (HDT/HFT). The input parame-
ters that has been used in this 3 ML models were concentration of the inhibitors, gas
composition and pressure. Each ML models theory and method will be explained
below in details.

2.1 Least Square Version of Support Vector Machine

(LSSVM)

LSSVM development is aimed at reducing the complexity of the Support Vector

Machine (SVM) and to eliminate complexity in the optimization process, linear
formulas are deciphered by linear encoding rather than solving quadratic encoding
equations to improve their speed. LSSVM algorithm’s output can be overcome by
fixed linear equality rather than inequality constraints
 [4].
 General
 formulation
  that
being used in approximating a given data set x1, y1 , x2, y2 , . . . , x N , y N into
138 S. D. Suresh et al.

Fig. 1 LSSVM model architecture [20]

consideration by employing following nonlinear function [14]:

f (x) = w, (x) + b (1)

where ., . represent dot product; (x) denotes the nonlinear function that performs
regression; b and w are bias terms and weight vector, respectively. The LSSVM model
structure can be seen in Fig. 1 where the K (xi , xj ) is defined as kernel function [20].

2.2 Extremely Randomized Trees (Extra Trees)

Tree models were identified by the assembly method named as Extra Trees, in which
this process produces multiple trees from the complete testing examples in terms
of regression or decision trees. Extra Trees is a top-down procedure where it forms
a joint of untrimmed choice or regression tress where the act of splitting nodes by
selecting action points randomly completes and uses the entire experiment model to
grow these trees, distinguishing it from other tree-based ensemble approaches [4].
The growth of the tree is in the top-down order where the numbering is from left
to right, as can be seen from Fig. 2. The value of a parameter is compared with the
Application of Machine Learning Models in Gas Hydrate Mitigation 139

Fig. 2 Study of Extra Trees framework [4]

threshold value of each internal node. If the shown inequality above each node is
satisfied, the path is either the left branch or the path is not the right branch [4].

2.3 Artificial Neural Network (ANN)

Artificial neural networks (ANN) are a method of calculation based on how the brain
conducts calculations [20, 21]. According to Samsudin et al. [20], this model has
the potential to carry out extensive input–output mapping and can create a network
that approximates nonlinear functionality. Figure 3 shows ANN feed-forward back-
ward propagation consists of three layers; data transmitted is the input (first layer),
interpreted data is hidden layer, and result provided is output (last layer) [20].
Output of the ANN assuming a linear output neuron j, a single hidden layer with
h sigmoid hidden nodes and the output variable (xt ) is given by [20]:
⎛ ⎞
h
 
xt = g ⎝ w j f s j + bk ⎠ (2)
j=1

where g(.) is the linear transfer function of the output neuron k and bk is its bias, wj
is the connection weights between hidden layers and output units, f (.) is the transfer
function of the hidden layer.
Below Fig. 4, shows the exact prediction of the distribution of temperature in the
gas-dominated system, with due regard to the exothermic effect of the composition
140 S. D. Suresh et al.

Fig. 3 Sample Structure of ANN [20]

Fig. 4 The interdependency

in the gas-dominated
environment between
hydrate formation, hydrate
deposition and temperature
variation

of hydrate and the resistive effect of the hydrate deposition layer on the rate of heat
transfer in the pipeline. The number on the Fig. 4 is decoded by (1) supplying hydrate
forming materials for deposition, (1’) reflecting the indirect effect by temperature,
pressure, and effective inner diameter variations. Whereas (2) signifies heat release
from hydrate formation by interaction of free water and gas; (2’) indicates the effect
of temperature differences on the hydrate formation sub-cooling. The heat tolerance
effect denoted by (3) of the hydrate layer on the heat transfer while (3’) indicates the
indirect effect influencing the rate of hydrate formation [13].
Application of Machine Learning Models in Gas Hydrate Mitigation 141

3 Results and Discussion

It is assumed that the accuracy and credibility of every statistical method is linked to
the universality and validity of the database used for its foundation. To compare the
results based on the several factors such as correlation coefficient, R2 , comparison of
the three ML models must be represented in terms of hydrate dissociation/ formation
temperature (HDT/HFT). All the data and information taken and referred from [4,
22] in order to obtain effective comparison and accuracy.
Based on article Yarveicy and Ghiasi [4], in the first step,1800 experimental data
points of different gas systems for liquid water-hydrate-vapor (LHV) or ice-hydrate-
vapor (IHV) balance were collected and divided into two classes called “training set”
and “test set” for Extra Trees and LSSVM models. 90% were used as training proce-
dures from the data point and the remaining 10% were used as testing procedures.
The primary goal is to approximate the HDT / HDFT of gases where electrolysis
and alcohol(s) are present in pure water or aqueous solutions, where it is corre-
lated as function of concentration of inhibitor(s) (C additive ), pressure (P), and the gas
composition (Zi ):

H D F T = f (Cadditive P, Z i ) (3)

The composition of gas (Zi ) consisting of C1 , C3 , i-C4 , i-C5 , n-C4 n-C6 , N2 ,

and CO2 in (mol %). Consequently, LSSVM model used radial basis function
(RBF) kernel function in its methodology with optimization of Coupled Simulating
Annealing (CSA). Two considerations were crucial to obtaining the HDT / HFT Extra
Tree algorithm, which was the sum of tress generated for calculating significance
and maximum depth of each tree using the test and error method. From trial and
error method, optimum number of tress obtained was 5 as well as maximum depth
for the Extra Trees algorithm changed from 3 to 40. By monitoring the error, best
performance can be seen at maximum depth value of 30 [4].
The objective of the ANN model was the same as LSSVM and Extra Trees, which
was to predict the temperature of the gas mixture (HFT / HDT), where 279 sweet
and sour gas experimental data points were collected. For the training procedure,
201 data points were used, 42 data points were used for testing set. The input was
a gas composition composed of H2 S, CO2 , and CH4 in moles percentage (mol%)
and pressure (P) in mPa, respectively [22]. For the three models [14], a comparative
study between the predicted hydrate formation temperature (HFT / HDT) values and
the experimental data for the train, validation and test data was developed. Below
seen Table 1, shows the results obtained from ANN, LSSVM and Extra Trees models
in terms of R2 and AARD in percentage (%).
Below mentioned are the formulas used to calculate the values placed in Table 1
which are coefficient of determination percent ((R 2 %), and average absolute relative
deviation percent (A A R D%) defined by Eqs. (3) and (4) in respective order, are
mathematical parameter used to estimate the HDT / HDF of gas hydrate as well as
142 S. D. Suresh et al.

Table 1 Error evaluation outcomes for the presented Extra Trees, LSSVM and ANN models
Parameter/model Extra trees LSSVM ANN
Train Test Total Train Test Total Train Test Total
R2 % 99.91 97.46 99.7 97.54 92.4 96.99 99.47 98.82 99.28
AARD% 0.01 0.35 0.04 0.28 0.61 0.32 0.1331 0.2137 0.1509

determining the reliability and efficiency of the built Extra Trees, LSSVM and ANN
models.
n
− ti )2
i (oi
R 2 % = 100 1 − n (4)
i (oi − average(ti ))
2

n
oi − ti
A A R D% = 100 /n (5)
i=1
ti

where n denotes the number of samples; oi and ti are the predictions of the model
and corresponding targets, respectively.
The error analysis results for the tabulated in Table 1 shows an excellent compli-
ance with the experimental hydrate deposition / forming temperature (HDT / HDF)
for the three distinct types of ML models that are acceptable outcomes [4]. Neverthe-
less, as the AARD percent is 0.04 relative to the other models with AARD percent of
0.32 and 0.1509 for LSSVM and ANN, respectively, the suggested tree-based model
has stronger prediction.

4 Conclusion

Interaction in this paper provides the significant comparison of ML based models

such as Extra Trees, LSSVM and ANN where their respective methodologies were
employed for phase equilibrium modelling of clathrate hydrates. From the coefficient
of determination percent, R2 is found to be greater than 96% for all three ML models.
Based on the analysis of the error study, it can be argued that the Extra Trees algorithm,
which was used for the first time in the supremacy of gas hydrate phenomenon
experimentation, is extremely capable of demonstrating the formation of gas hydrate
depending on the temperature, resulting in the Extra Trees providing more precision
than the LSSVM and ANN models in terms of AARD percent of 0.04, which is the
lowest compared to LSSVM and ANN models.

References

1. Bavoh CB, Lal B, Osei H, Sabil KM, Mukhtar H (2019) A review on the role of amino acids
in gas hydrate inhibition, CO2 capture and sequestration, and natural gas storage. J Nat Gas
Sci Eng 64:52–71
Application of Machine Learning Models in Gas Hydrate Mitigation 143

2. Bavoh CB, Partoon B, Lal B, Kok Keong L (2017) Methane hydrate-liquid-vapour-equilibrium

phase condition measurements in the presence of natural amino acids. J Nat Gas Sci Eng
37:425–434
3. Bomba J, Chin D, Kak A, Meng W (2018) Flow assurance engineering in deepwater offshore-
past, present, and future. In: Offshore technology conference
4. Ghiasi MM, Yarveicy H, Arabloo M, Mohammadi AH, Behbahani RM (2016) Modeling of
stability conditions of natural gas clathrate hydrates using least squares support vector machine
approach 223:1081–1092
5. Kazemeini M, Freidoonian F, Fattahi M (2012) Developing a mathematical model for hydrate
formation in a spray batch reactor 2:244–247
6. Khamehchi E, Shamohammadi E, Yousefi SH (2013) Predicting the hydrate formation
temperature by a new correlation and neural network
7. Khan MS, Bavoh CB, Partoon B, Nashed O, Lal B, Mellon NB (2018) Impacts of ammonium
based ionic liquids alkyl chain on thermodynamic hydrate inhibition for carbon dioxide rich
binary gas. J Mol Liq 261:283–290
8. Khan MS, Yaqub S, Manner N, Karthwathi NA, Qasim A, Mellon NB, Lal B (2018) Experi-
mental equipment validation for methane (CH4) and carbon dioxide (CO2) hydrates. IOP Conf
Ser Mater Sci Eng
9. Nashed O, Dadebayev D, Khan MS, Bavoh CB, Lal B, Shariff AM (2018) Experimental
and modelling studies on thermodynamic methane hydrate inhibition in the presence of ionic
liquids. J Mol Liq 249:886–891
10. Nasir Q, Lau KK, Lal B, Sabil KM (2014) Hydrate dissociation condition measurement of
CO2-rich mixed gas in the presence of methanol/ethylene glycol and mixed methanol/ethylene
glycol + electrolyte aqueous solution. J Chem Eng Data 59(11):3920–3926
11. Olajire A (2020) Flow assurance issues in deep-water gas well testing and mitigation strategies
with respect to gas hydrates deposition in flowlines—a review. J Mol Liq 114203
12. Partoon B, Sahith SJK, Lal B, Maulud ASB (2020) Gas hydrate models. In: Chemical additives
for gas hydrates. Springer, Cham, pp 67–85
13. Qasim A, Khan MS, Lal B, Abdullah MZ, Maulud AS, Simulation of hydrate phase boundary
for natural gas mixture with high CO2 content through simulation
14. Qasim A, Khan MS, Lal B, Shariff AM (2019) A perspective on dual purpose gas hydrate and
corrosion inhibitors for flow assurance. J Pet Sci Eng 183:106418
15. Qasim A, Lal B, Shariff AM, Ismail MC (2020) Role of ionic liquid-based multipurpose
gas hydrate and corrosion inhibitors in gas transmission pipeline. In; Nanotechnology-based
industrial applications of ionic liquids. Springer, pp 221–244
16. Qin H, Srivastava V, Wang H, Zerpa LE, Koh CA (2019) Machine learning models to predict
gas hydrate plugging risks using flowloop and field data. In: Offshore technology conference
17. Rafique D, Velasco L (2018) Machine learning for network automation: overview, architecture,
and applications [Invited Tutorial]. 10(10):D126–D143
18. Rebai N, Hadjadj A, Benmounah A, Berrouk AS, Boualleg SM (2019) Prediction of natural
gas hydrates formation using a combination of thermodynamic and neural network modeling.
182:106270
19. Sahith SAK, Pedapati SR, Lal B (2019) Application of artificial neural networks on
measurement of gas hydrates in pipelines
20. Samsudin R, Saad P, Shabri AJH, Sciences ES (2011) River flow time series using least squares
support vector machines 15(6)
21. Soroush E, Mesbah M, Shokrollahi A, Rozyn J, Lee M, Kashiwao T, Bahadori A (2015)
Evolving a robust modeling tool for prediction of natural gas hydrate formation conditions.
12:45–55
22. Yarveicy H, Ghiasi MM (2017) Modeling of gas hydrate phase equilibria: extremely
randomized trees and LSSVM approaches. J Mol Liq 243:533–541
Analysis of Signal Sensing with Adaptive
Threshold for Energy Detector
in Cognitive Radio Systems

Biji Rose and B. Aruna Devi

Abstract Overviewing the past the past period of 20 years, it is evident that in
certain geographical regions, the spectrum, lay unutilized whereas other areas face
spectrum crunch. To boost the spectral efficiency, unlicensed users (secondary users)
will be allowed to access unutilized parts of the spectrum and adapt their communi-
cations. It assists to utilize them while minimizing the interference on licensed users.
Such ability is the predominant feature of cognitive radio (CR) nodes, which needs,
algorithms and protocols for spectrum sensing, sharing, and management. The ratio-
nale is to facilitate optimization of the spectrum and make a conducive environment
between licensed and unlicensed users. The seamless method of spectrum sensing
is Energy Detection method, since it does not require to have priory knowledge of
Primary User (PU).In this paper, sequential sensing event at the Energy Detector is
used to decide on the optimal adaptive threshold value over Rayleigh fading channel.

Keywords Cognitive radio · Energy detector

1 Introduction

The year 2020 which is not yet over, went through health and economic crisis world-
wide due to Covid-19, faced lockdowns of months. All people during this pandemic
situation were connected and jointly fight the Covid-19 situation using digital tech-
nologies, which completely depends on mobile wireless network. At the outburst
of pandemic situation, an emergency meeting was called in starting week of April
and ITU emphasized on the importance of communication technologies in dissem-
inating timely critical information, supporting e-learning for more than 1.5 billion

B. Rose (B)
Department of IT, Dr. N.G.P. Institute of Technology, Coimbatore, Tamil Nadu, India
e-mail: [email protected]
B. A. Devi
Department of ECE, Dr. N.G.P. Institute of Technology, Coimbatore, Tamil Nadu, India

© The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 145
R. Ibrahim et al. (eds.), International Conference on Artificial Intelligence
for Smart Community, Lecture Notes in Electrical Engineering 758,
https://doi.org/10.1007/978-981-16-2183-3_13
146 B. Rose and B. A. Devi

students, training workers by digital means to improve productivity, and promoting

e-businesses [1].
The service mobile wireless network increased tremendously in 2020 due to
pandemic situation. Already in 2018 it was reported: the need of smartphones, tablets,
laptops, and other wireless devices that allow users to browse the Web, use email, and
download videos, multimedia, and applications, by 2023 will grow. Thus by 2023,
the average 5G connection speed will reach 575 Mbps, the average Wi-Fi speeds will
grow to 92 Mbps, and there will be 3.6 networked devices/connections per person
and nearly 10 devices and connections per household [2].
As a result, the growing demand for mobile wireless broadband is expected to
surpass the usable radio spectrum. Prior highlighting the concepts of the perfor-
mance analysis, standards of spectrum sensing include IEEE 802.22 WRAN (wire-
less regional area network) and its amendments, IEEE 802.11af for wireless LANs,
IEEE 1900.x series, and licensed shared access (LSA) for LTE mobile operators
[3]. Spectrum sensing technique achieves detection of channel state by (1) Energy
detection (ED), (2) Cyclostationary detection, (3) Matched filter-based detection, (4)
Waveform-based detection, (5) Covariance absolute value detection (CAV), and (6)
Eigen-values based detection (EVD). The spectrum sensing problem is solved as a
binary hypothesis test. The main attribute of interest is decision statistic, detection,
and false-alarm probabilities and the decision threshold. The Central Limit Theorem
(CLT) is used to achieve energy detection with prescribed performance level.
Studies of spectrum usage have shown that this spectrum allocation paradigm
often results in severe underutilization of the spectrum [4, 5]. In opportunistic or
dynamic spectrum access, a band of licensed spectrum can be used by unlicensed
users whenever it is not being used by licensed users. The licensed user is often
referred to as the primary user, whereas the unlicensed user is called the secondary
user. In addition, the secondary users must be capable of sensing the radio environ-
ment to determine the spectrum hole opportunities for dynamic spectrum access.
Such capabilities are realized in cognitive radio technologies [6].
In a cognitive radio system, a Markovian model of the primary user can be used for
state prediction, which in turn can prevent unintended collisions [7].A popular edge
detection technique uses the continuous wavelet transform (CWT) to decompose the
edge detector into multiple resolutions and multiplies the resolutions together, which
has a beneficial effect of reducing the noise [8]. While the novel Discrete Wavelet
Transform (DWT) based hybrid edge detection algorithm, where we use conven-
tional DWT technique in low noise variance scenario and DWT-Multiscale-Sum
(DWTMS) technique in high noise variance scenario depending on the Threshold
Value. The DWT Hybrid Edge Detection gives higher sensing accuracy, and more
noise reluctance for the secondary user in the cognitive radio system compared to
CWT [9].
Thus, an adaptive spectrum sensing algorithm in which the threshold is made
adaptive to the fading environment is the need of the hour [10].
Analysis of Signal Sensing with Adaptive … 147

2 Related Works

In the case of relatively low SNR,when the difference between noise and signal
energy is not distinct, single-threshold detection methods are easy to cause SUs to
perceive the existence of a PU incorrectly, which will reduce the detection probability
of the entire system. To improve the detection the single threshold is replaced over
the years by adaptive or dynamic threshold.
There are methods, named three-event ED, which makes decision considering the
event immediately before and the one immediately after current sensing event. The
primary users (PU) activity duty cycle value was exploited for tracking the changes
of the PU state [11–13].
Furthermore, an adaptive double threshold based on history energy detection is
proposed for Cooperative Spectrum Sensing. The weighting coefficient of thresholds
according to the SNR of all cognitive nodes is calculated during sensing period; thus
adjusting the upper and lower thresholds, if the average history energy is still between
two thresholds, the single-threshold method will be used for the end decision. Finally,
the fusion center aggregates the detection results of each node and obtains the final
cooperative conclusion through “or” criteria [14, 15].
However, the double threshold methods mentioned above are all no-decision
when the energy is between the high and low thresholds. The multilevel quanti-
zation method is adopted in the uncertain region of the traditional double-threshold
energy detection to solve the problem that the sensing information is ignored between
the two determined thresholds [16, 17].
In this method, an adaptive threshold method to overcome sensing failure at very
low SNR with uncertain noise power using a check parameter and double threshold
concept is proposed. It was noted that this method improved the detection probability
at a low SNR [18].
Nowadays, many Machine Learning algorithms are put forward for Optimization
of Adaptive Threshold in energy detection. The machine learning algorithm looks
for the optimum decision threshold, which is the most important parameter to decide
the presence or absence of the primary user, using historical detection data [19].

3 Methodology

Let the complex signal received at time t by the secondary user in the desired band
be [20]

n(t) : H0
y(t) = (1)
h(t)s(t) + n(t) : H1

where n(t) is an additive complex white Gaussian noise process, h(t) represents a
fading process (e.g., nonfading means h(t) = 1), and s(t) is a signal transmitted by
148 B. Rose and B. A. Devi

Fig. 1 Hypothesis model

the primary node [20]. The sensing decisions are made using N > > 1 samples of
y(t). A decision statistic T(y) is used for hypothesis testing, given by

T (y) < λaccept : H0
T (y) > λaccept : H1

where 0 < λ < ∞ is called the decision threshold. The reliability associated with the
decision rule (2) can be characterized by Probability of Detection Pd and Probability
of False Alarm Pf is shown in Fig. 1. Mathematically, it is a conditional probability
given as below:

Pd = Pr(T(y) > λ|H1 (2)

Pf = Pr (T(y) > λ|H0 (3)

The exact values of Pd and Pf depend on how T(y) is constructed using

received samples, channel estimates, propagation characteristics, the choice of the
threshold, and other information. The false-alarm and detection probabilities may be
approximated using CLT for large number of samples (N), as
 
λ − 2N
Pf = Q √ (4)
2 N
 
λ − 2N (1 + γ )
Pd = Q √ (5)
2(1 + γ ) N

where γ is the SNR and Q is the standard Gaussian upper tail probability function
∞ −x 2
given as Q(u) = 2π1
u exp 2
d x.
The probability of missed detection is given as Pm

Pm = 1 − Pd (6)
Analysis of Signal Sensing with Adaptive … 149

Therefore, the error decision probability Pe is written as

Pe = P f + Pm (7)

Substituting Eqs. (6) in (7), results in

Pe = P f + (1 − Pd ) (8)

From Eqs. (4) and (5) suitable sample size N can be calculated from Pf and Pd
as,
 2
N = γ −1 Q −1 P f − 1 + γ −1 Q −1 (Pd ) (9)

For constant detection probability (CDR) the normalized threshold λd (optimum

threshold) can be calculated as,
√  √ 
λd = 2 N Q −1 P f + N (10)

For constant false alarm rate (CFAR) the normalized threshold λf (optimum
threshold) can be calculated as,
√  √ 
λ f = 2(1 + γ ) N Q −1 P f + N (11)

Optimum Threshold is chosen to make Pd large (Pd ≥ 0.9) and Pf small (Pf ≤
0.1).CFAR and CDR is used for the selection of threshold [21]. It has been noted
that CFAR improves the throughput [22], but CFAR fails to give enough protec-
tion to the PU as compared to that of the CDR approach. To improvise the overall
throughput at low SNR, the combination of CFAR and CDR is utilized to select
the optimum threshold [23]. In this paper, considering all of the above aspects, the
optimal threshold is selected for ED which jointly achieve the targeted Pf and Pd at
low SNR.

3.1 Optimum Threshold Condition

The error decision probability Pe is given by

λ = (1 − δ)P f + δ(1 − Pd ) (12)

where δ (0 ≤ δ ≤ 1) is PU’s spectrum usage ratio and (1-δ) is PU’s channel is vacant &
noise occupied, (1-Pd ) gives probability of missed detection.
The adaptive threshold to minimize the error decision probability is
150 B. Rose and B. A. Devi
⎧   ⎫
⎪  (1−δ) ⎪
⎪
⎨  4(2 + γ )ln. (1 + γ ). δ ⎪ ⎬
(1 + γ ) 
λ = σn2 . . 1+ 1+ (13)
(2 + γ ) ⎪⎪ Nγ ⎪
⎪
⎩ ⎭

where σn2 = noise variance. Thus, optimum threshold setting should consider both
the probability of false alarm and the probability of mis-detection in order to benefit
both PU and SU.

4 Conclusion

The overall throughput at low SNR is improved by the combinational use of the
optimum threshold. Thus, the optimal threshold is selected for ED which jointly
achieve the targeted Pf and Pd at low SNR. While the SNR increases Pd is also
increases.

References

1. Special emergency session of the broadband commission pushes for action to extend internet
access and boost capacity to fight COVID19. https://www.itu.int/en/mediacentre/Pages/PR05-
2020 Broadband Commission-emergency-session-internet-COVID-19.aspx, online. Accessed
20 Apr 2020
2. Cisco visual networking index Global mobile data traffic forecast update, 2018–2023. https://
www.cisco.com/c/en/us/solutions/collateral/executive-perspectives/annual-internet-report/
white-paper-c11-741490.html
3. Gavrilovska L, Denkovski D, Rakovic V, Angjelichinoski M (2014) Medium access control
protocols in cognitive radio networks: overview and general classification. IEEE Commun
Surveys Tutor 16(4):2092–2124
4. FCC (2002) Spectrum policy task force. Technical report 02–135, Rep. ET Docket, Federal
Communications Commission
5. Shared Spectrum Company (2010) General survey of radio frequency bands: 30MHz to 3 GHz.
Technical report
6. Haykin S (2005) Cognitive radio: brain-empowered wireless communications. IEEE J Sel
Areas Commun 23(2):201–220
7. Liu Y, Jun LIANG, Xiao N, Hu Y, Hu M (2016) Dynamic double threshold energy detection
based on Markova model in cognitive radio. J Electron Inf Technol 38(10):2590–259
8. Tian Z, Giannakis GB (2006) A wavelet approach to wideband spectrum sensing for cognitive
radios. In: Proceedings of 1st international conference on cognitive radio oriented wireless
networks and communications (CROWNCOM), pp 1–5
9. Kumar A, Saha S (2019) Discrete wavelet transform based hybrid edge detection technique
for wideband spectrum sensing in CRNs. In: IEEE region 10 symposium (TENSYMP) pp
529–534
10. Verma P (2020) Adaptive threshold based energy detection over Rayleigh fading channel. Wirel
Person Commun 113(1):299–311
Analysis of Signal Sensing with Adaptive … 151

11. Vladeanu C, Nastase CV, Martian A (2016) Energy detection algorithm for spectrum sensing
using three consecutive sensing events. IEEE Wirel Commun Lett 5(3):284–287
12. Martian A, Al Sammarraie MJ, Vlădeanu C, Popescu DC (2020) Three-event energy detection
with adaptive threshold for spectrum sensing in cognitive radio systems. MDPI J Sens (Basel)
20(13):3614
13. Nasrallah A, Hamza A, Boukaba T, Baudoin G, Messani A (2018) Energy detection with
adaptive threshold for cognitive radio. In: IEEE international conference on communications
and electrical engineering, pp 1–5
14. Yu S, Liu J, Wang J, Ullah I (2020) Adaptive double-threshold cooperative spectrum sensing
algorithm based on history energy detection. Hindwai J Wirel Commun Mob Comput
15. Yang H, Fu Y, Lei J (2018) A new cooperative spectrum sensing algorithm based on double
threshold. In: IEEE 2nd advanced IMCEC conference, pp 588–592
16. Hu L, Cao N, Mao M, Liao H, Shi R, Wang S (2019) Dynamic adaptive double-threshold
cooperative spectrum sensing with multi-level quantization. In: IEEE 3rd advanced IMCEC
conference, pp 1381–1385
17. Sarala B, Devi SR, Sheela JJ (2020) Spectrum energy detection in cognitive radio networks
based on a novel adaptive threshold energy detection method. Elsevier J Comput Commun
152:1–7
18. Mahendru G, Shukla AK, Banerjee P, Patnaik LM (2019) Adaptive double threshold based
spectrum sensing to overcome sensing failure in presence of noise uncertainty. In: IEEE 6th
international conference on SPIN, pp 466–471
19. Kockaya K, Develi İ (2020) Spectrum sensing in cognitive radio networks: threshold
optimization and analysis: research square
20. Urkowitz H (1967) Energy detection of unknown deterministic signals. Proc IEEE 55(4):523–
531
21. Verma G, Sahu OP (2016) Intelligent selection of threshold in cognitive radio system.
Telecommun Syst 63(4):547–556
22. Koley S, Mirza V, Islam S, Mitra D (2015) Gradient based real-time spectrum sensing at low
SNR. IEEE Commun Lett 19(3):391–394
23. Verma G, Sahu OP (2016) Opportunistic selection of threshold in cognitive radio networks.
Wirel Person Commun 92(2):711–726
Collaborative Design in Concurrent
Engineering of Industry 4.0 Through
Virtual Reality Simulation in Achieving
Accelerated Time-To-Market During
COVID-19 (Coronavirus) Pandemic
Outbreak

A. F. Fudzin, M. Amin, and A. Mokhtar

Abstract In the era of industry 4.0, it is important for businesses to conduct design
and engineering activities via data exchange in cloud computing. Collaborative
design in concurrent engineering through virtual reality simulation using digital
mock up (DMU) presented in this paper is to construct a simulation in digital envi-
ronment in order to reduce development time, cost and improve quality of product
which will impact to accelerated time to market. However, due to current COVID-
19 (Coronavirus) pandemic outbreak, the activities could be interrupted. The virtual
reality design collaboration seems to be as a solution with the support of cloud storage
of product design and satisfy design issue in concurrent engineering. It is also over-
come the geographical limitation and social distancing constraint in “face to face”
collaboration. Thus, stakeholders’ activities are carried out as usual without inter-
ruption in the challenging period during the pandemic. Finally, virtual simulation of
collaborative design in concurrent engineering of previous related works conducted
by the author are briefly presented.

Keywords Industry 4.0 · Concurrent engineering · Collaborative engineering ·

Digital manufacturing · Simulation · COVID-19

1 Introduction

Competitive market has forced industry to improve new product development and
improve the speed of product launch through digitalization. The basic formulation
and solution of accelerated time to market in new product launch is a vision for the

A. F. Fudzin (B) · M. Amin · A. Mokhtar

Faculty of Mechanical, Universiti Teknologi PETRONAS, 32610 Seri Iskandar, Perak, Malaysia
e-mail: [email protected]
M. Amin
e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 153
R. Ibrahim et al. (eds.), International Conference on Artificial Intelligence
for Smart Community, Lecture Notes in Electrical Engineering 758,
https://doi.org/10.1007/978-981-16-2183-3_14
154 A. F. Fudzin et al.

realization of the product development digitalization process. In the global compe-

tition of new product introduction, the time required for the realization of products
as well as production is an absolute prerequisite in product development. For this
purpose, the virtual reality simulation of product development is an absolute require-
ment to achieve accelerated time to market and reduce cost [1]. On the other hand,
many considerations in product development is the involvement of various stake-
holder. Single stakeholder alone in development team will not be capable of meeting
the challenge. For this purpose, extensive cooperation among numerous stakeholder
will be necessary. Despite the fact that these issues have been brought to the attention
of many scholars and industrialists since the early stages of digitalization, the current
situation of global pandemic outbreaks has also come to global attention in recent
literature.
In current situation of COVID-19 (Coronavirus) global pandemic outbreak, the
face to face stakeholder collaboration is prohibited in many countries despite the
challenge of accelerated new product introduction [2]. In the event of catastrophe such
as COVID-19 pandemic, the industry was badly affected by forcing the workforce
to follow public health guidelines such as movement control order and work from
home (WFH) [3]. Therefore, it is essential for stakeholders or citizen to adhere to
the social distancing initiative by having virtual meeting from their location using
digital technology [4]. The launch of product need to be realized in timely manner
as planned ahead before the pandemic. Thus, this is the benefit of industry 4.0 by
having the cloud data storage which all the stakeholders be able to communicate
and retrieve the data for analysis. Subsequently, product development can still be
performing and interruption of time to market is less affected.
Nowadays, the digitalization reaching beyond boundary. Cloud-based digitaliza-
tion was introduced [5]. The emerging of technologies such as Industry 4.0 is known
as the fourth industrial revolution, which developed in Germany plays an impor-
tant major role in the design and development [6–8]. The ninth pillars of industry
4.0 which consists of big data, autonomous robot, simulation, system integration,
internet of things, cloud computation, additive manufacturing, augmented reality
and cyber security can help to achieve accelerated time to market [8, 9]. Virtual
simulation was used to perform an analysis such as clash interference check, tools
accessibility, workability, ergonomics and serviceability in the early stage of product
development. In the virtual simulation world, virtual commissioning may also be
carried out [10]. The advancement in new digital technology provides the advantage
of how industry reacts to the latest phenomena.
Before this pandemic shock all industrialist and engineering-related activities
normally getting in touch with stakeholders is a requisite. In order to have a common
emphasis on reducing product design and production time and costs and increasing
product quality, a lot of discussion with all stakeholders through effective collabora-
tion are needed [11]. As the purpose of the study in achieving uninterrupted product
launch and accelerated time to market in the current pandemic situation, virtual
people’s collaboration is highly essential to the success of concurrent engineering in
the Industry 4.0 platform. This work proposed a method to support the execution of
Collaborative Design in Concurrent Engineering … 155

collaborative design in concurrent engineering in industry especially in automotive

environment.

2 Collaborative Design in Concurrent Engineering

In collaborative design, a design communication and involvement of several stake-

holders such as marketing team, designer, process engineer, tools engineer and
sourcing team need to established in the early stage of development process. Design
coordination can be enhanced across multiple knowledge channels in multiple ways
compared to conventional settings as a one-way mechanism [5]. Nowadays, it is
also important for designers to co-develop components with various stakeholder in
different geographical locations [12]. Globalization of industry into several parts of
the world makes it difficult to gather all the stakeholder in one location, because of
geographical constraints. This initiative is a blessing in disguise by having virtual
collaborative design because of the COVID 19 (Coronavirus) pandemic problem.
Therefore, this phenomenon is a push factor for web-based virtual communication
method is taking place in rapid pace [13].
Concurrent engineering is defined an attempt to implement a parallelism in
product development. The realization of concurrent engineering through utilizing
DMU analysis in digital method helps to minimize the time taken from start of
project until start of series production. Figure 1 illustrate the DMU’s target process

Fig. 1 Product development process utilizing DMU

156 A. F. Fudzin et al.

concept of verification using digital product and verification using real objects [14].
The DMU process lead to significant reduction in the number of real object verifica-
tion and just limited for testing purpose only [15]. As the DMU’s goal are to reduce
development time, cost and improve quality, this will impact to accelerated time
to market of product. Furthermore, the number of design change due to assembly
problems and the amount of prototype validation needed for production during the
product validation stage will be reduced. The advantage of virtual simulation in the
design process is a key enabler for manufacturers to have a faster and more flexible
approach to solving complex issues in the collaborative design environment.

3 Methodology—Virtual Reality Simulation

In the early stage of design, face to face communication during collaboration design
of stakeholders will support faster feedback between designer and process engi-
neer. Figure 2 demonstrates the process of collaboration design in concurrent engi-
neering. The designer will issue design to process engineer for review process, so
that, product change request can immediately be considered for the change. At this
point of time, the industry 4.0 through cloud computing which shared platforms that
serve to multiple users (Shared Data Format) is very essential to support engineer
to review the same data and same format [9, 16]. Faster feedback can be achieved
from several parties in respond to the design issue aroused. It can be visualized as

Fig. 2 Collaboration design over the cloud storage

Collaborative Design in Concurrent Engineering … 157

designer and process engineer synchronized both the design and visualization data
formats on the same computer screen, eventhough they are from different part of the
world. This will help to reduce design-cycle time and cuts cost.
On design stage of product development, virtual computing analysis of manu-
facturing processes can be applied using 3D CAD/CAM modeling using CATIA
V5 (Dassault Systemes) [17]. It is a key enabler in the technology of Industry 4.0
and integrated process from Design for Manufacturing (DFM) [18] to visual and
augmented reality simulation can be achieved using seamless data transfer using
cloud computing [19]. The presented case study in automotive simulation analysis is
used in connection to 3D modeling. Three (3) stages of analysis investigated using
virtual simulation in development phase which are as followings;

3.1 Stage 1, Tools Accessibility Analysis

Tools accessibility analysis is an approach in assessing the tools require in production

assembly line [20]. This to ensure the correct tools to be purchased by the procurement
team. Thus, eliminate the use of incorrect tools for production.

3.2 Stage 2, Investigate Serviceability Using Human Manikin

Virtual Simulation

This investigation is to provide a visual simulation on how humans and objects

function together in term of serviceability. All potential problems of serviceability
are analyzed; human movements will work effectively with the proposed tools at an
early stage [21].

3.3 Stage 3, Tools Clash and Interference Check Using

Human Manikin Virtual Simulation

At this stage clash and interference check is using human manikin virtual simulation.
This investigation is to provide a visual simulation on how the inference of tools clash
with components during assembly and disassembly. It is very important for automo-
tive maintenance and service operation [15]. All possible clash and interference are
analyzed.
158 A. F. Fudzin et al.

4 Case Study—Virtual Reality Simulation

The described approach has been implemented in a practical application, which

supports the tools accessibility analysis, investigate serviceability using human
manikin virtual reality and tools clash and interference check using human manikin
virtual simulation.

4.1 Tools Accessibility Analysis

The case study of tools accessibility analysis is to verify assembly tools’ accessi-
bility applied during the assembly process during assembly process. Particularly, tool
placement constraints during tool applications are verified into the assembly product.
Tools accessibility analysis is an efficient approach in assessing the tools require in
production assembly line. Prior to the purchase of tools by the procurement team,
this activity may eliminate the use of incorrect tools for production. The analysis
utilized CAD data which retrieved from cloud data storage. Further, DMU CAD
models of the assembly product are imported into a virtual reality visualization tools
accessibility analysis.
The designer provided the components design to the cloud data storage and process
engineer as well as tool engineer have to analyze the tool accessibility to ensure that
the assembly process does not pose accessibility related problems. Prompt response to
designer in the case of accessibility issue can be rectified by the designer immediately.
In this case study, a virtual reality simulation is being utilized for the analysis as shown
in Fig. 3. This approach would detect any problem earlier in design phase.

Fig. 3 Virtual simulation of tools accessibility

Collaborative Design in Concurrent Engineering … 159

Fig. 4 Virtual simulation of human hand movements position 1 to positon 2 (Diassembly process)

4.2 Investigate Serviceability Using Human Manikin Virtual

Simulation

This investigation is adopted DMU is to provide a virtual simulation on how humans

and objects function together in term of serviceability. All potential difficulties during
serviceability are analyzed. Such a criteria during the procedure of disassembly
operation sequence is essential. Figure 4 shows virtually how human hand with tool
movements from position 1 to position 2 in disassembly process of oil filter from the
engine. The movement for disassembly of the oil filter can be preliminary identified so
it can be easily disassembly using dedicated tool. Others issues such as difficulties,
interference and clearance analysis with tools, human and product is analyzed. A
suitable tools need to be purchased, and in the case unavailability of tools, special
tool need to designed to ease the serviceability process.

4.3 Tools Clash and Interference Check Using Human

Manikin Virtual Simulation

At this stage clash and interference check is using human manikin virtual simulation.
The verification of the reach-ability of tools to perform a task is very crucial. This
investigation is to provide a visual simulation on how the reach-ability and inference
of tools clash with components. In the design phase all possible class and interference
are analyzed virtually. In Fig. 5, tool clash with the surrounding can be identified.
Feedback from process engineer to designer can immediately improve in the first
review session. The improvement of product design is validated in the second review
before proceed to the next phase. In order to support accelerated time to market, this
rapid and accurate assessment is needed.
Another example of reach-ability is the interference analysis for disassembly of oil
filter from the engine. As explained in previous section the tool used to disassembly
by turning the tool from position 1 to position 2. In continuation of that, human
hand is used to completely disassemble the oil filter. Figure 6 shows the virtual
simulation interference study during the disassembly of oil filter using hand. During
160 A. F. Fudzin et al.

Fig. 5 Virtual clash tool with product

Fig. 6 Virtual simulation interference study

the disassembly proceed we found the interference occurred between hand’s manikin
and sub frame. The red circle in the diagram clearly visualize the interference. With
help of simulation software this occurrence can immediately feedback to designer
in the form of product change request. It is proven that early detection during DMU
virtual simulation check will eliminate any cost during prototype stage. Figure 7
shows the proven real serviceability check where the interference was eliminated in
design phase.

5 The Role of Industry 4.0 in COVID-19 Pandemic

In automotive manufacturing environment, activities such as simultaneous engi-

neering, concurrent engineering and process engineering design, the role of Industry
4.0 is an important factor in fostering accelerated time to market. The deployment
of Industry 4.0 platform in the COVID-19 pandemic outbreak, helps the engineer
Collaborative Design in Concurrent Engineering … 161

Fig. 7 Real serviceability check

to continue moving forward using virtual reality simulation supported the design
review activities through the data exchange in cloud computing. It is a good tool
in order to increase social distancing and to adhere to the restriction on face to face
collaboration beside making simulation in economical ways. Based on the case study
presented in the previous section, stakeholders have been shown to be able to work
in same space of virtual environment without face to face collaboration as new norm
in working culture.
Collaborative design in concurrent engineering using DMU approach is part of
successful factor in virtual reality simulation where product introduction is unin-
terrupted due to pandemic. In the area of automotive design, the such operation
presented in the case study, the operations were checked on a physical vehicle must
now be checked by means virtual verification tools. The development of the virtual
verification is opens new approach of the analyses performed on the DMU. The
application not only allows designers to simulate the tools accessibility, service-
ability check, tools clash and interference check, but also offers a set of tools useful
for analyzing the DMU to detect potential modelling errors which could jeopardize
product functionality.
As virtual reality simulation has been presented during product development cycle
of the automotive industry, which it is known as an important tool in virtual veri-
fication. The advantage of virtual verification in development design phase, cost
reduction is achievable due to reduce and eliminate most of physical prototype. The
availability of Industry 4.0 as a key for future business is leading to these two goals,
as previously addressed in achieving uninterrupted product launch and accelerated
time to market in the current pandemic situation.

6 Conclusions

Today, automotive industries are required to transfer most of the activities in virtual
environment to meet the product launch time line in competitive market. In the avail-
ability of industry 4.0 platform, it is an important for businesses to conduct design
and engineering activities of the data exchange in cloud computing. The deployment
162 A. F. Fudzin et al.

of Industry 4.0 pillars in the COVID-19 pandemic outbreak, helps the engineer to
continue moving forward using virtual reality simulation supported the design review
activities through the data exchange in cloud computing. The Internet is becoming an
important element for concurrent engineering applications. The presented approach
supported a collaborative design in concurrent engineering through the virtual reality
simulation using DMU to construct a simulation in the digital environment in order
to reduce development time, cost and improve quality, which will impact to accel-
erated time to market of product. The stakeholders’ activities are carried out as
usual without interruption in the challenging period during the COVID-19 pandemic
outbreak. In this way, the virtual simulation design including cloud storage supports
product design and satisfy design issue in concurrent engineering. It is also over-
come the geographical limitation and social distancing constraint in “face to face”
collaboration in order to avoid possible infections among stakeholders. Therefore,
web-based virtual communication method is taking place in rapid pace. Finally, it
should be mentioned that the virtual simulation of collaborative design in concurrent
engineering of the previous related works has been implemented into the develop-
ment process of automotive manufacturer and finds its application in design phase.
In conclusion, with the presence of the latest Industry 4.0 application, there is no
significant impact on the disruption due to the COVID-19 pandemic in the virtual
reality simulation phase.

Acknowledgements The author would like to acknowledge Universiti Teknologi PETRONAS

support for this study.

References

1. Fragner A, Kreis A, Hirz M (2020) Virtual tools to support design and production engineering:
early detection of stone chips to optimize production processes. In 2020 IEEE 7th international
conference on industrial engineering and applications (ICIEA), 2020, pp 399–403
2. Avdiu B, Nayyar G (2020) When face-to-face interactions become an occupational hazard:
jobs in the time of COVID-19, ed: The World Bank
3. del Rio-Chanona RM, Mealy P, Pichler A, Lafond F, Farmer D (2020) Supply and demand
shocks in the COVID-19 pandemic: An industry and occupation perspective. arXiv preprint
arXiv:2004.06759, 2020
4. Katapally TR (202) A global digital citizen science policy to tackle pandemics like COVID-19.
J Med Int Res 22:e19357
5. Wu D, Rosen DW, Wang L, Schaefer D (2015) Cloud-based design and manufacturing: a new
paradigm in digital manufacturing and design innovation. Comput Aided Des 59:1–14
6. Zhou J, Li P, Zhou Y, Wang B, Zang J, Meng L (2018) Toward new-generation intelligent
manufacturing. Engineering 4:11–20
7. Xu LD, Xu EL, Li L (2018) Industry 4.0: state of the art and future trends. Int J Prod Res
56:2941–2962
8. Hermann M, Pentek T, Otto B (2016) Design principles for industrie 4.0 scenarios. In: 2016
49th Hawaii international conference on system sciences (HICSS), 2016, pp 3928–3937
9. Erboz G (2017) How to define industry 4. 0: the main pillars of industry 4. 0. In: 7th international
conference on management (ICoM 2017), At Nitra, Slovakia, 2017, pp 1–2
Collaborative Design in Concurrent Engineering … 163

10. Ružarovský R, Holubek R, Sobrino DRD, Janíček M (2018) The simulation of conveyor control
system using the virtual commissioning and virtual reality. Adv Sci Technol Res J 12
11. Vila C, Ugarte D, Ríos J, Abellán J (2017) Project-based collaborative engineering learning to
develop Industry 4.0 skills within a PLM framework. Procedia Manuf 13:1269–1276
12. Fuh JY, Li W (2005) Advances in collaborative CAD: the-state-of-the art. Comput Aided Des
37:571–581
13. Sułkowski Ł (2020) Covid-19 pandemic; recession, virtual revolution leading to de-
globalization? J Intercult Manage 12:1–11
14. Döllner G, Kellner P, Tegel O (2000) Digital mock-up and rapid prototyping in automotive
product development. J Integr Des Process Sci 4:55–66
15. Gynn M, Steele J (2015) Virtual automotive maintenance and service confirmation. In: SAE
technical paper 0148–7191
16. Zheng P, Sang Z, Zhong RY, Liu Y, Liu C, Mubarok K et al (2018) Smart manufacturing systems
for Industry 4.0: conceptual framework, scenarios, and future perspectives. Front Mech Eng
13:137–150
17. Ghionea IG, Devedžić G, Ćuković S (2015) Parametric modeling of surfaces using CATIA v5
environment. In: Applied mechanics and materials, 2015, pp 93–98
18. Myung S, Song K, Lee J (2002) Integration of DFM and virtual NC manufacturing process.
In: CIRP ISMS, pp 175–180
19. Posada J, Toro C, Barandiaran I, Oyarzun D, Stricker D, de Amicis R et al (2015) Visual
computing as a key enabling technology for industrie 4.0 and industrial internet. IEEE Comput
Graphics Appl 35:26–40
20. Kang X, Peng Q (2007) Analysis of tool accessibility in fixture setup planning. In: Interna-
tional design engineering technical conferences and computers and information in engineering
conference, 2007, pp 1007–1016
21. Bruno M, Milite A, Monacelli G, Pina M, Sessa F (2006) Serviceability analyses in virtual envi-
ronment for the automotive industry. In: DS 36: proceedings design 2006, the 9th international
design conference, Dubrovnik, Croatia, 2006
Stability Analysis of Semi-Markovian
Discrete-Time Neural Networks
with Time-Varying Leakage Delays

K. Maheswari, S. N. Shivapriya, and C. Ramkumar

Abstract This paper studies the stabilization of discrete-time semi-Markovian

neural networks with coupling time varying delays with unreliable communica-
tion links with sensor non-linearity in communication channels. There occurs the
time varying leakage delays with random parameters which obeys the uncorre-
lated Bernoulli-distributed sequences. By the proper construction of the Lyapunov
Krasovskii functional (LKF), sufficient conditions for stability analysis has been
attained that are derived by the means of linear matrix inequalities (LMI’s). Numerical
example with simulation is illustrated for the effectiveness of the proposed method.

Keywords Discrete-time NNs · Leakage delays · Stability analysis ·

Semi-Markovian neural networks

1 Introduction

Devoting much attention to the study of neural networks (NNs) in last few years
has tremendously impacted the application prospect in the real world scenario. NN’s
are extensively applied in varied areas of research in large data and artificial intelli-
gence which includes applications in pattern recognition, robotics, optimization etc.
Significant number of researches has also been carried out in dynamic analysis of
various NNs. Stability analysis via state estimation procedure is one of the hottest
research topic in the current trends of digital technology using discrete-time NNs.
In the areas of control and signal processing application problem procedures,
discrete-time NNs’ play a vital role along with time delays. There are a lot of

K. Maheswari (B)
Department of Mathematics, Kumaraguru College of Technology, Coimbatore, India
S. N. Shivapriya
Department of Electronics and Communications Engineering, Kumaraguru College of
Technology, Coimbatore, India
C. Ramkumar
Department of Biomedical Engineering, Dr. N. G. P. Institue of Technology, Coimbatore, India

© The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 165
R. Ibrahim et al. (eds.), International Conference on Artificial Intelligence
for Smart Community, Lecture Notes in Electrical Engineering 758,
https://doi.org/10.1007/978-981-16-2183-3_15
166 K. Maheswari et al.

approaches for the state estimation procedure in discrete time NNs that holds its
consideration with time delays several times and leakage delays at fewer number
of times. These time delays in the state estimation problems has drawn increasing
research attention in today’s digital world. This state estimation problem along with
Markovian jump parameters in discrete-time NN’s has been discussed in various
dynamical problems and the references therein [1–3]. Due to the strong modeling
ability of Markovian jump neural networks (MJNNs), greater attention has been
drawn in the field of hybrid dynamical systems like in the case of biology, medicine,
industry production, in aerospace Engineering.
But as is well known, in MJNNs, the transition matrix is converted into an invariant
matrix caused due to exponential distribution and obeys the stochastic process and is
not relevant with the past modes. This is one of the disadvantages of MJNNs. There-
fore, in order to overcome this shortcoming, semi-Markovian Jump Neural Networks
(SMJNNs) has been used where the probability density function matrix takes the key
role. Some of the recent researches have emphasized the synchronization controller
using SMJNNs given in [4] and the robust stochastic stability conditions have also
been incorporated in [5]. Inspired by all these works, in this paper, SMJNNs has been
applied to the model that has time varying coupling delays with randomly occur-
ring uncertainties [6]. By employing Lyapunov Karsovskii functional (LKF) with
different decision variables, sufficient conditions to attain the stability is achieved.
On the basis of the these discussions, the aim of this paper is to investigate the
asymptoticity for SMJNNs with delays that are mixed and time varying in nature.
By constructing suitable LKF’s and proceeding with the LMI technique, we derive
the feasibility of the stability conditions by the addressed NNs.
Notations:
Throughout the paper T and −1 represent the transpose and the inverse of the matrix
respectively. The n-dimensional Euclidean space is given by R n for the set of m x
n matrices in R mxn and λmax(min) (A) is the corresponding eigenvalue of the spectral
norm matrix H.

2 Formulation of the Problem


For  being the space in the probability P measured on F in the space is given by

(, F, P). Let (r(t), t > 0) be the semi-Markov process in the discrete-time state given
by S = {1, 2,…, N} with the generator  ∼ = γi j , i, j ∈ S

γi j (ĥ)ĥ + o(ĥ), i = j
Pr ob{r (t + ĥ) = j/r (t) = i} =
1 + γi j (ĥ)ĥ + o(ĥ), i = j

With lim o(ĥ)

ĥ
= 0 with γi j (ĥ) ≥ 0 for i = j when it moves from i at time t to
ĥ→0
mode j at time c with
Stability Analysis of Semi-Markovian … 167

γii (ĥ) = − γi j (ĥ) (1)
j∈S, j=i

Discrete-time delayed NNs with time-varying leakage delay as follows:


ς (t + 1) = B(ς (t − ψ(t))) + x(t) + C f (ς (t)) + D g(ς (t − d(t))) + J 

(2)

where ς (t) = [ς1 (t), ς2 (t), . . . , ςn (t)]T ∈ R n is the state vector and the activa-
   
tion functions are given as f (◦) = [ f (ς1 (t)), f (ς2 (t)), . . . , f (ςn (t))]T ∈ R n ,
   
g (ς (t)) = [g(ς1 (t)), g(ς2 (t)), . . . , g(ςn (t))]T ∈ R n and J  = [J1 , J2 , . . . , Jn ]T
is the input external vector. Each of ψ(t) in the system represents the delay due to
leakage which satisfies 0 < ψm ≤ ψ(t) ≤ ψ M where ψm is the lower bound and
ψ M is the upper bound of ψ(t) and d(t) represents the delay in the transmission
satisfying 0 < dm ≤ d(t) ≤ d M with the positive integers dm representing the lower
bound and d M representing the upper bound of d(t).

Assumption 1
 
For any c1 , c2 ∈ R, c1 = c2 , the activation functions f i (◦) and g i (◦) satisfy,

 
f (c1 ) − f i (c2 )
Mi− ≤ i ≤ Mi+ ,
c1 − c2
 
g (c1 ) − g i (c2 )
Ni− ≤ i ≤ Ni+ , i = 1, 2, . . . , n
c1 − c2

where Mi− , Mi+ , Ni− , Ni+ are the known constants.

Assuming ς ∗ = [ς1∗ , ς2∗ , . . . , ςn∗ ]T equation in (2) and letting


yi (t) = ςi (t) − ςi∗ , f i yi (t) = f i (ςi (t))

− f i (ςi∗ , ), gi (ςi (t − d(t)))
= g i (ςi (t − d(t))) − g i (ςi∗ )
 

Then, the NN system Eq. (1) can be written as,

y(t + 1) = B(y(t − ψ(t))) + C f (ς (t)) + Dg(ς (t − d(t))) (3)

where y(t) = [y1 (t), y2 (t), . . . , yn (t)]T ,

y(t − ψ(t)) = [y1 (t − ψ(t)), y2 (t − ψ(t)), . . . , yn (t − ψ(t))]T ,

168 K. Maheswari et al.

f (y(t)) = [ f (y1 (t)), f (y2 (t)), . . . , f (yn (t))]T ,

g(y(t − ψ(t))) = [g(y1 (t − ψ(t))), g
(y2 (t − ψ(t))), . . . , g(yn (t − ψ(t)))]T

By the above assumption given, it can be verified that for every f i (◦), gi (◦), i =
M − ≤ ( f i (c1 ) − f i (c2 ))/(c1 − c2 ) ≤ Mi+ ,
1, 2, . . . , n satisfy i− for any c1 = c2 and
Ni ≤ (gi (c1 ) − gi (c2 ))/(c1 − c2 ) ≤ Ni+
f i (0) = gi (0) = 0.
The model is associated with the initial condition as

y(c) ≤ θ (c), c = −δ, −δ + 1, . . . , 0 (4)

where δ = max{ψ M , d M }.

Lemma 1 In an open set A of Rm , let f 1 , f 2 , . . . , f N : R m → R then f i over D

satisfies.
 1  
min
 f i (l) = f i (l) + max gi, j (l) (5)
{βi |βi >0, βi =1}
i
βi i
gi, j (l)
i= j

given by
 
f i (l) gi, j (l)
g j,i (l) ≡ gi, j (l), ≥ 0. (6)
gi, j (l) f i j l)

Lemma 2 For any matrices ℘1 , ℘2 and ℘3 where ℘1T = ℘1 and ℘2T = ℘2 > 0, then
 
℘1 ℘3T
℘1 + ℘3T ℘2−1 ℘3 < 0 with < 0.
∗ −℘2

3 Main Results

In this section, we apply the semi-Markovian and investigate the delay dependent
sufficient conditions of LKF.

Theorem 3.1 The NNs system (3) is globally asymptotically stable, under the
assumption (1), if there exists the symmetric matrices Pi > 0, i = 1, . . . , 6, Q r >
0, r = 1, . . . , 4, R1 > 0, R2 > 0, T j > 0 with the diagonal matrices 1 > 0, 2 >
0 and matrices U j , j = 1 to 4 with weights W0 , W1 of appropriate dimensions, such
that following LMI holds with.
Stability Analysis of Semi-Markovian … 169
 
T1 T2
>0 (7)
T3 T4


where P̃i = γi j (h)P j ,
j∈S

1 = ẽ8 P̃1 ẽ8 + 2ẽ8 P̃1 ẽ1 + 2ẽ8 P̃3 (ẽ5 − ẽ7 )
+ 2ẽ8 P̃2 (ẽ2 − ẽ4 ) + 2ẽ8 P̃3 (ẽ11 − ẽ12 )
+ 2ẽ1 P̃2 (ẽ2 − ẽ4 ) + 2ẽ1 P̃3 (ẽ5 − ẽ7 )
+ (ẽ2 − ẽ4 )T P̃4 (ẽ2 − ẽ4 ) + 2(ẽ2 − ẽ4 )T 2ẽ1 P̃2 (ẽ2 − ẽ4 )

2(ẽ2 − ẽ4 )T P̃5 (ẽ5 − ẽ7 ) + 2(ẽ2 − ẽ4 )T P̃5 (ẽ11 − ẽ12 )
+ 2(ẽ5 − ẽ7 )T P̃6 (ẽ5 − ẽ7 )2(ẽ5 − ẽ7 )T P̃6 (ẽ11 − ẽ12 ).

2 = (ψ1 + 1)ẽ1T Q 1 ẽ1 − ẽ3T Q 1 ẽ3 , 3

= (d1 + 1)ẽ1T Q 2 ẽ1 − ẽ6T Q 2 ẽ6 ,
4 = d12 ẽ1T R1 ẽ1 + d1 ẽ5T U4 ẽ5 + d1 ẽ6T (U3 − U4 )ẽ6
− d1 ẽ7T U3 ẽ7 + d12 ẽ8T R2 ẽ8 ,
5 = −(ẽ11
T
R1 ẽ11 + 2ẽ11
T
U3 (ẽ6 − ẽ7 ))
+ (ẽ6 − ẽ7 )T (R2 + U3 )(ẽ6 − ẽ7 ) + ẽ12
T
R1 ẽ12 )
+ 2ẽ12
T
U4 (ẽ5 − ẽ6 ) + (ẽ5 − ẽ6 )(R2 + U4 )(ẽ5 − ẽ6 )),
6 = −ẽ1T (M1 1 )ẽ6 + 2ẽ1T (M2 1 )e9

− ẽ9T 1 ẽ9 , 7 = −ẽ6 (N1 2 )ẽ6

T

+ 2ẽ6T (N2 2 )e10 − ẽ10 T

2 ẽ10 , 8

= 2(ẽ8 W0T + ẽ1 W1T )(B ẽ3 + C ẽ9 + D ẽ10 − ẽ8T − ẽ1 ).

Proof The LKF is.

⎡ ⎤ ⎡ ⎤
y(t) y(t)
⎢ t−ψ ⎥⎡ P P P ⎤⎢ t−ψ ⎥
⎢  m −1
⎥ 1 2 3 ⎢  m −1
y(m) ⎥
⎢ y(m) ⎥⎣ ⎦ ⎢ ⎥
Vi (t) = ⎢ m=t−ψ M ⎥ ∗ P4 P5 ⎢ m=t−ψ M ⎥ (8)
⎢ t−dm −1 ⎥ ⎢ t−dm −1 ⎥
⎣  ⎦ ∗ ∗ P6 ⎣  ⎦
y(m) y(m)
m=t−d M m=t−d M

−ψm

t−1  
t−1
V2 (t) = y T (s)Q 1 y(s) + y T (s)Q 2 y(s) (9)
s=t−ψ(t) j=−ψ M +1 s=t+ j
170 K. Maheswari et al.


t−1 −dm
 
t−1
V3 (t) = y (s)Q 3 y(s) +
T
y T (s)Q 4 y(s) (10)
s=t−d(t) j=−d M +1 s=t+ j

−dm
 t−1 
 T   
y(s) R1 0 y(s)
V4 (t) = d1 (11)
η(s) 0 R2 η(s)
j=−d M +1 s=t+ j

where η(s) = y(s + 1) − y(s), d1 = d M − dm .

Calculating the difference of Va (t), a = 1, 2, 3, 4,

E{V1 (t)} = E{V (t + 1) − V (t)}

⎡ ⎤ ⎡ ⎤
η(t) + y(t) η(t) + y(t)
⎢ t−ψ ⎡ ⎤
⎥ P̃ P̃ P̃ ⎢ t−ψ ⎥
⎢  m −1
(η(m) + y(m)) ⎥ 1 2 3 ⎢
m −1
(η(m) + y(m)) ⎥
⎢ ⎥⎣ ⎢ ⎥
V1 (t) = ⎢ m=t−ψ M ⎥ ∗ P̃4 P̃5 ⎦⎢ m=t−ψ M ⎥
⎢ t−dm −1 ⎥ ⎢ t−dm −1 ⎥
⎣  ⎦ ∗ ∗ P̃6 ⎣  ⎦
(η(m) + y(m)) (η(m) + y(m))
m=t−d M m=t−d M
⎡ ⎤ ⎤⎡
y(t) y(t)
⎢ t−ψ ⎥⎡ P̃ P̃ P̃ ⎤ ⎢ t−ψ ⎥
⎢  m −1
⎥ 1 2 3 ⎢  m −1
y(m) ⎥
⎢ y(m) ⎥⎣ ⎦ ⎢ ⎥
− ⎢ m=t−ψ M ⎥ ∗ P̃4 P̃5 ⎢ m=t−ψ M ⎥,
⎢ t−dm −1 ⎥ ⎢ t−dm −1 ⎥
⎣  ⎦ ∗ ∗ P̃6 ⎣  ⎦
y(m)) y(m))
m=t−d M m=t−d M
⎡ ⎤ ⎡ ⎤
η(tˆ) η(tˆ)
⎢ y(tˆ) ⎥ ⎢ y(tˆ) ⎥
⎢ ⎥ ⎢ ⎥
⎢ t−ψ ⎥⎡ ⎢ ⎥
⎢  m −1
η(m) ⎥
⎤⎢ t−ψm −1
η(m) ⎥
⎢ ⎥ P̃1 P̃1 P̃2 P̃2 P̃3 P̃3 ⎢ ⎥
⎢ m=t−ψ M ⎥⎢ ⎥ ⎢ ⎥
⎢ t−ψ −1 ⎥⎢ P̃1 0 P̃2 0 P̃3 ⎢ m=t−ψ
0 ⎥⎢ t−ψ
M
⎥
⎢  m ⎥⎢ ⎥⎢ m −1 ⎥
⎢
V1 (t) = ⎢ y(m) ⎥ y(m) ⎥
⎥⎢ P̃2T P̃2T P̃4 P̃4 P̃5 P̃5 ⎥⎢ ⎥
⎢ m=t−ψ M ⎥⎢ ⎥ m=t−ψ M
P̃6 ⎦⎢ ⎥
⎥⎣ P̃3
T
⎢ t−d
m −1
P̃3T P̃5T P̃5T P̃6 ⎢ t−d
m −1 ⎥
⎢ ⎥ P̃ T ⎢ ⎥
⎢ η(m) ⎥ 3 0 P̃5T 0 P̃6 0 ⎢ η(m) ⎥
⎢ m=t−d M ⎥ ⎢ m=t−d M ⎥
⎢ t−d −1 ⎥ ⎢ t−d −1 ⎥
⎣  m ⎦ ⎣  m ⎦
y(m) y(m)
m=t−d M m=t−d M
Stability Analysis of Semi-Markovian … 171
⎡ ⎤
η(tˆ)
⎢ y(tˆ) ⎥
⎢ ⎥
⎢ t−ψ ⎥⎡
⎢  m −1
η(m) ⎥
⎤
⎢ ⎥ P̃1 P̃1 P̃2 P̃2 P̃3 P̃3
⎢ m=t−ψ M ⎥⎢
⎢ t−ψ −1
⎢ 
⎥⎢ P̃1
⎥⎢ 0 P̃2 0 P̃3 0⎥ ⎥
⎢
m
⎥
V1 (t) = ⎢ y(m) ⎥
⎥⎢ P̃2T P̃2T P̃4 P̃4 P̃5 P̃5 ⎥
⎢ m=t−ψ M ⎥⎢ ⎥
⎥⎣ P̃3 P̃6 ⎦
T
⎢ t−d
m −1
P̃3T P̃5T P̃5T P̃6
⎢ ⎥ P̃ T
⎢ η(m) ⎥ 3 0 P̃5T 0 P̃6 0
⎢ m=t−d M ⎥
⎢ t−d −1 ⎥
⎣  m ⎦
y(m)
m=t−d M
⎡ ⎤
η(tˆ)
⎢ y(tˆ) ⎥
⎢ ⎥
⎢ t−ψ ⎥
⎢  η(m) ⎥
m −1
⎢ ⎥
⎢ m=t−ψ M ⎥
⎢ t−ψ −1 ⎥
⎢  m ⎥
⎢ y(m) ⎥ = ϕ T (t)1 ϕ(t) (12)
⎢ m=t−ψ ⎥
⎢ ⎥
⎢ t−d ⎥
M
m −1
⎢ ⎥
⎢ η(m) ⎥
⎢ m=t−d M ⎥
⎢ t−d −1 ⎥
⎣  m ⎦
y(m)
m=t−d M

V2 (t) = (ψ1 + 1) ŷ T (t)Q 1 ŷ(t) − ŷ T (t − ψ(t))Q 2 ŷ T (t − ψ(t))

= T (t)2 (t) (13)

V3 (t) = (d1 + 1) ŷ T (t)Q 3 ŷ(t) − ŷ T (t − d(t))Q 4 ŷ T (t − ψ(t))

= T (t)2 (t) (14)

 T  
R1 0
ŷ(t) ŷ(t)
V4 (t) = d12
η̂(t)
0 R2 η̂(t)
 y(t) T  R1 0  y(t) 
t−dm −1 
− d1 (15)
η(t) 0 R2 η(t)
s=t−d M

where ψ1 = ψ M − ψm and d1 = d M − dm .
Also, by zero initial conditions, it is seen that,
172 K. Maheswari et al.

d1 y T (t − d(t))U1 y T (t − d(t)) − y T (t − d M )U1 y T (t − d M )

t−d(t−1
− η T (s)U1 (η(s) + 2y(t)) = 0,
s=t−d M

d1 ŷ T (t − dm )U2 ŷ T (t − dm ) − ŷ T (t − d(t))U2 ŷ T (t − d(t))

t−dm −1

− η T (s)U2 (η(s) + 2y(t)) = 0 (16)

s=t−d(t)

Also,

T (t)
⎡ T ⎤
y (t), y T (t − ψm ), y T (t − ψ(t)), y T (t − ψ M ), y T (t − dm ), y T (t − d(t)), y T (t − d M )
⎢ m −1 m −1
⎥
=⎢ ⎣ η(t) f T (y(t)) g T (x(t − d(t))
t−d t−d ⎥,
⎦
η(s) y(s)
s=t−d M s=t−d M

By differencing the above inequalities based on the expectation, we get,

 T  
R1 R2
ŷ(t) ŷ(t)
E{V4 (t)} = d12 {
R3 R4 η(t)
η(t)

t−d(t)−1  T   
ŷ(t) R1 R2 ŷ(t)
− d1
η(t) R3 R4 η(t)
s=t−d M


t−dm −1  T   
ŷ(t) ŷ(t) R1 R2
− d1
η(t) η(t) R3 R4
s=t−d(t)

+ d1 ŷ T (t − d(k))U1 ŷ T (t − d(k))
− ŷ T (t − d M )U1 ŷ T (t − d M )
+ ŷ T (t − dm )U2 ŷ(t − dm )
− ŷ T (t − d(t))U2 ŷ T (t − d(t))

t−d(t)−1
− η T (s)U1 η(s) + 2 ŷ(s)
s=t−d M


t−dm −1

− η T (s)U2 η(s) + 2 ŷ(s)}

s=t−d(t)

t−d(t)−1   T
d1  ŷ(t)
<  (t)4 (t) −
T
d M − d(t) η(t)
t=t−d M
Stability Analysis of Semi-Markovian … 173

 t−d(t)−1
  ŷ(t) 

R 1 U1
∗ R 2 + U1 η(t)
t=t−d M
⎛ ⎞

t−dm −1   TT
d1 ⎝ ŷ(t) ⎠
−
d(t) − dm t=t−d(t) η(t)
⎛ ⎞
  
t−d m −1  
R 1 U2 ⎝ ŷ(t) ⎠
∗ R 2 + U2 η(t)
t=t−d(t)

Letting, ∂1 = d M −d(t) , ∂2 = d M −d(t) and since ∂1 + ∂2 = 1, by Lemma 2, there

 
d1 d1
T T
exists matrix 1 2 > 0, such that
T3 T4
⎡   ⎤ T

t−d(t)−1
ŷ(t)
⎢ ⎥
⎢ t=t−d M ⎥
η(t)
⎢
V4 (t) ≤  (t)4 (t) − ⎢
T     ⎥
⎥
⎣ t−d m −1
ŷ(t) ⎦
t=t−d(t) η(t)
⎡ ⎤
R 1 U1 T1 T2
⎢ ∗ R1 + U1 T3 T4 ⎥
⎢ ⎥
⎣ ∗ ∗ R 1 U2 ⎦
∗ ∗ ∗ R 2 + U2
⎡   ⎤

t−d(t)−1
ŷ(t)
⎢ ⎥
⎢ t=t−d M η(t) ⎥
⎢    ⎥
⎢ t−d m −1
⎥
⎣ ŷ(t) ⎦
t=t−d(t) η(t)

V4 (t) ≤ T (t)4 (t)

⎡ ⎤T

t−d m −1

⎢ η(t) ⎥ ⎡R ⎤
⎢ t=t−d M ⎥ 1 U1 T1 T2
⎢ ⎥ ⎢
⎢ ŷ(t − d(t) − ŷ(t − d )) ⎥ ⎢ ∗ R 1 + U1 T3 T4 ⎥ ⎥
−⎢ M
⎥ ⎣
⎢ 
t−dm −1 ⎥ ∗ ∗ R 1 U2 ⎦
⎢ y(t) ⎥
⎣ t=t−d M ⎦ ∗ ∗ ∗ R 2 + U2
ŷ(t − dm ) − ŷ(t − d(t)))
174 K. Maheswari et al.
⎡ ⎤

t−dm −1

⎢ η(t) ⎥
⎢ t=t−d M ⎥
⎢ ⎥
⎢ ŷ(t − d(t) − ŷ(t − d M )) ⎥
×⎢ ⎥
⎢ 
t−dm −1 ⎥
⎢ y(t) ⎥
⎣ t=t−d M ⎦
ŷ(t − dm ) − ŷ(t − d(t)))
≤ T (t)(4 + 5 )(t) (17)

Considering the activation functions from the assumptions

 T   
ŷ(t) M1 1 −M2 1 ŷ(t)
≤ 0,
f ( ŷ(t)) −M2 1 1 f ( ŷ(t))
 T   
ŷ(t − d(t)) N1 2 −N2 2 ŷ(t − d(t))
≤0 (18)
g( ŷ(t − d(t)) −N2 2 2 g( ŷ(t − d(t))

where

1 = diag{α̃11 , α̃12 , . . . , α̃1n },

2 = diag{α̃21 , α̃22 , . . . , α̃2n },
M1 = diag{ M̃1− M̃1+ , M̃2− M̃2+ , . . . , M̃n− M̃n+ },
 
M̃1− + M̃1+ M̃2− + M̃2+ M̃n− + M̃n+
M2 = diag , ,..., , ,
2 2 2
N1 = diag{ Ñ1− Ñ1+ , Ñ2− Ñ2+ , . . . , Ñn− Ñn+ },
 
Ñ1− + Ñ1+ Ñ2− + Ñ2+ Ñn− + Ñn+
N2 = diag , ,..., , .
2 2 2

We obtain the zero equation, by introducing the relaxation matrix

 
2 η T (t)W0T + ŷ T (t)W1T [B ŷ(t − ψ(t))
+ C f ( ŷ(t)) + Dg( ŷ(t − d(k))) − ŷ(t) − η(t)] = 0 (19)

Combining all the equations,

 8 

V (t) ≤ T (t) s (t) < 0 (20)
s=1


8
If λ̃ = λmax () < 0,  = s with
s=1
Stability Analysis of Semi-Markovian … 175
 2
V (t) ≤ λ̃ ŷ(t) , ∀ ŷ(t) = 0. (21)

This implies that system (3) is globally asymptotically stable.

Hence, the proof is complete.

Remark These results can also be applied to the system of randomly occurring
parameter uncertainties, where the stability conditions does not depend upon the
transmission delays caused due to delay bounds but also depends upon the time-
varying leakage delays.

4 Numerical Example

Example 1 Let us take the uncertain NNs with the subsystem as:
     
0.8 0 −2.3 0 1.3 0.2
B= ,C = ,D = ,
0 0.7 1.7 0.4 0.1 −0.02
     
0.04 0.02 0.07 0 0.03 0.02
Na = , Nb = , Nc =
0 −0.03 0.04 0.03 0.01 −0.04

With the nonlinear activation functions given by,

   
tanh(0.6 ŷ1 (t)) + 0.6 tanh(0.3 ŷ1 (t))
f ( ŷ(t)) = , g(y(t)) =
tanh(−0.4 ŷ2 (t)) + 0.2 tanh(−0.4))

The time-varying
 delay
 and
 the leakage delay are given
 by.

d(t) = 21 7 + 5 sin tπ2 and ψ(t) = 21 5 + 3 sin tπ2 with corresponding
bounds given by dm = 3, d M = 8, ψm = 1, ψ M = 5. By solving the LMIs,
a solution is obtained.
Figure 1 showcases NNs converge smoothly in the absence of leakage delay and
there is some kind of dynamic behavioral change in the network in the presence of
leakage delay. Hence, the effectiveness of the given result is thus verified.

5 Conclusion

In this paper, the analysis of stability problem is discussed with leakage delays.
Sufficient conditions are encountered to ensure the delay-dependent conditions and
to ensure the robustness by using the LMI approach. Numerical example has also
been for the derived theoretical result.
176 K. Maheswari et al.

1.2
x (k)
1
1
x (k)
2
0.8

0.6

0.4
x(k)

0.2

−0.2

−0.4

−0.6

−0.8
0 1 2 3 4 5
Time k

Fig. 1 State responses of the system

References

1. Das S, Moura JMF (2015) Distributed Kalman filtering with dynamic observations consensus.
IEEE Trans Sign Process 63(17):4458—4473
2. Hu J, Wang Z, Gao H (2018) Joint state and fault estimation for time-varying nonlinear systems
with randomly occurring faults and sensor saturations. Automatica 97(2018):150–160
3. Maheswari K (2016) H∞ state estimation for discrete-time neural networks with multiple
missing measurements using second order reciprocal convex approach. Adv Intell Syst Comput
412:102–119
4. Rakkiyappan R, Maheswari K, Sivaranjani K (2018) Young Hoon Joo, Non-Fragile finite-time
l2 − l∞ state estimation for discrete-time neural networks with semi-Markovian switching and
random sensor delays based on Abel lemma approach. Nonlinear Anal Hybrid Syst 29:283–302
5. Rakkiyappan R, Maheswari K, Velmurugan G, Park JH (2018) Event-triggered H∞ state esti-
mation for semi-Markov jumping discrete-time neural networks with quantization. Neural Netw
105:236–248
6. Nagamani G, Ramasamy S (2016) Dissipativity and passivity analysis for discrete-time T-S
fuzzy stochastic neural networks with leakage time-varying delays based on the Abel lemma
approach. J Franklin Inst 353:3313–3342
Forecasting PM10 Concentration Based
on a Hybrid Fuzzy Time Series Model

Yousif Alyousifi and Mahmod Othman

Abstract Developing statistical models for air pollution forecasting is crucial

for managing air quality. Nevertheless, many researchers have concentrated on
improving the model’s accuracy when applying for data with many fluctuations in
the pollutant’s concentration. Also, they have attempted to address the uncertainty
analysis that might lead to inadequate outcomes. The fuzzy time series (FTS) is
considered one of the powerful models that are commonly applied in predicting air
pollution. However, most FTS models are not accurate in partitioning the universe of
discourse. Therefore, a new hybrid model based on the FTS-based Markov chain and
C-Means clustering technique with an optimal number of clusters is proposed in this
study. This hybridization is contributed to produce an adequate partition and improve
the model accuracy accordingly. The superiority of the proposed model is validated
using three common statistical criteria. The PM10 concentration data collected from
Melaka, Malaysia is used in this study. Results prove that the proposed model greatly
improved the prediction accuracy, for which it outperformed several fuzzy time series
models. Hence, we have concluded that the model proposed is a good option for
forecasting air pollution and any type of random data.

Keywords Air pollution · C-Means clustering techniques · Fuzzy time series ·

Markov transition matrix · PM10

Y. Alyousifi (B) · M. Othman

Department of Fundamental and Applied Sciences, Universiti Teknologi PETRONAS, Bandar
Siri Iskandar, 31750 Tronoh, Perak, Malaysia
e-mail: [email protected]
Y. Alyousifi
Department of Mathematics, Faculty of Applied Sciences, Thamar University, Thamar, Yemen

© The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 177
R. Ibrahim et al. (eds.), International Conference on Artificial Intelligence
for Smart Community, Lecture Notes in Electrical Engineering 758,
https://doi.org/10.1007/978-981-16-2183-3_16
178 Y. Alyousifi and M. Othman

1 Introduction

Prediction of air pollution is a crucial environmental problem in most countries as

it is beneficial to air pollution assessment where the result found can be used for
managing the air quality. Also, predicting the concentration of air pollutants is vital
for assessing the effects of air pollutants on human health [1, 2].
Many researchers have been applied the FTS models for forecasting air pollution
[3]. For instance, Cagcag et al. [4] have predicted air pollution in Ankara based on a
new seasonal FTS model. Koo et al. [5] have utilized some statistical models in order
to forecast the air pollution events and made a comprising to determine the adequate
model. Apart from that, the FTS model based on Fuzzy K-Medoids clustering is
suggested for minimizing the sensitivity of outliers in order to produce adequate
forecasts of air pollution [6]. Furthermore, in air pollution forecasting, an integrated
model of the fuzzy theory and advanced optimization algorithm is introduced [7]. A
hybrid FTS model is also suggested for predicting air pollution [8].
Since the hybrid models are performed better than their single models. For
example, the hybrid models in conventional time series have shown an improvement
in the model performance more than their single models such as the ARMA and
ARIMA models, which are combined between the autoregressive (AR) and moving
average (MA) models. These hybrid models have outperformed the singles models.
Accordingly, many researchers have proposed a variety of hybrid FTS models to
address some of these issues. For example, a hybrid fuzzy time series Markov chain
(FTSMC) model has been proposed by Tsaur [9], which is used for establishing
proper weights of the fuzzy relationships between the data points of the time series
pattern. Aripin et al. [10] have applied the FTSMC model for air pollution fore-
casting. FTSMC is superior to many methods available in the literature in terms of
model accuracy. Nevertheless, it has a drawback in utilizing an arbitrary partition
of intervals. So, it could not determine the appropriate length of intervals, which
is considered a deficiency in determining the effective length of the interval in this
hybrid model. Apart from that, Alyousifi et al. [11], proposed the use of the FTSMC
model based on the optimal grid partition method for modeling air pollution events
in Malaysia.
Furthermore, the hybrid model of the fuzzy time series model and the C-Means
clustering technique has outperformed the fuzzy time series models. For example,
Van Tinh et al. [12] proposed a hybrid fuzzy time series and K-Means clustering for
forecasting the enrolment of the University of Alabama. Likewise, Kai et al. [13]
proposed a forecasting model of fuzzy time series based on K-Means clustering.
However, it has a limitation in dealing with recurrences of the observations and
the number of clusters that have been chosen randomly, which can be led to an
insufficient length of intervals and insufficient forecasting accuracy. To address these
two limitations of the existing hybrid models that were previously mentioned, this
study proposed a new hybrid model (FTSMC-CMeans) by combining the fuzzy
time series, Markov chain, and C-Means clustering technique. It is anticipated that
the proposed model may overcome these limitations and produce a better model
Forecasting PM10 Concentration Based on a Hybrid … 179

forecasting and to produce better model accuracy. The proposed model is validated
by comparison with several existing models.

2 Methodology

FTS is a forecasting model based on the fuzzy set theory and Fuzzy Logic proposed
by Zadeh in 1965 [14]. FTS models are considered as advanced statistical time series
analysis applying to fuzzy sets. due to their advantages. Song and Chissom [15, 16]
are firstly introduced the first-order FTS model. To reach the high accuracy of the
forecasting model, developing FTS models have been done by enhancing the three
main stages of the FTS algorithm, which are fuzzification, fuzzy inference, and
defuzzification [16]. For instance, Chen [17] has improved the FTS model proposed
by Song and Chissom [15, 16] through the use of fuzzy logical relation tables for
the purpose of reducing the computational complexity of the model. Also, Huarng
[18, 19] have developed Chen’s model [17] by handling the issue of the length of the
interval.

2.1 C-Means Clustering Technique

The C-Means clustering technique [20, 21] is one of the famous unsupervised
learning algorithms. It is a partitioning clustering algorithm, which partitions a given
data set into a set of C clusters. The result of the C-Means clustering Technique
depends on the number of clusters. Apart from that, the main matter in partitioning
clustering is determining the optimal number of clusters in a data set. For instance,
the C-Means clustering requires the user to specify the number of clusters k in order
to be generated. Accordingly, in this paper, an optimal number of clusters is deter-
mined based on three different methods, which are elbow, average silhouette, and gap
statistic methods [22], through using two functions in R, which are fviz_nbclust() and
NbClust() [23]. By using these functions, the optimal number of clusters determines
the best partition selected. For further details about these methods see [24].

2.2 Hybrid Fuzzy Time Series Model

The algorithm of this study is adopted the arithmetic processes proposed by Tsaur
[9] in order to conduct the proposed model. The steps of the model are as follows:
Step 1. Define U.
Step 2. Divide U into subintervals based on the C-Means Clustering technique.
Step 3. Define the fuzzy sets Ai for each time series observation on U.
180 Y. Alyousifi and M. Othman

Step 4. Fuzzify the actual values of the time series (observations) into fuzzy
numbers based on the maximum membership value.

f Ai (u 1 ) f A (u 2 ) f A (u n )
Ai = + i + ··· + i (1)
u1 u2 un

where f Ai is the membership function of fuzzy set Ai ; f Ai : U → [0, 1]. f Ai (u r ) ∈

[0, 1] and 1 ≤ r ≤ n.
Step 5. Establish the fuzzy logical relationships (FLRs) and determine fuzzy
logical relation groups (FLRGs). Then, create the Markov transition probability
matrix. The matrix P is Pn×n , and Pi j is transition probability from state Ai to
state A j . Pi j can be calculated as follows

Ni j
Pi j = , i, j = 1, 2, . . . , n (2)
N i.


n
where Ni j is the frequencies and Ni. = Ni j is the total frequencies.
j=1
Step 6. Calculate the forecasted values by considering the following cases:
Case 1. If the FLRG of Ai is one-to-one, i.e., Ai → Ak , with Pik = 1 and Pi j =
0, j = k), then the forecasting of F(t) is m k , the midpoint of u k , k = 1, 2, ..n,
which is determined by using Eq. (3)

F(t + 1) = m k Pik = m k (3)

Case 2. If the FLRG of Ai is one-to-many, i.e., Ai → A1 , A2 , ...An , i =

1, 2, .., n). Thus, if the state is Ai for the actual value Y (t) at time t, the
forecasted value F(t + 1) is calculated according to Eq. (4)

F(t + 1) = m 1 pi1 + m 1 p12 + · · · + m i−1 pi(i−1) + Y (t) pii

+ m i+1 pi(i+1) + · · · + m n pin (4)

where m 1 , m 2 , . . . , m n are the midpoint of u 1 , u 2 , . . . , u n and m i replaced by

Y (t).
Step 7. Adjust the predicted values by adding the differences of actual values
Y (t). The adjusted forecasted values can be written by

F̂(t + 1) = F(t + 1) + diff (Y (t)) (5)

Forecasting PM10 Concentration Based on a Hybrid … 181

2.3 Model Evaluation

Three statistical criteria used in this study in order to validate the proposed model,
which are MAPE, RMSE, and Thiels’ U statistics, are given in Equations as follows:

N  
1  Yi − Fi 
MAPE = × 100, (6)
N i=1  Yi 

N
i=1 (Yi − Fi )2
RMSE = and (7)
N

N
i=1 (Yi − Fi )2
Theil’s U =   (8)
N N
i=1 Yi +
2 2
i=1 Fi

where Yi means the actual data, Fi the predicted values and N is the total number of
the actual data.

3 Results and Discussion

In this section, we will provide the results of the proposed model through the imple-
mentation and comparison of the model using two datasets. The algorithm of the
proposed model is implemented for predicting the concentration of PM10 . To vali-
date the proposed model, the datasets are used to assess the model performance and
compare it with other models.

3.1 Air Pollution Forecasting

For validation of the proposed model, the algorithm of the proposed model is also
applied for both training and testing datasets of PM10 concentrations are used to
evaluate the performance of the proposed model and compare it with some of the
existing FTS models. For implementing the Algorithm of the proposed model, we
start by defining U from data and then apply the C-Means clustering method.
As can be seen from Table 1 that the proposed model has been performed very well
with producing small errors. Figures 1 and 2 demonstrate the results of a compar-
ison of the proposed model and some existing models using the testing data of
PM10 concentrations. It observed that the model has predicted PM10 data well with
producing the smallest error. Besides, the proposed model outperformed the other
existing models. Also, the model shows its superiority as compared with the other
182 Y. Alyousifi and M. Othman

Table 1 A comparison of the hybrid model and some FTS models using PM10 data
Model Using training dataset Using testing dataset
RMSE MAPE U statistic RMSE MAPE U statistic
1 Chen’s model [17] 18.04 14.67 1.32 9.12 14.26 0.80
2 Sliva2 et al.’s model [25] 16.10 17.79 1.18 10.23 18.68 0.89
3 Yu’s model [26] 17.06 9.74 1.25 8.87 14.20 0.78
4 Cheng’s model [27] 16.70 10.88 1.22 8.67 13.69 0.76
5 Severiano et al.’s [28] 16.10 17.79 1.18 10.23 18.68 0.89
6 Sliva et al.’s model [29] 15.67 9.28 1.14 8.90 13.92 0.78
7 Sadaei et al.’s [30] 17.06 9.82 1.25 8.73 13.99 0.77
8 Tsuar’s model [9] 17.07 9.40 1.25 8.66 13.86 0.76
9 Hybrid Model 7.55 6.83 0.55 5.01 7.25 0.45

Fig. 1 A comparison of the hybrid model and some FTS models using the training PM10 Data

models. This implies the proposed model is an improved option for forecasting air
pollution events.

4 Conclusion

This study proposed a new hybrid fuzzy time series model, which is implemented
for predicting PM10 concentration. The model hybridization has been contributed in
Forecasting PM10 Concentration Based on a Hybrid … 183

Fig. 2 A comparison of the hybrid model and some FTS models using the testing PM10 Data

producing adequate partitioning and improving the model accuracy. The proposed
model was evaluated by comparison with eight FTS models, which are famous in
the literature. A comparison has demonstrated the ability of the model in avoiding
the arbitrary selection of intervals and dealing with recurrent observations, which
greatly improves model accuracy. The predicted values found by the model display
its flexibility in the FTS for predicting air pollution. Generally, the proposed model
has the flexibility to apply for many types of time series data.

References

1. Yan Y, Li Y, Sun M, Wu Z (2019) Primary pollutants and air quality analysis for urban air in
China: evidence from Shanghai. Sustainability 11(8):2319
2. Alyousifi Y, Othman M, Faye I, Sokkalingam R, Silva PC (2020) Markov weighted fuzzy
time-series model based on an optimum partition method for forecasting air pollution. Int J
Fuzzy Syst 22:1468–1486
3. Cheng CH, Huang SF, Teoh HJ (2011) Predicting daily ozone concentration maxima using
fuzzy time series based on a two-stage linguistic partition method. Comput Math Appl
62(4):2016–2028
4. Cagcag O, Yolcu U, Egrioglu E, Aladag CA (2013) Novel seasonal fuzzy time series method
to the forecasting of air pollution data in Ankara. American Journal of Intelligent Systems.
3(1):13–19
5. Koo JW, Wong SW, Selvachandran G, Long HV (2020) Prediction of Air Pollution Index in
Kuala Lumpur using fuzzy time series and statistical models. Air Q Atmosp Health 1–12
6. Dincer NG, Akkuş Ö (2018) A new fuzzy time series model based on robust clustering for
forecasting of air pollution. Eco Inform 43:157–164
184 Y. Alyousifi and M. Othman

7. Yang H, Zhu Z, Li C, Li R (2019) A novel combined forecasting system for air pollutants
concentration based on fuzzy theory and optimization of aggregation weight. Appl Soft Comput
105972
8. Wang J, Li H, Lu H (2018) Application of a novel early warning system based on fuzzy time
series in urban air quality forecasting in China. Appl Soft Comput 71:783–799
9. Tsaur RC (2012) A fuzzy time series-Markov chain model with an application to forecast the
exchange rate between the Taiwan and US dollar. Int J Innov Comput Inf Control 8(7):4931–
4942
10. Aripin A, Suryono S, Bayu S (2016) Web-based prediction of pollutant PM10 concentration
using Ruey Chyn Tsaur fuzzy time series model. AIP Conf Proc 17(46):20–46
11. Alyousifi Y, Othman M, Sokkalingam R, Faye I, Silva PC (2020) Predicting daily air pollution
index based on fuzzy time series Markov chain model. Symmetry 12(2):293
12. Van Tinh N, Vu VV, Linh TTN (2016) A new method for forecasting enrolments combining
time-variant fuzzy logical relationship groups and k-means clustering. Int Res J Eng Technol
3(3):1–32
13. Kai C, Fang-Ping F, Wen-Gang C v A novel forecasting model of fuzzy time series based on
k-means clustering. IWETCS, IEEE, 223–225
14. Zadeh LA (1965) Fuzzy sets. Inf Control 8(7):338–353
15. Song Q, Chissom BS (1993) Forecasting enrollments with fuzzy time series-Part I. Fuzzy Sets
Syst 54:1–10
16. Song Q, Chissom BS (1994) Forecasting enrollments with fuzzy time series-Part II. Fuzzy Sets
Syst 62(1):1–8
17. Chen SM (1996) Forecasting enrolments based on fuzzy time series. Fuzzy Sets Syst 81(3):311–
319
18. Huarng K (2001) Effective lengths of intervals to improve forecasting in fuzzy time series.
Fuzzy Sets Syst 123(3):387–394
19. Huarng K, Yu TH (2006) Ratio-based lengths of intervals to improve fuzzy time series
forecasting. IEEE Trans Syst Man Cybern Part B Cyber 36:328–340
20. Hartigan JA (1979) A K-means clustering algorithm: algorithm AS 136. Appl Stat 28(11):126–
130
21. Zhang Z, Zhu Q (2012) Fuzzy time series forecasting based on k-means clustering. Open J
Appl Sci 2:100–103
22. Rousseeuw PJ, Kaufman L (1990) Finding groups in data. Wiley, Hoboken
23. Charrad M, Ghazzali N, Boiteau V, Niknafs A (2014) NbClust: An R package for determining
the relevant number of clusters in a data set. J Stat Softw 61:1–36
24. Kaufman L, Rousseeuw PJ (2009) Finding groups in data: an introduction to cluster analysis.
John
25. Silva PC, Sadaei HJ, Guimarães FG (2017) Interval forecasting with fuzzy time series.
Conference: IEEE symposium series on computational intelligence
26. Yu HK (2005) Weighted fuzzy time series models for TAIEX forecasting. Physica A
349(34):609–624
27. Cheng CH, Chen TL, Chiang CH (2006) Trend-weighted fuzzy time-series model for TAIEX
forecasting neural information processing. Spring Berlin/Heidelberg Lect Notes Comput Sci
42(34):469–477
28. Severiano, C. A., Silva, P. C., Sadaei, H. J., Guimarães, F. G.: Very short-term solar forecasting
using fuzzy time series. In 2017 IEEE international conference on fuzzy systems (FUZZ-IEEE).
1–6 (2017).
29. Silva PC, Sadaei HJ, Guimarães FG (2019) Probabilistic forecasting with fuzzy time series.
IEEE Trans Fuzzy Syst 99
30. Sadaei HJ, Enayatifar R, Abdullah AH, Gani A (2014) Short-term load forecasting using a
hybrid model with a refined exponentially weighted fuzzy time series and an improved harmony
search. Int J Electr Power Energy Syst 62:118–129
Carbonated Water Injection for EOR
and CO2 Storage: A Simulation Study

A. Abdulrahman and J. Foroozesh

Abstract CO2 Enhanced Oil Recovery (EOR) techniques have gained massive
attention by oil companies as they target the oil industry’s two main concerns of CO2
contribution to the climate change and the decline in oil production. Carbonated Water
Injection (CWI) is a promising EOR technique that promotes safe and economic CO2
storage, moreover, it mitigates the pitfalls of CO2 injection which include low sweep
efficiency, early CO2 breakthrough, and the risk of CO2 leakage in subsurface forma-
tions. Upon the contact of Carbonated Water (CW) with oil inside the reservoir, CO2
molecules cross the interface between the two fluids moving towards the oil phase
due to the concentration gradient and its higher solubility in hydrocarbons. Subse-
quently CO2 molecules diffuse inside the oil until thermodynamic equilibrium is
reached. CO2 dissolution in oil causes it to swell and consequently leads to changes
in its physical properties (viscosity and density). Such changes in oil properties lead
to improved sweep and displacement efficiency and thus higher recovery factor.
Several experimental studies have been reported in the literature, but little work has
been done on the simulation of CWI due to the complex multi-physics nature of
this process. In this paper, Schlumberger equilibrium-based compositional simulator
(ECLIPSE300) has been used to investigate the oil recovery and CO2 storage during
CWI. The carbonated water has been simulated using two injector wells placed at the
same location where they inject free CO2 and water at a certain volumetric ratio to
account for the mass fraction of dissolved CO2 inside the carbonated water. CO2 SOL
option has been used to account for CO2 solubility inside the water phase and a 2D
cartesian model (x, y) has been considered to avoid the effects of gravity which might
reduce the amount of CO2 dissolved inside the oil phase. A sensitivity analysis on
CW injection rate, and the effect of CO2 diffusion have been investigated. It was
found that low injection rate promotes longer contact time thus more CO2 molecules

A. Abdulrahman
Petroleum Engineering Department, Universiti Teknologi PETRONAS, Perak, Malaysia
J. Foroozesh (B)
School of Energy and Electronic Engineering, University of Portsmouth, Portsmouth, United
Kingdom
e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 185
R. Ibrahim et al. (eds.), International Conference on Artificial Intelligence
for Smart Community, Lecture Notes in Electrical Engineering 758,
https://doi.org/10.1007/978-981-16-2183-3_17
186 A. Abdulrahman and J. Foroozesh

will get transferred to the oil leading to high oil recovery and CO2 storage. In addi-
tion, it was found that CO2 diffusion between grid cells has minimal impact on oil
recovery and CO2 storage in this case study.

Keywords Carbonated water injection · Compositional simulation · Enhanced oil

recovery · CO2 storage · CO2 diffusion · CO2 solubility in water

1 Introduction

Carbonated water injection is a promising CO2 based EOR technique that injectes
CO2 in an effective and efficient way. In this method, the CW is prepared by dissolving
CO2 inside the water at surface conditions using gas infusion generator [1], then
the CW mixture is injected into the formation to increase oil recovery and store
the dissolved CO2 [2]. When CW contacts the oil phase inside the reservoir, CO2
molecules gradually cross the interface between the two fluids due to the concen-
tration gradient (cross phases diffusion) and higher solubility of carbon dioxide in
hydrocarbons, and subsequently, CO2 molecules diffuse over time within the resident
oil until thermodynamic equilibrium is reached. CO2 dissolution in oil causes it to
swell and consequently leads to a change in its physical properties (density–viscosity)
[2–6].
Several experimental studies have been carried out to evaluate the additional
recovery factor during CWI, these experiments have been conducted using different
setups such as coreflood rigs, sand-packs, and high-pressure transparent micro-
models. It has been found that CWI in secondary mode (after primary depletion)
yields a higher recovery factor compared to CWI in tertiary mode (after water injec-
tion). Sohrabi et al. [7] have performed coreflood experiment in a water-wet reservoir
core (sandstone), in the first part of the experiment, CWI was injected in a secondary
mode and the recorded increment in the recovery factor was 19%. In the second
part of the experiment, CWI was injected in a tertiary mode, the recorded incre-
ment the recovery factor was determined to be 9%. Mosavat and Torabi [2] have
conducted a flooding experiment in a sand-pack apparatus filled with sand made of
99.88% silicon, it was found that the injection of CO2 -saturated water resulted in
improving the conventional water flooding oil recovery factor by about 19.0% and
12.5% of OOIP for the secondary and tertiary scenario, respectively. Riazi et al. [8]
have injected CWI (tertiary mode) in a micro model made of two flat glass plates,
the recorded recovery factor by water injection was 27% and the increment in the
recovery factor due to CWI (post water flood) was found to be 6%.
Carbonated water injection EOR has the potential to store CO2 safely and econom-
ically in the geological formations. Kechut et al. [9] have performed an experimental
investigation of CO2 storage in sandstone core samples. During secondary CWI, 45%
of the dissolved CO2 was stored while in tertiary CWI it was found that about 51%
of injected CO2 was stored. Tavakolian et al. [10] have performed series of coreflood
experiments using sandstone core plugs and crude oil at real reservoir conditions. The
Carbonated Water Injection for EOR and CO2 Storage … 187

experiments were carried out to compare the performance of CWI and CO2 flooding.
CWI resulted in a less and more controllable CO2 production profile compared to
CO2 flooding. Also, it was found that duringCO2 flood experiments, CO2 production
exceeded that of CO2 storage very quickly and after that large volume of CO2 had
to be reinjected (recycled), whereas CO2 production during CWI was significantly
less than CO2 storage even after a large volume of CWI.
The additional oil recovery reported during CWI can be attributed to the effect
of several mechanisms simultaneously. These recovery mechanisms include oil
swelling, viscosity reduction, wettability alteration, IFT reduction, and CO2 exsolu-
tion [3–6, 11–13]. Sohrabi et al. [4–6] have investigated oil swelling during CWI by
conducting many flooding experiments in a micro model using a light oil (viscosity
= 0.83 cp) and also a viscous oil (viscosity = 16.5 cp). They have found out that
the swelling in the light oil was 105% while the swelling in the viscous oil was 23%
which led to the conclusion of swelling degree is a function of oil viscosity. A series
of experimental studies have been conducted by Miller and Jones [3] to investigate
the change in oil viscosity during the CWI process. They have run the experiments at
different values of pressures and temperatures, different oil samples with a viscosity
of 10 and 17 API were used. It was found out that in all the cases the viscosity of the
oil in the absence of CO2 increased as the pressure increased, while the viscosities
of the oil with CO2 in the solution decreased as the pressure increased. Mosavat
and Torabi [13] have performed several experiments to investigate the effect of CO2
diffusion on wettability alternation. It was also concluded that once CO2 molecules
reach the rock surface (in contact with the oil droplet), the molecules started replacing
the hydrocarbon molecules were adsorbed on the surface. This gradual replacement
caused the rock surface to shift its wettability towards the water-wet condition. Yang
et al. [11] have conducted lab experiment to measure dynamic interfacial tension of
oil-brine systems compared to oil-brine CO2 systems at different values of pressures
and temperatures. They realized that in the crude oil-brine system, the dynamic inter-
facial tension reduction was primarily caused by the natural surfactant which was
possibly generated due to chemical reactions between the crude oil and the reservoir
brine. While For the crude oil-brine CO2 system, the dissolution of CO2 into crude
oil was an additional factor to reduce the dynamic interfacial tension. Riazi [12] have
conducted a fluid flow experiment on CWI using a high-pressure micro-model. They
have noticed an interesting phenomenon of gas nucleation subsequent to depressur-
izing of the CW flooded. This pressure reduction appreciably affected the production
process, leading to an additional oil recovery (due to CO2 exsolution).

2 Simulation Studies of CWI and Challenges Encountered

An accurate reservoir model with the ability to reliably predict the performance of
different recovery methods is vital for successful reservoir management. There are
black oil and compositional simulation approaches available depending on type of
reservoir fluids and recovery methods, Compositional simulation approach should
188 A. Abdulrahman and J. Foroozesh

be considered when simulating carbonated water injection since significant cross-

phase mass transfer of CO2 molecules occurs between carbonated water and hydro-
carbon leading to oil properties changes. In addition, hydrocarbon physical properties
(viscosity and density) are a function of both pressure and CO2 content. The following
paragraphs will explain how CWI is simulated in ECLIPSE300, the complex physics
required for successful simulation of CWI, the simulation studied that have been
performed and the limitations encountered.

2.1 Simulation of CWI in ECLIPS300 Compositional

Simulator

There is no unique option in ECLIPSE300 that for precise simulation of CWI.

However, a few researchers and industry experts tried to simulate CWI using
“CO2 SOL” which is a three-phase compositional option that accounts for CO2 solu-
bility in water and CO2 dissolution in oil phase [14]. In this option, it is assumed that
instantaneous equilibrium (instantaneous phase equilibrium) exists between oil, gas,
and water phases. Furthermore, water is not allowed to dissolve in gas or oil phase.
CO2 component can be present in the three phases, while hydrocarbon components
can only be present in gas and oil phases. “SOLUBILI” keyword is used in the PROPS
section to provide experimental CO2 in water solubility data. If the solubility data
is not provided by the user, built in correlations will be used to generate the CO2 in
water solubility data [14–16].
ECLIPSE300 has two diffusion models that can be used to represent CO2 diffusion
within the oil or gas phase (diffusion between two adjacent cells containing the same
phase); in the first model, diffusion is driven by the concentration gradient, while
in the second model diffusion is driven by the gradient of the chemical potential.
Diffusivity coefficient can be specified in two possible ways based on the mentioned
diffusion models; normal diffusion coefficients Di defined by keywords DIFFCOIL
and DIFFCGAS or activity corrected diffusion coefficient Dαi defined by keywords
DIFFAGAS and DIFFAOIL [15].

2.2 The Complex Physics Required for Successful

Simulation of CWI

Kechut et al. [9, 17, 18] used compositional simulators to simulate a set of carbonated
water coreflood experiments. The coreflood experiments had been performed using
two types of core (a reservoir water-wet core and Clashach water-wet sandstone core)
and two different oil samples: a light oil (n-decane) and a viscous stock-tank crude
oil with known composition. Secondary water injection, secondary CWI and tertiary
CWI were simulated using the ECLIPSE300 commercial reservoir simulators. First
Carbonated Water Injection for EOR and CO2 Storage … 189

water injection (WI) was simulated. The input water–oil relative permeability (Krw-
o) curve into the simulator was obtained based on Corey correlations when produc-
tion data was history matched. Next CWI was simulated. The ECLIPSE300 composi-
tional simulator was used with a tuned EOS, which could acceptably predict the PVT
properties of the oil-CO2 system. The Krw-o obtained for WI tests were input into
the ECLIPSE300 during CWI simulation. The simulation results by ECLIPSE300
showed an over prediction of oil recovery in comparison with experimental data
for all coreflood tests. They explained that compositional simulators were developed
based on the inherent assumptions of instantaneous equilibrium and complex mixing
while during carbonated water flooding, CO2 transferring process from water into the
oil phase happens gradually. The resistance against CO2 migration at the interface
between water and oil phases plays an important role, which prevents immediate
distribution of CO2 between the phases. Consequently, the assumption of instanta-
neous equilibrium may not be realistic during CWI process. To support their expla-
nation, they mentioned the micro-model observation reported by Sohrabi et al. [5],
who used image processing technique during the conducted visualisation tests. Riazi
et al. [8] have reported that in the micro-model system the swelling of oil blobs as
a result of CO2 diffusion happens gradually and it needs time to stabilise and reach
equilibrium state (170 h to reach its final shape corresponding to maximum swelling
value). Embid and Rivas [19] explained that this assumption of instantaneous equilib-
rium can lead to a large error in the cases where mass transfer resistances are large.
They mentioned that this may happen when there are short contact times for the
mass transfer process (laboratory displacement in cores) or large diffusion patterns
are available for components to diffuse through (field scale), moreover, if there are
slow diffusion velocities due to large viscosity. All these can lead to not having an
instantaneous thermodynamic equilibrium state.
The effect of CO2 diffusion within the oil phase is another challenge encountered
during the simulation of CWI. Molecular diffusion is defined as the movement of
molecules caused by Brownian motion or composition gradient in a fluid mixture
that can facilitate the dissolution of CO2 into the oil systems [20]. Injected CO2
during CO2 based injection projects could just pass through the reservoir, causing
early breakthrough, if there is no molecular diffusion. This would lead to lower
oil recovery prediction as some parts of the reservoir are not touched by the CO2
causing no benefit from the CO2 presence. This may lead to wrong prediction as
diffusion mechanism can play important roles during CO2 based injection projects
[21, 22]. Alfarge et al. [23] reported that most of the studies in the literature identified
molecular diffusion as an important mechanism for CO2 to enhance the oil recovery in
oil reservoirs. Grogan and Pinczewski [24] highlighted the importance of molecular
diffusion during the oilfiled life. As such disregarding the diffusion mechanism in the
model can lead to overestimation or underestimation of the recovered oil [22, 23].
190 A. Abdulrahman and J. Foroozesh

2.3 Nonequilibrium-Based CWI Simulation

Foroozesh et al. [25, 26] have come up with a novel 1D compositional simula-
tion approach whereby the assumption of instantaneous equilibrium was relaxed by
adding a new term to the governing equation called mass transfer coefficient (MTC).
This term used to capture the slow CO2 mass transfer from CW to oil within the
grid cell. Their kinetic based model was used to simulate a secondary CWI experi-
ment reported by Sohrabi et al. [7] for validation purposes. MTC was tuned to match
differential pressure (DP) and recovery factor (RF) data of the experiments and a
good match was achieved. Additionally, Foroozesh and Jamiolahmady [25] inves-
tigated the role of rock wettability and wettability alteration in the performance of
CWI process. They simulated CWI coreflood experiments in mixed wet sandstone
cores. The relative permeability, capillary pressure and MTC values were tuned and
a good matching was achieved. The also used their kinetic based model to map the
result of CWI in core to reservoir (scale-up practice). Although, Foroozesh et al.’s
simulation approach was a novel way of simulating CWI but they did not capture the
effect of CO2 diffusion within the phases.

3 Simulation Of CWI Using ECLIPSE300

In this study, a 2D two-phase compositional model has been constructed using

ECLIPSE300 to investigate the oil recovery and CO2 storage during CWI. The 2D
model is in the direction of the cartesian coordinates (x, y) to avoid the effects of
gravity (water underride and gas override) which might reduce the amount of CO2
dissolution inside the oil phase. CO2 SOL option (keyword) was used for this simu-
lation. It was not possible to set the concentration of CO2 in injected water stream by
using CO2 SOL. As such, the carbonated water has been simulated using two injector
wells placed at the same location where the first well injects free CO2 while the
second well injects water. The volumetric ratio was adjusted between the injected
water and CO2 to represent a CW with a mass fraction of 5% dissolved CO2 . The
following table and figure show the reservoir description, injection rate and relative
permeability curves (Table 1; Fig. 1).
The oil composition consists of Decane (C10) with a density of 43.6288 Ib/ft3 and
viscosity of 0.2780 cp at the reservoir initial conditions. The total simulation time was
about 8.3 years or 3000 days and no free CO2 was present inside the reservoir during
the simulation. The total oil recovery at the end of this simulation (base case) was
found to be 68.4% and all the amount of injected CO2 was stored safely inside both
oil and water phase as CO2 did not appear in the production stream. The following
figures show the oil saturation, free gas saturation and mass fraction of CO2 inside
oil phase at the end of simulation (Figs. 2, 3 and 4).
Carbonated Water Injection for EOR and CO2 Storage … 191

Table 1 Details of the

Number of grid blocks 50 × 50 × 1
created ECLIPSE300 model
Gridblock dimensions, ft3 100 × 100 × 1
Permeability, md (kx = ky) 500
Porosity, fraction 0.25
Reservoir temperature, F 212
Initial reservoir depth, psi 3500
Initial water saturation, fraction 0
Water injection rate, STB/day 4750
CO2 injection rate, MSCF/day 2.305

Fig. 1 a Oil water relative permeability and b gas oil relative permeability

Fig. 2 Oil saturation (C10)

after 3000 days

3.1 CW Injection Rate Sensitivity

A sensitivity analysis has been done on CW injection rate to investigate its impact on
oil recovery, CO2 storage and spatial spreading inside the reservoir. Three different
injection rates have been studies and the mass fraction of dissolved CO2 inside
injected CW for each scenario was 5%. In the first scenario, a total of 2000 STB/day
CW has been injected into the reservoir (1900 STB/day of water, 0.922 MSCF/day
192 A. Abdulrahman and J. Foroozesh

Fig. 3 Free gas saturation

after 3000 days

Fig. 4 Mass fraction of

dissolved CO2 inside the oil
phase at the end of
simulation

of free CO2 ) for 7500 days, while in the second scenario a total of 5000 STB/day
CW has been injected (4750 STB/day of water, 2.305 MSCF/day of free CO2 ) for
3000 days, and in the last scenario 8000 STB/day of CW has been injected (7600
STB/day of water, 3.688 MSCF/day of free CO2 ) for 1875 days.
It was found that the oil recovery factor for the injection rate of 2000 STB/day is
the highest (68.6%) and the wide spatial spreading of CO2 inside the reservoir led
to early CO2 breakthrough compared to the other scenarios thus less CO2 storage
(96%). This can be attributed to the long contact time between CW and oil phase
which allows more CO2 to be transfered based on the thermodynamic equilibrium
calculation. While for scenario three, less oil recovery was found (58%) and the
limited CO2 spatial spreading inside the reservoir caused 100% CO2 storage as CO2
did not reach the producer well. The flowing figures show the oil recovery, CO2
spatial spreading and CO2 storage for the three scenarios (Figs. 5, 6 and 7).
Carbonated Water Injection for EOR and CO2 Storage … 193

Fig. 5 Oil recovery factor at 70

the end of simulation for the
three different injection rates
of CW 65

RF %
60

55

50

2000 STB/day 5000 STB/day 8000 STB/day

Fig. 6 Mass fraction of dissolved CO2 inside the oil phase at the end of simulation for a 2000
STB/day, b 5000 STB/day, c 8000 STB/day

Fig. 7 Cumulative injected free CO2 and cumulative produced CO2 for the different scenarios of
CW injection rate
194 A. Abdulrahman and J. Foroozesh

Fig. 8 Oil recovery vs time for the 3 scenarios of different CO2 -Oil diffusion coefficients

3.2 Effect of CO2 Diffusion During CWI

As CO2 gets transferred from CW to the oil phase, CO2 molecules diffuse inside the
oil due to the concentration gradient until the equilibrium state is reached. DIFFCOIL
keyword added in PROPS section provides CO2 -Oil diffusion coefficient between
adjacent grid cells. To study the effect of CO2 diffusion, three scenarios have been
simulated where in the first scenario the diffusion coefficient was set to zero while in
the second and third scenario the diffusion coefficient was set to 0.007758103 ft2/day
(1d) and 0.02327 ft2/day (3d) respectively. It was found that CO2 diffusion has a
minimal (negligible) impact on both oil recovery and CO2 storage. The following
figure shows the oil recovery for all the scenarios (68.4%) thus led to the same CO2
storage (100% as CO2 did not reach production well) (Fig. 8).

4 Conclusion

Carbonated water injection is a promising EOR technique that promotes high oil
recovery and safe CO2 storage. Upon the contact of Carbonated Water (CW) with
oil inside the reservoir, CO2 molecules cross the interface between the two fluids
moving towards the oil phase due to the concentration gradient, and subsequently,
CO2 molecules diffuse inside the oil until thermodynamic equilibrium is reached.
CWI was studied numerically using ECLIPSE300 where the CO2 SOL option was
used to account for CO2 dissolution inside the water phase. A sensitivity analysis
on injection rate and CO2 diffusion coefficient was done to understand their impact
on oil recovery and CO2 storage. It was found that low CW injection rates leads to
Carbonated Water Injection for EOR and CO2 Storage … 195

a higher oil recovery and a wider spatial spreading of CO2 inside the oil thus a higher
storage, while in the case of high injection rates, a less oil recovery was found and
the limited spreading of CO2 causes a lower CO2 storage. Additionally, the effect of
CO2 diffusion was found to be minimal (negligible) on both oil recovery and CO2
storage in this case study.

Acknowledgements The first author (AA) gratefully acknowledges the financial support of the
Ministry of Higher Education Malaysia and Universiti Teknologi PETRONAS under FRGS grant
(Grant No.: 015MA0-030).

References

1. Mosavat N (2014) Utilization of Carbonated Water Injection (CWI) as a Means of Improved Oil
Recovery in Light Oil Systems: Pore-Scale Mechanisms and Recovery Evaluation. University
of Regina, Faculty of Graduate Studies and Research
2. Mosavat N, Torabi F (2014) Application of CO2 -saturated water flooding as a prospective safe
CO2 storage strategy. Energy Procedia 63:5408–5419
3. Miller, J.S. and R.A. Jones. A laboratory study to determine physical characteristics of heavy
oil after CO2 saturation. in SPE/DOE Enhanced Oil Recovery Symposium. 1981. Society of
Petroleum Engineers.
4. Sohrabi M et al (2011) Safe storage of CO2 together with improved oil recovery by CO2 -
enriched water injection. Chem Eng Res Des 89(9):1865–1872
5. Sohrabi M et al. (2009) Mechanisms of oil recovery by carbonated water injection. in SCA
annual meeting
6. Sohrabi M et al (2011) Carbonated water injection (CWI)–A productive way of using CO2 for
oil recovery and CO2 storage. Energy Procedia 4:2192–2199
7. Sohrabi M et al (2012) Coreflooding studies to investigate the potential of carbonated water
injection as an injection strategy for improved oil recovery and CO 2 storage. Transp Porous
Media 91(1):101–121
8. Riazi M, Sohrabi M, Jamiolahmady M (2011) Experimental study of pore-scale mechanisms
of carbonated water injection. Transp Porous Media 86(1):73–86
9. Kechut, N.I., M. Sohrabi, and M. Jamiolahmady. Experimental and numerical evaluation
of carbonated water injection (CWI) for improved oil recovery and CO2 storage. in SPE
EUROPEC/EAGE Annual Conference and Exhibition. 2011. Society of Petroleum Engineers.
10. Tavakolian, M., et al. Significant improvement in oil recovery and CO2 storage by carbonated
water injection (CWI). in Third EAGE CO2 Geological Storage Workshop. 2012.
11. Yang D, Tontiwachwuthikul P, Gu Y (2005) Interfacial tensions of the crude oil+ reservoir
brine+ CO2 systems at pressures up to 31 MPa and temperatures of 27 C and 58 C. J Chem
Eng Data 50(4):1242–1249
12. Riazi, M., Pore scale mechanisms of carbonated water injection in oil reservoirs. 2011, Heriot-
Watt University.
13. Mosavat N, Torabi F (2016) Micro-optical analysis of carbonated water injection in irregular
and heterogeneous pore geometry. Fuel 175:191–201
14. Schlumberger, ECLIPSE Technical Description. 2014.
15. Schlumberger, ECLIPSE Reference Manual. 2014.
16. Whitson, C.H. and M.R. Brulé, Phase behavior. Vol. 20. 2000: Henry L. Doherty Memorial
Fund of AIME, Society of Petroleum Engineers ….
17. Kechut NI, Jamiolahmady M, Sohrabi M (2011) Numerical simulation of experimental carbon-
ated water injection (CWI) for improved oil recovery and CO2 storage. J Petrol Sci Eng
77(1):111–120
196 A. Abdulrahman and J. Foroozesh

18. Kechut, N.I., et al. Tertiary oil recovery and CO2 sequestration by carbonated water injection
(CWI). in SPE International Conference on CO2 Capture, Storage, and Utilization. 2010.
Society of Petroleum Engineers.
19. Embid S, Rivas O (1994) Simulation of miscible displacement with interphase mass transfer
resistance. SPE Advanced Technology Series 2(01):161–168
20. Mohebbinia, S. and T. Wong. Molecular diffusion calculations in simulation of gasfloods in
fractured reservoirs. in SPE Reservoir Simulation Conference. 2017. Society of Petroleum
Engineers.
21. Bear, J., Dynamics of fluids in porous media. 2013: Courier Corporation.
22. Peksa, A.E., K.-H.A. Wolf, and P.L. Zitha. Molecular Diffusion of CO From Carbonated Water
(CW) Into the Oil-Experimental Observations. in SPE Asia Pacific Oil and Gas Conference
and Exhibition. 2013. Society of Petroleum Engineers.
23. Alfarge D, Wei M, Bai B (2017) Factors affecting CO2-EOR in shale-oil reservoirs: numerical
simulation study and pilot tests. Energy & Fsuels 31(8):8462–8480
24. Grogan A, Pinczewski W (1987) The role of molecular diffusion processes in tertiary CO2
flooding. J Petrol Technol 39(05):591–602
25. Foroozesh J, Jamiolahmady M (2016) Simulation of carbonated water injection coreflood
experiments: an insight into the wettability effect. Fuel 184:581–589
26. Foroozesh J, Jamiolahmady M, Sohrabi M (2016) Mathematical modeling of carbonated water
injection for EOR and CO2 storage with a focus on mass transfer kinetics. Fuel 174:325–332
Study of Efficient FIR Filter Architecture
for Fixed and Reconfigurable
Applications

B. Aruna Devi, V. Prasanna Devi, and S. Preethi

Abstract Transpose type Finite-Impulse Response (FIR) filters are implicitly

pipelined and endorse the technique of Multifold Constant Multiplications (MCM).
The studied model explores the possibility of understanding block FIR filters in the
transpose form pattern for the efficient realization of broad order FIR filters for both
fixed and reconfigurable applications in area-delay. A flow graph optimized registry
complexity is built based on the computational study of transpose type configuration
of the FIR filter. The proposed translate type block filter is an architecture based on a
multiplier. Analysis is done towards the need of an optimum low complexity design
using the MCM model for the block realization of fixed FIR filters. Since power
dissipation is directly linked to the hardware, the use of MCM algorithms indirectly
achieves a certain amount of power reduction. MCM is therefore largely found to be
sufficient for the implementation of Broad Order design research. Using Data Flow
Graph, which has been comprehensively iterated based on the ability to execute
algorithms and their enhancement with a high degree of parallel and asynchronous
operation, the computation process in MCM is likely to be well performed.

Keywords Reconfigurable architecture · FIR filter · Block processing ·

Finite-impulse response (FIR) filter · VLSI

1 Introduction

High Complexity in Field Programmable Gate Array (FPGA) is the major demerit
for the program to execute in a running period which increases the delay duration,
which makes the parameter of power to be high. The complexity in the number of
slices used in Field Programmable Gate Array(FPGA) is increased in large numbers
and gradually increases the storage memory space. Earlier, the duration period of
high Look Up Table (LUT) were introduced, so that delay is increased in execution

B. Aruna Devi (B) · V. P. Devi · S. Preethi

Department of Electronics and Communication Engineering, Dr. N.G.P Institute of Technology,
Coimbatore, Tamil Nadu, India
e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 197
R. Ibrahim et al. (eds.), International Conference on Artificial Intelligence
for Smart Community, Lecture Notes in Electrical Engineering 758,
https://doi.org/10.1007/978-981-16-2183-3_18
198 B. Aruna Devi et al.

period of field arithmetic and one of the methods used is Distributed Arithmetic
(DA). Here the multiplications are rearranged and for the explicit multiplications
using Read Only Memory lookups, in analysis of FPGA based implementation, with
advantages of reduction in memory in FPGA. The timing of the DA based FIR filter
models is done using minimum clock periods depending on the delay of the final
adder stage in the PSAT (Pipeline Shift-Add Tree).However, which depends on the
critical path calculation of the adder. The throughput rate is the amount of output
produced by the critical path over a period of time.
In this paper, the study of various models of Finite Impulse Response Filter
with energy efficient architecture for reconfigurable applications is done. The paper
presents a comparative study of optimum low complexity, reconfigurable digital FIR
filter for low power with reduction in delay implementation.

2 Literature Survey

Many studies have been undergone towards FIR efficient implementation for various
applications [1]. The implementation of filter aids in proposing various models for
efficacy of power, delay, area and test bed analysis.
The key study deals with distributed FIR filter using the least mean block square
algorithm using Look Up tables, reducing the number of flip flops and adder. It
provides n/6 adders with reduction in area delay product, with block size tables less
than four block size tables. The implementation of the parallel structure with weighted
adaptation phrase, output filter computation and modified distributed arithmetic helps
to reduce the logic and power consumption of the look up tables with less words to
be updated per output. The algorithm includes decomposition into smaller vectors of
both input and weight vector, and filtering analysis of the inner product computation
[2].
Shah et al. [3] has implemented a competent FIR filter towards reconfigurable
implementation for an coherent FIR filter for fixed and reconfigurable applications
by implanting delay and area efficient addition based carry selector by optimizing
the superfluous operations, by performing the carry process before the ultimate sum
function analysis. The ability of the transpose form FIR filter implementation resulted
in 26.6% of reduction in delay and area of 15% in reconfigurable models.
An architecture based on distributed arithmetic FIR filter for increased speed
and less power consumption was proposed using block least mean square algo-
rithm. The implementation design provided an efficiency of 17.3 times, lower power
consumption of 17.3 times and 95% higher in terms of through output. The variable
coefficient adaptive proposed filter units includes the following operations: Register,
partial product computation, weight update and error generation, simultaneously
with shift accumulator look-up tables, emulate multiply and accumulate functions.
A systolic array architecture with parallel data processing units has been suggested
for the architecture. The parallelism was found to be successful with the number of
clock cycles reduced [4].
Study of Efficient FIR Filter Architecture … 199

The least-square linear phase FIR filter (LLFE) is frequently used in the analysis
of ECG signals. It aids in removal of low frequency mode using discrete wavelet
transform.
Distributed Arithmetic (DA) is a technique that is bit-serial in nature. It can there-
fore appear to be slow. It turns out that when the number of elements in a vector is
nearly the same as the word size, DA is faster in nature, reinstate the explicit multi-
plications overdone by read only memory look up table, by ROM look-ups and an
effective model to realize and to implement on Field Programmable Gate Arrays
(FPGAs). The coherent model is found effective for estimating of inner product or
multiply and accumulate (MAC). In Digital Signal Processing algorithms, the MAC
operation is popular. Multiplications are reordered and combined in DA so that the
arithmetic becomes “distributed” rather than “lumped” through the structure. When
DA is implemented in FPGAs, to execute the MAC operation, one can take advantage
of memory in FPGAs. When DA was devised (in the 1970s), the comparisons given
were in terms of number of TTL ICs required for mechanization of a certain type of
filter. In particular, for an eighth order digital filter operating at a word rate close to
1 MHz, 72 ICs with a total power consumption of about 30 W was stated with the
DA approach while 240 ICs with a power dissipation of 96 W was indicated for a
multiplier-based solution [5].

3 Proposed Work

FIR filter transpose type block was suggested with decreased register complexity. A
low-complexity design methodology that can be optimized for greater throughput
in reconfigurable applications for the block realization of fixed FIR filters using the
MCM scheme is proposed. It can be optimized for higher output in application that
are reconfigurable. A low-complexity design approach for the block implementation
of fixed FIR filters using the MCM scheme It can be configured for greater throughput
in reconfigurable applications as shown in Fig. 1. Here presented an optimum low
complexity design using the MCM scheme for the block implementation of fixed FIR
filters. MCM technique minimizes the number of adders and subtractors. As a result
the computation process is reduced. Since Power dissipation is directly related to the
hardware, some amount of power reduction is indirectly achieved by using MCM

Fig. 1 Proposed model

200 B. Aruna Devi et al.

algorithms. Therefore MCM is suitable for implementation of Large Order Filters

with fixed coefficients. On simulation of serial in serial out register, merged data flow
graph and the power analysis of the partial generator unit, the power consumption
is 40.05 mW compared to traditional model to 42.8 mW. The area product delay is
reduced around 13 % and energy per sample improvement is 12.8%. Hence in this
method, the number of flip flops, adders and multipliers are reduced for large order
filter which reduces the power and delay.

4 Conclusion

Finite Impulse Response (FIR) digital filter is extensively used in various applica-
tions ranging from biomedical applications, to speech processing in various domains
of digital signal processing. Finite-Impulse Response (FIR) Reconfigurable Filter
whose filter coefficients alter dynamically during runtime has an important role. The
multiplier unit is the optical FIR filter’s building structure in its transposed form.
With the filter input realization, the multiplication filter coefficients can be intro-
duced, having a major effect on the complexity and design efficiency. Hence calcu-
lation of inner product or multiply and accumulate can be done efficiently. Hence we
have explored the possibility of realization of block FIR filter in transpose form
configuration for area-delay efficient realization of large order FIR filters for fixed
and reconfigurable applications.

References

1. Reddy KS, Suresh HN (2019) FPGA implementation of reconfigurable FIR filter using Vedic
design with CLA adder. Int J Adv Sci Technol 28(1):144–161. Retrieved from http://sersc.org/
journals/index.php/IJAST/article/view/228
2. Ahmed F (2019) The energy efficient fir filter for biomedical application using FPGA. Comput
Rev J 3:236–244
3. Shah S, Rajula S (2019) Design of FIR filter architecture for fixed and reconfigurable applications
using highly efficient carry select adder. Adv Intell Syst Comput 898:627–637 (2019)
4. Chowdari CP, Seventline JB (2020) Systolic architecture for adaptive block FIR filter for
throughput using distributed arithmetic. Int J Speech Technol 23:549–557 (2020). https://doi.
org/10.1007/s10772-020-09745-4
5. Mohanty BK, Kumar P (2013) A high-performance energy-efficient architecture for FIR adaptive
filter based on new distributed arithmetic formulation of block LMS algorithm. IEEE Trans
Signal Process 61(4):921–932
Improved Deep Learning Based
Prediction of Crop Yield Using
Bidirectional Long Short Term Memory

V. Saveetha, R. Kingsy Grace, S. Sophia, and S. Birundha

Abstract Yield estimation for crops is an important problem to be addressed with

respect to agricultural planning. Deep learning and machine learning provide feasible
solutions for the above issue. In the current research perspective yield prediction is
vastly addressed and more techniques evolve in the objective to improve accuracy.
This research work tries to analyze the predictive accuracy of machine learning
methods along with deep learning methods. The models are analyzed by metrics such
as the Mean absolute percentage error (MAPE), root-mean-square error (RMSE),
mean absolute error (MAE), and accuracy. The proposed Bidirectional LSTM deep
learning technique exhibits high accuracy and low error rates in order to predict
rice yield in INDIA The results show that Bi-LSTM have better accuracy when
compared to other models, Bi-LSTM and XGB obtain the lowest RMSE errors, the
lowest MAPE error percentage is for Bi-LSTM and the lowest MAE errors is also
for Bi-LSTM. Since Bi-LSTM exhibits crop yield model with better accuracy and
lowest errors it is suitable massive crop yield prediction in agricultural planning.

Keywords Artificial Intelligence · Neural networks · Deep learning · Recurrent

neural networks · Long short term memory

1 Introduction

Crop yield prediction (CYP), a challenging task, benefits farmers in order to take
managerial decisions and financial stands [1]. CYP is challenging because of factors
like environmental and field based features. Also, the food production of the world is

V. Saveetha (B)
Dr. N.G.P. Institute of Technology, Coimbatore, India
e-mail: [email protected]
R. Kingsy Grace · S. Birundha
Sri Ramakrishna Engineering College, Coimbatore, India
S. Sophia
Sri Krishna College of Engineering and Technology, Coimbatore, India

© The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 201
R. Ibrahim et al. (eds.), International Conference on Artificial Intelligence
for Smart Community, Lecture Notes in Electrical Engineering 758,
https://doi.org/10.1007/978-981-16-2183-3_19
202 V. Saveetha et al.

dependent on CYP for precise forecast which leads to timely decisions [2]. Machine
learning (ML), that act has a stem of Artificial Intelligence, has been implemented
for agricultural prediction enormously in literature [3] due to its behavior of treating
outputs with respect to inherent function of the input features. Precision Agricul-
ture, the field of CYP manages to monitor measure and produce optimized outputs
with response to variability in crops [4]. Current research contributes vastly to esti-
mate the CYP justifiably but always a better performance is preferred [5]. Recently
ML techniques like Clustering, Classification, Regression and Neural Networks are
employed for CYP. In our study data of crop yield and other factors including statis-
tics and Deep Neural Networks (DNN) enables to produce effectual forecasting of
yield. DNN with hidden layers act to be more powerful to reveal the fundamental
nonlinear relationship between input and response variables [6]. In this paper Recur-
rent Neural Networks (RNN) [1] is used to predict and check crop yield. RNN works
with multiple stacked layers that convert the raw data into meaningful representa-
tions. RNN are well suited for modeling functions for which the data is composed of
vectors that involve a time dependency between the values like CYP [7]. It includes
a feedback loop that it uses to learn from sequences, including sequences of varying
lengths. The long short-term memory (LSTM) type of RNN has received enormous
attention in the realm of sequence learning. LSTM a popular variant of RNN has
internal memory to allow long-term dependencies to affect the output [8]. The LSTM
is good for CYP because it is designed to confine to the time dependencies of envi-
ronmental factors, it enables capability to predict crop yield without drop in accuracy
and combined with the bi directional method, the model predicts effectively.
This research aims to evaluate methods to detect rice cultivation in India, using
the LSTM and Bidirectional LSTM (Bi-LSTM) methods. The study compared the
results based on LSTM and Bi-LSTM with traditional machine learning techniques:
Support Vector Machines (SVM), Random Forest (RF) and Gradient Boost(XGB)
Regression. The rest of the paper is structured as Sect. 2 describes related work of
various models used for CYP. Section 3 presents the proposed methodology. Section 4
discusses the results and finally, concluding remarks is discussed in Sect. 5.

2 Review

Many methods exist in literature for CYP and every time an attempt is performed
by researchers to enhance the performance of prediction. In this paper an ensemble
based classification algorithm using Artificial Neural Networks (ANN), Decision
Tree and SVM is designed and implemented to predict crop yield. They try to increase
the accuracy of classification. The authors modeled an learning technique based on
AdaBoost ensemble algorithm and enhancement is performed using genetic opti-
mization technique [9]. Khaki and Wang suggested a DNN model for CYP across
from 2008 to 2016 which was analysed along with methods like Lasso, shallow
neural networks, and decsion tree [10]. You et al. designed a hybrid model with
Convolution Neural Network(CNN) and RNN to predict soybean yield [11]. Kim
Improved Deep Learning Based Prediction … 203

et al. suggested a DNN model for CYP using optimized input features from satellite
products and meteorological datasets from 2006 to 2015 [12]. Wang et al. [13] devel-
oped a Deep Learning(DL) framework to predict soybean crop yields in Argentina
and achieved good results with a transfer learning approach to predict Brazil soybean
harvests. Yang et al. (2019) analyzed the working of CNN to predict rice grain yield
on images and concluded that CNN model gives robust yield forecast through the
ripening stage [14]. This study shares a model to perform soybean yield forecasts in
Brazil using LSTM. They compare performance of linear regression, random forest
and LSTM for forecasting yield and evaluate accuracy [15]. The authors propose
a CNN-LSTM model for soybean yield prediction at the county-level. The model
was evaluated by crop growth variables and environment variables weather data.
They show the results of the experiment that performance of the proposed deep
CNN-LSTM model outperforms the CNN or LSTM model [16]. The above review
instigates that Deep learning can be used for CYP.

3 Methodology

Decision making in CYP require crop response models that make accurate estima-
tions to predict yields [17]. The purpose of the study is to find out whether RNN
can effectually predict crop yield for varying seasons, assess model performance
and compare the same with other models. The machine learning algorithms already
examined with respect to DNN are LSTM and CNN [8]. CNN is a specific class of
neural networks mostly used in image recognition. In this work Bi directional LSTM
is proposed for CYP in order to increase the performance accuracy. A comparison
of machine learning models and feature levels was performed based on the yield
estimation. Models were trained to predict yields from 1997 to 2015 for each state.
Model performance was evaluated using Root Mean Squared Error (RMSE), Mean
Absolute Percentage Error (MAPE), Mean Absolute Error (MAE) and Accuracy.
The input variables for the prediction model include the Crop yield year, Climatic
data and Area Cultivated. The yield prediction is based on the equation y = f(x),
where y is the yield predicted and x is the input data with respect to time.
Figure 1 shows the flow of the research performed. Initially Data Acquisition is
performed and the dataset obtained from [18] is explored for information following
which the data related to rice is extracted. Data preprocessing like encoding cate-
gorical variables, filling missing values and scaling are performed on the derived
dataset. The machine learning model and deep learning model are implemented and
the accuracy and error metrics for the same are analyzed.
LSTM works exclusively for prediction of sequence problems. The different archi-
tectures of LSTM are LSTM, CNN-LSTM, Encoder-Decoder LSTM, Generative
LSTM and Bi-LSTM. LSTM a RNN has an internal state, models parallel input
sequence, process variable-length input to generate variable-length output (Fig. 2).
The memory cell is the computational unit of the LSTM and consists of weights and
gates.
204 V. Saveetha et al.

Fig. 1 Methodology used

Fig. 2 LSTM architecture

Huang et al. [19] The objective of a Bidirectional LSTM is to make forward and
backward passes on the features of CYP in order to predict the yield. The past and
future features for a given time help to process the given sequence and thus provide
efficient implementation of the BDLSTM model. This model is trained using back
propagation through time (BPTT) [20] and the passes front and back are carried over
by unfolding the hidden states for all time steps. In fields like phoneme classification,
traffic prediction and speech recognition Bi-LSTM model are significantly better than
unidirectional ones and so an effort is made to use it in CYP [21]. The Bi-LSTM
Improved Deep Learning Based Prediction … 205

Fig. 3 Bidirectional LSTM

model (Fig. 3) generate an output Yt which is calculated by Yt = σ(hft , hbt ) where σ

function combines the two output sequences from forward pass and backward pass.

4 Result Discussion

4.1 Data Set

Historical data of Indian agricultural production on various states and districts

acquired from the Indian government web page available in [18] is used. This dataset
is utilized to investigate and analyze production contribution to particular land area.
Dataset used is a Comma Separated Values file. The dataset contained observed
production for rice and other crops between 1997 and 2015 across 646 districts from
33 states.

4.2 Metrics Used

The following indexes are selected to evaluate the forecast model performance: Accu-
racy, Mean absolute percentage error (MAPE), root-mean-square error (RMSE) and
mean absolute error (MAE) [9]. The MAPE is a measure of prediction accuracy in
statistics and is the most common measure used to forecast error, and works best if
206 V. Saveetha et al.

Fig. 4 Actual versus

predicted trend

Table 1 Formula of
Performance metrics Formula
performance metrics
Accuracy Accuracy = P−A
P
N 
 
 Ak −Pk 
MAPE M AP E = 1
N  Ak 
k=1


N
RMSE RMSE = 1
N (|Ak − Pk |)2
k=1


N
MAE M AE = 1
N |Ak − Pk |
k=1

A Actual value, P Predicted value

there are no extremes to the data (and no zeros). RMSE is the standard deviation of
the residuals errors. Residuals tells how concentrated the data is around the line of
best fit. The RMSE is a measure of the differences between values predicted by a
model and the values actually observed. It is used to compare forecasting errors of
different models for a particular variable and not between variables, as it is scale-
dependent. The values for the above metrics are obtained from Actual Values and
Predicted Values of yield after training and testing the dataset. The sample Actual
versus Predicted trend is shown in Fig. 4. Table 1 shows the formula employed to
find the metrics.

4.3 Comparison

Table 2 shows the performance metrics values for the models which are compared
with Bi-LSTM. The accuracy graph shows good percentage for Bi-LSTM and LSTM.
The MAPE percentage error is minimum for Bi-LSTM. The RMSE error values
Improved Deep Learning Based Prediction … 207

Table 2 Performance metrics

Model Accuracy % MAPE RMSE MAE
%
SVM 88 10.6 0.4601 9.54
RF 90.8 9.02 0.3280 2.98
XGB 90.5 9.60 0.2500 2.65
LSTM 91.6 8.6 0.3350 2.42
Bi-LSTM 93.0 7.8 0.2450 2.11

Fig. 5 Performance metrics

are good for both XGB and Bi-LSTM. The MAE values are good for Bi-LSTM,
LSTM, XGB and RF. The results show that Bi-LSTM is better in predicting CYP
when compared with other methods with respect to accuracy and minimized error
metrics which is shown in Fig. 5.

5 Conclusion

This study analyzed the performance of crop yield prediction using different models
like SVM, RF, XGB, LSTM and Bi-LSTM of which four are machine learning models
and two are Deep Learning models. The yield prediction for test set is reasonable
high in accuracy. The Bi-LSTM model outperforms other models both in accuracy
and stability. The Bi-LSTM computationally scales better than SVM, RF and XGB.
These results contribute to a comprehensive understanding of the response of rice
208 V. Saveetha et al.

yield to variables and the comparative performance of crop yield forecasts across
different types of machine learning algorithms. Further improvement of the yield
prediction modeling could benefit from more explicit information on crop type. The
proposed method shows great potential in improving the accuracy of yield prediction
for other crops like corn, wheat, and potatoes at fine scales in the future.

References

1. Saeed K, Wang L (2019) Crop yield prediction using deep neural networks. Front Plant Sci 10
2. Horie T, Yajima M, Nakagawa H (1992) Yield forecasting. Agric Syst 40(1–3):211–236
3. Van Klompenburg T, Kassahun A, Catal C (2020) Crop yield prediction using machine learning:
a systematic literature review. Comput Electron Agricult 177:105709
4. Patrício DI, Rieder R (2018) Computer vision and artificial intelligence in precision agriculture
for grain crops: a systematic review. Comput Electron Agric 153:69–81
5. Filippi P, Jones EJ, Wimalathunge NS, Somarathna PD, Pozza LE, Ugbaje SU, Jephcott TG,
Paterson SE, Whelan BM, Bishop TF (2019) An approach to forecast grain crop yield using
multi-layered, multi-farm data sets and machine learning. Precision Agric 20(5):1015–1029
6. Yang X, Sun M (2019) A survey on deep learning in crop planting. In: IOP conference series:
materials science and engineering, vol 490, no 6. IOP Publishing, p 062053
7. Khaki S, Wang L, Archontoulis SV (2020) A cnn-rnn framework for crop yield prediction.
Front Plant Sci 10:1750
8. Kang Y, Ozdogan M, Zhu X, Ye Z, Hain C, Anderson M (2020) Comparative assessment of
environmental variables and machine learning algorithms for maize yield prediction in the US
Midwest. Environ Res Lett 15(6):064005
9. Koduri SB, Gunisetti L, Ramesh CR, Mutyalu KV, Ganesh D (2019) Prediction of crop
production using adaboost regression method. J Phys Conf Ser 1228(1):012005. IOP Publishing
10. Khaki S, Khalilzadeh Z, Wang L (2019) Classification of crop tolerance to heat and drought—a
deep convolutional neural networks approach. Agronomy 9(12):833
11. You J, Li X, Low M, Lobell D, Ermon S (2017) Deep gaussian process for crop yield prediction
based on remote sensing data. In: 31rst AAAI conference on artificial intelligence
12. Kim N, Ha KJ, Park NW, Cho J, Hong S, Lee YW (2019) A comparison between major
artificial intelligence models for crop yield prediction: case study of the midwestern united
states, 2006–2015. ISPRS Int J Geo-Inf 8(5):240
13. Wang AX, Tran C, Desai N, Lobell D, Ermon S (2018) Deep transfer learning for crop yield
prediction with remote sensing data. In: Proceedings of the 1st ACM SIGCAS conference on
computing and sustainable societies, pp 1–5
14. Yang Q, Shi L, Han J, Zha Y, Zhu P (2019) Deep convolutional neural networks for rice grain
yield estimation at the ripening stage using UAV-based remotely sensed images. Field Crop
Res 235:142–153
15. Schwalbert RA, Amado T, Corassa G, Pott LP, Prasad PV, Ciampitti IA (2020) Satellite-based
soybean yield forecast: Integrating machine learning and weather data for improving crop yield
prediction in southern Brazil. Agricult Forest Meteorol 284:107886
16. Sun J, Di L, Sun Z, Shen Y, Lai Z (2019) County-level soybean yield prediction using deep
CNN-LSTM model. Sensors 19(20):4363
17. Ji B, Sun Y, Yang S, Wan J (2007) Artificial neural networks for rice yield prediction in
mountainous regions. J Agricult Sci 145(3):249
18. Indian Agricultural Production. https://data.gov.in
19. Huang Z, Xu W, Yu K (2015) Bidirectional LSTM-CRF models for sequence tagging. arXiv
preprint arXiv:1508.01991
Improved Deep Learning Based Prediction … 209

20. Cui Z, Ke R, Pu Z, Wang Y (2020) Stacked bidirectional and unidirectional LSTM recurrent
neural network for forecasting network-wide traffic state with missing values. arXiv preprint
arXiv:2005.11627
21. Bolton DK, Friedl MA (2013) Forecasting crop yield using remotely sensed vegetation indices
and crop phenology metrics. Agric For Meteorol 173:74–84
Instrument Cluster IoT Enclosure Design
and Production Implementation in Self
Driven Vehicles

R. Thirumalai , M. Seenivasan, A. Sivakumar , M. Vivekraj, and V. Balaji

Abstract MNC’s has introduced a Self-driven two-wheeler in the metropolitan

cities, so that vehicle consists of numerous electronics and the major patented is self-
control through smartphones. The most important electronic PCB board is embedded
in an enclosure. The industry develops the enclosure design and production imple-
mentation with the help of Solid works, the 3D enclosure was designed into tow
sub-assemblies, one is cover part and another one is the PCB carrier. The material
for enclosure was selected as per IPC standards. (PA66 30%GF). The enclosure
was also designed as self-locking even without screw. The fasteners (TWIN FAST
SCEW) were selected as per the IPC standards. The metal to non-metal torque rate
was analyzed and selected as per requirements. The mechanical endurance such as
Hardness test and Salt Spray Test were conducted for the selection of screw. The
enclosure was produced using Hot injection molding process.

Keywords IoT · Self driven vehicles · PCB board · Block chain

R. Thirumalai (B) · V. Balaji

Department of Mechanical Engineering, Dr N G P Institute of Technology, Coimbatore, Tamil
Nadu, India
e-mail: [email protected]
M. Seenivasan
Department of Mechanical Engineering, KGISL Institute of Technology, Coimbatore, Tamil
Nadu, India
A. Sivakumar
Department of Mechanical Engineering, Kongu Engineering College, Perundurai, Tamil Nadu,
India
M. Vivekraj
Department of Mechanical Engineering, Research Scholar, Anna University, Chennai, Tamil
Nadu, India

© The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 211
R. Ibrahim et al. (eds.), International Conference on Artificial Intelligence
for Smart Community, Lecture Notes in Electrical Engineering 758,
https://doi.org/10.1007/978-981-16-2183-3_20
212 R. Thirumalai et al.

1 Introduction

The bounce vehicle principle is based on IoT, the major controls of the vehicle
depends on the IoT. From the powering ON to Powering OFF everything is controlled
through IoT. So the major essential electronic parts are being embedded in the instru-
ment cluster enclosure, that is designed and produced by Squraesri technologies.
IoT security: Review, blockchain solutions, and open challenges are discussed [1,
2]. Modeling of information processing in the internet of things at agricultural enter-
prises and solar powered buildings are more successful [3, 4]. Design and implemen-
tation of vehicle navigation system in urban environments, automatic garage doors
and weather stations using internet of things (IoT) are becoming familiar nowa-
days [5–7]. Managing IoT devices using blockchain platform is developed in many
countries [8, 9]. Secure integration of IoT and cloud computing is described with
several case studies [10]. IoT middleware survey on issues and enabling technolo-
gies is demonstrated [11]. Internet of Things (IoT) is the network of physical objects
embedded with sensors, and computers which will helpful in exchange of data and
information with several systems. There are too many applications with IoT avail-
able such as; Automobiles, Heart monitoring implants, Bio-chips etc., IoT consists
of web enabled devices to collect and send data across the environment using sensors
and processors. A machine to machine communication acts on the information they
get from each other. The lifecycle of IoT is to collect, communicate, analyze and
act on the data information. The sensors are useful in collecting the information in
the manufacturing plant and a cloud platform is created to send the data and events
through a network used in the industry. The analyzing of the data includes filtering
the data, building reports and visualizing the data. The final lifecycle of IoT—Act is
useful in communication to the other machine and send successful notifications to
the other systems effectively and efficiently. IoT works with RFID, Sensors, Smart
technology and Nano Technology. The applications of IoT is wider in the areas of
Medical and Healthcare systems, Manufacturing, Building and home automations,
Transportation, Energy management, Environmental monitoring and so on. Also
the IoT faces several challenges; technological standardization, complexity in the
software, wireless communication, power supply, data volumes and interpretation
and so on. IoT provides dynamic control of the industry and daily life, integrates
human society and physical systems, improves the resource utilization ratio, flexible
configuration, technological integrator.

2 Methodology

The electronic PCB board dimensions were taken into considerations with keeping
all the electronic components embedded into it. After that the area in vehicle which is
allocated for the installation of the enclosure is inspected. An 2D rough draft or sketch
were drawn of the enclosure in simple drawing sheet with all necessary dimensions
Instrument Cluster IoT Enclosure Design … 213

that were taken before. The 3D design is done with keeping 2D sketch as a reference,
in Solidworks software. When the deigning stage is over, the material for production
is selected. Preceding the screw for fastener as well as the manufacturing method
were selected as per the IPC standards. After all the above steps, the production
implementation of enclosure is done by the hot injection moulding method. Then
finally all the electronics are embedded into the enclosure and installed in the vehicle.
The process flow chart methodology is shown in Fig. 1.

PCB board dimensions were measured(all necessary electronics

included).

The area allocated for this enclosure to be

installed in vehicle is inspected.

A 2D rough draft/sketch
was drawn initially

The 3D design is drawn in

solidworks software.

Screw/ material/ manufacturing meth-

od were selected.

Production implementation

Final installation in vehicle

Fig. 1 Methodology
214 R. Thirumalai et al.

3 Overview of the Development of Enclosure in Solidworks

Software

While designing the 3D sketch in solidworks software, the 2D rough draft is kept
as the reference which was taken from various parameters while drawing. As earlier
mentioned, the enclosure was divided into two parts, and each part was designed sepa-
rately and then assembled together. So while designing in software several commands
were used to design the enclosure. One of the important step in the design process
is selection of the materials. This systematic selection of the materials will give the
minimization of the cost during the production process. The design engineer should
be aware of the properties of the material and behavior of the materials under several
working conditions. The thermal properties, machinability properties, strength, hard-
ness and other mechanical properties are to be well known the designer. The material
used for the production of the enclosure is PA66 30% GF. PA66 GF30 is a 30% glass
fibre reinforced PA 66(Poly Amide).Poly (hexamethyleneadipamide). Polyamide 66,
or Nylon 66 (PA66) is an engineering-grade thermoplastic.
PA66 provides lower impact resistance when compared to PA6. High strength,
creep strength, toughness are the mechanical properties which are the outstanding
properties of PA66. PA66 is a glass fibre reinforced about 30 percentage to form
PA66 GF30. Glass fibres exhibit high static load withstanding properties at elevated
temperatures.

4 Results and Discussion

MNC’s has introduced a Self-driven two-wheeler in the metropolitan cities, so that

vehicle consists of numerous electronics and the major patented is self-control
through smartphones. The most important electronic PCB board is embedded in
an enclosure. The industry develops the enclosure design and production imple-
mentation with the help of Solid works, the 3D enclosure was designed into tow
sub-assemblies, one is cover part and another one is the PCB carrier. The material
for enclosure was selected as per IPC standards. (PA66 30%GF). The enclosure
was also designed as self-locking even without screw. The fasteners (TWIN FAST
SCEW) were selected as per the IPC standards. The metal to non-metal torque rate
was analyzed and selected as per requirements. The mechanical endurance such as
Hardness test and Salt Spray Test were conducted for the selection of screw. The
enclosure was produced using Hot injection molding process. The material selected
is tested for hardness and salt spray test.
Instrument Cluster IoT Enclosure Design … 215

4.1 Hardness Test

Hardness is the mechanical property which resists deformation which is plastic in

nature when induced by mechanical abrasion. Materials are basically stronger and
harder than plastics and wood. Strong intermolecular bonds are generally present
in harder materials and therefore the behavior of solid material under force gives a
tough behavior under variable loads. The material selected for the enclosure design
provides high hardness and it is found satisfactory.

4.2 Salt Spray Test

For testing the corrosion resistance of the material, salt spray test is performed. The
materials tested for salt spray test are metallic in nature. These materials are also
finished with a surface coating for providing better test results. The material that is
to be tested provides a corrosion protection to the undercovering metal. Corrosive
attacks are produced in the Salt spray testing method. This salt spray test will evaluate
the correctness of the coated material for further usage. The viscosity is very sensitive
to temperature. Shrinkage is of the order of 0.01–0.02 mm/mm [1–2%]. The addition
of reinforcing glass fibers reduces the shrinkage to 0.2–1%. Differential shrinkage in
the flow and cross-flow directions is quite high. Mineral fillers yield more isotropic
moldings. PA66 is resistant to most solvents, but not to strong acids or oxidizing
agents. The material selected for the enclosure design provides sufficient success
rate and hence it is found satisfied.

5 Conclusion

The IoT consist of four layers namely; sensor layer, Gateway and Network layer,
Management service layer and application layer. Sensor layer collects and process the
real time information. Gateway and network layer allows the organizations to share
the information’s through the robust networks available in the industry. Management
service layer ensures security and privacy of data. Application layer provides a user
interface for the IoT. The enclosure for electronics to be embedded in it has been
successfully designed and produced by injection molding process. The enclosure has
been installed in vehicle without causing any interruption to other parts and system
in electric vehicle. Hence, the enclosure was designed initially from 2D sketch to
3D model by using Solidworks software, followed by selecting appropriate material
for production and also the fasteners as well and finally fabricated using injection
moulding process.
216 R. Thirumalai et al.

References

1. Khan MA, Salah K (2018) IoT security: review, blockchain solutions, and open challenges.
Fut Gener Comput Syst 82:395–411
2. Bagas AD, Budi WA (2018) IoT-based integrated home security and monitoring system. In:
IOP conferences series: journal of physics: conference series, vol 1140
3. Lvovich IY, Lvovich YE, Preobrazhenskiy AP, Preobrazhenskiy YP, Choporov ON (2009)
Modeling of information processing in the internet of things at agricultural enterprises. IOP
Conf Ser Earth Environ Sci 315(3):032029
4. Handayani TP, Hulukati SA, Jaya R, Tiandho Y, Abdullah R (2019) The prototype of solar-
powered building lighting IoT. IOP Conf Ser Mater Sci Eng 486(1):012079
5. Godavarthi B, Nalajala P, Ganapuram V (2017) Design and implementation of vehicle naviga-
tion system in urban environments using internet of things (IoT). IOP Conf Ser Mater Sci Eng
225:012262
6. Jianyun C, Yunfan S, Chunyan L (2017) Research on application of automatic weather station
based on internet of things. IOP Conf Ser Earth Environ Sci 104(1):012015
7. Uriawan W, Zulfikar WB, Sofa RM, Ramdhani MA (2018) Internet of things for automatic
garage doors using {ESP} 8266 module. IOP Conf Ser Mater Sci Eng 434:12057
8. Huh S, Cho S, Kim S (2017) Managing IoT devices using blockchain platform. In: International
conference on advanced communication technology (ICACT), IEEE, pp 464–467
9. Khan MA, Salah K (2018) IoT security: review, blockchain solutions, and open challenges.
Futur Gener Comput Syst 82:395–411
10. Stergiou C, Psannis KE, Kim BG, Gupta B (2018) Secure integration of IoT and cloud
computing. Futur Gener Comput Syst 78:964–975
11. Ngu AH, Gutierrez M, Metsis V, Nepal S, Sheng QZ (2016) IoT middleware: a survey on issues
and enabling technologies. IEEE Internet Things J 4(1):1–20
General Adversarial Networks: A Tool
to Detect the Novel Coronavirus from CT
Scans

R. Shriram, T. R. Kaushek Kumar, V. Samuktha, and R. Karthika

Abstract Detection of the novel Corona virus in the early stages is crucial, since
no known vaccines exist. Artificial Intelligence- aided prognosis using CT scans
can be used as an effective method to identify symptoms of the virus and can thus
significantly reduce the workload on the radiologists, who have to perform this task
using their eyes. Among the most widely used deep learning convolutional neural
networks, research shows that the Xception, Inception and the ResNet50 provide the
best accuracy in detecting Covid-19. This paper proposes that using General Adver-
sarial Network (GAN) as a data augmentation technique, in combination with these
models will significantly improve the accuracy and thereby increase the chances of
detecting the same. The paper also compares and contrasts how each of the three
GANs namely DCGAN, LSGAN, CoGAN, perform in association with the afore-
mentioned models. The main aim of this paper is to determine the most credible GAN
network to carry out the task of data augmentation as well to prove that involving
GANs would improve the existing accuracy of our model, paving way for an effective
approach to train the model.

Keywords CT scans · Artificial intelligence · RT-PCR · Data augmentation ·

Xception · Inception · ResNet-50 · GAN · Generative network · Discriminative
network · Generator and discriminator training · Deep convolutional GANs · Fully
connected layers · Batch normalization · Coupled GAN · Tuples of GANs ·
Marginal distribution · Least squares GAN · PyTorch · Validation accuracy Versus
epoch

1 Introduction

The giant tide of SARS-CoV-2 virus, widely known as Covid-19 or novel corona
virus has swept off all of mankind in one fell swoop and with no known vaccine,

R. Shriram · T. R. K. Kumar · V. Samuktha · R. Karthika (B)

Department of Electronics and Communication Engineering, Amrita School of Engineering,
Amrita Vishwa Vidyapeetham, Coimbatore, India
e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 217
R. Ibrahim et al. (eds.), International Conference on Artificial Intelligence
for Smart Community, Lecture Notes in Electrical Engineering 758,
https://doi.org/10.1007/978-981-16-2183-3_21
218 R. Shriram et al.

rapid detection and diagnosis of COVID-19 is of utmost importance in the efforts to

improve the chances for a quick recovery. Detection usually involves taking samples
from the nose, throat and lungs or carrying out a blood test. The standard test for
infection is the Reverse transcription- Polymerase chain reaction (RT-PCR) test and
takes up to two days for results to be received.
There are two major problems with this method-the finite number of detection
kits available and valuable time lost to obtain the results.
Fortunately, tools aided by artificial intelligence (AI) can help in detecting novel
corona virus outbreaks. Although these tools are being used in hospitals to detect
mild cases, it is highly unlikely that these AI tools will replace RT-PCR as the
primary diagnostic tool. A nucleic acid test can be carried out by taking a swab in
any isolated location as per their convenience, while a chest scan has to be performed
in an enclosed space with staff nearby. Also, most of these tools offer an accuracy
of around 96% [1] which, given the sheer amount of COVID positive cases, is not
good enough. In a bid to improve the accuracy, we can incorporate more such deep
learning tools and techniques.
One such technique is the use of GAN’s (General Adversarial Networks) to better
train state-of-the-art models like Xception [2], Inception [3] and ResNet-50, culmi-
nating in better accuracy. GAN is a class of deep learning framework that incorporates
neural networks named generator and discriminator, contesting with each other in a
game-like situation. to add images to the dataset which are as superficially authentic
as possible. This leads to improvement in accuracy. GANs will be dealt with in detail
in Sect. 21.1. Section 2 deals with related work that have been carried out, Sect. 2
elaborates on the different GAN architectures dealt with in this paper, followed by the
experiment performed and then the paper is concluded with the result and analysis.

2 Related Work

A plethora of Deep Neural Networks have been designed to aid in the process of
detection since every new model may train a specific dataset in a better way. Some
of them include using a light CNN architecture based on SqueezeNet [4] for classi-
fication of CT scans that averages a computation time of just 7.81 s. Abdul Waheed
et al. [5] proposed an ACGAN (Auxiliary Classifier Generative Adversarial Network)
based GAN, called CovidGAN to generate synthetic Chest X-Ray images for data
augmentation the training sets along with a CNN model for better detecting COVID-
19. Another Deep Learning framework called the CovNet was proposed to identify
Covid-19 and also distinguish it from other Community-acquired pneumonia (CAP)
and non-pneumonic lung diseases that it can be confused with [6].
Various regression models including Lasso Regression model, Exponential
Smoothing model, etc. have been used to predict the number of COVID patients
to be anticipated in the future [7]. But these tools perform well only when they have
sufficient usable information, accumulating which is not easy. CNN models have
been used not only in classifying CT scans, but are also used in other domains such
General Adversarial Networks: A Tool to Detect … 219

as labelled Waste management [8] and detecting Breast cancer from mammographic
images by using the retrieved similar cases as reference [9].
Not only CNN’s but also VGGNet (which is a significantly more accurate CNN
[10]) has been used for classification purposes. Since 4D imaging is significantly
better in detecting diseases in motion-reliant organs like heart, lungs, etc. than 3D,
a variety of synchronization mechanisms including wavelet de-noising is used to
reconstruct the said 4D image[11].

3 GAN Architectures

All GAN architectures have a unique generator portion or a discriminator portion,

connected in that order.

3.1 GAN

GANs are basically neural networks that are trained in an adversarial manner to
generate a probability distribution of data that mimics the original probability distri-
bution. It consists of two models namely, the generator and the discriminator. Both
these models compete with each other with an objective of mimicking the original
distribution. The discriminator model performs the classification of whether a given
image is fake (labelled 0) or real (labelled 1). The generator generates images to fool
the discriminator by synthesizing artificial images that have a close resemblance to
the real ones, making it difficult for the discriminator to distinguish the fake (the one
generated from generator) from the real which is sampled from the true dataset. In
essence, the generative network is trained to maximize the final classification error
(between true and generated data), while the discriminative network is trained to
minimize it. Equilibrium is reached when the generator produces samples that follow
the original probability distribution and the discriminator predicts fake or not-fake
with equal probability. It is highly important that both networks learn equally during
training and converge together. So, during the training of the discriminator, we ignore
the generator loss and focus on the discriminator loss, which penalizes the discrimi-
nator for misclassifying real images as fake or vice-versa. Generator’s weights are not
updated. During the generator training, we use the generator loss, which penalizes
the generator for failing to fool the discriminator and generating an image that the
discriminator classifies as fake. The discriminator is frozen during generator training
and only generator’s weights are updated through back propagation.
Consider the training dataset as uniform distribution, p(X), a real image drawn
from it is taken as x and a random image in Rd to be z. Let G(z) and F(z) be the
generative and discriminative models, respectively. The generative model generates
an image as its output, G(z). Let p(G) be the probability distribution of G(z). The
discriminative model is tasked with computing the probability that an input image
220 R. Shriram et al.

drawn is real. Ideally, F(x) = 1 if x is real and F(x) = 0 if x is generated. The

GAN framework corresponds to a mini-max, with the objective function given by
the equation 1 [12].
   
max min V ( f , g) = Ex−px − log f(x) + Ex−px − log(1 − f(g(z))) (1)
g f

E(x) is the Expected value (or First moment) of a function. f and g are the
respective functions of the generator and the discriminator.
However, in practice this is solved by alternating these two gradient update steps:
 
step 1 : θ t+1
f = θ tf − λt ∇θ f V f t , g t (2)

 
step 2 : θgt+1 = θgt − λt ∇θg V f t+1 , g t (3)

where λ is the learning rate and θ ƒ and θ g are the parameters of F and G. Goodfellow
et al. [12] show that the distribution, p(G), converges to p(X).

3.2 DCGAN

Deep Convolutional GANs are essentially tailor-made GANs that have certain archi-
tectural constraints that help in stabilizing the learning process of standard GANs
by learning a hierarchy of representations from object parts to scenes in both the
generator and discriminator. Standard GANs are known for being unstable to train.
DCGAN utilizes logical steps towards preventing this instability such as incor-
porating the all convolutional net which replaces spatial pooling functions (like
Max Pooling), the benefits of which are explained by Springenberg et al. [13] and
allowing the generator to determine the down-sampling on its own by providing
strided convolutional layers.
It also includes removing fully connected layers and applying Batch Normaliza-
tion (technique by which each input is transformed to have unit variance and zero
mean) to all but the input layer of the discriminator and output layer of the gener-
ator. Another unique feature is that generators employ the ReLu activation layer
whereas the discriminator uses the Leaky ReLu activation function. Generator and
Discriminator of a DCGAN architecture is shown in Figs. 1 and 2.
General Adversarial Networks: A Tool to Detect … 221

Fig. 1 Generator of a DCGAN [14] with strided convolutional layers

Fig. 2 Discriminator of a DCGAN with strided convolutional layers with the same dimensions

3.3 CoGAN

Coupled GAN (CoGAN) was proposed as an alternative approach to learn a joint

distribution of multi-domain images from the data. It is a function that assigns prob-
abilistic densities to every joint occurrence of an image in different domains like
images portraying the same scenery but using different modalities or images of the
same face portraying varying emotions like happy, sad, etc. Once this joint distri-
bution is obtained, novel tuples of images can be generated by the GAN. Tradi-
tional approach dictates that to obtain this joint distribution, each set of multi-
domain images must be fed individually as tuples of images. Building such a dataset
consisting of tuples of corresponding images can be quite cumbersome, which greatly
diminishes the application and usability of this concept. CoGAN comes to the rescue
since it does not have this correspondence constraint. Rather it just requires a set
of images to be drawn from the marginal distributions of each domain of images.
The architecture of a CoGAN is based on that of a GAN. CoGAN incorporates
222 R. Shriram et al.

Fig. 3 Weight sharing in CoGAN [15]

a tuple of GANs for each individual image domain, which under normal circum-
stances produces a product of marginal distributions instead of a joint distribution.
This inconvenience is overcome by using a weight-sharing constraint that favors a
joint distribution result over a product of marginal distributions result. This forces
the individual GANs, which decode the high- level semantics of the image to share
the weights with the other GANs. Then the layers of these GANs responsible for
decoding the low-level features translate the shared representation exclusively to fool
their corresponding discriminators. The structure of a CoGAN having two image
domains is shown in Fig. 3 of CoGAN proposal [14]. Although only two domains
are shown, it can be easily extrapolated to as many domains as required. As discussed,
the first few layers of the generators, which are responsible for extracting high-level
semantics, are forced to share the parameters from a common subset.
Also, the last few layers of the discriminators, which are responsible for encoding
the high-level semantics, are enforced to share the weights. Note that the flow of
information is in opposite directions. Thus, CoGANs effectively synthesizes pairs of
images without the need for correspondence in the input dataset and thereby aiding
in better training the models.
Here the forced weight sharing is represented using a dotted line.As mentioned
earlier, the weight sharing occurs in the first few layers of the generator and the last
few layers of the discriminator.

3.4 LSGAN

The Least Squares GAN is a slightly modified variant of the normal GAN. Standard
GAN employs the ‘Sigmoid Cross-entropy’ loss function in its discriminators while
classifying the synthesized images from the real ones. This loss function may lead to
this undesirable vanishing gradient problem. This problem arises in gradient based
learning methods that update the weights based on the partial derivative (gradient)
of the error function, where in some cases, the gradient becomes vanishingly small
and thereby effectively stops the training process. LSGAN provides a solution by
changing the loss function to the least squares loss function. Study [16] shows that
General Adversarial Networks: A Tool to Detect … 223

Fig. 4 a The decision boundaries of the two loss functions, b Fake images that may lead to vanishing
gradients, c Effect of LSGAN [15]

this approach is better than the standard GAN in two aspects-generating higher
quality images and learning in a more stable fashion.Least squares loss focuses
on the distance of the image from the discriminator boundary and thus penalizes
generated samples that, although appear on the right side of the boundary, are too
far away from the real data and subsequently the boundary and thus may lead to
vanishing gradients. Figure 4 shows how the LSGAN forces the generated samples
(in magenta) to move closer to the decision boundary.

4 Experiment

For this paper we have used an open source dataset of SARS-CoV-2 CT scan [17],
containing 2482 CT scans in total (1252 positive CT scans and 1230 negative CT
scans). This data has been collected from hospitals in the locality of Sao Paulo,
Brazil.
All the three different GANs were implemented with the help of pyTorch, an open
source ML library. Each scenario consists of a validation phase and testing phase.
In the validation phase, 15% of images were used whereas in the testing phase, 25%
of images were used. All images have a dimension of 75*75*3. A sample for the
fake images generated by the DCGAN is shown in Fig. 5 and their corresponding
generator and discriminator losses with respect to the iterations are shown in Fig. 6.
These images were generated after 5000 iterations. Better quality can be obtained
by increasing the number of iterations.
In addition to the deep transfer models, we added 3 dense layers with sigmoid
activation function in the last layer to classify outputs accordingly. Number of epochs
used to train is 100.
As a first step of approach, we trained all three deep transfer models using the
actual dataset (without GAN). Later, we trained them using the dataset obtained
by concatenating the actual dataset with the datasets obtained from various GANs
(artificial images) individually.
224 R. Shriram et al.

Fig. 5 Real versus generated images for DCGAN

Fig. 6 Discriminator and generator loss versus iterations (DC-GAN)

5 Results and Discussion

CT scan is a computerized imaging technique to produce cross-sectional images

using X-rays directed at a patient and interpreting the output signal. There are some
visible attributes in the lungs of patients with COVID- 19 signs, such as ground-
glass “opacities” (darkened and blurry patches in the lung) and zones of greater lung
density called consolidation which are more recurrent, wide-spread and often with
a rounded appearance and peripheral lung distribution when compared to a healthy
person.
For evaluation, we used the confusion matrix which can be obtained explicitly
using inbuilt libraries. The details of the confusion matrix are shown in Table 1.
The below mentioned parameters (4), (5), (6) are used for evaluation along with
accuracy.
General Adversarial Networks: A Tool to Detect … 225

Table 1 Confusion matrix

GAN\Models ResNetSO InceptionV3 Xception
output for various GANs
DC-GAN [[532 34] [[541 5] [[533 13]
[ 27 528]] [ 31 544]] [ 6 569]]
CoGAN [[536 10] [[531 15] [[541 5]
[ 20 555]] [ 12 563]] [ 12 563]]
LS GAN [[536 10] [[536 10] [[541 5]
[ 5 570]] [ 15 560]] [ 11 564]]

True Positives
Precision = (4)
True positives + False Negatives
True Positives
Recall = (5)
True positives + False Negatives
Precision*Recall
F1 Score = (6)
Precision + Recall

Here a True Positive refers to a Covid patient being correctly identified as one
whereas a True Negative refers to a healthy person identified correctly as one. A False
Positive is a healthy person incorrectly identified as a Covid patient whereas a False
Negative is a Covid patient was not identified as one. The plots of Validation Accuracy
Vs epoch and Validation Loss Versus epoch for all three of the deep transfer models
in each case (without GAN as well with each GAN) have been provided. Xception
has the most deviation among the models considered but it also has the best validation
accuracy among them. The series of tables drafted below shows the brief analysis of
the results obtained (Fig. 7).
Table 2 shows the results obtained by testing the deep transfer models trained
without GAN implementation. As we can say, Xception tops the list with its accuracy
being 97.26% which is a considerably good result.
Because of Xception’s complex network architecture, which simultaneously trains
3–4 small networks and concatenates all the results together for further processing,

Fig. 7 Validation accuracy and validation loss plot for deep transfer models without GAN
226 R. Shriram et al.

Table 2 Test results of 3

Model / Results ResNetSO InceptfonV3 Xception
deep transfer models without
using GAN architecture Testing accuracy 94.52 96.3 97.26
Precision 0.928 0.948 0.96
Recall 0.981 0.981 0.987
FI score 0.948 0.965 0.974

it can learn more details than ResNet50 consisting of 50 layers, whose architecture
is more of a repetitive one. Same analogy goes with InceptionV3 as its architecture
is less complex when compared to Xception’s but more than ResNet50’s.
Table 3 shows the results obtained by testing the deep transfer models trained with
the LS-GAN dataset along with the real dataset. This significant improvement in each
deep transfer model’s results attests the use of augmentation, especially using GAN.
Here, ResNet50 tops the list with an accuracy of 98.66%. It has improved its accuracy
to 4.14% which is to be noted. Due to the increase in dataset size, ResNet50 has
a greater exposure resulting in its improvement. Xception performed exceptionally
well with its accuracy being 98.57% which is much comparable with ResNet50, with
an improvement of 1.31%. InceptionV3 also showed improvement in its accuracy
with a difference of 1.47% which is more than Xception’s (Fig. 8).
Table 4 shows the results obtained by testing the deep transfer models trained with
CoGAN implementation. Here, Xception tops the list in this analysis with its accuracy
being 98.48%, followed by InceptionV3 and ResNet50. All the deep transfer models

Fig. 8 Validation accuracy and validation loss plot for deep transfer models with LSGAN

Table 3 Test results of 3

Model / Results ResNet50 InceptionV3 Xception
deep transfer models using
LS-GAN architecture Testing accuracy 98.66 97.77 98.57
Precision 0.991 0.973 0.98
Recall 0.982 0.982 0.991
FI score 0.986 0.977 0.985
General Adversarial Networks: A Tool to Detect … 227

showed improvement in its scores with ResNet50 being the highest with an increase
of 2.8% (Fig. 9).

Table 4 Test results of 3

Model / Results ResNetSO InceptionV3 Xception
deep transfer models using
CoGAN architecture Testing accuracy 97.32 97.59 98.48
Precision 0.964 0.978 0.978
Recall 0.982 0.973 0.991
FI score 0.973 0.975 0.9845

Fig. 9 Validation accuracy and validation loss plot for deep transfer models with CoGAN

Table 5 Test results of 3

Model / Results ResNetSO InceptionV3 Xception
deep transfer models using
DC-GAN architecture Testing accnracy 96.52 96.79 98.31
Precision 0.97 0.946 0.989
Recall 0.958 0.991 0.976
FI score 0.964 0.968 0.9825

Fig. 10 Validation accuracy and validation loss plot for deep transfer models with DC-GAN
228 R. Shriram et al.

Fig. 11 Accuracy comparison

Table 5 shows the results obtained by testing the deep transfer models trained with
DCGAN implementation. Here, Xception again tops the list in this analysis with its
accuracy being 98.31%, followed by InceptionV3 and ResNet50. Here, DCGAN
has the least incremented results among the GANs taken into consideration, yet
shows a significant improvement when compared to results obtained without GAN
implementation (Fig. 10).
From the above analysis, we can say that models trained with LSGAN augmented
images proved to be the best among the GANs considered in this proposal. This is
due to the ‘least squared’ loss function which solves the vanishing gradient problem
confronted by the other GANs here, resulting in a greater stability. LSGAN works
with the advantage of penalizing those samples that have no gradients though they
have been classified rightly (Fig. 11).

6 Conclusion

In this paper, the detection of the novel coronavirus proved to be more efficient by
augmenting the already available dataset [16]. The fabrication of artificial images
was primarily based on the adversarial network GAN. The main aim of this paper is
to determine the most credible GAN network to carry out the task of data augmen-
tation as well to prove that involving GANs would improve the existing accuracy of
our model, paving way for an effective approach to train the model. From our anal-
ysis, ResNet50 (98.66%) and Xception (98.57%) trained with LS-GAN augmented
General Adversarial Networks: A Tool to Detect … 229

images provides the best testing accuracy. The best improvement between archi-
tecture implemented without GAN and with GAN can be observed in ResNet50
implementation with LS- GAN (4.14%). The pre-processing phase involved here
solves the issue of limited datasets available in the biomedical areas of research and
also aids in training the Deep Learning models more efficiently. The work carried
out in this paper could be further experimented by venturing into other recent and
robust versions of GANs.

References

1. Ardakani A, Kanafi A, Acharya A, Rajendra U, Nazanin K, Afshin M (2020) Application of

deep learning technique to manage COVID-19 in routine clinical practice using CT images:
results of 10 convolutional neural networks. Comput Biol Med. https://doi.org/10.1016/j.com
pbiomed.2020.103795
2. Chollet F (2017) Xception: deep learning with depthwise separable convolutions. In: 2017 IEEE
conference on computer vision and pattern recognition (CVPR), Honolulu, HI, pp 1800–1807.
https://doi.org/10.1109/CVPR.2017.195
3. Szegedy C, Liu W, Jia Y, Sermanet P, Reed S, Anguelov D, Erhan D, Vanhoucke V, Rabinovich
A (2015) Going deeper with convolutions. In: 2015 IEEE conference on computer vision and
pattern recognition (CVPR), Boston, MA, pp 1–9. https://doi.org/10.1109/CVPR.2015.729
8594
4. Polsinelli M, Cinque L, Placidi G (2020) A light CNN for detecting COVID-19 from CT scans
of the chest. https://doi.org/10.1016/j.patrec.2020.10.001
5. Waheed A, Goyal M, Gupta D, Khanna A, Al-Turjman F, Pinheiro PR (2020) CovidGAN: data
augmentation using auxiliary classifier GAN for improved Covid-19 detection. IEEE Access
8:91916–91923. https://doi.org/10.1109/ACCESS.2020.2994762
6. Li L, Qin L, Xu Z et al (2020) Using artificial intelligence to detect COVID-19 and community-
acquired pneumonia based on pulmonary ct: evaluation of the diagnostic accuracy. Radiology
296(2):E65–E71. https://doi.org/10.1148/radiol.2020200905
7. Wang S, Kang B, Ma J, Zeng X, Xiao M, Guo J, Cai M, Yang J, Li Y, Meng X, Xu B (2020)
Deep learning algorithm using CT images to screen for Corona Virus 2 disease (COVID19).
medRxiv 2020.02.14.20023028. https://doi.org/10.1101/2020.02.14.20023028
8. Sidharth R, Rohit P, Vishagan S, Karthika R, Ganesan M (2020) Deep learning based smart
garbage classifier for effective waste management. In: 2020 5th international conference on
communication and electronics systems (ICCES), IEEE, pp 1086–1089. https://doi.org/10.
1109/ICCES48766.2020.9137938
9. Kiruthika K, Vijayan D, Lavanya R (2019) Retrieval driven classification for mammographic
masses. In: 2019 International conference on communication and signal processing (ICCSP),
pp 0725–0729. https://doi.org/10.1109/ICCSP.2019.8698044
10. Saiharsha B, Diwakar B, Karthika R, Ganesan M (2020) Evaluating performance of deep
learning architectures for image classification. In: 2020 5th International conference on commu-
nication and electronics systems (ICCES), pp 917–922. https://doi.org/10.1109/ICCES48766.
2020.9137884
11. Ennesai S, Narayanankutty KA, Ganesan M (2010) Temporal alignment of non-gated image
sequences for 4D cardiac imaging using wavelets. In: 2010 International conference on
computer and communication technology (ICCCT), pp 198–200. https://doi.org/10.1109/
ICCCT.2010.5640532
12. Goodfellow I, Pouget-Abadie J, Mirza M, Xu B, Warde-Farley D, Ozair S, Courville A, Bengio
Y (2014) Generative adversarial nets. NIPS
230 R. Shriram et al.

13. Springenberg JT, Dosovitskiy A, Brox T, Riedmiller M (2014) Striving for simplicity: the all
convolutional net. arXiv preprint arXiv:1412.6806
14. Radford A, Metz L, Chintala S (2016) Unsupervised representation learning with deep
convolutional generative adversarial networks
15. Liu M-Y, Tuzel O (2016) Coupled generative adversarial networks. Mitsubishi Electric
Research Labs (MERL)
16. Mao X, Li Q, Xie H, Lau RYK, Wang Z, Smolley SP (2017) Least squares generative adversarial
networks. In: 2017 IEEE international conference on computer vision (ICCV), Venice, 2017,
pp 2813–2821. https://doi.org/10.1109/ICCV.2017.304
17. Soares E, Angelov P, Biaso S, Froes MH, Abe D K (2020) SARS-CoV-2 CT-scan dataset: a
large dataset of real patients CT scans for SARS-CoV-2 identification. medRxiv. https://doi.
org/10.1101/2020.04.24.20078584
VLSI Implementation of Multipliers
for Artificial Intelligence Applications:
A Survey

R. Karthick, S. Shanmuga Raju, G. Abinaya, G. Ashwanth, and G. Kaviya

Abstract As the competition among different multiplier models keeps rising we

have explored a 2-speed, radix-4, serial-parallel (SP) multiplier for stimulated appli-
cations like Machine Learning and Artificial Intelligence. The designed multiplier is
a variant of the modified radix-4 with serial-parallel (SP) Booth multiplier that skips
over the zero operations and adds only the ones in Booth encodings, making the
latency dependent critical paths are taken for process so that throughput and latency
are invoked higher for a subnet of multiplier values. We show that for bit widths of
32 and 64, our optimizations can result in a 1.42 times to 3.36 times improvement
with the standard parallel (P) Booth multiplier in terms of time—area depending on
the incoming set.

Keywords Multiplier · Speed optimized · Dynamic CMOS · Booth multiplier

1 Introduction

Manipulating data on multiplication is the most important milestone in applications

like Artificial Intelligence(AI), digital signal processing (DSP) and machine learning
(ML), detailing the delay, area, power and overall performance of parallel(P) imple-
mentations. An extensive work on optimizing the multiplication circuits, however,
the problem is that Wallace or dadda tree in combination with higher radixes on
modified Booth algorithms. Parallel-Parallel (P-P), Serial-Parallel(S-P) and serial-
serial(S–S) in which any one of the multiplication that is generally performed in
digital circuits. Modified Booth algorithm, in which we target serial-parallel(SP)
two-speed multiplier (TSM) that rigidly adds only the ones encoded parts of the
multiplication and bounce over the zero encoded portions.
In AI, DSP and ML implementations, the improvement in performance of a design
depends on the turned down precision representations, to gain computational accu-
racy we strive for the bottom-most possible bit width. Exactness is usually fixed at

R. Karthick (B) · S. Shanmuga Raju · G. Abinaya · G. Ashwanth · G. Kaviya

Department of Electronics and Communication Engineering, Dr.N.G.P Institute of Technology,
Coimbatore, India

© The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 231
R. Ibrahim et al. (eds.), International Conference on Artificial Intelligence
for Smart Community, Lecture Notes in Electrical Engineering 758,
https://doi.org/10.1007/978-981-16-2183-3_22
232 R. Karthick et al.

design time, and therefore, any changes in the requirements entail that further modi-
fication involves reconfiguration of the implementation. In cases where a tiny bit
span would be sufficient, the design runs at bottom efficiency since least mandatory
computation is taken on. To lighten this, mixed-accuracy algorithms intended to be
used in a lower bit span some partition of time, and a substantial bit space when
necessary. Two data paths that make the operation at different precisions which is
normally implemented in the circuit.
The target of the paper is to remove a portion of the computation completely and
significantly during the runtime which is a dynamic control structure. The goal is
accomplished using modified serial Booth multiplier, which is independent of loca-
tion and skips over encoded entire-zeroes or entire-ones computation. A parallel bit
arrangement of all operands both multiplicand and multiplier that is designed to give
a smart block which can be injected into existing AI bots, DSPs and ML. For certain
intake sets, the multiplier hits considerable striking improvements in computational
system performance. The main benefits of this method are as mentioned below.
(1) The data path of the booth multiplier has two sub-circuits. These circuits operate
at different critical paths.
(2) Implementing two critical paths in booth multiplier provides enhancement in
bit-pattern calculation and increases the throughput. Therefore its results in
reduced time and improved latency than conventional multipliers.
(3) The FPGA implementation of the proposed multiplier model for evaluating
the performance and comparison of throughput with conventional multipliers
(Fig. 1).

Fig. 1 Multiplier
architecture
VLSI Implementation of Multipliers for Artificial Intelligence … 233

2 Related Works on Multipliers for AI

2.1 Booth Multiplier, Dadda Multiplier, Wallace Multiplier

Vasudeva et al. [4] proposed a “Study of 8 Bits Fast Multipliers for Low Power
Application”. In VLSI the design of multipliers is the most complex architecture
block. Designing the multiplier may face challenges in optimizing power, delay and
area. When the number of stages is increased the width of the multiplier is also
increased which leads to complexity in Areas. For accuracy and high sampling rate
wide bit width is required. So the requirement for designing and developing an IP
for high speed applications. The design is proposed utilizing the basic multiplier
building blocks. In this paper we analyze the different types of 8-bit multiplier and
compare their performance. We are comparing three multiplies namely modified
Booth, Wallace and Dadda.
Modified Booth’s Multiplier:
Booth’s algorithm is commonly used for multiplication of signed numbers because
it is the most powerful algorithm to perform multiplication of signed numbers. This
multiplier frequently performs addition of A or S with P and performs arithmetic
right shift in P register. The algorithm of booth multiplier as follows:
x—multiplicand and y—multiplier
Step 1: fill the A with x bits and y + 1 zeros, similarly S with −x and y + 1 zeros
also p with x + y + 1 zeros
Step 2: check the last two bit of p
If it is 01 perform P + A.
If it is 10 perform P + S.
If it is 00,11 do no operations.
Step 3: Shift the values obtained to one bit right.
Step 4: Repeat step 2 and three y times.
Step 5: Eliminate the MSB of P, and the resultant is obtained.
Wallace Multiplier:
In Wallace multiplier the delay is always equivalent to logarithm of operand size.
The operation of wallace multiplication follows:
Step 1: The bits are grouped for the product.
Step 2: Using carry save adders the product matrix is reduced to a two rows
Step 3: fast carry propagate adder will add the remaining two rows and the final
product is obtained.
The Wallace tree is a family of column compression multipliers (Fig. 2).
Dadda Multiplier:
The Dadda multiplier with carry look ahead reduces the number of rows in the
multiplication process. The counter numbers are less in this method. This reduction
in counter placement will provide a reduction in partial product. The figure shows
the flow of the Wallace multiplier (Fig. 3).
234 R. Karthick et al.

Fig. 2 Wallace multiplier

Fig. 3 Operation 8 × 8 bits

dadda multiplier

After analyzing the performance parameters like speed, area and power consump-
tion of Booth, Wallace, and Dadda multiplier it can be concluded that Dadda multi-
plier is best in considering area, Booth multiplier provides better power optimization
and Wallace tree multiplier provides better delay optimization. When the number of
interconnections between the building blocks increases there will be increase in the
intermediate stages of the multiplier. This paper shows that the Dadda multiplier has
a 76.5% increase in interconnections as it has more number of inter-stages.
VLSI Implementation of Multipliers for Artificial Intelligence … 235

2.2 Vedic Multiplier

Sri Lakshmi et al. [1] proposed “Design and Implementation of High-Speed Vedic
Multiplier Using Hybrid Full Adder”. A novel multiplier architecture is introduced
which includes the algorithms like Array, Modified Booth, Dadda multipliers etc. A
multiplier based on Vedic mathematics is discussed in this paper. Vedic mathematics
is an ancient method of sutras for mathematical calculations. The special purpose
of Vedic maths is it reduces the complicated calculations by calculating the partial
products in prior before the beginning of actual operations of multiplication so that
the optimization in time and delay can be achieved. A novel hybrid vedic multiplier
approach has been discussed. In this paper a Vedic Multiplier with CSA is used to
increase the speed. The half adders and full adders in existing Vedic multipliers are
replaced by hybrid adder and compressor adders and comparison has been performed.
Compressors are the logic circuits that can add 3 or more bits simultaneously. Hybrid
full adder is also a kind of full adder with a lesser number of gates so that the area
is reduced. Therefore, by observing all the above results, we can conclude that the
hybrid adder is more power and delay efficient. Though the compressor has more
power and area efficiency compared with normal conventional full adder design,
hybrid adder is giving even better results compared with compressors. So, hybrid
adders give better performance among these three adders circuits (Fig. 4).
Compared with conventional adder the hybrid adder has a delay of 7.04%, power
of 11.12% and area has 1.46% respectively. In comparison of hybrid adder with

Fig. 4 Vedic multiplier

236 R. Karthick et al.

compressors the delay, power and PDP got reduced by 2.02%, 1.43% and 3.74%
respectively. Finally conclude that hybrid adder with vedic is gives better efficiency
compared with conventional full adder and compressors.

2.3 Wallace Tree Multiplier with Kogge Stone Adder

Sundhar et al. [5] proposed the “performance analysis of wallace tree-multiplier with
Kogge stone Adder using 15–4 compressor” (Fig. 5).
Wallace multiplier using full adder and half adders to bring down the partial
product from three to two rows and then two partial products are added using the
final adder. In this paper an approximate 16 × 16 bit Wallace multiplier using 15–
4 compressor architecture is designed and if we compare the performance of this
multiplier with Kogge stone adder, it is using the same architecture of multiplier
using parallel adder. This architecture infers that kogge stone adder is faster when

Fig. 5 Overall architecture of 16 × 16 Bit wallace tree

VLSI Implementation of Multipliers for Artificial Intelligence … 237

we are compared to the multiplier with parallel adder since Kogge stone adder is
basically a parallel prefix adder. This adder has the fastest addition based on design
time. Hence this multiplier occupies about 15% of total coverage area and dissipated
power and delay are also highly reduced.

2.4 Array Multiplier

Devi et al. [6] proposed “low power optimized Array multiplier with reduced area”.
In this paper a 32-bit design of power and reduced area optimized unsigned array
multiplier is proposed. This multiplier is a 32bit array multiplier which makes use of
Carry Select Adders (CSAs) to execute partial product additions. This adder’s area
is expanded but delay is reduced. Ripple Carry Adders (RCAs) are the adders that
have a closely packed area but they are the slowest type of adders (Fig. 6).
On the other hand, carry look ahead adder (CLA) is the high-speed adder but they
are the worst from the area point of field of vision. Hence carry select Adder (CSA)
have been considered as the best comparing with RCA and CLA because they offer
a good counterbalance between compact areas of RCAs and short delay of CLAs.
The array multiplier with CSA uses modified booth algorithm with greatest power

Fig. 6 Array multiplier with carry select adder

238 R. Karthick et al.

compression technique to minimize the power dissipation upto 22% and reduction
in area is also achieved.

2.5 Array Multiplier with OTFC

Aruna et al. [2] proposed “A low power and high speed Array multiplier using on-
the-fly conversion”. In this multiplier design with single precision truncation, the
product of 2 N-bits produces 2 N but partial products, excluding these 2 N bit partial
products, are going to be divided into 2 N-(n/2) bits and N/2 bits. As a result finally,
2 N bits are created by adding the above bits using ripple carry adder (Fig. 7).
The array multiplier outlined in this paper is designed and enforced with no
truncation or addition approach; instead it is executed using a typical array multiplier
scheme. The proposed array multiplier produces the high order bit (MSB) of the
final product. The multiplier design leverages on the fly conversion converter that is
implemented as the tail end of the multiplier. This is to achieve the expedited carry
propagation in the last leg of multiplication. The OTF conversion logic is achieved to

Fig. 7 Array multiplier with OTFC

VLSI Implementation of Multipliers for Artificial Intelligence … 239

Table 1 Comparing existing multiplier

S. No. Multiplier Delay (ns) Power (mW) Area
1 Booth multiplier 3.75 3.24 214 LUTs
2 Dadda + CLA multiplier 2.56 5.65 878 µm2
3 Carry select multiplier 46.23 155.4 4034 LUTs
4 Variable carry skip multiplier 40.6 153.7 3970 LUTs

Table 2 Comparing recent multipliers

S. No. Multiplier Delay (ns) Power (mW) Area
1 Vedic + Full adder multiplier 13.64 173.2 3653 LUTs
2 Vedic + Hybrid adder multiplier 12.12 170.71 3396 µm2
3 Array multiplier 3.14 9.89 5349 LUTs
4 Radix-4 S-P multiplier 21.28 4.35 87 LUTs
5 Wallace multiplier 3.125 0.42 1570 µm2

provide constant delay irrespective of input bits processed. Hence the array multiplier
with OTF conversion has a minimum array, smaller area and less energy.

3 Comparison on Existing Multipliers

See Table 1.

4 Comparison on Recent Multipliers

See Table 2.

5 Resultant Efficiency Chart

6 Conclusion

The proposed multiplier is divided into two sub circuits which operate at different
critical paths. The design of CMOS Multiplier that will be highly adaptable to AI,
ML and DSP kind of applications. The Multiplier is designed for bit width of 32 and
64, expecting an improvement of 3 × and 3.56 × respectively. On comparing the
240 R. Karthick et al.

standard parallel multiplier, this design gives 3.64 × improvement in area and delay.
Upcoming works on application will take complete feature of two speed optimization.

References

1. Sri Lakshmi Y, Vigneswaran T (2020) Design and implementation of high-speed vedic multiplier
using hybrid full adder. Int J Adv Sci Technol 29(3):10663–10669
2. Aruna S, Venkatesh S (2019) A low power and high speed array multiplier using on fly
conversion. Int J Recent Technol Eng 7(5S4)
3. Moss DJM, Boland D, Leong PHW (2019) A two-speed, radix-4, serial-parallel multiplier. IEEE
Trans VLSI Syst
4. Vasudeva G, Cyril Prasanna Raj P (2015) Study of 8 bits fast multipliers for low power
application. Int J Soft Comput Eng 5(1). ISSN: 2231–2307
5. Sundhar A, Tharshini SD, Priyanka G, Ragul S Saranya C (2019) Performance analysis of
wallace tree multiplier with Kogge Stone Adder using 15–4 compressor. In: 2019 International
conference on communication and signal processing (ICCSP), Chennai, India, 2019, pp 0903–
0907
6. Devi P, Singh G, Singh B (2011) Low power optimized array multiplier with reduced area.
In: Mantri A, Nandi S, Kumar G., Kumar S. (eds) High performance architecture and grid
computing. HPAGC 2011. Communications in Computer and Information Science, vol 169
Automated Boneage Analysis Using
Machine Learning

J. K. Krithika, Yayha Norashikin, K. Porkumaran, and S. Prabakar

Abstract Bone age assessment is done to analyse the skeletal maturity according
to their chronological age. This is done by radiograph method considering the left
hand or wrist. Bone age assessment is done by comparing the chronological age to
assess the endocrine disorders and pediatric syndromes. Earlier the manual method
was used, where the radiologists compare the radiograph image with the atlas and
estimate the age of the bone. In this study, the analysis and classification of the x-ray
image of the left hand is experimented to determine the bone age. Here the bone
age analysis method involves the segmentation of the image, feature extraction and
classification using support vector machine(SVM). The results obtained is future
used to assess the skeletal abnormalities.

Keywords Support vector machine · Feature extraction · Hand X-ray image ·

Region of interest

1 Introduction

Bone is connective tissue. There are 300 soft bones which are formed during the birth,
then the cartilage grows and slowly replaced by hard bone. During the adolescence
period there are 206 bones formed. These bones have their own blood vessels and
made up of living cells. The longitudinal growth in long bones is due to process

J. K. Krithika (B)
Dr.N.G.P Institute of Technology, Kalapatti Road, Coimbatore, India
Y. Norashikin
UniversitiTeknologi Petronas, Seri Iskandar, Perak, Malaysia
e-mail: [email protected]
K. Porkumaran
Sri Sairam Engineering College, Chennai, India
S. Prabakar
Sona College of Technology, Salem, India
e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 241
R. Ibrahim et al. (eds.), International Conference on Artificial Intelligence
for Smart Community, Lecture Notes in Electrical Engineering 758,
https://doi.org/10.1007/978-981-16-2183-3_23
242 J. K. Krithika et al.

of endochrondral ossification and width growth is by the development of skeletal

tissue directly from the fibrous membrane [1]. Bone age assessment is done by
comparing the chronological age, which is used for diagnosis of endocrine disorders
and pediatric syndrome. Osteoporosis is a skeletal disease, which is caused due to
low mineral density in the bone. This is diagnosed by measuring the bone mineral
density (BMD) [2]. The Radiograph images are used as they are quite safe, the
exposed radiation is less. The use of age assessment in social development is to
identify the details about the refugee, this also used in criminal cases and also for
forensic reports [3].

2 Problem Statement

Bone age helps to evaluate growth but the children does not mature at the same time,
which makes it difficult to evaluate the growth. Usually the bone age assessment
depends on radiologist experience, it is carried out by visual evaluation of the skeletal
development using GP atlas. In manual approach, bone age assessment varies from
one scientist to another. It is also time consuming and may lead to errors.
In this research method, it will focus on analysis of the left hand x-ray image. This
is done by taking the ROI part and Extracting the image. This research will result
in development of automated bone age prediction algorithm using machine learning
technique.

3 Objectives

The assessments that will be carried out are about the bone age analysis from x-ray
image radiograph, to achieve few goals:
1. To analyze the bone from the x-ray image at various level of age.
2. To develop and evaluate bone age prediction algorithm by using machine
learning technique.

4 Literature Review

Osteoporosis is a skeletal disease due to low bone density; this is diagnosed by

measuring the bone mineral density. Bone age analysis was also used for evaluating
the growth and diagnosis of pediatric syndrome and endocrine diseases.
There were different methods for determining the bone age. CT was used but it
was not preferred as its radiation is high. MRI is being developed, but requires more
research. Ultrasonography is used but it is not accurate as radiographic methods [3].
In ultrasound method 2 transducers are used, one produce ultrasonic waves and other
Automated Boneage Analysis Using Machine Learning 243

one is the receiver. Radiograph method where considered as the predominant method
for bone age analysis. A study was conducted by using the mammography x-ray,
which has the high image quality. The method used was dual energy x-ray absorp-
tiometry (DEXA) and sonography. The DEXA is more accurate but it is expensive
and radiation emitted is more. It is used in spine and hip. Sonography method is
radiation free and its cost is low. It is usually used to assess the bone mineral density
of peripheral bones (heel, knee cap) [5].
Studies were not only done in hand bone but few studies done the bone age analysis
by dental maturity, it was used for forensic purposes, but not used in diagnosis of
endocrine diseases. The other method is by iliac bone, it was determined based on the
degree of maturation of the iliac crest apophysis, but this was a drawback one as the
ossification is not uniform in iliac crest apophysis. The other method is by femoral
head, which is the alternative way by assessing the depth of the epiphysial cartilage
of femoral head which is continuously being ossified. Ossification is complete and
most of the cartilage is replaced by bone remaining cartilage, which is called as
hyaline articular cartilage. The bone age analysis was first done manually in tooth,
later in bones such as skull, foot bones, knees, spinal cord, pelvis, rib, femur, carpals
and epiphyses of the ulna and radius. The main characteristics of the bone age
assessment are fusion of bones. The problems in manual approach is time consuming,
it varies from scientist to scientist, it may also lead to error and during surgery
over estimation or under estimation of bone length may lead to deformity. But in
automated bone age analysis, according to Dinesh M.S the radiograph must have
preprocessing, segmentation, image enhancement, rotation of bone pixels to proper
angel, measurement analysis and decision analysis [6].
An attempt using the SVM classification method has been introduced by Markus
Harmsen et al. [4]. The semi-automatic BA analysis developed by combining the
support vector machine with cross correlation to a prototype image for each class.
The evaluation of BAA is presented by comparing the nominal and real- valued
SVM with k nearest neighbor classification. According to this automated BAA,
there are some step that have present: (1) Extract 14 epiphyseal regions from the
radiographs; (2) retain image features using the image retrieval in medical application
framework; (3) Use these features to build a classifier model (training phase); (4)
Evaluate performance on cross validation (testing phase); (5) Classify unknown hand
image (application phase). These all steps have been implement by using C++ and
SQLite database version 2.8.17. Therefore, based on the 1097 hand radiographs
of 30 diagnostic classes in range 0–19 years, the nominal and real-value SVM are
resulting 91.57% and 96.16% accuracy respectively. However, in all the study done,
can observe that the SVM classification method have the higher accuracy and state
that the simple classification for the BA analysis. However, some improvement of
feature extraction is required as to achieve the BA assessment more efficient and can
classify to the age level categories.
244 J. K. Krithika et al.

Detection of ROI
(midfinger)

Middle finger
Calculation of the im-
age properties of mid-
finger

Classes (5-
12yrs)

SVM

Classes (13-
18yrs)

Fig. 1 Simplified performance model of automatic bone age analysis

5 Proposed Work

The detection and feature extraction of the hand bone x-ray of different image is
the significant work done in the bone age analysis. The input images are taken from
the website, and feature is extracted and the trained data is feeded into the classifier
(Fig. 1).

5.1 Input Image

The image dataset of the left hand is taken from the website kaggle.com. The test
dataset and trained datasets are downloaded, about 10,000 and above datasets for
each dataset. The dataset of the images is obtained with the CSV, which contains the
information regarding the person’s age, sex that are used for training.

5.2 ROI Extraction

The first stage used for segmentation. The middle finger is labeled using the toolbox
in mat lab for all 120 images. Middle finger is considered because of prominent
Automated Boneage Analysis Using Machine Learning 245

growth of bones. After labeling of the image the data is exported into the workspace
in tabular form.

5.3 Feature Extraction

Feature extraction is the next stage. From the ROI image the mean, standard deviation,
maximum value, energy, variance is determined. Normalization was also considered
and the output was compared between before normalization and after normalization.

5.4 Classification

The feature extraction will undergo the classification method. All the classifier was
used to find the better classifier for classification. From the work it was found that,
linear and quadratic svm has better classification accuracy.

5.5 Result of Training the Dataset for 120 Images

See Figs. 2, 3, 4, 5, 6 and 7; Table 1.

Fig. 2 Quadratic SVM after normalization

246 J. K. Krithika et al.

Fig. 3 Quadratic SVM before normalization

Fig. 4 Linear SVM after normalization

From the 120 images it is found that before normalization, the accuracy is 89.2%,
which is greater when compared to after normalization. Because during normal-
ization the raw data of the image may be missed. For less number of images, the
parameters used for feature extraction was enough and from this it is proven that
linear classifier gives the good result.
Automated Boneage Analysis Using Machine Learning 247

Fig. 5 Linear SVM before normalization

Fig. 6 Cubic SVM after normalization

5.6 Result of Training the Dataset for 336 Images

See Figs. 8, 9, 10, 11, 12 and 13; Table 2

For more no of images, the accuracy after normalization and before normalization
does not vary much, but comparing to 120 images, the accuracy obtained in this is
less. Because for higher no of images the parameters used such as mean, standard
248 J. K. Krithika et al.

Fig. 7 Cubic SVM before normalization

Table 1 Accuracy obtained

Classifier name Accuracy % (after Accuracy %
for 120 images
norm) (before norm)
Linear SVM 87.5 89.2
Quadratic SVM 86.5 88.3
Cubic SVM 85 85
Fine Gaussian SVM 86.7 86.7
Medium Gaussian 85 84.2
SVM

deviation, energy, minimum value, maximum value, variance for feature extraction
is not enough. The accuracy always dependent on the feature extracted.

6 Classifier Result

Once the extracted features are fed into classifier, it is trained and the result is gener-
ated as a code in the workspace. From the generated code, the Validation accuracy and
validation prediction of the images are obtained. For testing the obtained program,
more than 10 images are labeled again and feature is extracted from it and fed into
the code generated for the result Fig. 14.
Automated Boneage Analysis Using Machine Learning 249

Fig. 8 Linear SVM before normalization

Fig. 9 Linear SVM after normalization

From the figure the result is obtained. The accuracy for classification is 89.2% as
the code generated is by the linear classifier at accuracy 89.2%. 8 images for each
class was fed., Totally 16 images were fed into the generated code and the validation
accuracy is 0.8750 and for class 1, only 1 image was predicted wrong and for class
2, 1 images where predicted wrong.
250 J. K. Krithika et al.

Fig. 10 Quadratic SVM before normalization

Fig. 11 Quadratic SVM after normalization

7 Conclusions

In conclusion, the bone age reflects the skeletal maturity when differing significantly
from chronological age. There were a lot of method to automate the analysis of the
bone age that will be time consume for the radiologist in analysis the radiographs of
Automated Boneage Analysis Using Machine Learning 251

Fig. 12 Cubic SVM before normalization

Fig. 13 Cubic SVM after normalization

bone age but in this analysis, the reliable feature extraction is the most important key
point for analyze the development of the bone age. However, in the implementation
of the image processing for feature extraction as inspected ROI image have been
done as the proposed methodology. The method contains the detection of ROI using
252 J. K. Krithika et al.

Table 2 Accuracy obtained

Classifier name Accuracy % (after Accuracy %
for 336 images
norm) (before norm)
Linear SVM 81.5 81.3
Quadratic SVM 83.5 83.6
Cubic SVM 80.7 81
Fine Gaussian SVM 80.4 81.8
Medium Gaussian 81.0 81.5
SVM

Fig. 14 Classified result of the given test images

parameters like mean, variance, energy and then the images are trained in the SVM
classifier.

References

1. Lee H, Tajmir S, Lee J, Zissen M, Yeshiwas BA, Alkasab TK, Choy G (2017) Fully automated
deep learning system for bone age assessment. Springer, Aug 2017
2. Vicente G, Osman R, Hand bone age. Springer
3. Mughal AM, Hassan N, Ahmed A (2014) Bone age assessment methods. PMC, Jan–Feb 2014
4. Harmsen M, Fischer B, Schramm H, Seidl T, Deserno TM (2013) Support vector machine
classification based on correlation prototypes applied to bone age assessment. IEEE J Biomed
Health Inform 17(1):190–197
5. Huo Z, Jyang CY, Purdum M, Bone Mineral density assessment mammography system. United
States patent patent no.: us 7.965,813 b2 hu0 et al
6. Bakthula R, Agarwal S, Automated human bone age assessment using image processing
methods. Int J Comput
Multi-Class SVM Prediction Model
for Lung Cancer Diagnosis

D. Lakshmi, J. Sivakumar, and S. Ramani

Abstract Detection of lung abnormalities is essential in the field of medical diag-

nostics in order to reduce mortality. Early and correct diagnosis of these lung abnor-
malities leads to timely and appropriate treatment thereby reducing the mortality.
This area of research always faces a challenging task to differentiate the cancerous
tissues from non-cancerous tissues using CT images. This paper presents the SVM
prediction model for the characterization of lung tissues namely fibrosis, suspicious
of TB and carcinoma. This model is designed with three set of features formed with
Gray-Level Co-Occurrence Matrix (GLCM) and Gray-Level Run- Length Matrices
(GLRLM) and is evaluated using True Positive Rate, False Negative Rate, Posi-
tive Predictive Value, False Discovery Rate. The classifier model shows trustworthy
results with classification accuracy ranging between 93.9% and 98.6% respectively.

1 Introduction

In India, Non-Communicable diseases lead to mortality rate of 9% and 11% due to

cancer and chronic respiratory issues [1]. General clinical practice includes various
non-invasive techniques in the management and diagnosis of lung abnormalities.
These techniques offer great potential in the disease prediction and treatment. Devel-
oping countries conduct the lung cancer screening with a higher risk of diagnostic
procedures even for benign tissues and extreme treatment for over diagnosed cancers
[2]. Recently, integrated back scattering technique in endobronchial ultrasonography
has been used in the lung cancer diagnosis [3]. Bayesian Network (BN) predicts
better than SVM for two-year survival in non-small cell lung cancer patients with
missing variables. The logistic regression model to the BN has an Area Under Curve

D. Lakshmi (B) · J. Sivakumar

Department of Electronics and Communication Engineering, St. Joseph’s College of Engineering,
OMR, Chennai 600119, India
S. Ramani
Department of Computer Science and Engineering, Panimalar Engineering College, Poonamallee,
Chennai 600123, India

© The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 253
R. Ibrahim et al. (eds.), International Conference on Artificial Intelligence
for Smart Community, Lecture Notes in Electrical Engineering 758,
https://doi.org/10.1007/978-981-16-2183-3_24
254 D. Lakshmi et al.

(AUC) of 0.71, 0.75 and 0.76 for the Ghent, Toronto and Leuven datasets respec-
tively [4]. Selection of textural features is another major challenge in the character-
ization of lung tissues. Interobserver variability in texture analysis including lesion
morphology, image acquisition parameters, segmentation parameters, the number of
texture features and the category must be dealt that result in satisfactory outcome in a
clinical trial [5–9]. The role of non-linear classifiers in medical imaging diagnostics
is significant in the non-invasive characterization of lung abnormalities. Multiple
kernels based SVM has nowadays become effective in the characterization and
staging of lung diseases from CT images.

2 Literature Survey

Cancer is a multifaceted disease and many automated cancer detection systems are
proposed in literature. A predominant research work is based on classification of later
stages of lung tumors commonly termed as malignant. Early detection of cancer also
known as prognosis is not straight- forward as information about cancer biomarkers
are not known clearly. Detection of cancer from CT images involves feature extraction
and classification using machine learning algorithms such as ANN, Decision Tree,
Linear Discriminant Analysis, SVM, Nearest Neighbor classifier and Bayes Clas-
sifier. Recent literature reports that SVM classifier outperforms other conventional
classifier systems in lung cancer detection. SVM performs a two-class classification
by transforming the input data to higher dimensional feature space through a linear or
a non-linear mapping function. It identifies a decision boundary known as hyperplane
that maximizes the separation between the input variables in feature space. Keshani
et al. [10] proposed a SVM based lung nodule detection system from CT images. The
classifier was trained on 2D stochastic and 3D anatomical features to detect the nodule
areas. They reported an overall detection accuracy of 89% and false discovery rate of
7.3% per scan. Thomas et al. [11] proposed a data fusion technique to combine the
genomic and proteomic data for breast cancer detection. They proposed a weighted
average Least Squares SVM (LS-SVM) classifier for cancer classification. They
report classification performance of various conventional classifiers based on ROC
analysis. Shen et al. [12] proposed a bidirectional chain coding scheme and used SVM
classifier to segment the lung nodules from CT images. They reported a segmentation
accuracy of 92.6% with the proposed scheme. Huang et al. [13] compared the clas-
sification performance of standalone SVM and SVM ensemble classifier in breast
cancer analysis. They reported that SVM ensemble classifier based on RBF kernel
function outperformed other classifi-ers used in the analysis. Froz et al. [14] proposed
a combination of textural features based on artificial crawlers and used SVM classi-
fier for differentiating lung nodules from non-nodules of CT images. They reported
a mean accuracy of 94.3% and AUC value of ROC curve of 0.922 for RBF kernel
SVM classifier. A study done by Wang et al. [15] compared several classical machine
learning methods like Random Forest(RF), Support Vector Machine(SVM), Adap-
tive Boosting (Ada-Boost), Back Propagation Artificial Neural Network(BP-ANN)
Multi-Class SVM Prediction Model for Lung Cancer Diagnosis 255

with Convolution Neural Network(CNN) architectures for the classification of lung

nodules from metastasis of Non-Small Cell Lung Cancer(NSCLC). The effect of
textural features over diagnostic features in the discrimination of these lung nodules
is compared and the results are very attractive. Orozco et al. [16] presented RBF
kernel based SVM classifier utilizing wavelet features for the clinical data taken
from LIDC and ELCAP data base and obtained an accuracy ranging between 81 and
100. Makaju et al. [17] proposed SVM model to classify the lung nodules into benign
or malignant with a maximum accuracy of 86.6%. Hassen et al. [18] experimented
stepwise Forward Selection and Principal Component Analysis (PCA) based SVM
for the classification of lung diseases. Hybridizing PCA, FCM and SVM improves the
classification accuracy. Ajin and Mredhula [19] analyzed the behavior of combining
Radial Basis Function and Polynomial Function in the classification of different
Interstitial Lung Disease (ILD) patterns. Combining these two functions referred as
Hybrid kernel based SVM is able to differentiate these ILD patterns with accuracy
better than the ANN, K- Nearest Neighbor and Deep CNN. Optimizing the SVM
kernel function may greatly reduce the prediction time and [20] obtains the classifi-
cation accuracy of 99.17% for 70 number of nodules. A retrospective study done by
Dennie et al. [21] evaluated 55 patients differentiating primary lung cancer with gran-
ulotamous nodules and results in an accuracy of 64.8%. Mohana et al. [22] proposed a
method based on Nearest Neighbour for the detection of lung cancer. GLCM features
are extracted and using these features the cells are classified as benign and malignant.
Nearest Neighbour classifier achieved an accuracy of 98.76%. Senthil and Shubha
[23] implemented a method based on Back Propagation Network for lung cancer
detection. Seven features are selected using Principal Component Analysis (PCA).
The parameters and weights for BPN are optimized using Ant Lion Optimization
(ALO). Their experimental results shows an accuracy of 87.1%, 93.2% and 99.1%
in Stage I, Stage II and Stage III respectively.

3 Proposed System

3.1 Subject Selection and Dataset

Our Cohort includes 17 subjects of multi-slice chest Low-Dose CT images. The

specification of CT images includes tube voltage of 120kV, tube current ranging
between 120 mA and 220 mA and slice thickness between 0.3 mm and 0.7 mm and
is stored in Digital Imaging and Communications in Medicine (DICOM) format.
Dataset comprises of individual slices of each patient expressed as 512 x 512 matrix
with a 8- bit gray-level stored in Bit-Map format. The number of slices representing
healthy person is 123, fibrosis is 103, carcinoma is 185 and the suspicious of tuber-
culosis case is 50. The other cases include fibrosis and the suspicious of tuberculosis
cases. This work uses histopathological test as the ground truth for categorizing lung
tissues. This paper proposes SVM prediction model for the differentiation of various
256 D. Lakshmi et al.

lung tissues namely suspicious of TB, fibrosis and carcinoma. The work aims at the
implementation of highly accurate SVM prediction model with very less prediction
time.

3.2 Methodology

As shown in Fig. 1 our classification procedure begins with unsupervised segmenta-

tion of lung tissues from 440 number of low dose CT images comprehensively. GLCM
involves extracting first order features specifically contrast, correlation, homogeneity
and energy. These features are computed along the four directions given by angle θ
resulting in 16 number of features. Entropy is also calculated for the tissues inside the
lung parenchyma and outside the lung parenchyma defining model with 18 features.
Similarly, 11 GLRLM features are computed across 4 four directions defining
model with 44 features [24]. The features extracted are subsequently trained and
validated with the different SVM kernels namely linear, quadratic, cubic and gaussian
for characterization of the lung cancer with respect to other abnormalities as shown

Fig. 1 Methodology
Multi-Class SVM Prediction Model for Lung Cancer Diagnosis 257

in Fig. 1. The strengths of our proposed work are 1. Processes the low-dose CT
image comprehensively. 2. Prevents human intervention during diagnosis. 3. Prior
knowledge of ROI is not required.

4 Experimental Results and Discussion

Prediction of SVM model is evaluated in terms of accuracy and prediction time with
respect to kernels and type of feature set. Many researchers use SVM for binary
classification problem and shown satisfactory performance. But, the use of binary
SVM is restricted for multi-class problems. Our three class SVM classifier model
performance with the three feature sets namely GLCM, GLRLM, both GLCM and
GLRLM are evaluated by the Confusion matrix as in Tables 1, 2 and 3 respectively. It
is well observed that the feature set type and the size contribute the classifier perfor-
mance to a greater extent. Classifier Model with 62 features reduces misclassification
of carcinoma and other cases significantly.

4.1 Performance of SVM Classifier Model Based on Features

Our proposed system implements three class SVM prediction model trained with
GLCM based features, GLRLM based features and both GLCM and GLRLM based

Table 1 Confusion matrix

Actual/predicted Healthy Other cases Carcinoma
for model with 18 features
Healthy 113 – –
Other cases – 127 15
Carcinoma – 7 178

Table 2 Confusion matrix

Actual/predicted Healthy Other cases Carcinoma
for model with 44 features
Healthy 113 – –
Other cases – 135 7
Carcinoma – 14 171

Table 3 Confusion matrix

Actual/predicted Healthy Other cases Carcinoma
for model with 62 features
Healthy 113 – –
Other cases – 137 5
Carcinoma – 5 180
258 D. Lakshmi et al.

features. Table 4 shows the classifier performance parameters namely true positive
rate, positive predictive value, false negative rate and false discovery rate respec-
tively. Table 5 presents the accuracy and prediction time of SVM classifier with
gray-level co-occurrence matrix parameters and gray level run length matrix with
respect to different kernels like linear, quadratic, cubic and Gaussian kernels respec-
tively. Classifier accuracy is influenced by the selection of the different feature set
numbering 18, 44 and 62 respectively. Moreover, the maximum accuracy is obtained
for quadratic and cubic kernels in both the glcm and glrlm based feature set. The
model categorizes the three cases with an accuracy of 95, 95.2 and 97.7% respectively.

4.2 Prediction Time for Different Kernels

Our experimental results clearly demonstrate that the proper selection of the infor-
mation data from the training dataset and exploiting that information in the predic-
tion function optimization reduces the prediction time. Figure 2 shows Prediction
time of SVM with different kernel function gives interesting results that model with
glcm features consumes maximum time of 7.65s compared to 1.05s for model with
glrlm features. Figure 3 gives the Quadratic, Cubic and Gaussian kernel functions
with GLRLM based features able to predict with a less timing of 1.11s, 1.069s and
1.246s compared with GLCM based features of 1.269s, 1.49s and 1.167s respec-
tively. Combining both GLCM and GLRLM features shown in Fig. 4 does not yield
significant change on prediction time. Finally, SVM prediction model with Quadratic
and Cubic Kernels proves that GLRLM features is efficient in the characterization
of the lung tissues namely fibrosis, suspicious of TB and carcinoma resulting in an
accuracy of 98.6% each as shown in Fig. 5.

4.3 Multi-class SVM Classifier Model Prediction Accuracy

The classifier models are usually evaluated by their prediction accuracy. Our three
class SVM classifier model is assessed by its prediction accuracy and is also compared
with all the binary class SVM state of the art as shown in Table 6. Many researchers
implemented Binary class SVM classifier model with GLCM and GLRLM features
yielding less accuracy. Our proposed three-class SVM Classifier model shows
significantly high value of prediction accuracy.

5 Conclusion

We developed three class SVM prediction model with 5 fold cross-validation for
the differentiation of lung cancer from other lung tissues like suspicious of TB and
Table 4 SVM classifier performance parameters with three feature set
True positive rate Positive predictive value False negative rate False discovery rate
Performance Healthy Other Carcinoma Healthy Other Carcinoma Healthy Other Carcinoma Healthy Other Carcinoma
Parameters Cases (in Cases Cases (in Cases Cases Cases (in Cases (in Cases Cases (in Cases (in Cases Cases (in
%) (in%) %) (in%) (in %) %) %) (in %) %) %) (in %) %)
Model with 100 89 96 100 95 92 – 11 4 – 5 8
18 features
Model with 100 95 92 100 91 96 – 5 8 – 9 4
44 features
Multi-Class SVM Prediction Model for Lung Cancer Diagnosis

Model with 100 96 97 100 96 97 – 4 3 – 4 3

62 features
259
260 D. Lakshmi et al.

Table 5 SVM prediction response for GLCM based features with respect to kernel function
Kernel GLCM GLRLM Both GLCM and GLRLM
function Accuracy Prediction Accuracy Prediction
(in %) time (insec) (in %) Time (in Accurac y Prediction
sec) (in %) time (in
sec)
Linear 95 7.614 95.9 1.0518 98.2 7.2094
Quadratic 96.4 1.2694 98.6 1.1142 97.7 1.363
Cubic 96.4 1.492 98.6 1.0694 97.7 1.1665
Gaussia n 94.8 1.167 93.9 1.2463 94.3 1.313

Fig. 2 Prediction time of

SVM with glcm features for
different kernel functions

Fig. 3 Prediction time of SVM with glrlm features for different kernel functions

Fig. 4 Prediction time of

SVM with different kernel
functions for both glcm and
glrlm features
Multi-Class SVM Prediction Model for Lung Cancer Diagnosis 261

Fig. 5 Variation in accuracy of SVM Prediction model with three datasets

Table 6 Comparison of performance of SVM classifiers for different features

Author Features Images Classifier/class Accuracy (in %)
Hongkai Wang GLCM based PET/CT SVM/binary 69.23 ± 8.57
al et (2017) features
Hiram Madero Wavelet features CT SVM-RBF/binary 81–100
Orozco et al.
(2015)
Suren Makaju Area, perimeter, CT SVM/binary 86.6
et al. (2018) centroid, diameter,
eccentricity, mean
intensity
Donia Ben Circularity, Chest Hybrid 74.19
Hassen et al. skewness, Radiographs SVM/binary Fuzzy
(2016) kurtosis, entropy,
correlation,
homogeneity,
lacunarity,
localisation
Ajin M and Linear ternary CT Hybrid kernel 90.52
Mredhula L co-ocurrence based SVM/ binary
(2017) pattern (RBF and
Polynomial
function
Tao Zhou et al. Shape features, CT SVM/binary 99.17
(2016) intensity features,
texture features
Carole Dennie Correlation, sum Contrast SVM/binary 60
et al. (2015) of squares, sum Enhanced- CT
variance, sum
entropy, difference
Variance
(continued)
262 D. Lakshmi et al.

Table 6 (continued)
Author Features Images Classifier/class Accuracy (in %)
Our proposed GLCM based CT SVM/three class 95.0
work features – linear
– Quadratic 96.4
– Cubic 96.4
GLRLM based – Gaussian 94.8
features SVM/three class 95.9
– Linear
– Quadratic 98.6
– Cubic 98.6
– Gaussian 93.9
Both GLCM SVM/three class 98.2
features and – Linear
GLRLM – Quadratic 97.7
based features
– Cubic 97.7
– Gaussian 94.3

fibrosis. It is concluded that quadratic and cubic kernels for GLRLM based features
yield a maximum accuracy of 98.6% with prediction time of 1.23sec. Combining
these features and predicting with linear kernel increases the accuracy to 98.2%.
The main advantage of our Classifier model requires no prior knowledge of ROI
for segmentation, where the system process the low-dose CT comprehensively. The
work may be enhanced for the detection of metastasis of the lung abnormalities by
including dataset contributing images of different abnormalities.

References

1. http://www.who.int/nmh/countries/ind_en.pdf?ua=1
2. Veronesi G, Novellis P, Voulaz E, Alloisio M (2016) Early detection and early treatment of
lung cancer: risks and benefits. J Thorac Dis 8(9):E1060–E1062
3. Ito F, Kawasaki M, Ohno Y et al (2016) Noninvasive tissue characterization of lung tumors
using integrated backscatter intravascular ultrasound. An Ex Vivo Comp Study Pathol Diagnos
Chest 149(5):1276–1284
4. Hatt M, Majdoub M, Vallières M et al (2015) F-FDG PET uptake characterization through
texture analysis: investigating the complementary nature of heterogeneity and functional tumor,
volume in a multi-cancer site patient cohort. J Nucln Med 56:38–44
5. Jayasurya K, Fung G, Yu S, Dehing-Oberije C, Ruysscher D, Hope D (2010) A Comparison
of Bayesian network and support vector machine models for two-year survival prediction in
lung cancer patients treated with radiotherapy. Med Phys 37(4)
6. Fried DV, Tucker SL, Zhou S, Liao Z, Mawlawi O, Ibbot G, Court LE (2014) Prognostic value
and reproducibility of pretreatment CT texture features in stage III non–small cell lung cancer.
Int J Radiat Oncol Biol Phys 90:834–842
Multi-Class SVM Prediction Model for Lung Cancer Diagnosis 263

7. Doumou G, Siddique M, Tsoumpas C, Goh V, Cook GJ (2015) The precision of textural analysis
in 18 F-FDG-PET scans of oesophageal cancer. EurRadiol 25:2805–2812
8. Tixier F, Hatt M, Le Rest CC, Le Pogam A, Corcos L, Visvikis D (2012) Reproducibility of
tumor up take Heterogeneity characterization through textural feature analysis in 18 F- FDG
PET. J Nucl Med 53:693–700
9. Nyflot MJ, Yang F, Byrd D, Bowen SR, Sandison GA, Kinahan PE (2015) TU-AB- BRA-
04: quantitative radiomics:Sensitivity of PET textural features to image acquisition and
reconstruction parameters implies the need for standards”. Med Phys 42:3587
10. Keshani M, Azimifar Z, Tajeripour F, Boostani R (2013) Lung nodule segmentation and
recognition using SVM classifier and active contour modeling. A complete intelligent system,
Comput Biology Med 43:287–300
11. Thomas M, De Brabanter K, Suykens JAK, De Moor B (2014) Predicting breast cancer using
an expression values weighted clinical classifier. BMC Bioinformatics 15. https://doi.org/10.
1186/s12859 014- 0411–1
12. Shen S, Bui AAT, Cong J, Hsu W (2015) An automated lung segmentation approach using
bidirectional chain codes to improve nodule detection accuracy. Comput Biology Med 57:139–
4. https://doi.org/10.1016/j.compbiomed.2014.12.008
13. Huang MW, Chen CW, Lin WC, Ke SW, Tsai CF (2017) SVM and SVM ensembles in breast
cancer Prediction. PLoS ONE 2017;12:e0161501. https://doi.org/10.1371/journal.pone.016
1501
14. Froz BR, de Carvalho Filho AO, Silva AC, de Paiva AC, Nunes RA, Gattass M (2017) Lung
nodule classificatio using artificial crawlers, directional texture and support vector machine.
Expert Syst Appl. https://doi.org/10.1016/j.eswa.2016.10.039
15. Wang H, Zhou Z, Li Y, Chen Z, Lu P, Ang E, Wanyu Liu W (2017) Comparison of machine
Learning methods for classifying mediastinal lymph node metastasis of non-small cell lung
cancer from F-FDG PET/CT images. EJNMMI Res 7:11.https://doi.org/10.1186/s13550-017-
0260-9
16. Orozco HM, Villegas OOV, Sánchez VGC, Domínguezz HDJO. Alfaro (2015) Automated
system for lung nodules classification based on wavelet feature descriptor and support vector
machine, BioMedical Eng OnLine 14:9.https://doi.org/10.1186/s12938-015-0003-y
17. Makaju S, Prasad PWC, Alsadoon A, Singh AK, Elchouemi A (2018) Lung cancer detection
using CT scan images. Procedia Comput Sci 125:107–114
18. Hassen DB, Zakour SB, Taleb H (2016) A hybrid fuzzy-SVM classifier for automated lung
diseases diagnosis. Polish J Med Phys Eng 22(4):97–103.https://doi.org/10.1515/pjmpe-2016-
0017
19. Ajin M, Mredhula L (2017) Diagnosis of interstitial lung disease by pattern classification. In:
7th International conference on advanc in computing & communications, ICACC -2017, 22–24
August 2017, Cochin, India, Procedia Computer Science vol 115, pp 195–208
20. Zhou T, Lu H, Zhang J, Shi H (2016) Pulmonary nodule detection model based on SVM and CT
image feature-level fusion with rough sets. Hindawi Publishing Corporation BioMed Research
International, vol 2016, Articl ID 8052436, 13 p. https://doi.org/10.1155/2016/8052436
21. Dennie C, Thornhill R, Carolina V-V, Souza A, Gupta A, Maziak D, Bayanati H (2016) Role
of quantitative computed tomography texture analysis in the differentiation of primary lung
cancer and granulomatous nodules. Quant Imaging Med Surg 6(1):6–15
22. Mohana RM, Devi RDH, Bai A (2019) Lung cancer detection using nearest neighbour classifier.
Int J Recent Technol Eng (IJRTE) 8(2S11). ISSN: 2277-3878
23. Senthil S, Shubha BA (2019) Improving the performance of lung cancer detection at earlier
stage and prediction of reoccurrence using the neural networks and ant lion optimizer. Int J
Recent Technol Eng (IJRTE) 8(2). ISSN: 2277–3878
24. Lakshmi D, Santosham R, Ranganathan H (2014) Automated texture based characterization
of fibrosis and carcinoma using low-dose lung CT images. Int J Imaging Syst and Technol
24(1):39-44. Online ISSN: 1098-1098.https://doi.org/10.1002/ima.22077
Survey on Fire Safety Robot &
Implementation of Android Application

K. Sakthisudhan, S. Rubika, S. Sadhasivam, V. Suguna,

and Sri Muruga Rajan

Abstract Detection and extinguishing of fire are dangerous job that puts the life of
fire fighters in danger. Using robots to do these things will reduce the loss of lives.
The roles of current systems are limited in certain ways so it can be improvised to a
quicker and effective method. This paper offers an overview of a fire fighting robot
with toxic gas calculation. Several gas sensors are used to determine the volume of
poisonous gases such a carbon monoxide, carbon dioxide and ammonia. The fire is
detected using a flame sensor and the information is processed by the processor and
the procedure to extinguish it is initiated. Data and informatics is the essential factors
for effective firefighting operation.

Keywords Fire-fighting · Android application · Carbon dioxide · Carbon

monoxide and ammonia

1 Introduction

Fire battle is a dangerous job, but firefighters are still putting their lives on line against
fire. The mortality rate of firefighters in service is high. The temperature steadily
increases when a fire threat arises. The harmful gases such as carbon dioxide and
carbon monoxide appear to impact humans. Fire fighters can also be replaced by
robots that can do the same thing safely. This article reviews and describes prob-
lems and criteria for robot-based firefighting and presents prospects of productive
fighting in the future. The programmed robot locates the fire and automatically extin-
guishes it. If the fire is not detected, the mobile application may monitor it manually.
With mobile device, the robot can be run easily. It tests the safe concentrations of
poisonous gases such as carbon dioxide, carbon monoxide and ammonia in the burn.
This helps us to consider the strength of the fire and the wellbeing of victims. A
mobile network for autonomous firefighting can track and manage a particular area
for the occurrence of fire, and locate the exact place of the fire, extinguish the fire and

K. Sakthisudhan (B) · S. Rubika · S. Sadhasivam · V. Suguna · S. M. Rajan

Dr.N.G.P. Institute of Technology, Coimbatore 638401, India

© The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 265
R. Ibrahim et al. (eds.), International Conference on Artificial Intelligence
for Smart Community, Lecture Notes in Electrical Engineering 758,
https://doi.org/10.1007/978-981-16-2183-3_25
266 K. Sakthisudhan et al.

return safely after completion of the extinction. The obstacle monitoring is carried
out using A / D conversion in the microcontroller to convert analogue output from
the ultrasonic sensor to digital, and it is compared with the pre-specified threshold
value for the AFFMP to decide whether there are any obstacles too close [1]. The
Snake Fighter Anna Konda can drive the fire away from outer obstacles apart from
a flat floor, and can use the nozzle that is fixed to the front with a hydraulic medium
in joint action to extinguish the fire. It is a combined use of water to power, cool and
extinguish hydraulic joints [2]. Robots will reach the area of the enemy and submit
information via wireless camera for night vision. The fire sensor senses the fire inside
the warfare and sends the fire extinguisher information, so the water pump can be
triggered whenever possible. An ultrasonic sensor is used to detect all obstacles and
to detect the distance between the robot and the obstacle. Without causing damage,
ultrasonic sensors calculate the distance. When the robot identifies obstacles and
responds. The ultrasonic sensor is incorporated such that the entire world is laid out
[3]. The conventional robot firefighting IOT based firefighting robot can be replaced.
The fire alarm from the cloud is issued to an android application. Then, the auto-
matic receiver in the robot provides a live video of the fire position. The recipient
also allows the consumer and fire victims to connect. This helps the people trapped
in the position where fire has occurred to get guidance on evacuation routes. By
calculating the amount of carbon monoxide within the fire area, the possible health
effects can be calculated roughly [4]. A water tank and a wireless communication
operated pump are used for pumping water. The robot body moves by the micro-
controller’s performance. The robot is operated through a Bluetooth app. Remote
position IR flame sensor is used. Facilities provide wireless networking. Exactness
and performance are maximized after fire detection with a minimum amount of time
[5]. The robot is wirelessly operated and remotely controlled. The fire is sensed by
a sensor. The robot is loaded with a water tank and all motors are operated by the
Microcontroller. The fire and gas sensors are sent to the microcontroller when the
fire threshold is over 450C or the gas threshold is over 300 parts per million. The
message is received via Bluetooth technology on a mobile device [6]. On its own
axis the robot rotates to scan for fire at various angles in the building. If the sensor
LM35 detects heat, the robot stops and moves the fire by means of a centrifugal
pump that casts water, if the temperature increases. In darker areas, the fire flame
is best sensed. It is a robot that prevents fire because it can immediately detect and
extinguish [7]. Nodes where fire incidents happen sometimes are located in an indoor
environment. A sensor and microcontroller node is included. The information shall
be transmitted to the coordinator unit if the node detects fire. The co-coordinating
unit is the main unit linking all nodes. The coordinators unit performs two actions,
one is to send fire alarm message to the department and the other to send details to
a mobile robot consisting of an indoor map. An algorithm for the planning of a star
path designs a route from the current position to the fire location, which is called the
global path. Mobile robots planner [8] transforms the route information into separate
movement instructions. Using its onboard camera the robot will conduct monitoring.
The operator of the robot will direct it remotely around the entire site. The extinction
of Water and CO2 is conducted. Nebulous sprays are being used to avoid burns in the
Survey on Fire Safety Robot & Implementation … 267

robot. In harmful circumstances it becomes more helpful by bi-directional contact

with victims. A Tmega 2560 microcontroller [9] is used to monitor it primly. The
robot is generally automatic and can be changed to manual mode by remote control.
By using the Bluetooth module, smart phone is connected to the device. The robot
holds the fire away safely. The processing unit is Arduino microcontroller. Proteus
is used for robot research and simulation. Robot is operated from a remote site based
on an RF communication system, to ensure human protection. It is quick and easy
to implement the open-loop control technique. Improved output is the closed-loop
power. However, due to changes in the parameters [10], closed loop framework can
be unstable. For the intelligent burning robot, the search-and-find algorithm is used.
The robot began moving in the direction of the fire, when it was moved on. It recog-
nized a room without a fire and eventually discovered the room with a fire and turned
to the fire. It went back to where it began until the fire was extinguished. The robot
effectively fuses the fire through this algorithm [11]. To monitor the robot and to
acquire sensor data, the assembly language software is used. Using the wireless RF
controller, the monitoring computer and a remote supervised computer to operate
the mobile robot. An adaptive fusion algorithm is used to detect fires [12]. The robot
moving vertically is useful in high-level firefighting applications. In a reasonable
time, this robot can climb up a house. The handrail is used as a step on a balcony. It
can go up from 2.50 [m] to 3.10 [m] [13] to the building with its floor inteval.

2 Materials and Method

The purpose of the work proposed is to develop an Android-operated fire-fighting

robot to replace the conventional fire-fighting robot. The IR sensor data is read and
the robot moves forward or the robot spins and moves forward if no collision occurs.
The signal is sent to the microprocessor when the fire is detected by the fire sensor.
The data is interpreted by the microprocessor and then instructed to pump water in
the water pump. The pump works with the motor driver to extinguish the fuel. The
quantity of harmful gases is calculated during this process and the data is forwarded
to the customer. Multiple gas sensors can be used to detect gases such as carbon
dioxide, carbon monoxide and ammonia. A fire sensor is used in five channels, as it
can simultaneously sense fire in 120° and the fire can be detected more easily. The
potential health impacts of the victims can be measured approximately according to
the amount of gases and medical assistance is available in advance. A servo motor
is an electric device that can precisely push or rotate an object. In compliance with
the directions or inputs, the motor driver lets the motor shift. The control / processor
listens to the low voltage and regulates an actual engine that needs high input voltages.
A 3~6 V power supply will run the Submersible Pump engine. With a very low current
consumption of 220 mA, it can take up to 120 L per hour. It is used to put the fire out.
A flame detector module of 5 channels is used for detecting the wider area of flame
(>120°). Ammonia (NH3 ) has a strongly pungent odor, colorless gas or compressed
268 K. Sakthisudhan et al.

Fig. 1 Firefighting robot functional diagram

liquid. The detector of ammonia is based on an electrical theory. It disseminates

ambient air into the liquid electrolyte in the sensor via a membrane (Fig. 1).
A CO2 sensor or carbon dioxide sensor is a carbon dioxide measuring unit. Infrared
gas sensors (NDIR) and chemical gas sensors are the most general concepts for CO2
sensors. Carbon monoxide is an odorless, colorless and tasteless gas which cannot be
detected in the absence of an effective detector. MQ-7 gas sensor sensitive material is
SnO2 , which is less clean air conduction. The IR sensor is an electronic system that
transmits light to sensing an object of the surroundings, and it is used to detect CO at
low temperatures (heated to 1.5 V). An IR-sensor can measure and detect the heat of
an object. When there is not an object, there is no signal from the infrared receiver; the
infrared receiver receives signals from an object which blocks and reflects infrared
light. Node MCU is an open source firmware that provides open source prototyping
board designs. The term MCU Node (micro-controller unit) combines “node” and
“MCU”. The word “Node MCU” does not mean the related development kits but
strictly speaking, the firmware. It is cost-effective and can be integrated in support
of WIFI. The consumption of energy is poor.

3 Conclusion

This proposed work is aimed at using robots to detect and extinguish fires that are
minimally harmful to humans. It also tests the amount of poisonous gases including
carbon dioxide, carbohydrate and ammonia using gas sensors, as well as controlling
fire. The gas sensor data helps to predict approximately the potential health effects
of the injured, to ensure that medical assistance is prepared in advance.
Survey on Fire Safety Robot & Implementation … 269

4 Future Work

Water is used to extinct the fire in the proposed work, but the extinct material may
be modified by fire type. The water capacity in the water tank can be increased to
meet requirements. To find their safe levels, other gas sensors can be installed.

References

1. Khoon TN, Sebastian P, Saman ABS (2012) Autonomous fire fighting mobile platform. In:
International symposium on robotics and intelligent sensors 2012 (IRIS 2012). Elseiver
2. Liljeback P, Stavdahl O, Beitnes A (2006) Snake fighter—development of a water hydraulic
fire fighting snake robot. In: 9th International conference on control, automation, robotics and
vision. ICARCV ’06, pp 1–6
3. Kolambe K, Pote R, Jadhav A, Chennur V (2018) Spy robot with fire detection and water
sprinkling. In: 2nd International conference on electronics, communication and aerospace
technology (ICECA 2018) IEEE conference record # 42487; IEEE Xplore. ISBN:978-1-5386-
0965-1
4. Kanwar M, Agilandeeswari L, IOT based fire fighting robot. IEEE
5. Zaman HU, Khan TA, Falgunee SR, Rashid GMS, Talukder FH (2018) Autonomous firefighting
robot with optional bluetooth control. In: 2018 4th International conference on computing
communication and automation (ICCCA). IEEE
6. Bose JSC, Mehrez M, Badawy AS, Ghribi W, Bangali H, Basha A, Development and designing
of fire fighter robotics using cyber security. IEEE
7. Rakib T, Rashid Sarkar MA, Design and fabrication of an autonomous firefighting robot with
multi sensor fire detection using PID controller. In: 5th International conference on informatics,
electronics and vision. (ICIEV) –IEEE
8. Anantha Raj P, Srivani M (2018) Internet of robotic things based autonomous fire fighting
mobile robot. In; 2018 IEEE international conference on computational intelligence and
computing research (ICCIC)
9. Mittal S, Rana MK, Bhardwaj M, Mataray M, The fire fighting robot. In: IEEE international
conference on advances in computing, communication control and networking
10. Fahim SR, Das SK (2018) High-Performance open-loop tracking control of a small scale fire-
fighting robot. In: IEEE international conference on advancement in electrical and electronic
engineering 22–24 Nov 2018
11. Kim JH, Keller B, Lattimer BY (2013) Sensor fusion based seek-and-find fire algorithm
for intelligent firefighting robot. In: 2013 IEEE/ASME international conference on advanced
intelligent mechatronics (AIM) Wollongong, Australia, 9–12 July 2013
12. Su KL (2006) Automatic fire detection system using adaptive fusion algorithm for firefighting
robot. In: 2006 IEEE international conference on systems, Man, and cybernetics Oct 8–11,
2006.Taipei, Taiwan
13. Amano H, Osuka K, Tarn TJ (2001) Development of vertically moving robot with gripping
handrails for firefighting. In: Proceedings of the 2001 IEEE/RSI international conference on
intelligent robots and systems Mad, Hawaii, USA, Oct. 29 -Nov 03, 2001
Detection of Emergency Vehicles Using
Radio Frequency Identification (RFID)

Usha Mittal and Priyanka Chawla

Abstract Road traffic is a major problem in major cities. With the increasing number
of vehicles, the resolution of the problem is very crucial. Problems related to traffic
not limited to road congestion, but also difficult for the passage of emergency vehi-
cles. In order to build an intelligent traffic signal system, a complete traffic manage-
ment system must be built that deals with all traffic-related problems. The aim of this
article is to propose an efficient system for detecting and prioritising emergency vehi-
cles on the junction. In this paper, emergency vehicles are identified and prioritized
using Radiofrequency Identification (RFID). RFID tags serve to identify vehicles
in a unique way. In this article, ultra-high-frequency RFID readers are used with a
frequency of 865–928 MHz and cover a maximum distance of 9 m.

Keywords Emergency vehicle detection · RFID · Traffic lights · Priority · RF

receiver · RF transmitter

1 Introduction

Urbanization, industrialization and population growth have resulted in a considerable

increase in vehicular traffic. There is no strong traffic framework on the planet today,
an approach to transmit effective traffic frameworks is through the unique control of
traffic lights depending on the size of the traffic Furthermore, there are no priority
services for priority vehicles such as ambulances, police vans or fire trucks. As
a result, certain services other than regular services must be provided to priority
vehicles [1].
A traffic signal that runs on timers is the most common type of traffic control. You
are supposed to work somehow at a particular time of day, regardless of the traffic.
We are not getting rid of unnecessary congestion.
When you are stuck in a traffic jam, you can address the problem of emergency
vehicles with the means to clear the lane in which emergency vehicles travel. This

U. Mittal · P. Chawla (B)

Lovely Professional University Phagwara, Phagwara, India

© The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 271
R. Ibrahim et al. (eds.), International Conference on Artificial Intelligence
for Smart Community, Lecture Notes in Electrical Engineering 758,
https://doi.org/10.1007/978-981-16-2183-3_26
272 U. Mittal and P. Chawla

may be done as an emergency vehicle approaches with the nearest traffic light. During
an emergency, the RFID tag acts as a transducer and transmits signals to the lighting
system [2].
RFID is a technology that allows the automatic identification of an individual,
packet or element with radio signals [3]. It relies on RFID tags to do so. These
are small transponders that provide identity information when requested over a short
length. The majority of RFID labels have at least two parts. One is a built-in circuit that
stores certain information, modules and demodulates the signal for radio frequency
and other special functions. The second antenna serves the purpose of receiving and
transmitting the signal. There are primarily two types of RFID tags: active RFID tags
that contain batteries and passive RFID tags that do not contain batteries [4].

1.1 Different Ranges of RFID

To transmit, RFID labels and readers must be adjusted to a similar frequency. There
are various frequencies which may be used by an RFID gadget. Generally speaking,
the more widely recognized are.
• Low recurrence, or LF, (125–134 kHz)
• High recurrence, or HF, (13.56 MHz)
• Ultra-high recurrence, or UHF, (433, and 860–960 MHz).
On the various frequencies, the radio waves act in a contrasting manner, so it is
essential to choose the adjustment recurrence for the application.
Low-frequency labels, for example, have a frequency and may penetrate delicate
metallic surfaces even more appropriately. LF RFID frameworks are likewise ideal
for perusing high water content articles, for example, natural product or beverages,
however the read go is just centimeters or centimeters. Common place LF RFID
applications incorporate creature labeling and get to control.
High-frequency labels work very well for metal items and can be used to handle
goods containing medium to high water. HF RFID frames usually operate in inches,
but can have a greater reading range of about three feet (1 m). Following library
books, following the patient stream, and travel tickets are the average HF RFID
applications.
Generally, UHF frequencies provide essentially a better understanding of the
extent and can move information quickly (e.g., read many more labels every second)
than lower or high recurrence frequencies (inches to 50 + ft. as per the RFID frame-
work arrange- ment). Since, in all cases, the frequency of UHF radio waves is shorter,
their signal would be lowered (or weakened) and would not be communicated via
metal or water. Other basic uses of UHF RFID integrate the electronic assortment
of costs and the control of discontinuation of access because of the more stretched
scope of understanding.
Detection of Emergency Vehicles Using Radio Frequency… 273

2 Literature Survey

Because of wasteful traffic frameworks, fuel time and cash are incredibly lost, it
some of the time prompts lost life. So a great deal of work is being done to deal
with those issues. In 2009, the IEEE issued a smart traffic light framework that uses
RFID to filter the vehicle through the vehicle’s RFID label and collect its EPC data
[1]. The traffic volume, vehicle need and street and time are determined by sort of
vehicle. A guideline calculation that suggests how the traffic signals are worked is
given in the Decision Making Section (DMS).
In 2009 [5], the developer proposed a technique for the use of the system’s sensors
as well as the innovation implemented. Accordingly the exchanging of traffic signal
builds traffic and street limit and may prevent traffic jams [6].
In 2014 [7], a web-based traffic management system that could be managed
remotely and monitored simultaneously at different crossings was built using new
web server technologies. The primary node is the Central Traffic Management Unit
(CTMU), which uses Internet technology to remotely track and control individual
nodes.
The author used the RFID label in 2015 [8] and consists of three parts: automatic
monitoring of signals, clearance of vehicles and detection of stolen vehicles.
In [9], the utilization of RFID exercises controls is talked about, specifically those
recognized in the image arrangements and in impedance techniques, to keep up a
vital good ways from issues which generally emerge from standard movement control
sys- tems. This multi-vehicle, multilateral and multi-track convergence framework
is over- seen by RFID. It provides a beneficial time organization plan, in which, a
component schedule proves to be consistently for the segment of each movement
segment. The current use of the framework emulates the fundamental choice of a
functioning cop. The amount and monitoring of vehicles in each segment is the
property where estimates and decisions are completed. The disadvantage of this
examination is that the techniques used for the emergency unit and the control frame
of the moving panels are not attempted.
In 2013, the author [4] proposed a framework for RFID and GPS-enabled crisis
vehicles. The focal errand of this work is to lessen the time the crisis vehicle arriving at
the recuperating office by normally clearing up the course through which the salvage
vehicle continues before the moving signs are reached. This should be possible when
a crisis vehicle is at a specific division of its action convergence, turning the action
signal in transit toward the green. The use of RFID perceives the hand brake and non-
emergency cases, in which unnecessary movement blockage is seen. Correspondence
is delivered through the phone and GPS between the crisis vehicle and moving signs.
When development converges, the instrument is completely mechanical and does
not require human mediation. The restriction of the structure is that it needs all
the information about the start, the end of the tour. This that not work if the salvage
vehicle needs to follow an alternate course or if the beginning stage is unquestionably
not comprehended ahead of time.
274 U. Mittal and P. Chawla

Ongoing advances have been utilized to recognize a crisis vehicle nearness and to
figure the real traffic dense[10], for example, infrared (IR) and worldwide situating
framework (GPS). [11]. So as to recognize the nearness and check vehicles, RFID
labels were utilized to distinguish the nearness of crisis vehicles [12]. The most
widely recognized techniques for traffic recognition incorporate video, microwave,
radar identification, ultrasonic location and so on [9, 13].
For crisis vehicle detection [2, 14] and vehicle tracking [5], RFID innovation
has been recommended. So as to control traffic signals [15], RFID and GSM were
additionally proposed. Such projects, be that as it may, depend on specifically taking
care of various issues.

3 Proposed Methodology

Every vehicle has passive RFID labels with explicit Electronic Product Code (EPC)
RFID label number. No outer force source should be connected. The intricacies,
such as the vehicle number, the vehicle type and the owner’s data are also stored on
each RFID tag. The Ultra-High Frequency (UHF) recurrence band recognizes labels
in the RFID radio wire. The peruser module peruses this information and it is then
sent for additional preparing to the worker. Table 1 shows the details of emergency
vehicles taken for experiment. The proposed system includes 2 main units that are
linked together, namely: emergency vehicle detection and vehicle priority based on
the type of vehicle.

Table 1 Details of
EPC vehicle Vehicle type Owner’s Details priority
emergency vehicles
number
1011 XXX Ambulance XXX 1
1012 XXX Firebrigade XXX 2
1013 XXX Ambulance XXX 1
1015 XXX Ambulance XXX 1
1016 XXX Police XXX 3
1017 XXX Firebrigade XXX 2
1018 XXX Ambulance XXX 1
1019 XXX Police XXX 3
1020 XXX Police XXX 3
Detection of Emergency Vehicles Using Radio Frequency… 275

3.1 Detection of Emergengy Vehicle

This system is made up of an RFID tag in emergency vehicles installed. During an

emergency, it is used as a tracker. This means that the driver activates the RFID tag,
which is then detected by RFID readers located a few meters away at the junction,
whenever there is an emergency. These readers then continuously transmit the signals
to the intersection where the traffic lights are controlled. The reader starts detecting
signals and the conditions return to normal as soon as the emergency vehicle is tagged
into the junction.
Access control is used to detect IDs entering or leaving the RFID reader area. The
signals are passed to the junction unit following identification by RFID. The entire
detection unit is shown in the Fig. 1.

Detection of Emergency
Vehicle

Priority Assignment to Emergency

Vehicle

Multiple Yes
Emergency
Vehicles Prioritize
present at lanes
different
lanes?

No

Assign green light to priority

lane.

Fig. 1 Flow chart of proposed methodology

276 U. Mittal and P. Chawla

1. Signals sent by the RFID reader are collected by the transceiver. After recog-
nizing the received signals, system identifies the type of emergency vehicle and
assigns priority to it. For example, if more than one different or same emergency
vehicle are present at different lanes, then according to their type, a numeric
value say 1 for ambulance, 2 fire truck and 3 for police van.
2. System then prioritize the lanes according.

4 Results and Discussions

This paper has a great impact on the traffic problems faced in urban areas. Normally
the traffic is controlled by a predetermined signal light controlled system. The
working of the modules is based on the RFID tag information. In this paper, ultra high
frequency RFID tags are used. The only requirement of this paper to be implemented
is that the vehicles should be provided with a unique RFID tag. Proposed method
is implemented and simulated in Python and the results are shows in the Table 2.
Traffic light switching starts from north direction. Emergency vehicle present column
represents whether any emergency vehicle is present or not. Third column represents

Table 2 Results of proposed method

Traffic light Emergency Vehicle Number of Direction and Signal to be
switching Present? emergency details of vehicles provided
direction vehicles present from where
vehicles coming
North Yes 2 From west, Signal to be
Firebrigade (EPC: provided on
1017) From North, West
Police van (EPC:
1020)
West Yes 1 From North, Police Signal to be
van (EPC: 1020) provided on
North
North No NA NA Signal to be
provided on
South
South No NA NA Signal to be
provided on
East
East Yes 2 From North, Police Signal to be
van (EPC: 1016) provided on
From South, East
Ambulance (EPC:
1018)
Detection of Emergency Vehicles Using Radio Frequency… 277

number of vehicles present. Fourth column shows type and direction of the emer-
gency vehicle with RFID information. Last columns shows the decision of traffic
light controller i.e. in which direction next signal to be provided.
Any RFID label that exists on a path for over 20 min will distinguish an anomaly.
This might be an accident of the vehicle or a mishap by street. The vehi- cle with
this EPC number will be tried for issues utilizing the phone number of the proprietor
and, where pertinent, will give prompt help.

5 Conclusion

This paper presents a methodology for the arranging of traffic crisis vehicles. The
pro- posed crisis location and organizing framework depends on radio recurrence
ID. With the crisis vehicle location module it will guarantee that the crisis vehicles
will arrive at the goal as quickly as time permits. Contrasted with recently executed
plans, the program is both financially savvy and simple to actualize. The cost level
of the gad- get is exceptionally valuable. Each RFID tag is gathered progressing
by the system. That vehicle may utilize the RFID labels to spare significant data,
for example, vehicle number, proprietor’s information, vehicle model and number
of encroachments by that vehicle. The accompanying can likewise be utilized to
screen crooks, checking vehi- cles taken and e-ringing vehicles, assessing the most
recent details for different sorts of street vehicles, identifying bottlenecks, analyzing
territories for customary blockages and deciding the causes.

References

1. Bansal CS, Chandra A (2009) An intelligent traffic control system using RFID. IEEE Potentials
28(3):40–43. https://doi.org/10.1109/mpot.2009.932094
2. Sharma S et al (2013) Traffic light priority control for emergency vehicle using RFID. Int J
Innov Eng Technol 2(2):363–366
3. Hua LC (2010) Automatic vehicle identification system based on RFID. Anti-Counterfeiting
security and identification in communication (ASID) (2010), pp 281–284
4. Chellani N, Tahilyani C (2013) Traffic congestion detection and control using RFID technology.
Int J Eng Res Technol
5. Chavan SS (walke) et al, Design of intelligent traffic light controller using embedded system.
In: Proceedings of the second international conference on emerging trends in engineering and
technology, ICETET-09
6. Bharadwaj R et al (2013) Efficient dynamic traffic control system using wireless sensor
networks”. In: Proceedings of the IEEE international conference on recent trends in information
technology (ICRTIT). pp. 668–673
7. Pallavi P, Hussain MA (2014) Intelligent traffic control system using embedded web technology.
Int J Comput Sci Inf Technol Secur 4(2)
8. Sundar R, Hebbar S, Golla V (2015) Implementing intelligent traffic control system for conges-
tion control, ambulance clearance, and stolen vehicle detection. IEEE Sens J 15(2):1109–1113.
https://doi.org/10.1109/jsen.2014.2360288
278 U. Mittal and P. Chawla

9. Balamurugan A et al (2015) Automated emergency system in ambulance to control traffic

signals using IoT. Int J Eng Comput Sci 11533–11539
10. Farheena S, Chandak BM (2014) An approach towards traffic management system using density
calculation and emergency vehicle alert. IOSR J Comput Sci 4:24–27
11. Fan XH, Zhang YL (2009) A Design of bi-verification vehicle access intelligent control
system based on RFID. In: Proceedings of the ninth international conference on electronic
measurement & instruments (ICEMI’2009)
12. Zhou ZG et al (2008) Secure design of RFID tags in the new type license plates automatic
identification system. In: Proceedings of international symposium on computer sciences and
computational technology, TBD, pp 694–697
13. Andronicus F, Maheswaran PG, Scholar (2015) Intelligent ambulance detection system. Int J
Sci Eng Res 1462–1466
14. Mithun NC, Rashid NU, Mahbubur Rahman SM (2012) Detection and classification of vehicles
from video using multiple time-spatial images. IEEE Trans Intell Trans Syst 13(3):1215–1225.
https://doi.org/10.1109/tits.2012.2186128
15. Rifai I, Alj M, Yassine (2014) An energy-efficient WSN-based traffic safety system. 1–6.
https://doi.org/10.1109/IACS.2014.6841980
Big Data-Enabled Solutions for Covid-19

Dirisala Saikrishna, Bavisetti Sai Dhanush, Madhusudhan Rao,

Priyanka Chawla, and Usha Mittal

Abstract In the time of the pandemic like CORONA, Covid-19, everyone is ftghting
against this deadly virus. Besides, governments are looking for a barrier that stops
spread of virus until the vaccine is made. In modern era, technology plays an impor-
tant role. This paper brings the way by using a powerful technology called Big data.
Big data know for handling a large amount of data and provide powerful insights into
the data. Big data integrated with Artificial Intelligence is a powerful tool to ftght
against this pandemic. Many countries like Taiwan, China with the use of Big Data
stop this pandemic up to some extent. But the collection of data itself comes up with
the big challenge of PRIVACY AND SECURITY. In the recent times, the world has
seen the effect of data leaking whether by Facebook or by Google. Many European
countries due to this big challenge will not be able to use this technology.

Keywords Big data · Artificial intelligence · Privacy · Security · Data collection

1 Introduction

1.1 Bigdata

The term describes it all Big data refers to data which is in large volume and different
formats or variety. Generally, Big data is described using the following characteristics
also known as 3 V’s of Big Data. i.e., Volume, Variety, Velocity [1].
Volume—In Today’s world every day Over 2.5 quintillion bytes of data produced
in a single day i.e., approximate.1.7 MB of data in a single second. Big data helps to
manage this huge amount of data with the help of its various frameworks and tools.
Velocity—It refers to the rate at which data is generated and by that time only acts.
on that data. It is challenging but big data provides the various solution for this alsoby

D. Saikrishna · B. S. Dhanush · M. Rao · P. Chawla (B) · U. Mittal

Lovely Professional University, Phagwara, Punjab, India

© The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 279
R. Ibrahim et al. (eds.), International Conference on Artificial Intelligence
for Smart Community, Lecture Notes in Electrical Engineering 758,
https://doi.org/10.1007/978-981-16-2183-3_27
280 D. Saikrishna et al.

using Various frameworks like Apache Spark, Flink, Apache Storm we can perform
the data analytics as soon as data enters the system.
Variety—It refers to different types of data that are produced. It is been
Categorized into 3 formats i.e., structured, unstructured, semi-structured.
Structured—It consists of a traditional relational table. Data is represented in a
tabular manner.
Unstructured—The data which is not in a particular format is referred to as.
unstructured data. Audio, video is the type of unstructured data.
Semi-structured—The data which has format, but it not structured is called.
structured data. JSON, CSV formats are the common example of this type of data.

2 How Big Data Help in Fighting Against Covid

Data has a crucial role in knowing the hidden facts. By collecting Data of individuals
big data helps to provide the insights. Data which helps fight against this pandemic
is:

2.1 Medical Data

Data provided by hospitals will play an eminent role. Through recent research, it is
found that a person with heart disease, lung disease or a diabetic person will have
more affected by coronavirus by the prior knowledge of this type of person [2].
Govt. can target this type of person and provide them more instructions during this
coronavirus. Also, By the research, it is found that coronavirus has more effect on a
certain age group of people, i.e., children and old age people.

2.2 Mobile Data Location

Usage of location data is quite very useful [3]. Because as we have a medical history
of a patient, we can use the location of a person easily to provide sufficient aid to that
person. Also, by this Govt. also able to see people are following social distancing or
not.
Big Data-Enabled Solutions for Covid-19 281

2.3 Travel History

By getting this data, it is easy to identify which person has travel history (most
favorably out of the country) [4]. So that quarantine period is assigned to that person.
And a person is tested. By this government’s priority to handle these people first.

3 Privacy Preserving in the Healthcare System

As mentioned in abstract, the biggest challenge in data collection is privacy. Though

this type of data is sensitive data as it completely discloses the identity of an indi-
vidual. By medical data we know the disease that person is suffering. By location
data, there is no privacy of person. There will be a close eye on the govt. or hacker
on a person. Countries like Korea, Taiwan, and China collect that data but still the
privacy the fundamental rights of a person. How person will be protected is still a
blank. Democratic countries need a better solution to concur this challenge. One
solution for that is an individual can voluntarily share its data and authorities will
assure him that its identity will be secured [5]. And after using your data is destroyed.
But the problem here is how can one know its identity is not being compromised and
its data is being destroyed or not. In recent years, many companies said that they will
destroy the data. But they did not. Also, if there is an agreement and a person has
full trust in authorities. There is another big challenge for participation.

3.1 Big Data Approach to Handling These Challenges

Before stepping forward let us know what the difference between Privacy and security
is.
Privacy refers to control over how personal information is being collected or used.
It is the power of an individual to stop information about themselves disclose to the
public or using it for personal use.
Security refers to defend the information of an individual by using technology,
process and by using training data generation is categorized into two ways i.e., Active
Data Generation, Passive Data Generation. In Active data Generation data owner will
give its data to the third party while in passive data generation data is generated by
data owner while online activity and data owner do not know its data is being collected
by the third party. Both risks are minimized by access restriction or by falsifying the
data.
Access Restriction—In Access restriction, if data is collected by active data
generation user cannot give its data if it discloses its sensitive information. If it
is passive data generation data is restricted by using an anti- tracking extension or
adblocker and encryption tools.
282 D. Saikrishna et al.

Falsifying Data—Sometimes it is not possible to neutralize access to the data. For

that time data owners can distort the data before data gotten by a third party. Once
the data is distorted it is not easy to reveal the true information. Socket Puppet is one
of the falsifying tools that do its work by deception. Multiple Socket puppet the data
about an individual is regarded as having a place with multiple people by the data
collector.

4 Big Data Privacy When Data is Being Stored

Storing information is not hard in big data due to advancements in technology. But it is
high-risk work. Because once a big data database is being compro- mised the privacy
of an individual is disclosed. Moreover, Big data follows distributed approach data
is placed in several datasets are several locations so privacy is a priority for this. The
conventional security mechanism is divided into four categories Application- level
security, Database level Security, file- level security, and media level security [6].
3 V’s of big data depicts its storage mechanism to be scalable.

4.1 Privacy-Preserving Techniques in Big Data

Privacy is being a major issue not only now but also, in the past. So, some techniques
were used earlier.
De-Identification—It is a traditional technique for privacy- preserving, for
persevering individual privacy. first, we need to generalize the data then suppres-
sion we must give the data for data mining, following up de- Identification comes
with the concepts of K-anonymity, l-diversity, t- closeness to increase the privacy
of information [7]. Migrating de-identification in bigdata is a great deal for privacy
preserving. Thus, big data also offer many methods of re-identification so using
de-identification is not a great idea for privacy protection. For a description of
K-anonymity, l- diversity, t-closeness let us first discuss some important terms.
Privacy-Preserving Aggregation—It is a homographic encryp- tion technique
used for data collection. This works as follows given a homo- graphic public key
source can convert their data into cipher texts. Different sources can use the same
public key then the whole result is aggregated and recovered by using a specific
private key. But this technique is purpose spe- cific. We can only use this technique
for storing and collecting data. For data exploitation and mining, this technique has
no use. So, it does not help in big data analytics.
Big Data-Enabled Solutions for Covid-19 283

5 Recent Techniques in Privacy-Preserving

Differential Privacy Differential privacy is one of the recent technologies which is

used widely. This technique helps the analyst to provide useful information without
revealing the identity of the user. It generally works by adding some distraction in
data. This distraction is large enough to de- identified someone’s privacy and small
enough to perform analytics to data at the same time. So, both ways it is a useful
technique. As we have seen in deidentification by hiding some data it pretends like
data is de- identified and privacy is obtained but it is not following example support
above statement. In the mid-’90 s when the Commonwealth of Massachusetts Group
Insurance Commission (GIC) released the anonymous dataset by hiding the name,
age of the people. One Ph.D. Student by taking publicly available voters dataset
matches with the GIC dataset easily re- identified people and it destroys the privacy.
While DP does not work that way there is an intermediate software present in between
database and analyst. Only the information needed by the analyst is provided by the
database through the software. This intermediate software is also called a privacy
guard. It works as follows:
Step 1: Analyst makes the query to the database through the intermediate privacy
guard.
Step 2: Privacy guard evaluated the query and the previously evaluated queries
and evaluates privacy risk.
Step 3: Then the privacy guard will get the desired information from the database
by adding some distraction in it.
The level of distraction is proportional to the evaluated privacy risk. If the privacy
risk is high the distortion is high and vice versa.

2 1
Query

Database
3 4
Query Data

Privacy Guard Analyst

As Discussed earlier many algorithmic approaches of adding noise to the data but
the fact is chances of reidentification is still in a method. Meanwhile, this algorithm
also adds noise to the data, or we can say data to the noise, so it is quite hard. For the
hackers to find the original data as the same way of finding a needle in the haystack
284 D. Saikrishna et al.

concept or we can say finding salt in sugar. This will surely create a more cost of
computation for making a haystack (noisy data) to hide the needle. In Fig bellow, the
white circles are noisy data while the black circles are the original data. The service
provider adds a dummy as noisy data as original transactional data collected by the
data provider. Likewise, a unique code is assigned both dummies as well as original
data. This code is maintained by the service provider so that during the extraction of
data dummy data and original data got to separate [8].

6 Privacy-Preserving in the Healthcare System

The way in which medical records are being digitized gives a new face to the health-
care system. Thus, doingthis medical industry experiencing a heavy load of data. By
heavy lead of data then big data comes into the play. Due to the real-time availability
of records and more accurate test result helps doctors to take more constructive deci-
sions in treatment. The real-time monitoring of embedded sensors attached to the
patient gives the warning whenever there is an anomaly. Healthcare digitization with
integrated analytics is one of the next big waves in healthcare information technology
with electronic health records (EHR) being a crucial building block for this vision. By
this healthcare organization came to result that with the EHR incentive programs the
results are more accurate and thus eventually lead to improved healthcare of patients.
But with great power comes great responsibility. Like that the ever- changing risks
and vulnerabilities, security exploitations are expected to grow in the coming years.
Big Data-Enabled Solutions for Covid-19 285

Big data offers a comprehensive survey of different tools and techniques used in
pervasive healthcare in a disease-specific manner [9]. It provides a quick result by
finding major diseases and disorders to give a quick treatment. The adoption of big
data in healthcare significantly increases the security risks and patient privacy risks
[10]. As we have seen traditional techniques were not directly applied to the large
and inherently diverse datasets. In today’s world, we have seen there is a demand for
cloud-stored.

6.1 Real-Time Security Analytics

Today’s requirement in real-time analytics is predicting threat sources in real- time is

of utmost need in the healthcare industry. The most common attack in the healthcare
industry is distributed denial of service (DOS) attack to steal important information
and in real-time analytics, the delay is not accepted [11].

6.2 Data Quality

Health data is usually located from different sources, and with different databases,
this leads to anomalies in the data lot of missing values are there. These things should
be handled because dirty data leads to incomplete analysis [12].

6.3 Data Sharing and Privacy

Health data contains personal information. Their privacy of an individual is always

at risk. Healthcare information would be masked by de-identification techniques and
later that data is provided to the data analytics based on legal data sharing agreement.
But due to this, the quality of data I decreased that is not usable for analytics. So,
to take care of privacy, there is always a balance between privacy and providing
meaningful data [13].

7 Conclusion

In this paper, we discussed how big data is helpful in Fighting against COVID- 19 all
we need to take care of challenges that come with big data. We have seen different
techniques to secure the privacy of an individual. Big data leads to knowledge that
provide better analytics. Yet a small amount of data is being used for analytics
due to privacy issues. Privacy techniques with its advantages and disadvantages are
286 D. Saikrishna et al.

described in this paper. Concepts of differential privacy and other comprehensive

techniques are being used. MapReduce framework is used for anonymizing tech-
niques and it easily is scaled up using adding more mappers and reducers. As for
future more effective solutions need to be presented. In the field of healthcare as well
more scalable technique that can handle the true power of big data is needed. IoT is
a future integrating big data with IoT is gamechanger in every field. Forms of data
itself whether it is semi-structured or unstructured come with many challenges. As
such, there is a lot more to do in privacy-preserving in big data. For a time being to
fight against COVID-19 discussed techniques will improve big data analytics.

References

1. Oracle big data for the enterprise (2012) [online]. http://www.oracle.com/caen/technoloqies/

biq-doto
2. Big data versus COVID-19: opportunities and privacy challenges, [online] https://www.bru
egel.org/2020/03/bigdata-versuscovid-19-opportunitiesandprivacychallenges/?utm_content=
bufferfdf0a&utmmedium=social&utm_source=facebook.com&utm_campaign=buffer
3. Mohammadian E, Noferesti M, Jalili R (2014) FAST: fast anonymization of big data streams.
In: ACM proceedings of the 2014 international conference on big data science and computing,
article 1
4. Big data at the speed of business, [online]. http://www-01.ibm.com/software/data/bigdata/2012
5. Mehmood A, Natgunanathan I, Xiang Y, Hua G, Guo S (2016) Protection of big data privacy.
In: IEEE translations and content mining are permittedfor academic research
6. Lu R, Zhu H, Liu X, Liu JK, Shao J (2014) Toward efficient and privacy-preserving computing
in big data era. IEEE Netw 28:46–50
7. Li N et al (2007) t-Closeness: privacy beyond kanonymity and L-diversity. In: Data engineering
(ICDE) IEEE 23rd international conference
8. Abadi DJ, Carney D, Cetintemel U, Cherniack M,Convey C, Lee S, Stone- braker M, Tatbul N,
Zdonik SB (2003) Aurora: a new model and architecturefor data stream manag ement. VLDB
J 12(2):120–39
9. Jain P, Pathak N, Tapashetti P, Umesh AS (2013) Privacy preserving processing of data deci-
sion tree based on sample selection and singular value decomposition. In: 39th international
conference on information assurance and security (lAS)
10. Wang C, Wang Q, Ren K, Lou W (2010) Privacypreserving public auditing for data storage
security in cloud computing. In: Proceedings of IEEE international conference on infocom, pp
1–9
11. Sedayao J, Bhardwaj R (2014) Making big data, privacy, and anonymization work together in
the enterprise: experiences and issues. Big data congress
12. Fong S, Wong R, Vasilakos AV (2016) Accelerated PSO swarm search feature selection for
data stream mining big data. IEEE Trans Serv Comput 9(1)
13. Big data privacy: a technological perspective and review Priyank Jain*, Manasi Gyanchandani
and Nilay Khare
Smart Traffic Monitoring System

V. Radhika, V. Madhan Babu, and S. Jayaprakash

Abstract In the modern era, traffic congestion is one of the major issues in smart
cities. The existing traffic management is unable to handle the growing number
vehicles. The main aim of this work is to automate the traffic signal by monitoring
the entire system and allowing vehicles to pass through lanes based on the traffic
density. The existing method consists of a signal with fixed timing and also human
intervention for managing traffic. In this proposed work, a system is designed to
program the traffic signal timing and reassign the green signal glowing time based
on traffic density. The lane with higher traffic density will have maximum timing to
pass through and the lane with less traffic density will have minimum timing to pass
through. A system that gives highest priority to emergency vehicles is also developed.

Keywords Smart city · Traffic management · Traffic density

1 Introduction

Automation in traffic signals can be an effective transportation management system

utilized in smart cities. It uses the advancement in Internet of Things (IoT), sensor
technology and data analytic techniques to manage the traffic effectively and enhance
the traffic management system. The four major components in the traffic management
system are data collection, data transmission system, data analysis system and finally
the collected information is sent to the end user. The aim of the intelligent traffic
management is to create interactive transport system, manage the traffic and provide
the real time information about the traffic to the common users [1]. IoT can change
the transport system effectively and integrates the embedded sensors, actuators and
other gadgets to collect and provide the information about the real time traffic activity.
Traffic signals are generally used to ensure the safe and systematic movement of
vehicles. These signals ensures the pedestrians or the vehicles to cross the intersection

V. Radhika (B) · V. Madhan Babu · S. Jayaprakash

Department of Electronics and Instrumentation Engineering, Sri Ramakrishna Engineering
College, Coimbatore, Tamil Nadu, India
e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 287
R. Ibrahim et al. (eds.), International Conference on Artificial Intelligence
for Smart Community, Lecture Notes in Electrical Engineering 758,
https://doi.org/10.1007/978-981-16-2183-3_28
288 V. Radhika et al.

roads without any conflicts. But if a particular traffic signal fails it leads to increase
in accidents, waiting time, increase in greenhouse emissions also. If a traffic signal
in all the lanes have uniform timing it leads to more crowdie in a particular lane than
the other lanes. If the timing is allotted to lanes based on the density of vehicles and
also on any other specific criteria traffic jam can be avoided. So traffic signal timer
programmed based on the traffic density can reduce the traffic congestion problems.
Traffic light system consists of traffic lights and traffic light control system to
control the traffic signals. In the existing system the controllers used are based on
electro mechanical mechanism. These controllers are based on open loop mecha-
nism without any feedback mechanism and sensing network. So they lack in intel-
ligence to adapt to the change in traffic density and control the traffic signal effec-
tively. Traffic signal timer that works according to the traffic density by utilizing the
image processing technique is also proposed in earlier works. In those methods high
resolution cameras are used to sense the changing traffic density around the traffic
signal and programs the traffic signal timing accordingly. In these methods images
are captured on regular intervals, then the images are processed and given to the
controller. Controllers on receiving the input, compares it with predefined thresh-
olds and sends the required signal to traffic signal timer. But these methods need a
high resolution camera and images also to be captured continuously for predefined
time. So an alternative method to automate traffic signal that reduces the traffic jam
effectively and minimize the air pollution is needed for current busy scenario.
This paper proposes an intelligent traffic management utilizing IoT to handle
the different traffic situations. This proposed work reduces the flaws in the existing
system traffic system more effectively. This system utilizes the ultrasonic sensors to
detect the number of vehicles and the RFID tag placed in the emergency vehicles
like ambulance, VIP vehicles and fire brigades to detect and prioritize the emergency
vehicles. So this proposed system detect traffic congestion, provide predictable infor-
mation about the traffic condition can reduce the problems like traffic congestion,
accidents and parking issues.
In this paper Sect. 2 explains the related works, Sect. 3 elaborates the proposed
work, the results and discussion is explained in Sects. 4 and 5 gives the conclusion
and future scope.

2 Related works

This section presents an overview on the techniques and approaches used in the
related works considered for monitoring the traffic system.
For working with the traffic congestion issues, predictive type or real time knowl-
edge about the traffic is necessary. Traditional methods uses sensors like pneumatic
tube and automatic counter to obtain real time data on traffic signal data’s. These tech-
niques fail to detect the vehicle movement in multiple lanes. Video/image processing
techniques utilizing the number plate recognizing system to find the traffic density is
proposed. This technique fails due to non-uniformity in the number plates and visi-
bility problems in poor weather condition. Wireless Sensor Networks (WSN) and
Smart Traffic Monitoring System 289

Vehicle to Vehicle network is utilized to measure the traffic density is also proposed.
Also various related works using fuzzy logic and neural networks learning algorithms
are also utilized for controlling the traffic light is found in literature.
Combinational Interaction Testing (CIT) and Combinational interaction testing
with constraint approaches supports the prioritized based test generation for traffic
signal generation. Optimization algorithms were proposed for scheduling traffic light
conditions. Simulated annealing applied to CIT technique is proposed for computing
the test case generation for traffic signal conditions. Traffic analysis toolbox volume
IV: guidelines for applying corsim micro simulation modeling software and Direct
signal timing optimization: strategy development and results works uses Genetic
Algorithm (GA) for optimizing the programming cycle for traffic lights [2–4]. In
Intelligent Testing of Traffic Light Program approach a PFGM algorithm is proposed
to generate the prioritize test suite for a given a traffic feature model. In this work
an expert system with reduced set of scenarios with similar fault detection capacity
is developed with minimum cost. Developed expert system numerically quantify
and qualify the traffic cycle program for different traffic conditions. In the Isolation
niches particle swarm optimization applied to traffic lights controlling work, particle
swarm optimization algorithm is used for scheduling traffic lights. In this approach
one way road with two intersections was proposed. A multimodal traffic responsive
strategy using Particle Swarm Optimization (PSO) algorithm is proposed in this
approach multi-objective PSO was used. This work uses predictive control model
and simulations were carried out with private, and public vehicle model were carried
out on a virtual urban road network.

3 Proposed Work

Existing traffic signal management system consists of traffic signal that are
programmed such that the yellow, green and red signals glow for fixed interval
of time. The system proposed here identify the traffic density on different lanes and
adjust the traffic signal timing. But traffic jams happening in today scenario signs
that traffic signal must be programmed based on the density of road. So the proposed
system aims to reprogram the traffic signal based on the traffic density. This traffic
monitoring system has two modules such as hardware and software module. The
hardware module has ultrasonic sensor, RFID sensor, RF Module, LEDs, Arduino
UNO microcontroller and relay module [5, 6]. The software module uses Arduino
microcontroller programmed with Ardiuno IDE.
The primary element of this system is the ultrasonic sensor and RFID sensor.
The ultrasonic sensor is used to detect the density of vehicles in a lane and feed
the response to the centralized controller. Centralized controller make decisions to
regulate the traffic signal based on the data given by the ultrasonic sensor about the
traffic density. LEDs are used as traffic signal in this work. RFID Tag is fixed to
special vehicles like Ambulance, fire brigades and VIP vehicles. RFID is a technique
that automatically identifies the object through radio frequency waves. Various types
290 V. Radhika et al.

of RFID tags are available but majorly used tags are active and passive RFID tags in
this traffic management system. Passive tags don’t contain internal power source but
the life time of it is more than active tag. Active tags have internal battery as power
source and range of coverage is more than passive tags. Passive type tag is used in
this work. RFID tag has an integrated circuit for processing the data, modulating
and demodulating the RF signal and an antenna for transmitting and receiving the
data. The RFID reader module is fixed to the particular traffic signal and RFID tag
is placed to the emergency or special vehicles.
Arduino is an ATMega based microcontroller is used as a centralized traffic
controller and receives the ultrasonic sensor data and manages the timing of switching
of traffic light between red, green and yellow. The ultrasonic sensors fixed at four
sides of road in a four way lane that detects the vehicle density in a particular lane.
Ultrasonic sensor is generally used as a distance measuring sensor with a transmitter
and a receiver [7, 8]. Ultrasonic sensors are placed at regular interval of distance
across the lane. First ultrasonic sensor placed in the lane sends the ultrasonic wave
that hits the target surface and it reflects the wave back reaches the sensor and the
second sensor placed after 10 m of distance sends the ultrasonic wave get reflected
back to the sensor. The microcontroller counts the number of vehicle in a lane based
on the distance and timing for reflection of waves between sensors. RF module is
used for the communication between Arduino and the sensor. Figure 1 shows the
ultrasonic sensor placed in four lanes.

Fig. 1 Smart traffic management system with ultrasonic sensors

Smart Traffic Monitoring System 291

Fig. 2 Flow diagram for detecting vehicle density

The flow diagram for the count of the vehicles with ultrasonic sensor and
microcontroller is shown in Fig. 2.
Data from ultrasonic sensor is processed by the Arduino microcontroller and
programs the LEDs in traffic signal according to the density of vehicles. The
maximum time for tracking 5 min and minimum time for tracking the density is
1 min. Based on the tracking data received from the ultrasonic sensor the micro-
controller gives the priority to the lane with large vehicle density else the priority is
programmed based on Round- Robin method.
This system can consider the density of traffic of road by utilizing the ultrasonic
sensor and can also consider the priority of different types of vehicles [9, 10]. For
this to happen a RFID tag is fixed to the special vehicles or emergency vehicles.
RFID reader reads the RFID tag fixed to the special vehicles. If a special vehicle
enters the road with heavy traffic RFID reader gets the signal from RFID tag and
sends the signal to the microcontroller. Green signal is enabled in the particular lane
till the emergency or special vehicle passes away and traffic get sequenced. The
use of RFID tag differentiates the emergency and special vehicles from the other
292 V. Radhika et al.

Traffic

Special Vehicles RFID Reader Arduino Uno

RFID Tag

Fig. 3 Block diagram for detecting emergency vehicle

vehicles and avoids traffic congestion. The block diagram for the proposed work for
identifying the special vehicles is shown in Fig. 3.
The Arduino UNO is used as the microcontroller communicates with RFID reader
and the traffic signal. It establishes the communication between different modules and
generates the control signal for controlling the traffic light and different peripherals.
So the Microcontroller using RFID tag can alter the time of glowing of traffic lights
upon the arrival of the emergency signals and reduce the waiting time of those
vehicles.

4 Results and Discussion

Proposed work insights for controlling and managing the traffic by dynamically
changing the timing of traffic light. This proposed work reduces the congestion,
accident, waiting time and pollution. Results presented below shows the effective-
ness of proposed smart traffic monitoring system. The ultrasonic sensors along with
RF modules are placed at a regular intervals of distance in the road. The system is
placed near road as a stand-alone device. Microcontroller detects the number of vehi-
cles whenever the vehicle passes between the ultrasonic sensors in a particular lane.
The microcontroller control the timing of glowing of traffic light based on the vehicle
count and reduce the congestion at road intersections. Delays of traffic light are clas-
sified as LOW, MEDIUM and HIGH range. These ranges can be predefined based
on the vehicle count. Emergency vehicles are equipped with RFID tags and these
RFID tags are read by RFID reader placed at the traffic junction. If the emergency
vehicles are detected by the RFID reader it is given as an interrupt to the microcon-
troller. Microcontroller on receiving the interrupt changes the glowing conditions of
LEDs in the traffic signal. It makes the green LED in the particular lane to glow till
the emergency crosses the junction and after that system returns to normal working
based on the density of vehicles in each lane. The proposed system works effectively
in managing traffic at junctions of road and also gives priority based on the type of
vehicle. A prototype has been made to evaluate and show the arrangement of our
entire system.
Smart Traffic Monitoring System 293

Fig. 4 Placement of sensors

The Arduino UNO used as a microcontroller and programmed with C language.

Arduino is connected to different ultrasonic sensors and RFID reader. RF module
used for serial communication between microcontroller and sensors.
Placement of sensors is shown in Fig. 4 and traffic light setup is shown in Fig. 5.
Figure 6 shows the green signal for lane with more vehicles, Fig. 7 shows the red
signal for lane with less vehicles and Fig. 8 shows the green signal for an emergency
vehicle in a lane.

Fig. 5 Traffic light setup

294 V. Radhika et al.

Fig. 6 Green signal for vehicles

Fig. 7 Red signal for vehicles

5 Conclusion and Future Scope

In this work the traffic is managed based on the priority and the density of road
in different lanes of junction. The vehicle density is calculated as a first step in
this work and control the signal timing accordingly. The time of glowing for green
signal is reassigned. The lane with maximum density of traffic will receive maximum
timing for the vehicles to pass through. The lane with less traffic density will receive
Smart Traffic Monitoring System 295

Fig. 8 Green signal for emergency vehicle

minimum timing for the vehicles to pass through. Ultrasonic sensors and microcon-
troller plays a key role in this work to automate the traffic management system. This
work also gives preference to the emergency vehicles such as ambulance, fire vehicle,
and police cars by allowing traffic signal in the particular lane to glow to green till
the vehicle crosses the signal. RFID tag fixed to emergency vehicle and the RFID
reader at the traffic junction allows the emergency or special vehicle to pass through
with the help of microcontroller.
This work can be combined to a single system for detecting emergency vehicle
and density of vehicle in a particular lane without using separate system.

Acknowledgements Authors thank the Management, Principal and Head of the Department of Sri
Ramakrishna Engineering College for providing facilities and support to carry out this work.

References

1. Sanchez J, Gal M, Rubio E (2008) Applying a traffic lights ´evolutionary optimization technique
to a real case: ‘Las Ramblas’ area in Santa Cruz de Tenerife. IEEE Trans Evol Comput 12(2)
2. Ferrer J, García-Nieto J, Alba E, Chicano F (2016) Intelligent testing of traffic light programs:
validation in smart mobility scenarios. Hindawi Publishing Corporation, vol 3871046, no 1
3. Harrison C, Eckman B and R. Hamilton (2010) Foundations for smarter cities. IBM J Res Dev
54(1)
4. Mendonca M, Cowan D (2010) Decision-making coordination and efficient reasoning
techniques for feature-based configuration. Sci Comput Program 75:311
5. Sivasankar BB (2016) IoT based traffic monitoring using raspberry Pi. Intl J Res Eng Sci
Technol 1:2454
6. Perumalla BK, Sunil BM (2015) An intelligent traffic and vehicle monitoring system using
internet of things architecture. Int J Sci Res 6:391
296 V. Radhika et al.

7. Parkhi1 AA, Peshattiwar AA, Pande KG (2016) Intelligent traffic light system using vehicle
density. Int J Electr Electron Eng 8(172)
8. Lanke N, Koul S (2013) Smart traffic management system. Int J Comput Appl 75(93)
9. Tyagi V, Kalyanaraman S, Krishnapuram R (2012) Vehicular traffic density state estimation
based on cumulative road acoustics. IEEE Trans Intell Transp Syst 23:253
10. Shruthi KR, Vinodha K (2012) Priority based traffic lights controller using wireless sensor
networks. Int J Electron Signals Syst 1:2231
Harvesting Electrical Energy from Body
Heat to Power Bio Medical Wearables
M. Preethika, Mohamad Radzi Bin Ahmad, K. Porkumaran, S. Prabakar,
and R. Sundar

Abstract Energy harvesting has become a key factor in Biomedical applications.

Research on energy scavenging from body motion has been investigated to evaluate
the feasibility of powering a wearable or implantable systems. Energy from walking
has been previously extracted using generators placed on shoes, backpacks, and
knee braces while producing power levels ranging from milliwatts to watts. The
research presented in this paper examines the available power from body locations.
The finger, wrist, and palm were the chosen target localizations. There are many
untapped renewable energy resources such as wireless charging, thermal energy, and
vibration energy that can be used as an alternative to the conventional sources such as
solar, wind, and hydro. As we are moving towards Industry 4.0, new advancements
in technology has been introduced. Therefore, with a great advancement comes a
greater need for powering up devices by thermal technique. Implantable biomedical
equipment such as pacemaker and battery using biomedical wearables such as the
blood pressure monitor consumes low power and provides greater consistency. In
order to provide continuous supply to devices, a new approach of extracting power
from the heat of the human body is implemented thus avoiding unnecessary invasive
surgeries for implant.

Keywords Energy scavenging · Body motion · Power levels · Thermal technique ·

Implantable biomedical equipment · Body heat · Wearables

M. Preethika (B) · R. Sundar

Dr.N.G.P. Institute of Technology, Dr.N.G.P. Nagar, Coimbatore, Tamil Nadu, India
M. R. B. Ahmad
Universiti Tecknologi Petronas, Bandar Seri Iskandar, Tronoh, Perak, Malaysia
K. Porkumaran
Sri Sairam Engineering College, Chennai, India
e-mail: [email protected]
S. Prabakar
Sona College of Technology, Salem, India
e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 297
R. Ibrahim et al. (eds.), International Conference on Artificial Intelligence
for Smart Community, Lecture Notes in Electrical Engineering 758,
https://doi.org/10.1007/978-981-16-2183-3_29
298 M. Preethika et al.

1 Introduction

A fast-growing class of such devices is wearable, where sensor nodes are tightly
coupled with the human body, low power consumption is crucial in wearable systems
due to the light weight and size constraints for batteries, which severely limits
the energy that can be stored in the device. Wearables are; however, very tightly
constrained in terms of the size and weight and must also be compatable with the body.
Therefore, the possibilities for EH systems are more restricted than for other applica-
tions. Energy can be harvested from various environmental sources [1] including light
using photovoltaics [2], movement of the wearer [3], from radio frequency energy
(RF), or from temperature differences using thermoelectric generators (TEG). The
human body in contrast is a constant heat source. Typically, a temperature difference
exists between body core and the environment. Because of the voltages produced by
thermal harvesting are typically too low to power wearable electronics, a conversion
stage (DC-DC) with high conversion efficiency needs to be included into the wearable
system. A complete system analysis from the body heat to the application is required
to maximize both the output power and wearability. Thermal energy harvesting is
an another method of reducing the dependability on non-renewable sources. There
are various energy harvesting available, among them are electromagnetic, electro-
static, solar, and thermal methods. However, these mechanisms only produce a very
limited power. To meet the power necessity for most biomedical devices, we use
a boost converter to obtain the required output in order to satisfy the biomedical
wearable’s power requirement. The chosen target locations for power generation
here are the finger, palm, and wrist.

1.1 Problem Statement

Thermal energy from human body is rarely used to generate energy. The existing
thermal energy harvester is minimally used for Biomedical applications. Hence, there
is a need to develop an alternative charging method using thermal energy harvesting
to prolong the life of the battery.

1.2 Objectives

The objectives of this harvesting techniques are:

i. To explore the concept of thermal energy harvesting method for specific type
of Biomedical wearable devices.
ii. To model the thermal energy harvesting method and charge the Biomedical
device.
Harvesting Electrical Energy from Body Heat … 299

iii. To boost the minimial output using a boost converter for required power to
charge the biomedical device.
iv. A study of thermal energy harvesting method along the Electrical and
Electronical Engineering to produce a successful Energy Harvesting device.

2 Literature Review

2.1 Energy Generation from Human Body

Research on energy scavenging from body motion has been investigated to eval-
uate the feasibility of powering a wearable or implantable systems. Energy from
walking has been previously extracted using generators placed on shoes, backpacks,
and knee braces while producing power levels ranging from milliwatts to watts.
Thermal energy harvesting is an another method of reducing the dependability on
non-renewable sources. The energy harvested from heat uses Thermoelectric Gener-
ators (TEG), which is based on the principle of Seebeck effect [4]. According to [5],
it is a certain type of thermoelectric effect in which the temperature difference is
converted to electrical energy which can be represented by this equation:

E emf = Se T
or
V =∝ (TH − TC ) (1)

• Se / ∝ : Seebeck coefficient
• T: Temperature difference
• TH : Hot side temperature
• TC : Cold side temperature

A TEG comprises of two thermoelectric semiconductors n-type and p-type. These

semi-conductors are series-connected at the top and bottom of the junctions and are
subjected to two temperature difference which are hot and cold. The representation
is shown in Fig. 1.
Because of the different temperature at the top and bottom of the semiconductors,
the electrons and holes mobilize towards the opposite side, which gives rise to the
electric field between the cold and hot sides. As a result, an open circuit voltage is
produced at the terminals. To generate an output power, a load resister RL is connected
across the bottom of the cold side, i.e., the output terminal. Output power, PL can be
calculated from the voltage and current across the load resistor of the TEG. According
to [6], the electric power that is generated from the TEG is directly proportional to
the temperature difference between the surfaces of the TEG. Therefore, the internal
resistor, Rin = RL, a maximum power, Pmax is produced and can be expressed in
the equation as follows:
300 M. Preethika et al.

Fig. 1 Semi-conductor
thermocouple seebeck effect

∝2 T 2
PL max = (2)
4Rin
However, in [1], it is also known that the open-cicuit output voltage, Voc of a TEG
module is also directly proportional to the temperature difference between the two
surfaces. The equation for that can be expressed below:

VOC =∝ mT

Thermal energy harvesting is in DC. Thus, only a DC-DC converter is needed.

This is used to scale the voltage from the input to the output load. To develop an
efficient DC-DC voltage conversion, it is recommended to use an inductor-based
converter which stores the energy using an external inductor and then releasing it to
the load. However, since the voltage and current needs to be stepped up, usually a
boost convertor is introduced as a step-up transformer to overcome the drawbacks
of MOSFET switching; however, it is unable to sustain very low voltages (under
50 mV), which is the case with the thermoelectric generation.

2.2 Human-Based Energy Harvesters

According to [7], the typical self-powered wearable device technology’s power

requirement is in the range of 0.01–10 mW and is dependent on its applications.
It is found that the power levels from the thermal energy, generated by the human
body is greater than the power generated by the photovoltaic cells (Fig. 2).
Figure shows the heat transfer of a thermal generator for a TEG on the human
skin.
T = Temperature of the body core, skin, radiator, ambient air.
R = Thermal resistances.
W = Heat flow.
It is found statistically from the research that at the usual location of a watch, wrist,
thermal resistance at an ambient temperature of 22.7 °C is 440 cm2 K\W, while under
the watchstrap, which is the radial artery, it decreases to 120–150 cm2 K\W. Thus,
Harvesting Electrical Energy from Body Heat … 301

Fig. 2 Thermal circuit

equivalent of a TEG

Fig. 3 (a) Off-the-shelf

thermopile

the thermal resistance in a normally clothed person changes a little or not at all in the
trunk if the ambient temperature decreases as the average body’s thermal resistance
is 140 cm2 K\W.
Figures shows the miniaturized thermopile attached to the skin [2] (Fig. 3).
From Figure (a):
The miniaturized thermopile attached to the skin is shown in the figure provided
with 1. Ceramic plates which is a metal that acts as an outer covering for the ther-
mopile holding the 2. Thermocouple legs which are interconnected along the 3. Metal
interconnects altogether attached to the ceramic plate (Fig. 4).
From Figure (b):
4. The human skin is kept in contact with the 5. hot plate which is placed parallel
to the 6. cold plate from where the temperature difference T is noted between the
two plates. The 7. holding, encapsulating, and anti-shock elements are provided for
holding, covering, and providing resistance to the shock, respectively. Primarily the
302 M. Preethika et al.

Fig. 4 (b) Simulated TEG

on the human body

Fig. 5 (c) Simulated TEG

on the human body

temperature difference is recorded only with the placement of a 8. piece of clothing

between the skin and the two plates (Fig. 5).
From Figure (c):
Secondly, along with the arrangements of figure (b) 9, watchstrap is kept in
contact with the 5. hot plate instead of 8. a piece of clothing, and the tempera-
ture difference T is recorded. 10. Pins of radiator are the heat exchangers used to
transfer the thermal energy between [2] the human skin and two plates. Finally, the
11- Shock protection grid which is an integral component of electrical safety from
which the normal voltage will be identified and will provide the information needed
to determine the shock prevention method.
Harvesting Electrical Energy from Body Heat … 303

3 Proposed Work

Primarily, after understanding the concept behind the thermal energy harvesting,
I have decided to use a method for heat generation from the human body and to
construct a mathematical model for the thermal energy harvesting.
Secondly, based on the constructed mathematical model, a simulation was done
using MATLAB Simulink and that was to test whether it is feasible or not for the
required output, in order to charge the battery of Biomedical wearables. In order to
obtain the output in volts, I used a boost converter to boost the output obtained from
the thermal energy harvester and is then fed to the battery of the biomedical wearables.

3.1 Mathematical Equation

Thermal Energy Harvesting:

Voc = α × T (1)

where Voc is the output voltage, α is the Seebeck coefficient, and T is the difference
in temperature between the hot and cold surfaces.

P=η (2)

where P is power, η is the thermal efficiency, and Q is the heat generated.

From another equation,

VOC 2
P= (3)
4Re

where Rel is the electric resistance

Substituting Eqs. (2) in (3) we get,

VOC 2
Q= (4)
4ηRel

Substituting Eqs. (1) in (4) we get,

α × (T )2
Q= (5)
4η R

Thus Eq. (5) is the equation we simulate in MATLAB Simulink.

304 M. Preethika et al.

Fig. 6 MATLAB Simulink Block Diagram

3.2 Simulink Block

Thermal Energy Harvesting (Fig. 6).

3.3 Thermal Energy Harvesting

The output voltage of thermal energy harvesting is shown in figure based on the
simulation done using MATLAB Simulink.
Harvesting Electrical Energy from Body Heat … 305

According to the theoritical calculatuons of Eq. (1) of the output voltage, VOC , it
is found that VOC = 0.98 V. However, the experimental value based on the simulation
done shows that the output voltage, VOC = 0.976 V, which is close to the theoritical
value. Figure representing the output power, P of the thermal energy harvester.
306 M. Preethika et al.

According to the theoritical calculations of Eq. (2) of the output power, P, it is

found that P = 22.56 mW. However, the experimental value based on the simulation
done shows that the output voltage, P = 24.4 mW, which is much similar to the
theoritical value.

3.4 Thermocouple

The thermoelectric generator used as the thermal energy harvester is an off-the-shelf

20 mm × 20 mm thermocouple with a 10-cent coin as scale. The datasheet of the
thermocouple given by the supplier, where the parameters for the hot side temperature
is 200 °C and cold side temperature is 30 °C.
Thermocouple Output
The thermocouple is tested to see how much voltage it can produce by using a
voltmeter. As this will be used to power up a wearable device, multiple points on
the hand on two separate surfaces is tested to determine which will produce more
voltage. In addition to the obtained output. the boost converter circuit is attached to
increase the output voltage to power the biomedical wearable devices. Table 1 states
the maximum voltages obtained from the test.
Harvesting Electrical Energy from Body Heat … 307

Table 1 Output voltage results from multiple points in hand on two surfaces
Plastic lid output in V Counter top output in V
Part of hand Thermal Boosted Part of hand Thermal Boosted
output in V output in V output in V output in V
10.8 mV 4.95 V 17.8 mV 8.0 V

Finger Finger
17.88 mV 8.4 V 22.5 mV 9.94 V

Wrist Wrist

20.79 mV 9.7 V 22.45 mV 9.93 V

Palm Palm

3.5 Boost Converter Multisim

Considering the below boost converter circuit as a sample for the generation of output.
The values from the TEG are fed to the boost converter circuit for output appraisal
and then, it is fed to the biomedical wearables (Fig. 7).

4 Problem Solution

Based on the results in Table 1, it is found that when the thermocouple is on a counter-
top, the voltage produced is higher than the plastic lid. However, it is difficult for the
thermocouple to maintain its stability and temperature. Thus, the voltage fluctuates
and does not give a stable reading.
308 M. Preethika et al.

Fig. 7 Boost converter multisim

4.1 Boost Converter Output

Provided with the output of the finger, palm, and wrist readings by Multisim, the
given input is boosted to obtain the required output in order to charge the biomedical
wearables. The output will be generated by using the boost converter Multisim and
fed to biomedical wearable devices such as the blood pressure monitor, pulsoximeter,
heart rate sensor, temperature sensor, and pulse sensor.

5 Conclusions

In conclusion, it is found that the thermal energy harvesting simulation from

MATLAB Simulink produced an output voltage Voc of 0.98 V, which is very
much similar to the theoretical value 0.95 V. However, when conducting an experi-
ment at 3 points on the hand; fingers, wrist, and palm from the purchased TEG, it is
found that the output voltage produced a significantly lower power due to external
factors such as thermal equilibrium and fluctuating atmospheric temperature. The
maximum output voltage obtained from the fingers, wrist, and palm are 13.6 mV,
22.4 mV, and 17.2 mV, respectively. Therefore, the boost converter is used to increase
the voltage output from the TEG. Thus, the appraisal values are 4.95 V, 8.4 V, and
9.7 V, which is in volts. Thus, the obtained output is very much supportive and
effective to charge a Biomedical Wearable’s Battery.
Harvesting Electrical Energy from Body Heat … 309

References

1. Udalagama CJ (2010) Electrical energy generation from body heat. IEEE ICSET
2. Arabee SBM (2018) A study of wireless charging and thermal energy harvesting methods for
IOT applications
3. Thielen M, Sigrist L, Magno M, Hierold C, Benini L (2017) Human body heat for powering
wearable devices: from thermal energy to application. Energy Convers Manage 131:44–54
4. Alhawari M, Mohammad B, Saleh H, Ismail M (2013) A survey of thermal energy harvesting
techniques and interface circuitry. IEEE, pp 381–384
5. Ghosh A, Meenakshi Khalid S, Harigovindan VP (2015) Performance analysis of wireless body
area network with thermal energy harvesting. In: Proceedings of 2015 global conference on
communication technologies (GCCT 2015), pp 916–920
6. Ahiska R, Mamur H (2013) Design and Implementation of a new portable thermoelectric
generator for low geothermal temperatures. IET Renew Power Gener
7. Leonov V, Su J, Vullers RJM (2010) Calculated performance characteristics of micromachined
thermopiles in wearable devices. EDA Publishing/DTIP
The Significant Relationship Between
E-Governance System and Government
Operation Excellence

Almahdy Alhaj Saleh, Ahmed Rashid Alkhuwaylidee, Murugan Thangiah,

and Arif Razzaq

Abstract Each government is looking for to provide the best services to estab-
lish efficiency and quality of performance. This goal could be accomplished by
improving the service performance of entire sectors in society. The government of
Syria has realized the importance of moving in the direction of information tech-
nology. Therefore, E-governance initiatives were launched in Syria as a part of
overall country information technology in 20 s century. Each government sector has
since upgraded the performance by having its websites and e-services application.
However, there are gaps and loose connections exist among the sectors, which has
accordingly tarnished the image of Syrian E-governance. This has led to significant
questions about the requirement of modification and enhancement of such service.
Hence, the purpose of this research is to investigate and explore the factors that drive
the E-governance implementation and affect government performance as well as the
government-citizen relationship in Syria.

Keywords E-governance · Government of Syria · Government-Citizen

A. A. Saleh (B) · M. Thangiah

Postgraduate Centre Limkokwing University of Creative Technology Selangor, Selangor,
Malaysia
M. Thangiah
e-mail: [email protected]
A. R. Alkhuwaylidee
Ahmed Rashid Alkhuwaylidee Computer Engineer Technology Mazaya University College
Nassirya, Nassirya, Iraq
A. Razzaq
Faculty of Information and Communication Technology, Universiti Teknikal Malaysia Melaka,
Melaka, Malaysia

© The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 311
R. Ibrahim et al. (eds.), International Conference on Artificial Intelligence
for Smart Community, Lecture Notes in Electrical Engineering 758,
https://doi.org/10.1007/978-981-16-2183-3_30
312 A. A. Saleh et al.

1 Introduction

In recent years, new technologies and concepts have therefore altered government
contact with relation aspect such as organizations and citizens [1–3] by launching
original service styles, similar to e-learning, e-banking, e-health, e-commerce, e-
business, e-voting and others, the electronic government has been popularized with
the term E-governance.
This refers to the services supplied by governmental to citizens via providing new
information and communication technology technologies that may allow creating
provision faster and easier. It had been first presented at the ‘90 s, it is regarded as an
innovation introduced by using high-tech development from information technology
increase [4].
Most scholars have divided E-governance into three main categories which are:
Government to Government (G2G), Government to Citizens (G2C) and Government
to Business (G2B). According to Fang (2002) and a few scholars have invented a
fourth category which is Government to Employee (G2E) [5, 6], however, none of
the scholars has explored the interaction between the opposite categories beyond
the government. Category as they consider them to be the e-commerce a part of
the E-governance. For instance, Business to Citizen (B2C), Business to Employee
(B2E) and Employee to Citizen (E2C), don’t seem to be yet identified within the
E-governance system.
Moving closer to E-governance would provide several benefits to a republic’s
kind of services, collectively with finances procedure and control. Certainly, it would
enhance the processing of the government in various components for the improve-
ment of Government Operation Excellence (GOE). A range of these benefits is
decreasing expenses and growing activity thanks to the discount of printed paper
and acquisition of documents and area for storage [7–9].
Reducing the workload on employees is some other advantage to the work envi-
ronment due to the fact employees are concerned with the minimal quantity of citizens
and this can create an increased and serene environment. E-governance will advan-
tage country areas together with large cities since citizens can practice for services
lengthy as they’re a part of the republic and in some instances and services, citizens
may additionally practice for and request a service whilst they’re overseas. Further
advantages to the environment, knowledge, etc., should even be counted for the
improvement of the country [10]. Additionally, E-governance has some challenges
and downsides, and these should not be forgotten or ignored. Just to name a few of
these difficulties are illustrated underneath cost, unreachability, hyper-surveillance,
loss of transparency and accountability [11, 12].
The Significant Relationship Between E-Governance System … 313

2 The Proposed Method

The present research will use a quantitative approach to research given that it will
examine the relationship between several independent variables as well as the accep-
tance of E-governance system by users in Syria. The dependent variables of the
study are the Government Operation Excellence (GOE), while the independent vari-
ables are Citizens-centricity, facilitating conditions, performance expectancy, effort
expectancy, flexibility system. And the mediation variable are behavioral intention.
Given the quantitative leanings of the research, the study will thereby focus on
existing statistical data, facts and figures. The accompanying data collection tools
have been identified as questionnaires. Questionnaires were particularly effective in
collecting quantitative data, particularly if researchers are aware of how the variables
are to be tested [13]. In addition to this, questionnaires are also effective when it comes
to collecting data from a wide sample of respondents, dispersed across a number of
locations. Using a questionnaire also allows researchers to ensure that respondents
are presented with a uniform set of questions, this will later result in uniform sets of
data which are not challenging as far as analysis is concerned.
Questionnaires, if designed effectively are unlikely to be time-consuming, there-
fore allowing data to be collected in a relatively short space of time [13–15]. Given the
fact that the research sample is geographically dispersed across Syria, the researcher
relied on a questionnaire approach to investigate.

3 Research Method

According to [16], the method is a set of principles that in any particular situation has
to be reducing to a method uniquely suited to a given situation. The research method-
ology for this research, it presents a systematic and organized procedure that enables
the investigation of the relationship between the different variables attached, namely
E-governance, behavioral intention to use and government operation excellence.
In achieving the objectives, the research methodology deals with procedures
involved in this research. Questionnaire targeting to survey the adoption information
technology for promoting E-governance system among citizens in Republic of Syria
are including in the instrument.
Series of tests such as Structural Equation Model (SEM), and reliability analysis
were conducted to confirm that the instruments were valid (Reliability and Conver-
gent Validity). Also, the research methodology that was followed and adopted by
this research was detailed out. Includes all the procedures, process, and guidelines
that according to research literature is suitable for this kind of the investigation.
The main aim of this research is to study and investigate the mediation effect of
behavioral acceptance on the relationship between E-governance and government
operation excellence in Syria.
314 A. A. Saleh et al.

3.1 Research Framework

The review of the information system relevant literature indicates that the acceptance
of information systems, especially E-governance systems are influenced by three
factors, namely: (1) individual factors (2) system factors (3) environmental factors
[17–22].
Concerning theoretical contributions, the principal theories that are used in this
study are the Implications of DeLone and McLean model, Technology Acceptance
Model (TAM), the Task-Technology Fit (TTF) and Unified Theory of Acceptance
and Usage Technology (UTAUT). In this research we combining thus theories into
a new framework in order to achieve the optimum GOE for the Syrian citizens.
Hence, the evaluation of GOE will be supported by the equal common criterion
as is stipulated within the SERVQUAL approach. It will be classified into assurance,
tangibles, reliability, responsiveness and empathy [23].
The focus of this study is to examine the effect of E-governance on GOE.
H1: There is a significant relationship between an E-governance system and
Government Operation Excellence.
Conceptualization of Factors.
Information and communication technology review of the literature has shown
that a number of studies have been conducted to investigate users’ acceptance of
E-governance systems [24–26].
The majority of previous studies have focused on employees and citizens envi-
ronments within economically developed nations. While developing countries have
much to gain from utilising information technology, they usually receive little to no
attention in regards to research [24–30]. To do so, they need to improve upon three
aspects: the implementation environment, individual factor, and system factor. In
that respect, the current study will aim to develop an understanding of the various
factors which influence the acceptance of E-governance system by Syrian citizens.
Behavioral intention.
Behavioral intention to use: Citizens get motivated to use a certain system or
application if it was used or marketed by social groups, friends or family or any
matter that touches the person’s behavior or feeling [31]. A number of authors [31–
34] have attested to the fact that BI has been employed and utilized as a dependent
variable in the examination of E-government acceptance.
Compeau and Higgins [34] Argues that intention is the determinant of the theoret-
ical foundation in information system research in the sense that individual behavior
dictates whether or not to use or not use the technology. Also, BI refers to specific
prompts that cause individuals to act in certain ways or perform certain behav-
iors [35]. This study adding three related determinants of behavioral intent to use
E-governance. They are as follows: system flexibility, citizens-centricity, and facil-
itating conditions. They are added to reflect the individual factor because they have
a significant influence on the intention to use of E-governance.
The Significant Relationship Between E-Governance System … 315

H2: Behavioral Intention to Use moderates the relationship between an E-

governance system and Government Operation Excellence.
Performance Expectancy (PE).
A number of authors including [36, 37] outline PE as an individual’s expectations
of improvement, when using a specific system which could enhance their work
performance. A host of the literature suggests that performance expectancy is a key
variable in predicting user intention in information and communication technology
contexts [32, 37–40].
Adapting this construct (PE) to E-governance system suggests that Syrian citizens
think that using E-governance system will enhance their learning and consequently
will be reflected on their efficiency. Within an E-governance context, there are many
studies which suggest that PE directly influences user intent to use an E-governance
system [21, 41, 42].
H2: Performance expectancy will have a positive influence on behavioral intentions
to use E-governance services.
Effort Expectancy (EE).
Effort expectancy, as described by both [36, 37], is the level of ease felt by individ-
uals when using information systems. Several studies have demonstrated that effort
expectancy has a powerful effect on behavioral intention. That is to say, the intent to
use information system technology [36, 38, 43]. In that sense, then, effort expectancy
theory suggests that individuals, in this case, will usually accept a system if they feel
that said the system is easy to use. This study will aim to determine the degree to
which independent factors such as effort expectancy correlate with the intent to use
E-governance system.
H3: Effort expectancy will have a positive influence on behavioral intentions to
use E-governance services.
Citizens-Centricity (CC).
A citizen centric involves the availability of services from the top user’s point of
view instead of the attitude of the government department [44, 45]. E-governance
project should have a citizen centric perception. However, citizen-centric service
delivery is additionally an elaborate issue with many perspectives that require to be
considered at the very beginning of a change project. His provision of citizen centric
service has been identified by some researchers as a critical success factor [46].
The citizen-centric delivery E-governance services determinants in developing
countries would allow for better understanding of the citizens’ needs, desires, require-
ments and priorities that must be taken into consideration by governments to ensure
the success of services [47].
H4: Citizens-Centricity will have a positive influence on behavioral intentions to
use E-governance services.
System Flexibility (SF).
System flexibility within the context of E-governance refers to the level of freedom
that citizens’ feel E-governance permits them, in terms of time and place [48].
316 A. A. Saleh et al.

[49] Study concluded that flexibility greatly impacts behavioral intent. Moreover,
[50] explored the differing variables which influenced user satisfaction in regards
to online courses. [50] Concluded that, in most cases, perceived flexibility is vital
regarding user satisfaction. System Flexibility, which is the ability to implement
different processes and apply different facilities to achieve the same goals. It consists
of items such as: product volume flexibility, people flexibility, etc.; [51].
E-governance literature has also provided us with many studies that indicate the
predictive ability of flexibility of E- governance system in determining behavioral
intention [21]. In this research, Syrian citizens will adopt the E-governance system
if they perceive that they can use the system at any time from any place.
H5: System Flexibility will have a positive influence on behavioral intentions to
use E-governance services.
Facilitating Conditions (FC).
Facilitating conditions, as defined by Venkatesh (2003) are the factors which
lead users to believe that the technical and organizational infrastructure of Syria
governance can support E-governance application system. In that respect, facilitating
conditions can result in the rejection of information technology systems due to lack
of faith in the respective organizations and their systems themselves [37]. [39] Also
stated that facilitating conditions resources, such as money and time would motivate
a person to use a particular system. Several studies in the information technology
context have indicated that the facilitating conditions construct has a significant direct
influence on user behavior [52–54].
Adapting this construct to E-governance system suggests that Syrian citizens
perceive that technology infrastructure, and governance infrastructure will enable
them to use the E-governance system without any problem, and to interact with other
users [55].
H6: Facilitating conditions will have a positive influence on behavioral intentions
to use E-governance services.
SERVQUAL.
Jong and Wang [56] Refined their previous service quality evaluation model [57]
and created the SERVQUAL. In short, the authors suggest that a practical service
quality perception stems from the chasm between consumer expectations and there-
fore the performance of the service that’s actually delivered. Thus, finishing up new
studies, with new data collections and analysis, they were able to improve their
scale, reducing the initial ten dimensions to seven: (1) Tangibles; (2) Reliability; (3)
Responsiveness; (4) Communication, Credibility, Safety, Competence; 5) Courtesy;
6) Knowing the customer; and (7) Access.
During a second stage, which entailed another improvement of the study,
SERVQUAL was reduced to 5 dimensions (Table 1). According to [57], in spite
of the kind of service, consumers basically use an equivalent criterion to assess
quality. Service quality could be a general opinion the client forms regarding its
delivery, which is constituted by a series of successful or unsuccessful experiences.
It is difficult to live the standard of service operations because they need the
characteristic intangibility. Geared toward solving this problem, [57] developed a
The Significant Relationship Between E-Governance System … 317

Table 1 SERVQUAL dimensions

dimensions Definition
Tangibles The physical appearance of equipment and staff involved in the provision
of services
Reliability The ability to deliver the offered service in the scheduled date and time,
independently of any problems that may arise
Responsiveness The ability to help every client in an unhesitating way and to effectively
and positively promote the service
Assurance The ability to inspire trust, security and technical quality on the part of the
staff when delivering a service
Empathy The available and unhesitating assistance provided to the client,
individualized and focused on his main interests

Fig. 1 Research framework

strategy within which there’s a comparison between several orders of expectations

and perceptions of service quality by the patron (Fig. 1).

4 Results and Discussion

The measurement framework in this study was made up of 42 items to measure 11

constructs, namely: Citizens-Centricity (CC), Facilitating Conditions (FC) System
Flexibility (SF), Effort Expectancy (EE), and Performance Expectancy (PE).
Empathy (EMP), Responsiveness (RES), Reliability (REL), Assurance (ASSU),
Tangibles (TAN), and Behavioral Intentions (BI).
318 A. A. Saleh et al.

Table 2 Table type styles

Fit index Modified model Recommended Source
values
Df 1013
CMIN (χ2) 1238.809
p-value 0.000 > 0.05 [58, 59]
χ2/df 1.223 ≤ 5.00 [58]
GFI 0.881 ≥ 0.90 [59]
AGFI 0.867 ≥ 0.80 [60]
CFI 0.976 ≥ 0.90 [58, 61]
TLI 0.974 ≥ 0.90 [62, 63]
IFI 0.976 ≥ 0.90 [62, 63]
RMSEA 0.020 ≤ 0.10 [64]

4.1 Reliability and Convergent Validity

Once the uni-dimensionality of the constructs was achieved, each of the constructs
was assessed for their reliability and validity. Reliability is assessed using Cronbach’s
alpha, construct reliability (CR) and average variance extracted (AVE), while for
validity using construct, including convergent and discriminant Table 2. Represents
the result of Cronbach alpha and convergent validity for the modified measurement
model with 42 reminder items.
As shown in Table 2, the remaining indicators have high factor loadings ranging
from 0.863 to 0.915 indicating that these indicators have preserved the meaning of
the factors.
Table 2 also shows that the AVE, which reflects the overall amount of variance in
the indicators accounted for by the latent construct, was above the cut-off 0.5 for all
constructs as suggested by [65], ranged from 0.531 to 0.684.
The composite reliability values, which depict the degree to which the construct
indicators indicate the latent construct, exceeded the recommended value of 0.6 for
all constructs as recommended by [66], ranging from 0.797 to 0.915.
The Cronbach’s Alpha values, which describes the degree to which a measure
is error-free, range from 0.795 to 0.914 which were above the threshold of 0.7 as
suggested by [58]. Therefore, the achieved Cronbach’s Alpha for all constructs was
considered as sufficiently error-free.

4.2 Goodness of Fit Indices

Based on the current GOF results, the chi-square is substantial at the 0.010 level. The
normed fit index of lowest difference chi-square can be removed if the sample size
for the research is more than 200 [82, 83]. After correcting for degrees of freedom
The Significant Relationship Between E-Governance System … 319

Table 3 Type styles

Construct Internal reliability Convergent validity
cronbach alpha Average variance extracted Composite reliability
(AVE)a (CR)b
CC 0.795 0.567 0.797
EMP 0.843 0.531 0.819
RES 0.856 0.599 0.857
FC 0.854 0.576 0.844
REL 0.896 0.684 0.896
TAN 0.858 0.632 0.872
ASSU 0.809 0.597 0.815
BI 0.862 0.557 0.863
SF 0.862 0.611 0.863
EE 0.914 0.641 0.915
PE 0.892 0.577 0.891

in based on the number of variables, the adjusted GFI (AGFI) was 0.881, which was
greater than the recommended cut-off value of according to [84, 85]. The model
predicted 86 percent of the variances and covariance in the survey results, according
to the report. The model showed a satisfactory fit of collected data on the CFI, TLI,
and IFI indices with values more than the cutoff value of 0.9 (i.e., 0.976, 0.973, and
0.975, respectively) [86, 87]. In addition, the root-mean square error of approximation
(RMSEA) was 0.020, and that was lower than the suggested threshold of 0.1 [88, 89,
90]. Furthermore, the fact that the Comparative CMIN/df (1.224) has been less than
5 revealed the good fit of the model [89, 90]. See Table 3.

5 Conclusion

The study has distributed 600 questionnaires to Syrian citizens. From the total, only
423 questionnaires were returned, with only 393 questionnaires were usable (65.5%).
16 respondents returned the questionnaire with missing data of more than 30% for
each questionnaire. According to [67–70] the questionnaires that have missing data
of more than 25% should be omitted. Additionally, 14 of the respondents returned
the empty questionnaire. Therefore, only 385 cases have been used for the study
analysis.
A total of six hypotheses has been tested in achieving the objectives of the research.
From the total, five direct relationships have been found to be statistically significant
to users’ intention to use framework of E-governance system, since the study indi-
cates that the first significant relationship is between Performance Expectancy and
Behavioral Intention (0.000). The second significant relationship is between Effort
320 A. A. Saleh et al.

Expectancy and Behavioral Intention (0.000). The third significant relationship is

between Citizens-Centricity and Behavioral Intention (0.026). The fourth significant
relationship is between System Flexibility and Behavioral Intention (0.002). The fifth
significant relationship is between Facilitating Conditions and Behavioral Intention
(0.000).
When taking into consideration the indirect relationship, two relationships were
proposed to statistically affect between E-governance systems (EGOV) and Govern-
ment Operation Excellence (GOE). The coefficient parameters estimate of the struc-
tural model are examined to determine whether E-governance (EGOV) as indepen-
dent variables have positive indirect effects on Government Operation Excellence
(GOE). Through Behavioral Intention (BI).
The first important indirect relationship was proposed between E-governance
(EGOV) and Government Operation Excellence (GOE) with a total effect of (0.119).
The second important indirect relationship was proposed between E-governance
(EGOV) and Government Operation Excellence (GOE) with Satisfaction Services
with a total effect of (0.096). The third important indirect relationship was proposed
between E-governance (EGOV) and Government Operation Excellence (GOE)
through Satisfaction Services with a total effect of (0.184). The fourth important
indirect relationship was proposed between E-governance (EGOV) and Government
Operation Excellence (GOE) with a total effect of (0.119). The fifth important indirect
relationship was proposed between E-governance (EGOV) and Government Opera-
tion Excellence (GOE) with Behavioral Intention (BI) with a total effect of (0.088).
The sixth important indirect relationship was proposed between E-governance
(EGOV) and Government Operation Excellence (GOE) through Behavioral Intention
(BI) with a total effect of (0.031).

References

1. Lee M-C (2010) Explaining and predicting users’ continuance intention toward e-learning: an
extension of the expectation-confirmation model. Comput Educ 54:506–516
2. Rokhman A (2011) E-Governance adoption in developing countries; the case of Indonesia. J
Emerg Trends Comput Inform Sci 2:228–236
3. Gil-Garcia JR, Zhang J, Puron-Cid G (2016) Conceptualizing smartness in government: an
integrative and multi-dimensional view. Gov Inf Q 33:524–534
4. Grönlund Å, Horan TA (2005) Introducing E-Gov: history, definitions, and issues. Commun
Assoc Inf Syst 15:39
5. Bertot JC, Jaeger PT, Mcclure CR (2008) Citizen-Centered e-governance services: benefits,
costs, and research needs. In: Proceedings of the 2008 international conference on digital
government research, 2008. Digital government society of North America, pp 137–142
6. Carter L, Belanger F (2004) The influence of perceived characteristics of innovating on e-
governance adoption. Electron J E-Governance 2:11–20
7. Tolbert CJ, Mossberger K (2006) The effects of e-government on trust and confidence in
government. Public Adm Rev 66:354–369
8. Helbig N, Gil-García JR, Ferro E (2009) Understanding the complexity of electronic
government: implications from the digital divide literature. Gov Inf Q 26:89–97
The Significant Relationship Between E-Governance System … 321

9. Sharma SK, Govindaluri SM, Sharma RK (2013) A two-staged regression-neural network

model for understanding and predicting the quality determinants of e-governance services in
Oman. Int Found Res Dev (Ifrd) 57
10. Huang Z, Bwoma PO (2003) An overview of critical issues of e-governance. Issues Inf Syst
4:164–170
11. Bertot JC, Jaeger PT, Grimes JM (2010) Using Icts to create a culture of transparency: e-
governance and social media as openness and anti-corruption tools for societies. Gov Inf Q
27:264–271
12. Alzahrani AI (2011) Web-Based e-governance services acceptance for G2c: a structural
equation modelling approach
13. Sekaran U (2007) Metodelogi Penelitian Untuk Bisnis (Terjemahan). Edisi Empat, Jakarta,
Salemba Empat
14. Wozencroft A, Hardin R (2011) Examining counselor satisfaction in an outdoor therapeutic
recreation camp. Entertainment Manage
15. Alkhuwaylidee A, Almahdy AS (2019) Syrian e-governance framework toward government
excellence service. In: Paper presented at the technologies for developing information systems
Tris-2019
16. Checkland P (1981) Systems theory. Syst Pract
17. Zhang JQ, Craciun G, Shin D (2010) When does electronic word-of-mouth matter? a study of
consumer product reviews. J Bus Res 63:1336–1341
18. Saleh AA, Alyaseen IFT (2021) Successful factors determining the significant relationship
between e-governance system and government operational excellence. Bull Electr Eng Inf
10(6):3460–3470
19. Algharibi A, Keung LC, Niukyun S, Zhao L, Tyler E, Taweel A, Hobbs R (2012) Adapting
the unified theory of acceptance and use of technology (Utaut) as a tool for assessing future
technology adoption of e-trial software tools. In: Paper presented at the proceedings of the 10th
international conference on information communication technologies in health
20. Alzahrani ME, Goodwin RD (2012) Towards a utaut-based model for the study of e-governance
citizen acceptance in Saudi Arabia. In: Proceedings of world academy of science, engineering
and technology (No. 64). World academy of science, engineering and technology
21. Alkhuwaylidee AR (2019) Extended unified theory acceptance and use technology (utaut) for
e-learning. J Comput Theor Nanosci 16(3):845–852
22. Cody-Allen E, Kishore R (2006) An extension of the utaut model with e-quality, trust, and
satisfaction constructs. In: Paper presented at the proceedings of the 2006 ACM Sigmis CPR
conference on computer personnel research: forty four years of computer personnel research:
achievements, challenges & the future
23. Nanayakkara C, Widdett D (2005) A model of user acceptance of e-learning technology: a case
study of a polytechnic in New Zealand. In: The fourth international conference on information
system technology and applications. New Zealand, pp 180–189
24. Huai J (2011) Quality evaluation of e-governance public service. In: 2011 International
conference on management and service science, 2011. IEEE, pp 1–4
25. Chau PYK, Hu PJH (2002) Investigating healthcare professionals’ decisions to accept
telemedicine technology: an empirical test of competing theories. Inf Manage 39(4):297–311
26. Al-Khatib H, Lee H, Suh C, Weerakkody V (2019) E-Governance systems success and user
acceptance in developing countries: the role of perceived support quality
27. Prieto-González L, Matsuo MT, Colomo-Palacios R (2014) Electronic business and marketing
28. Susanto TD, Diani MM, Hafidz I (2017) User acceptance of e-governance citizen report system
(a case study of city113 app). Procedia Comput Sci 124:560–568
29. Taherdoost H (2018) Development of an adoption model to assess user acceptance of e-service
technology: e-service technology acceptance model. Behav Inf Technol 37:173–197
30. Maharjan M (2019) Citizens’ acceptance of e-governance services in Nepal: applying utaut
model. In: Advanced engineering and ict–convergence 2019 (ICAEIC-2019), p 122
31. Abbad MM et al (2009) Looking under the bonnet: factors affecting student adoption of e-
learning systems in Jordan. In: The international review of research in open and distance
learning, vol 10, no 2
322 A. A. Saleh et al.

32. Hung S-Y, Chang C-M, Kuo S-R (2013) User acceptance of mobile e-governance services: an
empirical study. Gov Inf Q 30:33–44
33. Alkhuwaylidee A, Almahdy AS (2019) Syrian e-government framework toward government
excellence service. Texnologii Pazpabotki Infopmacionnyx Cictem Tpic-2019
34. Compeau DR, Higgins CA (1995) Computer self-efficacy: development of a measure and initial
test. MIS Q 189–211
35. Hsia J-W, Tseng A-H (2008) An enhanced technology acceptance model for e-learning
systems in high-tech companies in Taiwan: analyzed by structural equation modeling. In:
2008 International conference on cyberworlds. IEEE, pp 39–44
36. Hasan H, Ditsa G (1999) The impact of culture on the adoption of it: an interpretive study. J
Glob Inf Manage (JGIM) 7(1):5–15
37. Barua M (2012) E-Governance adoption in government organization of India. Int J Manag
Public Sector Inf Commun Technol 3(1):1
38. Venkatesh, V., Morris, M. G., Davis, G. B. & Davis, F. D. 2003. User Acceptance Of Information
Technology: Toward A Unified View. Mis Quarterly, 425–478
39. Davis F, Bagozzi R, Warshaw P (1992) Extrinsic and intrinsic motivation to use computer in
the workplace. J Appl Soc Psychol 22(14):1111–1132
40. Naor, O., & Geri, N. (2008). Easy as e-mail? Probing the slow adoption of an online assign-
ment submission system. In: Proceedings of the chais conference on instructional technologies
research 2008, vol 5, pp 94–101
41. Taylor S, Todd P (1995) Understanding information technology usage: a test of competing
models. Inf Syst Res 6(2):144–176
42. Venkatesh V, Davis FD (2000) A theoretical extension of the technology acceptance model:
four longitudinal field studies. Manage Sci 46:186–204
43. Goussal DM, Lezcano MSU (2003) Synchronous distance learning and virtual classrooms: a
case study on student expectations and conditioning factors. Australas J Educ Technol 19
44. Sahin I, Shelley M (2008) Considering students’ perceptions: the distance education student
satisfaction model. Educ Technol Soc 11(3):216–223
45. Marchewka J, Liu C, Kostiwa K (2007) An application of the utaut model for understanding
student perceptions using course management software. Commun IIMA 7(2):93–104
46. Karim MA (2003) Technology and improved service delivery: learning points from the
Malaysian experience. Int Rev Adm Sci 69:191–204
47. Themistocleous M, Irani Z, Love PE (2005) Developing e-governance integrated infrastruc-
tures: a case study. In: Proceedings of the 38th annual Hawaii international conference on
system sciences. IEEE, pp 228–228
48. King S, Cotterill S (2007) Transformational government? The role of information technology
in delivering citizen-centric local public services. Local Gov Stud 33:333–354
49. Hermans CM, Haytko DL, Mott-Stenerson B (2009) Student satisfaction in web-enhanced
learning environments. J Instr Pedagogies 1
50. Davis FD, Bagozzi RP, Warshaw PR (1989) User acceptance of computer technology: a
comparison of two theoretical models. Manage Sci 35:982–1003
51. Elsheikh Y, Azzeh M (2014) What facilitates the delivery of citizen-centric egovernment
services in developing countries: model development and validation through structural equation
modeling. Int J Comput Sci Inf Technol 6(1):77
52. Hsia JW, Tseng AH (2008) An enhanced technology acceptance model for e-learning systems
in high-tech companies in Taiwan: analyzed by structural equation modeling. In: International
conference on cyberworlds, pp 39-44
53. Almahamid SM (2013) E-Governance system acceptance and organizational agility: theoretical
framework and research agendas. Int J Inf Bus Manage 5:4
54. Selim HM (2005) Critical success factors for e-learning acceptance: confirmatory factor
models. Comput Educ 49(2):396–413
55. Folorunso O, Ogunseye OS, Sharma SK (2006) An exploratory study of the critical factors
affecting the acceptability of e-learning in nigerian universities. Inf Manage Comput Secur
14(5):496–505
The Significant Relationship Between E-Governance System … 323

56. Jong D, Wang TS (2009) Student acceptance of web based learning system. In: 2009
International symposium on web information system and application (Wisa’09), pp 533–536
57. Tarcan E, Varol ES, Kantarcı K, Fırlar T (2012) A study on kazakh academicians’ information
technology acceptance. Bilig 62:205–230
58. Li, H. And Suomi, R., 2009. A Proposed Scale For Measuring E-Service Quality. International
Journal Of U-And E-Service, Science And Technology, 2(1), Pp.1–10
59. Nunnally JC, Bernstein IH (1994) Psychometric Theory. Mcgraw-Hill, New York
60. Bagozzi, R.P. And Yi, Y. (1988) ‘On The Evaluation Of Structural Equation Model’, Journal
Of Academy Of Marketing Science, Vol. 16, No.1, Pp.74–94
61. Hoyle RH (1995) Structural equation modeling: concepts, issues, and applications. Sage
Publications
62. Chau PYK, Hu PJH (2001) Information technology acceptance by individual professional: a
model comparison approach. Dec Sci 32(4):699–719
63. Byrne BM (2013) Structural equation modeling with lisrel, prelis, and simplis: basic concepts,
applications, and programming: psychology press
64. Hair JF, Black WC, Babin BJ, Anderson RE, Tatham RL (2006) Multivariate data analysis, 6th
edn. Pearson Prentice Hall, United State Of Amreica
65. Parasuraman A, Zeithaml VA, Berry LL (1988) Servqual: a multiple-item scale for measuring
consumer perc. J Retail 64:12
66. Parasuraman A, Zeithaml VA, Berry LL (1985) A conceptual model of service quality and its
implications for future research. J Mark 49:41–50
67. Ho R (2006) Handbook of univariate and multivariate data analysis and interpretation with
Spss. United States of American: Chapman & Hall/Crc, Taylor & Francis Group
68. Schumacker RE, Lomax RG (2010) A beginner’s guide to. Structural equation modeling, 3rd
edn. Taylor & Francis Group, New York
69. Razzaq A, Mohammed AA (2020) Cloud Erp in Malaysia: benefits, challenges, and opportu-
nities. Int J 9.5
70. Mohan NR et al (2014) Using cloud–based web application for universityes: a case of university
of technology in Iraq. Adv Comput 4.1:15–17
71. Saleh AA, Alkhuwaylidee AR, Thangiah M (2021) Successful factors determining the user
behaviour in information system. J Phys Conf Ser 1963(1):012161
Trajectory Tracking Control
of Industrial Robot Manipulator Using
VS Control System with STHWS Method

S. Nandhakumar , B. Shreeram , R. Thirumalai , K. Sakthisudhan ,

and P. Sivaprakash

Abstract In this paper, an experimental investigation has been made to evaluate the
effectiveness of proposed integration of single term haar wavelet series method with
a variable structure control system. The proposed method is applied to non linear
system to obtain solution for trajectory tracking problem by considering dynamics
of the robot as singular system time varying model. The aim is control the variation
or error occurs when the robot is working. In order to minimize the error, integration
of STHWS method with VS control has been designed and implemented in two dof
industrial robot. The experiment has been conducted with various time intervals and
results are analyzed. The obtained results reveals that the proposed integration of
STHWS method with VS control system works well and robust with singular system
time varying model for control of positions.

Keywords Trajectory tracking · Manipulator dynamics · Single term haar wavelet

series (STHWS) · Position control · Singular and non-singular systems · Time
varying and invariant · Variable structure control (VS)

1 Introduction

In today’s competitive world, it is very clear that the product manufactured should
satisfy the needs of the customer. This scenario increases demand of quality in manu-
facturing and it tends to improve the infrastructure to automation and robotics. This
demand increases the considerable research in robot dynamics and control. The
correct robot manipulator dynamic model plays a vital in control mechanisms. The
formulation of the dynamic model involves various internal and external factors. As

S. Nandhakumar (B) · B. Shreeram · R. Thirumalai · P. Sivaprakash

Department of Mechanical Engineering, Dr.N.G.P Institute of Technology, Coimbatore 641048,
India
K. Sakthisudhan
Department of Electrical and Communication Engineering, Dr.N.G.P Institute of Technology,
Coimbatore 641048, India

© The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 325
R. Ibrahim et al. (eds.), International Conference on Artificial Intelligence
for Smart Community, Lecture Notes in Electrical Engineering 758,
https://doi.org/10.1007/978-981-16-2183-3_31
326 S. Nandhakumar et al.

it operates, certain physical parameters like load and speed controls these forces [1–
4]. In order to achieve the proper control the exact dynamic model is very important.
In this proposed research, the dynamic model is formulated based on the Euler–
Lagrange principle [5]. To attain the required response and performance laws and
strategies are to be determined within the dynamic model.
Many literatures are available and the performance of robots has been studied
using differential equations. These differential equations are easy to solve and
provides a simple analytical solutions without non linearity’s [6–8]. For representing
any robot configurations where constraints causes rise in contact forces, differential
equations with singular systems are more suitable ([9] McClamroch 1986). Robots
that are applied for advanced manufacturing jobs, the complexity involved in building
a mathematical model is a direct indication. Hence, it is necessary to study the
behavior of proposed model robot arm and its performance for designing robustness
arm for industrial use. Moreover, it is essential to study various control techniques
to analysis problem and issues pertaining to control robot.
Due to various control issues and limitations of traditional controllers, the VS
control is highly desirable which has numerous interesting and significant properties
such as exact knowledge of the manipulator’s model is not needed and under certain
conditions, the system performance is insensitive to bounded external disturbances.
These properties are unique and cannot be easily obtained by other controllers which
makes VS control a perfect model for manipulator controls [10–14].
The objective of proposed work is to investigate the effectiveness of the proposed
integration of STHWS method with VS control by analyzing trajectories of each
joint of 2 DOF industrial pick and place robot. The analyses involves the tracing the
positions and velocities of each joint for find out the experimental results follows the
desired trajectories which is obtained by using cubic polynomial method with zero
speed at end point. During continuous operations, the trajectories has been traced and
plotted. Based on the task, the control system calculates the torque if a trajectory goes
beyond limits, the mathematical block programmed in the controller calculates the
new torque and controls the variations in position by varying input torque. The time
taken for completing one cycle is 48 s. The results are plotted and compared with the
reported values. From the results it has been observed that STHWS to VS control non
singular system time model and supports in attaining preferred trajectories within
the accuracy range expected.

2 Singular System Time-Varying Case Robot Arm Model

The dynamics of a robot arm can be represented in terms of system of non linear
equations by Nandhakumar et al. [15]
 
T = P(Q) Q̈ + R Q, Q̇ + S(Q) (1)
Trajectory Tracking Control of Industrial Robot Manipulator … 327

where:
P(Q)
 Isthe Coupled inertia matrix,
R Q, Q̇ Is the matrix of coriolis and centrifugal force
S(Q) Is the Gravity matrix,
T Is the Input torques applied at various joints.
For a manipulator with 2 DOF, under the assumption of lumped equivalent masses
and massless- links, the dynamic equations is indicated below

T1 = K11 K12 q1 + K122 (q2 )2 + K112 (q1 q2 ) + K1

T2 = K21 K22 q2 + K211 (q1 )2 + K2 (2)

where
K11 : (W1 + W2 ) d1 2 + W2 d2 2 + 2W2 d1 d2 cos(q2 )
K12 : W2 d2 2 + W2 d1 d2 cos(q2 )
K21 : W12
K22 : W2 d2 2
K112 : −2W2 d1 d2 sin(q2 )
K122 : −W2 d1 d2 sin(q2 )
K211 : W122
K1 : [(W1 + W2 )d1 sin(q1 ) + W2 d2 sin(q1 + q2 )]g
K2 : [W2 d2 sin(q1 + q2 )]g
A linear time-varying singular system discussed by Nandhakumar et al. [15],
which is in the following form

M(t) ẋ(t) = C(t)x(t) + D(t)u(t) (3)

with initial condition x(0) = x0 .

where x(t) is a n-state vector and
u(t) Is the p-input control vector and
M(t) Is a n × n singular matrix,
C(t) and D(t) are n × n and n × p matrices respectively.
The said algorithm has been reported to be used for solving position and velocity
errors in robot arm during operation [16–18].

3 Implementation on an Industrial Robot

The experiment is conducted using 2 dof industrial robot manipulator. The config-
uration and hardware details are discussed by Nandhakumar et al. [5] as shown in
the Fig. 1. In the experiment has carried out in four stages which include two rota-
tions and two linear movements. Initially the arm is in home position i.e. in the top
position. So it has to move to some degree in rotation (2 radians) followed by linear
328 S. Nandhakumar et al.

Fig. 1 Industrial Robot

Manipulator

down movement (0–10 cm) to pick the component. After picking up the object, it has
to move upward from (10–5 cm) and reaching the third position, it rotate and place
the object. During operations, each joint trajectory is traced. Whenever new position
is out of control, the proposed model evaluates the desired torque and compares it
with the existing one. If it goes out of control, the proposed integration estimates the
torque based on the varying positions, and compares with existing torque. Finally,
the exact trajectories have been achieved and results of velocity errors and positions
errors have been plotted and shown in Figs. 2, 3, 4, 5, 6, 7, 8 and 9. From these
figures, it has been observed that the proposed integration controls the errors and
provides the projected range of accuracy.

4 Results and Discussion

Figures 2, 4, 6 and 8 shows the positions and velocity profiles of Joint 1 and 2
respectively arising out of the proposed algorithm. From these profiles, it has been
observed that trajectories very closely follow exact trajectories which are obtained by
cubic polynomial method. The position errors and velocity errors are as shown in the
Figs. 3, 5, 7 and 9. The errors obtained are large than the numerical simulations [15].
The obtained level of accuracy is comparatively good with steady errors. A detailed
comparison of the results with [5] shows that the position and velocity error for joint
1 and joint 2 is less. From Fig. 3, at joint 1—Position error—4.0 × 10–3 rad,Velocity
error—3 × 10–2 and at joint 2—Position error 5.0 × 10–2 rad and Joint velocity
error 4 × 10–2 rad/s which is approximately 20% less than the model reported by
Nandhakumar el al 2013.
Trajectory Tracking Control of Industrial Robot Manipulator … 329

Fig. 2 Joint 1 position and velocity

Fig. 3 Joint 1 position and velocity errors

330 S. Nandhakumar et al.

Fig. 4 Joint 2 position and velocity

Fig. 5 Joint 2 position and velocity errors

Trajectory Tracking Control of Industrial Robot Manipulator … 331

Fig. 6 Joint 2 position and velocity

Fig. 7 Joint 2 position and velocity error

332 S. Nandhakumar et al.

Fig. 8 Joint 1 position and velocity

5 Conclusions

In this paper, an experimental investigation is made to analyze the effectiveness

proposed integration of STHWS with VS Control method. An appropriate law for
controlling the manipulator has been designed for integration and implemented in
two link robot arm. An experiment was carried out by taking into consideration
time varying model as singular system time. The objective was to exhibit the ability
of proposed integration of STHWS method with VS control system for practical
use with diverse arm dynamics. The results are compared with previous studies in
the yields good results and the performance of the manipulator good and fall within
estimated range of accuracy. The healthiness of proposed integration with VS control
system also justified from the results.
Trajectory Tracking Control of Industrial Robot Manipulator … 333

Fig. 9 Joint 1 position and velocity errors

References

1. Sekar S, Prabakaran K (2011) Numerical Solution for linear and nonlinear singular systems
using single term haar wavelet series, Int J Commun Comput Inf Sci 140:204–212
2. Sekar S, Prabakaran K, Paramanathen E (2012) Single-term Haar wavelet series technique for
time varying linear and non-linear singular systems. In: International conference on pattern
recognition, informatics and medical engineering (PRIME-2012), pp 364–369. https://doi.org/
10.1109/ICPRIME.2012.6208373
3. Nandhakumar S, Selladurai V, Sekar S (2009) Numerical investigation of an industrial robot
arm control problem using haar wavelet series. Am J Eng Appl Sci 2(4):584–589
4. Nandhakumar S, Selladurai V (2011) Haar wavelet approach to second order robot arm control
problem. Eur J Sci Res 56(4):573–583
5. Nandhakumar S, Muthukumaran V, Soorya Prakash K, Shunmughanaathan VK (2013) Position
control of industrial robotic manipulator using variable structure control system with single
term haar wavelet series method. J Vib Control 21(12):2465–2483
6. Ata AA, Johar H (2004) Dynamic simulation of task constrained of a rigid-flexible manipulator.
Int J Adv Robot Syst 1:61–67
7. Park BS, Yoo SJ (2018) A low-complexity tracker design for uncertain nonholonomic wheeled
mobile robots with time-varying input delay at nonlinear dynamic level. Nonlinear Dyn
89(3):1705–1717
8. Yamawaki T, Yashima M (2007) Effect of gravity on manipulation performance of a robotic
arm. In: IEEE international conference on robotics and automation, Rome, pp 4407–4413
9. Liu W, Li F, Wan Y et al (2020) Design of manipulator based on two-dimensional force
feedback. Intel Serv Robot. https://doi.org/10.1007/s11370-020-00321-y
10. Castri C, Messina A (2007) Exact modeling for control of flexible manipulators. J Vib Control
18(10):1526–1551
334 S. Nandhakumar et al.

11. Cheng-Yuan C, Hsu K-C, Chiang K-H, Huang G-E (2008) Modified fuzzy variable structure
control method to the crane system with control dead zone problem. J Vib Control 14(7):953–
969
12. Hannan MA, Hussin I, Ker PJ, Hoque MM, Hossain Lipu MS, Hussain A, Blaabjerg F (2018)
Advanced control strategies of VSC Based HVDC transmission system: issues and potential
recommendations. IEEE Access 6:78352–78369
13. Pai MC (2015) Sliding mode control of vibration in uncertain time-delay systems. J Vib Control
16(14):2131–2145
14. Quynh NX, Nan WY, Yen VT (2020) Design of a robust adaptive sliding mode control using
recurrent fuzzy wavelet functional link neural networks for industrial robot manipulator with
dead zone. Intel Serv Robot 13:219–233. https://doi.org/10.1007/s11370-019-00300-y
15. Nandhakumar S, Selladurai V, Muthukumaran V, Sekar S (2010) Haar wavelet approach to
time varing model robot arm control problem. Int J Acad Res 2(4):66–74
16. Bujurke NM, Salimath CS, Shiralashetti (2007) Numerical solution of stiff systems from
nonlinear dynamics using single-term Haar wavelet series. Int J Nonlinear Dyn 51(4):595–605
17. Hsiao CH (2004) Haar wavelet approach to linear stiff systems. Math Comput Simul 64(5):561–
567
18. Hsiao CH (2005) Numerical solution of stiff differential equations via Haar wavelets. Int J
Comput Math 82(9):1117–1123
Medical Fitness Device for Driving
Assessment
K. Chamundeswari, R. Saranya, M. Smitha, S. Prabakar,
and K. Porkumaran

Abstract This project is to check Medical Fitness for accidental Drivers and paral-
ysed patients health condition using pressure and heart beat sensor with RTC unit.
This can be helpful to analyze the accident person’s health improvement condition
or periodic monitoring of that person. In hospitals, accidental patients’ body condi-
tions (Leg and Hand) must be monitored constantly, which is typically done by
doctors or other paramedical staff. And when the conditions go abnormal then we
sense those values by Displaying in LCD and buzzer sound. The suitable parameters
which should be monitored are first determined and also the respective sensors are
connected to the patient‘s body and periodically checked by a PIC microcontroller.
And these parameters have fixed threshold levels which are monitored and also the
parameters are transmitted to a LCD for continues monitoring purpose.

Keywords Accidental drivers · Analyze · Parameters · PIC microcontroller

1 Introduction

An accident is knock or miss event that at times has convenient or more undesir-
able consequences, other times being inconsequential. The term implies that such
an incident might not be preventable since, their leads to circumstances go unrecog-
nized and unaddressed. Accident may lead to injuries which can affect the nervous
system. This may also lead to paralysis, trauma, stroke etc. Paralysis may be a loss
of muscle function partly of our body which may be temporary or permanent. This

K. Chamundeswari (B) · R. Saranya · M. Smitha

Dr.N.G.P Institute of Technology, Coimbatore, Tamil Nadu, India
e-mail: [email protected]
S. Prabakar
Sona College of Technology, Salem, India
e-mail: [email protected]
K. Porkumaran
Sri Sairam Engineering College, Chennai, India
e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 335
R. Ibrahim et al. (eds.), International Conference on Artificial Intelligence
for Smart Community, Lecture Notes in Electrical Engineering 758,
https://doi.org/10.1007/978-981-16-2183-3_32
336 K. Chamundeswari et al.

could flow from to attack within the nerves of the peripheral systema nervosum by
the body’s system which may eventually result in temporary paralysis of arms and
legs [Guillain –Barre syndrome]. The patients with such disabilities don’t seem to
be ready to perform their everyday actions like feeding, toilette usage, movement
through space and driving. After recovery, to assess their driving ability this project
are going to be helpful by measuring the parameters like pressure with which they
apply the break and therefore the time with which they react and heart beat. Since,
till now only the software and scans like MRI, CT, X-RAY are present to assess
the anatomical functions and therefore the overall fitness. Occupational therapists,
as well as certified driver rehabilitation specialist are required to create important
recommendation concerning the fitness to the drive of their clients on daily. To help in
formulating these recommendations they’re going for scanning technologies. Hence
the goal of our project is to assess the fitness without scanning. This project will have
high accuracy than the prevailing ones. This device are going to be helpful for simple
observation of patients exact health condition and improvement. it’s easy to control
and handle. It may be of great use thanks to it’s compactness and price effectiveness.

2 Related Works

A wireless and reliable health monitoring and warning system was proposed by
Khemapech [1] in that the vital signals has been recorded using the wearable device
and further analyzed for providing real time warnings. The heart rate and blood pres-
sure and pulse rate are the measured as the vital signals for depicting the patient’s
cardiovascular symptoms by comparing ranges with the standard values given by the
doctors. One of the key drawback they faced is the, range between wearable device
and the software of the system is the not in the limited communication aspects. Omer
Al-Salihi [2] described about an android mobile application which was successfully
built and consequently the data transmission using internet (WiFi) to both the appli-
cation and to the remote server by was done perfectly,. In addition to the current, the
next objective of this was achieved by examining the vital parameters remotely and
predicts the consequences. Although the results were far considerable there still exist
a problem in both hardware and software interface. Appelboom et al. [3] proposed a
project by implementing smart wearable sensors into the practice patient care could
improve the relationship between doctor and patient which increased the indepen-
dence and involvement of patients regarding to the healthcare aspects and remote
monitoring provides the most effective way of monitoring which brings a drastic
change in the healthcare organization.
Kumara et al. [4] designed a device to assist these patients fingers of the hand
that play a major role in this system. The system ensures the patient to precise his
needs by converting the finger bending into auditory speech. If the identical action
is repeated thrice, then there is an efficiency of the system to send a text message
to the concerned caretaker with the assistance of GSM module. In future more flex
sensors can be accustomated to perform multiple tasks and more health Parameter
may be included. Aminian and Naji [5] proposed a system which implements all
Medical Fitness Device for Driving Assessment 337

sufficient aspects of monitoring to measure the vital parameter for the multiple users
or the patient. In this work a coordinator node is assigned which is attached to the
body of the patient which in turn it collects all the vital information’s by the wireless
sensors and it transmits to the base station for future analysis. The placed sensors
on the body of the patient act as a wireless body sensor network (WBSN) and are
capable of sensing the vital parameters. In case any abnormalities in the measured
parameters the system indicates an alarm to the patient and message also has been
passed to the doctor’s end by an Email or SMS.
While comparing with the pervious findings there is a large amount of reduction in
the energy consumption and also there is a speed in the network lifetime which in turn
improves the communication coverage to extent the comfortness and independence
for ensuring the quality of life, these are considered as a most advantageous of the
proposed system. Deepasri et al. [15] proposed a monitoring system, for indicating
the patients health aspects. In this work the most efficient biosensors are used to
observe or record the physiological parameters such as blood pressure, pulse rate,
heart rate and measured data are continuously sent to the nurse through by using
GSM. Future it been processed in Microcontroller (MSP430). Roshdy et al. [14]
proposed a design for the paralyzed hand people a easy portable device with light
weight and low cost. Based on all these consideration we proposed a device for
monitoring the parameters after recovery of the paralyzed patient in the aspects of
driving.

3 Methodology

The project uses Pressure and heart beat sensor for transferring data to a micro-
controller unit fitted in doctors monitoring area. The pressure sensor value and
reaction time take by patient is displayed through LCD screen. Respective sensors
are connected to the patient‘s body whenever medical test is carried out and peri-
odically values are checked by a PIC Microcontroller. All these parameters have
fixed threshold levels which are monitored. And the parameters are transmitted to a
LCD for continues monitoring purpose. Depending upon the value treatment maybe
carryout further.
The project uses Pressure sensor for transferring data to a microcontroller unit
fitted in doctors monitoring area. The pressure sensor value and latent period take
by patient is displayed through LCD screen. Depending upon the worth treatment
maybe carryout further.

4 System Implementation

The proposed system makes use of two sensors namely flexi force or pressure sensor
and heart beat sensor to measure the appropriate parameters (Fig. 1).
338 K. Chamundeswari et al.

Fig. 1 Block diagram

4.1 PIC Micro Controller

In this system we use the PIC 16F877A microcontroller for the conversion of the
parameters. The pin of the this microcontroller is shown in the Fig. 2.

4.2 Pressure Sensor

FSR are used to continuously monitoring the pressure of the patients especially the
pressure applied to the peddle of accelerator and brake at the time of driving. The
Fig. 3 shows the FSR used in the proposed work.

4.3 Heartbeat Sensor

The heart sensor is placed on the patient’s body to continuously measure the heart
rate of the patient while driving. This is value is send to the physicians at the other
end. This is a digital value.
Medical Fitness Device for Driving Assessment 339

Fig. 2 Pin diagram of 16F877A

Fig. 3 Flexi force sensor

4.4 LCD Display

LCD (Liquid Crystal Display) is used to display the values of the pressure and the
heart rate. Here in this work we use a 16 × 2 LCD display is very basic module and
is very commonly used in various devices and circuits. The LCD display used in our
work is shown in the Fig. 4.
340 K. Chamundeswari et al.

Fig. 4 LCD display

4.5 Power Supply Description

The power supply we use in this work is ac voltage, typically 220V rms, is connected
to a transformer, which steps that ac voltage down to the level of the desired dc output.

4.6 Software Required

• Embedded C.
• PIC Boot loader.
• CCS C Complier.
Medical Fitness Device for Driving Assessment 341

5 Results and Discussion

The working modules of the proposed system are shown in Fig. 5 and Fig. 6 which
are very helpful for the paralytic patients and accidental people.
The system continuously monitors heart beat rate and blood pressure of patient.
The system also has the feature to check the physical fitnesss of the patient for
driving with a high accuracy. Patients exact health condition can be easily observed.
Due to its compactness and low cost this product can be used easily in all healthcare
centers, even in home. Also it is easy to operate and handle.

Fig. 5 Working module of

the proposed work

Fig. 6 Working module of

the proposed work
determining the heart rate
and BP
342 K. Chamundeswari et al.

6 Conclusion

This project presents the design and implementation of accidental health monitoring
system by using flexi force sensor module. The proposed system will be a helping
hand for accidental and paralytic patients. It is concluded that Programmable Inter-
face Controller (PIC) has been the low cost implementation used for recording and
transmitting the bio-medical signals and very useful to the patients. This system was
developed to minimize the device’s size and allow for daily life usage.

References

1. Khemapech I (2015) A real time health monitoring and warning system. TRON Forum
2. Omer RMD, Al-Salihi NK (2017) Healthmate: smart wearable system for health monitoring
(SWSHM). In: 2017 IEEE Students conference on electrical, electronics and computer science
3. Appelboom G, Camacho E, Abraham ME, Bruce SS (2014) Smart wearable body sensors for
patient self—assessment and monitoring. Archives of Public Health
4. Kumara KR, Kadam A, Rane N, Gouda A, Sensor based wearable system to assist paralytic
patient with continuous health monitoring. Int J Future Revolution Comput Sci Commun Eng
5. Aaminian M, Naji HR, A hospital healthcare monitoring system using wireless sensor networks.
J Health Med Inf
6. Nelson A, Shyam Kumar P (2013) Wearable multi sensor gesture recognition for paralysis
patients. In: 2013 IEEE, University of San Francisco
7. VasanthRao Y, Swamy Bhukya (2019) Gait Assistive exoskeleton device for semi paralyzed
stroke survivors. Int J Eng Adv Technol ( IJEAT) 8(5c)
8. Agrawal SK, Sangwan V (2007) Assessment of motion of a swing leg and gait rehabilitation
with a gravity balancing exoskeleton. IEEE Trans Neural Syst Rehabil Eng 15(3)
9. Yorozu A, Moriguchi T (2011) Improved leg tracking with data association considering gait
phase during turning motion in walk tests. In: 2015 IEEE International conference on advanced
intelligent mechatronics [AIM] July 7 2011
10. Milovanovic I (2008) Radial basis function (RBF) networks for improved gait analysis. In: 2008
IEEE 9 th Symposium on neural network applications in electrical engineering, NEUREL
11. Sayegh F, Fadhli F, Karam F, BoAbbas M, Mahmeed F, Korbane JA (2016) A wearable reha-
bilitation device for paralysis. In: IEEE 2016, American University of the Middle East Egaila,
Kuwait
12. Hasegawa Y, Mikami Y, Watanabe K, Firouzimehr Z (2008) Warble handling support system
for paralyzed patients. In: 2008 IEEE/RSJ International conference on intelligent robots and
systems acropolis convention center nice, France, Sept 22–26, 2008
13. Hasan HN (2019) A wearable rehabilitation system to assist partially hand paralysed patients
in repetitive exercises. Biomedical Engineering Department, College of Engineering Quality
Assurance , University of ThiQar, First International Scientific Conference Al-Ayen University
IOP. J Phys Conf Series 1279 (2019) 012040
14. Roshdy A, Al Kork S, Said S (2019) A wearable exoskeleton rehabilitation device for
paralysis—a comprehensive study. Adv Sci Technol Eng Syst J 4(1):17–26
15. Deepasri. T Gokulpriya. M Arun kumar. G (2017) Automated paralysis patient health care
monitoring system. Dep Electron Commun Eng South Asian J Eng Technol 3(2):85–92
Application of Artificial Intelligence
for Reservoir Storage Prediction: A Case
Study

Abdus Samad Azad, Pandian M. Vasant, José A. Gámez Vintaned,

and Junzo Watada

Abstract There are many relevant and interesting contributions using Artificial
Intelligence (AI) based techniques, with different purposes. It has been used as an
effective way for estimating the forecasted data of reservoir daily storage value. The
efficiency of various AI methods was explored in this article and later the best method
is selected for reservoir storage level prediction. In estimating reservoir storage levels
several regression algorithms and artificial neural network (ANN) approaches have
been evaluated. There is better agreement between the ANN model compared to
regression algorithms. The findings were demonstrated by significant correlation
coefficient (R2 ) rate among the expected and calculated training outcome variables
up to 0.91 and the highest validity outcome of Root Mean Square Error (RMSE)
was 5.1989. Consequently, this method is therefore adequate for robustness and
generalizability abilities and is ideal for forecasts.

Keywords Hydropower · Reservoir · Artificial intelligence technique · Reservoir

inflow · Forecast · Artificial neural network · Machine Learning

Nomenclature

ANN Artificial neural network

AI Artificial intelligence
PNN Back propagation neural network
ELM Extreme learning machine

A. S. Azad (B) · P. M. Vasant

Department of Fundamental and Applied Sciences, Universiti Teknologi PETRONAS, 31750
Tronoh, Perak, Malaysia
J. A. Gámez Vintaned
Department of Geosciences, Faculty of Sciences and Information Technology, Universiti
Teknologi PETRONAS, 31750 Tronoh, Perak, Malaysia
J. Watada
IPS Research Center, Waseda University, Tokyo, Japan

© The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 343
R. Ibrahim et al. (eds.), International Conference on Artificial Intelligence
for Smart Community, Lecture Notes in Electrical Engineering 758,
https://doi.org/10.1007/978-981-16-2183-3_33
344 A. S. Azad et al.

LM Levenberg marquardt
ML Machine learning
MSE Mean square error
R Correlation coefficient
RMSE Root mean square error

1 Introduction

A reservoir is a physical (artificial and natural) system used to preserve water for the
storage, management and regulation of the flow of water [1]. Reservoir level control-
ling parameters are inflows, water storage, water release, evaporation, and infiltration
from reservoir. These parameters are unpredictable and must be considered in the
operation of water supplies. The higher results can be achieved if the dam reservoir
level is calculated correctly [2]. Unfortunately, if a reservoir is full on the maximum
storage level due to excessive inflow, the water level rises and the discharge rate over
the drainage line rise. This event will lead to downstream flooding [3]. Therefore,
reservoir water level forecasting is important for flood caution, flood management,
river water potential assessment, hydroelectric power generation, domestic manage-
ment and for river water potential assessment [4]. Also, the significance of water level
forecasting would give the operator the advantage of optimum water level drawing
and a sustainable hydropower generation strategy [5].
There have been several methods used in the forecasting of hydrological activities
in recent years. Previously, the reservoir water level forecasting software used a tradi-
tional linear mathematical relationship approach focused on operator knowledge,
mathematical curves and instructions [6]. Overestimation of features in data and wide
missed variables, however, causes low performance and undermines computational
models because of difficulty and shortage of explosion to the data. With regard to these
problems, the use of Machine Learning (ML) techniques has been adopted, provided
that nonlinear process modelling and forecasting effects are better than conventional
models, such as moving average processes [7–9]. Machine Learning (ML) focuses on
allowing systems through a provided data to forecast, clustered, absorb classification
models, or render choices. The main importance of this modelling is the software’s
ability to visualize scenarios of data without the above-mentioned understanding
from variables influencing the projection variables [10, 11]. Studies have continued to
implement this ML to forecast a range of modeling methods and variables to increase
efficiency and model’s durability prediction description [12]. Regression Algorithms
is considered an alternative approach to ANN for implementing nonlinear predic-
tion. It has been increasingly used in hydrological prediction over the years [13]. In
graphical or mathematical equations, predictive methods are typically provided using
computed historical data and statistical analysis [14]. This research hence emphasizes
on predicting improvements to the amount of storage water through ML algorithms.
Application of Artificial Intelligence for Reservoir … 345

Key objective in this research to come up with reservoir operant by reliable storage
level predicting method and to investigate the simulation methodology of Regres-
sion Algorithms likely linear regression (LR) models, decision trees (DTs), Gaussian
process regression (GPR) models, support vector machines (SVMs), as well as Arti-
ficial Neural Network (ANN) and ensembles of tree models the models were trained.
Secondly, the efficacy of the best method will be used in studying the water level
parameter pattern. Lastly, the selected method will be used for industrial operations
and development and increases the reliability of forecasting for water level storage.
It will help to establish better river operating strategies, leading to better condi-
tions for water users, including agricultural operations, manufacturing practices,
and hydropower production. As a consequence, the exact prediction for the level
of stored water would have a positive effect on improved planning for all relevant
industrial and commercial operations. In a summary, the purpose over this analysis
is to predict volume of stored reservoir water thus it is an essential consideration for
decision-making or the reservoir operator to manage the planning of water supplies.

2 Materials and Methods

2.1 Study Area and Dataset

Mangla Dam is located on the Jhelum River in the Mirpur District in Pakistan.
Coordinates are latitude: 33.142083 °N and longitude: 73.645015 °E with Height
147 m (482 ft) and Length 3140 m (10,302 ft). The total capacity 9.12 km3 (7,390,000
acre ft) with catchment area 33,334 km2 . Turbine numbers are 10 MW × 100 MW,
installed capacity 1150 MW (15% overload) 1500 MW (max. planned). Figure 1
illustrate the Dam location.
Data for this study were 10 years Meteorological Historical data points, collected
for average monthly temperature, rainfall, wind speed, humidity, sun hours and UV
index from January 2009 to October 2020. Hydrological average monthly data of
inflow, irrigation demand and reservoir storage for 2014–2020 was collected. The
average monthly data was used for the study purpose (Fig. 2a, b).

2.2 Methodology

In this study, the dataset was split into training and test dataset, to avoid over-fitting
of the method. Using the dataset of the training section for various statistical models
and Neural Network Regression (NNR).
346 A. S. Azad et al.

Fig. 1 Location of Mangla Dam

Regression Learner App (RLA)

The RLA is a feasible framework available with MATLAB online and latest versions.
RLA models were developed for all the training, using influential hydrological
(inflow, irrigation demand) and meteorological variables (temperature, rainfall, wind
speed, humidity, sun hours, uv index), variables as independent variables, and the
reservoir storage water level of these sites as dependent variables.
Artificial Neural Network (ANN)
Three-layer ANN, consisting of one input layer, one hidden layer, and one output
layer, is widely preferred in practical engineering implementations because it can
approximate virtually any kind of complex relationships to excellent accuracy
between the inputs and target outputs. The sketch map of a standard three-layer
ANN is shown in Fig. 3.
Evaluation Criteria
The performance of the Regression Models and the ANN was evaluated using the
coefficient of determination (R2 ), mean–squared error (MSE) mean absolute error
(MAE) and Root Mean Square Error (RMSE). The value of R2 closer to 1 shows
good correlation between the data sets. Following equations are used to measure the
errors:
Application of Artificial Intelligence for Reservoir … 347

Fig. 2 a Meteorological data. b Hydrological data

Fig. 3 Structure of the ANN

348 A. S. Azad et al.

n  2
i=1 y pr ed,i − yex p,i
R =1−
2
n  2 (1)
i=1 y pr ed,i − ym

1  2
n
MSE = y pr ed,i − yex p,i (2)
n i=1

1  
n
M AE = y pr ed,i − yex p,i  (3)
n i=1

 n  
 y pr ed,i − yex p,i 2
RMSE = (4)
i=1
n

where, respectively, ypred and yexp are the predicted and experimental values. In
the data, N denotes the total number of values. The mean answer value is denoted
by ym.

3 Result and Discussions

3.1 RLA and ANN Modeling

The RLA enables system to instantly train several standard regression techniques
simultaneously time analyze their findings and thus choose the most appropriate one.
The RMSE, quadratic regression (R-square) parameter, mean absolute error (MAE),
mean squared error (MSE), predicted speed, and time of training of Regression
models and Artificial Neural Network are shown in Table 1. The table shows that ANN
had better accuracy in terms of errors evaluation than regression linear algorithms.
Therefore, ANN is selected for further study purpose.
The efficacy of the ANN model at various training protocols is listed in Table 2
here. The most efficient learning strategies for ANN training was back propagation
(BP) algorithm. This algorithm can be quickly applied and incorporated in functional
forecasting systems as given in early research works. It is possible to split back
propagation into two phases: A process of feed-forward and a backward step. In
current study feed forward BP algorithm had been used based on its minimum value
of MSE. The input information was delivered to the input layer, the hidden layer and
the output layer in order to the feed-forward stage, to receive the output information.
ANN was trained and tested further to reach the near-optimal relationship between
the input data set and the output data set.
It must be remembered, though, that the participation of several neurons in the
hidden layer may induce the over-adaptation, resulting in the lack of network gener-
alizability, and that a limited number of hidden neurons may also fall short of the
Application of Artificial Intelligence for Reservoir … 349

Table 1 Performance values of regression algorithms

Prediction model Results
RMSE R2 MSE MAE Prediction Training
Speed (obs/s) Time (s)
Linear Linear 7.34 0.8 53.89 5.71 ~800 1.835
Regression Interaction 169.39 9.72 28,691 65.876 ~1100 0.4768
Models Linear
Robust 7.3077 0.82 53.402 5.6857 ~2100 0.57047
Linear
Stepwise 7.7387 0.80 59.887 6.2356 ~2200 6.8793
linear
Regression Fine Tree 15.216 0.21 231.53 11.273 ~3400 1.137
Trees Medium 14.753 0.26 217.65 11.439 ~6200 0.1182
Tree
Coarse Tree 17.135 -0.00 293.62 14.898 ~6000 0.1274
Ensembles Boosted 18.269 0.14 333.75 16.436 ~1100 1.5164
of Trees Trees
Bagged 12.015 0.51 144.35 10.185 ~1400 1.0856
Trees
Support Linear SVM 8.0991 0.78 65.595 6.6522 ~2700 0.61451
Vector Quadratic 8.6011 0.75 73.98 7.019 ~5500 0.17838
Machines SVM
Cubic SVM 12.47 0.47 155.5 8.74 ~4900 0.13694
Fine 15.918 0.14 253.38 13.677 ~5100 0.17306
Gaussian
SVM
Medium 8.0131 0.78 64.209 6.4718 ~5100 0.13812
Gaussian
SVM
Coarse 11.814 0.52 139.58 9.5742 ~5300 0.13056
Gaussian
SVM
Gaussian Squared 8.5007 0.75 72.262 6.8081 ~3000 1.1217
Process Exponential
Regression Matern 5/2 7.8666 0.79 61.883 6.3482 ~4100 0.25802
Models GPR
Exponential 7.5524 0.81 57.038 6.085 ~4100 0.2701
GPR
Rational 7.7866 0.79 60.632 8.0132 ~4100 0.33501
Quadratic
GPR
ANN 5.1989 0.91 27.0289 9.1293 ~4900 5.6785
350 A. S. Azad et al.

Table 2 Trial-and-error results for development of reservoir efficiency ANN model

Algorithm Function Transfer function in Transfer function in Correlation
hidden layer output layer coefficient
Conjugate gradient traincgp Tansig Purelin 0.7217
backpropagation satlin
with purelin
Polak–Ribiere
updates
Levenberg trainlm Tansig Purelin 0.912
Marquardt satlin
backpropagation purelin
Gradient descent traingdx Tansig Purelin 0.164
with momentum and satlin
cascade-forward poslin
backpropagation
elman trainrp Tansig Purelin 0.427
backpropagation satlin
poslin
Feed-forward trainbfg Tansig Purelin 0.419
backpropagation satlin
poslin
Layer recurrent trainscg Tansig Purelin 0.539
satlin
poslin
Bayesian trainbfg Tansig Purelin 0.812
Regularization satlin
purelin

data and that the network will not then remember. Therefore, if the structure fails to
meet the experimental outcomes, the water treatment plant production value becomes
disorganized and the plant management challenging. The trial and error approach
were used to identify the most effective network model, the ideal number of cached
layers and the maximum possible number of neurons in the cached layer. There
were two maximum number of hidden layers. The addition of more hidden layers
increases the system’s complexity and thus simulator time and expense. Authors have
shown that a framework with a hidden layer will estimate any continuous function
if adequate freedom is given.
The required design generated both training and testing data with the minimum
of errors. Table 3 tests the output of each network model by evaluating the MSE for
each test performed to find the required architecture. Backpropagation network and
14–20 neurons have been shown to achieve minimum MSE levels, so that all models
are limited to 20 neurons and evaluated.
Simulated data was analyzed for the testing and validating of model networks with
only one low, one medium, one in high-value storage datasets. The findings for the
predictive process modelling in Table 3 show the number of neurons in hidden layer.
Several experiments for each group have been performed to achieve the required
Application of Artificial Intelligence for Reservoir … 351

Table 3 Number of neurons in hidden layer for modeling of prediction

Number of neurons MSE train MSE test R2 train R2 test
1 1132.40885e−0 1131.82874e−0 3.86626e−1 3.39020e−1
2 1130.51469e−0 1110.32990e−0 3.22761e−1 4.94977e−1
3 1221.4858e−0 825.41410e−0 2.89077e−1 3.77285e−1
4 719.83942e−0 871.38243e−0 6.81527e−1 6.18165e−1
5 585.61699e−0 714.74720e−0 7.27652e−1 6.59739e−1
6 663.92940e−0 858.15404e−0 7.00667e−1 5.64750e−1
7 608.52763e−0 910.74590e−0 7.17946e−1 6.12950e−1
8 625.97087e−0 845.29416e−0 7.22004e−1 6.03408e−1
9 512.81392e−0 439.05576e−0 7.71174e−1 7.44953e−1
10 568.93684e−0 410.62168e−0 7.36236e−1 8.17789e−1
11 589.04647e−0 617.56938e−0 7.30761e−1 7.43482e−1
12 581.47511e−0 691.45059e−0 7.55153e−1 6.71959e−1
13 447.71431e−0 823.96620e−0 8.05743e−1 5.85416e−1
14 403.09447e−0 651.54030e−0 8.33410e−1 7.23248e−1
15 636.77776e−0 561.76090e−0 7.23871e−1 7.49778e−1
16 635.03722e−0 576.82133e−0 6.97208e−1 7.53079e−1
17 541.03630e−0 976.30986e−0 7.49574e−1 6.13714e−1
18 488.40561e−0 784.90172e−0 7.85160e−1 6.83428e−1
19 409.98544e−0 703.54441e−0 8.29223e−1 6.46347e−1
20 187.03860e−1 740.59603e−1 9.22101e−1 6.57288e−1

network layout, before the appropriate learning rate, the number of hidden layers
and the number of neurons per hidden layer have been achieved.
If a model ANNs can precisely define the highest standards in the Correlation
coefficient (R2) and minimum standards in the Root Mean Square of Error (RMSE)
it can simply be the accurate and superior ANN model. The number is noticeable in
Fig. 4 that the result directions for training (R-value = 0.9221), validation (R value
= 0.82371) and testing (R value = 0.65729) are aimed very strongly. These values
may equate to a total R value of = 0.86872. In this scenario, the network reaction
has so far been optimistic, and simulation can be used to reach new inputs.
The Regression (R) plot highlights the assessment between the target that is
believed to be the optimal output and the ANN output that is known to be the actual
output, which indicates that the ANN output is exactly the same as the target output.
Notably, the Regression (R) value indicates the significance of outputs and priorities.
In fact, since R can exceed 1, this means an accurate linear relevance between goals
and outputs. The validation and the performance checked display values of (R) are
both greater than 0.90. It shows high performance of the model.
352 A. S. Azad et al.

Fig. 4 Regression plot by ANN

3.2 Prediction of Storage Data

The trained model was used for the prediction of storage data for water level of
January–November 2020. Figure 5 shows that the data was well predicted using the
trained ANN model. The prediction accuracy was 92% and RMSE value 5.1989. The
lower value of error shows that ANN is a reliable ML tool for prediction of reservoir
storage data provided with suitable meteorological and hydrological data.

4 Conclusion

The networks used to train and forecast water storage levels using the Matlab software
in this study. The learning algorithm, such as Levenberg–Marquardt, was qualified in
training to train the model. The results were shown by a high value of the correlation
coefficient (R2 ) between the expected and observed training variables of up to 0.91
and the highest validity efficiency value of the RMSE was 5.1989. Consequently, it
Application of Artificial Intelligence for Reservoir … 353

Test data Predicted data

380
375
Water flow (cumec)
370
365
360
355
350
345
Jan Feb Mar Apr May Jun July Aug Sep Oct Nov
Month

Fig. 5 Prediction by ANN

is concluded that for accuracy and generalization capability, this developed model
has been satisfactory and is suitable for prediction purpose.

Acknowledgements The research work has been supported by the Graduate Assistantship (GA),
Fundamental and Applied Sciences Department (FASD), Universiti Teknologi PETRONAS and
SGRG research group.

References

1. Hussain W, Ishak W, Ku-mahamud KR, Norwawi N (2014) Neural network application in

reservoir water level forecasting and release decision, pp 54–63 (2014)
2. Üneş F, Demirci M, Taşar B, Kaya YZ, Varçin H (2019) Estimating dam reservoir level
fluctuations using data-driven techniques. Pol J Environ Stud 28:3451–3462
3. Ashaary NA, Hussain W, Ishak W, Ku-mahamud KR (2015) Forecasting the change of reservoir
water level stage using neural network, pp 103–7 (2015)
4. Dogan E, Isik S, Toluk T, Sandalci M (2007) Daily streamflow forecasting using artificial
neural networks. In: International congress on river basin management, pp 449–59
5. Choi C, Kim J, Han H, Han D, Kim HS (2020) Development of water level prediction models
using machine learning in wetlands: a case study of Upo wetland in South Korea. Water
(Switzerland) 12
6. Valizadeh N, El-Shafie A, Mirzaei M, Galavi H, Mukhlisin M, Jaafar O (2014) Accuracy
enhancement for forecasting water levels of reservoirs and river streams using a multiple-
input-pattern fuzzification approach. Sci World J
7. Grenney WJ (1995) Neural networks for river flow prediction. J Comput Civ Eng 9:293–293
8. Saadon A, Abdullah J, Muhammad NS, Ariffin J (2020) Development of riverbank erosion
rate predictor for natural channels using NARX-QR Factorization model: a case study of Sg.
Bernam, Selangor, Malaysia. Neural Comput Appl 32:14839–49
9. Yang T, Asanjan AA, Welles E, Gao X, Sorooshian S, Liu X (2017) Developing reservoir
monthly inflow forecasts using artificial intelligence and climate phenomenon information.
Water Resour Res 53:2786–2812
354 A. S. Azad et al.

10. Najah A, El-Shafie A, Karim OA, Jaafar O (2011) Integrated versus isolated scenario for
prediction dissolved oxygen at progression of water quality monitoring stations. Hydrol Earth
Syst Sci 15:2693–2708
11. Hipni A, El-shafie A, Najah A, Karim OA, Hussain A, Mukhlisin M (2013) Daily forecasting
of dam water levels: comparing a support vector machine (SVM) model with adaptive neuro
fuzzy inference system (ANFIS). Water Resour Manag 27:3803–3823
12. Zantalis F, Koulouras G, Karabetsos S, Kandris D (2019) A review of machine learning and
IoT in smart transportation. Future Internet 11:1–23
13. Wang Q, Wang S (2020) Machine learning-based water level prediction in Lake erie. Water
(Switzerland) 12:1–14
14. Zhao T, Minsker B, Salas F, Maidment D, Diev V, Spoelstra J et al (2018) Statistical and hybrid
methods implemented in a web application for predicting reservoir inflows during flood events.
J Am Water Resour Assoc 54:69–89
Fabrication of Parallel Mechanism
Actuated by Pneumatic Artificial Muscle
for Rehabilitation Therapy

S. Krishnan, A. M. A. Rani, Laxamanan G. Kurappa,

and Sivajothi Paramasivam

Abstract Ankle–Foot injuries are amongst the most common injuries of the lower
limb and almost 25,000 people experience ankle injuries each day due to vigorous
activities. Traditionally, ankle injuries are rehabilitated via physiotherapy using
simple equipment like elastic bands and rollers, requiring intensive efforts of ther-
apists and patients. Currently, Stewart platform rehabilitation devices are actuated
by various methods including double acting pneumatic cylinder, hydraulic, electric
motor and shape memory alloy. The limitation of using these actuation methods is
that it provides lower range of motions and requires higher maintenances. A less
known type is Pneumatic Artificial Muscles (PAMs). PAMs are operated by pneu-
matic pressure and are contractible naturally upon inflation. The objective of this
study is to fabricate Stewart platform using Pneumatic Artificial Muscle, to analyze
the range of motion produce by various diameter of the PAMs. The diameters of PAM
used were 8, 10 and 12 mm. It was observed that Stewart Platform actuated by PAM
of diameter of 12 mm produced maximum platform angle of 31.73°, whereas PAM
with the diameter of 10 and 8 mm produced maximum platform angle of 28.62° and
25.31°, respectively.

Keywords Pneumatic Artificial Muscle · Stewart Platform · Range of motion

(ROM)

S. Krishnan (B)
Manipal International University (MIU), Nilai, Malaysia
e-mail: [email protected]
A. M. A. Rani
Universiti Teknologi PETRONAS (UTP), Tronoh, Malaysia
L. G. Kurappa
University Nilai, Nilai, Malaysia
S. Paramasivam
UOW Malaysia KDU University College, Glenmarie Campus, Shah Alam, Malaysia

© The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 355
R. Ibrahim et al. (eds.), International Conference on Artificial Intelligence
for Smart Community, Lecture Notes in Electrical Engineering 758,
https://doi.org/10.1007/978-981-16-2183-3_34
356 S. Krishnan et al.

1 Introduction

1.1 Background of Foot and Ankle

Ankle joint is one of the significant joints in human body which helps maintain body
balance during ambulation. In a human skeleton, human ankle joint is a very complex
bony structure. Ankle sprain is a common injury and the number of injuries is rising
day by day due to vigorous activities carried out by people and this can happen
to athletes and non-athletes, children and adults. Beside vigorous activities, it can
also happen when you simply step on an uneven surface, or step down at an angle.
Ligament play important role in human ankle where it holds the ankle bones and joint
in position. For all the abnormal movements-especially twisting, turning, and rolling
of the foot, the ligament provides protection to the ankle joint. Since ligament is an
elastic structure it has the ability go back to their normal positions after it stretches
within its limits. Sprain occurs when ligament is forced to stretch beyond its limit.
Actual tearing of the elastic fibers caused by the severe sprain. Apart from ankle
sprain, stroke is one of the most typical impairment which causes the patient difficult
to lift their foot due to weakness in the dorsiflexion muscles. Figure 1 shows anatomy
of human ankle.
Physical therapies method which rehabilitates patient one-to-one is known as
traditional rehabilitation. Obtaining desired results from this traditional rehabilita-
tion method solely depends on the therapist skills. This traditional method affects the
therapists in a negative way since the rehabilitation process contains repetitive and
intensive exercises and it takes longer period of time to cure. Hence, there will be
difficulties in delivering the same rehabilitation process by the therapist. Therefore,
from the viewpoint of patients, therapists and rehabilitation process by studies, the

Fig. 1 a Rotation and b joints of the ankle [1]

Fabrication of Parallel Mechanism Actuated … 357

Fig. 2 Movement of human ankle [2]

importance and requirements of robotic-based rehabilitation has been emphasized.

One of the advantages of using ankle rehabilitation robots is that it provides a precise
measurement of data to the therapists to records and observes every details of the reha-
bilitation process and makes changes of the diagnosis, prognosis, and customization
of conceivable treatment. Patient’s recovery can be accelerated by delivering more
accurate, repetitive and intensive therapies through this rehabilitation process using
ankle rehabilitation robots. Flexibility of the ankle joints can be increased by using
these rehabilitation robots and allow patient’s ankle undergo all four movement:
dorsiflexion, plantar-flexion, inversion and eversion. Figure 2 depicts movement of
human ankle.

2 Literature Review

2.1 Pneumatic Artificial Muscles

The braided pneumatics actuators were invented by McKibben to help the movement
of polio patients 1950. They are also called a McKibben Muscle or Pneumatic Muscle
which is actuated by compressed air. Pneumatic muscle was used as the original
applications such as, door opening arrangement and on industrial robot [3]. The
displacement movement performed by pneumatic muscle is shown in Fig. 3.
The main element of PAM is thin membrane which makes it lightweight and it
is an advantage during replacement of a defective muscle. There are different types
of braided sleeves discovered in the past literature such as latex and silk pneumatic
Muscle [5–7]. Another advantage of PAM in their inherent complaint behavior is
when force is exerted on the PAM, which gives in without increasing the force in the
actuation.
358 S. Krishnan et al.

Fig. 3 The displacement

movement carried out by
pneumatic muscle [4]

2.2 Application in Robotics

Wickramatunge and Leephakreeda [8] reported that the low assemble weight and
high power to ratio of PAMs is most desirable to be considered for use in mobile
robotics. In certain situation, the impotency of the PAMs is higher than the conven-
tional pneumatic cylinder. The development of a model gives effective description
to understand the mechanical behavior of the PAMs. The stiffness parameter of the
muscle denotes as K and it is considered a function of the actuated air pressure, P
and total length, LS . The elastic force is denoted by Felastic and the expression given
in Eq. 1.

Felastic = K (P, L s )L s (1)

In this study, the stiffness parameter K is taken as a second order polynominal

function of LS and P and it is expressed in Eq. 2. However, a universal approximation
method with fuzzy logic or neural network can be used as alternative choice for
accuracy viewpoint [9].

K = ao P 2 + a1 P L s + a2 L 2s + a3 (2)

where ao , a1 , a2 are constant cerfficient, which can obtained from experiment data.
Typical values of coefficient in K for operating on construction mode under low
pressure:

a0 = −2.977, a1 = −5.890, a2 = 3.203 and a3 = 104.824

Fabrication of Parallel Mechanism Actuated … 359

Dorsi flexion movement

Upper platform

Double acting
pneumatics cylinder
Lower platform

Fig. 4 The “Rutgers Angle”: the system view [10]

2.3 Parallel Mechanism

Recently there has been great interest in some research center around the world to
propose and develop automated system for ankle rehabilitation. Much of the work
focuses on the limitation of existing commercial rehabilitation devices, i.e. obtaining
combinations of basic ankle movement. Figure 4 shows the Rutgers ankle platform.
Rutgers ankle platform is a Stewart platform that generates displacement to the
ankle–foot of the patients during rehabilitation process by actuating six pneumatic
cylinders. The Stewart platform was designed to control the displacement of the
cylinder in six degrees of freedom movement of the foot–ankle.
The electro pneumatic controller is used to actuate the Stewart platform. The
electro pneumatic controller modulates the air pressure in the platform. The Rutgers
actuation is based on pneumatic system, and this has disadvantages in terms of
portability due to the heaviness of the device and sounds that is unsuitable for medical
rehabilitation environment.

3 Methodology

The schematic of Stewart platform is shown in Fig. 5 consisting of upper platform,

lower platform and Delrin panel. Basically, one end of the Pneumatic Artificial
Muscle is connected to upper platform and the other end is connected to pneumatic
fitting. When the Pneumatic Artificial Muscle is pressurized, the muscle will expand
and shorten leading to the displacement of the upper platform. The upper platform will
be deformed to the downward position which creates dorsi/plantar flexion movement.
The gyroscope is fixed in between the upper platform to measure the top angle and
force sensitive resistor (FSR-402) is fixed on top of the upper platform to measure
the force applied on the platform. Figure 5 shows the position of upper platform
before and after it is pressurized in 3D modeling.
360 S. Krishnan et al.

(a) 0° (b) 31.73°

Fig. 5 Stewart platform actuated by PAM a 0° and b 31.73°

The Pneumatic Artificial Muscle used in this experiment is enclosed with nylon
braided sleeves. The PAM is pressurized from 1 bar up to 5 bars and diameter of PAM
used is 12 mm. The gyroscope is fixed in between the upper platform to measure the
top angle and force sensitive resistor (FSR-402) is fixed on top of the upper platform
to measure the force applied on the platform. The developed Stewart platform is
depicted in Fig. 6.

Force Sensitive
Upper platform
Gyroscope

PAM

Delrin panel
Lower platform

Fig. 6 Developed Stewart platform actuated by PAMs

Fabrication of Parallel Mechanism Actuated … 361

Fig. 7 Force produced by 12mm 10mm

4
various diameters of PAM
3

Force (N)
2

0
1 2 3 4 5
Pressure Supplied (Bar)

4 Results and Discussion

4.1 Force Produced by PAMs

Preliminary experiment was conducted with different diameters of 8, 10 and 12 mm

to identify the force produced by the pneumatic artificial Muscle. Figure 7 shows the
comparison force produced by different diameters of Pneumatic Artificial Muscle.
It is observed that the PAM with the diameter of 12 mm produced maximum pulling
force of 3.5 N when PAM is pressurized maximum to 5 bars. The PAM with the
diameter of 10 mm has a pulling force of 3.0 N and 8 mm have a pulling force of
2.5 N when PAM is pressurized to 5 bars. According to PAM with the diameter of
12 mm the difference between maximum force and minimum force is 2.5 N followed
by PAM with the diameter of 10 mm and 8 mm which is 2 N and 1.5 N respectively.

4.2 Range of Motion Produced by the PAMs

Figure 8 presents the comparative study of Range of Motion produced by Pneumatic

Artificial Muscle with different diameters of 8, 10 and 12 mm. It was observed that
Pneumatic Artificial Muscle with the diameter of 8 mm produced minimum range

Fig. 8 Range of motion

Range of MoƟon (Degree)

40
produced by various 12mm 10mm
diameter PAMs 30

20

10

0
1 1.5 2 2.5 3 3.5 4 4.5 5
Pressure Supplied (Bar)
362 S. Krishnan et al.

of motion of 2.21° and the maximum range of motion of 25.31°. The difference
between minimum and maximum range of motion of the platform is 23.1°. PAM with
the diameter of 10 mm obtained minimum range of motion of 2.43° and maximum
range of motion of 25.31°. The difference between maximum and minimum range
of motion is 22.88°. There is a slight increase in range of motion when the diameter
increases. The maximum and minimum range of motion obtained for 12 mm PAM
is 31.73° and it was 24.8° ROM actuated by SMA [11, 12].

5 Conclusion

In this study, it is established that the use of Stewart platform actuated by Pneumatic
Artificial Muscle improves the range of motion (ROM) for patient’s ankle–foot reha-
bilitation in biomedical applications. As the PAM is pressurized, the muscle gets
expand and shorten in length pulling the upper platform downwards. Hence, dorsi-
flexion and plantar flexion movement is achieved. Pneumatic Artificial Muscle with
the diameter of 12 mm was fabricated in the Stewart platform. The Stewart platform
actuated by Ø 12 mm diameter of PAM produced maximum platform angle and force
of 31.73° and 3.5 N, respectively. The developed Stewart platform actuated by Pneu-
matic Artificial Muscle produced range of motion up to 31.73° which is acceptable
to promote dorsi/plantar flexion.

References

1. Krishnan S, Nagarajan T, Rani A, Ambaraj W, Ramiah R (2016) Rehabilitation for foot/ankle-

continuous passive motion (CPM) using shape memory alloy (SMA) actuated stewart platform.
ARPN J Eng Appl Sci 11:22
2. Brockett CL, Chapman GJ (2016) Biomechanics of the ankle. Orthopaedics Trauma 30:232–
238
3. Krishna S, Nagarajan T, Rani A (2011) Review of current development of pneumatic artificial
muscle. J Appl Sci 11:1749–1755
4. Deaconescu T (2007) Studies regarding the performance of pneumatic Muscles. In: Proceedings
of international conference on economic engineering and manufacturing systems, pp 271–275
5. Krishnan S, Nagarajan T, Rani AMA, Rao T (2012) Silk pneumatic artificial muscle (SPAM)
construction for bio-medical engineering application. In: 2012 IEEE business engineering and
industrial applications colloquium (BEIAC), pp 302–306
6. Nagarajan T, Krishnan S, Amirtham V, Abdul-Raniand AM, Rao T (2013) Experimental
investigation-natural fiber braided sleeve for pneumatic artificial muscles actuation. Asian J
Sci Res 6:596
7. Laksanacharoen S (2004) Artificial muscle construction using natural rubber latex in Thailand.
In: The 3rd Thailand and material science and technology conference, pp 10–11
8. Wickramatunge K, Leephakpreeda T (2009) Empirical modeling of pneumatic artificial muscle.
In: International multi conference of engineers and computer scientists, pp 1726–1730
9. Vanderborght B, Van Damme M, Van Ham R, Verrelst B, Lefeber D (2006) Pleaded pneumatic
artificial muscles for robotic application. In: Proceedings of the IEEE international conference
on robotics and automation, pp 4324–4328
Fabrication of Parallel Mechanism Actuated … 363

10. Krishnan S, Krishna A, Nordin A, Amirtham C, Rani A, Rao T (2020) Design and fabrication
of a parallel mechanism for foot/ankle rehabilitation therapy. In: Advancement in emerging
technologies and engineering applications. Springer, pp 133–141
11. Nagarajan T, Krishnan S, Rani A, Amirtham V, Rao T, Ambaraj W et al. (2015) A novel
approach of using SMA wires actuated stewart platform for continuous passive motion (CPM)
of ankle rehabilitation therapy. In: Advanced materials research, pp 494–498
12. Abdul-Rani AM, Krishnan S, Nagarajan T, Rao T, Ramiah R, Ambaraj W (2017) Early mobi-
lization using heat treated shape memory alloy (SMA) for rehabilitation phases. In: Key
engineering materials, pp 112–116
A Review of Artificial Intelligence-Based
Techniques to Estimate Atmospheric
Parameters Influencing the Performance
of Concentrating Photovoltaic/Thermal
Systems

F. Masood , P. Nallagownden , I. Elamvazuthi , J. Akhter ,

and M. A. Alam

Abstract Concentrating photovoltaic/thermal (CPV/T) technology is regarded as

the most auspicious part of renewable energy capable of reducing reliance on fossil
fuels due to its superior performance and hybrid output nature. CPV/T technology
aims to reduce the cost of the renewable systems by replacing the costly solar cell
material with relatively cheap optical devices that concentrate the light collected
from the sun to a small solar PV cell and simultaneously generating useful heat
energy for process heat applications. However, the electrical and thermal perfor-
mances of systems utilizing the methodology mentioned above get strongly affected
by atmospheric parameters like solar radiation, ambient temperature, and the solar
spectrum. In recent years, due to the advantages tendered by Artificial Intelligence
tools to solve ambiguous and non-linear problems, many authors have used intelli-
gent system-based techniques for the prediction of the above-mentioned atmospheric
parameters. This paper presents a review of artificial intelligence-based techniques,
including Artificial Neural Network, Genetic Algorithm, and their composite models
for the estimation of atmospheric parameters that significantly influence the working
of hybrid concentrating PV/thermal systems. The review demonstrates the feasi-
bility and accuracy of artificial intelligence-based tools for precise solar insolation
and ambient air temperature prediction.

F. Masood (B) · P. Nallagownden · I. Elamvazuthi

Department of Electrical and Electronics Engineering, Universiti Teknologi PETRONAS, 32610
Bandar Seri Iskandar, Malaysia
M. A. Alam
Department of Mechanical Engineering, Universiti Teknologi PETRONAS, 32610 Bandar Seri
Iskandar, Malaysia
J. Akhter
Department of Mechanical Engineering, University of Engineering and Technology Taxila,
Rawalpindi 47080, Pakistan
F. Masood
Department of Electrical Engineering, University of Engineering and Technology Taxila,
Rawalpindi 47080, Pakistan

© The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 365
R. Ibrahim et al. (eds.), International Conference on Artificial Intelligence
for Smart Community, Lecture Notes in Electrical Engineering 758,
https://doi.org/10.1007/978-981-16-2183-3_35
366 F. Masood et al.

Keywords Concentrating photovoltaic/thermal system · Artificial neural network ·

Genetic algorithm · Solar irradiance · Ambient temperature

1 Introduction

Owing to the rapid population growth and the fantastic advancement of the indus-
trial sector, energy consumption is rising globally at a much faster pace. Renewable
energy resources have emerged as a promising alternative to conventional energy
generation systems due to their abundance and free availability. They have gained
significant importance in overcoming the energy crisis [1]. Among all renewable
energy resources, solar energy has particularly drawn enormous consideration in
various engineering and industrial applications. Solar energy can be harnessed either
by employing solar photovoltaic panels, which directly convert sunlight into elec-
tricity, or by utilizing solar thermal collectors, which extract heat energy from the
solar radiation and transfer it to a working fluid, usually air or water, for further
use in thermal applications [2, 3]. The combination of photovoltaic panels, solar
thermal collectors, and reflective or refractive optical concentrators, also collectively
called concentrating photovoltaic thermal collectors, has attracted the attention of
researchers and developers in the field of solar energy due to their hybrid nature and
tremendous performance as compared to their constituent elements operating alone
[4]. The optics attached to the PV module also called the CPV system, is used to
concentrate sunlight on PV cells to generate the same amount of power with the rela-
tively lesser solar cell material, thus replacing costly solar cells with cheap optical
concentrators. In contrast, the purpose of solar thermal collector appended to the
back of PV cells, called photovoltaic/thermal system, is to remove heat from solar
cells for enhancing their efficiency on the one hand and producing thermal energy on
the other hand [5]. A CPV/T collector generates an output of hybrid nature consisting
of both electrical and thermal forms.
The CPV/T systems are usually classified on the basis of the concentration ratio of
the incident sun rays over the receiver [6]. The concentration ratio demonstrates the
extent to which sunlight is concentrated on the target. The concentration is sometimes
expressed in the form of ‘suns,’ which indicates the level up to which the solar power
is aggregated. Based on concentration ratio (CR), the CPV/T systems are grouped
into low (1 < CR < 10), medium (10 < CR < 100) and high concentration (CR > 100)
photovoltaic/thermal systems [7].
In a CPV/T system, the prediction of electrical and thermal attributes is mandatory
for the detailed assessment of operational constraints and the economic viability of
these systems to promote their market proliferation [8]. However, the modeling of
CPV/T systems is entirely dissimilar and more complicated than CPV and PVT
systems operating alone in real operating conditions due to diverse factors like their
dependence on spectral distribution, difficulty in measuring cell temperatures, and
solar flux distribution on the receiver surface [9].
A Review of Artificial Intelligence-Based Techniques … 367

Artificial intelligence-based techniques are incredibly beneficial in resolving

multifaceted problems and modeling complex systems [10]. For example, Artifi-
cial neural networks (ANN) have been extensively used in the field of photovoltaics
to estimate and forecast solar radiation and for modeling and simulation of both
standalone and grid-connected photovoltaic systems [10, 11]. Due to the superior
performance shown by ANNs for modeling of non-linear systems together with the
complicacy associated with the optical, electrical, and thermal modeling of CPV/T
systems, many researchers have applied ANNs for modeling, simulation, characteri-
zation, and performance estimation of both low concentration and high concentration
photovoltaic/thermal systems [10, 12]. The utilization of AI tools offers an alter-
nate solution to the problems that are still complicated from a fundamental physical
perspective due to the complexity of multi-physics phenomena associated with the
operation of these systems [13]. Such methods no longer require particularized infor-
mation regarding the materials needed for the fabrication of a given system. Besides,
advanced knowledge of optics, semiconductor physics, and thermal sciences is no
longer required as compared with most of the modern modeling techniques available
in the literature [14]. However, ANNs also exhibit some disadvantages as compared to
conventional analytical methods. For example, the inter-relationship between desired
input and output variables is sometimes unknown, and the historical data required
for modeling purposes must have high accuracy [15].
In this paper, a review of AI-based techniques to address the diverse issues related
to the estimation of atmospheric parameters affecting CPV/T systems is presented.
The primary aspiration is to prepare a summary of the published research papers
about the applications of AI tools like Artificial Neural Network (ANN), Genetic
Algorithms (GA), and their derivatives for the estimation of solar radiation as well
as ambient air temperature.

2 Application of ANNs for the Estimation of Solar

Radiation

The HCPV/T systems usually employ refractive optical elements to concentrate the
sunlight over solar cells. Such systems respond only to direct normal irradiance
abbreviated as DNI. Hence, this segment of sunlight is compulsory for the perfor-
mance prediction and characterization of HCPV/T structures. On the other hand, the
performance of LCPV/T systems depends on both direct and diffuse parts of solar
radiation due to their wide acceptance angles. The prediction of direct and diffuse
components of solar radiation is difficult when using analytical methods. Therefore,
many authors have employed ANNs for their prediction.
Lopez et al. [16] proposed an artificial neural network based on Bayesian function
for modeling of direct normal irradiance. The authors demonstrated that clearance
index and air mass are major parameters pertinent to the operation of the system.
368 F. Masood et al.

Alam et al. [17] used an ANN model to estimate beam solar radiation using a param-
eter called reference clearance index. The root-mean-square error for the model
fluctuates in the range of 1.66–2.78% for different states in India. In reference [18],
the authors established two ANN models for the prediction of DNI indirectly from
the clearance index. A close agreement between predicted and measured values was
observed for the Indian region. An adaptive alpha model for forecasting the hourly
global, diffuse, and direct irradiances was developed by Millet et al. [19]. However,
the results were found to be inferior compared to a feed-forward neural network.
Premalatha and Naveen [20] developed six models based on ANN using the training
algorithm, TRAINLM, and a transfer function, log-sigmoidal, to predict global solar
radiation. The authors used 32 combinations of different input parameters for predic-
tion purposes. Renno et al. [21] developed a multi-layer perceptron model for the
prediction of global radiation and DNI. Two ANN models were developed using
different parameters like climatic, astronomic, and radiometric variables. For hourly
DNI estimation, the root-mean-square error value was found to be 3.15%. Hameed
et al. [22] proposed a low-cost methodology for predicting solar insolation, a so-
called substitute to a pyranometer, based on ANN. The authors used the open-circuit
voltage and short circuit current of a solar PV module and then used these values
as input parameters to ANN to estimate the irradiance. The authors concluded that
the proposed method could produce solar irradiance values with fair approxima-
tion. Table 1 summarizes the AI base techniques used for the prediction of both
components of insolation, i.e., direct, and diffuse.
From the literature survey, it was observed that most authors employed a partic-
ular type of ANNs called feed-forward back-propagation artificial neural networks
to estimate the constituents of solar radiation like direct and diffuse components
and global irradiance. Most techniques were found to be dependent on atmospheric
parameters like temperature and humidity. Astronomic and radiometric parameters
like longitude, latitude, and zenith angle were also used as input variables. The value
of root-mean-square error for the selected papers ranges from 0.20 to 24.5%. Few
authors used a combination of ANN and GA for the prediction of solar insolation.
Rodriguez et al. [23] used the ANN model and GA to predict global solar radiation
for Spain. The predicted values had an accuracy of 99% and a root-mean-square error
of 6.74%. In another study, Rodriguez et al. [24] used a combination of ANN and
GA models to predict DNI. The improvement in estimation performance was found
to be 35%, whereas the root-mean-square error was found to be 24.29%.

3 Application of ANNs for the Estimation of the Ambient

Air Temperature

Another atmospheric parameter that strongly affects the performance of CPV/T

systems is the ambient temperature. An ANN model was used to predict, twenty-four
hours in advance, the average hourly value of ambient temperature for Saudi Arabia
A Review of Artificial Intelligence-Based Techniques … 369

Table 1 Summary of AI models for prediction of solar radiation

S. No Authors Type of AI tool Inputs Outputs Error
1 Lopez et al Bayesian clearance index, DNI –
Neural air mass
Network
2 Alam et al Feed-Forward Latitude, Ref RMSE = 2.7%,
Neural longitude, clearance MBE = 0.07%
Network altitude, month index
number
3 Mishra et al. Feed-Forward Latitude, Clearance RMSE = 5.4%
Neural longitude, index RMSE = 12%
Network, altitude, month
Radial Basis number, mean
Function sunshine,
Network rainfall, relative
humidity
4 Millet et al. Feed-Forward Temperature, Direct and R2 = 98.34%
Neural humidity, diffuse R2 = 98.2%
Network sunshine, radiation
advance hour
insolation
5 Premalatha and Feed-Forward Sunshine hours, Global RMSE = 3.97%
Naveen Neural extraterrestrial solar
Network radiation, daily radiation
max, and min
temperatures
6 Renno et al. Feed-Forward clearance index, DNI RMSE = 3.15%
Neural declination R2 = 0.992%
Network angle, hour angle
7 Hameed et al. Feed-Forward Open circuit DNI RMSE = 2.5%
Neural voltage, short
Network circuit current,
Irradiance

by Tasadduq et al. [25]. The percent difference between forecasted and measured
temperatures was 3.17, 4.18, and 2.84 for three years. For hourly temperature predic-
tion, Abdel-Al [26] developed an alternative machine learning model. The authors
used measured hourly temperature values for a period of seven years as input to the
model. The model was used to estimate hourly temperature for the following day
and the next hour temperature having a mean absolute error of 2.01 °F and 1.05 °F
respectively. In another study, Smith et al. [27] developed an ANN model for the
prediction of ambient air temperature for the entire year by using the actual weather
data set measured by an organization, namely the Georgia Automated Environmental
Monitoring Network. The present magnitudes of ambient air temperature together
with 24 h of advance values, solar insolation, relative humidity, wind speed, and rain-
fall were used as model inputs. The model was able to predict the values showing a
mean absolute error of only 0.516 °C for a one-hour horizon. To predict the average
370 F. Masood et al.

ambient air temperature in Denizli, Turkey, Dombayci, and Golcu [28] established an
ANN model. The authors used three years of temperature data noted by the Turkish
state metrological service to train the network. The average previous day tempera-
ture, the day, and the month were used as input to the model. The root-mean-square
value was found to be 1.96 for the testing data set. To estimate temporal and spatial
air temperature, Deligiorgi et al. [29] developed various ANN models. For one hour,
two hours, and three hours advance temporal air temperature prediction, the authors
trained different ANN models using Levenberg-Marquadrt back propagation algo-
rithm. The proposed ANN model performance was found to decrease with rising
forecasting lag. The mean absolute error was less than 1.4 °C in all cases. The radial
basis function network and the multi-layer perceptron non-linear feed forward neural
networks were compared for spatial forecasting. It was concluded that both models
produced accurate results for the air temperature.
An ANN model was developed by Almonacid et al. [30] for the prediction of
air temperature on an hourly basis. The daily maximum, minimum, and average air
temperature, latitude, and longitude of a site in Spain were used as inputs to the model.
The results were compared with those produced by conventional techniques. The
root-mean-square error value lies from 0.53 to 1.98 °C. Cobaner et al. [31] suggested
applying ANN together with Multiple linear regression models and adaptive neuro-
fuzzy inference system (ANFIS) for the prediction of the largest, smallest, and mean
temperatures for a whole month, taking into account the geographical coordinates and
the month number for a given site in Turkey. The input variables include longitude,
latitude and altitude, and the month number. When compared, the results obtained
by the ANFIS model were found to be the best, with a mean absolute error of
1.25% for the mean temperature. The value of R2 was found to be 0.966 for the
average temperature. The comparison of ANN models with ANFIS for predicting
air temperatures on a monthly basis was conducted by Kisi and Shiri [32]. For training
the network, data obtained from twenty different metrological stations located in Iran
were utilized. The longitude, latitude, altitude of the site, and the time span were used
as input parameters. The performance of the ANN model was found to be superior
to the ANFIS model in the test duration.
Venkadesh et al. [33] used a GA-based approach to adjudge the duration and
resolution of prior input data for an ANN model to predict air temperature. The
GA classified contributive roles of different input variables for the estimation of air
temperature for one to twelve hours prediction horizons. The authors concluded that
ANN models established based on the GA approach produced more accurate results
as compared to the usual ANN. The mean absolute error showed an improvement
with the proposed technique.

4 Conclusions

The atmospheric parameters like solar irradiance, ambient air temperature, and solar
spectrum strongly influence the performance of CPV/T systems. A comprehensive
A Review of Artificial Intelligence-Based Techniques … 371

review of artificial intelligence-based tools for predicting these atmospheric param-

eters was presented in this paper. From the review of published research articles,
it was found that most authors developed ANN-based models for the prediction of
solar radiation and ambient temperature. A few authors employed a combination
of ANN and GA for a more accurate prediction of irradiance and temperature. In
contrast, characterization of the solar spectrum for CPV/T systems using AI-based
models is still unexplored. Moreover, AI-based models were found to be robust
and accurate tools for the prediction of atmospheric parameters. Due to their wide
acceptance angles, low concentration systems possess the ability to collect direct
and diffuse components of solar radiation. In contrast, most research studies focused
on the prediction of direct solar irradiance only, so more accurate hybrid models
are required to be developed for the prediction of diffuse radiation and the solar
spectrum, particularly for LCPV/T systems.

Acknowledgements The authors highly acknowledge the support and resources provided by
Universiti Teknologi PETRONAS, for conducting this research under YUTP scheme (Cost Centre:
015LCO-024).

References

1. Baig H, Sarmah N, Heasman KC, Mallick TK (2013) Numerical modelling and experimental
validation of a low concentrating photovoltaic system. Sol Energ Mater Sol Cells 201–219
2. Chandan DS, Kumar PS, Reddy KS, Pesala B (2020) Optical and electrical performance
investigation of truncated 3X non-imaging LCPV/T systems. En Con Mgt 113056
3. Zhang G, Wei J, Wang Z, Xie H, Xi Y, Khalid M (2019) Investigation into effects of non-
uniform irradiance and photovoltaic temperature on performances of photovoltaic/thermal
systems coupled with truncated compound parabolic concentrators. Appl Enrg 245–56
4. Al-Nimr MA, Mugdadi B (2020) A hybrid absorption/thermo-electric cooling system driven
by a concentrated photovoltaic/thermal unit. Sus En Tech Assmnts 100769
5. Kim SM, Kim JH, Kim JT (2019) Experimental study on the thermal and electrical
characteristics of an air-based photovoltaic thermal collector. Energies 2661
6. Sharaf OZ, Orhan MF (2015) CPVT solar collector systems: part I—fundamentals, design
considerations and current technologies. Ren Sus En Rev 1500–65
7. Sharaf OZ, Orhan MF (2015) CPVT solar collector systems: part II—implemented systems,
performance assessment, and future directions. Ren Sus En Rev 1566–1633
8. Diwania S, Agrawal S, Siddiqui AS, Singh S (2019) Photovoltaic–thermal (PV/T) technology:
a comprehensive review on applications and its advancement. Int J En Env Eng 33–54
9. Almonacid F, Fernandez EF, Mellit A, Kalogirou S (2017) Review of techniques based on
ANNs for the electrical characterization of cpv technology. Ren Sus En Rev 938–953
10. Shahsavar A, Moayedi H, Al-Waeli A, Sopian K, Chelvanathan P (2020) Machine learning
predictive models for optimal design of BIPV/T collectors. Int J En Rsh 5675–95
11. Qazi A, Fayaz H, Wadi A, Raj RG, Rahim NA, Khan WA (2015) The artificial neural network
for solar radiation prediction and designing solar systems: a systematic literature review. J Cnr
Prdn 1–12
12. Motahar S, Bagheri-Esfeh H (2020) Artificial neural network based assessment of grid-
connected pv/t systems in heating dominated regions of Iran. Sus En Tch Assesments
100694
372 F. Masood et al.

13. Bermejo JF, Fernandez JF, Polo FO, Marquez AC (2019) A review of the use of artificial neural
network models for energy and reliability prediction. A study of the solar PV, hydraulic and
wind energy sources. Appl Sci 1844
14. Almaktar M, Rahman HA, Hassan MY, Saeh I (2015) Artificial neural network-based photo-
voltaic module temperature estimation for tropical climate of Malaysia and its impact on
photovoltaic system energy yield. Prog Phvltaics Res Appl 302–18
15. Almonacid F, Fernandez EF, Rodrigo P, Perez-Higueras PJ, Rus-Casas C (2013) Estimating
the maximum power of a HCPV module using an ANN. Energies 165–172
16. López G, Batlles FJ, Tovar-Pescador J (2005) Selection of input parameters to model direct
solar irradiance by using artificial neural networks. Energy 1675–1684
17. Alam S, Kaushik S, Garg SN (2006) Computation of beam solar radiation at normal incidence
using artificial neural network. Ren Enrgy 1483–1491
18. Mishra A, Kaushika ND, Zhang G, Zhou J (2008) Artificial neural network model for the
estimation of direct solar radiation in the Indian zone. Int J Sus En 95–103
19. Mellit A, Eleuch H, Benghanem M, Elaoun C, Pavan AM (2010) An adaptive model for
predicting of global, direct and diffuse hourly solar irradiance. En Cnvrsion Mgt 771–782
20. Premalatha M, Naveen C (2018) Analysis of different combinations of meteorological param-
eters in predicting the horizontal global solar radiation with ANN approach: a case study. Ren
Sus En Rev 248–258
21. Renno C, Petito F, Gatto A (2015) Artificial neural network models for predicting the solar
radiation as input of a concentrating photovoltaic system. En Con Mgt 999–1012
22. Hameed WI, Sawadi BA, Al-Kamil SJ, Al-Radhi MS, Al-Yasir YIA, Saleh AL, Abd-Al-
Hameed RA (2019) Prediction of solar irradiance based on ANN. Inventions 45
23. Linares-Rodriguez A, Ruiz-Arias JA, Pozo-Vazquez D, Tovar-Pescador J (2013) An artificial
neural network ensemble model for estimating global solar radiation from Meteosat satellite
images. Energy 636–45
24. Linares-Rodriguez A, Quesada-Ruiz S, Pozo-Vazquez D, Tovar-Pescador J (2015) An evolu-
tionary artificial neural network ensemble model for estimating hourly direct normal irradiances
from Meteosat imagery. Energy 264–73
25. Tasadduq I, Rehman S, Bubshait K (2002) Application of neural networks for the prediction
of hourly mean surface temperatures in Saudi Arabia. Ren Energ 545–554
26. Abdel-Aal, Radwan E (2004) Hourly temperature forecasting using abductive networks. Eng
App Artficial Intell 543–556
27. Smith BA, Hoogenboom G, McClendon RW (2009) Artificial neural networks for automated
year-round temp prediction. Comp Electro Agric 52–61
28. Dombaycı ÖA, Gölcü M (2009) Daily means ambient temp prediction using artificial neural
network method: a case study of Turkey. Ren Energ 1158–61
29. Deligiorgi D, Philippopoulos K, Kouroupetroglou G (2013) ANN based methodologies for the
spatial and temporal estimation of air temp. In: International conference on pattern recognition
applications and methods
30. Almonacid F, Pérez-Higueras P, Rodrigo P, Hontoria L (2013) Generation of ambient
temperature hourly time series for some Spanish locations by ANNs. Ren Energ 285–91
31. Cobaner M, Citakoglu H, Kisi O, Haktanir T (2014) Estimation of mean monthly air
temperatures in Turkey. Comp Electro Agric 71–79
32. Kisi O, Shiri J (2014) Prediction of long-term monthly air temperature using geographical
inputs. Int J Climato 179–86
33. Venkadesh S, Hoogenboom G, Potter W, McClendon R (2013) A genetic algorithm to refine
input data selection for air temperature prediction using ANNs. App Soft Comp 2253–60
On the Performance of MIMO-UVLC
System over Turbulence-induced Fading
Channels

Sajid Mumtaz, Azrina Abd Aziz, and Komal Masroor

Abstract Visible light communication in underwater environments is replacing the

RF and the acoustic signals due to its high bandwidth, low latency, and small atten-
uations. One of the performance degrading factors in underwater channels is the
turbulence which is caused by a continuous change in refractive index of water.
Spatial diversity can overcome the optical scintillations as well as the pointing errors.
This paper evaluates the ergodic capacity and the BER performance of MIMO as
well as SSK-MIMO UVLC system in presence of underwater turbulence, modelled
by Lognormal, Gamma, and Weibull distributions. The capacity analysis is based
upon varying number of transceivers and the water filling algorithm. Impact of
spatial correlation is also considered for the capacity analysis. The results find their
validation via Monte Carlo simulations.

Keywords Underwater visible light communication (UVLC) · Multiple input

multiple output (MIMO) · Optical turbulence · Ergodic capacity · Distribution
models · Space shift keying (SSK)

1 Introduction

Optical signals are currently in target stream of researchers due to their bandwidth
and power efficiency, least latency, and reliability compared to traditional RF and
acoustic signals. However, the transmission of optical signals underwater faces many
challenges, such as absorption, scattering, turbulence, noise and pointing errors, that
need to be resolved. Absorption and scattering of optical signals have been widely
studied and modelled. Moreover, the blue and green window of optical spectrum
offers minimal absorption and scattering [1], making it to be an optimized option

S. Mumtaz (B) · A. A. Aziz · K. Masroor

Department of Electrical and Electronic Engineering, Universiti Teknologi PETRONAS, Seri
Iskandar, Malaysia
K. Masroor
Department of Electronic and Telecommunications Engineering, NED University of Engineering
and Technology, Karachi, Pakistan

© The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 373
R. Ibrahim et al. (eds.), International Conference on Artificial Intelligence
for Smart Community, Lecture Notes in Electrical Engineering 758,
https://doi.org/10.1007/978-981-16-2183-3_36
374 S. Mumtaz et al.

for wireless communication, hence called underwater visible light communication

(UVLC).
Underwater turbulence, caused by rapid variations in temperature, salinity, air-
bubbles and pressure, severely degrades the optical signals [2–4] and it is still an open
challenge. Various statistical distributions, such as Lognormal, Gamma, Weibull,
Generalized Gamma, Gamma-Gamma (GG), and some mixture distributions have
been developed to characterize the scintillations due to optical turbulence underwater.
For weak turbulence-induced fading with scintillation index σi2 < 0.1, Lognormal
distribution is the most widely used because of the weighted sum approximation of its
random variables (RVs) [5–8]. For very low fading due to temperature gradients with
scintillation index σi2 < 0.001, Gamma distribution fits well with the optical intensity
fluctuations [9], and its mathematical simplicity provides tractable solutions. For
salinity-induced turbulence fading with scintillation index 0.1 < σi2 < 1, Weibull
distribution model is the most appropriate one [10].
The channel capacity for underwater VLC channels has been studied and analyzed.
Zhang [11] studied the ergodic capacity and the BER performance of UVLC system
with pointing errors (PEs) under different water types, and observed that water
turbidity and long travelling distance of photons affect both the capacity and the
BER performance. Li [12] an analytic and capacity model based upon the PEs vari-
ance for turbulence-induced UVLC links. Xu [13] analyzed the scintillation index
of plane wave propagation in anisotropic water turbulence under moderate-to-strong
turbulent conditions, and noticed the large receiving apertures and wavelengths to
improve significantly the UVLC quality and capacity.
The scintillations due to the optical turbulence are reduced using the methods,
like Adaptive optics [14], Rytov method [15], aperture averaging [16], and higher-
mode flat topped lasers [17]. MIMO is a proven technology to overcome fading, the
misalignments, and the possible blockage due to the obstructions in free space optical
(FSO) communications. Like RF, the capacity of MIMO Optical Wireless Commu-
nication (OWC) systems is also investigated rigorously. The ergodic capacity and
probability outage for MIMO-FSO links under Lognormal, Rayleigh and Gamma-
Gamma channels under different power constraints and modulation techniques have
been investigated and analyzed [18–20], and the benefits of MIMO are proved in
optical domain as well.
However, the performance of UVLC systems under spatial diversity techniques
received lack of attention. Jamali [5, 8] analyzed the performance of UVLC system
under MIMO technique, and proved that the spatial diversity remarkably improved
the system’s performance in weak underwater turbulence regimes. Abdallah and
Khalighi [21] modeled UVLC channel based upon angular MIMO scheme, and
found that compared to conventional MIMO, it withstood against the misalignment
conditions for the channel capacity. Li [22] analyzed the effect of spatial corre-
lation on imaging MIMO, and verified the fact that the correlation degraded the
UVLC system performance. Jiang [23] evaluated the performance of MIMO-UVLC
systems in combination with OFDM technique, and proved its worth to cope with
the underwater turbulence-induced fading.
On the Performance of MIMO-UVLC System over … 375

Moreover, for increased spectrum efficiencies, MIMO space shift keying (SSK)
technique has been widely applied to OWC systems; where instead of signal
constellations, the data bits are transmitted via antenna indices providing reduced
system complexities. Abaza [24] assessed the performance of optical SSK-MIMO
under atmospheric turbulence and observed an increase in spectral efficiency of 3
bits/s/Hz. Normally the existence of correlated signals in MIMO always reduced the
system’s performance. But Priyadarshani and Jaiswal [25] reported that the usage
of SSK-MIMO over correlated FSO channel provided a significant increased BER
performance up to 13 dB.
This paper evaluates the ergodic capacity for MIMO UVLC systems under
Lognormal, Gamma and Weibull distributed channels based upon their achieved
SNR. These models characterize the optical scintillations for a particular range of
underwater turbulence conditions. Sections 2 and 3 cast light over the employed
system and the channel models, respectively. Section 4 explains about the ergodic
capacity with MIMO technology with some light over the correlated cannels. For
enhanced MIMO spectrum efficiency, an overview about Space Shift Keying (SSK)
is provided in Sect. 5. Section 6 presents result and discussion, and Sect. 7 gives the
conclusion and the future work.

2 System Model

Consider an UVLC link with M number of transmitters (LEDs/LDs) pointing towards

N number of receivers (PDs). We assume that this link utilizes intensity modulation
direct detection (IMDD) with OOK modulation. We assume the performance of the
given system is affected by background radiations, thermal noise, shot noise, which
are normally modelled as independent and identically distributed (iid) additive white
Gaussian noise (AWGN). To obtain the uncorrelated optical signals at the receiver,
we assume that the separation between the receivers is greater than the irradiance
spatial coherence distance, and the receiver’s aperture diameter is less than the spatial
correlation width. If the scattering is not too high, then the correlated signals will be
received at the receiver, and that results in the reduced capacity and low data rates.
The received signal at the nth receiver aperture is

r (t) = ηs R Pt H s(t) + n(t) (1)

where s ∈ {0, 1} denotes the transmitted signal, and Pt is the transmitted power. η
represents the optical-to-electrical conversion coefficient, R is the receiver’s respon-
sivity, n(t) ∈ R M×N is the AWGN with zero mean, and N0 /2 noise power spectral
density (PSD). Moreover, H is the MIMO channel matrix comprising of row and
column vectors, and s(t) is the transmitted intensity OOK modulated optical beam.
376 S. Mumtaz et al.

3 Channel Model

For the analysis, we considered Lognormal, Gamma and Weibull distribution models
that can characterize the optical scintillations due to weak, temperature-as well as
salinity-induced turbulence, respectively.

3.1 Lognormal Distribution Channel

The intensity fluctuations caused by underwater weak turbulence are characterized by

the fading coefficient I , which can be modelled by the PDF of Lognormal distribution
given by
 
1 (ln(I ) − 2μx )2
f I (I ; μx , σx2 ) =  exp (2)
2I 2π σx2 8σx2

   
where μx = ln(I0 )/2 and σ 2 = E I 2 /E2 {I } − 1 is the logscale parameter (first
moment) and the scintillation index (second moment) of normal distributed log-
amplitude factor respectively, with I0 as the received irradiance in absence of the
considered channel.

3.2 Gamma Distribution Channel

The PDF for Gamma distribution is

 
1 (a−1) I
f I (I ; a, b) = I exp − (3)
(a)ba b

with a and b as shape and scale parameters. (.) is the Gamma function which
is related to the S.I by the relation b = 1/a = σ I2 . For UVLC links with very
low temperature-gradients, i.e., σ I2 ≤ 0.001, this simple distribution can be used to
characterize the fluctuations of received optical beam.

3.3 Weibull Distribution Channel

The PDF for Weibull distribution is given by

On the Performance of MIMO-UVLC System over … 377

Fig. 1 Lognormal, Gamma and Weibull distributions with zero mean and unit variance for intensity
fluctuations

    a 
a I (a−1) I
f I (I ; a, b) = exp − (4)
b b b

with a and b as shape and scale parameters, respectively. I , a and b are all
positive valued. The parameter b is related to the S.I by the relation σ I2 =
b2 {[1 + (2/λ)] − [1 + (1/λ)]}. It is noted that for a = 1, this distribution yields
f (I ; 1, b) = (1/b) exp(−(I /b)) which is negative exponential distribution, and
for a = 2, this distribution is converted to the Rayleigh distribution. The Weibull
distribution fits well for salinity-induced turbulence fading in underwater channels.
The graph between the probability of irradiance because of the turbulence and the
optical power per unit solid angle for the considered distribution models is shown in
Fig. 1.

4 Ergodic Capacity

The utmost mutual information that can be attained by changing the PDF of transmit
signal vector is called the capacity, which is an important parameter in characterizing
the performance of a wireless communication system. The instantaneous channel
capacity is defined as
 
Cinst = E log2 (1 + cI ) (5)
378 S. Mumtaz et al.

where c is a constant term, with c = 1 and c = e/2π for heterodyne detection as

well as intensity modulation and direct detection (IMDD) respectively [26].
We assume that the optical channel is memoryless, stationary, and perfect channel
state information (CSI) is available at both transmitter and receiver of the system. In
terms of MIMO channel, the capacity relation can be written as


C = E log det I M + (1/M)H H H (6)


where I M denotes M × M identity matrix, and H is an M × N channel matrix with
elements being Imn . Since the instantaneous SNR is a random variable (RV) with a
given PDF, so we assume the channel capacity as a RV. The average of this RV is the
ergodic capacity, and can be written as

∞
C= B log2 (1 + I ) f I (I )d I (7)
0

where B denotes the transmission bandwidth. By substituting the values of f I (I )

from Eqs. (2), (3), and (4) into Eq. (7), the ergodic capacity closed-form expressions
can be obtained.
The expression for ergodic capacity under MIMO-UVLC Lognormal distribution
channel is
∞  
1 (ln(I ) − 2μx )2
CL N = B log2 (1 + I )  exp dI (8)
2I 2π σx2 8σx2
0

The ergodic capacity for MIMO-UVLC Gamma distribution channel will be

∞  
1 (a−1) I
C Gam = B log2 (1 + I ) I exp − dI (9)
(a)ba b
0

Similarly, the ergodic capacity expression for MIMO-UVLC for Weibull distri-
bution channel can be expressed as

∞     a 
a I (a−1) I
CW b = B log2 (1 + I ) exp − dI (10)
b b b
0
On the Performance of MIMO-UVLC System over … 379

5 Space Shift Keying MIMO

Underwater communication severely needs the wireless systems to be power as

well as bandwidth efficient. Although MIMO technique can provide reduced signal
fading [8], yet it could not provide high spectrum efficiencies. Opting higher order
modulations can be a viable way out to attain spectrum efficient (SE) systems, but
they result in higher peak to average power ratios (PAPR) which is another problem.
There is an emerging method for achieving high SE in presence of PAPR is the spatial
modulation (SM) MIMO [27] which utilizes both signal constellations as well as
antenna indices at once. Some of the incoming bits are transmitted via signal space,
while the remaining are transmitted through the transmitter index, thus increasing
the SE. Being one antenna active at a time, power efficiencies are also achieved.
SSK is a special case of SM which uses only the antenna indices for data transfer.
SSK provides zero inter channel interference (ICI) and least detection complexities,
making the system simpler compared to SM [28, 29]. These features make SSK-
MIMO a good candidate for UVLC systems. This paper also analyzes the BER
performance of optical SSK-MIMO UVLC system under Lognormal, Gamma and
Weibull distribution models.

6 Results and Discussion

This section discusses the simulation-based results. As spatial diversity mitigates

with the fading and increase the data rates, so with increasing number of transmitters
and receivers, we increase the data rate, and obviously the capacity for the given
channel. Figure 2a, b depict the channel capacity under different values of SNR
with increasing number of transmitters and receivers. More receivers will receiver
different SNR values that can be combined by maximal ratio combining (MRC) and
equal gain combining (EGC) techniques, yielding net higher SNR, and offering huge

Fig. 2 a Channel capacity with varying number of receivers when number of transmitters is fixed,
b Channel capacity with varying number of transmitters when number of receivers is fixed
380 S. Mumtaz et al.

Fig. 3 a Channel capacity versus received SNR based on water filling algorithm, b compar-
ison between iid and correlated MIMO UVLC channel under lognormal, Gamma, and Weibull
distributions

channel capacity. For lower SNR values less than a specific threshold, an increment
in the number of receivers does not help in providing the higher capacity and the
data rates.
Figure 3a shows the channel capacity based on water-filling algorithm, where the
optical receivers adapt themselves according to the channel conditions to attain the
desired power points. When the channel state information (CSI) is not known at the
transmitter, then the transmitter throws the power isotopically towards the receivers.
Due to path losses, some of the power goes wasted, and the channel capacity is
affected. However, in presence of CSI known to the transmitter, the transmitted
power is then optimized, and it acts as a beamforming MIMO. For known CSI,
with increase of transmitters and receivers, the channel capacity increases and thus
offers high data rate transmission. Compared to SISO systems, MIMO provides an
increased capacity of up to 11 dB, that can further be increased employing a greater
number of transceivers, but at the cost of increased system complexities.
Much of the study is done for analyzing the capacity of MIMO channels when
all the optical signals are independent and identically distributed (iid). In a scat-
tering environment, the channel impulse response (CIR) has independent fingerprints
helping the receivers to identify and differentiate among the signals. On the receiver
side, this condition can be achieved by making separation between the receivers
more than the channel correlation length. However, in some practical scenarios, the
realization of this condition is not possible, and this makes the analysis of correlated
channels more crucial and important. In less scattering environments, there exists
the correlation among the signals, and hence the performance of the link is degraded.
Figure 3b shows the ergodic capacity of Lognormal, Gamma and Weibull distribu-
tion channels with varying SNR values. The results show that Gamma channel yields
higher capacity and data rates compared to Lognormal and Weibull channels. The
results also show that in presence of correlation among signals, the capacity as well
as the data rate under all channels reduce significantly.
On the Performance of MIMO-UVLC System over … 381

Fig. 4 BER versus SNR for optical SSK-MIMO UVLC system under Lognormal, Gamma and
Weibull distributions

The BER performance of optical SSK MIMO system under Lognormal, Gamma,
and Weibull distribution channels is manifested in Fig. 4. As SSK uses only the
antenna indices for data transfer, so the system and detection complexity are reduced,
and spectrum efficiency is achieved. BER versus SNR plot shows that OSSK-MIMO
under Gamma distribution channel outperforms Lognormal and Weibull channels.
However, its performance under Lognormal channel is much better than that of
Weibull distribution.

7 Conclusion and Future Work

This paper evaluates the ergodic capacity and BER performance of MIMO
underwater visible light communication (UVLC) under Lognormal, Gamma, and
Weibull distributions which can characterize the optical scintillations due to weak,
temperature-induced, and salinity-induced turbulence fading, respectively. More
transceivers can transmit and receive more power, and that can be combined using
MRC and EGC techniques to produce high SNR value, and hence that of the channel
capacity. To optimize the transmit power, CSI is required at the transmitter, because an
unknown CSI yields a reduced amount of channel capacity. Moreover, it is observed
that in presence of correlation among optical signals at the receivers, the ergodic
capacity for all channels reduces. The results also show that Gamma channel provides
higher capacity and data rates; however, it can model only low temperature gradient
382 S. Mumtaz et al.

turbulence induced underwater fading. The results also show that the performance
of SSK-MIMO under Gamma distribution channels is much better than those of
Lognormal and Weibull channels.
The mixture distribution models, named Exponential and Lognormal, Exponen-
tial and Gamma, Exponential and Weibull, Exponential and Generalized Gamma
(EGG), are also found to fit well with the experimental data for different under-
water turbulence conditions. The performance and capacity analysis of MIMO as
well as MIMO-OFDM UVLC systems under such channels with suitable coding
and modulation techniques are still open research issues in such regimes.

References

1. Duntley SQ (1963) Light in the sea. JOSA 53:214–233

2. Kaushal H, Kaddoum G (2016) Underwater optical wireless communication. IEEE Access
4:1518–1547
3. Tang S, Zhang X, Dong Y (2013) Temporal statistics of irradiance in moving turbulent ocean.
In: 2013 MTS/IEEE OCEANS-Bergen, pp 1–4
4. Andrews LC, Phillips RL (2005) Laser beam propagation through random media
5. Jamali MV, Nabavi P, Salehi JA (2018) MIMO underwater visible light communications:
comprehensive channel study, performance analysis, and multiple-symbol detection. IEEE
Trans Veh Technol 67:8223–8237
6. Huang A, Tao L, Wang C, Zhang L (2018) Error performance of underwater wireless optical
communications with spatial diversity under turbulence channels. Appl Opt 57:7600–7608
7. Peppas KP, Boucouvalas AC, Ghassemloy Z (2017) Performance of underwater optical wireless
communication with multi-pulse pulse-position modulation receivers and spatial diversity. IET
Optoelectron 11:180–185
8. Jamali MV, Salehi JA, Akhoundi F (2016) Performance studies of underwater wireless optical
communication systems with spatial diversity: MIMO scheme. IEEE Trans Commun 65:1176–
1192
9. Oubei HM, Zedini E, ElAfandy RT, Kammoun A, Abdallah M, Ng TK et al (2017) Simple
statistical channel model for weak temperature-induced turbulence in underwater wireless
optical communication systems. Opt Lett 42:2455–2458
10. Oubei HM, Zedini E, ElAfandy RT, Kammoun A, Ng TK, Alouini M-S et al (2017) Efficient
Weibull channel model for salinity induced turbulent underwater wireless optical communica-
tions. In: 2017 opto-electronics and communications conference (OECC) and photonics global
conference (PGC), pp 1–2
11. Zhang H, Dong Y, Hui L (2015) On capacity of downlink underwater wireless optical MIMO
systems with random sea surface. IEEE Commun Lett 19:2166–2169
12. Li Y, Zhang Y, Zhu Y (2019) Capacity of underwater wireless optical links with pointing errors.
Opt Commun 446:16–22
13. Xu G, Lai J (2020) Average capacity analysis of the underwater optical plane wave over
anisotropic moderate-to-strong oceanic turbulence channels with the Málaga fading model.
Opt Exp 28:24056–24068
14. Holohan M, Dainty J (1997) Low-order adaptive optics: a possible use in underwater imaging?
Opt Laser Technol 29:51–55
15. Ata Y, Baykal Y (2014) Scintillations of optical plane and spherical waves in underwater
turbulence. JOSA A 31:1552–1556
16. Yi X, Li Z, Liu Z (2015) Underwater optical communication performance for laser beam
propagation through weak oceanic turbulence. Appl Opt 54:1273–1278
On the Performance of MIMO-UVLC System over … 383

17. Yousefi M, Golmohammady S, Mashal A, Kashani FD (2015) Analyzing the propagation

behavior of scintillation index and bit error rate of a partially coherent flat-topped laser beam
in oceanic turbulence. JOSA A 32:1982–1992
18. Chakraborty K (2005) Capacity of the MIMO optical fading channel. In: Proceedings of
international symposium on information theory, ISIT 2005, pp 530–534
19. Luong DA, Thang TC, Pham AT (2013) Average capacity of MIMO/FSO systems with equal
gain combining over log-normal channels. In: 2013 fifth international conference on ubiquitous
and future networks (ICUFN), pp 306–309
20. Zhang J, Dai L, Han Y, Zhang Y, Wang Z (2015) On the ergodic capacity of MIMO free-space
optical systems over turbulence channels. IEEE J Sel Areas Commun 33:1925–1934
21. Ghazy AS, Hranilovic S, Khalighi M-A (2019) Angular MIMO for underwater wireless
optical communications: channel modelling and capacity. In: 2019 16th Canadian workshop
on information theory (CWIT), pp 1–6
22. Li Y, Qiu H, Chen X, Fu J, Musa M, Li X (2019) Spatial correlation analysis of imaging MIMO
for underwater visible light communication. Opt Commun 443:221–229
23. Jiang H, Qiu H, He N, Popoola W, Ahmad Z, Rajbhandari S (2020) Performance of spatial
diversity DCO-OFDM in a weak turbulence underwater visible light communication channel.
J Lightwave Technol 38:2271–2277
24. Abaza M, Mesleh R, Mansour A, Aggoune E-HM (2015) The performance of space shift
keying for free-space optical communications over turbulent channels. In: Broadband access
communication technologies IX, p 93870V
25. Priyadarshani R, Jaiswal A, Bhatnagar MR, Ghassemlooy Z, Zvanovec S (2018) Performance
of space shift keying over a correlated gamma-gamma FSO-MISO channel. In: 2018 11th
international symposium on communication systems, networks and digital signal processing
(CSNDSP), pp 1–6
26. Ansari IS, Alouini M-S, Cheng J (2014) On the capacity of FSO links under lognormal
and Rician-lognormal turbulences. In: 2014 IEEE 80th vehicular technology conference
(VTC2014-Fall), pp 1–6
27. Jaiswal A, Bhatnagar MR, Jain VK (2016) BER analysis of optical space shift keying with
gamma-gamma fading and pointing error. In: 2016 international conference on advanced
technologies for communications (ATC), pp 369–374
28. Di Renzo M, Haas H, Ghrayeb A, Sugiura S, Hanzo L (2013) Spatial modulation for generalized
MIMO: challenges, opportunities, and implementation. Proc IEEE 102:56–103
29. Jeganathan J, Ghrayeb A, Szczecinski L, Ceron A (2009) Space shift keying modulation for
MIMO channels. IEEE Trans Wirel Commun 8:3692–3703
Fabrication of One Degree of Freedom
Knee Exoskeleton for Nervous Disorder

J. Duwaraka, N. Sudharsana, S. Prabakar, K. Porkumaran,

Iraivan Ilamvaluthi, and Ramani Kannan

Abstract Instrumented Cane is a wearable sensory system for patient’s motion

intention estimation in order to control the trajectory of knee exoskeleton. This
Instrumented Cane will be a helpful assistive device in term of balance and motion
for those who suffering neurological disorder, in specific Paralysis and spinal cord
injury. This report describes the study on method utilizing Instrumented Cane as
a attainable intermediate between the patient and exoskeleton to enhance in terms
of movement in their day-to-day activities. The structure is designed with inertial
measurement unit integrated with wireless communication module in order to assist
real-time control of exoskeleton. This report aims the purpose of an Instrumented
Cane that able to contribute to the assisted locomotion technology as well as upgrade
the mobility of physically challenged people.

Keywords Knee · Wireless · One DOF · Paralysis · Serial communication ·

Instrumented cane and control trajectory

1 Introduction

Paralysis is a common term used to define the defective sensation due to following
damage to the nervous system. In medical terms accurate level of injuries are used
which is supportive in determining the parts of body that affected by paralysis and also
loss of function. Generally, this disease can be within a temporary period or even worst

J. Duwaraka · N. Sudharsana (B)

Biomedical Engineering, Dr. N. G. P Institute of Technology, Coimbatore, India
S. Prabakar
Sona College of Technology, Salem, India
K. Porkumaran
Sri Sairam Engineering College, Chennai, India
e-mail: [email protected]
I. Ilamvaluthi · R. Kannan
Electrical and Electronic Engineering, Universiti Teknologi Petronas, Seri Iskandar, Malaysia

© The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 385
R. Ibrahim et al. (eds.), International Conference on Artificial Intelligence
for Smart Community, Lecture Notes in Electrical Engineering 758,
https://doi.org/10.1007/978-981-16-2183-3_37
386 J. Duwaraka et al.

to be permanent. In common cases, paralysis affects half side of human body, called
as unilateral. Under classification of paralysis, there is a particular medical condition
defined as Paraplegia [4]. Paraplegia is defect affecting the lower extremities of
human body. This term describe the paralysis that affects both legs and partially the
trunk, but not the arms. Contracting the study of Paraplegia, there is another form
of paralysis that affects just one limb known as Monoplegia. Crucial Monoplegia is
scientific term for paralysis affecting the single lower limb [5]. Patients with these
disabilities have trouble in walking as well as reflecting their muscle. Conversely,
rapid development of technology has formed a solution with a design mechanism of
exoskeleton in conjunction of instrumented cane for the control.
Development in this field creates an enthusiasm towards patients to experience the
transaction from undergoing rehabilitation therapy to a “real world” walking prac-
tice. Human motion intention depends on certain bioelectrical signal of neurology
system. Because of variation on their bioelectrical difficulties in mobility. Research
community developed several wearable gait, exoskeleton, instrumented cane that
will compute and estimate the motion intention by disable. Nevertheless, there are
several constrains on the design mechanism and followings are:
1. Modelling with variable electronic hardware modules are composite.
2. In data transmission power supply for wireless is ineffectual.
3. Immense design and additional weight restrain, uncomfortable cane.
Movement transparency is critical when wearing a robot for gait rehabilitation.
In other words: When wearing the exoskeleton, its movement should be synchro-
nized and consistent with a patient’s natural movement. “If not, it exerts extra forces
on the human joint “And this extra force causes patient discomfort and unnatural
movements.”

1.1 Limitation in Current Design

Human motion intention depends on certain bioelectrical signal of neurology system.

Because of variation on their bioelectrical difficulties in mobility. Research commu-
nity developed several wearable gait, exoskeleton, instrumented cane that will
compute and estimate the motion intention by disable. Nevertheless, there are several
constraints on the design mechanism and followings are:
• Modeling with variable electronic hardware modules is composite.
• In data transmission power supply for wireless is ineffectual.
• Immense design and additional weight restrain uncomfortable cane.
Movement transparency is critical when wearing a robot for gait rehabilitation.
In other words: When wearing the exoskeleton, its movement should be synchro-
nized and consistent with a patient’s natural movement. “If not, it exerts extra forces
on the human joint” And this extra force causes patient discomfort and unnatural
movements.
Fabrication of One Degree of Freedom Knee Exoskeleton … 387

1.2 Objectives

• The hardware connection is simple with wireless transmission of information.

• The design is light weight to control the knee exoskeleton.
• The battery can be replaced easily and used for long duration.

2 Scope of Proposed System

The expected result of this project is to develop an instrumented cane that able
to communicate and control a lower limb knee exoskeleton together with control
algorithm that able to notice human motion trajectory. This work is to contribute in
the use of rehabilitation for remote assessments as well as for the disable with their
day-to-day activities. Throughout this project, research will be performed on the
required balance in motion, ability, upper body strength, and the use of arm strength
with the instrumented cane. Besides, the scope of this project is also to concentrate
on the wireless communication system to assist in real-time control of exoskeleton
via a standard I2C serial communication protocol.

2.1 Proposed Work

IMU sensor was attached to the instrumented cane to detect the movement trajectory
that is discharged. The output signal obtained from the sensor will be illuminated
using a microcontroller that is compatible with the sensor and wireless communi-
cation module. The data will then be mediate to the microcontroller used at the
exoskeleton in order to control the trajectory based on, the information is carried
via Bluetooth communication sensor. Instrumented cane and exoskeleton enhance
the compatibility using Bluetooth communication protocol as the best option. As the
information is conveyed to the exoskeleton at real-time, it begins to initiate the motor
to act as actuator in order to make motion. Fundamentally, humans target in motion
triggers first, in sequence to command the robotic structure make motion. The push
button will be assigned as a safety circumstance if there is any sudden detach or drop
of cane from the arm, it will not send any trajectory control data to the Exoskeleton.
This prevents the exoskeleton to make any false motion trajectory which could cause
weakness. In related parts for Exoskeleton section, the Microcontroller, Controller,
and Motor subsystem will be continuation from previous project.
388 J. Duwaraka et al.

Fig. 1 One degree of

freedom knee exoskeleton

3 Methodology

After a thorough study on several research papers, there are several different types
of instrumented cane with different design mechanism in order to control the trajec-
tory of exoskeleton. Two major factors is considered in this project, first the sensor
for angle reflective properties and second on the communication medium. Real-
time assessments on the motion intension can be establish using rotational attributes
computing sensing technology integrated with radio-communication technology in
order to control the trajectory of exoskeleton. Wireless communication medium plays
an important role as exoskeleton response based on acquired real-time data signal
that receive from the cane (Fig. 1).

4 Hardware Implementation

The development of exoskeleton design comprise of instrumented cane which control

the knee trajectory of one degree of freedom motion using the exoskeleton. The degree
of freedom that the exoskeleton can perform is flexion and extension. The instru-
mented cane consists of IMU (inertia measurement unit) sensor which is unavoidable
and provides best information about body motion. It detects the movement trajectory
of push button. It consists of two major specialities, compatible with Arduino and low
power consumption. Sensor consist of a single chip embedded with accelerometer,
gyroscope and provide information with high accuracy.
Fabrication of One Degree of Freedom Knee Exoskeleton … 389

The chip uses I2C (inter integrated circuit) protocol for serial communication to
function according to various body condition. The weight of sensor is 5 g which is
helpful in reducing the weight of the prototype. Arduino nano,interpret information
from MPU6050 and sends through I2C protocol to Bluetooth HC05.The board is
user friendly and smaller in size. The power requirement is 5 V. The Bluetooth
module HC05 act as the transmitter and has the capacity to act as both master and
slave according to the preference. In this work it performs as master and can be
paired with any slave device automatically. The consumption of power is low and
the average current fluctuation range of 25 mA.
The communication range of distance required is 1–3 m and it is applicable for
HC05 module. The HC06 bluetooth module in exoskeleton executes as the receiver
and performs as slave. It weighs 0.9 g and suitable for wireless communication. The
exoskeleton consist of controller boards Arduino UNO and MEGA. UNO act as the
intermediate and act as middle device and information from Bluetooth receiver and
convert into digital signals and relay it to Arduino MEGA. The coded program in the
mega rotates the servo/cylindrical motor. The controller circuit controls the torque
and provides the flexion and extension of the exoskeleton.
The design and development of electromechanical system for instrumented cane
was done by modifying the mechanical design of a regular cane and instrumented
with respecting sensor and microcontroller boards. The main objective is to able to
modify an off-the-shelf cane such that it can house position, orientation and force
sensor together with wireless microcontrollers to interpret and send data in order
to control the exoskeleton trajectory. Based on the design of instrumented cane as
illustrates in Fig. 2, additional weight factor on the cane was took into consideration
in order to achieve one of the targeted objective. Execution on this project believed
to deliverable an instrumented cane that able to sense the motion trajectory either
forward or backward by the user and sends the real-time data in order to initiate
trajectory of exoskeleton. This instrumented cane will be developing strong spirit to
experience “real world walking” especially for those who suffering in Cerebrovas-
cular Accidents (CVA) and Spinal Cord Injury (SPI). Critical consideration on the
design specification whereby the instrument cane should be wireless, portable, light
in weight as well as meets clinical requirements so that this cane is use comfortably.
In conclusion, this instrument cane technology target to assist disability patients as
all of the objectives is achievable.

5 Conclusion

IMU sensor was attached to the instrumented cane to detect the movement trajectory
that is discharged. The output signal obtained from the sensor will be illuminated
using a microcontroller that is compatible with the sensor and wireless communi-
cation module. The data will then be mediate to the microcontroller used at the
exoskeleton in order to control the trajectory based on, the information is carried via
390 J. Duwaraka et al.

Fig. 2 Block diagram of

instrumentation system

Bluetooth communication sensor. Instrumented cane and exoskeleton enhance the

compatibility using Bluetooth communication protocol as the best option (Fig. 3).

Fig. 3 Fabricated of one

degree of freedom knee
exoskeleton
Fabrication of One Degree of Freedom Knee Exoskeleton … 391

As the information is conveyed to the exoskeleton at real-time, it begins to initiate

the motor to act as actuator in order to make motion. Fundamentally, humans target
in motion triggers first, in sequence to command the robotic structure make motion.
The push button will be assigned as a safety circumstance if there is any sudden
detach or drop of cane from the arm, it will not send any trajectory control data to the
Exoskeleton. This prevents the exoskeleton to make any false motion trajectory which
could cause weakness. In related parts for Exoskeleton section, the Microcontroller,
Controller, and Motor subsystem will be continuation from previous project. Consid-
eration is made on the microcontroller that been used on exoskeleton, is the same
microcontroller that will be used for instrumented cane. This is important feature,
since to establish successful communication and suitable between two device in
terms of control.

References

1. SCI stats and facts. Available http://danceforparalysis.org/sci-stats-and-facts/

2. Ibrahim A, Lee KY, Kanoo LL, Tan CH, Hamid MA, Hamedon NMA et al (2013) Epidemiology
of spinal cord injury in hospital Kuala Lumpur. Spine 38:419–424
3. Selvathurai G (2016) Development of instrumented cane for control of exoskeleton of lower
limb. Dissertation, 18662
4. Osman M, Tageldeen M (2015) Development of a wearable exoskeleton for arm rehabilitation.
Dissertation, 17868
5. Huo W, Mohammed S, Amirat Y (2018) Impedance reduction control of a knee joint human-
exoskeleton system. IEEE J 1063_6536
6. Smith BY (2015, Apr 9) What is paraplegia? Available http://www.news-medical.net/health/
What-is-Paraplegia.aspx
7. Sadek A-R, Parmar NK, Sadek N-H, Jaiganesh S, Elkhodair S, Jaiganesh T Spontaneous upper
limb monoplegia secondary to probable cerebral amyloid angiopathy. Int J Emerg Med 5:1–1
8. Chang WH, Kim Y-H Robot-assisted therapy in stroke rehabilitation. J Stroke 15:174–181
9. Hassan M, Kadone H, Suzuki K, Sankai Y (2013) Wearable gait measurement system with an
instrumented cane for exoskeleton control. Sensors (Basel) 14:1705–1722
10. Claire P, Joyce F (2011) An instrumented cane devised for gait rehabilitation and research. J
Phys Therapy Educ 25(1):36–41
11. Kyle BH (2017) Design and control of a lower limb exoskeleton emulator for accelerated
development of gait exoskeleton. Proquest number 10283413
12. Mahmoodi SN, Hwan-Sik yoon KW, Haskew TA (2015) Design and control of power limb
prosthesis. Proquest number 10239909
13. Zetteberg JH (2015) Design and development of two degree of freedom haptic device for
neurorehabilitation. Proquest number 10020547. [13] Mussay SA (2016) Development and
assessment of a control approach for lower limb exoskeleton for use in gait rehabilitation post
stroke, May 2016. Proquest number 101551706
14. Gong C (2016) Development of robotic exoskeleton system for gait rehabilitation. Proquest
number 10310874
15. Hao MA (2015) Mechatronic design and control of assistive knee brace for gait rehabilitation,
September 2015. Proquest number 10297313
16. Luna CO (2016) Non linear control of a SVEN degree of freedom exoskeleton robot arm,
March 21, 2016. Proquest number 10673642
17. Pehlivan AU (2016) Subject adaptive control paradigm for robotic rehabilitation, May 2016.
Proquest number 10308368
392 J. Duwaraka et al.

18. Zhang J (2016) Towards systematic controller design in rehabilitation robot, August 2016.
Proquest number 10308368
19. Bing C (2017) Design and implementation of a lower extremity exoskeleton for motion
assistance to paraplegic patients
20. Jhawar V (2018) Design of knee exoskeletongait assistance, August 2018. Proquest number
10846724
Semi-Automatic Detection
and Measurement of Fetal Parameters
from Ultrasound Images and the Scope
Automatic System Using LabVIEW

S. Prabakar, K. Porkumaran, J. Samson Isaac, R. Karthikeyan, G. Gopu,

and Ramani Kannan

Abstract The idea of developing an automated fetal parameter detection and

measurement system using image processing techniques available in LabVIEW is
proposed in this research work. The six fetal parameters to be measured from the
ultrasound images are crown rump length, biparietal diameter, femur length, head
circumference, abdominal circumference and humerus length. We propose the semi-
automatic and automatic process using the vision assistant tool in LabVIEW. The
semi-automatic process is wherein the parameters are segregated in to six SUB VI’s
and are programmed into one single MAIN VI. This is then followed by manually
dragging the region of interest ROI which could be either circle or line annotation and
the values of the result is obtained. The automatic process is being performed only
for four parameters namely AC, HC, HL and FL. The trimester fetal scan image is
loaded and by applying certain threshold levels and performing convolution as well
as morphological operations using different icons such as contour extract, the search
direction and the contour selection are used to locate the images from left to right,
top to bottom, maximum length, etc., for automation. Here, it selects the appropriate
ROI and the values of the result are obtained. The main advantage of this work is that
all the conventional methods of fetal anatomical parameter detections are improved
with customized LabVIEW and the vision assistant tool. The conventional procedure

S. Prabakar (B)
Sona College of Technology, Salem, India
K. Porkumaran
Sri Sairam Engineering College, Chennai, India
e-mail: [email protected]
J. Samson Isaac
Biomedical Engineering, Karunya University, Coimbatore, India
R. Karthikeyan
Dr NGP Institute of Technology, Coimbatore, India
G. Gopu
ECE, Sri Ramakrishna Engineering College, Coimbatore, India
R. Kannan
EEE, Universiti Teknologi Petronas, Seri Iskandar, Malaysia

© The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 393
R. Ibrahim et al. (eds.), International Conference on Artificial Intelligence
for Smart Community, Lecture Notes in Electrical Engineering 758,
https://doi.org/10.1007/978-981-16-2183-3_38
394 S. Prabakar et al.

consumes 20–30 min for a single subject scanning whereas in case of these proposed
techniques the time taken for automatic detection is less than 5 min based on the
system used and the manual intervention also be reduced in diagnostic procedures.

Keywords Fetal parameters · Crown rump length · Biparietal diameter · Femur

length · Head circumference · Abdominal circumference · Humerus length

1 Introduction

Accurate fetal ultrasound measurements are one of the most important factors for
high quality obstetrics health care. Common fetal ultrasound measurements include
Biparietal diameter (BPD), Head circumference (HC), Abdominal circumference
(AC), Crown rump length (CRL), Femur length (FL), Humerus length (HL). These
measures are used to estimate the getational age(GA) of the fetus and are considered
as vital parameters in an important diagnostic tool to identify the viability of the
fetus.
In clinical settings, the manual measurement by specialists of BPD, HC, AC,
CRL, FL, HL present the following issues such as the quality of measurements are
user dependent, the exam can take more than 30 min and specialists can suffer from
Repetitive Stress Injury (RSI) due to these lengthy exams. Therefore, the automatic
system of these ultrasound measures has the potential of improving productivity
and patient throughput, enhancing accuracy and consistency of measurements, and
reducing the risk of RSI to specialists.

1.1 Existing Fetal Parameter Measurement

Automatic delineation and robust measurement of fetal anatomical structure in ultra-

sound images are a challenging task due to the complexity of the object appearance,
noise, shadow and quantity of information to be processed. The previous solu-
tions rely on explicit encoding of prior knowledge and formulate the problem as
a perceptual grouping task solved through clustering or variational approaches.
These methods are known to be limited in the validity of the underlying assump-
tions and cannot capture complex structure appearances. Approaches have been
used to distinguish between the appearance of the object of interest and background
by training a discriminative constrained probabilistic boosting tree classifier. This
system is able to perform effective segmentation of fetal abdomens.
A new class of nonlinear filter, the Combinational Non-Linear Mean Median
filter (CNLMM) filter as well as image detection using Ada Boost, The MH boosting
algorithm and probabilistic boosting tree were also proven efficient.
Semi-Automatic Detection and Measurement of Fetal Parameters … 395

Table 1 Details of
S. No Trimester scans Fetal parameters
parameters for all trimesters
1 First trimester Crown Rump Length
2 Second trimester BPD, HC, AC, HL and FL
3 Third trimester BPD, HC, AC, FL an HL

The detailed study of literature in the area of fetal parameters measurement and
medical image processing techniques and algorithms, especially the problem asso-
ciated with the automatic segmentation and feature extraction for clear visualization
issues.
The points inferred from results of literature review are:
• Loss of traits and application in outcomes of the measurement systems
• Fetal anatomy measurement requires manual intervention
• Long computing time for various parameter measurements.
This procedure is followed for all the six parameters and the parameters measured
during each trimester are enlisted below (Table 1).

2 Proposed Research Work

The proposed system for the measurement of fetal parameters such as BPD, AC, HL,
FL, HC and CRL is based on two processes namely semi-automatic and automatic
process which is performed with the help of the Image processing criterion in the
LabVIEW software known as VISION ASSISTANT.
The semi-automatic process is wherein the parameters are segregated into six
SUB VI’s and are programmed into one single MAIN VI. The trimester fetal scan
image is loaded and is converted into 1000 × 1000 resolution. This is then followed
by manually dragging the region of interest ROI which could be either circle or line
annotation or the values of the result is obtained.
The automatic process is being dealt only for four parameters namely AC, HC,
HL and FL. The trimester fetal scan image is loaded and by applying certain
threshold levels and performing convolution as well as morphological operations
using different icons such as contour extract, the search direction and the contour
selection are used to locate the images from left to right, top to bottom, maximum
length, etc. for automatic process. Here, it selects the appropriate ROI and the values
of the result are obtained.
396 S. Prabakar et al.

3 Semi-Automatic Fetal Parameter Measurement System

3.1 Process Involved in Semi-Automatic System

The functional blocks of semi-automatic fetal parameter measurement system is

shown in Fig. 1. The six parameters are divided into six different SUB-VI’s which
runs only when that particular parameter is selected. All these SUB-VI’s are together
linked into a single MAIN-VI. The three trimesters were studied to know which
parameter is measured during each trimester. The data is usually an ultrasound image
which is in the PNG form and is read and displayed using the IMAQ tool.
The while loop is programmed and allows the user to wait for drawing the ROI and
the invoke node is given to the input which stops the loop once the user uses the ROI.
It is because the code is done only for the ‘DRAW’ event. Once them ROI is drawn,
the ROI’s coordinates (pixel indices) will be sent out through the ‘COORDINATE’
terminal as an array. If the ROI is not drawn, it will return an array without any
elements. Once the ROI is drawn, the array with the coordinates of the ROI will be
sent. This stops the loop, if the coordinates array is not empty, which means the ROI
is drawn.
The property node is used for the processed input to read and write the properties
of reference. This is used to set the properties based on the applications. The ROI is

Fig. 1 Block diagram for semi-automatic system

Semi-Automatic Detection and Measurement of Fetal Parameters … 397

converted into a mask. A mask is nothing but a group of pixels with the programmer
specified value in it. Here the value is 255. The image is converted into an array with
the array elements corresponding to the pixel’s colour value/intensity. The output
will be a 2-D array of rows and columns pixel colour values of the input image.
Here, it is in the mask or other words, number of pixels in the ROI drawn by the user.
A function called unbundle by name is used. We use the element of global rect-
angle. The output from the global rectangle will be an array of 4 numeric elements
namely; ROI’s left, right, top and bottom coordinates respectively. Certain expres-
sions like negation, addition, subtraction, division, etc. are used. The DBL converts
a double-precision number to floating point number. The linear fit VI is used This
VI uses the iterative general Least Square method and the Levenberg–Marquardt
method to fit experimental data to a straight line of the general form described by
the following equation:

f = ax + b (1)

where x is the input sequence X, a is slope, and b is intercept. The expression node is
used to calculate the expressions that contain a single variable such as acosh, etc. The
calibration factor of 1000 × 1000 is done and multiplied, the sine of the expression
node is then divided to get the result. This is then linked with the main VI.

4 Scope for Automatic Fetal Parameter Measurement

System

The trimester fetal scan image is loaded and by applying certain threshold levels and
performing convolution as well as morphological operations are used to locate the
images from left to right, top to bottom, maximum length, etc., for automatic system.
Here, it selects the appropriate ROI and the values of the result are obtained.

4.1 Process Involved in the Automatic System

The image is created using IMAQ CREATE and IMAQ READ and a specific memory
is given to the image. This is then converted into an array to work with each pixel.
This is enclosed in a nested for loop, in which the outer loop runs according to the
number of rows in the pixels of the array and the inner loop runs for the number
of columns. A threshold is set inside the loop to convert grayscale image to binary
image.
Then the particles are removed from the image for proper segmentation of the
appropriate region from the surrounding particles. If there are large numbers of
particles it would be difficult to draw the region of interest ROI. The number of
398 S. Prabakar et al.

particles or pixels which has been removed from the image is identified with the help
of IMAQ PARTICLE ANALYSIS report. A filter of the nth order is used in which
the X axis coordinate is −1 and Y axis coordinate is 5. For every value in the X axis,
it is multiplied with 5 to define the organ of measurement as they are stronger in
densities. This filtering operation is done only for giving intensity for the appropriate
region.
Morphological operations such as dilation, erosion, opening and closing are
performed in the vision assistant tool of LabVIEW. Dilation is a transformation
that produces an image that is as the same shape as the original, but of different
size. Dilation stretches or shrinks the original figure. It increases the valleys and
enlarges the width of the maximum regions. Erosion is used to reduce the objects
in the image and known that erosion reduces the peaks and enlarges the widths of
minimum regions. Opening of an image is erosion followed by dilation with the same
structuring element and closing operation is the reverse of opening operation.
The forthcoming operation is performed to select the ROI automatically. Hence,
we use the IMAQ Extract Contour icon which extracts a single, best contour from an
image. If the input image is calibrated, the contour will be learned with calibration.

5 Result of Proposed Semi-Automatic Fetal Measurement

System

The main VI is the composition of all the six SUB-VI’s. The user or the sonographer
can select the type of trimester and perform the semi-automatic measurement. On
selecting the appropriate trimester, the parameters that are measured only remains
enabled while the other parameters are disabled. The main VI is displayed in Fig. 2.
The ultrasound scan image is loaded in the file path and the window shows two
sets of images. The first image shows the original image and the second image shows
the conversion of the original image to 1000 × 1000 resolution as shown in Fig. 3.
This is set as the standard resolution for all parameters for greater throughput and
accuracy.
In this semi-automatic process, the region of interest is drawn manually either
by using circular annotation or line annotation. The region of interest ROI of BPD,
HC and AC are drawn using circular annotation while the ROI for CRL, HL and
FL are drawn using line annotation. This ROI depends on the calibration factor. The
calibration factor for 1 pixel in the 1000 × 1000 resolution image is calculated as
0.016129 cm. The image is then converted back to array using Build array icon. This
is followed by the important mode where the values are then converted from degrees
to radians which are followed by other mathematical operations and the required
value for the result is obtained. The output of the ROI drawn manually is shown in
Fig. 4.
Semi-Automatic Detection and Measurement of Fetal Parameters … 399

Fig. 2 Main VI for semi-automatic system

Fig. 3 Conversion of resolution to 1000 × 1000

6 Conclusion

Thus, the fetal parameters are measured manually and are similar to the process
carried out by the sonographers and also LabVIEW being highly effective software
with very minimal time duration of 5 min per parameter. The manual intervention
drawbacks in this work is analyzed and hence to overcome these we proceed with
400 S. Prabakar et al.

Fig. 4 Semi-automatic ROI

estimation and measurement
of CRL

the automatic measurement system. The characteristics of different technologies

have been reviewed. This proposed method is an efficient process of measuring the
fetal parameters by semi-automatic system and further automating the measurement
system. The work contains 50 different individuals 120 ultrasound images per person
that cover various weeks of the trimester scans. The subjective appearance of the
output images are found good and the performance of the proposed technique are
good since they marginally overcome the drawbacks and inferences of the existing
ultrasound fetal parameter measurement systems and algorithms.

References

1. Carneiro G, Georgescu B, Good S, Comaniciu D (2008) Detection and measurement of fetal

anatomies from ultrasound images using a constrained probabilistic boosting tree. IEEE Trans
Med Imaging 27(9):1342–1355
2. Priestly Shan B, Madheswaran M (2010) A generalized despeckling filter for enhancing fetal
structures to aid automated obstetric pathologies. Int J Comput Theor Eng 2(3):1793–8201
3. Georgescu B, Good S, Comaniciu D (1983) Automatic fetal measurements in ultrasound using
constrained probabilistic boosting tree. Siemens Corporate Research, USA
4. Shi Z, Fung KB (1994) A comparision of digital speckle filters. In: Proceeding IEEE on IGARSS,
vol 4, pp 2129–2133
5. Schluter PJ, Pritchard G, Gill MA (2004) Ultrasonic fetal size measurements in Brisbane,
Australia, vol 48(4). Department of Radiology, pp 480–486
6. Loughna P, Chitty L, Chudleigh TET (1994) Fetal size and dating: charts recommended for
clinical obstetric practice. Academic Division of Obstetrics and Gynaecology, Nottingham
University
7. Jardim Hussain S, Figueiredo M (2005) Segmentation of fetal ultrasound images. Ultrasound
Med Biol 31(2):243–250
8. Tang S, Chen S-P (1969) A fast automatic recognition and location algorithm for fetal genital
organs in ultrasound images
A Framework for Enhancing Network
Lifetime in Internet of Things
Environment Using Clustering
Formation

Abdulrahman Aminu Ghali, Rohiza Ahmad, and Hitham Alhussian

Abstract Internet of Things (IoT) is one of the current technology that gains the
highest acceptance from various industries and academia nowadays. The growth
of this technology has streamlined the day-to-day activities of individuals. It can
be stated that day-to-day activities in the industries without the aid of the IoT is
becoming harder and harder due to the manual process. This study investigates the
issues of energy consumption in the IoT environment due to denial of service (DoS)
attack in the IoT environment. Besides, a solution that remedies the issues is proposed
in this study. The contributions of this study are categorized into two aspects. Firstly,
the study described the components of the IoT, which includes the sensor nodes.
Secondly, the study emphasizes in improving the energy of the nodes to enhance
the network lifetime in the IoT environment. To achieve the milestone the study was
implemented using MATLAB, and thereby considering the enabling technology such
as wireless sensor network (WSN) and the RFID in the perception layer of IoT. The
results reveal that the network lifetime has improved significantly with about 6%
improvements as compared with security-based modified LEACH (MS-LEACH).

Keywords IoT · Nodes · Network · WSN · Sensors · RFID and DoS

1 Introduction

Internet of things (IoT) is a collection of various objects (devices) that are intercon-
nected to perform a specific task without human endeavor. The connected devices
share data and information in order to achieve the intended goals in IoT application
areas [1]. Besides, IoT has been implemented in many application areas such as oil
and gas, transportation, health care, homes, education, and sports with the aim to
transform the industries and individual’s activities [2].
The widespread of using various technologies such as wireless sensor networks
(WSN) and smart devices that are equipped with several sensors, such as radio

A. A. Ghali (B) · R. Ahmad · H. Alhussian

Computer and Information Sciences Department, Universiti Teknologi PETRONAS, 32610 Seri
Iskandar, Perak Darul Ridzuan, Malaysia

© The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 401
R. Ibrahim et al. (eds.), International Conference on Artificial Intelligence
for Smart Community, Lecture Notes in Electrical Engineering 758,
https://doi.org/10.1007/978-981-16-2183-3_39
402 A. A. Ghali et al.

frequency identification (RFID) and near field communications (NFCs) have

contributed immensely in developing the new concept of IoT technology [3].
To understand the concept of IoT technology, sensors are tiny devices that are
connected wirelessly and coordinate to perform a specific task. Their small and their
ability to support a wide range of applications has rendered the IoT to be one of
the emerging wireless networking technology with extensive deployment across the
globe [4].
A sensor device consists of batteries, a radio antenna, and a board to accommodate
the aforementioned features. The sensor network consists of the ability to identify,
process, communicate, in IoT environment [5]. Therefore, users can communicate
in the environment using this network. Hence, it is easy to communicate with the
physical phenomena with the aid of connected devices and the sensor nodes unlike the
traditional ways. The wireless sensor network consists of tens to thousands of nodes.
These nodes captured information, process, and forward the information collected to
the cluster head (CH) and the CH will aggregate the data and sent to the base station
(BS). Additionally, the nodes have specific features such as short duty cycle, power
limitations, and battery life.
With all the advantages of IoT, on the other hand, it suffers a setback from DoS
attack. Addressing the aforementioned problem and setting up the proper opera-
tion of WSN to increase the lifespan of the network and thereby minimize energy
consumption is one of the main challenges. In the existing literatures varieties of
improvements have been made to minimize the energy consumption in the WSN,
and most of these approaches focusing only on the energy consumption improvement
and left the DoS attacks un attended which allowed the attacks to be escalated in
the IoT perception layer [6]. Therefore, solving the DoS attacks and placing the CH
over the grid and to find how many CHs are there in the cluster will improve the
efficiency of the network and thereby reduces the energy consumption. To present
the method in order to solve the challenges the study proposed a framework using
clustering formation to withstands with the challenges.
The study is organized as follows: Sect. 2 describes the application areas of the
IoT. Section 3 consists of the methodology. Section 4 discusses the results. Section 5
concluded the paper.

2 Application Area of IoT

The following sections will highlight the application area of IoT and its challenges.
These applications areas include transportation, health care, education, homes, sports,
and automobile. Table 1 summarizes the security challenges in the IoT environment
[6, 7].
Based on Table 1 the security challenges of the IoT application areas are summa-
rized with their security challenges and proposed recommendations. Thus, the secu-
rity challenges may caused the degradation of the network performance and commu-
nication denial between the sensor nodes and CHs. Hence, the study provides
A Framework for Enhancing Network Lifetime … 403

Table 1 Summary of the security challenges in the IoT application area of IoT and their solutions
Application area of IoT Security challenges Proposed solutions
Transportation DoS attack Authentication mechanism
Health care DoS and Zingbox Confidentiality and
authentication mechanism
Education DoS attack Authentication and
availability mechanism
Homes DoS and Replay attack Integrity and confidentiality
mechanism
Sports DoS and Snipping attack Communication security
Automobile Man in the middle and DoS attack Mutual authentication and
authorization technique

Fig. 1 Clustering formation arrangement

a framework that will improve the efficiency of the network lifetime in the IoT
environment.

3 Methodology

In this section the study presents the method on how clustering formation is formed.
Therefore, Fig. 1 displays the clustering formation method.
Initially the function name and number of the nodes are required to compute
the clustering formation arrangement. Next, the nodes are distributed based on the
number of scale design. After the nodes distribution, the CH is assigned. The calcu-
lation of the clustering rate is obtained to provide the adequate distribution between
the cluster and the CH (For further details see Sect. 5). The output of the formation
along with the CHs are displayed.

4 Results and Discussion

Based on the method presented, the result of the proposed method for enhancing the
network lifetime in the IoT environment is shown in Table 2. The illustration of the
figures and tables are displayed. Table 2 and Fig. 2 shows the nodes distributions.
404 A. A. Ghali et al.

Table 2 Nodes distributions

No Items Description parameter
1 Simulation area 350 ✕ 500
2 No. of points 350
3 Radius meters 50 by 50
4 Field area X by Y in meters

Fig. 2 Nodes distributions

As shown in Fig. 2, the initial field distributions of the nodes are distributed based
on 50 m ✕ 50 m. The nodes are randomly distributed in the clusters, before the CHs
are selected, the nodes are randomly installed into the scales waiting for the CHs
selection. The CHs can be placed at the center or the edge of the field scale but it
depends on the nodes selected to be the CH. The insertion of the CH in the center or
the edge is convenient to assist the node to finds its transmission range to CH. Table
3 and Fig. 3 shows the advertisement phase.
In Fig. 3 the nodes with the red colors are the two CHs. The CH sent an adver-
tising message to the BS for communication after a few rounds. The nodes usually
transfer their energy to CH in the initial data transmission rounds. Thus, the data are
aggregated and sent to the BS. Most importantly, the CHs uses their energy faster
than non-CH for data transmission. Table 4 and Fig. 4 display clustering formation.

Table 3 Advertisement
No Items Description parameter
phase
1 Simulation area 350 ✕ 500
2 No. of points 350
3 Radius meters 50 by 50
4 Field area X by Y in meters
5 Estimated no. of clusters 2
A Framework for Enhancing Network Lifetime … 405

Fig. 3 Advertisement phase

Table 4 Clustering
No Items Description parameter
formation
1 Simulation area 350 ✕ 500
2 No. of points 350
3 Radius meters 50 by 50
4 Field area X by Y in meters
5 Estimated no. of clusters 2

Cluster Head 9

Cluster Head 3

Fig. 4 Clustering formation arrangement

Figure 4 indicates the clustering formation for the nodes in the cluster. In this
figure, the formation are divided into two which are brown and blue. The brown color
indicates the clustering formation with a circle node representing its cluster group
as the CH. Therefore, having the formation in the cluster will reduce the number
of energy consumption from the nodes significantly. The more energy consump-
tion is reduce the more network lifetime increased. Figure 5 indicates the energy
consumption per nodes in the cluster.
406 A. A. Ghali et al.

Fig. 5 Energy consumption per node in the cluster

From Fig. 5, it is evident that the energy dissipation by the nodes in the cluster
is minimal. Among all the nodes, node nine (9), which served as the CH consumed
the highest amount of energy since all the remaining nodes in the cluster transmit
their data to the cluster head. The next highest node is (3), which is also the CH. This
signifies that the other nodes transmit their data to the CHs and this allow them to
have a rest in order to save their energy and by reserving their energy the network
performance increase significantly.

5 Conclusion and Future Work

This study investigated the issues of energy consumption in the IoT environment
and further provides solutions that remedy the challenges. Besides, the IoT security
challenges specifically the DoS attacks can be addressed from the early stage of
the IoT implementation. The contributions of this study are categorized into two
aspects. Firstly, the study described the initial components of the IoT, which includes
the sensor nodes. Secondly, the study provides a solution that improve the network
lifetime in the IoT environment using a method clustering formation. The simulation
results revealed that network lifetime is increase with 6% improvement as compared
MS-LEACH.

Acknowledgements The research was fully funded by the Center of Graduate Studies (CGS).
Author completely acknowledges the financial support given by Universiti Teknologi PETRONAS
(UTP) for this study.
A Framework for Enhancing Network Lifetime … 407

References

1. Mahmoud R, Yousuf T, Aloul F, Zualkernan I (2015) Internet of things (IoT) security: current
status, challenges and prospective measures. In: 2015 10th international conference for internet
technology and secured transactions (ICITST). IEEE, pp 336–341
2. Asghari P, Rahmani AM, Javadi HHS (2019) Internet of Things applications: a systematic review.
Comput Netw 148:241–261
3. Dehkordi SA, Farajzadeh K, Rezazadeh J, Farahbakhsh R, Sandrasegaran K, Dehkordi MA
(2020) A survey on data aggregation techniques in IoT sensor networks. Wirel Netw 26(2):1243–
1263
4. Ghali AA, Ahmad R, Alhussian H (2021) A framework for mitigating DDoS and DOS attacks
in IoT environment using hybrid approach. Electronics 10(11): 1282
5. Xia Y, Li W, Clark W, Hart D, Zhuang Q, Zhang Z (2019) Entangled radiofrequency-photonic
sensor network. arXiv preprint arXiv:1910.08825
6. Ghali AA, Ahmad R, Alhussian HSA (2020) Comparative analysis of DoS and DDoS attacks in
Internet of Things environment. In: 9th computer science on-line conference. Springer, Cham,
pp 183–194
7. Kouicem DE, Bouabdallah A, Lakhlef H (2018) Internet of things security: a top-down survey.
Comput Netw 141:199–221
An Empirical Analysis on Big Analytics
for e-Healthcare and Agriculture

N. Purandhar and S. Ayyasamy

Abstract There is a lot being said and done in the field of data analytics. Using large
amounts of data for analytics has become one of the rising trends in the business
world but, implementing this business intelligence into different sectors of govern-
ment hasn’t still progressed well. We have discussed two major applications of data
analytics in government sectors where the government and eventually the citizens
could benefit from all the available big data. The applications include (i) Agriculture,
where the big data analytics could result into better crop planning, yield analysis,
improved soil health and irrigation as well as reduce the support cost incurred. (ii) The
section on data analytics in healthcare mainly points out the importance of predic-
tive analytics in improving personalized healthcare and healthcare infrastructure as a
whole. It also talks about how the government can unlock value through big data and
machine learning to provide better health insurance than the existing ones and how
data analytics is helping with fraud detection while providing the health insurances.

Keywords Big data · Data analytics · Recommendation engine · Predictive

analytics

1 Introduction

Executive Vice President of Gartner, which is considered to be the world’s important

organization in Computer science and Technology Research and Advisory stated,
“Information is the oil of the 21st century and analytics is the combustion engine”
[1]. We need data analytics in e-governance in order to optimize performance in all
the fields that come under the roof of governance. Data analytics uses historical data
for providing intelligent solutions in order to make better decisions and strategies for
the betterment of the society. Therefore, the field of data analytics is not just relevant
but also very important for e-governance. The complete process of Data Analytics
insinuates the data collection, extraction of information from that data, transformation

N. Purandhar (B) · S. Ayyasamy

Department of CSE, Dr. N.G.P. Institute of Technology, Coimbatore, Tamil Nadu, India

© The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 409
R. Ibrahim et al. (eds.), International Conference on Artificial Intelligence
for Smart Community, Lecture Notes in Electrical Engineering 758,
https://doi.org/10.1007/978-981-16-2183-3_40
410 N. Purandhar and S. Ayyasamy

of data, analyzing and interpretation of the data and reporting. Monetizing this data,
i.e. creation of data to wealth is a challenge as well as an opportunity for the future.
For different objectives different types of analytics is to be carried out. For which we
need to classify data analytics.
The IBM organization in 2013 categorized analytics into three important ways
like descriptive, predictive and prescriptive analytics. To which in 2015 diagnostic
analytics was added.
A. Descriptive Analytics: Data mining techniques are used along with business
intelligence to inquire “What has happened?” That means it analyses the histor-
ical data for insights and uses that data to determine the future [2]. Though it
provides insights into historical data and trends, this technique is most time-
intensive and produces the least value of all the types. Therefore, descriptive
analytics alone is not used very often. But, combining it with any other form
of analytics can be practical and productive.
B. Diagnostic Analytics: It is a form of sophisticated analytics which examines
data or information to give solutions for the question “Why did it happen?”. It
uses historical data with other data to identify patterns and find dependencies
[3]. Diagnostic analytics is used for root cause analysis of a problem.
C. Predictive Analytics: Statistical models are used to forecast “What could
happen?” [2]. It uses the results of descriptive and diagnostic analytics to find
out patterns, clusters and outliers to estimate future trends. Predictive analytics
is widely used because of its numerous advantages but, it still depends largely
on the data quality and therefore it requires continuous optimization.
D. Prescriptive Analytics: Methods for optimization and simulations are used
to ponder “What should we do?” [2]. This is the action-oriented type and
focuses on taking any endeavors possible to deal with the problem at hand.
This advanced type of data analytics needs historical data coupled with infor-
mation from external sources for applying statistical algorithms. Other than that,
prescriptive analytics uses machine learning as well as business intelligence,
which makes it expensive to implement and manage [4].

2 Big Data and e-Governance

On Feb 1st 2018, while delivering his speech on the Union Budget, the Finance
Minister told the parliament that the government’s think tank—Niti Aayog, would
initiate a national program helping the government to focus on the development of
cutting edge technologies in digital space namely Big Data, Artificial Intelligence
(AI), Machine Learning, Internet of Things (IoT), 3D printing and others. Further
accentuating to invest in research, training and skill development in robotics, AI,
manufacturing, big data analysis and IoT, he announced to double the budget for
Digital India Programme to 3073 crores in 2018–19 [5]. This shows that the govern-
ment of India is keen on using the technological advancements in different fields for
good.
An Empirical Analysis on Big Analytics for e-Healthcare … 411

2.1 Big Data in Agriculture

Farming is the cornerstone of India’s wealth and this agriculture is responsible for
18% of India’s gross domestic product (GDP) and facilitates employment to more
than 50% of the countries’ manpower but, because of rapid development of the
industrial and service sectors the agricultural industry has declined to a great extent.
For improving the yield of small farmers, the government of India had an agreement
with IBM to develop an Artificial Intelligence based model for crop yield prediction
[6]. When IBM will employ Artificial Intelligence for developing the technological
model for improvement of agricultural output and productivity of various types of
crops and soil species for the specified districts, NITI will use this data insights
generated through these AI models to help farmers and other stakeholders [7]. Also,
by providing soil health cards the government has empowered the farmers to better
analyze the requirements of the water for irrigation and the crop to be sown.
To add value chain and increase profitability of the agricultural industry the
government is keen on employing technology. Also, for creating opportunities in
the field, the Ministry of Agriculture and Farmers ”Welfare had a deal signed with
the Indian Space Research organization (ISRO) to use satellites for geotagging the
agricultural assets [8]. The big data acquired through such agricultural infrastructure
will help develop vigorous responses for the issues pertaining to climate change, pest
and disease outbreaks, and land degradation. Applying these “big data” approaches
by the government for agriculture guarantees to search out new ways to build a
food-secure society.
Big data is seen as a prospective technology for selection of appropriate agri-
cultural inputs but, acquiring data is the biggest challenge. For any agribusiness
increasing the profitability means increased yield and decreased support costs. Big
data aids in enhanced yield analysis. Increased yield and decreased support costs
play principal role to gaining profitability and enhanced customer experience for
any small or large agribusiness. Big data analytics can enable improved analysis of
crop yield and other crucial methods for a profitable and in-depth root-cause analysis
resulting in actions for enhanced crop quality and reduced overall cost [9]. The data
associated with turnover, utilization of available resources and the effectiveness of
equipment, can be combined with further analysis for improved quality.
Predictive analytics may be used to foresee the demand for seeds, requirement of
fertilizers and animal feed needs, for empowering the agribusiness supplier to take
appropriate steps to match production to demand. Table 1 shows different fields of
agriculture and how big data is used for their management.
412 N. Purandhar and S. Ayyasamy

Table 1 Big data approaches in agriculture

Focus area Ref Proposed Possible improvements
techniques/solutions
Soil health [10] Geocoded soil data is Soil health cards are provided
generated using soil genome by the government of India to
sequencing and this data is all the farmers. They provide
characterized based on a comprehensive report of the
microbial diversity, further soil health [11]. The data
it is combined with AI to collected through the soil
estimate the soil health health cards could be
combining the ground level modeled with a ML
and above ground features algorithm to identify steps to
to describe soil microbial improve soil quality
diversity
[12] The soil data is compiled
based on different attributes
such as fertility, pH,
Nitrogen, phosphorous,
potassium contents and
others. Chi-square Map
Reduce method is used for
feature selection. An
Advanced Decision Tree
classifier is suggested for
pruning methods to build a
tree
[13] Information gain, Gain
Ratio and Chi Square are
calculated for feature
selection of the soil contents
using Map Reduce
Crop planning [14] Parameters like year and Farmers will be provided with
crop type are used in the soil health records, they can
clustering algorithms along study the soil management
with special attributes such techniques and accordingly,
as time or crop type to plan the upcoming crops [11]
identify which crop is to be
sown
The state of Andhra Pradesh This model could be
Productive irrigation is using the satellite data for implemented at various
predictive analytics to places in India where the
forecast water-stressed crops are needed to be rain-
farms and based on the crop fed. Also, innovative methods
type use rain guns as a for irrigation are needed to be
productive irrigation developed
technique
(continued)
An Empirical Analysis on Big Analytics for e-Healthcare … 413

Table 1 (continued)
Focus area Ref Proposed Possible improvements
techniques/solutions
Reduced support cost Many small and marginal To reduce cost of entire
farmers are utilizing agriculture value-chain the
M-trading platforms to sell farmers should be enabled
their farm produces right with data centric methods for
away to businesses across improving the efficiency of
the country through smart supply-chain and the
phones marketplace of agrarian items
Yield analysis [15] Different data mining With the soil health card, the
[16] techniques like linear and farmers can know the nutrient
multiple regression, contents of the soil better and
[17] clustering, decision tree and hence, get an idea of which
k-nearest neighbors have crops to invest in and what
been used with the big kind of fertilizers they need
datasets for finding the yield ultimately, increasing the
of the crop crop yield [11]. IBM”s AI
based model for crop yield
prediction is also in use by
small farmers
Crop protection and weed [18] Big data as well as machine There are no crop protection
control learning approaches have initiatives by the government
been suggested for available as yet
modeling and prediction of
herbicide/pesticide
resistance, detection of
non-native species plants
and flowers, decision
support system for crop
protection and using AI
based systems for plant and
flower control

2.2 Big Data in Healthcare

A real-time alerting system for instant care through wearable devices is fast growing
trend in the first world countries [19]. The world is rapidly moving from tradi-
tional medical practices to wearable medical devices but, India is still struggling for
providing primary healthcare to majority of population. Despite of the government
recommendations for standards of Electronic Medical Records (EMR), there are
no concrete steps taken for centralizing EMRs. The Ministry of Health and Family
Welfare suggests a standard for structured record fields to be inserted into the database
viz., Name of the patient, DoB, Residential addresses, Doctor’s name, Hospital name
and addresses, and medical treatment reimbursement codes for easier management
through a database system but, the lack of acceptance to field-code the data and
use of EMRs by the medical practitioners and caregivers is a major barrier [20].
414 N. Purandhar and S. Ayyasamy

Moreover, the traditional databases are inadequate to handle the large chunks of data
being produced every day. There are three dimensions to the role of government in
Healthcare in India.

2.3 Using Big Data and Predictive Analytics to Improve

Health Care by Selecting the Best Treatment Plans
and Improving Diagnostics for Personalized Healthcare

The major shortcoming of the government is its lack of insight for providing devel-
opment in personalized healthcare sector. The main focus needs to be on research and
development models for epidemiology research clubbed with big data. Big data can
help to identify problem areas where there is a need for improvement in personalized
healthcare. It can also be useful for statistically analyzing the course of treatment
employing the historical data. Whether it be biomedical sciences or pharmaceutical
researches data is everywhere. One-way Predictive analytics can help hospitals is
anticipate the number of incoming patients at different times during the day, week,
month and year. This can solve the difficulty caused by over-staffing or under-staffing
in hospitals because in the first scenario the cost of labor increases and in the latter the
patient services might be hampered. The other area where the predictive analytics can
help is by using data for identifying the cause of different diseases through available
data.

2.4 Using Big Data and Predictive Analytics to Improve

Health Care Infrastructure

The annual National Health Profile document; a publication of Ministry of Health

and Family welfare has six indicators for measuring the fundamental health informa-
tion. They are Demographic, Socio-Economic, Health Status, Health Finance, Health
Infrastructure and Human Resources [21]. Furthermore, the measures for Health
Infrastructure are categorized as educational infrastructure and service infrastructure.
The educational infrastructure lists out number of medical colleges, students enrolled
for diploma, undergraduate and postgraduate medical courses, as well as students
enrolled for dental schools, AYUSH Institutes, nursing courses and paramedical
courses. Health information indicators for Service infrastructure include details of
allopathic hospitals, number of beds in a hospital, Ayurvedic and Homeopathy hospi-
tals, sub centers, primary and community health centers, blood banks, mental health-
care hospitals and eye banks [21]. According to CRISIL”s opinion report, one sub-
centre for the population of 5000 in plain areas and for 3000 population in desert,
tribal and hilly areas is a norm. But, there is still a gap between the planned structure
and the present structure of the health and wellness sub-centers. Big data is central
An Empirical Analysis on Big Analytics for e-Healthcare … 415

to achieve this goal, by statistically measuring the population density, understanding

the demographics and thus making a predictive analysis about where a sub-centre
can be setup in a cost effective, efficient and impactful manner. The analysis will
also need to focus on the density of public and private hospitals in urban, sub-urban
and rural areas.

2.5 Using Big Data for Provisioning Health Insurances

Providing healthcare at affordable prices is one of the main objectives of the govern-
ment in India. Under the initiative of Ayushman Bharat, the government of India
is firm on contributing to better healthcare for all. The Pradhan Mantri Jan Arogya
Yojana under Ayushman Bharat aims to cater to more than 50 crore beneficiaries.
It covers not less than 10.74 crore underprivileged families from rural areas and
workers’ families from urban and suburban parts of the country as per 2011 Socio-
Economic Caste Census (SECC) data [23]. It has a provision of health cover of
rupees 5 lakhs for secondary and tertiary healthcare services. This scheme has a lot
to offer to private hospitals, insurance companies as well as the main stakeholders—
the beneficiaries. According to media reports of 13th Feb, 2019, there have been
discrepancies in some of the procedures, and a few hospitals receiving a greater
number of patients than the other. Therefore, the government has employed analytics
methods to ensure Ayushman Bharat remains without any scams. The analytics is
also necessary for detecting fraudulent behavior as the Aadhar card—the 12-digit
unique identity number were not mandatory for availing the services and any identity
cards were accepted in the initial stages. Utilizing the research evidence Recommen-
dation engines can be built based on collaborative filtering for creating healthcare
plans for patients with similar symptoms. These recommendation engines use best-fit
classification algorithms to classify patients based on different parameters like, age,
disease type, symptoms and seasonal changes among many. This would help to build
a better prediction model for precision medicine. The most commonly used strategies
for predictive analytics include Bayesian networks, Decision Tree learning, Markov
and Monte Carlo Simulations [24]. In the paper [25], the authors suggested a combi-
nation of Principal Component Analysis and Neural Network for sentiment analysis.
They had promising results when checked with different performance parameters.
A feedback mechanism can be developed, where the patient or the caregivers can
provide feedback on different parameters relating to the overall experience at the
hospital. A mechanism can be implemented by applying sentiment analysis on the
feedbacks, where (1) this analysis can be provided to the recommender systems for
personalized healthcare plans, (2) the hospital can utilize the data to improve the
infrastructure and (3) it can help to improve the efficiency of the nursing team by
providing useful insights.
416 N. Purandhar and S. Ayyasamy

3 Conclusion

The applications discussed in this paper specify the potential of data analytics in their
respective fields and the possible solutions using different methods of data analytics.
It presents the sources from which the data is acquired for various applications, and
how this data can help to improve the citizen services provided by the government.
In Agriculture domain the government does the soil analysis of each farm and gives
suggestions to the farmers for crop planning, soil health and yield analysis. But this is
not enough for entire growth of the agrarian market and therefore, the paper discusses
different big data analytics methods and machine learning algorithms for an inclusive
agri-business setup. Healthcare is the sector where data analytics could do wonders
with personalized medical plans, for infrastructure development and patient service
improvement. Recommendation engines could be built for personalized medicine
based on EMR’s. But, first there is a need to develop and deploy centralized electronic
health records systems. Once these systems are in place the data analysis could be
done to increase efficiency and save lives.

References

1. Sondergaard P (2019) Information is the oil of the 21st century, and analytics is the combustion
engine. Gartner Research, Datasciencecentral.com
2. Delen D, Demirkan H (2013) Data, information and analytics as services. Decis Supp Syst
55(1):359–363. Available https://doi.org/10.1016/j.dss.2012.05.044
3. Fleckenstein M, Fellows L (2018) Modern data strategy, 1st ed. Springer, Cham, Switzerland,
p 133. ISBN 9783319689937
4. 4 types of data analytics to improve decision-making. Scnsoft.com, 2019. [Online]. Available
https://www.scnsoft.com/blog/4-types-of-data-analytics
5. Ministry of Finance (2018) Budget speech 2018–19
6. Misal D (2019) How big data is the game changer for indian government in E-governance.
Analytics India Magazine
7. D’Monte L (2019) To improve weather forecasting for farmers in India, IBM is relying on AI.
https://www.livemint.com
8. Deoras S (2019) Top use cases where modi government used big data, AI for reform. Analytics
India Magazine
9. (2016) Improving manufacturing performance with big data. Oracle Enterprise Architecture
White Paper
10. Lesueur D, Burra D, Bui D, Nguyen M, Zhong D (2019) Geocoded soil data and AI to provide
an estimate and indicator for soil health. CGIAR Platform for Big Data in Agriculture
11. Soil Health Card|National Portal of India. India.gov.in, 2015
12. Rajeswari S, Suthendran K (2019) C5.0: advanced decision tree (ADT) classification model
for agricultural data analysis on cloud. Comput Electron Agric 156:530–539
13. Rajeswari S, Suthendran K (2018) Chi-square mapreduce model for agricultural data. J Cyber
Sec Mob 7(1):13–24
14. Sekhar C, Kumar J, Kumar B, Sekhar C (2018) Effective use of big data analytics in crop
planning to increase agriculture production in India. Int J Adv Sci Technol 113:31–40
15. Sellam V, Poovammal E (2016) Prediction of crop yield using regression analysis. Indian J Sci
Technol 9(38)
An Empirical Analysis on Big Analytics for e-Healthcare … 417

16. Fan W, Chong C, Xiaoling G, Hua Y (2015) Prediction of crop yield using Big Data. In: 8th
international symposium on computational intelligence and design
17. Rani S (2017) The impact of data analytics in crop management based on weather conditions.
Int J Eng Technol Sci Res 4(5):299–308
18. Ip R, Ang L, Seng K, Broster J, Pratley J (2018) Big data and machine learning for crop
protection. Comput Electron Agric 151:376–383
19. Firouzi F et al (2018) Internet-of-Things and big data for smarter healthcare: from device to
architecture, applications and analytics. Futur Gener Comput Syst 78:583–586
20. Raghupathi W, Raghupathi V (2014) Big data analytics in healthcare: promise and potential.
Health Inform Sci Syst 2(1)
21. Central Bureau of Health Intelligence (2018) National health profile—2018. Ministry of Health
and Family Welfare, Government of India, New Delhi
22. CRISIL Opinion, Bharat A (2018) Improvement in quality of government infrastructure and
leveraging private sector at right price to be crucial tasks for the scheme. CRISIL
23. Ministry of Family and Welfare, PIB Delhi (2019) Beneficiaries of Ayushman Bharat Yojana
24. Janke A, Overbeek D, Kocher K, Levy P (2016) Exploring the potential of predictive analytics
and big data in emergency care. Ann Emerg Med 67(2):227–236
25. Vinodhini G, Chandrasekaran RM (2014) Sentiment classification using principal compo-
nent analysis based neural network model. In: International conference on information
communication and embedded systems (ICICES2014). https://doi.org/10.1109/icices.2014.
7033961
Enhanced Approach in VANETs
for Avoidance of Collision
with Reinforcement Learning Strategy

A. Ganesh and S. Ayyasamy

Abstract VANET (Vehicular Ad Hoc Network) is the most prominent technology

that helps in the enhancement of performance and safety of the transportation system.
The ad hoc network established between the moving vehicles on the road defines
the VANET. VANET is distinguished by the mobility and self-organization of nodes,
vehicles are considered as nodes here. V2V (Vehicle-to-Vehicle), V2I (Vehicle-to-
Infrastructure) communications use VANET to warn vehicles nearby about emer-
gencies (i.e., traffic jams, accidents), usually of the order of a few meters. Latency,
secure transmission of messages, high-reliability criteria are the key challenges that
one faces when implementing VANET. The primary goal of this paper is to predict
the collision from the sensor readings and prevent the happening of collision and
V2V communication using ESP-NOW protocol. ESP-NOW helps in the transmis-
sion of wireless messages between the vehicle nodes which are controlled by ESP32
microcontroller. Reinforcement Learning (RL) with deep policy gradient networks
is chosen here. Policy gradient (PG) methods are widely used algorithms for RL.
Using Deep Reinforcement Learning (DRL) on the ESP32, the car node maintains
a safe distance from the nearest obstacle. ESP-NOW protocol helps in exchanging
the information between vehicle nodes, by finding all the MAC addresses of vehicle
nodes that are present in its range and communicates the information. The focus of
the proposed system is on taking the appropriate action using the DRL algorithm to
avoid the collision risk and enables V2V communication which helps to broadcast
the vehicle’s information to the other vehicles that are present in the V2V protocol
range to maintain collision-free environment.

Keywords MANET · VANET · Human machine interface · Deep reinforcement

learning · ESP32 microcontroller

A. Ganesh (B) · S. Ayyasamy

Department of CSE, Dr. N.G.P. Institute of Technology, Coimbatore, Tamil Nadu, India

© The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 419
R. Ibrahim et al. (eds.), International Conference on Artificial Intelligence
for Smart Community, Lecture Notes in Electrical Engineering 758,
https://doi.org/10.1007/978-981-16-2183-3_41
420 A. Ganesh and S. Ayyasamy

1 Introduction

VANET is essentially a subset of the Mobile Ad Hoc Network (MANET) [1–3]

established to create a wireless communication system for vehicles. Vehicles now-
a-days are equipped with intelligent and automated systems to work on real-time
safety requirements. VANET’s comprises of V2V, V2I, Vehicle-to-Grid (V2G) and
Vehicle-to-Anything (V2X) modes of communication. Vehicles in VANET are inte-
grated with OBU (On-Board-Unit), RSU (Road-side-Unit). OBU and RSU share
information through the wireless channel. OBU deals with the data transmission
from vehicle to vehicle, whereas RSU communicates the information from vehicles to
road-side network. The OBU talks to RSU and with other OBU via DSRC (Dedicated-
Short-Communication-Range). DSRC enables high speed, low latency V2V, and V2I
communications, using the IEEE 802.11p and WAVE (Wireless Access in Vehic-
ular Environment) standards. ITS (Intelligent Transportation system) is achieved by
employing VANET. ITS facilitates creative traffic management capabilities through
the use of various smart modes of communication and networks. Safe and secure
message transmission is the main challenge VANET face to have a smooth and safe
transportation system [4–6]. High vehicle mobility causes a wide range of prob-
lems, such as the disruption of the interaction between vehicles, in which vehicles
fail to establish safe communication between them. It makes difficult to transmit a
warning message to other devices in prior time before the deadline. Latency plays
a key role in the communication link in sharing alert messages or various physical
parameters. Minimum latency is the crucial requirement in providing the collision-
free environment in VANETs. Different sensors are integrated into the vehicle to
measure various physical parameters like acceleration, speed, and distance from the
nearby vehicle. These parameters are displayed on the Human Machine Interface
(HMI) device which is already connected with the vehicle [13]. Communication of
alert/warning messages, certain physical parameters helps to avoid wrong actions of
the driver which in turn prevents the occurrence of collisions.
The primary objective of VANET is to provide safety applications. In this paper,
V2V communication is shown between the vehicle nodes which are equipped with
ultrasonic sensor and LM393 speed sensor which gives two physical parameters
i.e. distance and speed. The wireless communication is done by employing ESP-
NOW protocol. Two-way communication is chosen here with the help of ESP-NOW
protocol. Training of the DRL model on the ESP32 microcontroller is done here to
avoid the nearest obstacle in the traveling path of a vehicle node [7, 8]. DRL involves
developing a deep learning model that enables function comparison between input
features and expected discounted reward values often referred to as Q values. Q-
learning is a model-free reinforcement learning algorithm to practice policy by asking
an agent what actions to take in what situations. The PG algorithm is used, which is
one of the RL family. PG network typically uses a neural network. The network here
consists of 1 input layer, it carries the distance to the nearest obstacle measured by
the ultrasonic sensor, the output layer corresponds to 5 possible directions at which
the car node can travel.
Enhanced Approach in VANETs for Avoidance … 421

Numerous vehicle applications include efficient traffic control, active road safety,
etc. The main goal is to eliminate car accidents and to provide a collision-free,
driver-safe environment by sharing information on the danger of collisions and obsta-
cles. To extend the driver’s perception by allowing him/her to respond much more
quickly is the fundamental idea. This is possible through the wireless communica-
tion of alert/warning messages. And also, the traffic jam situations are avoided by
maximizing the movement of vehicles. Discovering the nearby vehicles and sharing
the obstacle/hazard information at low latency is the best technique to maintain a
collision-risk free environment.

2 Sensor Modules Used

2.1 Ultrasonic Sensor

An Ultrasonic sensor measures the distance between sensor and object. It has trans-
mitter and receiver, in which transmitter sends an ultrasonic wave, if the wave comes
in to contact with any object, the wave gets reflected back to the sensor receiver. It’s
measuring range is 2–400 cm with the accuracy of 3 mm. Distance is measured using
the formula:
Distance (cm) = Speed * Time Speed = ultrasonic wave speed (0.034 cm/µs)
Time = time traveled by the wave to and from the object (consider half of the
time, time (μs)/2)

2.2 LM393 Speed Sensor

Figure 1 is the IR (Infrared) based LM393 speed measuring sensor. This is used to
measure the wheel rotation speed of a vehicle, the two vertical columns consist of IR

Fig. 1 LM393 Speed

Measuring Sensor
422 A. Ganesh and S. Ayyasamy

LED on one side and phototransistor on the other side. The speed is determined by
wheel encoder which passes in between the two vertical columns. RPM (Revolutions
per Minute) defines the completion of all the 20 slots passage between the two vertical
column slots. This gives the RPM.
From RPM, the speed (cm/s) is measured by: (2/r * rpm)/60, where r (cm) =
Radius of the wheel.

3 ESP-NOW Protocol

ESP-NOW protocol provides a fast, connectionless Wi-Fi communication involving

short packet transmission. ESP-NOW is perfect for remote control systems, sensors,
and other applications. There is a vendor-specific action frame in which the message is
embodied and gets transmitted from one device to another device, the transmission
is connectionless. For the information exchange between devices and to deliver a
broadcast message, adding a peer with the broadcast address is done. Local device
and peer device details are preserved in the low-level layer of ESP-NOW. Sending
and receiving call back functions informs us about the success/failure of packet
transmissions in broadcast/multicast communication.

4 Reinforcement Learning

RL consists of an agent, environment, action, reward, state [9, 10]. The agent performs
the action in an environment which results in the state transition from s to s´. s, s´
is the current state and next state respectively. Interaction of the agent with the
environment with certain actions leads to maximum reward. The combination of
artificial neural networks and RL architecture enables defined agents to determine
the best possible actions in a simulated environment to achieve their goals defines the
DRL. In other words, it combines function approximation and target optimization,
mapping state-action pairs to the predicted rewards.
Figure 2 shows the overview of the RL architecture for this application. The main
task of the agent here is to perform an action, so that vehicle is prevented from
the collision. The sensor readings are observed by the agent to get an idea about the
environment state. After observations from the environment, the agent has to generate
correct actions like moving forward/backward, left/right directions, or stopping the
vehicle, this prevents the vehicle involving in a collision/accident. The agent learns
a policy, which helps in the generation of correct actions from the observations. The
repeated attempts of the agent to have the collision-free environment maximize the
reward. Depending on the observations, actions, and rewards received, the training
algorithm adjusts the policy of the agent. After training, the vehicle will be able to
travel without involving in any collisions using tuned policy and sensor readings.
Enhanced Approach in VANETs for Avoidance … 423

Fig. 2 Depiction of reinforcement learning

4.1 Q-Learning

Q in Q-learning indicates quality. Quality here defines how effectively an action helps
in gaining some reward in the future. Mapping of state-action pairs to rewards is done
by Q. For each state-action pair, a Q-value is present in Q-table. Q-learning employs
Temporal Differences (TD) to determine the value of Qπ (s, a). π is the policy, it
maps the states to actions, where the actions guarantee the maximized reward. s and
a represents the state and action respectively.
Figure 3 depicts the Q-learning algorithm. Q-table has state-action pairs in which
each action is chosen corresponding to the current state. The reward is calculated for
each action and Q-table is updated until the maximum reward is attained.

4.2 Policy Gradient

PG methods are based on the principle “Observe and act” [11, 12]. A policy is trained
here, training makes actions with high rewards more probable or vice versa. PG holds
what is working and dumps what is not. It focuses on optimizing the parameterized
policies with respect to the anticipated cumulative reward by gradient descent. In
this, the agent analyzes the scenario, and instinctively understands what to do with
424 A. Ganesh and S. Ayyasamy

Fig. 3 Q-learning flow chart

it. The more time of training enables the policy to maximize the rewards. PG aims
to effectively optimize policy space. The neural network is used in PG to model the
probability of action directly. When the agent interacts with the environment, we
fine-tune the neural network parameters so that “good” actions are more likely to be
sampled in the future. This process is repeated until the policy network converges
towards an optimum policy.

5 Proposed Collision Avoidance Framework

The collision avoidance with and without DRL is conveyed here. The implementation
starts with the collection of sensor readings from the chosen ultrasonic sensor and
LM393 speed sensor. The sensors are mounted on the two-wheeler chassis which
has 2 DC motors got fixed for the movement of chassis. L298N motor driver is used
which enables to control speed and direction of 2DC motors. All these components
are connected to the ESP32 microcontroller and this setup is taken as a vehicle node
which is shown in Fig. 4, it is controlled by the code written into microcontroller
using the Arduino IDE platform.
Enhanced Approach in VANETs for Avoidance … 425

Fig. 4 Vehicle node setup for collision avoidance

First, let’s see how the simple collision avoidance code performs without any deep
learning technique. For this method, we use a servo motor on which the ultrasonic
sensor is mounted to rotate the ultrasonic sensor in the left and right directions from
the center. A threshold distance value is set for the ultrasonic sensor, if ultrasonic
measured distance > threshold value (i.e. obstacle is at safe distance from the sensor)
vehicle node moves in forward direction. Otherwise (i.e., measured value ≤ threshold
value) vehicle moves backward and checks the distance between sensor and obstacle
in the left and right directions. The sensor compares the distance in both left and
right directions and then the vehicle moves in the left or right direction which has
more space to move.
Collision avoidance with DRL starts with designing of Q-learning framework by
having state-action pairs, discrete states represent the distance range varying from 0
to 400 cms, actions are forward, backward, turn right and left, halting the vehicle.
In this DRL method, servo motor is not used, the ultrasonic sensor is set to a fixed
direction. Training is done in episodes. The PG network consists of an input layer of
dimension 1 × 1. Distance to the nearest obstacle is taken as input, the hidden layer
of 3 × 1 dimension with RELU (Rectified Linear Unit) activation, and the output
layer of dimension 5 × 1, with soft max activation, corresponding to 5 possible
directions at which the vehicle node can travel. PG network samples the 5 actions,
performs them, stores, and updates the gained reward. This sampling process is
encapsulated as the policy π, which is responsible for the implementation of action
dependent on past environmental states {s}. PG agent here specifically determines
an optimal strategy that enhances long-term rewards. By using both Q-learning and
policy gradient, Q-values are determined from the policy preferred actions to which
we apply Q-learning updates.
426 A. Ganesh and S. Ayyasamy

One can see the episode number, the number of actions sampled in the serial
monitor of a particular vehicle node and this Deep reinforcement algorithm helps
in maintaining the safe distance from the nearest obstacle. The vehicle node takes
appropriate action based on the training when it finds an obstacle. In both with and
without DRL algorithm methods, V2V communication is the same. It includes the
transmission of vehicle’s distance from the obstacle, speed of the vehicle which
is sensed by the LM393 speed sensor and, also MAC address of the transmitted
vehicle. These parameters are displayed on the serial monitor of all peer devices
which are present in the ESP-NOW protocol range. ESP-NOW is responsible for
V2V communication.

6 Results and Discussions

Figures 5 and 6 show the vehicle nodes V1 an V2 information display on arduino

serial monitor which depicts the V2V communication. COM3 and COM10 show the
display of V1 and V2 nodes respectively. COM3 showing the MAC address of the
vehicle, the distance, and speed of Vehicle V1, the parameters of Vehicle V2 which
are getting received from the Vehicle node V2 and vice-versa with the COM10. We
chose to broadcast the vehicle’s distance from the obstacle and speed to other vehicles
that are present in the protocol range. With an open field range of up to 220 m, the
exchange of sensor readings happens between two ESP32 boards. ESP-NOW holds
its peers MAC addresses to broadcast the message. Every ESP32 has its own unique
MAC address so that one can recognize every board from which the data is sent and
received using ESP-NOW. Two-way communication is seen in our results, i.e., the

Fig. 5 Vehicle V1 displaying it’s parameters and Receivingv2 parameters through ESP-NOW
protocol
Enhanced Approach in VANETs for Avoidance … 427

Fig. 6 Vehicle V2 displaying it’s parameters and receiving V1 parameters through ESP-NOW
protocol

vehicle V1 information is sent to the vehicle V2, and V2 information is received to

the serial port COM3, vice-versa with the COM10.
The sampling of actions, reward, and episode counter are the terms of the DRL
algorithm. As the training process is achieved in episodes. The episode counter is set,
it increments until it reaches the maximum training value, after that counter resets
to zero. Once the vehicle starts sensing the surrounding environment, if an obstacle
is detected, the algorithm samples all the actions (i.e. forward, backward, stop, turn
left, turn right), and optimized action which gets the highest reward is performed.
The goal is to find an effective action that maximizes the total cumulative reward
(sum of all the rewards received up to this point) of the agent that gets after execution
in every episode.
The chosen algorithm makes the vehicle nodes to travel by having minimum
distance from the obstacle. And ESP-NOW protocol enables in sending and receiving
the information of vehicle nodes which tells the V2V communication. V2V commu-
nication and chosen algorithm allow the vehicle to maintain a safe distance from the
obstacle and also to keep an eye on other vehicle movements that are present in the
protocol range, which enables the driver to drive carefully to ensure collision-free
environment.

7 Conclusion and Future Scope

VANETs are considered to be a significant research area and drawing the attention of
a noticeable number of researches. It is a vast area where there are many challenges
that are to be addressed. The training of ESP32 microcontroller with a deep learning
428 A. Ganesh and S. Ayyasamy

model is accomplished here using DRL. The proposed work is carried out with two
sensors and two vehicle nodes, but V2V communication can be extended by adding
GPS (Global Positioning System) module to the vehicle node setup which gives the
location information of a vehicle and number of vehicle nodes can be increased to
show the broadcasting of information effectively.

References

1. Kumar KV, Jayasankar T, Prabhakaran M, Srinivasan V (2017) Fuzzy logic based efficient
multipath routing for mobile adhoc networks. Appl Math Inf Sci 11(2):449–455. https://doi.
org/10.18576/amis/110213
2. Gopinath S, VinothKumar K, Jayasankar T (2019) Secure location aware routing protocol with
authentication for data integrity. Springer Cluster Comput 22:13609–13618. https://doi.org/10.
1007/s10586-018-2020-7
3. Ejmaa AME, Subramaniam S, Zukarnain ZA, Hanapi ZM (2016) Neighbor-based dynamic
connectivity factor routing protocol for mobile ad hoc network. IEEE Access 4:8053–8064
4. Wang M, Liu D, Zhu L, Xu Y, Wang F (2016) LESPP: lightweight and efficient strong privacy
preserving authentication scheme for secure VANET communication. Computing 98(7):685–
708
5. Chim TW, Yiu SM, Hui LC, Li VO (2011) SPECS: secure and privacy enhancing communi-
cations schemes for VANETs. Ad Hoc Netw 9(2):189–203
6. Bayat M, Barmshoory M, Rahimi M, Aref MR (2015) A secure authentication scheme for
VANETs with batch verification. Wirel Netw 21(5):1733–1743
7. Biswas SB, Iqbal MT (2018) Solar water pumping system control using a low cost ESP32
microcontroller. In: 2018 IEEE Canadian conference on electrical and computer engineering
(CCECE). IEEE, pp 1–5
8. Kushnir V, Koman B, Yuzevych V (2019, September) IoT image recognition system imple-
mentation for blind peoples using ESP32, mobile phone and convolutional neural network.
In: 2019 XIth international scientific and practical conference on electronics and information
technologies (ELIT). IEEE, pp 183–187
9. Wang W, Min M, Xiao L, Chen Y, Dai H (2019, May) Protecting semantic trajectory privacy
for VANET with reinforcement learning. In: ICC 2019–2019 IEEE international conference
on communications (ICC). IEEE, pp 1–5
10. Dai C, Xiao X, Xiao L, Cheng P (2018, December) Reinforcement learning based power control
for VANET broadcast against jamming. In: 2018 IEEE global communications conference
(GLOBECOM). IEEE, pp 1–6
11. Luo Y, Chiu CC, Jaitly N, Sutskever I (2017, March) Learning online alignments with contin-
uous rewards policy gradient. In: 2017 IEEE international conference on acoustics, speech and
signal processing (ICASSP). IEEE, pp 2801–2805
12. Liu S, Zhu Z, Ye N, Guadarrama S, Murphy K (2017) Improved image captioning via policy
gradient optimization of spider. In: Proceedings of the IEEE international conference on
computer vision, pp 873–881
13. Vishnupriya E, Jayasankar T, Venkatesh PM (2015) SDAOR: secure data transmission of
optimum routing protocol in wireless sensor networks for surveillance applications. ARPN J
Eng Appl Sci 10(16):6917–6931
VBlock—Blockchain-Based Traceability
in Medical Products Supply Chain
Management: A Case Study in VietNam

Tuan Khoi Nguyen Huynh, Tuan Anh Dao, Trong Nhan Van Pham,
Khang Hy Vuong Nguyen, Nghi Cong Tran, and Hoang Huong Luong

Abstract The COVID-19 is putting tremendous pressure on the medical facilities

supply, as the demand for facilities has significantly outweighed the production capa-
bility. Several rogue traders have taken advantage of this issue to distribute counterfeit
products. Moreover, some sellers advertise genuine products with unreasonably high
prices. Our team believes that fake or overpriced facilities will significantly compli-
cate the battle against COVID, thereby posing millions of lives to risk. That is why
our team is developing V-Block. V-Block is a supply chain management system that
harnesses the power of Blockchain. Its primary goals are to assist the government
in tracking the product’s distribution process and help customers avoid questionable
deals.

Keywords COVID-19 · V-Block · Blockchain · Supply chain

1 Introduction

The world has been witnessing a tremendous shortage of Personal Protective

Equipment (PPE) and other medical facilities since the beginning of the COVID
pandemic.

T. K. N. Huynh · T. A. Dao · T. N. Van Pham · K. H. V. Nguyen · H. H. Luong (B)

FPT University, Can Tho, Vietnam
T. K. N. Huynh
e-mail: [email protected]
T. A. Dao
e-mail: [email protected]
T. N. Van Pham
e-mail: [email protected]
K. H. V. Nguyen
e-mail: [email protected]
N. C. Tran
National Central University, Taoyuan, Taiwan

© The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 429
R. Ibrahim et al. (eds.), International Conference on Artificial Intelligence
for Smart Community, Lecture Notes in Electrical Engineering 758,
https://doi.org/10.1007/978-981-16-2183-3_42
430 T. K. N. Huynh et al.

WHO estimated that the world needs up to 89 million medical face masks each
month. Having acknowledged the scarcity of medical facilities, many rogue traders,
and sellers distributed counterfeit products to the community and even hospitals
to gain money. These facilities, as their quality is not guaranteed [1, 2], can pose
significant threats to patients. Also, some stores advertise genuine products at unrea-
sonably unreasonable prices. Our team believes that these illegal activities impede
the progress in dealing with COVID. Therefore, they need to be eliminated.
Our team believes that the solution to rogue business is to use a proper supply-
chain management platform. We define a proper supply-chain management platform
as a platform that is: fast, immutable, and decentralized. As for speed, if the whole
distributing process of a particular product is retrieved quickly, we can save much
time. If the government detects a facility’s problem, all entities related to that facility
can be traced within seconds. About immutability, all sensitive data needs to be
tamper-proof, meaning that rogue traders cannot alter what they have uploaded, and
hackers cannot break havoc. We define sensitive data as business related information
such as supply chain information, or companies’ credentials. Such data is critical for
the government to check business legality and products’ origin. Decentralization is
the last characteristic to mention. A decentralized platform does not store data in a
central place, but copies and distributes it across the network. This ensures that all
entities have a copy of the whole database, limiting the impact of a data loss at a
node [3].
Many researchers have suggested the use of Blockchain to develop such a plat-
form. Blockchain is a database system, commonly known as a distributed ledger. It
includes traceable records or blocks that are linked via cryptography. Blockchain is
intrinsically immutable and transparent [4]. Our team agrees with the researchers’
perspective, so we want to develop a medical supply chain management system using
Blockchain, which we name “V-Block”.
In this paper, we will introduce V-Block and explain how it can stop fake and
overpriced products. We will begin by describing some previous work and documents
on implementing Blockchain for the medical supply chain. We will discuss their
strengths as well as their shortcomings. From that analysis, we can suggest our ideas
on improving the implementations. After pointing out our ideas, we will present
our system. There are three main identities in our system: customers, members,
and agencies. Each identity will have distinct functions that suit their needs and
help enforce data safety. Our paper is expected to have three main benefits. First of
all, current Blockchain implementations are analyzed, and their shortcomings are
identified. Additionally, we propose our own system, which includes improvements.
Finally, our system can be used to assist the medical industry in the battle against
COVID-19.
The rest of the paper is organized as follows. Section 2 contains previous
Blockchain applications in various areas, especially in the supply chain. Then, in
Sect. 3, we briefly explain our Blockchain system. We will also get insight into
the system’s functions in Sect. 4 before describing experiments in Sect. 5. We will
summarize our working Sect. 6.
VBlock—Blockchain-Based Traceability in Medical … 431

2 Related Research

Today Blockchain is applied in many areas such as mobile payments [5] introduce a
device that was made up of an Arduino ATMEGA256 board, A TFT-LCD touch
screen, a smartcard socket, and a Bluetooth Low Energy (BLE) module named
Bluetooth Blockchain payment terminal (PBT). PBT is a new generation of open
payment terminals dedicated to Blockchain transactions. A removable secure element
is attached to the BPT to keep the card safe from the internet-connected device. They
inserted a smartcard with a level of security up to EAL5+ to store private keys and
compute signatures to the PBT to make transactions via BLE. Medical [6] this study
mentions some of the benefits of Blockchain in the medical field such as: Firstly,
smart healthcare management. It used a continuously updated decentralized database
to give all medical professionals and doctors access to a patient’s treatment history
with previous doctors as well as disease health information to get more informa-
tion and to have the right treatment for that patient. In addition, the decentralized
database helps to solve problems such as information that is not being updated,
which causes time to verify when needed. Secondly, collect medical information for
research purposes: Build a Blockchain system to help users update their health infor-
mation (heart rate, blood pressure, weight, sleep, used medicine, …). Researchers
can buy that information from users to research and find ways to cure new diseases.
Finally, Blockchain prevents counterfeit medicine: The drug is not in the correct dose
as provided, or the drug is produced in an impure way that can be dangerous to the
patient. Blockchain can be used to provide accurate information about the medicine.
Blockchain can be used to store transaction records in dental care [7] so that partici-
pants can access and control authority. It has the transaction id number generated by
the hash number mechanism provided by the Hyperledger editor. This application
creates a relationship of trust to establish the set of participants in an ecosystem
to share information and interact between clinics so that it can save time and a
variety of costs for the patients. Another application of Blockchain in security is the
recommended Blockchain-based framework (B2MDF) [8] For detecting malicious
apps in the App Store before being downloaded by a user. They proposed using two
private Blockchains inside and outside to form a double private Blockchain to make
the correct decision. Internal Blockchains store feature blocks that are analyzed by
static and dynamic analysis programs to increase malware detection rates. External
Blockchains store detected results as blocks for current versions of the application.
B2MDF shares information with third parties to help security companies deliver anti-
malware. B2MDF does not limit implementation to any particular machine learning
algorithm.
Blockchain plays an important role in the supply chain [9]. We have referenced
research on Blockchain applications in the supply chain “A Blockchain-based Supply
Chain Quality Management Framework” [10]. In this paper, we study the Blockchain-
based SCQI framework for improving supply chain quality management. It consists
of 4 layers as follows: The first layer is the IoT Layer. It uses IoT sensors to collect
Logistics Process Selling Process Manufacture and Quality Inspection Process data.
432 T. K. N. Huynh et al.

The information collected is captured using RFID technology or a bar code. The
second layer is the Data Layer. Blockchains and secure distributed ledgers are in
this layer. The distributed ledger contains the information Quality Data, Logistics
Data, Transaction, Data, Assets Data. The third layer is the Contract Layer. This
layer contains the data to be protected. So, this layer’s data needs to be protected
by means of digital identity access control. The last layer is the Business Layer.
This layer contains the business operations of the enterprise. With Blockchain,
businesses manage product quality in their supply chains more efficiently. “Supply
Chain Management using Blockchain” [11] from this paper, we study how to apply
Blockchain to product tracking from the raw material until it reaches the consumer,
and how to solve problems such as inaccuracies in profile, fraud problem, payment
problem. This study also shows that using Blockchain in product tracking can reduce
the cost, time, and manpower in monitoring and controlling raw materials and prod-
ucts. Besides, this study also suggests the use of smart contracts to handle product
quality problems such as handling compensation when factors are affecting product
quality such as temperature, humidity, not in accordance with smart contract require-
ments. We also looked at the case of Blockchain use in the agricultural supply chain,
“Blockchain-based Traceability in Agri-Food Supply Chain Management: A Prac-
tical Implementation” [12]. This article is about AgriBlockIoT, a fully decentralized
application used to monitor agricultural and food products from seeding to consump-
tion with the desire to monitor to improve food safety and maintain trust and reliability
throughout the supply chain. After reviewing the feasibility study for application in
Vietnam, we realized that AgriBlockIoT has some disadvantages when it comes to an
application as follows: farmers must be trained to store information into the system,
causing feeling constrained for farmers when being monitored by IoT sensors; Only
some common values, such as weight, were measured, but special indicators such as
the number of pesticides that were applied to agricultural products were measured.
To measure the plant protection drug index, specialized sensors are required with a
high cost and not suitable for practical implementation; IoT sensors cannot be widely
equipped because of the high cost of setting and maintaining. That causes false data
when farmers can affect agricultural products in areas without sensors. Assuming
the case of sensors being placed throughout the garden, the farmer can still cheat by
masking the sensors when using a pesticide so that the system recognizes “clean”
and high yielding produce.; The last problem, the sensor fails to measure correctly,
but the data is still within the allowable range and will not be detected and cause an
incorrect data supply. We have looked at the challenges of applying Blockchain to the
supply chain [13]. This article outlines the opinion that supply chain management
(SCM) is the key factor in achieving financial, environmental, and social benefits
in the supply chain industry. But the traditional SCM mechanisms are outdated and
have many weaknesses in information sharing, data retrieval, and reliability in finding
products. These problems can be solved by Blockchain due to Blockchain features
such as immutability, transparency, and decentralization. What have we learned from
this article? These are the requirements of supply chain management when adopting
Blockchain, the technical challenges of applying Blockchain to the supply chain,
VBlock—Blockchain-Based Traceability in Medical … 433

potential opportunities, and principles of supply chain management system design

based on Blockchain.

3 System Overview

V-Block is a system used to store and retrieve information about medical products’
supply chain. It differs from traditional systems in terms of data storage, as it uses
Blockchain technology. This storage implementation is expected to improve data
security, thereby assisting violation detection, and increasing business transparency.
In general, there are three main entities, including customers, manufac-
turers/distributors, and agencies. Manufacturers or distributors will upload data about
the origin and prices of products. Agencies and customers, when in need, can query
data and analyze it for anything questionable. Agencies have the right to view data
without restrictions.
The variety of users, along with different levels of rights and permissions, means
that V-Block requires a layer of security over data access and performed activi-
ties. For this reason, our team develops V-Block to be a permissioned Blockchain.
Permissioned Blockchain is a well-known model, acknowledged by enterprises and
researchers worldwide [14, 15]. The system has three main parts: users, access
control, and data storage.

3.1 Users

There are three kinds of users, namely: customers, members, and agencies. Customers
are the least-privileged identities in the system. They are hospitals or any other
institutions that purchase a medical product and want to check for its authenticity.
Customers are only able to query information about a particular product. Members
are manufacturers and distributors of a product. They are allowed to submit data to
the system and also view their uploaded data. Agencies refer to medical institutions
from the government. They are granted the most rights and permissions. They are
permitted to view any information as well as new changes made to the system. This
enables them to check the validity of a particular item or a manufacturer.

3.2 Access and Data Control

We acknowledge the importance of keeping uploaded data safe and private in supply
chain management [16]. That is why we implement different controls to ensure that
only allowed individuals can access private data. Normal customers do not have to
login or sign up to use our system, as they only want to query information about a
434 T. K. N. Huynh et al.

specific product. However, members must prove their identity before proceeding with
any operations. Upon registration, they first need to fill a form, in which they include
some personal details and business licenses. After this step, they will have to wait to
be checked and approved by agencies. What users can query is also controlled. For
example, only agencies can view all information about the system, while members
can only ask for information about their own products. This is done by implementing
smart contracts for searching and querying data [17]. Additionally, agencies can
create full nodes, but members only have light nodes. Having full nodes means that
a user can save everything on the system into their machines, as opposed to light
nodes, which only stores limited information.

3.3 Data Storage

Any data created will be stored on Blockchain. New data, after verified by peers,
will be kept in a block, which then is broadcasted via multiple network protocols [4].
Data input and output will be handled by different consensus algorithms [18] and
smart contracts. This serves the system’s functions and ensures the validity of data
being stored and submitted.

4 System Functions

The functions of our system can be summarized as below (Fig. 1).

Members (manufacturers and product distributors) first need to register with
the system before interacting with the database. Interaction with the data will be
controlled by smart contracts. When manufacturers produce a product, they will
upload the information about that product to the system. When they sell it to a
distributor or a reseller, that transaction is also uploaded to the Blockchain. By doing
this, the whole supply chain of a facility will be recorded (Fig. 2).
Customers can directly query information about the origin of a product from
our system. If they find some questionable trades, they can report to the agencies.
Agencies are in charge of verifying new members and handling reports from the
customers. They are able to retrieve any necessary information from the system.
Our system is expected to deal with illegal activities using the following methods.
Counterfeit Facilities
The V-Block system aims to combat this issue in two ways: assisting customers in
considering wisely before purchasing and supporting agencies to detect and punish
rogue traders.
Customers, before purchasing, can access our system and have a quick analysis
of the validity of the product. They can submit the code to the website by scanning
the QR code and then receive the distribution process of that product. Therefore,
VBlock—Blockchain-Based Traceability in Medical … 435

Fig. 1 System functions overview

Fig. 2 Chain formation

customers are able to know who the manufacturers are and where the facilities are
produced within a limited amount of time. If they find something suspicious about
the supply chain of any facility, they can report it to the agency. The report includes:
the name of the business, the name and the code of the product. Reports will be kept
away from companies, which means companies cannot know the reporters.
436 T. K. N. Huynh et al.

Fig. 3 Counterfeit facilities

If the agencies are reported about a potentially fake medical facility, they can
use our system as a great tool. The agencies can quickly get all information about
individuals and groups involved in the supply chain thanks to the origin-tracing
feature of V-Block. All they need to do is to submit the reported product’s code and
wait for the result. This will greatly speed up the investigation. Additionally, as all
data stored on our system is tamper-proof, agencies can use it as evidence against
any illegal trades (Fig. 3).
Overpriced facilities
V-Block allows users to analyze fluctuations in products’ prices. Manufacturers and
distributors will upload the product’s new cost after each stage in the supply chain. As
a result, price changes can be recorded and retrieved whenever needed. Agencies and
consumers can query the data and check if a distributor or a seller has unreasonably
increased the cost.
Furthermore, our system attempts to block suspicious cost-related data upload.
If a facility’s cost is raised 20% after only one stage in the supply chain, it will be
flagged, and the upload will be stopped (Fig. 4).
VBlock—Blockchain-Based Traceability in Medical … 437

Fig. 4 Overpriced facilities

5 Experiment

5.1 Introduction

We developed our Blockchain system using Ganache. We tested it on a device using

Ubuntu 16.04LTS, 4-core CPU, 8 GB RAM, and 50 GB HDD. The system was
analyzed in three scenarios: a user checking a product’s information, an agency
checking a product’s information, and a seller trying to upload an unreasonable price
for a product.
In order to start the test, we uploaded data about a product whose name was
‘product1’. It was manufactured by Tank Group. All data about individuals and
organizations involved were unreal and for testing purposes only. We also generated
438 T. K. N. Huynh et al.

one account for a customer, another for an agency, and the remaining one for a seller
(member).

5.2 Tested Scenarios

A user checks a product

In the first scenario, a user considered whether to buy a product. He scanned the QR
code to see who the manufacturer is and when the product was made. Our system then
processed the QR code to return data about the desired product. Besides displaying
basic information, V-Block also showed if the product was verified by the agency.
Other detailed data is hidden because the system knew that the person was just a
customer (he did not sign in) (Figs. 5 and 6).
An agency checks a product
In the second scenario, V-Block processed a query from the agency. In order to
receive the information, the agency needed to log in and then, like customers, had
to scan the QR code. The system returned detailed information about the facility
(Fig. 7).

Fig. 5 Example of QR
VBlock—Blockchain-Based Traceability in Medical … 439

Fig. 6 Example of product which was verified

Fig. 7 Example of showing a product’s information

A seller upload unreasonable price

The last scenario was a special case, as our system had to deal with an illegal attempt.
A seller wants to sell a product at a hefty price (which was 20% higher than permitted).
This will put a great financial strain on consumers, so it should be stopped promptly.
When our system processed such upload, the smart contract detected the violation
and showed a notification. V-Block successfully took the necessary action. The 20%
rate can be modified if required by the agencies (Fig. 8).

6 Conclusion

In this paper, we explained the issue of counterfeit and overpriced medical facilities.
We also described our solution to that issue, which is to develop a supply chain
management system using Blockchain, which we call ‘V-Block.’
440 T. K. N. Huynh et al.

Fig. 8 Example of reject because of the fault price

V-Block system will be the pioneering system for Blockchain application to

support the battle against the COVID pandemic. It helps users ensure medical
integrity, authenticity and provides an easy way to search for essential medical
information of medical manufacturers.

References

1. Cheng M (2001) A guide for the development of medical device regulation. Washington DC
2. Corciovă C, Andritoi D, Ciorap R (2013) Elements of risk assessment in medical equipment.
In: 2013 8th international symposium on advanced topics in electrical engineering (ATEE)
3. Xiao T, Choi TM, Edwin Cheng TC (2018) Pricing and benefit of decentralization for competing
supply chains with fixed costs. IEEE Trans Eng Manag 65(1)
4. Antonopoulos AM (2017) Mastering bitcoin, 2nd ed., June 2017
5. Urien P (2020) High security bare metal Bluetooth Blockchain payment terminal for trusted
ethereum transaction. In: 2020 IEEE 17th annual consumer communications and networking
conference (CCNC)
6. Mettler M (2016) Blockchain technology in healthcare: the revolution starts here. In: 2016 IEEE
18th international conference on e-health networking, applications and services (Healthcom)
7. Wutthikarn R, Hui YG (2018) Prototype of Blockchain in dental care service application
based on Hyperledger composer in Hyperledger fabric framework. In: 2018 22nd international
computer science and engineering conference (ICSEC)
8. Homayoun S, Dehghantanha A, Parizi RM, Choo KKR (2019) A Blockchain-based framework
for detecting malicious mobile applications in app stores. In: 2019 IEEE Canadian conference
of electrical and computer engineering (CCECE)
9. Caro MP, Ali MS, Vecchio M, Giaffreda R (2019) Blockchain-based traceability in agri-food
supply chain management: a practical implementation. In: 2018 IoT vertical and topical summit
on agriculture—Tuscany (IOT Tuscany) Wu H, Cao J, Yang Y, Tung CL, Jiang S, Tang B, Liu Y,
VBlock—Blockchain-Based Traceability in Medical … 441

Wang X, Deng Y (2019) Data management in supply chain using Blockchain: challenges and a
case study. In: 2019 28th international conference on computer communication and networks
(ICCCN)
10. Chen S, Shi R, Ren Z, Yan J, Shi Y, Zhang J (2017) A Blockchain-based supply chain quality
management framework. In 2017 IEEE 14th international conference on e-business engineering
(ICEBE)
11. Bhalerao S, Agarwal S, Borkar S, Anekar S, Kulkarni N, Bhagwat S (2019) Supply chain
management using Blockchain. In: 2019 international conference on intelligent sustainable
systems (ICISS)
12. Caro MP, Ali MS, Vecchio M, Giaffreda R (2018) Blockchain-based traceability in agri-food
supply chain management: a practical implementation. In: 2018 IoT vertical and topical summit
on agriculture—Tuscany (IOT Tuscany)
13. Wu H, Cao J, Yang Y, Tung CL, Jiang S, Tang B, Liu Y, Wang X, Deng Y (2019) Data
management in supply chain using Blockchain: challenges and a case study. In: 2019 28th
international conference on computer communication and networks (ICCCN)
14. Mitani T, Otsuka A (2019) Traceability in permissioned Blockchain. In: 2019 IEEE interna-
tional conference on Blockchain (Blockchain)
15. Islam MdA, Madria S (2019) A permissioned Blockchain based access control system for IOT.
In: 2019 IEEE international conference on Blockchain (Blockchain)
16. Roy A, Gupta AD, Deshmukh SG (2012) Information security in supply chains—a process
framework. In: 2012 IEEE international conference on industrial engineering and engineering
management
17. “SmartContracts”. https://solidity.readthedocs.io/en/v0.7.1/introduction-to-smart-contracts.
html
18. Chaudhry N, Yousaf MM (2018) Consensus algorithms in Blockchain: comparative analysis,
challenges and opportunities. In: 2018 12th international conference on open source systems
and technologies (ICOSST)
Development of Real-Time Internet
of Things (IoT) Based Water Quality
Monitoring System

Huzein Fahmi bin Hawari, Mohamad Nor Syahid bin Mokhtar,

and Sohail Sarang

Abstract This research involves discovering on real-time Internet of Things (IoT)

based water quality monitoring system using three sensor parameters which are
temperature, turbidity and pH. Using this, people can monitor the water quality,
collect all relevant data and information, decide and perform suitable water treatment
immediately. By performing comprehensive data analysis, people can measure water
pollution level referring to Water Quality Index (WQI). Other than that, this project
also involves cloud computing storing all data transferred remotely from the device
at anytime and anywhere. Furthermore, this project applies good power management
method to reduce power consumption and extend the battery life of the device within
a single charge for up to 31 days. The system also able to be deployed at 7 different
location in UTP and able to clarify monitoring data based on WQI. On the other
hand, mobile application is developed to monitor relevant data. Lastly, this project
has high reliability and practicality as it is applied with comprehensive techniques for
every step used and good response time to transmit data for water quality monitoring
in real-time.

Keywords Internet of Things · Water quality index · Pollution index

1 Introduction

The water quality status of rivers and lakes in Malaysia has always been a reason
of concern for various local authorities, government agencies and the public at large
scale. Water quality monitoring system is important for various applications that
requires a real time monitoring to determine the contents and substances present
inside the water. According to Departmental of Environmental (DOE) Malaysia, it is

H. F. Hawari (B) · M. N. S. Mokhtar

Department of Electrical and Electronic Engineering Department, Universiti Teknologi
PETRONAS, Bandar Seri Iskandar, Tronoh, Malaysia
S. Sarang
Faculty of Technical Sciences, University of Novi Sad, Trg Dositeja Obradovi´ca 6, 21000 Novi
Sad, Serbia

© The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 443
R. Ibrahim et al. (eds.), International Conference on Artificial Intelligence
for Smart Community, Lecture Notes in Electrical Engineering 758,
https://doi.org/10.1007/978-981-16-2183-3_43
444 H. F. Hawari et al.

stated that two main parameters are most likely to affect the river water quality that
are urbanization and development [1]. WQI is a method or a tool used to measure
the pollution level of water using SI calculation whereas National Water Quality
Standard (NWQS) is the national standards used to indicate the level of pollution
after measuring the water quality using WQI [2, 3].
IoT is defined as internetworking physical objects embedded with electronics,
sensors, network connectivity and actuators which can be used to collect and
exchange the data [4–6]. IoT can connect billions of devices over the network which
will eventually create much more need for data storage and much larger traffic to
control the connectivity of those devices using TCP/IP protocol stack to communicate
between network hosts for today’s Internet [7].
In this proposed IoT system, users can monitor the quality of water parameters
in real time using mobile application developed using suitable software and alert the
user for water quality alarm system. This alert system is important for the user to
alert them when the value of certain parameters exceed the threshold set. According
to UN Water, it is stated that around 780 million of people did not manage to get
safe and clean water while around 2.5 billion of people do not have suitable adequate
clean drinking water provided by UN-Water [8].
The problem with existing technology is water quality monitoring involves the
manual intervention of collecting water sample in different locations. This is followed
by further laboratory analytical techniques to characterizes the water quality based
on measurement parameters of WQI. This manual approach takes longer time and
is considered as less efficient for water quality monitoring. It is also not considered
as real-time as the person responsible must take samples from particular locations
from time to time.
Therefore, this project is implemented to develop and implement a real-time appli-
cation based on IoT for water quality monitoring system, to integrate water quality
sensors together with IoT communication platform for an efficient real-time water
quality monitoring system and to develop mobile application suitable in analysing
the water quality conveniently.

2 System Design and Architecture

2.1 Central Processing Unit

Arduino Nano is an open-source compact development board contains embedded

electronics parts such as micro-controller, voltage regulator, LED indicators and
power controller. This board is powered up using 7.4 V lithium-polymer battery,
operates at 5 V with 32 KB of flash memory, 16 MHz of clock speed and 19 mA of
current consumption (during active).
Development of Real-Time Internet of Things (IoT) … 445

2.2 Sensor Unit

A sensor unit consists of several sensors integrated together in a single system to

detect the predetermined water quality parameters to indicate water quality level. In
this project, three water quality parameters were selected as sensors unit which are
temperature, turbidity and pH to measure the water quality. Figure 1 shows the block
diagram of sensor unit constructed.
In marine ecosystems, climate changes will eventually affect the aquatic organ-
isms and its surrounding that lead to the quality of water for temperature [9]. Turbidity
is defined as the water cloudiness caused by many types of particles and suitable key
parameter in drinking water analysis and the presence of any suspended particles
that can cloud the water [10, 11]. pH is the measurement to describe the acidity and
alkalinity of the water. A microbial facilitated process called denitrification in which
the nitrate is reduced and produces molecular nitrogen through intermediate series
of gaseous nitrogen oxide products [12].
All the sensors and micro-controller use battery for their operation. The data of
water quality are being sensed by each sensor accordingly. Then, the sensor data are
processed in micro-controller by managing the real data that is understandable by
human. Table 1 shows the sensor response time and power consumption for each
sensor used and microcontroller (Arduino Nano).

Fig. 1 Block diagram of

sensor unit

Table 1 Current consumption and response time for sensors and Arduino Nano
Components/Sensors Current consumption (mA) Response time (ms)
Arduino Nano 19 0.9
Temperature 10 0.8
Turbidity 14 0.3
pH 30 0.4
446 H. F. Hawari et al.

Fig. 2 Wireless communication architecture

2.3 Wireless Communication

The data from sensors are transferred over the network using Wi-Fi connection. A
Wi-Fi module (EP8266) is integrated with Arduino Nano as controller to provide
the connectivity of sensor device with cloud computing by connection to a Wi-Fi
gateway to connect to Internet. It supports 802.11 b/g/n/e/i of network protocol with
Wi-Fi Direct (P2P). It is advantageous of using this module as it supports standard
network protocol, has antenna diversity and selection (to extend wireless range) and
relatively low power consumption which is about 170 mA during Transmit Mode
(Tx Mode). Figure 2 shows wireless communication architecture for this project.

2.4 Cloud Computing and Mobile Application

In the cloud, all data collected in the microcontroller are stored in the database through
cloud computing system. The cloud used for this project is Ubidots which provide
simple IoT implementation in cloud computing system. This cloud can provide suffi-
cient features needed to implement this project successfully. To enhance the porta-
bility and mobility of data monitoring system, mobile application for tablets and
smartphones is designed to capture those data stored in the cloud and visualized in
charts and gauges representation for water quality monitoring purposes. Real time
data will be measured and saved in the cloud to produce historical data for certain
period.

2.5 Power Management

This prototype uses two lithium-polymer batteries as the power sources to all elec-
tronic components involved. To manage the power consumption, 1st battery will
provide power to ESP8266 to set up the module and network. Once finished, 1st
battery is turned-off while 2nd battery is switched on providing power to sensors to
get measured sensor data from water. After all sensors data have been achieved, 2nd
battery is switched off while turning on 1st battery to send sensors data to cloud via
Development of Real-Time Internet of Things (IoT) … 447

Fig. 3 Process flow on device operation

standard wireless communication protocol. The power consumption estimated for

overall system is 159 mA.

2.6 Operation of Real-Time Water Quality Monitoring

Figure 3 shows the operation of prototype in water quality monitoring.

Sensors measure water quality from lakes and send to Nano. The sensor data will
then be processed by Nano and transmit them to cloud via Wi-Fi connectivity. Then,
data sent will be stored in cloud storage and will be used by mobile application in
iOS to display all data in several representations for water quality monitoring.

2.7 Overall System Architecture

Figure 4 shows the system architecture of the system starting with sensors getting all
the data from lakes and transmit to cloud via wireless connectivity and monitored in
mobile application.
448 H. F. Hawari et al.

Fig. 4 System architecture

2.8 Prototype Deployment in UTP Lake

The prototype is placed on top of water supported by three pillars anchored in the
water attached with stretchable strings. This is to prevent the device to flow away
from desired location. Figures 5 and 6 shows the location of prototype development
in UTP lakes and deployment technique of the device in UTP lake respectively:

Fig. 5 Prototype
deployment in UTP lake
Development of Real-Time Internet of Things (IoT) … 449

Fig. 6 Deployment location

in UTP lake

3 Results and Discussion

Data for water quality parameters include turbidity, pH value and temperature were
taken in two different locations in UTP. The duration for each data measurement is
1 h with 14 h of total data measurement per day for 7 days in a week.

3.1 Water Pollution Index

An equation is used to calculate water pollution index:

Pollution index = (0.152 × Temperature + 0.002 × Turbidity + 0.714 × pH)/3

(1)

The sensor data are inserted into the formula to get water pollution index value.
Table 2 shows the pollution index value for each pollution level.
The value of water pollution index in UTP lake is calculated based on formula
constructed with the value of temperature, pH and turbidity measured. Data for WPI
is calculated for overall 4 weeks of data collection to indicate the level of pollution
of UTP lake. From Fig. 7, we can observe that the trend of WPI is stable during 1st

Table 2 Pollution index table

Pollution level Water pollution index value
Polluted 0.00–1.40
Slightly polluted 1.41–2.70
Clean/Normal 2.71–5.90
Slightly polluted 5.91–7.30
Polluted 7.31–10
450 H. F. Hawari et al.

Fig. 7 Pollution index of

UTP lake in location A and B

week until 3rd week, but there is slight drop in WPI value in 4th week. This is due
to lower value of turbidity which contributes to lower WPI value in the calculation.
WPI increases during 5th week to 6th week because it has been measured in location
C where the water quality is found to be slightly polluted.

3.2 Temperature

Figure 8 shows the data for water temperature in UTP lakes for 6 weeks of data
collection:
The average of temperature value for location A is 29.50 °C, for location B is
29.10 °C and for location C is 29.45 °C. During 1st week until 6th week of data
collection, water temperature of location A, B and C shows nearly the same value
except for 4th week. During 1st week to 3rd week, the mean average water tempera-
ture is higher than during 4th week due to warmer ambient condition. This increases

Fig. 8 Water temperature in UTP lakes for 6 weeks

Development of Real-Time Internet of Things (IoT) … 451

the solubility and toxicity of certain compounds in the water. These compounds
include heavy metals elements such as zinc, lead, cadmium and ammonia. With
higher water temperature (more than 30 °C), oxygen consumption, metabolic rate
and tissue of aquatic organism permeability increases. This will reduce the survival
rate of aquatic organisms in the water as the lethal concentration 50 (LC50) drops for
a long period. The lower the value of LC50, the lower the concentration of chemical
inside the water, causing lower life time for aquatic organisms.

3.3 PH

Figure 9 shows the data for water pH in UTP lakes for 6 weeks of data collection:
The average pH value for location A is 6.84, for location B is 6.83 and for location
C is 8.65. Based on WQI, the permissible range of value for normal lake water pH
is 6.01–8. As the average of pH value for each week and total average of pH value
for 1st week until 4th week falls in between the permissible range of value, the water
pH measured contributes normal value of water pollution index of UTP lake.
Unlike water pH in location C where the average pH value is 8.65, it is included
in slightly polluted range. This value has possibility to contribute slightly polluted
water condition for water pollution index. When the pH value is below or higher
than the optimal value or permissible range of value, aquatic life become susceptible
to fungal infections and other physical damages thus it is not suitable for aquatic
organisms to stay in water that has non-optimal pH value.
This can harm those organisms and reduce water quality. Related to spectrum, high
pH levels can damage the gills and skin of aquatic organisms and can cause death at
pH level higher than 10.0. While the mean pH value of location C is still below that
pH level, the aquatic organisms are safe, but it also can harm those organisms inside
the water.

Fig. 9 Water pH in UTP lakes for 6 weeks

452 H. F. Hawari et al.

3.4 Turbidity

Figure 10 shows the data for water turbidity in UTP lakes for 6 weeks of data
collection. The average of turbidity value for location A and B contributes to normal
water pollution index while the average of turbidity value for location C contributes
to slightly polluted water pollution index. The lower the value of turbidity, the higher
the water quality based on water pollution index. This can be achieved by reducing
the Total Suspended Solids (TSS) present in the water which higher total suspended
solids causes the water to be turbid. TSS are particles that are larger than 2 µ, therefore
any particles in the water that are smaller than 2 µ are considered as Total Dissolved
Solids (TDS). TSS is contributed by algae and bacteria inside the water that can
cause water to be more turbid.
Usually, clear water is considered as clean water. Excessive value of turbidity
can impair the water quality for aquatic and human life. Other than that, higher
water turbidity will increase water temperature and reduce Dissolved Oxygen (DO).
Warmer water cannot hold DO more than cold water. Based on the data from those
locations, location C is showing least DO quantity compared to the other 2 locations
as it has highest water turbidity and included in slightly polluted range based on
WQI. This lead to higher heat absorption from solar radiation and transferred to
surrounding water by conduction. Therefore, location C is considered as slightly
polluted and not suitable for aquatic life and human usage.
On the other hand, pollutants are most likely to be found in high water turbidity.
The contaminants include bacteria, mercury, lead and other material that encourage
the development of harmful substances inside the water. Since those 3 locations are
used for recreational use, as turbidity reduces visibility of underwater structure, it
negatively affects the body’s recreational use for users in UTP.

Fig. 10 Water turbidity in UTP lakes for 6 weeks

Development of Real-Time Internet of Things (IoT) … 453

Fig. 11 Overall current

consumption for device

3.5 Power Consumption

Figure 11 shows the overall current consumption for device used in water quality
monitoring:
Analysis is made to calculate the battery life for both batteries (battery A and B)
using the formula shown below:

Capacity of battery A : 2200 mAh

Capacity of battery 2B 1200 mAh

Battery 1 = 2.2 Ah/[(122/3600 × 73) + (3478/3600 × 0.004)]

= 37.05 days

Battery 2 = 1.2 Ah/[(2/3600 × 75) + (2/3600 × 84) + (3596/3600 × 1.5)]

= 31.53 days

From the calculation made, the battery life for the prototype is estimated up to
30 days.

4 Conclusion

In conclusion, Internet of Things (IoT) assist and help people in deciding the best
solution for any problems faced in daily life especially water quality because water
is the main element required by human beings for their life every day. Without good
quality of water, people tend to have problems with their healthcare and welfare
because everyone need water to survive in any situation. Development of innovative
technology help them to monitor the water quality especially in the river as it is the
main source of water supply for users in entire world. Once water quality can be
454 H. F. Hawari et al.

monitored in real time, it is easy for the department responsible in maintaining the
water quality to get all required data of water quality, visualize them in most effective
representation and decide the best solution for water treatment. The performance of
this system is comprehensive with good techniques and methods used for every
scope included. The system has good response time sending all data successfully to
cloud with minimum communication error. This is a good practice as it increases
reliability and practicality of this project. Without this technology, people are forced
use manual intervention continuously by taking samples from site locations regularly
and analyse them in the laboratory using effective laboratory analysis methods. This
is time consuming and not considered as efficient water quality monitoring.

Acknowledgements I would like to thank Dr Radzi Ahmad from Electrical and Electronics
Engineering department, Universiti Teknologi PETRONAS who despite of being busy with their
schedule, managed to take time out to provide me with untiring assistance, direction, encourage-
ment, continuous guidance, and support indeed provides a huge contribution for me to complete
this project successfully.

References

1. Suratman S, Sailan MI, Hee YY, Bedurus EA, Latif MT (2015) A preliminary study of water
quality index in Terengganu River basin, Malaysia. Sains Malaysiana 19(44):67–73
2. Hossain MA, Sujaul IM, Nasly MA (2013) Water quality index: an indicator of surface water
pollution in eastern part of Peninsular Malaysia. Res J Recent Sci 10(2):10–17
3. Gazzaz NM, Yusoff MK, Ramli MF, Aris AZ, Juahir H (2012) Characterization of spatial
patterns in river water quality using chemometric pattern recognition techniques. Department
of Environmental Science, Faculty of Environment Studies
4. Khan R et al (2012) Future internet: the internet of things architecture, possible applications
and key challenges. DITEN Department University of Genova (UNIGE), 16145 Genova, Italy
5. Hawari HF, Zainal AA, Ahmad MR (2019) Development of real time internet of things (IoT)
based air quality monitoring system. Indonesian J Electr Eng CS 13(3)
6. Chung Hua BD, Fahmi H, Yuhao L, Kiong CC, Harun A (2018) Internet of things (IOT)
monitoring system for elderly. In: International conference on intelligent and advanced system
(ICIAS), p 201
7. Minerva R, Biru A, Rotondi D (2015) Towards a definition of the Internet of things (IoT). IEEE
Internet Initiative. Telecom Italia, Torino
8. UN-Water (2013) An increasing demand, facts and figures, coordinated by UNESCO in collabo-
ration with UNECE and UNDESA. Available: http://www.unwater.org/water-cooperation-201
3/en/
9. Vinegre C, Madeira D, Narciso L, Cabral NH, Diniz M (2012) Effect of temperature on oxida-
tive stress in fish: lipid peroxidation and catalase activity in the muscle of juvenile seabass,
Dicentrarchus labrax
10. Schveitzer R et al (2013) Effect of different biofloc levels on microbial activity, water quality
and performance of Litopenaeus vannamei in a tank system operated with no water exchange.
Aquacult Eng 56:59–70
11. Jonsson et al M (2013) Foraging efficiency and prey selectivity in a visual predator: differential
effects of turbid and humic water. Can J Fish Aquat Sci
12. Norshidah B et al (2014) Characterization of spatial patterns in river water quality using
chemometric techniques. Sains Malaysiana 9(43):1355–1362
Edge Computing Technology:
State-of-the-Art and Challenges
in an Internet of Things

Deepa Raghunathan and M. Krishnamoorthi

Abstract The Internet of Things (IoT) is a network of expedients that can intuiting,
collecting, managing and communicating data over the Internet without human inter-
vention. It is a concept that connects to nearby objects without user intervention, using
a wired or wireless network. In IoT, objects interact and share information to provide
the user or users with advanced, intelligent services. In the field of edge computing,
data engendered by IoT devices can be processed closer to where it was created,
and as an alternative to sending data to data centers or data centers in more ways.
New research and customer support in selecting seemly edge computing structures
for particular solicitations. This paper gives a complete overview of the methods of
introduced power adding structures, comparing Open Source devices through their
survival method. Finally, this work sheds light on smart home application tools using
Edge Computing Systems and also explores the challenges posed by Edge Computing
Systems.

Keywords IoT · Edge computing · WAN · Fog · Cloudlets

1 Introduction

Kevin Ashton first coined the term Internet of Things in the presentation tile, which
introduced sensor tagging for large consumer company products in 1999 to strengthen
supply chain management. Things to the Internet describe the ecosystem, where
sensors and actuators are connected to the Internet. All objects in the IoT system speak
the same language and understand and operate without human intervention. The
Internet of Things has been labeled too late, the concept of connected networks is not

D. Raghunathan (B)
Research Scholar, Department of Computer Science and Engineering, Dr.N.G.P Institute of
Technology, Coimbatore, India
M. Krishnamoorthi
Professor, Department of Computer Science and Engineering, Dr.N.G.P Institute of Technology,
Coimbatore, India

© The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 455
R. Ibrahim et al. (eds.), International Conference on Artificial Intelligence
for Smart Community, Lecture Notes in Electrical Engineering 758,
https://doi.org/10.1007/978-981-16-2183-3_44
456 D. Raghunathan and M. Krishnamoorthi

Fig. 1 Taxonomy of IoT

new and there are many developments like this in progress [1]. In the 1980s, the Coca-
Cola machine was connected to the Internet and monitored for availability, which
reduced unnecessary visits to vending machines. Sethi et al. The following statistics
review basic IoT systems, architectures, available protocols, and applications of IoT.
The Edge Computing model has gained a significant reputation in the academic
and industrial sectors over the past few years. It serves as the key to many future
technologies, such as vehicle-to-vehicle communication using 5G, IoT and cloud
computing services for future customers. It gives low-slung invisibility versatility
and arrangement mindfulness bolster for delay-sensitive solicitations. In the context
of the latest developments in cloud computing, mist computing, and versatile edge
computing, basic investigation within the space of edge computing, pioneered by
providing additional solutions to researchers in existing solutions and future appli-
cations [2]. Due to the characteristics and features of IoT services, smart home
applications have also become a major problem in the IoT application areas (Fig. 1).
Edge Computing speaks to an unused concept within the computing extract that
fetches the administrations and practicalities of cloud calculating closer to the conclu-
sion patron. It is characterized by two variables, such as quicker handling and speedier
application reaction time. Right now created Internet-facilitated applications such
as observing, cybernetic authenticity and instantaneous activity observing involve
quicker administration, speedier reaction times [3]. Conclusion clients ordinarily run
these applications on their portable gadgets whereas performing center administra-
tions and preparing on cloud servers. Cloud administrations are influenced by versa-
tile gadgets, coming about in tall idleness and versatility issues [4]. Gives an edge
computing application as per the prerequisites by moving the handling to the arrange
edge. Mist computing concerns can be illuminated by three distinctive authority
figuring simulations, such as cloudlets [5], mist computing [6] and versatile edge
computing [7].
This paper describes Edge Computing methods and models. Section 2 describes
the literature review of an IoT edge computing, fog computing, and mist computing.
Section 3 gives a brief computing model. Sections 4 and 5 provides conclusions
Edge Computing Technology: State-of-the-Art … 457

about the availability of equipment and, finally, Sect. 6 grants the supposition on this
work.

2 Literature Review

According to statista, Edge computing may be a fast-growing showcase, with the

figure worldwide income set to reach nine billion U.S. dollars by 2024. The capital
use on edge computing gadgets and framework is anticipated to sum to an amazing
146 billion U.S. dollars by 2028. A few businesses such as budgetary administrations
and fabricating have as of now made broad utilize of edge computing in their day-to-
day businesses, and nations like China are full-heartedly grasping the potential that
the modern innovation holds for them.
Granjal et al. [8] demonstrate a complete analysis of security and privacy according
to the protocols in each layer of the OSI model, and their results for the common IoT
domains. Batalla et al. [9] The Smart Home Domain focuses on this topic and gives
a detailed report on the most common threats to using the cloud platform in smart
home systems, and key concepts such as protocol analysis and security implications.
Jose has been introduced as an algorithm to get every location in a smart home and
tested in a factual environment with successful results. We often look at smooth
homegrown solutions for security issues, followed by instability, the costs we have
limited, which is an important aspect for consumers [10].
Tiwari et al. [11] asserted that there is greater clarity in IoT, enabling Artifi-
cial Intelligence, the IoT features to build more LinkedIn systems, and the Active
Engagement System customer interaction, technology optimization, waste reduction
and improved data collection. In addition, he noted that security, privacy, complexity,
flexibility and compliance are the main disadvantages of the Internet of Things.
Sharma et al. [12], Improved Security and Surveillance, Disability Assistance,
Automated Industrial Control, Smart Agriculture, Home and Building Automation
are important applications of the Internet of Things. However, it is not limited and
is currently the smart city for all IoT applications and has become an important area
of study.
The detailed review on challenges and issues of the Internet of things and catego-
rize the problems in global use, standards and protocols, operational and technical
and data and software. Privacy and security were the most significant challenges
when IoT enabled for universal use. Data integrity, Data Authentication, Data Confi-
dentiality, and Trust and Governance affects the global use of the Internet of things.
Also, detailed how interoperability, mobility, scaling, human–machine interaction,
and dependability are the significant operational and technical challenges of IoT
presented by Sarhan [13].
A periodical of the applications and remunerations of authority adding in a variety
of fields and also challenges such as service management, programmability, security,
naming, data capture, optimization metrics, and privacy presented by Zhang [14].
458 D. Raghunathan and M. Krishnamoorthi

Assuncao et al. [15] describes emerging technologies for distributed data stream
processing, as well as cloud resource management techniques for elasticity manage-
ment. But it is difficult to develop and implement stream processing applications
under different distribution structures.
Bilal et al. [16] offered an outline of edge computing motivation and edge tech-
nology. Concerns in this paper include resource management and allocation, resource
scalability, methods, data capture, fault tolerance, and privacy in service quality,
conservation and edge computing. Table 1 shows some of the detailed reviews about
edge computing in different application domains, also with future directions.
This review commends of an Edge Computing worldview is silent at a primary
organize, and the logical communal requirements additional period to get it in what
way to utilize its preferences to provide the administrations accessible to end-user
clients. In expansion, the study centers as it were on the focal points of a superiority
computing deprived of tending to great contests, such as the determination of the
finest procedures that IoT operators got to take after in arrange to attain great benefit.
Edge registering brings information handling power at the edge of the system, closer
to the wellspring of the information.

3 Edge Computing Models Enabling IOT

Edge computing is one way to improve data processing in cloud-based computing

systems at the network edge, which is closer to the source of the data. There are
many advantages to using sophisticated computing technology to help with latency
issues and bring bandwidth-intensive content closer to the user, as well as support
future network infrastructure. Let’s first look at the various state-of-the-art computing
technologies that have been implemented to help promote and monitor growth in the
use of IoT and 5G.

3.1 Portable Edge Computing

Portable edge computing may be an arrange design that empowers computing and
capacity assets to convey substance over the radio get to organize, and to conclu-
sion clients to make strides the viability of the network [25]. To do this, the MEC
adjusts the load on the radio connection to increase network performance and reduce
the need for long-range backhauling. It is hoped that network requirements will
further spread to IoT and 5G compliant technologies and devices. Mobile Edge
Computing enables operators to more efficiently meet traffic and resource demands
while laying the foundation for the future smart, next-generation networks. Mobile
Edge Computing supports services such as augmented reality, augmented reality,
and the Internet of Things, to give startups and startups the opportunity to adapt to
emerging technologies.
Edge Computing Technology: State-of-the-Art … 459

Table 1 Summary of edge computing reviews

Author Methodology Result and discussion Future scope
Alonso et al. [17] Developed a new edge The place engineering To implement RA in
computing is tried with IoT stage different scenarios
engineering dwell in in keen agro industry like IoT records in
safekeeping and to decrease transfer immense platform
secrecy by utilizing speed taken a toll plus smart cities and
block chain innovation between edge and brainy supervision
cloud also in healthcare
system
Luo et al. [18] Computation The COM approach Discover an
offloading method proposed was effective offloading procedure
(COM) in solving multi to realize most
objective optimization extreme vitality
problems utilization sparing in
versatile gadget
Alamgir Hossian et al. Layered approach A few quantitative and Executing
[19] subjective thinks circumstance
about are executed to mindfulness (SA)
preparing crude IoT framework in an IoT
information, which is based shrewd city
successful in terms of environment
inactivity and giving
circumstance
mindfulness choice to
clients
Ray et al. [20] Novel edge IoT based Created the novel To overcome the
architecture edgent empowered taking after issues like
edge IoT framework edge-IoT engineering
for e-healthcare and usefulness and
illustrated with test operational issues,
bed. The test was prerequisite capability
performed on two and determination
situation like client criteria
with fever and client
without fever
Dhavijay et al. [21] Wireless body area Gives the survey WBAN is the
network (WBAN) almost WBAN based effective patient
IoT healthcare session technology so
framework and it will be used in the
different advances of IoT and delivering
IoT arrange home health also
frameworks monitoring
(continued)
460 D. Raghunathan and M. Krishnamoorthi

Table 1 (continued)
Author Methodology Result and discussion Future scope
Zamora et al. [22] Precision agriculture Developed an open Experience the lack of
(PA) with three tier source three tier modularity and
platform platform like local delocalization. For
plane (cyber physical this reason future
system) to gather data, work aims to use
edge plane is for modular turnkey
monitoring and system to falling
managing PA tasks, industrial time
finally cloud plane in situation active and
will collect the current conservation
and past data records.
The system was
implemented for
tomato plants to
monitoring growth of
the crop
Sha et al. [23] Review based on IoT Checked on almost The virtual machine
security at edge layer edge centric IoT, edge and hypervisors are
based security plans basic for the security
and inquire about in entirety edge layer.
issues So these virtual
machine to be
lightweight conjointly
require secure
working framework
like SeL4
Khan et al. [24] Review edge Categorized the Tended restriction of
computing techniques up-to-date in edge edge computing
computing based on worldview. The
the application domain restrictions are
like cloudlet, mist and energetic charging
portable edge component, genuine
computing time application
bolster, security, joint
trade show, repetition,
adaptable design,
come up short over
capabilities and asset
administration

3.2 Fog Computing

Fog networking, or “fogging” directly, is a term used to describe a decentralized

computing system. Mist computing takes an edge of the system, retaining data,
handling, loading and submissions in a highly consistent and productive position
concerning the cloud network and the source of data [26]. Fog Intelligent aims
to expand mist computing and amenities on the control of the web and seek to
Edge Computing Technology: State-of-the-Art … 461

reduce the transmission, analysis and storage of data transmitted to the cloud. Data
collected from IoT sensors and other devices is conveyed to the mist for review and
dispensation. However, these sensors and devices are far from responsive, and cloud
calculations allow short-term analytics and processes at the arrange edge to shrinkage
the sum of information shown to the cloud [27].

3.3 Cloudlets

Cloudlets represent small-scale mist data hubs on the authority of the system and
the second tier of the three-tiered hierarchy: cloud-to-cloud phone or smart device.
Cloudlets aims to development resource-intensive and intelligently versatile appli-
cations by giving versatile gadgets with less idleness and more proficient computing
assets in geographical proximity [28]. This reduces the latency delay usually asso-
ciated with cloud computing in the WAN. While the 5G network is moving up on
the horizon, cloudlets have evolved to support resource-intensive applications such
as machine learning, speech recognition, and language processing applications. Due
to the increased demand, they provide the necessary support for the 5G network that
comes with them. Cloudlets need to be decentralized and spread widely to provide
the greatest amount of network coverage possible. It helps to get power from nearby
mobile computers.

3.4 Micro Data Centers

Microdata centers are small, access-level systems that provide all the necessary
components of a traditional data center. Microdata centers are more suitable for some
edge computing applications than traditional data centers because they are usually
smaller in size and can be deployed indoors and outdoors in harsh conditions. This
makes them an ideal solution for edge computing because they can be used locally
in a data source to suit the needs of those who choose to implement them [29].
Microdata centers may be more attractive to small and medium-sized ones—they
may not have data centers because large companies have more resources and do not
need such a clarification. While the growth of the Internet of Effects is driving the
expansion of new shrewd policies and IoT sensors, the souk for the micro data center
is estimated to be worth $32 billion over the next two-and-a-half years.
Some of the edge computing techniques discussed in the next generation network
infrastructure are discussed above. The Internet of Things and the continued advance-
ment of the 5G wireless networking network are viable when changes in other areas,
such as versatile edge and haze computing, are ready and provide a secure platform
for innovation. MEC, Fog Computing, Cloudlets and Micro Data Center play impor-
tant roles when it comes to the future of network computing, but their first test is
when 5G is ready for roll-out.
462 D. Raghunathan and M. Krishnamoorthi

In this segment we review a series of contemporary developments in edge

computing systems and design, programming models and applications [30]. Five
different application scenarios are discussed: general usage, vehicle information
analytics, shrewd domestic, video stream analytics and fundamental realism. In this
work, widespread use and smart home edge computing systems and equipment are
described below.

4 General Utility Applications

In general, the application landscape consists of ten different edge computing systems
with different end devices such as IoT devices, mobile devices, firewall nodes and
edge, as well as cloudlets, mobile devices, local servers, PCs, home gateway, router,
server, local cluster, FOG, MEC and seaport.

4.1 Cloudlet

Cloudlet could be a trusted, clever computer or computer cluster that is well connected
to the Web and is accessible for versatile gadgets within the encompassing region
[31]. It upgrades “Cloud to Mobile Device” from “Cloud to Mobile Device”, the basic
two-stage architecture of Mobile Cloud Computing. Cloudlet can support users as
a stand-alone cloud, which is currently becoming a “minor server” or “statistics
focus in the case”. However, cloudlet has not been developed under the name Edge
Computing, and its design and concept of edge computing are due to three key
features, such as softness, great resources, and proximity to customers. As an edge
computing system.

4.2 CloudPath

Cloudpath reduced the difficulty of designing and implementing Path Computing, in

order to maintain the most popular RESTful development model of the cloud appli-
cation. Cloudpath is based on the first observation that decomposition of RESTful
stateless properties is possible by the existence of a specific storage layer. Cloudpath
has facilitated the creation and deployment of computing applications by extending
traditional storage abstractions to the Data Center hierarchy embedded within the
geographic range of networked storage [32].
Edge Computing Technology: State-of-the-Art … 463

4.3 PCloud

PCloud [33] coordinating edge computing and capacity assets to bolster consistent
portable administrations. Basically, these assets are virtualized through the primary
virtualization layer called STRATUS [34]. Making a bunch of distributed resources
that can hunt for modern assets to memorize almost changes in assets. The runtime
method is capable for asset determination and allotment through the asset choice
interface; the runtime instrument chooses and interfaces fitting resources based on
diverse choice prerequisites. Beneath the Asset Get to Control Arrangement, the
reserves will be combined to supply comparative administrations for outside applica-
tions and make a modern worldview. In expansion, the PCLoud framework moreover
gives basic administration administrations such as authorization administration and
client information accumulation to control access to assets for other clients.

4.4 ParaDrop

The University of Wisconsin—Madison WiNGS Laboratory developed a computing

system called Paradrop [35]. It makes computing and capacity assets closer
to portable gadgets and makes information assets accessible to third-gathering
designers. The objective is to carry insights to the superiority of the network. Paradrop
Edge replaces the actual entry point for a computer network, which supports standard
server separation devices and utilities such as a multi-tenancy device. In addition,
Paradrop leverages lightweight container virtualization technology.

4.5 SpanEdge

SpanEdge [36] may be a investigate venture of an Institute in Sweden. It interfaces

central mist hub with the dominant edge hub to decrease arrange inactivity in WAN
associations and the programming environment. This permits the framework to run
close the source of the information. Designers ought to center on building gushing
applications without considering the area and dispersion of information sources.
The information midpoint in Span Advantage is made up of two planes: the mist
information center to begin with equal, and the edge information center moment
level. Fractional gushing preparing errands run on the internal edge to diminish
latency and increment execution.
Spilling handling is one of the foremost vital sorts of applications in edge
computing, where information from distinctive topographical sources are produced
by diverse information sources and persistently transmitted in stream shape. All
crude information is ordinarily transmitted to the WAN’s information middle waiter,
and brook handling frameworks such as Apache Stimulus and Flink are made and
464 D. Raghunathan and M. Krishnamoorthi

arranged for a consolidated information center. In any case, this strategy does not
switch enormous information created by numerous gadgets on the edge of orga-
nize impacts, and when the application requires less idleness and forecast, the
circumstance gets to be more awful.

4.6 Cloud-Sea Computing Systems

The Chinese Academy of Sciences is launching the Cloud-Sea Computing System

[37] in 2012. In cloud-sea figuring schemes, the term “cloud” denotes to data centers
and “sea” mentions to the fatal. The plan of this venture is spoken to at three stages: the
global framework design equal, the information center server and the capacity frame-
work near, and the computer chip-near [38]. The architecture of cloud computing in
the ocean has four distinctive structures, such as ternary computing by sea equip-
ment, cooperation with locality, scalability for ZB and trillion devices, and minimal
deployment to existing ecosystems.

4.7 Smart Home Applications

Another scenario that plays an imperative character in edge reckoning is the manage-
ment of IoT strategies in insolent home environments. One common factor is the
privacy issue of a variety of home appliances. There are two different edge computing
systems, Vigilia and HomePad.
Vigilia
In [39], Vigilia proposed tightening the smart home system by limiting network
access to system devices. An authentication framework has been introduced for
applications that reject default access rules and implement authentication at the
network level for API-granularity. Only check the time spent on the router for those
communications, thereby helping customers secure their home appliances.
HomePad
The HomePad system in [40] also proposed setting IoT applications and presenting
privacy-aware centers to diminish security issues. Home Pad permits clients to char-
acterize protection approaches for how apps get to and handle their information. By
requiring applications to utilize clear data streams, they can utilize Preamble rules to
confirm that they have the potential to damage the security approach indicated amid
establishment.
Edge Computing Technology: State-of-the-Art … 465

5 Open Challenges in Edge Computing

The complication of these edge computing frameworks has driven to numerous

specialized issues such as versatility administration, security and security, benefit
revelation, benefit conveyance and versatility, scalability, versatility, low-cost tolerant
deployment model reliability and resource management. Still, some important
challenges need to be addressed.
Xiao et al. [41] described an edge computing applications are affected by the
security attacks named as Distributed Denial of service (DDos) attacks, Cross Station
attacks, Malware inoculation attacks, Confirmation and Authorization attacks, Man
in the central bouts, Bad data inoculation attacks are pointing edge computing
substructures in real world. The idea behind this is that the author proposed direc-
tion to secure the Edge Computing System with an integrated outline. It consists
of 3 layers, designated as External Fine Granular Access Control Layer, Middle
Sanctuary Functional Sheet and Internal Hardware Discrete OS Layer.
Liu et al. [42] reviewed an Edge Computing security attacks targeting autonomous
vehicles that pose a threat to personal and public safety. Autonomous vehicle protec-
tion includes sensor protection, functioning scheme safety, and controller scheme
safety and communiqué security. This Autonomous Driving Edge covers various
layers of computing stack.

5.1 IOT and Edge Computing Challenges

IoT and Edge devices are used by lightweight Operating Systems and limited storage
capabilities projected for narrow, special tasks. For example AWS Lambda at the edge
platform the content delivery network (CDN) feature within Amazon Cloud Front
with benefits in the form of speed and the cost by pushing code closer the users
through limited edge compute executions.
Cost and Connectivity
IoT sensors and devices services are managed locally compared to the edge devices.
The edge devices are handled large services inside the infrastructure because of more
number of transactions are processed locally so that less number of data traverses to
and from of the cloud. The connectivity is a main problem while reaching out to an
on-premises data to the cloud.
Security
In distributed data processing, edge computing is the major challenge. The security
in edge computing usable for one edge computing application it’s not applicable
directly migratable to another situation because of variety of reasons for example
heterogeneity of edge devices and communication protocols. In this section, illustrate
that the Edge computing Challenges and Security issues in Table 2 [2].
466 D. Raghunathan and M. Krishnamoorthi

Table 2 Challenges and security issues

Challenges Security issues
Confidentiality and integrity Achieving confidentiality and integrity of data in transit to and
from the cloud-edge-device provider
Power consumption Mist usage added knobs
Power ingesting is higher than central cloud
Data security Data producing knobs are spread
Provided that confirmation and sanction scheme for the total
nodes is not an informal task
Fault tolerance Failure of a node should be immediately fixed
Architecture Programming architecture

6 Conclusion and Future Work

Edge Computing is a new model that gives users the ability to network and store
from remote cloud to IoT and 5G context. The Superiority Computing principle is
auspicious in providing a better user experience for better services and applications.
Newly developed edge computing systems and technologies normally decrease over-
head records dispensation and storage and increase the reliability and effectiveness of
transportable data analytics. This paper also discusses the basic concepts and systems
and tools related to edge computing and classifies them as advanced computing such
as cloudlets, mist, and portable edge computing and micro-data center application
domains. This work gives future researchers an excellent idea to understand the
superiority computing pattern and advance investigation to solve unexpected prob-
lems. Future research should explore investigation tendencies in edge-fog computing
systems.

References

1. Hassan N, Gillani S, Ahmed E (2018) The role of edge computing in internet of things. IEEE
Commun Mag 99:1–6
2. Yaqoob I, Ahmed E, Gani A (2016) Mobile ad hoc cloud: a survey. Wirel Commun Mob
Comput 16(16):2572–2589
3. Liu M, Yu FR, Teng Y (2019) Distributed resource allocation in blockchain-based video
streaming systems with mobile edge computing. IEEE Trans Wireless Commun 18(1):695–708
4. Pace P, Aloi G, Gravina R (2019) An edge-based architecture to support ancient applications
for healthcare industry 4.0. IEEE Trans Ind Inf 15(1):481–489
5. Shaukat U, Ahmed E, Anwar Z (2016) Cloudlet deployment in local wireless networks:
motivation, architectures, applications, and open challenges. J Netw Comput Appl 62:18–40
6. Bao W, Yuan D, Yang Z (2017) Follow me fog: toward seamless handover timing schemes in
a fog computing environment. IEEE Commun Mag 55(11):72–78
7. Ahmed E, Akhunzada A, Whaiduzzaman M, Gani A (2015) Network centric performance
analysis of runtime application migration in mobile cloud computing. Simul Model Pract
Theor 50:42–56
Edge Computing Technology: State-of-the-Art … 467

8. Granjal J, Monteiro E, Silva JS (2015) Security for the internet of things: a survey of existing
protocols and open research issues. IEEE Commun Surv Tutorials 17(3):1294–1312
9. Batalla JM, Vasilakos A, Gajewski M (2017) Secure smart homes: opportunities and challenges.
ACM Comput Surv 50(5):75:1–75:32
10. Jose AC, Malekian R (2017) Improving smart home security: Integrating logical sensing into
smart home. IEEE Sens J 17(13):4269–4286
11. Tiwary A (2018) Internet of Things (IoT) research, architectures and applications. Int J Future
Revolution Comput Sci Commun Eng 4(3):23–27
12. Sharma V, Tiwari R (2016) A review paper on IOT and it’s smart applications. Int J Sci Eng
Technol Res (IJSETR) 5(2):472–476
13. Sarhan QI (2018) Internet of things: a survey of challenges and issue. Int J Internet Things
Cyber-Assur1(1):40–75
14. Shi W, Cao J, Zhang Q, Li Y, Xu L (2016) Edge computing: vision and challenges. IEEE
Internet Things J 3:637–646
15. de Assuncao MD, da Silva Veith A, Buyya R (2018) Distributed data stream processing and
edge computing: a survey on resource elasticity and future directions. J Netw Comput Appl
103:1–17
16. Bilal K, Khalid O, Erbad A, Khan SU (2018) Potentials, trends, and prospects in edge
technologies: fog, cloudlet, mobile edge, and micro data centers. Comput Netw 130:94–120
17. Alonso RS, Corchado JM (2019) A review of edge computing reference architectures and a
new global edge proposal. Futur Gener Comput Syst 99:278–294
18. Luo Y, Peng K (2019) A computation offloading method over big data for IoT-enabled cloud-
edge computing. Futur Gener Comput Syst 95:522–533
19. Alamgir Hossain SK, Anisur Rahman MD (2018) Edge computing framework for enabling
situation awareness in IoT based smart city. J Parallel Distrib Comput 122:226–237
20. Ray PP, Dash D (2019) Edge computing for Internet of Things: A survey, e-healthcare case
study and future direction. J Netw Comput Appl 140:1–22
21. Dhanvijay MM, Patil SC (2019) Internet of Things: a survey of enabling technologies in
healthcare and its applications. Comput Netw 153:113–131
22. Zamora Izquierdo MA, Santa J (2019) Smart farming IoT platform based on edge and cloud
computing. Biosyst Eng 177:4–17
23. Sha K, Yang TA (2019) A survey of edge computing based designs for IoT security. Dig
Commun Netw (DCN) 1–12
24. Khan WZ, Ahmed E (2019) Edge computing: a survey. Futur Gener Comput Syst 97:219–235
25. Ahmed A, Ahmed E (2016) A survey on mobile edge computing. In: 10th International confer-
ence on intelligent systems and control (ISCO). pp 1–8. https://doi.org/10.1109/ISCO.2016.
7727082
26. Yi S, Li C, Li Q (2015) A survey of fog computing: concepts, applications and issues. In:
Proceedings of the workshop on mobile big data ACM. pp 37–42
27. Bao W, Yuan D, Yang Z, Wang S, Li W, Zhou BB, Zomaya AY (2017) Follow me fog: toward
seamless handover timing schemes in a fog computing environment. IEEE Commun Mag
55(11):72–78. https://doi.org/10.1109/MCOM.2017.1700363
28. Shaukat U, Ahmed E, Anwar Z, Xia F (2016) Cloudlet deployment in local wireless networks:
motivation, architectures, applications, and open challenges. J Netw Comput Appl 62:18–40
29. Yaqoob I, Ahmed E, Gani A, Mokhtar S, Imran M, Guizani S (2016) Mobile ad hoc cloud: a
survey. Wirel Commun Mob Comput 16(16):2572–2589
30. Liu F, Tang G (2019) A survey on edge computing systems and tools. Proc IEEE 107(8):1–24
31. Satyanarayanan M, Bahl V, Caceres R, Davies N The case for vm-based cloudlets in mobile
computing. IEEE Pervasive Comput
32. Mortazavi SH, Salehe M, Gomes CS, Phillips C, de Lara E (2017) Cloudpath: a multitier
cloud computing framework. In: Proceedings of the second ACM/IEEE symposium on edge
computing. ACM, p 20
33. Jang M, Schwan K, Bhardwaj K, Gavrilovska A, Avasthi A (2014) Personal clouds: sharing
and integrating networked resources to enhance end user experiences. In: Proceedings of IEEE
INFOCOM. IEEE, pp 2220–2228
468 D. Raghunathan and M. Krishnamoorthi

34. Jang M, Schwan K (2011) Stratus assembling virtual platforms from device clouds. In:
Proceedings of the IEEE international conference on cloud computing. IEEE, pp 476–483
35. Liu P, Willis D, Banerjee S (2016) Paradrop: enabling lightweight multi-tenancy at the networks
extreme edge. In: IEEE/ACM symposium on edge computing (SEC). pp 1–13
36. Sajjad HP, Danniswara K, Shishtawy A, Vlassov A (2016) Spanedge: towards unifying stream
processing over central and near the-edge data centers. In: IEEE/ACM symposium on edge
computing (SEC). IEEE, pp 168–178
37. Xu Z-W (2014) Cloud-sea computing systems: towards thousand-fold improvement in
performance per watt for the coming zettabyte era. J Comput Sci Technol 29(2):177–181
38. Fielding RT, Taylor RN (2000) Architectural styles and the design of network-based software
architectures, vol 7. University of California Irvine Irvine, USA
39. Trimananda R, Younis A, Wang B, Xu B, Demsky B, Xu G (2018) Vigilia: securing smart
home edge computing. In: Proceedings of the IEEE/ACM symposium on edge computing
(SEC). IEEE, pp 74–89
40. Zavalyshyn I, Duarte NO, Santos N (2018) Homepad: a privacy-aware smart hub for home
environments. In: Proceedings of the IEEE/ACM symposium on edge computing (SEC). IEEE,
pp 58–73
41. Xiao Y, Jia Y, Liu C (2019) Edge computing security: state of the art and challenges. Proc IEEE
107(8)
42. Liu S, Liu L, Tang J (2019) Edge computing for autonomous driving: opportunities and
challenges. Proc IEEE 107(8)
Comparison of Rectangular Tunnel
with Shield Jacking Support and Pipe
Roof Support in Finite Element Method

M. Y. M. Nasir, H. Mohamad, and H. Alarifi

Abstract Methods of the Gaussian distribution curve and Finite Element Method
(FEM) were applied in predicting ground settlement however, those equations are
still incomplete with a trough width parameter, K of soil condition that has evolved
from the developing of Kenny Hills tunnel in Malaysia. Therefore, this paper of
research analyses pipe roofing and shield support of rectangular tunnel using FEM
by PLAXIS 2D software to estimate in occurrences of ground deformations. Charts
of maximum ground surface settlement and trough width parameter are developed to
show the difference between both lining supports through parametric study comprises
variation of soil properties. Results show that pipe roofing support is suitable to be
adopted in weak soil condition while shield support is suitable in soil condition
with higher stiffness value. K can be concluded with value equal to 0.6 for the
rectangular pipe roofing support and 0.7 for rectangular shield support in Kenny
Hills soil formation.

Keywords Rectangular · Tunnel · Shield · Pipe · Support

1 Introduction

Rectangular tunnel is becoming popular and has been used in constructions especially
by developed country. Hence, for the past few decades improvements have been
done on its support systems in order to prevent deformation of ground especially on
a weak ground condition. Generally, shield and pipe roofing support were used for
constructing rectangular shaped tunnel.
Fang et al. [1] have go through on shield support and Koyama [2] has discussed
further on the benefits of the shield support. Koyama [2] found shield support can
be applied in different type of soil condition [2] due to rigid structure of shield while
pipe roofing support can prevent the settlement of ground above the tunnel using a

M. Y. M. Nasir · H. Mohamad (B) · H. Alarifi

Civil and Environmental Engineering Department, Universiti Teknologi PETRONAS, Seri
Iskandar, Malaysia
e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 469
R. Ibrahim et al. (eds.), International Conference on Artificial Intelligence
for Smart Community, Lecture Notes in Electrical Engineering 758,
https://doi.org/10.1007/978-981-16-2183-3_45
470 M. Y. M. Nasir et al.

Fig. 1 Settlement trough of Gaussian form

series of steel pipe by consolidate the ground stress and disperse the ground stress
to reduce the excavation stress during the tunnel excavation work [3]. The design
of the pipe roof support may vary between the project because it solely based on
experience [4].
Ground deformation is a major problem that crucial to be solved prior to develop
the underground systems which is not able to be seen directly through underground
layout. Furthermore, different properties and different characteristics of underground
soil at different coordinate makes prediction on ground deformation hardly to be
calculated manually. Therefore, simulation will be done to predict amount of risk
and to identify the affected surrounding and structure constructed in underground
space.
In order to determine the settlement above the tunnel, Peck [5] proposed an equa-
tion to develop transverse settlement above the tunnel known as Gaussian distribution
curve as shows in Fig. 1.
Equation 1 give a settlement at various point of the trough,

x2
Sx = Smax e− 2i x 2 (1)

where:
Sx the settlement profile at the surface
S max the maximum vertical settlement
ix the trough width parameter which, physically, is the distance from the tunnel
axis to the point of inflection of the curve.
Volume of the surface settlement profile is given by,
√
Vs = 2πi x Smax (2)
Comparison of Rectangular Tunnel with Shield Jacking Support … 471

i x = K z0 (3)

where

K constant value
z0 depth of the tunnel axis.

The width through parameter, K, varies with the type of soil. Kimura and Mair
[6] suggested that K values must be equal to 0.5 for clay soil. Meanwhile, O’Reilly
and New [7] recommend that K values must be between 0.2 and 0.3 for granular
soil with condition of tunnel depth less than 10 m, 0.4–0.5 for stiff fissured clay,
0.5–0.6 for Glacial deposits and 0.6–0.7 for silty clay deposits. O’Reilly and New
[8] also had suggested K values must be 0.4 for stiff clays, 0.7 for soft silty clays. In
contrast, Mair and Taylor [9] suggested differently in K values where K values that
they emphasized were 0.5 for all clay soil and 0.35 for granular soils. Khoo et al. [10]
stated that an appropriate K to be used was 0.5 for soil type encountered in Klang
Valley of Malaysia. On the other hand, Yeates [11] suggested that K value should
be from 0.2 to 0.3 for granular material above the water table. Moreover, Rankin
[12] had proposed K value which were 0.4–0.5 for stiff fissured clay, 0.5–0.6 for
glacial deposits, 0.6–0.7 for silty clay. Overall K is depending on soil properties that
significant to consider even though K value does not include in the Gaussian curve
distribution that mostly used by researchers, due to none of the research had found
K value that specifically use for rectangular or box tunnel. In addition, all trough
width parameter is produced by analyze the circular tunnel cross-section. Hence,
this research study aims at producing charts for the tunnel designer in estimating the
maximum settlement and searching tough width parameter value for the rectangular
tunnel with either shield support or pipe roof support located at Kenny Hill soil
condition.

2 Kenny Hills Soil Formation

Mostly, Kuala Lumpur, Malaysia covers Kenny Hill soil formation. This type of
soil consisting of interbedded shales, mudstone, siltstone, and sandstones. Kenny
Hills soil has undergone some metamorphic events resulting in changes of sand-
stone/siltstone to quartzite and schist/phyllite respectively. As stated by Ooi [13], the
depth of Kenny Hills soil layer will be more than 10 m below the existing ground
level and the soil formation becomes very hard with SPT greater than N = 50.
The investigation of the Kenny Hills engineering properties has been done by
Refs. [14, 15]. From their study, the measured bulk unit weight mostly ranged from
15.8 to 21.9 kN/m3 for residual soil and for highly weathered rock (Grade IV) is
24.0 kN/m3 .
For the effective shear strength parameters, [15] has stated that for the residual
soil with SPT ≤ 100, the range for cohesion, c is from 5 to 10 kN/m2 and for angle
472 M. Y. M. Nasir et al.

of friction, ϕ  is 28° while for the soil which has SPT greater than 100, the cohesion,
c is 15 kN/m2 and angle of friction, ϕ  is 29°. For highly weathered rock (Grade
IV), the equivalent Mohr–Coulomb strength parameters are 30 kN/m2 for cohesion,
c and 34° for an angle of friction, ϕ  [16].

3 Method of Analyses

The simulation is based on greenfield model with setting as follows:

Model Dimension = 40 m × 40 m
Surcharge Load = 0 kPa
Soil Layer = Homogenous Soil
Tunnel Depth = 15 m.
The boundary condition is fixed by standard fixities, where the side vertical bound-
aries are fixed in horizontal x-direction but free to move vertically, while the bottom
boundary is restrained from any movement in all directions.
For lining support, the steel pipe will be used in the simulation of pipe roof support
with an outer diameter of 813 mm with 16 mm thickness. The steel pipe will be used
as a pipe roof with grade S275. The tunnel shield support specification will set as in
Table 1.

3.1 Parametric Study

Each simulation in PLAXIS 2D software will have one variable parameter and
constants for other parameters by Hardening Soil method in as in Table 2.

Table 1 Shield support

Parameter Value
properties
Behavior type Elastic
Thickness (m) 0.35
Stiffness (MPa) 21,000
Specific weight (kN/m3 ) 38
Poisson’s ratio 0.3
Comparison of Rectangular Tunnel with Shield Jacking Support … 473

Table 2 A typical range of Kenny Hill soil for simulation

Soil parameters Study case Constant value Variable value
General properties yunsat (kN/m3 ) ysat = 20 kN/m3 yunsat = 15, 15.5, 16, 16.5, 17, 17.5,
c = 5 kPa 18
ϕ  = 28
E = 25 MPa
v = 0.3
ysat (kN/m3 ) yunsat = 19 kN/m3 ysat = 20, 20.5, 21, 21.5, 22
c = 5 kPa
ϕ  = 28
E = 25 MPa
v = 0.3
Strength parameter c (kPa) ysat = 20 kN/m3 c = 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 15,
yunsat = 19 kN/m3 20, 25
ϕ  = 28
E = 25 MPa
v = 0.3
ϕ  (o ) ysat = 20 kN/m3 ϕ  = 20, 22, 23, 25, 27, 28, 29, 30, 31,
yunsat = 19 kN/m 32, 33, 34, 35, 36, 38, 40
3

c = 5 kPa
E = 25 MPa
v = 0.3
Stiffness E (MPa) ysat = 20 kN/m3 E = 20, 30, 50, 75, 100, 150, 200, 250
yunsat = 19 kN/m3
c = 5 kPa
ϕ  = 28
v = 0.3
v ysat = 20 kN/m3 v = 0.1, 0.15, 0.2, 0.25, 0.3
yunsat = 19 kN/m3
c = 5 kPa
ϕ  = 28
E = 25 MPa

4 Result

4.1 Relationship Between Maximum Settlement and Soil

Parameter

The amount of soil settlements affects the degree damage of the ground surface.
Therefore, factors of settlement need to be identified at first in order to ensure least
ground deformation. Water content within the soil is one of the factors that determine
the level of settlement. Figure 2 proven a higher level of water in soil properties that
cause greater settlement from applying for rectangular shield supports. However,
rectangular pipe roof supports show a striking effect of constant settlement although
water content level is increasing. This condition occurs because pipe roof support
provide advance protection for soil deformation caused by tunnel excavation work
474 M. Y. M. Nasir et al.

Fig. 2 Relationship between maximum settlement and general properties

while shield support provide soil protection just before the excavation. Next, cohesion
and angle of friction as illustrated in Fig. 3 also affecting soil interaction when the
pattern of settlement is decreasing inversely to cohesion and angle of friction of
the soil with rectangular shield support. On the other side, rectangular pipe roofing
support will be maintaining a constant settlement with increasing value of cohesion
and angle of friction. Hence, constructing a rectangular pipe roofing support can

Fig. 3 Relationship between maximum settlement and shear strength properties

Comparison of Rectangular Tunnel with Shield Jacking Support … 475

Fig. 4 Relationship between maximum settlement and stiffness properties

minimize the soil interaction thus reduce soil deformation at any degree of cohesion
and angle of friction. Figure 4 illustrated a comparison between Young’s modulus
and Poisson’s ratio towards soil settlement and results show reducing settlement with
increasing Young’s Modulus value on rectangular pipe roofing support which similar
to the result of rectangular shield support. However, pipe roof supports produce less
settlement reduction compared to shield support because of steel pipe installation
for pipe roof support has disturbed the soil stiffness around the support’s perimeter.
Poison’s ratio shows no effect toward the ground surface settlement for both supports.

4.2 Relationship Between Trough Width Parameter and Soil

Parameter

By fitting the Gaussian distribution curve graph into the FEM graph, the value of the
trough width parameter can be produced as a reference to estimate ground surface
settlement in constructing any rectangular tunnel project which has approximate
similar soil condition. Hence, investigations to determine a trough width parameter,
K were done and mostly K value for rectangular pipe roofing support is 0.6 and K
value for rectangular shield support is 0.7. In this study, the range of the trough width
parameters are between 0.4 and 0.7 effected by saturated unit weight, angle of friction,
cohesion and Young’s modulus values used. Figure 5 illustrated the increment of
water level within soil due to lesser in K value for rectangular shield support, while
Fig. 6 resulted in inconsistent K value which between 0.6 and 0.7 for both angle of
friction and cohesion properties. Hence, it is significant to further study in angle of
friction and cohesion properties of soil that affected the K value inn understanding
476 M. Y. M. Nasir et al.

Fig. 5 Relationship between trough width parameter and general properties

Fig. 6 Relationship between trough width parameter and shear strength properties

the stress of soil. Meanwhile, Fig. 7 illustrated reduction on K value from 0.7 to
0.4 when the value of Young’s modulus increases for both types of supports with
approximately constant K value on Possion’s ratio properties. It can be concluded
that unsaturated unit weight and Poisson’s ratio properties give no effect to K value
for rectangular tunnel with shield of pipe roof support.
Comparison of Rectangular Tunnel with Shield Jacking Support … 477

Fig. 7 Relationship between trough width parameter and stiffness properties

5 Conclusion

Previous literatures are yet discovering the trough width parameter use in constructing
a rectangular tunnel cross-section. Hence, this simulation and analyses were done to
investigate on the maximum settlement and trough width parameter for rectangular
pipe roofing and shield support using Kenny Hill soil condition. The outcome of
applying rectangular pipe roofing resulted in the suitability of constructing this type
of support within the soil with properties of high-water content can still ensure the
reduction of ground surface settlement. Furthermore, the result obtains also shows
that by constructing a rectangular tunnel with pipe roof support, it is only suitable
to be applied for low soil stiffness because the affected area by the pipe excavation
in higher soil stiffness will increase soil plasticity area, hence reduction of ground
surface settlement is lower. Meanwhile, a rectangular tunnel with shield support gives
better respond in reducing the ground surface settlement in soil with high stiffness
as shield support only involves with jacking process. Besides that, stiffness of soil
is important to be considered prior to all construction works because it will affect
the soil deformation during the excavation either rectangular pipe roofing or shield
support. Finally, the trough width parameter value can be concluded with K equal to
0.6 for the rectangular pipe roofing support and 0.7 for rectangular shield support in
Kenny Hills soil formation.
478 M. Y. M. Nasir et al.

References

1. Fang Y, Chen Z, Tao L, Cui J, Yan Q (2019) Model tests on longitudinal surface settlement
caused by shield tunnelling in sandy soil. Sustain Cities Soc 47:101504. https://doi.org/10.
1016/j.scs.2019.101504
2. Koyama Y (2003) Present status and technology of shield tunneling method in Japan. Tunn
Undergr Space Technol 18(2–3):145–159. https://doi.org/10.1016/S0886-7798(03)00040-3
3. Wu K, Shao Z (2018) Effects of pipe roof support and grouting pre-reinforcement on the track
settlement. Adv Civil Eng.https://doi.org/10.1155/2018/6041305
4. Volkmann GM, Schubert W (2007) Geotechnical model for pipe roof supports in tunneling.
In: Proceedings of the 33rd ITA-AITES world tunneling congress. pp 755–760
5. Peck RB (1969) Deep excavations and tunneling in soft ground. In: Proceedings of the 7th
international conference soil mechanics and foundation engineering, Mexico City, State of the
art volume, pp 225–290
6. Kimura T, Mair RJ (1981) Centrifugal testing of model tunnels in soft clay. In: Proceedings of
the 10th international conference on soil mechanics and foundation engineering, Stockholm.
pp 319–322
7. O’Reilly MP, New BM (1982) Settlement above tunnels in the United Kingdom—their
magnitude and prediction. Tunnelling 82:173–181
8. New B, O’Reilly M (1991) Tunnelling induced ground movements: predicting their magnitudes
and effects. In: 4th International conference on ground movements and structures. Pentech
Press, Cardiff, pp 671–697
9. Mair R, Taylor R (1997) Theme lecture: board tunneling in the urban environment. In: Proceed-
ings of the 14th international conference on soil mechanics and foundation engineering.
Balkema, Hamburg, pp 2353–2385
10. Khoo CM, Idris NISI, Mohamad H, Rashid ASA (2018) Numerical evaluation of settlement
trough width parameter. MATEC Web Conf 203:1–9. https://doi.org/10.1051/matecconf/201
820304010
11. Yeates J (1985) The response of buried pipelines to ground movements caused by tunnelling
in soil. In: GEDDES JDF (ed) Ground movements and structures. Pentech Press, Plymouth,
pp 145–160
12. Rankin WJ (1988) Ground movements resulting from urban tunnelling: predictions and effects.
Geol Soc Eng Geol Spec Publ 5(5):79–92. https://doi.org/10.1144/GSL.ENG.1988.005.01.06
13. Ooi TA (1986) Design and construction problems of foundation for the high rise structure in
the Kuala Lumpur Areas. In: Proceedings of IEM-ISSMFE joint symposium on geotechnical
problem, Kuala Lumpur. pp 112–122
14. Toh CT, Ooi TA, Chiu HK, Chee SK, Ting WH (1989) Design parameters for bored piles in a
weathered sedimentary formation. In: Proceedings of the 12th international conference on soil
mechanics and foundation engineering, Rio de Janeiro, vol 2. pp 1073–1078
15. Wong J, Sing M (1996) Some engineering properties of Kenny Hill formation in Kuala Lumpur.
In: Proceedings of the 12th South-East Asian geotechnical conference, Kuala Lumpur. pp
179–187
16. Hoek E, Brown ET (1997) Practical estimates of rock mass strength. Int J Rock Mech Min Sci
34:1165–1186
The Effect of Skill Type on Skill-Gap
in the Nigerian Construction Industry

M. S. Aminu, S. U. Kunya, I. Y. Mohammed, and S. A. Bustani

Abstract Skills gap and its effects has become a serious problem requiring urgent
attention in the Nigeria construction industry. Basic literatures on construction prac-
tice left no doubts that the successful realization of construction projects requires
the committed action of professionals and vocational trade operatives. The objective
was to identify effects of skills-type on skills-gap related to Building projects in
Abuja, Nigeria. Field survey of some building projects sites in Abuja compliment
the secondary information on the existence of skill gap. The sample comprises both
public and private projects, using both criterion-based and stratified random sampling
techniques. 254 questionnaire was administered, 215 were returned, resulting to a
response rate of 85%. The result shows that more than 80% of the Nigeria construction
professionals are all male with years of experience ranges between 16 and 20 years,
with academic qualification ranges from bachelor degree to Doctorate Degree respec-
tively. Number of employees ranges within 10–50 people in most of the construction
sites. From the result it shows that stake holders participation on training and devel-
opment was not encouraging resulting to Poor workmanship, Construction waste,
Rework as some of the major effects of skills-gap. The study therefore recommend
that policy makers should develop skill acquisition programmed which include public
private partnership, developmental training, on the job training. Government should
come up with special scheme for the improvement of skills through the vocational
and certification program to assist and encourage younger generation to various
apprenticeships to reduce occupational segregation, provide the necessary require-
ment and standard training for both young men and women. The national direc-
torate of employment (NDE), the national skills qualification framework (NSQF),

M. S. Aminu
Department of Building Technology, Jigawa State Polytechnic Dutse, Dutse, Nigeria
e-mail: [email protected]
S. U. Kunya (B) · I. Y. Mohammed (B)
Department of Building Technology, Abubakar Tafawa Balewa University, Bauchi, Nigeria
e-mail: [email protected]
S. A. Bustani (B)
Department of Quantity Surveying, Bayero University Kano, Kano, Nigeria
e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 479
R. Ibrahim et al. (eds.), International Conference on Artificial Intelligence
for Smart Community, Lecture Notes in Electrical Engineering 758,
https://doi.org/10.1007/978-981-16-2183-3_46
480 M. S. Aminu et al.

the industrial training fund (ITF) and national board for technical education (NBTE)
should ready to face the challenges of the recent skills-gap economy as we enter the
third decade of the millennium for sustainability and productivity.

Keywords Construction industry · Effects · Nigeria · Skills-gap · Types

1 Introduction

The poor quality of project delivery can be attributed to the gradually diminishing in
size of competent skilled construction workers and the influx of unskilled, inefficient
and dissatisfied workers who see the sector as a last resort. There are indeed a few
that see construction crafts as career worth coming to. This scenario will inevitably
result in skill-gap, an issue of major concern in view of its invariable impact to the
construction industry, citizens and national economy [12]. There are no exact data
statistics of trade skill-gap in the construction industry, however, Medugu et al. [13]
established that, there is skilled craftsmen shortage that needed urgent attention in
the Nigerian construction industry.
Odediran and Babalola [15] observed that, the most needed large quantity
manpower for building construction in Nigeria and globally are artisans and
labourers. Okuntade [16] also observed that, despite the importance of craftsmen
to the construction industry, many of them still remain untrained, even though the
construction industry, all over the world have been implementing skills acquisition
programmed to meet the demand of occasional change in technologies. Bilau et al.
[3] suggested that, construction craftsmen are operatives who contribute skillfully
with their hands in the practical realization of projects in a construction industry.
Siboe [18] stated that, training in the construction industry has a specific purpose,
it should provide experience which develop the labor of employees in the area of
skills, knowledge and attitude. Overtime, training has been a precursor to human
resource planning which is a process for determining and ensuring that an organi-
zation will have enough number of skilled and experience persons available at the
right time, quantity/quality and place, carrying out jobs that can meet the desires of
the organization, which provide satisfaction for workers involved [18].

2 Statement of the Research Problem

Skill shortage is the major problem due to the rapid urbanisation in developing
nations as a result to substantial increase in construction activities. This has tremen-
dous consequences resulting to poor workmanship [10]. Lack of adequate training
centers and training programmed in Nigeria, dominantly utilizing unskilled labour to
maximize profit, resulting to constant rework leading to time and cost overruns, and
The Effect of Skill Type on Skill-Gap … 481

construction waste. Ameh and Itodo [2] assert that, most managers of the construc-
tion industry put little emphasis on the effects of construction waste contributed by
unskilled labour to construction projects.
The shortage of skilled craftsmen in Nigeria has also been attributed to boom in
the agricultural sector in recent times. Most craftsmen engage themselves on farming
being a lucrative business, living closer to their families, making them happier than
working for a long period of hours with a meager amount of pay. The quality shortage
was as a result of lack of passion for the job, seeing the sector as a last resort that
doesn’t get promoted, eager of finishing work on time irrespective of its quality.
This has drastic consequences on productivity in Nigeria. Construction waste sent
to landfills has a drastic effects of polluting the environment, reducing the quality of
construction, delay in construction activities leading to time and cost overruns in the
study area.
In spite of the contributions of skilled craftsmen to the development of the
construction industry, little efforts is being put on the issues that relate to the types
of skill-gap and skill-gap effects in the industry. As a result, this is rarely been
addressed and is little understood. There is need to address this problem by providing
a clear theoretical and conceptual framework to understands of the basics and related
concepts of skills-gap and its effects in the construction industry in Abuja Nigeria.

3 Research Aim and Objective

The aim of this research was to investigate the effects of skills-gap on types of
skills-gap in the construction industry in Abuja Nigeria.
To achieve the above aim, the objective was:
(i) To study the effects of skill-gap on types of skills-gap in the Nigerian
construction industry.

4 Skills-Gap and Its Effects in the Nigerian Construction

Industry

According to Kurushi [12], skills is what an individual possesses which may be learnt
informally and/or on-the-job. Skill-Gap is the difference in the skills required to do
the job and the actual skills possessed by the employees. A “Skill-Gap” arises when
a worker does not have all the skills necessary to do the job effectively. Therefore,
skill-gap refer to the situation where a company has employees but they are not
skilled enough to meet the organization’s objectives.
482 M. S. Aminu et al.

4.1 Types of Skill-Gap in the Construction Industry

The types of skill-gap in an organization can be attributed to Skills shortages

of masonry, Skills shortages of carpentry, Skills shortages of iron-bending, Skills
shortage of plumbing, Skills shortage of electrician, Skills shortages of tiling, Skills
shortages of painting, skills short-age of mixing/batching skills shortage of finishes,
and skills shortage of scaffolding [12].

4.2 Effects of Skills-Gap

Skills-gap has a drastic effects on construction resulting to: poor workmanship,

construction waste, rework, and client dissatisfaction, decrease in profitability of
projects, and so on.

4.2.1 Construction Waste

Most construction projects has frequent changes in design often result in construction
waste, time delays, cost overruns, quality defects and other negative impacts on
construction. In recent years, many researchers investigated the effects of design
change and more importantly the common causes behind the changes from different
perspectives. Reviewing the existing literature on design changes, causes and effects,
developing two taxonomies for change causes and change effects; and illustrating
how the taxonomies can be used during the design change management process to
reduce construction waste [20].

4.2.2 Rework

One of the most disturbing issues facing organizations in the construction industry
is their inability to become quality focused. As a result sub-standard products and
services often emanate, which inadvertently result in rework. Typically, rework is
caused by errors made during the design and construction process. These errors
appear downstream in the design process and therefore have a negative impact on
a project’s performance [17]. The lack of attention to quality, especially during the
design and construction process, has meant that rework has become an inevitable
feature of the construction work, and the costs have been found to be as high as
12.4% of total project costs. Such costs could be even higher because they do not
represent schedule delays, and other intangible costs of poor quality. To reduce the
cost and effect of rework, an understanding of its causal structure is needed so that
effective prevention strategies can be identified and the effects of rework reduced or
eliminated [17].
The Effect of Skill Type on Skill-Gap … 483

4.2.3 Client Dissatisfaction

The rate of client dissatisfaction is adversely affecting the timely delivery of construc-
tion projects. This assesses construction stakeholders’ perception of the causes of
client dissatisfaction and its effects on project delivery in a bid to proffer solution in
minimizing its occurrences. This suggest that client’s dissatisfaction related problems
are the main causes of delays while time and cost overruns are the major identifiable
effects of dissatisfaction in construction projects [1].

4.2.4 Decrease in Profitability of Projects

Change of design or nature of the work is one of the most significant causes of vari-
ation in construction projects, the client financial status can affect project progress,
and this status often leads to change in work schedules, causing delay, affecting
the quality of the project. Hindrance to prompt decision-making process leading to
decrease in productivity and profitability [22].

4.3 Conceptual Framework of the Study

A concept is a plan, vision, or a symbolic representation of an abstract idea. A

conceptual framework in research shows the researcher’s position on the research
problem, which gives direction to the study, and further shows the relationships that
exist between different constructs which the study investigate. It may be an adoption
of a model used in a previous study with modifications to suit the present investigation
[10] (Fig. 1).

5 Sample Size

Using the yiros formula of sample size calculation

z2 × P 2
S=
E2
where

Z is the z-score having a constant value of 1.96 at 95% confidence level [21].
P is the standard deviation ranges from 20 to 100 distributions (my P = 56).
E is the margin of error having a value of 4.

From the equation above

484 M. S. Aminu et al.

Causes of
Skills-gap

Current
Effect of Types of
measures of
skills-gap Skills-gap
bridging the gap

Challenges in
bridging the gap

Fig. 1 Conceptual framework of the study

1.962 × 562 3.84 × 3136 12, 042

S= 2
= = = 752
4 16 16
Sample size is 752 [11]; which was translated to 254 respondents and sampling
technique was simple random sampling method.
Therefore according to Krejcie and Morgan [11] the sample of this population is
254. The sample comprises both public and private projects, using both criterion-
based and stratified random sampling techniques. The choice of targeting building
construction sites and professionals in this area is done for the following reasons: The
proximity of the researcher’s state of residence to the study area (FCT, Abuja), Abuja
has the highest population of built environment professionals in the country, and it
has many on-going building construction projects, The location of the construction
activities in the same area makes the study more accurate in terms of cost and time.
Knowing the number of score and Ranking then in order from lowest to highest
the formula
P
R=
100(N + 1)

to calculate the percentile Rank [21].

6 Result and Discussions

Table 1 represent the respondents educational qualification, respondent with first

degree constitute the highest percentage (46.5%) while (20.9%) has master’s degree
and the least are those with PhD having (4.7%). Table 5 provides the results of
The Effect of Skill Type on Skill-Gap … 485

Table 1 Percentage of educational qualification of respondents

Qualification Number of respondents Percentage
National diploma 60 27.9
First degree 100 46.5
Master’s degree 45 20.9
PhD 10 4.7
Educational specialization Number of respondents Percentage
Builders000 85 39.5
Architects 40 18.6
Quantity surveyors 20 9.3
Civil engineers 60 37.2
Others 10 4.6
Years of experience No of respondents Percentage
1–5 years 25 11.6
6–10 years 30 13.9
11–15 years 70 32.5
16–20 years 70 32.5
Above 20 years 20 9.3
Organizational position No of respondents Percentage
CEO/owner 15 6.9%
Projects manager 90 41.8%
Site engineer 80 37.2%
Others 30 13.9%
Number of employee No of respondents Percentage
Less than 10 0 0.0
10–50 140 65.1
50–200 50 23.3
Above 200 25 11.6

the professionals that participated in the survey. Where Builders has the highest
percentage (39.5%), Architect (18.6%), Civil engineers (37.2%), Quantity surveyors
(9.3%) and other Professionals (4.6%).This indicates that the respondents that partic-
ipated in the survey are all professionals. Table 1 identify the working experience of
respondents are within the range of 11–15 years has the highest percentage (32.5%),
and those within (16–20) year’s also. and above 20 years (9.3%). Despite the highest
percentage goes to the group of 11–15 and 16–20 years of experience, therefore,
the perception of the experienced Expertise is respected in this research. Table 5
further represent the organizational position of respondents in their various compa-
nies, projects manager has (41.8%), (37.2%) for site engineer, and other positions has
13.9%. Table 5 further shows the number of employees in their various organization,
486 M. S. Aminu et al.

with (65.1%) between 10 and 50 workers, (23.3%) between 50 and 200 workers and
(11.6%) above 200 workers.

6.1 Research Objective: To Study the Effect of Skills-Gap

on Types of Skills-Gap in the Nigerian Construction
Industry

The objective was achieved by evaluating the measurement and structural models
using the Partial Least Squares Structural Equation Modeling (PLS-SEM) analysis.
The PLS-SEM approach involves a two-staged process of framework evaluation that
involved test of convergent validity, discriminant validity, collinearity diagnostic,
determining R2 and path coefficients [9]. The output of these criteria were gauged
against the recommended benchmarks provided in the extant literature. Table 2 shows
the recommended bench- marks for decision on the parameters.

6.2 Measurement Model Assessment

Figure 2 shows the measurement model for effects of skills-gap on types of skills-gap.
The model consist of five constructs measured in reflective mode [8]. The constructs
are linked with one-headed arrows which indicate causal effect between them. Being
a reflective model, the assessment of the measurement model involves testing the
validity of the model, the indicator loadings, Average Variance Extracted (AVE),
composite reliability and discriminant validity needs to be computed [8]. The main
purpose of assessing measurement model is to test for construct validity. According
to Hair et al. [9] construct validity is the degree to which the indicators of a latent
construct represent the underlying latent construct.
Construct validity is measured by testing the convergent and discriminant validity.
The convergent validity indicates the extent of correlation among the indicators of a
latent construct. On the other hand, discriminant validity test the degree of uniqueness
of a construct relative to other constructs in the research model [19].

6.3 Construct Validity/Convergent Validity

According to Hair et al. [9] convergent validity is the extent of correlation among
the indicators of a latent construct. Hair et al. [8] recommended the use of indicator
loading and AVE as criteria for judging convergent validity.
The Effect of Skill Type on Skill-Gap … 487

Table 2 PLS-SEM model evaluation criteria

Criteria Description Acceptable/Recommended
Measurement model assessment threshold
criteria
Convergent Indicator reliability Measures the variance Value ≥ 0.708 [9]
validity explained by the observed
variable towards underlying
latent construct
Cronbach’s alpha Measures the indicators α ≥ 0.70 [8, 14]
uni-dimensionality (inter-
correlation) with their latent
construct
Composite reliability It is a measure of internal ρc > 0.70 [8, 14]
consistency
Average variance Average percentage of AVE > 0.50 [6]
extracted (AVE) variation explained among
the items of a construct
√
Discriminant Fornell-Larcker Measures the degree to AVE > latent variables
valid criterion which two conceptually correlations [6]
similar concepts are distinct
(in such a way that each
latent variable shares more
variance with its own block
of indicators than with
another latent variable)
Cross-loadings A situation where a variable Variable loading on any
has two or more factor two or more factors at >
loadings exceeding the 0.40
threshold value deemed
necessary for inclusion in
the factor interpretation
process
Heterotrait-mono-trait Measures the degree to Correlation between any
(HTMT) ratio of which two conceptually pair of construct should be
correlations similar concepts are distinct < HTMT. [19]
Structural model evaluation criteria (in such a way that each
latent variable shares more
variance with its own block
of indicators than with
another latent variable)
Collinearity Measures the relationship VIF < 5 is considered mild
between two or more case of collinearity
independent variables
(continued)
488 M. S. Aminu et al.

Table 2 (continued)
Criteria Description Acceptable/Recommended
Measurement model assessment threshold
criteria
R2 (Coefficient of Determination) Amount of variance R2 = 0.20 is considered
explained in endogenous high in consumer
latent variable by all behaviour studies [9]
exogenous latent variable in R2 = 0.25, 0.50, 0.75 are
structural model considered weak,
moderate and substantial
respectively in [8]
R2 = 0.19, 0.33, 0.67 are
considered weak,
moderate and substantial
[5]
β coefficient Standardized regression Higher magnitude
coefficient that allows for a indicates more explanatory
direct comparison between power [7, 9]
coefficients as to their
relative explanatory power
of independent variable
Source Sinkovic et al. [19]

Fig. 2 Measurement model for effects of skills-gap on types of skills-gap

The Effect of Skill Type on Skill-Gap … 489

6.3.1 Indicator Reliability

Indicator reliability measures the absolute importance of individual measurement

items in defining the latent construct. It shows the relationship between the latent
construct and the measurement items [9]. A factor loading of 0.70 is considered
sufficient [4] although [9] pointed that values less than 0.70 are also acceptable
provided that the AVE is achieved.

6.3.2 Average Variance Extracted (Ave)

The second condition for ensuring convergent validity is the AVE. It is defined as
the overall mean value of the squared loadings of the indicators of a latent construct
[9]. The AVE measures the amount of variance a latent construct explained the
measurement items with respect to the amount explained by measurement error [6].
Hair et al. [9] recommended 0.50 and above as suitable to ensure that at least two-third
(2/3) of the variance in the indicator is explained by the construct.
Composite reliability (CR) measures the internal consistency of the measurement
model. Hair et al. [9] described it as the degree of inter-correlations among indicators
of a latent construct. The recommended value is 0.70. In an exploratory research,
however, Hair et al. [8] pointed that 0.60 is equally acceptable.
Table 3 shows the outer loading of the indicators, AVE and composite reliability
of the measurement model. As shown in the Table 3, the outer loadings, the AVE
and CR for the two constructs all fall within the recommended thresholds of the
respective parameter desired for convergent validity. The AVEs reported 0.884 while
the CR reported 0.991 for causes of Skill Gap. The AVEs reported 0.881 while
the CR reported 0.991 for challenges in bridging the gap. The AVEs reported 0.887
while the CR reported 0.992 for effects of skills-gap. The AVEs reported 0.895 while
the CR reported 0.989 for measures of bridging the gap. The AVEs reported 0.886
while the CR reported 0.987 for types of skills-gap. The values were well above the
recommended minimum of 0.50 and 0.70 for AVE and CR respectively. Similarly,
all outer loadings of the indicators measuring the constructs reported values above
0.70 which further indicate the achievement of convergent validity.

6.4 Discriminant Validity

Discriminant validity measures the uniqueness of latent construct with respect to

other constructs in the model. Discriminant validity is assessed using the Hetrotrait-
Monotrait (HTMT) criterion. According to Henseler et al. (2015) the HTMT is the
most conservative measure of discriminant validity when compared with the Fornell-
Larcker or cross loading criteria. The criterion is to compare the calculated result
with a value of 0.85 and 0.90 in such a way that any value in the table above the
recommended threshold implies that discriminant validity is not achieved. Table 4
490 M. S. Aminu et al.

Table 3 Outer loadings construct validity and average variance explained

Construct Measurement Outer loading T statistics P Values CR AVE
items
Causes CSGI 0.970 0.991 0.884
of_Skill_G ap CSGII 0.969
CSGIII 0.935
CSGIV 0.974
CSGIX 0.955
CSGV 0.928
CSGVI 0.957
CSGVII 0.961
CSGVIII 0.914
CSGX 0.876
CSGXI 0.913
CSGXII 0.967
CSGXIII 0.944
CSGXIV 0.894
Challenges CBGA 0.936 0.991 0.881
in_Bridging CBGB 0.965
the_Gap
CBGC 0.959
CBGD 0.943
CBGE 0.913
CBGF 0.913
CBGG 0.939
CBGH 0.951
CBGI 0.970
CBGJ 0.975
CBGK 0.936
CBGL 0.962
CBGM 0.969
CBGN 0.814
CBGO 0.923
Effect of_Skill ESGA 0.965 0.992 0.887
Gap ESGB 0.953
Q8C 0.896
ESGD 0.956
ESGE 0.878
ESGF 0.953
(continued)
The Effect of Skill Type on Skill-Gap … 491

Table 3 (continued)
Construct Measurement Outer loading T statistics P Values CR AVE
items
ESGG 0.959
ESGH 0.935
ESGI 0.911
ESGJ 0.948
ESGK 0.962
ESGL 0.938
ESGM 0.969
ESGN 0.958
ESGO 0.930
ESGP 0.956
Measures of MBGA 0.949 0.989 0.895
_Bridging MBGB 0.918
the_Gap
MBGC 0.934
MBGE 0.971
MBGF 0.956
MBGG 0.926
MBGH 0.962
MBGJ 0.934
MBGK 0.935
MBGL 0.959
MBGM 0.959
Type TSGI 0.945 0.987 0.886
of_Skill_Gap TSGII 0.916
TSGIII 0.947
TSGIV 0.922
TSGIX 0.952
TSGV 0.930
TSGVI 0.959
TSGVII 0.958
TSGVIII 0.921
TSGX 0.961

shows the result of the HTMT test as effects of skills-gap. The correlation matrix
indicates that the HTMT value reported is below 0.90 that is (0.889 and 0.892) which
implies that discriminant validity was achieved.
 492

Table 4 Discriminant validity HTMT

Cause_Skill_Gap Challenges_Bridging_Gap Effect_Skill_Gap Measures_Bridging_Gap Type_Skill_Gap
Cause_Skill_Gap
Challenges_Bridging_Gap 0.891
Effect_Skill_Gap 0.900 0.854
Measures_Bridging_Gap 0.900 0.900 0.889
Type_Skill_Ga p 0.886 0.897 0.892 0.870
M. S. Aminu et al.
The Effect of Skill Type on Skill-Gap … 493

Fig. 3 Structural model extracted from the PLS software

6.5 Structural Model Evaluation

After assessing the measurement model, next stage is to evaluate the effects of skills-
gap on types of skills-gap. The structural model is evaluated using collinearity assess-
ment, test of significance of path coefficients determination of R2 and effect sizes
[9]. Figure 3 shows the structural model extracted from the PLS software .

6.6 Path Coefficient and R2

The path coefficient measured the relative significance of the independent variables
in a model. According to Hair et al. [8], path coefficient ranges from +1 to −1 where
values tending towards 1 indicate stronger relationship. Apart from magnitude of the
path coefficient, it is equally important for the estimates to be statistically significant.
On the other hand, R2 is considered the overall measure of model performance. It is
an indicator of the combined effects of all the independent variables in explaining the
variation in the dependent variable in a model. Hair et al. [9] provided a threshold
for interpreting the R2 value where R2 of 0.75, 0.50 and 0.25 are considered to
be of substantial, moderate and weak effect respectively. However, in behavioural
research, consumer satisfaction studies and other related areas in management and
social science fields, R2 value of 0.20 are considered of high effect [9]. Table 5
presents the path coefficients for the relationships among the two constructs in the
model. The reported path coefficients showed a very strong positive and statistically
significant effect of the independent variable (Effects of skills-gap) on the types of
skills-gap (β = 0.765, t = 13.781, p = 0.000). As shown in the Table 5 explained
about 97% variation in Measures of bridging the skills-gap.
 494

Table 5 Path coefficients and R2

Path Beta Standard deviation T-statistics P-values R2
Direct relationship
Cause_Skill_Gap → Measures_Bridging_Gap 0.126 0.070 1.812 0.070 0.990
Challenges_Bridging_Gap → Measures_Bridging_Gap 0.496 0.042 11.862 0.000 0.994
Effect_Skill_Gap → Measures_Bridging_Gap 0.379 0.051 7.466 0.000
Cause_Skill_Gap → Type_Skill_Gap −0.269 0.064 4.225 0.000
Challenges_Bridging_Gap → Type_Skill_Gap 0.070 0.034 2.036 0.042
Effect_Skill_Gap → Type_Skill_Gap 0.765 0.055 13.781 0.000
Measures_Bridging_Gap → Type_Skill_Gap 0.434 0.046 9.353 0.000
Indirect relationship
Cause_Skill_Gap → Measures_Bridg- ing_Gap → Type_Skill_Gap 0.055 0.032 1.709 0.088
Challenges_Bridging_Gap → Measures_Bridging_Gap → Type_Skill_Gap 0.215 0.028 7.561 0.000
Effect_Skill_Gap → Measures_Bridg- ing_Gap → Type_Skill_Gap 0.164 0.026 6.408 0.000
M. S. Aminu et al.
The Effect of Skill Type on Skill-Gap … 495

6.7 Effect Size

Indicator reliability measures the absolute importance of individual measurement

items in defining the latent construct. It shows the relationship between the latent
construct and the measurement items [9]. A factor loading of 0.70 is considered
sufficient [4]. This shows that the relationship between the latent independent variable
(effects of skills-gap) and the measurement item (types of skills-gap) was considered
sufficient having a factor loading of 0.95 and the result was successfully achieved
(Table 6).

7 Conclusions and Recommendations

The findings from the research indicated that the skills shortage has directs effects
on all the types of skills-gap. Major effects of these skill-shortage were poor work-
manship, construction waste, rework which has definitely affects the productivity
and profitability of projects in the study area. Some current measures of bridging the
skills-gap include Establishing industry base skills acquisition centers, good wages
and remuneration to skilled artisans. The introduction of apprenticeship training
under the new national skills qualification framework (NSQF) across the country for
the informal skills sector should be upgraded, polytechnic should lead this National
Skills Qualification Framework intervention as trainers.
Poor workmanship apparent in buildings across the country make up skilling
existing Artisans and craftsmen imperative. On the job training, developmental
training will help to improve the situation. However, appropriate framework for
setting up vocational training, building trade schools/centers devoted to skill training
development no doubt will enhance workforce development.
Performance Enhancement: Employers should develop standard for improving
productivity of workforce with appropriate incentives. In addition, they need to have
in place clear and definite strategies, goals and performance metrics.
The Nigerian construction industry should diversify its awareness by establishing
a special link with their foreign counterparts in creating the best ways of handling
sustainable skills development internationally.
The federal government should formulate a regulation mandating all construction
project stake-holders to attend a compulsory workshop on the issues leading to skills
development, as well as their management principles.
The national directorate of employment (NDE), the national skills qualification
framework (NSQF), the industrial training fund (ITF) and national board for tech-
nical education (NBTE) should ready to face the challenges of the recent skills-
gap economy as we enter the third decade of the millennium for sustainability and
productivity.
 496

Table 6 Effect size

Cause_Skill_Gap Challenges_Bridging_Gap Effect_Skill_Gap Measures_Bridging_Gap Type_Skill_Gap
Cause_Skill_Gap 0.012 0.092
Challenges_Bridging_Gap 0.846 0.015
Effect_Skill_Gap 0.163 0.950
Measures_Bridging_Gap 0.314
M. S. Aminu et al.
The Effect of Skill Type on Skill-Gap … 497

References

1. Akinsiku O, Akinsulire A (2012) Stakeholder’ perception of the causes and effects of construc-
tion delays on project delivery. J Constr Eng Proj Manag 2. https://doi.org/10.6106/JCEPM.
2012.2.4.025
2. Ameh JO, Itodo ED (2013) Professionals’ views of material wastage on construction sites.
Organ Technol Manage Constr Int J 5(1):747–757
3. Bilau AA, Ajagbe AM, Kigbu H, Sholanke AB (2015) Review of shortage of skilled craftsmen
in small and medium construction firms in Nigeria. J Environ Earth Sci 5(1):35–42
4. Becker J, Klein K, Wetzels M (2012) Hierarchical latent variable models in PLS-SEM: Guide-
lines for using reflective-formative type Models. Long Range Plan 45:359–394. https://doi.org/
10.1016/j.lrp.2012.10.001
5. Chin WW (1998) The partial least squares approach to structural equation modeling. Mod
Methods Bus Res 295(2):295–336
6. Fornell C, Larcker DF (1981) Structural equation models with unobservable variables and
measurement error. Algebra Stat
7. Field A (2009) Discovering statistics using SPSS. 3rd Edition, Sage Publications Ltd., London
8. Hair J, Sarstedt M, Ringle C, Mena J (2012) An Assessment of the use of partial least squares
structural equation modeling in marketing research. J Acad Mark Sci 40:414–433. https://doi.
org/10.1007/s11747-011-0261-6
9. Hair JF, Hult GTM, Ringle CM, Sarstedt M (2014) A primer on partial least squares structural
equation modeling (PLS-SEM). SAGE Publication, Inc., Thousand Oaks, California
10. Ibrahim S (2016) Management of material waste and cost overrun in the Nigerian construction
industry: Unpublished PhD thesis submitted to Nelson Mandela Metropolitan University
11. Krejcie RV, Morgan DW (1970) Determining sample size for research activities. Educ Psychol
Measur 30(3):607–610
12. Kurushi P (2015) Construction practices and skills gap in the construction industry. In:
Proceedings of the 45 builder’s conference of the Nigerian institute of building
13. Medugu NI, Majid MR, Bustani SA, Bala K, Abdullahi U, Mbamali I (2011) Craft skills
availability in the Nigerian construction industry: perception of contractors and consultants.
IUP J Infrastruct 9(3):63–69
14. Nunnally JC, Bernstein IH (1994) The assessment of reliability. Psychom Theory 3:248–292
15. Odediran SJ, Babalola MO (2013) Assessment of business development strategies in the
Nigerian construction industry. J Bus Manag 2(1):34–45. https://doi.org/10.12735/JBM.V2I
1P34
16. Okuntade TF (2014) Building construction technician training: it’s relevance to modern
construction industry in Nigeria. Int J Technol Enhancements Emerg Eng Res 2(3):234–4289
17. Proverbs DG, Holt GD, Cheok HY (2000) construction industry problems: the views of UK
construction directors. In: 16th annual ARCOM conference, pp 6–8
18. Siboe IW (2016) Investigating the adequacy of construction planning in Kenya. Doctoral
dissertation, Jomo Kenyatta University of Agriculture and Technology
19. Sinkovics RR, Henseler J, Ringle CM, Sarstedt M (2016) Testing measurement invariance of
composites using partial least squares. Int Mark Rev
20. Sun M, Meng X (2009) Taxonomy for change causes and effects in construction projects. Int
J Project Manage 27(6):560–572
21. Vogt WP, Johnson RB (2011) Dictionary of statistics and methodology: a nontechnical guide
for the social sciences. Sage
22. Zaghloul SM (2006) Effect of poor workmanship and lack of smoothness testing on pavement
life-cycle costs. Transp Res Rec 1539(1):102–109
Propagation Model of Molecular
Communication Based Targeted Drug
Delivery for Atherosclerosis Disease
Therapy

Pradeep Murugesan, S. Prabakar, K. Porkumaran, and R. Karthikeyan

Abstract Atherosclerosis is one of the major cardiovascular disorder which causes

severe health issues in human body. It also ends with patient death, when improper
therapy takes place. Conventional therapy measures includes oral statin therapy, skin
mode application, Inhalation mode of therapy etc., has their own advantages. But
they cannot fully exploit the inflammation because of lower systemic bioavailability.
Nowadays, Nano molecular communication provides numerous solutions in the field
of targeted drug delivery system. The aim of this article is to propose a new analytical
model for the propagation process of molecules based on Brownian motion mecha-
nism by formulating the probability density of the Latency in blood medium. This
model is analyzed based on crucial parameters such as radius of the propagating
molecules, blood viscosity, drift velocity, distance between Nano-Tx and Nano-Rx,
temperature of the fluid medium with respect to various blood shear rates. Based
on simulation results, the latency is highly affected with molecular radius, distance
temperature, shear rate and drift velocity. Our future work is to apply this model for
various drug carrying molecules used in the treatment of the cardiovascular diseases
and to assess its propagation capacity under various conditions of blood medium.

Keywords Atherosclerosis · Molecular communication · Drug delivery ·

Nano-Tx · Nano-Rx

P. Murugesan (B)
Department of BME, Dr.N.G.P Institute of Technology, Coimbatore, India
S. Prabakar
Sona College of Technology, Salem, India
K. Porkumaran
Sri Sairam Engineering College, Chennai, India
e-mail: [email protected]
R. Karthikeyan
Department of ECE, Dr.N.G.P Institute of Technology, Coimbatore, India

© The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 499
R. Ibrahim et al. (eds.), International Conference on Artificial Intelligence
for Smart Community, Lecture Notes in Electrical Engineering 758,
https://doi.org/10.1007/978-981-16-2183-3_47
500 P. Murugesan et al.

1 Introduction

The research and development of Body Area Networks (BAN) has gained maturity
in recent years. Hence the researchers are now working for the In body networks
using Nano machines [1]. The first useful applications of Nano machines are in
medical technology, where they could be used to identify pathogens and toxins from
the samples of body fluid. The reduced size of Nano machines translates into higher
operational speed. They are designed according to their applications. Specialized
Nano machines called nanobots might be designed not only to diagnose, but to treat
the disease conditions, perhaps by seeking out invading bacteria and viruses and
destroying them. In the field of biomedicine, drug delivery and health monitoring
systems are achieved by biological Nano machines. Targeted Drug Delivery System
(TDDS) is an emerging technology that plays a major role in specialized thera-
peutic purposes. It promises to tackle the conventional hurdles in order to achieve a
controlled rate of drug release with precise location [2]. A new network Paradigm
called IoNT is developed for providing solutions in the drug delivery system. There
are four different types of communication in IoNT. The first type is Electromagnetic
communication. It enables Nano machines to communicate over electromagnetic
waves. The second type is Acoustic, where Acoustic energy is used for communi-
cation. The last type is called as Molecular communication. Here, the molecules are
used for communication between each other. This MC plays a major role in Nano
networks. In this article, focus is put forward on analysis of Diffusion based molec-
ular communication, its propagation model and finally numerical analysis is carried
out for MC-TDDS for Atherosclerotic disease therapy.

2 Methodology

2.1 General Propagation Model

Molecular communication takes place between Nano-transmitters and Nano-

receivers. The information molecule is propagated in the blood medium via different
process namely walkway-based, flow-based, and diffusion-based [3]. In walkway
based, there are pre-defined paths transmitting molecules to the communicating
transmitter and receiver. In flow-based and diffusion-based techniques, molecules
propagate over diffusion in a fluid medium. Most of the studies in the field of molec-
ular communication has been carried out by the diffusion propagation model based
on Brownian motion or Brownian motion with drift [4]. The diffusion process is
explained by Fick’s equations [5]. In the propagation process, propagation medium
is a fluid medium (blood) and the propagation path is from sender bio-nanomachine
to receiver bio-Nano machine. The propagation of messenger molecules is random,
that is to say, in all directions in the fluid medium. These random movements of
propagating molecules are modeled by the Brownian motion. The Brownian motion
Propagation Model of Molecular Communication Based … 501

is an important physical process that undergoes random movement of particles in

the fluid medium resulting from their collision with the fast moving molecules in
the fluid. This motion is generally approximated as a Gaussian model in accordance
with the probability theory [5].

2.2 Mathematical Modeling of Propagation Process

in MC-TDDS

Mathematical models are developed in order to compare the quality of the molec-
ular communication. Propagation period is calculated on the probability basis using
Gaussian distribution [5].
For Brownian motion in the blood medium, we apply the Gaussian distribution
as follows,
 
√ −d 2
f (t) =
1/2
4π Dt exp (1)
4Dt

This is known as the PDF of first hitting time (i.e., time when the molecule first
hit the Nano-receiver or absorption time).
Where D is the diffusion coefficient of the propagating molecules during the prop-
agation process and d is the distance between the Nano-transmitter and Nanoreceiver
in a one dimensional interval (−∞, d]. Diffusion coefficient represents the inclina-
tion of the propagating molecules during the propagation process through the fluid
medium and it can be obtained by the following formula [6]

Kb ∗ T
D= (2)
b
where Kb is a fixed value called the Boltzmann constant, T is the temperature of
the blood medium, and b is also a fixed value representing the drag constant of the
molecule in the fluid medium. The drag constant b is derived for two conditions
based on the size of the drug carrier molecule Scm and the size of the propagating
molecule Sfluid in the fluid (blood) medium [6],
When Scm = Sfluid

b = 4π ηrnm (3)

When Scm > Sfluid

b = 6π ηrnm (4)

η represents the viscosity of the fluid medium and rnm is the radius of the propa-
gating molecule in the fluid medium. According to author [7], liposomes play a major
502 P. Murugesan et al.

role in drug delivery systems. At the systemic level, drug bioavailability is increased
due to the high relative surface area of nano particles and it has been shown that
liposomes around 150–200 nm in diameter remains in the blood stream larger than
those with diameters less than 70 nm. Liposomes exhibit Brownian motion which is
directly proportional to temperature and inversely to the liposome size.
Since the diseased vessel has a blood flow with increased velocity, i.e., fluid
medium has a medium drift velocity, we go with Inverse Gaussian model [8] as
follows,

√  
νt − d 2
f (t) = 4π Dt exp −
d/2
3 (5)
4Dt

3 Numerical Analysis and Results

This part deals with the numerical analysis for the following conditions,
PDF of latency for different distances in blood,
• PDF of Latency for different shear rates with different Drift velocity
• Comparison of different shear rates.

3.1 PDF of Latency for Different Distances in Blood

Condition 1: In our scenario, we consider three different distances i.e., distance

between Nano-Tx and Nano-Rx as {2, 6, 10} in µm with drift velocity to be approx-
imately 1 µm for three different cases. We have some constant values in the PDF. The
Boltzmann constant (Kb) is equal to 1.3807 * 10ˆ−23. The radius of the liposome
molecule is considered into two different cases 75 nm and 100 nm respectively. The
temperature T reveals two different values 277 K as minimum human body temper-
ature and 310 K as maximum human body temperature. The blood viscosity for two
different shear rates 1sˆ−1 and 1000sˆ−1 have also been defined in the table.
Case 1: PDF of Latency for different distances {2, 6, 10}µm, constant drift
velocity 1 µm/s and rnm = 75 nm.
Case 2: PDF of Latency for different distances {2, 6, 10}µm, constant drift
velocity 1 µm/s and rnm = 100 nm (Figs. 1 and 2).
From the above graphs, it is very clear that when the distance between the nano-
Tx and Nano-Rx increases, the PDF decreases accordingly. This increase in distance
leads to the time delay in both the cases. i.e., the nano receiver should await for a
longer duration for the molecules in a blood medium. The temperature and blood
shear rate will also plays a major role in the propagation process. Blood medium with
 Propagation Model of Molecular Communication Based …

Fig. 1 PDF of Latency for different distances {2, 6, 10}µm, constant drift velocity 1 µm/s and rnm = 75 nm
503
 504

Fig. 2 PDF of Latency for different distances {2, 6, 10}µm, constant drift velocity 1 µm/s and rnm = 100 nm
P. Murugesan et al.
Propagation Model of Molecular Communication Based … 505

decreasing temperature and minimum shear rate has a greater chance for propagating
the molecule with very less time delay.

3.2 PDF of Latency for Different Shear Rates with Different

Drift Velocity

Condition 2: Minimum Distance:

The drift velocity inside the blood medium plays a major role in the propagation
process. Increase in the drift velocity has a greater efficiency in Propagation of the
drug carrying molecule. In this case, we have considered the latency in a blood
medium with drift velocity {1, 2, 2.5, 3} and the minimum distance 2 µm.
Case 1: PDF of latency for different drift velocity {1, 2, 2.5, 3}µm/s, minimum
distance −2 µm and rnm = 75 nm.
Case 2: PDF of latency for different drift velocity {1, 2, 2.5, 3}µm/s, minimum
distance −2 µm and rnm = 100 nm (Figs. 3 and 4).
Condition 3: Maximum distance:
In this condition, we have considered the latency in a blood medium with drift
velocity {1, 2, 2.5, 3} and the maximum distance.
Case 1: PDF of latency for different drift velocity {1, 2, 2.5, 3}µm/s, maximum
distance −10 µm and rnm = 75 nm.
Case 2: PDF of latency for different drift velocity {1, 2, 2.5, 3}µm/s, maximum
distance −10 µm and rnm = 100 nm (Figs. 5 and 6).
From both the conditions, it is very clear that when the drift velocity increases, the
first hitting time decreases for the propagating molecules inside the blood medium
with the same distance. In this condition, temperature and blood shear rates along
with the drift velocity has a greater influence in the propagation process. Higher the
drift velocity and lower temperature with minimum shear rate has a greater tendency
for the molecule to travel without any time delay. From the distance point of view,
though with the increased drift velocity in both cases, the PDF is high only in the
case 1. This confirms that the minimum distance between Nano-Tx and Nano-Rx
will be an suitable condition for a better propagation.

3.3 Comparison of Different Shear Rates

In this part, we compare different shear rates in a blood medium for the PDF of the
latency. Here, we have considered the latency in the blood medium with different
shear rates {1sˆ−1, 10sˆ−1, 50sˆ−1, 100sˆ−1, 1000sˆ−1}.
 506

Fig. 3 PDF of latency for different drift velocity {1, 2, 2.5, 3}µm/s, minimum distance −2 µm and rnm = 75 nm
P. Murugesan et al.
 Propagation Model of Molecular Communication Based …

Fig. 4 PDF of latency for different drift velocity {1, 2, 2.5, 3}µm/s, minimum distance −2 µm and rnm = 100 nm
507
 508

Fig. 5 PDF of latency for different drift velocity {1, 2, 2.5, 3}µm/s, maximum distance −10 µm and rnm = 75 nm
P. Murugesan et al.
 Propagation Model of Molecular Communication Based …

Fig. 6 PDF of latency for different drift velocity {1, 2, 2.5, 3}µm/s, maximum distance −10 µm and rnm = 100 nm
509
510 P. Murugesan et al.

Condition 4: Minimum distance = 2 µm and Minimum drift velocity = 1 µm.

PDF of latency for different shear rates {1s−1 , 10s−1 , 50s−1 , 100s−1 , 1000s−1 }.
Case 1: Minimum distance −2 µm, minimum drift velocity = 1 µm/s and rnm =
75 nm.
Case 2: Minimum distance −2 µm, minimum drift velocity = 1 µm/s and rnm =
100 nm (Fig. 7).
Condition 5: Minimum distance = 2 µm and Maximum drift velocity = 3 µm.
PDF of latency for different shear rates {1s−1 , 10s−1 , 50s−1 , 100s−1 , 1000s−1 }.
Case 1: Minimum distance −2 µm, maximum drift velocity = 3 µm/s and rnm
= 75 nm.
Case 2: Minimum distance −2 µm, maximum drift velocity = 3 µm/s and rnm
= 100 nm (Fig. 8).
Condition 6: Maximum distance = 10 µm and Minimum drift velocity = 1 µm
PDF of latency for different shear rates {1s−1 , 10s−1 , 50s−1 , 100s−1 , 1000s−1 }.
Case 1: Maximum distance −10 µm, minimum drift velocity = 1 µm/s and rnm
= 75 nm.
Case 2: Maximum distance −10 µm, minimum drift velocity = 1 µm/s and rnm
= 100 nm (Fig. 9).
Condition 7: Maximum distance = 10 µm and Maximum drift velocity = 3 µm.
PDF of latency for different shear rates {1s−1 , 10s−1 , 50s−1 , 100s−1 , 1000s−1 }.
Case 1: Maximum distance −10 µm, Maximum drift velocity = 3 µm/s and rnm
= 75 nm.
Case 2: Maximum distance −10 µm, Maximum drift velocity = 3 µm/s and rnm
= 100 nm (Fig. 10).
The latency is strongly affected by the shear rate of the blood medium. When the
shear rate of the blood medium increases, propagation time is more for the propa-
gating molecules. After the distance between the Nano-Tx and Nano-Rx increases,
the propagation time is always more for each shear rate of the blood at the same
temperature. From the above cases, it is very clear that case 2 has a greater.
Probability to reach the receiver at very low time delay. Hence blood is an efficient
medium for propagation of molecules during lower temperature and minimum shear
rate.

3.4 Comparison of Higher PDF Rate

In this section, the higher PDF rate obtained from different conditions using Inverse
Gaussian distribution methodology is compared for two different Liposome radii.
 Propagation Model of Molecular Communication Based …

Fig. 7 PDF of latency for different shear rates {1s−1 , 10s−1 , 50s−1 , 100s−1 , 1000s−1 } Minimum distance 2 µm and Minimum drift velocity 1 µm
511
 512

Fig. 8 PDF of latency for different shear rates {1s−1 , 10s−1 , 50s−1 , 100s−1 , 1000s−1 } for minimum distance 2 µm and Maximum drift velocity 3 µm
P. Murugesan et al.
 Propagation Model of Molecular Communication Based …

Fig. 9 Maximum distance = 10 µm and Minimum drift velocity = 1 µm

513
 514

Fig. 10 Maximum distance = 10 µm and Maximum drift velocity = 3 µm

P. Murugesan et al.
Propagation Model of Molecular Communication Based … 515

Table 1 Higher PDF rate comparison

Condition Temp (K) Shear rate PDF value for case1 PDF value for case 2
1 277 1sˆ−1 1.066 1.128
2 277 1sˆ−1 5.378 6.192
3 277 1sˆ−1 2.378 2.747
4 277 1sˆ−1 1.065 1.217
5 277 1sˆ−1 4.446 6.192
6 277 1sˆ−1 0.4608 0.531
7 277 1sˆ−1 2.377 2.747

This comparison predicts the role of molecular radius inside the propagation
medium. Let consider two different radius of the Liposome molecule 75 nm and
100 nm respectively.
From the Table 1, it is very clear that Liposome molecule with radius of 100 nm
has a greater efficiency to propagate inside the blood medium with less time delay. By
comparing the higher probability values for both cases, the higher most probability
takes place in the case 2 for all the conditions. Hence, molecule with increasing
radius will play a major role in the propagation process in targeted drug delivery
system.

4 Conclusion

Molecular communication, in other words the use of molecules to deliver drugs

among Nano machines is new communication paradigm. The most important expec-
tation of molecular communication is to provide new solutions for the treatment
of cancer. This thesis provided a novelty by applying molecular communication
in Atherosclerosis disease therapy. The proposed mathematical model is evaluated
under various conditions and compared with two different radii of the liposome
molecules. The results shows that, the latency is highly dependent on parameters
such as distance, molecular radius, blood viscosity, blood shear rate and drift velocity.
Hence blood becomes an efficient medium at lower shear rates in decreased temper-
atures. Our future work is to apply this model for various drug carrying molecules
used in the treatment of the cardiovascular diseases such as Lipid nanoparticles,
polymeric nanoparticles Micelles, CNT etc., and to assess its propagation capacity
under various conditions of blood medium.
516 P. Murugesan et al.

References

1. Tiwari G, Tiwari R (2012) Drug delivery system—an updated review. Pharma Investig
2. Nasrdlahzadeh M, Sajadi SM, Sajjadi M, Issaabadi Z (2019) An introduction to nanotechnology.
Interface Sci Technol
3. Akyildiz IF, Jornet JM, Pierobon M (2010) Propagation models for nano communication
networks. In: Proceedings of the fourth European conference on antennas and propagation.
pp 15–19
4. Kadloor S, Adve RS, Eckford AW (2012) Molecular communication using Brownian motion
with drift. IEEE Trans Nanobiosci 11(2):89–99
5. Berg HC (1993) Random walks in biology. Press Princeton University Press, Princeton
6. Kuran MŞ, Yılmaz HB, Tugcu T, Özerman B (2010) Energy model for communication via
diffusion in nanonetworks. Nano Commun Netw 1(2):86–95
7. https://www.pharmatutor.org/articles/liposomes-novel-drugdelivery-carrier
8. https://holisticprimarycare.net/topics/topicsn/healthy-aging/1299bloodviscosity-the-unifying-
parameter-in-cardiovascular-diseaserisk.html
Adaptation of Machine Learning
and Blockchain Technology
in Cyber-Physical System Applications:
A Concept Paper

Mujaheed Abdullahi, Hitham Alhussian, and Norshakirah Aziz

Abstract In recent years, Cyber-Physical Systems (CPS) have been adopted in

various sectors such as smart cities, smart industries etc. These types of systems
continuously generate a huge amount of data which increasingly attract cyber-
crimes. There are several existing approaches produced to overcome these issues
by using Blockchain Technology (BT) such as Public, Private, Construme, Hybrid
Blockchain-based on CPS applications and Machine Learning (ML) such as Support
Vector Machine (SVM), Linear Regression, and Decision Tree etc. With the rapid
increase in data size affix with cyber-crimes, such approaches become less effective
and therefore necessitate the invention of a more robust and self-trainable approach.
In this paper, we presented brief details on ML and BT and how they can be adopted
in CPS applications to solve security issues concerning cyber-crimes. The archi-
tecture was also presented to depict the proposed method. Moreover, technolo-
gies/techniques which can be implemented in CPS applications are discovered such
as industrial automation, smart buildings, medical systems, and vehicular systems.
We also have some future scope and conclusion.

Keywords Cyber-physical system (CPS) · Machine learning (ML) · Blockchain

technology (BT) · Security · Privacy

M. Abdullahi (B) · H. Alhussian · N. Aziz

Department of Computer and Information Sciences, Universiti Teknologi PETRONAS, Seri
Iskandar, Malaysia
e-mail: [email protected]
H. Alhussian
e-mail: [email protected]
N. Aziz
e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 517
R. Ibrahim et al. (eds.), International Conference on Artificial Intelligence
for Smart Community, Lecture Notes in Electrical Engineering 758,
https://doi.org/10.1007/978-981-16-2183-3_48
518 M. Abdullahi et al.

1 Introduction

From past few years, cybersecurity has become an increasing demand in computa-
tions devices, digital transformation in Cyber-Physical System (CPS) such as smart
manufacturing, smart homes, smart transportation and smart factory, where trust
has to be crucial aspect [1]. Those devices use wireless network as a means of
communication to broadcast their data, hence it is easy to be attacked during normal
network communication [2, 3]. Cybersecurity [4] was major concerned because the
technology moving is moving to industry 4.0 [5] where a large amount of data are
generated by CPS and others smart devices such as IoT [6] are usually used for
real-time monitoring and predictions failure [7, 8]. Many cyber attacked has been
accoutered in past few days which include Ukraine power plant in 2016 and Stuxnet
worm which target a nuclear power plant, also there is cyber-attacked on Australia’s
Maroochy water services in 2002 [9]. Also, those CPS applications are generating
a huge amount of data, storage this communications process is vulnerable against
attack. To overcome this problem Machine Learning and Blockchain can be used,
BT has a distributed database network with time series stamped of tamper-proof
record, while ML will analysis the data traffic through effective, efficient designing
of algorithms [10, 11].
Machine Learning (ML) in the field of Artificial Intelligence focus on building
applications that learn through experience and they make self-decision without
human interference such as Logistic Regression, Linear Regression, and Support
Vector Machine etc. ML was applied to various field of discipline such as statistics,
computational complexity [12, 13] smart devices [14] internet of things [15, 16] and
artificial intelligence to machines are getting smarter [17, 18]. While Blockchain
Technology (BT) are cryptography immutable set of records or distributed ledger
which is link together for auditing, by linking its valid blocks together the current
block contains the hash of the previous block. This process makes the Blockchain
traceable and resistant to change [10, 19].
Cyber-Physical System (CPS) was introduced by Helen Gill in 2006 at National
Science Foundation, they are integrated of both physical and cyber to enhance
the control and operation which include computational, communication, sensing
and actuation, in general, they are dynamics joint of computers, software, phys-
ical process, event processing [20] and networks. The CPS includes operating in
real-time, transactional services and secure communication between components
[21].

1.1 Adaptation of ML and BT in CPS

The ML processes can be applied to BT for CPS applications to improve their security,
privacy, monitoring [22] and making them smarter. It also gives an opportunity with
improvement in developing good models by using the decentralized architecture of
Adaptation of Machine Learning and Blockchain Technology … 519

BT [23]. We also proposed architecture for ML adaptation in BT for CPS applications

as shown in fig, whereby after the data was given which was generated by the CPS.
The BT will be included as an integral of the applications. At this point, the ML will
be applied for data analytics and real-time predictions [24] to reduce error and detect
anomaly activities [25].

1.2 Research Contribution

There are several research works which address the importance of combining ML
and BT, in this paper, we investigate how ML can be adopted in BT based on CPS
applications. The research contribution for this paper include the followings:
• We find out that ML and BT can be adopted together and a brief discussion on
ML with BT can be used together in CPS applications.
• We also proposed architecture on how ML and BT can be adopted and
implemented together.

2 Core Concepts and Rationale

The core concept and rationale of ML algorithms with BT can be adopted and imple-
mented in CPS applications to improve their performance in terms of security, moni-
toring and detection activities such as improved security and privacy were ML and
BT covers CPS applications such as data protection, data encryption, instruction
detection system (IDS), intrusion protection system (IPS). They provide an impor-
tant role in the monitoring and detection of malware. Furthermore, BT increased
the privacy between users for both private and public Blockchain through secure
authentication of data accessibility.
Improved efficiency and scalability increasing ability in multiple of CPS appli-
cations from multiple devices that consist of huge datasets due to their multiple
transactions through providing decentralize self-authorization, detection from ML.
Scalability form BT also provides capable of increment between users.
Secure of data sharing in CPS Applications was the BT contains chains blocks of
all data transaction for CPS applications by providing accountability, transparency
within the data access. ML components were included with its algorithms to provide
decentralised platform and removal of fraud fewer datasets.

3 Research Question and Methodology

In this section research question and methodology has been discussed, the research
question was (Can Machine Learning and Blockchain Technology be adopted
520 M. Abdullahi et al.

together in Cyber-Physical Systems Applications?). The question has been anal-

yses using a systematic literature review by using PRISMA method, we also provide
a comparison study from previous research.

3.1 Systematic Literature Review

PRISMA guideline was used to structure the systematic literature review by creating a
question to guide the review process (Can Machine Learning and Blockchain Tech-
nology be adopted together in Cyber-Physical Systems Application?). The search
strategy used for this systematic review, we developed a search strategy to identify
relevant literature. The search strategy was tailored by one database: Scope and search
terms used were the followings: “Machine Learning for Blockchain Applications”
and “Blockchain for Cyber-Physical Systems Application”. All search spanned from
the database inception were until 2020, and included journals articles with few review
papers published in English only.
The selection criteria were mainly focused on the mapping on existing literature on
Machine Learning for Blockchain Applications and Blockchain for Cyber-Physical
Systems Application in the field of computer science, Decision Science and math-
ematics. The search span was from 2020 to 2018. All articles before 2018 below
were excluded from the search. The also was mainly focused on the global search,
not a specific country or region. A total of 95 research articles were excluded at this
stage. There were 42 research article included and extracted at this stage. The quality
assessment is based on the original research and a few review papers. For maintain
the quality of the review, all duplicated were checked thoroughly.

3.2 Methodology

The methodology investigates process, steps to overcome this problem with rational
for CPS applications, by using a combination of two technologies which are ML and
BT with a proposed framework to support this research study (see Fig. 1).

4 Findings

Firstly, providing detection with prevention in CPS applications to handle data secu-
rity and cybersecurity issues within its application. Those methods are intrusion
detection system (IDS), intrusion prevention system (IPS) with ML classifiers. BT
also add additional support through linked blocks using cryptography for prevention
of datasets against attack.
Adaptation of Machine Learning and Blockchain Technology … 521

Internet of
Smart Devices
Things

Cyber Physical Systems

ApplicaƟons

Create Add to ExisƟng

VerificaƟon Verified
New Block Block

New
Block 1 Block 2 Block 3
Block
User
Blockchain

Machine Learning Model

Real-Ɵme AnalyƟcs Data AnalyƟcs

Fig. 1 Proposed architecture for ML and BT adaptation in CPS applications

Secondly, prediction with a response in CPS applications most methods used

for prediction of ML algorithms whereby providing security, improving the right
decision in datasets and data analysis. They also involve failures prediction of appli-
cations with taking security majors for necessary actions in terms of response. The
ML response through anomalies activities. Finally, they allow the CPS applications
to make self-decision to command and action using security policies.
Lastly, providing monitoring in CPS applications by using ML and BT through
verification of modelled datasets and external control. The CPS applications network
events is been analyzed which include input data. All records from events and datasets
are monitored in real-time control.

5 Conclusion and Future Scope

We conclude, that recent development of CPS application with their huge increasing
data generated this leads to the increase of cyber-attack. With the development of
ML and BT are helping to overcome these issues by adopting them together in
522 M. Abdullahi et al.

various CPS applications such as power systems, medical systems, smart buildings
and industrials automation. The research paper also presents brief details on ML
and BT with how they can be adopted together in CPS applications using a proposed
architecture. A study comparison has been made which shows methods, contributions
and limitation. We also have some suggestion for further scope.
The future scope on this topic is therefore recommended to provide in details on
how ML and BT can be adopted into specific CPS applications such as Industrial
Automation, Power Systems and Medical Systems or Internet of Things devices.

References

1. Kochovski P, Gec S, Stankovski V, Bajec M, Drobintsev PD (2019) Trust management in a

blockchain based fog computing platform with trustless smart oracles. Futur Gener Comput
Syst 101:747–759
2. Hasan M, Islam MM, Zarif MII, Hashem MMA (2019) Attack and anomaly detection in IoT
sensors in IoT sites using machine learning approaches. Internet Things 7:100059
3. Li D, Deng L, Liu W, Su Q (2020) Improving communication precision of IoT through behavior-
based learning in smart city environment. Futur Gener Comput Syst 108:512–520
4. Sri Devi R, Mohan Kumar M (2019) Cyber security affairs in empowering technologies. Int J
Innov Technol Explor Eng 8(10):1–7
5. O’Donovan P, Gallagher C, Leahy K, O’Sullivan DTJ (2019) A comparison of fog and cloud
computing cyber-physical interfaces for Industry 4.0 real-time embedded machine learning
engineering applications. Comput Ind 110:12–35
6. Jia P, Wang X, Zheng K (2020) Distributed clock synchronization based on intelligent clustering
in local area industrial IoT systems. IEEE Trans Ind Inform 16(6):3697–3707
7. Lv L, Wang W, Zhang Z, Liu X (2020) A novel intrusion detection system based on an optimal
hybrid kernel extreme learning machine. Knowl-Based Syst 195:1–17
8. Junejo KN (2020) Predictive safety assessment for storage tanks of water cyber physical systems
using machine learning. Sadhana—Acad Proc Eng Sci 45(1):1–6
9. Goh J, Adepu S, Tan M, Shan LZ (2017) Anomaly detection in cyber physical systems using
recurrent neural networks. IEEE Comput Soc 36:1–7
10. Tanwar S, Bhatia Q, Patel P, Kumari A, Singh PK, Hong WC (2020) Machine learning adoption
in blockchain-based smart applications: the challenges, and a way forward. IEEE Access 8:474–
488
11. Gupta S, Sabitha AS, Punhani R (2019) Cyber security threat intelligence using data mining
techniques and artificial intelligence. Int J Recent Technol Eng 8(3):6133–6140
12. Rahman M (2019) Cultural differences, strengths, weaknesses and challenges of Kentucky
fried chicken (KFC)—a comparison between United Kingdom (UK) and Malaysia. Double
Blind Peer Rev Int Res J Publ 19(1):7
13. Sharma R, Kamble SS, Gunasekaran A, Kumar V, Kumar A (2020) A systematic literature
review on machine learning applications for sustainable agriculture supply chain performance.
Comput Oper Res 119:104926
14. Gupta M, Abdelsalam M, Khorsandroo S, Mittal S (2020) Security and privacy in smart farming:
challenges and opportunities. IEEE Access 8:34564–34584
15. Shafiq M, Tian Z, Sun Y, Du X, Guizani M (2020) Selection of effective machine learning
algorithm and Bot-IoT attacks traffic identification for internet of things in smart city. Futur
Gener Comput Syst 107:433–442
16. Viejo A, Sánchez D (2020) Secure monitoring in IoT-based services via fog orchestration.
Futur Gener Comput Syst 107:443–457
Adaptation of Machine Learning and Blockchain Technology … 523

17. Bello A, Mahadevan V (2019) A cloud based conceptual identity management model for
secured Internet of Things operation. J Cyber Secur Mobil 8(1):53–74
18. Al-Saud M, Eltamaly AM, Mohamed MA, Kavousi-Fard A (2020) An intelligent data-driven
model to secure intravehicle communications based on machine learning. IEEE Trans Ind
Electron 67(6):5112–5119
19. Islam A, Young Shin S (2020) A blockchain-based secure healthcare scheme with the assistance
of unmanned aerial vehicle in Internet of Things. Comput Electr Eng 84:1–12
20. Roldán J, Boubeta-Puig J, Luis Martínez J, Ortiz G (2020) Integrating complex event processing
and machine learning: an intelligent architecture for detecting IoT security attacks. Expert Syst
Appl 149:1–22
21. Al-Hadhrami Y, Hussain FK (2020) Real time dataset generation framework for intrusion
detection systems in IoT. Futur Gener Comput Syst 108:1–10
22. Zhong CL, le Li Y (2020) Internet of things sensors assisted physical activity recognition and
health monitoring of college students. Meas J Int Meas Confed 159:1–9
23. Lu Y, Huang X, Dai Y, Maharjan S, Zhang Y (2020) Blockchain and federated learning for
privacy-preserved data sharing in industrial IoT. IEEE Trans Ind Inform 16(6):4177–4186
24. Lima E, Vieira T, de Barros Costa E (2020) Evaluating deep models for absenteeism prediction
of public security agents. Appl Soft Comput J 91:106236
25. Lin X, Li J, Wu J, Liang H, Yang W (2019) Making knowledge tradable in edge-AI enabled
IoT: a consortium blockchain-based efficient and incentive approach. IEEE Trans Ind Inform
15(12):6367–6378
Application of Artificial Neural Network
(ANN) and Adaptive Neuro Fuzzy
(ANFIS) Techniques for the Modelling
and Optimization of COD Adsorption
Process

Hifsa Khurshid, Muhammad Raza Ul Mustafa, and Yeek-Chia Ho

Abstract Artificial neural network (ANN) and adaptive neuro fuzzy (ANFIS)
modelling techniques have been applied in this study to model and optimize the chem-
ical oxygen demand (COD) adsorptive removal in produced water. The models were
well trained and showed minimum error values for predicted data when compared
to experimental data. The error values were 0.4035 and 0.2886 for sum of squared
error (SSE), 0.1628 and 0.0832 for mean square error (MSE) and 0.13 and 0.23%
for average relative error (ARE) using ANN and ANFIS, respectively. Error anal-
ysis and coefficient of determination (R2 ) of the models determined that ANFIS was
better than ANN for the prediction of COD adsorption on the biochar. Also, ANFIS
required minimum run time as compared to ANN. Both artificial intelligence (AI)
based techniques well predicted the optimized values of adsorption process, when
compared with the experimental values. It is concluded that the use of AI tech-
niques can inevitably pave the way in the water treatment sector using adsorption for
improved efficiency and process automation.

Keywords Adsorption · Artificial intelligence · Optimization · Automation ·

Water treatment · Green technology

1 Introduction

Adsorption is an efficient and economical method of elimination of pollutants from

wastewater and can be further improved by automation. The automation of wastew-
ater treatment systems is necessary to making it easier to run, saving both manpower
and resources [1]. However, it is very difficult to build an automatic wastewater
treatment plant when the attributes of commercial wastewater change significantly,
resulting in a drastic shift in the performance of the treatment plant. In fact, its

H. Khurshid (B) · M. R. U. Mustafa · Y.-C. Ho

Department of Civil and Environmental Engineering, Universiti Teknologi PETRONAS, 32610
Seri Iskandar, Perak Darul Ridzuan, Malaysia
e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 525
R. Ibrahim et al. (eds.), International Conference on Artificial Intelligence
for Smart Community, Lecture Notes in Electrical Engineering 758,
https://doi.org/10.1007/978-981-16-2183-3_49
526 H. Khurshid et al.

complex properties are very difficult to handle when it comes to the adsorption treat-
ment process [2]. Due to the interaction of several factors in adsorption process,
the resulting correlations are strongly non-linear [3]. This problem leads to inade-
quate understanding and false quantification of interactions between process inputs
and outputs, and it is thus difficult to explain by linear mathematical models. The
standard regression methods are also not sufficient for mimicking the experimental
adsorption data as the methodology does not grasp the physics [1]. For interrelating
the input and output parameters of adsorption process, advanced computer-simulated
models are therefore required [4].
Artificial intelligence (AI) methods e.g. fuzzy logic (FIS) [5], particle swarm opti-
mization (PSO) [6], genetic algorithm (GA) [7] and artificial neural networks (ANNs)
[8], have been applied recently in water treatment and adsorption optimization and
automation systems. ANNs are particularly well known for their high capacity to
research and arrange data [9]. Researchers have reported different studies using
ANN for modelling and optimization of adsorptive removal of various pollutants
e.g. chromium removal using cyanobacterial biomass [10], copper removal using
light expended clay aggregate [11], removal of various dyes using nanoparticles and
activated carbons [12, 13], removal of As (III) and various other metals [5]. Fuzzy
systems also have some advantages over traditional approaches, particularly where
ambiguous data or prior experience is involved [14]. Few studies have reported appli-
cation of ANFIS in adsorption studies for modeling and optimization of adsorption
process e.g. indium (III) [15], dyes and metals [5]. Meta-heuristic optimization algo-
rithms e.g. GA and PSO, are the products of social actions or natural phenomena
and used to optimize the various operations. The PSO is a well-known optimization
algorithm motivated by the action of a flock of birds. It is now widely used in system’s
optimization because it is not conveniently stuck in a local minimum.
Most of the studies in literature are performed on synthetic waters or stock solu-
tions containing single pollutant as an objective. These solutions do not completely
represent the industrial effluents, which contain multiple types of pollutants together.
Hence, practical application of the adsorptive materials and automation of treat-
ment plants becomes limited considering single pollutant studies. Therefore, the
purpose of this research work was to present the design, implementation, compar-
ison and evaluation of the ANN and ANFIS approaches to the adsorption of chemical
oxygen demand (COD) in produced water (PW). The wastewater obtained during
oil and gas exploration is known as PW. The COD in the water is caused by various
carcinogenic organic pollutants. Organic pollutants have been confirmed to be one
of the top priority pollutants to date including PAHs, dyes, and pharmaceuticals etc.
COD is an important parameter for interpreting the organic pollutants in wastewater
and has been used widely for effluents quality measurement [16]. Various studies
have reported efficient adsorption of COD in wastewater which also represents the
elimination of organic pollutants in the water [17, 18].
To the best of our knowledge, no study has been found in literature containing the
application of ANN and ANFIS for modelling and optimization of COD adsorption
process for predicting organic pollutants removal in PW. Tea waste (TW) had been
transformed into biochar for this purpose and was applied under the control of various
Application of Artificial Neural Network (ANN) … 527

parameters (pH, dosage and contact time). To improve the automated adsorptive
removal of COD from the wastewater sample, adsorption process was optimized
using the ANN-PSO and ANFIS techniques. The detailed objectives of this study
are: to establish ANN and ANFIS models capable of predicting adsorption efficiency
for the removal of COD from PW; Optimizing the COD adsorption process using
ANN-PSO and ANFIS; and to incorporate ANN and ANFIS models such that they
can be used effectively for PW treatment system automation. This study will help
the decision makers and wastewater treatment plant operators in decision making,
strategic designing of system, automation, and optimization of adsorption process.

2 Materials and Methods

2.1 Materials

Tea waste was obtained from the nearby restaurants in Malaysia. High range (HR)
COD vials were obtained from the laboratory stock of Avanti’s Malaysia. About 98%,
H3 PO4 was purchased in Malaysia from local supplier Merck. The water sample was
obtained from an oil and gas company operating in South East Asia.

2.2 Preparation of Adsorbent

The biochar was produced using the chemical activation technique. For 12 h, tea
waste was immersed in 10% phosphoric acid solution and then left to dry at room
temperature. Biochar was developed by pyrolysis at 700 °C using the dried waste
in the tube furnace. Afterward, the prepared biochar was cleaned with filtered water
and kept in the oven to dry. In this study, the biochar was only used for exploring the
application of AI based techniques in automation of adsorption process and mecha-
nism evaluation was not the objective. Therefor the material was not characterized
and used only for the experimental purpose.

2.3 Adsorption Batch Experiments

In this analysis, the commonly documented adsorption process was used to observe
the elimination of COD [19]. A significant segment of AI configuration is the selec-
tion of input variables from a variety of parameters that influence the mechanism
[1]. To select different important factors affecting the adsorptive removal of COD
from PW, batch adsorption tests were performed. From the experiments and litera-
ture review, it was found that initial pH of water, dosage of adsorbent and contact
528 H. Khurshid et al.

time were most significant input parameters effecting the adsorption process. There-
fore, the biochar was applied in varying dosage (25–300 mg/L) for 100 mL of PW at
varying initial pH (3–12) and shook for various contact times (10–60 min) at temper-
ature 20 ± 5 °C. The percentage of COD removal was detected by the formula given
[20]:

(Ci − Co ) × 100
Percentage removal (% ) = (1)
Ci

where, Ci and Co represent the initial and final concentrations of COD in mg/L,
respectively.

2.4 Artificial Neural Network (ANN)

A three-layer feed forward neural network was formed using the neural network
toolbox using programming codes written in MATLAB R2019b mathematical soft-
ware. The input layer was given three inputs i.e. pH, contact time and dosage of
biochar and one output e.g. removal efficiency of COD as shown in Fig. 1. Number
of neurons in hidden layer were selected based on hit and trial method using 1–10
neurons. Higher number of neurons were not tested to avoid the over fitting of model.
Cascade forward backpropagation (CFBP), Levenberg–Marquardt backpropagation
(LMBP), Bayesian Regularization (BR) and Scaled Conjugate Gradient (SCG) algo-
rithms were tested to find the best suitable algorithm. The most suitable algorithm

Fig. 1 Schematic diagram of ANN model

Application of Artificial Neural Network (ANN) … 529

was selected based on minimum value of mean square error (MSE). For all data sets
in ANN, the symmetric sigmoid transfer function (tansig) was used in the first layer
and the linear transfer function (purelin) was used at the output node for the simula-
tion and prediction of COD elimination. The backpropagation (BP-ANN) network
was selected for the tuning of parameters. It is a first-order gradient descent technique
to model experimental data [21]. Total 36 no. of data sets were taken through batch
experiments. 70% data was used for training of model and 30% data was used for
testing purpose. For validation of model and prediction of COD removal efficiency
13 more data values were taken through lab experiments using Box–Behnken Design
(BBD) of experiments.

2.5 ANN-PSO Hybrid

ANN and PSO hybrid was used for the optimization of input variables using ANN
predicted data with 1000 no of population, C1 = 1.5, C2 = 2.5 and 1000 iteration
values. The schematic diagram of hybrid model is shown in Fig. 2. MSE was taken
as function to be minimized. Tolerance value was taken as 1. Boundary violations
were taken as 1.5 and −1.5. The algorithm was run for optimization further. The

Fig. 2 ANN-PSO hybrid

530 H. Khurshid et al.

Fig. 3 Schematic diagram of ANFIS model

purpose was to obtain the input variables causing the maximum removal of COD in
PW.

2.6 Adaptive Neuro-Fuzzy Interference Systems (ANFIS)

The fuzzy inference method (FIS) applied to ANFIS was developed in the MATLAB
using Neuro-Fuzzy designer. The developed ANFIS was of the Sugeno type, which
consist of four hidden layers [15]: (i) the fuzzification layer with two membership
functions; (ii) the inference layer with weight rules; (iii) the de-fuzzification layer; and
(iv) the output layer (Fig. 3). Data was randomly divided into training (70%) and
testing data (30%). Three variables (pH, Dosage and contact time) were selected
as inputs and removal efficiency of COD was taken as target. Range of influence
was selected based on minimum MSE. To obtain the optimized results for COD
adsorption using ANFIS, surface plots were generated.

2.7 Error Analysis

The performance of the ANN and ANFIS techniques for predicted data was evaluated
using statistical equations i.e. (i) the coefficient of determination (R2 ), (ii) the sum
of squared errors (SSE), (iii) mean–squared error (MSE) and (iv) average relative
error (ARE) [22]. Following equations were used to measure the errors [23]:
Application of Artificial Neural Network (ANN) … 531

n  2
i=1 y pr ed,i − yexp,i
R =1−
2
n  2 (2)
i=1 y pr ed,i − ym

 n
 1  2
SS E =  y pr ed,i − yexp,i (3)
n i=1

1  2
n
MSE = y pr ed,i − yexp,i (4)
n i=1

100%  y pr ed,i − yexp,i

n
ARE = (5)
n i=1 yexp,i

where, ypred and yexp denote the predicted and experimental values, respectively. n
denotes the total number of values in data. Mean value of the response is denoted by
ym .

3 Results and Discussion

3.1 Modelling of ANN

The ANN structure was opted after analyzing the various parameters i.e. no. of
neurons and training algorithm. The algorithm and no. of neurons were selected
based on the best performance as shown in Fig. 4a, b. MSE was taken as perfor-
mance criteria. Total 10 no. of neurons and LMBP algorithm were selected based on
minimum MSE and simulation time. After selecting the structure of ANN, selected
model was trained by using trainlm training function. It was observed that, after
training, the output data had a correlation coefficient (R2 ) of 0.99 with the exper-
imental data (Fig. 4c). The MSE value was 0.0001 and training time was 400 s.
The lower value of error indicated efficient performance of selected ANN model for
the given dataset. For the validation and testing of model, correlation coefficients of
0.998 and 0.98 were obtained respectively (Fig. 4d).
Higher values of R2 showed that the model was well trained and could be used
effectively for the prediction of COD adsorption on biochar. The model was validated
further and a correlation coefficient of 0.99 was obtained for the predicted data when
compared with experimental data using 13 dataset points, as given in Table 1.
532 H. Khurshid et al.

Fig. 4 Modelling of ANN; a relationship between MSE and no. of neurons in hidden layer, b
relationship between CFBP, LMBP, BR and SCG algorithms and their performance, c ANN training,
d ANN testing

3.2 Optimization of COD Adsorption Process Using

ANN-PSO

The PSO technique was hybridized with the generated ANN model for optimizing
process parameters with the goal of maximizing the COD removal efficiency. The
performance of ANN-PSO model was analyzed on basis of MSE i.e., 5.27. Minimum
value was obtained after 500 runs and it took time of 4000 s. The optimum conditions
for the COD removal process were: 11.79 pH, 300 mg/L concentration of biochar
and 60 min contact time. The COD removal efficiency obtained under optimum
conditions was 89.80%. The value was confirmed in the lab and obtained as 89.3%.
It indicated the good performance of ANN-PSO in predicting the COD adsorption
optimized process parameters.
Application of Artificial Neural Network (ANN) … 533

Table 1 Experimental and predicted results of COD removal efficiency using ANN and ANFIS
No pH Dosage Contact time COD removal efficiency (%)
mg/L min Actual Predicted ANN Predicted ANFIS
1 10 162.5 60 53.123 52.31 52.123
2 3 162.5 60 36.611 36.12 36.59
3 6.5 25 60 69.269 69.01 69.16
4 6.5 300 60 78.044 78.77 78.12
5 6.5 162.5 35 63.918 64.66 63.91
6 10 25 35 54.296 54.21 54.3
7 6.5 25 10 65.574 65.57 65.57
8 3 162.5 10 36.892 36.72 36.89
9 3 25 35 39.169 39.23 39.18
10 10 162.5 10 44.941 45.01 44.9
11 3 300 35 49.431 49.43 49.39
12 10 300 35 60.831 60.85 61.02
13 6.5 300 10 72.958 72.8 72.8
R2 0.9991 0.9995
SSE 0.4035 0.2886
MSE 0.1628 0.0832
ARE 0.13% 0.23%

3.3 Modelling of ANFIS

Sugeno type sub-clustering ANFIS was generated for three inputs and one output.
For all input variables membership functions were taken as 35 in number and guassmf
type. Range of influence was selected as 0.00001 based on minimum MSE value,
when tested from 0–1 as shown in Fig. 5a. Whereas, squash factor, accept ratio and
rejection ratio were 1.25, 0.5 and 0.15, respectively. For training of generated FIS
structure hybrid optimization method was taken as method type due to lower MSE
(0.15721) as compared to backpropagation optimization method (2.32). Total 100 no
of epochs were taken for training of ANFIS model using 70% dataset. The model was
well trained and minimum MSE of 0.15721 was obtained at epoch 2 (Fig. 5b). The
run time of ANFIS was only 10 s for 100 epochs which was much lesser than ANN.
The model was further tested and checked using the 30% data. The MSE of 0.00017
and 0.0008 was obtained for the test and checked data, respectively (Fig. 5c, d).
After training, the ANFIS model was used for prediction of COD removal efficiency
using 13 input datasets. The predicted outputs were compared with the experimental
data as given in Table 1. A higher correlation of 0.9995 and lower MSE of 0.0832
was obtained for the predicted data. It showed that generated ANFIS model could
be used effectively for the optimization of COD adsorption process and selecting
suitable input parameters.
534 H. Khurshid et al.

Fig. 5 Modelling and contour plots for COD adsorption on biochar using ANFIS; a error analysis,
b ANFIS training data, c ANFIS testing data, d ANFIS checked data, e impact of pH and dosage
on COD removal efficiency, f impact of pH and contact time on COD removal efficiency, g impact
of dosage and contact time on COD removal efficiency

3.4 Optimization of COD Adsorption Process Using ANFIS

For the prediction of optimized values of adsorption data through ANFIS surface
plots were generated for pH range 3–12, dosage 25–300 mg/L and contact time 10–
60 min. Figure 5e–g shows that COD removal efficiency was highly affected by all the
three variables. It also shows the correlation between each variable. Maximum value
of COD removal was obtained at pH value of 12, dosage of 300 mg/L and contact
time of 60 min. The COD removal efficiency achieved at the optimized parameters
was 90.8%. The experimental value for the optimized input parameters was 89.3%.
It showed that the model was able to predict the optimized values with significant
accuracy.

3.5 Comparison of ANN and ANFIS

Developed ANN and ANFIS models were well trained for predicting the COD
removal efficiency using pH, dosage of biochar and contact time as input variables.
Both models well fitted the experimental data and predicted values were near to
Application of Artificial Neural Network (ANN) … 535

Table 2 Optimized values

Parameters Experimental ANN-PSO ANFIS
for COD adsorption
pH 12 11.79 12
Dosage (mg/L) 300 300 300
Contact time (min) 60 60 60
COD removal efficiency 89.30 89.80 90.80
(%)

experimental data (Table 1). Whereas ANN-PSO and ANFIS models also performed
very well for the optimization of adsorption process (Table 2). Further performance
of the models was analyzed using Eqs. 2–5 for error analysis of predicted data. The
analysis is given in Table 1. The value of R2 should lie between 0–1. Value near
to 1 shows good correlation between the data sets. SSE has values range from 0–1,
while best value is closer to 0 [4]. MSE minimum value is taken as better value
[24], whereas ARE is %age error and minimum value is recommended [25]. Table
1 shows that ANN and ANFIS had R2 values of 0.9991 and 0.9995, error values
of 0.4035 and 0.2886 for SSE, 0.1628 and 0.0832 for MSE and 0.13% and 0.23%
for ARE, respectively. Error analysis and coefficient of determination of the models
determined that ANFIS > ANN in performance for the prediction of COD adsorption
on the biochar. Also, for both methods no specific design of experiments was needed.
It helped the models to analyze the correlation between inputs and outputs in broader
range.

4 Conclusion

Tea waste biochar was synthesized and applied for treatment of produced water. Batch
tests were performed using pH, dosage, and contact time as input variables. The exper-
imental results were modelled and optimized using ANN and ANFIS models. ANN
model was modelled using 3 layered, feed forward Levenberg–Marquardt backprop-
agation algorithm and 10 no. of neurons. For optimization it was hybridized with
PSO. ANFIS was generated sugeno type using sub-clustering FIS type. The exper-
imental values were successfully optimized using surface plots correlations. Opti-
mization values were well matched with the experimental results. ANFIS required
minimum run time (10 s for 100 epoch) as compared to ANN model (400 s for
1000 epoch). Error analysis and coefficient of determination of three models deter-
mined that ANFIS was slightly better than ANN in performance for the prediction of
COD adsorption on the biochar. However, it can be stated that the ANN and ANFIS
methods are promising techniques and can be used effectively for the automation of
COD adsorption process for produced water treatment. Both methods have high capa-
bility to predict the adsorption efficiency of pollutants considering multiple factors
and can be implemented for sustainable and automated water treatment systems.
536 H. Khurshid et al.

Acknowledgements The authors would like to acknowledge the financial support provided for
this study under YUTP grant with cost center 015LC0-190.

References

1. Anupam K, Dutta S, Bhattacharjee C, Datta S (2016) Artificial neural network modelling

for removal of chromium (VI) from wastewater using physisorption onto powdered activated
carbon. Desalin Water Treat 57(8):3632–3641. https://doi.org/10.1080/19443994.2014.987172
2. Al Aani S, Bonny T, Hasan SW, Hilal N (2019) Can machine language and artificial intelligence
revolutionize process automation for water treatment and desalination? Desalination 458:84–
96. https://doi.org/10.1016/j.desal.2019.02.005
3. Khan T, Mustafa MRU, Isa MH, Manan TSBA, Ho YC, Lim JW et al (2017) Artificial neural
network (ANN) for modelling adsorption of lead (Pb (II)) from aqueous solution. Water Air
Soil Pollut Water 228(11). https://doi.org/10.1007/s11270-017-3613-0
4. Dutta S, Parsons SA, Bhattacharjee C, Bandhyopadhyay S, Datta S (2010) Development of
an artificial neural network model for adsorption and photocatalysis of reactive dye on TiO2
surface. Expert Syst Appl 37(12):8634–8638. https://doi.org/10.1016/j.eswa.2010.06.090
5. Dolatabadi M, Mehrabpour M, Esfandyari M, Alidadi H, Davoudi M (2018) Modeling of
simultaneous adsorption of dye and metal ion by sawdust from aqueous solution using of
ANN and ANFIS. Chemom Intell Lab Syst 181:72–78. https://doi.org/10.1016/j.chemolab.
2018.07.012
6. Fan M, Hu J, Cao R, Xiong K, Wei X (2017) Modeling and prediction of copper removal from
aqueous solutions by nZVI/rGO magnetic nanocomposites using ANN-GA and ANN-PSO.
Sci Rep 7(1):1–14. https://doi.org/10.1038/s41598-017-18223-y
7. Bhatti MS, Kapoor D, Kalia RK, Reddy AS, Thukral AK (2011) RSM and ANN modeling
for electrocoagulation of copper from simulated wastewater: multi objective optimization
using genetic algorithm approach. Desalination 274(1–3):74–80. https://doi.org/10.1016/j.
desal.2011.01.083
8. Mortula M, Abdalla J, Ghadban A (2010) Modeling phosphorus removal process using arti-
ficial neural network modeling approach. In: BALWOIS 2010—Ohrid, pp 1–7. Republic of
Macedonia
9. Igwegbe CA, Mohmmadi L, Ahmadi S, Rahdar A, Khadkhodaiy D, Dehghani R (2019)
Modeling of adsorption of methylene blue dye on Ho-CaWO4 nanoparticles using response
surface methodology (RSM) and artificial neural network (ANN) techniques. MethodsX
6:1779–1797. https://doi.org/10.1016/j.mex.2019.07.016
10. Sen S, Nandi S, Dutta S (2018) Application of RSM and ANN for optimization and modeling of
biosorption of chromium(VI) using cyanobacterial biomass. Appl Water Sci 8(5):1–12. https://
doi.org/10.1007/s13201-018-0790-y
11. Shojaeimehr T, Rahimpour F, Khadivi MA, Sadeghi M (2014) A modeling study by response
surface methodology (RSM) and artificial neural network (ANN) on Cu2+ adsorption optimiza-
tion using light expended clay aggregate (LECA). J Ind Eng Chem 20(3):870–880. https://doi.
org/10.1016/j.jiec.2013.06.017
12. Ghaedi AM, Karamipour S, Vafaei A, Baneshi MM, Kiarostami V (2019) Optimization and
modeling of simultaneous ultrasound-assisted adsorption of ternary dyes using copper oxide
nanoparticles immobilized on activated carbon using response surface methodology and artifi-
cial neural network. Ultrason Sonochem 51:264–280. https://doi.org/10.1016/j.ultsonch.2018.
10.007
13. Gadekar MR, Ahammed MM (2019) Modelling dye removal by adsorption onto water treatment
residuals using combined response surface methodology-artificial neural network approach. J
Environ Manage 231:241–248. https://doi.org/10.1016/j.jenvman.2018.10.017
Application of Artificial Neural Network (ANN) … 537

14. Sargolzaei J, Haghighi Asl M, Hedayati Moghaddam A (2012) Membrane permeate flux and
rejection factor prediction using intelligent systems. Desalination 284:92–99. https://doi.org/
10.1016/j.desal.2011.08.041
15. Franco DSP, Duarte FA, Salau NPG, Dotto GL (2020) Analysis of indium (III) adsorption
from leachates of LCD screens using artificial neural networks (ANN) and adaptive neuro-
fuzzy inference systems (ANIFS). J Hazard Mater 384:121137. https://doi.org/10.1016/j.jha
zmat.2019.121137
16. Malakahmad A, Law MX, Ng KW, Manan TSA (2016) The fate and toxicity assessment of
polycyclic aromatic hydrocarbons (PAHs) in water streams of Malaysia. Procedia Eng 148:806–
811. https://doi.org/10.1016/j.proeng.2016.06.572
17. Mohammad-pajooh E, Turcios AE, Cuff G, Weichgrebe D, Rosenwinkel KH, Vedenyapina
MD et al (2018) Removal of inert COD and trace metals from stabilized landfill leachate by
granular activated carbon (GAC) adsorption. J Environ Manage 228:189–196. https://doi.org/
10.1016/j.jenvman.2018.09.020
18. Bansode RR, Losso JN, Marshall WE, Rao RM, Portier RJ (2004) Pecan shell-based granular
activated carbon for treatment of chemical oxygen demand (COD) in municipal wastewater.
Biores Technol 94(2):129–135. https://doi.org/10.1016/j.biortech.2003.12.009
19. Ademiluyi F, Amadi S, Amakama N (2010) Adsorption and treatment of organic contaminants
using activated carbon from waste Nigerian Bamboo. J Appl Sci Environ Manage 13(3). https://
doi.org/10.4314/jasem.v13i3.55351
20. Azari A, Mahmoudian MH, Niari MH, Eş I, Dehganifard E, Kiani A et al (2019) Rapid and
efficient ultrasonic assisted adsorption of diethyl phthalate onto FeIIFe2IIIO4@GO: ANN-GA
and RSM-DF modeling, isotherm, kinetic and mechanism study. Microchem J 150:104144.
https://doi.org/10.1016/j.microc.2019.104144
21. Olawoyin R (2016) Application of backpropagation artificial neural network prediction model
for the PAH bioremediation of polluted soil. Chemosphere 161:145–150. https://doi.org/10.
1016/j.chemosphere.2016.07.003
22. Dutta M, Basu JK (2013) Application of artificial neural network for prediction of Pb(II)
adsorption characteristics. Environ Sci Pollut Res 20(5):3322–3330. https://doi.org/10.1007/
s11356-012-1245-x
23. Khoshsang H, Ghaffarinejad A (2018) Rapid removal of lead (II) ions from aqueous solutions
by saffron flower waste as a green biosorbent. J Environ Chem Eng 6(5):6021–6027. https://
doi.org/10.1016/j.jece.2018.09.020
24. Dil EA, Ghaedi M, Ghaedi A, Asfaram A, Jamshidi M, Purkait MK (2016) Application of
artificial neural network and response surface methodology for the removal of crystal violet by
zinc oxide nanorods loaded on activate carbon: kinetics and equilibrium study. J Taiwan Inst
Chem Eng 59:210–220. https://doi.org/10.1016/j.jtice.2015.07.023
25. Singh DK, Verma DK, Singh Y, Hasan SH (2017) Preparation of CuO nanoparticles using
Tamarindus indica pulp extract for removal of As(III): optimization of adsorption process by
ANN-GA. J Environ Chem Eng 5(1):1302–1318. https://doi.org/10.1016/j.jece.2017.01.046
An Evolutionary Stream Clustering
Technique Outlier Detection in Medical
Data

Nadilah Ayu Supardi, Said Jadid Abdulkadir, and Norshakirah Aziz

Abstract Clustering data streams from data mining become evident as one of the
most well-liked studies among the researchers correspond to their evolutionary field.
Numerous threats addressed by data streams on clustering like limited time, memory
and single scan data. In generic terms, data streams are defining an infinite sequence
of the element data and evolve without prior knowledge number of the clusters. Many
factors such as noise (outliers) have appeared periodically have a negative impact
on data streams environment. The density-based technique has demonstrated to be
an astonishing method in clustering data streams. It is computationally competent to
generate arbitrary shape clusters and detect noise instantaneously. The number of the
clusters are not required in advance to set as a parameter during the assessment begin.
In contradictory, traditional density-based clustering is not relevant to conduct in data
streams due to its own characteristics. Mostly all traditional density-based clustering
can be extended to the updated version of algorithms to achieve the objective of data
streams research. The idea is emphasizing on the density-based technique in the clus-
tering process dominate the restrain from data streams nature. The objective of this
paper intends a preliminary result on a density-based algorithm, named evoStream to
explore outlier detection on medical data sets, heart failure clinical records and gene
expression cancer RNA-seq. In due course, extensive evoStream later to develop to
optimize the model to detect outlier in data streams.

Keywords Data streams · Outlier · Density-based · evoStream · Data sets ·

Evaluation metrices

1 Introduction

Due to the fact, advance applications are practically implemented in related to various
field such as medical research. Often, data streams relate with as an enormous,

N. A. Supardi (B) · S. J. Abdulkadir · N. Aziz

Computer and Information Science Department, Universiti Teknologi PETRONAS, 32610 Seri
Iskandar, Perak, Malaysia
e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 539
R. Ibrahim et al. (eds.), International Conference on Artificial Intelligence
for Smart Community, Lecture Notes in Electrical Engineering 758,
https://doi.org/10.1007/978-981-16-2183-3_50
540 N. A. Supardi et al.

continuous and infinite data [1]. However, machine learning [2–4] address an issue
of interpretation data due to ambiguity from data streams. Data streams evolve with
time and complicate the stream processing. Data streams size increasing dispro-
portionately and leads to infinite data size [5]. Consequently, the transmission data
streams have influence memory space [6, 7]. Generally, the process to capture data
streams are extremely difficult to determine the number of the cluster and the shape
of the clusters where it considered unknown [8]. The outcome of clustering results
is changing consistently as the data arrives in streams [9–11]. The clustering process
and the shape of the clusters are mostly affected by outliers [11]. An outlier is defined
generally by [12, 13] which appears inaccurate information from the data.
This paperwork highlights the importance for clustering technique in outlier detec-
tion due to following factors [7, 14, 15]; (i) an outlier may distort data structures and
cluster result, (ii) the clustering methods can identify outlier detection as a small
cluster, (iii) the hybrid clustering algorithms are exemplifying the result for outlier
detection performs better than the existing algorithms.
In research studies, evoStream refers to data sets to conduct the clustering process
in the online step during idle periods and to optimize the outcome of the cluster
[16]. Thus, becomes a motivation to research problem. This research, however, was
primarily aimed at evaluating the outlier identification of evoStream implementation
with real datasets called clinical records as preliminary results.
The rest of this paper is arranged as follows: The related works is summarized in
Sect. 2. Methodology described in Sect. 3. The assessment is seen in Sect. 4, and the
result is outlined in Sect. 5.

2 Related Works

This section provides supplemental information literature review on density-based

clustering techniques.

2.1 Density-Based Clustering

Density-based clustering technique is one of the prominent applied by the researcher.

To identify density-based characteristics are between regions and density functions
[13]. Comprehensively, this technique takes place when data point calculates refer-
ring to density area. Otherwise stated further refines where data points are in a low-
density area perceives as outliers. Algorithms are mostly request parameter in order
to determine the radius from data point at a minimum number in a cluster [9]. In most
cases, the algorithms immediately tackle arbitrary-shaped clusters identification from
the data stream. Nearly the entire algorithms are highly able to scrutinize clusters
and form arbitrary shapes structure and detect outliers. Contrarily, this technique is
limited to a single scan from raw data [9, 17].
An Evolutionary Stream Clustering Technique Outlier Detection … 541

According to the recent study [18] found in the survey has made with distinct
approaches where density-based technique enhanced by practicing an efficient
density-based clustering along with maintaining the evolving medical data clusters.
evoStream algorithm from density-based clustering technique [16, 19] is imple-
mented on the following methodology in Sect. 3 to present the preliminary study
on outlier detection. The term of evoStream comprises from evolutionary algo-
rithms (EA) and DBStream [16]. evoStream algorithm brought together on pros
and cons from [19]. The advantage of evoStream are; (i) the performance of the final
cluster during online phase with high processing speed, (ii) removes computational
overhead. On the contrary, the disadvantage of evoStrem; (i) the application is not
suitable for multi-objective data stream, (ii) high dimensional datasets are impos-
sibly executed in evoStream. The role of evoStream exclusively focusing on the best
optimization solution for an online phase in the idle time to improve the final clus-
ters. Hence, a new approach to the purpose of this algorithm is mainly focusing on
investigating outlier detection from an inconsistency data set.

3 Methodology

An initial method to discover outlier detection in real data sets are implementing by
an algorithm named evoStream [16]. The purpose of this algorithm is mainly focusing
on investigating outlier detection from an inconsistency data object to differentiate
from a categorical problem.

3.1 Proposed Methodology

Benchmarking data sets are retrieving collected from an open source UCI Machine
Learning Repository. The results yielded from evoStream are the preliminary result
for outlier detection. The process of evoStream is comprehensively scrutinizing
through preprocessing data, calculating centers and weight of clusters, profiling clus-
ters results, optimizing weight vector, develop as well as finalizing the validity of
optimizing evoStream model.
The beginning process is data collection and preprocessing. The data sets imported
into the R language and the data sets implemented in evoStream algorithm. The
radius, r has been set to r = 0.05 and k which represents as macro-cluster has been
fixed to k = 3 with incremental generations = 1 as well as re-cluster generations =
1000. Then, initialize evoStream package to insert observations, n = 1000 to process
with calculating centers and weight for micro-clusters and macro-clusters. The plot-
ting result is described according to the centers and weight of micro-clusters and
macro-clusters. Additional generations, n = 2000 examine to observe the different
to improve from the current plot result. The numbers assigned from micro clusters
to macro-clusters according to the nearest neighbor of data points to cluster.
542 N. A. Supardi et al.

Weight vector optimization utilized to enhance the product of evoStream model by

integrating the Grey Wolf Optimization (GWO) algorithm for an outcome which later
implemented in due course. Measures weight vector will produce the optimization
result. As a result, the data objects will be clusters accordingly in the respective
groups.

3.2 Evaluation Metrics

Evaluation metrics are applied to evaluate the quality of statistical machine learning
model, well-known as cluster validation with known as internal measures and external
measures [8, 20]. The process cluster validation investigated under the measurement
of an algorithm to generate excellent results in data mining, particularly in clustering
techniques [20].
Seven (7) multitudes metrics SSQ and silhouette width are internal evaluation
metrics while purity, precision, recall, F-measure, NMI are external evaluation
metrics are selected to evaluate comparative analysis from [8, 10, 16]. Seven metrics
referring to Mathematically, the metrics are indicating with the following equations
except SSQ:
  
o Ci , o = o dist o, o
Silhouette = a(o) = (1)
|Ci | − 1
  
||.d
dh,l log dh clh,l
N M I =      (2)
hdh log ddh lCl log cdl

1
k
F − Measur e = Scor eCi (3)
k i=1

1
k
Purity = Pr ecisionCi (4)
k i=1

Visum
PrecisionCi = (5)
nci
Visum
RecallCi = (6)
Vitotal
An Evolutionary Stream Clustering Technique Outlier Detection … 543

4 Assessment

4.1 Experimental Setup

Programming language R used to evaluate the algorithm. All experiments are

performed on Windows Machine with intel i5-4200U CPU @ 1.60 GHz 2.30 GHz.
There are two real datasets utilize for this investigation.
The gene expression cancer RNA-Seq data set is a random extraction of gene
expressions of patients having different types of tumor: BRCA, KIRC, COAD, LUAD
and PRAD. This data set gathered from the University of Genoa. In addition, the heart
failure clinical records data set consist of the medical records of 299 patients who had
heart failure, collected during their follow-up period, where each patient profile has
13 clinical features and originally collected from Government College University,
Faisalabad, Pakistan.

4.2 Result

In this section, Figs. 1 and 2 show the final improved result from the preliminary
studies of two different data sets. Each of these benchmarking data sets employed
evoStream algorithm to report the studies outlier detection. The initial observations
of the online component n = 1000 required to complete at an early stage of the
investigation. During the idle time, the algorithm generates a macro-cluster when
there is no observation arrives, and the data point will move into their pertinent
clusters. Note that the performance of evolutionary step will not slow down the
processing time.
In general, evoStream did not employ re-clustering approach but applying re-
clustering steps when macro-clusters are requested and not incrementally Evalu-
ating generations does not affect underlying micro-clusters but does affect minor
improvement when implementing with 2000 generations into re-clustering. Hence,

Fig. 1 Gene expression

cancer RNA-Seq data set
544 N. A. Supardi et al.

Fig. 2 Heart failure clinical

records data set

Table 1 Internal and external

Data set/Quality Gene expression Heart failure
evaluation metrics:
metrics cancer RNA-Seq clinical records
micro-cluster
data set data set
SSQ 0.11744971 0.1075193
Silhouette 0.20763806 0.2518121
Purity 0.97747748 0.9761905
Precision 0.98170732 0.9821429
Recall 0.09895513 0.1012891
F1-measure 0.17978783 0.1836394
NMI 0.55517022 0.5560689

re-clustering testified the slightest improvement because additional generations

evaluated.
Moving on to the Tables 1 and 2 shown results from quality metrics of internal and
external cluster validation with types micro-clusters and macro-clusters. The target
minimum of median SSQ of micro-cluster achieved better results in heart failure
clinical. Contradictory to the macro-cluster from internal metrics where the median
SSQ achieve better result in gene expression cancer RNA-Seq data set. On the other

Table 2 Internal and external

Data set/Quality Gene expression Heart failure
evaluation metrics:
metrics cancer RNA-Seq clinical records
macro-cluster
data set data set
SSQ 1.2751618 1.7381368
Silhouette 0.5845640 0.5557643
Purity 0.9809140 0.9814815
Precision 0.9582071 0.9580764
Recall 0.9640463 0.9604450
F1-measure 0.9611179 0.9592593
NMI 0.9179111 0.9301648
An Evolutionary Stream Clustering Technique Outlier Detection … 545

hand, external metrics yielded different numbers according to their performance of

data set.
The values reveal the measurement quality metrics from evoStream algorithm
into two sample data sets. The finish result will be contributed for later development
and validation of an optimize evoStream model.

5 Conclusion

The incoming data made stream processing is difficult to track down when data are
processed in real-time and brings the challenge of understanding the hidden patterns
from data. For that reason, the clustering technique introduced to unveil the nature
of data streams, especially outliers. Thereby, density-based clustering technique is
selected and has proven the most decent model in handling this investigation studies.
The proposed methodology will be introduced to optimize an evoStream model
to encounter the current problem. The purpose of initial investigation to analyze an
optimum weight vector by clustering the data according to their characteristics. Clus-
tering analysis outstandingly comes up with an accurate understanding of relevant
variables in the clusters. All in all, determine a clustering model is validating the
significance of the research question to achieve a valid result.

Acknowledgements The authors acknowledge the support of this research by Yayasan Universiti
Teknologi PETRONAS Fundamental Research Grant (YUTP-FRG) under Grant 015LC0-119.

References

1. Twinkle JSD, Ankleshwaria B (2014) Mining data streams : a survey. 2(2):379–386

2. Abdulkadir SJ, Yong S-P (2013) Unscented kalman filter for noisy multivariate financial time-
series data. In: International workshop on multi-disciplinary trends in artificial intelligence.
Springer, pp 87–89
3. Abdulkadir SJ, Yong S-P, Marimuthu M, Lai F-W (2014) Hybridization of ensemble kalman
filter and non-linear autoregressive neural network for financial forecasting. In: Mining
intelligence and knowledge exploration. Springer, pp 72–81
4. Abdulkadir SJ, Yong S-P (2015) Scaled ukf–narx hybrid model for multi-step-ahead forecasting
of chaotic time series data. Soft Comput 19(12):3479–3496
5. Ahmed KN, Razak TA (2014) A comparative study of different density based spatial clustering
algorithms. Int J Comput Appl 99(8):18–25
6. Kumar A, Singh A (2017) Stream mining a review: tool and techniques. In: Proceedings of
the international conference on electronics, communication and aerospace technology ICECA
2017, pp 27–32
7. Vijayarani S, Jothi P (2016) A hybrid clustering algorithm for outlier detection in data streams.
Int J Grid Distrib Comput 9(11):285–396
8. Kokate U, Deshpande A, Mahalle P, Patil P (2018) Data stream clustering techniques,
applications, and models: comparative analysis and discussion. Big Data Cogn Comput 2(4):32
546 N. A. Supardi et al.

9. Ghesmoune M, Lebbah M, Azzag H (2016) State-of-the-art on clustering data streams. Big

Data Anal 1(1):13
10. Yeoh JM, Caraffini F, Homapour E, Santucci V, Milani A (2019) A clustering system for
dynamic data streams based on metaheuristic optimisation. Mathematics 7(12):1–24
11. Agrawal LS, Prof A, Adane DS (2016) Models and issues in data stream mining. Int J Comput
Sci Appl 9(1):6–10
12. Patel SP, Vala JAY (2015) A survey of outlier detection in data mining. Int J Adv Eng Res Dev
3(01):1–4
13. Bansal R, Gaur N, Singh SN (2016) Outlier detection: applications and techniques in data
mining. In: Proceedings of the 6th international conference—cloud system big data engineering
confluence, pp 373–377
14. Hassani M, Seidl T (2016) Clustering big data streams: recent challenges and contributions.
IT—Inf Technol 58(4)
15. Yogita, Toshniwal D (2013) Clustering techniques for streaming data-a survey. In: Proceedings
of the 3rd IEEE international advance computing conference IACC, pp 951–956
16. Carnein M, Trautmann H (2018) evoStream—evolutionary stream clustering utilizing idle
times. Big Data Res 14:101–111
17. Mousavi M, Bakar AA, Vakilian M (2015) Data stream clustering algorithms: a review. Int J
Adv Soft Comput Appl 7(3):1–15
18. Al-Shammari A, Zhou R, Naseriparsaa M, Liu C (2019) An effective density-based clustering
and dynamic maintenance framework for evolving medical data streams. Int J Med Inform
126:176–186
19. Aljibawi M, Zakree M, Nazri A, Othman Z (2018) A survey on clustering density based data
stream algorithms. 7:147–153
20. Palacio-Niño J-O, Berzal F (2019) Evaluation metrics for unsupervised learning algorithms
Analysis of Intelligent Agent and Its
Components for the SLA Negotiation
Process in Cloud Computing

Rishi Kumar, Mohd Fadzil Hassan, and Muhamad Hariz M. Adnan

Abstract Cloud service providers (CSP) are facing a competitive environment in

the cloud-computing world and CSP is huge in demand with the growing field of IoT,
fog, and cloud services. Consumers are always looking forward to acquiring services
as per requirement. Due to the competitive field in the market of cloud computing,
consumers face issues during the delivery of services. Consumer and CSP need an
intelligent agent to mediate which performs the negotiations process on behalf of both
parties and maintaining other responsibilities. An intelligent agent-based negotiation
system (IANS) is required to perform the negotiation process and measure the KPIs
during the process that helps to make quick decisions in minimum time. Researchers
suggested many learning algorithm-based agents that maximize the success rate and
speed of negotiation. IANS with behavior analysis of consumers and CSP helps to opt
for more reliable and secure services. IANS uses negotiation metrics, performance
factors, and evaluation measures for decision. Due to the dynamic environment of
cloud services, IANS is recommended for the negotiation process.

Keywords Agent · SLA · Cloud services · Negotiation · Intelligent system ·

Cloud computing

R. Kumar (B) · M. F. Hassan

Department of Computer and Information Sciences, Universiti Teknologi Petronas, Seri Iskandar,
Malaysia
e-mail: [email protected]
M. F. Hassan
e-mail: [email protected]
M. H. M. Adnan
Computing Department, Faculty of Arts, Computing, and Creative Industry, Universiti Pendidikan
Sultan Idris, Tanjung Malim, Malaysia
e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 547
R. Ibrahim et al. (eds.), International Conference on Artificial Intelligence
for Smart Community, Lecture Notes in Electrical Engineering 758,
https://doi.org/10.1007/978-981-16-2183-3_51
548 R. Kumar et al.

1 Introduction

Cloud computing services provide solutions to many ventures in form of infrastruc-

ture, software, and platform, and business process [1]. A cloud is a large group of
interconnected computers that provide the required services to any scale of the enter-
prise [2]. To obtain required services both consumers and providers must settle on
pre-established contracts called Service Level Agreements (SLA) [3]. The resources
assigned are based on SLA that contain Service Level Objectives (SLOs) for each
SLA parameter, like price, availability, response time, reliability, CPU utilization, etc.
[4]. Consumers and CSP have struggled over requirements; negotiation can provide
a solution to established deliverable SLA between them [5].
Negotiation is a process of settlement of SLA requirements between consumers
and CSP via the agent [6]. Consumer’s requirements are dynamic and change
according to time and demand, in this manner consumers cannot stick to one CSP [7].
To sustain those dynamic requirements with compound services, there is a necessity
of designing an intelligent agent-based negotiation system (IANS) [7, 8].
IANS can provide negotiation, decision-making, resource management, inter-
acting with other interconnected CSP also called inter-cloud, knowledge acquisition;
analyze behavior, monitoring, maximizing success rate in minimum time [9–11].
IANS maintains the requirements for cloud services within minimum response time,
maximum availability, reliable server, and services at the minimum price [3, 4].

2 Background Study

SLA is legal documents that are defined and modeled on which CSP and consumers
must agree, approved, and signed after SLA negotiation [12]. SLA violations may
lead to monetary penalties, credibility, and reduced reliability for services. SLA
management describes complete assistance to configure SLA negotiation metrics
information, automatic negotiation, and monitoring, speedy recovery over issues,
and quick response over violation [2, 13].
Negotiation between CSP and a consumer based on metrics mainly price (P),
availability (Av), the response time (RT), reliability (R) [3]. Performance of cloud
services, negotiation strategy, and the process can be measure by a utility function
(UF), negotiation speed (NS), success rate (S) [14]. Factors to evaluate the strategy
of the negotiation process are the number of resources (NoR), Deadline (DL), degree
of satisfaction (DoS) [3, 6, 15, 16].
In the Negotiation process, consumers want to fulfill maximum requirements and
on the other hand, providers want to maximize sales of services. IANS focus on the
cycle of negotiation, negotiation matrices, and performance metrics [17].
Negotiation metrics: Cloud consumers must understand the key performance
indicators (KPI) for better negotiation. In Table 1, there are eight major high-level
cloud service attributes and under each level, some major KPIs to evaluate and
Analysis of Intelligent Agent and Its Components … 549

Table 1 High-level cloud services attributes and KPIs [4]

S. No. Cloud service attributes Key performance indicators
1 Financial Price of service, free features, cost of security
2 Security and privacy Data center and location, encryption and decryption, privacy,
trust
3 Performance Response time, efficiency, latency, load balancing, network
quality, robustness, throughput, scaling latency
4 Agility Adaptability, elasticity, flexibility, portability
5 Usability Accessibility, compatibility, controllability, functionality,
maintainability, suitability, transparency, pooling
6 Accountability Auditability, integration support, sustainability
7 Assurance Reliability, availability, credibility, fault tolerance, service
stability
8 Management Data management, monitoring, value-added services

negotiate [4]. According to [3, 6, 11, 16], major negotiation metrics are price (P),
availability (Av), the response time (RT), reliability (R).

3 Framework and Workflow of Ians

The proposed IANS design in Fig. 1. Follow the life cycle of SLA. IANS helps to finds
out the suitable CSP concerning the consumer requirements [7]. Agent components
allow us to understand the behavior of CSP by learning algorithms like reinforcement
learning [15], which help to speed up the process of negotiation.
There are five steps for SLA satisfaction between CSP and consumer [2, 5].
1. SLA Definition: CSP provides SLA with available offers and resources. A
CONSUMER SLA covers the requirements of the resource. Both types of SLA
can customize later [2].
2. Published and Discovery: Consumer published requirement to IANS, which
translates and analyzes the SLA and discover suitable CSP from its database.
Discovery of CSP based on rank-based that depends upon security, trust, perfor-
mance, financial, etc. IANS analyzes behavior property based on the functioning
of KPIs [2, 4].
3. Negotiation: Negotiation process has five major steps, (i) assessment of SLA,
(ii) allowance computation, (iii) achievable solution development, (iv) produce
offer, and (v) negotiation termination [16].
(a) Assessment of SLA: IANS assess the SLA and analyze the requirements
of resources with price, response time, and availability. IANS computes
UF, if the value of UF maximum, the agent gives positive feedback
[16, 19].
550 R. Kumar et al.

Fig. 1 The framework of intelligent agent-based negotiation system [18]

(b) Allowance Computation: IANS computes discounts over the prices

of different resources. IANS computes the difference between previous
offers and current offers and produces satisfaction value [16].
(c) Achievable Solution Development: IANS keep threshold value (TV) for
KPI. If there is a vast difference concerning TV, then IANS produce nega-
tive feedback, in case of negligible difference, gives positive feedback,
which reduces ‘Ti’ and ‘Nr’ and increases NS [11, 16].
(d) Produce Offer: While comparing and analyzing the latest offer by CSP,
IANS makes the final call to produce an SLA offer. Results will mainly
depend upon the DoS of consumers and CSP for SLA and UF of consumers
[11, 16].
(d) Negotiation Termination: The termination of the negotiation process
has two values success and failure. Successes, if the process meets with
requirements before the deadline (DL) with a minimum number of nego-
tiation rounds. Failure, when define deadline is over before settlement.
In case of failure, the consumer offered a new SLA or new CSP for the
further process [11, 16].
4. Operationalization: The decision support system (DSS) of IANS will finalize
the agreement and start operation. Operationalization of SLA means negotia-
tion completed successfully. Operationalization includes monitoring of secure
service delivery, accounting of SLA, execution of SLA [2].
Analysis of Intelligent Agent and Its Components … 551

Fig. 2 Workflow diagram of negotiation process in IANS [16]

5. Termination: After the completion of the process, sharing of resources and all
activities ended [2].
In Fig. 2 [16], Workflow Diagram of Negotiation process showing communication
between consumer and CSP via IANS. IANS helps to make the process fast, secure,
and provide the best options for the consumer from a massive number of CSP [12].

4 Conclusion and Future Work

In this research study, IANS functioning explored dependency factors of SLA nego-
tiation process for cloud services. It is defined that cloud SLA negotiation process
with IANS for negotiation will understand the requirement of the client and produces
perfect match from the massive number of CSPs with maintaining KPIs price, avail-
ability, response time and reliable. The descriptive analysis of the agent, its key factor,
and functioning is to develop an intelligent intermediary for the negotiation process.
IANS can focus on the dynamic environment of the cloud market and consumers’
requirements. Management of SLA described understanding the roles and respon-
sibilities of SLA. The framework of IANS discussed the workflow diagram of the
negotiation process. In future aspects, IANS may explore various learning algo-
rithms to design agents, decision-making algorithms for optimizing solution, security,
monitoring solution, and add more attributes for analysis.
552 R. Kumar et al.

References

1. Bahsoon R et al (2018) A manifesto for future generation cloud computing. ACM Comput
Surv 51(5)
2. Voorsluys RBW, Broberg J (2011) Cloud computing : principles and paradigms table of contents
3. Shojaiemehr B, Rahmani AM, Qader NN (2018) Cloud computing service negotiation: a
systematic review. Comput Stand Interfaces 55:196–206
4. Nadeem F (2020) A unified framework for user-preferred multi-level ranking of cloud
computing services based on usability and quality of service evaluation. IEEE Access 1–1
5. Iyer GN (2016) Cloud testing: an overview
6. Shojaiemehr B, Rahmani AM, Qader NN (2019) A three-phase process for SLA negotiation
of composite cloud services. Comput Stand Interfaces 64:85–95
7. Sim KM (2018) Agent-based approaches for intelligent interCloud resource allocation. IEEE
Trans Cloud Comput 1
8. Elhabbash A, Samreen F, Hadley J, Elkhatib Y (2019) Cloud brokerage: a systematic survey.
ACM Comput Surv 51(6):1–28
9. Rajavel R, Thangarathanam M (2016) Adaptive probabilistic behavioural learning system for
the effective behavioural decision in cloud trading negotiation market. Futur Gener Comput
Syst 58:29–41
10. Rajavel R, Iyer K, Maheswar R, Jayarajan P, Udaiyakumar R (2019) Adaptive neuro-fuzzy
behavioral learning strategy for effective decision making in the fuzzy-based cloud service
negotiation framework. J Intell Fuzzy Syst 36(3):2311–2322
11. Hsu CY, Kao BR, Ho VL, Li L, Lai KR (2016) An agent-based fuzzy constraint-directed
negotiation model for solving supply chain planning and scheduling problems. Appl Soft
Comput J 48:703–715
12. De la Prieta F, Rodríguez-González S, Chamoso P, Corchado JM, Bajo J (2019) Survey of
agent-based cloud computing applications. Futur Gener Comput Syst 100:223–236
13. Paputungan IV, Hani AFM, Hassan MF, Asirvadam VS (2019) Real-time and proactive SLA
renegotiation for a cloud-based system. IEEE Syst J 13(1):400–411
14. Mubeen S, Asadollah SA, Papadopoulos AV, Ashjaei M, Pei-Breivold H, Behnam M (2018)
Management of service level agreements for cloud services in IoT: a systematic mapping study.
IEEE Access 6:30184–30207
15. Rajavel R, Thangarathinam M (2017) ADSLANF: a negotiation framework for cloud manage-
ment systems using a bulk negotiation behavioral learning approach. Turkish J Electr Eng
Comput Sci 25(1):563–590
16. Li L, Yeo CS, Hsu CY, Yu LC, Lai KR (2017) Agent-based fuzzy constraint-directed negotiation
for service level agreements in cloud computing. Cluster Comput 1–15
17. El-Matary DM, El-Attar NE, Awad WA, Hanafy IM (2019) Automated negotiation framework
based on intelligent agents for cloud computing. In: Proceedings of the 2019 international
conference on innovative trends in computer engineering ITCE. pp 156–161
18. Vallejo D, Castro-Schez JJ, Glez-Morcillo C, Albusac J (2020) Multi-agentarchitecture for
information retrieval and intelligent monitoring by UAVs in known environments affected by
catastrophes. Eng Appl Artif Intell 87:103243
19. Pouyllau H, Carofiglio G (2013) Inter-carrier SLA negotiation using Q-learning. Telecommun
Syst 52(2):611–622
Early Detection of Myocardial Infarction
Using Machine Learning with Maximum
Accuracy

S. Abirami Manisa, B. J. Abarna, V. Geethanjali, G. V. Hari Venkat,

and R. Karthikeyan

Abstract This paper presents a technique for detecting Myocardial Infarction (MI)
using Machine Learning through the ECG data. In today’s scenario, MI is one of the
major causes of demise worldwide. MI occurs due to coronary heart disease and if
detection or treatment is not done at appropriate time, the untreated MI may present
with serious late complications. For treating MI in earlier stage several methods have
been employed, but the parameters employed for those does not provide as much
accuracy. So, a system which detects the abnormality with maximum accuracy is
proposed. For that, Pan Tompkins algorithm is employed to filter and remove the
noise in the acquired ECG signal and S-T segment is extracted from it. The features
that are obtained from the extraction of S-T segment when given to the classifier
provided better results. Based on the application of the Machine Learning classifiers,
Naïve Bayes and Decision tree, the Decision tree classifier gave a higher accuracy
of 98.5% compared to Naïve Bayes of 93.9%.

Keywords Myocardial infarction · Pan Tompkins algorithm · Machine learning ·

Decision tree classifier · Naïve Bayes classifier · S-T Segment · ECG

1 Introduction

An Electrocardiogram (ECG) records the heart’s electrical activity, over a period of

time, by placing the electrodes on the patient’s body. From the captured ECG signal,
cardiologists observe the rhythm and functioning of heart. They can also identify the
cardiac abnormalities from ECG recordings [1]. Manual classification and detection
have its own drawbacks, which paved a way for an automated ECG classification
and earlier detection techniques. This paper deals with the detection of most pivotal
cardiac abnormality known as myocardial infarction, or heart attack.
Today, major cause of increasing cardiac death is myocardial infarction. MI is the
first symptom of coronary artery disease. Nearly 17.9 million people died in 2016

S. A. Manisa (B) · B. J. Abarna · V. Geethanjali · G. V. H. Venkat · R. Karthikeyan

Department of ECE, Dr. N.G.P Institute of Technology, Coimbatore, India

© The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 553
R. Ibrahim et al. (eds.), International Conference on Artificial Intelligence
for Smart Community, Lecture Notes in Electrical Engineering 758,
https://doi.org/10.1007/978-981-16-2183-3_52
554 S. A. Manisa et al.

Fig. 1 ECG waveform

due to CVDs, which accounts for 31% of all global deaths. Out of these deaths, 85%
are due to heart attack and stroke [2].
It is observed that ECG has three main components: The atrial depolarization is
represented by P wave; the depolarization of the ventricles is represented by the QRS
complex; and the repolarization of the ventricles is represented by the T-wave. It is
shown in Fig. 1. In between the period of ventricular ion change, blood is supplied to
the heart muscles, via coronary arteries. When the blood flow decreases or stops to a
part of the heart, causes myocardial infarction. So, the salient point in early detection
of MI is the ST segment. The ST segment duration indicates the blood supply to the
heart. MI has three stages namely injury, ischemia and infarction [3]. An elevated
ST segment occurs at the stage of ischemia. For detecting more precisely, Machine
Learning is opted due to its less complexity and training time [4].
Previous methods [5–8], of MI detection are supervised learning approaches,
in which they took only the parameters like Age, BP etc., They did not provide
that much accuracy. Some methods [9–12] process the ECG recordings, but good
beneficial results are not yielded because only the difference between labels of the
entire ECG is taken. They did not take the difference between labels of heart beats
within the ECG. By using these methods, a large data cannot be classified and it is to
be overtrained. Even though these methodologies [13–15] gave good classification
results, it had less accuracy.
As this paper aims at detecting MI by extracting ECG from a noisy signal, it
requires accurate detection. For that, the filtering is done by Pan Tompkins algorithm
and the proper extraction of the S-T segment is done and, that is given to classifier
to detect the abnormality.
Early Detection of Myocardial Infarction … 555

2 Proposed Methodology

The structure of MI detection in our work consists of three main phases, namely:
Signal acquisition, ECG feature extraction and classification. The proposed method
is as shown in Fig. 2.
The noisy raw input is denoised using Pan Tompkins QRS Detection algorithms.
The extracted noiseless ECG signal is given to S-T segment detection and then the
features are extracted. The classification is done using naive bayes and decision tree
classifier and MI is detected.

2.1 ECG Signal Acquisition

The ECG signal is obtained using the ECG Sensor AD8232 and the Arduino Uno
Controller. The ECG sensor which senses the electrical activity of the human heart.
The three electrodes are attached to the left arm and right arm and right abdomen
to detect the heartbeat. The sensor is very small in size, portable and accurate to
measure the continuous heartbeat.
The Arduino controller gets the data from the ECG sensor and transmits the signals
to the signal processing through the COM port.

ECG Signal Pan-Tompkins

Acquision QRS Detecon

S-T
Segment
Detecon

Classifier
Feature
Model
Extracon

Fig. 2 Block diagram of proposed system

556 S. A. Manisa et al.

Raw ECG
BPF Differenator Squaring
Signal

Adapve Moving
Decision
window
Thresholding
Integrator

Fig. 3 Block diagram of Pan Tompkins algorithm

2.2 Pan Tompkins Algorithm

For a remote ECG monitoring, QRS detection is the preliminary step for detecting the
heartbeat for the subsequent rhythm classification [16]. It is the most significant part
of the ECG analysis algorithm. The Pan Tompkins algorithm uses the bandpass filter
having low pass filter and high pass filter which reduces the noises. It has derivative
function that highlights the R wave and then a quadratic function is applied to enhance
the high frequency features of the complexes (Fig. 3).

2.3 S-T Segment Detection

In ECG, initially the PQRST points are obtained by setting the threshold values and
from which the S-T segments are detected. This is the segment where the depression
or elevation of it distinguishes the normal and abnormal ECG. The isoelectric period
during which the ventricles are in between depolarization and the repolarization is
represented by the S-T segment.

2.4 Feature Extraction

The statistical features such as mean, standard deviation, skewness and kurtosis.,
are extracted from the MATLAB Tool Diagnostic Feature Designer [17] through the
data sets obtained from PTB Diagnostic Database, MIT-BH Arrythmia Database and
over 330 data were collected and features were extracted and tabulated as shown in
Fig. 4.
Early Detection of Myocardial Infarction … 557

Fig. 4 Feature table

2.5 Classifier Model

The classifier model is built by using MATLAB classification learner app. The
extracted features are exported to the classifier model. Here, the classification is
done by using naive bayes and decision tree classification. The classifier is tested
and validated using the dataset obtained from the Physionet ATM database. Here the
hold out validation is given as 20% and then the dataset is trained.

3 Result and Discussion

3.1 ECG Signal Acquisition

The ECG Signal acquired through the sensor is plotted as shown in Fig. 5 using the
serial port, the system gets the analog value and plots that value using MATLAB.

3.2 Pan Tompkins Algorithm

The raw ECG signal after processed through the Pan Tompkins algorithm is found
as shown in Fig. 6 and the QRS peaks are detected as shown in Fig. 7.
558 S. A. Manisa et al.

Fig. 5 Raw ECG signal

Fig. 6 Pan Tompkins algorithm implementation

3.3 S-T Segment Detection

The PQRST points detected are in the ECG Signal are shown in Fig. 8 and S-T
segment plotting from the detected PQRST points are shown in Fig. 9.
The deviation in the normal and abnormal ECG are shown in Fig.10 which serves
as deciding factor for the Myocardial Infarction. The abnormal ECG may have either
elevated or depressed S-T segment.
Early Detection of Myocardial Infarction … 559

Fig. 7 QRS detection

Fig. 8 PQRST signal

560 S. A. Manisa et al.

Fig. 9 ST Segment detection from filtered ECG signal

Fig. 10 Normal and abnormal ECG signal ST segment

3.4 Classifier Model

The features extracted from the data sets are given to the Classification Learner App
and the scatter plot obtained for the Naïve Bayes and Decision Tree are shown in
Figs. 11 and 12. From which the accuracy for Naïve Bayes obtained as 93.9% and
Decision Tree as 98.5%
The confusion matrix for the both classifiers are as shown in Fig. 13.
Early Detection of Myocardial Infarction … 561

Fig. 11 Scatter plot for Naïve Bayes

Fig. 12 Scatter plot for decision tree

4 Conclusion and Future Work

Myocardial Infarction needs an accurate and timely treatment for patients. So, in
machine learning, Naïve bayes and Decision Tree were used and compared, in which
562 S. A. Manisa et al.

a) Naive Bayes b) Decision Tree

Fig. 13 Confusion matrix

the Decision Tree gave higher accuracy of 98.5%. In near future, we are trying
to improve the accuracy level and planning to give an additional support through
emergency assistance and transmit real-time data using IoT.

References

1. Salehi F et al (2017) The importance of electrocardiography parameters in healthy Iranian

children. ARYA Atherosclerosis 13(3):159–160
2. World Health Organization (2017, May 17) Fact sheets of ‘Cardiovascular diseases (CVDs).’
Retrieved from https://www.who.int/en/news-room/fact-sheets/detail/cardiovascular-diseases-
(cvds)
3. Thygesen K, Alpert JS, Jaffe AS, Chaitman BR, Bax JJ, Morrow DA et al (2019) Fourth
universal definition of myocardial infarction. Eur Heart J 40(3):237–269
4. Yan Y et al (2019) The primary use of artificial intelligence in cardiovascular diseases: what
kind of potential role does artificial intelligence play in future medicine?’ J Geriatric Cardiol
JGC 16(8):585–591
5. Celin S, Vasanth K (2018) ECG signal classification using various machine learning techniques.
J Med Syst 42:241
6. Pandey SK, Sodum VR, Janghel RR, Raj A (2020) ECG arrhythmia detection with machine
learning algorithms. In: Raju K, Senkerik R, Lanka S, Rajagopal V (eds) Data engineering
and communication technology. Advances in intelligent systems and computing. vol 1079.
Springer, Singapore
7. Sekhar BB, Likhitha V, Narendra I, Harika G (2019) Prediction and detection of heart attack
using machine learning and internet of things. Int J Eng Adv Technol (IJEAT) 8(4). ISSN:
2249–8958
8. Uddin S, Khan A, Hossain M et al (2019) Comparing different supervised machine learning
algorithms for disease prediction. BMC Med Inform Decis Mak 19:281
9. Purnendu SP (2017) Machine learning and IoT for prediction and detection of stress. In: 17th
international conference on computational science and its applications (ICCSA) 2017 IEEE
Early Detection of Myocardial Infarction … 563

10. Subashini A, Sai Ramesh L, Raghuraman G (2019) Identification and classification of heart
beat by analyzing ECG signal using Naive Bayes. In: 2019 Third international conference on
inventive systems and control (ICISC). pp 691–694
11. Verma AR, Gupta B, Bhandari C (2020) A comparative study of ECG beats variability
classification based on different machine learning algorithms. Augment Hum Res 5:16
12. Mudasir MK, Syed IA (2016) Prediction of heart disease using decision tree a data mining
technique. IJCSN 5(6)
13. Chamuleau SA, van den Brink RB, Kloek JJ, Broekhuis E, de Beaumont EM, Koster RW (2005)
Complicaties van een niet herkend hartinfarct; ‘Complications of an unrecognized myocardial
infarction’. Ned Tijdschr Geneeskd. 2005 Nov 19; 149(47):2593–9. Dutch. Erratum In: Ned
Tijdschr Geneeskd. 2006 Dec 30; 150(52):2898. PMID: 16355569
14. Kshirsagar P (2020) ECG signal analysis and prediction of heart attack with the help of
optimized neural network. ACJ 9(4):497–506
15. Priyan MK, Usha DG (2018) A novel three-tier internet of things architecture with machine
learning algorithm for early detection of heart diseases. Comput Electri Eng 65:222–235
16. Pan J, Tompkins WJ (1985) A real-time QRS detection algorithm. IEEE Trans Biomed Eng
BME-32 (3):230–223
17. Mali B, Zulj S, Magjarevic R, Miklavcic D, Jarm T (2014) Matlab-based tool for ECG and
HRV analysis. Biomed. Signal Process Control 10:108–116
Design and Development of a Real Time
Mouse for Human Computer Interaction

U. Vinothkumar , D. Kavya, M. Abinesh, M. Kavin, and P. Aravind

Abstract In this paper there are two parts, Transmitter and Receiver. The transmitter
side is designed with Peripheral interface controller (PIC), Bluetooth, limit switches
and Micro electro mechanical systems (MEMS). In the receiver side there are Blue-
tooth and Personal Computer with Visual Basic 6.0. In the transmitter, the signals
sensed by the accelerometer and limit switches, then it is given input to the micro-
controller. The Analog to Digital Converter in the microcontroller converts these
sensor input to digital code and fed to Bluetooth module via Universal asynchronous
receiver/transmitter (UART) which triggers the module to transmit the signals to the
computer. At the computer, Bluetooth captures these signals. Thus, according to the
movement of hand, cursor is moved and according to limit switch press, functions
analogous to clicking the button is carried out in Personal Computer with the help
of Visual Basic 6.0.

Keywords MEMS · PIC · UART · Visual Basic 6.0 · ADC · Bluetooth

1 Introduction

1.1 Overview

Nowadays the human computer interaction is a need for survival. Most of industries,
military, medical and domestic application require this type of interaction. Mainly
interactions such as speech and gestures have been used for communication. But
speech recognition lacks accuracy and probable to make errors [1] . So we require
a wireless real time user friendly device which can avoid such incidents or a system
that is accessible to gestures. In this paper therefore we go with the hand glove
gesture recognizing system, which can perform all the function that a real time
mouse does like dragging, selecting, scrolling the pointer to 3600 etc. A wearable

U. Vinothkumar (B) · D. Kavya · M. Abinesh · M. Kavin · P. Aravind

Dr.N.G.P. Institute of Technology, Coimbatore, Tamil Nadu, India
e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 565
R. Ibrahim et al. (eds.), International Conference on Artificial Intelligence
for Smart Community, Lecture Notes in Electrical Engineering 758,
https://doi.org/10.1007/978-981-16-2183-3_53
566 U. Vinothkumar et al.

wireless electronic glove that transforms hand and finger movement to real time
data for different applications. This paper replaces the standard computer mouse for
controlling a computer mainly used for power point presentations etc. Project uses
Microchip’s microcontroller IC named Peripheral Interface Controller PIC16F877A
and Microchip’s ‘Integrated Development Environment’, MPLAB, to simulate and
assemble the written code. Why we call such system user friendly? Because these
are some methods that we require to implement in a world where physical inabilities
of humans are a cause of concern the normal condition like backache, half paralyzed,
deaf and dumb etc. So we call them user friendly.

1.2 Objectives

• The objective of this paper is to design the wearable, real time and user friendly
device for human computer interaction.
• To mimic all the gestures performed in the real world wearing the data glove in
real-time with high accuracy.

1.3 The Scope of the Paper

The scope of the paper in future deals with the gaming sector which shall be further
more improvised for virtual environment applications. Another field where this can
be used is medical field, where paralyzed persons can be installed with this equipment
for their communications.
This paper replaces the standard computer mouse for controlling a computer
mainly used for power point presentations through wireless connection.

2 Existing Methods

In 2004 a social anthropologists ’Edward T Hall’ claimed that over 60% of human
interaction that we today is nonverbal in nature i.e. most of the communication can
be possible by gestures [2]. In this paper we use gestures which function indepen-
dently on speech i.e. we use autonomous gestures for communication. For a accurate
interaction the clarity of gestures is also a case of concern so we must select compo-
nents which can be user friendly can be accessible and reduce cost. Considering
the case where magnetic induction coils are used as sensors [3]. Here we require
separate sensors for each of the fingers respectively to improve degree of freedom
but looking for economic consideration we could go for better. Therefore we go for
TRI-axis MEMS sensor which have all the ability that magnetic induction coils and
have reduced cost [4].
Design and Development of a Real Time Mouse for Human Computer Interaction 567

2.1 Gesture Recognition

Here we use autonomous gestures for communication purpose where motion of

hand is selected as co-ordinate axis x, y, z axis respectively. Considering other kind
of gesture recognition system example as for robotic interaction, where gestures are
identified using the rule based approach and combination of three skin like regions
at a particular image frame for satisfy a particular defined condition respectively [5].

2.2 Tri Axis Control

The tri axis control can be obtained by using MEMS sensors. It gives opportunity for
free arm rotation. For each x, y, z axis program is written and it is being executed by
PIC using VB software. The tri axis control overcomes the disadvantages of magnetic
induction coil as it require less capital cost.

3 Methodology

Methodology section will served as a ‘technical blue print’ for this project to who
has an interest to upgrade or modify this project in the future. In this section, it is
important to know that all technical aspect regarding to this project will be included.
For this section, it will divided into two parts which consist of software and hardware.
For software sub-section, need MP-lab programming to program the programmable
integrated circuit and for interfacing the PIC module with pc visual basic 6.0 had
been used. In hardware sub-section, it will more on designing and developing the
circuit using strip board, fit in and solder the components on it and communicate
with microcontroller with the laptop through RS232 interface and Bluetooth. The
functional block diagram of Hand Data Glove: A New Generation Real-Time Mouse
for Human–Computer Interaction is shown in Fig. 1.

3.1 Data Glove

Here we uses a glove-type device which could detect hand position and movement.
In this approach user require to wear a glove like device, which uses sensors that
can sense the movements of hand and pass the information to the computer. These
approaches can easily provide exact coordinates of palm’s location and orientation,
and hand configurations. The main advantage of these approach are high accuracy
and fast reaction speed.
568 U. Vinothkumar et al.

Fig. 1 Top and bottom view of hand data glove

3.2 Bluetooth Module

The Bluetooth module are used for wireless transmission of data from wearable hand
glove device to personal computers. Bluetooth used here have a range of 10 m with
an available frequency range of 2.4 GHz. As per requirement we can set variable
distance by using amplifiers up to 100 m and so on. The Bluetooth is selected so that
it is low cost, low power, wireless mesh network standard (Fig. 2).

Fig. 2 Bluetooth module

Design and Development of a Real Time Mouse for Human Computer Interaction 569

Fig. 3 MEMS
accelerometer

3.3 Limit Switches

The limit switches used here to obtain different function done by mouse. Each
switches act for functions such as left button, right button and center click. In visual
basic 6.0 each of the functions are described as special character respectively. i.e.
for left button- dollar ($),right button-hash tag (#), center click-asterisk (*) char-
acters are used respectively. For each function when we click on limit switch the
following symbols are executed by PIC16F877A and the corresponding movements
are generated by windows application.

3.4 MEMS Accelerometer

The MEMS accelerometer (Fig. 3) is placed above the wearable data glove. The
accelerometer measures acceleration forces. These forces may be static, like the
constant force of gravity, or they could be dynamic—caused by moving or vibrating
the accelerometer. Here we use tri—axis MEMS which recognize the movement of
arm in the form of respective co-ordinates of x, y, z respectively. Here the moment
of palm is recognized.
For each positions the function generated in visual basic 6.0 i.e. for x, y, z different
coding is written through visual basic 6.0 and is been executed in PIC16F877A.

3.5 Proposed Method

See Fig. 4.
570 U. Vinothkumar et al.

Transmitter side:

Receiver Side:

Fig. 4 Functional block diagram

4 Results and Discussion

Thus the Hand data glove is able to perform all functions of optical mouse with more
degree of freedom (Fig. 5).

5 Conclusion

The hand data glove can now mimic all the gestures performed by a real time mouse.
The added advantage given by degree of freedom of hand data glove is more than
that of mouse. The Bluetooth act as an interfacing device between the two hardware
sections. The detection of palm movement is sensed by using tri-axis accelerometer
Design and Development of a Real Time Mouse for Human Computer Interaction 571

Fig. 5 Real time mouse

interaction with computer

MEMS. The results also indicates left, Right and Center click operations effectively
done by using accelerometer MEMS and limit switches.

6 Future Scope

The scope of the project in future deals with the gaming sector which shall be further
more improvised for virtual environment applications. Another field where this can
be used is medical field, where paralyzed persons can be installed with this equipment
for their communications.
This project replaces the standard computer mouse for controlling a computer
mainly used for power point presentations through wireless connection.

References

1. Piyush K, Jyoti V, Shitala P (2012) Hand data glove: a wearable real-time device for human
computer interaction. Int J Adv Sci Technol 43
2. Shiratori T, Hodgins JK (2008) Accelerometer-based user interfaces for the control of a
physically simulated character. ACM Trans Graph 27(5):1–9
3. Chin-Shyurng F, Herman S (2005) Development of a data glove with reducing sensors based on
magnetic induction. IEEE Trans Indus Electron 52(2)
4. PRIME Faraday Partnership” An Introduction to MEMS” January (2002) Wolfson School of
Mechanical and Manufacturing Engineering Loughborough University, Loughborough. ISBN
1–84402–020–7. http://www.primetechnologywatch.org.uk.
5. Pranathi N, Ahmed SM (2013) Tri-Axis motion detection using MEMS for unwired mouse
navigation system in the future generation machines. Int J Adv Res Comput Commun Eng 2(9)
Design and Implementation of Low Cost
Energy Meter Using MIT App Inventor

P. Gajendran, S. N. Deepa, and N. Rajasingam

Abstract Rapid development of low-cost sensors and smart sensors in this twentieth
century makes the electronic based industries to get into the plenty of new solutions in
all the sectors. In this line software tools also support in nice a way to implement the
task with less effort. This proposed work going to enable such kind of an industrial
need application, namely low cost smart energy measurement is the popular area
in many accepts like Data control, management, saving and so on. Producing the
one unit of electricity is equal to saving half unit is the common thumb rule in the
electricity sector. Such a wonderful area utilizes all the advanced technologies to
implement its requirement.

Keywords Low cost energy measurement · Sensors and smart devices · Open
source softwares

1 Introduction

An open source hardware and software platform triggers the smart devices imple-
mentation with less effort. Every designer now-days tasting the open source tools
to implement their needs with short span of time [1]. There are two main reasons
behind that one is cost of overall implementation and second one is very important

P. Gajendran
Karpagam College of Engineering, Coimbatore, India
S. N. Deepa
Anna University Regional Campus, Coimbatore, India
N. Rajasingam (B)
Dr. N. G. P. Institute of Technology, Coimbatore, India

© The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 573
R. Ibrahim et al. (eds.), International Conference on Artificial Intelligence
for Smart Community, Lecture Notes in Electrical Engineering 758,
https://doi.org/10.1007/978-981-16-2183-3_54
574 P. Gajendran et al.

Fig. 1 Block diagram of the proposed work

one in this past decade this method accelerate the development work in the electronic
industries that is online community support. This one accelerate everyone simulation
work into real hardware implementation work [2]. This paper going to focus such
tools like Arduino and MIT APP inventor these two open source tools enable our
ideas to function in the real world environment.
A block diagram shown in Fig. 1 contains only two components in it. That the
Bi-directional communications between the server and client. In conventional energy
measurement topologies are defined as sensing and monitoring unit. Now these two
terms are slightly modified into server and client model [3]. Server going to plays
the vital role in thus proposed work. The functions are listed as follows.
• Sensing unit
• Data Manipulation
• Communication between Server and Client (Fig. 2)
Now the sever portion is clearly illustrated in the above block diagram. The heart of
the server unit is Arduino UNO, powerful hardware implementation. In this decade its
growth is tremendous in the area of embedded systems. Arduino community provides
hardware and software support to implement any low-cost embedded solutions.
Other major components are P.T. (Potential Transformer), C.T. (Current Trans-
former), ZCD (Zero Crossing Detector) and Bluetooth module. Using P.T. and C.T.
the important quantities required for the energy measurement in sensed and fed into

Fig. 2 Block diagram of server unit

Design and Implementation of Low Cost Energy Meter … 575

Fig. 3 Block diagram of

client unit

the microcontroller unit (Arduino UNO). Then the same voltage and current signal
are tapped from the output of the P.T. and C.T. and gives as the input to zero crossing
detectors [4, 5].
By receiving these signal ZCD provides the output as theta, that is phase angle
difference between V and I. To compute P (real power), Q (reactive power), Power
factor and energy. Finally, the computed signals were transmitted via Bluetooth
module (Fig. 3).
Client Unit is simply a smart phone using Bluetooth protocol the message packets
transmitted from the Server unit is received in the smart phone. Client Unit imple-
mentation is going to carried out with the help of MIT App Inventor. Again, this
one is for the open source cloud-based software support for the App Inventor. The
android app development is the one role in the client design. The Implementation
is clearly illustrated in the Fig. 4.This app has access security feature; to access the
particular “App Fingerprint Authentication” is added as the gateway to enter into
their application. Once the authentication is verified, it allows the user to get into the
network. In this way, server and client talks to each other. The calculated parameters
are then transmitted from server unit to receiver. Then the same will be displayed in
the dashboard window in the mobile App.

2 Implementation of the Proposed Work

2.1 Hardware Selection

Hardware units are only present in the server client. They are,
1. P.T.
2. C.T.
3. ZCD
4. Bluetooth Module
576 P. Gajendran et al.

Fig. 4 Flow charts of server unit and client unit function

2.2 App Implementation

MIT App inventor tool is used to develop the android App required for this work. It
has two stages one is screen development and another one is the code blocks. There
are three screens used in this android app. First and second screen are used for user
verification using finger print sensor. Then third screen has main dashboard window
there user can take the readings of the energy meter.
Figure 5 clearly shows the components used in the screen-1 window. Here using
inbuilt fingerprint sensor in the smart phone is used to read the user input. Also the
Fig. 6 shows the code blocks required for finger print authentication.
Design and Implementation of Low Cost Energy Meter … 577

Fig. 5 Finger print gateway

Fig. 6 Code blocks used in the screen-1 and Screen-2, Screen-3 view
578 P. Gajendran et al.

Fig. 7 Outputs from Screen-1 and Screen-2 window

As shown in Fig. 6 Screen-2 will allow entering the next level once finger print
authentication is success. In screen-2 authentication confirmation is displayed then
provides the entry to reach the main dashboard window that is screen-3. Here in
screen-3 Bluetooth connectivity option is given to the user to pair with the server
Bluetooth module. Once it’s done it will start receive the transmitted parameters from
the server to display it.

3 Output from MIT App Windows

At first download the App and install it in android based smart phone. One important
condition the phone must have fingerprint sensor. Then open the App the first screen
will appear as shown in the Fig. 7. Here the user needs to click the Scan Fingerprint
button followed by scan the finger. Once it’s done, it allows the user to enter in to the
second screen. There it will give Authentication Succeeded message. Then user need
to click the Go To Meter Dashboard button, immediately screen three will appear as
in the Fig. 8.
Finally the screen-3 window will shows the actual dashboard window of the low
cost energy meter. Here using Click to connect button user can pair the mobile
Bluetooth with server Bluetooth module that is HC-05. Once its paired successfully,
the dashboard will shows the various the electrical quantities.

4 Conclusion

In this way the proposed work is designed and implemented using Arduino and MIT
App inverter tools. This one serves the requested functions as mentioned in the initial
chapters. At present this work focuses on only single phase AC domestic loads. In
Design and Implementation of Low Cost Energy Meter … 579

Fig. 8 Outputs from Screen-3 window

future work the same need to be extended to three phase and DC measurements. Also
online cloud storage needs to implement to get in to the internet world.

References

1. Morello R, De Capua C, Fulco G, Mukhopadhyay SC (2017) A smart power meter to monitor

energy flow in smart grids: the role of advanced sensing and IoT in the electric grid of the future.
IEEE Sens J 17(23):7828–7837
2. Barkana BZ, McDonough W (2019) AP computer science principles: designing the hourly app
in MIT app inventor. In: 2019 IEEE long island systems, applications and technology conference
(LISAT) 2019 May 3, IEEE, pp 1–6
3. Pokress SC, Veiga JJ (2013) MIT app inventor: enabling personal mobile computing. arXiv
preprint arXiv:1310.2830
4. Albu MM, Sănduleac M, Stănescu C (2016) Syncretic use of smart meters for power quality
monitoring in emerging networks. IEEE Trans Smart Grid 8(1):485–492
5. Burunkaya M, Pars T (2017) A smart meter design and implementation using ZigBee based
wireless sensor network in smart grid. In: 2017 4th international conference on electrical and
electronic engineering (ICEEE) 2017 Apr 8, IEEE, pp 158–162
Detection of Sybil Attack in Internet
of Things

C. Vinothini, J. Priya, P. Anitha, and T. S. Reshmi

Abstract The technology so called Internet of Things (IoT) has been knocking
around discussion circles since few years back. Internet of Things, the technology
behind the term is actually a simple idea, allowing the physical or dumb devices to
get communicate with one another via the internet. Although the Internet of Things
appears like it is just a concept at present, it will get to most of our lives very finely
and ever soon. Also various industries will be affected distinctly. Though massive
applications are there in IoT, data are getting transferred digitally in this modern world
every day which paves the way for hackers to steal our private information. Hence the
major grief is about digital security in todays’ existence. Enormous research in this
security area is in progress which includes privacy threats, vulnerabilities and secured
communication etc,. Among these issues the crucial aspect is security and privacy.
IoT network architecture get varies with respect to different applications which are
vulnerable to threats of different types used by the hackers to make use of the IoT
network. Sybil Attack is one of them which affect the communication module of the
whole network. In this paper, a speculation of detection and mitigation of the Sybil
attack has proposed. In contrast to other existing schemes, this proposed conception
will work better in each and every situation with less time consumption.

Keywords Sybil attack · Prevention · Security · Privacy · Internet of things

1 Introduction

Internet of Things (IoT) is one of the evolving technologies employed in enormous

applications in real time includes medical, agricultural, disaster management, envi-
ronmental monitoring, and goes on. IoT network comprises of number of actuators

C. Vinothini (B) · P. Anitha

Department of Computer Science and Engineering, Dr.N.G.P Institute of Technology,
Coimbatore, India
J. Priya · T. S. Reshmi
Department of Information Technology, Bannari Amman Institute of Technology, Sathy, India
e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 581
R. Ibrahim et al. (eds.), International Conference on Artificial Intelligence
for Smart Community, Lecture Notes in Electrical Engineering 758,
https://doi.org/10.1007/978-981-16-2183-3_55
582 C. Vinothini et al.

and sensor nodes to sense and monitor the environmental or physical conditions.
The characteristics of sensors are used to determine the performance in IoT network,
natch. Few characteristics are: i. Communication failures, if anyone of the node
in the network get fails while exchanging of information with others, which must
be informed without any retardation. ii. Utilization of sensors, the way the sensors
should produce the maximum performance. iii. Heterogeneity of sensor nodes means
the sensors of distinct varieties should work in a cooperative manner. iv. Scalability,
hundreds and thousands of sensors are getting employed in a network therefore the
IoT should be designed highly scalable.
In spite of many more advantages, the conjugations of IoT devices present a pivotal
challenge to industries in terms of intense security risks emerging from unmonitored
and unsecured devices connected to the IoT network. As applications of IoT network
grow continuously, data are getting transferred digitally in this modern world every
day which paves the way for hackers to steal our private information. Therefore, it
is understood that how IoT security is important in terms of security strategy and
protection schemes that significantly protect from the chances of security attacks
and even cyber-attacks on IoT devices that are interconnected to the network built
for specific applications and functionalities. Without strong and healthy security,
any connected physical device is vulnerable to getting compromised, breached and
controlled by an attacker.
Presently, though, there is not much cause for distress. Whereas very few people
are using IoT devices in a full-fledged manner, so there won’t be as many attackers
now to get access for them. This above-said statement is not to declare that this will
be the situation always. Soon, IoT will become a major part of our life in the near
future, and as days pass, it is essential to be monitored to the maximum extent and
protected even more carefully.
One of the crucial challenges faced with the evolving IoT applications is privacy
and security, and majority of the IoT smart devices are not structured to deal with
security and privacy issues [1]. Hence the increase in number of connected devices,
the chances for malicious attacks and other security threats is unavoidable.
Therefore the major grief is about digital security in todays’ existence. Here comes
the remaining section of this paper, Sect. 2 briefs about Sybil attack and its causes.
Related works are done relevant to the proposed work given in Sect. 3. Whereas the
Proposed work and Conclusion explained in Sects. 4 and 5 are respectively.

2 Sybil Attack

Physical and Network attacks, Software attacks and Encryption attacks are few
attacks that are vulnerable to Internet of Things. It includes DoS (Denial of Service)
attack, Node tampering, Node injection or Malicious code, Physical damage, etc.,
Under physical and network attacks, Sybil Attack is the one [2]. Earlier, the peer-
to-peer network category experiences this type of attacks in the beginning in which
Detection of Sybil Attack in Internet of Things 583

a peer or node in the network directs multiple identities effectively simultaneously

and erodes the authority and power in legitimate systems.
The foremost intent of this Sybil attack is to obtain a greater influence in the
network to carry out illicit (w.r.to the instructions and protocols set to the network)
steps in the system. A single commodity (a host) has the potential to create and direct
multiple identities (like user credentials, IP address details). At the spectators from
outside, the multiple fake identities created because of the Sybil attack appear to be
the real and unique identities.
The hope and belief in most IoT networks are presumed in a tacit fashion. This
implicit assumption of hope and belief can be ill-treated by assailants to interrupt the
network and falsify the reputations of trusted devices. Spoofing devices’ identities
or forging new identities allow assailants to impersonate them as legitimate devices.
Devices with these sorts of illicit and forged identities are termed as Sybils [3].
As fundamental to Sybil attacks, harmful nodes try to interrupt and spoil the
network operations by creating several fake and manipulated identities. The Sybil
attack is a huge ruinous attack against the IoT network where numberless real iden-
tities with fabricated identities are used to get an illicit entry into the IoT network
[4–6].

3 Related Works

As above-noted point, for the first time, the Sybil attack was identified in the peer-
to-peer network [4]. In a research work related to Sybil attack, the author framed
taxonomy for different formations of Sybil attack which targets any of the networks
to interrupt or damage that classifies into i. Direct and Indirect Communication ii.
Fabricated and Stolen Identities. iii. Simultaneity [1]. In all of these classifications,
this Sybil attack intrudes in the network catches a legitimate node and makes it into
an illegal node or a legal node by “malicious node” and gives identity to itself and
then releasing several forged IDs or fabricated IDs from other true nodes.
The fake or fabricated Identities will look similar to the legitimate node identities
but that is forged for non-existing nodes. This process will happen in the network is
getting affected by the Sybil attack. As a result of the Sybil attack, the legal nodes in
the network will think they have many legitimate peers [7, 8]. Once after misleading
the legitimate node into believing malicious nodes as their neighbor nodes this attack
breaks the condition of mapping between the identity and host of nodes [9].
Numerous works has been done for detecting and mitigating Sybil attacks in the
network. Few methodologies and schemes discussed here for reference. They are:
Message Authentication and Passing, Random Password Comparison (RPS),
Channel-based detection, Compare and Match Approach and still the progress goes
on.
Though these many techniques are there in existing, some limitations also to be
considered where the attack gets compromised.
584 C. Vinothini et al.

For Example in the RPS method there is no repair mechanism for route if the route
failure is the case. In Channel- Based Detection and RADS UWB methods focuses
only on stationary networks and the indirect Sybil attacks are get detected whereas
in the approach of Compare and Match has to know about complete information
about all the existing nodes else it fails. And in the RFID method the communication
overhead is high [10].
In a further Sybil attack detection mechanism proposed in [3] relies on the Ultra-
Wide Band (UWB) ranging capabilities in the physical layer and the distance between
each node and its neighbors are monitored periodically by every node which is a time-
consuming process and also the mobility of the nodes were not considered [3]. Also,
come across an algorithm named USAS (Unpredictable Software-based Solution)
for detecting compromised nodes. Here the administers attest few selected nodes
randomly to reduce the computation time of checksum, therefore there is a chance to
miss nodes that are far from the control node for attesting [4]. Another work proposed
in [5] for WSN architecture. The lightweight proposed method similar to message
authentication and passing algorithm employs watchdog nodes that overhear Hello
Packets exchanges between nodes. Then in the form of state diagram every watchdog
nodes produce detected information partially and generates the final Sybil nodes list
by using a detection rule defined already.
By using other personal nodes identities the Sybil activity can be created is well-
known to everyone and now almost 70% of research methods in existing for Sybil
attack detection are deals with Message authentication and passing method [3] that
requires some modification for time consumption and cost-effectiveness.
A secure RPL routing has been proposed in this paper by detecting and mitigating
Sybil attack the Rank attack. Based on the parameters used for the detection of
attackers in the IoT network, the trust values of direct and indirect for all parent and
child nodes were estimated [1].
One other author described the protocols named LEACH and LEACH-E. The
author states that the energy consumption is high when the cluster node and the BS
are getting communicated while comparing non-cluster nodes in WSN networks.
Therefore it is understood that the number of clusters-heads is directly proportional
to the energy consumption which leads shorten the lifetime of the corresponding
network. Hence, number of head nodes to be considered [4]. Here this concept can
be related to IoT network in terms of the number of sensors and actuators connected
to a host.
The light-weight identity certificate method is an authentication-based approach
to defeat Sybil attacks that use Merkle hash trees and a one-way key chain scheme.
Since Merkle hash tree is being used, this method consumes more memory storage
for information storage [5].
Interestingly another idea in [6] to mitigate Sybil attack in VPNETs also defined
and proposed a novel scheme called Local Sybil Resistance (LSR) for effective
defending against Sybil attack in privacy-preserving VPNETs. Here, in LSR scheme,
the sensitive information of all vehicles including identity and location privacy can
be well protected.
Detection of Sybil Attack in Internet of Things 585

A trust-based security design for opportunistic IoT network called T_CAFE

framed by the author in [7] that computes the direct trust of a node based on parame-
ters such as correct packet ratio (CPR), amiability (Am), Frequency and forwarding
ratio (FR) and ER detects and mitigates Sybil attack and also provides the additional
benefit of isolating the attackers from the packet forwarding procedure.

4 Proposed Method

Our proposed algorithm is based on the authentication scheme for detecting and
mitigating Sybils and this work also uses the LEACH protocol [9]. This proposed
algorithm supports any kind of physical system. Here the network type chosen is IoT
as a physical system, sensor, and actuator nodes are get registered to the network and
generate their unique identities, then the credentials are shared to a particular host that
controls a network of devices in an application. Consider a high storage actuator as
a head node of the sensors which are connected to it. While sending and transferring
data to the host system via the actuator, the host will be using cryptography hash
function for the security purpose.
With reference to the KANGAROO CHF [8] scheme, the algorithm has been
proposed for the significant IoT network.
According to the proposed algorithm, If the sensor nodes and actuator nodes
are authenticated successfully once after the registration and generation of unique
Identities through the KANGAROO CHF scheme then the nodes are normal and
considered to be a legitimate and real node otherwise it must be a Sybil node.
Few parameters are considered here to calculate and by the result of it, the Sybil
attackers can be detected and mitigated [9]. The parameters are Threshold and Calcu-
lated Throughput of the receiver similarly the packet delivery ratio and total energy
consumed by the IoT network.
Once the Sybil attack is detected in the network then immediately the respective
Sybils should be mitigated from the IoT network. To mitigate it, first step is to
remove the identified malicious node by using any one of the existing algorithm like
RSSI. Let us assume a threshold values of throughput and energy are TDTh, TDEn
respectively for the IoT network. Calculate the total energy consumed by the whole
network at the deployment stage. Similarly calculate the Packet delivery ratio and
Throughput of the IoT network in the same manner. And once after designed IoT
application has been deployed for a certain need, Calculate again the parameters
throughput, packet delivery ratio and the energy consumption as said above. Now
to identify and to detect the Sybil attack, the calculated values of threshold, packet
delivery ratio and energy consumption at two different stages are to be compared.
Based on the difference in the values of these parameters calculated above deter-
mines the presence of Sybil attack and that will be shown as a major threat to the
entire system.
Thus, our proposed algorithm detects the malicious node and mitigates the same
from the IoT network. In our IoT network the Sybil attack attacks the communication
586 C. Vinothini et al.

layer or network layer of the IoT system. The procedure has been detail explained
in the below given algorithm.

Pseudo-code: Sybil Attack Detection

Prerequisite: Threshold and Calculated value of Throughput and Energy of N
nodes.
Step 1: Frame a physical system or a network with N nodes.
Step 2: Based on clustering scheme, determine the Cluster Head.
Step 3: Generate unique Identity for n nodes using LEACH protocol.
Step 4: Connect Cluster Head and Base Station of the network.
Step 5: If calculated value of
Throughput is less than the threshold value of Throughput and the calcu-
lated value of Energy is less than the threshold value of Energy then there
is no Sybil in the network.
Step 6: If Step 5 fails, then Sybil is there in the network
Step 7: Apply mitigation algorithms like RSSI to remove Sybil node from
the network.

5 Conclusion

Internet of Things (IoT) is one of the evolving technologies employed in enor-

mous applications in real time includes medical, agricultural, disaster management,
environmental monitoring, and goes on.
The major grief is about digital security in todays’ existence. Though many
techniques are there in existing, some limitations also to be considered where the
attack gets compromised. Hence our proposed algorithm based on the authentication
scheme, detects and mitigates Sybil and protects the IoT network from the Sybil
attack to the greatest extent. The throughput, energy and delay are calculated results
in better values when Sybil Attack is mitigated. In future digital signature based or
Blockchain based scheme may be proposed to detect and mitigate Sybil Attack in
Internet of Things.

References

1. Tandon F (2019) Trust-based enhanced secure routing against rank and sybil attacks in IoT. In:
2019 twelfth international conference on contemporary computing (IC3), pp 1–7
2. Priya J, Gunasekaran M (2019) Security-aware and privacy-sensitive of internet of things (IoT ):
a review. In: 5th International conference on advanced computing and communication systems
(ICACCS), pp 225–230
Detection of Sybil Attack in Internet of Things 587

3. Patel ST (2017) A review: sybil attack detection techniques in WSN. In: 4th international
conference on electronics and communication systems (ICECS), pp 184–188
4. Suriya U, Kumar R, Vayanaperumal R (2015) Detecting and preventing sybil attacks in wireless
sensor networks using message authentication and passing method. Scientif World J 1–8
5. Amuthavalli R, Bhuvaneswaran RS (2015) Genetic algorithm enabled prevention of sybil
attacks for LEACH-E. Mod Appl Sci 9(9):41–49. https://doi.org/10.5539/mas.v9n9p41
6. Zhang Q, Wang P, Reeves DS, Ning P, Carolina N (2005) Defending against Sybil attacks
in sensor networks. In: 25th IEEE international conference on distributed computing systems
workshops (ICDCSW’05), pp 1–7
7. Lin X, Member S (2013) LSR : mitigating zero-day Sybil vulnerability in privacy-preserving
vehicular peer-to-peer networks. IEEE J Selected Areas Commun/Supplement. 31(9):1–10
8. Kandhoul N, Dhurandher SK, Woungang I (2019) T _ CAFE : a trust based security approach
for opportunistic IoT. IET Commun 13:3463–3471. https://doi.org/10.1049/iet-com.2019.0657
9. Bertoni G, Daemen J, Peeters M, Van Assche G, Van Keer R (2018) In: KangarooTwelve : fast
hashing based on Keccak- p, Applied cryptography and network security book, pp 1–23
10. Lim J, Yu H, Gil J (2017) Detecting sybil attacks in cloud computing environments based on
fail-stop signature, symmetry. pp 1–12. https://doi.org/10.3390/sym9030035
Diagnosis of Brain Tumor Using ANN
with Spatial Fuzzy Clustering and Active
Level Set Contour

Hafiza Akter Munira and Md Saiful Islam

Abstract In medical science, image segmentation is a crucial and demanding factor.

It has massive usage in the diagnosis of brain tumors. Magnetic resonance imaging
(MRI) scans are used to detect the unnatural tissues of the brain efficiently which
is always a powerful tool. The brain tumor identification and segmentation process
are very steady and long delayed operation, because of the natural physical compo-
sition of human body. This paper deals with the classification of brain disease from
MRI neuro-cell images. For the image segmentation and detection of brain tumor, a
further combination of efficient method is proposed in this paper which uses inbuilt
fuzzy C-means clustering which is further rectified by spatial clustering. Along with
this segmentation, the proposed technique follows anisotropic diffusion filter and
histogram of images for image enhancement. To meet the goal of separating tumor
from normal brain cells, along with spatial fuzzy C-means, active level set contour
is used for level-based image segmentation. The extraction of features from brain
tumor is done by using DWT. With these features, the Artificial Neural Network
classifier is trained. The accuracy of the proposed method is 96.7% on the patients
699 MRI data images. The result of this proposed method has been evaluated based
on accuracy, sensitivity and specificity which is better than many existing methods.

Keywords MRI brain images · Spatial Fuzzy-C means clustering · Active level set
contour · Anisotropic diffusion filter · Histogram of image · Artificial neural
network

1 Introduction

With the rapid growth of technological advancement in image processing, the

improvement of medical science is also noticeable. Different techniques are now
commonly used, such as X-beam, ultrasonography, Computed Tomography (CT)
and Magnetic Resonance Image (MRI). Amongst them, MRI is widely used for

H. A. Munira · M. S. Islam (B)

Chittagong University of Engineering and Technology, Chattogram, Bangladesh

© The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 589
R. Ibrahim et al. (eds.), International Conference on Artificial Intelligence
for Smart Community, Lecture Notes in Electrical Engineering 758,
https://doi.org/10.1007/978-981-16-2183-3_56
590 H. A. Munira and M. S. Islam

brain tumor diagnosis. The identification of tumor as a benign or non-benign is also

very crucial. This can be classified as benign and malignance.
In this modern era, automated image processing is playing a crucial part in various
applications. To decrease the image noise, pre-processing is used which is the most
basic procedure of image analysis. The RGB to grayscale conversion is performed
in conjunction with image resizing [1]. Through the median filter, the elimination of
noise from the image data is performed. The anisotropic diffusion filter is then used
for reducing the contrast between consecutive pixels. Image enhancement is done
using histogram of images [2]. The enhanced image is then delivered to the spatial
fuzzy C-means clustering [3, 4]. The extraction of features is carried out using DWT
[5]. An important segmentation method called the active contour level set is used in
this proposed approach to monitor the boundaries of image. To improve the level set
active contour segmentation, there are many hybrid systems of fuzzy clustering [6,
7]. Actually, fuzzy clustering based on image intensity is used by such algorithms.
In this paper, fuzzy clustering incorporates with spatial information which elim-
inates the need of morphological operation. The rest of the paper is structured in
following way: Sect. 2 provides a proper discussion of this proposed method; Sect. 3
provides the result analysis. Finally, in Sect. 4 the conclusion of this proposed method
has been described.

2 Proposed Tumor Classification Methodology

The proposed methodology is implemented through the following steps, in which

each step contains the elaborate description of each process. The entire process of
the proposed implementation is shown in Fig. 1.

Dataset Pre-processing Post-processing ANN classifi-

er
High Pass Filter Spatial FCM
Median Filter Active level set con- Glioma
tour
Anisotropic Diffu-
sion Filter Feature Extraction Malignant

Image Histogram

Fig. 1 Block diagram to show the entire process of the implementation

Diagnosis of Brain Tumor Using ANN with Spatial Fuzzy Clustering … 591

2.1 Dataset

The required datasets for MRI images have been obtain from https://github.com/
sartajbhuvaji/brain-tumor-classification-dataset. Totally 699 MRI images of patients
comprised of glioma and meningioma tumor had been undertaken.

2.2 Pre-Processing

In order to get more assurance and simplicity in distinguishing the tumor [8, 9], the
motivation behind these methods is fundamentally to improve the image quality.
High Pass Filtering. It is used for edge detection and noise highlighting if any.
The algorithm steps for this pr-processing is given as: hp_fil = [−1 2 −1; 0 0 0; 1 −
2 1];
Median Filter. Median Filter is being utilized in the subsequent stage to eliminate
any undesirable clamor whenever recognized by the HPF.
Anisotropic Diffusion Filter. Anisotropic filter blurs the images without
removing any edge. This filter can be expressed as,

∂I
= div (c(x, y, t), ∇ I ) = ∇c · ∇ I (1)
∂t
where, c (x, y, t) denotes the coefficient of diffusion. Through this the rate of
diffusion is managed and is selected to preserve the image’s edges. The diffusion
co-efficient is suggested as,
 
||∇ I || 2 1
c(||∇ I ||) = exp − or c(||∇ I ||) =  2 (2)
K ||∇ I ||
1+ K

Figure 2 represents a filtered image using anisotropic diffusion filter which shows
that the image edges are not blur due to filtering.
Histogram of Images. The image was enhanced using histogram of images [9].
In Fig. 3, the enhanced image is shown.

Fig. 2 Pre-processed filter

image
592 H. A. Munira and M. S. Islam

Fig. 3 Pre-processed
enhanced image

2.3 Post-Processing

The improved image from pre-processing then reaches the post-processing phases.
Spatial Fuzzy C-means. In medical problems, fuzzy C-means clustering is widely
used because of its popularity [10, 11]. On the other hand, a membership function
u mn is used to denote the proportion of integration of nth object to the mth cluster in
fuzzy C-means. The cost function of FCM is given by:


N 
C
J= μlmn [[in −Vm ]]2 (3)
n=1 m=1

where, l (>1) is a constant through which the obtained segmentation fuzziness is

observed. The following constraints are responsible for membership function:


C 
N
μmn = 1; 0 ≤ u mn ≤ 1; μmn > 0 (4)
m=1 n=

where, μmn denotes the proportion of integration of nth object to the mth cluster
in FCM. The absence of spatial data [4, 5, 9] is the biggest issues of conventional
FCM algorithms for segmentation of image. The output of FCM segmentation will
decrease due to the image noise and artifacts. That’s why, spatial FCM algorithm is
used. A spatial FCM algorithm was suggested by Chuang [4] which can be defined
by:
p q
 μmn h mn
μmn = C p q
(5)
k=1 μkn h kn

where, the respective controlling parameters are p and q. The spatial information
is carried by h mn , can be expressed as:

h mn = μnk (6)
k∈Nn

where, the image pixel is centered at a local window given by Nn .

Diagnosis of Brain Tumor Using ANN with Spatial Fuzzy Clustering … 593

Active level set contour. Effective contrast boundaries for an image segmentation
is used in active level set contour [12, 13]. With the help of active contour level set,
image segmentation is a popular process [14]. Using the zero-level tracking  t, the
evolution of active contours can be given by:

ϕ(t, x, y) less than 0 : when, value of (x, y) is placed in  t

ϕ(t, x, y) equal to 0 : when, value of(x, y) remains at  t.
ϕ(t, x, y) greater than 0 : when, value of(x, y)located outside  t.

Feature Extraction. For feature extraction, the proposed approach follows

discrete wavelet transform (DWT) for statistical feature extraction. From the MRI
input images, the following features were extracted: contrast, mean, energy, variance,
correlation, skewness, kurtosis, inverse directional moment, standard deviation and
entropy.
ANN Classifier. ANN classification is carried out after extraction of the necessary
features. The number of iterations of the network classifier is 1000. The initial weights
of the network are randomized. The number of hidden layers of this network is 20
and training function is Resilient backpropagation (trainrp).

3 Results Analysis

The implementation of both spatial FCM algorithms and active level set contour has
been done in MATLAB 2020a. Through this proposed approach, initial segmentation
of images is possible with level set active contour and the accuracy of the trained
network on the 105 tested image is 96.7%. The network is trained on 489 MRI images
and validation and testing image are 105.
In Fig. 4, the segmented image after using spatial FCM (a, c) and active level set
contour image with edges (b, d) are shown.
In Fig. 5, the confusion matrix of the training is shown and Table 1 represents
the performance of the network based on sensitivity, accuracy and specificity. The
comparison of this proposed method with others is shown in Table 2.

Fig. 4 Spatial FCM segmentation (a, c) and active level set contour after 100 iterations (b, d)
594 H. A. Munira and M. S. Islam

Fig. 5 Confusion matrix of training

Table 1 Classification
ANN classifier result
performance
Sensitivity Specificity Accuracy
97.38% 95.42% 96.7%
Diagnosis of Brain Tumor Using ANN with Spatial Fuzzy Clustering … 595

Table 2 Comparison of this proposed approach with existing ones

Method Accuracy (%) Result
Region based segmentation + 63 It cannot detect the edges smoothly,
watershed segmentation + ANN [8] sometimes the high-density tumor
tissues edge detection becomes difficult
K-means segmentation + Gabor 88 Not only the affected tissue is
filter + ANN segmented, also the normal tissue is
segmented with this segmentation
Fuzzy C-means segmentation [3] 92 Due to the noise and artifacts, it is not
strong enough for medical image
segmentation
Proposed method 96.7 It gives initial segmentation of image
with proper edge detection

Thus, the improved spatial FCM segmentation is esteemed for trying to merge
intensity and spatial information altogether. This algorithm esteemed less noise and
it is perfect for initial level set contour in medical image processing.

4 Conclusion

For brain tumor detection, an automated image segmentation is proposed in this paper
using enhanced spatial fuzzy C-means with active level set contour. The proposed
approach contains the pre-processing steps where the images are being ready for
further segmentation using noise removal median filter, anisotropic diffusion filter
and histogram of images. Then, the images are sent to post-processing steps where
the spatial fuzzy C-means is used with active level set contour for better detection of
boundaries of abnormal tissues. The statistical extracted features are trained using
ANN where the accuracy is found 96.7% on test data which is far better than many
existing methods.

References

1. Loges Wari T, Karnan M (2010) In: International conference on signal acquisition and
processing, an enhanced implementation of brain tumor detection using segmentation based
on soft computing
2. Hemanth DJ et al (2009) Effective fuzzy clustering algorithm for abnormal MR brain image
segmentation. In: Advance computing conference 2009. IACC 2009, IEEE international. pp
609–614
3. Suganya R, Shanthi R (2012) Fuzzy C-Means algorithm—a review. IJSRP 2(11. ISSN 2250–
3153
4. Aslam A, Khan E, Beg MMS (2015) Improved edge detection algorithm for brain tumor
segmentation. In: Second international symposium on computer vision and the internet
(VisionNet’15), Elsevier
596 H. A. Munira and M. S. Islam

5. Lakra A, Dubey RB (2015) A comparative analysis of MRI brain tumor segmentation technique.
Int J Comput Appl 125:5–14 (0975–8887)
6. Sudharani K, Sarma TC, Prasad KS (2015) Advanced morphologicaltechniqueforautomat-
icbraintumordetectionandevaluationof statistical parameters. In: International conference on
emerging trends in engineering, science and technology (ICETEST)
7. Dhage PM, Phegade MR, Shah SK (2015) Watershed segmentation brain tumor detection. In:
International conference on pervasive computing (ICPC)
8. Kumbhar U, Patil V, Rudrakshi S (2013) Enhancement of medical images using image
processing in MATLAB. Int J En Res Technol 2(4):2359–2364
9. Chuang KS, Hzeng HL, Chen S, Wu J, Chen TJ (2006) Fuzzy c-means clustering algorithms
with spatial information for image segmentation. Comput Med Imaging Graph 30:9–15
10. Cai W, Chen S, Zhang D (2007) fast and robust fuzzy c-means clustering algorithms
incorporating local information for image segmentation. Pattern Recogn 40:825–838
11. Osher S, Fedkiw R (2003) Level set methods and dynamic implicit surfaces. Springer, New
York
12. Chan TF, Vese LA (2001) Active contours without edges. IEEE Trans Image Process 10:266–
277
13. Varuna shree N, Kumar TNR (2018) Identification and classification of brain tumor MRI images
with feature extraction using DWT and probabilistic neural network. In: Springer corporate
information
14. Havaei M et al (2017) Brain tumor segmentation with deep neural networks. Med Image Anal
35:18–31
Current Overview of Machine Learning
Application for Predicting Steam Huff
and Puff Injection Production
Performance

Mohammad Galang Merdeka, Syahrir Ridha, Berihun Mamo Negash,

and Suhaib Umer Ilyas

Abstract Thermal Enhanced Oil Recovery (EOR) is one of the main contributors
to EOR worldwide production. Steam huff and puff injection, one of its methods, is a
technique in which steam is injected in a cyclical manner alternating with oil produc-
tion. Reservoir simulation is considered as the most reliable solution to evaluate the
reservoir performance and designing an optimized production scheme. However, it
still remains time-consuming and expensive. Applying machine learning to build a
predictive proxy model is a suitable solution to deal with the issue. Presently, there
have been a limited number of studies covering the topic of proxy model development
to estimate production performance for this injection method. This study provides a
review of the machine learning implementations for estimating steam huff and puff
injection production performance, starting with an introductory explanation about
the method, followed by the currently deployed machine learning models along with
the challenges and future prospects.

Keywords Enhanced oil recovery · Machine learning · Oil and gas · Proxy
model · Reservoir engineering · Steam huff and puff

1 Introduction

As most of the giant oil fields are all currently at the mature stage, Enhanced Oil
Recovery (EOR) is considered a viable solution to recover the remaining oil after the
primary and secondary recovery stages. It involves the injection of a fluid or fluids
into a reservoir intending to supply the energy needed to displace oil. Furthermore,
the injected fluids interact with the reservoir rock and fluid to alter the physical
properties and create an advantageous condition to improve oil recovery [1].

M. G. Merdeka (B) · S. Ridha · B. M. Negash

Petroleum Engineering Department, Universiti Teknologi PETRONAS, Seri Iskandar, 32610
Perak, Darul Ridzuan, Malaysia
S. Ridha · S. U. Ilyas
Institute of Hydrocarbon Recovery, Universiti Teknologi PETRONAS, Seri Iskandar, 32610
Perak, Darul Ridzuan, Malaysia

© The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 597
R. Ibrahim et al. (eds.), International Conference on Artificial Intelligence
for Smart Community, Lecture Notes in Electrical Engineering 758,
https://doi.org/10.1007/978-981-16-2183-3_57
598 M. G. Merdeka et al.

Thermal EOR is one of the main EOR methods applied worldwide operating
on the mechanism involving reduction of oil viscosity by the injection of thermal
energy. As reported by the International Energy Agency, it accounts for more than
40% of total EOR production in 2015 [2]. The target oil is a high-density and high-
viscosity oil. Heavy oils are such example for the target, and several reports on the
thermal EOR implementations showed that almost all of the reservoirs are heavy oil
reservoirs.
In evaluating the performance of a reservoir and designing an optimized produc-
tion scheme, reservoir simulation is conducted beforehand. It is considered as the
most reliable solution. However, the whole process of reservoir simulation still
remains time-consuming and expensive due to the complexity of a reservoir system
and the difficulty of building a reliable full reservoir model. Building predictive
proxy model using machine learning algorithms is considered a suitable solution to
deal with the issue, where proxy models will estimate the reservoir simulation results
using lighter mathematical models [3]. In reservoir engineering, the proxy model has
been mainly used for sensitivity analysis, risk analysis, history matching, produc-
tion forecasting and production optimization [4]. The advantages of the proxy model
over reservoir simulation are the computational speed and significantly economical
solution, at a tradeoff of slightly lower accuracy.
This paper will provide a review of the research works related to machine learning
applications in predicting steam huff and puff injection method. It is aimed to summa-
rize the findings and highlight the advantages and disadvantages of applying machine
learning approaches in this research area. The discussion will start with an introduc-
tory explanation of steam huff and puff method, including the how-to and past field
successes. Then, the currently deployed machine learning models for this method will
be explored. Finally, the challenges and future prospects to utilize machine learning
coupled with current reservoir engineering technologies will be discussed.

2 Steam Huff and Puff Injection Method

Steam huff and puff injection, also known as cyclic steam stimulation, is a method in
which steam is cyclically injected into the well. After injection, the well is shut-in,
allowing the steam to “soak” into the reservoir. The high temperature of the steam will
reduce oil viscosity near the steam-oil interface. This is called the soaking period.
The well is opened again after the soaking period, producing oil at a higher rate.
After some time, the heat dissipates, and the near-well oil viscosity returns to its
original value. The production declines and after reaching an economic limitation
rate, steam is reinjected into the well, starting a new injection cycle. This whole
process of injection, soaking and production is called the huff and puff method and
illustrated in Fig. 1. Several screening criteria of this method had been developed in
the past [5–7]. In general, a typical steam huff and puff process is well suited for
heavy oil reservoirs with high porosity and permeability, shallow depths and high oil
saturation.
Current Overview of Machine Learning Application … 599

Fig. 1 Steam huff and puff injection schematic by Green and Willhite [1]

Even though the recovery factor is still smaller in comparison to the other steam
injection methods, this injection method is still highly attractive due to small capital
investment and quick payout. It is one of the most implemented thermal EOR methods
with numerous reported commercial success in the past, for examples: Cold Lake
Field in Alberta, Canada [8, 9], Tia Juana Field in Venezuela [10, 11], Midway Sunset
Field in California [12], Liaohe Field in China [13] and Duri Field in Indonesia [14–
16]. Furthermore, it is often used as a precursor before conducting a full field scale
steam injection. To predict the performance of steam huff and puff injection, several
semi-analytical models had been developed in the past [12, 17–20]. However, the
drawbacks of those models are the limitations imposed by the assumption of the
underlying physical equation and the need to have a reasonable history matching
before using the models for prediction.

3 Machine Learning Applications in Predicting Steam Huff

and Puff Injection Production Performance

3.1 Introduction

Machine learning (ML) is currently gaining a lot of attention to the recent break-
through in data analysis. It is a tool that gives computers the ability to learn without
being explicitly programmed. In oil and gas, especially reservoir engineering, most
of the ML problem types were supervised learning problems, e.g. assisted history
matching and forecasting and optimizing production [4, 21]. Those problems can be
600 M. G. Merdeka et al.

challenged by applying machine learning algorithms to develop a proxy model, which

is a mathematically or statistically defined function that approximates a real system
or its simulation. Based on the objectives, the developed model can be utilized for
specific field case or presented as a general or universal model [22]. For specific case
model, the obtained field data are used to train the model and solve problems related
to that field. Conversely, the universal model is used to solve a category of prob-
lems regardless of the variation of reservoir properties. It is trained using synthetic
data obtained from reservoir simulation results by running numerous simulation
experiment cases with varied inputs.

3.2 Existing Proxy Models for Steam Huff and Puff Injection

Although there were many research works related to steam huff and puff performance
prediction, a very limited number related to the use of machine learning to develop
proxy model was found. In predicting steam huff and puff performance, the most
used machine learning algorithm was the Artificial Neural Network (ANN).
ANN is a model that emulates the biological neural system. It consists of nodes
which are similar to neurons in the human brain. A node receives signals from
adjacent nodes and processes them to produce an output. A generalized ANN model
structure is presented in Fig. 2.
It consists of an input layer, an output layer, and one or multiple hidden layers in
between. Each neuron from a layer is connected to the neurons from other nearby
layers with a connection called weight, which represents an influence of corre-
sponding input on the connected neuron. Then, to transfer the contained information,
an activation (or transfer) function is used. This is called forward propagation and
it is conducted for every connection in the network. After defining the model, the

Fig. 2 Artificial neural

network structure by Arpaci
[23]
Current Overview of Machine Learning Application … 601

backpropagation algorithm is used to readjust the weight and minimize the error
between the output and the database.
Several existing researches on ANN proxy models for steam huff and puff injec-
tion are presented below. They were presented as a general model and some are
the continuation and improvements from previous works over the years. ANN was
deployed to predict the production performance based on input parameters such as
reservoir properties and operating conditions.
Arpaci developed ANN models for steam huff and puff injection using horizontal
wells in naturally fractured reservoir [23]. The inputs of the model are reservoir rock
and fluid properties, operation design parameters and fracture design parameters.
The outputs are the number of cycles, cycle duration, oil flowrate, and cumulative
oil production. Synthetic data were generated using the help of a thermal reservoir
simulator. In addition, inverse models were developed to predict operation parameters
and fracture design parameters, using performance indicators such as oil production
and production period as additional inputs. The errors of each built models are less
than 5%. However, most of the reservoir parameters were assumed constant, such as
oil density, relative permeability, anisotropy and capillary pressures.
Sun and Ertekin developed a data-driven proxy model for steam-assisted gravity
drainage (SAGD) and steam huff and puff injection in naturally fractured reservoir
[24]. There are two kinds ANN models that were developed: a supervisory model
that predicted the number of cycles and a sub-model that predicted the reservoir
performance. It was set that the cycle automatically switches when the oil production
rate hits a certain economic limit, thus the total number of cycles in the end of
simulation period is not certain for every simulation case. The inputs will first visit
the supervisory model to determine the number of cycles, then the sub-model will
be predicting the reservoir performance, using both inputs from the database and the
result of the supervisory model. The developed models show a good match with the
database obtained from the numerical simulation. In addition to the ANN models,
a Graphical User Interface was developed to facilitate easier use of the models to
the end-users. Although this study had covered the disadvantage of less reservoir
parameters from the previous study, the effect of the heterogeneity, anisotropy and
capillary pressures were still neglected. The other limitations came from the fluid
model assumption where the fluid composition was dead oil and water, and the
number of predicted cycles in the sub-model were limited to 10 cycles.
Ersahin and Ertekin continued the previous study by Sun and Ertekin and devel-
oped ANN models for steam huff and puff injection in naturally fractured reservoir
[25]. In this study, there were two main performance indicators that became the
output of the model, which are cumulative production profile and viscosity contours.
The simulations were performed to understand the heat-exchange mechanisms in
the stimulation zone around the wellbore. Improvements were also carried out to the
input parameters, e.g., the relative permeability table was generated using correla-
tions specifically designed for sandstone rocks, as most of the thermal EOR projects
were performed in sandstone rocks [26]. The results were good for predicting the
outputs, with average error for cumulative production profile and viscosity contours
were 3.64% and 5.32 respectively.
602 M. G. Merdeka et al.

3.3 Challenges and Future Prospective

A limited number of research works encourage further explorations on this topic. The
challenges would be coming from the difficulties of building a reliable model itself
and finding the effective use of machine learning alongside conventional reservoir
engineering practices.
Current models were able to be developed due to the huge amount of available data.
Finding the data to train the model would be difficult if there are less implementation
of the methods themselves, especially in the future, where new innovations on steam
huff and puff injection method will emerge. The other challenge is the limitation of
the model. Proxy model is basically an interpolation model. Hence, it can only be
utilized when all the values of input parameters fall within the range of the training
data. Having more data with larger intervals of inputs will help in developing a better
general model.
Although proxy model provides a faster and cheaper computational solution, the
weakness of the technology should not be ignored. Conventional reservoir simulation
is still the most reliable tool in reservoir engineering and proxy model should not
be used as a replacement. Instead, it should be used as an assisting tool to achieve
reservoir engineering analysis more effectively. For example, as a starting guide in
creating a field or well development plan.
The future prospects of this research topic lie on the technological innovations of
steam huff and puff methods itself, as well as the advancements of machine learning
and its subset, deep learning. Thermal EOR is projected to be still one of the highest
contributors of EOR production in 2040, second only to miscible EOR [2]. There
have been several studies and field trials of the new modifications to steam huff and
puff injection, such as including additives to steam, conducting injection in different
well configurations and combining cyclic injection with fracturing [27]. Therefore,
one can expect more innovations in this method and their field implementations to
gather more data for modelling purposes. On the other side, improvements in machine
learning may highly contribute to determine the best approach to develop a proxy
model for performance prediction. For instance, choosing the best training algorithm
or determining the best adjustments for hyperparameters of the proxy model.

4 Conclusion

This study provided an overview of the steam huff and puff injection methods and
the application of machine learning to predict the production performance. Machine
learning helps to improve the reservoir simulation approaches in predicting the
production performance by developing a proxy model which is highly efficient in
reducing the required time and cost, at a tradeoff of slightly lower accuracy. However,
it should be noted that the proxy model is not a replacement for reservoir simulation,
but rather as an assisting tool that effectively helps in reservoir engineering practice.
Current Overview of Machine Learning Application … 603

For future prospective, the technological innovations of steam huff and puff method
and the advancements of machine learning are expected to explore further towards
more viable solutions.

References

1. Green DW, Willhite GP (1998) Enhanced oil recovery, 1st edn. Society of Petroleum Engineers
2. McGlade C, Sondak G, Han M (2018) Whatever happened to enhanced oil recovery?. Inter-
national energy agency (2018). https://www.iea.org/commentaries/whatever-happened-to-enh
anced-oil-recovery. Last Accessed 28 Oct 2020
3. Negash BM, Tufa LD, Ramasamy M et al (2017) System identification based proxy model of
a reservoir under water injection. Model Simul Eng 7645470
4. Zubarev DI (2009) Pros and cons of applying proxy-models as a substitute for full reservoir
simulations. In: SPE annual technical conference and exhibition, New Orleans, Lousiana, USA
5. Taber JJ, Martin FD, Seright RS (1997) EOR screening criteria revisited—part 1: introduction
to screening criteria and enhanced recovery field projects. SPE Reserv Eng J 12(3):189–197
6. Farouq Ali SM (1974) Current status of steam injection as a heavy oil recovery method. JPT
13(1):54–68
7. Sheng J (2013) Enhanced oil recovery field case studies. Gulf Professional Publishing
8. Buckles RS (1979) Steam stimulation heavy oil recovery at cold lake, Alberta. In: SPE
California regional meeting, 18–20 April, Ventura, California
9. Beattie CI, Boberg TC, McNab GS (1991) Reservoir stimulation in the cold lake oil sands.
SPE Reservoir Eng
10. De Haan HJ, Van Lookeren J (1969) Early results of the first large-scale steam soak project in
the Tia Juana. J Petrol Technol 21
11. Puig F, Schenk L (1984) Analysis of the performance of the M-6 area of the Tia Juana Field,
Venezuela, under primary, steam-soak, and steamdrive conditions. In: Society of petroleum
engineers AIME, SPE, vol 1. pp 263–276
12. Jones J (1977) Cyclic steam reservoir model for viscous oil, pressure depleted, gravity drainage
reservoirs. In: SPE California regional meeting California
13. Liu W-Z (1997) Steam injection technology to produce heavy oils. Petroleum Industry Press
Beijing, China
14. Gael BT, Gross SJ, McNaboe GJ (1995) Development planning and reservoir management in
the Duri steam flood. In: SPE western regional meeting, 8–10 March, Bakersfield, California
15. Pearce JC, Megginson EA (1991) Current status of the duri steamflood project Sumatra,
Indonesia. In: SPE International thermal operations symposium, 7–8 February, Bakersfield,
California
16. Bae WS, Masduki A, Permadi AK, Abdurrahman M (2017) EOR in Indonesia: past, present,
and future. Int J Oil and Gas Coal Technol 16(3):250
17. Boberg TC, Lantz RB (1966) Calculation of the production rate of a thermally stimulated well.
J Petrol Technol 18(12):1613–1623. Society of Petroleum Engineers
18. Gontijo JE, Aziz K (1984) A simple analytical model for simulating heavy oil recovery by cyclic
steam in pressure-depleted reservoirs. In: SPE annual technical conference and exhibition,
16–19 September, Houston, Texas
19. Sylvester NB, Chen HL (1988) Improved cyclic steam stimulation model for pressure-depleted
reservoirs. Soc Pet Eng AIME, SPE, pp 213–226
20. Gozde S, Chhina HS, Best DA (1989) Analytical cyclic steam stimulation model for heavy oil
reservoirs. In: SPE California regional meeting, 5–7 April, Bakersfield, California
21. Jaber AK et al (2019) A review of proxy modelling applications in numerical reservoir
simulation. Arabian J Geosci. Saudi Society for Geosciences
604 M. G. Merdeka et al.

22. Ertekin T, Sun Q (2019) Artificial intelligence applications in reservoir engineering: a status
check. Energies. MDPI
23. Arpaci B (2014) Development of an artificial neural network for cyclic steam stimulation
method in naturally fractured reservoirs. Master Thesis, The Pennsylvania State University
24. Sun Q, Ertekin T (2015) The development of artificial-neural-network-based universal proxies
to study steam assisted gravity drainage (SAGD) and cyclic steam stimulation (CSS) processes.
In: SPE Western regional meeting, 27–30 April, Garden Grove, California, USA
25. Ersahin A, Ertekin T (2019) Artificial neural network modeling of cyclic steam injection process
in naturally fractured reservoirs. SPE Reservoir Evaluation and Engineering. SPE
26. Alvarado V, Manrique E (2010) Enhanced oil recovery: an update review. Energies
27. Alvarez J, Han S (2013) Current overview of cyclic steam injection process. J Petroleum Sci
Res 2(3)
Enhanced Cluster Head Based Data
Gathering (ECHGS) Technique in IoT
Based Smart Irrigation System

K. Padmanaban, A. M. SenthilKumar, A. K. Velmurugan, and E. S. Madhan

Abstract Internet of Things (IOT) is an emerging technique which is used in several

applications like home automation, office automation, hospital monitoring, network
device monitoring, industrial related problems and smart agricultural monitoring
system. In IoT environment, sensors are dynamically connected with the network
to monitoring and routing the data to the server. Energy utilization and maximizing
throughput are important issues in the Internet of Things (IOT). Energy plays a vital
role in increasing network lifetime. Throughput improves the network performance
and it is calculated based on average successful data delivered at a time instance.
Various models such as cluster tree, clustering and Tree Cluster based Data Gathering
Scheme (TCDGS) are used for handling maximizing energy and throughput. In this
paper, a novel technique “Enhanced Cluster Head based Data gathering” (ECHGS)
is presented to focus the energy and delay in the Internet of Things (IOT) based
smart irrigation system for agriculture. In this proposed technique, the cluster head
for the group of sensor nodes is selected based on the workload and energy. The
node which has high residual energy and less workload is selected as a cluster head
to collect and deliver the data to the server node effectively. The proposed technique
results are evaluated and compared to the existing techniques in terms of residual
energy and throughput. The residual energy is saved by 4.54% higher than the Tree
Cluster based Data Gathering Scheme (TCDGS) and the throughput is maximized
by 11.57% higher than the Tree Cluster based Data Gathering Scheme (TCDGS).

K. Padmanaban (B) · A. K. Velmurugan

Department of Computer Science and Engineering, Koneru Lakshmaiah Education Foundation,
Vaddeswaram, Guntur, Andhra Pradesh, India
A. M. SenthilKumar
Department of Computer Science and Engineering, Vel Tech Rangarajan Dr. Sagunthala R & D
Institute of Science & Technology, Chennai, Tamil Nadu, India
E. S. Madhan
Department of Computational Intelligence, SRM Instititute of Science and Technology,
Kattankulathur, Tamil Nadu, India

© The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 605
R. Ibrahim et al. (eds.), International Conference on Artificial Intelligence
for Smart Community, Lecture Notes in Electrical Engineering 758,
https://doi.org/10.1007/978-981-16-2183-3_58
606 K. Padmanaban et al.

Keywords Internet of things · Tree cluster based data gathering scheme

(TCDGS) · Enhanced cluster head based data gathering (ECHGS) · Residual
energy

1 Introduction

Different physical devices such as washing machines, refrigerators, air conditioners

and television are connected to a heterogeneous network called the Internet of Things
(IoT). Different devices connected to an IoT have different topology and battery
functionality makes a complex network [1]. In IoT, the wireless sensor is an important
part. Actuators of the sensors are used to sense the data and send the information
to the server [2]. This data can be analyzed based on the problem specified in the
application. Various sensors located inland are used to transmit the data to the server
with the help of work station. Energy and lifetime are important aspects of the sensors
[3].
Energy plays an important role in the sensors which are used in IoT based appli-
cations. Residual energy can increase the lifetime of the network. Residual energy
is calculated based on the energy remains in the sensor nodes after transmitting the
sensed data and information to the servers. Any node’s failure due to less energy can
lead the entire network to become inactive. In the IoT based applications, another
major factor to be considered is a throughput. All the sensed data which are trans-
mitted from the sensor nodes must reach the server with less time. If any data delivered
with a high failure rate to the servers cause the poor performance of the entire IoT
application [4].
Various techniques are used to focus the residual energy and throughput in the IoT
applications during the data collection, data sensing and data transmitting phases.
Cluster tree, Clustering and Tree Cluster based Data Gathering Scheme (TCDGS) are
some of the existing techniques used to focus the energy and throughput. In all these
techniques, sensors are clustered together into the group to minimize the coverage
area to save the energy and minimize the delay.
The major contribution of this paper is discussed below,
1. The proposed Enhanced Cluster Head based Data gathering (ECHGS) technique
improves the residual energy level.
2. The proposed technique maximizes the throughput during the data transmission
between the sensors and the server.
3. The Proposed technique results are validated and compared with the existing
techniques Cluster Independent Data gathering Technique (CIDT) and TCDGS.
The rest of the paper is organized as Sect. 2 details the literature review of the
existing techniques. Section 3 shows the proposed technique. Section 4 includes the
experimental setup and results. Section 5 concludes the work.
Enhanced Cluster Head Based Data Gathering (ECHGS) Technique … 607

2 Literature

Wang et al. [5] proposed a path planning algorithm to reduce the mobile nodes count.
In this technique, an efficient algorithm is designed to reduce the nodes on the path by
path optimization. The results of the path planning algorithm are outperformed than
the existing technique CSWEEP in terms of travel length and number of nodes. Lee
et al. [6] presented a model based on Neural Network with input and output data as
working hours, temperature, weather, humidity and electrical consumption for IoT
applications. Model is verified with test data and performance is evaluated. Energy
consumption is optimized in this work. Mean squatted error method is applied to
the obtained data and simulation is done. Ke et al. [7] introduced a novel method
called Joint Optimization method for Data Offloading, Renewable energy-aware,
and Bandwidth allocation for IoT devices based on deep Reinforcement Learning
(JODRBRL). This method minimizes the total cost which includes data delay cost,
bandwidth cost and energy consumption cost by adaptive learning from the dynamic
IoT environment. The JODRBRL results are better than the existing techniques
Double DQN (DDQN), Dueling DQN and greedy policy.
Guo et al. [8] proposed a comprehensive energy estimation model for IoT
networks. This proposed model analyzes the insufficient random-access procedure
and utilize the power ramping and preamble picking for optimizing energy utilization.
Also, Multi-Agent Reinforcement Learning (MARL) algorithm based on Win-or-
Learn-Fast Policy Hill-Climbing (WoLF-PHC) is introduced in this work to decrease
the algorithm complexity. The results are evaluated and compared with the random
picking method. The proposed WoLF-PHC method results outperformed the random
picking method results. Padmanaban et al. [4] presented a method called Tree Cluster
based Data Gathering Scheme (TCDGS) in wireless sensor networks. In this work,
nodes are clustered and form a tree-like structure. Cluster head node is identified
for each cluster to maximize the residual energy. Data gathering is done by the data
gathering node in the network and data is sensed from the cluster head instead of
sensing all the nodes in the cluster and transfer to the base station. This particular
module increases the life time of the network. The TCDGS results are better than
the existing method of CIDT results.
All these existing works [9–17] have not considered residual energy and time
delay in the IoT based agriculture smart irrigation system. In this proposed work,
the residual energy and average time delay are considered. The residual energy is
maximized for all the nodes and average time delay is minimized when compared to
the existing methods CIDT and TCDGS.
608 K. Padmanaban et al.

3 The Proposed IoT Model for Agriculture Smart

Irrigation System

The proposed architecture for the smart irrigation system in agriculture is shown in
Fig. 1. The agriculture lands are shown in the diagram with circles. The clusters are
formed based on the agriculture lands. Entire agriculture area is divided into lands.
Each land among the total agricultural area is considered as a cluster. There are many
sensors placed in each cluster. The sensors are placed in each agricultural land to
collect the water level information from the land. In this paper, the Enhanced Cluster
Head based Data gathering” (ECHGS) technique is used to collect the data from
sensors effectively by proposing a new module to identify the cluster head. In each
cluster of sensors, the cluster head is identified by conducting the poll based on the
load and energy parameters of sensor.
The poll is conducted by substation periodically and the sensor node which has
maximum energy and minimum load is considered as cluster head. The elected cluster
head is responsible for collecting water-related data and send it to the substation.
The substation forwards the data which is received from the cluster head to the base
station. The substations are placed in all the agriculture land. The base station which
sends the data that is received from the substation to the data server. The data server
stores the water data and calculates the irrigation level for each agriculture lands.
This processed information about water irrigation is given as output to the users who
are accessing the data server to know the level of water irrigation to their lands. Each
sensor’s load and energy values are varying in every time instance. Current cluster
head’s load and energy level is changed after transmission of data. To maintain the
efficiency of the IoT model after every data transmission from sensors, the proposed
algorithm suggests an idea to identify the alternate cluster head which has more
energy and less load is to be elected as cluster head by the substation.

Fig. 1 Architecture diagram

Enhanced Cluster Head Based Data Gathering (ECHGS) Technique … 609

The energy in a sensor is calculated based on transmitting and receiving the data
in Eqs. (1) and (2). The formula used to calculate the energy is given as,

T x E (m, l) = E ∗ m + λ ∗ m ∗ l 2 (1)

Rx E (m) = E ∗ m (2)

where, l is the length between two sensors, E is the consumption of energy factor
denoting the power per bit during transmission and receiving the data and λ is the
amplifier coefficient to send a single bit.
The proposed flowchart of ECHGS method is shown in the Fig. 2. Initially the
clusters are formed based on the land area among the total agriculture. Sensors are
placed equally in all the clusters. Cluster head is selected in each cluster based on
the high energy and less load values.
Data gathering is done by the cluster head and send to the substation. Substation
forwards the data to the data server for analyze the level of irrigation to each cluster.
These analytical reports can be given as output to the users when they use the IoT
based smart irrigation system.

4 Results

The proposed method results are analyzed using Network Simulator version-2 (NS2).
The nodes in the simulation environment are distributed in the communication
network. The channel type used in this simulation is a wireless channel, simula-
tion time considered is 100 s, the number of nodes considered are 50, MAC type
of the protocol is 802.11, traffic model used is CBR, transmission range is 250 m,
network interface type used is wirelessPhy and mobility model used is random way
model. In this simulation environment, the proposed method ECHGS results are
evaluated in terms of residual energy and throughput. The proposed ECHGS method
results are outperformed the results of TCDGS and CIDT.

4.1 Residual Energy (RE)

The proposed model in the IoT environment aims to increase the network lifetime
by maximizes the throughput. The Residual Energy (RE) is calculated as the sum of
remaining energy available in all the nodes as shown in Eq. (3) Residual Energy in
each node is the amount of energy remaining in a sensor node at the current moment
of time.
610 K. Padmanaban et al.

Fig. 2 ECHGS flow

diagram


M
RE = R Ei
i=1
(3)

Residual Energy is the remaining energy stored in the node. Figure 3 shows
the residual energy comparison of proposed system ECHGS and existing methods
TCDGS and CIDT. The proposed method ECHGS residual energy is higher than the
existing methods of TCDGS and CIDT. Residual is maximized gradually from the
Enhanced Cluster Head Based Data Gathering (ECHGS) Technique … 611

Fig. 3 Residual energy

simulation time 0 s to 100 s. The proposed ECHGS method residual energy is higher
by 4.54 and 14.59% when compare to TCDGS and CIDT methods.

4.2 Throughput (Th)

The proposed model in IoT environment aims to increase the performance by maxi-
mizes the throughput. Throughput is the total successful data delivered to the server
in the IoT environment. It is calculated Eq. (4) as the total successful data transferred
(RP) in a unit of time (T).


N
Th = R P j/T
j=1 (4)

Figure 4 shows the throughput of the proposed method. The throughput is eval-
uated for the proposed system ECHGS and compared with the existing methods
TCDGS and CIDT. Throughput of the ECHGS method is higher than the existing
methods TCDGS and CIDT. Throughput is increased gradually when the simulation
time 0 s to 100 s. The proposed ECHGS method throughput is higher by 11.57 and
50.00% when compare to TCDGS and CIDT methods.
612 K. Padmanaban et al.

Fig. 4 Throughput

5 Conclusion

In this paper, the IoT based smart irrigation system using ECHGS method is discussed
to address the solutions to increase the network lifetime and improve the network
performance. The sensor nodes are clustered and a cluster head for each node is
selected based on the essential parameters such as residual energy and load. The
effective way of selecting a cluster head in each cluster leads to the better results
from sensor node data collection step to data transfer up to the server. The proposed
ECHGS method results are evaluated and compared with the existing methods of
TCDGS and CIDT in terms of residual energy and throughput. The proposed ECHGS
method results are better than the existing methods. In future, IoT based smart irriga-
tion system can be implemented in real-time with other parameters packet delivery
ratio and packet loss ratio.

References

1. Rapate GS, Naveen NC (2018) Energy and routing efficiency in IoT: proposal for combined
approach. In: 2018 International conference on electrical, electronics, communication,
computer, and optimization techniques (ICEECCOT), Msyuru, India, pp 451–454
2. Atzori L, Iera A, Morabito G (2010) The internet of things: a survey. Comput Netw
54(15):2787–2805
3. Deshpande Niranjan R, Vadane Pandurang M, Sangle Sagar D, Dighe MS (2016) A IOT-based
modern healthcare system using body sensor network (BSN). Int J Innov Res Comput Commun
Eng 4(11):19540–19546
4. Padmanaban K, Jagadeesh Kannan R (2016) Tree cluster based data gathering scheme
(TCDGS) in wireless sensor networks. Int J Comput Technol Appl 9(61):2809–2818
Enhanced Cluster Head Based Data Gathering (ECHGS) Technique … 613

5. Wang C, Ma H (2011) Data collection in wireless sensor networks by utilizing multiple mobile
nodes. In: 2011 seventh international conference on mobile ad-hoc and sensor networks,
Beijing, pp 83–90
6. Lee S, Cha J, Kim KS (2019) Data gathering and application to building energy optimization
with sensitivity analysis for IoT applications. In: 2019 International SoC design conference
(ISOCC), Jeju, Korea (South), pp 184–185
7. Ke H, Wang J, Wang H, Ge Y (2019) Joint optimization of data offloading and resource allo-
cation with renewable energy aware for IoT devices: a deep reinforcement learning approach.
In: IEEE Access vol 7. pp 179349–179363
8. Guo Y, Xiang M (2019) Multi-agent reinforcement learning based energy efficiency optimiza-
tion in NB-IoT networks. In: 2019 IEEE globecom workshops (GC Wkshps), Waikoloa, HI,
USA, pp 1–6
9. Appala Raju V, Sri Harsha V, Bhanu Deepthi N, Prasanth N (2018) Zonal stable election
protocol for heterogeneous wireless sensor networks. Int J Eng Technol (UAE) 7:725–728
10. Dhage MR, Vemuru S (2018) A effective cross layer multi-hop routing protocol for heteroge-
neous wireless sensor network. Indonesian J Electri Eng Comput Sci 10(2):664–671. https://
doi.org/10.11591/ijeecs.v10.i2.pp664-671
11. Goutham Chand K, Sidhendra M, Hussain MA (2018) Soil nutrient measurement in paddy
farming using IoT. Int J Eng Technol (UAE) 7:356–358
12. Gupta P, Satyanarayan KVV, Shah DD (2018) Development and testing of message scheduling
middleware algorithm with SOA for message traffic control in IoT environment. Int J Intell
Eng Syst 11(5):301–313. https://doi.org/10.22266/IJIES2018.1031.28
13. Rao MV, Rama Krishna TV, Ganduri R, Roohi A (2018) An effective energy management
system for smart office cubicles using IoT. J Adv Res Dynam Control Syst 10(2 Special
Issue):338–347
14. Krishna MNV, Harsha NS, Kasula VDK, Swain G (2017) Optimization of energy aware path
routing protocol in wireless sensor networks. Int J Electri Comput Eng 7(3):1268–1277. https://
doi.org/10.11591/ijece.v7i3.pp1268-1277
15. Rajakumar R, Amudhavel J, Dhavachelvan P, Vengattaraman T (2017) GWO-LPWSN: grey
wolf optimization algorithm for node localization problem in wireless sensor networks. J
Comput Netw Commun. https://doi.org/10.1155/2017/7348141
16. Gupta P, Shah DD, Satyanarayana KVV (2016) An IoT framework for addressing parents
concerns about safety of school going children. Int J Electri Comput Eng 6(6):3052–3059.
https://doi.org/10.11591/ijece.v6i6.10448
17. Rao KR, Kumar TR, Venkatnaryana C (2016) Selection of anchor nodes in time of arrival for
localization in wireless sensor networks.https://doi.org/10.1007/978-81-322-2671-0_5
Estimation of State of Charge Using EKF
and SVR in Li-Ion Battery

P. Sannihith Reddy and R. Shanmughasundaram

Abstract Precise State of Charge (SoC) estimation is an essential requirement in

electric vehicle energy management and also helps in the optimal control system
design. But it is not a direct and easy task to be done. Batteries, which have a
distinct nonlinear behavior because of various internal and external conditions and
being complex electrochemical devices makes SoC estimation a very difficult task.
Therefore, a wide range of methods is currently in place for estimating the battery
SoC in real-time. This paper proposes a hybrid model which uses Extended Kalman
Filter (EKF) and Support Vector Regression (SVR) to estimate the SoC. Firstly, an
EKF algorithm that is based on a first-order Li-ion battery model is developed and
SoC of the model is estimated. The estimated SoC value is then used as an input
parameter to the SVR algorithm which acts as closed feedback to the model and
helps in accurate prediction of SoC.

Keywords State of charge · EKF · SVR and Li-ion battery

1 Introduction

As air pollution due to fossil-fueled vehicle emissions has been raising at an alarming
rate, the automotive sector has been looking forward to finding other efficient ways
to power the engines. Electric Vehicles (EVs) seems to be a great solution to address
this problem. According to the International Energy Agency (IEA) in 2018 electric
car sales has been increased by 68% globally which is almost double the year-on-
year average sales growth. This shows the impact of EVs in the transportation sector.
However, the expert agrees on the need for the improvement in battery technology
which is a major bottleneck for EVs.
Proper operation of EVs depends on fuel cells (battery pack) and their management
systems. So for ideal energy transformation, stable and reliable power batteries are
very important [1]. SoC is considered as an important parameter in the Battery

P. Sannihith Reddy (B) · R. Shanmughasundaram

Department of Electrical and Electronics Engineering, Amrita School of Engineering, Amrita
Vishwa Vidyapeetham, Coimbatore, India

© The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 615
R. Ibrahim et al. (eds.), International Conference on Artificial Intelligence
for Smart Community, Lecture Notes in Electrical Engineering 758,
https://doi.org/10.1007/978-981-16-2183-3_59
616 P. Sannihith Reddy and R. Shanmughasundaram

Management System (BMS). The BMS monitors SoC and provides information
about power level of the battery. The inefficient control of Li-ion battery charging
process can affect the battery life and might also damage the battery. The SoC can
be defines as the level of charge measured in comparison to its total capacity.
Out of the many available methods to estimate the SoC, one simple and common
way is to build linear relation between battery open circuit voltage (OCV) and it’s
SoC. Despite, its simplicity this method is not preferred for dynamic applications.
Because OCV takes long time (few hours) to settle down after a charge or discharge
cycle [2]. The other notable methods include Ampere-hour, coulomb-counting, modi-
fied coulomb-counting method etc., that are used for estimating the SoC. However,
these methods suffer from accumulation error with increasing time. Few advanced
techniques like neural network method require huge training data and many complex
calculations are involved in the process [3].
The model proposed in this paper makes use of EKF and SVR for estimating
the SoC. EKF is one of the finest estimation methods available for a non-linear (but
differentiable) state space model of a physical system. In EKF, SoC is estimated by
comparing model outputs with the measurements acquired from the actual system.
The physical system is modelled in a way that Gaussian white noises affect state and
output equations [4]. The support vector regression model developed for estimating
state of charge is made from the utilization of the right battery data to train and
selection of suitable kernel function. The data set used in this model was collected
from the CALCE battery group [5].

2 Battery Model

A mathematical model of the battery is required to build battery internal state

variables and external quantitative relation to electrical characteristics. Then using
external variables like terminal voltage, current and temperature internal state vari-
ables like SoC can be calculated. In this paper, the Thevenin model is chosen over
its ability to accurately simulate the Li-ion battery dynamic characteristics.
In Fig. 1, the following parameters are defined as:
R0 = Battery Internal Resistance.
C1 = Double Layer Capacitance.
R1 = Electrochemical Polarization Resistance.
C2 = Capacity of the Battery.
V1 = Voltage across RC branch.
V2 (SoC) = Open Circuit Voltage (Function of SOC).
Vt = Terminal Voltage.
I(t) = Battery current.
Applying Kirchhoff’s voltage law to the model

Vt = V2 − V1 − R0I (1)
Estimation of State of Charge Using EKF and SVR in Li-Ion Battery 617

Fig. 1 Thevenin battery model

Substituting V2

Vt = fn(SOC) − V1 − R0I (2)

From the above equations, the following system equations can be derived:
   
I
SoC
= kC2
(3)
Vt − R1IC1 Vt + 1
C1
I

V2 = k SoC + Vt + I R0 (4)

Equation (3) represents state equation and Eq. (4) represents observation equation
of the battery model. The discretization to get continuous equation and calculation
of SoC using EKF are explained next part.

3 EKF Implementation

EKF algorithm implementation can be divided into two steps [6]:

1. Prediction step: This step involves anticipating the system future state by using
present state.
2. Updating step: In this step, the value of present state estimate is updated on the
basis of system output value.
618 P. Sannihith Reddy and R. Shanmughasundaram

From Eqs. (3) and (4), state equation of the system can be expressed as:
 T
x = x1 x2 , x1 (t) = SoC, x2 = Vt (5)

Consider system input u(t) = I and output y(t) = Vt , then battery model will be
represented as below:

ẋ = f (x, u) + w (6)

ẏ = g(x, u) + v (7)

where w = process noise and v = measurement noise.

In Eq. (6) and (7)
 
u
f (x, u) = kC2
(8)
− R1IC1 x2 + 1
C1
u

g(x, u) = kx1 + x2 + u R0 (9)

To linearize the model, Taylor series expansion is needed on Eq. (6) and (7). After
linearization:

ẋ = At x + Bt u + w (10)

ẏ = Ct x + Dt u + v (11)

In Eqs. (10) and (11)

∂ f (x, u)
At = (12)
∂x
∂ f (x, u)
Bt = (13)
∂u
∂g(x, u)
Ct = (14)
∂x
∂g(x, u)
Dt = (15)
∂u
The discretized model is expressed as follows:

xk+1 = Ak xk + Bk u k (16)
Estimation of State of Charge Using EKF and SVR in Li-Ion Battery 619

yk = Ck xk + Dk u k (17)

EKF algorithm calculation steps are given as:

• Initially set value of k = 0
• Initial condition is set to x*k
• P is error covariance matrix
• Q is noise covariance matrix
Update state estimation:

x̂k = Axk−1 + Bu k (18)

Update error covariance:

Pk = A Pk−1 A T + Q (19)

Kalman gain matrix:

 −1
L k = Pk CkT Ck Pk CkT (20)

Update the estimate:

  
x̂k = x̂k + L k yk − Ck x̂k + Dk u k (21)

Update error covariance:

Pk = (I − L k Ck )Pk−1 (22)

Using the above formulae, EKF calculates the system state. The main step of
the EKF is to adjust the kalman gain according to error magnitude by comparing the
predicted value with the measured and this gain is used to compute the further values.
The bigger the error, greater the gain and greater the correction rate of estimated
value; lesser the error, smaller the gain, and lesser the correction rate of estimated
value. This provides a lucid reason as to why the EKF method is good in real-time
computation performance [7].

4 Simulation

The simulation of EKF algorithm has been carried out in Matlab Simulink based on
the battery model. The capacity of Li-ion battery used in this model is 2000mAh and
cut-off voltage is 4.2 V.
Simulation parameters are:
620 P. Sannihith Reddy and R. Shanmughasundaram

Fig. 2 SoC estimated value and actual value

Fig. 3 Error in estimation

 
1
• Initial State:
0
 
10
• Initial Covariance:
01
• Q = 0.1
• R = 100
To validate the result of extended kalman filter algorithm, the simulation results
were compared with the experimental data collected from CALCE battery group.
Figure 2 shows the estimated SoC values by EKF to the actual values. Error in
the estimation can be observed from Fig. 3.

5 SVR Implementation

Usually, Support Vector Machine (SVM) is used as a classification algorithm. SVM

classifies the data points by shifting data to a high-dimensional space. In a situation
where linear separation of data is not possible, the data is transformed to enable
drawing a hyperplane between various classes. Then, features of new data is used to
predict the class to which a new data point belongs to. SVR uses similar principles as
SVM, with only a few changes. In SVR, since the output is a real number predicting
Estimation of State of Charge Using EKF and SVR in Li-Ion Battery 621

the information is not a direct task. However, the main idea remains same: to minimize
error, identify the hyperplane, which maximizes the margin, keeping in mind that
part of the error is tolerated [8, 9].
The accuracy of the SVR model majorly depends on the proper data set, kernel
function and model parameters. The model parameters should be chosen wisely or
else there is a chance that the model may not behave as it is intended to. Also, select
the input features that has considerable effect on the required target parameter. This
helps to avoid unnecessary complexity and improves the accuracy [10].
Here in this paper SVR is implemented using python sklearn library. The data
required to the model is collected from CALCE battery data group. The kernel func-
tion used is Radial Bias Function (RBF). Because, the optimal predictive outcome
is better for a nonlinear kernel compared to a linear one. The model parameter C,
which is the penalty parameter used to control the trade-off between training error
and a testing error. This helps in generalizing the classifier to unconsidered data. C
is set to 0.1 in this model. The value of C is selected by comparing the accuracy with
the various C values.
Figure 4 shows the variation in model prediction accuracy for various C values.
It is found that at C = 0.1 model is getting saturated. So C is set to 0.1 to avoid
further complexity by increasing the C value. The other model parameter gamma is
set to 1, gamma determines how quickly boundary dissipation happens; larger values
decrease the effect of any individual support vector. Figure 5 shows how variation in
c and gamma values effect the classification model.
The idea is to develop a model that has the ability to estimate SoC accurately in
real-time. It is known fact that EKF is highly dependent on the battery model. But
creating a battery model which can address all real-time scenarios is not an easy
task. So this might affect the EKF model estimation accuracy in actual situation. To

Fig. 4 C versus prediction accuracy

622 P. Sannihith Reddy and R. Shanmughasundaram

Fig. 5 SVR parameters (c & gamma) effect [11]

address this and improve the prediction accuracy, a kind of closed feedback SVR
model is developed by giving SoC estimated by the EKF algorithm as one of the input
parameter along with voltage, current and discharge capacity to the SVR model. This
acts as indirect closed feedback loop and help the model to correct the errors in EKF
model estimation, which results in better accuracy.
Figure 6 shows the variation in SoC values predicted by SVR algorithm vs actual
SoC values. The SVR model prediction accuracy is found to be 99%, which is 4%
more than EKF. The increase in accuracy may seem to be small but the model is
capable of handling real-time scenarios in a better way when compared to EKF
model alone.

6 Conclusion

A hybrid model for estimating the SoC based on EKF and SVR algorithm has been
proposed in this paper. Initially required battery data set for the model has been
obtained from CALCE battery group, and then a mathematical model of the battery
with the same specification as of data set collected is developed. Then using the model
Estimation of State of Charge Using EKF and SVR in Li-Ion Battery 623

Fig. 6 SoC prediction using SVR

system state equation and observation equation are obtained. Now using EKF algo-
rithm equations iteratively system state has been calculated and updated constantly.
Finally, an accurate estimation value is achieved from EKF. Now this estimated SoC
value along with other battery parameters is used to develop the SVR model, which
resulted in accurate SoC prediction.

References

1. Ding-Xuan Y, Yan-xia G (2003) SOC Estimation of Lithium-Ion battery based on kalman filter
algorithm. Appl Mech Mater 347–350
2. Ciortea F, Rusu C, Nemes M, Gatea C (2017) Extended Kalman Filter for state-of-charge
estimation in electric vehicles battery packs. In: 2017 International conference on optimization
of electrical and electronic equipment and 2017 Intl aegean conference on electrical machines
and power electronics, Brasov, pp 611–616
3. Muthumanikandan S, Shanmughasundaram R (2016) Estimation of state of charge of lithium
ion battery using artificial neural networks. Int J Control Theory Appl 9(10):4331–4338
4. Taborelli C, Onori S (2014) State of charge estimation using extended Kalman filters for battery
management system. In: IEEE international electric vehicle conference (IEVC), Florence, pp
1–8
5. https://web.calce.umd.edu/batteries/data.html
6. Krishnakumar A, Shanmughasundaram R (2018) Simplified SOC estimation by EKF in Li-Ion
cell. JARDC 3:616–622
7. Kitagawa M, Shibuya H, Takehara J (1992) Development of battery state of charge indicator
for electric vehicles. In: Proceedings of the 11th international electric vehicle symposium
8. Saimurugan M, Ramachandran KI, Sugumaran V, Sakthivel NR (2011) Multi component fault
diagnosis of rotational mechanical system based on decision tree and support vector machine.
Expert Syst Appl 38:3819–3826
9. Sabareesh GR, Sugumaran V, Ramachandran KI (2006) Fault diagnosis of a taper roller
bearing through histogram features and proximal support vector machines. In: IEEE inter-
national conference on signal and image processing, December, B V Bhoomaraddi college of
Engineering and Technology, Hubli
624 P. Sannihith Reddy and R. Shanmughasundaram

10. Haq IN, Saputra RH, Edison F, Kurniadi D, Leksono E, Yuliarto B (2015) State of charge
(SoC) estimation of LiFePO4 battery module using support vector regression. In: Proceedings
of the joint international conference on electric vehicular technology and industrial, mechanical,
electrical and chemical engineering, Surakarta, pp 16–21
11. https://scikitlearn.org/stable/_images/sphx_glr_plot_rbf_parameters_001.png
Evolving Spiking Neural Network
for Prediction Problems in Healthcare
System

Tasbiha Ibad, Said Jadid Abdulkadir, and Norshakirah Binti Ab Aziz

Abstract This paper highlights the role of evolving spiking neural networks (an
enhanced version of SNN) for predicting medical diagnosis. This article aims to
focus on regression problems under a supervised learning strategy. In this paper, we
have trained and tested eSNN on benchmarking datasets. Among the three datasets,
one is the ICU Dataset which helps in predicting the recovery ration of patients who
stayed in ICU. Another dataset is Plasma_Retinol which predicts the risk of cancer-
related to certain carotenoids. Dataset pharynx is a part of a study conducted in the
USA to determine the success rate of two radiation types. The selected datasets are
those which were previously used for BioMedical Engineering related tasks. Later
the evaluation was conducted using Regression Metrics. From experiment results, it
is concluded that eSNN with standard parameters without optimization performed
well but there is still space available for improvement to achieve the highest possible
prediction scores.

Keywords eSNN · Regression · Optimization · Prediction

1 Introduction

Artificial neural networks can be appraised as one of the vigorous classifiers. Because
of their ability to handle noisy data skillfully [1–4]. ANN solves problems that
are established on approved algorithmic techniques [5–7]. SNNs (Spiking Neural
Networks), the 3rd generation of ANNs, plays a significant role in the processing of
biological information [8]. The initial two prototypes of SNN are conspicuous which
are recognized as the Threshold and Conductance model [9]. There is one derived
model of SNN which is known as Evolving Spiking Neural Network (eSNN), which
was first presented by [10]. eSNN is the one that evolves through learning. For
encoding techniques, eSNN utilizes Population encoding. As for the training, the

T. Ibad (B) · S. J. Abdulkadir · N. B. A. Aziz

Department of Computer and Information Sciences, Universiti Teknologi PETRONAS Seri
Iskandar, Perak, Malaysia
e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 625
R. Ibrahim et al. (eds.), International Conference on Artificial Intelligence
for Smart Community, Lecture Notes in Electrical Engineering 758,
https://doi.org/10.1007/978-981-16-2183-3_60
626 T. Ibad et al.

one-pass learning method has been chosen and it fits well. The target of the training
is to establish an output neuron repository with class labels [11]. The objective
of eSNN is to serve as a fast-learning process that is suitable for both offline and
online tasks. eSNN is widely applied in classification tasks, such as face recognition
[12] ecological problems [13], taste recognition [14] and has accomplished better
results than conventional methods. But for regression tasks, the contribution of eSNN
is not highlighted. Therefore, this study aims to provide a brief review of eSNN
performance in Regression Problems.

2 Literature Review

In the literature study of eSNN, there is a wide range of its application for classifi-
cation tasks. Such as the Integration of eSNN with the Harmony search algorithm
[15]. Firefly Algorithm was used to optimized eSNN for classification problems [9].
Multi-Objective Differential Evolution algorithm is combined with eSNN in [16],
to achieve the best accuracy results for classification. Differential evolution (DE)
was integrated with eSNN in [19] to optimize it for Classification Problems. A new
hybrid K-DESNN for clustering issues was established by integrating D.E and K-
means with evolving spiking neural networks (K-means eSNN) in [17]. eSNN was
optimized with the Harmony Search Algorithm (HSA) in [18]. To tackle the issue
related to hyperparameter sensitivity of eSNN, an optimizer was needed by eSNN
for parameter optimization. Therefore, this article proposed the combination of HSA
as an optimizer and ESNN as a classifier for the optimization of parameters. There
were several attempts to solve the issue of manual parameter tuning. To compare
the results of the optimized eSNN model there is a standard eSNN with parameter
values: C: 0.9, Sim: 0.1, and Mod: 0.75 introduced in [19]. As stated earlier, mostly
optimized eSNN frameworks are for classification tasks. Therefore, this study aims
to evaluate the standard eSNN for Regression Problems. For this paper, we have
trained eSNN on popular datasets from literature, for Regression Problems.
Evolving Spiking Neural Network for Prediction … 627

3 Methodology

3.1 Evolving Spiking Neural Network

The training begins with the initialization of 3 eSNN parameters in the interval [0,
1]: C (proportion factor) and Sim (similarity value), and Mod (modulation factor).
Each sample i that is from a similar class l is encoded into numerous pre-synaptic
input neurons named as j. The weight wj is calculated using Mod i and or der j .
Where, Mod i represents modulation factor from Thorpe neural model. The or der j
defines the rank of an emitted spike from neuron j. The output neuron similarity is
estimated according to the Euclidean distance between the neuron weight vectors. If
a particular neuron is believed to be very similar to others, it will combine with the
truly similar one. Thus, the merging procedure involves the calculation of average
weight vectors along with the threshold value. N represents the figure of samples
formerly used to upgrade output neuron k.

3.2 Datasets

Benchmarking datasets from the OpenMl.org dataset repository, are used to train and
test eSNN see Table 1. The chosen datasets are related to health sciences. And were
collected to be used for a specific study based on regression analysis. The ICU dataset
contains information of 200 survived patients after staying in an adult intensive care
unit (ICU). Plasma Retinol, datafile comprises of 315 observations on 14 variables.
628 T. Ibad et al.

Table 1 Datasets description

Datasets Number of instances Number of attributes
ICU 200 21
Plasma Retinol 315 14
Pharynx 195 12

The pharynx data is a small portion of a large medical trial conducted in the United
States for patients with cancer (squamous carcinoma).

3.3 Validation Metrics

The validation of eSNN conducted using the following famous regression metrics.
• R-Squared: It is a statistical degree of indicating how well the regression
predictions estimated the real data values. It is also called the Coefficient of
Determination
• Mean Absolute Error: It calculates the average amount of the errors in a set of
calculations, without focusing their direction. Concisely, it is the average absolute
difference between X and Y.
• Root Mean Square Error: It is the measure to tell how the residuals (i.e., predic-
tion errors) are spread out over the regression line. It is the standard deviation of
the residuals.

4 Result and Discussion

This section demonstrates the findings of the analysis based on prediction accuracy
and other measures. Additionally, the experiments are carried out 10 times for both
pieces of training as well as testing on the whole data sets. From a total number of
samples, 80% of the data is the training dataset and the rest is used for testing. The
experiment was conducted on ICU, Plasma-Retinol, and Pharynx Datasets from the
field of Medical and Health Sciences. The results have been evaluated based on their
prediction performance. According to the results presented in Table 2 the average
R-square value is negative for Plasma_Retinol. It is conceivable to get a negative
R-square for calculations that do not comprise of the constant period. In the case of

Table 2 Average prediction

Datasets R2 MAE RMSE
rate of eSNN
ICU 0.32 0.49 1.13
Pharynx 0.5 96 186
Plasma_Retinol −0.80 116 132
Evolving Spiking Neural Network for Prediction … 629

negative results, R-square cannot be interpreted as the square of a correlation [20]. For
ICU and Pharynx datasets R-square values are positive but very low. In general, the
higher the R-squared, the better the model fits the data. Statistically, Mean Absolute
Error (MAE) is one of the several metrics for reviewing and evaluating the excellence
of a machine learning model. The unit of MAE value is the same as the target variable.
The average MAE value for ICU 0.49, for Plasma_Retinol it is 96, and for Pharynx
data it is 116. This value indicates that the prediction error for ICU is very low. But for
the rest of the two datasets, error values are high which means eSNN did not predict
accurately. RMSE tells us how focused the data is across the line of finest fit. If the
RMSE will be 0, it implies that the points lie on the regression line (and therefore
there are no inaccuracies). The unit of RMSE value is also the same as the target
variable. But the higher the value the greater the prediction error, which indicates the
poor performance of the model. According to average values from Table 2, it can be
seen that the RMSE value for Plasma_Retinol is 132, and for Pharynx data it is 186
which describes that eSNN did not predict well for these two datasets. As the rate of
error is high. But for the ICU dataset, the error rate of eSNN prediction performance
is low i.e., 1.13. This shows good eSNN performance. The overall results of eSNN
prediction performance is a combination of good and bad fit. In this study, eSNN
was with standard parameters, so it could be hypothesized that eSNN can improve
its performance with its optimized parameters (Fig. 1).

5 Conclusion and Future Directions

In this study, we have tested standard eSNN for regression problems. The model is
trained on widely used datasets from OpenML.org. During the literature analysis, it
was found that eSNN was widely used and tested for classification problems. But
regression problems are neglected except for a few stdies. So, this study aims to test
eSNN for the regression task. Here we used eSNN with standard parameter values.
As stated, earlier eSNN is sensitive to its parameters. Therefore, the selection of
parameter values is a crucial task. Here eSNN with standard values of parameters is
trained and tested. According to the results eSNN performed not so well for regres-
sion datasets. Hence there is wide room for improvement for the same problems.
Future work could be the optimization of eSNN with several well-known Optimiza-
tion Algorithms. So that eSNN would achieve the best possible results in solving
regression problems.
630 T. Ibad et al.

Fig. 1 Graphical representation of prediction score at every 10 iterations

Acknowledgements The authors acknowledge the support of this research by the Yayasan Univer-
siti Teknologi PETRONAS Fundamental Research Grant (YUTP-FRG) under Grant 015LC0-119.

References

1. Morris A, Josifovski L, Bourlard H, Cooke M, Green P (2000) A neural network for classi-
fication with incomplete data: application to robust ASR. In: Proceedings ICSLP 2000 (No.
CONF)
2. Abdulkadir SJ, Suet-Peng Y, Foong OM (2013) Variants of particle swarm optimization in
enhancing artificial neural networks. Austral J Basic Appl Sci
3. Abdulkadir SJ, Yong S, Zakaria N (2016) Hybrid neural network model for metocean data
analysis. J Inform Math Sci 8(4):245–251
4. Abdulkadir SJ, Yong S-P, Marimuthu M, Lai F-W (2014) Hybridization of ensemble Kalman
filter and non-linear autoregressive neural network for financial forecasting. Mining Intell
Knowled Explor 72–81
5. Abdulkadir SJ, Yong SP (2014) Empirical analysis of parallel NARX recurrent network
for long-term chaotic financial forecasting. In: 2014 International conference on computer
and information sciences ICCOINS 2014—a conference of world engineering science and
technology congress ESTCON 2014—proceedings
6. Abdulkadir SJ, Yong SP (2015) Scaled UKF–NARX hybrid model for multi-step-ahead
forecasting of chaotic time series data. Soft Comput 19(12):3479–3496
7. Abdulkadir SJ, Yong S-P (2013) Unscented Kalman filter for noisy multivariate financial time-
series data. In: International workshop on multi-disciplinary trends in artificial intelligence
8. Ghosh-Dastidar S, Adeli H (2009) Third generation neural networks: spiking neural networks.
In: Advances in computational intelligence. pp 167–178
9. Roslan F, Hamed HN, Isa MA (2017) The enhancement of evolving spiking neural network
with firefly algorithm. J Telecommun Electron Comput Eng
Evolving Spiking Neural Network for Prediction … 631

10. The ECOS framework and the ECO learning method for evolving connectionist systems. J Adv
Computat Intell 2(6):195–202
11. Lobo JL, Laña I, Del Ser J, Bilbao MN, Kasabov N (2018) Evolving spiking neural networks
for online learning over drifting data streams. Neural Netw 1(108):1–9
12. Wysoski SG, Benuskova L, Kasabov N (2008) Fast and adaptive network of spiking neurons
for multi-view visual pattern recognition. Neurocomputing 71(13–15):2563–2575
13. Schliebs S, Platel MD, Worner S, Kasabov N (2009) Quantum-inspired feature, and parameter
optimization of evolving spiking neural networks with a case study from ecological modeling.
In: 2009 international joint conference on neural networks 2009 Jun 14 IEEE
14. Soltic S, Wysoski SG, Kasabov NK (2008) Evolving spiking neural networks for taste recogni-
tion. In: 2008 IEEE international joint conference on neural networks (IEEE World Congress
on Computational Intelligence) 2008 Jun 1. IEEE, pp 2091–2097
15. Yusuf ZM, Hamed HN, Yusuf LM, Isa MA (2017) Evolving spiking neural network (ESNN)
and harmony search algorithm (HSA) for parameter optimization. In: 2017 6th international
conference on electrical engineering and informatics (ICEEI) 2017 Nov 25. IEEE, pp 1–6
16. Saleh AY, Shamsuddin SM, Hamed HN (2015) Multi-objective differential evolution of
evolving spiking neural networks for classification problems. In: IFIP international conference
on artificial intelligence applications and innovations 2015 Sep 14. Springer, Cham
17. Saleh AY, Hameed HN, Najib M, Salleh M (2014) A novel hybrid algorithm of differential
evolution with evolving spiking neural network for pre-synaptic neurons optimization. Int J
Adv Soft Comput Appl 6(1):1–6
18. Saleh AY, Hamed HN, Shamsuddin SM, Ibrahim AO (2017) A new hybrid k-means evolving
spiking neural network model based on differential evolution. In International conference of
reliable information and communication technology 2017 Apr 23. pp 571–583
19. Saleh AY, Shamsuddin SM, Hamed HN (2016) A memetic harmony search algorithm based on
multi-objective differential evolution of evolving spiking neural networks. Int J Swarm Intel
Evol Comput 5(130):2
20. Kitchenham BA, Pickard LM, MacDonell SG, Shepperd MJ (2001) What accuracy statistics
measure. In: IEEE proceedings-software. Jun 1. vol 148(3). pp 81–5. Menard S. Applied logistic
regression analysis. Sage, 2002
Prototyping Pro-Active Wearable Gadget
for the Surveillance of Coal Miners
in Pakistan

Rahul Kumar, Ramani Kannan, Jayasankari Ganasan,

Ghulam E Mustafa Abro, Abdul Sattar, Vipin Kumar, and Nirbhay Mathur

Abstract After an advent of internet of things (IoT), the entire dynamics of this
world has been changed so rapidly. Being a student of science and technology, it is
very much essential for us to think about the potential alternatives for social issues.
In this regard, one may notice that there are thousands of people who die annually in
the coal mines of Pakistan and rescue team cannot help in any way. This is one of the
most emerging issues right now in other developing countries too where extraction
of minerals from in depth mines is going on. This paper identifies this issue and
suggests the prototyping of pro-active IoT-enabled wearable gadget for the smart
surveillance of coal miners within the mines of Pakistan. This wearable IoT device
not only monitor the hazardous gas, pulse rate of miner, humidity and temperature
conditions within and exact altitude where miner is working currently but it will also
provide a pro-active control to miners to generate any sort of query they face while
working within the dark and depth mines using error palette. This entire data will be
received at a specific Internet protocol (IP) using internet broker. This will also help
the rescue officers to decide their rescue flow accordingly.

Keywords Coal miners · Internet of things (IoT) · Internet protocol · Proactive

approach · Surveillance

R. Kumar · R. Kannan · N. Mathur

Department of Electrical and Electronics Engineering, Universiti Teknologi PETRONAS, 32610
Perak, Malaysia
J. Ganasan (B) · A. Sattar
Department of Computer and Information Sciences, Universiti Teknologi PETRONAS, 32610
Perak, Malaysia
e-mail: [email protected]
G. E. Mustafa Abro
Department of Electrical and Electronics Engineering, Hamdard University, Karachi 74600,
Pakistan
V. Kumar
Department of Civil and Environmental Engineering, Universiti Teknologi PETRONAS, Perak,
Malaysia

© The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 633
R. Ibrahim et al. (eds.), International Conference on Artificial Intelligence
for Smart Community, Lecture Notes in Electrical Engineering 758,
https://doi.org/10.1007/978-981-16-2183-3_61
634 R. Kumar et al.

1 Introduction

In the era, where internet of things (IoT) has evolved the dynamics of every part and
corner of the world, there are still some areas where the utilization of IoT-enabled
systems is still at the stage of infancy. One of such areas is the coal mining industry
where various Alliances of support workers and coal miners witnessed the loss of
their loved ones. In the list of recent incidents, one may see the tragedy of Sanjdi
coal mine of Baluchistan, Pakistan where more than 10 people died, and 8 people
were in critical situation because of the leakage of some poisonous gas.
So far, one may see the serious concerns of International Industrial Union for
coal miners and specifically the focus of Pakistan central mines labor federation
(PCMLF) to resolve such issues and they both are in favor of IoT enabled systems
i.e. wearable gadgets. The proposed IoT-enabled pro-active system in this paper
cannot be implemented within coal mine due to constraints of communication but if
managed on a wireless sensor nodes or network than thousands of the lives may be
secured easily. Even in the case of any uncertainty i.e. collapse of a mine, the rescue
team can perform their duties more efficiently without wasting unnecessary time.
The suggested system in this manuscript proposes an IoT-enabled system, that
comprises of several sensors as discussed in the section of methodology later and this
make this pro-active gadget as one of its kinds. There are most of the cases where these
rescue professionals start their operation at a very unknown place where there is very
less success probability. By utilizing the proposed system, these professionals can
save maximum coal miners because this system will not only be used for surveillance
purpose unlike in [1]. This gadget is pro-active in nature which means now the coal
miners can intimate their location to base station if they are fortunately alive and as
per altitude the rescue team initiate the operation.

2 Literature Review

This section brief some of the latest strategies that were previously used in the active
sensorization and detection of physical parameters. These same techniques will be
discussed in general to produce a statement that these sensors can be interfaced
with any sort of micro-controller to design a pro-active wearable gadget. This pro-
active wearable gadget will not only sense the physical parameters within the local
mines of Pakistan to update the base station [1] but this will also enable the miners
to generate a query too. The brief study of literature can be witnessed under this
section that can acquire the exact whereabouts of a miner and let us know about
his surroundings easily. Moreover, previously people used such techniques only for
monitoring and surveillance. One can see the ubiquitous and pervasive technology for
data acquisition [2]. This approach found to be suitable because of less computational
capability but less interaction between the system and user. Some of the researchers
have proposed the single dimension audio signal communication but this will increase
Prototyping Pro-Active Wearable Gadget … 635

the requirement of storage device. These techniques if utilized properly then the
rescue team will not dig at unknown place and instead of this, they will know exactly
the spot from where they can rescue the miner [3]. Researchers of today’s era are
still engaged in proposing the wearable and easy to carry devices, but all are limited
to surveillance [4].
After studying several research contributions, it has been noted that previously
proposed techniques i.e. An alert system design based on android applications [5]
were either limited to just monitoring or to generate the alert only. The one of the
major shortcomings of such strategies where they were entirely unable to share the
exact location of miner who is in trouble within the debris.
One may raise the questions regarding the accuracy of using such sensors; thus,
it is to share that these same sensors have been proposed in the design of various
sensitive incubators where exact data of babies are communicated using wireless
communication [6]. As a researcher, one may see various people have contributed
various alternatives like this to resolve many critical as well as sensitive issues [7].
One can see the use of such sensors in the design of sewer-bot that is designed to clear
the drain line and senses the temperature, humidity, and leakage of any hazardous
gas [8] (Table 1).
The leakage of hazardous gases results in various tragedies thus, there have been
many sensors and ways to identify the leakage such that average slope-based multipli-
cation [9]. In the catalogue of sensors that may detect the leakage of such hazardous
and explosive gases, one may find number of sensors that have made our work

Table 1 Summary of all previously proposed techniques

Sr. Proposed Technique Cons Refs.
1. Smart wearable jacket design Miners cannot generate any query in [1]
case of emergency or uncertainty
2. Pervasive computing strategies The algorithm is complex and less [2]
flexible
3. 2-D signal communication Needs more storage and there is more [3]
distortion
4. Smart wrist watch Mobile communication cannot done be [4]
within mines
5. Android application based systems Every miner cannot afford the smart [5]
mobile phones
6. Mean square and average slope The process time is huge and in our [6]
multiplicative algorithms scenario it misfits because of process
delays
7. Magnetic low frequency These low magnetic frequencies causes [10]
techniques cancer
8. MEMS systems Still there is no full range of sensors that [11]
are required for our case
9. Intelligent robotic applications The mapping constraints of mines as [12]
every environment is different
636 R. Kumar et al.

easier to detect the leakage of such explosive or poisonous gases i.e. methane [10].
The main constraint for utilizing such sensors is to establish the communication
network within the underground environment. This type of communication is already
established using low frequency magnetic fields and their immediate transmission
approach [11]. While studying the literature related to such utilization of sensors,
one may come across with MEMS based sensors and devices that are very robust
when it comes to use them within mines or underground environment but this require
the digital communication followed by the standard IEEE 802.15.14 protocol [12].
This IEEE standard is proposed to communicate at low rate within a wireless area
network. Almost in all sort of research contributions so far, researchers are somehow
more tilted towards monitoring of an unknown area and generating the alerts with the
detection of some physical parameters. Some researchers have also proposed some of
the robots for this task i.e. the famous wolverine version 2 which was introduced for
the first time in the year of 2001 by mine health and safety administration (MHSA),
United States of America [13]. This is not the first time that robots are introduced in
such critical applications, they have been used before such that ground hog, Gemini
scout and numbat proposed for the similar issues. One may see the Table 1, that
highlights the summary of such techniques and systems proposed for the similar
cases.
The mentioned Table 1 presents a brief comparison among the proposed tech-
niques. One may see most of the proposed strategies are based generating either the
alert or just for monitoring purpose. In addition to this, majority of the techniques
cannot communicate their global positioning satellite (GPS) location, temperature,
humidity, exact depth and lastly, the status of miner whether he is alive or not.

3 Methodology

In this section, the hardware implementation for the pro-active wearable gadget is
proposed. This manuscript presents some of the efficient sensors to be embedded
with the system to work as per objectives as highlighted in the below mentioned
Table 2.
In this entire pro-active wearable gadget, the user will first turn on the system
by pressing the single pole single through switch (SPST). After just turning it on,
the sensors will start gathering the results of all physical parameters and send it to
Arduino board.
Once Arduino Mega board will receive it, this will further be processed it to send
it to static internet protocol (IP) using ESP8266 Wi-Fi module. One can see the
entire data i.e. temperature, humidity, pulse rate, altitude, and exact location on a
single IP-oriented screen. The role of error palette is to provide a pro-active access
to miner. This is because if anything happens in front of him for which he needs to
update the base station then this smart pro-active wearable gadget may provide him
this facility to communicate easily using this palette. This error palette comprises
of 01 switch and 01 Led dully connected with the same main system. This switch
Prototyping Pro-Active Wearable Gadget … 637

Table 2 Proposed component list with their purpose

Sr. Proposed sensor Purpose
1. DHT 11s For detection of runtime temperature and
humidity values
2. MQ-2 sensor To detect the leakage of any hazardous gas
within mine
3. Pulse sensor To acquire the pulse of miner working within
the mines
4. GPS module Proposed to know the exact longitudinal and
latitudinal values of miner
5. BMP 180 sensor To acquire the altitude of miner at what depth
he is working
6. ESP8266 Wi-Fi module To link the sensor values with internet of
things
7. Arduino mega board To process the entire sensed values
8. Error palette To send the queries from miner’s side to base
station
9. Single pole single through switch SPST To turn on the system

is there for generating an emergency query within the mine and to indicate for help
from base station. One may see the entire block diagram as mentioned in Fig. 1. The
entire prototype as shown in the Fig. 1 can be initiated in acrylic sheet [1] or may
be mounted on any cloth fabric or one may opt for a safety helmet and embed this
system with in it. In this manuscript the proposed system is not mounted on anything
to show that it can be adjusted into any form. The entire schematic diagram is shown
in Fig. 2. Whereas the hardware prototype is illustrated in Fig. 3.

4 Results and Discussion

The device has been tested for the results through different ways. In order to check out
the change in altitude and location, this device had been moved to different height and
positions and the entire results were updated accordingly on serial monitor of Arduino
integrated development environment shown in Fig. 4. Moreover, for checking the
pulse values, the sensor has been given an extended wire so that the miner may place
it on of his finger. The entire data has been sent to IP and one can see the display as
illustrated in Fig. 5.
In above figures, the altitude is shown in feet one may convert it into meters too.
The minus (−) sign shows the depth at which the miner is working.
After putting the values on LatLong.net, we come across the exact location of the
webinar as shown in Fig. 6 given below:
Whereas the user can also utilize the error pallet and may generate the query.
After generating the query, the base station may send the help or rescue team to help
638 R. Kumar et al.

Fig. 1 Block diagram of proposed pro-active wearable gadget

Fig. 2 Schematic diagram of proposed pro-active wearable gadget

Prototyping Pro-Active Wearable Gadget … 639

Fig. 3 Proposed pro-active wearable gadget

Fig. 4 All physical parameters and their display on serial monitor

640 R. Kumar et al.

Fig. 5 All physical parameters and their display on IP

Fig. 6 Display of location on LatLong.net via values acquired from device

Prototyping Pro-Active Wearable Gadget … 641

Fig. 7 Display of an emergency query generated by error palette

the miner within the mines. The query will be also be generated on IP display unit
as shown below in Fig. 7.

5 Conclusion

This manuscript provides a non-conventional strategy in the form of a pro-active

device that aware almost every physical parameter within the mines that can affect
the life of coal miners. In addition to this, provide an easy way to monitor the environ-
ment remotely from base station. Unlike other strategies, this prototype of pro-active
wearable gadget provide a facility to miners as well to generate the emergency query
so that if sensors are failed in recording any sort of uncertainty even then the rescue
team and essential help may be obtained within short span of time.
The manuscript presents a system design that along with the detection of several
parameters such that miner’s pulse rate, availability of hazardous or poisonous gas,
updating exact altitude and GPS positions make this subject one of its kind. The only
thing due to which it is much more efficient and reliable is the pro-active facility
provided within through which now the miner can also communicate the base station
to ask for any sort of help in emergency or uncertain conditions. In future, the authors
are trying to work out on data visualization using a dashboard to monitor the live
moment for the coalminers.

Acknowledgement This is to acknowledge the support and facility provided for this project by
department of electrical and electronic engineering, Hamdard University, Karachi Sindh Pakistan.
642 R. Kumar et al.

References

1. Abro GEM, Shoaib AS, Safeeullah S, Ghulam A, Kun- dan K, Fiaz A (2018) Prototyping IOT
based smart wearable jacket design for securing the life of coal miners. In: 2018 International
conference on computing, electronics and communications engineering (iCCECE), IEEE, pp
134–137
2. Lukowicz P, Baker MG, Paradiso J (2010) Guest Editors’ introduction: hostile environments.
IEEE Pervasive Comput 9:13–15
3. Kwon GH, Smith-Jackson TL, Bostian CW (2011) Sociocognitive aspects of interoperability:
understanding communication task environments among different organizations. ACM Trans
Comput-Human Interac 18:1–21
4. Cernea D, Mora S, Perez A, Ebert A, Kerren A, Divitini M, Gil de La Iglesia D, Otero N (2012)
Tangible and wearable user interfaces for supporting collaboration among emergency workers.
In: Herskovic V, Hoppe HU, Jansen M, Ziegler J (eds) Collaboration and technology. Springer,
Berlin, Heidelberg, pp 192–199
5. Divitini M, Farshchian BA, Floch J, Mathisen BM, Mora S, Vilarinho T (2012) Smart jacket as
a collaborative tangible user interface in crisis management. In: Proceedings of the workshop
on ambient intelligence for crisis management
6. Chen W, Nguyen ST, Coops R, Oetomo SB, Feijs L (2009) Wireless transmission design for
health monitoring at neonatal intensive care units. In: Applied sciences in biomedical and
communication technologies, ISABEL 2009. 2nd international symposium on, IEEE, pp 1–6
7. Özmen A, Tekce F, Ebeoˇglu MA, Tas, altin C, Öztürk ZZ (2012) Finding the composition of
gas mixtures by a phthalocyanine-coated QCM sensor array and an artificial neural network.
Sensors Actuators B Chem 115(1):450–454
8. Markham A, Trigoni N (2012) Magneto-inductive networked rescue system (MINERS):
taking sensor networks underground. In: Proceedings of the 11th international con ference
on information processing in sensor networks, ACM, pp 317–328
9. Tanmoy M, Das PS, Mukherjee M (2011) Rescue and protection system for underground mine
workers based on ZigBee. Int J Adv Comput Eng Architect 1:101–06
10. Reddy AH, Kalyan B, Murthy CSN (2015) Mine rescue robot system–a review. Procedia Earth
Planet Sci 11:457–462
11. https://www.cdc.gov/niosh/mining/content/emergencymanagementan-dresponse/commtrack
ing/advcommtrackingtutorial3.html
12. https://pdfs.semanticscholar.org/d60c/0fd4782ab327ff61625f0863cba0e20bf67f.pdf
13. Jing L, Yingchun S, Wenhui C, Xufang B (2012) Application research on gas detection with
artificial olfactory system. Sensors Actuators 3:233–254
Design of a Self-Tuning PID Controller
for a Temperature Control System Using
Fuzzy Logic

Md. Tauhidul Islam, Ariful Islam, Rahul Kumar, Ghulam E Mustafa Abro,
Sourav Majumdar, and Vipin Kumar Oad

Abstract Temperature plays a significant role in industrial realm, especially in

metallurgy and chemical industries. As maintaining optimal conditions in manufac-
turing can make a huge difference to its profitability and efficiency. Without providing
a proper temperature control, the entire production lines can fall apart. Every product
needs to be manufactured in an individual and specific conditions. If the tempera-
ture is too high or too low then it can affect the output product, potentially it may
ruin completely and leads to wasting of raw materials. To maintain the stability of
temperature requiring in any situation a temperature control system is needed. Typi-
cally classical conventional PID controllers are being used but these controllers have
some lack of efficiency in controlling. These controllers exhibit higher overshoot,
long rise time and settling time. Self- tuning Fuzzy PID controllers are more efficient
than those classical PID control- lers. In this paper we proposed a design of Fuzzy
Logic based self-tuning PID controller to reduce overshoot and conquer long rise
time and settling time. This design provides more suitable and effective control to
the temperature control system in industries.

Md. Tauhidul Islam

Department of Electronics and Telecommunication Engineering, Chittagong University of
Engineering and Technology, Chittagong, Bangladesh
A. Islam
Department of Fundamental and Applied Sciences, Universiti Teknologi PETRONAS, 32610
Bandar Seri Iskandar, Perak Darul Ridzuan, Malaysia
R. Kumar (B)
Department of Electrical and Electronic Engineering, Universiti Teknologi PETRONAS, 32610
Bandar Seri Iskandar, Perak, Malaysia
G. E. Mustafa Abro
Department of Electrical and Electronic Engineering, Hamdard University, Karachi, Pakistan
S. Majumdar
Department of Electronics and Communication Engineering, Khulna University of Engineering
and Technology, Khulna, Bangladesh
V. K. Oad
Department of Civil and Environmental Engineering, UTP, Perak, Malaysia

© The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 643
R. Ibrahim et al. (eds.), International Conference on Artificial Intelligence
for Smart Community, Lecture Notes in Electrical Engineering 758,
https://doi.org/10.1007/978-981-16-2183-3_62
644 Md. Tauhidul Islam et al.

Keywords First self-tuning PID controller · Fuzzy interference system (FIS) ·

Temperature · MATLAB · Conventional PID

1 Introduction

Temperature control systems are being used in a wide variety of industries to manage
various operations or processing of manufacture. A temperature controller controls
temperature so that the process value matches the set point, but the response differs
due to the characteristics of the controlled object and the controlling method of the
temperature controller.
Various strategies can be exerted to the temperature control system, like PI, PD,
PID [1, 2], Artificial Intelligence (AI) [3], Fuzzy Logic [3], Genetic Algorithm (GA)
[1, 4], Fuzzy Self-tuning PID [3] etc. Recently, some strategies based on PID control-
ling and tuning method have been proposed in order to ameliorate system perfor-
mances. Gani et al. have proposed an ameliorated design using Genetic Algorithm
(GA) on optimal PID tuning of temperature system [1]. Salsbury proposed a feed
forward control method as a replacement of conventional PI feedback control [5].
Some has also done with self-tuning fuzzy PID controllers [3]. However those clas-
sical conventional PID controllers are less effective and some with modern strategies
like GA or AI, there is no doubt about their effectiveness but these strategies are too
hard to implement and costly too. Among these Self-tuning Fuzzy PID controller’s
algorithm is the easiest strategy to implement and it has better effectiveness [4,
6]. This paper predominantly focuses on designing a new self-tuning Fuzzy PID
controller to maintain required temperature.
The proposed design of self-tuning Fuzzy PID controller is mainly based on Fuzzy
Interference System (FIS). The procedure of fuzzy interference system includes three
phases: (i) Fuzzification: In this phase, crisp inputs are transposed into degrees of
membership. The degree of membership is discerned by plugging crisp inputs into
the membership function affiliated to the fuzzy set. (ii) Rule evaluation: In this phase,
each and every fuzzy rule is imposed with a strength value. The strength is discerned
by the degrees of memberships of the crisp inputs in the fuzzy sets of preceding
portion of the fuzzy rules. (iii) Defuzzification: In this phase the fuzzy outputs are
transposed into crisp outputs [7, 8] (Fig. 1).

2 Model of the System

In this paper we are modeling a temperature control system of an industry.

Considering a first order transfer function for the controlling system:

K
G(s) = e−τ s (1)
Ts + 1
Design of a Self-Tuning PID Controller for a Temperature … 645

Fig. 1 Block diagram of fuzzy interference system

where, K = 1 is defined as the static gain and τ = 2.8 as pure lag time, T = 3 as
time-constant of the controlled object [9, 10].

2.1 Conventional PID

PID controller is vastly used in every controlling system. The first letter of the name
of three controlling terms make up unitedly PID. It is actually a total combination of
proportional, integral and derivative terms of controlling system. Here PID controller
is used for controlling initially for the system. Equation of conventional PID can be
expressed as follows:

de(t)
u(t) = K P e(t) + K i e(t) + K d (2)
dt

where, K P is proportional gain, K i is integral gain, K d is derivative gain and e is error

present in the controller. These are the tuning parameters of the controller (Fig. 2)
[11, 12].

Fig. 2 Block diagram of conventional PID

646 Md. Tauhidul Islam et al.

Fig. 3 Block diagram of self-tuning -fuzzy PID

2.2 Self-Tuning Fuzzy PID

The block diagram of the proposed system is shown in Fig. 3. The system is designed
in MATLAB through Simulink. Firstly, a Fuzzy Interference System is designed by
defining membership functions, universe of discourse and fuzzy rules in MATLAB
through Fuzzy Logic Toolbox. Then this FIS system has been implemented in the
de- signed block system in Simulink.
In this paper Mamdani method of Fuzzy Interference system (FIS) is used to
process the fuzzy system. In MATLAB, there is a tool named Fuzzy logic toolbox
which is used to create membership functions, universe of discourse and fuzzy rules.
Here, in this interference system, the crisp inputs are system error (e) and rate of
system error changing (ec) and the crisp outputs are proportional gain (K P ) integral
gain (K i ) and derivative gain (K d ). The range for the system error (e) is taken as
[−1 1], rate of change of error (ec) is [−1 1] and range for the output is considered
as proportional gain (K P ) at [1 5], integral gain (K i ) at [0 1] and derivative gain
(K d ) at [0 1.5]. These ranges are obtained from the observations of conventional PID
controller since the conventional tuning gives as K P 1.285, K i as 0.27 and K d as 0.9
(Fig. 4) and (Table 1).

3 Result and Analysis

The proposed design has been simulated in MATLAB through Simulink. Figure 5
shows the step response of both the conventional PID controller and proposed PID
controller. The step response is determined at the step input temperature of 500 °C.
The comparison of performance and robustness between these two controller is
shown in Table 2.
The analysis of the simulation result and comparison table exposes that the
proposed self-tuning controller overshoots 0.63% with peak value of 503.16 whereas
the conventional PID controller overshoots 13.82% with peak value of 569.13. The
proposed method reduced the overshooting problem as well as providing better
dynamic performance and robustness. So, It is clear that the proposed self-tuning
Fuzzy PID controller for temperature control system provides better efficiency in
controlling.
Design of a Self-Tuning PID Controller for a Temperature … 647

Fig. 4 FIS system and input–output membership functions

Table 1 Fuzzy rules

ec
e L− M− S− Z S+ M+ L+
L− Kp L+ L+ M+ M+ S+ S+ Z
Ki S+ S+ Z Z Z L+ Z
Kd L− L− L− M− M− Z Z
M− Kp L+ L+ M+ M+ S+ Z S−
Ki S− S− S− S− Z Z M−
Kd L− L− M− M− S− Z Z−
S− Kp M+ M+ M+ S+ Z S− M−
Ki L− L− M− S− Z S+ M+
Kd M− M− S− S− Z S+ S+
Z Kp M S+ S+ Z S− M− M−
Ki L− M− M− S− Z S+ M+
Kd M− M− S− Z S+ S+ M+
S+ Kp S+ S+ Z S− S− M− M−
Ki L− M− S− S− Z S+ S+
Kd S− S− Z S+ S+ M+ M+
M+ Kp S+ Z S− M− M− M− L−
Ki M− S− S− S− Z S+ S+
Kd Z Z S+ M+ M+ L L+
L+ Kp Z S− S− M− M− L− L−
Ki S+ Z Z Z Z L+ L+
Kd Z S+ S+ M+ L+ L+ L+
Where (L−) = Negatively large, (M−) = Negatively Medium, (S−) = Negatively Small, Z=Zero,
(S+) =Positively Small, (M+) = Positively Medium, (L+) = Positively Large. And these are the
linguistic variables.
648 Md. Tauhidul Islam et al.

Fig. 5 Response graph after simulation

Table 2 Comparison table of

Parameter Conventional PID Fuzzy PID
performance and robustness
Rise time 2.43 s 2.99 s
Settling time 8.87 s 4.56 s
Overshoot 13.82% 0.63%
Peak 569.13 503.16
Peak-Time 5.21 s 7.66 s

4 Conclusion

In this paper a self-tuning fuzzy PID controller has been designed for controlling
temperature system for industrial management. Fuzzy interference system (FIS)
has also been described elaborately in this paper. After simulating the system in
MATLAB, the results expose that this self-tuning fuzzy PID has almost zero over-
shoot with a better dynamic performance than conventional PID. Analyzing the
proposed algorithm and results of this paper, it is concluded that fuzzy PID is more
effective and it can be replaced with conventional PID. This system can be amelio-
rated by farther modification in the designing algorithm like membership functions,
rules and gains which impacts the controlling performance by extenuating both the
overshoot and dynamic response.
Design of a Self-Tuning PID Controller for a Temperature … 649

Acknowledgement This is to acknowledge the support provided from the state art of facility from
Power Electronic Engineering Lab of Hamdard University, Karachi Sindh Pakistan.

Sponsorship
This work is not sponsored by any of the organization.

Conflict of Interest
There is no conflict of interest among the authors.

References

1. Gani MM, Islam MS, Ullah MA (2019) Optimal PID tuning for controlling the temperature of
electric furnace by genetic algorithm. SN Appl Sci 1:880
2. Anwar AZ, Jabeen B (2019) Prototyping non-holonomic hovercraft for path planning and
obstacle avoidance. Sir Syed Univers Res J Eng Technol 9(1)
3. Baogang H, Hao Y (2001) Review of fuzzy PID control techniques and some important issues.
Automatic Sinica 27(4)
4. Rehman NU, Kumar K (2018) Implementation of an autonomous path planning and obstacle
avoidance UGV using SLAM. In: 2018 International conference on engineering and emerging
technologies (ICEET), IEEE, pp 1–5
5. Salsbury TI (1998) A temperature controller for VAV air-handing units based on simplified
physical models. HVAC and R Res
6. Abro GEM, Jabeen B, Ajodhia KK, Rauf A, Noman A (2019) Designing smart Sewerbot for
the identification of sewer defects and blockages
7. Hassan MMM (2013) Current studies on intrusion detection system, genetic algorithm and
fuzzy logic. Int J Distrib Parallel Syst (IJDPS) 4(2)
8. Mugisha JC, Munyazikwiye B, Karimi HR (2015) Design of temperature control system using
conventional PID and intelligent fuzzy logic controller. In: Proceedings of 2015 international
conference on fuzzy theory and its applications (iFUZZY) The Evergreen Resort Hotel (Jiaosi),
Yilan, Taiwan, November 18–20
9. Abro GEM, Bazgha J, Manan A (2019) Stabilization of Non-ho- lonomic 03 dof hovercraft
using robust RST control design. Sukkur IBA J Emerg Technol 2(1): 45–50
10. Diwakar TK, Vivek Y, Raut KH (2014) PID tuning rules for first order plus time delay system.
Int J Innov Res Electri Electron Instrument Control Eng 2(1)
11. Wu X, Wang X, He G (2020) A fuzzy self-tuning temperature PID control algorithms for 3D
bio-printing temperature control system. In: Chinese control and decision conference (CCDC)
by IEEE. ISSN: 1948–9447
12. Jiang W, Jiang X (2012) Design of an intelligent temperature control system based on the fuzzy
self-tuning PID. In: International symposium on safety science and engineering in China, 2012
(ISSSE-2012), Procedia Engineering, vol 43. pp 307–311
Design of a Low-Cost High-Speed
and Large-Capacity Data Logger Using
MicroSD Card

Muhammad Sajjad , Mohd Zuki Yusoff , and Junaid Ahmad

Abstract Data logging is an important process for evaluating the functionality and
performance of many electronic systems such as internet of things (IoT). Most of the
mixed-signal systems digitize the signals from the sensors to generate input signals
and then perform some digital signal processing on these input signals to generate
output signals. Data logging of these input and output signals helps to evaluate the
performance of the systems. Many of these systems contain an FPGA for digital
signal processing. The FPGA can also be used to facilitate data logging. This paper
presents a data logging and data extraction system which is a high capacity, low cost
and ensures data integrity over gigabytes of the data. The system has been evaluated
by logging data in a system, having four inputs and one output signal each having
a sampling rate of 1 kHz, for five hours. Data integrity has been observed for each
sample of the data.

Keywords MicroSD · VerilogHDL · SPI interface · FPGA · State machine ·

MATLAB · Data logging

1 Introduction

Sensors are an integral part of electronic systems. Systems such as IoT can have a
lot of sensors. They sense some physical quantities and convert them into electronic
signals. After digitization, these signals are passed to signal processing modules
as input signals. These modules perform digital signal processing on the input
signals and generate some output signals. Most of the electronics systems contain

M. Sajjad (B) · M. Z. Yusoff

Universiti Teknologi PETRONAS, Perak, Malaysia
e-mail: [email protected]
M. Z. Yusoff
e-mail: [email protected]
J. Ahmad
Ghulam Ishaq Khan Institute of Engineering Sciences and Technology, Swabi, Pakistan

© The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 651
R. Ibrahim et al. (eds.), International Conference on Artificial Intelligence
for Smart Community, Lecture Notes in Electrical Engineering 758,
https://doi.org/10.1007/978-981-16-2183-3_63
652 M. Sajjad et al.

programmable logic devices like FPGA as a signal processing module. Data logging
of the input and the output signals is necessary for evaluating the performance of
the systems and verifying the functionality of the signal processing techniques in
real-time. In case of reliable data logging, the inputs can be passed to verified high-
level software routine to generate new signals. These signals can be compared to
the recorded output signals for verifying the functionality of the signal processing
modules.
One way for data logging is writing data to NOR-flash, NAND-flash, or EEPROM
which are normally present in the system electronics [1]. But these memories are
small in capacity and can hardly store a few megabytes of data [2, 3]. So these devices
are not suitable for data logging in high throughput systems. They also require erasing
before writing which reduces their data writing speed and add another task for the data
logging controller. Normally these devices are mounted on PCBs by the soldering
process and cannot be separated from the electronic systems. Because of this, the
electronics systems have to be moved in the lab to recover logged data for offline
analysis. Another way is using commercial off-the-shelf (COTS) data recorders [4].
These data recorders can store data in gigabytes and are interfaced with RS-232/RS-
422 ports [5]. They record data to file system of the attached USB flash or microSD
card. Main drawback of these COTS recorders is that they take electrical power and
some space and hence are not a suitable choice for critical missions.
The proposed data logging system consists of a microSD card connected to
four general-purpose digital pins (for implementing SPI) of a FPGA-based main
processing board. Raw data from the sensors and the processed data are written on
sectors of the microSD card. The digital design architecture ensures recording of
each sample, handling of contention arising from simultaneous read and write to the
buffers and matching of cyclic redundancy checks (CRC) during write operations to
the microSD card. Most of the existing data loggers use parity bit for error detection
which only applies to byte level errors in contrast to CRC which applies to the whole
data of the sectors. The data logging system has been tested for five hours duration
and no missing or corrupted data have been observed.
The paper is organized as follows. Section 2 discusses SD card Interfacing.
Section 3 presents implementation detail, followed by the Data and Information
Retrieval in Sect. 4. Results and Interpretation is presented in 5. Finally, the
conclusion is presented in Sect. 6.

2 SD Card Module

The Secure Digital card is widely used for data logging systems. It has flash memory
and a microcontroller embedded in it. The flash memory control operations such as
writing, reading, and erasing are performed inside the SD card. The data transfer
from the host system to an SD card is in the form of data blocks which are in
units of 512 bytes. Two modes can be used for data communication between a host
controller and an SD card i.e. SD mode and SPI mode. By default, the SD card is
Design of a Low-Cost High-Speed and Large-Capacity … 653

in SD mode. Both modes have their pros and cons. SPI communication protocol is
a bit simpler as compared to its default mode because the SD card can be attached
to any host controller using some GPIO pins. Therefore, this mode is most suited
for cost-effective embedded systems. Moreover, the SD card requires two different
clock frequencies for its initialization mode (100–400 kHz), and data transfer (up to
25 MHz) mode as specified in the SD card product manual whereas the commercially
available loggers record data from a RS422/RS485/RS232 which can hardly manage
data transfer up to 430600 Hz. Keeping in view all the challenges of the SD card, a
state machine has been designed to cater to the data archival on the microSD card.
After powering up the system, configure SPI clock frequency between 100–400 kHz
and put MOSI and CS pins to the high state. Send 74–80 clock pulses to the serial
clock pin of the SD card to make it ready for communication. Put CS to low and
send CMD0 to reset the card. The card will respond with 0 × 01 when it will be in
the idle state. After the idle state is achieved, send CMD8 with proper argument and
CRC before the start of the initialization process. The proposed data logging system
architecture comprises of a control unit, a microSD card module, two multiplexers,
and two block RAMs.

2.1 Control Unit

A Control unit is like a brain for the data logging system. A state machine is designed
to control operations. The flow chart for the state machine is shown in Fig. 1. At first,
the control unit will get the SD card ready flag from the SD card module to start its
operation. After the SD card is ready, the control unit will cater to the data flag from
sensors. After getting the data flag, the Control Unit will provide an 8-bit address
to a Buffer and store the sensors data on Block RAM. When block RAM is full,
the control unit will generate a RAM full flag to the SD card module to store the
respective RAM data on the sector of the SD card. The control unit also stops the
data writing process by providing an End data transmission signal to the SD card
module.

2.2 Block RAM

Single port Block Rams are implemented using IP cores. Each Block RAM has a size
of 500 bytes, having a 9-bit address, 8-bit data-in, 8-bit data-out, and its Read–write
operation is controlled by the control unit. The reason for its size of 500 bytes size is
that the sector length of the SD card is 512 bytes and the data packet after each 1 ms
is 25 bytes. The reason for using two block RAMs is that when the first RAM is full
and is being dumped on the SD card, the incoming data from sensors is stored on the
second RAM to avoid any data loss. Write address as it will take data from sensors,
and then write it on RAM. Whereas the address from the SD card is named as the
654 M. Sajjad et al.

Fig. 1 Architecture of the data logging system

Read address as it will read the stored data from RAM. The initialization process of
the card will start when CMD1 is received by the card. This initialization process
will end when the response is changed from 0 × 01 to 0 × 00. After initialization
data block length is set in units of 512 bytes by sending CMD16 to the card. CMD25
is then sent to the card to configure it for multiple block writing. After the command
is accepted by the card, the host can send data packets to the card.

3 Implementation

Data logging on a PC using serial cables is not possible in remote locations or for the
case of a drone flight. Acquiring real-time data in such a scenario is a tedious task.
The proposed data logging system is Implemented using four sensors connected
to the 4 channels of a 24-bit ADC, which is interfaced with an FPGA using SPI
protocol as shown in Fig. 2. The data from sensors is stored in an internal buffer of
FPGA. When the buffer is full, it generates a trigger to inform the control unit about
Design of a Low-Cost High-Speed and Large-Capacity … 655

Fig. 2 Experimental setup of the system

data availability. The reason for using a buffer is that its access time is very small
as compared to external memories. Data is also available to actuators to perform
necessary actions and is then stored on the SD card for offline analysis. This data
from sensors can also be made available for further processing.
S1(n), S2(n), S3(n), and S4(n) represents the sensor outputs of first, second, third
and fourth sensor respectively (shown in Fig. 2), so, S(n) is the output of algorithm
which is running on these sensors’ outputs. The algorithm in this proposed system is
just an integration of all sensors. Hence, every sample of sensors data is used in this
integrator algorithm ensuring data integrity. The equation for this integration-based
algorithm is (1).

S(n) = S(n − 1) + S1(n) + S2(n) + S3(n) + S4(n) (1)

As ADC is a 24 bit, so each sensor’s data is a 24-bit data. After applying a scale
factor on the sensors data, it becomes a 32-bit data i.e. 4Bytes. So, there are 16 Bytes
of sensors data, and 4 bytes of last iteration summation result hence, making it a data
of 20 Bytes. This data is then framed by a start byte, 2 bytes of checksum, and 2
bytes of stop bytes to make it a data packet of 25 Bytes. After 1 ms, a data trigger
is generated that store these 25 bytes on Block RAM. Block RAM is 500 Bytes,
so after 20 ms Ram is full and is then ready to dump its data on the sector of the
SD card. The last 12 locations of each sector of the SD card are filled with 0xFF.
Sensors data is written in LSB to MSB format. Clock frequency for initialization
and data transfer mode has been set to 384.615 kHz in its native mode so it takes
approximately 1.33 ms to write 512 bytes on a sector. After every 20 ms, a sector
is written on the SD card. So, in 5 h duration, it writes on 900,000 sectors of an SD
card. Each sector is of 512 Bytes, so, the total data is 450 MB.
656 M. Sajjad et al.

4 Data and Information Retrieval

For missions involving high data rates for a long duration, data logging in a single
file is not a feasible solution because the software used for post-processing or offline
analysis crashes due to limited memory. To avoid this situation, some computer
programs like hex editors can be used. With the help of MicroSD Card Reader, data
written on the sectors of the microSD card can be viewed and manipulated. A data
file can be created by specifying addresses for both the start sector and the end sector.
In this way, multiple files, having a manageable size for post-processing software,
can be generated through a hex editor. These files contain sector addresses and
bytes of data written by data logging setup. So, we cannot use them without further
processing. To make the files usable for analysis, these files can be further processed
by specially designed software routines on Integrated Development Environment
(IDE) like Microsoft Visual Studio or LabVIEW. Such routines can process these
files on the byte level. So, tasks like checksum matching, removing header or footer,
and concatenating relevant bytes to make a float, double, or integer type data values
can be performed easily.

5 Results and Interpretation

For example, Fig. 3 shows a comparison of offline and online (or real-time) perfor-
mance of the algorithm and hence the electronic system. Figure 3 also shows outputs

Fig. 3 Online and offline algorithm comparison

Design of a Low-Cost High-Speed and Large-Capacity … 657

of the same algorithm running in MATLAB (offline mode) and the FPGA (online or
real-time mode). This algorithm has been fed with the same data from the sensor and
ideally the graphs in Fig. 3 should track each other. But this type of mismatch does
not imply any data logging or algorithm error. This kind of deterministic error is due
to different type of truncation modes in double-precision floating-point computations
in MATLAB and FPGA. Since the algorithm consists of integration, which sums up
all the errors arising in the whole mission, so the error tends to increase with time.

6 Conclusion

This paper has presented a data logging setup for recording digitized input and
output signals in an electronic. The presented method ensures that every sample of
the signals is intact even for the missions with data size reaching to gigabytes. By
using the proposed data logging setup, large-capacity data can be recorded with full
integrity at a very low cost. This can help to speed up the evaluation and development
of IoT.

Acknowledgements This work is supported by the Graduate Assistantship Scheme of Universiti

Teknologi PETRONAS, and in part by the YUTP Fund under Grant 0153AA-H39.

References

1. Kim GJ, Baek SC, Lee HS, Lee HD, Joe MJ (2006) LGeDBMS: a small DBMS for embedded
system with flash memory. In: VLDB, CiteSeer, pp 1255–1258
2. Lee SW, Moon B (2007) Design of flash-based DBMS: an in-page logging approach. In:
Proceedings of the 2007 ACM SIGMOD international conference on management of data,
pp 55–66
3. Kim J, Kim JM, Noh SH, Min SL, Cho Y (2002) A space-efficient flash translation layer for
compactflash systems. IEEE Trans Consum Electron 48(2):366–375
4. De-Leon HL, Quiros RE (2005) Self-contained flight data recorder with wireless data retrieval.
United States Patent 6,898,492
5. Yamawaki Y (1995) Traffic accident data recorder and traffic accident reproduction system.
United States Patent 5,446,659
Machine Learning in Healthcare:
Current Trends and the Future

Usman Ahmad Usmani and Jafreezal Jaafar

Abstract Today, an abundance of electronically stored medical image data and DL

algorithms can be used to recognize and detect patterns and anomalies in this kind of
dataset. Computers and algorithms can interpret the imaging data as a very qualified
radiologist can see irregular skin, lesions, tumours and brain bleeds. Consequently,
the use of AI/ML tools/platforms to help radiologists is poised to grow exponen-
tially. This approach addresses a vital issue in the healthcare sector as well-trained
radiologists are challenging to come by worldwide. These professional professionals
are, in most cases, under tremendous pressure due to the influx of digital medical
data. We analyze and address the current state of A.I. applications in healthcare.
A.I. can be applied to various healthcare data forms (structured and unstructured).
Popular A.I. techniques include machine learning for structured data such as classic
support vectors and neural networks, modern in-depth learning unstructured data
natural language processing.

Keywords Healthcare trends · Machine learning tools · ML disease detection ·

Robotic surgery

1 Introduction

Identifying diseases rare or hard to diagnose also depends on detecting what are
known as’ edge cases. As this type of ML system relies on large datasets of raw
pictures (and different transformations) of these conditions, they are more precise
than humans for this type of detection [1]. They should improve automation produc-
tivity and intelligent decision making in primary/tertiary and public health care. This
would have the most significant effect on A.I. instruments since the quality of life
will increase for thousands of people worldwide [2].

U. A. Usmani (B) · J. Jaafar

Universiti Teknologi Petronas, Seri Iskandar, 32610 Perak, UTP, Malaysia
e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 659
R. Ibrahim et al. (eds.), International Conference on Artificial Intelligence
for Smart Community, Lecture Notes in Electrical Engineering 758,
https://doi.org/10.1007/978-981-16-2183-3_64
660 U. A. Usmani and J. Jaafar

The Microsoft Project Inner Eye, which uses ML methods for segmentation and
detecting tumours with 3D images, is an excellent test case. It can assist in the pre-
case preparation, navigation and effective tumour contours for radiation prepara-
tion. M.R.I.s and other sophisticated imaging systems are fitted with ML algorithms,
increasingly used for early cancer detection. ML tools often offer significant value by
increasing the surgeon’s display with information such as cancer’s location during
robotic procedures and other image-led procedures. Therefore, the use of AI/ML
software/networks to help radiologists should be exponentially expanded. Unfortu-
nately, these details are messy and unstructured too. In particular, patient data can not
be used for basic computational and analytical modelling instead of regular business
transaction data [3]. The need for the hour is a powerful and agile AI-enabled plat-
form, which can link and analyze a diverse mix of data types (i.e. blood pathology,
genomics, radiological images, medical history) [4]. These systems should also dig
extensively and uncover the secret trends through the analyses [5]. Furthermore,
doctors and other health practitioners should understand and understand their find-
ings to achieve their success with great confidence and clearness. These bills are
appropriate for the interpretable A.I. and distributed ML systems and can fulfil these
systems’ requirements within a short period. Surgery robots can provide human
surgeons with unique assistance, enhance their vision and manoeuvrability during
operations and make precise and minimally invasive incisions [6]. The optimal stitch
geometry and wound cause less pain. There are inspiring opportunities to apply
AI/ML to such robots in digital operations.

1.1 The Motivation for ML in Healthcare

The following article provides a detailed summary of this. Software-centred robot

communication with the aid of massively distributed computing. Insights and guid-
ance based on the history of Surgery (personal and machine) and findings (favourable
or not). A.I. has built space for augmented reality guidance and guidance in real-time.
Telemedicine and remote surgery prospects for reasonably easy operations. Daily
citizens in the United States have had a long and bitter debate about actual health
care costs and complexities. A.I. and similar methods powered by data are unique to
solve some of the root causes of the problems—long waiting, concerns of unreason-
able charges, a long, overly complicated appointment process, and lack of access to
the right physician [7]. The best thing about this kind of A.I. application is that data
security is not a big issue: a daunting and challenging healthcare system problem
[8]. The organizational challenge often contains no sensitive patient data concerning
illness, diagnosis or medicine. Still, it consists of data concerning funding, money,
marketing or human resources concerns as is the case for any modern organization
[9, 10]. These same problems plague conventional organizations for several decades,
and AI/ML techniques are now part of a solution. This is because massive databases
and smart search algorithms, which are a vital part of A.I. systems, are excellent
for specific patterns or optimization problems. Hospitals and public health agencies
Machine Learning in Healthcare: Current Trends and the Future 661

should also use advanced AI/ML technologies and strategies in their organizational
aspects every day. Such programmes’ core purpose should be to build AI-supported
networks to improve healthcare outcomes for the most extensive section of ordinary
citizens. The ultimate aim of current processes in conventional organizations is to
maximize benefit. Robust A.I. methods for managing medical operations need to
differentiate themselves from traditional systems by mixing empathy with income.
Increasingly, important names in the pharmaceutical industry are selecting A.I. and
ML technologies to resolve the challenging issue of useful drug testing [11].

1.2 Artificial Intelligence Techniques for Detection

of Diseases

This article offers several famous examples—like Sanofi, Genentech, Pfizer. These
case studies cover all sorts of clinical fields—metabolism, cancer therapies,
immunooncology. Beyond the traditional long-term method, A.I. techniques are
being used more and more to speed up the fundamental processes of early selection
and mechanism discovery. The Berg biotechnology company uses its A.I. method for
study in its research. For start-ups, the evaluation of multi-channel data, including
research papers, patents, clinical trials and patients records, may include state-of-
the-art inferential approaches, Markov chain models, enhanced learning and natural
language (N.L.P.s). The key objectives are to identify patterns and establish high-
dimensional representations stored in the cloud and used during drug discovery [12,
13]. Also, A.I. instruments can help shape why and under what conditions diseases
are more likely to arise. Physicians can direct and plan to intervene (personalized)
even before they begin to display symptoms. The aim is extraordinarily complicated
and challenging. It is to identify exact treatment options for an individual based on
an ongoing personal medical history, decisions about behaviours, genetic data and
pathological tests. The ongoing COVID-19 crisis has shown the value of performing
hundreds of simultaneous development and therapeutic trials of vaccines. Traditional
statistical modelling approaches are almost impossible to use, designed for small
experiments to collect data and distinguish patterns from all these various sources,
often delivering high uncertainty levels. For such a planetary scale, A.I. techniques
must bear the challenge of solving. A wide range of exciting and forward-looking
approaches to AI/ML and platforms have been investigated in the field of health
care. Sujets from radiology assistants to intelligent medical management, person-
alized medicine and digital public health monitoring have been explored [14–16].
Data protection issues and legal frameworks are still recognized as barriers to the
complete adoption of these systems. The type of data that third parties can access and
use legally (for example, the owners of the A.I. and ML software, physical devices
or platforms) may be quite challenging to decide. As technologists and IA/ML prac-
titioners, it is essential to have a bright future in which the AI-algorithms support
ordinary citizens’ trillions to improve their primary health and wellbeing.
662 U. A. Usmani and J. Jaafar

2 ML Potential Applications in Healthcare

With the rapid population growth, it seems impossible to record and evaluate large
patient information quantities. Machine learning gives us the ability to automatically
identify and process this data, making the healthcare system more complex and
resilient. Machine education in healthcare combines into a common thread two types
of fields: informatics and medical science [17–19]. Machine learning methodology
gives medical research development and analyses complex medical data for further
study.
Many researchers are working to develop new dimensions and features in this area.
Google recently developed an algorithm of learning machines to identify cancerous
tumour mammograms. Stanford also has a deep learning algorithm to assess cancer
of the skin [20, 21]. There are annual conferences, such as Machine Learning for
Healthcare, which seek to explore new automated medical technologies for better
service [22, 23]. Machine education aims to increase the machine’s prosperity, effi-
ciency and reliability. However, the machine learning method in a health system is the
doctor’s brain and experience what a patient requires a human touch and concentra-
tion [24, 25]. This can not be replaced by machine learning or any other technology.
An automated computer can provide better service. Here are the top 10 machine
learning frameworks for healthcare.

2.1 ML-Based Diagnosis for Heart Diseases

The heart is one of the primary organs of our body. We also have several heart
disorders like Coronary Artery Disease (CAD), CHD, etc. (Fig. 1).

Fig. 1 Diagnosis of heart disease

Machine Learning in Healthcare: Current Trends and the Future 663

Many researchers work on algorithms for the diagnosis of cardiovascular disease.

It’s a really hot problem of research in the world. One of the main advantages of
machine learning in health care is an integrated cardiovascular diagnostics method
[26–28]. Researchers use various tracked machine learning algorithms as a teaching
algorithm, such as the Support Vector Machine (SVM) or Naive Bayes. The Dataset
for U.C.I. Heart Disease can be used for training or assessment. For data analysis,
the WEKA data mining tool can be used. Alternatively, cardiac disease’s diagnostic
method can be established using the Artificial Neural Network (ANN) approach. One
of the most common and dangerous diseases is diabetes. This is also a significant
cause of other extreme disorders and deaths. This condition can affect our different
body parts, including the heart and kidney nerves.

2.2 ML-Based Prediction of Diabetes

A machine learning approach aims to predict diabetes and early-stage patients.

Random forest, K.N.N., Decision Tree and Naive Bayes can be used to construct
a prediction system as a classification algorithm. Naive Bayes beats other accurate
algorithms (Fig. 2).
Because of its excellent efficiency and less time for estimation, the diabetes dataset
can be downloaded from here. It comprises 768 data points, each with nine func-
tions [29, 30]. The liver is the second most crucial internal organ in our body. It
plays a significant role in metabolism. Many liver diseases such as cirrhosis, chronic
hepatitis, liver cancer can be targeted (Fig. 3).

Fig. 2 Prediction of diabetes

664 U. A. Usmani and J. Jaafar

Fig. 3 Liver disease prediction using ML techniques

Recent concepts have been used significantly in machine learning and data mining
to prevent liver disease. Disease with extensive medical records is challenging to
predict.
However, researchers use machine learning principles like grouping, clustering
and many more to solve this problem. Dataset (I.L.P.D.) can be used for hepatic
disease prediction. This dataset comprises ten variables or liver diseases. A classifier
can be used with the Vector Machine Support (SVM). You can use M.A.T.L.A.B. to
create a prediction system for liver disease.

2.3 ML-Based Robotic Surgery

One of the primary qualifications for healthcare training is robotic Surgery. This
application will soon be a promising field. This application can be divided into four
subcategories, including automatic suturing, surgical capacity assessment, surgical
materials, and surgical modelling. Automation of sutures minimizes machine life
and fatigue (Fig. 4).
The Raven Surgical Robot, for example. Researchers attempt to use a machine
learning approach to test the surgeon’s performance in the least invasive robot proce-
dure. University of California researchers, San Diego (UCSD), explore machine
learning technology to improve surgical robotics. Since robots can not function effi-
ciently in neurosurgery, the manual operational workflow requires time and can not
automatically provide input. The use of machine learning is a benchmark appli-
cation of machine learning in medicine in drug discovery [31]. The Microsoft
Project Hanover aims to implement precise medicine learning machine technolo-
gies. Many Organizations are currently using drug discovery machine learning tech-
nology. Benevolent AI, for example. It aims to use artificial intelligence (A.I.) to
Machine Learning in Healthcare: Current Trends and the Future 665

Fig. 4 Robotic surgery using ML

discover drugs. Machine learning has many advantages in this field, including accel-
erating the process and reducing the failure rate. Machine learning also optimizes
the development and expense of drug discovery (Fig. 5).

Fig. 5 Prediction and detection of cancer using ML

666 U. A. Usmani and J. Jaafar

2.4 Machine Learning in Radiology

An intelligent electronic health record system can boost computer training, including
classification documents and optical character recognition. This programme aims
to create a plan to sort patient requests via email or convert a manual recording
system into an automated system. This application seeks to develop a secure and
easily accessible framework. Electronic health records’ rapid growth has increased
patient healthcare information storage and advancement. For example, duplicating
data removes data errors. It could be used as a classifier or artificial neural network
(ANN) to create an electronic health recorder managed learning algorithm, like the
Support Vector Machine (SVM). Aidoc provides radiologists with machine learning
software to speed up detection. The purpose is to interpret the medical picture to
recognize the body’s abnormalities understandably. The supervised algorithm for
machine learning is used mainly in this field. Machine learning technology is used
for medical image segmentation. Segmentation is the process by which structures
are represented in an image. The graph cutting method is used primarily for image
segmentation. For analysis, standard Radiology Text File Language Processing is
used. The use of computer radiology can also improve patient care. The clinical trial
may be a series of studies requiring responses for biomedical or pharmaceutical effi-
ciency and safety. This research aims to reflect on the ongoing progress of treatment
(Fig. 6).
It takes a lot of time for this clinical trial. In this area, machine learning has an enor-
mous impact. An ML system can provide full service and real-time monitoring. The
advantage of using machine learning technology is that it can be remotely controlled
in clinical trials and studies. Machine learning also provides patients with a healthy
clinical environment. The clinical trial’s efficiency can be improved with supervised
computer education in healthcare.

Fig. 6 Machine learning in radiology

Machine Learning in Healthcare: Current Trends and the Future 667

Regarding machine learning performance, more knowledge almost always

provides better outcomes, and the health sector relies on multiple details. McKinsey
aims to deliver as much as $100B a year for big data and machine learning in pharma-
ceutical and medicine focused on improved decision making, optimized creativity,
research/clinical testing and emerging technology for doctors, consumers, insurers
and regulators.
What is the source of this fundamental knowledge? If we look at named data
points, we will see research and growth, physicians, nurses, patients, caregivers,
etc. The range of various sources (at present) forms part of the synchronization and
use of this information to enhance the health system and treatments. The biggest
challenge at the intersection of machine learning and healthcare today is finding
ways to collect and use various kinds of data to help analyze, prevent and manage
people. Burgundy ML applications in pharmaceuticals and medicine offer hope for a
future in which data synchronicity, research and ingenuity are a daily fact. At Emerj,
we research A.I.’s effect on the pharmaceutical industry as part of our A.I. Landscape
Opportunity service. Global drug companies have used A.I. Incentive Environments
to assess where A.I. works in their business, and what applications drive its value.
Disease disorder identifying and diagnosing is a crucial element of medical ML
science. In 2015, more than 800 cancer medicines and vaccines were investigated,
according to a survey conducted by Pharmaceutical Research and Manufacturers of
America (Fig. 7).
In an interview with Bloomberg Science, research scientist of the Knight Institute
Jeff Tyner pointed out that it was also a struggle to find ways to use all the findings
while it was interesting. “The concept of a biotechnologist working with computer
scientists and informatics is so interesting,” Tyner said. It is no wonder that great
players were the first to jump on the car, especially in areas of significant need, such
as detecting and treating cancer. I.B.M. Watson Health announced I.B.M. Watson

Fig. 7 Cancer detection using ML

668 U. A. Usmani and J. Jaafar

Genomics in October 2016, a joint project with Quest Diagnostics to support accurate
medicines using cognitive computing, genomic tumours, and a biopharmaceutical
company located in Berg. Current research programmes include intravenous tumour
dosing experiments and prostate cancer identification and management. In Google,
the Deep Mind Health disclosed several cooperation with the United Kingdom last
year, including Moorfields Eye Hospital in London, where macular degeneration
technology is developed in older eyes. Currently, supervised learning allows clini-
cians to choose from selective diagnoses or predict patient risk from symptoms and
genetic data, for example. I.B.M. Watson Oncology is a pioneering medical reform
and history institution to expand treatment possibilities. This personalized treat-
ment approach has significant consequences for patients to boost their wellbeing and
minimize overall healthcare costs. For instance, reductions in health care costs will
fluctuate and decrease as more people stick to medication and treatment plans.
Behavioural adaptation is also a significant driver in the prevention system,
addressed in December in an interview with Emerj by Catalia Health’s Cory Kidd.
And there are many start-ups with varying levels of success in cancer diagnosis,
prevention and care (for example). Somatix—a B2B2C software platform provider
with an ML database analysis that “recognizes mouth-tohand movement,” especially
when smoking stops. SkinVision is the first to build the software for ‘SkinVision’
which was DermCheck for ‘Just, Just C.E. Cert.’ The first time this app was devel-
oped, DermCheck sent images to dermatologists (people rather than machines) via
phone to return to a personalized treatment plan. In this area, the frontier appears to be
precision medicine with the detection of ‘multifactorial’ mechanisms and alternate
therapeutic routes. This study is unaccompanied, often mainly limited to detecting
data patterns without prediction (recently included in supervised learning). Machine
learning has many practical applications that can help develop and conduct clin-
ical trials. Advanced predictive analytics can identify clinical problems and thereby
contribute to lower, quicker, and cheaper testing, for example, for social media,
doctor visits and genetic information for target groups. According to McKinsey,
several more ML applications for improved clinical trial performance have been
developed, including determining sample sizes for better quality, addressing and
adjusting patient recruitment discrepancies, and using electronic health records to
remove information errors (e.g. double entry). DeepMind Wellness, Google, with
University College London Hospital, collaborated on machine-learning algorithms
to identify different levels of safe and cancerous radiation. DeepMind and U.C.L.H.
work to distinguish healthy and cancerous tissue variations on ML applications. More
knowledge is covered in our radiology machine learning applications. Classification
of documents (sorting patient requests by email, for example) with the aid of vector
machines and optical character recognition (transforming a digitized character of
cursive or another manuscript). The MATLAB ML and the handwritten character
recognition network of the Google Cloud Vision API provide only two examples of
creativity.
Intellectual and electronic health records are being developed over the next decade
by the M.I.T. Clinical Machine Learning Community. Diagnostics, clinical deci-
sions and personal care guidelines, including integrated ML/AI, where supported.
Machine Learning in Healthcare: Current Trends and the Future 669

The M.I.T. research site notes that ‘the need to build robust machine learning algo-
rithms will learn from smaller data on the labelling, interpretation and generalization
of natural languages in medical and institutional environments.’ ML and A.I. tech-
nology will monitor and predict all outbreak ovarian diseases through satellite-based
information, historical Web-based information and sociological information. Today’s
opioid crisis is a clear example of A.I. Vector support. Artificial neural networks were
used to predict malaria outbreaks based on temperature, mid-month precipitation,
total positive cases, and other data points. Epidemic severity predictions are urgently
needed, especially in third-world countries lacking in medical facilities, education,
and healthcare access. ProMED-mail is a software that monitors and detects emerging
diseases in real-time.
Significant issues still have to be addressed when integrating ML technology in
pharmaceuticals and medicinal products: data processing is one of the most pressing
problems today. Patient information is also sensitive and not easily accessible. It
seems safe to believe that most of the population is concerned with sharing infor-
mation rather than privacy concerns. It is noteworthy to note, in a survey carried out
in the United Kingdom by the Welcome Foundation in March 2016, that only 17%
of the respondents never decided to distribute their anonymized information to third
parties other than study.
The more open algorithms should comply with the strict drug development regu-
lations; the causes of machine findings must be seen and understood. It is essential in
the pharmaceutical industry to attract data science talent and develop a deep pipeline
of Skills. The breakdown of data silos and the promotion of a “data-centered perspec-
tive” across industries is critical in helping the industry to transform their thinking into
long-term progressive changes and to find value. Pharmaceutical firms have refused
to update or fund research programmes in history unless there is a considerable and
immediate monetary benefit. The relocation of already chaotic electronic records
through networks will be a first step towards creating tailor-made care solutions.

3 Natural Language Processing in Healthcare

Natural Language Processing provides exciting opportunities for swimming through

and using enormous data currently unused to improve results, reduce costs and
increase care efficiency. We found the N.L.P. drivers in healthcare in the first part
of this two-part series. For the increasing amount of data already produced in siLos
and daily systems, the A.I. sector appears necessary. The paper discusses the factors
which contribute to the growth and implementation in the health care sector of natural
language processing, the potential benefits of performance and the future of artifi-
cial intelligence and machine learning in the healthcare industry. It is a synthetic
intelligence division which mainly focuses on perception and human data or speech.
The technology involves a range of sub disciplines, including language questions,
natural language creation and language processing. Firstly, the technology employed
in health care is double: analyzing individuals’ speech and extracting significance.
670 U. A. Usmani and J. Jaafar

The two leading roles of technology in machine learning and N.L.P. in healthcare
are the most important. Over the years, N L.P. has advanced with validated R.O.I.
recognition, enabling clinics to transcribe notes with useful E.H.R. information using
natural language treatments in healthcare. Front-end speech recognition eliminates
medical records’ role rather than cure, while back-end technology works to identify
and correct transcription errors before moving them to human evidence. While the
market is almost overwhelmed by language recognition technology, many companies
disrupt the field with deep learning algorithms in mining applications that give more
significant opportunities. Clinical recording developments—Healthcare computer
training has resolved clinical recording and free doctors from a complicated E.H.R.
manual system. This was possible due to the dictation from speech to text and the
wording of the data entry capturing structured data while processing. As machine
learning progresses in healthcare, essential knowledge can be gathered from other
new sources, and P.H.M. and V.B.C. initiatives can be powered by analytics. Study of
data mining—The integration of data mining into healthcare systems allows organi-
zations to reduce decision-making subjectivity and provide valuable medical knowl-
edge. Data mining can be a cyclic technology until every HCO has started developing
a successful business plan for better patient care. Computer-assisted coding—NLP-
driven C.A.C. promises to prove coder accuracy im-. Computer-aided code collects
details and information on therapy to capture and optimize any code. The research
showed that current suppliers in the production of C.A.C. solutions on the market
must redirect their answers to a value-driven paradigm’s demands and ensure they
function as expected. Many health I.T. programmes have a regulatory burden because
metrics such as ejection fractions are not stored as separate values. Health systems
have to determine if an expulsion fraction (Fig. 8).
Health systems have to determine if an expulsion fraction is reported in a note and
save any value in a format which can be used for accurate reporting on the organi-
zational automated analytical reporting platform. New N.L.P. cases, such as clinical
study matching, N.L.P. and machine education, have an immediate impact. Some
companies are now trying to solve N.L.P. engine problems to suit the test. N.L.P.
simplifies and turns the test match process into a seamless one with innovations.
Prior approval—A report showed that the charge for doctors’ last approval needs is
growing. These standards improve overall activities and interfere with the delivery
of treatment. The question as to whether payers earned and approved reimbursement
was not rendered after a while as a result of language processing. I.B.M. Watson
and Anthem are now working on a payer N.L.P. module for quick prior authoriza-
tion assessment. Medical decision-making encourages physicians’ correct option
for natural languages production and engineering. Some healthcare areas require
extra monitoring, such as medical errors. Risk Adjustment and Hierarchical Cate-
gory Category Coding, a risk adaptation model, was initially developed to predict
patients’ potential costs. N.L.P. is also used for the control and evaluation of symp-
toms by physicians. H.C.C. coding is increasingly prevalent in value-based payment
models. H.C.C. uses ICD-10 coding to assign each patient risk ratings. The natural
language processing will allocate a risk factor to patients and estimate healthcare
costs.
Machine Learning in Healthcare: Current Trends and the Future 671

Fig. 8 N.L.P. in healthcare

Environmental Virtual Scribe—Language software must substitute human scribes

full of clinical documentation. N.L.P. and Artificial Health Intelligence will turn
clinical documentation into a wholly changed game in that scenario. The N.L.P.
can also enable physicists to teste patients in machine phenotyping and biomarker
development. Compared to the specialist’s expertise, N.L.P. recognizes phenotypes
through the patient’s present conditions. N.L.P. may also be used to determine speech
patterns and diagnoses for neurocognitive injuries such as Alzheimer’s and dementia.
An application for the N.L.P. will eventually identify a subset of ethnic or racial
groups in population surveillance to monitor and map health differences. There is no
672 U. A. Usmani and J. Jaafar

granularity for defining significant socio-cultural disparities and population tracking

in the current administrative databases.
However, N.L.P. offers a strong basis for further research and development. The
N.L.P. will use health systems to improve its managing treatment and solutions
approach. Machine-learning firms can boost healthcare physicians’ workflows and
patient experiences. Natural language processing solutions in the field of healthcare
are complete: enhancing patient experience with the doctor and E.H.R. Language
recognition systems may help close the distance between complicated medical issues
and the perception of the wellbeing of patients. N.L.P. can be an ideal way to deal
with E.H.R. Many clinicians take N.L.P. as an alternative form of typing and writing
notes, including growing patient health knowledge. Although patients have access
to their health information through an E.H.R. system, the specifics are difficult for
most people to understand. Just a fraction of the patients can choose their medical
history. This changes as machine learning is implemented in healthcare. The N.L.P.
tools help to measure and improve treatment outcomes while improving healthcare
quality. Value-based pay will allow medical facilities to assess doctors’ efficiency and
recognize care inequalities. N.L.P. algorithms can help HCOs detect potential treat-
ment errors. The N.L.P. algorithms can gather useful data from extensive databases to
classify sensitive patients and provide healthcare practitioners with the best methods
for coping with challenging patients. For predictive healthcare research, identifying
high-risk patients and the development of the diagnostic process can be carried out
using Pervasive Analytics and natural speech processing and predictional healthcare.
For emergency rooms, complete data is essential.
For example, delays in diagnosing Kawasaki disease cause significant complica-
tions if they are in some way, overlooked or violated. As scientific evidence shows,
an NLP-based algorithm has classified risk patients with Kawasaki’s disease with
93.6% sensitivities and 77.5% accuracy than the clinician’s manual test. Many French
researchers are working on another NLP-based patient algorithm to map, diagnose
and prevent HAI. The N.L.P. generated unstructured data to classify early indications
and clinicians as such. Further tests have also been carried out to analyze and predict
heart failure-treated patients’ likelihood automatically. Natural language processing
was used to assess a patient’s risk of hospital admission and death for the past 30
days and provide free text updates for the last 24 hours. At the end of the successful
experiment, the algorithm was better than expected, and the overall positive predic-
tive performance of the model reached 97.45%. The advantages of using N.L.P. will
undoubtedly be used to evaluate and forecast these conditions in patients by applying
a wide range of algorithms. The promotion of superior care decisions and patient
health outcomes includes natural language processing and other ML approaches.

4 Conclusion and Discussion

If your child has lost its wisdom teeth, individual doses of Vicodin are more likely
to be given. They may have Bactrim for an infection in the urinary tract (U.T.I.).
Machine Learning in Healthcare: Current Trends and the Future 673

Few patients, hopefully, will ever receive the same dose of any drug in the not
too distant future. Few people use the same medication if we know enough about
the patient’s biology and history. The promise of personalized medicine is a world
in which guidelines are customized to everyone’s wellbeing and disease remedies,
their medical history, genetic features, past ailments, nutrition, levels of stress and
much more. However, it might eventually expand to minor conditions (i.e. to give
somebody a smaller dose of bacteria for U.T.I., or an unparalleled version of bacteria
formulated for the Hooman Hakami from Medtronic, a Medtronic and I.B.M. diabetes
video) Medtronic would like to see insulin control pumps function independently,
monitor blood glucose levels and, if appropriate, administer insulin w While Western
Medicine has maintained its primary focus on the treatment and improvement of
diseases, proactive treatment and intervention are essential. This application is being
advanced by the first wave of IoT devices (particular FitBit). Machine training may be
used to track the worker’s performance or tension and seek beneficial improvements
in risk groups (not just to re-live or cure symptoms after reversals). Ethical concerns
regarding “increasing” human physical and (particularly) mental capacity are severe
and probably will become more critical as the technology of improvement in the next
15 years.
On the other hand, any human expert could easily defeat the narrower computer
vision application (if the model has been adequately trained). Federal “red tape”
or HIPAA also can make sports more “goliath” for the medical profession than
“david.” New social networks will catch teenagers up and beat Snapchat and Face-
book because of their virality, marketing and user interface. You only need a dozen
engineers at the right time, just like Instagram, but it is unlikely that a dozen engi-
neers have the industry, although they have earned ten million dollars. This labyrinth
can demand more money, contacts and know-how than any small Silicon Valley
company can collect. It seems like I.B.M. or Medtronic can have a healthy lead in
medical innovation because of these factors alone. In recent years, ML has contributed
to several areas, including the vision and processing of natural languages. Many
existing training data (e.g. images in vision or sentences in natural language) can
benefit from more complex models. We are now on the verge of a substantial change
in HE. Through applying ML to electronic health data, including genomic data, more
and more accessible, epidemiologists in health care can better understand the under-
lying risk for acquiring infectious diseases and transmission mechanisms and develop
specific strategies for reducing HAI. It is important to note that, while powerful, ML
can not recognize relationships that do not exist in the data.
Furthermore, ML does not substitute for regular statistical or random screening.
Instead, ML can be used to extend the existing HE toolbox. Interdisciplinary teams
would have the most significant influence on the future and work together to interpret
the data.
674 U. A. Usmani and J. Jaafar

References

1. Lipton RB, Scher AI, Steiner TJ, Bigal ME, Kolodner K, Liberman JN, Stewart WF (2003)
Patterns of health care utilization for migraine in England and the United States. Neurology
60(3):441–448
2. Islam MN, Inan TT, Rafi S, Akter SS, Sarker IH, Islam AKM (2020) A survey on the use of
A.I. and ML for fighting the COVID-19 pandemic. arXiv preprint arXiv:2008.07449
3. Lu C, Strout J, Gaudreau R, Wright B, Marcus FBDC, Buch V, Andriole K (2020) An overview
and case study of the clinical A.I. model development life cycle for healthcare systems. arXiv
preprint arXiv:2003.07678
4. Kaur J, Mann KS (2017) AI-based healthcare platform for real-time, predictive and prescriptive
analytics using reactive programming. J Phys: Conf Ser 933:012010
5. Tripathi G, Ahad MA, Paiva S (2020) S2HS-A blockchain-based approach for the smart
healthcare system. Healthcare 8:100391
6. Yoon JE, Suh CJ (2019) Research trend analysis by using text-mining techniques on the
convergence studies of A.I. and healthcare technologies. J Inf Technol Services 18(2):123–141
7. Gil-Lacruz M, Gracia-Pérez ML, Gil-Lacruz AI (2019) Learning by doing and training
satisfaction: an evaluation by health care professionals. Int J Environ Res Public Health
16(8):1397
8. Das S, Sanyal MK, Application of A.I. and soft computing in healthcare: a review and
speculation 8:21
9. Clifford GD (2020) The future A.I. in healthcare: a tsunami of false alarms or a product of
experts? arXiv preprint arXiv:2007.10502
10. Troncoso EL (2020) The greatest challenge to using AI/ML for primary health care: mindset
or datasets? Front Artif Intell 3:53
11. Lysaght T, Lim HY, Xafis V, Ngiam KY (2019) AI-assisted decision-making in healthcare.
Asian Bioethics Rev 11(3):299–314
12. Drysdale E, Dolatabadi E, Chivers C, Liu V, Saria S, Sendak M, Wiens J, Brudno M, Hoyt A,
Mazwi M (2019) Implementing A.I. in healthcare
13. Johnson SL (2019) A.I., machine learning, and ethics in health care. J Legal Med 39(4):427–441
14. Stanfill MH, Marc DT (2019) Health information management: implications of artificial
intelligence on healthcare data and information management. Yearb Med Inform 28(1):56
15. Pawar U, O’Shea D, Rea S, O’Reilly R (2020) Explainable A.I. in healthcare. In: 2020 interna-
tional conference on cyber situational awareness, data analytics and assessment (Cy- Bersa).
pp 1–2
16. Panesar A (2019) Machine learning and A.I. for healthcare. Springer
17. Farroha J (2019) Security analysis and recommendations for A.I./ML-enabled automated cyber
medical systems. Big Data: Learn Anal Appl 10989:109890
18. Adadi A, Berrada M (2020) Explainable A.I. for healthcare: from black box to interpretable
models
19. Reddy S, Allan S, Coghlan S, Cooper P (2020) A governance model for the application of A.I.
in health care. J Am Med Inf Assoc 27(3):491–497
20. Hernandez-Boussard T, Bozkurt S, Ioannidis J, Shah NH (2020) MINIMAL (MINimum Infor-
mation for Medical A.I. Reporting): Developing reporting standards for artificial intelligence
in health care. J Am Med Inf Assoc
21. Hunter P (2019) The advent of A.I. and deep learning in diagnostics and imaging: machine
learning systems have the potential to improve diagnostics in healthcare and imaging systems
in research. EMBO Rep 20(7):e48559
22. Greco L, Percannella G, Ritrovato P, Tortorella F, Vento M (2020) Trends in IoT based solutions
for health care: moving A.I. to the Edge. Pattern Recognit Lett
23. Henriksen A, Bechmann A (2020) Building truths in A.I.: making predictive algorithms doable
in healthcare. Inf Commun Soc 23(6):802–816
24. Halminen O, Tenhunen H, Heliste A, Seppälä T (2019) Factors affecting venture funding of
healthcare A.I. companies. ICIMTH 268–271
Machine Learning in Healthcare: Current Trends and the Future 675

25. Terry N (2019) Of regulating healthcare A.I. and robots. Available at SSRN 3321379
26. Srivastava SK, Singh SK, Suri JS (2020) State-of-the-art methods in healthcare text classifica-
tion system: A.I. paradigm. Front Biosci (Landmark edition) 25:646–672
27. Morley J, Machado C, Burr C, Cowls J, Taddeo M, Florida L (2019) The debate on the ethics
of A.I. in health care: a reconstruction and critical review. Available at SSRN 3486518
28. Tan Y, Jin B, Yue X, Chen Y, Vincentelli AS (2020) exploiting uncertainties from ensemble
learners to improve decision-making in healthcare A.I. arXiv preprint arXiv:2007.06063
29. Jiang F, Jiang Y, Zhi H, Dong Y, Li H, Ma S, Wang Y, Dong Q, Shen H, Wang Y (2017) Artificial
intelligence in healthcare: past, present and future. Stroke Vascular Neurology 2(4):230–243
30. Ellahham S, Ellahham N, Simsekler MCE (2020) Application of artificial intelligence in the
health care safety context: opportunities and challenges. Am J Med Qual 35(4):341–348
31. Fritchman K, Saminathan K, Dowsley R, Hughes T, Cock MD, Nascimento A, Tere- Desai
A (2018) Privacy-preserving scoring of tree ensembles: a novel framework for A.I. in health-
care. In: 2018 IEEE international conference on big data (Big Data), pp 2413–2422
Predictive Maintenance for a Turbofan
Engine Using Data Mining

Ismaila Mahmud, Idris Ismail, and Zuhairi Baharudin

Abstract Airplane safety remains one of the crucial areas that must have a robust
maintenance strategy due to its impact in the transportation of human beings and
goods. Predictive maintenance is a vital means of ensuring complex system such
as turbofan engines in airplane are being used safely and optimally. The advent of
information and communication technologies provide ways to collect useful data for
maintenance strategies and decision making. The acquired data are unstructured and
may contained incomplete information. Data mining transform the data to become
meaningful and useful for machine learning application. In this paper, data mining
techniques for predictive maintenance are presented, with machine learning algo-
rithms applied to predict the maintenance conditions of a turbofan engine and the
results are compared. The results show that support vector machine has a slightly
better accuracy than the other methods.

Keywords Predictive maintenance · Machine learning · Turbofan engine · Data

mining

1 Introduction

Maintenance is a process that ensure tools, machine (simple or complex) are kept in
a good working condition and ensuring that it is repaired when the need arises as a
result of breakdown within the possible time. Thus, maintenance has been around
since the existence of machineries. It has also evolved with the changes that accom-
pany new innovations in machine systems. Traditionally, maintenance tends to be
reactive-respond to failures in equipment or devices. And as a result, when failure
occurs, it becomes very catastrophe. Human life and properties have been lost due

I. Mahmud (B) · I. Ismail · Z. Baharudin

Department of Electrical/Electronic Engineering, Universiti Teknologi PETRONAS, Seri
Iskandar, Malaysia
e-mail: [email protected]
I. Mahmud
Department of Electrical Engineering, Ahmadu Bello University, Zaria, Nigeria

© The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 677
R. Ibrahim et al. (eds.), International Conference on Artificial Intelligence
for Smart Community, Lecture Notes in Electrical Engineering 758,
https://doi.org/10.1007/978-981-16-2183-3_65
678 I. Mahmud et al.

to insufficient maintenance and safety strategies in the aviation industry. A well

planned maintenance procedure could decrease the rate of failure, ensure system
availability, reduce the cost of maintenance and increase system reliability [1, 2].
The complexity of airline systems makes conventional maintenance methods not
to meet the requisite maintenance requirements for the aviation industry. With the
advent of smart systems, gives room to have some high level machine flexibility,
availability and high quality of services and products [3]. The advancement brings
about new form of challenges which is the complexity of systems and processes.
However, digitization in major sectors and couple with advancement in information
and communication technologies (ICT), brings new prospect for maintenance strate-
gies [4]. Maintenance measures could be predicted through the analysis of real time
data and recorded data. Predictive maintenance means having smart way of handling
maintenance from the usual reactive done before [5]. The lack of structured data pose
as a challenge in implementation of predictive maintenance strategies [6, 7] but the
use of Data mining allows unstructured data to become useful.
Prognostics have been used in the area of maintenance for quite sometimes. Prog-
nostic and health management (PHM) is a means where failure rates of machine
or equipment are estimated which enables monitoring its overall effect on the entire
system and provides ground for decision making of maintenance actions to be taken or
otherwise [8]. PHM could be achieved through the following methods; model-based
or data-driven methods. In the model-based approach, it has the ability to incorporate
physical knowledge of the system. However, there are usually difficulties in finding
the mathematical representation of complex systems. Unlike the model-based, the
data-driven method has the ability to utilize high dimensionality data to perform
prediction. Still, a dataset that has a run-to-failure in it, to be used for modelling are
quite not available. This is as result of most data does not have fault evolution due
to the nature of design of the system and managerial demands until failure occur.
Moreover, time and cost could hind obtaining actual system fault evolution. Propriety
issue is another factor that could make access to available data hard [9]. All these
constitute challenges and limits wide research in this area. However, an alternative
has been initiated through the use of run-to-failure data simulation and thus, rein-
forced research in the field. Data-driven methods build the prediction model through
the use of machine learning algorithms. Examples of classical learning algorithms
include approaches based on artificial neural networks (ANNs), k-means clustering,
logistic regression, Decision Trees (DTs) and support vector machines (SVMs) [10].
In addition, data-driven algorithms sometimes make more accurate predictions than
physics-based models because mathematical models cannot accommodate enough
complexity. This paper investigates data mining and machine learning application
on simulated run-to-failure data from NASA turbofan engine dataset. Four machine
learning algorithms such as decision tree, support vector machine, random forest and
gradient boosted tree were applied, and their accuracies compared. The organization
of the paper is as follows; Sect. 2 covers the related work. In Sect. 3, data miming
is being described with its features and the machine learning algorithms used in the
paper are highlighted. Case study and results are presented in Sect. 4 and Sect. 5
respectively. Finally, the conclusion is in Sect. 6.
Predictive Maintenance for a Turbofan Engine Using Data Mining 679

2 Related Work

There have been some research works reported in the literature on predictive mainte-
nance. Researchers have proposed novel methodologies to produced high accuracy
in predictive maintenance. In Behera et al. [11], the authors proposed a data-driven
prognostic technique for predictive maintenance based on tree ensemble algorithms.
They used NASA turbofan engine dataset and found out that gradient boosted trees
performed better than random forest. Alberto et al. [12] proposed a data-driven
predictive maintenance process that perform data acquisition and processing for a
real machine process. The data was obtained from computer numerical control (CNC)
turning centre with two tool posts. Machine learning models were used in assessing
the remaining useful life of the machining tool. The concept shows that predictive
maintenance approach could be used in areal machine process where preventive
maintenance remain the feasible maintenance strategy adopted. Accorsi et al. [13]
developed a data mining framework for condition-based maintenance with different
machine learning classifiers. The models were applied to real-world case study, the
fault prediction accuracy was quantified and compared. The results show that random
forest having slightly better accuracy than other methods. In Bukhsh et al. [14], the
authors propose three tree-based classification methods for predictive maintenance
of railway switching systems. The models were able to trigger the activity type and
predict the maintenance need of the railway switches. Lee et al. [15] applied predictive
maintenance to two machine tool system elements using artificial intelligence algo-
rithms the two machine tool systems consist of spindle and cutting tools. The compo-
nent’s conditions were obtained from flank wear and bearing’s remaining useful life
for cutting tool and spindle respectively. Artificial neural network and support vector
machine were used in classifying the tool’s condition. Okoh et al. [16] proposed a
statistical techniques-based data-driven prognostic approach for predictive mainte-
nance. The model was conducted on a synthetic data validated by industrial experts.
In this paper, data mining based predictive maintenance is used for turbofan engine
with machine learning algorithms to predict the engine conditions.

3 Data Mining

Data mining provide means where hidden trends, patterns, and associations on
datasets are discovered. It involves many methods such as statistics and machine
learning to solve wide range of the problem at hand [17]. In a typical classification
problem, neural network uses gradient descent training to solve it, while a decision
tree uses greedy search algorithms to solve it. The main purpose of data mining is
model creation [18]. A framework to execute Data Mining similar to standard one
like Cross Industry Standard Process for Data Mining (CRISP-DM) [19] with modi-
fication to suit our classification model case is developed. The steps in the framework
includes; Data Preparation, Feature Selection and Prediction model.
680 I. Mahmud et al.

3.1 Data Preparation

Row data are often not useful due to having presence of noise or too weak and in
some instance very large. Such data have to be process and reduced to a form, it could
be useful and easily be applied on predictive model. Some of the actions taking to
improve the signal quality include data quality assessment, operating condition oper-
ation discovery and data normalization. The tools used in data preparation include
filtering, amplification, data compression, data validation and de-noising.

3.2 Feature Selection

The data goes to another form of process to make it more suitable for use. At this stage,
features are extracted and selected. The aim is to extract indictors or features with
fault or patterns and also having minimal subset features from problem domain while
maintaining original feature structure with high accuracy. The purpose of feature is
to select relevant and informative features effectively.

3.3 Prediction Models

Decision tree, support vector machine, random forest and gradient boost tree are the
machine learning algorithms adopted for the prediction models in this paper. These
are supervised methods which build its models through establishing links between
input variables and target variables.
Decision Tree (DT). A decision tree is a machine learning algorithm that built
its classification model using a tree structure flowchart. Decision tree algorithm is
constructed with three set of nodes; root node, internal node and the leaf which
represents the class label. Among the algorithms used in developing decision tree
are C4.5 [20], classification and regression tree (CART) [21]. These algorithms are
implemented base on divide and conquer strategy where the training set is being
recursively partition.
Support Vector Machine (SVM). SVM is a statistical learning concept use to
perform classification through obtaining optimal hyperplane between two classes.
The hyperplane is the linear separator and in case of nonlinear classes, a kernel
function is used [22]. Kernel is a mathematical function that transformed nonlinear
data to be classified by a linear hyperplane and is given by
    2 
K x, x  = exp −γ x − x   (1)
Predictive Maintenance for a Turbofan Engine Using Data Mining 681

where ||x – x  || is the Euclidean distance between two data points x and x  . γ is
the control parameter that maintain error trade-off between variance and bias in the
model. Cost function is used in training SVM for error minimization

1  n
||w||2 + C ξk (2)
2 k=1

perform under these constraints

 
yk w T xk + b ≥ 1 − ξk and ξk ≥ 0 (3)

where w is weight vectors, C is the penalty parameter, ξ k is the slack parameters

used in handling non-separable data, T is matrix transpose, x k is the training data
points, b is a constant and yk is the class label of the training data.
Random Forest (RF). Random forest is one of the ensembles of decision tree. In
random forest, N subsets of training data are generated from the original dataset using
bootstrap sampling approach. Decision tree T i is constructed from the training of the
subset. Random forest is built from using N—decision trees [14]. The prediction of
random forest is the average predictions of the trees given as;

1 
N
f (x) = Ti (x) (4)
N i=1

Gradient Boost Tree (GBT). Gradient boosted tree is also an ensemble of deci-
sion tree that produces weak tree classifiers and it use boosting approach to build
the base model. In the boosting approach, multiple models are built through resam-
pling of the training instance that previously could not be estimated and by having
some loss function (sum of squared errors or cross entropy) minimized [14]. If the
gradient overall ensemble function is F m (x) and it generates a new model F m+1 (x)
for the base-learner h(x). For a given squared error loss function, the algorithm is
express as:


N

ρm = argmin p ζ (yi , Fm−1 (xi ) + ρh(xi )) (5)
i=1

Fm (x) = Fm−1 (x) + ρm h(x) (6)

where N is the total number of training instances, ζ is the loss function and ρ is the
gradient step-size.
682 I. Mahmud et al.

4 Case Study

The prognostic and health management dataset used in this paper is from NASA
prognostic research center. It is a simulation of turbofan engine under several contin-
uous conditions. The data was captured using C-MAPSS (Commercial Modular Aero
Propulsion System Simulation) software. The turbofan engine architectural structure
with its primary components is shown in Fig. 1.
Air goes into the engine through the fan and low pressure compressor. Thereafter,
the air passes through high pressure compressor that mixed it with fuel and ignited
by a combustor. This results in lifting up discharge air velocity by the combust fuel
in the high pressure compressor (HPC) that controls the low pressure turbine (LPT)
and high pressure turbine (HPT). A 9000 lbs thrust class engine with various sensors
installed on it was used for the data readings collection. Several operating conditions
were used in gathering the engine information such as temperature variation (within
the range of − 51–39 °C) and the engine’s altitudes from sea level to 40,000 ft [11].
Table 1 provides some details of the four subsets of testing and training datasets in
the C-MAPSS dataset.
There is k-by-26 matrix-order stored in each subset where 26 is the number of
input attributes and k is the row count of the dataset. The three operational settings
have their combine data represented in every tuple of the dataset. The dataset is put
such that it start from engine number, followed by the cycle number, columns 3–5 are
the values of the operational settings and the remaining are the 21 different sensor
reading values attached on the aircraft engine [23]. It is considered that in each tuple,

Fig. 1 Architecture of turbofan engine

Table 1 Details of C-MAPPS dataset

Datasets FD001 FD002 FD003 FD004
Training trajectory 100 260 100 249
Test trajectory 100 259 100 248
Operating conditions 1 6 1 6
Fault condition 1 1 2 2
Predictive Maintenance for a Turbofan Engine Using Data Mining 683

the training and testing subset hold the total life-cycle time of the engine and the
last entry is the condition where the engine is on the brink of failure or declared
unhealthy [24, 25]. The major purpose of RUL estimation is to identify the condition
when the engine might need inspection or replacement. The RUL attribute targets
could be define as normal, warning and critical depending on the particular value in
the RUL attribute.

5 Results and Discussion

In this section, the FD001 dataset is used in building and testing the predictive
maintenance model for the turbofan engine. The four machine learning algorithms;
decision tree, support vector machine, random forest and gradient boosted tree are
trained to monitor and predict the conditions of turbofan engine.
Confusion matrix is used in evaluating the performance of the models’ prediction
ability. In the matrix, each row and column show the true and predicted values of the
model and can reveal which classifier is confused. A true positives are positive cases
correctly classified as positive class and true negatives are negative cases correctly
classified as negative class. Conversely, a false positives are negative cases incorrectly
classified as positive class and false negatives are positive cases incorrectly classified
as negative class. True positive rate is the proportion of correctly classified positives
while false negative rate is the proportion of misclassified positive cases (error rate
per class).
The input dataset for the models is extracted from FD001 dataset of NASA prog-
nostic and health management dataset. It is 20626 × 26 dataset with 26 input features.
21 features represent sensor signals employ in monitoring the conditions of the
turbofan engine. The first 5 attributes consisting of; engine numbers (100), number
of cycles (time) and three operating conditions. In the pre-processing of the dataset,
14 features are selected from the 21 sensor signals as being meaningful for the predic-
tion models. Principal component analysis (PCA) is applied to the selected features
to reduce the dimensionality of the features [15]. PCA is a mathematical process of
mapping multi-dimensional dataset to lower dimension and maintaining its original
information. The mapped dataset is used as the input for the models. The predic-
tion conditions are labelled as; critical, warning and normal which corresponds to
the remaining useful life of the dataset. Table 2 shows the accuracy obtained by
each model which measures the overall performance of the models. The results are

Table 2 Accuracy of the

Model Accuracy (%)
predictive maintenance by the
models Decision tree (DT) 83.8
Support vector machine (SVM) 88.6
Random forest (RF) 86.0
Gradient boosted tree (GBT) 83.6
684 I. Mahmud et al.

Fig. 2 Confusion matrix for decision tree

obtained using hold-out validation, by using 75% for training and the remaining 25%
for testing.
SVM had the better result with 88.6% accuracy while GBT had the least result of
83.6%. Figures 2, 3, 4 and 5 show the confusion matrices of the models.
From the Fig. 3, it shows that SVM correctly predict critical condition 87%,
warning condition 96% and normal condition 38%. The SVM maintained better
prediction among the models. Though, RF perform relative similar with the SVM in
all the indicators as shown in Fig. 4

6 Conclusion

In this paper, machine learning algorithms through data mining were employed in the
development of predictive maintenance of turbofan engine. Four models; decision
tree, random forest, gradient boosted tree and support vector machine were trained.
Three engine’s conditions (normal, warning and critical) represent the remaining
useful life of system. The results show that SVM had better overall prediction accu-
racy than the other three decision tree-based models. In the future, we intend to
enhance the performance of the decision tree-based model by having it hybridize
with a metaheuristic algorithm.
Predictive Maintenance for a Turbofan Engine Using Data Mining 685

Fig. 3 Confusion matrix for support vector machine

Fig. 4 Confusion matrix for random forest

686 I. Mahmud et al.

Fig. 5 Confusion matrix for gradient boosted tree

Acknowledgements The authors gratefully acknowledge and appreciate Universiti Teknologi

PETRONAS support toward this work through its Yayasan Universiti Teknologi PETRONAS grant
(Cost centre: 015LCO - 006).

References

1. Li Z, Wang K, He Y (2016) Industry 4.0-potentials for predictive maintenance. Adv Econ, Bus
Manag Res
2. Li Z, Wang Y, Wang K-S (2017) Intelligent predictive maintenance for fault diagnosis and
prognosis in machine centers: industry 4.0 scenario. Adv Manuf 5(4):377–387
3. Fitouri C, Fnaiech N, Varnier C, Fnaiech F, Zerhouni NJIP (2016) A decison-making approach
for job shop scheduling with job depending degradation and predictive maintenance. FAC-
PapersOnLine. 49(12):1490–1495
4. El Maraghy W, El Maraghy H, Tomiyama T, Monostori L (2012) Complexity in engineering
design and manufacturing. CIRP Ann 61(2):793–814
5. Khoshafian S, Rostetter C (2015) Digital prescriptive maintenance. In: Internet of things,
process of everything, BPM everywhere, pp 1–20
6. Colledani M et al (2014) Design and management of manufacturing systems for production
quality. CIRP Ann 63(2):773–796
7. Roy R, Stark R, Tracht K, Takata S, Mori M (2016) Continuous maintenance and the future–
foundations and technological challenges. CIRP Ann 65(2):667–688
8. Lee J, Wu F, Zhao W, Ghaffari M, Liao L, Siegel D (2014) Prognostics and health management
design for rotary machinery systems—reviews, methodology and applications. Mech Syst
Signal Process 42(1–2):314–334
Predictive Maintenance for a Turbofan Engine Using Data Mining 687

9. Saxena A, Goebel K, Simon D, Eklund N (2008) Damage propagation modeling for aircraft
engine run-to-failure simulation. In: 2008 international conference on prognostics and health
management. IEEE, pp 1–9
10. Kotsiantis SB, Zaharakis I, Pintelas P (2007) Supervised machine learning: a review of
classification techniques. Emerg Artif Intell Appl Comput Eng 160:3–24
11. Behera S, Choubey A, Kanani CS, Patel YS, Misra R, Sillitti A (2019) Ensemble trees learning
based improved predictive maintenance using IIoT for turbofan engines. In: Proceedings of the
34th ACM/SIGAPP symposium on applied computing, pp 842–850
12. Jimenez-Cortadi A, Irigoien I, Boto F, Sierra B, Rodriguez G (2020) Predictive maintenance
on the machining process and machine tool. Appl Sci 10(1):224
13. Accorsi R, Manzini R, Pascarella P, Patella M, Sassi S (2017) Data mining and machine learning
for condition-based maintenance. Procedia Manuf 11:1153–1161
14. Bukhsh ZA, Saeed A, Stipanovic I, Doree AG (2019) Predictive maintenance using tree-
based classification techniques: a case of railway switches. Transp Res Part C: Emerg Technol
101:35–54
15. Lee WJ, Wu H, Yun H, Kim H, Jun MB, Sutherland JW (2019) Predictive maintenance of
machine tool systems using artificial intelligence techniques applied to machine condition
data. Procedia CIRP 80:506–511
16. Okoh C, Roy R, Mehnen J (2017) Predictive maintenance modelling for through-life
engineering services. Procedia CIRP 59:196–201
17. Kovacs K, Ansari F, Geisert C, Uhlmann E, Glawar R, Sihn W (2019) A process model for
enhancing digital assistance in knowledge-based maintenance. In: Machine learning for cyber
physical systems. Springer, pp 87–96
18. Rokach L, Maimon OZ (2008) Data mining with decision trees: theory and applications. World
scientific
19. Shearer C (2000) The CRISP-DM model: the new blueprint for data mining. J Data Warehousing
5(4):13–22
20. Quinlan J (1993) C4. 5: programs for machine learning Morgan Kaufmann San Francisco. CA,
USA
21. Breiman L, Friedman J, Stone CJ, Olshen RA (1984) Classification and regression trees. CRC
press
22. Tong MT (2019) Using machine learning to predict core sizes of high-efficiency turbofan
engines. J Eng Gas Turbines Power 141(11)
23. Babu GS, Zhao P, Li X-L (2016) Deep convolutional neural network-based regression approach
for estimation of remaining useful life. In: International conference on database systems for
advanced applications. Springer, pp 214–228
24. Peel L (2008) Data driven prognostics using a Kalman filter ensemble of neural network models.
In: 2008 international conference on prognostics and health management. IEEE, pp 1–6
25. Wang T, Yu J, Siegel D, Lee J (2008) A similarity-based prognostics approach for remaining
useful life estimation of engineered systems. In: 2008 international conference on prognostics
and health management. IEEE, pp 1–6
Cyber Security Breaches
and the Long-Run Effect on Firms’
Market Value: A Conceptual Framework

Syed Emad Azhar Ali and Fong-Woon Lai

Abstract Incidents of cyber security breaches are on the rise and pose a poten-
tial threat to businesses. Especially for publicly traded firms as they could create
a long-lasting influence on their financial performance and, thus, the market value
(MV) of a firm. Following the footsteps of the efficient market hypothesis, previous
studies have examined only the short-run impact on MV ensuing to security breach
announcements. Therefore, this study aims to conceptualize the impact of security
breaches on MV as manifested by long-run abnormal returns those firms. The study
is expected to provide a meaningful insight to investors and managers on the long-run
interconnection between cyber security breaches and firms’ market value.

Keywords Efficient market hypotheses · Firms’ market value · Long-run

abnormal returns · Security breaches

1 Introduction

The Global Risk Reports by World Economic Forum (WEF) [1, 2] perceived a higher
risk for cyber-attacks with very high likelihood and impact following the enormous
risks associated with natural disasters. The increasing dependency on IoT and IR4.0
is making the business operations more vulnerable for incidents of cyber-attacks
such as spyware, malware, Denial of Service (DOS) attacks, phishing, ransomware,
and other vulnerabilities [3–5].
Measuring and analyzing the economic impact of security breaches has been
a point of interest from the scholars’ perspective. The effect of a security breach
on the trust and confidence of business stakeholders, including the investors in the
stock market [3, 6, 7]. A firms’ MV is one indicator to assess the change in the
confidence of investors. Therefore, various studies have studied the effect of security
breaches on the MV of a firm using theory for Efficient Markets [6, 8–14]. In most

S. E. A. Ali (B) · F.-W. Lai

Department of Management and Humanities, Universiti Teknologi PETRONAS, Darul Ridzuan,
32610 Perak, Malaysia
e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 689
R. Ibrahim et al. (eds.), International Conference on Artificial Intelligence
for Smart Community, Lecture Notes in Electrical Engineering 758,
https://doi.org/10.1007/978-981-16-2183-3_66
690 S. E. A. Ali and F.-W. Lai

of these studies, the negative impact was witnessed on MV manifested by negative

abnormal returns. These studies were underpinned by the ‘Semi-Strong Efficient
Market Hypothesis (EMH) [15, 16], according to which stock prices adjust quickly
to all new information. Accordingly, the short-run examination of a security breach
event has been performed in these studies by analyzing the stock price behavior (i.e.,
from few days before an event to few days after an event), exhibiting an immediate
analysis of MV. While the short-run analysis helps get the market’s quick reaction to
an event announcement, it is imperative to explore the long-run analysis to estimate
a more realistic economic impact of a breach announcement. Studies addressing this
problem are unclear whether an event of security breach will affect the long-run MV
of firms.

1.1 Long-Run Impact on MV

According to the Global Cost of Data Breach Report [17], the time to identify and
contain a data breach has been continuously increasing from 257 days in 2017 to 280
days in 2020, with estimated response cost ranging around $1 million for each firm.
The announcements such as security breaches indicate that the firms are reluctant to
disclose complete details of a breach on their first announcement. In recent times,
we have witnessed events of security breaches where the details concerning a breach
are disclosed months after the first breach announcement. For instance, in July 2019,
Equifax was penalized with US$ 700 million by the Federal Trade Commission
(FTC) and Consumer Financial Protection Bureau, after an enormous data breach
in 2017. From the investors’ perspective, they will probably have a close watch on
subsequent disclosures and announcements by the firm concerning a security breach
announcement that can influence their investment decision making. Hence, the impact
of a security event, especially concerning MV, cannot be judged by analyzing the
announcement effect only using the traditional event study methodology, as examined
by [7, 18]. Especially in cases where the announcement and economic impact vary.
Considering the above arguments, probably, the abnormal effect of a security
breach announcement by a listed firm can have a meaningful impression on future
cash flows, the required rate of return, financial distress, and credit rating. All these
concerns will be manifested on MV in the long run as well. Overall, it can be hypothe-
sized that the security breach announcement might lead to abnormality in the long-run
indicators of MV in the stock market, such as long-run abnormal returns. To the best
of our knowledge, the long-run impact on MV after a security breach has not been
addressed in the literature. An unfavorable abnormality might manifest that influence
in its long-run stock returns.
Cyber Security Breaches and the Long-Run Effect … 691

2 Literature Review and Hypothesis Development

An influential research group has discussed various issues relating to the manage-
ment of security risks, such as security investments [19, 20], institutional stimulus
on innovation, and security [21]. Another line of research focuses on the market
implications of disclosures linked to the security of information [67, 68] and secu-
rity breaches [11, 22–28] by having an underpinning of ‘efficient market theory’ and
the methodology of an event study. Among all event studies on security, the current
paper is compatible with those scholarly works which have inspected the impact of
the security breach on the stock market. Among all event studies on security, the
current paper is compatible with those scholarly works which have inspected the
impact of the security breach on the stock market [18, 29–31]. It can be elucidated
from most of the studies that the security breach will have a significant unfavor-
able influence on the overall confidence of stock investors as signaled by negative
abnormal returns within 1 or 2 days of the breach announcement. All these system-
atic research works have based only on the short-run impacts of breaches while only
anecdotally exploring the long-run implications of security breaches.

2.1 Studies on Long-Run Impact:

The review of existing works reveals that the implications of events for the firm’s
long run prospects, especially the firms’ MV have become a common concern for
researchers. Long-run effects have been mostly explored with the BHAR model,
1–3 years after the incident. Within the information management avenue, the BHAR
model was used to address the long-run value of the company about the impact
of Capability Maturity Model (CMM) [32], Enterprise Resource Planning (ERP),
Supply Chain Management (SCM), and Customer Relationship Management (CRM)
[33, 34]. As far as we know, the impact of a security breach on the company’s long-
run MV has not yet been researched in the avenue of security. This research is the first
to conceptualize the effects of the security breach on the firms’ MV in the long-run.
Hypothesis Development and Conceptual Framework.
Long-run Abnormal Returns in the period before the breach announcement. Despite
breached firms’ intention to delay the announcement, there could be signs by which
security breaches can somewhat be anticipated. For example, the firm’s webpage
might become inaccessible for their customers and other stakeholders, a very sluggish
internet browsing, customers might face access issues on the firm’s website, pointless
popping up of messages at the time of customer’s log-in, abrupt changes to system
passwords or accounts, browser warning of errors. Moreover, the concerned users
might face infection alarms through antivirus toolkits installed in their systems. Thus,
these signals of system malfunction can negatively influence the trust and confidence
of customers. Likewise, the firms’ MV in the stock market might be influenced, and
692 S. E. A. Ali and F.-W. Lai

Fig. 1 A conceptual framework for analyzing the impact of security breaches on long-run MV

they might have allocated a likelihood of a security breach for the concerned firm.
It is also probable that the financial impact of a security breach might have already
incorporated by investors in their stock returns even before the actual security breach
announcement. Hence, it can be hypothesized that:
H1: In the period before the announcement, abnormal returns of security breached firms will
be negative.

Long-run Abnormal Returns in the period after breach announcement. To fully

evaluate the financial influence of security breaches, it is also vital to analyze stock
price behavior in the post-announce era. Studies in recent times have also reported
statistically unusual long-run stock price reaction ensuing to the announcement of
an event. Examples include repurchase offers, spin-offs, initiations and omissions on
dividends,open market repurchases, stock splits, seasoned equity offerings, share-
holder fights, and initial public bidding (see [35] all such studies). Likewise, security
breaches can also impact firm financials even after the announcement as consumers
and suppliers respond to security breaches. Since these impacts must be part of
quantifying the overall economic impact of security breaches, stock returns should
be measured over extended timelines ensuing an announcement. Accordingly, there
are equal possibilities of positive and negative returns in the post-announcement
period. Our second hypothesis is:
H1b: In the period after the announcement, abnormal returns of security breached firms will
be negative.

The conceptual framework of this study is exhibited in Fig. 1. The announcement

of a security breach is an exogenous construct. At the same time, the endogenous
construct is demonstrated by long-run MV.

3 Methodology

The methods and estimation techniques we have proposed in measuring the long-run
abnormal returns are different from those usually used in event studies when the short-
run impact of events is analyzed on the stock market. As the methodology of event
study frequently provides skewed approximations of both the ultimate economic
influence and the test statistics [36, 37]. Our proposed methodology is based on robust
Cyber Security Breaches and the Long-Run Effect … 693

and more accurate methods, which were newly established and used frequently in
studies [34, 38–42].

3.1 Assessing the Long-Run Abnormal Returns

The literature argues as to how to measure long-run abnormal returns [35, 36].
The first problem is the relevant variables to be controlled for calculating abnormal
long-run returns. Previous studies on long-run stock valuation have been primarily
controlled for a company’s systemic risk (or beta). According to literature, the size,
market-to-book ratio, and previous performance are imperative predictors of stock
returns [43–45]. The present consensus, therefore, appears to be that abnormal returns
have to be determined after controlling size, market-to-book ratio, and previous
performance [38].
Buy-and-hold abnormal returns (BHARS) using one-to-one match samples:
Our proposed approach is one-to-one matching in which each sample firm is
compared to the appropriate control firm having similar size, market-to-book ratio,
and prior performance. Then it is recommended to create two individual samples of
one-to-one:
i. Choose a firm nearest in size to the sample firm (size-matched).
ii. Choose a firm nearest to the sample firm in terms of its previous performance
from the sample firm (performance-matched).
The abnormal return is the gap between the sample firm’s BHARs and its
matched control firm, ensuring consistency with the null hypothesis. Accordingly,
both parametric and non-parametric tests can be used for statistical inferences. This
paper provides an estimate of abnormal buy-and-hold-abnormal-returns (BHARs)
by daily return statistics. The raw returns and the return of the sample company
are compounded first through the frame for BHARs measurements. The abnormal
return is the disparity between the sample firm’s cumulative returns and its benchmark
(Fig. 2). BHARs are determined as:

Fig. 2 One-to-one matched sampling methodology [36]

694 S. E. A. Ali and F.-W. Lai

Fig. 3 Mapping of calendar time for assessment of MV


T 
T
B H A Ri = (1 + Rit ) − (1 + Rbt ) (1)
t=1 t

1 
N
B H A Rit = B H A Ri (2)
N i=1

3.2 Mapping of Security Events for Assessment of MV

To collect findings over time, we conceptualize to map the calendar to date of the
occurrence for each entity in our sample. The day of the announcement is day 0, the
next day for trade is day one, and the day before the announcement is day-1, etc. It
is advised to forecast abnormal returns for at least 18 months, beginning six months
prior announcements through two periods of six months each after the security breach
notification. It consists of 250 trading days each year.
Furthermore, it contemplated deducting from both sides a 2-week duration (10
trading days) while measuring abnormal changes in stock returns. Moreover, security
breaches may, in the period before the disclosure, have already impacted stock prices,
and thus, the MV. Therefore, before the actual announcement, abnormal changes
may occur (Fig. 3). In the following three periods of six months each, we degree
buyand-hold abnormal returns and standard deviation from daily stock returns.
• Six months pre-announcement: trading days -135 to -11,
• Six months post-announcement: trading days 11–135,
• Twelve months post-announcement: trading days 136–260.

4 Conclusion

Findings in future studies based on the proposed framework might emphasize the
significance of security and determine that it cannot be disregarded. Events of security
Cyber Security Breaches and the Long-Run Effect … 695

breach do not occur unwittingly, and over time, users have a lower level of tolerance
for such events. Also, the news or media attention exposes the presence of flaws in
the firm’s security, rendering the firm more likely to be the victim of such assaults in
the future. Taking SONY as an example, this reveals that SONY’s loss of MV of 293
million dollars a year after the incident and total loss of 937 million two years after
the event would result from a security breach event. It implies that investors would
reevaluate the worth of the company in the stock exchange as they face tangible and
intangible risks from security breach incidents.

Acknowledgements The authors would like to acknowledge the efforts of the Department of
Management & Humanities-Universiti Teknologi PETRONAS (UTP) and Faculty of Business &
Economics-Universitas Islam Indonesia (UII), under grant cost center: 015MEO-114.

References

1. Collins A (2019) The global risks report 2019. World economic forum, Geneva, January 15
2019. Available: https://www.weforum.org/reports/the-global-risks-report-2019
2. Collins A (2018) The global risks report 2018. World Economic Forum, Geneva, January 17
2018. Available: https://www.weforum.org/reports/the-global-risks-report-2018
3. Smith KT, Jones A, Johnson L, Smith LM (2019) Examination of cybercrime and its effects
on corporate stock value. J Inf Commun Ethics Soc 17(1):42–60
4. Jansen J, Junger M, Montoya L, Hartel P (2013) Offenders in a digitized society. In: Cybercrime
and the Police. Eleven International Publishing, Hague, ND, pp 45–59
5. Lukonga I (2018) Fintech, inclusive growth and cyber risks: focus on the MENAP and CCA
regions. In: IMF working papers, vol 18. Available: https://doi.org/10.5089/978148437490
0.001
6. Cavusoglu H, Mishra B, Raghunathan S (2004) The effect of internet security breach announce-
ments on market value: capital market reactions for breached firms and internet security
developers. Int J Electron Commer 9(1):70–104. https://doi.org/10.1080/10864415.2004.110
44320
7. Bose I, Leung ACM (2014) Do phishing alerts impact global corporations? A firm value
analysis. Decis Support Syst 64(August):67–78
8. Ettredge M, Richardson VJ (2002) Assessing the risk in e-commerce. In: Presented at the
annual Hawaii international conference on system sciences. Hawaii, US
9. Campbell K, Gordon LA, Loeb MP, Zhou L (2003) The economic cost of publicly announced
information security breaches: empirical evidence from the stock market. J Comput Secur
11(3):431–448. https://doi.org/10.3233/JCS-2003-11308
10. Hovav A, D’Arcy J (2004) The impact of virus attack announcements on the market value
of firms. Inf Syst Secur 13(3):32–40. https://doi.org/10.1201/1086/44530.13.3.20040701/830
67.5
11. Kannan K, Rees J, Sridhar S (2007) Market reactions to information security breach announce-
ments: an empirical analysis. Int J Electron Commer 12(1):69–91. https://doi.org/10.2753/JEC
1086-4415120103
12. Malhotra A, Kubowicz Malhotra C (2011) Evaluating customer information breaches as service
failures: an event study approach. J Service Res 14(1):44–59
13. Sinanaj G, Muntermann J (2013) Assessing corporate reputational damage of data breaches:
an empirical analysis. In: Presented at the 26th Bled E-Conference—E-Innovations challenges
and impacts for individuals, organizations and society. Bled, Slovenia
696 S. E. A. Ali and F.-W. Lai

14. Tweneboah-Kodua S, Atsu F, Buchanan W (2018) Impact of cyberattacks on stock performance:

a comparative study. Inf Comput Secur 26(5):637–652. https://doi.org/10.1108/ICS-05-2018-
0060
15. Fama EF, Fisher L, Jensen MC, Roll R (1969) The adjustment of stock prices to new
information. Int Econ Rev 10(1):1–21. https://doi.org/10.2307/2525569
16. Schwartz RA (1970) Efficient capital markets: a review of theory and empirical work:
discussion. J Financ 25(2):421–423
17. Ponemon (2020) Cost of a data breach report, vol 2020. IBM Security, Michigan, USA.
Available: https://www.ibm.com/security/digital-assets/costdata-breach-report/#/
18. Hovav A, Han JY, Kim J (2017) Market reaction to security breach announcements: evidence
from South Korea. Data Base Adv Inf Syst 48(1):11–52. https://doi.org/10.1145/3051473.305
1476
19. Gordon LA, Loeb MP (2006) Budgeting process for information security expenditures.
Commun ACM 49(1):121–125
20. Gordon LA, Loeb MP (2002) The economics of information security investment. ACM Trans
Inf Syst Secur 5(4):438–457
21. Hsu C, Lee JN, Straub DW (2012) Institutional influences on information systems security
innovations. Inf Syst Res 23(3–2):918–939. https://doi.org/10.1287/isre.1110.0393
22. Lowry PB, Cao J, Everard A (2011) Privacy concerns versus desire for interpersonal awareness
in driving the use of self-disclosure technologies: the case of instant messaging in two cultures.
J Manag Inf Syst 27(4):163–200. https://doi.org/10.2753/MIS0742-1222270406
23. Hovav A, D’Arcy J (2005) Capital market reaction to defective IT products: the case of computer
viruses. Comput Secur 24(August):409–424
24. Deane JK, Goldberg DM, Rakes TR, Rees LP (2019) The effect of information security certi-
fication announcements on the market value of the firm. Inf Technol Manage 20(3):107–121.
https://doi.org/10.1007/s10799-018-00297-3
25. Wang T, Kannan KN, Ulmer JR (2013) The association between the disclosure and the realiza-
tion of information security risk factors. Inf Syst Res 24(2):201–218. https://doi.org/10.1287/
isre.1120.0437
26. Gordon LA, Loeb MP, Sohail T (2010) Market value of voluntary disclosures concerning
information security. MIS Q 34(3):567–594. https://doi.org/10.2307/25750692
27. Goldstein J, Chernobai A, Benaroch M (2011) An event study analysis of the economic impact
of IT operational risk and its subcategories. J Assoc Inf Syst 12(9):606–631. https://doi.org/
10.17705/1jais.00275
28. Cao J, Calderon T, Chandra A, Wang L (2010) Analyzing late SEC filings for differential
impacts of IS and accounting issues. Int J Account Inf Syst 11(3):189–207. https://doi.org/10.
1016/j.accinf.2010.07.010
29. Rosati P, Deeney P, Cummins M, Van der Werff L, Lynn T (2019) Social media and stock
price reaction to data breach announcements: evidence from US listed companies. Res Int Bus
Financ 47(September):458–469
30. Jeong CY, Lee SYT, Lim JH (2019) Information security breaches and IT security investments:
impacts on competitors. Inf Manag 56(5):681–695. https://doi.org/10.1016/j.im.2018.11.003
31. Johnson MS, Kang MJ, Lawson T (2017) Stock price reaction to data breaches. J Financ Issues
16(2):1–12
32. Filbeck G, Swinarski M, Zhao X (2013) Shareholder reaction to firm investments in the
capability maturity model: an event study. Eur J Inf Syst 22(2):170–190
33. Dehning B, Richardson VJ, Zmud RW (2007) The financial performance effects of IT-based
supply chain management systems in manufacturing firms. J Oper Manag 25(4):806–824
34. Hendricks KB, Singhal VR, Stratman JK (2007) The impact of enterprise systems on corporate
performance: a study of ERP, SCM and CRM system implementations. J Oper Manag 25(1):65–
82
35. Fama EF (1998) Market efficiency, long-term returns, and behavioral finance. J Financ Econ
49(3):283–306
Cyber Security Breaches and the Long-Run Effect … 697

36. Barber BM, Lyon JD (1997) Detecting long-run abnormal stock returns: the empirical power
and specification of test statistics. J Financ Econ 43(3):341–372. https://doi.org/10.1016/
S0304-405X(96)00890-2
37. Kothari S, Warner JB (1997) Measuring long-horizon security price performance. J Financ
Econ 43(3):301–339
38. Lyon JD, Barber BM, Tsai CL (1999) Improved methods for tests of long-run abnormal stock
returns. J Financ 54(1):165–201. https://doi.org/10.1111/0022-1082.00101
39. Lee I, Loughran T (1998) Performance following convertible bond issuance. J Corp Finan
4(2):185–207
40. Hendricks KB, Singhal VR (2014) The effect of demand–supply mismatches on firm risk. Prod
Oper Manag 23(12):2137–2151
41. Hendricks KB, Singhal VR (2005) An empirical analysis of the effect of supply chain disrup-
tions on long-run stock price performance and equity risk of the firm. Prod Oper Manag
14(1):35–52
42. Ali SEA, Khurram S (2017) Impact of demographic and health factors on GDP growth of South
Asian Countries. Int J Acad Res Bus Soc Sci 7(3):2222–6990
43. Carhart MM (1997) On persistence in mutual fund performance. J Financ 52(1):57–82
44. Fama EF, French KR (1996) Multifactor explanations of asset pricing anomalies. J Financ
51(1):55–84
45. Jegadeesh N, Titman S (1993) Returns to buying winners and selling losers: implications for
stock market efficiency. J Financ 48(1):65–91
46. Ali SEA, Lai F-W, Dominic PDD, Brown N, Lowry PB, Ali RF (2021) Stock market reactions to
favorable and unfavorable information security events: A systematic literature review. Comput
Secur 110(November):102451
47. Ali SEA, Lai F-W, Hassan R (2020) Socio-economic factors on sector-wide systematic risk of
information security breaches: Conceptual framework. Paper presented at the 9th International
Economics and Business Management Conference, Melaka, Malaysia
Towards Early Distribution
of Container-Based Microservices
in Cloud Computing Environment

Abdul Saboor, Ahmad Kamil Mahmood, Mohd Fadzil Hassan,

Syed Nasir Mehmood Shah, Muhammad Aadil Siddiqui,
Saeed Ahmed Magsi, and Muhammad Junaid

Abstract Microservices gained the popularity for use in scalable cloud application.
The application based on microservices uses intensive network communication to call
other microservices or to pass on the messages to hundreds of microservices running
inside the cloud/edge nodes. This study addressed the early distribution of container-
based microservices and proposed two distribution strategies named as Random
Distribution and Design Pattern Distribution. In random distribution approach the
microservices are assigned arbitrarily to the available data centers. While in design
pattern distribution the microservices are grouped together on the basis of behavioral
design patterns which identifies common communication patterns among objects.
The proposed solution was tested using custom built simulation environment and the
results showed that the early distribution of microservices according to the design
pattern of the application resulted in significant reduction of network calls to the
microservices hosted at other network nodes or data centers.

Keywords Cloud/edge computing · Microservices · Container · Optimization

A. Saboor (B) · A. K. Mahmood

High Performance Cloud Computing Center (HPC), Universiti Teknologi Petronas, 32610 Sri
Iskandar, Malaysia
e-mail: [email protected]
A. K. Mahmood
e-mail: [email protected]
M. F. Hassan
Institute of Autonomous System for Autonomous Facilities, Universiti Teknologi Petronas, 32610
Sri Iskandar, Malaysia
e-mail: [email protected]
S. N. M. Shah
KICSIT, Institute of Space Technology (IST), Islamabad, Pakistan
M. A. Siddiqui · S. A. Magsi · M. Junaid
Universiti Teknologi Petronas (UTP), 32610 Sri Iskandar, Malaysia

© The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 699
R. Ibrahim et al. (eds.), International Conference on Artificial Intelligence
for Smart Community, Lecture Notes in Electrical Engineering 758,
https://doi.org/10.1007/978-981-16-2183-3_67
700 A. Saboor et al.

1 Introduction

In the last decade cloud computing has gained interest of the industries and research
institutes. The Cloud Computing enables ubiquitous computing and provides the
convenient on-demand access to the shared pool of configurable resources, such as
storage, network, server, applications and services [1]. The cloud data centers mainly
relies on the concept of virtualization to provide access to the shared resources and
cloud services [2]. For the last few years there had been a shift towards the container-
ization. The concept of containers creates lightweight and flexible environment by
allowing applications to share operating system. Recent advancement in container
technology have encouraged organization and researchers to use the containers in the
cloud environment. The adoption of containers helped in realization of microservices
at cloud level. Microservices offer number of advantages such as scalability, relia-
bility, re-usability, fast response time, reduced cost, and many more [3, 4]. Microser-
vices also comes with number of challenges. The microservices adoption challenges
include security vulnerabilities [5, 6], finding the right size and number of services
[3].
As microservices are independent units and often needs to interact with other
microservices. The large number of microservices and their interaction thus increases
the communication overhead. One way to provide reliable and fast communication
is by decreasing the communication among physical machine hosted at multiple data
centers. This study suggests different ways to distribute the microservices on physical
machines so that the microservices have to use lesser communication network. The
microservices are ordered/managed in such a way that they mostly interact with the
microservices available at same physical machine and do not have to interact most
of time with the microservices hosted on other physical machines.

2 Literature Review

In cloud computing one of the most critical factor is to provide quality of service
(QoS) within the defined constraints. Traditionally, heuristic scheduling algorithms
had been used to provide QoS by minimizing the processing time and schedule tasks
efficiently [7], but the design shift of cloud services from monolithic applications to
loosely coupled micro-services brought in new challenges such as increased make-
span, heavy network utilization, and increased cost. These challenges are addressed
by the researchers time to time. According to Zahang [8] container placements on
VM is an issue. The VM placement and container placement should not be addressed
individually, therefore Zahang proposed a solution named Container-VM-PM. The
solution resulted in optimization of the placement of new containers and minimized
the number of physical machines and resource wastage.
In a study by Lin [9], the container-based microservice scheduling was done based
on multi-objective optimization method using an ant colony optimization algorithm.
Towards Early Distribution of Container-Based Microservices … 701

The solution took into account the utilization of computing and memory resources of
physical machines However, the study considered microservice scheduling in single
container only, and did not considered the producer consumer relationship of the
microservices. Similarly, Lv et al. [10] proposed the container distribution strate-
gies to address the problem of heavy communication workloads among containers.
For container placement they proposed the worst fit decreasing algorithm, and
for container assignment proposed a two stage Sweep and Search algorithm. The
experimental results showed the reduction in communication overhead among the
containers.
Zhou et al [11] proposed the optimal placement schema for containers in the
cloud environment. They introduced one-shot algorithm which work as placement
schema for the container cluster, and an online algorithm which breaks down the
online decision making into on-spot decision depending on resource price.
Recently another study by Sampaio et al [12] proposed an adaption mechanism,
named as REMaP (Run-time Microservices Placement). In this model the automatic
run-time placement of microservices is based upon usage history of resources and
mciroservices affinity.
The literature review showed one common concern that heavy network commu-
nication and latency is one of the major issue in container based cloud environment.
Especially the decomposition of a particular application into many small microser-
vices results into large round trips between service calls. The possible solution could
be aggregating multiple services or batch service call in a single round trip, or
replacing expensive inter-process communication call with language-level function
calls [13]. This study propose an early distribution of container-based microservices
in cloud environment such that the microservices are grouped in such a way that it
results in lower number of network calls. The solution addressed the distribution of
microservices and containers at early stages of application deployment.

3 Methodology

For this paper the distribution of microservices is considered in such a way that
one container hosts one and only one microservice in the cloud/edge environ-
ment. This section describes two proposed microservice distribution approaches i.e.
Randomized Distribution, and Design Pattern Distribution.

3.1 Randomized Distribution

The random distribution and assignment approach does not take into account any
performance parameters while assigning the containers to randomly chosen phys-
ical machine of the data centers, refer Fig. 1. The services are deployed arbitrarily
702 A. Saboor et al.

Microservice 4

Pool 1
Microservice 1
Microservice 7
Microservice 2
Microservice 3
Microservice 3 Cloud Data
Center

Pool 2
Microservice 4 Microservice 11

Microservice 5 Microservice 8

Microservice 6
Microservice 6

Pool 3
Microservice 7 Cloud Data
Microservice 2
Microservice 8 Center
Microservice 9
Microservice 9
Microservice 10

Pool 4
Microservice 10
Microservice 1
Microservice 11 Cloud Data
Microservice 5
Center

Fig. 1 Random selection and distribution of microservices to the data centers

i.e. without any consideration of performance parameters. Therefore this way of

assignment can also be termed as a classical approach.
The microservices are assigned randomly and the assigned containers are
distributed across random location. Thus resulting in no consideration of design
pattern or microservices interaction.
The random distribution approach is shown in Fig. 1 where the randomize function
selects a microservices on random basis and assign it to the random pools. Then these
pools of microservices are randomly assigned to the available data centers.
In such a random distribution there is high chance that a microservice assigned
to on data center is totally dependent on a microservice available at other data
center. Thus, the execution of such microservices is likely to place multiple network
communication calls between services for successful execution.

3.2 Design Pattern Distribution

The term pattern used by software designers describes the software abstraction, thus
patterns used across other procedures, subroutines, and objects may also combine
for more common abstraction [14]. According to Erl [15] it is set of related patterns
which provides the building blocks and these building blocks are built upon the other
blocks thus defining the pattern application sequence. For this study we considered
the behavioral patterns as these design patterns are concerned with communication
between objects.
Towards Early Distribution of Container-Based Microservices … 703

Microservice 1 Microservice 2

Microservice 2 Microservice 4

Microservice 3 Microservice 7
Cloud Data
Microservice 4 Microservice 9 Center
Microservice 5 Microservice 1

Microservice 6 Microservice 6

Microservice 7 Microservice 3
Cloud Data
Microservice 8 Microservice 8 Center
Microservice 9 Microservice 5

Microservice 10 Microservice 10

Microservice n Microservice n
Cloud Data
Center

Fig. 2 Arrangement of microservices according to their interaction dependencies and assignment

to the data centers

The proposed design pattern distribution schema is shown in Fig. 2, where

microservices are assigned according to the relation and interaction defined in soft-
ware design. If two or more microservices are dependent on each other for the execu-
tion of services, then those microservices are pooled together. Then these pools are
assigned to the respective data centers.

3.3 Results and Discussion

To assess the proposed techniques, we designed the simulation program using Java
to produce the microservices utilization data sets. To get the insights of the produced
data sets we used an online tool called plotly [16]. Plotly provided online analytic
and visualization of the data sets. The data sets were produced for one hour of
microservices calls, and for the assessment of the proposed system it was considered
that maximum of 1000 calls can be originated for the microservices at any given time.
The anatomized simulation process however reached maximum of 729 simultaneous
microservices calls at a given time. The number of service calls originated from the
system are shown in Fig. 3. They are the simulated calls for microservices which will
be used for testing the random and design pattern distribution. To get the test results
for classical random approach, the service calls were originated for the microservices
distributed randomly at different location i.e. data centers. The number of network
calls placed by the microservices in random distribution are shown in Fig. 4. It clearly
704 A. Saboor et al.

Fig. 3 Number of microservices calls originated from the system over a period of one hour

Fig. 4 Number of network calls in random distribution of microservices

shows that heavy network traffic is generated by such a distribution. On average 189
network calls were placed at a given time.
In the second simulation the design pattern distribution strategy was tested and the
results showed that the number of network calls reduced significantly (refer Fig. 5).
On average 84 network calls we placed at a given time, which are very low when
compared to random distribution. The comparison of both distribution strategies is
shown in Fig. 6.

Fig. 5 Number of network calls in design pattern distribution of microservices

Towards Early Distribution of Container-Based Microservices … 705

Fig. 6 Comparison of random distribution (RD) and design pattern distribution (DPD) network
calls

During the multiple simulation run of the system, it was noted that for couple
of time the design pattern based distribution of network calls exceeded the random
distribution network calls, but it was rare. This was due to the ideal match of the
microservices calls and the distribution done during the random distribution. Thus
less network calls placed in such a configuration were by coincidence.

3.4 Conclusion and Future Work

The design shift of cloud services from monolithic applications to loosely coupled
micro-services brought in many challenges. This study addressed the communication
issues and proposed two microservices distribution strategies to reduce the network
calls. The random distribution and design pattern distribution were suggested, and
the simulation tests showed that the early distribution of microservices according
to the design pattern of application significantly reduce the network call to other
microservices.
The proposed system was tested using a custom built data sets and the system
further needs to be tested using real cloud traces. In future the study will also
be extended by defining other early distribution strategies such as priority based
distribution and weighted mean average distribution.

References

1. Mell P, Grance T, et al (2011) The nist definition of cloud computing

2. Gong C, Liu J, Zhang Q, Chen H, Gong Z (2010) The characteristics of cloud computing. In:
2010 39th international conference on parallel processing workshops. IEEE, pp 275–279
3. Amaral M, Polo J, Carrera D, Mohomed I, Unuvar M, Steinder M (2015) Performance
evaluation of microservices architectures using containers. In: 2015 IEEE 14th international
symposium on network computing and applications. IEEE, pp 27–34
706 A. Saboor et al.

4. Malavalli D, Sathappan S (2015) Scalable microservice based architecture for enabling dmtf
profiles. In: 2015 11th international conference on network and service management (CNSM).
IEEE, pp 428–432
5. Newman S (2015) Building microservices: designing fine-grained systems. ”O’Reilly Media,
Inc.”
6. Yu D, Jin Y, Zhang Y, Zheng X (2019) A survey on security issues in services communication of
microservices-enabled fog applications. Concurrency Comput: Pract Experience 31(22):e4436
7. Juarez F, Ejarque J, Badia RM (2018) Dynamic energy-aware scheduling for parallel task-based
application in cloud computing. Futur Gener Comput Syst 78:257–271
8. Zhang R, Zhong AM, Dong B, Tian F, Li R (2018) Container-vm-pm architecture: A novel
architecture for docker container placement. In: International conference on cloud computing.
Springer, pp 128–140
9. Lin M, Xi J, Bai W, Wu J (2019) Ant colony algorithm for multi-objective optimization of
container-based microservice scheduling in cloud. IEEE Access 7:83088–83100
10. Lv L, Zhang Y, Li Y, Xu K, Wang D, Wang W, Li M, Cao X, Liang Q (2019) Communication-
aware container placement and reassignment in large-scale internet data centers. IEEE J Sel
Areas Commun 37(3):540–555
11. Zhou R, Li Z, Wu C (2019) An efficient online placement scheme for cloud container clusters.
IEEE J Sel Areas Commun 37(5):1046–1058
12. Sampaio AR, Rubin J, Beschastnikh I, Rosa NS (2019) Improving microservicebased
applications with runtime placement adaptation. J Internet Services Appl 10(1):1–30
13. Richardson C (2019) Microservices patterns: with examples in Java. Manning publications
14. Coplien JO (1998) Software design patterns: common questions and answers. In: The patterns
handbook: techniques, strategies, and applications, vol 13, pp 311
15. Erl T (2008) SOA design patterns (paperback). Pearson Education
16. Sievert C, Parmer C, Hocking T, Chamberlain S, Ram K, Corvellec M, Despouy P (2017)
Plotly: create interactive web graphics via ‘plotly. js’. R package version 4(1):110
An AI-Based Chicken Disease
Management System

Luyl-Da Quach , Nghi Pham Quoc , Nhien Huynh Thi ,

Nhan Le Thanh , Xuan Truong Thi Thanh , Duc Chung Tran ,
and Mohd Fadzil Hassan

Abstract Food stability has always received worldwide attention, especially in the
development of the poultry industry. However, the poultry’s diseases have caused the
loss of the poultry population and direct income of the owners. In this work, a system
comprising of a website and a mobile application has been developed to support
disease identification in chickens; the disease spread locations can be managed and
traced with maps. In addition, the system allows farmers to make contact and get
support from experts through chat, voice calls, and video calls. It also has some
e-commerce functions for improving sales of poultry’s owners. The system uses the
improved ResNet-50 model, with an accuracy of about 93.56%.

Keywords Disease identification in chickens · Identify disease spots · Application

for livestock · ResNet-50

L.-D. Quach (B) · N. P. Quoc · N. H. Thi · N. Le Thanh · X. T. T. Thanh

Software Engineering Department, FPT University, Cantho city, Vietnam
N. P. Quoc
e-mail: [email protected]
N. H. Thi
e-mail: [email protected]
N. Le Thanh
e-mail: [email protected]
X. T. T. Thanh
e-mail: [email protected]
D. C. Tran
Computing Fundamental Department, FPT University, Hanoi, Vietnam
e-mail: [email protected]
M. F. Hassan
Department of Computer and Information Sciences, Universiti Teknologi PETRONAS, Seri
Iskandar, Perak, Malaysia
e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 707
R. Ibrahim et al. (eds.), International Conference on Artificial Intelligence
for Smart Community, Lecture Notes in Electrical Engineering 758,
https://doi.org/10.1007/978-981-16-2183-3_68
708 L.-D. Quach et al.

1 Introduction

According to the World Health Organization (WHO), as of September 2020, the

world has about 30 million patients infected with COVID-19, of which 940,000
deaths. This shows the rapid spread of the epidemic in recent years. Therefore, to
limit the spread of disease, governments of countries have introduced policies to
restrict travel, avoid crowding, isolate society, etc. However, the demand for food is
continuously increasing, especially livestock and agriculture.
According to the United States Department of Agriculture statistics on July 10,
2020, despite the epidemic of COVID, the demand for food imports from China was
increasing, especially chicken products. Despite the decline from the US and China,
production is still about 1% higher than the world (volume reached 100 million tons)
[1]. The demand for poultry products is increasing. According to the forecasts 2016–
2025 of The Organization for Economic Co-operation and Development (OECD),
the output of poultry meat from consumer demand in 41 countries and 12 regions
could increase from 115,247 million tons (in 2006) to 131,225 million tons (in 2025)
[2].
The poultry industry’s development also entails many risks, not only causing
economic losses but also affecting the ability to expand the scale of livestock produc-
tion. According to statistics from the World organization for animal health, in the
period from July 10 to July 30, there are about 39,161 birds destroyed in Africa
and Asia with the H5N8 avian flu standard [3] appeared. Some new countries,
such as the Philippines and South Africa. In 2014, the H5N2 disease was incu-
bated between November 26–28 and December 2, which resulted in the destruc-
tion of over 48 million birds in a short period [4]. Also, poultry pathogens such as
Escherichia coli, Salmonella, and Campylobacter jejuni can directly affect humans
by consuming poultry products [5]. Therefore, developing technology for monitoring
and monitoring is a very urgent matter.
In the world, there has been researched on diagnosing and detecting diseases in
plants and animals. Astill [6] conducted research on the identification and detection
of avian influenza virus based on new big data technology. Meanwhile, Hemalatha [7]
identify chickenpox with machine learning methods such as Support Vector Machine
in conjunction with Gaussian Radial Basis Function and Extreme Learning Machine,
the uses in this study using grey-level matrices with the calculation of mean, stan-
dard deviation, noise, etc. for statistics. Based on the humming sound signals in
chickens, Rizwan [8] utilized the extreme machine learning method and support
vector machine to classify healthy and active chickens. Specifically, disease data is
recorded in 20 min per day and labeled of 25 consecutive days. The results showed
potential automatic sound monitoring of flock health. The above studies gave us that
there have been both accuracy and feasibility in diagnosing and detecting diseases.
However, the urgency of having an automatic system of identification and diagnosis
of diseases in chickens using artificial intelligence applications is also implied more
and more clearly. Breeders can identify the problems related to their chickens timely
and promptly. According to Astill [9], some standard methods of detecting disease
An AI-Based Chicken Disease Management System 709

in poultry, such as biosensor, wearable sensor, image analysis, surveillance robot,

etc. were developed. They were unable to promote infectious diseases and limited
to having to install on individual instances, though. Since then, it was essential to
develop a new system based on monitoring, machine learning, and history tracking.
The identification of chicken diseases based on machine learning and deep
learning methods has been carried out in many nations and territories, consisting of
Vietnam. Previously, Quach [10] performed the experimental research for utilizing
the deep learning model ResNet-50, and VGGNet to evaluate four diseases in chicken
with 74.1% accuracy. Currently, the group has examined to integrate the ResNet-50
model, SURF characteristics, and K-means clustering method to improve the accu-
racy in identifying disease in chicken. The results proved that this method helped
increase the model’s accuracy to 93.56% and the number of features available for
training [11].
In this research, a system for identifying and managing disease zones in poultry
using artificial intelligence was built. The system formed two basic components,
such as:
• Mobile application for smallholder farmers using the trained ResNet-50 model
focuses on disease identification in poultry through images. The application also
has additional functions such as helping users purchase livestock products such
as feed, veterinary medicine, and vaccines from stores linked to the system. An
important and indispensable function is to contact veterinarians for their advice.
• The website manages the poultry disease zone based on the user’s coordinates
when they identify the disease. The coordinates will be displayed on the map to
help experts evaluate the disease area for timely treatment measures. Besides, the
website needs an interface for experts to advise users when needed.

2 Methodology

2.1 Research Methods

The methods are applied include as follows:

• Document study method: Based on documents about models, algorithms, project
implementation, and libraries.
• Experimental methods: Collect data, train models, write software, and test the
system.
• Statistical method: Evaluating the accuracy of the model, assessing the model’s
reliability.
710 L.-D. Quach et al.

2.2 Research Content

Scrum is an Agile software development model used in small to medium projects

[12]. By dividing the project into small phases to easily manage and adapt to changes
in the development process. The stages used in the study included:
• Stage 1: Overall research planning, implementation planning, and development
goals to be achieved.
• Stage 2: Analyzing the overall requirements of the project, thereby identifying
the problem to be built, the level of prediction, and the system to be met, etc.
• Stage 3: Data collection and training. This stage was done using:
– Programming language: Python.
– Libraries: Tensorflow, Sklearn.
• Stage 4: Design and program the system based on phases 1, 2, and 3. This stage
was done using:
– Programming language: Python, Javascript.
– Framework: Flash, ReactJS, React Native, Bootstrap4.
– API: Mapbox, Facebook.
– Database management: Sqlite.

3 Application

To build the system, research based on the requirements, the process of finding
out management problems in practice, and building the model include the Main
functional model of the system (use case model), relational database model, business
model, and application development.

3.1 Disease Identification System

The study collected 492 images, divided into four disease classes: avian pox, bird-flu,
Marek, and infectious laryngotracheitis. Details of the number and illustrations of
the disease are shown in Table 1 and Fig. 1.
After obtaining the data, the study conducted steps aimed at data transformation,
including the following steps:
• Step 1: Use the SURF feature [12] to perform characteristic extraction from
diseased chickens. Each SURF feature has 64 vectors with parameters 5000 and
10,000. The image is depicted as Fig. 2 and the quantity in Table 2.
• Step 2: Research using the K-Means algorithm [13] with parameter k = 4 to filter
the pixels from the obtained feature vector. As a result, the study obtained a result
An AI-Based Chicken Disease Management System 711

Table 1 Statistics of chicken

Labels Total species
disease image number
Avian pox 138
Bird-flu 147
Marek 96
Infectious laryngotracheitis 111
Total 492

Fig. 1 Illustrative image of chicken diseases, from left to right: Bird-flu, Avian pox, Infectious
laryngotracheitis, Marek

Fig. 2 Extract the SURF feature with parameters 5000 and 10,000

Table 2 Statistic number of

Labels Parameters Number of features
characteristics collected from
chicken disease data Avian pox 5000 2568
collection Bird-flu 5000 5272
Marek 5000 2191
Infectious laryngotracheitis 10,000 1310
Total 11,341
712 L.-D. Quach et al.

Table 3 Set features after

Labels Total species
remaining options
Avian pox 472
Bird-flu 475
Marek 326
Infectious laryngotracheitis 120
Total 1393

Fig. 3 Examples of samples selected after step (ii), from left to right, are diseases: Bird-flu, Avian
pox, Marek, Infectious laryngotracheitis

set of 1393 characteristics of the four obtained diseases described in the table. The
results obtained include 1393 characteristics to train with four different diseases
with the number as Table 3 and illustrations Fig. 3.
• Step 3: The research using the ResNet-50 model [14] to train the disease identifi-
cation dataset. The study conducted experiments, selected parameters and input
size 64 × 64 × 3 with default. The reseach got the best result was 93.56%, with
a learning rate of 0.01, and the worst result on our test was 90.93, with a learning
rate of 1. Efficien-cy The best yield of our adjusted model on training and test
sets, learning rate = 0.01, is shown in Fig. 4. Research results are used to illustrate
the process of disease identification for the system.

3.2 Main Use Case Model of the System

The use case diagram is used to represent the relationship between the use case
and the system agent. The actors in the diagram are users (people), some internal
or external applications. Use cases are a major function of the system [15]. In the
system of research (see Fig. 5), parts of the use case diagram include:
• Target users: Administrators, owners of livestock products shops, veterinary
establishments, and smallholder farmers.
• Use cases that describe the system’s main functions are disease identification in
chickens, contact advice with experts from veterinary facilities, e-commerce, map
zone management, and user decentralization.
An AI-Based Chicken Disease Management System 713

Fig. 4 Training results with learning rate = 0.01, epochs = 1000, input size = 64 × 64 × 3

Fig. 5 Use case diagram in poultry disease management systems

3.3 Relational Database Model

The relational database model is the model that builds a digital database based on
data relationships, as proposed by the E. F. Codd in 1970. A software system will
714 L.-D. Quach et al.

use this database to avoid data redundancy and ensure data consistency is described
[17]. Therefore, research building a basic management system includes the entities.
The system functions to meet the requirements set out include store management,
products, users, etc.

3.4 Business Model

Based on the relational database model and use case diagram, the research team has
built up the system’s operation diagram with essential functions on two platforms.
Dissipation includes:
1. Business Model of the Website System
The system includes a home page to login; disease and density zoning maps;
consulting contact by message, call, or video call. On the management website
of the system, the research team used the streets-v9 map from Mapbox GL-JS
[17] with the map center coordinates of [16] and the zoom level of 4.8 to allow
the territory of Vietnam to be displayed immediately when users access and
direct consulting function on the website (see Fig. 6).
2. Business Model of Mobile Application:
The mobile phone’s application includes authentication of users through
registration and login, homepage, image-based diagnostics through library or
camera, contact consultants such as system website, purchase, and farm manage-
ment. Based on analytical models, the research team has developed a mobile
application that includes functions for small poultry farmers (see Fig. 6).

Fig. 6 Video demo of the

proposed system
An AI-Based Chicken Disease Management System 715

4 Conclusion

With the available data, research has shown the feasibility of identifying the disease
in chickens. With the combination of the SURF feature determination method, the
K-means algorithm has helped the ResNet-50 model improve the training and recog-
nition accuracy and avoid the limitations of the CNN model, such as saturation and
overfitting 93.56% accuracy with four common diseases. The demo video of the
application is shown in the QR code (Fig. 6). Through the application of the image
recognition model and the deployed system, the system can develop in the following
directions:
• Collecting data directly from users: The purpose is to increase data, improve
model accuracy, and apply other CNN models.
• By zoning the data using the K-means algorithm, the model found that an indi-
vidual chicken can suffer from many diseases. Therefore, the application of this
model helps to detect any problems on the individual body.
• Improved model accuracy through a combination of image recognition and
descriptive (text) data.
• Develop a farm management model according to world standards such as
AsianGAP, GlobalGap, etc. from here; it is possible to query the origin of clean
poultry meat to serve consumer needs.
• Develop an ecosystem to link farmers with veterinary facilities, livestock product
suppliers to make e-commerce easier.
• The disease management system helps trace the origin of the disease and helps
the regulator timely detect and localize the infection to avoid spreading to other
areas and affecting people.

References

1. Kuberka L, Vuillemin J, Mezoughem C (2020) Livestock and poultry: world markets and trade.
US Department of Agriculture Foreign Agricultural Service, Washington, DC, p 16
2. OECD (2016) Table 3.A1.4—World meat projections. In: OECD-FAO Agricultural Outlook
2016–2025. OECD Publishing, Paris. https://doi.org/10.1787/agr_outlook-2016-table102-en
3. World Organisation for animal health (2020) Highly pathogenic avian influenza (HPAI)
report. No. 12 https://www.oie.int/fileadmin/Home/eng/Animal_Health_in_the_World/docs/
pdf/OIE_AI_situation_report/HPAI_asof30072020.pdf. Accessed 08/2020
4. Shriner SA, Root JJ, Lutman MW, Kloft JM, VanDalen KK, Sullivan HJ et al (2016) Surveil-
lance for highly pathogenic H5 avian influenza virus in synanthropic wildlife associated with
poultry farms during an acute outbreak. Sci Rep 6:1–11. https://doi.org/10.1038/srep36237
5. Mellata M (2013) Human and avian extraintestinal pathogenic Escherichia coli: infections,
zoonotic risks, and antibiotic resistance trends. Foodborne Pathog Dis 10(11):916–932
6. Astill J, Dara RA, Fraser E, Sharif S (2018) Detecting and predicting emerging disease in
poultry with the implementation of new technologies and big data: a focus on avian influenza
virus. Front Vet Sci 5:263. https://doi.org/10.3389/fvets.2018.00263
716 L.-D. Quach et al.

7. Hemalatha MS, Maheswaran R (2014) Recognition of poultry disease in real time using extreme
learning machine. In: Procceedings of the internation conference on inter disciplinary research
in engineering & technology 2014 [ICIDRET 2014], pp 44–50
8. Rizwan M, Carroll BT, Anderson DV, Daley W, Harbert S, Britton DF, Jackwood MW (2016)
Identifying rale sounds in chickens using audio signals for early disease detection in poultry.
In: 2016 IEEE global conference on signal and information processing (GlobalSIP). USA, pp
55–59. https://doi.org/10.1109/GlobalSIP.2016.7905802
9. Astill J, Dara RA, Fraser ED, Sharif S (2019) Corrigendum: detecting and predicting emerging
disease in poultry with the implementation of new technologies and big data: a focus on avian
influenza virus. Front Vet Sci 5:337
10. Quach L-D, Quoc NP, Tran DC, Hassan MF (2020) Identification of chicken diseases using
VGGNet and ResNet models. In: EAI INISCOM 2020—6th EAI international conference on
industrial networks and intelligent system, springer publishing (presented)
11. Quach L-D, Quoc ND, Thi NN, Tran DC, Hassan MF (2020) Using SURF to improve ResNet-50
model for poultry disease recognition algorithm. In: International conference on computational
intelligence 2020. Malaysia
12. Bay H, Tuytelaars T, Gool L (2006) SURF: speeded up robust features. In: Leonardis A, Bischof
H, Pinz A (eds) ECCV 2006, vol 3951. LNCS, pp 404–417
13. Steinley D, Brusco MJ (2007) Initializing k-means batch clustering: a critical evaluation of
several techniques. J Classif 24(1):99–121
14. He K, Zhang X, Ren S, Sun J (2016) Identity mappings in deep residual networks. European
conference on computer vision. Springer, Cham, pp 630–645
15. Khurana N, Chhillar RS, Chhillar U (2016) A novel technique for generation and optimization
of test cases using use case, sequence. Act Diagram Genet Algorithm. JSW 11(3):242–250
16. Paredaens J, De Bra P, Gyssens M, Van Gucht D (2012) The structure of the relational database
model. vol 17. Springer Science & Business Media. https://scholar.google.com/scholar?hl=
en&as_sdt=0%2C5&q=The+structure+of+the+relational+database+model&btnG=
17. Eriksson O, Rydkvist E (2019) An in-depth analysis of dynamically rendered vector-based
maps with WebGL using Mapbox GL JS
Automatic Polyp Segmentation
in Colonoscopy Images Using Single
Network Model: SegNet

Chin Yii Eu, Tong Boon Tang, and Cheng-Kai Lu

Abstract Colorectal cancer is the third most common diagnosed cancer world-
wide. Early detection and removal of adenoma during the colonoscopy examination
may increase the survival probability. A novel computer-aided tool for automated
polyp segmentation in colonoscopy images is described in this work. SegNet, a deep
convolutional neural networks has been chosen to map low resolution features with
the input resolution for automated pixel-wise semantic polyp segmentation. Publicly
available databases, CVC-ClinicDB, CVC-ColonDB, and ETIS-LaribPolypDB were
used to train and to test the model. The outcome demonstrated the proposed method
is feasible as it attains an average of 81.78, 92.35% for mean intersection over union,
and dice coefficient, respectively for testing on a combination of the aforementioned
datasets.

Keywords Colorectal cancer · Computer-aided diagnosis · Convolutional neural

network · Polyp segmentation

1 Introduction

Accounting all types of cancer, colorectal cancer is the third most common diag-
nosed cancer worldwide [1]. As the name indicates, colorectal cancer is a malignant
tumor found in colon or rectum and often due to the specific type of polyp [2].

C. Y. Eu
Department of Electrical and Electronic Engineering, Universiti Teknologi PETRONAS, 32610
Seri Iskandar, Malaysia
T. B. Tang
Centre of Intelligent Signal and Imaging Research (CISIR), Universiti Teknologi PETRONAS,
32610 Seri Iskandar, Malaysia
C. Y. Eu (B) · C.-K. Lu
Smart Assistive and Rehabilitative Technology (SMART), Institute of Health and Analytics
(IHA), Universiti Teknologi PETRONAS, 32610 Seri Iskandar, Malaysia
e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 717
R. Ibrahim et al. (eds.), International Conference on Artificial Intelligence
for Smart Community, Lecture Notes in Electrical Engineering 758,
https://doi.org/10.1007/978-981-16-2183-3_69
718 C. Y. Eu et al.

Currently, optical colonoscopy is the gold standard to screen colon for polyp detec-
tion and prevention of colorectal cancer. However, the polyp miss rate in colonoscopy
screening is approximately 10–30% [3]. Missed and misdiagnosed polyp are due to
various polyp structure [4]. Thus, computer-aided diagnosis (CAD) tool needs to
be developed to help medical expert in enhancing the accuracy and promptness of
medical diagnosis.
Automated polyp segmentation task can be presented in various methods which
essentially divided into two domain, traditional handcrafted method and modern deep
learning method. Those traditional methods often use features for polyp segmenta-
tion. For example, Sánchez-González et al. [5] used texture and color features for
training classifiers. The author utilize image windowing to increase generalization
and effectiveness. Active contour segmentation method was proposed by Dutta et al.
[6], where the active contour framework was modified to work without edges.
Deep learning techniques have been massively initiated to the image analysis
domain and outperform the old traditional ways consequently in current era. In
Vázquez et al. [7], they used Fully Convolutional Network—8 (FCN-8) to up-sample
the image to recover spatial resolution for per pixel prediction task. Huang et. al. [8]
proposed and compared two deep neural networks, DeepLab_v3 with Long Short-
Term Memory networks (LSTMs) and Segnet with LSTMs. The LSTMs are utilized
for keeping localization information that could slowly vanish along the encoder
networks. Another two proposed methods namely Dilated ResFCN and SE-Unet
were also compared in Guo and Matuszewski [9].
In this paper, a SegNet based CAD tool was proposed for polyp segmentation. The
proposed method has been validated in publicly available datasets, CVC-ClinicDB,
CVC-ColonDB and ETIS-Larib. The rest of the paper presented as follows. The
method used in this algorithm are demonstrated in Sect. 2, and the results obtained
are further discussed in Sect. 3. Lastly, Sect. 4 concludes this paper.

2 Material and Method

2.1 Polyp Datasets

Three publicly available polyp databases namely CVC-ClinicDB, CVC-ColonDB,

and ETIS-Larib polyp were used to train and evaluate the proposed method. Out
of 1187 images, 70% of the databases were used to train the segmentation model
network; while the rest were used to test the model. CVC-ClinicDB contains 612
standard definition still images of 384 × 288 resolution. CVC-ColonDB database has
379 images of 574 × 500 resolution. ETIS-LaribPolypDB is a polyp database that
has 196 high-definition still images with a resolution of 1225 × 966. Each image was
associated with own individual polyp masks. The ground truth segmentation masks
were manually drawn and verified by expert clinicians.
Automatic Polyp Segmentation in Colonoscopy Images … 719

Fig. 1 Left image represents original image and right image represents pre-processed image

2.2 Image Pre-Processing and Normalization

Feature extraction through any image processing is greatly affected by the image
quality itself. In this research work, thresholding was performed on image datasets
as a method for pre-processing. Image pixel threshold values were set for both red
channel and green channel as 50 and 15 respectively. Those pixel values that lower
than pre-defined threshold values were set to zero as the non-polyp background pixel.
The model network trained can be more biased to the polyp features rather than the
non-polyp background features since the image was mostly contributed by non-
polyp pixels. Next step is to normalize the image size to meet the input requirement
of network. The image datasets have been resized to 360 × 480. Figure 1 shows the
original image and the pre-processed image.

2.3 Proposed Model—SegNet

SegNet with VGG-19 framework architecture was used as our model. Datasets in
RGB image format were pre-processed, followed by model network training and
in the end test the model with testing images. The general steps involved in this
automated system are presented in Fig. 2.
SegNet is an architecture that was designed to map low resolution features with
the input resolution for pixel-wise semantic segmentation [10]. Encoder network
of SegNet is essentially the same as typical Convolutional Neural Network (CNN).

Input Pre- Network Segmentation

Normalization
Image processing Training Result

Fig. 2 Overall research methodology

720 C. Y. Eu et al.

Input Layer
3×3 Conv + BN + ReLU

Max-pooling
Max-unpooling
Softmax

Pixel Classification
Layer

Fig. 3 Overview of model network, SegNet

Encoder network consists of input layer, convolution layer (Conv), activation layer
and pooling layer. In convolution layer, a filter will be convolving around receptive
field [11]. Each of these filters can be thought of as feature identifiers. The filters on the
first layer convolve around the input image will be getting high values through convo-
lution when a filter found its desired feature from the input image. The same thing will
be iterated through all the location of input layer to form feature map. We included
batch normalization (BN) after each convolutional layer to normalize the output,
allowing for a larger learning rate that accelerate the training procedure. Trainable
parameters that remain untrained previously are able to start training. The following
layer is Rectified Linear Unit (ReLU), an activation layer which can limit the gener-
ated output. Pooling layer which usually comes after activation layer, is applied to
decrease the spatial size of the convolved feature. By doing this, the computational
power required to process the data can be greatly reduced. Moreover, it is useful for
extracting dominant features which are rotational and positional invariant. There is
a decoder network in SegNet which corresponded to the encoder network. It utilizes
max-pooling indices from previous encoder network to up-sample the input feature
map encoded from the encoder network [10]. At the final layer before SegNet last
output layer, there is one layer called Softmax. This layer works to segment the polyp
by activates output neurons via generating discrete probability distribution across all
the output neurons with the help of cross-entropy loss calculation. Overview of the
entire model network has been shown in Fig. 3.

3 Result and Discussion

In this project, pre-trained VGG-19 network was used as the framework architecture.
The network has been modified into SegNet by adding deconvolution network to meet
this project objectives. The training datasets were fed into the network together with
ground truth images for machine learning process. Testing images will be used to
evaluate the performance of the network in terms of mean Intersection over union
(mIoU) and dice coefficient. The network was trained with training options set to
Automatic Polyp Segmentation in Colonoscopy Images … 721

100

80
Accuracy

60

40

20

0 500 1000 1500 2000 2500 3000 3500 4000

Iteration

Fig. 4 SegNet training process

0.9 momentum, 0.001 learning rate, 20 epoch and mini batch size of 4. The network
was trained using single CPU, Intel® Core TM i7-2600 CPU @ 3.40 GHz, 8 GB
DDR3 RAM. Figure 4 shows the graph of training process. Our method results are
detailed in Table 1. This method was evaluated by comparing the performance with
other proposed work and the results surpass them.
According to Fig. 5a, the results show that entire polyps were successfully
segmented. Some over-segmented polyps can be seen in these results prove that
this model is poor in determining the edges of polyps. From Fig. 5b, the results
obtained were moderate. The outcome shows more over-segmented polyp regions
than the good results from Fig. 5a. The model did missed segment some true polyp
region and false polyp segmentation on light-reflected region. Some poor performed
outcomes also determined and presented in Fig. 5c. The model totally missed the

Table 1 Result comparison

Research work Mean intersection Dice coefficient
with previous research works
over union
Sánchez-González – 87.31%
et al. [5]
Akbari et al. [12] – 81.00%
Nguyen and Lee [13] – 89.60%
Huang et al. [8] 77.64% –
Guo and – 82.93%
Matuszewski [9]
Yu et al. [14] 83.40% 89.10%
Bagheri et al. [15] – 82.00%
Guo et al. [4] 96.00% 98.00%
Proposed Method 81.87% 92.35%
722 C. Y. Eu et al.

Fig. 5 First row of images show segmentation results, second row of images show ground truth. a
Good segmentation results. b Moderate segmentation results. c Poor segmentation results

location of polyp and most of segmented non-polyp pixels were found. These poor
results indicate that the model is not intelligent enough and unable to tackle more
complicated structure polyp and poor condition colon.
Our proposed model obtained 81.87% mIoU, lower than [4] and slightly lower
than [14] by 1.53%. Our dice coefficient reaches 92.35%, surpassed all previous
works that are all under 90% except [4] which is 98.00%. Our model outperforms
[14] on dice coefficient by 3.25%. Our model performed poorer than [4] due to our
model contains only one single network, while [4] proposed an ensemble of three
different networks into one final model. Our model is weak in classifying edges
of polyps, possibly due to some of polyp features being eliminated across deeper
encoder path especially convolution and max-pooling operations.

4 Conclusion

In this paper, SegNet was proposed as segmentation network. SegNet is an encoder-

decoder network that map low resolution features with the input resolution for pixel-
wise semantic segmentation. The proposed method were tested on CVC-ClinicDB,
CVC-ColonDB, and ETIS-LaribPolypDB databases. Our algorithm achieves results
of 81.78, 92.35% for mIoU and dice coefficient, respectively. There were certainly
some flaws in our model indicated there are rooms to improve further. In future work,
we planned to do research on color spaces that give more dominant polyps features
beside RGB image. We will also try to solve the problem of useful polyp features
being eliminated down the encoder path by preserving features and feed to decoder
path. Development of CAD tool in polyp segmentation would surely give beneficial
impact on future colorectal cancer diagnosis and management.

Acknowledgements This work was supported in part by Ministry of Higher Education, Malaysia,
under Grant FRGS/1/2020/TK0/UTP/02/23, and partially supported by YUTP-Fundamental
Research Grant (015LC0-002).
Automatic Polyp Segmentation in Colonoscopy Images … 723

References

1. “Cancer” World Health Organization (2018). Available: http://www.who.int/news-room/fact-

sheets/detail
2. Cao Y, Liu D, Tavanapong W, Wong J, Oh JH, de Groen PC (2006) Automatic classification of
images with appendiceal orifice in colonoscopy videos. In: 2006 international conference of
the IEEE engineering in medicine and biology society. Available: https://www.ncbi.nlm.nih.
gov/pubmed/17946108
3. Zheng Y, et al (2018) Localisation of colorectal polyps by convolutional neural network features
learnt from white light and narrow band endoscopic images of multiple databases. In: 2018
40th annual international conference of the IEEE engineering in medicine and biology society
(EMBC). Honolulu, HI, pp 4142–4145. https://doi.org/10.1109/EMBC.2018.8513337
4. Guo X, et al (2019) Automated polyp segmentation for colonoscopy images: a method based
on convolutional neural networks and ensemble learning. In: 2019 American association of
physicists in medicine, vol 46, no 12. doi.org/https://doi.org/10.1002/mp.13865
5. Sánchez-González A, Garcia-Zapirain B, Sierra-Sosa D, Elmaghraby A (2018) Colon polyp
segmentation using texture analysis. In: 2018 IEEE international symposium on signal
processing and information technology (ISSPIT). Louisville, KY, USA, pp 579–588. https://
doi.org/10.1109/ISSPIT.2018.8642748
6. Dutta S, Sasmal P, Bhuyan MK, Iwahori Y (2018) Automatic segmentation of polyps in endo-
scopic image using level-set formulation. In: 2018 international conference on wireless commu-
nications, signal processing and networking (WiSPNET). Chennai, pp 1–5. https://doi.org/10.
1109/WiSPNET.2018.8538615
7. Vázquez D et al (2017) A benchmark for endoluminal scene segmentation of colonoscopy
images. J Healthcare Eng 2017(4037190):9. https://doi.org/10.1155/2017/4037190
8. Huang C, Xiao W, Chang L, Tsai W, Liu W (2018) Automatic tissue segmentation by deep
learning: from colorectal polyps in colonoscopy to abdominal organs in CT exam. In: 2018
IEEE visual communications and image processing (VCIP). Taichung, Taiwan, pp 1–4. https://
doi.org/10.1109/VCIP.2018.8698645
9. Guo YB, Matuszewski BJ (2019) GIANA polyp segmentation with fully convolutional dilation
neural networks. In: Proceeding of the 14th international joint conference on computer vision,
imaging and computer graphics theory and applications, vol 4. Czech Republic, GIANA, pp
632–641. ISBN 978-989-758-354-4, https://doi.org/10.5220/0007698806320641
10. Badrinarayanan V, Kendall A, Cipolla R (2017) SegNet: a deep convolutional encoder–decoder
architecture for image segmentation. IEEE Trans Pattern Anal Mach Intell 39(12):2481–2495.
https://doi.org/10.1109/TPAMI.2016.2644615
11. Deshpande S (2016) A beginner’s guide to understanding convolutional neural networks. Avail-
able: https://adeshpande3.github.io/adeshpande3.github.io/A-Beginner’s-Guide-To-Understan
ding-Convolutional-Neural-Networks/. Accessed 12 July 2019
12. Akbari M, et al (2018) Polyp segmentation in colonoscopy images using fully convolutional
network. In: 2018 40th annual international conference of the IEEE engineering in medicine
and biology society (EMBC). Honolulu, HI, pp 69–72. https://doi.org/10.1109/EMBC.2018.
8512197
13. Nguyen N, Lee S (2019) Robust boundary segmentation in medical images using a consecutive
deep encoder-decoder network. IEEE Access 7:33795–33808. https://doi.org/10.1109/ACC
ESS.2019.2904094
14. Yu J, Pan H, Yin Q, Bian X, Cui Q (2019) Fully convolutional dense nets for polyp segmentation
in colonoscopy. In: 2019 IEEE 35th international conference on data engineering workshops
(ICDEW). Macao, Macao, pp 306–311. https://doi.org/10.1109/ICDEW.2019.00010
15. Bagheri M, et al (2019) Deep neural network based polyp segmentation in colonoscopy images
using a combination of color spaces. In: 2019 41st annual international conference of the IEEE
engineering in medicine and biology society (EMBC). Berlin, Germany, pp 6742–6745. https://
doi.org/10.1109/EMBC.2019.8856793
An Agile System for FiO2 Regulation
in Ventilators
S. Prabakar, J. Samson Isaac, R. Karthikeyan, K. Porkumaran,
and Ramani Kannan

Abstract Respiratory failure cases are many flowing and filling the ICUs of the
hospitals at any point of time. In cases of respiratory failure, either short-term or long-
term ventilation is required for the patient for survival. Maintaining of O2 saturation
in patients is of prime objective in mechanical ventilation which in turn contributes
for their recovery. Patients connected to the ventilator are constantly monitored for
O2 saturation. Any fall in O2 saturation requires immediate intervention of clinician
to introspect the problem and sort out the issues. One of the main reasons for low
O2 saturation is low value FiO2 settings on the ventilator. This FiO2 setting has to be
increased or decreased manually as per monitored SpO2 analyzer. There are always
possibilities that exist, that patient on ventilators may go unnoticed or get less atten-
tion during emergencies cases entering the ICU. In this project we approached to
automatically blend Air/O2 mixer of an ICU ventilator to maintain a target SpO2
Saturation on the patient, say 98%. Algorithm mechanism is designed in such a way,
continuous FiO2 changes will lead to achieve targeted SpO2 level of the patient.

Keywords SpO2 · FiO2 · ICU ventilator · Automatic intelligence · Blender

correction · Half the difference

S. Prabakar (B)
Sona College of Technology, Salem, India
J. Samson Isaac
Biomedical Engineering, Karunya University, Coimbatore, India
R. Karthikeyan
Dr. N.G.P. Institute of Technology, Coimbatore, India
K. Porkumaran
Sri Sairam Engineering College, Chennai, India
e-mail: [email protected]
R. Kannan
EEE, Universiti Teknologi Petronas, Perak, Malaysia

© The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 725
R. Ibrahim et al. (eds.), International Conference on Artificial Intelligence
for Smart Community, Lecture Notes in Electrical Engineering 758,
https://doi.org/10.1007/978-981-16-2183-3_70
726 S. Prabakar et al.

1 Introduction

Respiratory failure cases are flowing and filling the ICUs of the hospitals every now
and then. This respiratory distress cases range from road accidents to snake bite and
suicide cases. All the above cases require either short-term or long-term ventilation
for their survival. For the fast recovery of patients, maintaining O2 saturation level
is the prime objective of mechanical ventilation.
Normally, Alveolar Ventilation is unconsciously regulated to maintain constant
arterial blood gas tensions, despite variable levels of oxygen consumption and CO2
production. Many drugs and techniques used in anesthesia interfere with control
or mechanics of ventilation, and it is the Anesthetists responsibility to ensure the
adequacy of ventilation during the preoperative period. Equipment related to venti-
lation is consequently of great importance to the Anesthetist and Anesthetic Techni-
cian. Correct use of the equipment relies on a good understanding of basic respiratory
physiology as well as how the individual ventilator operates.
Patients connected to the ventilator are constantly monitored for O2 saturation.
Any fall in O2 saturation requires immediate intervention of clinician to introspect
the problem and to sort out the issues. The main reason for Low O2 Saturation is
low FiO2 settings on the Ventilator. So there are always some possibilities that exist,
such as patients under ventilation are left unnoticed or given less attention that may
lead to even death for patients under ventilation.
FiO2 is Fractional Inspired Concentration of Oxygen. It is a mixer of Air and
Oxygen, blend at a specific proportion, delivered to the patient through ventilator.
Air/O2 Mixer or Blender is one of the main components in any ventilator that mixes
Air and O2 at a particular proportion that is delivered to the patient to maintain his/her
O2 Saturation. The range of mixing is from 21 to 100% of Oxygen as atmospheric
air which always contains 21% Oxygen by nature.

1.1 Scope and Significance of Proposed Work

The main objective is to blend Air/O2 mixer of an ICU Ventilator automatically to

maintain a target SpO2 Saturation on the Patient, say 98%. The SpO2 feedback of
the patient, that is monitored will automatically titrate the FiO2 mechanism on the
Ventilator to achieve the target SpO2 level. The corresponding algorithm is designed
for the continuous FiO2 changes which will lead to attain the intended SpO2 level of
the patient.
According to the algorithmic process, the continuous feedback system will lead
to control the flow of ventilator’s FiO2 according to the patient’s SpO2 . Automating
the Blending process gives a great relief to the clinician in patient management and
allows them to concentrate on other areas that need more time. This is a very safe
method gives mental relief to doctors and nursing staffs and at the same time patient
is also ensured of minimum O2 Saturation for his good life.
An Agile System for FiO2 Regulation in Ventilators 727

The importance of ventilation and oxygenation have analyzed through previous

work analysis. The problems of the manual setting of FiO2 in ventilators which may
lead to many medical accidents due to clinician’s unawareness represented through
many papers. These papers help to analyze about the problems faced while the patient
is under ventilation and also lead to an innovative approach.

1.2 Significance of the Proposed Work

The proposed research work aims to develop an efficient hardware-software approach

to continuously receive and monitor the sensor data received from fingertip based
pulse oximetry which is connected to a mobile app via Bluetooth. The app contin-
uously receives the data and checks it against a standard value. If the value starts
to fall into the abnormal range, a text will be sent so that the person may take their
oxygen supplement before the saturation of oxygen in the blood falls too below the
normal values.
In this project we have used Reflectance pulse oximetry which is an alternative to
transmissive pulse oximetry. The advantage of this method over transmissive method
is that it does not require a thin section of the person’s body and is therefore well
suited to a universal application. Continuous monitoring pulse oximeters are typically
used for long period (1 h or more) oxygen (SpO2 ). Continuous pulse oximeters can
be used for spot checks (quick readings), sleep studies and constant monitoring.

2 Proposed System Design

A complete design and approach for the automatic intelligence and feedback system
are described. It includes the block diagram and algorithmic approach in depth.
Figure 1 shows the input, processing and output modules of the automatic FiO2
regulation in ventilators.
After the analysis of the deviation in the O2 saturation, FiO2 value is newly
computed according to the algorithm designed. The computed FiO2 value is feedback
to the ventilator as an input. Depending on that new value, the SpO2 value will
gradually increase and gives relief to the patient. This is done repeatedly until the
patient attains the normal O2 saturation level.

2.1 Working Process of the Proposed System

The SpO2 sensor detects the oxygen saturation from the patient. That percentage
output as a feedback as well as the set FiO2 from the ventilator are given as inputs to
728 S. Prabakar et al.

Fig. 1 Block diagram of automatic FiO2 regulation in ventilators

the ARDUINO hardware via the input channels. The flow control algorithm is fed
to the ARDUINO hardware to give the relevant output. The algorithm as follows:
Step 1: Set the FiO2 value in ventilators and sense the SpO2 level from the patient
using sensors.
Step 2: Calculate the difference between the optimal and measured SpO2 and
determine the “Half the difference” for blender correction mode.
Step 3: Compute the new FiO2 value for the ventilator to supply on patient.
Step 4: Then again sense the SpO2 and repeat the same steps to attain the desired
SpO2 value on the patient.
According to the above algorithm a new FiO2 value is given as a output from the
ARDUINO hardware. That FiO2 value is given as a feedback to the ventilator. Thus
the SpO2 value is automatically modified according to the new FiO2 value.

2.2 Automatic Intelligence Algorithm for FiO2 Regulation

The algorithm is developed from the SpO2 deviation on the patient under ventilation
in order to compute the desired FiO2 automatically. Figure 2 provides the flow of
FiO2 computation from the obtained values.
Initially we will consider two values such as FiO2 and SpO2 . FiO2 is the fractional
inspired concentration of oxygen, which is the set value in the ventilator (for say 50%)
and SpO2 is saturation of O2 %, which is measured from the patient through the sensor.
An Agile System for FiO2 Regulation in Ventilators 729

Fig. 2 Steps involved in

new FiO2 computation

The deviation between the standard SpO2 value, which is the normal level of oxygen
saturation i.e. 98% and the measured SpO2 , which is theme assured value from the
patient through sensor is determined. The determined value is the difference between
the standard SpO2 value and the measured SpO2 value. According to the difference,
the FiO2 will be changed according to the at present SpO2 value. A new FiO2 value
will be computed as a result it is given to the ventilator as a feedback input. According
to that FiO2 value which is newly computed, the SpO2 value gets modified. So this
progress is done continuously and the corrections are done gradually in loops (Table
1).
We are here taking a situation that a patient connected to a Ventilator with FiO2
set value of 50% results in the patient’s saturation of O2 80%. But, normal SpO2 to
be achieved is 98%.
Here there is a difference between normal and achieved saturation (98–80%)
which is 18%. This is an indication that the FiO2 set on the ventilator is in adequate
and it has to be increased. This increase has to be systematic and it should tend
to increase the O2 saturation towards the normal level of SpO2 . Hence the FiO2
(Air/O2 mixer) is increased to a level that is “Half of the Difference” between target
and achieved SpO2 , i.e., 9%. Now the new (Air/O2 mixing) FiO2 setting has to be
increased from 50 to (50 + 9) 59%.
730 S. Prabakar et al.

Table 1 FiO2 estimation for regulation in ventilators

Set FiO2 in Measured Required Deviation Blender corr. New FiO2
(%) FiO2 in (%) SpO2 in (%) [C-B] mode (changes setting on
reqd. in FiO2 ) vents
A B C D E F
50 50 98 48 24 74
50 55 98 43 21.5 71.5
50 60 98 38 19 69
50 65 98 33 16.5 66.5
50 70 98 28 14 64
50 75 98 23 11.5 61.5
50 80 98 18 9 59
50 85 98 13 6.5 56.5
50 90 98 8 4 54
50 95 98 3 1.5 51.5
50 98 98 0 0 50
50 99 98 −1 −0.5 49.5
50 100 98 −2 −1 49

Let us assume that this increase in FiO2 is taking the patient to a higher O2
saturation of 88% from 80%. Still the Patient’s SpO2 is below normal by a value of
10% (98–88%), that is deficit. This monitored value gives a feedback to the Air/O2
mixer to increase the FiO2 to another “Half of the Difference” that is 5%.Now the
new FiO2 on the ventilator would be 59 + 5% i.e., 64%. This process will tend to
continue till the patient achieves the targeted SpO2 level of 98%.The above process
is the systematic way of maintaining a patient’s O2 saturation through “Intelligent
Feedback System”.

3 Result Analysis and Discussion

This proposed approach delivered results in great relief to the clinicians as it is

automatically titrating the FiO2 . To understand clearly, consider that the patient
connected to an ICU ventilator are managed by clinicians and attended by nursing
staff. We are here taking a situation that a patient is connected to a ventilator with
FiO2 setting of 50% that results in the patient’s saturation of SpO2 80%. Normal
SpO2 to be achieved is 98% and so the algorithm will generate the desired computed
value.
Feedback from the SpO2 monitor is interfaced with the blender of the ventilator
to achieve an optimum SpO2 level of 98%. Any deviation in the SpO2 value will
trigger the blender to increase or decrease the FiO2 to maintain the optimum 98%
An Agile System for FiO2 Regulation in Ventilators 731

of O2 saturation. Patient saturation is automatically maintained to the optimum level

throughout the time period, he/she connected to the ventilator 24 × 7. Human error
or failure to re-titrate the FiO2 to achieve the target SpO2 is avoided. It is ensured
that any point of time patient is not de-saturated. The above process is the systematic
way of maintaining a patient O2 saturation through “Intelligent Feedback System”.
This proposed approach for designing and prototyping using Arduino hardware
is reducing the complexity on the implementation of our real world project since it
delivers the powerful hardware and software technology. In the software simulation
the coding is done accordingly for the implementation of the algorithm for the auto-
matic feedback system. The coding inclusive of LCD display, algorithmic calculation
for automation process and stepper motor implementation for triggering the value
according to saturation level measured from the patient and required delay time is
given for the progress to be done systemically at required interval period according to
the situation of the patient. The LCD display is connected for displaying the current
SpO2 and computed FiO2 value for our verification.
After the software part done then the coding is checked under virtual simulation
board and ensured the coding done is working as per the expectation. Then the coding
is fed into the Arduino board, the Arduino board is connected to the external power
supply, SpO2 sensor for the saturation measurement, stepper motor for triggering of
the value, relay for the switching the stepper motor and LCD display for viewing the
required values. Initially the saturation is measured from the patient according to the
that value, if any fall occurs then the relay board starts to trip and makes the stepper
motor to rotate until it acquires an desired value and gives that value as a feedback
the ventilator.
Thus this blending is done repeatedly until the required SpO2 is obtained. Thus
by using this technology the automatic FiO2 regulation can be achieved. We need not

Fig. 3 Intelligent system for FiO2 regulation in ventilators

732 S. Prabakar et al.

require any manual adjustments of clinician and nurses. Whenever there is a fall in
O2 saturation occurs the ventilator automatically blends the FiO2 settings according
to the patient’s oxygenation (Fig. 3).
Automating the Blending process gives a great relief to the clinician inpatient
management. They can also over-ride this mechanism in case they want to have
direct control of the patient. This mechanism allows the clinicians to concentrate on
other areas that need more time. This is a very safe method gives mental relief to
doctors and nursing staffs and at the same time patient is also ensured of minimum
O2 Saturation for his good life.

4 Conclusion

Thus the proposed work concludes that automating the Blending process gives a great
relief to the clinician in patient management. Thus patient needs not require contin-
uous monitoring and manual correction. They can also over-ride this mechanism in
case they want to have direct control of the patient. The automatic regulation can be
done not only for FiO2 but also for all the blood gas values in different algorithmic
procedures like PaO2 , PaCO2 etc. Thus the system will lead to complete automatic
operation on ventilator without clinicians’ periodical check and also ensures the
patients good care. This mechanism allows the clinicians to concentrate on other
areas that need more time. This is a very safe method gives mental relief to doctors
and nursing staffs and at the same time patient is also ensured of minimum O2
Saturation for his good life.

References

1. Zhou F (2009) Component-based design of software for embedded control systems: the medical
ventilator case study. In: Guan W, Sierszecki K, Angelov C (eds) Embedded software and system,
ICESS’09, international conference. Mads Clausen Institute for Product Innovation, University
of Southern Denmark, Soenderborg, pp 157–163
2. Ionescu CM (2009) Flow controlled artificial ventilation of a COPD patient. In: De Keyser R
(ed) Control conference (ECC). System and Automation Department, Ghent University, Gent-
Zwijnaarde, Belgium pp 2475–2481
3. Jafari MM (2005) Senior member, IEEE, Control systems engineering group, Tyco Healthcare
Puritan Bennett R&D—“Robust Feedback Design for Proportional Assist Ventilation-System
Dynamics and Problem Definition” decision and control, 2005 European Control Conference,
CDC-ECC’05, 44th IEEE Conference, pp 4389–4844
4. An P (2010) High-precision ventilator with compensation algorithm and automatic regulation.
In: Song JT (ed) Database technology and application (DBTA), 2nd international workshop.
Schema of Electronics and Information Engineering, Ningbo University of Technology, Ningbo,
China, pp 1–4
5. Tehrani FT (2007) A new decision support system for mechanical ventilation. In: Engineering
in medical and biology society, 29th annual international conference of the IEEE. California
State University, Fullerton, pp 3569–3572
An Agile System for FiO2 Regulation in Ventilators 733

6. Tzavaras A (2011) Locating of the required key-variables tobe employed in a ventilation manage-
ment decision support system. In: Weller PR, Prinianakis G, Lahana A (eds) Engineering
in medicine and biology society EMBC, Medical Instrumentation Technology Department,
Technological Educational Institute (TEI) of Athens, Greece, pp 112–115
7. Yi W (2009) Fuzzy control for regulating level of pressure support ventilation. In: Zhang Q,
Wang Y, Xiong J (eds) Measuring technology and mechatronics automation, ICMTMA’09,
international conference, vol 2. Mechatronics and Automatic Schema, Natural University of
Defense Technology, Changsha, China, pp 562–565
8. Li Y (2005) Investigation on dynamics of the flow control unit in ventilator systems and its funda-
mental performance limitations. In: American control conference, vol 3. VIASYS HealthCare,
Yorba Linda, CA, USA, pp 2181–2186
Computer-Aided Diagnostic Tool
for Classification of Colonic Polyp
Assessment

Win Sheng Liew, Tong Boon Tang, and Cheng-Kai Lu

Abstract Colorectal cancer is the third most common malignancy and the fourth
leading cause of cancer-related deaths worldwide. This paper presents a combination
of techniques (e.g., pre-processing, transfer learning, principal component analysis,
and support vector machine) to detect the polyp during colonoscopy. In particular, we
carefully choose the pre-trained deep convolutional neural networks (i.e., AlexNet,
GoogLeNet, ResNet-50, and VGG-19) according to their performance extracting
features. A publicly available database, Kvasir, is used to train and test the detection
model. The result indicates that to use ResNet-50 as a pre-trained network provides
the best results among the rest. Our proposed model achieves an accuracy of 99.39%,
and its sensitivity and specificity are 99.39, 99.41, and 99.38%, respectively.

Keywords Colorectal cancer · Transfer learning · Polyp · Deep convolutional

neural network · Support vector machine

1 Introduction

Colorectal cancer (CRC) is one of the leading causes of cancer deaths based on
the analysis of the American Cancer Society (ACS) as the number of cancer deaths
was increased year by year from 1997 to 2015 [1]. The small and tiny polyp might
be missed [2], which caused the chance of getting converted into a giant neoplastic
polyp. Hence, the discovery of colonic polyps is essential as the polyp can be inhibited
by removing it and reduce the incidence of CRC, whereas the detection of malignant
polyps at an early stage may result in a curative intervention [3]. However, it is not
as difficult to detect polyp as before [1] because of the elegant automated computer-
aided diagnosis (CAD). Those CAD tools can detect the polyp and provide polyp
location depending on the curvature of polyps. Several CAD systems have been
developed for various cancer diseases such as prostate cancer, brain cancer, lung

W. S. Liew · T. B. Tang · C.-K. Lu (B)

Department of Electrical and Electronic Engineering, Universiti Teknologi PETRONAS, 32610
Seri Iskandar, Malaysia
e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 735
R. Ibrahim et al. (eds.), International Conference on Artificial Intelligence
for Smart Community, Lecture Notes in Electrical Engineering 758,
https://doi.org/10.1007/978-981-16-2183-3_71
736 W. S. Liew et al.

cancer, etc. These CAD systems obtain the images through the hardware renders and
save them as digital movies using image processing software. Therefore, doctors
can accurately carry out the diagnosis and decision based on the CAD system’s
information.
With the CAD scheme’s rapid development, the detection system is continuously
optimized and improved with AI techniques. AI is considered as a machine emulation
with human thinking. Thus, deep learning (DL) applications in AI allow the systems
to become intelligent, learning, and self-organizing [4]. Pre-processing is vital for
every detection system, as every image has different sizes and features (i.e., color,
density, shape, and texture) [5]. The images need to be pre-processed to reduce the
image’s noise and degradation; normalization is required to resize the images into the
same size. Besides, the images need to undergo feature extraction and selection to
carry out classification task with a different type of classifiers such as ensemble clas-
sifier [6], support vector machine (SVM) [7], Naïve Bayes classifier [8, 9], random
forest [9] and K-Nearest Neighbors (KNN) [9].

2 Material and Method

In the research, a publicly available database, Kvasir, is used to train, validate, and
test the detection model. All images are color colonoscopy images in JPG format
with a resolution of 720 × 576 pixels.
A total number of 2000 images in the dataset consisted of 1000 normal cases (non-
polyp) and 1000 abnormal cases (polyp). The colonoscopy images are split into 70%
for training and 30% for testing. The model is developed through the MATLAB
platform on a Windows 10 64-bit operating system with Intel® Core i5-2430 M
CPU @ 2.40 GHz, 8 GB RAM, and NVIDIA GeForce GT 525 M GPU. For this
research work, each step is illustrated in Fig. 1.

Fig. 1 Overview of the research methodology

Computer-Aided Diagnostic Tool for Classification … 737

2.1 Pre-processing

During colonoscopy screening, the images will be affected due to the visual condition
in the bowel. This is because there is no ambient light in the bowel, but only the
capsule on-board light as the illuminating source. Under this condition, the images
captured by the camera will have some artifacts, vignettes, and illuminations [5, 10].
To enhance the polyp feature in the image, the thresholding method is used to separate
the polyp from the background, which indicates the image consists of two classes of
pixels (foreground and background) [11]. A color image has three channels, which
are red (R), green (G), and blue (B); the threshold values of R and G channels are
scaled to 0 if the values are below 50 and 15, respectively. On the other hand, the B
channel is not utilized for thresholding due to pixel values of polyp and background
are almost the same.
Furthermore, the images needed to resize into the same scale through normaliza-
tion to feed into the neural networks for training. Nevertheless, different networks
have different input sizes; AlexNet can only process the images that are 227 × 227
pixels, while GoogLeNet, ResNet-50, and VGG-19 process images with 224 × 224
pixels. The normalization equation is computed as follows:

y−y
y  (Resi zed I mage) = (1)
σ
where y is the original pixel vector, y is the mean of the pixel vector, and σ is its
standard deviation.

2.2 Transfer Learning

Deep learning (DL) is potent as it helps us a lot in complex tasks. However, it is not
easy to create a DL model because it requires generous data to train the deep neural
network from scratch. To resolve this issue, transfer learning (TL) is a powerful and
effective solution to limited data, especially in the medical research field [12].
In addition, TL can improve the accuracy of the network while reducing its
training time [13]. The original architecture of pre-trained networks such as AlexNet,
GoogLeNet, ResNet, and VGG can be used and fit into own task by replacing the last
three layers with the new specific task from the database. Besides, the neural networks
can learn from training to progress their proficiency [14]. Nevertheless, different
neural networks have various performances in terms of accuracy and training time
due to the number of deeper layers and the training database’s size.
738 W. S. Liew et al.

2.3 Classification

From the view of machine learning, one of the classification methods is supervised
learning. Support vector machine (SVM) is a classifier with supervised learning [15],
and it provides remarkable performance and accuracy with less computation power.
SVM divides data into two classes (normal and abnormal cases) using hyperplanes
in a high dimensional feature space.
However, SVM has a higher dimensionality, which may lead to model overfitting.
Therefore, principal component analysis (PCA) is utilized to reduce data dimen-
sionality [16] to the maximum extent. To achieve dimensionality reduction, PCA’s
feature elimination is one of the techniques as it remains interpretability of variables
[17].

3 Result and Discussion

3.1 Performance of Models

The models are evaluated based on the confusion matrix with true positive (TP),
true negative (TN), false positive (FP), and false negative (FN), where TP: a polyp
is detected in a frame that contains a polyp; TN: No polyp is detected in a frame
without a polyp image; FP: a polyp is detected in a frame without a polyp image;
FN: A polyp is missed in a frame that contains a polyp. The accuracy, sensitivity,
and specificity are calculated from the confusion matrix as follow:

TP +TN
Accuracy( ACC) = (2)
T P + T N + FP + FN
TP
Sensitivit y(T P R) = (3)
T P + FP
TN
Speci f icit y(T N R) = (4)
T N + FP

Table 1 Performance of each neural network

CNNs Accuracy Sensitivity Specificity AUC
AlexNet 0.9635 0.9645 0.9625 0.9935
GoogLeNet 0.9727 0.9704 0.9750 0.9964
ResNEt-50 0.9939 0.9941 0.9938 0.9988
VGG-19 0.9791 0.9645 0.9938 0.9986
Computer-Aided Diagnostic Tool for Classification … 739

The summary of results is provided in Table 1. Within these four deep neural
networks, ResNet-50 is better than the rest of the networks with an accuracy of
99.39%, sensitivity of 99.41%, and specificity of 99.38%.
In Table 1, the area under receiver operating characteristics (AUROC) is measured
based on the receiver operating characteristics (ROC) curve in Fig. 2. The ROC curve
is a graph of true positive rate (TPR) versus false positive rate (FPR) to show a clas-
sification model’s performance at all the classification thresholds. TPR is sensitivity;
while FPR is calculated as follows:

F P R = 1 − Speci f icit y (5)

Although every neural network has different computation time for data pre-
processing and training, but the exact training iteration time is still depending on
the hardware or the size of mini-batch set. Figure 3 shows the graph of accuracy
versus training time for four neural networks. AlexNet is the fastest network, but it
has the lowest accuracy. In contrast, VGG-19 takes the longest time for training and
its accuracy is still lower than ResNet-50.
Due to the network complexity, the number of convolution layer directly relates to
the performance of networks. ResNet-50 has the most convolution layers, and it tends
to overfit the training data. As a result, it has the highest accuracy for the colonic polyp
detection task and higher training time than the AlexNet and GoogLeNet. In contrast
with training time, the number of learnable parameters plays an important role.

Fig. 2 ROC curve for each neural network

740 W. S. Liew et al.

ResNet-50
Accuracy (%)

GoogLeNet
VGG-19

AlexNet

Relative Prediction Time Using GPU

Fig. 3 Accuracy against relative training speed for each neural network

VGG-19 takes the longest time for the training because it has the biggest learnable
parameters among the four networks.
Based on the discussion above, ResNet-50 is chosen for this research work. The
results obtained by using the proposed method are compared with the state-of-the-art
techniques. The comparison results have been made with previous works in Table 2.
Our proposed method has better performance evaluation scores than other studies. In
this way, the proposed method increases CAD systems’ success rate with the most
effective medical image detection method for the colonic polyp.

Table 2 Comparison of our

Name of study Accuracy (%) Sensitivity Specificity
results with previous studies
(%) (%)
Bae and Yoon 89.00 95.00 81.00
[6]
Xu and 91.00 95.00 90.00
Suzuki [17]
Fan et. al [18] 93.70 95.40 74.70
Fiori et. al 94.00 95.00 88.00
[19]
Proposed 99.39 99.41 99.38
study
Computer-Aided Diagnostic Tool for Classification … 741

Fig. 4 Sample of colonoscopy images. a With polyp. b Without polyp

3.2 Graphical User Interface (GUI)

After the training process, the network’s parameters are saved as a MAT file so that
there is no training required again during the testing of the image. For the testing,
a graphical user interface (GUI) is created to load the images, and it is designed as
in Fig. 1. There are two buttons available in the GUI, which are machine learning
and load image. First, the user requires to choose a model trained by four different
CNN architectures. Once a network is chosen, all the parameters will be shown in
the GUI, and it is ready to load the image for testing. Then, the classification result
will be shown in the identified class column of the GUI.

3.3 Dataset

From the dataset, there are two categories of images, which are with polyp and
without polyp. These 2000 colonoscopy images have different visual quality and
features such as the brightness, distribution of pixel intensity, position, and polyp
size. To prove the diversity of data, we have various types of images in Fig. 4, which
allows the detection model to learn with multiple features.

4 Conclusion

To conclude, an automated polyp detection system was developed using pre-

processing, principal component analysis, transfer learning as a feature extractor,
and SVM as a classifier. The presence of polyps is detected with an accuracy of
99.39% from the Kvasir publicly accessible database. The sensitivity and speci-
ficity of polyp detection are 99.41%, and 98.75, 99.38%, respectively. The methods
developed so far are therefore promising as the basis for subsequent expert human
assessment.
742 W. S. Liew et al.

Acknowledgements This work was supported in part by Ministry of Higher Education, Malaysia,
under Grant FRGS/1/2020/TK0/UTP/02/23, and partially supported by YUTP-Fundamental
Research Grant (015LC0-002).

References

1. Sareena, Mittal A, Kaur M (2016) Computer-aided-diagnosis in colorectal cancer: a survey

of state of the art techniques. In: 2016 international conference on inventive computation
technologies (ICICT), pp 1–6
2. Tajbakhsh N, Chi C, Gurudu SR, Liang J (2014) Automatic polyp detection from learned
boundaries. In: 2014 IEEE 11th international symposium on biomedical imaging (ISBI), pp
97–100
3. Paik DS, Beaulieu CF, Rubin GD, Acar B, Jeffrey RB, Yee J et al (2004) Surface normal
overlap: a computer-aided detection algorithm with application to colonic polyps and lung
nodules in helical CT. IEEE Trans Med Imaging 23(6):661–675
4. Bose BK (1994) Expert system, fuzzy logic, and neural network applications in power
electronics and motion control. Proc IEEE 82(8):1303–1323
5. Gueye L, Yildirim-Yayilgan S, Cheikh FA, Balasingham I (2015) Automatic detection of
colonoscopic anomalies using capsule endoscopy. In: 2015 IEEE international conference on
image processing (ICIP), pp 1061–1064
6. Bae S, Yoon K (2015) Polyp detection via imbalanced learning and discriminative feature
learning. IEEE Trans Med Imaging 34(11):2379–2393
7. Romain O, Histace A, Silva J, Ayoub J, Granado B, Pinna A, et al (2013) Towards a multimodal
wireless video capsule for detection of colonic polyps as prevention of colorectal cancer. In:
13th IEEE international conference on bioinformatics and bioengineering, pp 1–6
8. Gautam A, Singh P, Raman B, Bhadauria H (2016) Automatic classification of leukocytes
using morphological features and Naïve Bayes classifier. In: 2016 IEEE region 10 conference
(TENCON), pp 1023–1027
9. Devika R, Avilala SV, Subramaniyaswamy V (2019) Comparative study of classifier for
chronic kidney disease prediction using Naive Bayes, KNN and random forest. In: 2019
3rd international conference on computing methodologies and communication (ICCMC), pp
679–684
10. Mamonov AV, Figueiredo IN, Figueiredo PN, Tsai YR (2014) Automated polyp detection in
colon capsule endoscopy. IEEE Trans Med Imaging 33(7):1488–1502
11. Jeyavathana B, Ramasamy B, Pandian A (2016) A survey: analysis on pre-processing and
segmentation techniques for medical images. Int J Res Sci Innov (IJRSI)
12. Shie C-K, Chuang C-H, Chou C-N, Wu M-H, Chang E (2015) Transfer representation learning
for medical image analysis, pp 711–714
13. Felzenszwalb PF, Girshick RB, McAllester D, Ramanan D (2010) Object detection with
discriminatively trained part-based models. IEEE Trans Pattern Anal Mach Intell 32(9):1627–
1645
14. Ansari A, Bakar AA (2014) A comparative study of three artificial intelligence techniques:
genetic algorithm, neural network, and fuzzy logic, on scheduling problem. In: 2014 4th inter-
national conference on artificial intelligence with applications in engineering and technology,
pp 31–36
15. Ratheesh A, Soman P, Nair MR, Devika RG, Aneesh RP (2016) Advanced algorithm for polyp
detection using depth segmentation in colon endoscopy. In: 2016 international conference on
communication systems and networks (ComNet), pp 179–183
16. Ibrahim MFI, Al-Jumaily AA (2016) PCA indexing based feature learning and feature selection.
In: 2016 8th cairo international biomedical engineering conference (CIBEC), pp 68–71
Computer-Aided Diagnostic Tool for Classification … 743

17. Xu J-W, Suzuki K (2014) Max-AUC feature selection in computer-aided detection of polyps
in CT colonography. IEEE J Biomed Health Inform 18(2):585–593
18. Fan L, Song B, Gu X, Liang Z. Semi-supervised graph embedding-based feature extraction
and adaptive kernel-based classification for computer-aided detection in CT colonography. In:
2012 IEEE nuclear science symposium and medical imaging conference record (NSS/MIC),
pp 3983–3988
19. Fiori M, Musé P, Aguirre S, Sapiro G (2010) Automatic colon polyp flagging via geometric and
texture features. In: 2010 Annual international conference of the IEEE engineering in medicine
and biology, pp 3170–3173
Live Obstacle Detection Audio System
for Visually Impaired People

P. Anitha, C. Vinothini, J. Priya, and S. Guhan

Abstract Object detection is the complex task for the visually impaired people
when they are on move. This paper proposes smartphone’s capability more useful
for visually impaired people to overcome day to day problems with live navigation.
Depth sensing camera is used to differentiate foreground objects from background
objects which is used to place real world objects in different layers. As soon as the
user moves to the particular layer the fore coming objects are detected and notified
to the user by audio instructions. Information from the internet can be collected
using the mobility services. The headphone with mic is used by the system for
receiving user’s voice input. Emergency tasks like sending SMS and location are
sent to the subject’s guardian while the subject is in danger with the help of player.
This concept allows visually impaired people to walk both indoors and outdoors
independently. Experimental results obtained show that the navigation system eases
the user’s travelling experience in complex indoor and outdoor environments. Hence,
it serves as challenged people-friendly device for helping the visually impaired people
to travel hassle-free and safely. The goal is to provide an effective means of navigation
for the visually-impaired.

Keywords Obstacle detection · Audio navigation · Visually impaired

1 Introduction

The number of visually impaired people is on the raise over the recent years. Report of
the world health organization (WHO) claims about 253 million people live with vision
impairment. Among them, 36 million people are visually impaired and 217 million
have moderate to severe vision impairment. Most of the visually impaired people

P. Anitha (B) · C. Vinothini · S. Guhan

Department of Computer Science and Engineering, Dr. N.G.P Institute of Technology,
Coimbatore, Tamilnadu, India
J. Priya
Department of Information Technology, Bannari Amman Institute of Technology,
Sathyamangalam, Tamilnadu, India

© The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 745
R. Ibrahim et al. (eds.), International Conference on Artificial Intelligence
for Smart Community, Lecture Notes in Electrical Engineering 758,
https://doi.org/10.1007/978-981-16-2183-3_72
746 P. Anitha et al.

is aged 50 or older and live in the developing countries struggling when they are
alone. Every day these visually impaired people face problems in understanding and
interacting with the surroundings, particularly those that are unfamiliar. It is hard for
a visually impaired person to go out alone and there are not many available products
that can assist them. However, research has been going on for decades for developing
an effective device for visually impaired people. In this paper, a device with new
features that can aid in multiple tasks while maintaining low development cost.
The device can easily guide the visually impaired people and is able to give proper
directions. The most valuable thing for a disabled person is gaining independence. A
blind person can lead an independent life with some specifically designed adaptive
things for them. There are lots of adaptive equipment that can enable a blind person
to live their life independently but they are not easily available in the local shops or
markets. A blind person needs to hunt and put much effort to get each equipment
that can take them one step closer towards independence.

2 Related Works

Various methodologies have been proposed for the welfare of vision-impaired people
using the concept of AI and ML. Here are some of them in brief.
Li et al. [1] proposed an approach pioneering effort of indoor robotic naviga-
tion, NavChair used sonar sensor to detect objects. A vector field histogram (VFH)
algorithm and later Minimal VFH were proposed to provide effective sonar-based
obstacle detection and avoidance. For updating a local map with detected obstacle
in the form of uncertainty grids, VFH used sonar readings. It was used to calcu-
late a polar histogram to represent the geometric information of the obstacles. At
last, NavChair found a direction with minimum VFH cost and followed it to bypass
obstacles. The RGB-D indoor navigation system. An emerging research focus for the
robotics community to improve the mobility of blind and visually impaired people is
Intelligent assistive navigation. Numerous studies have been carried out for indoor
navigation on mobile devices.
Bai et al. [2] proposed a system that includes a depth camera for acquiring
the depth information of the surroundings, an ultrasonic rangefinder consisting of
an ultrasonic sensor, a MCU (Microprogrammed Control Unit) for measuring the
obstacle distance, an embedded CPU (Central Processing Unit) board acting as main
processing module, which does such operations as depth image processing, data
fusion, AR rendering, guiding sound synthesis, etc., pair of AR glasses to display
the visual enhancement information and an earphone to play the guiding sound.
Feng Lan, Shanghai Jiao Tong University, Shanghai, China [3] proposed a smart
glass prototype system based on Intel Edison. The Intel Edison is a tiny computer
offered by Intel as a development system for wearable devices. The board’s main SoC
is a 22 nm Intel Atom “Tangier” (Z34XX) that includes two Atom Silvermont cores
running at 500 MHz and one Intel Quark core at 100 MHz (for executing RTOS Viper
OS). The SoC has 1 GB RAM integrated on package. There is also 4 GB eMMC
Live Obstacle Detection Audio System for Visually Impaired People 747

flash on board, Wi-Fi, Bluetooth 4 and USB controllers [4]. Edison is powerful in
computing but small in shape that makes it perfect to build a wearable device. Public
signs in cities can act as guides for people. Following the public signs, people can find
public infrastructure, such as public toilets, bus stations, subway stations, hotels and
so on. So finding and recognizing the public signs outsides would help someone who
is visually impaired or visually impaired gain increased independence and freedom.
Therefore, they have implemented an application of public signs recognition in this
developed smart glass system. This application can automatically detect, analyse and
recognize all kinds of public signs around the visually impaired and visually impaired
and give corresponding voice hints through wireless bone conduction headphones.
With the help of this smart glass system, the visually impaired may find bus stations,
subway stations, restaurants, hotels and so on. The visually impaired may even go
for a trip alone.
Bourne et al. [4] proposed a project to solve instrumentation problems encountered
by visually impaired employees or potential employees who are not suited to (nor
do they require) complex microcomputer or synthetic speech circuitry for practical
solutions. A common problem encountered by custom cabinetmakers, as well as
individual homemakers, is finding the structural members within walls (“studs”)
onto which heavy shelves and wall hangings can be mounted. There have been
many devices for finding “studs,” but these are traditionally visual instruments. They
have successfully modified a commercial stud finder to present auditory feedback
to the visually impaired cabinetmaker and handyman”. The Zircon “Stud Sensor”
(also marketed by Radio Shack as the “Archer Stud finder”) is a relative capacitance
meter that senses hidden structures by their capacitance effect; it presents these
relative indications to a sighted user via a column of LEDs. By reverse engineering, a
“relative capacitance” analog signal was found which could be used to drive a voltage-
controlled oscillator. Thus, with their modified instrument the visually impaired user
hears a distinct rise in pitch when the capacitance increases—which happens when
a stud is found. Four prototypes are being made for field trials, and several visually
impaired individuals who have tested the unit have found it effective and easy to use.
Rentschler et al. [5] proposed a system to test the walker when run on a two-drum
test machine. The walker was then run through the functionality test and visually
inspected for cracks or damage. This approach helps the visually impaired navigate
hassle free through streets.
Elaiyaraja et al. [6] proposed a system in which the prototype consists of an optical
glass with head-mounted camera that captures images and offloads the computation
to the database. The database is built using Microsoft Azure API and consists of
details of all known persons to the visually impaired. Based on the computations,
the database performs facial recognition and identification. If it is a face that already
exists in the database, then the person’s name and face features will be returned.
If it is an unknown face, then it returns that the person is unknown to the visually
impaired. With the help of identification information obtained from the unknown
person, he/she can be added to the database. The Text to Speech API converts the
input text into human-sounding custom speech and returned as an audio file.
748 P. Anitha et al.

Some of the implemented guide systems like smart guide proposed by Manufali
[7] gives the user a prototype of an intelligent guide for the blind person is successfully
designed, implemented and tested. The prototype device is facilitating the movement
of blind person by warning him/her about any nearby obstacles in order to help
him/her during daily activities. The guidance will be provided in the form of audio
instructions through the headset and based on real-time situation for both indoor and
outdoor environments. But it cannot detect what is against the user and it is also low
battery storage which cannot long last for long duration.

3 Proposed Framework

3.1 Overview

The Audio navigation system proposed can be implemented in the smart phone
itself. It is more battery efficient and portable. A dedicated voice assistant directs
the visually impaired with help of google maps. Notifications will be given through
headphones via voice commands so that it cannot confuse the user. It will guide the
person during walk so one can make sure he is stable while stepping. This requires
the internet connection; however, most of the processing is done offline. This system
also allows the user to do basic mobile functionalities such as calling and messaging
through voice commands. Figure 1 represents the overview of the flow in the system.

3.2 Proposed System

Obstacles in the visually impaired’s pathway are captured with live relay using the
phone’s camera system shown in Fig. 2. The data is then processed and used in
guiding the visually impaired via an obstacle free path. This is implemented with
the help of TensorFlow powered API. Initial process of this audio navigation system
is to find the obstacles before the user. It is required to guide the user how to avoid
them without encountering. A depth-sensing camera with IR sensor in a mobile can
provide us the data how far an object is away from the camera. Input data is processed
through tensor flow and open-cv. Detected objects with its accuracy of detection will
be the outcome of this module.
Depth sensing camera helps with identification of objects in layers which conveys
which object is to encountered first. The camera also detects the floor using the floor
level deduction which helps to identify pits and peaks. Floor detection is done by the
barometer which is installed default in the phone.
Audio navigation is required to notify the person “How to move” and “What is
before them”. First objective of this module is filter what to be notified and what
must be neglected.
Live Obstacle Detection Audio System for Visually Impaired People 749

Fig. 1 System overview

diagram

Fig. 2 Mobile camera

system

The input to this module will be from the text output of camera vision. The detected
objects distanced detected by depth cam, which is shown in Fig. 2, and the object
itself will be formed, as a sentence through a program. It will be played as audio to
the user. The text to audio conversion is done by gTTS engine service. In terms of
storage, data is stored internally.
750 P. Anitha et al.

Objective of this module is to get users audio input and direct them to the places
where they like to move. Audio input is processed as text by gTTS module. User’s
location is obtained from the device as input. The input is given to “DIRECTIONS
API” to find the directions for their given place from the user’s location. The reply
from the API is used by the user to guide them all the way. Nearby API locates the
nearby location to their location.
A visually impaired person downloads the developed application in his/her smart-
phone and performs the initial setup with the guardians help. The person then connects
the headphones to the smars application helps to navigate to the desired destination
with the help of our assistant providing audible directions. In addition to this, the
assistant also provides recommendations for nearby eateries, shops, hospitals etc. In
order to use these features the person has to long press the headphone button.

3.3 Applications

The main task of the application is to detect lively the obstacles like roadcrosses,
pits, persons and other objects. It can also identify an object when the user captures
it. Once it captures the objects it conveys the user by the voice commands. If an user
needs to know the nearby locations like hospital it can be provided by asking the
system. It provides the hospitals nearby. If the user needs any other locations like
hotels or shops it provides them those location details within 1 km. Basic mobile
functionalities such as call, message and checking battery status. It locates one to
their home with live routing. It says the way to their home in steps. In an emergency
it can send SMS to a friend or doctor with the current location details. It can also
make call to their Friend or doctor immediately once the system hears the keywords.
It allows the visually impaired navigate through cities independently. It can be used
for live tracking globally with some constraints. This system will be trained for an
individual user which gives them more accessibility.

4 Experimental Result

The camera analyzes and recognize objects. It then conveys this information to the
visually impaired via the headphone. As an example consider the above two images.
Figure 3 shows how the camera recognize the human faces along with the certainty
that the object is a human. Similarly, Fig. 4 shows how camera recognize the bottle.
Figure 5 shows the voice assistant interface which will converted as background
process. The interface consist of five parts: Results, Response, Get_gps, Live tracking
and input. In input field the user’s voice is taken as input. Get_gps gives the user
location. Input can be given through voice or text. Live tracking can give the obstacle
detection lively when the user points at any direction. Audible reply is given by the
Live Obstacle Detection Audio System for Visually Impaired People 751

Fig. 3 TensorFlow vision

detecting persons

Fig. 4 TensorFlow vision

detecting bottle

assistant and also the response field will carry the reply. The app logo within an app
drawer is shown in Fig. 6.
Most of the audio comments are given by google assistant’s default voice with
the pitch rate and speed of 20%.It can be adjust in accordance with google account
maintenance (Table 1).
On real time comparison between the existing systems and our system we came
up with a prediction amount shown in the below matrix. The system detection rate
of persons is most accurate as it uses object detection principle for classification of
objects. The existing system 1 is down in accuracy because it uses ultra-sonic sensors.
These sensors may miss the person’s body when there gaps in between them. The
existing system 2 uses an object detection which is less accurate to recognize a
person. Our system and existing system 2’s detection rate of glass objects is slightly
less because of the light diffraction whereas ultrasonic sensors don’t miss glass
object because of light. Pits detection is completely missed by existing system 1 as
it considers it as far away object. Our system considers it by using depth sensing
752 P. Anitha et al.

Fig. 5 Application interface

which detects floor level and varies a pit. Whereas in system 2 it misses the accuracy
without using floor level detection. We use TensorFlow most accurate algorithm and
online api for detecting road crosses where as other systems lacks it.

5 Conclusion

A majority of the proposed system involve the use of the 3d glass or an obstacle
detection system to assist the visually impaired for navigation. The systems are often
complex to operate and expensive to implement. This paper describes an approach
similar to ones discussed but differs in terms of simplicity and cost. The system
proposed requires basic smartphone, dedicated service with internet connectivity
and a headset to achieve live navigation for visually impaired effectively. Obstacle
detection and audible guidance using voice assistant gives an easier way of access.
The system developed is portable in the sense that it requires a smartphone with
internet connectivity. Thus, a cost effective and simplistic solution is made possible.
Live Obstacle Detection Audio System for Visually Impaired People 753

Fig. 6 Application icon in

the app-drawer

Table 1 Comparative analysis on performance of various supportive systems for visually impaired
Obstacles Predicted (%)
Smart audio system Smart guide [7] Smart guiding glasses [2]
Actual (%) Persons 95 75 85
Glass 60 80 60
Pits 85 2 80
Road cross 92 0 20
Average (%) Overall 83 39.25 61.25

Requires internet connection at all times to ensure proper functionality. Problems of

usage in places with limited network connectivity. Sometimes, radiation from mobile
device imposes health hazards for the life.
754 P. Anitha et al.

6 Future Enhancement

Live detection is to be made more accurate and fully offline capable. This is similar
to a person walking beside the visually impaired person and providing direction.
This approach will be implemented with machine learning. The image processing is
done using TensorFlow. The ultimate goal is to guide the visually impaired through
cities effectively and hands-free.

References

1. Li B, Munoz JP, Rong X (2019) Vision-based mobile indoor assistive navigation aid for blind
people. IEEE Trans Mobile Comput 18(3):702–714
2. Bai J, Lian S, Liu Z, Wang K, Liu D (2018) Smart guiding glasses for visually impaired people
in indoor environment. IEEE Trans Consum Electron 63(3):258–266
3. Lan F, Zhai G, Lin W (2015) Lightweight smart glass system with audio aid for visually impaired
people. In: TENCON, IEEE, region 10 conference
4. Bourne RRA, Flaxman SR, Braithwaite T, Cicinelli MV, Das A, Jonas JB, et al (2017) Vision
Loss Expert Group. Magnitude, temporal trends, and projections of the global prevalence of
visually impairedness and distance and near vision impairment: a systematic review and meta-
analysis. Lancet Glob Health 5(9):e888–e897
5. Rentschler AJ, Cooper RA, Blasch B, Boninger ML (2003) Intelligent walkers for the elderly:
performance and safety testing of VA-PAMAID robotic walker. J Rehabil Res Dev 40(5):423–
432
6. Elaiyaraja, Aishwarya, Akshaya, Monikha M (2019) AJNA—a smart assistance for the visually
and cognitively impaired. Int J Adv Res Electric, Electron Instrument Eng
7. Manoufali M, Aladwani A, Alseraidy S, Alabdouli A (2011) Smart guide for blind people
international conference and workshop on current trends in information technology
Reducing the Aleatoric Uncertainties
of Failure Prediction Using Singular
Value Decomposition

Ahmad Kamal Mohd Nor, Srinivasa Rao Pedapati, and Masdi Muhammad

Abstract Uncertainty is a vital indicator in evaluating the prediction of real-

world deep learning models. Without uncertainty, important decision-making process
linked to safety, security and investment cannot be executed correctly by the users.
While probabilistic deep learning methods have shown promise in quantifying predic-
tion uncertainty, very few works in managing these uncertainties can be found in the
literature. In failure prediction especially, it is favourable that the model produces the
lowest level of prediction uncertainty, i.e., the highest level of confidence in order
to initiate any degree of problem-solving mechanism. In this work, we present a
technique to reduce Aleatoric uncertainty associated with noisy data using Singular
Value Decomposition (SVD). Sensor data from industrial assets are presented in
SVD matrix form and higher SVD modes more susceptible to contamination are
eliminated to denoise the data. We compare the uncertainty level between the orig-
inal dataset with the ones denoised by SVD and several known denoising methods
in a Remaining Useful Life (RUL) prediction problem. Our results show that SVD
denoising treatment outperforms the other denoising methods in reducing prediction
uncertainty and improving prediction performance.

Keywords Uncertainty · Prognostic · PHM · RUL · CMAPSS

1 Introduction

Long term exploitation of aerospace assets is one of the goals sought by the industry.
Maintenance efforts to ensure operational longevity have evolved steadily from
simple corrective maintenance, to preventive measures and more recently to a more
systematic framework called Prognostic and Health Management (PHM). PHM aims
to optimize maintenance activities by accurately predicting failure event. It has the
effect of minimizing maintenance operations, reducing workload and decreasing

A. K. M. Nor (B) · S. R. Pedapati · M. Muhammad

Universiti Teknologi Petronas, 32610 Perak, Malaysia
e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 755
R. Ibrahim et al. (eds.), International Conference on Artificial Intelligence
for Smart Community, Lecture Notes in Electrical Engineering 758,
https://doi.org/10.1007/978-981-16-2183-3_73
756 A. K. M. Nor et al.

maintenance interventions in hazardous places, resulting to significant maintenance

cost reduction [1–4].
Remaining Useful Life (RUL) estimation, where the health of engineered system
is predicted, is one of the core activities in PHM [5, 6]. In this domain, we witness
the rise of data-driven prognostic in the last 2 decades, fuelled by its simplicity
of deployment, the ever-increasing volume of data and the advancement wireless
technology. Amongst all, deep learning has emerged as one of the most powerful
tools due to its ability to map nonlinear relation of the system’s variables.
Deep learning methods can be categorized as either being frequentist or proba-
bilistic in nature [7, 8]. Looking into RUL prognostic angle specifically, frequentist
methods predict point estimates RUL while the probabilistic ones predict the distri-
bution or intervals of RULs [9, 10]. The latter thus has the advantage of predicting
with uncertainties. In real world application, where ground truth RUL is non-existent,
the only indication to trust the model’s prediction is uncertainty quantification. This
is especially important when manpower or investment decision need to be made
following a failure prediction.

1.1 Uncertainty Reduction

Aleatoric uncertainty in deep learning occurs due to noise or stochasticity captured

in the input data, which is a common situation in real world applications [11–13].
This uncertainty thus capture noise inherent to the observation. In an ideal failure
prognostic model, the uncertainty should be higher when the prediction deviates
from the ground truth and stays low in the case of correct prediction.
In real world sensor signal, it is hard to differentiate between the noise and the
clean data. Though many denoising and smoothing algorithm have been proposed,
most of them need the intervention of domain expert to choose the appropriate level
of denoising and smoothing that minimizes the risk of information loss.
To this end, SVD has emerged as a formidable tool in denoising data without the
intervention of domain expert [14–17]. By representing noisy data in SVD matrix,
we can eliminate the higher SVD modes which are more prone to contamination than
the lower ones, retaining only the clean SVD modes that form a clean signal.
In this paper, we present an uncertainty reduction technique based on SVD signal
denoising. We demonstrate this ability using CMAPSS gas turbine dataset. To the
best of our knowledge, this is the first paper of its kind dedicated in investigating
the ability of denoising methods in reducing deep learning uncertainty. This work is
important as in real world condition, having the lowest uncertainties, i.e., the highest
confidence in failure prediction is the goal for an accurate decision making. This
paper is organized as follows. In the next section, related publications are reviewed.
In Sect. 3, we elaborate the SVD denoising methodology and performance evaluation.
Then, in Sect. 4, a case study is described before we present the results and discussion
of our work in Sect. 5. Finally, we give our concluding remarks in the last section.
Reducing the Aleatoric Uncertainties of Failure Prediction … 757

2 Related Work

SVD has become a standard tool in denoising data in general due to its ability to
eliminate singular values residing in noisy subspaces and retaining only the clean
ones. In Wang et al [18] a combination of Wavelet Transforms and SVD is used
to denoise image data. The wavelet transform is used to divide the image into
low and high frequencies divisions. Then SVD is employed to denoise the high
frequency parts to enable a clean image reconstruction. In Zhang et al. [19], a two-
division recursion SVD is proven to perform best in improving contaminated signals
amongst the various SVD model variants. The methods also outperform other signal
decomposition techniques such as Empirical Mode Decomposition (EMD), ensemble
EMD (EEMD) and Variational Mode Decomposition (VMD). SVD is deployed on
matrix of image patches in He et al. [20], and by doing so, separating the signal
subspace from the noise subspace. Then, SVD denoising is performed on each group
of similar patches that were previously combined as training samples, producing
leading singular vectors candidates that represent the corresponding patch. Singular
vectors coming from noise populated samples and noise subspace are then elimi-
nated, thus leaving the clean singular vectors that form the local basis of each patch.
In Wang et al. [21], the Singular Value Difference (SV) between the noisy and clean
image is exploited to denoise image in wavelet domain. SVD is applied to denoise
SVs in high frequency parts of the noisy image.
Apart from image denoising, SVD also performs well with general data. In Liu
et al. [22], seismic data is denoised by SVD in the frequency domain. Raw seismic
data is first presented in the frequency domain by Fourier transform. Then SVD is
employed to decompose the frequency spectra into eigen images, which will be then
cleaned by selecting clean singular values. In Zhao et al. [23], a reweighted Singular
Value Decomposition (RSVD) framework which has the advantage of improving
weak features while denoising is proposed. The periodic modulation intensity, an
indicator that represents the quality of Singular Components (SCs) information level,
is introduced to rank the SCs. Finally, the clean signal is reconstructed based on the
information level presented by the SCs.

3 Methodology

3.1 LSTM Architecture with Probabilistic Layer

500 evaluations of Bayesian hyperparameter optimization are employed to find the

best hyperparameters for the proposed LSTM architecture [24]. The evaluations are
done in the hyperparameter search space presented in Table 1. The Health Index
(HI), corresponding to the normalized value of the RUL target for each cycle is used
as the training target. The model output layer is set as a Normal distribution layer
which produces simultaneously the mean and standard deviation of the HI prediction.
758 A. K. M. Nor et al.

Table 1 Bayesian optimization search space

BayesOpt search space
Hidden units Dense layer size Dropout rate Mini batch size Max epoch Learning rate
10–1000 10–500 0.0–0.5 26–130 10–2000 5e-4 to 1e-3

The training thus intends to minimize the gap between the mean of the predicted HI
distributions with the HI target values while at the same time predicting variable
standard deviations.
In LSTM, gating functions acts as data flow regulator, preserving or forgetting
important or unimportant memory, respectively. LSTM consists of Input Layer,
Hidden Layer and Output Layer with a recurrent mechanism enabling the output
of the Hidden Layer to be added with new input. This made LSTM particularly
powerful in modelling sequential data. Below is the mathematical representation of
the model.
LSTM gating equation:
 
f t = σ W f .X t + R f .h t−1 + b f (1)

i t = σ (Wi .X t + Ri .h t−1 + bi ) (2)

ot = σ (Wo .X t + Ro .h t−1 + bo ) (3)

Mt = ϕ(Wc .X t + Rc .h t−1 + bc ) (4)

Ct = f t ∗ Ct−1 + i t ∗ Mt (5)

h t = ot ∗ ∅(Ct ) (6)

W ∗ = Input Weight.
R∗ = Recurrent Weight.
b∗ = Bias.
σ, ϕ, ∅ = Activation Functions.
Looking at the equations above, Eq. (1)–(3) are Forget Gate, Input Gate and
Output Gate equations. Equation (4) is used to calculate the candidate value that
will be added to the new state “t” (5) together with the old state “t-1”, regulated
by the Forget Gate. Equation (6) represents the final output of LSTM. “.” is matrix
multiplication while “” is point to point multiplication [25].
Reducing the Aleatoric Uncertainties of Failure Prediction … 759

3.2 Singular Value Decomposition (SVD)

We apply the new SVD denoising methodology as presented in [26, 27]:

1. A is the matrix of clean data expressed as:

A ∈ RT ×D

with T , the time steps and D the number variables, i.e., column of features
and with the condition that T < D.
2. A can be expressed in a reduced singular value decomposition form as:


T

A = U SV  = u k sk v k (7)
k=1

With U ∈ RT ×T and V ∈ R D×T as orthogonal matrix and S ∈ RT ×T as a diagonal

matrix. The kth SVD mode is composed of:

uk ≡ U1:T,k , (Le f t Singular V ector )

sk ≡ Skk , (Singular value)

v k ≡ V1:D,k , (Right Singular V ector )

3. In the same form, noisy data  can be express as:

 ∈ RT ×D

and E is the matrix of noise containing random noise with standard deviation, ∈
that can be expressed as:

E ∈ RT ×D

with their respective SVD representations and the same descriptions as mentioned
before:

 
T

 = Û Ŝ V̂ = ûk ŝk v̂ k (8)
k=1

   
T
  


E = USV = uk sk vk
k=1
760 A. K. M. Nor et al.

4. Writing  in function of A and E:

 = A + E (9)

3.3 Noise Filtering with SVD

Lower SVD modes in noisy data stay quite clean, thus filtering the higher effected
ones could give a cleaner signal. We assume the noise to be independent, identically
distributed (i.i.d.) Gaussian noise.
1. We project noisy time series data à as SVD matrix to obtain ũ k ,
sk and ṽk .
2. By fitting a Marchenko-Pastur distribution to the tail of noisy singular values
s̃, we can calculate the measurement error ∈. For full calculation, please refer
to [14].
3. We can then estimate the root mean square error (RMSE) of the modes with:
⎧ ⎡ ⎤ 21 ⎫
⎪
⎪ ⎪
⎪
⎪
⎪  ⎥ ⎪ ⎪
⎪
⎪ ⎢ ⎪
⎪
⎨ ⎢ ⎥ ⎪ ⎪
2 ∈⎢ D − w w  T λ̃ m 3λ̃ k − λ̃ m ⎥ ⎬
R M S E(ṽk ) ≈min , ⎢ +   ⎥ ⎥
⎪
⎪ D s̃k ⎢
⎢ D D λ̃m − λ̃k ⎥ ⎪ ⎪
⎪
⎪ ⎣ m=1 ⎦ ⎪ ⎪
⎪
⎪ ⎪
⎪
⎪
⎩ ⎪
⎭
m = k

with λ̃k ≡ s̃k (10)

4. For i.i.d Gaussian noise, the spatial correlation parameter, f and the effective
smoothing window width, w, should be 1 [26]:

f =w=1

5. Once RMSE of the modes obtained, we can calculate the parameter tk , which
represent the cleanness of a mode.
 
log(R M S E(ṽk )) − log 2
D
tk =   (11)
log(R M S E(ṽ1 )) − log 2
D

6. After that, we estimate the rank for minimum loss data reconstruction with the
rank rmin as:

rmin ≡ maximumo f ksuchthattk > 5% (12)

7. We then reconstruct the clean singular values:

Reducing the Aleatoric Uncertainties of Failure Prediction … 761
  2
s̃k − ∈ ŝk (k < kc )
sk = (13)
0 (kc ≤ k)

8. With ∈ ŝk as the Marchenko-Pastur distribution and kc as the minimum index

k that satisfies:

s̃k <∈ ŝk (14)

9. Finally, the clean signal can be reconstructed with rmin and s k obtained
previously:


rmin
Ar = ũk s k ṽ k (15)
k=1

3.4 Performance Evaluation

The prediction uncertainty is presented by the RUL distribution’s rolling standard

deviation. Since the premier objective of the work is to evaluate the level and
behavior of uncertainty, comparisons are made between the full feature dataset with
data denoised by SVD, Simple Moving Average (MOVAVG), Exponential Weighted
Moving Average (EWMA), MATLAB Empirical Mode Decomposition (EMD) and
MATLAB Wavelet denoising. We use the default MATLAB settings for EMD and
Wavelet.
For moving average methods (simple and exponential), we choose the denoising
window size that filter noise whose characteristic best resemble a white noise:
• 0 mean residual noise value.
• Monotonic rolling standard deviation.
• Low autocorrelation.
Secondly, we compare the root mean square error (RMSE) between the mean of
RUL distribution and the ground truth [28, 29].

 M  2
1 
RMSE =  RU L (i) (i)
tr uth − Mean pr ed (16)
M i=1

With RU L (i) (i)

tr uth as the ground truth RUL for gas turbine i, Mean pr ed as the mean
of RUL distribution for gas turbine i and M as the total number of gas turbine.
762 A. K. M. Nor et al.

Fig. 1 Simplified turbofan

diagram

4 Case Study

4.1 CMAPSS Turbofan Dataset

Turbofan failure dataset from NASA Commercial Modular Aero Propulsion System
Simulation (CMAPSS), published by Nasa Prognostic Centre (PCoE) of Ames
Research Centre is used in this study [30]. The simplified diagram of the studied
turbofan is shown in Fig. 1.
The chosen training and testing data consist of 100 recorded turbofan degrada-
tions. Each Training and Testing data correspond to a turbofan whose health condition
deteriorate after certain cycle [30]. Each fleet training and testing data are time series
sequence comprising of Time (Cycle), 3 Operating Conditions (OC) and 21 sensors
measurements as in Table 2. The OC refers to different operating regimes combina-
tion of Altitude (O-42 K ft.), Throttle Resolver Angle (20–100), and Mach Number
(0–0.84) [31]. High levels of noise are incorporated, and the faults are hidden by the
effect of various operational conditions [31]. However, no RUL information is given
in the training dataset. Only the ground truth RUL for each turbofan, are given. Data
pre-processing is thus applied to select the appropriate sensors, visualize the sensor
trends, and extract each cycle HI target.

4.2 Sensor Selection

Only strictly monotonic sensors are selected [32]. These sensors are useful as
they best represent single trending degradation contrary to irregular and unchanged
signals. 14 sensor signals, corresponding to sensors 2, 3, 4, 7, 8, 9, 11, 12, 13, 14,
15, 17, 20, 21 are used.
Reducing the Aleatoric Uncertainties of Failure Prediction … 763

Table 2 Sensors description

Sensor Ref Description Unit
S1 T2 Total temperature fan inlet 0R

S2 T24 Total temperature at LPC 0R

outlet
S3 T30 Total temperature at HPC 0R

outlet
S4 T50 Total temperature at LPT 0R

outlet
S5 P2 Pressure at fan inlet psia
S6 P15 Total pressure in psia
bypass-duct
S7 P30 Total pressure at HPC psia
outlet
S8 Nf Physical fan speed rpm
S9 Nc Physical core speed rpm
S10 Epr Engine pressure ratio N/A
(P50/P2)
S11 Ps30 Static pressure at HPC psia
outlet
S12 Phi Ratio of fuel flow to Ps30 Pps/psi
S13 NRf Corrected fan speed rpm
S14 NRc Corrected core speed rpm
S15 BPR Bypass ratio N/A
S16 farB Burner fuel–air ratio N/A
S17 htBleed Bleed enthalphy N/A
S18 Nf_dmd Demanded fan speed rpm
S19 PCNfR_dmd Demanded corrected fan rpm
speed
S20 W31 HPT coolant bleed lbm/s
S21 W32 LPT coolant bleed lbm/s

4.3 Piece-Wise Linear Degradation

To obtain RUL sequence labels for training, piece-wise linear degradation model
is employed [33, 34]. Each fleet health is thus considered stable in the beginning
until the start of deterioration which is the failure start point that initiate a linear
degradation until failure.
Each fleet is associated with time series corresponding to the total operational
duration of the fleet and the last cycle indicate the final instant before failure. Thus
initially, we model the RUL of a fleet to start with the value of the last cycle and
degrade linearly until 0 as shown in Fig. 2.
764 A. K. M. Nor et al.

Fig. 2 Initial HI assessment for fleet 1

The failure start point in each sensor is obtained by using Cumulative Sum
(CUSUM) anomaly detection technique, which returns the first index of the upper
or lower cumulative sums of each sensor signal that have drifted beyond 5 standard
deviations from the target mean. This indicates the initiating points of degradation
[35]. Then, the mean of the failure start points for each sensor can then be calculated.
From these means, we can calculate the mean of the means of the sensor’s detection,
which is our failure start point. Combining the linear degradation obtained earlier
and the failure start point, the transformed Fleet 1 RUL sequence is presented in
Fig. 3.

4.4 Data Sequence Length and RUL Values

We show the sequence length plot of the training data in Fig. 4. From this plot, we
derive the RUL plot as shown in Fig. 5 following steps mentioned in Sect. 73.3.
In Fig. 6 we show the sequence length of testing and data its associated expected
RUL in Fig. 7 following our previous reasoning in Sect. 73.3.
However, as we can see, the ground truth RUL, shown in Fig. 8 are different from
the expected ground truth. Some testing data have long sequence length while at the
same time having very short RUL. We thus anticipate that the model to perform more
poorly on these data.
Reducing the Aleatoric Uncertainties of Failure Prediction … 765

Fig. 3 Final HI assessment for fleet 1

Fig. 4 Training data sequence length

766 A. K. M. Nor et al.

Fig. 5 Training RUL after threshold

Fig. 6 Test data sequence length

Reducing the Aleatoric Uncertainties of Failure Prediction … 767

Fig. 7 Expected test data RUL prediction

Fig. 8 Ground truth RUL

768 A. K. M. Nor et al.

Fig. 9 Full noisy dataset

5 Result and Discussion

5.1 Aleatoric Uncertainty

Window size 2 was found appropriate for MOVAVG while 3 for EWMA. The
uncertainty measurement according of the full dataset and respective methods are
presented. For illustration purposes, we take the case of Fleet 1 as an example (Fig. 9).

5.2 RMSE Result

As we can see from the illustrations given, the lowest degree of uncertainty is accorded
to SVD denoising in Fig. 10 while the highest uncertainty come from EWMA and
EMD, shown respectively in Fig. 12 and 14. The curve of rolling standard deviation
for SVD is also smoother than the rest of the methods. As for RMSE, again, SVD
surpasses other denoising method in improving the probabilistic LSTM performance
(Table 3).
MOVAVG and EWMA, as respectively shown in Fig. 11 and 12, cannot denoise
the data properly as we are applying the same simplistic rule (window k and fixed
weights) to denoise the whole data. As for Wavelet Transform and EMD, presented
here in Fig. 13 and 14, the usage of default parameter setting, due to absence of
domain expert, is obviously not sufficient to fulfil its intended function. To execute
effective denoising in Wavelet Transform, manual manipulations need to be done first
to determine the mother wavelet and decomposition level [9]. EMD denoising, on
Reducing the Aleatoric Uncertainties of Failure Prediction … 769

Fig. 10 SVD denoising

Table 3 RMSE results for each method

RMSE results
Method Noisy data SVD MOVAVG EWMA WAVELET EMD
RMSE 34.86 34.82 34.99 35.37 35.53 34.97

Fig. 11 Moving average

770 A. K. M. Nor et al.

Fig. 12 Exponential weighted moving average

Fig. 13 Wavelet transform

the other hand, possess various decomposition modes that need deep understanding
by the user to be successful [9]. SVD however treat each data sites individually by
removing higher modes more prone to contamination, thus it denoise more effec-
tively. We show the complete HI prediction for the original data and after SVD
treatment in Fig. 15 and 16, respectively.
Reducing the Aleatoric Uncertainties of Failure Prediction … 771

Fig. 14 Empirical mode decomposition denoising

Fig. 15 Original data HI prediction with uncertainty

6 Conclusion

In this paper, SVD denoising treatment is applied to gas turbine failure dataset
to reduce the Aleatoric uncertainty of failure prediction. A probabilistic LSTM is
employed to predict the RUL distribution of the gas turbines based on sensor data. Our
proposed method show superiority in reducing the prediction uncertainty compared
to various other denoising techniques such as simple moving average, exponential
weighted moving average, empirical mode decomposition and wavelet transform.
772 A. K. M. Nor et al.

Fig. 16 HI prediction after SVD treatment

The results also indicate that SVD treatment improves the performance of prediction
better than the other mentioned techniques.

Acknowledgements We would like to thank Yayasan Universiti Teknologi Petronas (YUTP) for
funding this research

References

1. Li R, Verhagen W, Curran R (2019) A systematic methodology for prognostic and health

management system architecture definition. Reliab Eng Syst Saf 193:106598. https://doi.org/
10.1016/j.ress.2019.106598
2. Vogl GW, Weiss BA, Helu M (2019) A review of diagnostic and prognostic capabilities and
best practices for manufacturing. J Intell Manuf 30:79–95. https://doi.org/10.1007/s10845-
016-1228-8
3. Qu Y, Ming X, Qiu S, Zheng M, Hou Z (2019) An integrative framework for online prognostic
and health management using internet of things and convolutional neural network. Sensors
19:2338
4. Wang F, Pan S, Xiong Y, Fang H, Wang D (2017) Research on software architecture of prog-
nostics and health management system for civil aircraft. In: 2017 international conference on
sensing, diagnostics, prognostics, and control (SDPC). Shanghai, pp 510–513. https://doi.org/
10.1109/SDPC.2017.102
5. Li X, Xie G, Liu H, Wang W (2018) Predicting remaining useful life of industrial equipment
based on multivariable monitoring data analysis. In: 2018 Chinese automation congress (CAC).
Xi’an, China, pp 1861–1866. https://doi.org/10.1109/CAC.2018.8623249
6. Zhang H, Hu C, Fan H, Zhang W, Gao Y (2015) A new remaining useful life prediction approach
based on Wiener process with an adaptive drift. In: 2015 Chinese automation congress (CAC).
Wuhan, pp 2052–2056. https://doi.org/10.1109/CAC.2015.7382842
7. Jun S (2016) Frequentist and Bayesian learning approaches to artificial intelligence. Int J Fuzzy
Logic Intell Syst 16:111–118. https://doi.org/10.5391/IJFIS.2016.16.2.111
Reducing the Aleatoric Uncertainties of Failure Prediction … 773

8. Alaa A, Schaar M (2020) Frequentist uncertainty in recurrent neural networks via blockwise
influence functions
9. Kim M, Liu K (2020) A Bayesian deep learning framework for interval estimation of remaining
useful life in complex systems by incorporating general degradation characteristics. IISE Trans.
https://doi.org/10.1080/24725854.2020.1766729
10. Peng W, Ye Z, Chen N (2020) Bayesian deep-learning-based health prognostics toward prog-
nostics uncertainty. IEEE Trans Industr Electron 67(3):2283–2293. https://doi.org/10.1109/
TIE.2019.2907440
11. Shaker MH, Hüllermeier E (2020) Aleatoric and epistemic uncertainty with random forests.
In: Berthold M, Feelders A, Krempl G (eds) Advances in intelligent data analysis XVIII. IDA
2020. Lecture notes in computer science, vol 12080. Springer, Cham. https://doi.org/10.1007/
978-3-030-44584-3_35
12. Ståhl N, Falkman G, Karlsson A, Mathiason G (2020) Evaluation of uncertainty quantification
in deep learning. In: Lesot MJ et al (eds) Information processing and management of uncer-
tainty in knowledge-based systems. IPMU 2020. Communications in computer and information
science, vol 1237. Springer, Cham. https://doi.org/10.1007/978-3-030-50146-4_41
13. Hüllermeier E, Waegeman W (2019) Aleatoric and epistemic uncertainty in machine learning:
a tutorial introduction
14. Deng W, Ye B, Wu J, Wang X, Bao J, Chen C (2019) Sparse denoising of eddy current signals
from conductive material defects based on K-SVD dictionary learning. In 2019 IEEE 8th data
driven control and learning systems conference (DDCLS). Dali, China, pp 656–660. https://
doi.org/10.1109/DDCLS.2019.8909047
15. Seyedebrahim M, Mansouri A (2017) Non-local means denoising based on SVD basis images.
In: 2017 3rd international conference on pattern recognition and image analysis (IPRIA).
Shahrekord, pp 206–210. https://doi.org/10.1109/PRIA.2017.7983047
16. Yang Y, Rao J (2019) Robust and efficient harmonics denoising in large dataset based on random
SVD and soft thresholding. IEEE Access 7:77607–77617. https://doi.org/10.1109/ACCESS.
2019.2921579
17. Zhao Z, Wang S, Wong D, Guo Y, Chen X (2020) The sparse and low-rank interpretation of
SVD-based denoising for vibration signals. In: 2020 IEEE international instrumentation and
measurement technology conference (I2MTC). Dubrovnik, Croatia, pp 1–6. https://doi.org/10.
1109/I2MTC43012.2020.9129272
18. Wang M, Li Z, Duan X, Li W (2015) An image denoising method with enhancement of the
directional features based on wavelet and SVD transforms. Math Probl Eng 2015:1–9. https://
doi.org/10.1155/2015/469350
19. Zhang G, Xu B, Zhang K, Hou J, Xie T, Li X, Liu F (2020) Research on a noise reduction
method based on multi-resolution singular value decomposition. Appl Sci 10:1409. https://doi.
org/10.3390/app10041409
20. He Y, Gan T, Chen W, Wang H (2011) Adaptive denoising by singular value decomposition.
IEEE Signal Process Lett 18:215–218. https://doi.org/10.1109/LSP.2011.2109039
21. Wang M, Yan W, Zhou S (2018) Image denoising using singular value difference in the wavelet
domain. Math Probl Eng 2018:1–19. https://doi.org/10.1155/2018/1542509
22. Liu B, Liu Q (2020) Random noise reduction using SVD in the frequency domain. J Petrol
Explor Prod Technol 10:3081–3089. https://doi.org/10.1007/s13202-020-00938-w
23. Zhao M, Jia X (2017) A novel strategy for signal denoising using reweighted SVD and its
applications to weak fault feature enhancement of rotating machinery. Mech Syst Signal Process
94:129–147. https://doi.org/10.1016/j.ymssp.2017.02.036
24. Sherstinsky A (2018) Fundamentals of recurrent neural network (RNN) and long short-term
memory (LSTM) network
25. Zheng S, Ristovski K, Farahat A, Gupta C (2017) Long short-term memory network for
remaining useful life estimation pp. 88–95. https://doi.org/10.1109/ICPHM.2017.7998311
26. Epps B, Krivitzky E (2019) Singular value decomposition of noisy data: noise filtering. Exp
Fluids 60:1–23. https://doi.org/10.1007/s00348-019-2768-4
774 A. K. M. Nor et al.

27. Epps B, Krivitzky E (2019) Singular value decomposition of noisy data: mode corruption. Exp
Fluids 60:1–23. https://doi.org/10.1007/s00348-019-2761-y
28. Wang Q, Zheng S, Farahat A, Serita S, Gupta C (2019) Remaining useful life estimation using
functional data analysis. In: 2019 IEEE international conference on prognostics and health
management (ICPHM). San Francisco, CA, USA, pp 1–8. https://doi.org/10.1109/ICPHM.
2019.8819420
29. Ge Y, Sun L, Ma J (2019) An improved PF remaining useful life prediction method based on
quantum genetics and LSTM. IEEE Access 7:160241–160247. https://doi.org/10.1109/ACC
ESS.2019.2951197
30. Ramasso E, Saxena A (2014) Performance benchmarking and analysis of prognostic methods
for CMAPSS datasets. Int J Prognostics Health Manag 5:1–15
31. Saxena A, Goebel K, Simon D, Eklund N (2008) Damage propagation modelling for aircraft
engine run-to-failure simulation. Int Conf Prognostics Health Manag. https://doi.org/10.1109/
PHM.2008.4711414.M
32. Gao Y, Zhou J, Wu K, Zhao G, Hu C (2019) Construction method of turbine engine health
indicator based on deep learning. In: 2019 prognostics and system health management confer-
ence (PHM-Qingdao). Qingdao, China, pp 1–6. https://doi.org/10.1109/PHM-Qingdao46334.
2019.8943055; Wang Q, Zheng S, Farahat A, Serita S, Gupta C (2019) Remaining useful life
estimation using functional data analysis. In: 2019 ieee international conference on prognostics
and health management (ICPHM). San Francisco, CA, USA, pp 1–8. https://doi.org/10.1109/
ICPHM.2019.8819420
33. Li J, Li X, He D (2019) Domain adaptation remaining useful life prediction method based
on AdaBN-DCNN. In: 2019 prognostics and system health management conference (PHM-
Qingdao). Qingdao, China, pp 1–6. https://doi.org/10.1109/PHM-Qingdao46334.2019.894
2857
34. Detect small changes in mean using cumulative sum. MATLAB MathWorks® https://www.
mathworks.com/help/signal/ref/cusum.html
35. Heimes F (2008) Recurrent neural networks for remaining useful life estimation. In: IEEE
international conference on prognostics and health management
Reducing Uncertainty in Failure
Prediction Using Singular Value
Decomposition Feature Selection

Ahmad Kamal Mohd Nor, Srinivasa Rao Pedapati, and Masdi Muhammad

Abstract Uncertainty is one of the indicators to evaluate the prediction of deep

learning models. In real-world prognostic domain, predicting failure with uncertainty
provides users with powerful argument when important decisions related to safety,
security and monetary must be made. While quantifying uncertainty has slowly
becoming a norm in deep learning, no work has been dedicated in formulating the
appropriate feature selection framework that lessens the prediction uncertainties.
Much of the available techniques have only been proven to improve point estimate
predictions, without considering its effect on uncertainty. In this paper, a feature selec-
tion method based on elimination of noisy data is proposed to reduce the Aleatoric
uncertainty in a Remaining Useful Life (RUL) prediction problem. Singular Value
Decomposition (SVD) technique is employed to denoise sensor data in SVD matrix
by filtering higher SVD modes susceptible to contain noise. Then, the “cleaned”
Signal to Noise Ratio (SNR) of each feature is calculated and ranked. Features with
low SNR, thus with higher noise, are eliminated. We compare the uncertainty level
and behavior between a full feature dataset and different percentage of features
selected using SVD and SNR. The same comparison is done between our approach
and other feature selection methods such as Pearson and Spearman correlations as
well as F Regression. The results show that our approach achieved a lower uncer-
tainty degree with generally better prediction performance than the other mentioned
methods.

Keywords Uncertainty · Feature selection · PHM · RUL · CMAPSS

1 Introduction

Sudden failure is the worst imaginable situation involving complex engineering

systems. It is a risk to be minimized to safeguard the interest of safety, security, and
investment in the industrial environment. Throughout the decades, concerted efforts

A. K. M. Nor (B) · S. R. Pedapati · M. Muhammad

Universiti Teknologi Petronas, 32610 Perak, Malaysia
e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 775
R. Ibrahim et al. (eds.), International Conference on Artificial Intelligence
for Smart Community, Lecture Notes in Electrical Engineering 758,
https://doi.org/10.1007/978-981-16-2183-3_74
776 A. K. M. Nor et al.

have been put by researchers and reliability practitioners to improve operational

longevity of industrial assets.
Recently, Prognostic and Health Management (PHM) has emerged as a strong
force in providing frameworks for managing engineering system health [1–3].
PHM aims to predict impending failure, optimizing maintenance activities, reducing
workload, improving safety, availability and reducing maintenance cost [4–7].
Remaining Useful Life (RUL) prognostic is one of the core activities in PHM
[8, 9]. It consists of predicting the health state of engineering system often complex
in nature. Currently, deep learning prognostic methods gain much attention due to
the ease of deployment, the increasing data volume, and the sophistication of data
transmission technology.
Deep learning methods can be divided into frequentist and probabilistic in
nature. While the premier predicts or classifies point estimates, the latter produces
distribution or interval predictions, thus predicting with uncertainty.

1.1 Aleatoric Uncertainty in Deep Learning

Uncertainty related to the quality of input data is called Aleatoric uncertainty. This
uncertainty occurs when sensor data, used as input to deep learning model, is
subjected to measurement error or noise contamination [10–13]. This uncertainty
cannot be reduced by collecting more data [12, 13].
In the real-world applications, where ground truth RUL is not available, uncer-
tainty is often the main indicator for users to trust the model’s prediction. Security,
safety, and investment strategies depend on prediction uncertainty.

1.2 Data Denoising Using SVD

Singular Value Decomposition (SVD) is a technique to represent signal in SVD

matrix form. SVD matrix is composed of modes and singular values. When noisy
data is represented in SVD matrix form, higher modes are found to be more prone to
contamination while the lower modes remain relatively clean [14, 15]. By filtering
the higher modes, the clean signal can be reconstructed.

1.3 Feature Selection’s Impact on Uncertainty

Since Aleatoric uncertainty concerns the quality of input data, it is natural to assume
that certain features contribute more to uncertainty than others. However, there have
been no notable work dedicated to study this effect. Available feature selection tech-
niques are suitable only to improve frequentist point estimates prediction without
Reducing Uncertainty in Failure Prediction Using Singular Value … 777

considering the impact it can cause uncertainty. It is obvious that in the real-world
deep learning applications, the lowest uncertainty in prediction is preferred.
In this paper, a feature selection technique to reduce Aleatoric uncertainty is
proposed. Noisy data is denoised using SVD. Then, the Signal to Noise Ratio (SNR)
for each clean feature is calculated and ranked. Signals with low SNR, i.e., high noise
are eliminated according to certain percentage.
We demonstrate our approach in a gas turbine Remaining Useful Life (RUL)
prediction problem. A probabilistic LSTM is employed to predict the gas turbine’s
RUL sequence distributions. We compare the impact of our approach on prediction
uncertainty with full feature dataset. Comparison is also done with several others
commonly used feature selection techniques such as Pearson and Spearman correla-
tions and F Regression. We believe that this is the first such work concerning deep
learning uncertainty reduction using feature selection. This paper is very important
as uncertainty management is crucial for real-world deep learning application.
This paper is organized as follows: In Sect. 2, we mention the related literature to
our work. In the next section, the methodology of SVD and SNR feature selection is
explained. Then in Sect. 4, we describe the case study before we present and discuss
our findings in Sect. 5. Finally, we present our concluding remarks in the last section.

2 Related Work

2.1 Uncertainty Reduction in Deep Learning

Most of the works concerning uncertainty in deep learning relate to its quantification.
Very few work touches about uncertainty management. One such publication is [16]
where Fuzzy Deep Neural Network (FDNN) is proposed to reduce data ambiguity
and noise for classification task through the fusion of fuzzy rules and neural network
logic representation. In Boskos et al. [17], Wasserstein ambiguity sets are employed
in dynamic data collection process. Ambiguity sets enables the characterization of
the true random variable’s probability distribution’s uncertainty, thus exploiting high
confidence guarantees to minimize uncertainty. In Du et al. [18], to minimize the
uncertainty in domestic solar energy harvesting, only predictions of photovoltaic
(PV) output with large probability are chosen. These probabilities help to construct
plausible scenarios enabling accurate home energy scheduling. In Rodemerk et al.
[19] context-based indicators are used to reinforce the prediction confidence, i.e.,
minimize the uncertainty of car driver’s behavior at intersections. This is done by
combining context information, such as head pose and gaze direction with vehicle
data to predict the intention of the driver.
778 A. K. M. Nor et al.

2.2 SVD Denoising

SVD is a popular tool for denoising data without the intervention of domain expert.
It is widely used in denoising image data. In Yang et al. [20] SVD, together with
non-local self-similarity technique are used to denoise image data. Block matching
techniques construct matrix based on similar patches that will be denoise by SVD.
The denoised patches will be then be combined to reconstruct the clean image.
In Leal et al. [21], sparse representation and SVD are employed to denoise MRI
image for enhancing medical diagnosis. Sparse representations are obtained from
the image subdivisions. Then, new subdivisions based on the representations are
produced through SVD denoising process. These clean subdivisions will be used
later to reconstruct the clean image. The same concept is applied in Yang et al.
[22] where SVD is used to separate and denoise Computed tomography (CT) image
components.
SVD is also used to denoise general noisy data. In Zhu et al. [23], an algorithm
based on generalized S transform (GST) and SVD is employed to denoise echo
signals in ultrasonic pulse-echo testing. The GST’s function is to transform the echo
signal to time–frequency matrix before going through SVD denoising treatment. In
Zhang et al. [24], noisy signal is presented by SVD in Henkel matrix form where
clean subspaces and noisy subspaces are separated. By only taking singular values
residing in clean subspace, clean signal can be reconstructed. In Schanze [25] a signal
compression and denoising method via objective transformation and singular values
denoising is presented. The noisy signal is first mapped into matrix form where
denoising and compression process are done. Then, inverse mapping is executed to
reconstruct the clean signal. The application of this technique is demonstrated using
biomedical signals.

3 Methodology

3.1 LSTM Architecture with Probabilistic Layer

LSTM architecture which is a powerful tool for sequential data modelling is

employed. We execute Bayesian hyperparameters optimizations to find the best set
of hyperparameter in the search space shown in Table 1 [26]. 500 Bayesian optimiza-
tions are done to minimize the validation error. For the sequential training targets, we
use Health Index (HI) that corresponds to the normalized value of the RUL of each
cycle. A Normal distribution layer is incorporated as the output layer of the LSTM
network, outputting both mean and standard deviation of the HI prediction.
LSTM gating equation:
 
f t = σ W f .X t + R f .h t−1 + b f (1)
Reducing Uncertainty in Failure Prediction Using Singular Value … 779

Fig. 1 Simplified turbofan

diagram

i t = σ (Wi .X t + Ri .h t−1 + bi ) (2)

ot = σ (Wo .X t + Ro .h t−1 + bo ) (3)

Mt = ϕ(Wc .X t + Rc .h t−1 + bc ) (4)

Ct = f t ∗ Ct−1 + i t ∗ Mt (5)

h t = ot ∗ ∅(Ct ) (6)

With W ∗ as the Input Weights, R ∗ as the Recurrent Weights, b∗ as bias, and σ, ϕ, ∅

as activation functions.
In LSTM, gating mechanism enables the preserving and forgetting of data,
resulting to a longer memory retention [27]. It consists of Input Layer, Hidden Layer
and Output Layer with a recurrent mechanism enabling the output of the Hidden
Layer to be added with new input. This permits the modelling of sequential data.
Equation (1)–(3) are Forget Gate, Input Gate and Output Gate equations. Equa-
tion (4) is used to calculate the candidate value that will be added to the new state “t”
Eq. (5) together with the old state “t-1”, regulated by the Forget Gate. Equation (6)
represents the final output of LSTM. “.” is matrix multiplication while “” is point to
point multiplication [28].

3.2 Singular Value Decomposition (SVD)

The new SVD denoising method as presented in [14, 15] is used:

1. A is the matrix of clean data expressed as:

A ∈ RT ×D
780 A. K. M. Nor et al.

with T , the time steps and D the number variables, i.e., column of features and
with the condition that T < D.
2. A can be expressed in a reduced singular value decomposition form as:


T
A = U SV  = uk sk v kT (7)
k=1

With U ∈ RT ×T and V ∈ R D×T as orthogonal matrix and S ∈ RT ×T as a diagonal

matrix. The kth SVD mode is composed of:

uk ≡ U1:T,k , (Le f t Singular V ector )

sk ≡ Skk , (Singular value)

v k ≡ V1:D,k , (Right Singular V ector )

3. In the same form, noisy data  can be express as:

 ∈ RT ×D

and E is the matrix of noisy data containing random noise with standard deviation,
∈ that can be expressed as:

E ∈ RT ×D E ∈ RT ×D

with their respective SVD representations and the same descriptions as mentioned
before:

 
T

 = Û Ŝ V̂ = ûk ŝk v̂ k (8)
k=1

   
T
  

E = U S V = uk sk vk
k=1

4. Writing  in function of A and E:

 = A + E (9)
Reducing Uncertainty in Failure Prediction Using Singular Value … 781

3.3 Noise Filtering with SVD

Lower SVD modes in noisy data stay quite clean, thus filtering the higher effected
ones could give a cleaner signal. We assume the noise to be independent, identically
distributed (i.i.d.) Gaussian noise.
1. We project noisy time series data à as SVD matrix to obtain ũ k ,
sk and ṽk .
2. By fitting a Marchenko-Pastur distribution to the tail of noisy singular values
s̃, we can calculate the measurement error ∈. For full calculation, please refer
to [14].
3. We can then estimate the root mean square error (RMSE) of the modes with:
⎧ ⎡ ⎤ 21 ⎫
⎪
⎪ ⎪
⎪
⎪
⎪  ⎥ ⎪ ⎪
⎪
⎪ ⎢ ⎪
⎪
⎨ ⎢ ⎥ ⎪ ⎪
2 ∈⎢ D − w w  T λ̃ m 3 λ̃ k − λ̃ m ⎥ ⎬
R M S E(ṽk ) ≈min , ⎢ +   ⎥ ⎥
⎪
⎪ D s̃k ⎢
⎢ D D λ̃m − λ̃k ⎥ ⎪ ⎪
⎪
⎪ ⎣ m=1 ⎦ ⎪ ⎪
⎪
⎪ ⎪
⎪
⎪
⎩ ⎪
⎭
m = k

withλ̃k ≡ s̃k (10)

4. For i.i.d Gaussian noise, the spatial correlation parameter, f and the effective
smoothing window width, w, should be 1 [14]:

f =w=1

5. Once RMSE of the modes obtained, we can calculate the parameter tk , which
represent the cleanness of a mode.
 
log(R M S E(ṽk )) − log 2
D
tk =   (11)
log(R M S E(ṽ1 )) − log 2
D

6. After that, we estimate the rank for minimum loss data reconstruction with the
rank rmin as:

rmin ≡ maximum o f k such that tk > 5% (12)

7. We then reconstruct the clean singular values:

  2
s̃k − ∈ ŝk (k < kc )
sk = (13)
0 (kc ≤ k)
782 A. K. M. Nor et al.

With ∈ ŝk as the Marchenko-Pastur distribution and kc as the minimum

index k that satisfies:

s̃k <∈ ŝk (14)

8. Finally, the clean signal can be reconstructed with rmin and s k obtained
previously:


rmin
Ar = ũk s k ṽ k (15)
k=1

3.4 Signal to Noise (SNR) Calculation

For amplitude-based signal, the SNR can be calculated by Eq. (16) with Sclean as the
clean signal and Snoise is the noise signal.
 
Sclean
S N R S = 20log10 (16)
Snoise

3.5 Performance Evaluation

The prediction uncertainty is presented by the RUL distribution’s rolling standard

deviation. Since the premier objective of the work is to evaluate the level and
behaviour of uncertainty, comparisons are made between the full feature dataset
with only 85 and 75% of the full features selected by SVD + SNR, Pearson, and
Spearman correlations as well as F Regression feature selection techniques.
Secondly, we compare the root mean square error (RMSE) between the mean of
RUL distribution and the ground truth [29, 30].

 M  2
1 
RMSE =  RU L (i)
tr uth − Mean (i)
pr ed (17)
M i=1

With RU L (i) (i)

tr uth as the ground truth RUL for gas turbine i, Mean pr ed as the mean
of RUL distribution for gas turbine i and M as the total number of gas turbine.
Reducing Uncertainty in Failure Prediction Using Singular Value … 783

Table 1 Bayesian optimization search space

BayesOpt search space
Hidden units Dense layer size Dropout rate Mini batch size Max epoch Learning rate
10–1000 10–500 0.0–0.5 26–130 10–2000 5e-4 to 1e-3

4 Case Study

4.1 Gas Turbine Dataset

The data we use is from CMAPPS (Commercial Modular Aero Propulsion System
Simulation) Turbofan run-to-failure dataset, published by Nasa Prognostic Centre
(PCoE) of Ames Research Centre [31]. The diagram of the studied turbofan is shown
in Fig. 1.
Run-to-failure time series data from 100 recorded turbofan degradations is used
for training and testing. Each set of data corresponds to a turbofan whose health
condition begin to deteriorate after certain cycle [31]. Each set comprises of Time
(Cycle), 3 Operating Conditions (OC) and 21 sensors measurements as in presented
in Table 2. The OC refers to different operating regimes combination of Altitude
(O-42 K ft.), Throttle Resolver Angle (20–100), and Mach Number (0–0.84) [32].
High levels of noise are incorporated, and the faults encountered are hidden by the
effect of various operational conditions [32]. Training data RUL targets are absent,
only the ground truth RUL for each turbofan, are given. We thus need to do data
preparation to select the appropriate sensors, visualize the sensor trends and extract
each cycle RUL target.

4.2 Sensor Selection

From the 21 sensors mentioned above, only 14 strictly monotonic sensors, namely
sensor 2, 3, 4, 7, 8, 9, 11, 12, 13, 14, 15, 17, 20 and 21 are selected [33]. Mono-
tonicity of these sensors represent single trending degradation contrary to irregular
and unchanged sensor signals.

4.3 Piece-Wise Linear Degradation

To obtain RUL sequence labels for training, piece-wise linear degradation model
is employed [34, 35]. Each fleet health is thus considered stable in the beginning
until the start of deterioration which is the failure start point that initiate a linear
degradation until failure.
784 A. K. M. Nor et al.

Table 2 Sensors description

Sensor Ref Description Unit
S1 T2 Total temperature fan inlet 0R

S2 T24 Total temperature at LPC 0R

outlet
S3 T30 Total temperature at HPC 0R

outlet
S4 T50 Total temperature at LPT 0R

outlet
S5 P2 Pressure at fan inlet psia
S6 P15 Total pressure in psia
bypass-duct
S7 P30 Total pressure at HPC psia
outlet
S8 Nf Physical fan speed rpm
S9 Nc Physical core speed rpm
S10 Epr Engine pressure ratio N/A
(P50/P2)
S11 Ps30 Static pressure at HPC psia
outlet
S12 Phi Ratio of fuel flow to Ps30 Pps/psi
S13 NRf Corrected fan speed rpm
S14 NRc Corrected core speed rpm
S15 BPR Bypass ratio N/A
S16 farB Burner fuel–air ratio N/A
S17 htBleed Bleed enthalphy N/A
S18 Nf_dmd Demanded fan speed rpm
S19 PCNfR_dmd Demanded corrected fan rpm
speed
S20 W31 HPT coolant bleed lbm/s
S21 W32 LPT coolant bleed lbm/s

Table 3 SNR measurement

Sensor S2 S3 S4 S7 S8 S9 S11
SNR 68.4 57.8 56.8 65.0 75.8 74.9 56.1
Sensor S12 S13 S14 S15 S17 S20 S21
SNR 66.8 75.8 74.1 55.2 55.2 54.5 54.5

Table 4 SNR feature selection

SNR + SVD
85% S2 S3 S4 S7 S8 S9 S11 S12 S13 S14
75% S2 S3 S4 S7 S8 S9 S11 S12
Reducing Uncertainty in Failure Prediction Using Singular Value … 785

Table 5 Pearson and spearman feature selection

Pearson and spearman correlation
85% S2 S3 S4 S7 S8 S9 S11 S13 S15 S17 S20 S21
75% S2 S3 S4 S7 S8 S9 S15 S17 S20 S21

Table 6 F regression feature selection

F regression
85% S2 S3 S4 S7 S8 S11 S12 S13 S15 S17 S20 S21
75% S2 S3 S4 S7 S11 S12 S15 S17 S20 S21

Each fleet is associated with time series corresponding to the total operational
duration of the fleet and the last cycle indicate the final instant before failure. Thus
initially, we model the RUL of a fleet to start with the value of the last cycle and
degrade linearly until 0 as shown in Fig. 2.
The failure start point in each sensor is obtained by using Cumulative Sum
(CUSUM) anomaly detection technique, which returns the first index of the upper
and lower cumulative sums of each sensor signal that have drifted beyond 5 stan-
dard deviations above and below a target mean, indicating the initiating points of
degradation [36]. Then, the mean of the failure start points for each sensor can then
be calculated. From these means, we can calculate the mean of the means of the
sensor’s detection, which is our failure start point. Combining the linear degradation

Fig. 2 Initial HI assessment for fleet 1

786 A. K. M. Nor et al.

Fig. 3 Final HI assessment for fleet 1

obtained earlier and the failure start point, the transformed Fleet 1 RUL sequence is
presented in Fig. 3.

4.4 Data Sequence Length

As mentioned before, we apply a piece wise linear degradation with maximum

threshold base on the time series sequence length to estimate the RUL target for
the training data. As such, the RUL value is proportionate to the sequence length like
shown in Figs. 4 and 5. With the same reasoning, we can estimate the RUL value of
the test data from its sequence length as shown in Figs 6 and 7.
However, looking at the ground truth RUL in Fig. 8, some of the RUL values
are not proportionate to its data sequence length. We believe that the model will not
perform well on these data sequences.

5 Results and Discussion

5.1 SNR Results and Feature Selection

See Tables 3, 4, 5, 6.
Reducing Uncertainty in Failure Prediction Using Singular Value … 787

Fig. 4 Training data sequence length

Fig. 5 Training data RUL

788 A. K. M. Nor et al.

Fig. 6 Testing data sequence length

Fig. 7 Expected test data RUL prediction

Reducing Uncertainty in Failure Prediction Using Singular Value … 789

Fig. 8 Ground truth RUL

Fig. 9 Noisy data uncertainty

5.2 Aleatoric Uncertainty

The noisy data uncertainty is presented in Fig. 9. Here we present the result of our
test. For illustration purposes, we only show the uncertainty result for Fleet 1.
790 A. K. M. Nor et al.

Fig. 10 SVD + SNR 85% features uncertainty

5.3 RMSE Results

The result of SVD + SNR with 85% features is shown in Fig. 10. As we can see
in Fig. 11, SVD + SNR is the only technique that achieves the lowest uncertainty
measurement in the order of 1e-3 with 75% features. Pearson and Spearman corre-
lation feature selection with 85% features also produces lower uncertainty than the
original data as shown in Fig. 12.

Fig. 11 SVD + SNR 75% features uncertainty

Reducing Uncertainty in Failure Prediction Using Singular Value … 791

Fig. 12 Pearson and spearman correlation 85% features uncertainty

Table 7 RMSE results according to denoising treatment

Methods Noisy Data SVD 85% SVD 75% Pearson 85% Pearson 75%
RMSE 34.86 34.26 34.48 34.37 34.89
Methods Spearman 85% Spearman 75% F reg 85% F reg 75%
RMSE 34.37 34.89 35.01 35.19

Looking at the RMSE result in Table 7, we can also see that the prediction perfor-
mance is generally better with SNR + SVD than correlation-based feature selection
methods. We can understand this phenomenon as only SVD + SNR that has the effect
of denoising the data, thus improving the prediction uncertainty. In addition, SVD
is known to be amongst the best denoising technique available. Thus, the effect of
SVD denoising is really apparent with 25% less noisy features. Pearson and Spearman
correlations only optimize the features in term of linear and nonlinear correlation,
avoiding redundancy between correlated features. F Regression on the other hand,
only evaluate the correlation between features and target. This also indicates that
commonly used feature selection techniques such as Pearson and Spearman Corre-
lation as well as F Regression could result to a higher prediction uncertainty like
shown in Figs. 13 and 15. To have a clearer picture, we present in Figs. 16 and 17 the
difference in uncertainty between prediction with noisy data and with SVD + SNR
feature selection.
792 A. K. M. Nor et al.

Fig. 13 Pearson and spearman correlation 75% features uncertainty

Fig. 14 F regression 85% features uncertainty

6 Conclusion

In this paper, we present a feature selection technique based on SVD + SNR where
more noisy features are eliminated to reduce the Aleatoric uncertainty in an RUL
prediction problem. A probabilistic LSTM that predicts RUL distributions of gas
turbines is employed. From the evaluations that have been done, we demonstrate
the superiority of this method compared to correlation-based feature selection tech-
niques in reducing deep learning Aleatoric uncertainty. We also show that our feature
Reducing Uncertainty in Failure Prediction Using Singular Value … 793

Fig. 15 F Regression 75% features uncertainty

Fig. 16 Noisy data HI prediction with uncertainty

selection method yields a generally more accurate failure prognostic than the other
tested techniques. Finally, this work illustrates that usual feature selection techniques
popularly used such as Pearson and Spearman as well as F Regression could result
in higher deep learning prediction uncertainty.
794 A. K. M. Nor et al.

Fig. 17 HI prediction with uncertainty after SVD + SNR feature selection

Acknowledgements We would like to thank Yayasan Universiti Teknologi Petronas (YUTP) for
funding this research

References

1. Tsui KL, Chen N, Zhou Q, Hai Y, Wang W (2015) Prognostics and health management: a
review on data driven approaches. Math Prob Eng 2015(793161):17. https://doi.org/10.1155/
2015/793161
2. Xu M, Han M (2012) Research of prognostics and health management for Aircraft Propul-
sion System. In: Proceedings of the IEEE 2012 prognostics and system health management
conference (PHM-2012 Beijing). Beijing, pp 1–4. https://doi.org/10.1109/PHM.2012.6228891
3. Zhang A, Cui L, Zhang P (2013) State prognostics based health management for complex
systems. In: 2013 international conference on information technology and applications.
Chengdu, pp 458–461. https://doi.org/10.1109/ITA.2013.111
4. Das S, Hall R, Herzog S, Harrison G, Bodkin M, Martin L (2011) Essential steps in prog-
nostic health management. In: 2011 IEEE conference on prognostics and health management.
Montreal, QC, pp 1–9. https://doi.org/10.1109/ICPHM.2011.6024332
5. Sutharssan T, Stoyanov S, Bailey C, Yin C (2015) Prognostic and health management for
engineering systems: a review of the data-driven approach and algorithms. J Eng 7:215–222.
https://doi.org/10.1049/joe.2014.0303
6. Wilkinson C, Humphrey D, Vermeire B, Houston J (2004) Prognostic and health management
for avionics. In: 2004 IEEE aerospace conference proceedings (IEEE Cat. No.04TH8720), vol
5. Big Sky, MT, pp 3435–3447. https://doi.org/10.1109/AERO.2004.1368149
7. Shin I, Lee J, Lee JY et al (2018) A framework for prognostics and health management appli-
cations toward smart manufacturing systems. Int J Precis Eng Manuf-Green Tech 5:535–554.
https://doi.org/10.1007/s40684-018-0055-0
8. Wang J, Wen G, Yang S, Liu Y (2018) Remaining useful life estimation in prognostics using
deep bidirectional LSTM neural network. In: 2018 prognostics and system health management
conference (PHM-Chongqing). Chongqing, pp 1037–1042. https://doi.org/10.1109/PHM-Cho
ngqing.2018.00184
Reducing Uncertainty in Failure Prediction Using Singular Value … 795

9. Wang H, Liu Y, Liu Z, Wang Z, Huang H (2013) Remaining useful life estimation for degrada-
tion and shock processes. In: 2013 international conference on quality, reliability, risk, mainte-
nance, and safety engineering (QR2MSE). Chengdu, pp 1762–1764. https://doi.org/10.1109/
QR2MSE.2013.6625917
10. Ghoshal B, Tucker A, Sanghera B, Wong WL (2019)Estimating uncertainty in deep learning
for reporting confidence to clinicians when segmenting nuclei image data. In: 2019 IEEE 32nd
international symposium on computer-based medical systems (CBMS). Cordoba, Spain, pp
318–324. https://doi.org/10.1109/CBMS.2019.00072
11. Shaker MH, Hüllermeier E (2020) Aleatoric and epistemic uncertainty with random forests.
In: Berthold M, Feelders A, Krempl G (eds) Advances in intelligent data analysis XVIII. IDA
2020. Lecture notes in computer science, vol 12080. Springer, Cham. https://doi.org/10.1007/
978-3-030-44584-3_35
12. Kendall A, Gal Y (2017) What uncertainties do we need in Bayesian deep learning for computer
vision? In: Proceedings of the 31st international conference on neural information processing
systems (NIPS’17). Curran Associates Inc., Red Hook, NY, USA, pp 5580–5590
13. Depeweg S, Hernández-Lobato JM, Doshi-Velez F, Udluft S (2018) Decomposition of
uncertainty in bayesian deep learning for efficient and risk-sensitive learning. ICML
14. Epps B, Krivitzky E (2019) Singular value decomposition of noisy data: noise filtering. Exp
Fluids 60:1–23. https://doi.org/10.1007/s00348-019-2768-4
15. Epps B, Krivitzky E (2019) Singular value decomposition of noisy data: mode corruption. Exp
Fluids 60:1–23. https://doi.org/10.1007/s00348-019-2761-y
16. Deng Y, Ren Z, Kong Y, Bao F, Dai Q (2017) A hierarchical fused fuzzy deep neural network for
data classification. IEEE Trans Fuzzy Syst 25(4):1006–1012. https://doi.org/10.1109/TFUZZ.
2016.2574915
17. Boskos D, Cortes J, Martinez S (2020) Data-driven ambiguity sets with probabilistic guarantees
for dynamic processes. IEEE Trans Autom Control 1–1. https://doi.org/10.1109/TAC.2020.301
4098
18. Du Z, Wang W, Zhang J, Zhang Y, Xu X, Liu J (2020) A data-driven home energy scheduling
strategy under the uncertainty in photovoltaic generations. IEEE Access 8:54125–54134.
https://doi.org/10.1109/ACCESS.2020.2980850
19. Rodemerk C, Winner H, Kastner R (2015) Predicting the driver’s turn intentions at urban
intersections using context-based indicators. In: 2015 IEEE intelligent vehicles symposium
(IV). Seoul, pp 964–969. https://doi.org/10.1109/IVS.2015.7225809
20. Yang G, Wang Y, Xu B, Zhang X (2019) An image denoising algorithm based on singular value
decomposition and non-local self-similarity. In: Vaidya J, Zhang X, Li J (eds) Cyberspace safety
and security. CSS 2019. Lecture notes in computer science, vol 11983. Springer, Cham. https://
doi.org/10.1007/978-3-030-37352-8_44
21. Leal N, Zurek E, Leal E (2020) Non-local SVD denoising of MRI based on sparse
representations. Sensors 20:1536
22. Yang W, Hong JY, Kim JY, Paik SH, Lee SH, Park JS, Lee G, Kim BM, Jung YJ (2020) A novel
singular value decomposition-based denoising method in 4-dimensional computed tomography
of the brain in stroke patients with statistical evaluation. Sensors (Basel) 20(11):3063. https://
doi.org/10.3390/s20113063 PMID:32481740;PMCID:PMC7309118
23. Zhu Y et al (2019) Denoising ultrasonic echo signals with generalized s transform and singular
value decomposition. Traitement du Signal 36:139–145
24. Zhang X, Tang J, Zhang M, Ji Q (2016) Noise subspaces subtraction in SVD based on the
difference of variance values. J Vibroeng 18:4852–4861. https://doi.org/10.21595/jve.2016.
16745
25. Schanze T (2018) Compression and noise reduction of biomedical signals by singular value
decomposition. IFAC-PapersOnLine 51:361–366. https://doi.org/10.1016/j.ifacol.2018.03.062
26. Ruben M (2014) BayesOpt: a Bayesian optimization library for nonlinear optimization,
experimental design and bandits. J Mach Learn Res 15(1):3735–3739
27. Sherstinsky A (2018) Fundamentals of recurrent neural network (RNN) and long short-term
memory (LSTM) network
796 A. K. M. Nor et al.

28. Zheng S, Ristovski K, Farahat A, Gupta C (2017) long short-term memory network for
remaining useful life estimation 88–95. https://doi.org/10.1109/ICPHM.2017.7998311
29. Zhang X et al (2019) Remaining useful life estimation using CNN-XGB with extended time
window. IEEE Access 7:154386–154397. https://doi.org/10.1109/ACCESS.2019.2942991
30. Bruneo D, De Vita F (2019) On the use of LSTM networks for predictive maintenance in
smart industries. In: 2019 IEEE international conference on smart computing (SMARTCOMP).
Washington, DC, USA, pp 241–248. https://doi.org/10.1109/SMARTCOMP.2019.00059
31. Ramasso E, Saxena A (2014) Performance benchmarking and analysis of prognostic methods
for CMAPSS datasets. Int J Prognostics Health Manag 5:1–15
32. Saxena A, Goebel K, Simon D, Eklund N (2008) Damage propagation modelling for aircraft
engine run-to-failure simulation. Int Conf Prognostics Health Manag. https://doi.org/10.1109/
PHM.2008.4711414.M
33. Gao Y, Zhou J, Wu K, Zhao G, Hu C (2019) Construction method of turbine engine health
indicator based on deep learning. In: 2019 prognostics and system health management confer-
ence (PHM-Qingdao). Qingdao, China, pp 1–6. https://doi.org/10.1109/PHM-Qingdao46334.
2019.8943055; Wang Q, Zheng S, Farahat A, Serita S, Gupta C (2019) Remaining useful life
estimation using functional data analysis. In: 2019 IEEE international conference on prognos-
tics and health management (ICPHM). San Francisco, CA, USA, pp 1–8. https://doi.org/10.
1109/ICPHM.2019.8819420
34. Li J, Li X, He D (2019) Domain adaptation remaining useful life prediction method based
on AdaBN-DCNN. In: 2019 prognostics and system health management conference (PHM-
Qingdao). Qingdao, China, pp 1–6. https://doi.org/10.1109/PHM-Qingdao46334.2019.894
2857
35. “Detect small changes in mean using cumulative sum” MATLAB MathWorks® https://www.
mathworks.com/help/signal/ref/cusum.html
36. Heimes F (2008) Recurrent neural networks for remaining useful life estimation. In: IEEE
international conference on prognostics and health management
Unsupervised Classification of Acoustic
Emission Signal to Discriminate
Composite Failure at Low Frequency

Noor A’in A. Rahman, Zazilah May, and Muhamad Shazwan Mahmud

Abstract The use of acoustic emission (AE) for damage assessment and detection
technique in structural engineering is widely used and has earned a reputation as
one of the reliable non-destructive techniques. AE source is produced based on the
elastic wave propagation through the specimen which converted into the electrical
AE signal by the AE sensors. Certain AE features belong to the signal allow their
use to discriminate mode of damage in a composite material. However, the challenge
encountered during analysis of AE signals attributed to the techniques like pattern
recognition and classification method. In this paper, several orientation of laminated
fiber specimens are undergoing tensile test. According to the information from tensile
load test, significant features were monitored namely amplitude and energy in order to
execute the classification method. The AE data are successfully cluster by k—means
algorithm.

Keywords Acoustic emission · Composite fiber · Clustering · Low frequency

monitoring

1 Introduction

Carbon fiber laminates composite are widely used in many structure and engi-
neering application due to their excellent properties such as good corrosion resistance,
high modulus elasticity, light weight accompanied by the high strength. Despite all
these advances, the monitoring and prediction of the failure experience by a fiber
composite remain major problems. Understanding damage on fiber composite is very
challenging due to the complex process had occurred during the process.

N. A. A. Rahman (B) · Z. May

Electrical and Electronics Department, Universiti Teknologi PETRONAS, 31260 Bandar Seri
Iskandar, Perak, Malaysia
M. S. Mahmud
Mechanical Engineering Department, Universiti Teknologi PETRONAS, 31260 Bandar Seri
Iskandar, Perak, Malaysia

© The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 797
R. Ibrahim et al. (eds.), International Conference on Artificial Intelligence
for Smart Community, Lecture Notes in Electrical Engineering 758,
https://doi.org/10.1007/978-981-16-2183-3_75
798 N. A. A. Rahman et al.

In-service monitoring approach by Acoustic Emission (AE) is a reliable technique

to monitor and distinguish these failures. AE events can be monitored with low
frequency (audible sound range) up to high range frequency (ultrasonic range). When
critical stress occurs on a structure, the strain energy stored is suddenly released
thus create mechanical stress waves which spread concentrically around the source
of occasion. The release energy can be detected with suitable sensors (mechanical
information converted into the electrical signal).
AE from material damage in composite is relatively easy to monitor and record,
but the efficient and effective analysis is still challenging [1]. One of the common
analysis is clustering. Clustering can be divided into three major parts consists of
choosing the significant features, identifying the right number of clusters and lastly
the modifying technique of the clustering to achieve. Different AE features can
be used for the first part namely amplitude, duration, energy, peak frequency and
many more. Extra features can be calculated from combination of those features. For
instance Bakhtiary and his co-workers [2] define their algorithm by ratio of strain
energy and acoustic energy to identify the onset of delamination and its progression
in composite fiber.
Whenever multiple phenomena are involved and thus different signal clusters,
unsupervised pattern recognition is also sometimes used to address the problem of
labeling the class/cluster. The term unsupervised pattern recognition is on the other
hand used to describe the complete methodology consisting of procedures for cluster
analysis, descriptor selection and cluster validity, whenever none of the information
is available [3].
The k-means algorithm is one of the most widely used methods to cluster any
signal including AE signal. Ech-Choudany et. al [4] applied the k-means combined
with Incremental Clustering (IC) in order to construct a proper labelled learning
database of AE signals. However, they conclude AE signals are more relevant to
study in the frequency domain in order to extract more relevant descriptor.
A lot of works are available at finding the mode of failure on composite fiber with
a wide range of frequency up to 600 kHz [1, 5]. Based on the dominant study on
failure mechanism in the composite, matric cracking has occurred in almost low-
frequency [6–8]. The resulting output from their research showed the AE event able
to be clustered and based on well-defined frequencies. Rather than frequency, other
AE waveform parameters should be therefore considered in clustering analysis. One
of the main concerns during acquisition AE signal is capturing the signal by a proper
frequency range of sensor. This work will be focused on a low frequency range of
detection on composite failures. This can be done by using a proper AE sensor which
having the working frequency range below than 100 kHz.
This paper presents the laboratory test results of AE measurement during tensile
test on composite laminate fiber specimen. The AE features from laboratory test
will be clustered by k-means algorithm. The unsupervised techniques are used for
improving the clustering/classification of failures on composite fiber for automated
clustering. Clustering generally requires user to identify the number of clusters in
Unsupervised Classification of Acoustic Emission Signal … 799

Table 1 Specimen specification

Specimen type Lay-ups Length (mm) Width (mm) Thickness (mm)
T30 30° 320 250 2.5
T45 45° 320 250 2.5
T90 90° 320 250 2.5

second part of the process. There are many cluster validity techniques for determina-
tion the cluster. However, in this paper the mechanical failure of composite structure
will be verified by the load diagram which obtain from the tensile testing.

2 Experimental Procedure

2.1 Materials and Specimen Preparation

Composite laminates were composed using 2 mm High Strength Carbon Fibre Sheet
which a cured carbon fiber plate supplied by Easy Composites Ltd. Staffordshire,
United Kingdom. High Strength Carbon Fibre Sheet is produced from in-house
manufacturing proses with specially design platen press of 3 layers of XPEG XC130
from the same company. XPEG XC130 is epoxy based 2 × 2 twill weave prepreg
carbon of Pyrofil TR30S high strength 3 k carbon. The specimens were prepared into
three types of lay-up which were 30°, 45° and 90° lay-ups with final dimensions of
(270 × 250 × 2.5) mm using a diamond coated abrasive cutter. Each lay-out was
cut by referring to the initial 0° of the vertical direction of the first corner of 2 × 2
twill in weave. Prior to the tensile test, all specimens underwent pre-conditioning at
23 °C for 16 h in a desiccator to stabilize the humidity to satisfy the ASTM standard.
Table 1 summarizes the lay-ups, dimension and abbreviated name used in the rest of
paper.

2.2 Tensile Test

The mechanical testing was performed using a tensile test following the ASTM
D3039/D, standard testing method to generate damage modes of composite material.
The specimens were fixed between the jaws of the machine to undergo testing.
These tests were conducted using a universal tensile machine (UTM) manufactured
by Gotech Testing Machine Inc. (AI-7000L). The testing was used for inducing
external force to the specimen at 15 kN capacity with a constant speed 2 mm/s, at
room temperature. For each type of lay-up, at least five specimens were tested. The
average value and the standard deviation were then recorded.
800 N. A. A. Rahman et al.

2.3 AE Equipment

Acoustic emission software AEWin and an acquisition system supplied by Physical

Acoustic Corporation (PAC) with a sampling rate of 10 MHz were used to record
AE event in this study. A piezoelectric sensor type R6I from PAC was used to sense
event during tensile testing. The sensor was place in the middle of the specimen,
coupled with grease to provide good acoustic coupling between them. The sensors
had a resonance frequency of maximum 100 kHz. Pencil lead break (PLB) test was
performed on the specimen in an arrangement that the source and the sensor is
passes through in the middle of the specimen. PLB is a well-established procedure
for generating simulated AE sources as recommended by ASTM standard (E976-99).
Procedure were repeated for the rest of 15 specimens during tensile tests.
The typical AE waveform is shown in the following figure where the interested
features are labelled (Fig. 1).

3 Methodology

A big data was collected during acquisition from the hardware. The hardware was
set at 40 dB threshold value in order to cut off low amplitude signal related to
environmental noise. The methodology used in the discrimination of failure modes
in post composite laminates is by using unsupervised clustering techniques. The
software used in implementation this methodology is KNIME Analytics Platform
version 4.0.2. KNIME is increasingly used by researchers in various areas of data
mining and machine learning research work. It is use graphical sequence based to
add the node for each algorithm. The pattern recognition techniques selected for the
present study is k-means. The data were normalized prior to the clustering procedure.

Fig. 1 Typical AE
waveform generated on
carbon fiber composite
during acquisition
Unsupervised Classification of Acoustic Emission Signal … 801

Fig. 2 k-means algorithm workflow in KNIME

3.1 k-means

The k-means algorithm aims to minimize the sum of the squared distances between
all the vectors of a cluster and its center. The cluster number k is specified based
on the mechanical test information. Process for executing the k-means algorithm in
KNIME is shown in the following diagram (Fig. 2).
The k-means performs a crisp clustering that assigns a data vector to exactly
one cluster. Thus, the algorithm terminates when the cluster assignments do not
change anymore. Basically, the clustering algorithm uses the Euclidean distance on
the selected attributes. The AE features information were normalized prior to the
clustering process.

4 Result and Discussion

The investigation of the current effort has been carried out entirely experimentally
by laboratory tests. The results are segregated to subsections to make it easier for
understanding the insight of each effective variables.

4.1 Correlations of AE Results with Tensile Test

Figure 3 represents an evolution of stress–strain and AE features for each types

of specimen. Practically no difference is observed between the stress–strain curve
obtained in the different types of specimen. Each curve shows a linear behavior from
initial up to the higher load applied, at the end of test. Then the stress drastically drops
when the UTS is achieved (not shown in diagram). According to UTS values, state
in Table 2, the average tensile strength (for all 5 specimen) increase by increasing the
orientation degree of laminated fiber. T30 specimen reaches the break at the same of
802 N. A. A. Rahman et al.

Fig. 3 Load curve and 0.3 110

significant AE features T30
0.25 100
(energy and amplitude) for

Amplitude, dB
each specimen 90
0.2

Stress, kN
80
0.15
70
0.1
60
0.05 50
0 40
0 50 100 150 200 250
Time, sec
0.3 5000
0.25 T30
4000
0.2

Energy, eu
Stress, kN

3000
0.15
2000
0.1
0.05 1000

0 0
0 50 100 150 200 250
Time, sec
0.3 110
T45 100
0.25

Amplitude, dB
90
Stress, MPa

0.2
80
0.15
70
0.1 60
0.05 50
0 40
0 100 200 300
Time, sec
0.3 5000
0.25 T45
4000
Stress, MPa

0.2
Energy, eu

3000
0.15
2000
0.1
0.05 1000

0 0
0 100 200 300
Time, sec
Unsupervised Classification of Acoustic Emission Signal … 803

Fig. 3 (continued) 30 110

T90 100
25

Amplitude, dB
90

Stress, MPa
20
80
15
70
10 60
5 50
0 40
0 50 100 150 200 250
Time, sec

30 5000
T90
25 4000
Stress, MPa

20

Energy, eu
3000
15
2000
10
5 1000

0 0
0 50 100 150 200 250
Time, sec

Table 2 Average ultimate

Specimen type UTS (MPa) Tensile strain (%)
tensile strength (UTS) and
tensile strain for each test T30 454.6 1.48
specimen T45 484.4 1.62
T90 507.9 1.23

the T90 specimen followed by T45 later. Figure 3 also show the evolution of AE signal
energy and amplitude during tensile load carried out on each specimen. Generally,
the matrix cracking and some interfacial failures are recorded. Since the interested
frequency (< 100 dB) are recorded. This is logical because, as the load increase, not
only is there much more extensive matrix cracking but also many more delaminations
occur at the previous loading steps. Moreover, as the loading continued, the crack
propagation spread to the top, rupturing of the rest of specimen, while fiber pull-out
gradually occurred.
According to Fig. 3, as expected the highest energy value recorded for all spec-
imen are right before the specimen were break. Generally, the release elastic energy
produced i.e. by matrix cracking and delamination is detected from the beginning
of the tensile load test. Moreover, these failures producing a very ‘loud’ sound to be
listened by the sensor. The highest recorded energy before fiber break for specimen
T30, T45 and T90 are 4051 eu, T45 3094 eu and T90 4508 eu respectively. When
the load increased, trends of the AE energy of the specimen were similar regardless
the fiber orientation of the composite. Similar pattern is recorded by the amplitude
804 N. A. A. Rahman et al.

of AE signal. However, an almost linear behavior of AE peak amplitude is recorded

in each of the specimen towards end of experiment. Unfortunately, the amplitude
of AE signal is not much different for each laminated fiber. According to literature,
the predominance of AE events occurring at an amplitude less than 60 dB possibly
correlates with fiber breakage [9]. Thus, according to the test on this specimen,
the fiber breakage occurs on the specimen mostly in the middle of the tensile load
testing. In conclusion, it is evident that with continuous loading amplitude continued
increasingly, up to the maximum value (approximately 100 dB) while approaching
the UTS.

4.2 Clustering

AE cluster behavior were assessed and compared to each of the test specimen exposed
to tensile loading. AE clusters were based on the normalized amplitude and energy
feature of each AE event. Practically, the clustering is based on the damage evolution
during tensile load testing. If one damage type is the exclusive source of an AE
cluster, both damage and the accumulation of event cluster would start together [1].
In this study, the k value is based on the type of mechanical during tensile load.
According to stress diagram, the mode of failures can be divided into two major
cases namely delamination and matrix cracking (occurred at the beginning of test)
and fiber breakage fiber pull-out towards end of experiment. As mention previously,
only a low frequency of damage is counted, based on the type of AE sensor. Variations
in the damage sources associated with the same mechanism most probably exist. The
clustering results assigning the AE features (amplitude and energy) data as shown in
Fig. 3 (Fig. 4).

5 Conclusion

Different laminated fiber orientation composite material tensile strength experiments

were analyzed, and research on failures mode based on tensile load was conducted.
Moreover, the clustering analyzes the potential AE features which can be indicates
the mode of failure during monitoring. The following can be concluded:
(1) The tensile strength of laminated fiber composite material is significantly
increase in order T30 > T45 > T90.
(2) The selected fiber orientation had almost similar behavior of amplitude and
energy pattern in the laminated composite specimen.
(3) k-means algorithm able to cluster the amplitude and energy of AE signal by
using KNIME data mining software.
Unsupervised Classification of Acoustic Emission Signal … 805

T30
Amplitude

Energy
Peak frequency Peak frequency

T45
Amplitude

Energy

Peak frequency

T90
Amplitude

Energy

Peak frequency Peak frequency

Fig. 4 k-means clustering output

Acknowledgements The authors acknowledge Petroleum Research Fund–Board of Trustees

(Grant number 0153AA-E55) (PRF-BOT) and Yayasan Universiti Teknologi PETRONAS (YUTP)
for providing the research grant YUTP-FRG 2016 to perform this research and Universiti Teknologi
PETRONAS for their supports and financial support.
806 N. A. A. Rahman et al.

References

1. Sawan HA, Walter ME, Marquette B (2015) Unsupervised learning for classification of acoustic
emission events from tensile and bending experiments with open-hole carbon fiber composite
samples. Compos Sci Technol 107:89–97
2. Bakhtiary Davijani AA, Hajikhani M, Ahmadi M (2011) Acoustic emission based on sentry
function to monitor the initiation of delamination in composite materials. Mater Des 32(5):3059–
3065
3. Calabrese L, Galeano M, Proverbio E, Di D, Cappuccini F, Donato A (2016) Monitoring of
13% Cr martensitic stainless steel corrosion in chloride solution in presence of thiosulphate by
acoustic emission technique. Corrosion Science
4. Ech-Choudany Y, Assarar M, Scida D, Morain-Nicolier F, Bellach B (2017) Unsupervised
clustering for building a learning database of acoustic emission signals to identify damage
mechanisms in unidirectional laminates. Appl Acoust 123:123–132
5. Koruk H, Genc G (2015) Investigation of the acoustic properties of bio luffa fiber and composite
materials. Materials Letter 157:166–168
6. Gutkin R, Green CJ, Vangrattanachai S, Pinho ST, Robinson P, Curtis PT (2011) On acoustic
emission for failure investigation in CFRP: Pattern recognition and peak frequency analyses.
Mech Syst Signal Process 25(4):1393–1407
7. Ramirez-Jimenez CR, Papadakis N, Reynolds N, Gan TH, Purnell P, Pharaoh M (2004) Identi-
fication of failure modes in glass/polypropylene composites by means of the primary frequency
content of the acoustic emission event. Compos Sci Technol 64(12):1819–1827
8. De Groot PJ, Wijnen PA, Janssen RB (1995) Real time frequency determination of acoustic
emission for different fracture mechanisms in carbon/epoxy composites. Compos Sci Technol
55:405–412
9. Gholizadeh S, Leman Z, Baharudin BTHT, Inayatullah O (2018) Acoustic emission analysis
for characterisation of damage mechanisms in glass fiber reinforced polyester composite. Aust
J Mech Eng 16(1):11–20
Finding the EEG Footprint of Stress
Resilience
Rumaisa Abu Hasan, Syed Saad Azhar Ali, Tong Boon Tang,
and Muhamad Saiful Bahri Yusoff

Abstract Work stress faced by adults can lead to decreased job performance,
reduced mental and physical wellbeing, and other detrimental health problems.
Researchers are reporting resilience as a key factor in determining a person’s vulner-
ability towards mental stress disorders. Psychosocial measures of resilience conven-
tionally use the self-assessment approach which is susceptible to potential biases
caused by self-reporting and concerns of social stigma. With increasing emphasis
of its role in mental health, researchers are using fMRI modality to identify the
brain activity of stress resilience. But this approach is costly and lack practicality
when evaluating stress resilience in daily tasks. The EEG modality provides a cost-
efficient alternative with better practicality and high temporal resolution in studying
the brain activity of stress resilience. However, EEG-based literatures on stress
resilience are limited to brain activity during resting state. With reference to the cogni-
tive affective conceptual stress model, we define stress resilience as an adaptation
process, involving cognitive appraisal, physiological arousal and coping behaviour,
that utilizes individual resources to cope with stress. This paper proposes an approach
to identify the features of EEG-neural correlates of stress resilience through brain
rhythms, hemispheric asymmetry and brain network.

Keywords Resilience · Stress · EEG feature

R. A. Hasan (B) · S. S. A. Ali · T. B. Tang

Centre for Intelligent Signal and Imaging Research (CISIR), Electrical and Electronic Engineering
Department, Universiti Teknologi PETRONAS, 32610 Bandar Seri Iskandar, Perak, Malaysia
e-mail: [email protected]
M. S. B. Yusoff
Department of Medical Education, School of Medical Sciences, Universiti Sains Malaysia,
Kelantan, Malaysia

© The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 807
R. Ibrahim et al. (eds.), International Conference on Artificial Intelligence
for Smart Community, Lecture Notes in Electrical Engineering 758,
https://doi.org/10.1007/978-981-16-2183-3_76
808 R. A. Hasan et al.

1 Introduction

In year 2002, the World Health Organization published a monograph concerning the
alarming mental health problem as one of the leading causes of disability and dis-
ease world-wide [1]. The increase in work-related stress among adults is a raising
concern throughout the world [2–4]. This problem has been recognized as detrimental
in both poor and rich countries, and persist in population regardless of age, gender and
social strata. In Malaysia, the percentage of clinically-diagnosed adults with stress-
disorders have increased [5], with a 29.9% overall prevalence of work-related stress
[2]. This indicates that one of three adults are suffering from work-related stress.
Despite being an “unseen” condition, work-related stress has its pathogenicity [2]
and leads to mental health deterioration [6].

1.1 Role of Resilience in Mental Health

Stress resilience is a measure of an individual’s vulnerability towards stress in

perceiving adverse events as minimally threatening and developing adaptive physio-
logical and psychological responses [7, 8]. Recent findings from researches on brain
responses towards stress have reported resilience as a key factor for perceiving and
reacting positively, both emotionally and behaviorally, towards stressful conditions.
Davydov et al. identified resilience as a defense mechanism of the body that protects
it from the negative effects of stress response [9]. It is a measure of how the body
adapts or maladapts to internal and external stressors [10]. Adapting towards stress
build resilience and has been associated with positive emotions and good emotion-
regulating abilities [11]. Connor and Davidson emphasized the need to quantify
and validate the measure of resilience as it can be used in gauging efficiency of
psychotherapy and to investigate which stress coping strategies are adaptive and
maladaptive to the individual [10].

1.2 Limitation of Psychological Assessment for Stress

Resilience

Researchers in the psychosocial fields have proposed the multidimensional charac-

teristic of resilience [7, 9–11]. The resilience is viewed as an integration of multiple
constructs as it combines the biology, psychology and social factors of an individual
[9]. Although these assessments have been reported to have valid psychometric prop-
erties in measuring resilience and has been used in literatures on different cohorts,
there is a limitation of non-physiological approach which only considers character-
istics or level of resilience and not the resiliency adaptation process [10]. In addition,
self-assessment approach in psychotherapy has the disadvantage of being vulnerable
Finding the EEG Footprint of Stress Resilience 809

to potential biases caused by self-reporting and concerns of social stigma. In studies

which use associated factors to measure resilience, researchers are required to use
a number of measuring instruments to ensure that the multidimensional character-
istic of resilience is addressed [12, 13]. Hence, there is a need for a physiological-
based assessment of resilience that allow objective monitoring of the body responses
towards stressful life events.

1.3 Brain Activity of Stress Resilience

As resilience is the adaptation of mental and physical responses to stress, the brain
also exhibits neural activities that reflect this mechanism. When continuously ex-
posed to stress, a region within the brain limbic system known as amygdala under-
goes structural changes through dendritic growth and spine synapses formation,
whereas medial prefrontal cortex has been observed to undergo dendritic short-
ening [14]. Neural activity within the prefrontal cortex has also been reported to
influence the reactivity of amygdala in emotional arousal [15]. The hypothalamic–
pituitary–adrenal (HPA) axis is another neural circuitry which plays a key role in
determining the degree of individual stress resilience from the gestational stage [8].
Stress stimuli alter the neuroendocrine activity of this axis, triggering the cascading
effects of neuronal modulation and hormonal secretion that affects the brain. More-
over, the reciprocal communication of both the HPA axis and medial prefrontal
cortex with the dorsal raphe nucleus forebrain region highlights the functional impli-
cation of prefrontal cortex on the HPA axis [8]. Franklin et al. (2012) highlighted
an interesting hypothesis on the role of ventral medial prefrontal cortex in acquiring
resilience through the cognition of progressive learning. These findings on the func-
tional connectivity of frontal activity and both amygdala and the HPA axis emphasize
the relevance for a brain-based measure for stress resilience as this characteristic has
been strongly associated to vulnerability to mental illness [8, 9].

1.4 Limitation of FMRI-Based Application on Stress

Resilience

Findings from functional magnetic resonance imaging (fMRI) –based literatures

have identified the association of stress resilience with the anterior cingulate cortex
[16] and prefrontal cortex [13, 17, 18]. However, the use of fMRI modality limits
its applicability in clinical and research purposes due to its immobility and high cost
required for the machine, procedures and setup [19]. The procedure of collection
fMRI data itself involves noisy scanner and are highly susceptible to physiological
noise due to motion [20]. The fMRI scanner limits the movement of a person as they
are required to be in supine position throughout the procedure and therefore cannot
810 R. A. Hasan et al.

be used in collecting data involving any task performance. In a review by Thibault

et al., the minimum initial setup cost for fMRI modality is USD 500,000 whereas the
maximum initial setup cost for electroencephalography (EEG) modality is only USD
50,000 [21]. In addition, the EEG modality has the advantage of being user-friendly
especially for claustrophobic patients that cannot use the fMRI, portable and has high
temporal resolution [21]. Findings from fMRI studies highlight the significance of
the anterior cingulate and prefrontal cortices for resilience in both resting [16] and
building [13, 17, 18] states. As brain activity in these locations can be monitored
using EEG-modality, this suggests the relevance of EEG-modality in identifying the
neural correlates for stress resilience. By identifying the EEG-neural correlates for
stress resilience, it can be used as a risk marker for mental illnesses at a much higher
applicability rate in both clinical and non-clinical populations.

1.5 EEG-Based Approach in Identifying Stress Resilience

The role of resilience in stress coping is evident [8, 16–18]. However, the EEG-based
literature investigating the stress resilience is still at its infancy despite the increasing
findings from fMRI studies highlighting the role of prefrontal cortex in resilience.
Findings from studies that focus on cohorts with specific life experience may not be
generalized to normal healthy populations. The association of EEG-brain waves with
resilience at resting state of the brain may also differ when the brain is responding
to a stressful event. These limitations highlight the need to identify EEG-neural
correlates for resilience in a stress-inducing state. These findings can further be used
in resilience training program for healthy population.

1.6 Rationale of Study

The identification of EEG-neural correlates of stress resilience can be used as brain

markers to assess risk of individual in developing mental illness. By establishing the
correlation between these EEG-neural correlates for stress resilience with established
psychological assessments, the brain markers can be used in cognitive-behavioural
therapy in both clinical and non-clinical populations. With the use of EEG-modality,
these findings can be further explored in research focusing on resilience as it is a
more cost-efficient alternative than the fMRI modality.

2 Proposed Methodology

In order to identify EEG-neural correlates for resilience building, there is a need to

present the brain with a stress stimulus. This can be observed in the stress stimuli
Finding the EEG Footprint of Stress Resilience 811

presented to subjects in the fMRI-based literatures [13, 17, 18]. The stimuli trigger
the stress resilience mechanism which attempts to cope with the brain activity
changes induced by stress. Although many studies on stress resilience using the
EEG modality evaluated the brain activity at resting state [11, 22, 23], these findings
provide directions for further analysis in identifying the features for stress resilience.

2.1 Current Findings of EEG for Stress Resilience

Compared to the literature on stress responses, the EEG brain markers of stress
resilience are less studied [23]. Paban et al. (2019) has observed significant corre-
lation between resilience level and brain network flexibility in the delta, alpha and
beta waves. Flexibility index is a measure of how often one node (i.e. a brain region)
within a network changes its affiliation with a module (i.e. a set of strongly connected
brain regions performing specific function [24]) over time [23, 25, 26]. Flexibility
of delta, alpha and beta brain networks during 10 min of resting with eyes closed
was negatively correlated with the psychological resilience score [23]. Network flex-
ibility of specific brain regions significantly decreased with higher resilience score,
suggesting that the brain core network is less flexible in resilient subjects.
Zhang et al. (2018) also investigated the association of resilience of network at
resting-state of EEG signal between patients with major depressive disorders and
healthy participants [22]. In the study, network resilience reflected the degree of
tolerance when the functional brain network is subjected to unstable events. Patients
with MDD had lower inter-nodal mean and positive correlation coefficients in the
alpha wave of the functional brain network, which suggests weaker density of node-
pairing connections compared to healthy participants. This finding strengthens the
postulation by Paban et al. (2019) that people with higher psychological resilience
level have a more stable rich-club network.
In a study investigating biomarker for attentional control using theta/beta ratio,
Putman et al. observed that higher power density of the ratio is significantly associated
with lower self-reported trait attentional control [27]. Findings from this study was
concluded to contribute in predicting an individual’s resilience to effects of stress-
inducing cognitive performance anxiety on the self-reported state attentional control.
Although this conclusion is unclear as to its association with stress resilience, the
beta band may be linked to the brain activity reflecting stress resilience as observed
by [23].
Curtis and Cicchetti studied the resilience and its associated hemispheric EEG
asymmetry during resting eyes-open and closed in maltreated and non-maltreated
children [11]. Findings suggested that resilient children had greater left hemispheric
activity in the central cortical area, C3 and C4. Asymmetrical activity in the parietal
cortices, favoring P3 location, was also observed in non-maltreated children.
812 R. A. Hasan et al.

Select ive at t ent ion

Sit uat ional Dem and Cognit ive Appraisal Physiological Arousal Inst rum ent al and coping behaviour

Fig. 1 Cognitive affective conceptual stress model adapted from [7]

2.2 Stress Resilience as a Flexible Adaptation Process

The cognitive affective stress management training program is a brief six-session

stress program designed to promote resilience using the “induced affect” procedure
by imagining stressful situations [7]. The procedure allows patients to experience and
learn to control the negative effects of stress, and thus enhance stress resilience. The
theoretical framework of this training program can be explained using the cognitive
affective conceptual stress model [28], shown in Fig. 1, that relates the cognition,
physiological responses and behaviour towards stress [7]. When a person faces a
stressful situation, the brain responds by paying selective attention to important
stimuli of the situation and appraising how demanding and relevant the situation
is to their well-being, and what resources do they have to cope with the situation.
Physiological arousal is the response to what has been perceived earlier and involves
the mobilization of resources to deal with the stressful situation. During this process,
the level of emotional intensity is also fed back to cognitive appraisal and therefore
influence the appraisal component. In response to the cognitive appraisal and physi-
ological arousal, a person takes behavioural actions as a coping mechanism towards
the stressful situation. The actions can orient towards palliative emotion-regulating
or direct problem-solving techniques [28–30].
Throughout this process, stress resilience can be defined as the flexible adapta-
tion process where an individual mobilizes their resources to healthily cope with
current and future adversities. We hypothesize that the EEG-features extracted from
these three components correlate with the psychological measure of stress resilience:
cognitive appraisal, physiological arousal, and behavioural tendencies.

2.3 Proposed EEG-Signal Analysis

With reference to the cognitive affective conceptual stress model, we suggest

inducing stress to observe the elicitation of stress resilience using EEG brain activity.
This paper proposes the pretest–posttest quasi experimental study design where the
brain activity will be recorded from subjects at three time-points; (1) at relax-state
as-signed as control condition, (2) during stress-induced training as intervention
condition, and (3) at relax-state post-intervention. This study design is similar to the
within-subjects experiment where the same subjects are assigned to both control and
Finding the EEG Footprint of Stress Resilience 813

intervention conditions [31]. The sample size estimated for this experiment is N =
29, calculated as follows [32]:
 2
2 z α + z β (1 + (n − 1)ρ)
N=
nδ 2
N = 29

where
N = sample size.
Zα = 1.96 at level confidence of 95%
Zβ = 0.842 at 80% power.
n = 3, number of time points.
ρ = 0.5, assumed correlation of repeated measure.
δ = 0.6, effect size.
To address our hypothesis, the flow of EEG-signal analysis shown in Fig. 2 is
proposed for our study. We aim to identify the prominent EEG bands and hemispheric
asymmetries that may reflect the adaptation process of stress resilience. We also
aim to identify the brain networks between stress-inducing event and stress-induced
training using EEG source-space connectivity method.

Fig. 2 Proposed approach

Start
for EEG features of stress
resilience

EEG data on stress-induced training

Extract significant frequency band

Compare hemispheric asymmetry

Identify brain network

Correlation with psychological resilience score

Feature extraction of stress resilience

End
814 R. A. Hasan et al.

3 Expected Findings of EEG Footprint for Stress Resilience

In understanding the neural activity of resilience, our approach of identifying the

EEG features from brain rhythm, hemispheric asymmetry and brain network can
be used to provide a physiological measure for resilience. The brain rhythms and
network extracted from EEG signal have been strongly documented in literatures to
reflect different brain activities such as aversive response towards stress [33–35] and
coping behaviour [34, 35]. Hence, we hypothesize that resilience at resting state and
during training will elicit different EEG bands, and that these significant bands are
functionally link and form the brain networks of stress resilience.

4 Summary

Prolonged stress in working environment can lead to reduced quality of job perfor-
mance and quality of life. Literatures have found that mental health is associated
with lower stress level and higher resilience. As resilience is defined as the ability to
cope and adapt to stressful situation, there is a need to identify the EEG-neural corre-
lates of this psychological parameter when stress in being induced. In addition to
its advantages of being non-invasive, having high temporal resolution, low cost and
ease-of-portability, the EEG modality has been extensively used in monitoring stress
responses. Through this study, we would be able to identify the features of stress
resilience and correlate them to the conventional psychological measures. Similar to
the application of EEG-feature of stress response, these potential features may be
used in clinical practices for risk assessment, diagnosis and interventions.

Acknowledgements This research is supported by Ministry of Higher Education Malaysia (MoHE)

under Higher Institution Centre of Excellence (HICoE) Scheme awarded to Center for Intelligent
Signal and Imaging Research (CISIR).

References

1. Harnois G, Gabriel P (2002) Mental health and work: impact, issues and good practices. World
Health Organization
2. Kassim MSA, Ismail A, Ismail R (2018) A review of occupational stress prevalence and
its predictors among selected working populations in Malaysia. Malays J Publ Health Med
18(2):1–6
3. Fink G (2016) Stress: Concepts, Cognition, Emotion, and Behavior, ser. In: Handbook of Stress
Series. Academic Press, vol. 1
4. Feltoe G, Beamish W, Davies M (2016) Secondary school teacher stress and coping: insights
from Queensland, Australia. Int J Arts Sci 9(2):597–608
5. Menon J (2016) Depression on the rise as Malaysians burn out from stress, expert warns.
MalayMail. [Online]. Available: https://www.malaymail.com/news/malaysia/2016/03/24/dep
ression-on-the-rise-as-malaysians-burn-out-from-stress-expert-warns/1086087
Finding the EEG Footprint of Stress Resilience 815

6. Thye TSLL (2016) Work stress and mental health. New Straits Times. [Online]. Available:
https://www.nst.com.my/news/2016/04/142074/work-stress-and-mental-health
7. Smith RE, Ascough JC (2016) Promoting emotional resilience: cognitive-affective stress
management training. The Guilford Press, New York
8. Franklin T, Saab B, Mansuy I (2012) Neural mechanisms of stress resilience and vulnerability.
Neuron 75(5):747–761. [Online]. Available: http://www.sciencedirect.com/science/article/pii/
S0896627312007532
9. Davydov DM, Stewart R, Ritchie K, Chaudieu I (2010) Resilience and mental health. Clin
Psychol Rev 30(5):479–495. [Online]. Available: http://www.sciencedirect.com/science/art
icle/pii/S0272735810000437
10. Connor KM, Davidson JR (2003) Development of a new resilience scale: the connor-davidson
resilience scale (CD-RISC). Depression Anxiety 18(2):76–82. [Online]. Available: https://onl
inelibrary.wiley.com/doi/abs/https://doi.org/10.1002/da.10113
11. Curtis WJ, Cicchetti D (2007) Emotion and resilience: a multilevel investigation of hemispheric
electroencephalogram asymmetry and emotion regulation in maltreated and nonmaltreated
children. Dev Psychopathol 19(3):811–840
12. Burton NW, Pakenham KI, Brown WJ (2010) Feasibility and effectiveness of psychosocial
resilience training: a pilot study of the ready program. Psychol Health Med 15(3):266–277,
pMID: 20480432. [Online]. Available: https://doi.org/10.1080/13548501003758710
13. Keynan JN, Cohen A, Jackont G, Green N, Goldway N, Davidov A, Meir-Hasson Y, Raz
G, Intrator N, Fruchter E, Ginat K, Laska E, Cavazza M, Hendler T (Jan 2019) Electrical
fingerprint of the amygdala guides neurofeedback training for stress resilience. Nat Hum Behav
3(1):63–73. [Online]. Available: https://doi.org/10.1038/s41562-018-0484-3
14. McEwen BS (2006) Protective and damaging effects of stress mediators: central role of the
brain. Dialogues Clin Neurosci 8(17290796):367–381. [Online]. Available: https://www.ncbi.
nlm.nih.gov/pmc/articles/PMC3181832/
15. Goodman RN, Rietschel JC, Lo L-C, Costanzo ME, Hatfield BD (2013) Stress, emotion regu-
lation and cognitive performance: the predictive contributions of trait and state relative frontal
EEG alpha asymmetry. Int J Psychophysiology 87(2):115–123. [Online]. Available: http://
www.sciencedirect.com/science/article/pii/S0167876012005934
16. Kong F, Wang X, Hu S, Liu J (2015) Neural correlates of psychological resilience and their
relation to life satisfaction in a sample of healthy young adults. NeuroImage 123:165–172.
[Online]. Available: http://www.sciencedirect.com/science/article/pii/S1053811915007296
17. Peres JF, Foerster B, Santana LG, Fereira MD, Nasello AG, Savoia M, Moreira-Almeida A,
Lederman H (2011) Police officers under attack: resilience implications of an fMRI study. J
Psychiatr Res 45(6):727–734. [Online]. Available: http://www.sciencedirect.com/science/art
icle/pii/S0022395610003225
18. Reynaud E, Guedj E, Souville M, Trousselard M, Zendjidjian X, Khoury-Malhame ME, Fakra
E, Nazarian B, Blin O, Canini F, Khalfa S (2013) Relationship between emotional experi-
ence and resilience: an fmri study in fire-fighters. Neuropsychologia 51(5):845–849. [Online].
Available: http://www.sciencedirect.com/science/article/pii/S0028393213000134
19. Meir-Hasson Y, Keynan JN, Kinreich S, Jackont G, Cohen A, Podlipsky-Klovatch I, Hendler
T, Intrator N (May 2016) One-class fmri-inspired EEG model for self-regulation training. PLOS
ONE 11(5):e0154968. [Online]. Available: https://doi.org/10.1371/journal.pone.0154968
20. Scarapicchia V, Brown C, Mayo C, Gawryluk JR (2017) Functional magnetic resonance
imaging and functional near-infrared spectroscopy: insights from combined recording studies.
Frontiers Human Neurosci 11:419. [Online]. Available: https://www.frontiersin.org/article/
https://doi.org/10.3389/fnhum.2017.00419
21. Thibault RT, Lifshitz M, Birbaumer N, Raz A (2015) Neurofeedback, self-regulation, and brain
imaging: clinical science and fad in the service of mental disorders. Psychother Psychosom
84(4):193–207. [Online]. Available: https://www.karger.com/DOI/https://doi.org/10.1159/000
371714
816 R. A. Hasan et al.

22. Zhang M, Zhou H, Liu L, Feng L, Yang J, Wang G, Zhong N (2018) Randomized EEG
functional brain networks in major depressive disorders with greater resilience and lower rich-
club coefficient. Clin Neurophysiol 129(4):743–758. [Online]. Available: http://www.scienc
edirect.com/science/article/pii/S1388245718300324
23. Paban V, Modolo J, Mheich A, Hassan M (2019) Psychological resilience correlates with
EEG source-space brain network flexibility. Netw Neurosci 3(2):539–550. [Online]. Available:
https://doi.org/10.1162/netn_a_00079
24. Bassett DS, Wymbs NF, Porter MA, Mucha PJ, Carlson JM, Grafton ST (2011) Dynamic
reconfiguration of human brain networks during learning. Proc National Acad Sci 108(18):
7641–7646. [Online]. Available: https://www.pnas.org/content/108/18/7641
25. Mattar MG, Betzel RF, Bassett DS (July 2016) The flexible brain. Brain 139(8):2110–2112.
[Online]. Available: https://doi.org/10.1093/brain/aww151
26. Braun U, Schäfer A, Walter H, Erk S, Romanczuk-Seiferth N, Haddad L, Schweiger JI, Grimm
O, Heinz A, Tost H, Meyer-Lindenberg A, Bassett DS (Sept 2015) Dynamic reconfiguration
of frontal brain networks during executive cognition in humans. Proc National Acad Sci US
Am 112(26324898):11678–11683. [Online]. Available: https://www.ncbi.nlm.nih.gov/pmc/art
icles/PMC4577153/
27. Putman P, Verkuil B, Arias-Garcia E, Pantazi I, van Schie C (Jun 2014) EEG theta/beta ratio as
a potential biomarker for attentional control and resilience against deleterious effects of stress
on attention. Cogn Affect Behav Neurosci 14(2):782–791. [Online]. Available: https://doi.org/
10.3758/s13415-013-0238-7
28. Lazarus RS, Folkman S (1984) Stress. Springer Publishing Company, Appraisal and Coping
29. Kyriacou C (2001) Teacher stress: directions for future research. Educ Rev 53(1):27–35
30. Research H (2015) Supporting teachers during times of change and stress. Hanover Research
31. Jhangiani RS, Chiang I-CA, Price PC (2015) Research Methods in Psychology. BCcampus
Open Education
32. Hedeker D, Gibbons RD, Waternaux C (1999) Sample size estimation for longitudinal designs
with attrition. J Educ Behav Stat 24:70–93
33. Seo S-H, Lee J-T (2010) Stress and EEG. In: Crisan M (ed) Convergence and hybrid information
technologies. Intech, ch 27, pp 413–426
34. Brouwer A-M, Neerincx MA, Kallen V, van der Leer L, ten Brinke M (2011) EEG alpha
asymmetry, heart rate variability and cortisol in response to virtual reality induced stress. J
Cyberther Rehabil 4(1):27–40
35. Flo E, Steine I, Blågstad T, Grønli J, Pallesen S, Portas CM (2011) Transient changes in frontal
alpha asymmetry as a measure of emotional and physical distress during sleep. Brain Res
1367:234–249. [Online]. Available: http://www.sciencedirect.com/science/article/pii/S00068
99310021566
Modeling and Tuning of PID Controller
for Continuous Stirred Tank Reactor

A. Suguna, S. N. Deepa, and N. Rajasingam

Abstract Continuous Stirred tank reactor is a chemical reactor system which

exhibits complex non-linear dynamic characteristics. The quality of final product
is based on the design of the controller. The mathematical modeling of CSTR is
designed based on first principle method. Conventional PID controllers Ziegler-
Nichols, Tyreus-Luyben, Cohen-Coon and IMC based PID have been imple-
mented and the performance analysis of different PID tuning methods is done. The
performance of the PID controller is analyzed in MATLAB simulation.

Keywords CSTR · PID controller · Tuning

1 Introduction

Continuous stirred tank reactor which is widely used equipment mainly in the chem-
ical industry, exhibits highly nonlinear behavior. A CSTR is a exothermic irreversible
reaction, a fluid stream is continuously fed to the reactor. Since the reactor is perfectly
mixed the exit stream has the same concentration and temperature as the reactor fluid.
The jacket surrounding the reactor has feed and exit streams. The jacket is assumed
perfectly mixed, energy passes through the reactor walls into the jacket, removes
the heat generated by reaction. The mathematical model is developed from material
balance equation [1]. The two state variables are the reactor concentration and reactor
temperature. PID control has three modes i.e., proportional, integral and derivative.
The proportional term is proportional changes for error (the difference between the
set point and process variable) to the control output. The integral term examines
the process variable over time and it corrects the output by reducing the offset from

A. Suguna
Government College of Technology, Coimbatore, India
S. N. Deepa
Anna University Regional Campus, Coimbatore, India
N. Rajasingam (B)
Dr. N. G. P. Institute of Technology, Coimbatore, India

© The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 817
R. Ibrahim et al. (eds.), International Conference on Artificial Intelligence
for Smart Community, Lecture Notes in Electrical Engineering 758,
https://doi.org/10.1007/978-981-16-2183-3_77
818 A. Suguna et al.

process variable. Derivative control mode monitors the rate of change of the process
variable and therefore changes the output when there are unusual variations. Each
parameter of the three control functions adjust to obtain the desired performance
from the process [2]. Easy implementation and maintenance due to their simple struc-
ture, PID controllers are the most extensively used controllers in the process control
industry. Different types of controllers are available like sliding mode control, robust
control. But the PID controllers still used in industries, Conventional tuning methods
of PID controller used for this CSTR process.

2 Mathematical Model of CSTR

The mathematical model for the Continuous Stirred Tank Reactor is derived from
mass balance and energy balance equation [3]. Mass balance equation is given by
the rate of accumulation for material is equal to difference between rate of material
inflow and outflow.
dV C A
= FC A0 − FC A − r V (1)
dt
Energy balance equation is given by
   
d v P C p (T − T r e f ) /dt = Fρ C p T f − T r e f − Fρ C p (T − T r e f )
+ (−H )Vr − U A(T − Ti ) (2)

State variable form of equations can be written as

dC A F 
f 1 (C A , T ) = = CAf − CA − r (3)
dt V
 
dT F  −H UA
f 2 (C A , T ) = = Tf − T + r− (T − Ti ) (4)
dt A pC p vpC p

The reaction rate per unit volume r (Arrhenius expression) is given by

 
−E
r = K 0 ex p CA (5)
RT

State variables are concentration and temperature. At steady state dC A

dt
= 0 and
dT
dt
=0
 
F  −E
f 1 (C A , T ) = 0 = C A f − C A − K 0 ex p CA (6)
V RT
Modeling and Tuning of PID Controller for Continuous Stirred … 819

   
F  −H −E UA
f 2 (C A , T ) = 0 = Tf − T + K 0 ex p CA − (T − Ti )
V pC p RT vpC p
(7)

The controlled variable is temperature which is changed by regulating the coolant

flow rate. CSTR is a non-linear system, the linearization has be done using Jacobian
matrix form
   
∂ f1 ∂ f1 
∂ x1 ∂ x2
− VF − K S −K S C AS
A = ∂ f2 ∂ f2 =
∂ x1 ∂ x2
H
pC p S
K − VF − VUp CAp + (−H
pC p
) 
KSCA
 
−E (8)
K S = K 0 ex p
RT
 ∂ KS
kS =
∂T
   
∂ f1
0
B = ∂u 1
∂ f2 = UA (9)
∂u 2 vpC p

where, x1 - Reactor concentration, x2 - reactor temperature and u2 - coolant flow

rate.
Substituting the steady state operating data the state space matrices can be obtained
as follows
 
0.0121 −0.0000 0
A= B= C = 0 1 D = [0]
2.2291 −0.0080 7

The Steady state operating data are available in [4].

3 PID Controller

3.1 Ziegler-Nichols Tuning (ZN)

This paper presents ZN closed loop tuning method. It requires the determination of
the ultimate gain and ultimate period [5] By adjusting the controller gain (Ku) till the
system goes sustained oscillations (ultimate gain) this can be achieved, while main-
taining the integral time Constant (Ti) at infinity and the derivative time constant
(Td) at zero. Note down ultimate gain-Ku (gain at oscillation occurs) and ultimate
period—Pu. The tuning parameters Proportional gain, Integral Time and Derivative
Time are calculated from prescribed ZN tuning rules as mentioned in the Table.1 [6].
820 A. Suguna et al.

Table 1 Tuning rules of PID

Tuning methods KP TI Td
controller
Ziegler and Nichols 0.6Ku 0.5Pu 0.125Pu
Ku Pu
Tyreus- Luyben 2.2 2.2Pu 6.3

32+ 6τ
1 T 4 τ T
Cohen and coon K τ 3 4T τ τ 4
13+ 8τ
T 11+ 8τ
T

(τ +0.5T ) τT
IMC based PID K p (λ+0.5T ) τ + 0.5T 2τ +T )

3.2 Tyreus-Luyben Method (TL)

B.D Tyreus and W.I Luyben in 1997, proposed this method. To calculate the ultimate
gain Ku and ultimate period Pu the same Ziegler Nichols tuning procedure have to
be followed. Controller parameters calculated from the rules given below.

3.3 Cohen and Coon Method (CC)

The other name of Cohen and Coon method is Process reaction curve method, the
step input is given to the open loop response of the system, the S shaped curve
obtained. From this curve gain, dead time and time constant calculated, Substituting
this parameters into the tuning rules controller tuning parameters computed.

3.4 IMC

To model CSTR process by fitting the open-loop step test data as a first order function
with time delay, the internal model control-PID (IMC-PID) tuning rules have the
advantage of using a single tuning parameter to achieve a clear tradeoff between
closed-loop performance and robustness to model inaccuracies.
The IMC-PID controller provides good set-point tracking but has a sluggish distur-
bance response [7]. The different tuning rules for PID controller as shown in Table
1.

4 Simulation and Results

Ziegler-Nichols, Tyreus-Luyben, Cohen-Coon and IMC based PID methods conven-

tional PID parameters are determined using the tuning rules given in Table 1. The
values of PID controller parameters for different tuning rules are identified in Table
Modeling and Tuning of PID Controller for Continuous Stirred … 821

Table 2 Tuning parameters

Tuning methods KP TI Td
of PID controller
ZN 10.8 1.5 0.375
TL 12.1 1.9 0.25
CC 3.7 1.5 0.0388
IMC based PID 2.109 12.607 0.0157

Table 3 Time domain

Tuning methods Rise time Peak overshoot Settling time
specifications of PID
controller ZN 0.23 25.8 2
TL 0.0852 30.8 1.5
CC 0.0359 33.6 0.488
IMC based PID 0.121 5.61 0.41

2. The unit step response for the close loop control of CSTR is given to get the output
response.
Closed loop response parameters are analyzed using the time domain specifica-
tions of all PID tuning methods and shown in Table 3. To use these empirical tuning
rules to CSTR process, we need to fit the data to a first order with dead time. First
order with dead time system obtained the approximation of the CSTR by giving step
signal input.
Figures 1, 2, 3 and 4 represents the closed loop response of the CSTR process based
on the PID tuning parameters of Ziegler-Nichols, Tyreus—Luyben, Cohen-Coon and
IMC based PID methods.

Fig. 1 Response of
CSTR-Ziegler Nicholas
822 A. Suguna et al.

Fig. 2 Response of
CSTR-Tyrus Luyben

Fig. 3 Response of
CSTR-Cohen and Coon

5 Conclusion

Comparison of different tuning methods of PID controller for temperature control

of Continuous Stirred Tank Reactor has been implemented. The tuning methods
are implemented in the CSTR process transfer function obtained from first order
principle method. Simulation results are analyzed by giving the step input to get the
output response using MATLAB. These tuning methods are analytical; Based on the
Modeling and Tuning of PID Controller for Continuous Stirred … 823

Fig. 4 Response of
CSTR-IMC based PID

analyze IMC based PID controller provide good transient response for the CSTR
process.

References

1. Ogata K, Yang Y (2002) Modern control engineering. Prentice hall, India

2. Tyreus BD, Luyben WL (1992) Tuning PI controllers for integrator/dead time processes. Ind
Eng Chem Res 31(11):2625–2628
3. Bequette BW (2003) Process control: modeling, design, and simulation. Prentice Hall Profes-
sional
4. Kamala N (2011) Studies in modeling and design of controllers for a nonideal continuous stirred
tank reactor
5. Cominos P, Munro N (2002) PID controllers: recent tuning methods and design to specification.
IEE Proc-Control Theory Appl 149(1):46–53
6. George S (1984) Chemical process control: An introduction to theory and practice. PTR Prentice
Hall, Inc. 3(9015):03983
7. Shamsuzzoha M, Lee M, Lee J (2005) IMC-PID controller tuning for improved disturbance
rejection of unstable time delay processes. Theories and application of Chemical Engineering,
vol 11
Adaptive Security for Cognitive Robotic
Process Automation in Enterprise
Computing Using AI-Powered Analytical
Engine

Mohamed Ibrahim Beer Mohamed and Mohd Fadzil Hassan

Abstract The robotic process automation (RPA) in enterprise computing refers to

software bots that are capable of mimicking most of the human–computer interac-
tions to carry out day-to-day business operations. This RPA is targeted for automating
rule-based repetitive and high-volume tasks with higher accuracy which eventually
reduces the operation cost and processing time. The cognitive automation which is
being developed as part of enterprise computing automation best utilizes Artificial
Intelligence (AI) and Machine Learning (ML) techniques to enhance the process
model of RPA in terms of improved accuracy, reliability, and consistency in taking
intelligent business decisions with no or minimal human interventions. The majority
of the communications between computing nodes and bots are being provisioned
through service-oriented enterprise computing interfaces, which are built based on
Service-Oriented Architecture (SOA). The SOA by itself does not possess any secu-
rity layer and it defaults to the Open Systems Interconnection (OSI) model for security
that is inadequate in this modern era of process automation and interfacing. In this
paper, the security concerns of Cognitive RPA for enterprise computing are analyzed,
and a novel approach is presented for adding-up the security layer for Cognitive RPA.
This is an adaptive approach which works on the predict-prevent-learn pattern for
effective proactive security as differed from traditional reactive security, where the
Artificial Intelligence (AI) techniques are used for ‘predict’ part for predicting poten-
tial security threats, the Artificial Neural Networks (ANN) techniques are applied
for ‘learn’ part on unsupervised learning of anticipated security vulnerabilities, and
security prevention algorithms are equipped for ‘prevent’ part to defend against the
security threats on Cognitive RPA systems.

Keywords Artificial Intelligence · RPA · Security · Cognitive automation · ANN

M. I. Beer Mohamed (B) · M. F. Hassan

Department of Computer and Information Sciences, Universiti Teknologi PETRONAS, 32610
Seri Iskandar, Malaysia

© The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 825
R. Ibrahim et al. (eds.), International Conference on Artificial Intelligence
for Smart Community, Lecture Notes in Electrical Engineering 758,
https://doi.org/10.1007/978-981-16-2183-3_78
826 M. I. Beer Mohamed and M. F. Hassan

1 Introduction

The robotic process automation (RPA) is being focused on enterprise computing

to minimize the operational cost and processing time. The cognitive RPA allows
enterprises to automate tasks that usually need human intervention for interpretation
and decision making. As a result of Cognitive RPA, the processes which have been
performed by humans are assigned to software modules, named “bots”. These “bots”
are equipped with Artificial Intelligence (AI) techniques for real-time analysis and
thus to take cognitive decisions resulting in minimal or no human interactions. While
RPA is more focused on data-intensive processes, the Cognitive RPA concerns with
logic-based outcomes [1].
RPA is effective for the automating processes which are (i) rule-based, (ii)
repeated/pre-defined trigger, (iii) definitive inputs/outputs, and (iv) large-scale in
volume. With the Cognitive RPA, an ‘intelligent’ part is added to these processes
which result in the following significant benefits [2, 3]. Moreover, software bots are
easy to deploy and integrate with existing business application modules seamlessly.
• Minimal human intervention: The AI-powered software bots replaced human
interactions for performing intellectual tasks.
• Better accuracy: The results of software bots are more accurate in analytics and
decision makings on big data consistently.
• Improved compliance: The software bots are more controlled and operated in
compliance with organizational regulations and policies.
• Quick wins on cost savings: As a quick win, most of the organizations experience
more than 60% cost saving in less than 12 months of time after implementation
of Cognitive RPA [4]. This cost savings rate increased over the years as only it
requires minimum operating costs.
• Scalability: Comparing to the human-based operating model, bots are scalable on
an as-needed basis where they can be added or reduced based on workload.
A comprehensive cognitive RPA bot should be equipped with an array of
functional and non-functional entities such as reliability, scalability, speed, light-
weighted, manageability, intelligence, and more importantly security [3, 4].
Despite the extended benefits, cognitive ROA introduces a new attack surface
to enterprise computing in addition to the existing security risks of integration on
process automation which results in the paramount importance of having RPA secu-
rity. Applying security for cognitive RPA is not as simple as implementing a reactive
way of the security establishment. It requires that the security should be applied across
the entire product life cycle from design, architecture, interfacing, and implementa-
tion to underlying business operations satisfying essential security goals, including
integrity, traceability, confidentiality, and control.
In this paper, the security concerns of the cognitive RPA are studied from an enter-
prise computing perspective. As a novelty, a security component named “Intelligent
Security Engine (ISE)” is constructed based on Artificial Intelligence (AI) techniques
for risk prediction and Artificial Neural Networks (ANN) for unsupervised learning
Adaptive Security for Cognitive Robotic Process Automation … 827

of security vulnerabilities for Cognitive RPA development and implementation, and

the same is presented as the adaptive way forward security layer for Cognitive RPA.
The remaining sections of this paper are organized as the rationale and motivation
of the study is briefed in Sect. 2, the proposed AI-based security approach is presented
in Sect. 3, a brief discussion on the Proof-of-Concept is given in Sect. 4, and Sect. 5
concludes the paper.

2 Rationale for the Study

This section briefs on the rationale and motivation behind the conducted research
study. The security concerns and the potential vulnerabilities of the cognitive robotic
process automation are reviewed in terms of available literature, and the mandatory
need for a proactive security model is highlighted as the traditional reactive security
model is inadequate in this modern era of cognitive automation with the advent of
related software and hardware technologies.

2.1 Security Risks in Cognitive RPA

In addition to the common enterprise application integration risks, the potential risks
for cognitive RPA are listed below [5, 6].
• Privilege escalation: It is the risk of abusing the given access privilege given to a
bot.
• Data tampering: It is the risk of disclosing sensitive data that are tampered with
during inter-process communication.
• Denial of Service: It is the risk of stressing with the bot with a greater number of
fake requests so that it will deny or delay in servicing to even legitimate requests.
• Application vulnerabilities: It is the risk of enacting the security threat using the
existing vulnerabilities in the application components and/or supported software
modules that are deployed on bots.
• Abusing the digital identity and access control: It is the risk of illegitimately
applying digital federated identity and access privileges that are given to specific
bots.
• Data identification and protection: It is both business and operations risk of sensi-
tive data handling by robotics and automation tools adhering to compliance with
the governance body.
828 M. I. Beer Mohamed and M. F. Hassan

2.2 Security Vulnerabilities on Cognitive RPA

In addition to the technology and implementation constraints, multiple business and

managerial aspects could lead to security vulnerabilities, including the following [7,
8].
• Inadequate security controls on governing bots
• Lack of business continuity on workflow management
• Single point of failure, service unavailability, and machine errors on processing
• Unfortunate vulnerabilities in data processing and insufficient data protection
• Inconsistency on results among processes
• Inability of handing exceptions during the automation process
• Insufficient authentication/authorization controls on inter-process communica-
tions
• Unpredictable patterns on compliance and regulatory implications

2.3 Adaptive Security Approach for Cognitive RPA

The potential security threats for RPA are increasing as the attack surface is widening
due to the advent of the latest automation and supported technologies. Hence, the
security for RPA became a mandatory element in the business operating model. The
traditional reactive way of security threat identification and perimeter defense strate-
gies are obsolete right now [9–11]. It should be the paradigm shift with continuous
monitoring and response to threats, shifting from reactive to a proactive approach to
handling and preventing security attacks with the support of Artificial Intelligence
(AI) and related technologies.
An adaptive security approach applies a layered way of detecting and preventing
security threats. Hence it is focused to sense, detect, and respond to the prevention
of security threats before the potential vulnerabilities are being used by the attackers
on performing attacks for the exploitation of enterprise assets.
The adaptive security approach focuses on security aspects before, during, and
post-execution of automated processes, targeted for (i) reducing the attack surface
and attack velocity, (ii) real-time event-based monitoring on identifying security
breaches, (iii) assurance of reliability and availability of automation resources, (iv)
ensuring reliability in automation processes, (v) prevention of data and identity loss,
and (vi) anticipated security thereat for earlier prevention.
The artificial neural networks (ANN) approach differs from conventional compu-
tation techniques in problem-solving as the conventional techniques apply the stan-
dard procedural and algorithmic approach with a pre-defined set of instructions where
the ANN is being used to solve the problems which cannot be expressed exactly in
terms of algorithmic steps. Security for cognitive-based solutions is one kind of
such problem that can be effectively solved only through AI-based analytics and
implementation [12].
Adaptive Security for Cognitive Robotic Process Automation … 829

3 Proposed Security Approach

The proposed adaptive security approach is built on the Predict-Prevent-Learn

pattern, as depicted in Fig. 1. In the ‘predict’ stage, the security risks are identi-
fied and anticipated in the underlying cognitive RPA implementation, the ‘prevent’
stage hardens the system defending against the identified and anticipated security
attacks without jeopardizing the industrial security standards. The ‘learn’ stage built
the knowledge base on the security risks and vulnerabilities which is sourced by
unsupervised learning processes.
The adaptive approach is applied and built as a software module, named “Intel-
ligent Security Engine (ISE),” which can be attached to any computing node
with minimal configuration. Basically, this ISE is an AI-powered analytical engine
equipped with threat modeling and prevention. The cross-cutting view of ISE is given
in Fig. 2 that applies the existing security standards for authentication/authorization,
analytical rule-based decision engine, organization-specific security policies, and
token management which are powered by AI and ANN implementations.
The data for analytical decision making on the severity of security threats and
protection are provisioned by knowledge base which is sourced by Artificial Neural
Networks (ANN) learning. The computational model of an artificial neuron is
inspired by the biological neuron. A comparison of a biological neuron and an artifi-
cial neuron is given in Fig. 3, where the biological neuron works on electrochemical
signals, and the artificial neuron works based on the mathematical model. In a biolog-
ical neuron, the cell body is called “Nucleus”, branching input as “Dendrite”, and
branching output as “Axon”. The synapses connect dendrites to axons for propa-
gating electrochemical signals through the nucleus. The neuron fires only its input

Fig. 1 Adaptive security

approach for cognitive RPA
830 M. I. Beer Mohamed and M. F. Hassan

Fig. 2 Cross-cutting view of Intelligent Security Engine (ISE)

Fig. 3 Biological neuron versus artificial neuron

Adaptive Security for Cognitive Robotic Process Automation … 831

signals meet the threshold limit in a short period. The transmitted electrochemical
signal from dendrites to an axon via the nucleus will act as an input dendrite signal
for another neuron.
The artificial neuron collects input signals as values that are being multiplied by
the factor, named weights which represent the strength of the respective input signals.
Then, these values are fed to the mathematical function (aka transfer function) which
determines whether to activate the neuron or not. If the weights to the input signals
are high, then the corresponding neuron is strongly connected. The mathematical
function is equivalent to the nucleus in a biological neuron, and the weights are similar
to synapses which can hold positive, negative, or neutral (zero) values. The learning
in ANN is the process of adjusting these weights for desired functionality such as
threat identification and pattern matching. For this research study, the hyperbolic
tangent sigmoid mathematical model is taken as a transfer function which has the
representation as below.

ex − e−x
f(x) =
ex + e−x
∂E
The gradient of each weight concerning the desired output is measured as, ∂w ij
where E represents the sum of input values which are multiplied by their corre-
sponding assigned weights, and wij represents the actual assigned weight on an edge
which is connecting node i to node j. The backpropagation performs this gradient
measurement on each edge and adjusts the weights incrementally using stochastic
∂E
gradient descent as wi j = wi j − η ∂w ij
for all the nodes in the defined neural network.
The process flow of the proposed AI-powered threat prevention solution involves
a systematic approach as outlined in Fig. 4.
The proposed security component “ISE” as depicted in Fig. 2 embeds this process
model as given in Fig. 4. This software component can be easily added with any
RPA processing nodes through an interceptor design pattern with a simple config-
uration where the input/output messages will be intercepted for threat monitoring
and prevention. Though this model is good enough on learning the existing security
vulnerabilities and threats, the training part will require a significant amount of time
to build the knowledge base. Instead, the data mining techniques can be used to
construct the initial knowledge base where it will be further refined by neural nets.

4 Result and Discussion

For proof-of-concept (PoC) purposes, the proposed security approach is implemented

in a large-scale banking environment with real-time data on cognitive robotic process
automation of customer onboarding and cheque clearance modules. The new to
bank (NTB) customers are onboarded to the bank by the automated bots without
any human intervention, and bots are automated to perform the entire customer
832 M. I. Beer Mohamed and M. F. Hassan

Collect data on Cognitive

RPA Security Threats

Build the neural network & Train the network with

set the expectation backpropagation

Process data for pattern Test the result, Re-train &

matching & modelling Benchmark

Analysis & Interpret data Add Neural Nets as a security

for threat categorization component to Threat Model

Fig. 4 Process flow of the proposed AI-powered security approach

on-boarding journey from due diligence checking, fraud monitoring, KYC updates,
opening up the banking accounts, and registration for debit/credit cards, up to instruct
the customer upon the account is ready for operation. The cheque clearance bot
performs an intelligent check on signature matching, verification of the given details
in the cheque, validation of transactions, and perform the fund transfer.
The artificial neural network model with back-propagation is chosen for the
learning part and implemented for fraud and security risk assessment purposes, as
part of PoC implementation. This ANN-based model produces reliable results when
compared to other forecasting mechanisms.
As obtained from the conducted PoC, the statistical data on security attack and
prevention ratio with existing (without AI) and proposed (AI-powered) security solu-
tions on the chosen attack vector on cognitive RPA is given in Table 1. The ratio is
represented as a failure of protection to protect from security attacks, for example, the
ratio of 0.6000: 0.4000 represents 60% of failure on protection and 40% of protection
from the underlying security attacks.

Table 1 Analysis of proof-of-concept (PoC) result

S.No Attack vector Existing security Proposed security
1 Privilege escalation attack 0.5771: 0.4229 0.1083: 0.8917
2 Data tampering attack 0.5469: 0.4531 0.1854: 0.8146
3 Denial of Service attack 0.3442: 0.6558 0.0964: 0.9036
4 Man-in-the-middle attack 0.5922: 0.4078 0.2357: 0.7643
5 Abusing digital identity 0.6230: 0.3770 0.1323: 0.8677
6 Broken access control 0.2123: 0.7877 0.2066: 0.7934
Adaptive Security for Cognitive Robotic Process Automation … 833

By analyzing the PoC results, the observed sensitivity rate falls above 75% for
the proposed AI-powered security solution. Except for the security protection for
broken access control, the protection rate of the proposed security solution is at least
35% higher than the existing security solutions. Moreover, a higher protection rate
for broken access control can be achieved through customized machine learning
processes. The graphical representation of the obtained result as given in Table 1
is portrayed in Figs. 5, 6 and 7. With this analysis, it is clear that the proposed AI-
powered security solution for cognitive RPA supersedes the existing non-AI based
security solutions with a higher protection rate.

Fig. 5 Security protection view of PoC result

Fig. 6 Security attack view of PoC result

834 M. I. Beer Mohamed and M. F. Hassan

Fig. 7 Security protection view of PoC results in a line graph

5 Conclusion

The cognitive robotic process automation is the modern trend in enterprise computing
and automation targeted for cost reduction and higher throughput in a minimum
amount of time. With cognitive RPA, the automated tasks are getting completed
with more accuracy and expected scalability. However, the security boundaries are
getting expanded due machine to machine interfacing and end-to-end automation
of business processes without any human interventions. In this paper, the major
security risks on cognitive RPA are studied in the view of available literature, a navel
security software component is constructed and presented in an adaptive way forward
approach. This security approach is equipped with artificial intelligence and related
technologies and best suited for this current era of business process automation. The
proposed approach is practically implemented and tested as part of proof-of-concept,
the obtained results are analyzed, and proved that the proposed AI-powered security
solution provides at least 35% higher security than the existing non-AI based security
solutions.

References

1. Martínez-Rojas A, Barba I, Enríquez JG (2020) Towards a Taxonomy of Cognitive RPA

Components. In: The proceedings of Springer international conference on business process
management, pp 161–175
2. Lamberton C, Brigo D, Hoy (2017) Impact of robotics, RPA, and AI on the insurance industry:
challenges and opportunities. J Financ Perspect 4(1)
3. Teli R, Prasad SK (2020) Delivering value in procurement with robotic cognitive automation
(RCA) Services. In: The IGI global proceedings of robotic systems: concepts, methodologies,
tools, and applications, pp 1773–1785
4. Danger JL, Guilley S, Hoogvorst P, Murdica C, Naccache D (2012) Low-cost countermeasure
against RPA. In: The proceedings of Springer international conference on smart card research
Adaptive Security for Cognitive Robotic Process Automation … 835

and advanced applications, pp 106–122

5. Syed R, Suriadi S, Adams M, Bandara W, Leemans S (2020) Robotic process automation:
contemporary themes and challenges. Computers in Industry, pp 1–55
6. Zaharia-Radulescu AM, Pricop CL, Shuleski D, Ioan AC (2017) RPA and the future of
workforce. In: Proceedings of the international management conference, vol 11(1), pp 384–392
7. Santos F, Pereira R, Vasconcelos JB (2019) Toward robotic process automation implementation:
an end-to-end perspective. Bus Process Manage J
8. Leshob A, Bourgouin A, Renard L (2018) Towards a process analysis approach to adopt
robotic process automation. In: IEEE 15th international conference on e-business engineering
(ICEBE), pp 46–53
9. Hassan MF (2016) Construction of customizable SOA security framework using artificial neural
networks. Jurnal Teknologi 78:69–75
10. Rudman R, Bruwer R (2016) Defining Web 3.0—opportunities and challenges. The Electronic
Library
11. Willcocks L, Hindle J, Lacity M (2019) Keys to RPA success. Executive Research Report,
Knowledge Capital Partners
12. Park S, Park H ANN based intrusion detection model. In: Workshops of the Springer
international conference on advanced information networking and applications, pp 433–437
Multidirection Features Based Image
Inpainting with Color-Direction
Patch-Sparsity

B. Vidhya and M. Nikhil Madhav

Abstract To preserve the texture clarity, coherence of structure and surrounding

consistence of the inpainted areas of an image in a better way, patch sparsity based
image inpainting approach through color direction is proposed. To approximate
the multi direction characteristics of a degraded image, super- wavelet transform
is utilized in this method. It combines with color details to build the weighted color-
direction distance (WCDD) for the determining the difference among two patches.
WCDD is used to interpret the color direction structure sparsity which gives a more
satisfying order of filling and more matching candidate patches for selection. Then,
the target regions are sparsely characterized by the numerous candidate patches
under surrounding consistency by considering the multi direction and color spaces
constrictions together. Experimental results prove that the removal of scratch, text
and objects are performed effectively through the proposed inpainting method.

Keywords Image inpainting · Multi direction feature · Color direction structure

sparsity · Wavelet transform

1 Introduction

Image inpainting is the method of restoring the target regions that are scratched,
missed or damaged. This process was first introduced to reconstruct the deteriorated
old pictures or artworks. It is also called as image disocclusion or completion. Image
inpainting has become a dynamic research subject in the area of computer vision and
image processing since the recent advancement of digital image usage. The ultimate
goal of image inpainting is to restore the lost region of a damaged picture in an
unnoticeable manner. Nowadays image inpainting are widely used for reconstructing
old pictures and restoring digital images.

B. Vidhya (B)
Dr. N. G. P Institute of Technology, Coimbatore, India
M. N. Madhav
Sri Shakthi Institute of Engineering and Technology, Coimbatore, India

© The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 837
R. Ibrahim et al. (eds.), International Conference on Artificial Intelligence
for Smart Community, Lecture Notes in Electrical Engineering 758,
https://doi.org/10.1007/978-981-16-2183-3_79
838 B. Vidhya and M. N. Madhav

The main classification of image inpainting methods are: the exemplar based,
the sparse based and the diffusion-based techniques. Partial Differential Equations
(PDE) are used in the diffusion based approach where the missing regions are filled
by the diffusion of image details from the source regions. The filling of missing
regions was achieved by the diffusion of the image details along the isophote direc-
tion by Bertalmio et al. [1]. Later Bertalmio et al. [2] presented the Navier–Stokes
equivalence in fluid dynamics into the process of image inpainting. T. F. Chan et al [3]
utilized the Total Variation (TV) model for reconstructing the damaged regions. Then,
T. F. Chan et al [4] anticipated the curvature-driven diffusion (CDD) equivalence to
improve the TV model that underperformed in the connectivity principle.
P-Harmonic energy minimization was introduced in the TV model for filling the
target regions by Zhang et al. [5], while other methods normally use surrounding
pixels to fill the target pixels. Telea used first marching method to complete the
target region by filling the known pixels of the neighbour from exterior to interior
region [6]. Weickert et.al [7] proposed a unified framework for interpolation built on
elliptic partial differential equations. Image was restored by Takeda et al. [8] through
the association of the nonparametric statistics field and expanded kernel regression
concepts. Many examinations were conducted by the researchers in the process of
image inpainting to attain an improved restoration result.
Reconstruction of non textured missing regions that are smaller in size achieved
by the diffusion-based algorithms have proven results. But its performance is poor
for smaller textured or structured portions because of inadequacy of semantic
texture/structure combination, and are unsatisfactory for larger missing regions.
The issues in image inpainting methods were addressed by the image sparse
representation. Here the illustration of the image is performed by the sparse union
of an over complete dictionary using DCT, wavelet, Contourlet, wavelet and so on
[9–13] and the approximation of the missing pixels are corrected adaptively by the
sparse representation. As an illustration, the damaged picture is separated into the
texture layers and cartoon layers, that are correspondingly sparsely represented by
Curvelet transforms and the Discrete Cosine Transform [9].
The problems in image inpainting can be addressed by the sparse representation
of the image. Guleryuz [14] and Fadili et al. [15] characterized an image by its
sparse combination using DCT, wavelets, Contourlet, Curvelet in the overcomplete
dictionary approach. Pixels that are missing are calculated dynamically by restoring
these sparse representation periodically. The target region or a damaged region in
an image is segregated into textural regions and structural regions that are character-
ized sparsely into Curvelet Transform and Discrete Cosine Transform respectively
[13]. Fadili et al. [16] presented an approach of expectation maximization (EM)
for image inpainting in Bayesian framework for restoring the coefficients by the
indorsing of prior penalty sparsity. All the above methods have limited generalization
of algorithms, that are improved by the adaptive dictionary learning approach.
The proposed approach concentrates on the image inpainting technique based on
matching performed by the patch sparsity. Here, multi-direction feature (MDF) using
patch sparsity for inpainting the image is proposed for preserving the clarity of the
textures and coherence of the structures. Patch sparse representation and structure
Multidirection Features Based Image Inpainting with Color … 839

sparsity are used to represent the patch sparsity in the images. Based on the concept
of algorithms used in the image inpainting algorithms, super-wavelet transform is
implemented for extracting the multi-direction feature of an image.
It will be blended with the color details to form a weighted color-direction distance
(WCDD) to estimate the homogeneity of patches. The value of WCDD calculated
between the known and target patches also focuses on preserving the structure coher-
ence through the color-direction structure sparsity (CDSS) function. For yielding
subspace for patch sparse representation, numerous reference patches are checked
using WCDD.
Additionally, an enhancement calculation with local patch uniformity limitations
in both the color and the multi direction spaces is created to attain sparse linear
combination factors of reference patches. The ultimate aim of this technique is to
preserve the consistency of the inpainted regions with respect to the neighboring
patches and to enhance the clarity of the structures and the textures of the images.
The key contributions of this article are: (i) to present the multi-direction feature
to the image inpainting algorithm; (ii) to implement weighted color-direction interval
to explore appropriate reference patches; (iii) to claim CDSS to get a strong order
of filling; and (iv) to integrate the multi-direction and color limitations into the
optimization measure for getting superior inpainting outcomes. This paper is ordered
in the following order. Section 2 explains the method of extracting multi-direction
characteristics of an image and the suggested exemplar based inpainting technique.
Experimental results are shown in Sect. 3 and possible future work is proposed in
the conclusions.

2 Proposed Algorithm

2.1 Extraction of Multi-direction Features in Image

Inpainting

Image inpainting is the process of filling the target regions based on the details of
neighboring regions. Image inpainting is mainly performed to preserve the structural
consistency and surrounding consistency. If the details of the source regions known
is higher, then the outcomes of the inpainting process will be better. For corrupted
images, gradient features or divergence characteristics were used by the existing
methods of image inpainting. Image consists of varied direction characteristics, so
the gradient function that defines only 2 directions will not be able to completely
define the details of an image.
The multidirectional and multiscale characteristics can be estimated using super-
wavelet transform. Hence in the proposed algorithm, for obtaining the multidi-
rectional attributes of an image, the super-wavelet transform is used. Among the
different types of super-wavelet transforms, Curvelet transform [17], is selected
840 B. Vidhya and M. N. Madhav

in the proposed system as it has proven results of extracting valid details accu-
rately than other transforms like Contourlet transform [18] (CT), Non-subsampled
Contourlet transform [19] (NSCT). The method of extracting eight direction matrices
is explained with an illustration as given below.
Images are considered in YUV color space compared with RGB color space for
coding efficiency. Curvelet forward transform is taken for the Y component of the
image (IY) for obtaining the multi direction and multi scale decomposition. The
coefficient matrix set of an image IY, is given by the Curvelet forward transform
(T + ) of the Y component of the image.

Q = T+ (IY) (1)

where Q = {Qs, d } gives the coefficient matrix set of image IY, and s and d in the
matrix indicates the scale and the direction of the image respectively. If the multi
scale curvelet coefficient is illustrated, it will resemble the structure as shown in the
Fig. 1, where the first scale square matrix is present at the centre of the figure.
The subsequent squares observed from inside to outside of the structure represents
the second to fifth scale coefficient matrices in the increasing order. Varied number of
Curvelet coefficient matrix can be observed for all scales of coefficient matrix except
for first scale. This is because of the reason that first scale coefficient matrix represents
the low frequency characteristics whilst other scale coefficients are high frequency
components. The second to fifth scale coefficient matrices comprises of the direction
matrices of the size sixteen, thirty two, thirty two and sixty four respectively. As a
starting point to figure out the coefficient matrix of every scale layer, 45° is considered
as a reference point and is measured in anti-clockwise direction. Figure 2 shows
the representations of fourth scale layer of the image consisting of 32 coefficients
matrices.

Fig. 1 Distribution and

direction of Curvelet
coefficients
Multidirection Features Based Image Inpainting with Color … 841

Fig. 2 Partition sets of

directional Curvelet
coefficients

{Qs, d} values as {Q4, 1}, {Q4, 2}, {Q4, 3},… denotes the distinct coefficient
matrices for varied directional characteristics of an image. These coefficient matrices
are used to calculate the Multi Direction Features of an image. The second to fifth
scale layers of the Curvelet coefficient matrices are divided into eight sets based on
the direction of angles given by 157.5°, 135°, 112.5°, 90°, 67.5°, 45°, 22.5° and 0°
in anti- clockwise direction that are represented by Z 1 , Z 2 ,..., and Z 8 respectively.
By using the following rule, nth directional multi-direction matrix An is given by

An = T− (Cn(Q)) (2)

where T − represents Curvelet inverse transform, n = 1, 2, … 8 and Hn (Q)

represents the coefficient matrix set.

{Qsd}, Qs, d ∈ Z n
Cn(Q) = (3)
0, otherwise

This equation provides eight directional matrices An for eight directions. Image
size IY.
and nth directional multi-direction matrix An are of same size and every point An
(i, j) represents the direction intensity corresponding to the image pixel IY (i, j) that
is observed in the nth direction. If the value of An (i, j) is, the higher, then the value
of the pixel IY (i, j) will have richer feature details in the nth direction.
According to the direction partition rule based on the Fig. 1, the direction for every
coefficients set is assigned with the higher limit of the angle interval. Example, if
the coefficients set is given by (0°–22.5°), then the direction is assigned to be 22.5°.
Similarly for (22.5°–45°) and (157.5°–180°) coefficient set direction is 45° and 0°
respectively. Eight directional matrices are obtained by applying the above rule as
842 B. Vidhya and M. N. Madhav

Fig. 3 a Original image. b The multi-direction matrices An of eight directions

illustrated in the Fig. 3b. The multi-direction matrices An of the first row of Fig. 2b
illustrates the images obtained through the direction features at 0°, 22.5°, 45° and
67.5° from left to right and second row of illustrates the images obtained through
the direction features at, the image at 90°, 112.5°, 135° and 157.5° from left to right
respectively. For understanding the direction partition rule, consider the first image
A1 , where the angle of the upper white stripe is around 175° but the angle of the
lower white stripe is nearly 160°. This image is assigned with direction 0° according
to the partition rule and only distinguished direction features are obtained.
The direction features that are extracted from the previous step are applied to
the image inpainting technique. Initially the multi directional features are integrated
with the color details to obtain the WCDD for estimating the difference between
neighboring patches.
Then the color direction structure sparsity is obtained through the patch selection
procedure for calculating the filling order. Using WCDD procedure, the best candi-
date patches are selected and multi directional features are combined with color
consistency constraints for the representation of patch sparsity in the optimization
equation.
Multidirection Features Based Image Inpainting with Color … 843

2.2 Distance Measurement

Accurate process of image inpainting can be obtained only if the distance between
the two patches is captured correctly. Distance measurement is a very important step
because it is used to select the best matching patch as well as to find the filling order
through the estimation of CDSS value. These two steps are necessary to perform
image inpainting through patch matching process.
To estimate the difference of two patches, the multi color features are combined
with the color details. Consider a corrupted image I that has a missing region . If
p and q are the two patches that are centred at pixels p and q respectively. Using
super wavelet transform, An can be estimated for sparse representation coefficients
and the preliminary image Ip that is repaired can be obtained by the considering the
size of the neighborhood N(p) as 25 × 25. A local search criterion is used to find the
multiple candidate patches. The size of the patch p that is centred at pixels p and
the size of the multi directional patch Gp that is centred at pixels p in An are same.
Similarly the size of the multi directional patch Gq and q are same.
To estimate the difference between the patches p and q, a weight color direc-
tion distance (WCDD) is performed which is denoted by dw that is given by the
following expression.

dw = η . d (p, q, ) + (1 − η) . d(p, q, ) (4)

where η represents the weight coefficient that balances the details of the color
and direction features. The function d (p, q) represents the mean square distance
between p and q. To estimate WCDD, only the known values of the patches p,
q and Gp, Gq are considered in the Eq. (4) which makes the estimation to be precise
for measuring the difference of color and direction features and WCDD better.
To illustrate the performance of the proposed WCDD algorithm in image
inpainting, the test results were compared with SSD technique as shown in Fig. 4c,
d. In this demonstration, all the same procedures were followed except the applica-
tion of SSD and proposed WCDD algorithm for image inpainting to fill the missing
region by the best matching patches. Figure 4c shows some structures in the inpainted
image and WCDD method give better results of image inpainting. Similarly, these
tests were conducted over wide number of images for analysing the performance of
the proposed system with SSD method. Table 1 gives the PSNR value of the five
different test images for the SSD and WCDD method. From the results, it is proven
that the proposed WCDD algorithm has higher PSNR value and reproduces the image
that is more similar to the original image than SSD method. From the results of Table
1, it is inferred that the proposed WCDD method provides better PSNR value and
used for filling the filling order through CDSS function.
844 B. Vidhya and M. N. Madhav

Fig. 4 Analysis of Image inpainting techniques, a Original image, b corrupted image, c SSD, d
WCDD

Table 1 Analysis of image

Method SSD WCDD
inpainting method using
PSNR Rose image (Fig. 4) 30.67 dB 38.15 dB
Plane image 33.51 dB 38.97 dB
Duck image 32.45 dB 37.98 dB
Baby image 36.02 dB 38.91 dB
Water falls 34.71 dB 39.12 dB

2.3 Estimation of Filling Order

The structure patch must be filled first and texture of the patches must be preserved
whist filling the missing patches from the neighboring pixels. To accomplish this,
a sturdy filling order must be calculated which is usually defined by the product of
color direction structure sparsity S(p) and the confidence term C(p). This product is
termed as the priority function V(p).

V(p) = S(p) · C(p) (5)

This procedure is performed to make sure that priority of filling the patch is given
to the structure patch compared to other patches by S(p) parameter. S(p) parameter
is used to preserve the texture and smooth patches in the patch filling.
A. Criminisi et.al [16, 17] defined the confidence term C(p) as a degree of reli-
able details neighboring the pixel in an image. Z. Xu and J. Sun [20] measured
the confidence of a patch present at the structure by introducing the structure spar-
sity. This produced a better filling order as that of the method explained in [16, 17,
21, 22]. But few constraints were identified like patches in complex structural and
Multidirection Features Based Image Inpainting with Color … 845

textural regions or the missing regions with less or absence of nearby patch details
or consistent confidence value is not provided by the structure sparsity.
To overcome the above limitations, a color direction structure sparsity (CDSS)
based on weighted color-direction distance (WCDD) is used to isolate the structures
and the textures of an image. A weighted color-direction similarity function built on
WCDD is given by the following expression
 
1 dw
X p,k = exp − 2 (6)
Z ( p) σ

where Z(p) is a normalization factor and summation of X p,k will be equal to 1, σ

= 5 and dw value is obtained from Eq. 4. The similarity function gives the similarity
measures between two patches that are present in the known and missing regions.
The color-direction structure sparsity (CDSS) S(p) is expressed by the following
way.

S( p) = (1 − θ ) X 2p,k + θ (7)
k

where θ is a correcting factor that can take up the value between 0 and 1. In the
proposed system, value of θ is assigned as 0.21 after evaluating using different values.
θ is mainly used to rise the minimum value of S(p) to fill the target patches exactly as
texture patch propagates poorly. CDSS uses color and MDS details to estimate the
similarity of the patches like the structure sparsity techniques and also more accurate
results are obtained.

3 Experimental Results

Various experiments were conducted to analyse the performance of the proposed

method. Initially the performance of image inpainting is analysed for different super
wavelet transforms like Curvelet and CT for four and eight direction features as
shown in the Fig. 5. From the results, it has been observed that the eight direction
curvelet transform provides higher PSNR value compared to other methods.
The proposed method is also applied for the application of block removal. Figure 6
is used to illustrate the results of image inpainting technique for the purpose of object
removal. The image is artificially corrupted by introducing objects or blocks in the
original image.
In this study, about 7 methods are used to remove the objects. But in the Fig. 6c,
d the reconstructed image from Exempler based image inpainting and the proposed
method is alone showcased. IN the analysis of output produced, it was found that
the TV-P method, sparsity method, Xu method, Wangs method produced the images
846 B. Vidhya and M. N. Madhav

Fig. 5 Comparative results of image inpainting by wavelet transforms with directions. a Four
direction feature extracted by Curvelet Transform and. b Contourlet transform –CT. c Eight direction
feature extracted by Curvelet Transform and. d CT

which has the traces of objects in the restored images. But the proposed method,
produced an image similar to the original image with higher PSNR values.
Multidirection Features Based Image Inpainting with Color … 847

Fig. 6 Analysis of results of image inpainting techniques for object removal. a Original image. b
corrupted image. c Exempler method [23]. d Proposed method

4 Conclusion

An image inpainting algorithm is proposed in this article to preserve the structural and
textural coherence characteristics along with the neighboring consistency through
process of adopting color and multi directional features in the process of image
inpainting. The proposed method can be used for various applications like object
removal, text and scratch removal. First, the multidirectional features are introduced
in the process of image inpainting and WCDD is estimated to measure the distance
848 B. Vidhya and M. N. Madhav

and similarity of the patches. Then to find the filling order of the patches, CDSS
which is a function of structure sparsity is used. To achieve sharp inpainting results,
limitations of the color and multi-direction features are combined in the optimization
measure. It has been proved that the proposed method provides better structural
coherence over the other existing inpainting techniques.

References

1. Bertalmio M, Sapiro G, Caselles V, Ballester C (2000) Image inpainting. In: Proceeding 27th
Annual Conference ACM SIGGRAPH, pp 417–424
2. Bertalmio M, Bertozzi AL, Sapiro G (Dec 2001) Navier–Stokes, fluid dynamics, and image
and video inpainting. In: Proceedings IEEE Computer Society Conference on Computer Vision
and Pattern Recognition, pp I-355–I-362
3. Chan TF, Shen J (2001) Mathematical models for local nontexture inpaintings. SIAM J Appl
Math 62(3):1019–1043
4. Chan TF, Shen J (2001) Nontexture inpainting by curvature-driven diffusions. J Vis Commun
Image Represent 12(4):436–449
5. Zhang HY, Wu B, Peng QC, Wu YD (2007) Digital image inpainting based on p-harmonic
energy minimization. Chin J Electron 3(3):525–530
6. Telea A (2004) An image inpainting technique based on the fast marching method. J. Graph.
Tools 9(1):23–34
7. Weickert J, Welk M (2006) Tensor field interpolation with PDEs. Visualization and processing
of tensor fields. Springer, Berlin, Germany, pp 315–325
8. Takeda H, Farsiu S, Milanfar P (2007) Kernel regression for image processing and reconstruc-
tion. IEEE Trans Image Process 16(2):349–366
9. Miyoun J, Bresson X, Chan TF, Vese LA (2011) Nonlocal Mumford-Shah regularizers for
color image restoration. IEEE Trans Image Process 20(6):1583–1598
10. Wen Y-W, Chan RH, Yip AM (2012) A primal–dual method for total-variation-based wavelet
domain inpainting. IEEE Trans Image Process 21(1):106–114
11. Mainberger M et al (2012) Optimising spatial and tonal data for homogeneous diffusion
inpainting. Scale space and variational methods in computer vision. Springer, Berlin, Germany,
pp 26–37
12. Yu G, Sapiro G, Mallat S (2012) Solving inverse problems with piecewise linear estimators:
from Gaussian mixture models to structured sparsity. IEEE Trans Image Process 21(5):2481–
2499
13. Elad M, Starck J-L, Querre P, Donoho DL (2005) Simultaneous cartoon and texture image
inpainting using morphological component analysis (MCA). Appl Comput Harmonic Anal
19(3):340–358. 9
14. Guleryuz OG (2006) Nonlinear approximation based image recovery using adaptive sparse
reconstructions and iterated denoising—Part I: theory. IEEE Trans Image Process 15(3):539–
554, 11 Mar 2006
15. Fadili MJ, Starck JL (Sep 2005) EM algorithm for sparse representation- based image
inpainting. In: Proceedings IEEE International Conference Image Processing, pp II-61–II-64.
10
16. Criminisi A, Perez P, Toyama K (Jun 2003) “Object removal by exemplar- based inpainting. In:
Proceedings IEEE Computer Society Conference on Computer Vision and Pattern Recognition,
pp II-721–II-728
17. Criminisi A, Perez P, Toyama K (2004) Region filling and object removal by exemplar-based
image inpainting. IEEE Trans Image Process 13(9):1200–1212
Multidirection Features Based Image Inpainting with Color … 849

18. Wu J-Y, Ruan Q-Q (2009) A novel exemplar-based image completion model. J Inf Sci Eng
25(2):481–497
19. Do MN, Vetterli M (2005) The contourlet transform: An efficient directional multiresolution
image representation. IEEE Trans Image Process 14(12):2091–2106
20. Xu Z, Sun J (May 2010) Image inpainting by patch propagation using patch sparsity. IEEE
Trans Image Process 19(5):1153–1165, CT
21. Guleryuz OG (2006) “Nonlinear approximation based image recovery using adaptive sparse
reconstructions and iterated denoising—Part II: adaptive algorithms. IEEE Trans Image Process
15(3):555–571, 12 Mar 2006
22. Wong A, Orchard J (Oct 2008) A nonlocal-means approach to exemplar-based inpainting. In
Proceeding 15th IEEE International Conference Image Processing, pp 2600–2603
23. Zhang Q, Lin J (2012) Exemplar-based image inpainting using color distribution analysis. J
Inf Sci Eng 28(4):641–654
Classification of Sub-frequency Bands
Based Two-Class Motor Imagery Using
CNN

Muhammad Ahsan Awais , Mohd Zuki Yusoff , and Norashikin Yahya

Abstract EEG has been primarily used in both clinical and research applications.
Brain-computer system (BCI) is one of the leading EEG research applications that
offer special users a new means of communication. Previous studies have reported
the occurrence of MI patterns in mu and beta rhythms, but that does not provide
in-depth knowledge of the frequency range. This paper focuses on the classification
of 2-class Motor Imagery using several frequency sub-bands in the mu and beta
range. “EEG motor imagery dataset from the Physionet database,” has been used for
validation purposes. Although this data includes both imagery and real movements,
we have just used the imagination data. Data is collected from 109 healthy subjects,
but we have only used the first 15 subjects in the study. The study aims to divide
the data into multiple frequency bands to study the motor imagery classification
behaviour over different frequencies. Afterward, a CNN-based deep learning model
with two convolutional layers has been used to classify the left and right classes for
different types of same data. The study seeks to compare the results from various
sub-frequency bands.

Keywords EEG · BCI · CNN · Motor imagery

1 Introduction

The number of people with disabilities who have lost their mobility is substantial.
In addition to their damaged body, they have a working sound brain in their bodies.
The brain should be used to allow them the freedom to move easily by using their
minds to power the wheelchair. The brain-computer interface offers a forum for
people to create a communication relationship between users and computers [1]. It
helps us to interact with the outside world using our brains independently. The BCI
framework does not require any actual muscle movement commonly involved in the

M. A. Awais (B) · M. Z. Yusoff · N. Yahya

Centre for Intelligent Signal and Imaging Research (CISIR), Electrical and Electronic
Engineering Department, Universiti Teknologi PETRONAS, Seri Iskandar, Malaysia
e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 851
R. Ibrahim et al. (eds.), International Conference on Artificial Intelligence
for Smart Community, Lecture Notes in Electrical Engineering 758,
https://doi.org/10.1007/978-981-16-2183-3_80
852 M. A. Awais et al.

communication process. The worldwide scientific community is designing numerous

BCI-based technologies, supporting patients who are unable to speak and provide
mobility. Several tools are available for capturing electrical signals that match every
person’s mental or physical reaction. These techniques include electrocorticography
(ECoG), functional near-infrared spectroscopy (fNIRS), magnetoencephalography
(MEG), functional magnetic resonance imaging (fMRI) and electroencephalography
(EEG) [2, 3].
Electroencephalography is non-invasive, electrophysiological monitoring that
records electrical neural activities of the brain using small metal discs (i.e., elec-
trodes). Most of the brain-computer interface systems incorporate the EEG tech-
nology because of its portable, cost-effective and non-invasive nature. Each action
or thought emits a special brain signal that can vary between individuals.
Throughout the work, the brain signal undergoes signal treatment (pre-processing)
and signal processing (feature extraction and classification) to provide the state
machine’s feature for several applications such as a wheelchair. The integration of
robotic wheelchair and EEG technologies will enhance mobility, boost the quality
of lifestyle and reduce caregiver expenses for differently-abled people. To provide
basic mobility for the disabled people to move around by themselves, brainwave
controlled electrical wheelchair is the best solution. There are many factors like BCI
strategy (Motor Imagery, SSVEPs, etc.), EEG acquisition system, feature extraction
and classification schemes on which the quality of any BCI wheelchair depends [4].
Significant effort has been made by researchers in the field of EEG based BCI
wheelchair on which BCIs are categorized into different types based on the EEG brain
activity patterns. These types are P300 component of event-related potentials (ERPs),
steady-state visual evoke potentials (SSVEP), Motor Imagery (Event-related desyn-
chronization/synchronization) [5]. For each category, a different type of experimental
paradigms is developed and utilized by researchers to collect the EEG data.
The most common and feasible type of EEG datasets for BCI wheelchair lies in the
category of motor imagery. Motor imagery can be interpreted as a dynamic state in
which a person mentally simulates an action without actually performing it physically
[6]. Researchers have collected MI-based EEG data under multiple classes. The
data is divided into different classes based on actions (left-hand movement, feet
movement, etc.) against which imagery data has been collected.
Studies [7, 8] revealed the existence of MI patterns in the mu (8–13 Hz) and
beta band (13–32 Hz). But this does not provide in-depth knowledge about the MI
patterns. For this purpose, we have proposed a system in which the mu and beta
frequency range (i.e., 8–32 Hz) has been divided into several sub-frequency bands
to study the effect of frequency bands on classification accuracy.
Classification of Sub-frequency Bands Based Two-Class Motor … 853

2 Methodology

The research will develop a brain-computer interface system which aims to introduce
the use of several sub-bands of motor imagery-based EEG data. CNN with frequency
sub-bands seeks to achieve better performance, reliability and flexibility.

2.1 Dataset

EEG MI dataset from the Physionet database has been used in the proposed system
[9]. In producing these records, the BCI2000 instrumentation system has been used.
The data collection comprises above 1500 recordings of one and two minutes from
109 subjects. The dataset incorporates both actual and motor imagination tasks
from the subjects. The experimental paradigm comprised 14 runs, where subjects
performed different tasks including rest state, the right fist/left fist movement, the
imagination of right fist/left fist movement, both fists/ both feet movement, and the
imagination of both fists/ both feet movement. The EEG signals were recorded from
64 electrodes as per the international 10–10 system. The data was sampled at 160 Hz
for all subjects. In this study, only the MI trials (imagination of left and right fist
movement) of the first 15 subjects were used. For motor imagery, subjects were
instructed to perform the imagination task as the target stimulus appears on the left
side of the screen or the right side of the screen. The subjects imagine the fist to
be opened and closed before the target stimulus disappears. Then the subject goes
to the rest state. Every subject recorded three sessions for each type of task. The
single-session comprises 7–8 random trials of each class, i.e., left or right movement
imagery. Each trial is carried out for 4 secs, followed by the rest period of 4 s ± 5%.

2.2 Pre-processing

It is essential to remove the noise or unwanted components from the signal. The
undesired signals in our case can be EMG signals (real hand movement) and the
frequency components of the EEG signal other than the motor-imagery task.
The 15 subject’s data has been filtered using BrainStorm in MATLAB. Initially,
the data has been passed through 3 steps including, DC offset correction, notch filter
at 60 Hz in order to eliminate the electrical interference, and the bandpass filter from
8 to 32 Hz, to eliminate other frequencies like EMG (50–250 Hz) and other unwanted
signals.
854 M. A. Awais et al.

2.3 Sub-band Frequency Division

The pre-processed data from the previous stage has been divided into nine sub-
frequency bands using the same platform, i.e., Brainstorm, in order to study the
data in detail. The sub frequency bands include (8–12 Hz), (12–16 Hz), (16–20 Hz),
(20–24 Hz), (24–28 Hz) and (28–32 Hz), (8–16 Hz), (16–32 Hz) and (8–32 Hz).

2.4 Convolutional Neural Networks (CNN)

In this study, CNN based deep learning approach has been adopted to identify the
effect of MI pattern on different frequency ranges.
CNN is a Deep Learning algorithm that can take in any input and process it
efficiently to classify the different classes. CNN uses a system that has been designed
for reduced processing requirements. The layers of a CNN consist of an input layer,
an output layer, and a hidden layer containing multiple convolutional layers, pooling
layers, fully connected layers and normalization layers.
The overall 15 subject’s data is divided into training (10 subjects) and validation
(5 subjects).
The CNN model (Fig. 1) proposed in this research comprises 2 convolutional
layers with 5 kernels per layer. 1st layer performs convolution and the output data is
passed through batch normalization. ReLU activation function is used in the model
and the output is reduced in the max-pooling layer. The same process is applied
to the 2nd layer as well. The output of the 2nd layer is passed through the dropout
layer in which we have used the default value, i.e., 50%. The data is fed to the fully
connected layer after flattening it into 1-D. The output is generated by a softmax
layer with many neurons that fit the number of categories in the data to be classified.
In our case, the output is classified as the left and right class.
The CNN model is composed of 2 convolutional layers and the model is trained
for 10 epochs. The initial learning rate is set to 0.001, while the validation frequency
is 100.
The stated CNN model has been used for two different cases. In case 1, the model
has been implemented on all the 64 channels using 9 sub-frequency bands stated in
Sect. 80.3. Whereas in Case 2, all 9 sub frequency bands have been implemented
using only 3 channels (i.e., C3, Cz and C4) from the central region of the brain.

3 Result and Discussion

The CNN model has been implemented on the basis of two cases. Table 1 illustrates
the results derived from both cases (i.e., use of 64 channels data and the use of only
3 EEG channels). The accuracy of the model is calculated over 10 epochs.
Classification of Sub-frequency Bands Based Two-Class Motor … 855

Fig. 1 Block diagram of the

proposed CNN model Input

2D Convoluon
Batch Normalizaon
Acvaon Funcon: ReLU
2D Max Pooling

2D Convoluon
Batch Normalizaon
Acvaon Funcon: ReLU
2D Max Pooling

Dropout
Fully Connected

Somax

Le Class Right Class

Table 1 Average validation

Frequency Band (Hz) Accuracy (64 Accuracy (3
accuracy against
channels) (%) channels) (%)
sub-frequency bands
08–16 73.29 73.03
08–32 65.97 68.46
16–32 61.11 59.91
08–12 77.93 77.82
12–16 76.88 75.39
16–20 61.92 61.01
20–24 63.21 61.24
24–28 63.33 61.95
28–32 62.08 62.27
856 M. A. Awais et al.

Fig. 2 Classification comparison of different sub-frequency bands for both cases

In case 1, the maximum accuracy, i.e., 77.93%, was reported by the 8–12 Hz
frequency band. In contrast, the frequency band 16–32 Hz resulted in the lowest
validation accuracy, i.e., 61.11%. Whereas in case 2, the maximum accuracy, i.e.,
77.82% has been reported by 8–12 Hz. On the other hand, the minimum validation
accuracy for case 2 has been obtained by the sub-band 16–20 Hz. Figure 2 provides
a graphical representation of the classification comparison of cases 1 and 2.
Among 9 different frequency bands, the 8–16 Hz main band and its two sub-
bands, including 8–12 Hz and 12–16 Hz, has shown better results for 2-class MI
prediction. Figure 2 also demonstrates that frequency bands ranging from 16–32 Hz,
including sub-bands and the main band, has given the lowest classification accuracies.
Thus, it shows that there is no effect of dividing the frequency if the main band is not
giving a better performance. One of the major limitations of getting low classification
accuracies might be the use of a small dataset. Deep learning algorithms are enhancing
efficiency by reducing complexity. But still, it lacks the ability to handle the small
datasets with the same level of efficiency being addressed in the case of large EEG
datasets.

4 Conclusion

The objective of this research is the analysis of sub-frequency bands for the devel-
opment of the BCI system using 2-class motor imagery. CNN has been used for the
classification of two classes using several sub-frequency bands for 15 subjects’ data.
The maximum classification accuracy does not exceed 77.93% as the data is consid-
ered to be a small dataset in order to be processed in a deep learning algorithm. In
Classification of Sub-frequency Bands Based Two-Class Motor … 857

the future, we will be using a sufficient dataset for deep learning algorithms in order
to achieve the maximum classification results effectively.

Acknowledgements This research is supported by the Ministry of Education Malaysia under the
Higher Institutional Centre of Excellence (HICoE) Scheme awarded to the Centre for Intelligent
Signal and Imaging Research (CISIR).

References

1. Abiri R, Borhani S, Sellers EW, Jiang Y, Zhao X (2019) A comprehensive review of EEG-based
brain-computer interface paradigms. J Neural Eng 16(1):011001
2. Biasiucci A, Franceschiello B, Murray MM (2019) Electroencephalography. Curr Biol
29(3):R80–R85
3. Awais MA, Yusoff MZ, Yahya N, Qamar MU(2020) Brain controlled wheelchair: a smart
prototype. In: J Phys Conf Series 1529:042075
4. Andrade MK et al.,(2020) An EEG brain-computer interface to classify motor imagery signals.
In: Biomedical signal processing. Springer, pp 83–98
5. Aoh Y et al (2019) Event-related desynchronization/synchronization in Spinocerebellar Ataxia
Type 3. Front Neurol 10:822
6. Padfield N, Zabalza J, Zhao H, Masero V, Ren J (2019) EEG-based brain-computer interfaces
using motor-imagery: techniques and challenges. Sensors 19(6)
7. Tacchino G, Coelli S, Reali P, Galli M, Bianchi AM (2020) Bicoherence interpretation, in
EEG, requires signal to noise ratio quantification: an application to sensorimotor rhythms. IEEE
Transactions on Biomedical Engineering
8. Tariq M, Trivailo PM, Simic M (2020) Mu-Beta event-related (de) synchronization and EEG
classification of left-right foot dorsiflexion kinaesthetic motor imagery for BCI. Plos one
15(3):e0230184
9. EEG Motor Movement/Imagery Dataset https:/www.physionet.org/physiobank/database/eeg
mmidb/. Last Accessed 25 Oct 2020
Prediction of Methane Hydrate
Formation Rate in Multiphase System
using Artificial Neural Network

Wan Adli Nuaim Bin Wan Ahmad Nadzri, Omar Nashed, Bhajan Lal,
Khor Siak Foo, and Khalik Mohd Sabil

Abstract The research of the hydrate formation has advanced over the last few
decades. Several models have been developed to understand the hydrates formation
kinetics and conditions. In this study, computer based model is used to predict the
gas hydrate formation rate. Since the hydrae formation is stochastic phenomenon
and it is common to get inconsistent data, Artificial Neural Network (ANN) model
has potential to outstand other conventional kinetic models. ANN used to predict the
methane hydrate formation rate in multiphase system. The liquid phase composed
of water + drilling oil + nonionic surfactants was used to form methane hydrates
at pressure 8.80 MPa and temperature of 274.15–277.15 K. This research would
essentially assess the effectiveness of ANN model for the kinetic modeling of the
formation of gas hydrate from the acquired regression analysis. The result of this
research revealed that ANN model with 16 number of hidden neurons had a better
prediction as the highest regression value R was found to be 0.9956.

Keywords Gas hydrate · Artificial neural network · Formation rate

W. A. N. B. W. A. Nadzri · B. Lal (B) · K. S. Foo

Chemical Engineering Department, Universiti Teknologi PETRONAS, 32610 Bandar Seri
Iskandar, Perak, Malaysia
e-mail: [email protected]
O. Nashed
Petroleum Engineering Department, College of Energy Technologies, Jikharrah, Libya
K. S. Foo
PTTEP, Level 26–30, Tower 2, Petronas Twin Towers, Kuala Lumpur City Centre, 50088 Kuala
Lumpur, Malaysia
K. M. Sabil
PETRONAS Research Sdn Bhd, Kawasan Institusi Bangi, Malaysia, Jalan Ayer Itam, 43000
Bandar Baru Bangi, Selangor, Malaysia

© The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 859
R. Ibrahim et al. (eds.), International Conference on Artificial Intelligence
for Smart Community, Lecture Notes in Electrical Engineering 758,
https://doi.org/10.1007/978-981-16-2183-3_81
860 W. A. N. B. W. A. Nadzri et al.

1 Introduction

Gas hydrates are an ice-like inclusion compound, which mainly consist of low
molecular weight gas molecules entrapped into hydrogen bonded water cages [1,
2]. Gas hydrate formation in gas production may hinder the flow and eventually lead
to serious operational problem and huge financial loss. Despite imposing negative
impacts in the oil and gas industries, gas hydrate has some potential application in
some industries [3, 4]. Gas hydrate-based technology could be applied for water
desalination, gas separation, transportation, and storage [5]. Not to forget that, gas
hydrates are available naturally at deep marine and permafrost area [6, 7].
Several modeling studies available are focusing mainly on the thermodynamics
properties of gas hydrate, e.g. phase equilibria and molecular structures [8]. However,
less papers are available in the literature discussing and predicting the time-dependent
behaviour of gas hydrate formation and dissociation [9]. This is due to the naturally
occurring intrinsic stochastic phenomena of gas hydrate which lead to difficulty in
developing a well-established model [2, 10, 11]. Some of the hydrate growth kinetic
models available in the literature include reaction kinetics based model [12], mass
transfer based model [13, 14], combination of reaction kinetics and mass transfer [15,
16], heat transfer based model [17] as well as model based on combination of heat
transfer and reaction kinetic [18]. Gas hydrate formation kinetics are very complex
process, and it is almost difficult to display a definite outcome from the kinetic model
as hydrates form are very uncertain and formation is based on various driving forces
[19]. In addition, the gas hydrate growth rate depends on the phases composition as
well as testing conditions. Recently, artificial neural network (ANN) has attracted a
massive attention due to its ability to deal with scattered and big data to solve complex
problems. ANN is a technique used nowadays for prediction of data set which use
the structure and function of neuron [9, 20–22]. In gas hydrate studies, most of the
papers utilizing ANN focusing on the prediction of the phase equilibrium. However,
less research conducted on the kinetic of the hydrate formation.
Zahedi et al. used two different methods with 203 experimental data points from
literature to predict the hydrate formation temperature. It was found that the best
predictor of hydrate formation temperature was MLP with seven neurons [21].
Moreover, Chapoy et al. estimated hydrate dissociation pressure in the presence of
inhibitor. In their studies, 19 input variables with 35 neurons in one hidden layer were
trained, but various numbers of hidden layers and neurons have not been examined
to find the best structure in the neural network [23].
In this work, ANN was employed to predict methane hydrates formation rates in
multiphase system in the presence of nonionic surfactants. The HLB values were used
to identify the surfactant and used as input along with temperature and surfactants’
concentration.
Prediction of Methane Hydrate Formation Rate in Multiphase … 861

2 Methodology

The data used in this work were collected from experimental study of the gas hydrate
formation kinetics in multiphase system (water + oil + gas) by using 6 different
chemicals: Span 20, Span 40, Span 80, Tween 20, Tween 40, and Tween 80. The
experimental data is presented in Table 1. The network was trained to minimize the
errors between the desired target values and the values computed from the model
based on iterative gradient using MATLAB. The inputs are the types of non-ionic
surfactant represented by Hydrophilic-Lipophilic Balance (HLB), temperature, and
concentration. Meanwhile, the output is the rate of CH4 hydrate formation. The data
set were split into two subset which were for training as well as for validation/testing.
By default, the percentage of each set were 70% training, 15% testing and 15%
validation. For the training algorithm, it was selected according to the quantity of
the data set so that the network could simulate a task with minor error. Usually, the
Levenberg–Marquardt algorithm is the most used algorithm to automatically adjusted
the biases and weights of the neural network [9].
However, for this model, Bayesian Regularization algorithm was selected because
this algorithm could result in good generalization for difficult, small, and noisy

Table 1 Raw experimental data from laboratory

Sample HLB Concentration (wt%) Temperature (°C) Rate of gas hydrate formation
(mol/min)
Sp20 8.6 1 4 0.021247
Sp20 8.6 2 4 0.016504
Sp40 6.7 1 4 0.030413
Sp40 6.7 2 4 0.030926
Sp80 4.3 1 4 0.006789
Sp80 4.3 2 4 0.016452
Tw20 16.7 1 4 0.004801
Tw20 16.7 2 4 0.004642
Tw40 15.6 1 4 0.004651
Tw40 15.6 2 4 0.004477
Tw80 15 1 4 0.005277
Tw80 15 1.5 4 0.004744
Tw80 15 2 4 0.004837
Tw80 15 2.5 4 0.004719
Tw80 15 3 4 0.004433
Tw 80 15 2.5 1 0.001705
Tw80 15 2.5 2 0.002763
Tw80 15 2.5 3 0.003638
Tw80 15 2.5 4 0.004719
862 W. A. N. B. W. A. Nadzri et al.

datasets which suits our input data sets. Next, the number of hidden layers were
manipulated to get the least error. The number of hidden layers were determined
through a process of trial an error [22]. In this paper, the number of hidden layers
were decided through the optimization of regression. A correlation coefficient (R) was
chosen to evaluate the prediction as 1 means a perfect correlation while 0 indicates
for random correlation.

3 Result and Discussion

Dividing the data to training, validation, and testing data set was the first step for
applying ANN prediction. This ANN model was built from 19 number of data of
six chemicals, each taken from different temperature, HLB value and concentration.
In ANN model, the number of hidden neurons were manipulated starting from 1 to
20 hidden neurons. Trial-and-error results showed that optimal hidden neurons 16.
Figure 1 illustrates the ANN architecture with respect to performance function.
Regression analysis was carried out to determine the correlation between the
experimental and predicted data by the ANN models. The highest R is found to be
0.9956 for the ANN model with 16 number of neurons. It is evident there is excel-
lent agreement between the predicted and experimental data for the same training
conditions as shown in Fig. 2. However, ANN model used in this work could have
insufficient number of data with only 3 inputs. Additional data might be required to
get more practical ANN model to predict the gas hydrate formation rate. Addition-
ally, another limitation of ANN model prediction is that the model valid only within
the studied range of HLB, concentration and temperature for the prediction of gas
hydrate formation rate. Consequently, ANN model is not applicable for extrapolation.
Lastly, it is worth mentioning that using too many hidden neurons increases the
complexity of the model which lead to over-fitting. In other hands, different problem
occurs when the network is assumed to has very few hidden neurons causing the
network unable to fit or in other words, under-fitting. The number of hidden neurons
in the hidden layers determines to a network’s capability to compute the output
results.

Fig. 1 Multilayer Perceptron with 16 hidden neurons

Prediction of Methane Hydrate Formation Rate in Multiphase … 863

Fig. 2 Correlation coefficient between experimental and predicted data by ANN with 16 number
of neurons

4 Conclusion

ANN was used to perform the numerical study of kinetic behavior of methane gas
hydrate formation rate in the presence of nonionic surfactants. It was found that
the predicted results by ANN with 16 number hidden neurons is the closest to the
experimental data. The correlation coefficient value for 16 hidden neurons is 0.99556.
Therefore, ANN could be an efficient method for predicting the hydrate formation
rate.

References

1. Bavoh CB, Lal B, Keong LK (2020) Introduction to gas hydrates, in chemical additives for
gas hydrates. In: Lal B, Nashed O (eds). Cham, Springer International Publishing, pp 1–25.
https://doi.org/10.1007/978-3-030-30750-9_1
2. Yin Z, Khurana M, Tan HK, Linga P (2018) A review of gas hydrate growth kinetic models.
Chem Eng J 342:9–29. https://doi.org/10.1016/j.cej.2018.01.120
3. Yin Z, Chong ZR, Tan HK, Linga P (2016) Review of gas hydrate dissociation kinetic models
for energy recovery. J Nat Gas Sci Eng 35:1362–1387. https://doi.org/10.1016/j.jngse.2016.
04.050
864 W. A. N. B. W. A. Nadzri et al.

4. Yaqub S, Lal B, Mellon NB, Sufian BS (2018) Effect of the natural green materials on methane
hydrate formation kinetics. IOP Conf Ser: Mater Sci Eng, 458:012074. https://doi.org/10.1088/
1757-899x/458/1/012074
5. Nashed O, Lal B, Shariff AM, Sabil KM (2020) Gas hydrate promoters. In: Lal B, Nashed O
(eds), Chemical additives for gas hydrates. Cham, Springer International Publishing, pp 47–65.
https://doi.org/10.1007/978-3-030-30750-9_3
6. Khan MS, Lal B, Bustam MA (2020) Gas hydrate inhibitors. In: Lal B, Nashed O (eds),
Chemical additives for gas hydrates. Cham, Springer International Publishing, pp 27–46. https://
doi.org/10.1007/978-3-030-30750-9_2
7. Yaqub S, Lal B, Shariff AM, Mellon NB (2019) Unraveling the effect of sub-cooling temper-
atures on the kinetic performance of biopolymers for methane hydrate. J Nat Gas Sci Eng
65:68–81. https://doi.org/10.1016/j.jngse.2019.03.002
8. Manakov AY, Penkov NV, Rodionova TV, Nesterov AN, Fesenko EE Jr (2017) Kinetics of
formation and dissociation of gas hydrates. Russ Chem Rev 86:845–869. https://doi.org/10.
1070/rcr4720
9. Foroozesh J, Khosravani A, Mohsenzadeh A, Mesbahi AH (2014) Application of artificial
intelligence (AI) in kinetic modeling of methane gas hydrate formation. J Taiwan Inst Chem
Eng 45:2258–2264. https://doi.org/10.1016/j.jtice.2014.08.001
10. Naeiji P, Varaminian F, Rahmati M (2019) The kinetic modeling of methane hydrate growth
by using molecular dynamic simulations. Int J Heat Mass Transf 142:118356. https://doi.org/
10.1016/j.ijheatmasstransfer.2019.07.006
11. Sayani JK, Pedapati SR, Lal B (2020) Phase behavior study on gas hydrates formation in gas
dominant multiphase pipelines with crude oil and high CO2 mixed gas. Sci Rep 10:1–2. https://
doi.org/10.1038/s41598-020-71509-6
12. Vysniauskas A, Bishnoi PR (1985) Kinetics of ethane hydrate formation. Chem Eng Sci
40:299–303. https://doi.org/10.1016/0009-2509(85)80070-1
13. Englezos P, Kalogerakis N, Dholabhai PD, Bishnoi PR (1987) Kinetics of formation of methane
and ethane gas hydrates, vol 42, pp 2647–2658. https://doi.org/10.1016/0009-2509(87)870
15-X
14. Skovborg P, Rasmussen P (1994) A mass transport limited model for the growth of methane and
ethane gas hydrates. Chem Eng Sci 49:1131–1143. https://doi.org/10.1016/0009-2509(94)850
85-2
15. Hashemi S, Macchi A, Servio P (2007) Gas hydrate growth model in a semibatch stirred tank
reactor. Ind Eng Chem Res 46:5907–5912. https://doi.org/10.1021/ie061048+
16. Henning RW, Schultz AJ, Thieu V, Halpern Y (2000) Neutron diffraction studies of CO2
clathrate hydrate: formation from deuterated ice. J Phys Chem A, Article 104:5066–5071.
https://doi.org/10.1021/jp0001642
17. Freer EM, Selim MS, Sloan ED Jr (2001) Methane hydrate film growth kinetics. Fluid Phase
Equilib 185:65–75. https://doi.org/10.1016/S0378-3812(01)00457-5
18. Bollavaram P, Devarakonda S, Selim MS, Sloan ED Jr (2000) Growth kinetics of single crystal
sII hydrates. Elimination of mass and heat transfer effects, Annals of the New York Academy
of Sciences 912:533–543. https://doi.org/10.1111/j.1749-6632.2000.tb06808.x
19. Partoon B, Sahith J, Lal B, and Maulud ASB (2020) Gas hydrate models. In: Lal B, Nashed O
(eds), Chemical additives for gas hydrates. Cham: Springer International Publishing, pp 67–85.
https://doi.org/10.1007/978-3-030-30750-9_4
20. Sahith J, Pedapati SR, Lal B (2019) Application of artificial neural networks on measurement
of gas hydrates in pipelines. Test Eng Manag 81:5769–5774
21. Zahedi G, Karami Z, Yaghoobi H (2009) Prediction of hydrate formation temperature by both
statistical models and artificial neural network approaches. Energy Convers Manage 50:2052–
2059. https://doi.org/10.1016/j.enconman.2009.04.005
Prediction of Methane Hydrate Formation Rate in Multiphase … 865

22. Ghavipour M, Ghavipour M, Chitsazan M, Najibi SH, Ghidary SS (2013) Experimental study of
natural gas hydrates and a novel use of neural network to predict hydrate formation conditions.
Chem Eng Res Des 91:264–273. https://doi.org/10.1016/j.cherd.2012.08.010
23. Chapoy A, Mohammadi AH, Richon D (2007) Predicting the hydrate stability zones of natural
gases using artificial neural networks. Oil Gas Sci Technol 62:701–706. https://doi.org/10.
2516/ogst:2007048
Chemometrics Analysis and Wavelength
Biomarker Identification Using Fourier
Transform Infrared Spectroscopy
for Lard Adulteration

Muhammad Aadil Siddiqui, Mohd Haris Md Khir, Gunawan Witjaksono,

Muhammad Junaid, Saeed Ahmed Magsi, and Abdul Saboor

Abstract Food authenticity is a major concern that is closely linked to public health
and safety in the period in which the long food supply chain is the norm. One of the
key problems facing customers in meat markets is the detection of non-halal meat.
While several studies have been carried out to identify biomarkers for adulterated
meats such as lard, still these studies are at an early stage, and there is no authenticated
biomarker available which can produce specific identification as a result. The purpose
of this research is to utilize the technique of Fourier Transform Infrared Spectroscopy
(FTIR) to analyze pig, cow, lamb, and chicken fat to identify potential biomarkers
for the identification of lard. FTIR studies have shown that chicken and lard fat have
particular peaks relative to lamb and beef fats at wavenumbers 1600, 1750, 2750 and
2920 cm–1 . Furthermore, PCA shows promising clustering and correlation between
the species which can lead to development of a model for successful adulteration
detection.

Keywords Near infrared spectroscopy · FTIR · Lard · Halal · Chemometric

1 Introduction

In the food industry around the world, food authenticity is a significant concern. Food
fraud is still a problem because of the glut of packaged food with a long supply chain
on the market. First coined by Spink and Moyer which stated that food fraud is defined
as alteration of the true food labeling ingredients in which costly, fewer accessible raw
materials are substituted by cheaper alternatives [1]. Some of these adulterations may
impact the public mass financially only, but others may be more seriously affected
by some adulteration, i.e. food allergy [2], food poisoning, religious views [3], etc.
Although food tampering is not a recent issue, some of these tampering was quite
poisonous. For instance, incidences such as addition of sawdust to make white bread

M. A. Siddiqui (B) · M. H. M. Khir · G. Witjaksono · M. Junaid · S. A. Magsi · A. Saboor

Department of Electrical and Electronic Engineering, Universiti Teknologi PETRONAS, Seri
Iskandar, Malaysia
e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 867
R. Ibrahim et al. (eds.), International Conference on Artificial Intelligence
for Smart Community, Lecture Notes in Electrical Engineering 758,
https://doi.org/10.1007/978-981-16-2183-3_82
868 M. A. Siddiqui et al.

[4], formula milk adulteration with melamine [5], adulteration of consumable oil with
non-consumable [6]. Not only dangerous to health, but some incidences also involved
misrepresentation of food ingredients, such as adulteration of meat with horse meat
in 2013 in the UK [7]. Because the price difference between pork and other red
meats, halal and kosher consumers are always concerned with the authentication of
halal and kosher foods [8]. Global halal market is worth billions, and it is projected
to increase to Trillions by 2021 [9]. Therefore, the industry is very interested to
strengthen consumer’s trust in the halal brand. Market confidence is necessary to
protect this [10]. Several authentication techniques can be employed to ensure halal
and kosher brand food products adhere to the standards, such as DNA based methods
[11]. Majority of approaches still require considerable time for preparing sample or
getting vague results due to extreme sensitivity to adulterants which results in overall
poor accuracy. Nowadays, spectroscopy based on vibrations is popularized due to
its effectiveness for adulteration detection in various food [12]. This is because the
simplicity of this nondestructive method reduces time and cost for overall experiment.
FTIR spectroscopy can separate, easily and with high accuracy. For instance, pork
meat in meatball broth [13], chocolate [14], vegetable oils [15], etc. Some research
starts to veer to Near Infrared (NIR) spectroscopy mainly because the feasibility
of this would open possibilities to make food authentication instrumentation setup
to be portable [16]. However, the research done on this reported lower accuracy in
discrimination of adulterants. Although FTIR is quick and relatively inexpensive
with easier sample preparation and non-destructive process [17], it is still limited in
portability due to its bulky equipment. The aim of this research paper is to improve
the accuracy of NIR spectrometer by means of optimizing post-processing analysis,
and to quantify the level of lard adulteration in mixtures.

1.1 Food Authentication Based on Vibrational Spectroscopy

In food safety, the production of food authentication is important since this is the
mechanism that will verify whether a certain food follows its label definition. [18].
This can include the origin of the food (species, gene, geographical), production
method (conventional, organic, free range, etc.), and processing method (frozen,
etc.). Vibrational spectroscopy is based on the idea that inside molecules, atom-to-
atom bonds vibrate at frequencies that can be represented and are therefore subject
to measurement by the laws of physics. In molecules, during the compression step,
the respective electron clouds of the two bound atoms, as well as the charges on the
nuclei, limit the approach of the nuclei, forming an energy barrier. The bond will
inevitably split at the extension of the stretch when the vibrational energy exceeds
the energy of dissociation. Table 1 shows the different stretching frequencies.
Chemometrics Analysis and Wavelength Biomarker Identification … 869

Table 1 Important IR stretching corresponding to their respective frequency and bond [19]
Chemical bond Spectrum location (Cm−1 ) Absorbance strength
C≡N 2260–2220 Medium
C≡C 2260–2100 Medium to weak
C=C 1680–1600 Medium
C=N 1650–1550 Medium
1500–1430 Strong to weak

C=O 1780–1650 Strong

C−O 1250–1050 Strong
C−N 1230–1020 Medium
O−H 3650–3200 Strong
(alcohol)
O−H 3300–2500 Strong
(carboxylic acid)
N−H 3500–3300 Medium, broad
N−H 3300–2700 Medium

1.2 Chemometrics

The use of statistical transformation techniques such as Principal Component Anal-

ysis (PCA), a classical statistical approach for translating a dataset’s attributes into a
new collection of uncorrelated attributes called main components (PCs), is the basis
of chemometrics techniques. The principle is that this allows the data to be reduced to
a smaller number of dimensions, with low loss of information, simply by discarding
some PCs. Each PC is a linear combination of the original inputs and each PC is
orthogonal, reducing the collinearity problem. Chemometrics is based on Beer’s law
that directly relates the concentration of a sample with the intensity in absorbance.
Beer’s law states that the absorbance of a key band of an analyte is proportional to
its concentration. The law can be expressed as.

A = ε cd (1)

where A is the spectra observances, ε is the molar absorption coefficient in lt/mol

cm, c is the concentration in dm2 /mol, and d is the cell thickness in cm.
870 M. A. Siddiqui et al.

2 Materials and Methods

2.1 Meat Sample

Samples were purchased from a local market at Teronoh, Malaysia. After that, the
meat was cut into small element (1 cm × 1 cm) and keep it at –2C°.

2.2 Preparing Mixture Samples

Lard was adulterated with fats of lamb, Beef and Chicken to obtain a series of standard
sets of ten pure and thirty samples containing 10–50% v/v of lard in Lamb, Beef,
Chicken samples. This method follows according to Rohman et al. [18]. The mixture
samples are mixed according to Table 2.

Table 2 Meat Samples with composition percentage for each species

Mixture samples Pork% Lamb% Beef% Chicken% Number of samples
1 10 90 2
2 20 80 2
3 30 70 2
4 40 60 2
5 50 50 2
6 10 90 2
7 20 80 2
8 30 70 2
9 40 60 2
10 50 50 2
11 10 90 2
12 20 80 2
13 30 70 2
14 40 60 2
15 50 50 2
Total mixture samples 30
Chemometrics Analysis and Wavelength Biomarker Identification … 871

2.3 Data Analysis

The data analysis was performed using MATLAB and Spectrograph. Spectrograph
1.1 has been used to extract information from spectrum results, where the data is pre-
processed as needed. To further examine the findings from pre-processing, MATLAB
R2017b was used. To analyze the consistency of lard adulteration, the Principle
Components Analysis (PCA) technique was used.

3 Results and Discussion

3.1 FTIR Spectra of Pure Fats

The FTIR Spectra of pure fats is shown in Fig. 1. This spectrum consists of four
regions: the first region from 4000 to 2500 cm–1 , the second region from 2500 to
2000 cm–1 , the third region from 2000 to 1500 cm–1 , and finally the fingerprint
region from 1500 to 800 cm–1 . For second region, there are no peaks showing no
functional group present in this range. Moreover, 1st and 3rd region have the potential
to differentiate between the sample species.
Figure 2 represents PCA results for all the samples, at 1-D PCA projection, some
of the samples such as Chicken and Pig are already starting to cluster together, even
though there is still overlap between samples of different species. PCA results at 2-D
projection shows that Chicken and Pig samples clustered together without overlap,
Lamb samples have one outlier, but the beef samples proved challenging since results
are still scattered. The 3-D PCA shows better results though with all pig samples
lined around third PCA Component.

Fig. 1 FTIR Spectrum of all the meat species with region identification
872 M. A. Siddiqui et al.

Fig. 2 1D, 2D and three-dimensional representation of chicken, beef, lamb, and lard samples using
principal components

4 Conclusion

Identification of lard discrimination from chicken, lamb, and cow fats was demon-
strated in this paper. FTIR studies have shown that by having a peak at a wavenumber
1600 cm–1 , while other species of fats do not have peak at this stage, lard differenti-
ates itself from other species of fats. In addition, shoulder peak was seen for lard in
the range of wavenumber 2950–2990 cm–1 , while cow and lamb fats showed sharp
peaks, and double peaks showed chicken. In comparison to the usual peaks seen for
the other fats, at wavenumber 2750 cm–1 a reversed peak was detected for lard. For
future work we can further develop this method by analyzing the adulterated samples
as well.

References

1. Spink J, Moyer DC (2011) Defining the public health threat of food fraud. J Food Sci 76:R157–
R163
2. Guardian T (2017) Allergic teenager’s death after eating kebab was accidental, rules coroner.
The Guardian.
3. Li DK (2017) Toddler allergic to dairy dies after pre-school serves him grilled cheese. In: New
York Post, ed
4. Barlass T (2015) Child aged 10 dies after drinking coconut drink as importer admits label
charges. In: The Sydney Morning Herald ed
5. FSA (2013, 7/2/2018) Timeline on horse meat issue. Available: http://webarchive.nation
alarchives.gov.uk/20150403184406/, http://www.food.gov.uk/enforcement/monitoring/horse-
meat/timelinehorsemeat
6. Rohman A, Che Man YB (2012) Analysis of pig derivatives for Halal Authentication Studies.
Food Rev Int 28:97–112
7. Tähkäpää S, Maijala R, Korkeala H, Nevas M (2015) Patterns of food frauds and adulterations
reported in the EU rapid alert system for food and feed and in Finland. Food Control 47:175–184
Chemometrics Analysis and Wavelength Biomarker Identification … 873

8. Abaitua Borda I, Philen RM, Posada de la Paz M, Gomez de la Camara A, Diez Ruiz-Navarro
M, Gimenez Ribota O et al (1998) Toxic oil syndrome mortality: the first 13 years. Int J
Epidemiol 27:1057–1063
9. Regenstein JM, Chaudry MM, Regenstein CE (2003) The kosher and halal food laws. Compr
Rev Food Sci Food Saf 2:111–127
10. Chuah L-O, He XB, Effarizah ME, Syahariza ZA, Shamila-Syuhada AK, Rusul G (2016)
Mislabelling of beef and poultry products sold in Malaysia. Food Control 62:57–164
11. Reuters T (2016) State of The Global Islamic Economy Report 2016/17
12. Barnett J, Begen F, Howes S, Regan A, McConnon A, Marcu A et al (2016) Consumers’
confidence, reflections and response strategies following the horsemeat incident. Food Control
59:721730
13. Schmutzler M, Beganovic A, Böhler G, Huck CW (2015) Methods for detection of pork
adulteration in veal product based on FT-NIR spectroscopy for laboratory, industrial and on-site
analysis. Food Control 57:58–267
14. Vlachos A, Arvanitoyannis IS, Tserkezou P (2016) An updated review of meat authenticity
methods and applications. Crit Rev Food Sci Nutr 56:1061–1096
15. Danezis GP, Tsagkaris AS, Camin F, Brusic V, Georgiou CA (2016) Food authentication:
techniques, trends & emerging approaches. TrAC Trends Analyt Chem 85:123132
16. Kurniawati E, Rohman A, Triyana K (2014) Analysis of lard in meatball broth using Fourier
transform infrared spectroscopy and chemometrics. Meat Sci 96:94–98
17. Meza-Márquez OG, Gallardo-Velázquez T, Osorio-Revilla G (2010) Application of mid-
infrared spectroscopy with multivariate analysis and soft independent modeling of class
analogies (SIMCA) for the detection of adulterants in minced beef. Meat Science 86:511–519
18. Rahmania H, Sudjadi, Rohman A (2015) The employment of FTIR spectroscopy in combination
with chemometrics for analysis of rat meat in meatball formulation. Meat Sci 100:301–305
19. Bruice PY (2016) Organic Chemistry, 8th edn. PEARSON
Deep Learning Approach for Divergence
Behavior Detection at High Density
Crowd

M. U. Farooq, M. N. Mohamad Saad, Y. Saleh, and S. Daud Khan

Abstract At high-density crowd gatherings, people naturally escape from the region
where any unexpected event happens. Escape in high-density crowds appears as a
divergence pattern in the scene and timely detecting divergence patterns can save
many human lives. In this paper, we propose to physically capture crowd normal
and divergence motion patterns (or motion shapes) in form of images and train a
shallow convolution neural network (CNN) on motion shape images for divergence
behavior detection. Crowd motion pattern shape is obtained by extracting ridges
of Lagrangian Coherent Structure (LCS) from the Finite-Time Lyapunov Exponent
(FTLE) field and convert ridges into the grey-scale image. We also propose a diver-
gence localization algorithm to pinpoint anomaly location(s). Experimentation is
carried out on synthetic crowd datasets simulating normal and divergence behav-
iors at the high-density crowd. Comparison with state-of-the-art methods shows
our method can obtain better accuracy for both divergence behavior detection and
localization problems.

Keywords Divergence · FTLE · LCS · Motion estimation · Image shape

1 Introduction

Divergence detection at the high-density crowd is a tough task due to several chal-
lenges involved in high-density crowd videos e.g., few pixels available per head,
extreme occlusion, cluttering and noise, and perspective problems, etc. If crowd

M. U. Farooq (B) · M. N. Mohamad Saad

Center for Intelligent Signal and Imaging Research (CISIR), Universiti Teknologi PETRONAS,
Tronoh, Malaysia
Y. Saleh
Electrical Engineering Department, Faculty of Engineering and Islamic Architecture, Umm
Al-Qura University, Makkah, Saudi Arabia
S. Daud Khan
Department of Computer Science, National University of Technology, Islamabad, Pakistan

© The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 875
R. Ibrahim et al. (eds.), International Conference on Artificial Intelligence
for Smart Community, Lecture Notes in Electrical Engineering 758,
https://doi.org/10.1007/978-981-16-2183-3_83
876 M. U. Farooq et al.

N1
N2

(a) (b)

N1
D1
N2
D2

(c) (d)
Fig. 1 Demonstration of crowd divergence at Love parade 2010: a crowd walking under normal
conditions with low density b crowd walking paths N1, N2 under normal conditions c high-density
crowd within the same region d crowd diverging through paths D1 and D2

divergence is not detected earlier at its development stage, it may lead to larger
disasters like a stampede. Figure 1 shows an example of high-density crowd diver-
gence behavior (Love parade 2010 musical festival [1]) where divergence eventually
leads to disastrous stampede. Figure 1a, b demonstrate high-density crowd normal
behavior following paths N1 and N2, whereas in Fig. 1c, d, a critical situation is
shown where the incoming crowd is blocked by a stationary crowd and diverge
through D1, D2 paths. Such divergence situations are common in mass gatherings
where the whole crowd is marching towards a common destination and with an
increase of density, ends up with a half-stationary half-moving crowd segments that
result in divergence behavior.
Previous divergence detection methods [2, 3] learn manual motion features for
every individual in the crowd from optical flow (OPF) including location, direction,
magnitude, etc. An inherent problem with such methods is with an increase of crowd
density, it is almost impossible to capture individual-level motion information and
one must learn global crowd features. Later several methods have been developed to
capture global crowd motion information e.g., optical flow with pathline trajectories
[4–7], pathlines with Lagrangian particle analysis [8], streakflow [9–13], etc. These
methods performed well in capturing crowd global motion information under normal
Deep Learning Approach for Divergence … 877

behavior scenes only. Unfortunately, no results are reported in the literature for
abnormal behavior detection at very high-density crowd levels.
In this work, we solve divergence detection in the high-density crowd by directly
capturing crowd global motion in form of images and learn crowd normal and diver-
gent motion shapes through a neural network that predicts crowd behavior for the
unknown scene. We also propose a novel divergence localization algorithm to pin-
point divergence location with the help of a bounding box. Finding a source of
divergence can help to efficiently deploy crowd management staff right at the critical
locations.

2 Related Work

Motion is one of the key ingredients in the crowd scene analysis and the success of the
behavior prediction scheme greatly relies on the efficiency of the motion estimation
(ME) method. Therefore, we provide a comprehensive review of ME techniques and
the corresponding abnormal behavior detection methods with emphasis on their capa-
bilities for ‘high’ density crowded scenarios. OPF is considered to be one of the most
fundamental motion flow model [14–17] that has been widely employed for motion
estimation [18, 19], crowd flow segmentation [20], behavior understanding [21–23]
and tracking in the crowd [24]. However, OPF methods suffer from various prob-
lems like motion discontinuities, lack of spatial and temporal motion representation,
variations in illumination conditions, severe clutter and occlusion, etc.
To overcome problems of OPF ME, researchers employ particle advection
concepts from fluid dynamics into the computer vision domain [8] and obtain long-
term “motion trajectories” under the influence of the OPF field. We et al. [7] employ
chaotic invariants on Lagrangian trajectories to determined either the behavior of
the crowd is normal or not. They also perform localization of anomaly by deter-
mining the source and size of an anomaly. Unfortunately, no results were reported
for the high-density crowd. Similarly, Ali et al. [8] obtain Lagrangian Coherent
Structures (LCS) from particle trajectories by integrating trajectories over a finite
interval of time termed as Finite-Time Lyapunov exponent (FTLE). LCS appears as
ridges and valleys in the FTLE field at the locations where different segments of
the crowd behave differently. Authors perform crowd segmentation and instability
detection in the high-density crowd using LCS in FTLE, however actual anomalies
of the high-density crowd like crowd divergence, escape behavior detection, etc.
are not performed. Similarly, authors in [10, 11] obtain particle trajectories using
high accuracy variational model for crowd flow and perform crowd segmentation
tasks only. Mehran et al. [9] obtain streakflow by spatial integration of streaklines
that are extracted from particle trajectories. For anomaly detection, they decompose
streakflow field into curl-free and divergence-free components using the Helmholtz
decomposition theorem and observe variations in potential and streak functions used
with SVM to detect anomalies like crowd divergence/convergence, escape behavior,
878 M. U. Farooq et al.

Fig. 2 High-density crowd motion estimation by state-of-the-art methods

etc. However, results are reported for anomaly detection and segmentation at low-
density crowd and efficacy is still questionable for anomalies at the high-density
crowd. Eduardo et al. [25] obtain long-range motion trajectories by using the farthest
point seeding method called streamline diffusion on streamlines instead of spatial
integration.
Behavior analysis is performed by linking short streamlines using Markov
Random Field (MRF). However, only normal behavior detection and crowd segmen-
tation results are reported. Although particle flow methods discussed above are
better candidates for ME of the high-density crowd, but they are rarely employed
for abnormal behavior detection at high density crowded scenes. Figure 2 provides
a comparison of ME methods for high-density crowd performing Tawaf around
Kabbah. Conventional object tracking based ME methods [26, 27] (Fig. 2b, c) works
best at low crowd density but completely fails at high crowd density. The OPF method
from Brox et al. [15] can estimate motion at high density but motion information is
short-term. SFM [28] method can provide better motion estimation in low-density
crowd areas but at high density, the performance of SFM also degrades. Streakflow [9]
method also performs similarly to the SFM method at a high-density crowd. Unfor-
tunately, all these methods are unable to provide a clean motion-shape for the crowd.
FTLE method [9] (Fig. 2g) produce clear ridges at cowd boundaries and can be best
to describe high-density crowd motion. Therefore, in this work, we utilize the FTLE
method to obtain crowd motion-shape and translate it to a single channel greyscale
image (Fig. 2h) for both normal and abnormal behavior analysis. Our framework for
divergence detection is shown in Fig. 3 (top portion). It consists of two main phases:
Phase 1: low-level FTLE feature extraction and conversion into a grey-scale motion
shape image; Phase 2: behavior classification using a CNN. Motion shape images
are also used for divergence localization process.
Deep Learning Approach for Divergence … 879

Fig. 3 Framework for divergence behavior detection in high-density crowd

3 Divergence Detection with Motion Shape and Deep

Convolution Neural Network

3.1 Data Preparation

Due to the unavailability of a very high-density crowd dataset with divergence

behavior, we generate synthetic data by simulating crowd in Massmotion software
[29]. We model two crowd scenarios: Stampede at Loveparade 2010 and Tawaf
around Kabbah. Example snapshots of normal and divergence crowd behaviors are
shown in Figs. 4 and 5.

3.2 Global Motion Estimation and Shape Extraction

In this work, high-density crowd motion is computed by the Finite-Time Lyapunov

Estimation (FTLE) method [8, 30]. Lagrangian Coherent Structure (LCS) appears as
ridges in the FTLE field where two crowd segments behave differently. We extract
LCS from FTLE field FTLE using the field-strength adaptive thresholding (FFSAT)
scheme and convert it into a grey-scale image. At every integration step in the FTLE
pipeline, maximum Eulerian distance (dmax) is calculated between LCS absolute
peak value and average FTLE field strength, and a threshold (ffsat_thr) is set for dmax
(65% in our work). LCS values crossing ffsat_thr are extracted and converted into a
single-channel grey-scale image. FFSAT algorithm ensures only strong magnitude
LCS values from the FTLE field are extracted and noise is filtered out.

3.3 Deep Network for Divergence Detection

A deep CNN network developed for normal and divergence classes in the high-density
crowd is shown in Fig. 6.
880 M. U. Farooq et al.

a. Normal behavior b. Diverging crowd

c. K13 view d. K5 view e. K12 view

Fig. 4 Synthetic crowd data for Love parade 2010 disaster: a, b are camera-top views; c, d, e- same
crowd with perspective views

a. Normal behavior b. Diverging behavior

Fig. 5 Synthetic crowd data for Kabbah Tawaf—Normal behavior and divergent crowd

The greyscale image is first rescaled to 50 × 50 pixels at the input layer. A

convolution layer is used (24 filters) with ReLU activation. The purpose of using a
large number of convolution filters is to ensure all important receptive fields of CNN
are excited about a given motion-shape. ReLU is adopted as the activation function
because of its good performance for CNNs [31] and Max pooling is used for each
2 × 2 region. Finally, two fully connected layers are used and the softmax layer is
used for the classification of normal or divergent behavior.
Deep Learning Approach for Divergence … 881

Fig. 6 Deep CNN for divergence behavior detection

3.4 Divergence Localization Algorithm

We propose a novel divergence localization algorithm that analyzes changes in

motion shape blob to search for the region of divergence. It was noticed that motion-
shape also exhibit undesired local variations (Fig. 7 top row) that could lead to false
divergence region detection. These changes are occurred due to the to-and-fro motion
experienced by the crowd at high densities crowd [32]. As these oscillatory motions
propagate and reach the crowd boundary, the shape does not remain consistent in
every frame. Whereas the initial occurrence of divergence also appears as a small
shape change and progressively increases in size (as shown in Fig. 7 bottom row). To
cater to undesired local shape changes, a blob processing pipeline is implemented
shown in Fig. 8.

Fig. 7 Top row: Undesired motion-shape variations due to crowd oscillatory motion; Bottom row:
Real shape change due to divergence
882 M. U. Farooq et al.

Fig. 8 Baseline blob extraction pipeline for normal and divergence behaviors

Baseline blob extraction pipeline extracts a baseline blob from the normal and
divergence motion shapes and input to divergence localization algorithm is shown in
Table 1. The divergence localization algorithm indicates divergence location(s) with
the bounding box.

4 Experimentation Results

We evaluate proposed methods of crowd divergence behavior detection and diver-

gence localization using crowd datasets of Love parade and Kabbah (data preparation
details in Sect. 83.1). A detailed qualitative and quantitative analysis is provided for
both methods on two selected scenarios. We also compare our methods with OPF
from Brox et al. [15] by converting the OPF field in binary images.

4.1 Divergence Behavior Detection

For divergence behavior detection at two scenarios, the crowd is simulated to diverge
from 25 different locations in each scenario and 1000 motion-shape images are
captured (total images for 25 divergence locations = 25 × 1000 = 25,000 divergence
images for each scenario). One thousand images for each divergence location are
generated to train CNN with minor local motion changes contributed by crowd
oscillatory motion. Similarly, 2500 images are generated for normal crowd behavior.
The dataset for each scenario is split into two parts: randomly 20 divergence locations
data (20 × 1000 = 20,000 images) are used for training/validation purposes, whereas
the remaining random 5 divergence locations data (completely unseen to CNN) is
used for prediction. Figure 9 provides a confusion matrix of divergence behavior
detection for both scenarios and performance is compared with the OPF method.
For both the Love parade and Kabbah scenario, our method can achieve 100%
accuracy. However, in both scenarios, OPF was able to detect approx. 50% of diver-
gence behaviors only. Motion-shapes obtained through the OPF method are not
as smooth and consistent as produced by our method; hence OPF performance
degradation is evident.
Deep Learning Approach for Divergence … 883

Table 1 Algorithm for divergence localization with bounding box

Algorithm: Divergence localization using the bounding box
Init N: Total video frames, M: Total divergence frames to analyze=50, s_n:No. of frames for spa-
tial averaging =5, t_n: No. of frames for temporal averaging = 5, st_thr: FTLE field spatio-
temporal threshold = 0.5, blb_cnt: number of small blobs to process = 5, i_t: transition frame
index
Step 1 Search normal to divergence-starting Transition frame (global features analysis)
FOR frames i=1:N
-Obtain FTLE motion-shape images for two frames i(previous) and i+1(current),
-Run blob processing pipeline (fig. 8) on previous and current frames FTLE im-
ages and obtain large single blobs for both frames
-Subtract the current large blob from the previous large blob to obtain several
smaller blobs
-Sort blb_cnt small blobs according to the area in descending order
-Blob shape analysis for ROI
FOR blob_num=1: blb_cnt
Compare blob shape with shape-list [square, rectangle, triangle, pentagon,
hexagon, semi-circle or nearly circular]
IF blob shape matches with any of shape-list
Transition frame ‘i_t’ with the first ROI, Exit this loop and go to
step 2
ELSE
Select next blob in area sorted list
END
-Increment frame number i
END
Step 2 Bounding box search for divergence location (global features analysis)
FOR frames j=i_t:i_t+M
- Obtain FTLE motion-shape images for two frames j(previous) and j+1(current)
- Run blob processing pipeline (fig. 8) on previous and current frames FTLE im-
ages and obtain large single blobs for both frames
- Subtract current large blob from previous large blob to obtain several smaller
blobs
- Sort blb_cnt small blobs according to the area in descending order
- Blob shape analysis for ROI
FOR blob_num=1: blb_cnt
IF blob shape matches with any of shape-list
Mark this blob as possible ROI and update ROI_list
ELSE
Select next blob in area sorted list
END
- Jump to step 3 to obtain divergence location bounding box coordinates for all
ROI of the current frame
- Increment j
END
(continued)
884 M. U. Farooq et al.

Table 1 (continued)
Step 3 FTLE field search for refined divergence location bounding box (local features anal-
ysis)
FOR ROI=1:length(ROI_list)
- Spatial Averaging: For each ROI boundary pixel, perform 8-connected spatial
averaging of FTLE field (for all frames from j-s_n to j+s_n frames)
- Temporal averaging: For each ROI boundary pixel in j-t_n to j+t_n frames,
perform temporal averaging.
- FOR all ROI boundary spatio-temporal averaged values
IF spatio-temporal averaging value >= st_thr,
keep pixel as valid
ELSE
discard pixel
END
- Obtain bounding box coordinates of current ROI by calculating maximum and
minimum (x,y) coordinates from selected valid pixels list in above step
END
Return bounding box coordinates for all ROIs in ROI_list

Fig. 9 Confusion matrices for divergence behavior detection: Love parade scenario a Proposed
method b OPF method; Kabbah scenario c Proposed method d OPF method

4.2 Divergence Localization

The performance of the divergence localization algorithm is evaluated by calculating

the Intersection over Union (IoU) area of the predicted bounding box and ground truth
bounding box for each divergence region. Ground truth bounding boxes are obtained
by hand labeling divergence regions of each abnormal frame. IoU is calculated using
Eq. (1).
Deep Learning Approach for Divergence … 885

Area of overlap
IoU = (1)
Area of union
Generally, an IoU score greater than 0.5 (50% overlap) is considered a good
prediction by any bounding box (b. box) detection algorithm [33]. In this work,
the IoU score is calculated for N post i_t frames. IoU score of six selective frames
(out of N = 50 post i_t frames) for the Love parade scenario is shown in Fig. 10.
The green color b.box represents ground truth and the red color b.box represents
prediction by our algorithm. The Final IoU score is obtained by averaging N frames
IoU scores. The average IoU score of our algorithm for the Love parade scenario
is 0.501 (50% overlap). We also perform divergence region detection using OPF
motion images. The average IoU score with the OPF method is found to be 0.15
(15% overlap) which proves our method performs well than OPF for divergence

i_t+1 i_t+5

i_t+10 i_t+15

i_t+20 i_t+25

Fig. 10 Divergence region localization with our proposed method. IoU scores shown for six post
i_t frames
886 M. U. Farooq et al.

region localization. Similarly, the average IoU score for the Kabbah scene with our
algorithm is 0.63 (63% overlap) and 0.18 (18% overlap) for the OPF method.

5 Conclusion

In this work, we propose a deep CNN-based divergence behavior detection frame-

work that extracts high-density crowd motion shapes in form of images to train deep
CNN. Experimentation results show that the proposed method can achieve close
to 100% accuracy for divergence detection in challenging Loveparade and Kabbah
crowding scenarios. Similarly, a novel divergence region detection algorithm effi-
ciently detects divergence regions with IoU of more than 50%. However, we notice
there are few limitations of our proposed methodology of converting crowd motion
into images using the FTLE method. Motion shape analysis is inefficient in the situ-
ations when a crowded segment in high density gets stationary due to any reason.
Since there is no more movement at the stationary crowd segment, FTLE is unable
to predict crowd motion shape at static crowd portions and results in incomplete or
broken motion-shapes. Therefore, for our framework to work efficiently, the crowd
needs to keep moving (for consistent motion-shape) that is always not true. Secondly,
in the FTLE method, LCS ridges appear only at crowd boundaries, if any anomaly
takes place at interior portions of the crowd (far from crowd boundaries towards
the center), FTLE is unable to provide any information there. Therefore, in future
work, we shall improve our method by incorporating spatial and temporal crowd
density variations to capture static crowd behavior. And predict crowd behavior in
all segments of the crowd, either crowd is stationary or in motion.

References

1. Helbing D, Mukerji P (2012) Crowd disasters as systemic failures: analysis of the Love Parade
disaster. EPJ Data Sci 1(1):1–40. https://doi.org/10.1140/epjds7
2. Chen CY, Shao Y (2015) Crowd escape behavior detection and localization based on divergent
centers. IEEE Sens J 15(4):2431–2439. https://doi.org/10.1109/JSEN.2014.2381260
3. Wu S, Wong HS, Yu Z (2014) A bayesian model for crowd escape behavior detection. IEEE
Trans Circuits Syst Video Technol 24(1):85–98. https://doi.org/10.1109/TCSVT.2013.2276151
4. Andrade EL, Blunsden S, Fisher RB (2006) Modelling crowd scenes for event detection. Proc
Int Conf Pattern Recognit 1:175–178. https://doi.org/10.1109/ICPR.2006.806
5. Courty T, Corpetti N (2007) Crowd Motion Capture. Comput Animat Virtual Worlds
18(September 2007):361–370
6. Nam Y, Hong S (2014) Real-time abnormal situation detection based on particle advection in
crowded scenes. J Real-Time Image Process 10(4):771–784. https://doi.org/10.1007/s11554-
014-0424-z
7. Wu S, Moore BE, Shah M (2010) Chaotic invariants of lagrangian particle trajectories for
anomaly detection in crowded scenes. In: Proceedings of the IEEE computer society conference
on computer vision and pattern recognition, pp 2054–2060
Deep Learning Approach for Divergence … 887

8. Ali S, Shah M (2007) A Lagrangian particle dynamics approach for crowd flow segmentation
and stability analysis. In: IEEE conference on computer vision and pattern recognition, pp 1–6
9. Mehran R, Moore BE, Shah M (2010) A streakline representation of flow in crowded scenes.
Lect Notes Comput Sci (including Subser Lect Notes Artif Intell Lect Notes Bioinformatics)
LNCS 6313(PART 3):439–452. https://doi.org/10.1007/978-3-642-15558-1_32
10. Wang X, Gao M, He X, Wu X, Li Y (2014) An abnormal crowd behavior detection algorithm
based on fluid mechanics. J Comput 9(5):1144–1149. https://doi.org/10.4304/jcp.9.5.1144-
1149
11. Wang X, Yang X, He X, Teng Q, Gao M (2014) A high accuracy flow segmentation method in
crowded scenes based on streakline. Opt Int J Light Electron Opt 125(3):924–929. https://doi.
org/10.1016/j.ijleo.2013.07.166
12. Wang X, He X, Wu X, Xie C, Li Y (2016) A classification method based on streak flow for
abnormal crowd behaviors. Opt Int J Light Electron Opt 127(4):2386–2392. https://doi.org/10.
1016/j.ijleo.2015.08.081
13. Huang S, Huang D, Khuhro MA (2015) Crowd motion analysis based on social force graph
with streak flow attribute. J Electr Comput Eng 2015. https://doi.org/10.1155/2015/492051
14. Horn BK, Schunck BG (1981) Determining optical flow. Artif Intell 17(1981):185–203
15. Brox T, Papenberg N, Weickert J (2004) High accuracy optical flow estimation based on a
theory for warping. In: Computer Vision - ECCV 2004, vol 4, no May, pp 25–36. https://doi.
org/10.1007/978-3-540-24673-2_3
16. Lucas BD, Kanade T (1981) An iterative image registration technique with an application to
stereo vision. Proc Imaging Underst Work 130:121–130
17. Fortun D, Bouthemy P, Kervrann C, Fortun D, Bouthemy P, Kervrann C (2015) Optical flow
modeling and computation: a survey. Comput Vis Image Underst 134:1–21
18. Lawal IA, Poiesi F, Anguita D, Cavallaro A (2016) Support vector motion clustering. IEEE
Trans Circuits Syst Video Technol X(X):1–1. https://doi.org/10.1109/TCSVT.2016.2580401
19. Cheriyadat AM, Radke RJ (2008) Detecting dominant motions in dense crowds. IEEE J Sel
Top Signal Process 2(4):568–581. https://doi.org/10.1109/JSTSP.2008.2001306
20. Ali S, Shah M (2007) A lagrangian particle dynamics approach for crowd flow simulation and
stability analysis
21. Hu MHM, Ali S, Shah M (2008) Learning motion patterns in crowded scenes using motion
flow field. In: 2008 19th international conference pattern recognit, pp 2–6. https://doi.org/10.
1109/ICPR.2008.4761183
22. Solmaz B, Moore BE, Shah M (2012) Identifying behaviors in crowd scenes using stability
analysis for dynamical systems. IEEE Trans Pattern Anal Mach Intell 34:2064–2070. https://
doi.org/10.1109/TPAMI.2012.123
23. Chen DY, Huang PC (2011) Motion-based unusual event detection in human crowds. J Vis
Commun Image Represent 22(2):178–186. https://doi.org/10.1016/j.jvcir.2010.12.004
24. Hu W, Xiao X, Fu Z, Xie D, Tan T, Maybank S (2006) A system for learning statistical
motion patterns. IEEE Trans Pattern Anal Mach Intell 28(9):1450–1464. https://doi.org/10.
1109/TPAMI.2006.176
25. Pereira EM, Cardoso JS, Morla R (2016) Long-range trajectories from global and local motion
representations. J Vis Commun Image Represent 40:265–287. https://doi.org/10.1016/j.jvcir.
2016.06.020
26. Dalal N, Triggs B (2005) Histograms of oriented gradients for human detection. In: EEE
computer society conference on computer vision and pattern recognition (CVPR’05), pp 886–
893
27. Barnich O, Van Droogenbroeck M (2011) ViBe: a universal background subtraction algorithm
for video sequences. IEEE Trans Image Process 20(6):1709–1724. https://doi.org/10.1109/TIP.
2010.2101613
28. Mehran R, Oyama A, Shah M (2009) Abnormal crowd behavior detection using social force
model. In: IEEE computer society conference on computer vision and pattern recognition work.
CVPR Work 2:935–942. https://doi.org/10.1109/CVPRW.2009.5206641.
29. https://www.oasys-software.com
888 M. U. Farooq et al.

30. Shadden SC, Lekien F, Marsden JE (2005) Definition and properties of Lagrangian coherent
structures from finite-time Lyapunov exponents in two-dimensional aperiodic flows. Phys D
Nonlinear Phenom 212(3–4):271–304. https://doi.org/10.1016/j.physd.2005.10.007
31. Zeiler MD et al (2013) On rectified linear units for speech processing New York University,
USA Google Inc ., USA University of Toronto , Canada. In: IEEE international conference on
acoustic speech and signal processing (ICASSP 2013), pp 3–7
32. Krausz B, Bauckhage C (2012) Loveparade 2010: automatic video analysis of a crowd disaster.
Comput Vis Image Underst 116(3):307–319. https://doi.org/10.1016/j.cviu.2011.08.006
33. Ahmed F, Tarlow D, Batra D (2015) Optimizing expected intersection-over-union with
candidate-constrained CRFs. In: Proceedings of the IEEE international conference on computer
vision, vol 2015, pp 1850–1858. https://doi.org/10.1109/ICCV.2015.215
Anomaly Localization at High-Density
Crowd Using Motion Shape Image (MSI)

M. U. Farooq, M. N. Mohamad Saad, S. Daud Khan, and Y. Saleh

Abstract Anomaly localization plays a critical role if a disaster occurs in a high-

density crowd to efficiently rescue the crowd from the right location. This paper
enriches anomaly localization by introducing new localization features, in addition to
the ‘only’ localization feature of source/start point detection existing literature offers.
New features introduced include crowd density estimation in localized regions and
direction/angle of a localized region. A motion shape image (MSI) based approach is
introduced for localization and features detection and experimentation is performed
on benchmark and our proposed high-density crowd datasets.

Keywords Anomaly · Localization · Divergence · Motion shape image · FTLE ·

LCS

1 Introduction

It is well known that automated anomaly detection in surveillance videos has attracted
a lot of attention due to efficient methods developed in the past. However, anomaly
detection alone is less effective without localizing it. Anomaly localization is critical
information that can help to efficiently deploy rescue-resources at the right place to
save many precious lives.
There have been several methods developed in the past focusing anomaly local-
ization at low to medium density crowd [1–3], however, localizing anomaly at high-
density is challenging and there are not many methods available in the literature [4–7].

M. U. Farooq (B) · M. N. Mohamad Saad

Center for Intelligent Signal and Imaging Research (CISIR), Universiti Teknologi PETRONAS,
Tronoh, Malaysia
S. Daud Khan
Department of Computer Science, National University of Technology, Islamabad, Pakistan
Y. Saleh
Electrical Engineering Department, Faculty of Engineering and Islamic Architecture, Umm
Al-Qura University, Makkah, Saudi Arabia

© The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 889
R. Ibrahim et al. (eds.), International Conference on Artificial Intelligence
for Smart Community, Lecture Notes in Electrical Engineering 758,
https://doi.org/10.1007/978-981-16-2183-3_84
890 M. U. Farooq et al.

Therefore, this work aims to propose an anomaly localization method at high-density

crowd gatherings. Representative anomaly localization methods are discussed next.
Wu et al. [4] localize anomaly in the crowd by finding the position of abnormal
representative trajectories. Wu et al. [5] localize crowd escape anomaly by deter-
mining the divergence center locations using a reversible jump Mteropolis-Hasting
algorithm. Similarly, Chen et al. [1] localize escape behavior anomaly by applying K-
nearest search and distance segmentation methods on optical flow features of magni-
tude, speed, and direction. Chaker el at. [6] localize anomaly using a social network
model that separates rare actions from dominant actions, however, their moving
object in a crowd, not localizing crowd anomaly. Mehran et al. [7] localize crowd
divergent or convergent behavior by computing potential functions from streakflow.
Ongun et al. [2] localize crowd anomaly by clustering finite-time Lyapunov expo-
nent heat maps and applying thresholding on clusters to detect an anomaly, however,
anomaly localization details are not clearly stated in this work. Ren et al. [3] perform
anomaly detection using an entropy model and anomaly is localized by detecting
high behavior entropy (BE) value pixels in an image.
The methods discussed above suffer from two types of important limitations: (1)
Majority of existing art localize anomalies at low or medium-density crowds, to the
best of our knowledge, no method is developed to localize anomalies at high/very
high-density crowds. (2) the localization information is limited i.e., all methods only
show the source or start of the anomaly. There is no updated information on an
anomaly at the localized region later in the video. Hence, limits the application of
existing methods in real-life anomaly localization tracking/updates.
In this paper, we propose a motion shape image (MSI) based approach for the
anomaly localization method. The proposed method not only locates start/source
points of an anomaly but also provides more useful pieces of information about
anomaly localized region throughout the video and hence is useful for practical
anomaly localization at high-density crowds. Specifically, the main contributions of
this paper are as follows:
1. An anomaly localization method, that generates localization masks for normal
as well as abnormal crowd segments,
2. Provides anomaly localization updates each frame i.e., if anomaly location is
moved in the video from initial detection, the method will provide an updated
location of the anomaly at later times,
3. Crowd density estimate is provided inside anomaly localized regions,
4. The angle and direction of the localized region are estimated with the help of
the angle of the localized anomaly region measured about the normal crowd
segments.
In this work, we focus on escape or divergence anomaly at a high-density crowd.
However, the definition of divergence in the existing literature is not clear and we
define the divergence of the crowd as the ‘significant change in crowd motion pattern
from its normal walking motion pattern’. Usually a high-density crowd exhibit
two types of normal motion patterns i.e., straight and circular. And the significant
Anomaly Localization at High-Density Crowd … 891

change in motion pattern from the normal straight or circular patterns in any direc-
tion is considered as divergence. In the next subsection, the localization method for
divergence anomaly at the high-density crowd is discussed in detail.

2 Proposed Method

A detailed block diagram of our anomaly detection and localization framework for
divergence behavior at a high-density crowd is shown in Fig. 1. The top portion of the
block diagram shows an anomaly detection network using an MSI. The performance
evaluation of the proposed anomaly detection network is discussed in detail in [8]. The
proposed anomaly detection network outputs scores for two classes i.e., normal and
anomalous (divergent). A divergent class score greater than a pre-defined threshold
triggers the divergence localization process, shown at the bottom of Fig. 1.

Fig. 1 Divergence detection and localization framework

892 M. U. Farooq et al.

2.1 Divergence Localization and Source/start Point Detection

Our divergence localization scheme is simple, it takes the difference of incoming

MSI from a reference motion shape as shown in Fig. 2. The differencing process
is started upon the indication of divergence detected by CNN and the motion shape
at that time instance is saved as a reference shape for the subsequent subtraction
process. However, motion shapes obtained at every integration step are not similar
and exhibit both local and global shape variations.
Figure 3 shows samples of shape variations for normal (top row) and divergence
(bottom row) behaviors for the crowd at the Loveparade scenario (crowd on the
ramp entering from left and right tunnels and heading towards ramp top for exiting).
Regions marked with squares in Fig. 3 (upper row) are motion shape variations
experienced during normal behavior frames. Simply differencing normal and diver-
gence motion shapes (in the presence of shape variations) produce many undesired
blobs that are not actual divergence sources and can lead to false divergence-source
detections.
As mentioned above, shape variations occur locally and globally at the high-
density crowd. Local motion shape variations are due to to-and-fro motion [9] and
global crowd motion shape changes are due to the segment(s) of the crowd become
‘stationary’ and results in no FTLE field. For the former problem, the crowd naturally
starts oscillating left and right at their central axis, generating crowd waves. As the
waves reach boundaries, the crowd naturally expands and shrink. As a result, the
FTLE field at the crowd boundary also expands/shrinks causing local (minor) motion
shape variations.

Fig. 2 Divergence localization method for source/starting point detection and divergence mask
Anomaly Localization at High-Density Crowd … 893

Fig. 3 Top row: undesired motion-shape variations due to crowd oscillatory motion; Bottom row:
real shape change due to divergence

Global motion shape variation occurs if any crowd segment stops that reduce
crowd velocities to zero, results in no OPF/ FTLE field in that crowd region. Thus,
LCS ridges in the FTLE field vanishes from stationary crowd segments and only
appear at the moving crowd segments, causing an undesired global motion shape
variation. Due to the undesired changes in motion shape, raw MSI’s cannot be utilized
for further analysis and we need to extract baseline motion shape for both normal
and divergence behaviors. A baseline motion shape represents a shape that is most
common throughout all the specific behavior frames. Figure 4 shows the pipeline for
baseline motion shape extraction that extracts baseline motion shape for both normal
and abnormal video frames.
Baseline motion shapes obtained for normal and divergent motion patterns differ in
the shape in the way that divergent motion shape will contain extra blobs representing
diverging crowd. The reference normal baseline motion shape and baseline divergent
motion shape are then subtracted that generates smaller blobs. The resulting small

Fig. 4 Pipeline for baseline motion shape extraction for normal and divergence behaviors
894 M. U. Farooq et al.

Fig. 5 Left side: Crowd circulatory normal motion with crowd diverging in the top second image
(lower row shows corresponding MSI’s); right side images for synthetic love parade dataset crowd
diverging from different locations (lower row are corresponding MSI’s)

Fig. 6 Top row: MELAN dataset with crowd diverging from left and right gates; bottom row:
PILGRIM dataset with two normal scenarios and last image showing divergent crowds for two
normal crowd scenarios

blobs contain the actual divergent region(s) and many noisy blobs. The noisy blobs
are filtered out by applying temporal and spatial filtering on OPF and FTLE fields
on current and past N filter frames.
The blobs obtained from the above process represent divergent regions or diver-
gence masks around the diverging crowd. However, we need to find an exact starting
point or source of divergence from the divergent mask detected above. Initially, the
Anomaly Localization at High-Density Crowd … 895

contour of the divergence mask is obtained and the Euclidean distance of contour
pixels from normal baseline motion shape pixels is computed. The contour pixels
having a distance less than a threshold is/are marked as divergence source or starting
point(s). The experimentations have shown that a threshold of 10 pixels is most
optimal for the majority of scenes and can represent correct source points. Euclidean
distances more than 10 pixels are generally far from the actual source point and can
represent false source point detections.

2.2 Crowd Density Estimation (CDE) in the Localized Mask

Our anomaly localization process discussed above primarily generate a localization

mask around the diverging crowd. Crowd density inside the divergence mask is
an important indicator of the severity of divergence i.e., a high crowd density crowd
inside the divergence mask indicates a life-threatening situation that forces the crowd
to take a divergence path, etc. In this work, crowd density is estimated using the
CSRNet method from [10] (Keras implementation is available at [11]). The model
is pre-trained on the ShanghaiTech dataset and further fine-tuned on our PILGRIM
dataset and on high-density crowd images from Hajj (a detailed discussion on crowd
datasets is provided in section III). To estimate the crowd density inside the divergence
mask, initially, a density map and crowd density count is generated by CSRNet for
the whole image. The density map and crowd count are then cropped according to
the divergence mask produced by the divergence localization algorithm.

2.3 Direction and Angle of the Localized Mask

The diverging crowd direction and the angle of divergence are critical information for
first responders to effectively manage the crowd. To quantify the crowd divergence
information, crowd “dominant motion” needs to be computed. Dominant motion
represents the motion of the major segments of the crowd in the video. In the case of
divergence, we have to compute dominant flows for both normal and divergent crowd
segments to show how far the divergent crowd moves away from the normal walking
crowd. In this work, dominant motion flow is computed using the method from [12].
Summarizing the method of [12] for dominant motion flow, initially, video is divided
into several segments and particle advection is performed to obtain small size raw
tracklets in each segment. Raw tracklets provide short duration motion information
and crowd starting (source) and ending (sink) points cannot be determined. Local
long tracks are identified in each video segment by using an unsupervised hierar-
chical clustering algorithm that utilizes the Longest Common Sub-sequence (LCS)
similarity measure to combine smaller tracklets into local large trajectories. To find
dominant flow throughout the video, local trajectories from different video segments
are also combined using the same clustering and similarity method to produce global
896 M. U. Farooq et al.

dominant flows. This method is also able to identify the source and sink point of a
dominant motion.
The above algorithm is run to obtain dominant motion flows in both normal and
divergence masks. To quantify the divergent dominant motion from normal dominant
motion, we need to compute the mean angle for both dominant flows and the angle
difference between the two flows would show how far the divergent crowd moves
away from the normal crowd. A dominant motion trajectory represents the positions
of a particle Pi it takes from source to sink in video and there is associated angle
information (∅i ) at every position of a particle. The particle angle (∅i ) of each particle
Pi in the dominant flow trajectory is computed using OPF velocities given in Eq. 1.
v
∅i = tan−1 (1)
u
where u and v are optical flow horizontal and vertical velocities respectively.
It is important to notice here that the OPF used in the particle advection process
is computed at an evenly spaced grid whereas particle positions of dominant flow
trajectory are at off-the-grid indexes. And we do not have OPF at off-grid particle
positions. Therefore, linear interpolation is implemented to obtain OPF at off-grid
particle positions and ∅i is computed from interpolated OPF u and v velocities. The
angle ∅i computed above lies in the range of [0, π]. To know the exact quadrant in
which a particle lies, ∅i is mapped to a unit circle with the angle from [0, 2π]. We
use the method in [13] to map the linear angle into a circular angle on the unit circle.
The circular angle shows the actual quadrant of a particle it lies in. A similar process
is repeated for all particles by mapping all on the unit circle. Then a mean circular
angle is computed that represents the mean direction of a dominant flow. The mean
angle for both normal and divergent dominant flows is computed and the divergence
angle is calculated by taking the difference of two mean angles.

3 Proposed Crowd Dataset

The majority of existing datasets containing divergent scenes consist of either

low-density crowd or contain normal behavior only A new large-scale dataset is
developed to evaluate our method under realistic high-density crowd divergence
scenarios. Mainly we construct three datasets: (1) Synthetic (SYN) dataset (2)
MELAN CONCERT dataset and (3) PILGRIMS dataset. A summary of the datasets
is provided in Table 1 (more details for the proposed dataset can be seen at [8]).
Table 1 Video datasets comparison (existing benchmark and ours)
Dataset No, of Frames No.of scenes Resolution Frame rate Density level Type (Real/synthetic) Normal motion type
Normal Abnormal
UCF 110 140 1 480×360 25 Low Real-Ped Straight
UMN 500* 100* 11 320×240 30 Low Real-Ped Circular
PETS2009 52* 52* 8 768×576 7 Low-medium Real-Ped Straight
Anomaly Localization at High-Density Crowd …

NGSIM 250* 241* 8 640×480 10 Low Real-Traffic Straight

OUR-SYN 10000* 10000* 100 1280×720 24 High Syn-Ped Straight/circular
OUR-MELAN 553* 553* 4 1920×1080 50 High Real-Ped Straight
CONCERT
OUR-PELGRM 1500* 1500* 3 1280×1080 50 Hight Real-Ped Straight
*Number of frames averaged for the redpective number of scenes (normal abnormal)
897
898 M. U. Farooq et al.

4 Experimentation and Discussion

4.1 Parameters Settings

Divergence class score value from CNN lies between 0 and 1 and the threshold
at the class score is set to 0.65 (65%) to trigger the localization detection process.
Residual blobs are filtered out through temporal averaging filtering and spatially
through connected neighbors filtering. The number of images for temporal averaging
N filter is set to 10 i.e., averaging previous 10 frames OPF and FTLE fields. Spatial
filtering is applied by performing an 8-connected neighbor on residual blob(s).

4.2 Localization–Qualitative Evaluation

In qualitative evaluation, we compare our proposed divergence localization scheme

with two state-of-the-art divergence localization methods [5] and [1]. The method in
[5] estimates divergent centers by placing potential destination points over normal
image sequences. For the escape scene, a set of divergent points is initialized and
the foreground velocity patches start from the divergent points reach potential desti-
nations are marked as divergent centers. The Method in [1] analyzes intersections
of foreground velocity vectors and uses the distance segmentation method and knn
search to locate the divergent center. Intersections are obtained with the assumption
that there are three moving objects in the neighbor of the desired divergent center and
straight lines of moving objects should intersect to declare the point as the divergent
center.
A qualitative comparison of three methods is shown in Fig. 7. Divergent centers are
marked with triangles for three methods and ground truth. Scenarios in the first two
rows of Fig. 7 depict divergent center detections at the low-density crowd. Existing
state-of-the-art methods accurately detect divergent center’s at low crowd density
whereas our method shows few false detections at the low crowd density divergence.
At low crowd density, the crowd motion shape is broken and produces many resid-
uals after the difference from the reference motion shape. The residual blobs cause
marking of false divergent centers by our method. Also, since our method analyzes
variations at crowd boundaries, the divergence points detected by our method at low
crowd density are slightly offset from the ground truth. The last two images in the
third row of Fig. 7 show divergence center detections at the high-density crowd. It
can be seen that existing art failed to detect divergence sources at the high-density
crowd.
The method at [5] analyzes variations in foreground velocity patches from high (in
non-escape) to low (in escape) case. However, at high density, foreground velocity
concentration is the same in both behaviors, no divergence center points are reaching
potential destinations. However, the method is still able to detect divergence center
points at the crowd boundary as patches at boundaries experience variations in
Anomaly Localization at High-Density Crowd … 899

Fig. 7 Divergence localization comparison with state-of-the-art methods. Legends: Red triangle
[5], Green triangle [1], Orange triangle our method, Blue triangle Ground truth. The top two rows
show divergence localization at low-density scenes. False divergent center detections by our method
at 2nd , 3rd image at the first row and 1st image, second row. Existing art failed to detect divergent
centers at high density (last two images in the third row) while our method can accurately detect
divergent centers

velocity. Similarly, the method in [1] shows poor performance at high-density diver-
gence. Assumption of three moving objects and solving three straight-line equations
is possible at low density but, at the high-density crowd. As there can be hundreds
of people in the neighbor of a divergent center and solving straight line equations
for hundreds of moving objects is computationally very expensive. The method with
existing model settings is unable to detect actual divergent center locations at high
density. However, the model can be improved by obtaining intersections of a large
number of neighboring moving objects with reasonable compute. False detections by
existing art on the high-density crowd can be seen in the last two images of the third
row in Fig. 7. Our method perform well at high density scenario and can idenfity
divergence points at correct locations.
As mentioned earlier our localization algorithm also identifies divergence regions
(or divergence mask) evolves temporally. Figure 8 shows three examples of diver-
gence mask detected by our algorithm. The first row is a crowd divergence sequence
taken from the PETS2009 dataset, the second row shows people diverging from the
normal path of walking over the zebra crossing, and the last row shows people
900 M. U. Farooq et al.

Fig. 8 Divergence region mask changes over time. Divergence regions: first row PETS2009;
second-row UCF; and third-row Synthetic Kabbah datasets. The first column is the region at the time
of divergence detection (td ) by CNN. The second column represents divergence region detection
n_frames after td , the third column represents divergence region detection 2*n_frames after td, and
so on. The value of n_frames is different for each sequence

diverging from circular Tawaf after completing seven laps of Tawaf. Images in
columns show the temporal progression of divergence where images are taken every
n_frames after the time instant (t d ) divergence is detected by CNN. Divergence mask
size can be seen increasing over time indicating variations in the size of divergence
and also shows the direction in which divergence is leading.

4.3 Localization—Quantitative Evaluation

As mentioned earlier our divergence localization algorithm not only identify diver-
gence source points but also detect region(s) of divergence. To qualitatively evaluate
the performance of our algorithm, we compute Intersection over Union (IoU) between
the predicted divergence region and ground truth divergence region. Ground truth
regions are obtained by hand labeling divergence regions at each abnormal frame.
IoU score is calculated using Eq. (2),

Area of Overlap
IoU = (2)
Area of Union
Anomaly Localization at High-Density Crowd … 901

where,
Area of overlap is the overlap area between the predicted region and ground truth
region.
Area of union is the area encompassed by both the predicted bounding box and
ground truth bounding box.
IoU score greater than 0.5 (50% overlap) is generally considered a good prediction
by algorithm [14]. Two samples of divergent regions detected overlaid with ground
truth divergence region are shown in Fig. 9. Again it is clear from Table 2 that at low-
density datasets (UMN, PETS2009, NGSIM), the divergence shape is not smooth
and is broken, resulting in a low IoU score, whereas the IoU score at high densities is
better meaning our algorithm is better able to detect divergent regions at high crowd
densities. IoU scores for divergent datasets used in this work are provided in Table
2.

Fig. 9 Examples of divergence region by our method (yellow) compared to ground truth (green)

Table 2 IoU score obtained

Dataset IoU score
by our method for divergence
regions UCF 0.70
UMN 0.55
PETS2009 0.63
NGSIM 0.68
SYN-KABBAH 0.82
SYN-LOVEPARADE 0.89
CONCERT 0.76
PILGRIM 0.80
* IoU score is averaged for dataset contain multiple sequences
902 M. U. Farooq et al.

4.4 Crowd Density Estimation Inside Localization Mask

We use pre-trained CSRNet from [10] on high-density crowd images from Kabbah
and Hajj and perform experimentation on benchmark datasets. Figure 10 shows a
crowd density map and crowd count inside divergence masks for various datasets.
The crowd count inside the divergence mask is compared with the ground truth count.

4.5 Angle of Localization

In this section, experimentation is performed on divergence at low-density datasets

(UCF and PETS2009) and divergence at high-density crowd datasets (PILGRIM and
CONCERT).
Figure 11 shows the normal crowd motion mask (green color) and divergent crowd
mask (red color) for UCF. Tracklets obtained after the particle advection process are
also plotted on the same image (different segments tracklets superimposed on a
single image). Noisy tracklets whose length is shorter than a given threshold δ are
removed
 1 1  by computing the Euclidean distance between start and end points i.e.,
xi , yi − xiT , yiT 2 < δ. Dominant flows inside each mask are also shown with
start and endpoint market with circle and arrow respectively.
One dominant flow from the normal mask and a single dominant flow of the diver-
gent mask is selected for subsequent computation of divergence angle. Initially, the
particle angle ∅i for all particles in a dominant flow are computed from interpolated
u and v velocities of OPF. ∅i are then mapped on the unit circle to find the exact
quadrant in which the dominant flow lies. Figure 11 shows the mapping of each
particle circular angle ∅i of a dominant flow onto the unit circle (first column) with a
mean circular angle (angles in radian). It also shows the quadrant in which particular
dominant flow lies and hence shows the direction of the normal and diverging crowd.
Similarly, the second column represents a circular angle histogram (in degrees) with
a mean circular angle plotted as well. The angle of divergence () is computed by
taking the difference of mean circular angle of normal dominant
  ( ∅n ) and mean
flow
 
circular angle of divergent dominant flow (∅d ) i.e.,  = ∅n − ∅d . Figure 11 also
shows  computed for UCF divergent scenario found to be diverging crowd drifts
aways 24.82° from the actual path of the normal walking crowd. Results for PETS,
PILGRIM, and CONCERT datasets are shown in Fig. 12.

5 Conclusion

In this paper, we introduce a more extended view of looking into anomaly localiza-
tion to make this information more useful for real-life rescue resource deployments.
a localization mask is initially generated and features of source/starting point, crowd
Anomaly Localization at High-Density Crowd … 903

Fig. 10 Crowd density map and density count inside divergence mask. 1st row: UCF; 2nd row:
PETS2009; 3rd row: PILGRIM left-divergence; 4th row: PILGRIM right divergence; 5th row: two
divergences at CONCERT
904 M. U. Farooq et al.

Fig. 11 Divergence (localized) direction and angle estimation for UCF dataset

density estimation, and direction/angle of the localized region are computed. Exper-
imentation results show that the proposed method outperforms the existing state-of-
the-art method for source/starting point detection at the high-density crowd and CDE
and direction/angle estimations are close to the ground truth.
Anomaly Localization at High-Density Crowd … 905

Fig. 12 Direction and angle identification of localized region for PETS (1st row), PILGRIM (2nd
row), and MELAN (3rd row)

Acknowledgements The Massmotion crowd simulation software is supported by the Center for
Intelligent Signal for Imaging Research (CISIR) under PO number 3920089787/30.10.2017.

References

1. Chen CY, Shao Y (2015) Crowd escape behavior detection and localization based on divergent
centers. IEEE Sens J 15(4):2431–2439. https://doi.org/10.1109/JSEN.2014.2381260
906 M. U. Farooq et al.

2. Ongun C, Temizel A, Temizel TT (2014) Local anomaly detection in crowded scenes using
finite-time lyapunov exponent based clustering. 11th IEEE international conference advance
video signal-based surveillance, AVSS, pp 331–336. https://doi.org/10.1109/AVSS.2014.691
8690.
3. Ren WY, Li GH, Chen J, Liang HZ (2012) Abnormal crowd behavior detection using behavior
entropy model. In International conference on wavelet analysis and pattern recognition, pp
212–221. https://doi.org/10.1109/ICWAPR.2012.6294781
4. Wu S, Moore BE, Shah M (2010) Chaotic invariants of lagrangian particle trajectories for
anomaly detection in crowded scenes. In: Proceedings of the IEEE computer society conference
on computer vision and pattern recognition, pp 2054–2060
5. Wu S, Wong HS, Yu Z (2014) A bayesian model for crowd escape behavior detection. IEEE
Trans Circ Syst Video Technol 24(1):85–98. https://doi.org/10.1109/TCSVT.2013.2276151
6. Chaker R, Al Aghbari Z, Junejo IN (2017) Social network model for crowd anomaly detection
and localization. Pattern Recognit 61:266–281. https://doi.org/10.1016/j.patcog.2016.06.016
7. Mehran R, Moore BE, Shah M (2010) A streakline representation of flow in crowded scenes.
Lect Notes Comput Sci (including Subser Lect Notes Artif Intell Lect Notes Bioinformatics),
vol 6313 LNCS, no PART 3, pp 439–452. https://doi.org/10.1007/978-3-642-15558-1_32
8. Muhammad Farooq U (2020) Divergence behavior detection and localization using Motion
Shape Image (MSI). Universiti Teknologi PETRONAS
9. Krausz B, Bauckhage C (2012) Loveparade 2010: automatic video analysis of a crowd disaster.
Comput Vis Image Underst 116(3):307–319. https://doi.org/10.1016/j.cviu.2011.08.006
10. Li Y, Zhang X, Chen D (2018) CSRNet: dilated convolutional neural networks for under-
standing the highly congested scenes. Proc IEEE Comput Soc Conf Comput Vis Pattern
Recognit February: 1091–1100. https://doi.org/10.1109/CVPR.2018.00120
11. https://github.com/ZhengPeng7/CSRNet-Keras
12. Khan SD, Bandini S, Basalamah S, Vizzari G (2016) Analyzing crowd behavior in naturalistic
conditions: Identifying sources and sinks and characterizing main flows. Neurocomputing
177:543–563. https://doi.org/10.1016/j.neucom.2015.11.049
13. Berens P, Velasco MJ (2009) Journal of statistical software. Wiley Interdiscip. Rev. Comput.
Stat. 1(1):128–129. https://doi.org/10.1002/wics.10
14. Ahmed F, Tarlow D, Batra D (2015) Optimizing expected intersection-over-union with
candidate-constrained CRFs. Proc IEEE Int Conf Comput Vis Inter 1850–1858. https://doi.
org/10.1109/ICCV.2015.215
EEG Based Brain Controlled RC Car
with Attention Level

Muhammad Talha Ejaz , Ammara Zahid ,

and Muhammad Mudassir Ejaz

Abstract The brain-controlled interface is gaining popularity in the academic and

research industry due to its promising performance in many fields. In this research
work, a prototype is mentioned to help the Quadriplegic patients move the wheelchair
from their thinking decision. The proposed method used the BCI technology that
takes the input data in an EEG signal using a Neurosky headset. The signal is then
processed and classified into five significant actions. The movement of the remote-
controlled (RC) car is controlled through attention levels. We incorporate the Internet
of Things (IoT) for wireless communication that transmits the data using an android
application. For safety purposes, we have designed a mobile application that allows
the user to control the RC car manually. The proposed method was implemented on
the RC car, and the attention level of 5 different subjects was recorded for 3 min. A
video of our experiments can be found at https://www.youtube.com/watch?v=UzP
Gdy54AZw.

Keywords EEG · BCI · Nuerosky · Quadriplegia

1 Introduction

Spinal Cord Injuries (SCI) are increasing rapidly. According to the study of the
National Spinal Cord Injury Statistical Center (NSCISC) at UAB. The most promi-
nent cause of SCI is vehicular accidents, which are approx. 39.3%, then by falling
31.8% and so on. One of the US studies shows that around 1.25 million died in road

M. Talha Ejaz (B)

Department of Mechatronics Engieering, PAF—Karachi Institute of Economics and Technology,
Karachi, Pakistan
A. Zahid
Department of Biomedical Engineering, NED University of Engineering and Technology,
Karachi, Pakistan
M. Mudassir Ejaz
Department of Electrical and Electrincs Engineering, Universiti Teknologi PETRONAS (UTP),
Seri Iskandar, Malaysia

© The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 907
R. Ibrahim et al. (eds.), International Conference on Artificial Intelligence
for Smart Community, Lecture Notes in Electrical Engineering 758,
https://doi.org/10.1007/978-981-16-2183-3_85
908 M. Talha Ejaz et al.

accidents in the world [1]. Quadriplegia is generally caused by road accidents, which
becomes a global catastrophe and leading cause of death by injury. Quadriplegia is
one of the spinal cord injuries in which the lower body becomes paralysis. It is caused
by damage to the spinal cord at a level of C1–C7 or sometimes brain injury.
Quadriplegia caused several complications, including pain, blood clots, pressure
sores, related injuries, spastic muscles, respiratory problems, automatic dysreflexia,
and loss of bladder and bowel control. The most important complication is the loss of
the motor sensing control of limbs and sometimes hands also. These complications
make a person handicapped. Different types of wheelchairs, including manual hand-
controlled, hand gesture-controlled, voice recognition control, eye motion, finger
movement control, and remote control, are available in the market. Still, some of
them are not feasible for Quadriplegia patients, and some are expensive.
Brain-controlled wheelchairs are designed for those patients who are unable to
use the powered wheelchairs for mobility. After the successful advancements in
robotics, artificial intelligence, and sensor technology, more advanced methods come
into existence for designing a smart wheelchair. EEG based controlled wheelchairs
are very popular nowadays as they are cost-effective and easy to operate. EEG or
electroencephalography is a process to record and detect brain activity in terms of
signals in real-time. The human brain comprises 100 billion neurons, and signals
are transmitted from one neuron to another through electric impulses. Alpha, Beta,
Theta, Delta, and Gamma are five different waveforms found in the human brain, and
each one is different from another in terms of frequency and amplitude. Electrodes
are attached on a scalp, and conductive gel is used to get better results to record or
monitor an EEG signal. It takes around 30 min for the setup, which is not suitable
to drive a wheelchair. To consider this issue, Neurosky is being used to detect the
brain signals, which does not need any gel before using it, and this method is fast and
cost-effective. However, this method is not accurate as detecting the signals using
electrodes with gel.
Brain-Computer Interface (BCI) is an approach where a physical device is operates
using brain activity. Attention level and eye blink are the main methods used in the
BCIs application. In [2–4], attention level was used to analyze human activities such
as sleep patterns and recognized the degree of human attention. When a person blinks
an eye, a deflection in a waveform is detected. By capturing that deflection, several
applications were proposed in [5–8]. In [9, 10], methods were discussed for the
wheelchair for quadriplegia patients using the head movement and retina movement
utilizing the image processing and accelerometer technique. These methods were not
suitable for a person, as it causes fatigue while moving the head. A neuro handset is
used to record the attention level for driving a wheelchair that mitigates the issues.
It makes a system faster and allows the patient to control the wheelchair by focusing
the directions.
In this paper, we proposed a conceptional idea where RC car is drive using the
attention level. This method helps the handicapped especially quadriplegia patients
to move from one place to another on their own. We also incorporate IOT in the
proposed method, where signals are transmitted to an android application.
EEG Based Brain Controlled RC Car with Attention Level 909

The rest of the paper is organized as follows. Section 2, covers the background
knowledge of different types of EEG signals and recording methods. The proposed
methodology and results are discussed in Sects. 3, 4 respectively. Section 5 covered
the conclusion.

2 Background

Electroencephalography, abbreviated as EEG, is a process to monitor the brain

activity of a brain. Signals are taken using the electrodes attached to a scalp. Different
EEG headsets are used to monitor the different number of channels. It is currently
widely used for Brain Controlled Interface (BCI) applications as it is a non-invasive
method with high time resolution and a low cost. Countless research [11, 12] has
been conducted for the Motory Imagery BCIs, where users imagine the movements
in a brain, generates a signal that capture by an EEG headset. More the channels
in the headset mean more accurate results will be. The cells in the brain called
neurons produced an action potential by communicating with each other. The opening
and closing of the ion channels (Na+ and K+ ) in the neuron’s membrane cause this
discharge known as the action potential.
One main advantage of using an EEG headset is that it is less weight and a
portable device, allowing users to wear it and control the wheelchair. However, two
main challenges are faced by EEG signals, such as artifacts caused by the noise and
pose determination. EEG signals are mostly disturbed when the noise is present in
the surroundings, and it also affects the position of a body.
Alpha wave (α) is generated when a person is relaxed and profoundly found in the
frontal lobe or at the temporal lobe. It is more activated when a person thinks while
eyes are open, and the range of the frequency starts from 7.5 Hz to 12 Hz. Beta wave
(β) has the second-highest frequency, ranging from 12 to 30 Hz, and activated during
the stress and working a hazard scenario. It is further classified into two bands, such
as β1 and β2. It is also visible during sleep and rapid eye movement. Theta wave (θ)
is mainly shown when a person is emotionally disturbed and the time when a person
awake from a deep sleep. It lies between 3.7 and 7.5 Hz. Delta wave (δ) has the lowest
frequency bands, and it is activated when a person becomes unconscious or thinks
very deeply during sleeping. The frequency range of this waveform iss 0.5–3.5 Hz
[13]. A comparison is mentioned in Table 1.
Signal acquisition of MI BCIs is not expensive, but it’s a challenging task. EEG
signals are not stable, which, due to which sometimes it fails to classify the wave, and
many classifiers fail to consider time-series information [14]. The general structure
of MI BCIs application is depicted in Fig. 1.
910 M. Talha Ejaz et al.

Table 1 Frequency description of EEG bands

Type Frequency (Hz) Location Normally
Alpha (α) 7.5–12 Posterior regions of head, both Relaxed/reflecting closing the
sides, higher in amplitude on eyes
dominant side. Central sites
(c3–c4) at rest
Beta (β) 12–30 Both sides, symmetrical Alert/working active, busy or
distribution, most evident anxious thinking, active
frontally. Low amplitude waves concentration
Theta (θ) 3.7–6.7 – Young children drowsiness or
arousal in older
Delta (δ) 0.5–3.5 Frontally in adults, posteriorly Adults slow wave sleep in
in children; high amplitude babies
waves
Gamma (γ) 30–100 – Certain cognitive or motor
functions

Fig. 1 A signal processing flowchart of MI BCIs applications

3 Methodology

The novelty of the proposed method is that Neurosky was used as a sensor. It is
an EEG headset for acquiring the EEG signal. This setup not reduced the cost of
the overall system but increase efficiency. Neurosky is a one channel band with
a sampling rate of 512 Hz. It is the only sensor used to drive an RC car in four
directions: right, left, forward, and backward. The specification of Neurosky (Mind
wave) is mentioned in Table 2.
Neurosky is placed on a frontal side of a head precisely at the FP1 location. It has
a dry electrode, which senses the EEG and gives the value during attention and eye
blink. The Neurosky and the site where it is placed is shown in Fig. 2.
We dealt with the attention level of a person. Attention was calculated by acquiring
a signal from a device and passed to the signal processing unit. EEG is recorded as a
sampling frequency of 512 Hz with a 16-bit quantized level. Fast Fourier Transform
(FFT) is being used to convert time signal into a frequency domain with a low pass
filter with a cutoff frequency of 50 Hz.

Table 2 Frequency description of EEG bands

Channels Electrode type Bandwidth Sampling rate Communication Weigth
1 Dry 3–100 Hz 512 Hz Bluetooth 90 g
EEG Based Brain Controlled RC Car with Attention Level 911

(a) (b)

Fig. 2 a shows the Neurosky headset and b depicts the position where Neurosky attached to the
scalp

3.1 Design Flow

In this section, we are discussing the hardware and software flow of the proposed
method. (Appendix shows the hardware and software flow of a system). The flow
starts by calibrating the headset, as all the movement depends upon the data acquired
from the headset. After obtaining the data, it was propagated to the micro-controller
to perform the actions. L298N motor driver was also attached to the microcontroller
that received the command and then powered the motors for the actions. The actions
were also monitored in the mobile application, which was designed in an android
studio for this particular application. For communication between the RC car and
application was done by the Wi-Fi module (ESP8266). In the proposed method,
four basic movements were considered: forward, backward, left, and right. One
more additional command, stop, was considered for the safety purpose. The RC car
moved when the attention level was greater than 10. The actions were classified as
mentioned in Table 3.

Table 3 Action classification

Attention level Action
according to the attention
level 10≤ Attention level Stop
10≤ Attention level ≤40 Forward
40≤ Attention level ≤50 Reverse
50≤ Attention level ≤80 Right
80≤ Attention level ≤100 Left
912 M. Talha Ejaz et al.

4 Experimental Results

We had conducted the experiments on the RC car, we named it as SpyShip. Figure 3

shows the model of SpyShip. The communication between the headset and actuators
of the car was done wirelessly using a Bluetooth module. The attention level was
measured for three minutes of 5 subjects, and they were asked to focus on the direc-
tions to move the car. Figure 4 depicts each subject’s attention level curves at each
time step. The result shows that the proposed method can capture the attention level
without any disturbance and then derived the car in real-time.
Figure 5 shows that the right direction was chosen mostly by all the subjects and
then the forward direction. The maximum value of the right direction taken by a
subject is 108. Whereas for forwarding is 75, backward is 52, and for the left is 62.
The mean value of each direction of five different subjects is calculated and depicted
in Table 2. The average mean value of the right direction is 83.4, and the p-value of
the data set is 0.0000367. Figure 6 shows the Android application interface and the
action selected by the subject.
The subjects mostly have their attention level between 50 and 80. It can also be
seen that the proposed method can capture the data in real-time and then classify

Fig. 3 SpyShip, side view

of the RC car

Fig. 4 Attention level of each subject at each time step

EEG Based Brain Controlled RC Car with Attention Level 913

Fig. 5 Illustration of actions selections by each subject

(a) (b)

Fig. 6 a shows the Andriod application interface and b depicts the action selected by the subject

it accordingly. For wireless communication, we have designed an android app that

transmits and receives the data in real-time. It was possible by using the NodeMCU
that supports ESP8266 SOC. It acts as a LAN HTTP server or LAN web server. This
research work used to convert IP base links into serial communication to process in
both station and hotspot mode. (see Appendix).
914 M. Talha Ejaz et al.

5 Conclusion

In this article, a conceptual idea is presented that can be used for Quadriplegic
patients to control the wheelchair using the EEG signals. The signals have extracted
using the Neurosky headset placed at the FP1 position of the user’s head. Actions
are classified based on attention levels such as forward, backward, left, right, and
stop. The proposed method was tested on five different subjects and results shown
that all the subjects were able to classify the actions according to the attention level.
An IoT-based approach was also used in the proposed method that transmits and
receives the data wirelessly with the help of NodeMCU. An android app was also
developed that shows the actions, and again, a user can control the car manually if
malfunctions happen. In the future, the idea will be implemented in the wheelchair
and also incorporate the GPS for tracking.

Appendix

See Figs. 7 and 8; Tables 4 and 5.

Fig. 7 Hardware flow

EEG Based Brain Controlled RC Car with Attention Level 915

Fig. 8 Software flow

916 M. Talha Ejaz et al.

Table 4 Station mode IP

No Links Serial output
base link to serial characters
1 http://192.168.1.12/?pin=Forward F
2 http://192.168.1.12/?pin=Backward B
3 http://192.168.1.12/?pin=Right R
4 http://192.168.1.12/?pin=Left L
5 http://192.168.1.12/?pin=Stop S

Table 5 Hotspot mode IP

No Links Serial output
base link to serial characters
1 http://192.168.43.12/?pin=Forward F
2 http://192.168.43.12/?pin=Backward B
3 http://192.168.43.12/?pin=Right R
4 http://192.168.43.12/?pin=Left L
5 http://192.168.43.12/?pin=Stop S

References

1. World Health Organization Website (2017) [Online]. Available: http://www.who.int/mediac

entre/factsheets/fs358/en/
2. Liu NH, Chiang CY, Chu HC (2013) Recognizing the degree of human attention using EEG
signals from mobile sensors. Sensors 13(8):10273–10285
3. Van Hal B, Rhodes S, Dunne B, Bossemeyer R (2014) Low-cost EEG-based sleep detection.
In: Proceeding of 36th annual international conference of the IEEE engineering in medicine
and biology society (EMBC), pp 4571–4574
4. Poltavski DV (2015) The use of single-electrode wireless EEG in biobehavioral investigations.
Methods Mol Biol 1256:375–390
5. Jiang Y; Lee H, Li G, Chung W-Y (2016) High performance wearable two-channel hybrid BCI
system with eye closure assist. In: Proceeding. of 38th annual international conference of the
IEEE engineering in medicine and biology society (EMBC), pp 5869–5872
6. Ang MS, Zhang ZG, Hung YS, Mak JNF (2015) A user-friendly wearable single-channel EOG-
based human-computer interface for cursor control. In: Proceeding of IEEE 7th international
IEEE/EMBS conference on neural engineering (NER), pp 565–568
7. Varela M (2015) Raw EEG signal processing for BCI control based on voluntary eye blinks.
In: Proceeding. of IEEE thirty fifth central american and panama convention (CONCAPAN
XXXV), pp 1–6
8. Szibbo D, Luo A, Sullivan TJ (2012) Removal of blink artifacts in single channel EEG. In:
Proceeding of annual international conference of the IEEE engineering in medicine and biology
society (EMBC), pp 3511–3514
9. Baig MZ, Aslam N, Shum HPH, Zhang L (2017) Differential evolution algorithm as a tool for
optimal feature subset selection in motor imagery EEG. Expert Syst Appl 90:184–195
10. Oikonomou VP, Georgiadis K, Liaros G, Nikolopoulos S, Kompatsiaris I (2017) A comparison
study on EEG signal processing techniques using motor imagery EEG Data. In: Proceeding
of the IEEE 30th international symposium on computer-based medical systems (CBMS),
Thessaloniki, Greece, 22–24 June 2017
11. Cheng D, Liu Y, Zhang L (2018) Exploring motor imagery EEG patterns for stroke patients
with deep neural networks. In Proceeding of the IEEE international conference on acoustics,
speech and signal processing (ICASSP), Calgary, AB, Canada, 15–20 April 2018
EEG Based Brain Controlled RC Car with Attention Level 917

12. Awais MA, Yusoff MZ, Yahya N, Ahmed SZ, Qamar MU (2020) Brain controlled wheelchair:
a smart prototype. J Phys: Conf Seri 1529(4):042075
13. Khan D, Kamel N, Muzaimi M, Hill T (2020) Effective connectivity for default mode network
analysis of alcoholism. Brain Connectivity
14. Guo X, Wu X, Gong X, Zhang L (2013) Envelope detection based on online ICA algorithm
and its application to motor imagery classification. In Proceeding of the 6th international
IEEE/EMBS conference on neural engineering (NER), San Diego, CA, USA, 6–8 November
2013
Automated Pill and Syringe Dispenser

R. Paviya, S. Prabakar, K. Porkumaran, and Abu Bakar Sayuti Saman

Abstract This paper presents the Automated Fabrication of Pill and Syringe
Dispenser. As it is necessary to provide medication for the elderly people on time
a medication remainder is already developed. So, in addition to help the diabetic
patient, automatic pill and syringe dispenser is recommended. It is designed in such
a way that, based on the real time clock it dispenses the pill and syringe for the patients
on required time and it acts as a reminder. In addition to this, automatic syringe
dispenser is developed, which loads the insugen automatically. The major compo-
nents of the combination of pill and syringe dispenser are DS3231 RTC, Arduino
UNO, Servo motor, LCD, LED, Buzzer, and in addition to this, the software called
BLYNK, is used to pass the information through Bluetooth and the medication is
dispensed on the required time for the elderly ones. And the components used for
automatic syringe dispenser are Arduino UNO, lead screws, L293d, DVD stepper
motor board, servo motor, piezo buzzer.

Keywords Medication dispenser · Automatic dispenser · Pill and Syringe · Micro

controller · Servo motor · Alzheimer disease · Android application · Syringe
dispenser

R. Paviya (B)
N.G.P Institute of Technology, Kalapatti road, Coimbatore, India
S. Prabakar
Sona College of Technology, Salem, India
e-mail: [email protected]
K. Porkumaran
Sri Sairam Engineering College, Chennai, India
A. B. S. Saman
Universiti Teknologi PETRONAS, Seri Iskandar, Perak, Malaysia
e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 919
R. Ibrahim et al. (eds.), International Conference on Artificial Intelligence
for Smart Community, Lecture Notes in Electrical Engineering 758,
https://doi.org/10.1007/978-981-16-2183-3_86
920 R. Paviya et al.

1 Introduction

As there are many elderly individuals who need constant help, to take their medi-
cations on time and to prevent any illness timing is must. So Pill Dispenser has
been developed. Also, many elders face problems like Alzheimer’s, poor vision and
Diabetics etc.., and most of the death occur due to the intake of wrong medications.
So, in addition to help the diabetic patient, Syringe dispenser is combined with the
Pill dispenser. In this combination of pill and syringe dispenser the patient themselves
has to load the insugen. So, the extension of Syringe dispenser is made. Where the
syringe is automatically loaded to the insugen and reminds the people to take their
drug on time.

1.1 Pill Dispenser

These are items which release pills at specified times.

• These devices are proposed in a motive to help the people with Alzheimer
disease. The main function of pill dispenser is to remind the patients to take
their medications on time.
• When the predetermined time comes to take the medication, the device automat-
ically releases the pill and it sounds a loud warning signal that it is time to take
the medication.

1.2 Syringe Dispenser

These are items which release syringe at specified times. And automatic Syringe
dispenser is developed. This mainly helps in loading the insugen for the required
amount and it dispenses the syringe at the predefined time. This is done with the help
of stepper motor, servo motor, lead screws, L293d IC, 3D print software for holding
the syringe.
• Their purpose is to help the Diabetic patients those who are insulin dependent, to
take their medications on time.
• It dispenses Syringe at a set interval of time, especially for the Diabetic patient.

2 Literature Survey

(i) “Smart drugs: Improving Healthcare using Smart pill box for Medicine remainder
and monitoring system” Diaa Salama Abdul Minaam, Mohamed Abd-elfattah [1].
Automated Pill and Syringe Dispenser 921

The paper “Smart drugs: Improving Healthcare using Smart pill box for Medicine
remainder and monitoring system” published in the journal Science Direct in 2018.
This has the concept, design and creation of pill box prototype which has nine separate
sub boxes to solve the deficiency in medical area. It enables the care takers to take
the determined amount of pill and timing to take pill on every day. It requires the
client to stack the box consistently in order to refill it with medicines.
(ii) “Smart Medication Dispenser” Suraj shinde, Nitin bange, Monika kumbhar,
Snehal patil [2].
The paper “Smart Medication Dispenser” published in the International Journal
of advanced research in Electronics and Communication Engineering (IJARECE)
in 2017. This uses 4 × 4 matrix keypad, GSM module, LCD display and real time
clock (RTC) and this helps to provide information automatically to patients to take
their right dosages according to the prescription at appropriate time.
(iii) “Construction of a Smart Medication Dispenser with High Degree of
Scalability and Remote Manageability” Jugeon Pak, Keehyun Park [3].
The paper “Construction of a Smart Medication Dispenser with High Degree of
Scalability and Remote Manageability” published in the Journal of Biomedicine and
Biotechnology in 2012. In this each medication have a medication tray dispenser.
The time is set in RTC and when the pre-determined time reaches the time, it gives
some kind of notification and then the user has to push a button and the medication
is dispensed from the medication tray (MDT).

3 Methodology for Both Pill and Syringe Dispenser

3.1 Introduction

For the combination of pill and syringe dispenser same methodology is used and this
dispenses the syringe and pill and it reminds the patient to take their drug on time
and in this, the patient has to load the insulin into the needle.

3.2 Block Diagram

Construction: For both the pill and syringe dispenser, same methodology is used.
The Arduino UNO has two-way input, one from the power supply and another from
the ‘BLYNK’ app via Bluetooth. The time is set in BLYNK as well as in RTC and it
is compared. When the RTC (DS3231) reaches the BLYNK time, it starts dispensing
the medications on the predefined time. Thus both the pill and syringe is released for
the patients use, for the required time (Fig. 1).
Working: Once the power supply is fed into Arduino UNO and the information
from BLYNK is received, and it sends the signal to DS3231 and it again receives as
922 R. Paviya et al.

Fig. 1 Block diagram of pill

and syringe dispenser Power supply BLYNK

Bluetooth

LCD Display

DS3231
Arduino UNO

Servo motor Buzzer

an input to the Arduino UNO and then the Arduino processes the program to run a
servo in order to dispense the pill and syringe on the required time. An Alarm gives
off with the buzzing action which is also controlled by the Arduino and then the LED
glows.

4 Methodology for Automatic Syringe Dispenser

4.1 Introduction

In this, the extension of syringe dispenser is made as, the Syringe holder is made
by using the 3D printer and the syringe moves forward and backward as the syringe
holder is placed on the CD driver setup. This setup consists of the lead screws with
stepper motor which helps in the linear motion and IC L293d is used, it mainly helps
to run 2 DC motors with same IC. It can control both the speed and direction.

4.2 Block Diagram

Construction: Arduino UNO is connected to the IC L293d and also connected to the
stepper motors of CD drive. The syringe holder is made by using the 3D printer and
the servo setup moves front and back with the help of lead screws in CD driver. And
the syringe gets dispensed and it starts to buzzer. This process continues (Fig. 2).
Working: When the power supply is fed into Arduino UNO and then it initiates
the IC L293d and the IC starts moving the CD driver motor forward and backward
with the help of lead screws and the syringe moves forward into the insugen bottle
and then the back holder of syringe moves backward, so that the insugen is sucked
inside the syringe and then both the setup moves backward and then the micro servo
Automated Pill and Syringe Dispenser 923

Fig. 2 Block diagram of

automatic syringe dispenser Start

Initialize Arduino UNO and L293d

NO
Timer
If t=y?
YES

Servosetup1, Servo setup Servo run

2 forward 2 reverse

Buzzer Stop Dispense syringe

pushes the syringe out and it starts buzzing off and notifies the patient to take the
syringe.

5 Results and Discussion

5.1 Overall Performance

The overall performance of this device is good. First the combination of both pill and
syringe dispenser (Fig. 3) has been developed. This helps to remind the patients or the
care taker to take the drug on time. It dispenses the Pill followed by the Syringe. In
this the syringe is dispensed without suction of insugen. So an extension of syringe
dispenser is done (Fig. 4). In this extension of syringe dispenser, it automatically
sucks the insugen with the help of the CD driver setup and IC. And the syringe gets
dispensed and alarm goes on and notifies the patient. This mainly helps the patient
in not missing their medications and it will not lead to any medication error.

6 Future Improvements and Scope

There are some improvements in this device to be made in future. It includes auto-
matic refilling of syringe after usage, sending message to the caretaker if the medi-
cation is not taken this can be done using the GPS, fixing the insugen bottle with the
needle, and also an input panel will be set so that the amount of insugen to be sucked
924 R. Paviya et al.

Fig. 3 Combination of pill and syringe dispenser

Fig. 4 Automatic syringe dispenser

by the needle can be entered easily and time can be set easily by the care taker. This
would be very useful for the patients and also it reduces stress for the care takers.

7 Conclusion

There are many pill boxes available which are bit complicated and most of them are
costlier. But the combination of pill and syringe dispensers is not available, so it was
developed. In this fabrication of pill and syringe dispenser, with the help of BLYNK
app the time is set and when the time reaches the pre-determined time the pill gets
dispensed followed by the syringe. In this the syringe is not loaded with insugen.
Automated Pill and Syringe Dispenser 925

So, the extension of this process has been developed with the help of CD drivers
and IC L293d the syringe automatically loads the insugen and it is released with the
help of micro servo. This is user friendly and it helps the patients in not missing
the medications. It is less weight and portable, it reduces the medication errors. This
can be used in hospitals, homecare, and other old age homes. The main goal of the
system is to provide healthy, tension free life to those who are taking pills regularly
and to provide it at an affordable cost. This product is easy to design. So, thus it
requires less maintenance.

References

1. Abdul Minaam DS, Abd-Elfattah M (2018) Smartdrugs: improving healthcare using smart pill
box for medicine reminder and monitoring system. Sci Dir 29 November (2018)
2. Shinde S, Bange N, Kumbhar M, Patil S (2017) Smart medication dispenser of a proceedings
paper: IJARECE
3. Pak J, Park K (2012) Construction of a smart medication dispenser with high degree of scalability
and remote manageability. J Biomed Biotech
Multi-classification of Brain Tumor
Images Using Transfer Learning Based
Deep Neural Network

Pramit Dutta, Khaleda Akhter Sathi, and Md. Saiful Islam

Abstract In recent advancement towards computer-based diagnostics system, the

classification of brain tumor images is a challenging task. This paper mainly focuses
on elevating the classification accuracy of brain tumor images with transfer learning
based deep neural network. The classification approach is started with the image
augmentation operation including rotation, zoom, horizontal flip, width shift, height
shift, and shear to increase the diversity in image datasets. Then the general features
of the input brain tumor images are extracted based on a pre-trained transfer
learning method comprised of Inception-v3. Finally, the deep neural network with 4
customized layers is employed for classifying the brain tumors in most frequent brain
tumor types as meningioma, glioma, and pituitary. The proposed model acquires an
effective performance with an overall accuracy of 96.25% which is much improved
than some existing multi-classification methods. Whereas, the fine-tuning of hyper-
parameters and inclusion of customized DNN with the Inception-v3 model results
in an improvement of the classification accuracy.

Keywords Image augmentation · Transfer learning · Inception-v3 · Deep neural

network · Brain tumor classification

1 Introduction

A brain tumor is defined as an uncontrolled and unnatural growth of neural cells.

According to the world cancer report, in this year around 18,020 adults may die
from primary brain cancerous disease [1]. Therefore, the early classification of brain
tumors into their particular types plays an imperative role to treat the tumor efficiently
for reducing the human death rate. In this case, the implication of deep learning
methods to classify the tumor images can accelerate the treatment process more
effectively. Recently numerous researches have been conducted on deep learning-
based classification method to increase the classification accuracy of the brain tumor

P. Dutta · K. Akhter Sathi · Md. Saiful Islam (B)

Department of Electronics and Telecommunication Engineering, Chittagong
University of Engineering and Technology, Chittagong 4349, Bangladesh

© The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 927
R. Ibrahim et al. (eds.), International Conference on Artificial Intelligence
for Smart Community, Lecture Notes in Electrical Engineering 758,
https://doi.org/10.1007/978-981-16-2183-3_87
928 P. Dutta et al.

images. For instance, Sajjad et al. [2] employed an unsupervised learning method
called the convolutional neural network (CNN) algorithm for the classification of the
brain tumor images in different classes. The accuracy was found to be almost 94.58%
for the classification of the multiple categories of tumors. Moreover, Amin et al. [3]
employed a fusion process using the discrete wavelet transform (DWT) method
to extract a more informative tumor region. Then the noise removal process was
applied based on a partial differential diffusion filter (PDDF) before segmentation.
After that, the CNN model was utilized for classifying the tumors as cancerous and
non-cancerous. In another study, Sultan et al. [4] developed an approach for the
multi-classification of brain tumors. The classification of the brain tumors images
in multiple classes was based on the CNN algorithm. The accuracy was found to be
almost 96.13% for the classification of the multiple categories of tumors. Abiwinanda
et al. [5] proposed a classification model based on CNN for the multi-classification
of brain MRI images. The architecture of the CNN model consisted of different
layers i.e., convolution, max-pooling, flattening, and fully connected one hidden
layer. The classification method based on CNN provided a classification accuracy of
approximately 94.68%.
The major contribution of this paper is to develop an efficient classification method
using CNN with the aid of the inception-v3 transfer learning method. Moreover, a
suitable learning rate and batch size are employed to make the designed model
robust and also faster and smooth the training process. In addition, the effectiveness
of the designed method is analyzed by comparing it with some existing classification
methods in terms of classification accuracy.
The rest of the paper is organized as follows: Sect. 2 provides the methodology
of the proposed system with a detailed explanation of each of the steps. Section 3
represents the results with analysis and the comparative study is also conducted in
this Section. Finally, Sect. 4 shows concluding remarks.

2 Proposed Multi-classification Approach

The process of the proposed multi-classification method is divided into three steps as
illustrated in Fig. 1. In the first step, the image preprocessing is performed using the
image augmentation technique to increase the amount of total image dataset. After
the preprocessing step, the image features are automatically generated by employing

Image Preprocessing using Im- Image Feature Image Classification us-

age Augmentation (rotation, Extraction using ing Customized Four
zoom, horizontal flip, width Inception-v3 Layer Deep Neural Net-
shift, height shift, and shear) work

Fig. 1 The process of proposed classification method

Multi-classification of Brain Tumor Images … 929

the Inception-v3 transfer learning method. Finally, the extracted features are feed
into the modified DNN to classify the three brain tumor types.

2.1 Tumor Dataset

The dataset used for this model is obtained from 233 patients with three categories of
brain tumor images at different slices (a) 994 axial images, (b) 1045 coronal images,
and (c) 1025 sagittal images. This T1-weighted contrast-enhanced image dataset is
provided by Cheng [6]. It comprised of 3064 brain tumor MRI images. The datasets
are formulated with 1426 meningioma images, 708 glioma images, and 930 pituitary
images. Each of the images has a size of 512 × 512 in pixels. For decreasing the
computational time and dimensionality the images are resized to 150 × 150 pixels.
Then the total image datasets are splitting into training (80% of the total dataset)
and validation (20% of the total dataset) dataset. The validation dataset is employed
to estimate the proposed classifier model. Moreover, the details explanation of each
step of the classifier model are described as following.

2.2 Image Augmentation

Before feeding the classifier model, the image augmentation process is performed
because of the lower number of the image dataset. For this case, various opera-
tion including rotation, zoom, horizontal flip, width shift, height shift, and shear is
performed to enhance the diversity of the brain tumor dataset.

2.3 Classifier Model Using Inception-v3 and DNN

The architecture of the proposed classifier model is comprised of Inception-v3and

customized deep neural network as shown in Fig. 2a to classify the brain tumor
classes effectively.

2.3.1 Feature Extraction Using Inception-v3

After image augmentation, the features are automatically extracted by using the
Inception-v3 based transfer learning model. The model utilizes two or three layers
of a small convolutional layer based on factorized convolution operation instead of a
large convolution layer that reduces the parameter without reducing the efficiency of
the model. The factorization process is represented by the block as shown in Fig. 2a.
This model also employed a grid size reduction technique for mapping the features
930 P. Dutta et al.

Inception-v3
(a)

(b) (c)

Fig. 2 a The architecture of proposed classification model, b The detail of block, and c The detail
of block_1

as shown in Fig. 2b. For this factorization process, the 42 layers Inception-v3 model
with fewer parameters is much more efficient than the VGGNet [7]. Moreover, the
default weights are used for this particular model. This model contains many filters to
detect simple feature that is very effective for classification problem. For this reason,
the images are convoluted to extract the desired features.

2.3.2 Classification Using DNN

In order to make the classifier more effective, the output of the Inception-v3 model
is coordinated with the customized DNN that results in fine tuning of the transfer
learning model. Figure 3 shows the network architecture of DNN with four layers
where the first layer is used to attend the output of the inception model and the second
layer consisting of 1024 neurons and the next layer is utilized to dropout. Finally,
the output layer is designed to generate the output with 3 neurons that represents the
brain tumor types of glioma, meningioma, and pituitary respectively.
For training the network, the Adam is employed as an optimizer and categorical
cross-entropy is utilized as a loss function with the learning rate of 0.00003 and
batch size of 32. The training process of the designed model is performed in the
Multi-classification of Brain Tumor Images … 931

Fig. 3 The architecture of deep neural network

Google Colab environment using the Graphics Processing Unit (GPU). The training
process takes about 6 min for 19 epochs (callback stop) with 20 s per epoch. During
the training, the model occupied 3.06 GB RAM and 2.53 GB GPU in the colab
environment.

3 Results Analysis

After completing the training process, the loss and accuracy of the designed classifier
model are evaluated at different epochs as shown in Fig. 4. The training and validation
loss are found 0.0614 and 0.1468 that ensure optimum performance of the classifier
model since the validation loss shows a decreasing outline with the increasing of
the number of epochs. Moreover, the accuracy is obtained 97.80% and 96.25% for
training and validation sets that ensure a good fit of the classifier model because both
sets show an increase of accuracy value for every epoch.

Fig. 4 The loss and accuracy of the proposed classification model

932 P. Dutta et al.

Table 1 Precision and recall of the proposed classifier set at different epochs
Epoch At epoch At epoch At epoch At epoch At epoch At epoch At epoch
1 4 7 10 13 16 19
Training 0.7859 0.9166 0.9490 0.9524 0.9644 0.9726 0.9816
precision
Validation 0.8094 0.9373 0.9522 0.9343 0.9161 0.9195 0.9639
precision
Training 0.6663 0.9013 0.9412 0.9466 0.9621 0.9718 0.9755
recall
Validation 0.7276 0.9266 0.9429 0.9282 0.9086 0.9135 0.9592
recall

For evaluating the performance of the proposed model, various metrics including
precision, recall are obtained at different epochs shown in table 1. The precision
and recall of the training set are found 98.16% and 96.39% and for validation set
are found 97.55% and 95.92% respectively. Moreover, the F1 Score for training
and validation set is calculated 97.85% and 96.15% respectively depending on the
following equation.

precision * Recall
F1 Score = 2 × × 100% (1)
Pricision + Recall

Table 2 shows the comparison of the proposed work with some existing research
works based on the classification method and performance parameter. From this table,
it is shown that the accuracy is quite improved by using the transfer learning-based
CNN classifier proposed in this work. Therefore, the proposed method is proficient to
significantly improve the multi-classification accuracy comparing with other models
presented in the table.

Table 2 Comparison of the

References Classification method Accuracy (%)
proposed model with the
existing model [1] Modified VGG-19 CNN 94.58
[4] CNN 96.13
[5] CNN 94.68
[8] Deep transfer 95.23 ± 0.6
learning with ResNet-50
[9] VGG-16 with LSTM 84.00
This work Inception-v3 with 96.25
customized DNN
Multi-classification of Brain Tumor Images … 933

4 Conclusion

This paper proposed a method of multi-classification of brain tumor images based on

inception-v3 transfer learning with customized DNN in MRI images. The fine-tuning
transfer learning model is employed after the image augmentation can improve the
accuracy of the classifier for precisely classifying the tumor images as glioma, menin-
gioma, and pituitary respectively. The proposed deep neural network-based classifier
provides satisfactory performance in terms of accuracy of 96.25%. Therefore, the
improved classification accuracy has a great impact on the computer aided diagnosis
of brain cancer that assists the physicians to make an exact decision for the treatment
of the neurological patient.

References

1. Siegel RL, Miller KD, Jemal A (2020) Cancer statistics. CA A Cancer J Clin 70:7–30 https://
doi.org/10.3322/caac.21590
2. Sajjad M, Khan S, Muhammad K, Wu W, Ullah A, Baik SW (2019) Multi-grade brain tumor
classification using deep CNN with extensive data augmentation. Journal of computational
science 1(30):174–182
3. Amin J, Sharif M, Gul N, Yasmin M, Shad SA (2020) Brain tumor classification based on DWT
fusion of MRI sequences using convolutional neural network. Pattern Recogn Lett 1(129):115–
122
4. Sultan HH, Salem NM, Al-Atabany W (2019) Multi-classification of brain tumor images using
deep neural network. IEEE Access. 27(7):69215–69225
5. Abiwinanda N, Hanif M, Tafwida Hesaputra S, Handayani A, Mengko TR (2018) Brain tumor
classification using convolutional neural network. In: World congress on medical physics and
biomedical engineering 2018. Springer, Singapore, pp 183–189
6. Cheng J (2017) Brain tumor dataset (version 5) 2017. https://doi.org/10.6084/m9.figshare.151
2427.v5.
7. Tsang S-H (2018) Review: inception-v3 — 1st Runner Up (Image Classification) in
ILSVRC 2015. https://sh-tsang.medium.com/review-inception-v3-1st-runner-up-image-classi
fication-in-ilsvrc-2015-17915421f77c. Accessed 10 September 2018
8. Talo M, Yildirim O, Baloglu UB, Aydin G, Acharya UR (2019) Convolutional neural networks
for multi-class brain disease detection using MRI images. Comput Med Imaging Graph. https://
doi.org/10.1016/j.compmedimag.2019.101673
9. Shahzadi I, Tang TB, Meriadeau F, Quyyum A (2018) CNN-LSTM: cascaded framework for
brain Tumour classification. In: 2018 IEEE-EMBS conference on biomedical engineering and
sciences (IECBES) 2018 Dec 3. IEEE, pp 633–637
A Meta Model Based Particle Swarm
Optimization for Enhanced Global
Search
Rasel Ahmed , Shuhaimi Mahadzir ,
and Nor Erniza Mohammad Rozali

Abstract The performance of a metaheuristic algorithm depends on the appropriate

selection of its behavioral parameters. A good selection of parameters increases the
search ability of an algorithm and avoids premature convergence. Particle swarm opti-
mization (PSO) is swarm-based metaheuristic algorithm, which needs few parameter
adjustments and less computational time. Meta-optimization has been used to tune
the parameters and to get better results. Previously, authors applied meta optimiza-
tion techniques to specific problems to tune the parameters and to get better results
for specific case studies in different fields, but the application of meta optimization
in benchmark functions are limited. The present study proposes meta optimization-
based PSO to minimize the computational effort required for manual trial and error-
based parameter selection. The proposed algorithm is tested for 14 benchmark func-
tions (8 unimodal and 6 multimodal), and its efficiency and robustness are assessed
via statistical analysis. The algorithm outperforms other renowned established algo-
rithms (GA, PSO), and its performance remains consistent with increasing modality
and dimensionality.

Keywords Particle swarm optimization · Genetic algorithm · Evolutionary

algorithms · Swarm intelligence · Computational intelligence · Metaheuristic
algorithms · Meta optimization · Stochastic optimization

1 Introduction

Population-based evolutionary algorithms belong to a higher level of stochastic tech-

niques, which do not need detailed information, high computational resources and
expert knowledge of a system [1, 2]. They improve the quality of a solution step by
step by utilizing the basic features such as selection, variation, updating velocity and
position [3, 4]. These all involve parameters; whose values need to be chosen prop-
erly to achieve the best results for a specific problem. Parameters could be set in two

R. Ahmed (B) · S. Mahadzir · N. E. Mohammad Rozali

Department of Chemical Engineering, Universiti Teknologi Petronas, 32610 Perak, Malaysia

© The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 935
R. Ibrahim et al. (eds.), International Conference on Artificial Intelligence
for Smart Community, Lecture Notes in Electrical Engineering 758,
https://doi.org/10.1007/978-981-16-2183-3_88
936 R. Ahmed et al.

ways: tuning of parameters and parameter control [5]. Traditionally, researchers used
to do fine tuning of algorithm parameters manually to get better performance of the
algorithm. Shi and Eberhart conducted studies to manually improve the performance
of algorithms [7, 8]. Some researchers analyzed particle swarm optimization (PSO)
and showed how the parameters affect the convergence of particles to the global
optimum [9, 10]. However, these approaches are sometimes not systematic and have
limited scope. Another widely used path followed by researchers is to develop new
variants of algorithms [11, 15]. Two common ways to this are: hybridize one algo-
rithm with another for better performance and controlling the particles contraction in
the swarm [11–13]. This article presents a meta-optimization of PSO and its evalua-
tion on 8 unimodal and 6 multi-modal benchmark functions (Table 1) to test both of
its exploration (global search) and exploitation (local search) ability. The proposed

Table 1 Details of 8 unimodal and 6 multi-modal benchmark functions [15]

Sl. Function name Function Dimension Global Search Initialization
No optima range range
Uni-modals
1 Sphere F1 30 0 [−100, [−100, 100]
100]
2 Rosenbrock F2 30 0 [−5, 10] [−5, 5]
3 Dixon and price F3 30 0 [−10, [−10, 10]
10]
4 Zakharov F4 30 0 [−5, 10] [−5, 10]
5 Beale F5 2 0 [−4.5 [−4.5, 4.5]
,4.5]
6 Rotated_Hyper_Ellipsoid F6 30 0 [−100, [−100, 100]
100]
7 Branin F7 2 5/4Pi [−5, 15] [−5, 15]
8 Modified schaffer3 F8 30 0.001566 [−100, [−100, 100]
100]
Multi−modals
9 Levy F9 30 0 [−10, [−10, 10]
10]
10 Schwefel F10 30 0 [−500, [−500, 500]
500]
11 Ackley F11 30 0 [−32, [−32, 16]
32]
12 Griewank F12 30 0 [−600, [−600, 200]
600]
13 Rastrigin F13 30 0 [−5.12, [−5.12, 2]
5.12]
14 Shubert F14 2 −186.7 [−10, [−10, 10]
10]
A Meta Model Based Particle Swarm … 937

algorithm showed significant improvement in terms of quality of solution and conver-

gence and outperforms other two well-known robust optimization algorithms such
as the genetic algorithm (GA) and PSO.

2 Meta Model Based Particle Swarm Optimization

Meta optimization is a systematic way of finding optimal parameters of an algorithm

for a specific optimization problem [6]. Meta optimization is a two-step algorithm,
where the super swarm (i.e., outer algorithm) tries to find the optimum values of
algorithm parameters for use by the sub swarm (i.e., inner second algorithm) to
optimize the objective function under investigation. Getting the optimum values
of algorithm parameters by this way is an established concept in the literature to
significantly improve the performance of an algorithm. Meta optimization-based
GA, DE and PSO have already been applied to different fields of engineering. The
main elements of a meta optimization for PSO are a) super swarm, b) sub swarm
and c) objective function [14]. Super swarm particles are related to the sets of sub
swarm, in order to determine the quality of the former; the sub swarm is run up to the
number of pre-defined iterations, then the optimum final values are transferred to the
super swarm. The super swarm parameter values are selected based on the analysis
from the ranges in the literature, namely, S = 20 to 70, w = −0.5 to 0, c1 = −0.5 to
0 and c2 = 0 to 4 [6]. The details procedure of the meta model based PSO algorithm
has been shown in the following flow chart (Fig. 1).

3 Experimental Result and Discussion

To demonstrate the robustness of our meta optimized particle swarm optimization

algorithm, we tested it against fourteen renowned benchmark functions as minimiza-
tion problem. Among them the first eight functions are unimodal (F1–F8) and the last
six functions (F9-F14) are multi-modal. Among these functions F11 has a narrow
valley, F12 also has a narrow global optimum and many trivial local optima, F10
is a complex multi-modal function with many local optima. The parameters values
assigned to the three algorithms are similar such as the iteration number and particle
numbers are considered as 100, number of dimensions is 30 for all problems where
F5, F7, F14 are considered as 2-dimensional problems (except few inherent prop-
erties of GA). The search area (lower and upper bounds) are kept similar for all
three algorithms. Each algorithm has been repeatedly run for 10 times and the mean,
standard deviation and the best (minimum) values are calculated from the MATLAB
simulation environment. The details summary of the benchmark functions and their
comparative performances are summarized in Tables 1 and 2.
The unimodal benchmark functions (F1–F8) are well known for testing exploita-
tion ability of an algorithm, whereas the exploration ability of an algorithm is judged
938 R. Ahmed et al.

Input parameters

Initialize Super swarms

n =1
If n < Niteration

j =1
No Yes
J < Nsuper_swarm

n=n+1
c1, c2, w, ns

J=J+1

Run sub swarms

Update velocity, position

Update links

Evaluation

Evaluate fitness of particles

Fig. 1 Working principles of meta model base PSO [14]

by multi-modal (F9–F14) functions. In addition to this the local optima avoidance of

an algorithm can also be validated by multi-modal (F9–F14) functions. The proposed
algorithm outperforms other two algorithms in terms of exploitation, exploration and
local minima avoidance. In case of the convergence of algorithms, meta PSO shows
better convergence ability than PSO and GA for 13 of the 14 benchmark func-
tions, it also outstripped other two robust algorithms for both 30 dimensional and 2
dimensional problems simultaneously.
Among the 14 benchmark functions, for 8 benchmark functions the best
(minimum) value achieved by the meta PSO is better than other algorithms best
values. The meta PSO reached the exact global optimal solution for 6 (F5, F7, F8,
F12, F13, F14) functions where GA, PSO reached the exact global optima for 4 and 5
of these function but the GA, PSO couldn’t reach the global solution for the function
F12. For few functions F4, F6, F11 meta PSO performs significantly better than the
GA, PSO. It has been noticed from the figures (Figs. 2, 3, 4, 5, 6, 7 and 8) that for all
the considered cases meta PSO converge faster than GA, PSO except for F10 where
GA performs better than both PSO meta PSO. In terms of the mean, best and standard
deviation value (Table 2), the values that we get from the meta PSO is better than the
other two algorithms, which means that the results generated by meta PSO is more
accurate and stable. In terms of total computational time the meta PSO requires more
time to complete 10 runs (with 100 iteration each run) than GA, PSO. It happened
because of the complex nature of meta PSO algorithm and it’s wide spread search
ability where information is exchanged between the subswarms and super swarms.
Table 2 Statistical analysis and comparison of result for Algorithms
Func Name of Par GA PSO Meta-PSO Func Name of par GA PSO Meta-PSO
F1 Mean 1.3889 5.56E−07 0.0012462 F8 Mean 0.0022568 0.0015669 0.0015669
Std 0.99309 2.53 E−08 0.0021579 Std 0.0010892 3.21 E−17 0
Best 0.49275 5.28 E−07 3.11 E−07 Best 0.0015688 0.0015669 0.0015669
F2 Mean 89.0693 25.7472 18.3969 F9 Mean 1.50 E−32 1.4998 E−32 1.4998 E−32
Std 22.8912 0.60685 1.3046 Std 0 0 0
Best 66.3609 25.0824 17.1719 Best 1.50 E−32 1.50 E−32 1.4998 E−32
F3 Mean 8.0144 0.67802 0.54096 F10 Mean 3820.7129 5382.7844 3913.8624
A Meta Model Based Particle Swarm …

Std 4.3274 0.019657 0.22578 Std 458.5762 883.9675 370.0767

Best 3.2016 0.66667 0.2803 Best 3326.9552 4402.6275 3604.3912
F4 Mean 1.42 E−32 2.63 E−58 1.67 E−88 F11 Mean 0.29216 1.3116 0.068682
Std 2.30 E−32 3.68 E−58 2.90 E−88 Std 0.06441 0.32935 0.029537
Best 2.04 E−34 4.55 E−59 2.41 E−97 Best 0.2231 0.9313 0.034592
F5 Mean 5.00 E−13 0 0 F12 Mean 0.031249 0.0049307 0
Std 8.66 E−13 0 0 Std 0.039196 0.0042701 0
Best 7.33 E−21 0 0 Best 0.007396 0 0
F6 Mean 80.3431 1.36 E−05 3.15 E−11 F13 Mean 0 0 0
Std 126.9302 1.76 E−05 5.18 E−11 Std 0 0 0
Best 5.8177 1.42 E−07 3.83 E−13 Best 0 0 0
F7 Mean 0.39789 0.39789 0.39789 F14 Mean −186.7309 −186.7309 −186.7309
Std 0 0 0 Std 2.01 E−14 2.01 E−14 0
Best 0.39789 0.39789 0.39789 Best −186.7309 −186.7309 −186.7309
939
940 R. Ahmed et al.

(a) (b)

Fig. 2 Convergence curve for Sphere and Rosenbrock benchmark functions

(a) (b)

Fig. 3 Convergence curve for Dixon Price and Zakharov benchmark functions

It agreed with the no free launch theorem, that if we want to emphasize any specific
property of any algorithm, we need to sacrifice some other properties as a tradeoff.

4 Conclusion

The main objective of this article is to analyze the performance of meta model based
PSO algorithm in case of optimization benchmark functions. It outperforms the other
algorithms in terms of efficiency, convergence and robustness, it also overcomes
the limitations of manually selecting optimal behavioral parameters. The proposed
A Meta Model Based Particle Swarm … 941

(a) (b)

Fig. 4 Convergence curve for Beale and Rotated_hyper_ellipsoid benchmark functions

(a) (b)

Fig. 5 Convergence curve for Branin and Modified Schaffer3 benchmark functions

meta optimization increases the exploration and exploitation ability of the algorithms
by continuously updating the parameter values and by maintaining a good balance
between exploration and exploitation, that makes the meta model based PSO algo-
rithm to outperform other renowned algorithms and keep its performance consistent
with increasing modality and dimensionality.
942 R. Ahmed et al.

(a) (b)

Fig. 6 Convergence curve for Levy and Schwefel benchmark functions

(a) (b)

Fig. 7 Convergence curve for Ackley and Griewank benchmark functions

A Meta Model Based Particle Swarm … 943

(a) (b)

Fig. 8 Convergence curve for Rastrigin and Shubert benchmark functions

References

1. Łapa K (2019) Meta-optimization of multi-objective population-based algorithms using multi-

objective performance metrics. Inf Sci 489:193–204. https://doi.org/10.1016/j.ins.2019.03.054
2. Smit SK, Eiben AE (2009) Comparing parameter tuning methods for evolutionary algorithms.
In: 2009 IEEE congress on evolutionary computation. IEEE, pp 399–406. https://doi.org/10.
1109/CEC.2009.4982974
3. De Jong K (2016) Evolutionary computation: a unified approach. In: Proceedings of the 2016
on genetic and evolutionary computation conference companion, pp 185–199
4. Eiben AE, Smith JE (2003) Introduction to evolutionary computation. Natural Computing
Series Springer. https://doi.org/10.1007/978-3-662-05094-1
5. Eiben ÁE, Hinterding R, Michalewicz Z (1999) Parameter control in evolutionary algorithms.
IEEE Trans Evol Comput 3(2):124–141. https://doi.org/10.1109/4235.771166
6. Pedersen MEH, Chipperfield AJ (2010) Simplifying particle swarm optimization. Appl Soft
Comput 10(2):618–628. https://doi.org/10.1016/j.asoc.2009.08.029
7. Shi Y, Eberhart RC (1998) Parameter selection in particle swarm optimization. In: Porto VW,
Saravanan N, Waagen D, Eiben AE (eds) Evolutionary Programming VII. EP 1998. Lecture
Notes in Computer Science, vol 1447. Springer, Berlin, Heidelberg. https://doi.org/10.1007/
BFb0040810
8. Eberhart RC, Shi Y (2000) Comparing inertia weights and constriction factors in particle swarm
optimization. In: Proceedings of the 2000 congress on evolutionary computation. CEC00 (Cat.
No. 00TH8512), vol 1. IEEE, pp 84–88. https://doi.org/10.1109/CEC.2000.870279
9. Clerc M, Kennedy J (2002) The particle swarm-explosion, stability, and convergence in a
multidimensional complex space. IEEE Trans Evol Comput 6(1):58–73. https://doi.org/10.
1109/4235.985692
10. Trelea IC (2003) The particle swarm optimization algorithm: convergence analysis and param-
eter selection. Inf Process Lett 85(6):317–325. https://doi.org/10.1016/S0020-0190(02)004
47-7
11. Keane AJ (1995) Genetic algorithm optimization of multi-peak problems: studies in conver-
gence and robustness. Artif Intell Eng 9(2):75–83. https://doi.org/10.1016/0954-1810(95)957
51-Q
12. Lovbjerg M, Krink T (2002) Extending particle swarm optimisers with self-organized criti-
cality. In: Proceedings of the 2002 congress on evolutionary computation. CEC’02 (Cat. No.
02TH8600), vol 2. IEEE, pp 1588–1593. https://doi.org/10.1109/CEC.2002.1004479
944 R. Ahmed et al.

13. Krink T, Løvbjerg M (2002) The lifecycle model: combining particle swarm optimisation,
genetic algorithms and hillclimbers. In: International conference on parallel problem solving
from nature. Springer, Berlin, Heidelberg, pp 621–630. https://doi.org/10.1007/3-540-45712-
7_60
14. Atashnezhad A, Wood DA, Fereidounpour A, Khosravanian R (2014) Designing and opti-
mizing deviated wellbore trajectories using novel particle swarm algorithms. J Nat Gas Sci
Eng 21:1184–1204. https://doi.org/10.1016/j.jngse.2014.05.029
15. Xu X, Tang Y, Li J, Hua C, Guan X (2015) Dynamic multi-swarm particle swarm optimizer
with cooperative learning strategy. Appl Soft Comput 29:169–183. https://doi.org/10.1016/j.
asoc.2014.12.026
HCShare: Blockchain Technology
in Healthcare System

R. Sangeetha and M. Krishnamoorthi

Abstract On this generous growing world, the medical healthcare system is expe-
diting by driving new modernizations in the healthcare system by providing contem-
poraneous varieties for an inmate’s management. The enhancement of the medical
system is done by collecting enormous preceding data, accommodating those records,
scrutinizing of the data, and interchanging them with the different medical system
layers. In the headway to enhance the medical system with further preservation
of the health records, the Blockchain Technology takes the part in the system. The
prospective system likewise supports to transfer or interchange the data of the inmates
between different healthcare system, which is in the guaranteed way. HCShare is a
confidential, protected, reliable and translucent data interchanging scheme between
the different healthcare organizations in a scattered decentralized mechanism. In
which some of the cryptographic algorithms are used to ensure the protection of the
inmate’s confidential records stored in the ledger and to enhance the data integrity
of the records.

Keywords Healthcare system (HCShare) · Blockchain technology · SHA-256. ·

Data integrity

1 Introduction

As the Blockchain technology has a power to change dramatically as it is transforma-

tional, it is considered by the most healthcare organizations. As the generous growth
in the blockchain system, extensive quantity of endeavor has done [1]. The health
records of the patients are kept and maintained in the different entities such as their
local, centralized servers of the hospice, pharmaceutical places, and sometimes in
the physician’s places according to the need and storage facility of the records over
cloud platform or maybe on local paper based records, which are scattered all around
the world depends on the intention of the data [2, 3]. Hence the improvement in the

R. Sangeetha (B) · M. Krishnamoorthi

Dr. N.G.P. Institute of Technology, Coimbatore, India

© The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 945
R. Ibrahim et al. (eds.), International Conference on Artificial Intelligence
for Smart Community, Lecture Notes in Electrical Engineering 758,
https://doi.org/10.1007/978-981-16-2183-3_89
946 R. Sangeetha and M. Krishnamoorthi

Fig. 1 Blockchain network structure

development of the healthcare system has delivered by the most looming sciences
like Blockchain Technology, Internet of Things, and Edge Computing techniques.
These are the major technologies drives the healthcare system in the smartest way
with better security and integration among the other healthcare organization’s data
[4]. HCShare Healthcare system is a confidential, protected, reliable and translucent
data interchanging scheme between the different healthcare organizations in a scat-
tered decentralized mechanism [5]. HCShare system helps consistently to distribute
and digitization [6] of the records between multiple medical organizations [6]. The
transaction of the records of the inmate are transferred from one of the hospitals to
other hospitals can be done by the blockchain network. As shown in the Fig. 1, every
data that are stored in the medical organizations are shared in the scattered distributed
blockchain network, in which all the data are stored in the secured mechanism by
encrypting the data of the inmates [2].
The blockchain is referred to as the amalgam of couple of decrepit automation
equally like cryptography [6] and peer-peer communication mechanisms [6]. It is a
kind of scattered storage of records in a distributed-ledger-technology [6] in which
the ledger can be increased generously within a massive collection of data directory.
In the HCShare System model, the blockchain technology helps in the transaction of
the medical records of the patients over the different medical institutions or over the
different clinics depends on the patient’s need about the records. For example, if the
patient has been treated from hospital 1 for over the one years of duration, then in
the critical situation, they have to claim another hospice for a quick treatment. In that
situation the patient can’t able to get all their previous medical reports again by doing
all the test cases of the hospital to ensure his problem, or they can’t go the hospital A
for records of their medical reports. This HCShare system model helps the patients
to get their medical data reports whenever they need through the blockchain system.
HCShare: Blockchain Technology in Healthcare System 947

2 Related Works

In the HS-BC [7] for the electronic medical record [8], that provides the permission
to approach the patients records in a secured manner, by implementing the KUNodes
algorithm [7]. Similarly, in the [9], for securing the data of the inmates the algorithms
likely SHA-256 [9] which is to calculate the hash key value, ECC (Elliptic Curve
Cryptography) algorithm [9] for creating irregularity hash key values and for the
encryption of the records of the patients, the AES (Advanced Encryption Standard)
Algorithm is used. In [9] there have been some of the inadequacy such as absence
of integrity of the records, and the system can’t be able to perform massive requests
simultaneously [2]. When the system uses cloud transaction via a blockchain tech-
nology, it is found that there is no security for the confidential data of the inmate’s
records. In which there is lack of the data integrity such as problems in the correct-
ness of the records, lack of accuracy in the data, lack of consistency of the records.
In some of the data sharing system, also used cloud environment, in which the data
can be presented in the different layers, along with the accessibility to the data [2]. In
such cases, the third party can filch the data, and can misuse the confidential data of
the inmate’s records. The SHealth works on the different layers such as government,
user, and the IoT layers [2]. In the ChainSDI [3], the home-based medical service
sharing programming introduced to manage the confidential data of the inmates [3].
In the Sshealth [4] system, massive records are analyzed also the difficulties in the
management to provide the secured storage for the inmate data preservation [4]. In
the [5], the objections in the blockchain scalability are well focused and found the
issues like size and number of transactions of a block, and the solutions to sort it out
was classified into storage optimization and redesigning blockchain [5].

3 Proposed Model

The proposed research work on the healthcare system is to yield the security of
the inmate data, from the third-party attacks. Figure 2 provides the structure of the
blocks [6] hash values in the blockchain system [10], in which the hash values are
ordered in the Merkle tree [6]. In the HCShare the physician depot the medical reports
in the corresponding servers of the specific hospice, in which the only authorized
administrator who were allowed creates and depot the private passcode to ensure
the safety of every dossier, in the permissioned [8] blockchain system to accomplish
secured scattered ledger of hospice records. So that, every data within the ledger can
be available for the hospice, physician and for the inmate of the hospice can access
them in anytime from anywhere.
948 R. Sangeetha and M. Krishnamoorthi

Fig. 2 Merkle tree block

structure [6]

3.1 Blocks

In the HCShare System model, blocks in the blockchain transaction consists of the
couple of things, which are called as block header [11] and block body [11], in which
it contains the version numbers represented in the blocks, and hash content from the
Merkle tree as shown in Fig. 2, timestamp value and the parental hash key values.
These are shown in Fig. 3.
In the Merkle tree, the hash key generation starts from the nodes in the leaf level,
in which the parental node hash value is calculated from its combination of leaf
nodes, in which manner all the node’s hash value is calculated. For example, assume
that the leaf node of the Merkle tree as H0 , H1 , H2 , H3 , and H4 , in which the parental
node hash value is calculated as follows,
HV01 = Hash_VALUE [HV0 + HV1 ].
HV23 = Hash_VALUE [HV2 + HV3 ].
HV0123 = Hash_VALUE [HV01 + HV23 ].
In which the each and individual hash values of the leaf node contains the hash
value of the transactions. As it contains every transaction details of the blocks, the
complete Merkle tree contains the overall information of all transactions.

Fig. 3 Blocks in the blockchain system [11]

HCShare: Blockchain Technology in Healthcare System 949

Fig. 4 KUNodes algorithm

[7]

3.2 Encryption Algorithms

For the better secured data transaction among the multiple blockchain system in the
healthcare system, every blocks of the transaction contains the confidential private
key along with the hash values and the transaction details of the blocks. In the
transaction details the inmate’s records are stored according the specific hospice. For
the encryption of the inmate’s records, algorithms such as SHA-256, AES, KUNodes
algorithms are used to ensure the safety and security of the inmate records from the
malicious attacks.
In the encryption process of the HCShare System, pair of different keys has been
introduced to improve the security and safety of the ledger, they are as Priority
hash address key values and Secondary updated hash value. Where, with the help of
priority hash key, one can only view their records, which were updated in the system
by the reputed authorized hospice. So that the inmate and their family can view their
data, when it is needed in the emergency situations. Another secondary updated hash
value will help the hospice authorized admin or the physician to upload the records
the inmates, where only with the help of the secondary hash key value, one can access
completely of the ledger of particular inmate’s data.
In the HCShare system model, we propose SHA-256 for generating the priority
hash address key values, and to generate the secondary key values, KUNodes (Fig. 4),
AES algorithm is used to ensure the data integrity of the patient’s records.

3.3 Ethereum Smart Contracts

Every block in the blockchain system has its own unique key, in which each node of
the blockchain system is connected to the prior blocks of the system via a individual
hash key data [12]. Ethereum [12] is most widely used blockchain technology to give
a better transaction over the blockchain system. In which the smart contracts [12]
are used to secure the transaction details of the inmates. Smart contracts are used
to execute set of programs in which the program is executed automatically when
it is needed. These smart contracts manage the transaction of the ledger and the
cryptocurrencies [12] of the blockchain system. These programs can be encountered
without depending any other data, which performs and control overall transaction
950 R. Sangeetha and M. Krishnamoorthi

Fig. 5 Inmate record mapping with database

in the blocks. In the HCShare System model, we propose Ethereum blockchain

technology for the transaction among the hospice, physicist as well as to the inmates.
Figure 5 shows the mapping of the hash key values to access and retrieve the records
from and to the database.

4 Existing Model Analysis

SHealth [2] Multilayer-signature [2] based healthcare system designed

with sufficient functionalities like including and eradicating
of the data in the system design
Sshealth [4] External Edge is used here to manage and maintain the
efficiency of the system before every transaction occurs [4]
EHR system [1] Concussions on the blockchain system evaluated using
hybrid fuzzy-ANP TOPSIS [1] to implement the top
blockchain system for the better HER maintenance
ChainSDI [3] Home-based medical data management via smart contracts
[3], and CORD [3]
Scalability challenges [5] It focuses on the solution to the scalability challenges in size
of the block, high amount of transaction, number of nodes,
protocol used [5]. These objections were solved by the
couple of solutions as Storage Optimization [5] and
Redesigning Blockchain [5]
HS-BC [7] Efficient data security of the inmate is maintained by using
update key by KUNodes algorithm [7], which provides
another step of encryption to the data [7]. From the result of
the [7], it is proven that very cost effective
Decentralized Authentication [11] Efficient transaction duration for registering data and with
less access time for the data obtained and evaluated with the
help of NS-2V2.35 simulator [11]

In Fig. 6, the expected authentication delay of the HCShare system will be lesser
than the prior existing system, as the number of transaction increases over the commu-
nication of the inmates with the hospice, the response time is efficient as compared
to other systems.
HCShare: Blockchain Technology in Healthcare System 951

Fig. 6 Expected proposed

HCShare system

5 Conclusion

In this enormous world, the blockchain became most noteworthy technology in the
internet security [13], where the it gives all kinds of facility in the smart technological
world like in the fields of smart houses, smart grid [13] applications, as well as in
the food safety application [13], and so on. In this proposed HCShare system model,
the system gives better accessibility of the records in the blocks, by attribute-based
encryption [14] technology, in which the people without having the confidential keys
cannot have an access to the data of the inmates. So that the records of the inmates
will be stored in secured manner in the blockchain technology. We propose HCShare
system by using the Ethereum technology for the transaction of the medical reports
over the different medical institutions. Our research work plans to get the better data
integrity of the medical reports of the inmates in which all the records of the inmates
should be transferred from and to different hospice, with the complete records in
a secured manner. This HCShare is to give the better truncation efficiency over a
different block to ensure the performance of the data transaction. The lack of data
integrity is needs to be solved in the HCShare system, in which to overcome these
kinds of issues the data are secured using pair of different addresses to ensure the
safety and integrity of the data.

References

1. Zarour M et al (2020) Evaluating the impact of blockchain models for secure and trustworthy
electronic healthcare records. IEEE Access 8:157959–157973. https://doi.org/10.1109/ACC
ESS.2020.3019829
2. Zghaibeh M, Farooq U, Hasan NU, Baig I (2020) SHealth: a blockchain-based health system
with smart contracts capabilities. IEEE Access 8:70030–70043. https://doi.org/10.1109/ACC
ESS.2020.2986789
952 R. Sangeetha and M. Krishnamoorthi

3. Li P et al (2020) ChainSDI: a software-defined infrastructure for regulation-compliant home-

based healthcare services secured by blockchains. IEEE Syst J 14(2):2042–2053. https://doi.
org/10.1109/JSYST.2019.2937930
4. Abdellatif A, Al-Marridi AZ, Mohamed A, Erbad A, Chiasserini CF, Refaey A (2020) ssHealth:
toward secure, blockchain-enabled healthcare systems. IEEE Netw 34(4):312–319. https://doi.
org/10.1109/MNET.011.1900553
5. Mazlan A, Mohd Daud S, Mohd Sam S, Abas H, Abdul Rasid SZ, Yusof MF (2020)
Scalability challenges in healthcare blockchain system—a systematic review. IEEE Access
8:23663–23673. https://doi.org/10.1109/ACCESS.2020.2969230
6. Chukwu E, Garg L (2020) A systematic review of blockchain in healthcare: frameworks,
prototypes, and implementations. IEEE Access 8:21196–21214. https://doi.org/10.1109/ACC
ESS.2020.2969881
7. Su Q, Zhang R, Xue R, Li P (2020) Revocable attribute-based signature for blockchain-based
healthcare system. IEEE Access 8:127884–127896. https://doi.org/10.1109/ACCESS.2020.
3007691
8. Niu S, Chen L, Wang J, Yu F (2020) Electronic health record sharing scheme with searchable
attribute-based encryption on blockchain. IEEE Access 8:7195–7204. https://doi.org/10.1109/
ACCESS.2019.2959044
9. Li H, Zhu L, Shen M, Gao F, Tao X, Liu S (2018) Blockchain-based data preservation system
for medical data. J Med Syst 42(8) Art. no. 141153
10. Zhuang Y, Sheets LR, Chen Y-W, Shae Z-Y, Tsai JJP, Shyu C-R (2020) A patient-centric health
information exchange framework using blockchain technology. IEEE J Biomed Health Inform
24(8):2169–2176. https://doi.org/10.1109/JBHI.2020.2993072
11. Yazdinejad GS Parizi RM, Dehghantanha A, Choo K-KR, Aledhari M (2020) Decentralized
authentication of distributed patients in hospital networks using blockchain. IEEE J Biomed
Health Inf 24(8):2146–2156. https://doi.org/10.1109/JBHI.2020.2969648
12. Jaiman V, Urovi V (2020) A consent model for blockchain-based health data sharing platforms.
IEEE Access 8:143734–143745. https://doi.org/10.1109/ACCESS.2020.3014565
13. Li C-T, Shih D-H, Wang C-C, Chen C-L, Lee C-C (2020) A blockchain based data
aggregation and group authentication scheme for electronic medical system. IEEE Access
8:173904–173917. https://doi.org/10.1109/ACCESS.2020.3025898
14. Sun J, Yao X, Wang S, Wu Y (2020) Blockchain-based secure storage and access scheme
for electronic medical records in IPFS. IEEE Access 8:59389–59401. https://doi.org/10.1109/
ACCESS.2020.2982964
SAZZ Converter Fed Fuzzy Logic Speed
Controlled BLDC Motor Drive

A. Senthilnathan, P. Palanivel, and R. Balakrishnan

Abstract BLDC motors possess high torque than the same rated Induction motor.
Due to electronic commutation and the absence of mechanic commutation and
brushes, BLDC motors are widely used in many industrial applications. The main
disadvantage of BLDC motor is high torque ripple. Commutation Ripple reduc-
tion of brushless direct current (BLDC) motor torque using the SAZZ converter is
proposed in this work; torque ripple is proposed to carry out under both conduction
and commutation intervals in this proposed work. The fuzzy logic speed controller
is used for generating reference pulse in the conduction interval by processing the
speed error. The performance of torque and speed based on the proposed control
method was analyzed. MATLAB/Simulink platform is used for implementing as
well as showing effective implementation of the proposed technique. PI and Fuzzy
logic control are used for evaluating the performance. SAZZ (Snubber Assisted Zero
Voltage and Zero Current Transition) topology was proposed to reduce switching
losses. The main switch snubber circuit reduces the voltage during turn off, and the
auxiliary circuit discharges the snubber before the next switching cycle.

Keywords BLDC motor · SAZZ converter · Fuzzy logic controller

1 Introduction

Brushless DC motors possess low electromagnetic interferences, high sensitivity,

and more efficiency than a brushed motor, making BLDC motor more reliable in
most industrial applications. So BLDC motors are most commonly recommended in
industries for medium power applications such as electric traction, air conditioning,
robotics, etc. One of the major drawbacks of a BLDC motor is torque ripples due to

A. Senthilnathan (B) · R. Balakrishnan

EEE Department, Dr. N.G.P. Institute of Technology, Coimbatore, India
e-mail: [email protected]
P. Palanivel
EEE Department, E.G.S. Pillay Engineering College, Nagapattinam, India

© The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 953
R. Ibrahim et al. (eds.), International Conference on Artificial Intelligence
for Smart Community, Lecture Notes in Electrical Engineering 758,
https://doi.org/10.1007/978-981-16-2183-3_90
954 A. Senthilnathan et al.

Fig. 1 Energizing of two phases in a BLDC motor

cogging, commutation, and switching power devices [1, 2]. Many research works are
progressing to reduce current and torque ripples in PMBLDC motor. In the six-step
inverter conduction topology, at any time, only two phases are in conduction [3–
5]. Generally, the rectangular phase current is considered to be ideal for two-phase
conduction mode. Due to commutation, the incoming phase current’s slew rate time
is not equal to the outgoing phase current, which results in current pulsation [6].
This current pulsation produced torque ripple in the PMBLDC motor. Fuzzy logic
controlled SAZZ converter fed BLDC motor will create fewer torque ripples than
PI- Controller.

1.1 BLDC Motor

A permanent magnet BLDC is a synchronous motor with electronic commutation

rather than a mechanical one. A position sensor in a rotor such as a hall sensor is
used to locate the rotor position. Using rotor position sensor firing pulses for each
phase current of the Brushless DC motor is synchronized. Instead of mechanical
commutation, electronic switches are provided for proper commutation. Figure 1
shows the energizing of two winding when two switches are fired one from the upper
limb next from the lower limb.

1.2 SAZZ Converter

SAZZ (Snubber Assisted Zero Voltage and Zero Current Transition) converter is
used to reduce switching losses. The main switch snubber circuit is used to reduce
the voltage during turn off and the auxiliary circuit discharges the snubber before the
next switching cycle. Figure 2 Shows the block diagram of the SAZZ converter fed
SAZZ Converter Fed Fuzzy Logic Speed Controlled BLDC Motor Drive 955

SAZZ
3 Phase BLDC
Battery Con-
Inverter motor
verter

MOSFET Driver Rotor

7667 position
sensor

Power
Supply PI/Fuzzy logic controller

Fig. 2 Block diagram of SAZZ converter fed BLDC motor

BLDC motor. The battery is used as the input source supply and the SAZZ converter
is used to step-up DC voltage and fed the supply to the inverter. Six step inverter is
used to energize the BLDC motor. At any point, only two phases are energized as
shown in Fig. 1. The Hall sensor is used to sense the rotor position and the PI and
fuzzy controller is used to control the MOSFET driver circuit. The error and change
in error are taken as two inputs to the controller and the output of the controller is
used to fire switches in SAZZ converters. The block diagram of proposed work is
shown in Fig. 2.

2 PI Controller Based SAZZ Converter for BLDC Motor

The reference voltage of the inverter is produced based on the speed received from
the brushless DC motor. The inverter’s output voltage is compared with the refer-
ence voltage, and an error voltage is generated. This error is used to control the PI
controller.
The PI controller has a proportional gain and an integral gain. The proportional
gain (KP = 0.8) supports the analog error function and an integrator (KI = 1.1)
to reduce the constant level error value to zero. Based on the fault voltage, the PI
controller produces pulses that regulates the MOSFET. Although the PMDC motor
current follows the change in torque and change in voltage, the armature current from
Fig. 3a has different deformations, and Fig. 3b and c back EMF and torque are more
distorted and take longer to resolve.
956 A. Senthilnathan et al.

Ia (A) a) Armature current

Time in sec

b) Back EMF
Ea (V)

Time in sec
c) Torque
Te(N-m)

Time in sec
Fig. 3 Performance of PI controller based SAZZ converter fed BLDC motor

3 FLC Operation

Fuzzy logic is the usually available control approach for the above two decades. Fuzzy
logic consists of fuzzification, inference engine, and defuzzification. The Mamdani
fuzzy control system is used here. The input will be given to a fuzzifier [7–9]. A
fuzzy rule base contains the fuzzy rule set which we have to design. This input is
given to a fuzzy interface engine that compares the input with the fuzzifier inputs
to give the required output to the defuzzifier. The defuzzifier gives the output after
defuzzification and gives the respective output to the next system. The same input is
obtained at the output but after fuzzification i.e. the character of the output does not
SAZZ Converter Fed Fuzzy Logic Speed Controlled BLDC Motor Drive 957

Fig. 4 Fuzzy partition

change. This output is given to the converter to boost up the voltage and obtain the
required output. The Fuzzy Logic Control is designed by MATLAB R2014a.
Where,
• N B- negative Big
• N M- negative Medium
• N S- negative Small
• Z E- zero
• P S-positive Small
• P M- positive Medium
• P B- positive Big
Figure 4 shows the Fuzzy partition. Each input and output set is assigned with
seven linguistic variables and forty-nine rules are enclosed in FLC. Mamdani algo-
rithm is used to design the set of rules for the FLC. The fuzzy partition is shown in
Fig. 4 which includes all the 7 sets of variables that use the IF–THEN logic to design
the 49 sets of fuzzy rules.
From Fig.5a it is very clear that stator current is almost rectangular in nature.
From Fig. 5c it is very clear that percentage average speed error is less than 1% and
the speed follow the command as fast as possible. Moreover the step change from
200V to 150 V at t = 1 sec produce a change in speed within 0.25 s.

4 Conclusion

A Fuzzy Logic Controller based SAZZ converter for BLDC motor has been designed
in this paper. This technique has been implemented as it proves to be more efficient
than any other method. By employing the Fuzzy Logic Controller (FLC), the varying
outputs can be brought in control to get a constant output. Here two sets of input
blocks have been used where several voltage variations have been obtained on the
input side in an FLC. Thus the simulation and its outputs have shown that even
though voltage variations occur to a large extent, the output tends to remain the same
throughout the whole process.
958 A. Senthilnathan et al.

Ia (A) a) Armature current

Time in sec

b) Back EMF
Ea

Time in sec

c) Speed
Speed (rpm)

Time in Sec

Fig. 5 Performance of Fuzzy logic controller based SAZZ Converter fed BLDC motor

References

1. Al Nabulsi A, Dhaouadi R (2012) Efficiency optimization of a DSP-based standalone PV system

using fuzzy logic and dual-MPPT control. IEEE Trans Ind Inf 8(3)
2. Alajmi BN, Ahmed KH, Finney SJ, Williams BW (2013) A maximum power point tracking
technique for partially shaded photovoltaic systems in Microgrids. IEEE Trans Ind Elect 60(4)
3. Chen B-Y, Lai Y-S (2012) New digital-controlled technique for battery charger with constant
current and voltage control without current feedback. IEEE Trans Ind Electron 59(3)
4. Camara MB, Gualous H, Gustin F, Berthon A, Dakyo B (2010) DC/DC converter design for super
capacitor and battery power management in hybrid vehicle applications—polynomial control
strategy. IEEE Trans Ind Electron 57(2)
SAZZ Converter Fed Fuzzy Logic Speed Controlled BLDC Motor Drive 959

5. Vasallo Vázquez MJ, Andújar Márquez JM, Segura Manzano F (2008) A methodology for
optimizing stand-alone PV-system size using parallel-connected DC/DC converters. IEEE Trans
Ind Electron 55(7)
6. Senthilnathan, Palanivel P (2020) A new approach for commutation torque ripple reduction of
FPGA based brushless DC motor with outgoing phase current control. Microprocess Microsyst
75
7. Veerachary M, Senjyu T, Uezato K (2003) Neural-network-based maximum-power-point
tracking of coupled-inductor interleaved-boost-converter-supplied PV system using fuzzy
controller. IEEE Trans Ind Electron 50(4)
8. Muyeen SM, Al-Durra A (2013) Modeling and control strategies of fuzzy logic controlled
inverter system for grid interconnected variable speed wind generator. IEEE Syst J 7, No. 4,
Dec. 2013.
9. Mutoh N, Ohno M, Inoue T (2006) A method for MPPT control while searching for parameters
corresponding to weather conditions for PV generation systems. IEEE Trans Ind Electron 53(4)
Pedestrian Re-identification in Video
Surveillance System with Improved
Feature Extraction

Sina Salehian, Patrick Sebastian, and Abu Bakar Sayuti

Abstract In this work, we present a comparison between using different pedes-

trian re-identification (re-id) architectures. We have investigated the advantages of
using more complex and deeper convolutional neural networks (CNNs) at the feature
extraction stage. The re-id network is based on the summary network presented by
(Ahmed and Marks 2015) which we have modified and enhanced. The comparison is
done by replacing the feature extraction portion of the network. The newer improved
models performed better than the baseline model and resulted in an accuracy of
above 96% on our dataset and an accuracy of 92.09% on CUHK03 test dataset. The
network takes 2 images as input and, outputs a confidence level indicating whether or
not the 2 images depict the same person. The 2 images both go through a CNN with
shared weights and the resulting 2 feature maps are used to compare and classify the
2 images as a positive or a negative match.

Keywords Re-id · Pedestrian · Re-identification · CNN · Deep convolutional

1 Introduction

1.1 A Subsection Sample

Research in the area of full body pedestrian Re-identification (re-id) has been thriving
in recent years. Pedestrian re-id aims at identifying full body shots of people in
non-overlapping camera views [1–7]. There are a number of challenges involved in
identifying people in a new location from a different angle, such as drastic appearance,
lighting and image quality changes. These types of variations can make the task of

S. Salehian (B) · P. Sebastian · A. B. Sayuti

Universiti Teknologi PETRONAS, 32610 Seri Iskandar, Perak, Malaysia
P. Sebastian
e-mail: [email protected]
A. B. Sayuti
e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 961
R. Ibrahim et al. (eds.), International Conference on Artificial Intelligence
for Smart Community, Lecture Notes in Electrical Engineering 758,
https://doi.org/10.1007/978-981-16-2183-3_91
962 S. Salehian et al.

person re-id seem quite challenging. The main goal of any re-id system is to identify
a subject based on a few sample images that it has seen of that subject before, these
images could be from a different camera at a different location. Pedestrians re-id is
of great importance for applications such as surveillance, security systems, activity
analysis and tracking. Some of the other non-obvious challenges are: pose, occlusion,
activity, angle and background variations in two images from two different cameras
[8–12].
In this work, we are investigating the video-based re-id, by implementing the state-
of-the-art image-based re-id techniques. Image based re-id is a more challenging task
compare to video-based re-id due to the lack of information in the very few images
available (in some cases only a single image) for each identity and also the lack of
spatial and temporal information. However, an image-based re-id model would be
more practical in a scenario with a small database of each ID. Moreover, the public
datasets that are typically used by the image-based re-id models are much more
diverse in terms of pose variation, lighting condition and angle variation, as compare
to the datasets used by the video-based re-id models. Thus, alleviating biases in our
model. Hence, our motivation for applying image-based re-id to perform re-id on a
video [1–3, 13, 14].
The first step in re-id is to create descriptors that characterize the appearance of a
person based on color and textural information of the person’s clothes. The descriptor
has to have high discriminative power in order to distinguish different people, but at
the same time be able to handle variations in pose, angle, lighting and etc. [15].
Typically, the images have large variation of poses and are low quality; conse-
quently, it has been proven by both, machine learning methods and also the conven-
tional image processing methods that color and texture are the two most important
features for the task of person re-id [15] and [16].
Recent work in the area of pedestrian re-id has been mainly on either creating
discriminative features that are view, lighting, pose and background in variance [4, 6,
8, 11] or producing a distance metric between two positive or negative pairs of images
[5, 9, 10]. In some cases, both of these methods are being implemented together, to
achieve a better performance [1–3, 7, 12, 17]. Distance metric learning has proven
to outperform discriminative feature learning in most cases. Hence, the focus of this
work is on model utilizing both, distance metric and discriminative feature learning.
Using a single dataset that is bias to a particular pose, feature, background or
appearance would not generalize well to real world scenarios. Thus, it is important
to use a well-mixed batch of training data, in order to avoid bias towards a particular
category. For example, most of the pedestrian re-id datasets suffer from this problem
because either they are captured in an environment with a particular group of people
such as CUHK1 [18] and CUHK3 [19] in which most the subjects are university
students that are carrying backpacks or contains constant background information
like crosswalks in PRID [20] dataset. VIPeR [21] dataset is a good example of cases
where there is significant difference in resolution between the images. One solution
is to use a combination of these datasets to create a more diverse dataset which in
turn would help the model learn more robust features [22] the other solution is to use
better and deeper models.
Pedestrian Re-identification in Video Surveillance … 963

2 Related Work

Currently There are 2 main categories of pedestrian re-id, machine learning

approaches and traditional image processing. The traditional approaches are usually
carried out in two stages. Finding hand crafted features such as color histogram
[23], local maximal occurrence representation (LOMO) [24] and hierarchal Gaus-
sian descriptor (GOG) [25]. The handcrafted features are chosen in a way to contain
as much discriminative features as possible. The second stage is similarity metric
learning which has to determine if two images are a match or not. A number of
metric learning methods have been proposed over the past few years [26–29]. The
metric learning stage indicates how similar a pair of images are based on the features
extracted from the first stage. It is expected that two images of the same person
should have a smaller distance between them as compare to images from different
people. The problem with such systems is that the two stages cannot be optimized
simultaneously resulting in a sub-optimal system.
The deep learning based approaches have seen a lot of attention in recent years
due to their success in various computer vision tasks. Recently, the power of deep
convolutional networks in various tasks, including pedestrian re-id, has been demon-
strated. In [19] a Siamese network was proposed that takes two images as inputs and
compares them to one another. The visual features are extract using convolutional
layers and the result will indicate the similarity score. Most common networks can
be broken down into 2 main sub-network: one for feature extraction and another one
for matching. The feature extraction sub-network can be either a shallow network
with one or two convolutional and pooling layers or a deep network such as VGG,
AlexNet and etc. The deeper network have to first be retrained over large data sets
such as ImageNet dataset and then fine-tuned for person re-id. The matching sub-
network has to select the features that minimize the misalignment between the learnt
features and the ground truth. In short, the matching sub-network is at least one of
the loss functions that the system has to minimize. Pairwise loss [30, 31] and triplet
loss and its variations are the most commonly used loss functions in computer vision
tasks.
In our work the feature extraction stage and the metrics learning stage are
trained end to end simultaneously, which results in better optimization and higher
performance.

3 Methodology

This work is aimed to be used in a distributed surveillance camera system with

non-overlapping camera views. In a typical surveillance system there are multiple
cameras at different locations and backgrounds. This makes it impossible to use
simple tracking algorithms to follow each person around. Hence the need for a
pedestrian re-id network. The first step is to select/auto select a person in one of the
964 S. Salehian et al.

cameras the subject is then detected and tracked throughout the field of view of that
camera. The cropped images generated by the detector are extracted, labeled and
stored in a database.
The pedestrian detector estimates the ROI (region of interest), which would be
a bounding box around the area with a person in a given frame. The tracker keeps
track of the identity of each box in consecutive frames in a video [32]. There are now
multiple images from subject in our database.
At this stage the database contains images of the subject, these images can now
be used to identify that subject in different cameras and locations. The new images
are passed to the re-id network to be compared against the database to be identified.
As shown in Fig. 1, the two images are resized appropriately to the input size of
the network, tabulated in Table 1. Each image is then fed to a CNN, which would
result in a feature map. The two feature maps are then stacked on top of one another
to form a cross-neighborhood difference layer based on the method described in
the original paper [1]. The result is then fed into a fully connected layer to make a
final classification. A positive match indicates that the two input images are from a
same individual and a negative match indicates that they are not depicting the same
person. The final stage is a SoftMax classification which would result in a probability
between 0–1 for each class.
In this paper we are going to focus on the re-id block which consists of 2 parts.
• Feature extraction stage
• Difference learning stage

Fig. 1 Network block diagram

Table 1 Network parameters

Network Baseline AlexNet DenseNet InceptionV3
Batch size 150 150 12 12
Input size 160 × 60 227 × 227 160 × 160 160 × 160
CNN layers 2 5 121 126
CNN output feature depth 25 256 1024 2048
Total variables 2,322,192 4,180,052 8,430,356 24,400,636
Total size (bytes) 18 MB 32 MB 65 MB 186 MB
Pedestrian Re-identification in Video Surveillance … 965

We have implemented the difference learning stage similar to the work done
by [1], by computing cross-input neighborhood differences from the output of the
feature extraction stage, which is then fed to a patch summary feature layer which
summarizes the neighborhood difference maps by producing a holistic representation
of the differences in each block. It is then followed by another convolutional layer
in order to incorporate the spatial information into the model. This output is then
connected to a fully connected layer, to perform the final classification of whether
or not the 2 images belong to the same ID.
In this work we are going to mainly focus on the feature extraction stage, which
in the case of our baseline is just 2 convolutional layers. We have experimented with
3 different convolutional neural networks (CNNs). Some of the main parameters
of these networks are sown in Table 1. With the goal of achieving a much higher
accuracy of re-id by learning a deeper feature map from the input.
A pair of images are fed through the network, they each go through the same CNN
network and a feature map is extracted from each image. These 2 feature maps are
then passed to the difference learning stage. We have replaced the CNN used in the
baseline with AlexNet, DenseNet121 and InceptionV3. The networks have to first
be truncated because the difference learning stage expects a feature map and all of
the networks mentioned above are designed as classifiers with fully connected layers
at the end. In all 3 cases the output was taken from the last layer before the fully
connected layers (no top). In the case of AlexNet the original network has a max
pooling layer before the fully connected layers in order to reduce the dimensions of
the convolutional feature map so for AlexNet the output was taken from the max
pooling layer. The network with no top is then connected to the rest of the summary
network. The dimension of the feature map is different for each of the networks refer
to Table 1 for input, feature map dimensions.
The database of new IDs that are saved during registration contains a 50 image for
each subject. When a new image is being evaluated it has to be checked against all the
images in the database. Every comparison result in a probability score which is then
used to calculate an average for the 50 images that belong to each ID. The ID with
the highest average probability score is chosen as the final prediction. The accuracy
score is calculated based on the rate of correctly identified images in a batch.
In CUHK03, the dataset that is used for training, there are 13,164 images of
1,360 pedestrians. Therefore, the number of negative pairs is much greater than the
positive pairs. If the model is trained on this dataset, it will just learn to identify every
image as a mismatch which is an example of a data imbalance and overfitting. We
performed a 2D translation on the dataset to produce 5 images from every image,
based on [1]. However, the dataset is still imbalance and the negative pairs have to be
down sampled to just twice the number of positive pairs. We randomly divide 1360
identities into non-overlapping train (1160), test (100), and validation (100) sets.
966 S. Salehian et al.

4 Results

The performance of these models was evaluated on the test data from the training
dataset and also our own dataset that we created. Our dataset is divided into 2 Batches
as follows: In batch number 1, all the videos have the same background with two IDs
moving from side to side, recorded from different distances of 4, 6 and 8 m. There
are also different conditions such as single subjects moving from left to right and
vice versa, complete occlusion of either subjects in the middle of the frame by the
other subject and subjects following each other. For batch number 1 all the videos
are recorded 1280 × 720 pixels resolution. In batch number 2, there are multiple
backgrounds and for each background there are multiple videos of two subjects
(different from batch number 1) walking from side to side at different distance (3,
5 or 7 m). However, there are no special conditions. All the videos in this batch are
recorded with a resolution of 320 × 240 pixels. Figure 2 depicts some sample images
from our dataset.
In order to select the best Re-id dataset for the purpose of training, the baseline
model was trained and on 3 different datasets and tested on our own dataset. The
result from Table 2 shows that Market1501 performed well on ID0-2 but it had a
very bad performance on ID3. CUHK03 had an overall higher accuracy as compare

Fig. 2 Example of image from our dataset

Pedestrian Re-identification in Video Surveillance … 967

Table 2 Training result on

ID Number of images Dataset Accuracy (%)
multiple datasets
0 925 CUHK01 100
CUHK03 100
Market1501 96.61
1 837 CUHK01 99.64
CUHK03 93.44
Market1501 100
2 1105 CUHK01 80.68
CUHK03 95.55
Market1501 99.05
3 1309 CUHK01 90.15
CUHK03 90.12
Market1501 54.03

to CUHK01 hence, it was chosen as the primary training dataset. Choosing a diverse
dataset is an important factor in order to reduce unwanted biases, which could lead
to an underperforming re-id system.
Figure 3 also shows the accuracy for each ID which is a measure of how many
times the model correctly identified that ID for the given number of images.
Finding a perfect dataset with no bias is nearly impossible. Looking at the test
result from Fig. 3 we can observe that selecting a better model can alleviate some
of these biases. Ideally, we expect that the accuracy of the re-id be closer to a single
value. However, we can see that in the case of the baseline and DenseNet121 model
there is a 12% difference between the highest and lowest values. Whereas in AlexNet
this difference is only 6%.
The second important criteria in evaluating an identification system is space
between each user. In short, the difference in prediction probability of the model for a

Fig. 3 Result of training on 102

multiple networks 100
98
96
94
92
90
88
86
0 1 2 3
Baseline AlexNet
DenseNet121 IncepƟonV3
968 S. Salehian et al.

Table 3 Probabilities of different networks for each ID

ID Network Prob ID_0 Prob ID_1 Prob ID_2 Prob ID_3
0 Baseline 96.50 39.15 3.49 0.61
AlexNet 95.76 22.42 1.833 0.01
DenseNet121 99.06 14.84 1.08 0.01
InceptionV3 92.65 5.35 0.7 0.02
1 Baseline 25.17 95.44 62.87 52.97
AlexNet 17.8 83.29 4.87 1.41
DenseNet121 17.83 87.99 13.39 2.43
InceptionV3 6.14 97.17 36.37 19.7
2 Baseline 3.46 52.11 98.54 95.87
AlexNet 0.08 1.84 89.51 73.86
DenseNet121 0.63 7.31 99.52 82.64
InceptionV3 0.22 19.21 94.01 84.96
3 Baseline 0.65 32.97 86.50 93.34
AlexNet 0.00 1.70 70.48 93.38
DenseNet121 0.13 2.52 65.17 97.49
InceptionV3 0.04 18.06 70.40 89.44

match and non-match should be maximized. Table 3. Shows the average probability
of each ID compared to all the IDs in the database.
For example, for ID 0 the average probability of all those images when they
pass through the baseline model is 96.50% meanwhile, the second highest average
probability for the same images is for ID_1 at 39.15% which means that the network
is very confident in distinguishing between images of ID 0 and all the other 3 IDs
in the database. However, the same cannot be said about ID2 where the average
probability of ID_2 and ID_3 are very close 98.54% and 95.87% respectively. This
is mainly caused by the bias in CUHK03 dataset that different subjects wear clothes
with different colors and ID 2 and 3 have similar color clothes which increases the
chance of misclassification. By optimizing the models, we can see that in all 3 cases
AlexNet, DenseNet121 and InceptionV3 the differentiation between the IDs is more
pronounced.
The probability column shows the average probability of those images belonging
to each of the 4 IDs. Ideally the correct ID should have a probability of 100% and the
rest be 0%. Therefore, a larger difference between these probabilities would ensure
a lower chance of misclassification.
Typically, the datsets are recorded either in a university, a shopping mall or in the
street and the subjects are wearing various color clothes and they appear in different
angles and distance. This is good for training but in order to investigate what are
the shortcomings of the system an engineered dataset like the one used in this paper
helps with analyzing each parameter separately.
Pedestrian Re-identification in Video Surveillance … 969

Fig. 4 Result of Training on 101

Multiple Networks at 98
Different Distances
95
92
89
86
83
80
77
8m 4m 8m 4m 7m 5m 7m 5m
ID0 ID0 ID1 ID1 ID2 ID2 ID3 ID3

Baseline AlexNet
DenseNet121 IncepƟonV3

All the networks are trained on CUHK03 and the test accuracies are listed in Table
7. DenseNet121 performed the best with an accuracy of 92.09 as compare to the 88.7
which is the highest accuracy reported in the literature.
Table 6. AlexNet has the highest accuracy with InceptionV3 and baseline coming
at second and third place respectively and DenseNet121 had the lowest accuracy on
our dataset.
Figure 4 shows the accuracy of re-id of different models at different distances. Our
dataset is recorded at different distances which enables us to investigate the effects of
distance on performance of re-id networks. The database of existing IDs that every
new image is compared against is consist of images from ID0 at 8 m and right side
towards the camera, ID1 at 8 m and left side towards the camera, ID2 at 5 m and
right side towards the camera and ID3 at 5 m and right side towards the camera
As evident by Fig. 4 and Table 4 for ID 2 and 3 the accuracy is higher at 5 m
as compare to 7 m. This is mainly due to the lower quality of the images which
causes the far objects to be a lot blurrier than the images at 5 m. However, for ID2
the difference between the average probabilities are much closer at 5 m than they are
at 7 m. This is true for all of the models which means the models are struggling to
distinguish between images of ID2 and ID3 at 5 m. The main reason for this unusual
behavior is the fact that the images of ID 2 and 3 at 5 m have a lot of more similar
features which results in a higher probability on both IDs.
The accuracy of the re-id for different point of view where in one case the camera
sees the right side of the person and another that sees the left are shown in Fig. 5.
As mentioned earlier the database contains only images from one side therefore this
test can demonstrate the importance of having images from different angle in the
database.
Looking at Fig. 5 we can see a pattern where the accuracy is either higher or similar
when evaluating images with the same view. We can also observe that InceptionV3
970 S. Salehian et al.

Table 4 Accuracy of re-id for different architectures at different distances

ID Range (m) Images Network 2nd Highest Prob Highest Prob
0 8 447 Baseline 12.37 93.99
AlexNet 22.52 86.34
DenseNet121 12.02 98.96
InceptionV3 3.33 91.96
4 451 Baseline 12.05 93.68
AlexNet 19.47 94.75
DenseNet121 13.79 99.07
InceptionV3 3.49 94.31
1 8 447 Baseline 23.93 91.74
AlexNet 12.22 82.86
DenseNet121 12.72 88.79
InceptionV3 33.41 97.56
4 451 Baseline 34.61 85.00
AlexNet 23.55 83.56
DenseNet121 24.1 86.73
InceptionV3 30.86 96.4
2 7 753 Baseline 84.3 95.89
AlexNet 62.1 84.98
DenseNet121 76.78 99.44
InceptionV3 80.34 91.81
5 352 Baseline 92.59 97.07
AlexNet 90.31 95.84
DenseNet121 90.85 99.63
InceptionV3 91.43 97.07
3 7 886 Baseline 63.93 82.86
AlexNet 62.33 89.57
DenseNet121 66.17 96.73
InceptionV3 67.72 87.57
5 423 Baseline 59.32 85.48
AlexNet 81.89 98.72
DenseNet121 63.77 98.57
InceptionV3 74.16 92.05
Pedestrian Re-identification in Video Surveillance … 971

Fig. 5 Result of training on 101

multiple networks at 99
different angles 97
95
93
91
89
87
85
right le right le right le right le
ID0 ID0 ID1 ID1 ID2 ID2 ID3 ID3

Baseline AlexNet
DenseNet121 IncepƟonV3

has a more uniform characteristic which results a similar accuracy for each ID,
regardless of the side.
Table 5 shows that the difference between the average probabilities are above
10% for AlexNet, DenseNet121 and InceptionV3 which is an improvement over the
7% of the baseline model. The models are capable of handling variations in viewing
direction which is evident by both the accuracy result and the difference in average
probability values (Tables 6 and 7).
As we mentioned earlier, our dataset is consisting of 2 batches that are quite
different in a few ways. Table 6 shows that all the 4 networks perform relatively
good on batch number 1 with 3 of them having an accuracy of above 97% and only
DenseNet121 is lower at around 95.6%. This could be either because batch 1 uses a
similar background for all the images or because the 2 IDs are wearing distinctive
clothing or just the fact that batch 1 has images with higher quality. As for batch 2,
AlexNet performed the best with an accuracy of almost 95% which is only 3% lower
than its batch 1 accuracy. InceptioV3 came second, with an accuracy around 91% in
this case 8% lower than its batch 1 accuracy. The baseline and DenseNet121 had an
accuracy around 90, with their accuracies dropping by 7 and 6% compare to batch 1
respectively. Because AlexNet had a high performance on both batches it has a much
higher overall accuracy than the other networks.
In terms of difference in average probability all three networks outperformed the
baseline based on values in Table 3.
As shown in Table 7 all three networks AlexNet, DenseNet121 and Incep-
tionV3 have outperform previous networks in the literature. This is mainly because
these designs are much deeper and more computationally expensive than the other
networks. Which is a tradeoff between speed and performance. DenseNet + SumNet
achieved the highest accuracy with AlexNet + SumNet in the second place and
InceptionV3 + SumNet in third.
Looking at Figs. 3, 4 and 5 we can see a more uniform pattern in results from
AlexNet and InceptionV3 where the results from different IDs are closer to one
another.
972 S. Salehian et al.

Table 5 Accuracy of re-id for different architectures at different angles

ID Side Images Network 2nd Highest prob Highest Prob
0 Right 414 Baseline 13.17 93.86
AlexNet 25.85 95.87
DenseNet121 12.70 98.96
InceptionV3 4.75 92.85
Left 512 Baseline 12.78 93.80
AlexNet 19.48 95.66
DenseNet121 16.66 99.14
InceptionV3 5.87 92.47
1 Right 360 Baseline 30.37 89.38
AlexNet 20.07 82.32
DenseNet121 20.36 84.38
InceptionV3 40.46 95.77
Left 477 Baseline 27.29 89.59
AlexNet 15.86 84.12
DenseNet121 15.67 91.09
InceptionV3 32.86 98.37
2 Right 539 Baseline 88.80 96.32
AlexNet 75.12 90.27
DenseNet121 81.32 99.45
InceptionV3 86.62 93.87
left 566 Baseline 86.70 96.45
AlexNet 72.60 88.75
DenseNet121 83.96 99.59
InceptionV3 83.31 94.15
3 Right 646 Baseline 59.41 80.76
AlexNet 67.84 93.80
DenseNet121 61.63 96.43
InceptionV3 69.26 88.18
Left 663 Baseline 80.17 87.14
AlexNet 73.12 92.97
DenseNet121 68.72 98.56
InceptionV3 71.54 90.69
Pedestrian Re-identification in Video Surveillance … 973

Table 6 Overall accuracy of

Batch Images Network Accuracy (%)
re-id
1 1762 Baseline 97.54
AlexNet 97.26
DenseNet121 95.67
InceptionV3 99.88
2 2414 Baseline 90.21
AlexNet 94.94
DenseNet121 89.84
InceptionV3 91.74
Overall 4176 Baseline 93.88
AlexNet 96.1
DenseNet121 92.76
InceptionV3 95.81

Table 7 Overall
Method Rank-1 (%)
accuracy comparison with
other re-id methods IDLA [1] 54.74
CAN [3] 77.6
MLS [4] 87.50
PAR [5] 85.4
MSDL [6] 76.87
DCAF [8] 74.21
PersonNet [10] 64.80
LDNS [12] 62.55
Gated Siamese [14] 68.1
PDC [33] 88.7
DPFL [34] 86.70
DeepAlign [35] 85.40
Spindle [36] 88.50
JLML [37] 83.20
AlexNet + SumNet 91.89
DenseNet121 + SumNet 92.09
InceptionV3 + SumNet 90.18

The networks were trained using the following parameters, softmax cross-entropy
loss function, Momentum Optimizer, learning rate 0.001 with a batch size of 12
for DenseNet121 and Inception, learning rate of 0.01 with a batch size of 150 for
AlexNet. The networks were all pretrained on ImageNet which helps them converge
faster. AlexNet was trained end to end for 100,000 iteration but for DenseNet121
974 S. Salehian et al.

and Inception only the top of the network was trained for the first 100,000 iterations
by freezing the weights in the CNN portion of the network, it was then trained end
to end for an additional 100,000 iterations.

5 Conclusion

We have investigated and demonstrated the advantages of using different datasets at

the training stage. We have also looked how a deeper network with a deeper feature
depth can improve the performance of pedestrian re-id systems.
Based on the result we can also conclude that a deeper network can pick up on
more robust and discriminative features. The deeper networks can also help reduce
the effects of biases in the data that they were trained on. This is evidence based
on the improved accuracy and increased distance between the average probability
values in deeper models such as AlexNet, DenseNet121 and InceptionV3.
Evaluating the network on an engineered dataset, similar to the one used in this
paper, enables us to compare and test cases that do not exist on typical publicly
available datasets.
During training the baseline and AlexNet models converged within 100,000 iter-
ations however, the DenseNet121 and IceptionV3 models took much longer (more
than 150,000 iterations) to converge which is expected, since they are much deeper
networks.
It is important to note that the training batch size for the DenseNet121 and Incep-
tionV3 models had to be reduced in order to reduce the overall size so that it would
fit on a GPU with 11 GB of memory.

References

1. Ahmed E, Jones M, Marks T (2015) An improved deep learning architecture for person re-
identification. In: 2015 IEEE conference on computer vision and pattern recognition (CVPR),
Available: https://doi.org/10.1109/cvpr.2015.7299016
2. Lukezic A, Vojir T, Zajc LC, Matas J, Kristan M (2017) Discriminative correlation filter
with channel and spatial reliability. In: 2017 IEEE conference on computer vision and pattern
recognition (CVPR)
3. Liu H, Feng J, Qi M, Jiang J, Yan S (Jul. 2017) End-to-end comparative attention networks for
person re-identification. IEEE Trans Image Process 26(7):3492–3506
4. Guo Y, Cheung N-M (2018) Efficient and deep person re-identification using multi-level
similarity. 2018 IEEE/CVF conference on computer vision and pattern recognition
5. Zhao L, Li X, Zhuang Y, Wang J (201) Deeply-learned part-aligned representations for person
re-identification. In: 2017 IEEE international conference on computer vision (ICCV)
6. Qian X, Fu Y, Jiang Y-G, Xiang T, Xue X (2017) Multi- scale deep learning architectures for
person re-identification. In: The IEEE international conference on computer vision (ICCV)
7. Chen W, Chen X, Zhang J, Huang K (2017) Beyond triplet loss: a deep quadruplet network for
person re-identification. In: The IEEE conference on computer vision and pattern recognition
(CVPR
Pedestrian Re-identification in Video Surveillance … 975

8. Li D, Chen X, Zhang Z, Huang K (2017) Learning deep context-aware features over body and
latent parts for person re-identification. In: The IEEE conference on computer Vi- sion and
pattern recognition (CVPR)
9. Chen W, Chen X, Zhang J, Huang K (2017) A multi-task deep network for person
re-identification. In: AAAI conference on artificial intelligence (AAAI)
10. Wu L, Shen C, van den Hengel A (2016) PersonNet: person re-identification with deep
convolutional neural networks. Comput Vis Pattern Recogn
11. Matsukawa T, Okabe T, Suzuki E, Sato Y (2016) Hierarchical gaussian descriptor for person
re-identification. In: Proceedings of the IEEE conference on computer vision and pattern
recognition, pp 1363–1372
12. Zhang L, Xiang T, Gong S (2016) Learning a discriminative null space for person re-
identification. CVPR
13. Xiao T, Li H, Ouyang W, Wang X (2016) Learning deep feature representations with domain
guided dropout for person re-identification. In CVPR
14. Varior RR, Haloi M, Wang G (2016) Gated siamese convolutional neural network architecture
for human re- identification. In: European conference on computer vision (ECCV)
15. Xiong F, Gou M, Camps O, Sznaier M (2014) Person re-identification using kernel-based
metric learning methods. Comput Vis ECCV 1–16
16. Liao S, Hu Y, Zhu X, Li S (2015) Person re-identification by local maximal occurrence represen-
tation and metric learning. In: 2015 IEEE conference on computer vision and pattern recognition
(CVPR)
17. Li Z, Chang S, Liang F, Huang T, Cao L, Smith J (2013) Learning locally-adaptive decision
functions for person verification. In: 2013 IEEE conference on computer vision and pattern
recognition
18. Guillaumin M, Verbeek J, Schmid C (2009) Is that you? Metric learning approaches for face
identification. In: 2009 IEEE 12th international conference on computer vision
19. Chen D, Yuan Z, Hua G, Zheng N, Wang J (2015) Similarity learning on an explicit polynomial
kernel feature map for person re-identification. In: 2015 IEEE conference on computer vision
and pattern recognition (CVPR)
20. Chen J, Zhang Z, Wang Y (2014) Relevance metric learning for person re-identification by
exploiting global similarities. In: 22nd international conference on pattern recognition
21. Li W, Zhao R, Xiao T, Wang X (2014) DeepReID: deep filter pairing neural network for person
re-identification. In: CVPR
22. Lukežič A, Vojíř T, Čehovin Zajc L, Matas J, Kristan M (2018) Discriminative correlation filter
tracker with channel and spatial reliability. Int J Comput Vis 126(7):671–688
23. Xiong F, Gou M, Camps O, Sznaier M (2014) Person re-identification using kernel-based metric
learning methods. Comput Vis ECCV, pp 1–16, 2014. Available: https://doi.org/10.1007/978-
3-319-10584-0_1
24. Liao S, Hu Y, Zhu X, Li SZ (2015) Person re-identification by Local Maximal Occurrence
representation and metric learning. In: 2015 IEEE conference on computer vision and pattern
recognition (CVPR), Boston, MA, pp 2197–2206. https://doi.org/10.1109/CVPR.2015.729
8832
25. Matsukawa T, Okabe T, Suzuki E, Sato Y (2020) Hierarchical gaussian descriptors with applica-
tion to person re-identification. IEEE Trans Pattern Anal Mach Intell 42(9):2179–2194. https://
doi.org/10.1109/TPAMI.2019.2914686
26. Li Z, Chang S, Liang F, Huang TS, Cao L, Smith JR (2013)Learning locally-adaptive decision
functions for person verification. In: 2013 IEEE conference on computer vision and pattern
recognition, Portland, OR, pp 3610–3617. https://doi.org/10.1109/CVPR.2013.463
27. Guillaumin M, Verbeek J, Schmid C (2009)Is that you? metric learning approaches for face
identification. In: 2009 IEEE 12th international conference on computer vision, Kyoto, pp
498–505. https://doi.org/10.1109/ICCV.2009.5459197
28. Chen D, Yuan Z, Hua G, Zheng N, Wang J (2015) Similarity learning on an explicit polynomial
kernel feature map for person re-identification. In: 2015 IEEE conference on computer vision
and pattern recognition (CVPR), Boston, MA, pp 1565–1573. https://doi.org/10.1109/CVPR.
2015.7298764
976 S. Salehian et al.

29. Chen J, Zhang Z, Wang Y (Dec. 2015) Relevance metric learning for person re-identification
by exploiting listwise similarities. IEEE Trans Image Process 24(12):4741–4755. https://doi.
org/10.1109/TIP.2015.2466117
30. Shi H et al (2016)Embedding deep metric for person re-identication a study against large
variations. Available: https://arxiv.org/abs/1611.00137v1
31. Varior RR, Shuai B, Lu J, Xu D, Wang G (2016) A siamese long short-term memory architecture
for human re-identification. In: Leibe B, Matas J, Sebe N, Welling M (eds) Computer vision—
ECCV 2016. ECCV 2016. Lecture Notes in Computer Science, vol 9911. Springer, Cham.
https://doi.org/10.1007/978-3-319-46478-7_9
32. Salehian S, Sebastian P, Sayuti AB (2019)Framework for pedestrian detection, tracking and
re-identification in video surveillance system. In: 2019 IEEE international conference on signal
and image processing applications (ICSIPA), Kuala Lumpur, Malaysia, pp 192–197
33. Su C, Li J, Zhang S, Xing J, Gao W, Tian Q (2017) Pose-driven deep convolutional model for
person re-identification. In: 2017 IEEE international conference on computer vision (ICCV),
Venice, pp 3980–3989. https://doi.org/10.1109/ICCV.2017.427
34. Chen Y, Zhu X, Gong S (2017) Person re-identification by deep learning multi-scale repre-
sentations. In: 2017 IEEE international conference on computer vision workshops (ICCVW),
Venice, 2017, pp 2590–2600, https://doi.org/10.1109/ICCVW.2017.304
35. Zhao L, Li X, Wang J, Zhuang Y (2017) Deeply-learned part-aligned representations for person
re-Identification. In: The IEEE international conference on computer vision (ICCV), pp.3219–
3228
36. Zhao H, et al (2017) Spindle net: person re-identification with human body region guided
feature decomposition and fusion. In: 2017 IEEE conference on computer vision and pattern
recognition (CVPR), Honolulu, HI, pp 907–915. https://doi.org/10.1109/CVPR.2017.103
37. Li W, Zhu X, Gong S (2017) Person re-identification by deep joint learning of multi-loss clas-
sification. In: Proceedings of the 26th international joint conference on artificial intelligence,
pp 2194–2200
Smart Children Management Using Data
Analytics, Machine Learning and IoT

Md Jobair Hossain Faruk and Muhamad Hariz Muhamad Adnan

Abstract One of the biggest challenges faced by parents is managing the difficult
or defiant behavior of children of the twenty-first century. Technology can help
to provide support in managing children’s activities and helping parents to be in
control. Recent advances in IoT have sparked the interest of researchers to develop
a smart management system for children. This paper presents a review of the smart
management of children using the internet of things (IoT), machine learning, and
Data Analytics. The objective is to identify potential approaches regarding smart
management for children based on IoT for future researchers. The reviews were
conducted from secondary sources, where conference proceedings and journals have
been analyzed from Scopus and other databases. The finding indicates that IoT-based
smart system has robust accuracy and performance. Thus, exploiting IoT for the
development of a smart management system for children will provide significant
benefits to the future generation.

Keywords Smart children management · IoT · Machine learning · Data analytics ·

Healthcare

1 Introduction

Children’s programs are now more important to parents as they are encouraged to
engage their children in as many activities as possible to help them plan for their
future [1, 2]. Many studies have described various methods used by families to handle
all the work they need to do around the overlapping duties of work, education, home,
and enrichment events [1–7]. Managing the challenging or rebellious actions of

Md J. H. Faruk
College of Computing and Software Engineering, Kennesaw State University, Marietta, USA
e-mail: [email protected]
M. H. M. Adnan (B)
Computing Department, Faculty of Art, Computing and Creative Industry, Universiti Pendidikan
Sultan Idris, Tanjung Malim, Malaysia
e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 977
R. Ibrahim et al. (eds.), International Conference on Artificial Intelligence
for Smart Community, Lecture Notes in Electrical Engineering 758,
https://doi.org/10.1007/978-981-16-2183-3_92
978 Md J. H. Faruk and M. H. M. Adnan

twenty-first century children is one of the greatest challenges faced by parents [8].
As information and communication technologies (ICT) and smart devices can be an
integral part of self-management plans, technology can help support the management
of children’s activities and help parents to be in charge [1–9].
The concept of a smart system has been proposed by many scholars where the
generic term ‘smart system’ involves several classes of systems composed of sensors,
actuators, a network, and a signal processing unit that delivers definite signals and
fulfill subtasks or complete tasks [10]. Smart systems may visualize as an integration
of the ideas of smart people, smart environments, and smart living and it expands into
not only IoT but also data analytics and machine learning which is contextualized
with an open and comprehensive network of intelligent objects that can auto-organize,
share information, data and resources while reacting and acting in face of situations
and changes in the environment [4–13].
In this paper, we present potential approaches towards an ideal smart management
system for children that shall help future researchers to get a comprehensive knowl-
edge of plenty of existing techniques. Figure 1. presents some potential approaches
for the smart management of children.
The paper is organized as follows: In the second section, potential methods of
IoT based on the smart management system for children are presented followed by
presenting the discussions. The last section provides the conclusion.

Fig. 1 Illustrations of potential techniques towards smart management of children

Smart Children Management Using Data … 979

2 Smart Management of Children Review

2.1 Internet of Things (IoT)

Bed-Activities and Vital Signs Monitoring System: Monitoring sleep and physical
activities are vital because of their uncertain moves and emerging technologies offer
new opportunities for parents to monitor children regardless of their position and
pursuit. Besides, monitoring vital signs also significant as indicators of one’s health
condition and assurance of proper circulatory, respiratory, neural and endocrinal
functions towards keep track of the children’s health condition and it refers to the
parameters of the body’s basic functions [14]. A group of researchers from the Univer-
sity of Georgia proposed vibrations based on a real-time, contactless bed-mounted
monitoring system that can estimate vital signs, body movement, falls from bed,
posture changes and on/off bed detection using during sleep time [15, 16].
One of its primary components is Raspberry Pi 3 that connects with a seis-
mometer to get real-time measurement data, maxima statistics and an instanta-
neous property-based approach towards enhancing the quality of monitoring sleep
condition and posture. Envelop based approach is integrated to estimate BPMh for
measuring heart rate (HR) [16]. System evaluation indicates its robust accuracy of
both approaches to monitoring bed activities and detecting vital signs and body move-
ments. This approach can be one of the potential components of a smart management
system because of its robust filtering techniques to estimate the heart rate (HR) and
respiration rate (RR) as well as other sleep activities.
Health Care Monitoring System: According to the National Academy of Science
[17], In addition to supporting this viewpoint, wellness during childhood sets the
stage for adult health, but also provides an important legal, social and economic
obligation to ensure that all children are as healthy as they can be. Amrita University
researchers [18] present an IoT-based smart and secure health care monitor system
for children to monitor behavioral and health issues and predict child disorders.
C4.5 algorithm has been adopted to predict the disorder of children by analyzing
collected therapeutic games and bodily parameter data. Besides. an android game
has been presented that enables the device to evaluate the disorders of the child using
visuospatial and strop test, sustained in attention-al blindness, and ADD/ADHD test.
A case study is carried out among 500 children from schools and health centers. The
result indicates its capability in predicting disorders in the child with 93.4% accuracy.
Contact-free and privacy-preserving solutions for monitoring bed activities need to
be emphasized towards the smart management of children.

2.2 Machine Learning

Activity Recognition and Diagnostic Approach: Physical activity of children recog-

nition plays a significant role and value to measure towards analyzing, predicting,
980 Md J. H. Faruk and M. H. M. Adnan

child-development monitoring, energy-expenditure estimation, child-obesity preven-

tion, child safety in and around the home, etc. [19, 20]. The activity recognition
approach is one of the important methods for the SMART management system
[21], thus a group of researchers (2020) [22] proposed a real-time and continuous
activity recognition-based knowledge-driven system with aims to provide an activity
recognition approach for smart homes (SH).
The proposed method adopted an ontology-based recognition model and semantic
subsumption reasoning algorithms towards recognition of activities for SH. To trace
the activity of daily living (ADL), the conceptual activity model has been structured
for the system. The agent-based integrated system architecture was conceived for
a real-time, continuous activity recognition system. The proposed system has been
evaluated in various scenarios and satisfactory accuracy in activity recognition has
achieved which is 94.44% while the average run time was measured 2.5 s.

2.3 Data Analytics

Emotion-Aware Big Data Framework: Recognition of feelings is an important topic

in different fields of activity in particular for futuristic SMART management systems
and it may be one of the influential fields for this area of research [23]. The ability to
recognize and understand emotions portrayed as non-verbal cues is essential to the
development of related approaches [24]. Hossain and Muhammad, [25] introduced a
5G enabled, emotion-aware connected, healthcare-based big data framework towards
classifying the data to recognize the emotions of the patient; particularly whether
they feel pain or not.
The work integrated Bluetooth technology for the positioning, local binary pattern
(LBP), and an interlaced derivative pattern (IDP) to detect the signals. The study
also adopted sensor technology towards recognizing speech and image signals of the
patient as input data separately within a smart home scenario. Input data is used to esti-
mate the final score to detect emotion where it processes the speech and signals where
center-symmetric LBP (CSLBP) has been utilized to calculate the histogram, Fig. 2.
The proposed approach has demonstrated by conducting experiments among fifty
trained university students where the system shows the highest accuracy (99.87%)
of detecting emotion.
Reading Emotional Parameters: A group of researchers [26] investigated the possi-
bility of reading the emotional parameters of humans where they carried out an exper-
imental study to identify six emotional parameters including engagement, excite-
ment, focus, stress, relaxation, and interest using brain signals. Supervised learning
classifiers, Naive Bayes, and Linear Regression were adopted to show the accuracy
and competency of well-known electroencephalography (EEG) based Emotiv Epoc+
Neuroheadset. Experimental studies indicate 69 % and 62 % improved accuracy in
reading six emotional parameters of the human subjects and such study can be a
focal point for the smart children management.
Smart Children Management Using Data … 981

Fig. 2 Big data flow

3 Discussion

A review of smart management of children based on IoT, Machine Learning and Data
Analytics has been presented where potential methods and techniques proposed by
researchers have been highlighted. The summary of the findings from the literature
review is illustrated in Table 1. Throughout the studies, we discovered the approaches
from different areas and a combination of suitable methods that may lead us towards
the development of a new smart management system for children.
Considering the challenges that the smart management applications for children
faced some common approaches have been identified. Eye-tracking data analysis,
emotion recognition, facial expression, infant activity monitoring and real-time inter-
active growth management are some of the areas that have huge potentials for the
development of a smart management system for children. Similarly, the vital signs
monitoring approach is an important technique that can be further investigated and
integrated into a smart management system to keep track of the children’s health
condition. The development of such kind of SMART systems for children may
initiate a huge impact on the people globally, especially among the people in Asia.
We conducted a preliminary survey among 22 parents aged 35–45 from Malaysia and
Bangladesh to estimate the guardian’s approaches to monitoring their children. A self-
administered questionnaire was utilized to collect the data from the parents and there
were asked to answer and complete the questionnaire online where 48.2% are female.
The study indicates that most respondents have medium knowledge of technologies,
and they are more familiar and dependent on conventional techniques. A 14% of
respondents utilize the existing monitoring approaches, for instance, CCTV cameras,
and tracking systems. However, such a tracking system was not efficient since each
component needs to monitor separately which leads the users dispassionate.
982 Md J. H. Faruk and M. H. M. Adnan

Table 1 Initial findings of selected approaches

Approaches Objectives Outcome
Bed-activities and vital signs Development of a real-time, Utilizing verities of novel
monitoring system contact-free, privacy-protective non-intrusive methods, the
sensor-based system that shall system demonstrated robust
capable to monitor not only performance in three
sleeping activities, but also environments in the principle
vital signs, and fall detection of measuring heart and
of an individual’s bedtime respiration rate, on-bed, off-bed
and sitting movements; finally,
posture changes. The accuracy
of detecting and alerting falls
from the bed, monitoring heart
and respiration rate is
significantly higher than 95%
Health care monitoring Introduces a novel, IoT based Both health care and eHealth
system healthcare approach to monitor approaches indicate their
behavioral and health issues potentials and can be assets for
and predict child disorders. future researchers.
Using an extended approach, Combinations of two novel
the system may be facilitated healthcare approaches may
by robot assistant diabetic initiate a significant impact on
treatment for children a smart management system for
children where guardians shall
be able to identify and monitor
their children’s health issues;
particularly, a few well-known
disorders. Such as ADHD,
mental health, and diabetics
Emotion-aware big data Introduces an emotion-aware The outcome of an emotion
framework connected, big data-based recognition approach is
framework to recognize the interesting particularly for its
emotions of the patient; accuracy that demonstrated
particularly whether they feel higher than 99.87%. Although
pain or not the system is evaluated among
university students; however,
such a system can be huge
potentials for smart
management of the system of
children because of the demand
and needs to ensure the
children’s health issues

4 Conclusion

In this paper, we reviewed and discussed the potentials of smart management of

children. The study shows a huge expectation of parents for the SMART system
towards monitoring their children in a way that is efficient, convenient, and that
parallels with the expectations. By utilizing IoT, a complete and robust smart child
Smart Children Management Using Data … 983

management system may be proposed. As similar to smart homes and smart cities,
not only sensors-based IoT but also data analytics and machine learning shall be the
revolutionary technique towards developing a smart management system. However,
it is expected to conduct future research to design and develop a fully functional smart
management system by interpreting the methods and approaches, and hundreds of
thousands of people shall be benefited using the futuristic system.

Acknowledgements We would like to thank Universiti Pendidikan Sultan Idris (UPSI)

Malaysia, the Ministry of Higher Education (MOHE), Malaysia (FRGS Grant No: FRGS
/1/2020/ICT02/UPSI/02/2) and Kennesaw State University, USA and for supporting this study.

References

1. Davidoff S, Lee MK, Yiu C, Zimmerman J, Dey AK (2006) Principles of smart home control.
Lect Notes Comput Sci (including Subser Lect Notes Artif Intell Lect Notes Bioinformatics).
4206 LNCS:19–34
2. Lee MK, Davidoff S, Zimmerman J, Dey AK (2007) Smart bag: managing home and raising
children. Proc 2007 Conf Des Pleasurable Prod Interfaces, DPPI’07. 434–437.
3. Lee MK, Davidof S, Zimmerman J, Dey A (2006) Smart homes, families, and controls. 2006;
Available from: https://pdfs.semanticscholar.org/b751/6530ebffe5b7202f0ac201aefc7850b87
a52.pdf?_ga=2.123838666.1183646419.1597302163-2076110809.1586498474
4. Yang H, Lee W, Lee H. IoT smart home adoption: the importance of proper level automation.
J Sens (2018)
5. Taylor AS, Swan L (2005) Artful systems in the home. CHI 2005 Technol Safety, Community
Conf Proc - Conf Hum Factors Comput Syst 641–650
6. Tolmie P, Pycock J, Diggins T, MacLean A, Karsenty A (2002) Unremarkable computing. Conf
Hum Factors Comput Syst - Proc 4(1):399–406
7. Darrah CN, English-Lueck JA, Freeman JM (2000) Living in the eye of the storm: controlling
the maelstrom in silicon valley, 1–10
8. Childmind. Managing problem behavior at home: a guide to more confident, consistent and
effective parenting. Child Mind Inst Inst [Internet]. Available from: https://childmind.org/art
icle/managing-problem-behavior-at-home/
9. Sendra S, Parra L, Lloret J, Tomás J (2018) Smart system for children’s chronic illness
monitoring. Inf Fusion. 40:76–86
10. Kanoun O, Universit T, View E, Kanoun O (2014) Smart systems and devices : innovative key
modules for engineering applications
11. Lanting C, Lionetto A (2015) Smart systems and cyber physical systems paradigms in an IoT
and industrie/y4.0 context. S5002
12. Madakam S, Ramaswamy R, Tripathi S (2015) Internet of things (IoT): a literature review. J
Comput Commun 03(05):164–173
13. Apanaviciene R, Vanagas A, Fokaides PA (2020) Smart building integration into a smart city
(SBISC): development of a new evaluation framework. Energies 13(9)
14. Teixeira CC, Boaventura RP, Souza ACS, Paranaguá TT de B, Bezerra ALQ, Bachion MM et al
(2015) Aferição de sinais vitais: Um indicador do cuidado seguro em idosos. Texto e Context
Enferm 24(4):1071–1078
15. Clemente J, Li F, Valero M, Song WZ (2019) Demo: contactless device for monitoring on-bed
activities and vital signs. Proc - 2019 IEEE Int Conf Smart Comput SMARTCOMP 2019.
472–474.
984 Md J. H. Faruk and M. H. M. Adnan

16. Clemente J, Valero M, Li F, Wang C, Song WZ (2020) Helena: real-time contact-free monitoring
of sleep activities and events around the bed. In: 18th annual IEEE international conference on
pervasive computing and communications
17. (2005) Children’s health, the nation’s wealth: assessing and improving child health. Choice
Rev Online 42(10):42–5896–42–5896
18. Binu PK, Akhil V, Mohan V (2017) Smart and secure IoT based child behaviour and health
monitoring system using hadoop. In: 2017 International conference on advances in computing,
communications and informatics (ICACCI) 2017. 2017-Janua:418–423
19. Boughorbel S, Breebaart J, Bruekers F, Flinsenberg I, Ten Kate W (2011) Child-activity
recognition from multi-sensor data. ACM Int Conf Proceeding Ser
20. Sindhu P (2018) Toddler activity recognition using machine learning. Loyola eCom-
mons [Internet]. 2018; Available from: https://ecommons.luc.edu/cgi/viewcontent.cgi?article=
4755&context=luc_theses
21. Guo J, Li Y, Hou M, Han S, Ren J (2020) Recognition of daily activities of two residents in a
smart home based on time clustering. Sensors (Switzerland) 20(5)
22. Rawashdeh M, Al MG, Samarah S, Hossain MS, Muhammad G (2020) A knowledge-driven
approach for activity recognition in smart homes based on activity profiling. Futur Gener
Comput Syst 107:924–941
23. Dzedzickis A, Kaklauskas A, Bucinskas V (2020) Human emotion recognition: review of
sensors and methods. Sensors (Switzerland) 20(3)
24. Sivasangari A, Ajitha P, Rajkumar I, Poonguzhali S (2019) Emotion recognition system for
autism disordered people. J Ambient Intell Humaniz Comput
25. Hossain MS, Muhammad G (2018) Emotion-aware connected healthcare big data towards 5G.
IEEE Internet Things J 5(4):2399–2406
26. Hossain Faruk MJ, Valero M, Shahriar H (2021) An investigation on non-invasive brain-
computer interfaces: emotiv epoc+ neuroheadset and its effectiveness. In: 2021 IEEE 45th
Annual Computers, Software, and Applications Conference (COMPSAC), pp 580–589. https://
doi.org/10.1109/COMPSAC51774.2021.00086
Optical Wireless Communication Based
Wireless Body Area Network for Remote
Patient Monitoring

Komal Masroor , Sajid Mumtaz, Micheal Drieberg , and Varun Jeoti

Abstract Internet-of-Things (IoT) based Wireless Body Area Networks (WBANs)

are likely to change the traditional healthcare infrastructure by allowing remote moni-
toring of patients outside the conventional clinical settings e.g. in homes. However,
technology remains a key enabler for such a system realization. Currently, Radio
Frequency (RF-) based Zigbee networks are being employed for data transmission
and reception, nevertheless, with the widespread use of IoT devices, the unlicensed
Industrial, Scientific and Medical (ISM) spectrum may exhaust soon. Consequently,
Visible Light Communication (VLC) technology is being considered as an alternative
to RF owing to its extremely high unregulated bandwidth and low cost. Additionally,
it is considered safe for human beings. Several studies in the literature have reported
the effects of transceiver related parameters on a VLC link, however, in this study,
we examine the impact of various parameters that are pivotal for VLC channel char-
acterization in healthcare settings. We show that reflectivity of room surfaces, angle
of irradiance of light emitiing diodes (LEDs), and on-body receiver heights have a
significant impact on the received power (RXPWR). Furthermore, we validate that
increase in number of transmitters (Txs ) reduce the fluctuations in RXPWR by 75%.

Keywords Visible light communication (VLC) · Optical body sensor network

(OBSN) · Remote patient monitoring (RPM) · Uniform receiver power

K. Masroor (B) · S. Mumtaz · M. Drieberg

Universiti Teknologi PETRONAS, Seri Iskandar, Malaysia
e-mail: [email protected]
K. Masroor
NED University of Engineering and Technology, Khi, Pakistan
V. Jeoti
Faculty of Technical Sciences, University of Novi Sad, Novi Sad, Serbia

© The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 985
R. Ibrahim et al. (eds.), International Conference on Artificial Intelligence
for Smart Community, Lecture Notes in Electrical Engineering 758,
https://doi.org/10.1007/978-981-16-2183-3_93
986 K. Masroor et al.

1 Introduction

Visible Light Communication (VLC) technology is becoming increasingly popular

as an alternative to Radio Frequency (RF-) based communication in many applica-
tions. Remote Patient Monitoring (RPM) is one such application where VLC enables
health monitoring inside indoor spaces via the pre-installed lighting infrastructure
[1]. The use of VLC in RPM can be attributed to its high unregulated bandwidth, low
cost, resistance to electromagnetic interference (EMI), simplicity of the transceiver
design, and its inherent features like security and safety for human body. Gener-
ally, communication in optical body sensor network (OBSN) is possible between the
coordinator and on-body motes or between coordinator and the optical access point
(OAP). In both the scenarios, reliable transmission and reception of patient’s data is
of utmost importance.
VLC enabled healthcare monitoring is in its early stage and investigations are
being done to prove its viability for such systems. The initial study to demonstrate
the use of optical technology for patient monitoring was presented in Torkestani et al.
[2]. Following this, several studies have established the use of VLC/ infrared (IR) in
monitoring applications. For instance, in Scully et al. [3], experimental demonstration
of use of mobile phone camera to derive various physiological parameters from skin
colour has been performed. The authors in Khalid et al. [4] used bipolar signaling
and high pass filtering to avoid unwanted background ambient light interferences in
a diffused mobile IR-based WBAN system along with power savings. Line-of-sight
(LoS) is usually a hindrance when it comes to mobility; however, use of this type
of link design to offer mobility has been investigated in Torkestani et al. [5]. The
inclusion of first and second order reflections along with the major LOS component
to evaluate system performance has been examined in Alyan et al. [6]. Effects of
using LoS and diffuse configurations for patient monitoring in indoor environment
using error coding scheme have been studied in Torkestani et al. [7]. Mostly, non-
LoS (NLoS) i.e. diffuse links are employed to ensure maximum coverage indoors,
nevertheless, there are several other factors that affect the OBSN link performance
including room size, transceiver locations, number of transceivers, semi-angle at half
power (SAHP) of the LED, field-of-view (FOV) of the receiver (Rx ), reflectivity (ρ)
of room surfaces and body-shadowing. In our previous study [8], we highlighted
the importance of transmitter (Tx ) locations on uniform distribution of RXPWR for
single and four transmitter scenarios. In this study, we use a single Tx scenario to
study the effects of SAHP and reflectivity on RXPWR. Subsequently, we use five
Txs to evaluate the impact of number of Txs on uniform distribution of RXPWR.
Besides, the effects of different on-body Rx heights are also invesigated (Fig. 1).
Optical Wireless Communication Based ... 987

Fig. 1 System model for OBSN

2 System Description

2.1 System Model

The indoor environment is assumed to be an empty room with dimensions 5 × 5 ×

3 m3 which is consistent with the literature [9, 10]. The LED transmitters, facing
vertically downward, are located on the ceiling as described in Masroor et al. [8] with
an additional Tx located at the centre of the room with coordinates (0, 0, 3). Also, a
photodetector (PD) is assumed to be placed around the patient’s arm and is oriented
towards the ceiling such that the LEDs and PD are facing each other at all times. The
simulations to determine the RXPWR have been performed in MATLAB using the
ray tracing algorithm as described in Barry et al. [11]. The simulation parameters
have been summarized in Table 1.

2.2 Channel Model and Received Power

An intensity modulation and direct detection (IM/DD) channel model has been
considered for VLC-based LED transmitters, given by Ghassemlooy et al. [12]:

y = Hγ x + n (1)
988 K. Masroor et al.

Table 1 Summary of
Parameters Values
simulation parameters
For transmitter
Center luminous intensity per LED chip 0.73 Cd
Lambert’s order 1
Number of LEDs per array 3600 (60 × 60)
For receiver
Field-of-View 85o
Physical detector area 1 cm2
Gain of an optical filter 1.0
Refractive index of lens 1.5
Height (from the floor) → Sitting 0.85 m
Height (from the floor) → Standing 1.2 m

where y is the received signal power, H is the channel static gain, γ is the receiver
responsivity, x is the transmitted signal, and n denotes the Additive White Gaussian
Noise (AWGN) with double-sided noise power spectral density N 0 .
The total received power (Pr-total ) due to LoS (H (0)) and first-order NLoS (Hnlos
k
)
reflections can be calculated as [12]:
 ∞


Pr −total = H (0) + k
Hnlos Pt (2)
k=1

where k is the order of reflections assumed to be 1 in this study.

3 Results and Discussion

In this study, simulations were carried out for a 5 × 5 × 3 m3 room to examine the
effects of physical parameters i.e. the reflectivity of walls and SAHP of LEDs on the
performance of an OBSN link. Illustrated in Fig. 2 is the RXPWR obtained in Watts
for LoS configuration at different angles of irradiance. It can be observed in Fig. 2a
that with single transmitter (Tx ), the RXPWR is mainly concentrated at the centre
when angle of irradiance is 10º while it is more evenly distributed when LED transmits
at 60º (see Fig. 2b). Although the RXPWR has decreased in the latter due to the
increasing distance of the PD from the Tx , it provides better coverage. Consequently,
to ensure maximum coverage, further analysis will be based on LED transmissions
at SAHP = 60º. Furthermore, Fig. 2c illustrates the effects of reflectivity of walls on
the RXPWR via NLoS link for two values i.e. for ρ min and ρ max corresponding to 0.1
and 0.8, respectively. It can be observed that the reflected power reduces considerably
with low reflectivity and vice versa indicating that reflectivity of surfaces contributes
Optical Wireless Communication Based ... 989

(a) (b)

(c) (d)

Fig. 2 Received Power for Single Transmitter with a SAHP 10º and b 60º, c NLoS distribution at
ρ = [0.1, 0.8] and d SAHP = 10º

significantly to RXPWR computations. Besides, Fig. 2d depicts how a small value of

SAHP significantly reduces the RXPWR. Therefore, when designing an environment
for patient monitoring using a VLC link, a high value of SAHP and ρ is desirable.
Figure 3 illustrates the impact of number of Txs on RXPWR. From Fig. 3a, it is
evident that a single Tx is not sufficient to provide uniform coverage to the entire
area and there could be communication dips at several points in the room with
RXPWR spread over ~ 20dBm range; thereby, resulting in significant fluctuations.
Conversely, in Fig. 3b depicting the five transmitters’ scenario, the range has reduced
to ~ 5 dBm implying that the RXPWR fluctuations have reduced by 75%. Based on
these results it can be said that use of multiple Txs is inevitable to establish a reliable
communication link with uniformly spread RXPWR. Apart from the number of
Txs , the Rx heights also play a key role in determining the performance of the OBSN
network. As illustrated in Fig. 4, it can be seen that the values of RXPWR obtained
for standing at different points in the room are higher than those obtained for sitting.
This difference can be attributed to the varying distances (d TR ) between the Txs and
Rx i.e. when d TR is high, the RXPWR would be less and vice versa.
990 K. Masroor et al.

(a) (b)

Fig. 3 Received power for a Single and, b Five transmitters

Fig. 4 Received power for

varying PD heights

4 Conclusion

This paper emphasized the importance of different transceiver parameters that

affect the performance of an OBSN link to carry out the communication effec-
tively. It was found that reflectivity of surfaces and angle of irradiance of the LED
sources play an important role in determining the power received at the PD. Also,
it was shown that a single Tx leads to significant fluctuations in RXPWR, therefore,
multiple Txs must be used to establish uniform distribution. Moreover, it was demon-
strated that as the distance between Txs and Rx reduces, the RXPWR will be better
and vice versa.
Optical Wireless Communication Based ... 991

References

1. Julien-Vergonjanne A, Sahuguede S, Toumieux P (2015) Remote mobile patient monitoring

using diffuse optical transmissions, pp 553–557
2. Torkestani SS, Julien-Vergonjanne A, Cances J-P (2010) Mobile healthcare monitoring in
hospital based on diffuse optical wireless technology. In: 21st Annual IEEE international
symposium on personal, indoor and mobile radio communications. IEEE
3. Scully CG et al (2012) Physiological parameter monitoring from optical recordings with a
mobile phone. IEEE Trans Biomed Eng 59(2):303–306
4. Khalid A, Cossu G, Ciaramella E (2013) Diffuse IR-optical wireless system demonstration for
mobile patient monitoring in hospitals. In: Transparent optical networks (ICTON), 2013 15th
international conference on 2013. IEEE
5. Torkestani SS et al (2012) Indoor optical wireless system dedicated to healthcare application
in a hospital. IET Commun 6(5):541–547
6. Alyan E, Aljunid S, Anuar M (2015) Mobile optical wireless system for healthcare continuous
monitoring using IR technology. J Theor Appl Inf Technol 78(3):353
7. Torkestani SS, et al (2011) Performance and transmission power bound analysis for optical
wireless based mobile healthcare applications. In: Personal indoor and mobile radio commu-
nications (PIMRC), 2011 IEEE 22nd international symposium on 2011. IEEE
8. Masroor K, Jeoti V, Drieberg M (2019) Analyzing the effects of LED lamp arrangements on
performance of an indoor visible light communication system. In: 2019 IEEE 14th Malaysia
international conference on communication (MICC) 2019. IEEE
9. Komine T, Nakagawa M (2004) Fundamental analysis for visible-light communication system
using LED lights. IEEE Trans Consum Electron 50(1):100–107
10. Gismalla M, Abdullah M (2019) Performance evaluation of optical attocells configuration in
an indoor visible light communication. Indones J Electr Eng Comput Sci 14(2):668–676
11. Barry JR et al (1993) Simulation of multipath impulse response for indoor wireless optical
channels. IEEE J Sel Areas Commun 11(3):367–379
12. Ghassemlooy Z, Popoola W, Rajbhandari S (2019) Optical wireless communications: system
and channel modelling with MATLAB® 2019. CRC Press
Development of an IoT-Enabled Stroke
Rehabilitation System

Huzein Fahmi Bin Hawari and Safuan Bin Abu

Abstract Electromyography (EMG) signal plays a crucial role in rehabilitation

training of the lower limb. An-IoT-enabled stoke rehabilitation system was imple-
mented in this study what was based on microprocessor, sensor, cloud and mobile
application. Some of the main concerns that we can fix for stroke rehabilitation real
time monitoring for long-term. Currently, there are a lot of focus on rehabilitation for
upper limb muscle as this is more important as part of the stroke patient’s survival
skills to use their hand for eating, drinking etc. However, the lower limb is also impor-
tant skill for stroke patient to improve their mobility. Thus, this research is to help
explore lower limb stroke rehabilitation. The prototype was designed by combining
muscle sensor and accelerometer for real time monitoring through mobile application
and server. This prototype suitable apply and use in the house for home therapy. With
this prototype, doctor can monitor their patient health status in real time monitoring.
Besides, can connect patient health status for the right information.

Keywords Internet of Things · Stroke · Rehabilitation · Mobile application ·

Cloud

1 Introduction

One of the effective methods to bounce back from stroke effectively is by taking
prevention steps such as identifying and understanding the signs of a stroke, although
prompt care can save lives and significantly affect the chance of full recovery. Unfor-
tunately, in stroke survivors the risk of a second stroke experiencing increase. About
25% of stroke survivors will suffer a second stroke, according to The National Stroke
Association. The probability of a second stroke is around 40% higher within the first
5 years after the first stroke. Apparently, it is predicted that about 80% of all secondary
strokes are preventable with improvements in lifestyle and medical care.

H. F. B. Hawari (B) · S. B. Abu

Department of Electrical and Electronic Engineering Department, Universiti Teknologi
PETRONAS, Bandar, Tronoh, Seri Iskandar, Malaysia

© The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 993
R. Ibrahim et al. (eds.), International Conference on Artificial Intelligence
for Smart Community, Lecture Notes in Electrical Engineering 758,
https://doi.org/10.1007/978-981-16-2183-3_94
994 H. F. B. Hawari and S. B. Abu

By 2019, Malaysia be the 113 in the world according to stroke patients. This type
of disease is very dangerous and silent killer among Malaysia citizen. Due to slowly
growth disease in human body this stroke disease almost unnoticeable until disease
attack at the paralyzed stage. he goal of long-term recovery is to help insure that the
victim of the stroke is as healthy as possible. Theoretically this is achieved in a way
that keeps the patient comfortable and inspires them to learn the new simple skills
like bathing, eating, sitting and walking.
Rehabilitation usually begins in hospital following a stroke. If your diagnosis is
stable, recovery will begin during two days of the stroke and proceed after discharge
from hospital. It can often take several weeks, months, or even years for anyone
after a stroke recovery period is dissimilar. Few people completely recover while
others have long-term or lifelong disabilities. Rehabilition may include consultation
with speech therapists, physiotherapists and occupational therapists. Speech therapy
supports individuals who have trouble generating expression or hearing expression.
Because of the stroke, physical therapy uses exercises to help you relearn balance
and agility skills you may have lost. The aim of occupational therapy is to change
everyday habits, such as feeding, drinking, dressing, bathing, reading and writing.
A stroke rehabilitation system can be divided to upper limb muscle and lower limb
muscle [1]. The top limb or upper limb is the limb of a vertebrate extending from
the deltoid area to and including the neck, including the head, axilla and shoulder
Although, in the general sense of the term, the human leg is the entire lower limb
for the human body, including the foot, groin, and even the hip or gluteal region.
Nonetheless, in human anatomy the term only applies to the part of the lower limb
extended from the knee to the ankle, also known as the crus. Legs are used for
balancing, and all types of locomotion, including leisure activities such as dancing,
are an important part of the mass of a individual. Figure 1, show that the people using
armband upper limb stroke rehabilitation system. In past several past years, stroke
rehabilitation system widely have been invent to detect the movement of upper limb
muscle. A stroke rehabilitation system is a device that can detect lower limb muscle
expand and contract and in the same time, able to analyses the movement of lower
limb muscle.
Basically a stroke rehabilitation system consists of the 2 major parts. The first
part consists of the stroke rehabilitation equipment usually in a form of machine
exercise. Secondly, a wearable device that can detect the movement of lower limb
muscle during the stroke rehabs. In past several years, stroke rehabilitation system
widely been invent to detect the movement of upper limb muscle. From this concept,
it shows that there is a lot of room of possible application of stroke rehabilitation
system that can apply on lower limb of the body. In this research, we will develop an
stroke rehabilitation system that will be able to detect the movement of lower limb
muscle and able to provide useful information through the analysis of the stroke
rehabs to diagnose the movement of lower body muscle by using muscle sensor. By
looking at several criteria from lower limb muscle movement, such as how much the
muscle expand and contract.
Development of an IoT-Enabled Stroke Rehabilitation System 995

Fig. 1 Upper limb stroke

rehabilitation system

The severity of stroke complications and each person’s ability to recover vary
widely. Researchers have found that people who participate in a focused stroke reha-
bilitation program perform better than most people who don’t have stroke rehabili-
tation. For example, these skills can include coordinating leg movements in order to
walk or carrying out the steps involved in any complex activity [2]. According to Joel
Stein, MD, a physician under department of Rehabilitation Machine of Columbia
University’s College, repetitive exercising is the key to stroke rehabilitation. Since
there is no commercial device that is currently available to monitor real time stroke
rehabs at home, it is important to develop a system that a user is able to monitor their
health performance and in the same time, its data being monitor by medical expertise
to ensure the if abnormal sign happens in data.
In IoT, health information of the user can be stored in cloud and the data accessible
by third party like doctor and or medical officer to review the data and analyze the
condition of the user without having a appointment between them. This exchange
information of the user and the doctor will improve the medical check-up quality and
able to track the rehabs activity [3, 4]. With IoT, it is hope the application of stroke
rehabilitation system rehabs diagnostic will help society to boost the quality of life
and the same time, provide user as alternative device to monitor their health.
996 H. F. B. Hawari and S. B. Abu

2 System Design and Architecture

2.1 Central Processing Unit

NodeMCU is the controller which have an open source system for IOT devices and
application which make this controller special the device equipped with ESP8266
WiFi SoC make the device ready to be used for IOT loop from devices and to the
cloud based system [5]. This controller is improving version from Arduino type of
controller. The functioning and software for this controller are same with Arduino
and the only different is WIFI part inside the controller that makes this controller
ready to connected unlike Arduino need to buy Wifi Module and write a programmer
to activate the Wifi by pairing the IP Address.

2.2 Myoware Muscle Sensor

Advancer Technologies muscle module monitors the behavior of muscles by quanti-

fying the electrical potential produced by the muscle cells. This is called electromyo-
graphy (EMG). The sensor magnifies and processes a muscle’s dynamic electrical
action and transforms it into a simple analog signal that can be interpreted efficiently
by any microcontroller with an analog-to-digital converter (ADC), such as an A-
Star or Arduino-or even a Maestro servo controller as we demonstrate [6]. When
the target group of muscles is flexing, the output voltage of the sensor is steadily
increasing. An onboard gain potentiometer can be used to fine-tune the exact rela-
tionship between the output voltage and the muscle operation. he MyoWare Muscle
Sensor is an upgraded version of the older Muscle Sensor v3 of Advancer Tech-
nologies with these upgrades, notably single-supply operation (no need for nega-
tive voltage supply) and built-in snap connectors for electrodes. Certain additional
features include a raw EMG performance, safety against reverse power, a control
switch and LED indicators. To be attached to the skin, the module needs three elec-
trode (not included) which snap into the snap-style connectors of the sensor, making
it simple to attach and disconnect electrodes. Two connectors are located directly on
the PCB, and at the end of the connected reference electrode cable is the third one.

2.3 Accelerometer Sensor

Accelerometers seem to be instruments that evaluate acceleration, which is the rate

of change of an object’s velocity [6]. It calculate in meters per second squared (m/s2 )
or in G-forces (g). For us here on planet Earth, a constant G-force is equal to 9.8 m/s2 ,
although this ranges signficantly with the height (and may be a different magnitude
on different planets due to gravitational pull variations). Accelerometers are useful
Development of an IoT-Enabled Stroke Rehabilitation System 997

for vibration detection in devices or for guidance implementations. Accelerometers

can be plotted on one, two, or three axes. 3-axis platforms are now more popular
as the production costs for them significantly reduce. In fact, accelerometers have
internal capacitive plates. Most of these are set while some are connected to tiny
springs which shift internally as the forces of acceleration operate upon the sensor.
When these plates shift in relation to each other, their capacitance varies among them.
The acceleration can be calculated from certain capacitance changes.

2.4 Thing Speak Cloud

In information technology (IT) sectors, cloud computing is an efficient, scalable, and

costless way to employ and deploy the organization’s resources that include software,
platforms, and infrastructures [7–9]. ThingSpeak allows scientists and engineers to
test and build IoT systems without configuring servers or designing web applications.
An IoT system is one of the main parts of an IoT program. ThingSpeak is one of that
kind widget toolkit that provides a variety of different of features [10, 11]. At the heart
of ThingSpeak is a portal that is used to store and process information from ‘things’.
ThingSpeak also offers additional cloud services, other APIs and social networks
collaboration apps and provides the opportunity to develop the applications as plug-
ins. It is a fantastic forum with vast options to explore how the Internet of Things is
embedded.

2.5 System Block Diagram

The major process flow of whole project as per Fig. 2. Myoware muscle sensor
sense the muscle movement and the accelerometer give 3 axis data of the motion
and send the analog signal to the microcontroller which is NodeMCU powered by
battery. The data that been read by sensor will be connected to the ESP8266 which
a module Wi-Fi which sends the data to the mobile application and store inside
the cloud storage. Within this research, the most significant and one of the latest
technologies is a mobile application, based on extracting real-time information from
medical instruments attached to calculate the cardio details of patients [12].
As in Fig. 3, the wiring diagram show the wiring part and the wireless part is
being described. This project base an IoT technology show that the device portable
and connected to the smartphone and cloud database and smartphone via WiFi
technology.
998 H. F. B. Hawari and S. B. Abu

Fig. 2 Block diagram of the lower limb stroke rehabilitation system

Fig. 3 Circuit diagram of the lower limb stroke rehabilitation system

Development of an IoT-Enabled Stroke Rehabilitation System 999

2.6 System Prototype

As in Fig. 4, the description of the development of the lower limb stroke rehabilitation
system is highlighted. The prototype of this project is attached with the e-box that
contain microcontroller, muscle sensor, accelerometer and battery. The dimension of
the system prototype will be 360 × 290 × 180 mm. The user is required to place their
leg on the rehabilitation platform where they are required to step down the pedal for
a number of tries. The output of accelerometer and muscle sensor will be displayed
real time in the Blynk mobile application and Thing Speak cloud.

3 Results and Discussion

The system prototype test is conducted at old folk’s home at Rumah Seri Kenangan,
Seri Iskandar Perak. The field testing is tested on random elderly at Rumah Seri
Kenangan. The testing is conducted using the whole system communicate together
with Thing Speak Cloud and main central system using wifi module. The data is
then observed through the Thing Speak Cloud. The objective of this field testing is
to study the effectiveness and the efficiency of the overall system.
Person X is a male, 42 years old person with related to asthma problem. Based
on Fig. 5, the muscle sensor attached to the leg is able to provide a good response
whenever the person tries to perform the steps using the system prototype where the
result is posted real time on the Thing Speak cloud. For the accelerometers, it also
good response on different axis, indicating good movement on the leg.
Figure 6 shows the Think Speak for Person Y. Person Y is a female, 46 years old
person with related to diabetic problem and osteoporosis history. The result for the
muscle sensor shows that the muscle movement is less compared to Person X. The
first response is almost similar to Person X, but as the Person Y continue to perform
more steps on the prototype, the movement is slightly stuttered possibly due to her
health condition. As a result, the muscle sensor did not show good response for the
remaining steps. Similarly, for the accelerometer the x axis indicate similarity for
person X, but for Y and Z axis, the results are not different due to stuttered movement
of the leg.
Person Z is an elderly male, 65 years old person with related to diabetic and gout
problem. Person Z has problem with his leg and has problem to perform the steps
movement on the prototype. Based on Fig. 7, the muscle sensor attached to the leg
show that person Z try to perform a step. However, due to his leg problem, he was
only to perform partial step and take about 10 min to perform it. From the Think
Speak cloud, the accelerometer response also shows difficulty in the leg movement
through indication of linear response for different x, y and z axis.
1000 H. F. B. Hawari and S. B. Abu

Fig. 4 Lower limb stroke system prototype

Development of an IoT-Enabled Stroke Rehabilitation System 1001

Fig. 5 Lower limb stroke system prototype think speak result for person ‘X’

Fig. 6 Lower limb stroke system prototype think speak result for person ‘Y’
1002 H. F. B. Hawari and S. B. Abu

Fig. 7 Lower limb stroke system prototype think speak result for person ‘Z’

4 Conclusion

As a conclusion, this project is one of the important innovations needed in our

rehabilitation facilities with real time monitoring technology which are very benefit
for rehabilitation department to monitor citizen. The IoT based stroke rehabilitation
system will gives important information about the patient’s improvement for post
stroke patient. Stroke patient are required to do constant rehabilitation in the hospital.
In rural area there are low accessibility for rehabilitation facilities. Currently, the
conventional method where it requires patient to visit the hospital. At the same time,
the process is time consuming where it requires medical expert to process the data
from human body and identify the illness or disease of the patient. Through the
research, it will give benefits to the parents, caretakers doctors and therapist where
they can monitor the post-stroke patient’s online in real time. By combining the
stroke rehabilitation equipment and muscle sensor with online database system will
improve the efficiency in medical diagnostic where accurate data delivery which are
useful for rehabilitation department to diagnose the movement of the muscle. The
status of the project currently able to calibrate and measure the muscle performance
with a low data error produced.
Development of an IoT-Enabled Stroke Rehabilitation System 1003

Acknowledgements I would like to thank the personnel and elderly from the Rumah Seri
Kenangan, Seri Iskandar Perak who despite of being busy with their schedule, managed to take
time out to provide support on the testing.

References

1. Muth CC (2016) Recovery after stroke. JAMA 316:2440

2. Islam SMR, Kwak D, Kabir MH, Hossain M, Kwak KS (2015) The Internet of things for health
care: a comprehensive survey. IEEE Access 3:678–708
3. Ghosh AM, Halder D, Hossain SKA (2016) Remote health monitoring system through IoT. In:
Paper presented at the 2016 5th international conference on informatics electronics and vision
ICIEV 2016, pp 921–926
4. Chung Hua BD, Fahmi H, Yuhao L, Kiong CC, Harun A (2018) Internet of Things (IOT)
monitoring system for elderly. In: International conference intelligent advance system ICIAS
2018, vol 201
5. Alankus G, Proffitt R, Kelleher C, Engsberg J (2010) Stroke therapy through motion-based
games. In: A case study proceedings of the 12th international ACM SIGACCESS conference
on computers and accessibility (2010), pp 219–226
6. Lu Z, Chen X, Zhao Z, Wang K (2011) A prototype of gesture-based interface. In: Proceedings
conference human-computer interaction mobile devices services (DBLP), pp 33–36
7. Yadav N, Jin Y, Stevano LJ (2019) AR-IoMT mental health rehabilitation applications for smart
cities. In: Smart cities: improving quality of life using ICT & IoT and AI (HONET-ICT) 2019
IEEE 16th international conference on, pp 166–170
8. Dohr A, Modre-Osprian R, Drobics M, Hayn D, Schreier G (2010) The Internet of Things for
ambient assisted living. In: Seventh international conference on information technology, pp
804–809
9. Fan YJ, Yin YH, Da Xu L, Zeng Y, Wu F (2014) IoT-based smart rehabilitation system. IEEE
Trans Ind Informat 10(2):1568–1577
10. Del Din S, Patel S, Cobelli C, Bonato P (2011) Estimating fugl-meyer clinical scores
instroke survivors using wearable sensors. In: Proceeding annual international conference IEEE
engineering med biolsoc EMBS
11. Hawari HF, Zainal AA, Ahmad MR (2019) Development of real time internet of things (IoT)
based air quality monitoring system. Indonesian J Electr Eng CS 13(3):1039–1047
12. Zheng X, Chen W, Cui B (2011) Multi-gradient surface electromyography (SEMG) movement
feature recognition based on wavelet packet analysis and support vector machine (SVM). In:
Proceeding 5th International Conference IEEE Bioinformatics Biomedical Engineering, pp
1–4
MRI Brain Tumor Classification
Technique Using Fuzzy C-Means
Clustering and Artificial Neural Network

Angona Biswas and Md Saiful Islam

Abstract Brain tumor is a deathly disease and it is indispensable to point out the
tumor immediately. Detection of brain tumor from MRI with higher accuracy has
become a major research region for the medical sector. In this paper, an automatic
brain tumor classification procedure applying fuzzy C-means and artificial neural
network is proposed which provides higher precision. For this proposed technique,
inputted MRI images are resized and then sharpening filter is used for preprocessing.
After that, fuzzy C-means cluster process is chosen for image segmentation. At
the next step, discreate wavelet transform is utilized for feature extraction and then
features quantity are reduced by principal component analysis. Furthermore, reduced
features are taken to artificial neural network for brain tumor classification. An
effective training function Levenberg–Marquardt is used for neural network. This
proposed method provides 99.8% accuracy, 100% sensitivity and 99.59% specificity
which is comparatively better than other existing detection techniques.

Keywords Brain tumor · Fuzzy C-means · Discreate wavelet transform · Artificial

neural network · Levenberg–Marquardt function

1 Introduction

Brain tumor is another disease which is playing a dominant role to cause death.
ASCO published that 18,020 adults die from brain tumor in a year [1]. Cancerous
brain tumors can snatch life within 5 years from the diagnosis time [1]. Brain tumors
are mainly abnormal growth of cells that fill up the necessary brain space [2]. 120
types of brain tumors exist but they said all tumor types are not cancerous [2]. Glial
cells are the point from where Glioma tumor arises. Pituitary tumor exists in pituitary
gland [3]. A patient needs to detect the tumor type quickly to initiate treatment. It is
already proved that Artificial Intelligence (AI) has enough skill to detect object more
perfectly than manual process [4]. Again, without trained-experienced radiologists,

A. Biswas · M. S. Islam (B)

Department of Electronics and Telecommunication Engineering, Chittagong University of
Engineering and Technology, Chattogram, Bangladesh

© The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 1005
R. Ibrahim et al. (eds.), International Conference on Artificial Intelligence
for Smart Community, Lecture Notes in Electrical Engineering 758,
https://doi.org/10.1007/978-981-16-2183-3_95
1006 A. Biswas and M. S. Islam

MRI cannot be diagnosed perfectly [4]. Magnetic Resonance Imaging (MRI) is an

efficient way for classification. In that theme, the proposed technique is based on
artificial neural network (ANN) and image segmentation is done by fuzzy C-means
(FCM) clustering.
Different research work was done previously to detect brain tumor types using AI
and several clustering techniques were trialed. Authors N. Nandha, Dr. Karnan was
found 98.87% rate for tumor detection by using FCM with Particle Swarm Optimiza-
tion (PSO). They said their method provided a best result than Genetic Algorithm with
FCM approach comparatively [5]. Authors Dipali M., Dr. N. Rana used histogram
equalization for pre-processing and thresholding for segmentation. They used Neuro
Fuzzy classifier and within eight tests they got three wrong results [6]. MRI brain
tumor image segmentation analysis was done by author Heena Hooda, Om Prakash.
They used K-means clustering and FCM. They claimed that FCM can detect the
tumor area more perfectly than k-means [7]. FCM was used for image clustering,
features were extracted by Grey level matrix (GLRLM) and the SVM was utilized for
image categorization by Parveen and Amritpal Singh [8]. The authors found 91.66%
accuracy by using Kernel Linear Function with 120 images [8]. Malathi and Sinthia
[9] did an improved research work. They used Spatial FCM and Back Propagation
Algorithm [9]. They detected area, defected cells, detection time of tumor. Two
clustering algorithms were used combinedly by Rasel and Foisa [10] to reduce the
error. Templated K-means and improved fuzzy C-means (TKFCM) method provided
97.1% accuracy.
This paper is proposed an effective classification technique using FCM and ANN.
The novelty of the proposed method is to bring out 99.8% accuracy by utilizing proper
preprocessing, clustering algorithm and effective ANN training function from an
immense data. This paper arranges the remaining part as follows. Proposed frame-
work and briefly explanation are presented in Sect. 2. Obtained results are explained
using images, tables in Sect. 3. Finally, the proposed system is concluded in Sect. 4.

2 Methodology

The classification technique of proposed paper includes FCM and ANN to explore
the higher accuracy. The total process of classification can be divided into four steps:
pre-processing steps, image segmentation step, feature extraction and reduction and
finally, classification step with ANN. The workflow of technique is exhibited in
Fig. 1.

2.1 Pre-processing

This proposed technique applied resizing and sharpening filter for preprocessing:
MRI Brain Tumor Classification Technique … 1007

Fig. 1 Proposed
classification technique flow
chart

Image Resizing. Dataset consisted of different size of images. To work with all
images at a time, input images were resized by 200 × 200 pixels which means to
change the pixel information.
Sharpening Filter. Sharpening filter was used to strengthen the contrast shade of
the bright and dark area of the input image. This also enhanced the object’s edge.
Sharpening filter provides better accuracy in case of image processing.
1008 A. Biswas and M. S. Islam

2.2 Fuzzy C-Means (FCM) Clustering

In 1981, Bezdek is the person who introduced FCM algorithm at first and FCM is an
improved version than previous clustering algorithms [11] and FCM allows two or
more clusters in one pixel [12]. FCM is unsupervised clustering process where data
points and centers of cluster’s distance create clusters [13]. It performs better than
k-means (result section). FCM works based on optimizing Objective Function [11].


n 
c
 m  2
J (U, V ) = µi j xi − v j  (1)
i=1 j=1

where, Objective Function = J (U, V), data points = n, fuzziness index = m, member-
ship = µi j , jth cluster center = v j , No. of
 cluster center = c, Euclidean distance of ith
particular to jth mid-point of cluster = xi −v j , k = iteration step, di j = Euclidean
distance.
FCM Algorithm
Step 1: Cluster center ‘c’ is selected randomly
Step 2: Calculation of Fuzzy membership function ‘µi j ’
 c
  2 −1
µi j = 1 di j − dik ( m ) (2)
k=1

Step 3: Computation of fuzzy centers, v j

 n  n 
 m  m
vj = µi j xi µi j , v j = 1, 2, . . . , c (3)
i=1 i=1

Step 4: Repetition of step 2 and 3 before achieving the minimum Objective

Function.
From taking input tumor MRI to segmentation procedure of proposed method is
demonstrated in Fig. 2. Input MRI was first preprocessed and then clustered by FCM.

(a) (b) (c) (d) (e) (f)

Fig. 2 a Glioma MRI. b Pituitary MRI. c Sharpening filtered image of Glioma MRI. d Sharpening
filtered image of Pituitary MRI. e FCM clustered image. f FCM clustered image
MRI Brain Tumor Classification Technique … 1009

Here, Fig. 2a is Glioma and Fig. 2b is Pituitary MRI images. Contrast was strength-
ened by sharpening filter, shown in Fig. 2c and d. Then in Fig. 2e and f, fuzzy C-means
clustering was applied after filtering to segment tumor from the image background
properly.

2.3 Feature Extraction and Reduction

Two-dimensional DWT was used in this proposed method and that was three-level
decomposition wavelet. Discreate Wavelet Transform is employed to extract multiple
features from the filtered image. Fourier transformation has a serious disadvan-
tage as it doesn’t contain any information about time domain [14]. LL, LH, HH,
and HL are the four sub-bands for each scale [15]. Principal Component Anal-
ysis (PCA), an efficient way of feature reduction was used to reduce the quan-
tity of features after feature extraction to reduce computational time and memory.
Extracted multiple features were: Contrast, Energy, Kurtosis, Mean, Entropy, Corre-
lation, RMS, Variance, Homogeneity, Smoothness, Standard Deviation, Skewness,
IDM.

2.4 Proposed Artificial Neural Network Structure

for Classification

Input of this ANN was collection of 13 features, hidden neurons were 10 in hidden
layer and output layer had 2 neurons to indicate glioma and pituitary classes of
brain tumor. This neural network was constructed with the fastest training function
‘Levenberg–Marquardt’. ANN is dexterous tool for classifying brain tumors and
makes flawless decisions. ANN took 70% data for training firstly, 15% sample was
used for validation and then 15% was used for prediction in the testing stage.

3 Experimental Result and Analysis

3.1 Dataset

512 × 512 pixels dimension, T1-weighted data was obtained from [16]. Dataset
consisted of total 489 MRI images where 246 glioma and 243 brain tumors. Axial
and coronal plane images are used for this experiment and dataset was unbiased.
1010 A. Biswas and M. S. Islam

(a) (b) (c) (d) (e) (f)

Fig. 3 a Glioma MRI. b K-means clustered of Glioma. c FCM clustered image of Glioma. d
Pituitary MRI. e K-means clustered of Pituitary. f FCM clustered image of Pituitary

3.2 Selection of Perfect Clustering Algorithm

Figure 3a and d is sequentially Glioma and Pituitary MRI. Sequentially, Fig. 3b is

the k-means segmented image of Fig. 3a and e is the K-means clustered image of
Fig. 3d. Then, Fig. 3c and f are representing the FCM clustered image of Fig. 3a and
d respectively. The result of Fig. 3 makes clear that k-means algorithm was unable to
find out tumor location perfectly. But FCM had the ability to locate perfectly both of
the Glioma and Pituitary tumors which is shown in Fig. 3c and f. In case of k-means
clustering Fig. 3e, it was totally unable to indicate the Pituitary tumor. So, FCM was
selected for methodology and features values were taken for neural network training.

3.3 Classification Accuracy of Proposed Method

The best result confusion matrix is shown in Fig. 4 where 99.8% accuracy was found
for 18 iterations. Figure 4 proves that the proposed methodology can classify glioma
and pituitary brain tumors flawlessly.

Fig. 4 Obtained all

confusion matrix of best
result
MRI Brain Tumor Classification Technique … 1011

Table 1 Comparison of proposed technique with appurtenant work

Reference Classification technique Classification accuracy (%)
[4] Region of interest (ROI) + DWT + gabor 91.9
filter + BPNN
[8] Image enhancement + FCM + SVM 91.66
classifier
[9] K-means and SFCM (KISFCM) + 93.28
morphological filter + BPNN
[10] Templated- K-means + improved FCM 97.1
(TKFCM) + ANN
[15] GoogLeNet (learn features) + deep 98
CNN-SVM
Proposed method Resizing + sharpening filter + FCM + 2D 99.8
DWT + PCA + ANN

Accuracy = (T N + T P)/(T P + T N + F N + F P)100 (4)

Sensitivit y = T P/(T P + F N )100 (5)

Speci f icit y = T N /(T N + F P)100 (6)

where, TP = True Positive, TN = True Negative, FP = False Positive, FN = False

Negative. Accuracy = 99.8%, Sensitivity = 100%, Specificity = 99.59% are elevated
from the confusion matrix of Fig. 4 and Eqs. (4)–(6). Network Performance was found
0.0033 from target data and output data.

3.4 Comparison of Proposed Technique

In Table 1, different applied methods of authors and obtained result of proposed

approach are compared. Table 1 makes it clear that the introduced approachment
is tremendously successful. The proposed method is simple and obtain the highest
accuracy of 99.8%. Reasons behind this higher precision are choosing FCM based
segmentation and using ‘Levenberg–Marquardt’ training function.

4 Conclusion

A unique and flawless technique is proposed based on FCM clustering and ANN
classifier to classify glioma tumors and pituitary tumors using MRI. This method the
removes previous limitation of low accuracy by using 489 patient’s data. MRI images
1012 A. Biswas and M. S. Islam

were pre-processed by sharpening filter to enhance the contrast. Then better clustering
algorithm ‘fuzzy C-means (FCM)’ was selected for segmentation. DWT was utilized
to extract features from clustered images and for feature reduction, PCA was applied.
13 features were collected from images and then used for artificial neural network
training and testing. The fastest ANN function ‘Levenberg–Marquardt’ was used
for proposed network training. Test result provided 99.8% accuracy at 18 iterations,
100% sensitivity, 99.59% specificity. This proposed method is able to provide 99.8%
accuracy which is much better than some previous classification techniques. This
result is enough to claim that proposed technique can classify glioma and pituitary
tumor types magnificently. This improved accuracy is beneficial for medical sector
to classify automatically.

References

1. Brain Tumor (2020) American Society of Clinical Oncology (ASCO) Publication

2. Brain Cancer Types (2020) Cancer Treatment Centers of America
3. Gliomas (2020) Johns Hopkins Medicine
4. Mustafa R, Ikhlas A (2018) Brain tumor classification via statistical features and back-
propagation neural network. IEEE Xplore
5. Gopal N, Karnan S (2011) Diagnose brain tumor through MRI using image processing clus-
tering such as FCM along with intelligent optimization techniques. In: International conference
on intelligence and computing research. IEEE Xplore, pp 112–116
6. Dipali J, Rana S, Misra T (2010) Classification of brain cancer using ANN. In: 2nd international
conference on electronic technology (ICECT 2010). IEEE, pp 112–116
7. Heena Hooda F, Om Verma S, Tripti T (2014) Brain tumor segmentation: a performance
analysis using K-Means, FCM and region growing algorithm. In: International conference on
communication and computing technologies (ICACCCT). IEEE, pp 1621–1626
8. Parveen F, Amritpal S (2015) Detection of brain tumor in MRI images, using combination of
FCM and SVM. In: 2nd international conference on signal processing and integrated networks
(SPIN). IEEE, pp 98–102
9. Malathi MF, Sinthia P (2020) MRI brain tumor segmentation using hybrid clustering and
classification by BPNN. Asian Pac J Cancer Prev 19
10. Rasel A, Md Foisal H (2016) Tumor detection in brain MRI image using template based K-
means and fuzzy C-means clustering algorithm. In: 2016 international conference on computer
communication and informatics (ICCCI-2016). IEEE
11. Preetha R, Suresh FG (2014) Performance analysis of FCM in automated detection of brain
tumor. In: World congress on computing and technologies. IEEE, pp 30–33
12. Shasidhar MF, Raja SSV, Kumar BVT (2011) MRI brain image segmentation using modified
fuzzy C-means clustering algorithm. In: 2011 international conference on communication
systems and network technologies. IEEE, pp 473–478
13. Brundha BF, Nagendra MS (2015) MR image segmentation of brain to detect brain tumor and
its area calculation using K-means clustering and fuzzy C-means algorithm. Int J Technol Res
Eng 2(9)
14. Zhang Y, Wu LS (2012) An MR brain images classifier via principal component analysis and
SVM. Prog Electromagnet Res 130:369–388
15. Deepak S, Ameer PM (2019) Brain tumor classification using deep CNN features via transfer
learning. Comput Biol Med 111
16. Brain tumor dataset into figshare (2017) https://figshare.com/articles/brain_tumor_dataset/151
2427
An IoT Based Automatic Vehicle
Accident Detection and Rescue System

K. Gayathri Devi , K. Yasoda, B. Rajesh, R. Sowmiya, and S. S. Vishalidevi

Abstract A large variety of precious lives are lost because of road traffic accidents
each day that may occur due to the mistake of driver and delayed response from the
place of accident to the emergency services. The most important aspect of driver is
to own a good road accident detection and data communication system in place to
save injured persons. A system that sends information messages to close emergency
services regarding the accident location for timely response is completely essential.
The accident detection process uses raspberry pi, in detection of the data from the
vibration sensor and accelerometer during accident. The process of rescue system
uses GSM, in the process of sending message to the neighboring medical Centre or
relatives. GPS is used to track the location of the vehicle and used in the calculation
of the speed of the vehicle based on the position of (latitude, longitude) and time
difference. GSM in the system is used to send the location of accident through
message. Vehicle ad-hoc network (Wi-Fi) is used to transfer of the data from the
controlling unit to the application server in the later retrieval of data. Hyper Text
Transfer Protocol is used to transfer web page from remote server after internet
connection is established. The FireBase database is used in the storage of collected
data, and uploads over in the web-page and in the process informing the accident
location and in alerting.

Keywords Accident detection · Human rescue system · Global positioning system

(GPS) · Global system for mobile (GSM) · Vehicle ad-hoc network

K. Gayathri Devi (B) · B. Rajesh · R. Sowmiya · S. S. Vishalidevi

Dr. N.G.P Institute of Technology, Coimbatore, Tamilnadu, India
e-mail: [email protected]
K. Yasoda
SNS College of Technology, Coimbatore, Tamilnadu, India

© The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 1013
R. Ibrahim et al. (eds.), International Conference on Artificial Intelligence
for Smart Community, Lecture Notes in Electrical Engineering 758,
https://doi.org/10.1007/978-981-16-2183-3_96
1014 K. Gayathri Devi et al.

1 Introduction

Nearly 1.25 million individuals die in road crashes annually, on average 3287 deaths
on a daily basis. An additional 20–50 million are injured or disabled. Over 1/2 all
road traffic deaths occur among young adults ages 15–44. There are 1012 deaths in
2019. This can be 74 deaths (+7.9%) over an equivalent amount in 2018. Within the
12 months complete October 2019 there have been 1209 road deaths. Throughout
calendar 2018 there have been 1135 road deaths, 7% not up to in 2017. A recent
World Health Organization (WHO) report showed that each year 1.35 million people
die and 50 million place get injured. Road accidents are ranked because the eighth
leading reason behinds deaths (up from ninth its previous report in 2015), with the
Association for Safe International Road Travel (ASIRT) predicting that it’s going to
rise to the fifth leading reason behind deaths within close to future, unless forceful
changes occur. Moreover, because the social damage caused by road traffic accidents,
there is a big price. ASIRT estimates that between one and two percent of the annual
budget of each country is spent on road accidents. Now-a-days, it has become terribly
troublesome to understand that an accident has occurred and to find the position
wherever it’s happened. There is no system of identification and intimation relating
to an accident in the past. For intimation purpose messaging services is used. GPS
makes use of identification of location and GSM is built for the usage of intimation.
The most intention of an accident detection is to locate the accident spot anywhere
and intimating the emergency through the GPS and GSM networks. The GPS based
vehicle accident identification module contains Mostly of different sensors such
as accelerometer, MEMS etc., GSM module and a GPS modem connected to the
microcontroller. Global System for Mobiles (GSM) technology is employed for a
proper cellular mobile communication. Position of the vehicle accident location is
traced down by GPS.

2 Literature Survey

The self-collision detection sensor is used to detect the accident and the information
is uploaded to cloud based database server collision vehicle accident recognition,
and a connected emergency notification is provided [1]. The detection of accidents
supported by monitored speed and send the accident location acquired from the GPS
together with the time by GSM and therefore the speed by utilizing the GPS network
[2], sent to a OpenGTS server, providing a time period OpenStreetMap visualization
of traffic scenario. MongoDB appraise the close mobile APPs with alert messages
[3]. Auto continuous collision accident collision mechanism, exploitation physical
phenomenon preview iterative algorithm, trace examination reconstruction algo-
rithm, serial collision contact position reconstruction localization algorithm [4], self-
tuning iterative hard thresholding (ST-IHT) algorithm for learning distributed spatio-
temporal features and a weighted extreme learning machine (W-ELM) [5], Sudden
An IoT Based Automatic Vehicle Accident Detection … 1015

Lane Departure Avoidance Module and Rear-End Collision Avoidance Module [6]
is used over in the process of detection of accident. The accident is detected by
the in-Vehicle sensor and updates in database, verifies and compares the severity of
accident supported to different accident data [7] and notifies the emergency services,
uses remote cut off mechanism [8] and using GPS and GSM intimates the emergency
services [9, 10] and provides over an crash path reconstruction using Kalman filter
and GPS [11]. Data continuously collected from the smartphones accelerometer and
analyzed using Dynamic Time Warping (DTW) [12], inputs from the accelerometer,
magnetometer, gyroscope and on vehicle crash detection [13] are analyzed and inti-
mated as e-Call with owner’s information. The local prediction is exchanged from
each sensor with neighboring vehicles aggregates the native predictions it receives
using a weighted majority aggregation rule to get a final prediction [14], crash signals
at different locations of a vehicle and their implications on crash analyzes the severity
of accident [15], the system also sends a warning message. The system uses CAD-
CVIS consists of varied sorts of accident types, atmospheric conditions and accident
location, a deep neural network model YOLO-CA supported CAD-CVIS and deep
learning algorithms to detect accidents and in alerting [16]. The system detects and
assesses the severity of accidents with the assistance of an on-board unit and the
TestBed analyses different accident scenarios and helps in rejecting false alert and
serially detects the rollover and force [17]. Besides over these many systems, they
are costly and not applicable in every cars and lacking over in the process of saving
location for later use.

3 Proposed Model

This works focuses on designing and developing an accident detection using vibration
sensors and wireless modules. The block diagram of the proposed system is shown in
Fig. 1. In order to enhance real time accident detection and monitoring, the system has
Vibration sensor SW-420, GPS and GSM module, which is connected to Raspberry
pi controller, which works with Wi-Fi in transfer of data. The Vibration sensor SW-
420 produces an digital output, on the detection of any collision or accident. GPS is
used to track over the location in which the vehicle travels. GSM is used to send the
location of accident as an URL link with the latitude and longitudinal information on
the place of accident. Raspberry pi controlling unit is used to sense the information
and to perform the entire controlling functions with database through Wi-Fi. For the
enhancement of real time accident monitoring, the system with sensors which works
on IoT basis [18, 19]. By using IoT, status on speed and the location of accident is
recorded and the database manages over the data and uploads over in the webpage,
in tracking of vehicle and its speed.
The connection and the flow diagram of the Vibration Sensor SW420 with Rasp-
berry pi is given in Figs. 2 and 3. The accident detection system operates with the
vibration sensor producing digital pulse output on the detection of any accident or
collision. It produces over an output based on the threshold which is been set over
1016 K. Gayathri Devi et al.

Power Supply
CLOUD

Vibration Database-Firebase
Sensor

Raspberry
GPS Pi 3
Tracks Location
and Speed
Reset GSM
Button
Update Location
Message

Fig. 1 Block diagram of proposed system

Fig. 2 Connection of the vibration sensor SW420 with Raspberry pi

Fig. 3 Flow diagram of the

vibration sensor SW420 with
Raspberry pi
An IoT Based Automatic Vehicle Accident Detection … 1017

in the potentiometer. The sensor is tightly fitted over in any part of the car. The
variations in the output of the sensor are trapped and those are fed as input to the
micro-controller.
GPS is commonly used for both tracking and navigation in vehicles. Tracking
systems helps the base station to keep track of the vehicles in the way of travel without
the intervention of the driver whereas, the system also helps the driver to reach the
destination by navigation. The architecture of navigation or tracking system is more
or less similar. The connection and the flow diagram of the GPS with Raspberry pi
is given in Fig. 4. During the occurrence of accident in any place, tracking on the
position of the vehicle by GPS system and send the information by alerting person
through call or SMS by using GSM.
The connection and the flow diagram of the GSM with Raspberry pi is given in
Fig. 5. Now based on these variations, the micro-controller is programmed to take
decision which in turn.
• Sent as an message an through the GSM to relative or nearby emergency service.
• Data is updated over in the database through wireless transmission.
• Webpage provides over the alert message and update the location in the map.

Fig. 4 Connection and flow diagram of GPS with Raspberry pi

Fig. 5 Connection diagram

of GSM with Raspberry pi
1018 K. Gayathri Devi et al.

3.1 Wireless Transmission in IoT

Wireless communication system acts as a bridge for duplex communication for data
collection and control message delivery and is an essential part for IoT infrastruc-
ture. Various IoT applications can be applied using Wireless communication system,
which includes health care monitoring and home automation etc. and gains a lot of
attention in connecting devices to internet.

3.2 Device-To-Cloud Communication

Internet cloud services like an application service provider is used to exchange data
and manage the traffic of messages that will be connected directly to IoT devices
in a device-to-cloud communication model. There are numerous advantages when
a connection is established between the device and the IP network through cloud
service when compared with the existing communications mechanisms like Ethernet
or Wi-Fi connections.
The Device-to-cloud communication model that is proposed in our work is shown
in Fig. 6 and maintained by more consumer IoT devices like the nest labs learning
devices and many SmartTv and Smart phones. The Nest Learning Thermostat uses the
devices to broadcast data to a cloud database where the data can be used to analyzed
and problems can be identified. Cloud connection may help or enable user to obtain
remote accessing of data through smartphones or web interface and supporting in
some software updates to device.
The SmartTV technology make use of the Internet connectivity for the process
of transmitting the information viewed by the user to do further analysis by the
service provider and this enables the voice recognition features. The user can expand
capability of device further than its original features to device-cloud module with
respect to certain cases. Interoperability problem exist when there is an amalgamation
of device detail from many diverse manufacturer it can be avoided if the authorized

Fig. 6 Device-to-cloud communication model diagram

An IoT Based Automatic Vehicle Accident Detection … 1019

data protocols are used between the device and the cloud services. The device owner
or user can be coupled to a precise cloud service in order to prevent them the usage
of alternative service providers. This is referred to as “vendor lock-in”, a term the
compasses other facts of the relationship with the provider such as ownership of
and access to the data. In mean time, users can have confidence that devices can be
integrated based on the design for the specific platform.
IoT has made the change in the evolution today’s world by connecting things
over the Internet, which makes very much easier remote access. IoT is the emerging
and trending technology that can be implemented over in the safety and security
monitoring systems and in rescuing.

4 Results and Discussion

The Raspberry pi controls the functionality of the above mentioned sensor and
modules. Programming for the controller is done using Python coding. After comple-
tion of the entire program, debugging is carried out. Finally, the program is dumped
to the controller. Then, the working mechanism is verified. Figure 7 gives the snap-
shot of the prototype developed with major components. IoT plays a major role in
collection of database, from the user. Each user can reach the server using their own
e-mail id. This is achieved by means of mobile-to-cloud communication. When the
user access the server, any variation in sensor detection is updated in form of database
with the date and time. Therefore, real time accident detection is achieved by the
implementation of IoT.

Vibration sensor GPS

GSM SIM800A

Raspberry pi 3

Fig. 7 Hardware implementation

1020 K. Gayathri Devi et al.

Fig. 8 Location of accident

occurrence

Figure 8 gives the snapshots of the location of the accident that were transmitted
through GSM. In this prototype firebase help in cloud based data storage and helps
in easy retrieval of data. Firebase provides developers in the platform of mobile and
web app development and which provides a plenty of tools and services which help in
developing high-quality apps, grow their user base, and earn more profit without any
back end server. Figure 9 gives the snapshot of the updation of location and speed in
the database along with the information on accident occurrence through wi-fi using
HTTP protocol. Webpage provides over the view in which location the vehicle travels,

Fig. 9 Updation of accident occurrence in the database

An IoT Based Automatic Vehicle Accident Detection … 1021

Fig. 10 Accident detection message in webpage

Table 1 Result comparison between existing and proposed system

Parameter Existing system [20] Proposed system
Accident detector Li-Fi system between cars Vibration sensor in every car
System adaption Hard Easy
Alert system Implemented using connected cars Implemented using IoT
Data collection Not implemented Implemented and stores the data over
in database

and provides over a pop-up information on the accident occurrence. Figure 10 gives
the snapshot of the webpage during accident occurrence which gathers information
from the database. The information based on the location helps in easy retrieval of
victim people.
The Comparison between the proposed and existing system is given in Table 1.
The existing system was designed with need to transmit the accident location using
IoT technology to cater the needs of people who are undergoing accident.

5 Conclusion

This proposed system with advanced features reduces manual work and allows
detecting the accident and reporting to the emergency services and updating data
in database. The IoT technology used in this system helps in treating the victim
people instantly in case of any emergencies and the route to the location of accident.
1022 K. Gayathri Devi et al.

References

1. Chang WJ, Chen LB, Su KY (2019) DeepCrash: a deep learning-based internet of vehicles
system for head-on and single-vehicle accident detection with emergency notification. IEEE
Access 7:148163–148175
2. Amin MS, Jalil J, Reaz MBI (2012) Accident detection and reporting system using GPS, GPRS
and GSM technology. In: IEEE international conference on informatics, electronics & vision
(ICIEV), pp 640–643
3. Celesti A, Galletta A, Carnevale L, Fazio M, Lay-Ekuakille A, Villari M (2017) An IoT cloud
system for traffic monitoring and vehicular accidents prevention based on mobile sensor data
processing. IEEE Sens J 18(12):4795–4802
4. Lang W, Biao G, Tao C (2013) Vehicle continuous collision accident reconstruction system
development. Procedia-Soc Behav Sci 96:1659–1669
5. Yu Y, Xu M, Gu J (2019) Vision-based traffic accident detection using sparse spatio-temporal
features and weighted extreme learning machine. IET Intel Transp Syst 13(9):1417–1428
6. Chang TH, Hsu CS, Wang C, Yang LK (2008) Onboard measurement and warning module for
irregular vehicle behavior. IEEE Trans Intell Transp Syst 9(3):501–513
7. Fogue M, Garrido P, Martinez FJ, Cano JC, Calafate CT, Manzoni P (2013) A system for auto-
matic notification and severity estimation of automotive accidents. IEEE Trans Mob Comput
13(5):948–963
8. Srabanti S, Asaduzzaman M, Mokter MKB, Anannya TT, Tumpa SN, Afroze L, Karim
MM (2018) A proposed system for automatic vehicle monitoring and accident detection in
Bangladesh. In: 2018 international conference on computer, communication, chemical, material
and electronic engineering (IC4ME2). IEEE, pp 1–5
9. Kumar A, Jaganivasan V, Sathish T, Mohanram S (2018) Accident detection and alerting system
using GPS & GSM. Int J Pure Appl Math 119(15):885–891
10. Taj FW, Masum AKM, Reza ST, Chy MKA, Mahbub I (2018) Automatic accident detection and
human rescue system: assistance through communication technologies. In: 2018 international
conference on innovations in science, engineering and technology (ICISET). IEEE, pp 496–500
11. Aldimirov M, Arnaudov R (2018) Method for automated reconstruction of a car’s path during
crash from GPS/INS data using a Kalman filter. Adv Eng Softw 115:386–390
12. Aloul F, Zualkernan I, Abu-Salma R, Al-Ali H, Al-Merri M (2015) iBump: smartphone
application to detect car accidents. Comput Electr Eng 43:66–75
13. Fernandes B, Alam M, Gomes V, Ferreria J, Oliveira A (2016) Automatic accident detection
with multi-modal alert system implementation for ITS. Veh Commun 3:1–11
14. Canzian L, Demiryurek U, van der Schaar M (2015) Collision detection by networked sensors.
IEEE Trans Signal Inf Proc Over Networks 2(1):1–15
15. Chan CY (2002) On the detection of vehicular crashes-system characteristics and architecture.
IEEE Trans Veh Technol 51(1):180–193
16. Tian D, Zhang C, Duan X, Wang X (2019) An automatic car accident detection method based
on cooperative vehicle infrastructure systems. IEEE Access 7:127453–127463
17. Khaliq KA, Raza SM, Chughtai O, Qayyum A, Pannek J (2018) Experimental validation of an
accident detection and management application in vehicular environment. Comput Electr Eng
71:137–150
18. Kokila J, Gayathri Devi K, Dhivya M, Haritha Jose C (2017) Design and implementation of
IoT based waste management system. Middle-East J Sci Res 25(5):995–1000
19. Hou Z, Chen Y (2017) A real time vehicle collision detecting and reporting system based on
internet of things technology. In: 2017 3rd IEEE international conference on computer and
communications (ICCC). IEEE, pp 1135–1139
20. Krishnan P (2018) Design of collision detection system for smart car using Li-Fi and ultrasonic
sensor. IEEE Trans Veh Technol 67(12):11420–11426
Automatic Firefighting System Using
Unmanned Aerial Vehicle

K. Gayathri Devi , K. Yasoda, and Maria Nithin Roy

Abstract In the world of technological advancements, Internet of Things had

opened a vast opportunities and gained attention in academics and in industrial
automation. This paper focuses on the general representation of the design of the
Unmanned Aerial Vehicle with fire extinguishing balls that can be applied for fire-
fighting situation. The paper highlights the design, and the purview of the experiment
conducted with respect to fire extinguishing balls that is attached to the drone. The
observation of the experiments clearly signifies that the size of the fire extinguishing
balls connected to Unmanned Aerial Vehicle (UAV) system should be lesser to put
off the fire outbursts. The proposed work has implemented an drone of required
specification to extinguish the fire using fire balls attachable to drones. The hard-
ware consists of Navio 2 and Raspberry pi 3 as flight controller interfaced with drone
(UAV). The system design is a combination of internet of things and dedicated server
for continuous storage and monitoring system using Amazon web services.

Keywords Amazon web services (AWS) · Drones · Fire extinguishing balls ·

Remote sensing · Unmanned aerial vehicle (UAV) · Wildfires · Navio · Raspberry
pi

1 Background

The implementation of the systems with drone and various remote sensing tech-
nologies is rapidly increasing to provide a mobile, and low-cost alternative for the
previous systems such as ground-based systems, air controlled vehicles, and satellite-
systems [1]. The images captured by the satellites with low resolution and noise are
not adequate to provide effective result. It is also proves ineffective in populated areas

K. Gayathri Devi (B) · M. N. Roy

Dr. N.G.P Institute of Technology, Coimbatore, Tamilnadu, India
e-mail: [email protected]
K. Yasoda
SNS College of Technology, Coimbatore, Tamilnadu, India

© The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 1023
R. Ibrahim et al. (eds.), International Conference on Artificial Intelligence
for Smart Community, Lecture Notes in Electrical Engineering 758,
https://doi.org/10.1007/978-981-16-2183-3_97
1024 K. Gayathri Devi et al.

[2, 3] since there exist problem to classify the region of interest and there may be
misclassification. The constrained surveillance ranges affect measurement of ground
equipment [4]. The early stages of fire detection UAV system with image capturing
is now replaced by infrared cameras for detecting the radiation emitted by fires [5].
A direct view of the radiation is required by the infrared sensors for effective results
[3]. Visual cameras, a small size high resolution camera have been used for detection
of smoke produced by fire in daylight [6]. Visual cameras provide the details such as
flame height, angle, location and width. These cameras produce different contrast,
texture, and motion analysis [3]. Another mentioned is the usage of LIDAR devices
to identify smoke particles [7].
Holman et al. [4] proposed that UAV system can be used to capture images from
a camera fixed to systems in a real time environment. It was concluded that image
processing quality of topographic images which are in existence and the meta data
of the specific image is rectified and stored for later usage purposes. The paper
discusses the quality of resolution of images and how they are processed frame by
frame. He constructed a similar exiting prototype of unmanned aerial vehicle (UAV)
with a quadcopter setup to collect test images for specific enhancements of low cost
and flexibility. The autonomous functions in the UAV is capable of excellent ground
station. The accuracy of images from ground position could be found with accuracy.
Kumar et al. [8] had developed a optimization model for the detection of fire
and to extinguish the fire with fire suppressing fluid for a region where there are no
habitant with multiple UAV. Two step approach is considered, the first step represents
the fire by a two dimensional shape with a boundary, it requires all the UAV to work
in coordination within the boundary and there should be no collision. The second
step proposes a model as a function of fire and time. This approach aims at reducing
the Euclidean distance between the quadcopter and the fire at any instant of time.
There are few studies that considered fire suppressants other than water. They tested
this theory with master model with use of 10 drones via simulation. The suppressant
fluid is not practical. The undergraduate project designed as a quadcopter with a
ball-dropping mechanism that can drop a single unit of fire-extinguishing ball by
using compressed springs.
Wallar et al. [9] suggested a approach for the surveillance of risky areas by using
three quadcopters to evaluate the coverage area of the sensors on board, the quality
of the data collected with reference to various altitudes and risk involved with ground
level. The two stage approach is implemented in which the movement of the quad
copter to maximize the coverage area is considered in first stage and the second
age optimizes the risk involved. It was intended for the examination of regions to
diminish the detection risk and to ensure the data provided by the sensor is of high
quality. This approach is more costly as more number of quad copters are used.
Mois et al. [10] has focused on the data transmission, security and success rate of
huge amount of environmental parameter data with three the Internet of Things (IoT)
based protocol, the cloud computations, and cyber-physical systems. Three types of
IoT-based wireless sensors for monitoring the environment and the surroundings
were analyzed in this paper. The web based communication established with User
Datagram Protocol (UDP) was not reliable as the number of packets that was lost
Automatic Firefighting System Using Unmanned Aerial Vehicle 1025

in each minute and only the success of the transmission was 95.4% and it is mainly
due to the network and is not related with the hardware. The success rate of the
transmission was slightly increased, there exist a reduced battery life time when
the communication was established through Wi-Fi and Hypertext Transfer Protocol
(HTTP).The packets transmitted are received only when the devices is within a
certain range in the case of third one using Bluetooth Smart. This paper concludes
with various advantages and disadvantages of using the above three communication
protocol with respect to the factors such as power consumption and life time of the
devices, Internet facility, success and reliability of the transmission and revealed that
all three can be used for web based application. Wi-Fi and BLE are two technologies
suited for ZigBee protocol for monitoring applications that can successfully compete
with the concept established.
Zhou et al. [11] proposed a tracking framework and a systematic road detection
approach in video footage using UAV. Initially Graph cut algorithm is applied for the
extraction of the road regions from other non road regions. After the initialization,
for quick retrieval of the road region homography estimation is used. The proposed
framework has a precision of 97.4% and processing 35 frames per second for 1080
× 600 videos on average. Author has published a systematic approach for the identi-
fication of road by applying the relevant algorithm for different time instant in UAV
videos.
Lockheed Martin et al. conducted and demonstrated a collaborative system
comprising of drone and helicopter. The helicopter is designed to drop water exactly
on the hot spots identified by the UAV [12]. Companies such as Aerones, Nitrofirex,
and Singular Aircraft are focusing to design and implement drone systems that
make use of water to suppress fires [13–15]. Phan and Liu [16] developed a system
consisting of airship, UAV and unmanned ground vehicles (UGVs). The peak level
of the hierarchy the airship generates a mission plan by utilizing fire outbursts, UAV,
and UGV planned models [16].
The continuous research of developing UAV system for the detection and construc-
tion of wildfire implemented with drones, application program to program the hard-
ware and relevant algorithms for detection and tracking is in progress day by day
[4]. The main focus is different from the previous establishments or papers which
deliberately work in developing and resulting a remote-sensing capability which will
give an alarm not only for fires, but it also in detection of the fire head and flanks’
spreading. Thus remote sensing is designed to detect fires caused in buildings, fire-
fighters, and based on that will recognize the risk of the fire spreading to any other
area or a precise location parameters. The purview of this particular paper is designed
for this remote-sensing system.
Howden et al. developed a system with multiple that overlooks complex areas for
fires [17]. This system will split the area under surveillance into cells and the parame-
ters such as velocity, pheromone factor and the repulsion factor will be continuously
updated for every iteration. The simulation result for various parameters and giving
priority to the cell was also investigated in detail. Their algorithm proved successful
when implemented in real time environment as one of the UAV may have less fuel
and refilling is needed suddenly, one of them may crash and one may become failure.
1026 K. Gayathri Devi et al.

Pham et al. also developed a similar algorithm for multiple UAV system that iden-
tifies the fire that is spreading across the boundaries [18]. It ensures collisions are
avoided between UAVs. Yuan et al. implemented a forest fire detection system for
a forest fire when the fire is spreading in regions [19]. The video captured by an
aircraft in real time environment and the detection proves the systems reliable and
accurate. Similarly, a decentralized autonomous system of multiple UAVs to monitor
and govern the perimeter of the forest fire and containment was presented by Alexis
et al [20].
The detailed analysis of the implementation of various system for fire detection
and methods to extinguish the fire motivated to implement the drone of required
specification to extinguish the fire using fire balls attachable to drones. This system
is implemented on low power chip Raspberry Pi 3 and Navio 2. The Raspberry
Pi 3 is interfaced with Navio 2 module for better specific controls such as barom-
eter, accelerometer etc. The hardware modules interfaced with raspberry pi 3 are 4
BLDC motor, 4 ESC, Power distribution board, a transceiver, and a thermal sensor.
The system design is a combination of internet of things and dedicated server for
continuous storage and monitoring system using Amazon web services.

2 Proposed System

Figure 1 is the representation of implemented wireless UAV system in two stages.

Stage 1 is for the detection and to extinguish the fire. The stage 2 of Fig. 1 highlights
the blocks used for transmitting the fire outburst information. Figure 2 shows the
interconnection of the various hardware of the UAV System. The central unit of the
drone is flight controller Raspberry pi3 connected to brushless motor and electronic

Fig. 1 Proposed UAV system

Automatic Firefighting System Using Unmanned Aerial Vehicle 1027

Fig. 2 Connection diagram of the hardware of the system

speed controller (ESC) for the stable movement. The purpose of using electronic
speed control or ESC is for varying the speed of Brushless DC electric motor (BLDC)
motor, its direction and also to act as a dynamic brake. The Raspberry Pi 3 is build
with powerful processor, ten times speeder than the first generation, wireless LAN
and with Bluetooth connectivity.
The drone frame used in the system is built from quality glass fiber and polyamide
nylon. The main setup of the UAV is connecting the motors with the power distri-
bution board where the power to the motors is controlled by the electronic speed
controller. The power distributed board is connected to the flight controller where
the input output to the ESC is connected to the flight controller. Then the LIPO
battery is connected to the power distribution board and tested for the function of
the motors. For wireless communications a transmitter is connected to the raspberry
pi. On bottom of the UAV the fire ball drop is fixed. The balls effectiveness and
robustness for building fires is maintained by conducted experiments. The fire extin-
guisher ball (AFO) which is approximately 0.7 kg, is used. The AFO was dropped
directly into fire, whereas they were located in bottom of the drone frame which can
be released the fire a few meters above.
The stage 2 focuses upon the software connection using dedicated server. Here
the raspberry pi is loaded with a pre-configured operating system and connected to
amazon web services (AWS) which is used to monitor and track the UAV movements.
Here the operating system image configures the flight controller, so the dumping of
operating system is done on a memory card. The installation takes place containing
1028 K. Gayathri Devi et al.

Fig. 3 Prototype of the developed drone

all services, kind of a meeting point for nodes. Hence, the nodes are discoverable and
communicated to each other Raspberry Pi. There are different set of drivers within
one node.
Considering fire is around two meter diameter, 0.7 kg AFO ball was dropped, the
possibilities of which if the ball exploded but was not able to extinguish the fire and the
fir outbursts continues or if it did not explode and the fire passed an approximate area
of three meters in diameter, the fire was going to be put out manually by the firefighters
These conditions were set for safety purposes. Figure 3 gives drone developed. The
drone is connected to the ground station and controlled via internet. Using the AWS
console a new account is created with the following services such as Lambda Service,
Simple Storage Service, Cloud watch, Cloud alarm, Simple Notification Services and
some additional services for external image processing. The servos are backed up by
the power to so a Bit Error Correction (BEC) should be present. BEC would serve as
back-up power supply to Navio + incase the main power source supply is intervened.
ESCs will heat each other if many ESC central wire connected to Navio 2.

3 Results

Drone configurations are usually connected to motors and move when the downward
thrust of UAV system is equal to the gravitational pull working against it. When the
rotors provide an upward force greater than gravity, they automatically descend, and
the UAV is directed in the desired direction using pitch, yaw, and roll. When the
rotors provide an upward force greater than gravity, they automatically descend, and
the UAV is directed in the desired direction using pitch, yaw, and roll. Drones are a
microcosm because they are dependent to provide two-way communications for the
Internet of Things. Testing of the drone must be done for safety and reliability for
minimum cost and requirements. They must have fail-proof software that is certified
for safety and effective in communication systems. They also should have efficient
Automatic Firefighting System Using Unmanned Aerial Vehicle 1029

power management and ability to operate in challenging and difficult environments.

The raspberry pi acts as an interface between the Navio and AWS console. The
mission planner software a GPS location is being fixed and controlled. The use of
IoT core enables us to dump software into the microcontroller. The transmitter is an
additional safe-switch.
The amazon web services console is a web application which enables the users or
organizations to establish their application worldwide. The management of console
consists of root user where certain permissions are allowed and denied. This enables
us to monitor in the AWS console which will be ground control station. Initially the
connection was checked using the putty window. In this window the reading that are
set in the program were tested. The drone works like any other drone flying up to an
extent and capturing views. But, in this case the drone flies above the fire affected
areas and helps to put off the fire. This drone works with the help of controller where
signals are connected through a transmitter. The drone carries fire extinguisher balls
along with it and drops it from a certain height. The ball further rolls down to the
center of the fire and blasts where it puts off the fire this step can be followed until
the fire is completely put off, their efficient in terms of time to activate id done by
AFO ball. Time taken is 280.9 s to explode. The 0.7 kg size ball extinguished a circle
of 1-m diameter fire outbursts yet the 1.5 kg balls are assumed to extinguish a radius
of 1.5 m. There is only need of little help from man to control the drone. AWS was
connected to Raspberry Pi to capture the images from the drone.

4 Discussion

Consider the assumption that the UAVs could be re-loaded with a new method, it
could be possible to prevent fire outbursts from spreading in multiple directions, or
to extinguishing fires igniting the surrounding objects in large amount. The system
should be increased exponentially if there is relation between increases in swarm size
to the number of balls per UAV increase. The Table 1 depicts the idea of required
components.
Power will be 200 amps combination of continuous current and 400 amps peak
current. Motor power capability can pull up to the amps which is 25 amps continu-
ously without any disruptions. The thrust tests, each motor provides of 3 kg with 1867
carbon propellers, at 22.3 V, 24.1 amps. A ball-release mechanism was designed
and built as a mechatronic system attachable to the firefighting UAV. The system
consists of electronic components such as microcontroller, power supply, sender and
receiver, and motor as well as a mechanism to carry and release the ball and its
connections to the fire-fighting UAV. The ball-releasing mechanism was attached
with the mechatronic components.
1030 K. Gayathri Devi et al.

Table 1 Table captions

Component Description
should be placed above the
tables Frame Quadcopter frame 550 mm
Motors 2200 kV BLDC
ESCs 30 Amp
Flight controller Navio 2 + Raspberry Pi
GPS Navio 2
Telemetry 2.4 GHz transceiver
RC transmitter FS-i6S 2.4 GHz transceiver
Battery 2200 mAh 3S 20C Lipo
Propellers 10 × 5 carbon fiber
Raspberry pie Raspberry Pi 3
Thermal sensor DS18B20 temperature sensor
Camera Ardu-cam 5 MP 1080p sensor OV5847 mini
camera

5 Conclusion and Future Work

The project can be further improved by widening the obstacle detection range with
ultrasonic detectors. Further improvement can be made for this system. This system
can be improved with less latency and more accuracy. Deep learning algorithms and
Machine learning algorithms could be initiated. This proposed work with advanced
features reduces manual work and allows people to move to that place immedi-
ately and independently. The advancements in software side can reduce the latency
and produce high throughput. In the hardware side adding high resolution cameras
providing high sustainable circuit boards which have high durability. Therefore, real
time UAV is achieved by the implementation of web services in IOT.
These experiments were constructed and developed for a research project which
the main objective is to design a system of UAVs which can be automatic connected
to web server and monitored through internet and minimize the fire by extinguishing
the AFO balls above them. The idea can be simple if a single system of UAV could
drop them to optimal points, in optimal numbers, on time then there can be multiple
systems simultaneously working at same wavelength. Remote-sensing and image
capturing technology will be needed to determine points of attack and the web server
interfaced should be a dedicated one which cannot be used for other services. The
size of the fire-extinguishing balls depends upon the drones frame and the thrust it
can lift along with its own weight is a parameter that requires further investigation
for the usage of this system in real time applications. This is a promising result for
further studies that will be conducted for developing the proposed drone-assisted
fire-fighting system.
Automatic Firefighting System Using Unmanned Aerial Vehicle 1031

References

1. Rome NY (2015) Lockheed martin conducts collaborative unmanned systems demonstration;

Unmanned KMAX helicopter, Stalker XE Small Unmanned Aircraft System (UAS), and UAS
Traffic 25 Management (UTM) Systems Work Together to Extinguish Fire
2. Lee W, Kim S, Lee Y, Lee H, Choi M (2017) Deep neural networks for wildlife detection with
unmanned aerial vehicle. In: Proceedings of the IEEE international conference on consumer
electronics (ICCE). Las Vegas, NV, USA, pp 8–10
3. Martínez-de Dios JR, Merino L, Caballero F, Ollero A (2011) Automatic forest-fire measuring
using ground stations and unmanned aerial systems. Sensors 11(6):6328–6353
4. Holman RA, Katherine LB, Nicholas JS (2017) Surf zone characterization using a small
quadcopter: Technical issues and procedures. IEEE Trans Geosci Remote Sens 55(4)
5. Laurenti A, Neri A (1996) Remote sensing. Communications and information technologies for
vegetation fire emergencies. In: Proceedings of TIEMEC 96. Montreal, QC, Canada, pp 28–31
6. Beltran C (2013) Unmanned aerial vehicle with fire extinguishing grenade release and inspec-
tion system: 100% report (Doctoral dissertation, Dissertation, Florida International University.
https://mme.fiu.edu/wp-content/uploads/2013/12/F13-SR)
7. Krider EP, Noggle RC, Pifer AE, Vance DL (1980) Lightning direction-finding systems for
forest fire detection. Bull Am Meteor Soc 61(9):980–986
8. Kumar M, Cohen K, HomChaudhuri B (2011) Cooperative control of multiple uninhabited
aerial vehicles for monitoring and fighting wildfires. J Aerosp Comput Inf Commun 8(1):1–16
9. Wallar A, Plaku E, Sofge DA (2015) Reactive motion planning for unmanned aerial surveillance
of risk-sensitive areas. IEEE Trans Autom Sci Eng 12(3):969–980
10. Mois G, Folea S, Sanislav T (2017) Analysis of three IoT-based wireless sensors for
environmental monitoring. IEEE Trans Instrum Meas 66(8):2056–2064
11. Zhou H, Kong H, Wei L, Creighton D, Nahavandi S (2014) Efficient road detection and tracking
for unmanned aerial vehicle. IEEE Trans Intell Transp Syst 16(1):297–309
12. Yuan C, Zhang YM, Liu ZX (2015) A survey on technologies for automatic forest fire moni-
toring, detection and fighting using UAVs and remote sensing techniques. Can J Forest Res
45(7):783–792
13. Lin H, Liu, Z, Zhao T, Zhang Y (2014) Early warning system of forest fire detection based on
video technology. In: International symposium on linear drives for industry applications, pp
751–758
14. Sharifi F, Zhang YM, Aghdam AG (2014) A distributed deployment strategy for multi-agent
systems subject to health degradation and communication delays. J Intell Rob Syst 73(1–
4):623–633
15. Celik T, Demirel H (2009) Fire detection in video sequences using a generic color model. Fire
Saf J 44(2):147–158
16. Phan C, Liu HH (2008) A cooperative UAV/UGV platform for wildfire detection and fighting.
In: 2008 Asia simulation conference-7th international conference on system simulation and
scientific computing. IEEE, Santa Clara, CA, USA, pp 494–498
17. Howden D, Hendtlass T (2008) Collective intelligence and bush fire spotting. In: Proceedings
of the 10th annual conference on genetic and evolutionary computation. Atlanta, GA, USA, pp
41–48
18. Pham HX, La HM, Feil-Seifer D, Deans M (2017) A distributed control framework for a team
of unmanned aerial vehicles for dynamic wildfire tracking. In: 2017 IEEE/RSJ international
conference on intelligent robots and systems (IROS), pp 6648–6653
19. Yuan C, Liu Z, Zhang Y (2017) Aerial images-based forest fire detection for firefighting
using optical remote sensing techniques and unmanned aerial vehicles. J Intell Robot Syst
88(2–4):635–654
20. Alexis K, Nikolakopoulos G, Tzes A, Dritsas L (2009) Coordination of helicopter UAVs for
aerial forest-fire surveillance. In: Applications of intelligent control to engineering systems.
Springer, Dordrecht, The Netherlands, pp 169–19
State-of-the Art: Short Text Semantic
Similarity (STSS) Techniques in Question
Answering Systems (QAS)

Zaira Hassan Amur, Yewkwang Hooi, Irum Naz Sodhar, Hina Bhanbhro,
and Kamran Dahri

Abstract Semantics can be used to assess responses in question answering systems

(QAS). The responses are typically short sentences. Assessing short sentences for
similarity with the expected answer is a challenge for Artificial Intelligence. Unlike
long paragraphs, short texts lacks the adequate and accurate semantic information.
Existing algorithms don’t work well for short texts due to insufficient semantic infor-
mation. This Paper provides the state of art on semantic similarity techniques and
proposed the research framework to enhance the accuracy of short texts.

Keywords NLP · Text summarization · Semantic similarity · Short text ·

Information content · Information retrieval · Text assessment · Accuracy QAS ·
Semantic information

1 Introduction

Natural language processing the semantic similarity put its tremendous impact on
question answers systems. Whereas short answers/texts observed to be most import

Z. H. Amur (B) · Y. Hooi

Department of Management and Information Sciences, University Teknologi Petronas, Seri
Iskandar, Malaysia
e-mail: [email protected]
Y. Hooi
e-mail: [email protected]
I. N. Sodhar · H. Bhanbhro
Department of Information Technology Shaheed, Benazir Bhutto University, Shaheed Benazir
Abad, Pakistan
e-mail: [email protected]
H. Bhanbhro
e-mail: [email protected]
K. Dahri
Department of Information Technology, University of Sindh, Jamshoro, Pakistan
e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 1033
R. Ibrahim et al. (eds.), International Conference on Artificial Intelligence
for Smart Community, Lecture Notes in Electrical Engineering 758,
https://doi.org/10.1007/978-981-16-2183-3_98
1034 Z. H. Amur et al.

techniques. Which commonly recognized as STSS. These short texts include social
posts, question/answers, conversations, keywords and comments etc. In STSS short-
questions provide a useful understanding of specific concepts in a particular subject
domain. Short texts semantic similarity (STSS) helps to measure 10–20 words long.
Due to limited information these texts consider the most challengeable for number
of applications. These texts contains more than one meaning so they are generally
difficult to understand. Short texts contains lots of textual errors and produce result
ambiguity. These Texts generate semantic errors like grammatical and real world
errors. The main reason of these errors are user’s insufficient knowledge. Analyzing
and fixing these errors in short-texts improves the effectiveness of learning environ-
ment. Semantic calculation of short texts influenced the text words that causes the
poor accuracy [1]. Usually, understanding the hidden semantic meaning of word is
difficult task.
In natural language processing the text represents sentence. Each sentence/texts
has set of words, these words counted with different Parts of speech. Machines only
understand the language of symbols rather than text. To recognize the particular
and intended meaning of words or sentences it’s highly needed to convert text into
symbols. Part-of-Speech (POSs) determines that each word in the sentence contains a
specific sense of information. Generally machines don’t understand the dual meaning
of words/sentences. So therefore, natural language processing enables machines to
learn and seek the human natural language.
Text understanding plays a vital role in NLP. Most of the current methods are task
oriented such as information retrieval, relationship extraction and Q&A systems. The
main reason of using text in natural language processing make machines to know the
human languages. The aim of this paper is to summarize STSS techniques to identify
the trend of research.

1.1 Structure of the this Paper

The Paper is organized as follows: Section 2 is on literature study; Sect. 3 analyses the
various similarity approaches; Sect. 4 proposed the research framework and Sect. 5.
Gives the conclusion where section 6 provides the references of study.

2 Literature Review

Journal articles and proceeding papers from ISI and Scopus-indexed publications or
reputable institutions on STSS research in the past years have been carefully selected.
The papers are further filtered with focus on ML applications for A&Q systems. The
literature review identifies interesting development and trends of the study.
Refer to Table 1.1 in Attachment I. The table characterizes the STSS methods and
show the years and trends.
State-of-the-Art: Short Text Semantic Similarity (STSS) Techniques … 1035

It is important that short texts (subjective answers) must be understandable and

disambiguated so that learner can find the accurate information. However, Short text
similarity also related to textual entailment (TE) and paraphrasing techniques. Which
mostly use in many natural language processing tasks. These techniques differs
from each other in different context. TE (Textual entailment) uses the direct relation
among text fragments by using the hypothesis techniques, whereas paraphrasing
used to recognize the same meaning of text. These both techniques works on yes/no
decisions. STSS use these techniques to rate the semantic relatedness among groups
of words/sentences [1].
There are so many traditional methods used in NLP and the most common are,
Bag of words, Vector space model, and BM25 [2], helps to generate the words in
text. In NLP usually traditional methods are poor methods that can’t properly detect
the semantics of text at conceptual levels. Because of the limited amount of text, it
is not easy to calculate semantic similarity with these methods.
Several other methods like n-gram based, word-based methods, long based
distance so on and so forth are used to fixing the real word errors to improve the
accuracy for short answers [3]. There are some limitation in the existing intelligent
subjective questions. Same similarity often repeated among the number of questions
during the assessment process, increases the problem of duplication as well as low
accuracy [4].
This text similarity not only because of synonyms but the similarity can also be
possible because of antonyms or other similarity domains with the different learning
context. When text reduced it increases the difficulty for the semantic similarity.
Number of approaches have been based on traditional methods of test items and
words similarity [5, 6]. But these methods are not applicable to our problem domain.
Moreover, some algorithms contain high time complexity because it uses number
of semantic relations. One of study [7, 8] mentions the relationship diagram short
texts known as knowledge graph and random walk method to improve the concepts
of optimal word segmentation.
Another study [9] provides the attention neural network approach for reading the
text. This methods distinguish the text summarization and conceptualization semantic
information of text. However, the Paper [10] proposed a method that provides the
input transformation to set the optimal changes to the model architecture.
Meanwhile a study [11] introduces a new Q&A system for selecting grounded and
commonsense information from Concept another study [12] reviewed how to gain
new facts about the world based on knowledge graphs. Most of these methods are
task oriented. Although these methods have good expressiveness on particular tasks
but they still lack understanding and improving the accuracy of text semantics. Some
studies focuses on external knowledge to upgrade the quality of topic identification
in short texts.
The textual similarity has been proposed in 2006, where only small amount of text
were supposed to be included but after that the research becomes enhanced from short
to long and long to individual words. [1] It also works well in many web applications
like ontology generation, keyword extraction, and entity disambiguation. Like, Phan
et al. [13] find out the hidden Model. According Zhang [14], suggest that when two
1036 Z. H. Amur et al.

or more words are collocated, there must be one word that works as the center word
and the other word as the collocation word. Parts of speech works well to construct
the center word.
However, Kim [15] suggest that related features can also help with disambigua-
tion. Furthermore, they worked to capture semantic relations between terms using
the novel approach LDA, this approach helps to improve the accuracy of short text
conceptualization by using context semantics into consideration. Whereas POSs like
verbs, adjectives, and other attributes, can also help to detect instance ambiguity.
Moreover, the study [16] introduces the framework for short texts that detect the errors
from text. More specifically, the work has divided into three subtasks to understand
the short text such as text segmentation, type detection, and concept labeling.
Wang [4] present an approach to solve the problem of semantic similarity in test
papers, with the help of density entropy they selected the test papers to screen the
various question papers from item bank and then applies the calculation of semantic
similarity to detect the intelligent test paper from the corpus.
Furthermore, Huang [17] developed the algorithm to improve the performance of
STSS with low time complexity. This algorithm incorporated the different WordNet-
based word measures to address the word pairs with specific POSs that helps to
enhance the evaluation of semantic similarity of unknown words in shorts texts.
Hesham [1, 13] represented the work, that focusing one Semantic Textual Similarity
(STS) of a question pair. In this study they finds if two questions have the same
answers, then they are semantically equivalent.
Otherwise, if the answers are different to one another, or partially different then
(the answer of the first question is included in the answer of the second question but it
considered the incomplete answers, or the questions share a portion of their answers
yet different otherwise) then these answers subsequently considered the semantically
nonequivalent.
To compute semantic similarity for short texts is important in many areas. Many
approaches have been proposed that use linguistic analysis. These methods determine
whether the words in two short texts look alike, in terms of largest common substring
[18]. These approaches usually work for trivial cases.
This kind of methods creates a major problems, because the complex features are
derived from other NLP tools and components which use for other purposes. Some
application depends on parse trees that are completely restricted to syntactically
well-formed texts. More ever, the error from these tools also creates the problems in
Text semantic similarity models that are very hard to solve individually.
Furthermore Younes [19] proposes a technique for text summarization they
focused on identifying significant parts of text to create summary. To solve the
problem they represented an optimized algorithm that helps to produce summary
with the lexical chains and WordNet dictionary.
However, the Study [20], proposed methodology that consists on rules, based on
NLP algorithms like sentence splitting, tokenization and POS tagging. These rules
were applicable on user provided text which he want to summarized. In Algorithm
each word is assigned a score based on its frequency, this algorithm will run sentence
by sentence.
State-of-the-Art: Short Text Semantic Similarity (STSS) Techniques … 1037

There are number of techniques that uses Part-Of-Speech (POS) as syntactic

information to compute similarity between sentences. Subsequently, the work present
[21], gives the concept of weighting strategy based on POS. They represent the idea
that many POS and certain relation between POSs are more important and beneficial
than others. In particular they combine the weighting strategy with bag-of-words
approach to computer semantic (sentences) similarity in texts.

3 Techniques for Textual Similarity

In this section, this study proposes the methods that have potential for further devel-
opment and the rationals. There are various approaches that used to measure the
semantic similarity. This section represents the proposed method for extracting simi-
larity in short-text questions in detail. These approaches needs to acquire similarity
from semantic and syntactic information that exist in the compared natural language
sentences. In a natural language processing (NLP) a sentence can be defines as a set
of words with Part-of-speeches (POSs) [22] (Fig. 1).

3.1 Corpus/Statistical-Based Semantic Similarity

A large variety of the proposed approaches in words similarity or primarily corpus

based. In this approach valuable data is extracted from analyzing an enormous

Fig. 1 Techniques for measuring the short text semantic similarity (STSS) [33]
1038 Z. H. Amur et al.

corpora. Furthermore, analyzing words using Eq. 1 would allow the massive corpus
helps to extract the words similarity accordingly. Corpus-based approaches also
known as statistical-based approach. Corpus/statistical based similarities can be
measured by two.
Different techniques, normal statistical analysis (LSA) and the other one is deep
learning. A large corpora analyzed statistically through counting the words within
the corpus and documents. Tf-idf is a vital statistical object for the corpus analysis
which is used to employ as a word co-efficient [23].

3.2 Latent Semantic Analysis/Approach (LSA)

In this technique, words usually expressed as vectors. In LSA, the word matrix and
vectors constructed by analyzing the large text. The word matrix represents in the
form of rows and columns. In this corpus based analysis, usually cosine similarity
is highly used for calculating the word similarity text, tables and figures. Substantial
changes in content, e.g., new results, corrected values, title and authorship, are not
possible and can-not be processed.

3.3 Deep Learning Analysis/Approach

This approach also known as word embedding technique which used to find the
words semantically. In the semantic space, a very big corpus is required to extract
the word representation. This technique is specifically used to guess the word from
the given surrounding words, usually known as bag of words model. This model also
helps to understand the vector representation. Deep learning also use the Skip gram
method to understand the context meaning of word. It also helps to understand the
word/vector semantically. For example if we perform some calculation on vectors
semantically,

[VECT (KING) − VECT (MAN) + VECT (Woman)]

The result of this vector shows the word “queen” which is generally very close
to the specific semantic calculation of word representation. Deep learning also use
web based similarity.

3.4 String-Based Similarity Approach

String-based also knowns as lexical based word similarity. It uses the

comparison technique between the sequences of characters. There are various
State-of-the-Art: Short Text Semantic Similarity (STSS) Techniques … 1039

methods/approaches used in String-based similarity to calculate the similarity of

words like, levenshtein distance, longest common sub string Q-gram and jaccard
distance. As for as LD, it computes the distance between two strings by using different
operations like, insert, delete or replace. And change the one form of string to another.
Whereas, longest common substring (LCS) finds the longest substring of string.
Example

LC Subtr (S1 , S2 ) = max LC Su f f (S11 ...i, S2... j),

1<i< m .1< j≤n

3.5 Topological/Knowledge Based

These methods works well to understand the meaning of word to process compu-
tationally. It actually use the WordNet for the lexical relationships. There are three
methods that helps WordNet to extract the similarity. These can be categorized as
node-based/Ic, edge-based and Glass-based measure.

3.6 Node-Based/Ic (Information Content)

This approach uses concepts to extract the similarity. This measure includes the
hierarchical structure. In this method every node uses IS-A taxonomy that are kept
in on set. There are also other measures that considers the node/Ic based measures
for example, RES, LIN and JCN [1]. Before implementing the node based model the
concepts can be divide into words(c) and classes defined as classes (w) [24].

classes (w) = {c/w ∈ wor ds (c)}.

Here in the whole hierarchy the class works as a subtree and the class (w) contains
all sets of possible words [24].

3.7 Edge/Distance Based Measures

This is the direct approach for computing the similarity. This technique counts the
edges of those nodes that are correspond to each other. Distance based measures used
the hierarchal concept for WordNet where the PATH length can be used to measure
the similarity There are three other approaches based on Distance measures PATH,
WUP, LCH [1].
1040 Z. H. Amur et al.

3.8 Gloss Based Measure

These measures uses the glosses between the concepts of WordNet. Vector and LESK
are the type of this measure [1]. Table 1 Shows the Summary of various methods and
dataset used for the semantic similarity of short texts. This table presents the latest
work. Various datasets has been used by the different approaches the most common
is Microsoft paraphrase corpus. This dataset contain 5801 sentence pairs that can
extract from news, documents and various social network platforms.

4 Proposed Research Framework

This study proposed Research framework (Fig. 2) that has one independent variable
(Text assessment) whereas the other’s work as dependent variables. In this Frame-
work the text assessment will be done by the student’s essays. The evaluation will be
based on examiners or expert essays that will be available to check the relatability
among essays. These short texts divided into text extraction process that actually
extract the core keyword from text through the machine learning process. Further
the machine normalize the text through stemming and lemmatizing techniques to
remove the suffixes. In summarization the machine performs the text prepressing
that helps to reduce the noise from text to detect the stop words or punctuation
marks from Texts. Moreover the Intent recognition check the similarity among the
short texts to identify the relatable Keywords that has the highest similarity. This
technique helps to give the rank to the text (answers). Whereas (Table 2) Text assess-
ment, summarization, as well as similarity all these effects on the accuracy of short
texts. The measurable variables in framework are: Text Assessment, Text extraction,
intent recognition and accuracy.
In Q &A systems usually the Reponses works independently. So if the student’s
and examiner’s responses relate with each other than machine will understand the
context of word/sentence and indicates the predicate feature that has the higher
similarity. If the predicate feature has high similarity then the accuracy of STSS will
increase. If the predicate or candidate word failed to provide the accurate response
then the accuracy of answer will not match to the specific answer.

5 Conclusion and Future Possibilites

The main consideration of this study is to give brief overview on short texts. There are
various approaches like string similarity, corpus based or knowledge based. Usually
these approaches provide good impact on text representation. From the literature and
previous work we have identified and noticed that corpus based and knowledge based
measure are most suitable measures for text summarization. Corpus based measure
Table 1 Summary of various methods and datanet used for the Semantic similarity of short texts
Research Methods
WordNet LCS Word embedding Structure based Vector Based PMI-IR LSA Base method Dataset
Huang et al. [3], ✓ ✓ ✓ String Based MSRPar(Microsoft
2019 paraphrase corpus)
Islam et al. [25], ✓ ✓ ✓ Combined (String & MSRPar(Microsoft
2009 Corpus based) paraphrase corpus)
Zhao et al. [26], ✓ ✓ Corpus based Gigaword &
2019 DUC-2004
Suleiman et al. [27], ✓ ✓ Corpus based OSAC
2019
Pawar et al. [28], ✓ ✓ Knowledge based PILOT
2018
Hien et al. [29], 2019 ✓ ✓ Corpus based MRPC
P4PIN.STS2015
Cai et al. [30], 2020 ✓ ✓ ✓ String & knowledge M&C, R&G, WS-
State-of-the-Art: Short Text Semantic Similarity (STSS) Techniques …

based 353
Zhiguo et al. [30], ✓ Corpus based QASent, MSRP
2016
Avani et al. [28], ✓ ✓ Corpus based Kaggle
2019
1041
1042 Z. H. Amur et al.

Fig. 2 Research framework

Table 2 Variables of proposed research framework

Variables Independent variable Dependent variable Technique (Measurable)
Student’ versus ✓ Text assessment
examiner’s essay
Summarization ✓ Text extraction
Similarity ✓ Intent recognition
Short Text ✓ Accuracy

provides the advantage of cosine similarity. These approaches helps to minimize the
risk of errors and improve the chances of accuracy between short texts. Morever, this
study proposed the research Framework that helps to find out the similarity among
short texts (answers). For the future work, there is a lot of work yet to be done on
unsupervised learning in semantic similarity.

References

1. Huang PS, Chiu PS, Chang JW, Huang YM, Lee MC (2019) A study of using syntactic cues
in short-text similarity measure. J Internet Technol 20(3):839–850
2. Jimenez S, Cucerzan SP, Gonzalez FA, Gelbukh A, Dueñas G (2018) BM25-CTF: improving TF
and IDF factors in BM25 by using collection term frequencies. J Intell Fuzzy Syst 34(5):2887–
2899
State-of-the-Art: Short Text Semantic Similarity (STSS) Techniques … 1043

3. Song X, Min YJ, Da-Xiong L, Feng WZ, Shu C (2019) Research on text error detection and
repair method based on online learning community. Procedia Comput Sci 154:13–19
4. Wang H, Yang W (2019) An intelligent test paper generation method to solve semantic similarity
problem. In: IOP conference series: earth and environmental science Vol 252, No 5. IOP
Publishing, p 052126
5. Gashaw I, Shashirekha HL (2020) Machine learning approaches for amharic parts-of-speech
tagging. arXiv preprint arXiv:2001.03324
6. Croft D, Coupland S, Shell J, Brown S (2013) A fast and efficient semantic short text similarity
metric. In: 2013 13th UK workshop on computational intelligence (UKCI) (pp. 221–227).
IEEE
7. Chen HH, Lin MS, Wei, YC (2006) Novel association measures using web search with double
checking. In: Proceedings of the 21st international conference on computational linguistics and
44th annual meeting of the association for computational linguistics, pp 1009–1016
8. Martinez-Gil J, Pichler M (2014) Analysis of word co-occurrence in human literature for
supporting semantic correspondence discovery. In: Proceedings of the 14th international
conference on knowledge technologies and data-driven business, pp 1–7
9. Wang H, Zeng M, Xiong Z, Yang F (2017) Finding main causes of elevator accidents via
multi-dimensional association rule in edge computing environment. China communications
14(11):39–47
10. Radford A, Narasimhan K, Salimans T, Sutskever I (2018) Improving language understanding
by generative pre-training
11. Bauer L, Wang Y, Bansal M (2018) Commonsense for generative multi-hop question answering
tasks. arXiv preprint arXiv:1809.06309
12. Nickel M, Murphy K, Tresp V, Gabrilovich E (2015) A review of relational machine learning
for knowledge graphs. Proc IEEE 104(1):11–33
13. Phan XH, Nguyen LM, Horiguchi S (2008) Learning to classify short and sparse text & web
with hidden topics from large-scale data collections. In: Proceedings of the 17th international
conference on World Wide Web, pp 91–100
14. Tachbelie MY, Abate ST, Besacier L (2011) Part-of-speech tagging for underresourced and
morphologically rich languages—the case of Amharic. HLTD 2011:50–55
15. Song Y, Wang H, Wang Z, Li H, Chen W (2011) Short text conceptualization using a prob-
abilistic knowledgebase. In: Proceedings of the twenty-second international joint conference
on artificial intelligence-volume volume three, pp 2330–2336
16. Kim D, Wang H, Oh A (2013) Context-dependent conceptualization. In: Proceedings of the
twenty-third international joint conference on artificial intelligence, ser. IJCAI’13. AAAI Press,
pp 2654– 2661. [Online]. Available: http://dl.acm.org/citation.cfm?id=2540128 2540511
17. Hua W, Wang Z, Wang H, Zheng K, Zhou X (2015) Short text understanding through lexical-
semantic analysis. In: 2015 IEEE 31st international conference on data engineering. IEEE, pp
495–506
18. Al-Bataineh H, Farhan W, Mustafa A, Seelawi H, Al-Natsheh HT (2019). Deep contextualized
pairwise semantic similarity for Arabic language questions. In: 2019 IEEE 31st international
conference on tools with artificial intelligence (ICTAI). IEEE, pp 1586–1591
19. Slam A, Inkpen D (2008) Semantic text similarity using corpus-based word similarity and
string similarity. ACM Trans Knowled Discov Data (TKDD) 2(2):1–25
20. Jaafar Y, Bouzoubaa K (2018) Towards a new hybrid approach for abstractive summarization.
Procedia comput sci 142:286–293
21. Vuong La CS (2019) Artificial intelligence versus natural stupidity: evaluating AI readiness
for the vietnamese medical information system. J clin med, 8(2):168
22. Otter DW, Medina JR, Kalita, JK (2020) A survey of the usages of deep learning for natural
language processing. IEEE Tran Neural Netw Learn Syst
23. Farouk M (2019) Measuring sentences similarity: a survey. arXiv preprint arXiv:1910.03940
24. Majumder G, Pakray P, Gelbukh A, Pinto D (2016) Semantic textual similarity methods, tools,
and applications: a survey. Computación y Sistemas 20(4):647–665
1044 Z. H. Amur et al.

25. Islam A, Inkpen D (2009) Semantic similarity of short texts. Recent Advances in Natural
Language Processing V 309:227–236
26. Zhao F, Quan B, Yang J, Chen J, Zhang Y, Wang X (2019) Document summarization using
word and part-of-speech based on attention mechanism. In: Journal of Physics: Conference
Series, Vol 1168, No 3. IOP Publishing, p 032008
27. Suleiman D, Awajan AA (2019) Using part of speech tagging for improving word2vec model.
In: 2019 2nd International Conference on new trends in computing sciences (ICTCS). IEEE,
pp 1–7
28. Wang Z, Mi H, Ittycheriah A (2016) Sentence similarity learning by lexical decomposition and
composition. arXiv preprint arXiv:1602.07019
29. Sakhapara A, Pawade D, Chaudhari B, Gada R, Mishra A, Bhanushali S (2019) Subjective
answer grader system based on machine learning. In: Soft computing and signal processing.
Springer, Singapore, pp 347–355
30. Cai Y, Pan S, Wang X, Chen H, Cai X, Zuo M (2020) Measuring distance-based semantic
similarity using meronymy and hyponymy relations. Neural Comput Appl 32(8):3521–3534
31. Tsatsaronis G, Varlamis I, Vazirgiannis M (2010) Text relatedness based on a word thesaurus.
J Artif Intell Res 37:1–39
32. Jin O, Liu NN, Zhao K, Yu Y, Yang, Q (2011) Transferring topical knowledge from auxiliary
long texts for short text clustering. In: Proceedings of the 20th ACM international conference
on information and knowledge management, pp 775–784
33. Han M, Zhang X, Yuan X, Jiang J, Yun W, Gao C (2021) A survey on the techniques, applica-
tions, and performance of short text semantic similarity. Concurrency Comput Pract Experience
33(5):e5971
Modeling of Wireless Power
Transmission for Pacemaker Battery
Charging in MATLAB Simulink
Environment

G. P. Kiruthiga, Mohdfakhizan B. Romlie, K. Porkumaran, and S. Prabakar

Abstract Cardiovascular pacemaker is a little gadget that utilizes electrical driving

forces conveyed to the terminal by contracting heart muscles and they can direct
the thumping of the heart. It requires the energy where ordinary wires are badly
arranged and dangerous to human body. When the life expectancy of the pacemaker
battery is finished, continued changing of pacemaker will influence the patients.
To build the lifetime of the implantable heart pacemaker. By utilizing remote force
transmission for cardiovascular pacemaker depends on inductive coupling between
the two curls has been proposed in this paper. The essential curl is to be put on the
body, while auxiliary loop is to be set inside the body. From this, sending and getting
loop can move power remotely, and beneficiary circuit will get the air conditioner
sign and convert it into DC signal for the battery of the cardiovascular pacemaker.
The reproduction results show that the moved productivity is about 80.01% at the
distance 2 cm.

Keywords Wireless power transmission · Cardiac pacemaker battery · Inductive

couplin

G. P. Kiruthiga (B)
Dr.N.G.P Institute of Technology, Kalapatti road, Coimbatore, India
M. B. Romlie
UniversitiTeknologi Petronas, Seri Iskandar, Perak, Malaysia
e-mail: [email protected]
K. Porkumaran
Sri Sairam Engineering College, Chennai, India
S. Prabakar
Sona College of Technology, Salem, India
e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 1045
R. Ibrahim et al. (eds.), International Conference on Artificial Intelligence
for Smart Community, Lecture Notes in Electrical Engineering 758,
https://doi.org/10.1007/978-981-16-2183-3_99
1046 G. P. Kiruthiga et al.

1 Introduction

These days the world without electric force is troublesome thing to live. For instance,
the day-day items like electronic vehicle, versatile, PC, clothes washer and so on,
requires electric force. Like that force move through the little attachment containing
numerous wires may cause fire mishaps and demise [1]. Pacemaker is only it is a little
gadget that utilizes electrical motivation conveyed to the terminal by contracting heart
muscles after that they can direct the pulsating of the heart [5]. In 1958, the pacemaker
was embedded first with battery-powered battery made of nickel cadmium. In 1973,
lithium battery was utilized to control the pacemaker. A pacemaker is put in the chest
(just underneath the collarbone) to help manage moderate electrical issues with the
heart.
Pacemaker leads might be situated in the chamber (upper chamber) or ventricle
(lower chamber) or both, contingent upon the clinical condition [6]. Inherent coro-
nary illness in conceived babies needs to the implantation of pacemaker from their
introduction to the world to whole lifetime. And furthermore utilized in mature age
individuals to aid appropriate working of heart. Boundaries of pacemaker are; capac-
itor, timing circuit, terminal, beat generator, battery. In this paper we zeroed in on to
charge the battery.

1.1 Battery Technology

In pacemaker battery generally they are utilizing Lithium batteries in light of longer
life expectancy and 5 V has been utilized, battery substitution is needed after 3.3 V.
Mercury batteries likewise utilized however 3–5 cells are utilized to make one
mercury battery, it makes costly. Natural force source battery is as yet in explo-
ration not yet emerged. Plutonium based batteries were utilized in times past, and
its life expectancy around 30 years however plutonium is atomic receptive and
undependable, harmful subsequently not utilized.

1.2 Wireless Power Transmission

Remote Power Transfer (WPT) is a strategy for sending the electrical energy starting
with one point then onto the next in a remote medium (without actual association).
Two sorts in WPT,
• Near field it is non radiative and the force can be communicated to the short
distance and inductive coupling is utilized.
• Far field it is radiative and the force can be communicated to the significant
distance and capacitive coupling is utilized. For remote force transmission close
documented is utilized on the grounds that it is non radiative (Fig. 1).
Modeling of Wireless Power Transmission for Pacemaker Battery … 1047

Fig. 1 Basic wireless power

transmission

Table 1 Battery technology

Battery type Life span
Nickel Cadmium 7 years
Mercury battery 10 years
Biological power source Still in research
Lithium Iodide 10—12 years

From the above circuit when the power can be feed into the transmitter side then
by through the coupling devices the power can be goes to the receiver side and then
voltage is transferred to the battery. This is the basic wireless power transmission [7]
(Table 1).

1.3 Current Issue

Once the life span of battery is over, repeated changing of pacemaker would affect
the patients in these ways,
• Re-implantation of whole device cause defect to the surrounding tissues.
• Frequent surgeries for every 10 years interval for the new born babies with
congenital heart disease may lead to weakness of the patient body.
• As well as senior citizens who are above 60 years cannot opt surgery method
frequently [4].
To overcome the above issues, we adopted wireless power transmission to charge
the cardiac pacemaker can be the alternative for surgery. In this paper we designed
a simple model using MATLAB Simulink software for pacemaker wireless charger
that constitutes amplifier, oscillator, inductive power transfer coupler, rectifier, LC
filter and resistance load which represents pacemaker battery.
1048 G. P. Kiruthiga et al.

2 Proposed Work

In the wireless power transmission, the output circuit consists of a coil, a transmitter
which transmits the power and receiver receives the power and send it to the battery.
Whereas the transmitter side consist of an input voltage that is an AC voltage source,
amplifier and the oscillator. The receiver circuit consists of rectifier, monitoring
circuit and battery. All my simulation and output were generated using MATLAB
(Fig. 2).

2.1 Amplifier

From the transmitter side foremost component is class A amplifier with Darlington
pair. Moreover, in biomedical devices, the most commonly used type of amplifier is
the class A amplifier, it is the simplest form of power amplifier during one complete
cycle of input signal waveform producing minimum distortion and maximum ampli-
tude of the output signal. Amplifier is used to increase the power of the signal.
The current gain of the two-transistor individual is multiplied together along with
collector currents are compared to the single transistor circuit (Fig. 3).
From the above circuit the input power 5 V and the frequency is 50 Hz then the
supply goes to the capacitor C1 and value is 0.22 µF and it reaches the resistor
R1, R2 and value is 2000, 7250 k and then supply goes to the Darlington pair
of collectors and the process continues along the Darlington pair and the amplified
voltage is approximately 9 V (Fig. 4).

Fig. 2 Proposed work flow

Modeling of Wireless Power Transmission for Pacemaker Battery … 1049

Fig. 3 Class A amplifier with Darlington pair

Fig. 4 Output for class A amplifier

2.2 Oscillator

The acceptable range of frequency for a human body is 300 kHz and 3.55 MHz.
Hence the 5 V is given as the input voltage and obtaining the frequency in the range
300 kHz after oscillation (Fig. 5).
After we get the amplified voltage form the amplifier it reaches the oscillator. The
oscillator can oscillate the frequency. In the above circuit 555 timer is an 8 pin IC.
Pin 2 trigger is connected to the capacitor and the value is 0.1 nF, then the discharge
pin is said as pin 7 and pin 4 is said as reset pin, this is connected to the input DC
voltage source which is 5 V and control voltage pin is pin 5, which is grounded with
capacitor value 0.1 nF. We are using the Astable mode 555 timer as an oscillator and
1050 G. P. Kiruthiga et al.

Fig. 5 Oscillator by using 555 timer

it can generate the pulse. Generally, the Astable mode generates a square wave, in
order to obtain a Sine wave, we add L resonance of value 100 µH and C of value
2.5 nF along with the Astable mode 555 timer. The obtaining frequency range after
simulation is about 318 kHz (Fig. 6).

f = 1 ÷ 2π )Lc

Then substituting the values for above equation then the frequency is about, f =
318 kHz.

2.3 Receiver Side

The output from the oscillator goes the capacitor which is connected parallel. The
AC voltage from the capacitor goes to the transmitter coil and the consists of mutual
inductance L1, L2 = 95.7 nH, by through the magnetic coupling signal goes to the
receiver coil. In the receiver side rectifier is present, here half rectifier is used, because
Modeling of Wireless Power Transmission for Pacemaker Battery … 1051

Fig. 6 Output for oscillator

the secondary coil is placed inside the heart so component should not be complex.
Full wave rectifier is not used here because it is still in research. Rectifier is used
to convert alternating current into direct current. To the charge the battery, the input
will be in DC. The output from the rectifier is appeared as pulsated DC, for charging
of the battery pulsated DC is enough. By adding LC filter pure DC is formed. This
is more appropriate compared to pulsated DC (Fig. 7).

Fig. 7 Half wave rectifier with LC filter

1052 G. P. Kiruthiga et al.

3 Results and Discussion

The above circuit mentioned is the overall circuit. When the input here is DC of
about 5 V then it goes to the amplifier circuit then it amplifies voltage about 9 V
then it goes to oscillator, and it converts DC voltage into AC voltage and it shapes
the frequency which is about 318 kHz and it goes to the transmitter coil through
the coupling device, it goes to the receiver coil, in that rectifier converts AC to DC
voltage. After rectification the output voltage is about 4 V, in spite of the loss occurred.
The simulation results show that the transferred efficiency is about 80.01% at the
distance 2 cm (Figs. 8 and 9).

Fig. 8 Overall circuit

Fig. 9 Overall output for WPT

Modeling of Wireless Power Transmission for Pacemaker Battery … 1053

4 Future Work

More experimental analysis should be done with different background of cases in

order to ensure the quality and effectiveness.

References

1. Gore VB, Gawali DH (2016) Wireless power transfer technology for medical applications
2. Peng K, Liu X, Huang P Study on the wireless energy supply system in the implantable cardiac
pacemaker
3. Campi T, Cruciani S, Palandrani F, De Santis V, Hirata A, Feliziani M Wireless power transfer
charging system for AIMDs and pacemakers
4. Lafrance A Who killed the rechargeable pacemaker? The Atlantic, 4 Feb 2014
5. Ali KK, Jarndal AH Remote-monitoring and wireless-powering system for pacemaker patients
6. Khan MU, Jafar A, Karimov KhS, Feroze S A proposed optimized solution for wireless power
transfer using magnetic resonance coupling
7. Hu Y, Zhang Y, Xu C, Lin L,Snyder RL, Wang ZL Self-powered system with wireless data
transmission
8. Vulfin V, Sayfan-Altman S, Ianconescu R Wireless power transfer for a pacemaker application
9. https://en.wikipedia.org/wiki/Wireless_power_transfer
10. Sung SH, Kim YS, Joe DJ, Mun BH, You BK, Keum DH, Hahn SK, Berggren M, Kim D,
Lee KJ Flexible wireless powered drug delivery system for targeted administration on cerebral
cortex
11. https://en.wikipedia.org/wiki/Artificial_cardiac_pacemaker
12. https://www.hopkinsmedicine.org/healthlibrary/test_procedures/cardiovascular/pacemaker_
insertion_92, p 07980
Recent t-way Test Generation Strategies
Based on Optimization Algorithms:
An Orchestrated Survey

Ammar K. Alazzawi, Helmi Md Rais, Shuib Basri, Yazan A. Alsariera,

Abdullahi Abubakar Imam, Saad Adnan Abed,
Abdullateef Oluwagbemiga Balogun, and Ganesh Kumar

Abstract In software testing, test case generation is the most challenging activities
and expensive tasks. Where has a considerable impact on the ability to produce
a desired or intended result (i.e., quality and efficacy) of software testing. As a
result, several researchers have developed a number of t-way test case generation
strategies (where t points the interaction strength between parameters) due to the
market demand to the various types of tests based on different approaches. This paper
presents an orchestrated survey of the latest test case generation strategies such as
Binary Black Hole (BBH), Sine Cosine Variable Strength (SCAVS), Combinatorial
Testing Based Jaya Algorithm (CTJ), deterministic genetic multi-parameter-order
(GAMIPOG) and Hybrid Artificial Bee Colony (HABC). This survey illustrates the
strengths and weaknesses of each strategy, and indicates potential research studies
in the field for future work.

Keywords Meta-heuristics · Combinatorial testing · Optimization algorithms ·

t-way testing · Software testing

1 Introduction

The researchers in the software testing have sophisticated a number of helpful tech-
niques in order to avoid bugs and promote bug detection the last few decades.
Although the sophisticated techniques are useful. However, the manufactured soft-
ware produced has no bug-free guarantee [1]. Thus, the only procedure that can
demonstrate the achieved quality and recognize the remaining problems, and risks
is software testing.

A. K. Alazzawi (B) · H. M. Rais · S. Basri · A. A. Imam · S. A. Abed · A. O. Balogun · G. Kumar

Department of Computer and Information Sciences, Universiti Teknologi PETRONAS, Bandar
Seri Iskandar, 32610 Perak, Malaysia
e-mail: [email protected]
Y. A. Alsariera
Department of Computer Science, Northern Border University, Arar 73222, Saudi Arabia

© The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 1055
R. Ibrahim et al. (eds.), International Conference on Artificial Intelligence
for Smart Community, Lecture Notes in Electrical Engineering 758,
https://doi.org/10.1007/978-981-16-2183-3_100
1056 A. K. Alazzawi et al.

In spite of the exhaustive testing, is desirable, but often is impossible due to

limitations on time and resources. As a result, a new sampling technique named t-way
test case generation based on combinatorial testing (where t points the interaction
strength between parameters) has begun to appear as an alternative to exhaustive
testing to systematically reduce the test cases to a manageable one. For example,
BTS [2–4], GS [5], IPOG families [6–8], PSTG [9] and ABCVS [10, 11]). In the last
20 years, several new t-way test case generation strategies have developed in order
to find the most efficient strategies are able to produce optimal test cases for each
configuration system (i.e., every single combination covered at most once a time).
The t-way test case generation strategies based on meta-heuristics have gained
a huge deal of attention in software testing as part of a general interest in recent
years. In spite of, is desirable, most of the current existing t-way test case generation
strategies do not support the Input output relations, variable interaction strength,
constraint, high interaction strength t > 3 and etc. This paper is an attempt to survey
the current t-way test case generation strategies and collect the knowledge (i.e., the
strengths and weaknesses) of some current strategies that published recently.
The rest of paper structured as follows. Sect. 2 describes the theoretical back-
ground of t-way testing. The t-way test case generation strategies based on
meta-heuristics in Sect. 3. Finally, the conclusion given by Sect. 4.

2 Theoretical Background of t-way Testing

Every configuration system under test (CSUT) is consists of a different set of elements
called parameters (p), and each parameter is associated with different values (v).
Table 1 shows a simple software system to easily understand the interaction between
parameters and the associated values (i.e., t-way testing). Where this software system
consists of 3-parameters associated with 2-values each, 1-parameter associated with
3-values each and the interaction strength t = 2. This software system can be
expressed as covering array CA (N; 2, 23 31 ).
The exhaustive test suite size in this software system is 24 test cases (3 × 2×2
× 2 = 24) in order to test this system exhaustively as shown in Fig. 1. By executing
2-way testing, where interaction strength between parameter’s t = 2. The possible
combinations are 6 combinations between parameters (AB, AC, AD, BC, BD and
CD) respectively. The optimal test suite size achieved is 6 test cases (See Fig. 1) that

Table 1 Simple software system

Input parameters
A B C D
System configurations 0 0 0 0
1 1 1 1
2
Recent t-way Test Generation Strategies Based on Optimization … 1057

Fig. 1. 2-way test suite generation

able to cover each interaction between parameters. Therefore, t-way testing achieved
a minimization from 24 test cases to 6 test case only (75% of total exhaustive testing).

3 Current t-way Test Case Generation Strategies

The generation of optimal test suite size can be considered as an optimization problem
[12]. Test case generation strategies based meta-heuristics (i.e., optimization algo-
rithm) always outstanding in this regard. Numerous of the existing meta-heuristics
algorithms based strategies are using the one-test-at-a-time (OTAT) approach exclu-
sively. In general, all the test case generation strategies based meta-heuristics generate
non-deterministic solutions due to the randomization. In comparison, most studies
are limited to small interaction intensity values for pure computation-based equiva-
lents (e.g., 2 ≤ t ≤ 6). In addition, there was also inadequate investigation of support
for variable strength and constraints.
Binary Black Hole (BBH) is one of the most recently developed test case gener-
ation strategy based on meta-heuristic algorithms called Black Hole (BH) algorithm
[13, 14]. The inspiration of BBH strategy came by the interaction of the phenomenon
of the black hole and the star’s behaviour together. In case of the star become close
to the black hole, the black hole will swallow it. Where the algorithm will begins to
generate a new solution (i.e., new star) randomly and will be searched again by the
algorithm. BHH is an attractive strategy to solve the t-way testing problem because
it is a tuning-free. BHH does not need multiple parameters for balancing between
the global and local search like other generation strategies based on meta-heuristics.
These features make it simple and easy for implementation. Up until now, BBH
address small values of interaction strength (i.e., t ≤ 4). In fact, BBH do not address
variable strength interactions and do not provide the support for constraints.
1058 A. K. Alazzawi et al.

Sine Cosine Variable Strength (SCAVS) strategy [15] based on Sine Cosine Algo-
rithm (SCA) Algorithm for t-way variable test suite generation. SCA is meta-heuristic
based on exploit the mathematical properties of both sine and cosine trigonometric
functions. SCAVS is as any other algorithm starts first to produce random solutions.
SCAVS supports the interaction strength up to t = 6 and addressing variable strength
interaction. Although SCAVS addressing variable strength interaction, SCAVS does
not support the constraints.
Hybrid Artificial Bee Colony (HABC) strategy [16–19] is an extended version
based on Artificial Bee Colony (ABCVS) [10, 11] strategy in order to overcome the
limitations of the original Artificial Bee Colony (ABC) algorithm. HABC based on
hybridization of two meta-heuristic algorithms that can help the search capabilities of
each by compensating for the limitations of one algorithm with the strength of others.
HABC strategy based on merging the advantages of the original Artificial Bee Colony
(ABC) algorithm with the advantages of an original Particle Swarm Optimization
(PSO) algorithm. Unlike other strategies, HABC does support the high value of
interaction strength (i.e., t ≤ 6) and addresses the variable strength interactions.
HABC also addresses the support for constraints through its variant, called PhABC
[20, 21].
Unlike BBH, SCAVS and HABC that generate non-deterministic solutions,
GAMIPOG is deterministic strategy. Deterministic Genetic Multi-Parameter-Order
(GAMIPOG) strategy [22], based on Modified Input Parameter Order General
(MIPOG) for test case generation [23]. GAMIPOG overcomes many desired chal-
lenges by combined them together in one strategy in order to take the benefits of
each one such as meta-heuristics, one-parameter-at-a-time and one-test-at-a-time. In
addition, GAMIPOG comparing to other existing strategies have less complexity,
manageable test size and fast execution time. Where GAMIPOG works to find the
global solution by breaking the problem into sub-problems merging all solution’s
together iterative steps. Thus far, GAMIPOG addresses high values of interaction
strength (i.e., t ≤ 15) and addresses the variable strength interactions. Furthermore,
GAMIPOG does not provide the support for constraints.
Unlike all the above-mentioned strategies, CTJ is support input–output based
relation (IOR) for test case generation [24]. Combinatorial Testing based on Jaya
Algorithm (CTJ) strategy is one of the recently proposed strategies that support
input–output based relation (IOR) for test case generation. Where CTJ adopts the
Jaya algorithm as the main engine for generating the solutions. CTJ has two common
control parameters: (1) maximum iteration and (2) population size. This feature
makes CTJ performance to produce the solutions easily than other existing strategies.
Thus far, CTJ addresses small values of interaction strength (i.e., t ≤ 3) and addresses
the input–output based relation (IOR). Furthermore, CTJ does not provide the support
for constraints and variable strength interactions.
In conclusion, Table 2 gives a brief statement of the current t-way variable test
generation. Referring to Table 2, all the existing strategies are non-deterministic due
to the randomization (i.e., different test suite size in every implementation) except
GAMIPOG strategy. There was a lack of investigation of constraints support, input
output relations and as well as variable strength. GAMIPOG is the only strategy
Recent t-way Test Generation Strategies Based on Optimization … 1059

Table 2 Current t-way test case generation strategies

Strategies
Non-deterministic Deterministic Interaction Variable Constraints Input–output
strength strength relations
√ √
SCAVS X t≤6 X X
√ √
CTJ X t≤3 X X
√ √
GAMIPOG X t ≤ 15 X X
√
BBH X t≤4 X X X
√ √ √
HABC X t≤6 X

support the addressing of high interaction strength (t ≤ 15). Unlike GAMIPOG,

SCAVS and HABC support the interaction strength up to t ≤ 6. Unlike existing
strategies, HABC only has the ability to support constraints. All strategies support
the variable interaction strength, except CTJ and BBH. In contrast, CTJ addresses
the input output relations unlike the other existing strategies.

4 Conclusions

This paper summarizes the current existing t-way test case generation strategies based
on meta-heuristics and emphasises the strengths and weaknesses of each strategy. The
software market is massive, and therefore, the test is needed to ensure compliances
are varied; so new test ideas and test suite development are still needed. There are
also numerous other meta-heuristics algorithms that still do have good potential and
have not been adopted for t-way test suite generation.

References

1. Alsewari AA, Alazzawi AK, Rassem TH, Kabir MN, Homaid AAB, Alsariera YA, Tairan NM,
Zamli KZ (2017) ABC algorithm for combinatorial testing problem. J Telecommun Electron
Comput Eng (JTEC) 9:85–88
2. Alsariera YA, Zamli KZ (2015) A bat-inspired strategy for t-way interaction testing. Adv Sci
Lett 21:2281–2284
3. Alsariera YA, Majid MA, Zamli KZ (2015) Adopting the bat-inspired algorithm for interaction
testing. In: The 8th edition of annual conference for software testing, pp 14
4. Alsariera YA, Majid MA, Zamli KZ (2015) A bat-inspired strategy for pairwise testing. ARPN
J Eng Appl Sci 10:8500–8506
5. Esfandyari S, Rafe V (2018) A tuned version of genetic algorithm for efficient test suite
generation in interactive t-way testing strategy. Inf Softw Technol 94:165–185
6. Lei Y, Kacker R, Kuhn DR, Okun V, Lawrence J (2007) IPOG: a general strategy for t-way
software testing. In: Engineering of computer-based systems, 2007. ECBS’07. 14th annual
IEEE international conference and workshops on the, IEEE, pp 549–556
7. Lei Y, Kacker R, Kuhn DR, Okun V, Lawrence J (2008) IPOG/IPOG-D: efficient test generation
for multi-way combinatorial testing. Softw Testing Verif Reliabil 18:125–148
1060 A. K. Alazzawi et al.

8. Younis MI, Zamli KZ, Isa NM (2008) MIPOG-modification of the IPOG strategy for T-
Way software testing. In: Proceeding of the distributed frameworks and applications (DFmA),
Penang, Malaysia
9. Ahmed BS, Zamli KZ (2010) PSTG: a t-way strategy adopting particle swarm optimization. In:
The fourth asia international on mathematical/analytical modelling and computer simulation
(AMS), IEEE, pp 1–5
10. Alazzawi AK, Rais H, Basri S (2019) ABCVS: an artificial bee colony for generating variable
t-way test sets. Int J Adv Comput Sci Appl 10:259–274
11. Alazzawi AK, Rais HM, Basri S (2018) Artificial bee colony algorithm for t-way test suite
generation. In: 2018 4th international conference on computer and information sciences
(ICCOINS), IEEE, pp 1–6
12. Floudas C, Pardalos P, Adjiman C, Esposito W, Gumus Z, Harding S, Klepeis J, Meyer C,
Schweiger C (1999) Handbook of test problems in local and global optimization. Kluwer
Academic Publishers, Dordrecht
13. Nsaif HN, Jawawi DNA (2020) Binary black hole-based optimization for t-way testing. In:
IOP conference series: materials science and engineering, IOP Publishing, pp 012073
14. Al-Sammarraie HNN, Jawawi DN (2020) Multiple black hole inspired meta-heuristic searching
optimization for combinatorial testing. IEEE Access 8:33406–33418
15. Altmemi JM, Othman R, Ahmad R (2021) SCAVS: implement sine cosine algorithm for gener-
ating variable t-way test suite. In: IOP conference series: materials science and engineering,
IOP Publishing, pp 012011
16. Alazzawi AK, Rais HM, Basri S, Alsariera YA (2020) Pairwise test suite generation based on
hybrid artificial bee colony algorithm. In: Advances in electronics engineering, Springer, pp
137–145
17. Alazzawi AK, Rais HM, Basri S (2019) HABC: hybrid artificial bee colony for generating
variable T-Way test sets. J Eng Sci Technol 7:13
18. Alazzawi AK, Rais HM, Basri S (2019) Parameters tuning of hybrid artificial bee colony search
based strategy for t-way testing. Int J Innov Technol Explor Eng (IJITEE) 8
19. Alazzawi AK, Rais HM, Basri S (2019) Hybrid artificial bee colony algorithm for t-way
interaction test suite generation. In: Computer science on-line conference. Springer, pp 192–199
20. Alazzawi AK, Rais HM, Basri S, Alsariera YA, Balogun AO, Imam AA (2021) A hybrid
artificial bee colony strategy for t-way test set generation with constraints support. J Phys:
Conf Ser 042068. IOP Publishing
21. Alazzawi AK, Rais HM, Basri S, Alsariera YA (2019) PhABC: a hybrid artificial bee colony
strategy for pairwise test suite generation with constraints support. In: 2019 IEEE student
conference on research and development (SCOReD). IEEE, pp 106–111
22. Younis M (2020) Gamipog: a deterministic genetic multi-parameter-order strategy for the
generation of variable strength covering arrays. J Eng Sci Technol 15:3142–3161
23. Younis MI, Zamli KZ (2011) MIPOG-an efficient t-way minimization strategy for combinato-
rial testing. Int J Comput Theory Eng 3:388
24. Younis MI, Alsewari ARA, Khang NY, Zamli KZ (2020) CTJ: input-output based relation
combinatorial testing strategy using jaya algorithm. Baghdad Sci J 17
Research and Accomplishments
in Applications of Non-negative Matrix
Factorization

Phong Cao Nguyen, Cao Hong Nga, and Huong Hoang Luong

Abstract Since its introduction in the 1990s, non-negative matrix factorization

(NMF) has captured a great amount of attention due to its capability and effec-
tiveness in processing data in such a way that few earlier methods could perform,
partly due to its non-negative constraint. This paper first briefly presents the basic
NMF algorithm and concerns with the algorithm itself, then demonstrates its power
with three applications in three different fields, namely face recognition in Com-
puter Vision, distance prediction in Networking and molecular pattern discovery in
Genetics. The paper ends with a quick look at other applications of NMF and recent
developments that researchers have made.

Keywords Non-negative matrix factorization applications · Face recognition ·

Network distance prediction · Molecular pattern discovery

1 Introduction

Non-negative matrix factorization was popularized, to a large extent, by D. D. Lee

and H. S. Seung when they introduced some algorithms and related proofs [1] after
the idea was first described in [2] under the slightly different name positive matrix
factorization (PMF). Since then, many applications of NMF have been found in
different fields, from Informatics to Biology or Astronomy and beyond. This paper
looks at some among that large collection of its uses. It starts with a scenario which
reveals the need for such an efficient data analysis algorithm and touches briefly
on the challenges accompanying the algorithm (Sect. 2). It then presents how NMF
has been applied in Image Processing and Face Recognition, Network Distance
Estimation, and Molecular Pattern Discovery by reviewing several typical works

P. C. Nguyen · H. H. Luong (B)

FPT University, Can Tho, Vietnam
e-mail: [email protected]
C. H. Nga
National Central University, Taoyuan City, Taiwan

© The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 1061
R. Ibrahim et al. (eds.), International Conference on Artificial Intelligence
for Smart Community, Lecture Notes in Electrical Engineering 758,
https://doi.org/10.1007/978-981-16-2183-3_101
1062 P. C. Nguyen et al.

in the respective subjects in Sects. 3–5. Finally, it ends by describing how NMF is
being applied in even more fields, and how research to enhance NMF’s performance
in current applications is being carried out in Sect. 6, which is followed by a brief
conclusion.

2 Non-negative Matrix Factorization Overview

Methods for summarizing and analyzing mass datasets are needed in order to effec-
tively utilize the huge amount of data available nowadays. For example, an online
movie platform should be able to recommend new movies to users based on their
past ratings, other users’ ratings, similarities between movies and so on. Consider
a simple case where users’ ratings (m users on n movies) are stored in the m × n
matrix X . The (i, j)-entry of X denotes the rating of user i towards movie j. Such
a matrix could easily be large; one million users and 10,000 movies constitute 10
billion entries. What NMF does is, for a relatively small integer k, finding an m × k
matrix W and a k × n matrix H such that every entry in W and H is non-negative
and X ≈ W H . The W and H matrices are usually called basis and weight matrices
respectively because each column in X can be viewed as a linear combination of the
rows in W with coefficients provided by columns in H [1] (see Fig. 1). The n movies
are also said to be divided into k clusters.
The NMF problem itself is NP-hard [3], which means algorithms for exact NMF
have exponential time complexity, thus too slow for real-world scenarios. Conse-
quently, most practical algorithms are heuristic with acceptable accuracy. One of the
most common algorithms is that presented by Lee and Seung in [1], on which many
variants have been based. The core of the standard algorithm lies in the iterative
updates of W and H [1]. The main steps of the algorithm are briefly listed below.
1. Define a function that measures the quality of the approximation, called a cost
function. A simple measure is the Euclidean distance E(A, B) between two m × n
matrices A and B:

Fig. 1 Reconstructing original data using W and H

Research and Accomplishments in Applications … 1063


 m 
n
 2
E(A, B) =  Ai j − Bi j (1)
i=1 j=1

Alternatively, the divergence of A from B can be used:

 n 
m 
Ai j
D (A||B) = Ai j log − Ai j + Bi j (2)
i=1 j=1
Bi j

The goal is to minimize E(X, W H ) or D (X ||W H ), depending on what function

was chosen.
2. Initialize W0 and H0 . The initialization should depend on the specific conditions
of each problem.
3. At each iteration, update W and H according to the multiplicative rules.

(a) If the Euclidean distance was chosen, the (i, j)-entry of the matrices are updated
as follows.  
X HT ij
Wi j ← Wi j   (3)
W H HT ij
 
WT X
ij
Hi j ← Hi j  T
 (4)
W W H ij

(b) If the divergence was chosen, the (i, j)-entry of the matrices are updated as
follows.
H jk X ik
k (W H )ik
Wi j ← Wi j (5)
l H jl

Wki X k j
k (W H )k j
Hi j ← Hi j (6)
l Hli

Iteration continues until the value of the chosen function falls within the pre-specified
desired range.
The authors proved the convergence property of the algorithm, by which contin-
uous iterations would converge to a locally optimal factorization [1].
The algorithm is not without its flaws, however. Certain concerns regarding the
use of heuristic algorithms like this have been raised, including the non-uniqueness of
W and H (multiple satisfactory pairs may exist, in which case initialization must be
carefully determined) and the choice of rank k (an understanding of the underlying
structure of the objects may be required). Research into these issues is still on-
going, along with the development of variants of the standard algorithm like graph
regularized NMF [4] or hierachical NMF [5], just to name a few. Readers interested
in such research are referred to [1, 3, 6] and the mentioned works.
1064 P. C. Nguyen et al.

3 Applications of Non-negative Matrix Factorization

in Image Processing and Face Recognition

Image processing is one of the earliest applications of NMF. A collection of n images

containing m pixels each is represented by the m × n matrix X . The jth image—
the jth column of X —can be approximated as the linear combination of k basis
images—k columns in W (see Fig. 2).
If the image is about a human face, the basis images are usually facial parts like
the eyes, the nose or the mouth. An advantage of NMF becomes clear in this case:
it expresses the face as a sum of facial parts. The non-negative restriction on the
matrix entries is useful intuitively; it is hard to interpret what a negative weight for
the mouth component means.
Lee and Seung applied this method to a collection of n = 2429 images contain-
ing m = 19 × 19 pixels each [7]. They compared NMF with Principle Components
Analysis (PCA) and Vector Quantization (VQ), which are holistic approaches; that
is, in these methods, the basis images are whole faces rather than parts of a face.
PCA approximates the original image with only one prototypical basis image, while
VQ allows negative weights, which the authors deemed not intuitively meaning-
ful [7]. These two algorithms were chosen for comparison primarily to highlight
the effectiveness of the non-negative and part-based properties of NMF over the
whole-object-based property of other existing algorithms.
The part-based property of NMF has also been utilized in face recognition. Guil-
lamet and Vitriá compared NMF against PCA and two commercial techniques—
FaceIt and Bayesian—in terms of their ability to recognize faces under different
conditions [8]. FaceIt and Bayesian are considered leading techniques in computer
vision, thus comparison with these two techniques aid the community in determin-
ing if NMF can replace these commercial techniques and work well in a production
environment. The authors used the well-known AR face database [9]. The database
consists of images of 116 people taken in two sessions two weeks apart. Features
like facial expressions and external conditions were also present in the collection for
each session (see Table 1) [9]. The images with neutral expression (labelled 01) were
used as the training set.

Fig. 2 NMF on a collection of images

Research and Accomplishments in Applications … 1065

Table 1 Features in each collection

Neutral
01 Neutral expression
Facial expression
02 Smile
03 Anger
04 Scream
Light
05 Left light on
06 Right light on
07 Both lights on
Sunglasses
08 Sunglasses
09 Sunglasses, left light on
10 Sunglasses, right light on
Scarf
11 Scarf
12 Scarf, left light on
13 Scarf, right light on
14–26: second session

For each of PCA and NMF, training was conducted with dimensions of 50, 100
and 150 to reveal, if any, the relationship between dimensionality and the results of
face recognition. Initialization is effectively random. The experiments pointed out
that NMF performed better than PCA at recognizing smiles and was comparable
with FaceIt when it came to anger. However, scream proved too difficult for both
NMF and PCA. For the illumination set, NMF, when used in a high dimensional
space (150), even outperformed both FaceIt and Bayesian [8].
Under the effect of sunglasses, recognition rates for all approaches decreased
remarkably, implying the importance of eyes in recognizing faces [8]. NMF did best
when there was no light, but when lighting conditions affect the photos, it became
less capable. Similar results were observed for the scarf set, suggesting NMF was
good for a certain level of occlusions without another change in the scene [8].
NMF’s ability to recognize face could be enhanced by adding various sets of
constraints to the matrices W and H , enforcing these constraints by using different
cost functions, which in turn produce variants of pure NMF. A number of such
variants has been proposed, including local NMF (LNMF), Fisher NMF (FNMF)
or PCA NMF (PNMF). For example, LNMF adds three constraints to pure NMF,
which are (1) the number of bases should be minimized, which essentially means a
basis component should not be decomposed any further, (2) bases must be made as
orthogonal as possible, which is to minimize redundancy among bases and (3) only
1066 P. C. Nguyen et al.

bases with the most important information should be retained. LNMF consolidates
these constraints into its divergence function (specific limits for i and j have been
omitted for simplicity):

 Xi j  
D (X ||W H ) = X i j log − X i j + (W H )i j +α Ui j − β Q ii
i, j
(W H )i j i, j i
(7)
where α, β > 0 are constants, U = W T W and Q = H H T .
The reader is referred to [10] for a detailed description of these variants. Other
than the mentioned constraints, the core steps should remain more or less the same
to what was described in Sect. 2. Experiments with different databases demonstrated
the improved performance of FNMF and PNMF against NMF or LNMF in certain
circumstances due to the ability of these methods to better localize basis images to
the certain facial parts [10].

4 Applications of Non-negative Matrix Factorization

in Network Distance Estimation

Network distances (including, but not limited to, round-trip time between hosts or
network latency) have a great effect on the operation of distributed applications. A
user who wants to download a file is able to choose the fastest server to fetch the file
from if information about network distances is available. However,
  obtaining all pair-
wise distances for N hosts is expensive: it would take O N 2 measurements. Thus,
an issue of interest is estimating unknown distances from O(N ) measurements and
many studies have been dedicated to solving this problem. Early models are based on
Euclidean distance, in which hosts are represented as coordinates  in a d-dimensional 
systems.
 The distance
 between H i and H j represented by Hi 1 Hi 2 , . . . , Hi d and
,
H j1 , H j2 , . . . , H jd respectively is

 d
  2
Di j =  Hik − H jk (8)
k=1

Examples include Global Network Positioning (GNP), in which estimations are

made by obtaining known distances from predetermined hosts called landmarks [11]
and Vivaldi, one of the most widely-used distributed system [12]. Such systems
have two common major drawbacks. First, the distances predicted are symmet-
ric; that is, Di j = D ji . This is sometimes not true as some studies have indicated
[13, 14]. Second, distances predicted this way have to observe the triangle inequality:
Di j + D jk ≥ Dik . H. Zheng et al. have shown that triangle inequality violations are
not rare in networks [15], and other studies have analyzed the effect they can have
on the accuracy of Euclidean distance-based systems [16, 17].
Research and Accomplishments in Applications … 1067

Fig. 3 Matrix factorization in estimating network distances

A system based on matrix factorization, called Internet Distance Estimation Ser-

vice (IDES), was introduced in [18]. In IDES, N × d matrix W and d × N matrix
H are found such that the N × N distance matrix D is approximated by W H . Then,
the distance from host i to host j can be approximated as the dot product of the ith
row of W and the jth column of H T (see Fig. 3) [18]:

→ −→
−
Di j = Wi · H jT (9)

Under this model, distances do not have to observe the triangle inequality, and
Di j is not necessarily equal to D ji . Hence, it better predicts real-life distances com-
pared to earlier Euclidean distance-based models. In the paper, the authors proposed
two factorization algorithms—Single Value Decomposition (SVD) and NMF—and
found that NMF had a major advantage over SVD: it could “fill” missing values
in the distance matrix with a slight modification in the algorithm, while SVD had
to “skip” hosts with missing values [18]. The predictions were closer to the actual
values, but were still negatively affected by landmarks that often produced incorrect
values. Chen et al. [19] introduced weight to the system—landmarks that provided
correct values more often would play a larger part in calculating the prediction. They
demonstrated a significant decrease in relative error on predictions of certain datasets
when weight was added compared to the IDES model in [18, 19].
The NMF algorithms behind IDES and Phoenix both try to minimize the error
function E, which is the Euclidean error function presented in Sect. 2 [18, 19]:


N 
N
 2
E= Di j − D ji (10)
i=1 j=1

Chai et al. showed that higher accuracy in predicting network distances can be
achieved if the error function used is the divergence of D from R = W H , also
mentioned in Sect. 2 [20]:

 N 
N 
Di j
K (D||R) = Di j log − Di j + Ri j (11)
i=1 j=1
Ri j
1068 P. C. Nguyen et al.

This class of NMF is usually called NMF using KL divergence (KL-NMF), and
it was also discussed in [1]. Apart from this difference, the steps in the algorithm
of [20] are similar to those in [18], making use of landmarks or reference hosts. Chai
et al. observed a 10–14.2% increase in prediction accuracy on four datasets when
compared with the version of NMF used in IDES [20]. The authors attributed the
improved performance to the ability of KL divergence to model network delay as
having a Poisson distribution rather than the normal distribution that the Euclidean
distance-based error function assumes [20].

5 Applications of Non-negative Matrix Factorization

in Molecular Pattern Discovery in Genetics

Some genes in our body belong to the same group and contribute to an expression
or a disease. Clustering a large number of genes into a small number of metagenes
and analyzing the expression patterns of these metagenes are useful in discovering
hidden or unclear biological structure or relationships between groups of genes and
certain expressions. These clusters are comparable to facial parts in face recognition
applications. Brunet et al. applied NMF on Leukemia, Medulloblastoma and Cen-
tral Nervous System Tumors datasets to classify cancer subtypes [21]. This section
describes the experiment with the Leukemia set.
N genes and their expression levels in M samples (in the Leukemia dataset [24],
M = 38) are stored in the N × M matrix X . When an approximation X ≈ W H
is found (W is an N × k matrix), each column in W defines a metagene that is a
positive linear combination of the N genes, and each column in H represents an
expression pattern of each sample as a positive linear combination of the metagenes.
This factorization also divides M samples into k clusters; sample j belong to cluster
i if the (i, j)-entry of H is the largest in column j (see Fig. 4) [21].
Acute leukemia is generally divided into two types—acute myelogenous leukemia
(AML) and acute lymphoblastic leukemia (ALL), with the latter being further divided
into two subtypes—T and B cells. The authors compared NMF with two other meth-
ods, namely hierarchical clustering (HC) and self-organizing maps (SOM). HC is
essentially a process of clustering that involves merging the closest pair of clusters

Fig. 4 NMF as a clustering method

Research and Accomplishments in Applications … 1069

at the previous step into a larger one until a strong cluster, which consists of all
objects, are obtained [22]. The result is a tree-like structure of clustering, hence the
name hierarchical. An SOM consists of nodes that are initially placed at random
into k-dimensional space, whose positions are then adjusted by moving the points
toward a random data point selected for that iteration. After the specified number
of iterations (about 20,000–50,000), the points that are near each other are likely to
form a related cluster [23]. Among the three, better methods would correctly classify
the samples into the 2 or 3 abovementioned clusters (depending on the rank used in
the algorithm) and provide more stable results (sample j belongs or does not belong
to cluster j in almost all of the runs of the algorithm with different initial conditions).
The findings were as follows [21].
HC
(1) HC split ALL-B samples into two subgroups (when they should have been
together).
SOM
(2) While SOM could divide the samples into two clusters, such classification was
not stable as the two groups could be either [AML] vs [ALL] or [AML + ALL-T] vs
[ALL-B].
(3) SOM with k = 3 classes did not correctly divide the samples into the three known
classes (ALL-T and ALL-B were mixed).
NMF
(4) NMF correctly split the samples into two clusters with remarkable stability regard-
less of initialization.
(5) When used with rank k = 3, NMF correctly distinguished ALL-T from ALL-B.
(6) Under higher ranks, NMF found a 4-cluster division quite robust while showing
considerable dispersion when k = 5, implying that many clusters were not likely.
The authors concluded NMF had multiple strong points when compared with HC
and SOM, and was better at discovering biological substructure [21]. sThis was con-
sistent with the observation that HC imposes strict hierachical structure to objects,
which may not always be the case, and at times needs subjective clustering definition,
while SOM, though imposes only a partial structure on the objects, which is more
appropriate, is unstable and may produce different clusterings with different initial-
izations [21]. Y. Gao and G. Church later demonstrated the enhanced performance
of sparse NMF, a variation, over classic NMF [25].
The non-uniqueness of W and H , however, may produce inconsistent results
between various runs of NMF. Yang and Seoighe studied some post-processing meth-
ods to improve the stability of the algorithm [26]. They standardized W and H using
1-norm, 2-norm, 3-norm and maximum norm in addition to quantiles and sample
standard deviation. Normalization maps the matrix W to a diagonal matrix D and
updates the solution as follows: W  = W D −1 and H  = D H . The authors also tried
applying filters to the resulting matrices so that irrelevant genes were removed before
running the algorithm again. Irrelevant genes were defined as those that had similar
1070 P. C. Nguyen et al.

values in the clusters, and thus it was not clear as to which metagenes they belonged
to [26]. When performed on the Leukemia dataset, maximum norm remarkably made
no errors even when there were two well-known outliers in the dataset [21]. In con-
trast, both outliers were incorrectly categorized when basic NMF was used. For most
of the other datasets in the experiments, NMF with maximum norm was also singled
out as the best classifier, and a filter was also proved to be able to boost performance
in certain cases [26].
In medical, NMF is also used to detect new subtypes of diseases. Chalise et al.
proposed combining NMF and similar network-based clustering (nNMF) across mul-
tiple data sources to recognize latent clustering structures in the data [31]. In this
method, each consensus matrix constructed from each dataset represents the simi-
larity between patients. These matrices are integrated into a robust consensus matrix
which represents the network of patient similarities and the nNMF clusters. Since
consensus matrices are calculated separately, they are not affected by the discrepan-
cies in the distributions and scales of the datasets. Consensus matrices are built on a
cross-validation approach that partially solves the overfitting and underfitting issues
of unsupervised methods. This method is applied on both real and simulated data
and proves the applicability of this method.

6 Other Applications and Recent Developments

Applications of NMF are not limited to the above three fields. Texts of hundreds of
thousands of words are clustered into a few topics [27]. In astronomy, Ren et al. used
NMF to extract images of astrophysical objects out of direct imaging data [28] and
Berne et al. analyzed the emission of dust particles from spectro-imagery data [29].
New recommendation systems using NMF are being proposed [30]. Along with
improvements in the algorithm for NMF itself, new applications in existing and new
fields are certain to be discovered.

7 Conclusions and Future Works

NMF in three different fields—computer vision, networking and genetics, demon-

strates the ability of NMF to both act as a predicting/estimating model and a clustering
method. The results of the experiments indicate certain edges NMF has over existing
approaches due to its constraint of allowing only non-negative entries in the matrices.
It is certain that improved versions of these applications as well as new applications
will emerge, thanks to the continuous effort researchers put into the field. As part
of our future research, we will focus on a particular use of NMF, namely recom-
mendation systems. Specifically, we will build new models by using session-based
and streaming session-based in conjunction with the NMF and set up a new recom-
mendation system using our models. In addition, we will carry out some research
Research and Accomplishments in Applications … 1071

to overcome the drawbacks of face recognition in some conditions like screaming

faces, face with sunglasses, or face in different light conditions by combining NMF
with deep neural networks.

References

1. Lee DD, Seung HS (2001) Algorithms for non-negative matrix factorization. Adv Neu Inf Proc
Syst 14:556–562
2. Paatero P, Tapper U (1994) Positive matrix factorization: a non-negative factor model with
optimal utilization of error estimates of data values. Environmetrics 5(2):111–126. https://doi.
org/10.1002/env.3170050203
3. Gillis N (2014) The why and how of nonnegative matrix factorization. In: Suykens JAK, Sig-
noretto M, Argynou A (eds) Regularization, optimization, kernels, and support vector machines.
Chapman and Hall/CRC, New York, NY, USA, pp 257–291
4. Cai D, He X, Han J, Huang TS (2011) Graph regularized nonnegative matrix factorization for
data representation. IEEE Trans Patt Anal Mach Intell 33(8):1548–1560. https://doi.org/10.
1109/TPAMI.2010.231
5. Li Y et al (2013) Hierarchical non-negative matrix factorization (hNMF): a tissue pattern
differentiation method for glioblastoma multiforme diagnosis using MRSI. NMR in Biomed
26(3):307–319. https://doi.org/10.1002/nbm.2850
6. Berry MW, Browne M, Langville AN, Pauca VP, Plemmons RJ (2007) Algorithms and applica-
tions for approximate nonnegative matrix factorization. Comput Stat Data Anal 52(1):155–173.
https://doi.org/10.1016/j.csda.2006.11.006
7. Lee DD, Seung HS (1999) Learning the parts of objects by non-negative matrix factorization.
Nature 401(6755):788–791. https://doi.org/10.1038/44565
8. Guillamet D, Vitria J (2002) Non-negative matrix factorization for face recognition. Catalonian
Conf Artif Intell 336–344
9. Martínez AM, Benavente R (1998) The AR face database. Comput Vis Center Tech Rep
#24:1998
10. Wang Y, Jia Y, Hu C, Turk M (2005) Non-negative matrix factorization framework for
face recognition. Int J Patt Recogn Artif Intell 19(4):495–511. https://doi.org/10.1142/
S0218001405004198
11. Ng TSE, Zhang H (2002) Predicting Internet network distance with coordinates-based
approaches. In: Proceeding twenty-first annual joint conference of the IEEE computer and
communications societies, New York, NY, USA, pp 170–179 vol 1. https://doi.org/10.1109/
INFCOM.2002.1019258.
12. Dabek F, Cox R, Kaashoek F, Morris R (2004) Vivaldi: a decentralized network coordinate sys-
tem. ACM SIGCOMM Comput Commun Rev 34(4):15–26. https://doi.org/10.1145/1030194.
1015471
13. Paxson V (1997) End-to-end routing behavior in the Internet. IEEE/ACM Trans Netw 5(5):601–
615. https://doi.org/10.1109/90.649563
14. Lakshminarayanan K, Padmanabhan VN (2003) Some findings on the network performance
of broadband hosts. In: Proceeding 3rd ACM SIGCOMM Conference Internet Measurement,
pp 45–50. https://doi.org/10.1145/948205.948212
15. Zheng H, Lua EK, Pias M, Griffin TG (2005) Internet routing policies and round-trip-times.
In: International workshop on passive and active network measurement, pp 236–250. https://
doi.org/10.1007/978-3-540-31966-5_19
16. Lee S, Zhang ZL, Sahu S, Saha D (2006) On suitability of euclidean embedding of Inter-
net hosts. ACM SIGMETRICS Perform Eval Rev 34(1):157–168. https://doi.org/10.1145/
1140103.1140296
1072 P. C. Nguyen et al.

17. Wang G, Zhang B, Ng TE (2007) Towards network triangle inequality violation aware dis-
tributed systems. In: Proceeding 7th ACM SIGCOMM conference internet measurement, pp
175–188. https://doi.org/10.1145/1298306.1298331
18. Mao Y, Saul LK, Smith JM (2006) IDES: an internet distance estimation service for large
networks. IEEE J Select Areas Commun 24(12):2273–2284. https://doi.org/10.1109/JSAC.
2006.884026
19. Chen Y et al (2011) Phoenix: a weight-based network coordinate system using matrix factor-
ization. IEEE Trans Netw Serv Manage 8(4):334–347. https://doi.org/10.1109/TNSM.2011.
110911.100079
20. Chai L, Luo X, Zhao F, Li M, Liu S (2017) Network coordinate system using non-negative
matrix factorization based on KL divergence. In: 2017 19th international conference on
advanced communication technology (ICACT), Bongpyeong, pp 193–198. https://doi.org/10.
23919/ICACT.2017.7890082
21. Brunet JP, Tamayo P, Golub TR, Mesirov JP (2004) Metagenes and molecular pattern discovery
using matrix factorization. Proc. Nat Acad Sci 101(12):4164–4169. https://doi.org/10.1073/
pnas.0308531101
22. Johnson SC (1967) Hierarchical clustering schemes. Psychometrika 32(3):241–254. https://
doi.org/10.1007/BF02289588
23. Tamayo P et al (1999) Interpreting patterns of gene expression with self-organizing maps:
methods and application to hematopoietic differentiation. Proc Nat Acad Sci 96(6):2907–2912.
https://doi.org/10.1073/pnas.96.6.2907
24. Golub TR et al (1999) Molecular classification of cancer: class discovery and class prediction
by gene expression monitoring. Science 286(5439):531–537. https://doi.org/10.1126/science.
286.5439.531
25. Gao Y, Church G (2005) Improving molecular cancer class discovery through sparse
non-negative matrix factorization. Bioinformatics 21(21):3970–3975. https://doi.org/10.1093/
bioinformatics/bti653
26. Yang H, Seoighe C (2016) Impact of the choice of normalization method on molecular cancer
class discovery using nonnegative matrix factorization. In: PloS One 11(10):e0164880. https://
doi.org/10.1371/journal.pone.0164880
27. Hassani A, Iranmanesh A, Mansouri N (2019) Text mining using nonnegative matrix factor-
ization and latent semantic analysis. arXiv:1911.04705
28. Ren B, Pueyo L, Zhu GB, Debes J, Duchêne G (2018) Non-negative matrix factorization:
robust extraction of extended structures. Astrophys J 852(2). https://doi.org/10.3847/1538-
4357/aaa1f2
29. Berne O et al (2007) Analysis of the emission of very small dust particles from Spitzer spectro-
imagery data using blind signal separation methods. Astronomy Astrophys 469(2):575–586.
https://doi.org/10.1051/0004-6361:20066282
30. Benzi K, Kalofolias V, Bresson X, Vandergheynst P (2016) Song recommendation with non-
negative matrix factorization and graph total variation. In: IEEE international conference acous-
tics, speech and signal processing (ICASSP). Shanghai 2439–2443. https://doi.org/10.1109/
ICASSP.2016.7472115
31. Prabhakar C, Ni Y, Fridley BL (2020) Network-based integrative clustering of multiple types
of genomic data using non-negative matrix factorization. Comput Biol Med. https://doi.org/
10.1016/j.compbiomed.2020.103625
Feature Selection Using Correlation
Matrix on Metagenomic Data with
Pearson Enhancing Inflammatory Bowel
Disease Prediction

Huong Hoang Luong, Trong Thanh Tran, Ngoc Van Nguyen, An Duc Le,
Huyen Thi Thanh Nguyen, Khoi Dinh Nguyen, Nghi Cong Tran,
and Hai Thanh Nguyen
Abstract The Fourth Industrial Revolution has brought up a vast amount of new
innovative implementations. These can be used for numerous areas to make wealth
and to improve human ways of living. Our point of view is to consider medical
problems for enhancing prediction first. In this study, we like to rise a question about
whether if we could enhancing Inflammatory Bowel Disease (IBD) prediction for
early detect related sickness by feature selection on metagenomic data. Over the
last few years, the prediction has been a challenge. Because of rare information and
lacking data, the problem is not well considered enough. To bring back the subject, in
this work, we propose a new way of enhancing Inflammatory Bowel Disease (IBD)
prediction by using the Correlation Matrix with the Pearson on Metagenomic Data.
Our implications have the purpose of finding out whether we could do predictions
better using a specific amount of features selected by Pearson correlation coefficient.
The result with the proposed method is quite promising, when we address some high
correlation features out, the model can predict better comparing to randomly select
features.

Keywords Feature selection · Pearson coefficient · Metagenomic data ·

Inflammatory bowel disease

H. T. Nguyen
College of Information and Communication Technologies (CICT),
Can Tho University, Can Tho, Vietnam
e-mail: [email protected]
H. H. Luong (B) · T. T. Tran · N. Van Nguyen · A. D. Le · H. T. T. Nguyen · K. D. Nguyen
Information Systems Department, FPT University, Can Tho, Vietnam
e-mail: [email protected]
N. C. Tran
National Central University, Taoyuan, Taiwan

© The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 1073
R. Ibrahim et al. (eds.), International Conference on Artificial Intelligence
for Smart Community, Lecture Notes in Electrical Engineering 758,
https://doi.org/10.1007/978-981-16-2183-3_102
1074 H. H. Luong et al.

1 Introduction

In recent years, the quantity of deaths related to the gastrointestinal tract is increas-
ing. Inflammatory Bowel Disease (IBD) is not a doom disease but worth of mention
because of its part to develop some deadly diseases. IBD is also one of the primary
causes of colorectal cancer (CRC) [1]. IBD [2] is a term for two main categories
including Crohn’s disease (CD) [3] and Ulcerative colitis (UC) [4], which is char-
acterized by chronic inflammation of the digestive tract. Digestive tract damages
could be caused by prolonged inflammation which is also a major risk factor for
the formation and development of malignancies. Generally, cancer [5–7] is a dis-
ease characterized by an uncontrolled nature division and the existence of abnormal
cells. There are abnormal growth occurs in the colon or rectum, it is called colorectal
cancer [8, 9].
There are many factors that result in the disease including lifestyle, genetics,
and environmental surroundings. According to WHO statistics in 2018 [10], CRC
is the second most deadly cancer and the third most commonly diagnosed cancer
in the world. There are 1.80 million new cases and 862 000 deaths. In the United
States, CRC is the second most common cancer diagnosis. The American Cancer
Society’s [11] estimate for 2020 is 147,950 new cases. It is expected to cause around
53,200 deaths by 2020. The dangers if the person has this disease: cancer cells only
form and grow in the colon area in an early stage. Therefore, treatment is not too
difficult. Moving into the second stage, the pathogenic cells attack other parts of
the colon. This is a milestone that marks the development of cancer in the colon.
If cancer cells are not controlled in time, they begin to work harder, spreading to
lymph nodes. More lymph nodes that are attacked, the more severe your condition
becomes, and the harder it is to control. Finally, when the disease enters a terminal
stage, almost every organ of the patient is affected by the cancer cells. Treatment
does not bring much hope for the patient. The risk of death is high. For prevention
and early detection of the disease, an improved approach is required to cut back the
number of individuals who are becoming this illness, further leading patients into a
brand new healthy lifestyle.
Later to improve patients’ experiences with the growth of personalized medicine.
We see that patients are diagnosed with the disease in less time, assess the risks,
evaluate the optimal treatment, and last but not least help them improve health while
reducing expenses. On the other side, medical manufacturers also are provided a
chance to develop agents that are targeted to patient groups that do not respond to
medications as intended. New value assessments for personalized medicine products,
together with return-on-investment models, will be required as these new strategies
for pharmaceutical and diagnostic products emerge. For the direct side, hospitals,
health care providers, and health plan sponsors, personalized medicine represent yet
one more challenge in uncertain times. Innovation in provider and benefits manage-
ment, together with clarity in regulatory and legal constructs are required, even as
new national insurance reforms begin to emerge. The applications of PM are diag-
nosis and intervention, drug development and usage, respiratory proteomics, cancer
Feature Selection Using Correlation Matrix on Metagenomic … 1075

genomics, and population screening. And through that use metagenomic to research
methods for personalized medicine. Metagenomics is the research of metagenomes
that aims to obtain genetic material directly from samples within the real living envi-
ronment. This field is included in environmental genetics, ecological genetics, or
simply genetics generally. Due to its important role within the discovery of microbial
diversity, metagenomics are often considered as a prism to assist us better understand
the planet of small organisms, contributing to humanity’s understanding of the entire
living world. In medicine, personalized medicine could be a type of using informa-
tion a couple of person’s genes or proteins, from the data above wont to diagnose,
prevent, or treat a disease that someone may have. For that individual field of study,
aimed by the expansion of machine learning technology. We will combine medi-
cal data and machine learning by determining the concentration of bacteria that are
characteristic of the chance of making tumors that cause sickness.
Some current applications of personalized medicine and machine learning tech-
nology are extremely helpful not only within the way of supporting doctors in early
diagnosis but also for the sake of all stakeholders [12]. Machine learning has created
an unlimited amount of applications in healthcare. These are a number of imple-
menting machine learning in healthcare like diagnosis and identifying diseases, drug
discovery, medical imaging diagnosis, etc. Personalized medicine is one in these
applications which is referenced during this article [13].
During this particular article, we illustrate a way approaches for using the correla-
tion matrix to boost Inflammatory Bowel disease prediction. Our goal is to represent a
brand new efficient set of methods thanks to the trustworthy results of the prediction.
Our paper consists of 4 main sections. In the next Sect. 2, we present some high-
lights of related work. Afterward, the methodology Sect. 3 where we introduced to
you the methods which we use on the topic later. The next Sect. 4 will be details about
the Experiments. Finally, in the final Sect. 5, we conclude our paper and summarize
the point of our experiments.

2 Related Work

Numerous research studies have applied Machine Learning to analyze and visual-
ize metagenomic datasets. In particular, authors have done a study on the method
to increase disease prediction based on Metagenomic using Unsupervised Binning
Approaches. Their research aims to present one-dimensional (1D) using a binning
and scaler approach. They also use Multi-layer Perceptron (MLP) and traditional
artificial neural networks. In the paper, their purpose is to divide the bin into differ-
ent which includes frequency, width, and proposed breaks. From there, the process of
learning with MLP is to classify diseases with multi-fold (K-fold). The Binning clas-
sification method was implemented in the test with 10 bins. The implemented MLPs
are the Rectified Linear Unit (ReLU) [14]. Additionally, the authors investigate many
methods for visualizing features of metagenomic datasets relate to colorectal cancer.
Their objective is to propose an efficient method to present data distribution visually
1076 H. H. Luong et al.

and use for classification. They proposed three methods including Image Generation
and Models for Evaluation, Visualizations Based on Dimensionality Reduction Algo-
rithms, and Visualization Based on Data Density. Image Generation and Models for
Evaluation use Species Bins and Quantile Transformation Bins to generate the image,
then CNN is used to evaluate the considered visualization and models. Visualiza-
tions Based on Dimensionality Reduction Algorithms apply dimensionality reduc-
tion algorithms such as t-SNE, Principal Component Analysis, Random Projection,
Spectral Embedding, Isomap, Locally Linear Embedding, Multidimensional scaling
and Non-negative matrix factorization. They are used to shaped high-dimensional
data in 2D images for improving prediction results. Visualizations based on the Data
Density illustrate the data based on the data density with Fill-up [17]. Besides, data
visualization is a challenge for many areas. With measurement data, it is character-
ized by the size of the data that is difficult for humans to interpret. In diseases that
use measurement data to predict, DL is often less effective than Classical Machine
Learning in predicting CRC. Authors present an approach using Manifold Learning
with t-SNE [15] and Spectral embedding to visualize digital data into images and
using DL to improve performance in predicting CRC. Results come from 5 different
areas used to combine visualization with Deep Learning to improve performance in
diagnostics.
In the last few years, Metagenomics is a study that includes many genomes at
the same time. This Metagenomics can be collected from a variety of environments.
For example, Bacteria in the human intestine are estimated to have between 500 and
more than 1000 species in the human intestine. However, investigating the disease
remains a major challenge due to inconsistent results in disease prediction and dis-
ease complexity. High-dimensional researchers have difficulty understanding data.
Many studies use 2D illustrations to interpret data and find patterns in the data. To
explain metagenomic data, many studies devise methods to visualize data. Krona
[16] allows to display data in order and use the zoom feature of multi-layered pie
charts. The survey study shows that: Manifold Learning-based approach shows the
abundance of bacteria that cause colon cancer through data samples, from which 2
out of algorithms are selected. The best is Spectral embedding (SE) and t-SNE [15]
for further investigation in preparation for the filling method and based on the mean
magnitude of the feature. Visualization based on the fill approach: arrange and fill
the values of the features into a matrix in order of right to left and top to bottom. The
author fills in using phylogenetic ordering and random sorting. Then, the proposed
visuals are loaded into a DL algorithm for prediction. In short, this article uses Man-
ifold Learning with 2 methods of embedding SE and t-SNE to put data into a Deep
Learning algorithm to predict and then Fill-up [17, 18]. Moreover, authors show
how to use the CNN network on data that has no original visual structure. The first
is to map the data and define it to a 1D or 2D image. Based on that description they
apply the CNN network to predict the disease. The collection of data in the medical
field has created a revolution in medical research, application and biotechnology.
These data are called omics data, they reflect aspects of the biology of the system.
More and more data sets are available and models are using this information to make
decisions. Today, the most successful techniques applied to this omics data set are
Feature Selection Using Correlation Matrix on Metagenomic … 1077

mainly Random Forest (RF) and sparse regression. In this paper, authors use DL
directly onto 6 measurement datasets to reflect the abundance of bacterial species,
their presence in the intestines of sick and healthy patients. This technology works
well in image classification we focus on using CNN. The purpose of the paper is to
propose an effective, concise presentation in images and demonstrate the DL tech-
niques as an effective tool for the metagenomics prediction tasks. Methods include
the Visualization of image-based representations. Its steps are 1 color is chosen and
applied according to the abundance of distribution. The division can be done on a lin-
ear or logarithmic scale. The features are displayed as images in 1 of 2 ways, namely
phylogenetic-sorting (PLG) or t-SNE. In short, this paper proposes the MET2IMG
method to predict the patient’s disease using measurement data. Authors use two
main methods to construct synthetic images which are Fill-up [17] and t-SNE [19].
The Fill-up [17] used small and simple images. The T-SNE [19] feature is embedded
in two-dimensional space using classic embedding in ML [20].

3 Methodology

3.1 Feature Selection Using Correlation Matrix with Pearson

In our study, we select features using the Correlation Matrix with Pearson [21]. It is
a widely used method to classify and visually. We use the Correlation Matrix with
Pearson which measures linear dependence between two variables [22]. Later we
will be able to pick out some of the features that have a high correlation coefficient.
The datasets as mention in 4.1 Datasets Description, consisting of 6 sets of data.
Each of the datasets includes the number of rows corresponding to be called, the
subjects. These next columns contain the code of the bacteria and their value Metage-
nomic Data corresponding to the subjects. After implementing the Correlation Matrix
with Pearson, we get a matrix n · n, with n is the number of features. Matrix values
are the Correlation Coefficient values range from −1.0 to +1.0 with −1.0 is a total
negative correlation, 0.0 is no correlation and +1.0 is a total positive correlation.
Regardless of the sign correlation, we only take absolute values. These absolute val-
ues will be considered to be greater than n to let the features to be selected. In the
study, we set n to 0.1, 0.2 and 0.3.

3.2 Feature Selection Randomly for Later Comparison

For the comparison purpose, we used randomly select features to compare the Feature
Selection using the Correlation Matrix with Pearson for the final result. We used the
IBD datasets mentioned in the previous Sect. 3.1 as well. With this method, we based
on the number of features that using Feature Selection using Correlation Matrix with
1078 H. H. Luong et al.

Pearson has filtered out to random. From there we will have the Feature Selection
randomly correspond to Feature Selection using Correlation Matrix with Pearson in
quantity terms. The later process is the same.

3.3 Predict Using K-Fold Cross Validation Method on Both

3.1 and 3.2

For both methods, randomly selection or feature selection using Correlation Matrix
with Pearson, we apply K-Fold cross-validation [23] after finding out Feature Selec-
tion using Correlation Matrix with Pearson or randomly which one predict better.
K-Fold cross-validation is a statistical technique to estimate efficiency of ML model
to independent datasets. The goal of this method is to test the model’s skill to eval-
uate how accurately the model’s prediction result can generate with an unknown
dataset with the purpose of avoiding overfitting [24] or selection bias [25]. Due to
the amount of data is not abundant, we assigned K with values 5 which means divides
datasets into 5 groups. In every group of data, the train data and the test data was
divided by the algorithm so that we can evaluate the effectiveness of the model, and
the remaining groups for training the model. Based on the metrics mention in 4.2.

4 The Experiments

4.1 Dataset Description

Data is originally published on 3 February 2016 [26] which is well-known research

about Fungal microbiota dysbiosis in IBD. In this original research, the authors
extract data from 235 patients with IBD and 38 healthy subjects (HS) was deter-
mined using clinical 16S and ITS2 sequencing [26]. Later on, a paper published on
6 September 2017 provided the data with classification on IBD datasets includes 6
sets of data [27]. In this paper, we use all of them as follows:
• HS_UCr - Healthy (HS) and Ulcerative Colitis (UC) in remission
• HS_iCDr - Healthy (HS) and ileal Crohn’s disease (iCD) in remission
• HS_UCf - Healthy (HS) and Ulcerative Colitis (UC) in flare
• HS_iCDf - Healthy (HS) ileal Crohn’s disease (iCD) in flare
• HS_CDr - Healthy (HS) Crohn’s diseasen(CD) in remission
• HS_CDf - Healthy (HS) Crohn’s diseasen(CD) in flare.
As the purpose of learning these data to predict whether a subject is determined as
IBD patient. We will use one set for train UCr, and tests on iCDr, UCf, iCDf, CDr
and CDf. Information about these sets shows in the Table 1.
Feature Selection Using Correlation Matrix on Metagenomic … 1079

Table 1 Datasets information detail

Information Dataset
UCr iCDr UCf iCDf CDr CDf
Total 238 258 251 248 258 260
features
IBD 44 59 41 44 77 60
patients
Healthy 38 38 38 38 38 38
subject
Total 82 97 79 82 115 98
subjects

4.2 Data Division and Scoring Metrics

In the previous Sect. 4.1, we have been explored 6 informative datasets. Further to the
problem, we divide these datasets into two groups: one group for train and another
group for tests. In the first group, we choose UCr for feature selection later on for the
train. In the other group, five others set later on for tests. For evaluating the results
we used three scoring metrics:
• Accuracy (ACC): ACC metrics results in range from 0.0 to 1.0. The value 1.0
indicates perfect prediction and value 0.0 means that prediction all false. Depend-
ing on the number of correct predictions n corr ect and total predictions n total the
equation accuracy be:
n corr ect
ACC =
n total

• Matthews correlation coefficient (MCC): a widely used measurement in biomed-

ical research [28]. MCC measures result in range −1.0 to +1.0. The value +1.0
shows that prediction is true one hundred percent, the value +0.0 means predic-
tion is no better than random and −1.0 indicates prediction and observation has
nothing to do with each other [29].
• Area Under Curve (AUC): another widely used measurement in biomedical
research. AUC measures the entire area under the curve, values result in range
from 0.0 to 1.0 [30].

4.3 Using Correlation Matrix with Pearson to Select Features

We describe three scenarios for selecting features using Correlation Matrix with
Pearson on the dataset UCr. When all features have the absolute correlation coefficient
greater than n, these features later will be used for the trained model and give out
prediction.
1080 H. H. Luong et al.

Fig. 1 Features with the absolute correlation coefficient >0.1

In the first scenario, we select all features which have n > 0.1 (159 features are
selected). That means a feature has a positive correlation or negative correlation
which is greater than 0.1 will be selected. The matrix is shown in Fig. 1
In the second scenario, we select all features which have n > 0.2 (30 features
are selected). That means a feature has a positive correlation or negative correlation
which is greater than 0.2 will be selected. The matrix is shown in Fig. 2.
In the final scenario, we select all features which have n > 0.3 (10 features are
selected). That means a feature has a positive correlation or negative correlation
which is greater than 0.3 will be selected. The matrix is shown in Fig. 3.

4.4 Prediction Result with ACC, MCC, AUC Metrics

Following Sect. 3, we got three sets of selected features from three scenarios : cor-
relation >0.1 (159 features are selected), correlation >0.2 (30 features are selected)
and correlation >0.3 (10 features are selected). We also take three sets of random
features with the same size, correspondingly. Totally, we got 6 sets from UCr. These
Feature Selection Using Correlation Matrix on Metagenomic … 1081

Fig. 2 Features with the absolute correlation coefficient >0.2

sets will be used for the trained model, then test on the other datasets (iCDr, UCf,
iCDf, CDr and CDf).
After predict using K-fold cross-validation method we got results in three scoring
metrics. These metrics include Accuracy (ACC), Matthews correlation coefficient
(MCC) and Area Under Curve (AUC). The results shown in Tables 2 and 3.

5 Conclusion

As the prediction result has shown, we could state that using the Correlation Matrix on
Metagenomic Data with Pearson for Feature Selection has an extraordinary improve-
ment during predicting IBD patients compared to random features selection. This is
made evident by the number that same size random features are not well suited for
the train prediction model. The way of using the Correlation Matrix to take the high
relate mutually feature to train is better for enhancing the prediction. Further research
based on that premise could lead people to a healthy life with early detection of any
kind of illness not just only IBD.
1082 H. H. Luong et al.

Fig. 3 Features with the absolute correlation coefficient >0.3

Table 2 Results with metrics corresponding ACC, MCC and AUC

Dataset
Scenario Features Metrics iCDr UCf iCDf CDr CDf
corr > 0.1 159 ACC 0.95258 0.97975 0.99268 0.93750 1.00000
MCC 0.90300 0.95957 0.98543 0.87500 0.96875
AUC 0.95540 0.97972 0.99210 0.96875 1.00000
Random 159 ACC 0.90928 0.92911 0.94390 0.87500 0.93750
MCC 0.82009 0.85999 0.88830 0.78125 0.87500
AUC 0.91699 0.92978 0.94377 0.90625 0.93750
corr > 0.2 30 ACC 0.99381 0.99747 1.00000 1.00000 1.00000
MCC 0.98734 0.99499 1.00000 1.00000 1.00000
AUC 0.99491 0.99756 1.00000 1.00000 1.00000
Random 30 ACC 0.87422 0.90886 0.84375 0.78125 0.87500
MCC 0.75353 0.82036 0.71875 0.59375 0.75000
AUC 0.88443 0.90969 0.87500 0.81250 0.87500
corr > 0.3 10 ACC 1.00000 1.00000 1.00000 1.00000 1.00000
MCC 1.00000 1.00000 1.00000 1.00000 1.00000
AUC 1.00000 1.00000 1.00000 1.00000 1.00000
Random 10 ACC 0.72371 0.80253 0.87500 0.75000 0.87500
MCC 0.50249 0.61778 0.78125 0.56250 0.75000
AUC 0.75227 0.80552 0.87500 0.78125 0.87500
Feature Selection Using Correlation Matrix on Metagenomic … 1083

Table 3 Top 10 features with highest correlation score

Feature name Score
k__Bacteria.p__Firmicutes.c__Clostridia.o__Clostridiales.f__.g__ 0.46782
k__Bacteria.p__Firmicutes.c__Clostridia.o__Clostridiales.f__ .Mogibacteriaceae..g__ 0.45053
k__Bacteria.p__Firmicutes.c__Clostridia.o__Clostridiales.f__ Christensenellaceae.g__ 0.38269
k__Bacteria.p__Firmicutes.c__Clostridia.o__Clostridiales.f__ Dehalobacteriaceae.g__ 0.55617
Dehalobacterium
k__Bacteria.p__Firmicutes.c__Clostridia.o__Clostridiales.f__ Lachnospiraceae.g__Coprococcus 0.40403
k__Bacteria.p__Firmicutes.c__Clostridia.o__Clostridiales.f__ Ruminococcaceae.g__ 0.34405
k__Bacteria.p__Firmicutes.c__Clostridia.o__Clostridiales.f__ Ruminococcaceae.Other 0.38699
k__Bacteria.p__Firmicutes.c__Clostridia.o__Clostridiales.Other. Other 0.31592
k__Bacteria.p__Proteobacteria.c__Deltaproteobacteria.o__ Desulfovibrionales.f__ 0.41116
Desulfovibrionaceae.g__Desulfovibrio
k__Bacteria.p__Tenericutes.c__Mollicutes.o__RF39.f__.g__ 0.30068

References

1. Kim ER, Chang DK (2014) Colorectal cancer in inflammatory bowel disease: the risk, patho-
genesis, prevention and diagnosis. World J Gastroenterol 20(29):9872–9881. https://doi.org/
10.3748/wjg.v20.i29.9872
2. Centers for Disease Control and Prevention: Inflammatory bowel disease (IBD), from https://
www.cdc.gov/ibd/what-is-IBD.htm
3. NIH-U.S. National Library of Medicine: Crohn’s Disease, from https://medlineplus.gov/
crohnsdisease.html
4. National Institute of Diabetes and Digestive and Kidney Diseases (NIDDK): Ulcerative Colitis,
from https://www.niddk.nih.gov/health-information/digestive-diseases/ulcerative-colitis
5. NHS: Overview Cancer, from https://www.nhs.uk/conditions/cancer/
6. NIH-National Cancer Institute: Understanding Cancer, from https://www.cancer.gov/about-
cancer/understanding/what-is-cancer
7. American Cancer Society: What Is Cancer? from https://www.cancer.org/cancer/cancer-
basics/what-is-cancer.html
8. (ASCRS) American Society of Colon & Rectal Surgeons: The Colon: What it is, What it
Does and Why it is Important: Overview Cancer, from https://fascrs.org/patients/diseases-
and-conditions/a-z/the-colon-what-it-is,-what-it-does
9. Innerbody: Rectum, from https://www.innerbody.com/image_digeov/dige14-new3.html
10. World Health Organization: Cancer. Retrieved September 28, 2020, from https://www.who.
int/news-room/fact-sheets/detail/cancer
11. American Cancer Society: Key Statistics for Colorectal Cancer. Retrieved September 28, 2020,
from https://www.cancer.org/cancer/colon-rectal-cancer/about/key-statistics.html
12. Vogenberg F, Isaacson Barash C, Pursel M (2010) Personalized medicine: Part 1: Evolution
and development into theranostics. Retrieved September 27, 2020, from https://www.ncbi.nlm.
nih.gov/pmc/articles/PMC2957753/
13. Top 10 Applications of Machine Learning in Healthcare - FWS. (n.d.). Retrieved September
27, 2020, from https://www.flatworldsolutions.com/healthcare/articles/top-10-applications-
of-machine-learning-in-healthcare.php
14. Nguyen TH, Zucker J (2019) Enhancing metagenome-based disease prediction by unsupervised
binning approaches. In: 2019 11th international conference on knowledge and systems engi-
neering (KSE), da nang, Vietnam, 2019, pp 1–5. https://doi.org/10.1109/KSE.2019.8919295
1084 H. H. Luong et al.

15. Nguyen TH, Nguyen TN (2019) Disease prediction using metagenomic data visualizations
based on manifold learning and convolutional neural network. Lecture Notes in Computer
Science, vol 11814. Springer, Cham. https://doi.org/10.1007/978-3-030-35653-8_9
16. O ndov BD, Bergman NH, Phillippy AM (2011) Interactive metagenomic visualization in a web
browser. BMC Bioinform. 12:385. https://doi.org/10.1186/1471-2105-12-385. (ISSN:1471-
2105)
17. Nguyen TH et al (2018) Disease classification in metagenomics with 2D embeddings and deep
learning. In: The annual French conference in machine learning (CAp 2018). France: Rouen;
June 2018. arXiv: 1806.09046
18. Thanh-Hai N, Thai-Nghe N (2020) Diagnosis approaches for colorectal cancer using manifold
learning and deep learning. SN COMPUT. SCI. 1:281
19. Laurens van der Maaten GH (2008) Visualizing data using t-sne. J Mach Learn Res 9:8
20. Nguyen T, Chevaleyre Y, Prifti E, Sokolovska N, Zucker J (2017) Deep learning for metage-
nomic data: using 2D embeddings and convolutional neural networks. arXiv: 1712.00244
21. Benesty J, Chen J, Huang Y, Cohen I (2009) Pearson correlation coefficient. In: Noise reduction
in speech processing. Springer Topics in Signal Processing, vol 2. Springer, Berlin, Heidelberg.
from https://doi.org/10.1007/978-3-642-00296-0_5
22. Correlation Test Between Two Variables in R. (n.d.). Retrieved October 13, 2020, from http://
www.sthda.com/english/wiki/correlation-test-between-two-variables-in-r
23. The ‘K’ in K-fold cross-validation: davide anguita, Luca Ghelardoni, Alessandro Ghio, Luca
Oneto and Sandro Ridella https://www.elen.ucl.ac.be/Proceedings/esann/esannpdf/es2012-
62.pdf
24. Overfitting (2020) Retrieved 13 October 2020, from https://en.wikipedia.org/wiki/Overfitting
25. Selection bias (2020) Retrieved 13 October 2020, from https://en.wikipedia.org/wiki/
Selection_bias
26. Sokol H, Leducq V, Aschard H et al (2017) Gut 66:1039–1048
27. Fioravanti D et al (2018) Phylogenetic convolutional neural networks in metagenomics. BMC
Bioinformatics 19.S2 (2018): n. pag. Crossref. Web
28. Boughorbel S, Jarray F, El-Anbari M (2017) Optimal classifier for imbalanced data using
Matthews Correlation Coefficient metric. PLoS ONE 12(6): e0177678. https://doi.org/10.1371/
journal.pone.0177678
29. En.wikipedia.org (2020) Matthews correlation coefficient. Retrieved October 12, 2020, https://
en.wikipedia.org/wiki/Matthews_correlation_coefficient
30. Huang J, Ling CX (2005) Using AUC and accuracy in evaluating learning algorithms. IEEE
Trans Knowl Data Eng 17(3):299–310. https://doi.org/10.1109/TKDE.2005.50