Fault Diagnosis of Centrifugal Pump

using Deep Learning

By
Aman Anand (150103008)
P. Gopal (150103052)
Somkuwar Chirag (150103067)

Guided by
Prof. Rajiv Tiwari
Dhruba Jyoti Bordoloi
 Overview

● Binary classification with faults - cavitation and healthy pump.

● Multiclass classification with faults - cracked impeller, cover plate, flow blockage
and healthy pump.
● Used data from centrifugal pump operating on single speed.
● Used Artificial Neural Network, Decision Tree and Random forest models.
● Used data from the CP operating on 8 different speeds(30Hz - 65-Hz)
● Obtained the classwise(faultwise) training and testing accuracy of different faults.
● Built Forecasting Methodology to predict data at all speeds & predicted faults for
the same.
Laboratory Study: Experimental Set-up

● Machine Fault Simulator

● Tri-axial accelerometers
● Line current probes
● Constant DC power source unit
● Data acquisition system
● Centrifugal pumps

Image source : Ref 2

 Laboratory Study: Experimental Set-up (block diagram)

DAU
Faulty Impeller Cover plate (Data Acquision Unit)

Suction valve Discharge valve

Water Vibration signal

Tank in ‘time’ domain

6-step flow control valve Image source : Internet

Introduction: Faults in Centrifugal Pumps
Cavitation Suction Blockage

Impeller
Cover plate

Image source : Internet

 Types Of Classification

Binary Classification Multi Class Classification

Healthy Cracked Cover Flow Healthy

Cavitation Pump Impeller Plate blockage Pump

Image source : Internet

 DATA
PREPROCESSING
Raw Data from the DAQ
Sensor Data (acquired from Accelerometer sensors)

● 3-axis Accelerometers mounted on the monoblock casing and bearing.

● Line Currents.

● The data is in time-frequency domain

● For vibrations are caused by defects and these defects are independent from each other.
Sensor Data for Binary Classification of Cavitation
Defects
Pump in normal function characteristics

Pump with indicated cavitation

 Binary
Classification
 Raw Data

Data Cleaning

Training Model - Classification & Regression Trees (CART)

Predicted Value
Confusion Matrix
Actual Value

0 1 Accuracy
0 138307 71693
64.78%
1 76206 133794
Laboratory Study: Initial Data Analysis
Raw Data

Training Model Training Model

Simple Artificial Neural Network model Classification & Regression Trees (CART)
Laboratory Study: Initial Data Preprocessing

Data acquisition Feature Extraction Feature Selection

Two 3-Axis Raw sensor Data can’t be used Feature selection techniques
Accelerometer are to train the Decision Model help you reduce large data
mounted on the because of the dimensionality sets by eliminating features
centrifugal pump and redundancy of the sensor that are irrelevant to the
data analysis you are trying to
At monoblock pump perform
body and bearing housing Continuous Wavelet
Transform (CWT) is used to Wrapper model is used to
Data is acquired through extract statistical features select relevant features from
a DAQ and saved for the from the sensor data in the features available.
preprocessing time_frequency domain
Precision ( 1/σ )
Mean ( μ )
Standard Deviation ( σ )
Artificial Neural Network

Hidden activation : relu

Output activation : sigmoid

Input Layer Hidden Layer Output Layer

After feature extraction
 MULTICLASS
CLASSIFICATION
Various Faults Speeds considered

class0 - impeller defect Speed 1 - 30Hz

class1 - impeller with flow blockage Speed 2 - 35Hz
class2 - impeller defect with suction blockage
class3 - cover plate defect Speed 3 - 40Hz
class4 - cover plate with flow blockage Speed 4 - 45Hz
class5 - cover plate with suction blockage Speed 5 - 50Hz
class6 - healthy pump with suction blockage Speed 6 - 55Hz
class7 - healthy pump
class8 - healthy pump with flow blockage Speed 7 - 60Hz
Speed 8 - 65Hz
ARTIFICIAL
NEURAL
NETWORK
Artificial Neural Network

Hidden activation : relu

Output activation :
softmax

Dropout
Regularization:0.2

27 20 15 15 9 units

Hidden layers
 ANN

Testing Data by ANN

(Accuracy - 99.30%)
DECISION
TREE
Optimization : K-fold cross
validation used and cp value 1e-09
 Decision Tree

Training Data by Decision Tree Testing Data by Decision Tree

(Accuracy - 96.3047) (Accuracy - 94.55)
RANDOM
FOREST
 Random Forest

Testing Data by Random Forest Node size : 25

(Accuracy - 98.59%) Number of trees : 400
 CLASSWISE
PREDICTIONS
AT ALL SPEEDS
Training Dataset Average Accuracy at Different
Speeds
Speed Average
Accuracy

1 97.40%

2 97.63%

3 97.34%

4 98.47%

5 98.04%

6 97.95%

7 98.47%

8 98.27%
Training Dataset Classwise Accuracy Prediction at
Different Speeds
Training Dataset Classwise Accuracy Prediction at
Different Speeds
Training Dataset Confusion Matrix at Different
Speeds
Training Dataset Confusion Matrix at Different
Speeds

● Non Zero entries other than

Diagonal entries are close to
the Diagonal.

● Even we are getting close to

97.5% accuracy penalty is less.

● If let a pump has 5 defects it

means model spotted 4 defects
or 6 defects, hence penalty is
less.
Testing Dataset Average Accuracy at Different
Speeds
Speed Average
Accuracy

1 97.84%

2 97.57%

3 97.24%

4 98.67%

5 98.05%

6 98.34%

7 99.12%

8 97.91%
Testing Dataset Classwise Accuracy Prediction at
Different Speeds
Testing Dataset Classwise Accuracy Prediction at
Different Speeds
Testing Dataset Confusion Matrix at Different Speeds
Testing Dataset Confusion Matrix at Different Speeds

● Non Zero entries other than

Diagonal entries are close to
the Diagonal.

● Even we are getting close to

97.5% accuracy penalty is less.

● If let a pump has 5 defects it

means model spotted 4 defects
or 6 defects, hence penalty is
less.
 Testing confusion matrix at all speeds

S1 S3

S2 S4
S5 S7

S6 S8
Training Accuracy at all speeds
Speed Acc

1 93.32%

2 94.34%

3 94.23%

4 96.01%

5 96.90%

6 95.93%

7 95.15%

8 94.42%
Testing Accuracy at all speeds

Speed Acc

1 97.09%

2 98.62%

3 98.38%

4 98.78%

5 99.35%

6 98.3%

7 98.94%

8 94.94%
Classwise Testing Accuracy at all speeds
 AFTER
MIDSEM
Forecasting Methodology & Faults
Prediction
 What is Forecasted? ~DATA

Feasible Dimensions out of 9 Dimensions at at all the Speeds

between 30Hz - 65Hz. Basically we are predicting data at all
Speeds without experiments.

What is used?
Experimental data at 8 Speeds;
[ 30, 35, 40, 45, 50, 55, 60, 65 ]
Why to Forecast? ~IMPACT

● Till now we were able to predict faults in pumps operating at

just 8 different speeds for which experimental data was
available.
● Using this method we can predict faults at all the Speeds lying
in the range 30Hz-65Hz.
● ~Results~ - Fault Prediction will have lower accuracy
comparable to the one we got at experimental
Speeds(~99.50%)
What is used throughout the
Process? ~Requisites~
● Measure of Central tendency(Median) of all the 9 Dimensions
of each class of Faults(9 Classes).
● Graphs of Readings Vs Speed of all the 9 Dimensions of each
class of Faults. So in total there are 36 Graphs.
● Euclidean Distance calculation[Concept of Clustering]...
How it is done? ~Algorithm

Part A - Forecasting
~Actual Data~
At
30HZ
Step1: We
took median
value of all
columns for
each Class at
all Speed.
Step2: Now for each
Class we analysed all
the Dimension(
Accl1_XAxis,
Accl1_YAxis,
Accl1_ZAxis,
Accl2_XAxis,
Accl2_YAxis,
Accl2_ZAxis,
Line_Current1,
Line_Current2,
Line_Current3) using
graphs for all Speeds.
 Let’s say we have to
forecast at 36Hz. We
will find the
corresponding points
using the patterns of
each graph. If we don’t
get any pattern then
we reject that graphs.
The analysis of
rejection was carried
out for all speed range
& is mentioned in Table
in Slide 43.
The Predictability from
the graph decides
whether to consider
the Parameter or not.
Step3: Dimensions Rejection
~To increase Accuracy~
Ex- For 47Hz we take just
Accl1_Y,Accl2_Y,Accl1_Z,
Accl2_Z as has all Y in Table
Suppose if in previous case of
47Hz rather than just 4
dimensions we want to
increase dimensions for better Accuracy can be
accuracy. Thus it can be seen increased by
in table that most of the compensation lower
dimensions are rejected just number of faults to be
because of Class1. So if we classified.
didn’t consider Class1 than
we can increase dimensions
from 4 to 8.
Part B - Fault Classification
Given a pump operating forecast the data using
at 36.5 Hz. Graphs.
● We will look into the ● Thus as there are 9
table. Determine classes of faults there
how many & which will be 9 points in 9D.
parameter to ● Now calculate the
consider. Euclidean distance of
● It can be seen that data taken from the
all 9 dimensions can pump to be tested
be taken. We will from these 9 points.
 ~Visualisation~

Lower the Euclidean

distance more will be
the probability of pump
belonging to that class.
Result & Conclusion
1. Data Preprocessing & Feature extraction. Standard Deviation & Precision are the most important
features. Binary Classification considering Healthy Pump & Pump with Cavitation.

2. Multiclass Classification considering cracked impeller, cover plate, flowblockage and healthy
pump & their combinations in a pump.

3. Applied Deep Learning & Machine Learning Algorithm, Artificial Neural Network, Decision Tree
and Random forest models to achieve accuracy closed to 99%. Observed that the testing accuracy
was highest of Artificial Neural Network.

4. Trained model for all the Speeds using both Random Forest & ANN. We calculated the accuracy
of all the predicted faults at each speed classwise from both the model.

5. We have figured out the method using which for any Speed in the range 30-65 Hz, given a pump
for which we have to check which Fault it has, we can check, even if we don’t have experimental
data taken on that particular Speed.
References
1. Y Zhou, C Liu, P Zhao, B SUN, W. HONG,"Fault diagnosis methods for centrifugal pump based on autoregressive and
continuous hidden Markov model" in Proceeding of the CSEE, 2008

2. S Rapur, R Tiwari, "Multi Fault Diagnosis of Centrifugal Pumps with Time, Frequency and Wavelet-Based Features Using
Support Vector Machines", in Jan 2019 (Proceedings of the 10th International Conference on Rotor Dynamics – IFToMM)

3. W Abdulkareem, R A Kannan, and K F. Al-Raheem, "Centrifugal Pump Impeller Crack Detection Using Vibration
Analysis", 2nd International Conference on Research in Science, Engineering and Technology (ICRSET’2014), March 21-
22, 2014 Dubai

4. Y. Luo, H. Sun, S. Yuan, and J. Yuan, "Research on Statistical Characteristics of Vibration in Centrifugal Pump",Revista
Técnica de la Facultad de Ingeniería Universidad del Zulia, vol. 38, 1, pp.49 - 61

5. Zouari R, Sieg-Zieba S, Sidahmed M. "Fault Detection system for centrifugal pumps using neural network and neural-fuzzy
techniques". In: Surveillance 5 CETIM Senlis 11-13 October 2004

6. N. R. Sakthivel, V. Sugumaran, and B. B. Nair, Comparison of decision tree-fuzzy and rough set-fuzzy methods for fault
categorization of monoblock centrifugal pump", Mechanical Systems and Signal Processing, vol. 24, 8, pp. 1887-1906, 2010.

7. S. Farokhzad, "Vibration Based Fault Detection of Centrifugal Pump by Fast Fourier Transform and Adaptive Neuro-Fuzzy
Inference System", Journal of Mechanical Engineering and Technology, vol. 1, 3, pp. 82-87 2013
THANK YOU