PME 3132 - Introduction To CM
PME 3132 - Introduction To CM
James K. Kimotho
SoMMME
Department of Mechanical Engineering
Content
Motivation
Motivation
Question: Is it possible to anticipate failures during operation?
Answer: Yes through condition monitoring
What is condition monitoring?
[KAt]
5
0.2 Threshold 0.2 Threshold
Vibration / g
Vibr. Ampl. / g
Vibr. Ampl. / g
0
0.1 0.1
-5
0 0
0 0.5
0 100 200 0 100 200
Time / s Frequency / Hz
Frequency / Hz
Introduction
Advancement of condition monitoring
– Diagnostics: detection, isolation and identification of faults
0.2 Threshold
Vibr. Ampl. / g
0.1
0
0 100 200
[KAt] Frequency / Hz
Introduction
Diagnostic Methods
Diagnosis is an assessment of the current health state of a
system based on observed features from condition monitoring
data.
– It involves identifying anomalous behavior in a system, locating
the component with anomalous behavior and identifying the type
of anomaly in that component.
– In some instances, diagnosis can imply the isolation of a faulty
component, a failure mode, or a failure condition
Two main categories:
– Data driven diagnostics
– Model based diagnostics
Diagnostic Methods
Data-driven Diagnostics
Two sub-categories
– Expert knowledge with signal processing
– Machine learning algorithms
Advanced signal processing coupled with expert knowledge of
the system is used to identify possible faults.
Graphical tools such as power spectrum graph, cepstrum
graph, spectrogram, etc, are employed.
In rotating machinery, fault characteristic frequencies such as
bearing frequency characteristics, gear mesh frequencies,
shaft frequencies and their harmonics, etc, have been
extensively used in fault diagnosis
Diagnostic Methods
Application Example - Rolling Element Bearing Faults
Diagnostic Methods
Diagnostic Methods
Data-driven Diagnostics – Machine Learning
Converts sensor data into useful Advantages
information – No expert knowledge of system
Application of machine learning – Easily adaptable to different
algorithms systems
– Classification Disadvantages
– Clustering – Requires huge amount of data for
- Fault class
- Health States training
– Susceptible to overfitting
Training
Target
Machine
Training Feature
Learning Model
Data Extraction
Algorithm
Online application
New Feature
Model Output
Data Extraction
Diagnostic Methods
Model based Diagnostics
Detailed mathematical models Advantages
– Physics of failure – High accuracy
Empirical based models – Can be implemented at
development stage
Disturbances Faults Disadvantages
𝑑𝑑(𝑡𝑡) 𝑓𝑓(𝑡𝑡) – Requires expert knowledge of
system
Inputs Actual Output
– Computationally intensive
𝑢𝑢(𝑡𝑡) 𝑦𝑦(𝑡𝑡)
– System specific
system
Residual
+ 𝑟𝑟(𝑡𝑡)
𝐺𝐺[𝑢𝑢 ∙ , 𝜃𝜃]
Diagnostics
Model -
Expected Output
𝑦𝑦�(𝑡𝑡)
Prognostic Methods
Model Reliability
based based
Prognostic Methods
Model based prognostics
Detailed mathematical models Advantages
– Physics of failure – High accuracy
Empirical based models – Can be implemented at
development stage
Disturbances Faults Disadvantages
𝑑𝑑(𝑡𝑡) 𝑓𝑓(𝑡𝑡) – Requires expert knowledge of
system
Inputs Actual Output
– Computationally intensive
𝑢𝑢(𝑡𝑡) 𝑦𝑦(𝑡𝑡)
– System specific
system
Residual
+ 𝑟𝑟(𝑡𝑡) 𝑅𝑅𝑅𝑅𝑅𝑅
𝐺𝐺[𝑢𝑢 ∙ , 𝜃𝜃] Damage Damage
Diagnostics
Model - Estimation Propagation
Expected Output
𝑦𝑦�(𝑡𝑡)
Prognostic Methods
Model based prognostics
Prognostic Methods
Data driven based prognostics
Converts sensor data into useful Advantages
information – No expert knowledge of system
Application of machine learning – Easily adaptable to different
algorithms systems
– Classification Disadvantages
– Regression - RUL – Requires huge amount of data for
- Health Index training
- Health States
– Susceptible to overfitting
Training
Target
Machine
Training Feature
Learning Model
Data Extraction
Algorithm
Online application
New Feature
Model Output
Data Extraction
Prognostic Methods
Reliability based prognostics
Utilizes distributions of historical failure times of similar systems
Usually for low risk systems with no sensor network
Commonly employed distributions:
– Weibull
– Exponential
– Lognormal
– Normal
Extensively applied in prognostics of electronic systems
Advantages
– No condition monitoring data is required
– Easy to implement
Disadvantages
– Low accuracy
– No information on incipient faults
– Operating conditions are not taken into account
– Difficult to implement at component level
19 PME 3132/ Condition Monitoring/ J.K. Kimotho / 20.06.2019
SoMMME
Department of Mechanical Engineering
Prognostic Methods
Reliability Based Prognostics
Estimate TTF
-3
10
1 6
Failure PDF
0.8
2
0
0.6 0 100 200 300 400 500
Time (days
Reliability
0.4
0.2
0
0 50 100 150 200 250 300 350 400 450 500
Time (days
Tutorial
Tutorial 3.1
Tutorial
Tutorial 3.2
References
Thank You
for
Your Kind Attention!
Dr.-Ing. James Kuria Kimotho Department of Mechanical Engineering
[email protected] JKUAT
P.O. Box 62000-00200
Tel: Nairobi