automatic fault diagnosis of rolling element bearings...
TRANSCRIPT
Automatic Fault Diagnosis of Rolling Element
Bearings Using Wavelet Based Pursuit Features
Hongyu Yang
Bachelor of Engineering (DUT)*
Master of Engineering (DUT)
* Dalian University of Technology, China
Thesis submitted in total fulfilment of the requirements of the degree of
Doctor of Philosophy
School of Mechanical, Manufacturing, and Medical Engineering
Faculty of Built Environmental Engineering
Queensland University of Technology
8th October 2004
STATEMENT OF ORIGINALITY
____________________________________________________________________
I
STATEMENT OF ORIGINALITY
This thesis contains no material which has been accepted for the award of any other
degree or diploma in any university, and to the best of my knowledge it contains no
material previously published or written by another person, except where due
reference is made in the text of the thesis.
Signed :_____________ Date:____________
Hongyu Yang
School of Mechanical, Manufacturing, and
Medical Engineering
Queensland University of Technology;
Gardens Point, Brisbane, Queensland,
4001
Australia
ACKNOWLEDGEMENTS
____________________________________________________________________
II
To my loving parents, Yang Yutian, and Jiang Guiyun.
ACKNOWLEDGEMENTS
____________________________________________________________________
III
ACKNOWLEDGEMENTS
I wish to express my sincere appreciation to the following people for their support
and contributions:
My families for their emotional support and encouragement.
My supervisors, Prof. Joseph Mathew and Dr. Lin Ma, for their technical
supervision, and general support and advice.
Dr Vladis Kosse for providing access to the bearing test rig located in the School of
Mechanical, Manufacturing and Medical Engineering at Queensland University of
Technology.
My friends for their encouragement.
Hongyu Yang
School of Mechanical, Manufacturing, and
Medical Engineering
Queensland University of Technology;
Gardens Point, Brisbane, Queensland,
4001
Australia
TABLE OF CONTENTS
____________________________________________________________________
IV
TABLE OF CONTENTS
STATEMENT OF ORIGINALITY ..........................................................................I
ACKNOWLEDGEMENTS......................................................................................II
TABLE OF CONTENTS......................................................................................... IV
LIST OF FIGURES ..............................................................................................VIII
LIST OF TABLES ................................................................................................XIII
NOMENCLATURE..............................................................................................XIV
ABSTRACT ....................................................................................................... XVII
CHAPTER 1. GENERAL INTRODUCTION ........................................................ 1
1.1 INTRODUCTION.............................................................................................. 1
1.2 OBJECTIVES.................................................................................................. 1
1.3 SIGNIFICANCE ............................................................................................... 2
1.4 SCOPE OF RESEARCH..................................................................................... 3
1.5 ORIGINALITY OF RESEARCH.......................................................................... 4
1.6 ORGANISATION OF THESIS............................................................................ 6
CHAPTER 2. LITERATURE REVIEW ................................................................. 8
2.1 INTRODUCTION.............................................................................................. 8
2.2 BACKGROUND............................................................................................... 8
2.2.1 CONDITION MONITORING AND FAULT DIAGNOSIS................................................................8
2.2.2 ARTIFICIAL INTELLIGENCE .................................................................................................12
2.3 LITERATURE ON FAULT DIAGNOSIS OF ROTATING MACHINERY.................... 15
2.3.1 FEATURE EXTRACTION FOR FAULT DIAGNOSIS OF ROTATING MACHINERY .......................16
2.3.2 ARTIFICIAL INTELLIGENCE TECHNIQUES IN FAULT IDENTIFICATION ..................................42
2.3.3 ARTIFICIAL INTELLIGENCE & WAVELET TRANSFORM FOR FAULT DIAGNOSIS...................56
2.4 CONCLUSION OF REVIEW ............................................................................ 59
CHAPTER 3. ADVANCED FEATURE EXTRACTION TECHNIQUES .. ...... 61
TABLE OF CONTENTS
____________________________________________________________________
V
3.1 INTRODUCTION............................................................................................ 61
3.2 BEST BASIS DISCRETE WAVELET PACKET ANALYSIS ................................. 62
3.2.1 INTRODUCTION TO DIFFERENT WAVELET FUNCTIONS........................................................66
3.2.2 SELECTION OF BEST BASIS .................................................................................................74
3.3 ADAPTIVE APPROXIMATION WITH PURSUIT................................................ 77
3.3.1 FUNDAMENTALS OF ADAPTIVE APPROXIMATION WITH PURSUIT........................................77
3.3.2 MATCHING PURSUIT ...........................................................................................................84
3.3.3 BASIS PURSUIT ...................................................................................................................85
3.4 SUMMARY ................................................................................................... 89
CHAPTER 4. AUTOMATIC DIAGNOSIS SCHEMA.............. .......................... 91
4.1 INTRODUCTION............................................................................................ 91
4.2 AUTOMATIC FAULT DIAGNOSIS USING SPECTRUM ANALYSIS .................... 92
4.3 AUTOMATIC FAULT DIAGNOSIS BASED ON DWPA..................................... 93
4.4 AUTOMATIC FAULT DIAGNOSIS USING PURSUIT......................................... 96
4.5 DESIGN OF THE FEED FORWARD NEURAL NETWORK CLASSIFIER............... 97
4.5.1 FEED FORWARD NEURAL NETWORKS.................................................................................97
4.5.2 DESIGN OF THE STRUCTURE OF THE FEED FORWARD NEURAL NETWORK CLASSIFIERS...100
4.6 SUMMARY ................................................................................................. 101
CHAPTER 5. SIMULATION AND EXPERIMENT ............... .......................... 103
5.1 INTRODUCTION.......................................................................................... 103
5.2 SIMULATED SIGNALS ................................................................................ 103
5.3 TEST RIG AND EXPERIMENT PROCEDURE.................................................. 104
5.4 FAULT SIMULATION .................................................................................. 106
5.5 SUMMARY ................................................................................................. 109
CHAPTER 6. RESULTS AND DISCUSSION.................................................... 110
6.1 INTRODUCTION.......................................................................................... 110
TABLE OF CONTENTS
____________________________________________________________________
VI
6.2 ANALYSIS USING DWPA, MATCHING PURSUIT, AND BASIS PURSUIT....... 111
6.2.1 TIME-FREQUENCY ANALYSIS OF SIMULATED SIGNALS ....................................................111
6.2.2 TIME-FREQUENCY ANALYSIS OF THE SIGNALS OF BEARINGS WITH MARGINAL FAULTS ..121
6.2.3 BASIS PURSUIT AND BEST BASIS DWPA AND MATCHING PURSUIT – A COMPARISON.....128
6.2.4 SEVERITY OF BEARING FAULTS ANALYSED USING BASIS PURSUIT..................................130
6.2.5 BASIS PURSUIT DENOISING...............................................................................................135
6.3 AUTOMATIC FAULT DIAGNOSIS ................................................................. 137
6.3.1 SPECTRUM BASED AUTOMATIC FAULT DIAGNOSIS..........................................................137
6.3.2 DWPA FEATURE BASED AUTOMATIC FAULT DIAGNOSIS................................................146
6.3.3 MATCHING PURSUIT FEATURE BASED AUTOMATIC FAULT DIAGNOSIS............................155
6.3.4 BASIS PURSUIT FEATURE BASED AUTOMATIC FAULT DIAGNOSIS....................................161
6.4 DISCUSSION AND CONCLUSION................................................................. 166
6.5 SUMMARY ................................................................................................. 167
CHAPTER 7. CONCLUSION.............................................................................. 168
7.1 INTRODUCTION.......................................................................................... 168
7.2 THE IMPROVED DWPA, MATCHING PURSUIT, AND BASIS PURSUIT ......... 168
7.3 AUTOMATIC DIAGNOSIS USING SPECTRUM............................................... 169
7.4 AUTOMATIC DIAGNOSIS USING DWPA.................................................... 169
7.5 AUTOMATIC DIAGNOSIS USING MATCHING PURSUIT................................ 170
7.6 AUTOMATIC DIAGNOSIS USING BASIS PURSUIT........................................ 170
CHAPTER 8. FUTURE RESEARCH ................................................................. 171
8.1 SIGNAL PROCESSING TECHNIQUES FOR FEATURE EXTRACTION ................ 171
8.2 ARTIFICIAL INTELLIGENCE FOR DIAGNOSIS.............................................. 171
8.3 INCIPIENT FAULT DETECTION USING TIME-FREQUENCY ANALYSIS
TECHNIQUES......................................................................................................... 172
8.4 AUTOMATED DIAGNOSIS OF PROCESS MONITORING AND MATERIAL
DEGRADATION. .....................................................................................................172
TABLE OF CONTENTS
____________________________________________________________________
VII
8.5 AUTOMATIC DIAGNOSIS OF TRANSMISSION SYSTEMS................................ 172
8.6 COMMERCIALIZING AN INTEGRAL INTELLIGENT DIAGNOSTIC TOOLBOX.... 173
REFERENCES....................................................................................................... 174
PUBLICATIONS ................................................................................................... 195
GLOSSARY .......................................................................................................... 196
APPENDIX .......................................................................................................... 199
LIST OF FIGURES
____________________________________________________________________
VIII
LIST OF FIGURES
Figure 2.1: Machine life bathtub curve........................................................................ 9
Figure 2.2: Overall levels of a bearing of continuing phases of failure..................... 10
Figure 2.3: Spectral characteristics of different stage bearing faults ......................... 12
Figure 2.4: Conceptual representation of a pattern recognition problem................... 14
Figure 2.5: Architecture of a Neural Network ........................................................... 14
Figure 2.6: Fault diagnosis -an overview................................................................... 15
Figure 2.7: An overview of feature extraction techniques......................................... 18
Figure 2.8: An overview of time domain feature extraction techniques.................... 19
Figure 2.9: An overview of fault detection and identification techniques................. 44
Figure 2.10: A Recurrent Neural Network................................................................. 46
Figure 2.11: One Dimensional Self Organising Map ................................................ 47
Figure 2.12: Two dimensional Self Organising Map................................................. 47
Figure 2.13: A Zero-Order Sugeno Fuzzy Model...................................................... 54
Figure 3.1: Filter bank representation of DWT and DWPA [149] ............................ 66
Figure 3.2: An example tree of wavelet packet decomposition [91] ......................... 67
Figure 3.3: Harr Wavelet............................................................................................ 67
Figure 3.4: Meyer wavelet function........................................................................... 69
Figure 3.5: Coiflets wavelet function......................................................................... 70
Figure 3.6: Daubechies function ................................................................................ 71
Figure 3.7: Symlet function ....................................................................................... 72
Figure 3.8: An example of the best tree DWPA decomposition of a signal (with depth
position index).................................................................................................... 76
Figure 3.9: An example of the best tree DWPA decomposition of a signal (with
Entropy value index).......................................................................................... 76
Figure 3.10: An example of a time-frequency atom plot [151] ................................. 82
Figure 4.1: Automatic fault diagnosis procedure using spectrum, spectrogram with
NN...................................................................................................................... 92
LIST OF FIGURES
____________________________________________________________________
IX
Figure 4.2: Automatic fault diagnosis procedure using DWPA and NN................... 94
Figure 4.3: Symlet8 function with a few scales and locations................................... 95
Figure 4.4: Automatic fault diagnosis procedure using Basis Pursuit (or Matching
Pursuit) and NN ................................................................................................. 96
Figure 4.5: Sigmoid function [152]............................................................................ 98
Figure 4.6: A single output FFNN ........................................................................... 100
Figure 4.7: A multi output FFNN ............................................................................ 101
Figure 5.1: Experimental apparatus ......................................................................... 104
Figure 5.2: Test rig with V-belt load........................................................................ 105
Figure 5.3: Test rig without load.............................................................................. 105
Figure 5.4: Experimental apparatus ......................................................................... 106
Figure 5.5: Fault simulation of SKF 6205 ............................................................... 107
Figure 5.6: Fault simulation of KOYO 6201 RS ..................................................... 107
Figure 6.1: Simulated impulse signal 1y ................................................................. 112
Figure 6.2: Basis Pursuit TF plane of 1y ................................................................. 112
Figure 6.3: Best basis DWPA TF plane of 1y ......................................................... 113
Figure 6.4: Matching Pursuit TF plane of 1y .......................................................... 113
Figure 6.5: Simulated impulse signal 1y with noise................................................ 114
Figure 6.6: Basis Pursuit TF plane of 1y with noise................................................ 115
Figure 6.7: DWPA plane of 1y with noise .............................................................. 115
Figure 6.8: Matching Pursuit Plane of 1y with noise .............................................. 116
Figure 6.9. Simulated impulse signal 2y ................................................................. 117
Figure 6.10: Basis Pursuit TF plane of 2y ............................................................... 117
Figure 6.11: DWPA TF plane of 2y ........................................................................ 118
Figure 6.12: Matching Pursuit TF plane of 2y ........................................................ 118
Figure 6.13: Simulated impulse signal 2y with noise ............................................. 119
Figure 6.14: Basis Pursuit TF Plane of signal 2y with noise ................................... 120
Figure 6.15: DWPA TF plane of signal 2y with noise............................................. 120
LIST OF FIGURES
____________________________________________________________________
X
Figure 6.16: Matching Pursuit TF plane of signal 2y with noise ............................. 121
Figure 6.17: Vibration signal of a bearing with normal condition........................... 122
Figure 6.18: Basis Pursuit TF plane......................................................................... 122
Figure 6.19: Best basis DWPA TF plane.................................................................123
Figure 6.20: Matching Pursuit TF Plane.................................................................. 123
Figure 6.21: Vibration signal of a bearing with IRF................................................ 124
Figure 6.22: Basis Pursuit TF plane......................................................................... 124
Figure 6.23: Best basis DWPA TF plane.................................................................125
Figure 6.24: Matching Pursuit TF Plane.................................................................. 125
Figure 6.25: Vibration signal of a bearing with ORF.............................................. 126
Figure 6.26: Basis Pursuit TF plane......................................................................... 126
Figure 6.27: Best basis DWPA TF plane.................................................................127
Figure 6.28: Matching Pursuit TF plane .................................................................. 127
Figure 6.29: The time waveform and its Basis Pursuit TF plane obtained from the
bearing with 0.07 inch EDM IRF..................................................................... 131
Figure 6.30: The time waveform and its Basis Pursuit TF plane obtained from the
bearing with 0.14 inch EDM IRF..................................................................... 131
Figure 6.31: The time waveform and its Basis Pursuit TF plane obtained from the
bearing with 0.21 inch EDM IRF..................................................................... 132
Figure 6.32: The time waveform and its Basis Pursuit TF plane obtained from the
bearing with 0.28 inch EDM IRF..................................................................... 132
Figure 6.33: The time waveform and its Basis Pursuit TF plane obtained from the
bearing with a crack in inner race ....................................................................133
Figure 6.34: The time waveform and its Basis Pursuit TF plane of the bearing with
ORF: 0.07inch.................................................................................................. 133
Figure 6.35: The time waveform and its Basis Pursuit TF plane of the bearing with
ORF: 0.14 inch................................................................................................. 134
Figure 6.36: The time waveform and its Basis Pursuit TF plane of the bearing with
ORF: 0.21inch.................................................................................................. 134
Figure 6.37: The time waveform and its Basis Pursuit TF plane of the bearing with
LIST OF FIGURES
____________________________________________________________________
XI
ORF: crack ....................................................................................................... 135
Figure 6.38: The time waveforms and the Basis Pursuit denoised signals: (a), (b)
normal, (c), (d) with IRF (e), (f) with ORF...................................................... 136
Figure 6.39: Spectrum of the signal of a normal bearing......................................... 138
Figure 6.40: Spectrum of the signal of a bearing with IRF...................................... 138
Figure 6.41: Spectrum of the signal of a bearing with ORF .................................... 139
Figure 6.42: Spectrum of the signal of a bearing with REF..................................... 139
Figure 6.43: Feature vectors based on Spectrum of the signal of a normal bearing 140
Figure 6.44: Feature vector based on Spectrum of the signal of a bearing with IRF140
Figure 6.45: Feature vector based on spectrum of the signal of a bearing with ORF141
Figure 6.46: Feature vector based on Spectrum of the signal of a bearing with REF141
Figure 6.47: Spectrogram of the signal of a normal bearing.................................... 142
Figure 6.48: Spectrogram of the signal of a bearing with IRF................................. 142
Figure 6.49: Spectrogram of the signal of a bearing with ORF............................... 143
Figure 6.50: Spectrogram of the signal of a bearing with REF ............................... 143
Figure 6.51: Feature vector based on spectrogram of the signal of a normal bearing144
Figure 6.52: Spectrogram feature of the signal of a bearing with IRF .................... 144
Figure 6.53: Spectrogram feature of the signal of a bearing with ORF................... 145
Figure 6.54: Spectrogram feature of the signal of a bearing with REF ................... 145
Figure 6.55: The DWPA of the signal of a normal bearing..................................... 148
Figure 6.56: The DWPA of the signal of a bearing with IRF.................................. 148
Figure 6.57: The DWPA of the signal of a bearing with ORF ................................ 149
Figure 6.58: The DWPA of the signal of a bearing with REF................................. 149
Figure 6.59: Features based on wavelet packets: Mean Value ................................ 150
Figure 6.60: Features based on wavelet packets: Variance...................................... 150
Figure 6.61: Features based on wavelet packets: Skewness .................................... 151
Figure 6.62: Features based on wavelet packets: Kurtosis ...................................... 151
Figure 6.63: Features based on wavelet packets: Energy ........................................ 152
LIST OF FIGURES
____________________________________________________________________
XII
Figure 6.64: Features based on wavelet packets: Root Mean Square...................... 152
Figure 6.65: Features based on wavelet packets: Crest Factor ................................ 153
Figure 6.66: Features based on wavelet packets: Matched Filter ............................ 153
Figure 6.67: The Matching Pursuit of the vibration signal of a bearing under
condition: Normal ............................................................................................ 157
Figure 6.68: The Matching Pursuit of the vibration signal of a bearing under
condition: ORF................................................................................................. 158
Figure 6.69: The Matching Pursuit of the vibration signal of bearing under condition:
IRF ................................................................................................................... 158
Figure 6.70: The Matching Pursuit of the vibration signal of a bearing under
condition: REF ................................................................................................. 159
Figure 6.71: The Matching Pursuit (MP) coefficients of vibration signals of bearings
under conditions: (a) Normal (b) ORF (c) IRF (d) REF .................................. 159
Figure 6.72: The Matching Pursuit Features of Vibration Signals of bearings under
conditions: (a) Normal (b) ORF (c) IRF (d) REF ............................................ 160
Figure 6.73: The Basis Pursuit of the vibration signals of bearings under condition:
Normal ............................................................................................................. 163
Figure 6.74: The Basis Pursuit of the vibration signals of bearings under condition:
ORF.................................................................................................................. 163
Figure 6.75: The Basis Pursuit of the vibration signals of bearings under condition:
IRF ................................................................................................................... 164
Figure 6.76: The Basis Pursuit of the vibration signals of bearings under condition:
REF .................................................................................................................. 164
Figure 6.77: The Basis Pursuit coefficients of vibration signals of bearings under
conditions: (a) Normal (b) ORF (c) IRF (d) REF ............................................ 165
Figure 6.78: The Basis Pursuit features of vibration signals of bearings under
conditions: (a) Normal (b) ORF (c) IRF (d) REF ............................................ 165
LIST OF TABLES
____________________________________________________________________
XIII
LIST OF TABLES
Table 2.1: An overview of frequency techniques and time-frequency techniques.... 34
Table 2.2 Neural Networks applied in diagnosing rotating machinery faults ........... 48
Table 3.1 Summary of Wavelet Families and Associated Properties (Manual of
wavelet toolbox in matlab)................................................................................. 73
Table 5.1 Fault specification for the bearings KOYO 6201 RS. ............................. 108
Table 5.2 Drive end bearing KOYO 6201 (Size in mm) ......................................... 108
Table 5.3 Drive end bearing: 6204 SKF, deep groove ball bearing......................... 108
Table 5.4 Drive end bearing: 6205-2RS JEM SKF, deep groove ball bearing (Size in
inches) .............................................................................................................. 109
Table 5.5 Fault Specifications for 6204-2RS JEM SKF (All dimension in inches) 109
Table 6.1 Fault severity specifications in Figures 6.29-37 ...................................... 136
Table 6.2 Signal to Noise Ratio of the Original and the BP Denoised signals ........ 137
Table 6.3 Classification performance of automatic diagnosis based on Spectrum and
Spectrogram ..................................................................................................... 146
Table 6.4 Performance of the single output FFNN using DWPA features.............. 154
Table 6.5 Performance of the multi output FFNN using DWPA features............... 155
Table 6.6 Classification performance of different procedures using Matching Pursuit160
Table 6.7 Classification performance of different procedures using Basis Pursuit . 166
NOMENCLATURE
____________________________________________________________________
XIV
NOMENCLATURE
AI Artificial Intelligence
ANN Artificial Neural Network
x(t) A time signal
( )ωF Fast Fourier Transform
STFT Short Time Fourier Transform
( )ftSPECx , Spectrogram
µ Mean value
sN Sampling length
ix i th sample of the series of vibration data
RMS Root Mean Square
σ Variance
E Energy
ψ Wavelet
cf Crest Factor
Mfrms Matched Filter
Ai(first) Amplitude of the i th frequency component of the first data set
S A scale parameter
U Translation
g(x) Activation function
P Vector
R Dimension
CWT Continuous Wavelet Transform
NOMENCLATURE
____________________________________________________________________
XV
DWPA Discrete Wavelet Packet Analysis
j Different levels of wavelets
k Number of wavelets in each level
E(s) Entropy
γφ Elements
Γ An over complete dictionary
γ Index of a setΓ
γα Coefficient of the elementγφ
m Order of decomposition
( )mR A residual
( )tγψ Wavelet atoms
( )ωψ γˆ Fourier transform of ( )tγψ
ξ Demodulation
nξ Frequency parameter
0ξ Constant
( )kpj ,,=γ Index of wavelet packet dictionary
Φ Dictionary
O(n) Complexity
MP Matching Pursuit
BP Basis Pursuit
BPFF Back Propagation for Feed Forward Networks
FFNN Feed Forward Neural Network
RNN Recurrent Neural Network
RBF Radial Basis Function
NOMENCLATURE
____________________________________________________________________
XVI
MLP Multi Layer Perceptron
SOM Self Organising Maps
LVQ Learning Vector Quantization
SVM Support Vector Machines
rd Diameter of rolling elements
cd Diameter of the cage
outd Diameter of the outer race
ind Diameter of the inner race
α Contact angle between the rolling elements and rolling surfaces
BPFI Ball-pass frequency on the inner race
BSF Rotational frequency of the rolling elements
BPFO Ball-pass frequency on the outer race
IRF Inner Race Fault
ORF Outer Race Fault
REF Rolling Element Fault
ABSTRACT
____________________________________________________________________
XVII
ABSTRACT
Today’s industry uses increasingly complex machines, some with extremely
demanding performance criteria. Failed machines can lead to economic loss and
safety problems due to unexpected production stoppages. Fault diagnosis in the
condition monitoring of these machines is crucial for increasing machinery
availability and reliability.
Fault diagnosis of machinery is often a difficult and daunting task. To be truly
effective, the process needs to be automated to reduce the reliance on manual data
interpretation. It is the aim of this research to automate this process using data from
machinery vibrations. This thesis focuses on the design, development, and
application of an automatic diagnosis procedure for rolling element bearing faults.
Rolling element bearings are representative elements in most industrial rotating
machinery. Besides, these elements can also be tested economically in the laboratory
using relatively simple test rigs.
Novel modern signal processing methods were applied to vibration signals collected
from rolling element tests to destruction. These included three advanced time-
frequency signal processing techniques, best basis Discrete Wavelet Packet Analysis
(DWPA), Matching Pursuit (MP), and Basis Pursuit (BP). This research presents
the first application of the Basis Pursuit to successfully diagnosing rolling
element faults. Meanwhile, Best basis DWPA and Matching Pursuit were also
benchmarked with the Basis Pursuit, and further extended using some novel
ideas particularly on the extraction of defect related features.
The DWPA was researched in two aspects: i) selecting a suitable wavelet, and ii)
choosing a best basis. To choose the most appropriate wavelet function and
decomposition tree of best basis in bearing fault diagnostics, several different
wavelets and decomposition trees for best basis determination were applied and
comparisons made. The Matching Pursuit and Basis Pursuit techniques were effected
by choosing a powerful wavelet packet dictionary. These algorithms were also
studied in their ability to extract precise features as well as their speed in achieving a
result. The advantage and disadvantage of these techniques for feature extraction of
bearing faults were further evaluated.
ABSTRACT
____________________________________________________________________
XVIII
An additional contribution of this thesis is the automation of fault diagnosis by
using Artificial Neural Networks (ANNs). Most of work presented in the current
literature has been concerned with the use of a standard pre-processing technique -
the spectrum. This research employed additional pre-processing techniques such as
the spectrogram and DWPA based Kurtosis, as well as the MP and BP features that
were subsequently incorporated into ANN classifiers. Discrete Wavelet Packets and
Spectra, were derived to extract features by calculating RMS (root mean square),
Crest Factor, Variance, Skewness, Kurtosis, and Matched Filter. Certain spikes in
Matching Pursuit analysis and Basis Pursuit analysis were also used as features.
These various alternative methods of pre-processing for feature extraction were
tested, and evaluated with the criteria of the classification performance of Neural
Networks.
Numerous experimental tests were conducted to simulate the real world environment.
The data were obtained from a variety of bearings with a series of fault severities.
The mechanism of bearing fault development was analysed and further modelled to
evaluate the performance of this research methodology.
The results of the researched methodology are presented, discussed, and evaluated in
the results and discussion chapter of this thesis. The Basis Pursuit technique proved
to be effective in diagnostic tasks. The applied Neural Network classifiers were
designed as multi layer Feed Forward Neural Networks. Using these Neural
Networks, automatic diagnosis methods based on spectrum analysis, DWPA,
Matching Pursuit, and Basis Pursuit proved to be effective in diagnosing different
conditions such as normal bearings, bearings with inner race and outer race faults,
and rolling element faults, with high accuracy.
Future research topics are proposed in the final chapter of the thesis to provide
perspectives and suggestions for advancing research into fault diagnosis and
condition monitoring.
Keywords: Rolling element bearing, Fault diagnosis, Feature extraction, Discrete
Wavelet Packet Analysis (DWPA), Matching Pursuit, Basis Pursuit, Neural Network.
CHAPTER 1. GENERAL INTRODUCTION
____________________________________________________________________
1
CHAPTER 1. GENERAL INTRODUCTION
1.1 Introduction
Today’s industry uses increasingly complex rotating machines, some with extremely
demanding performance criteria. Attempting to diagnose faults in these systems is
often a difficult and daunting task for operators and plant maintainers. Machine
failure can lead to economic loss and safety problems due to unexpected and sudden
production stoppages.
Rotating machinery is a common class of machinery in industry. The root cause of
faults in rotating machinery is often faulty rolling element bearings. One way to
increase operational reliability and thereby increase machine availability is to
monitor faults in these bearings.
Fault diagnosis techniques are crucial for monitoring conditions in bearings. Current
fault diagnosis techniques have a variety of limitations. Methods that are more
effective need to be researched and developed for industrial machinery diagnostic
activities.
This work presented in the adjoining sections of this thesis is a thorough
investigation into application of selected wavelet based automatic fault diagnostic
techniques to non-stationary signals collected from bearings.
This chapter provides the description of the objectives, significance, and scope of
this research. The originality of this work and its contribution to the overall field of
fault diagnosis is also presented.
1.2 Objectives
The main aim of this research was to develop novel signal processing methods to
enable automated diagnosis of rotating machinery. The work program comprised
synthesis of techniques from the fields of pattern recognition and Neural Networks,
with application to condition monitoring. Further detailed objectives of the research
were:-
(1) Novel techniques in the field of signal processing were developed and applied in
feature extraction for machinery faults diagnosis, which enabled features to be
CHAPTER 1. GENERAL INTRODUCTION
____________________________________________________________________
2
extracted effectively and clearly to reduce the dependency of fault diagnosis on well-
trained technicians. In particular, wavelet packets based methods were investigated,
improved, and applied from the perspective of:
• attaining better time-frequency resolution, which assists in avoiding missing
information and increasing information accuracy;
• wavelets that are suitable for vibration analysis. These wavelets were selected on
the basis of their resemblance to real characteristics of vibration signals of
rotating machinery.
Transient characteristics of vibration data of rolling element bearings are usually
represented more comprehensively using wavelet based methods than conventional
time or frequency analysis techniques. In this study, features in signals were
extracted and selected out, which were often impulsive components in signals. These
impulsive components were generated by defects on the inner and outer races, and
rolling elements of bearings. The wavelet based methods were targeted to extract
these components from an overall signal containing noisy components. The
resonance and defect related frequencies of bearings were distinguishable for rotating
machinery fault diagnosis.
(2) Automated fault diagnosis techniques were developed to try to reduce the
dependency on human interpretation for the task of fault diagnosis. Broadly, Neural
Networks were used to classify faults in bearings. Algorithms derived from
Objective (1) were subsequently combined with Neural Networks in the development
of the automatic diagnosis procedure.
1.3 Significance
Activities for condition monitoring and fault diagnosis such as observation and
periodic maintenance are labour-intensive and often unreliable. Some faults, for
instance, lack a flexible, an easy to use and efficient apparatus for fault detection and
can only be distinguished through the observation or audition of experts. This makes
maintenance costly and increases the possibility of undetected faults. More seriously,
a run-to-failure program and periodic maintenance often unnecessarily interrupt
production, thus dramatically reducing industry profits. Machines need to be
CHAPTER 1. GENERAL INTRODUCTION
____________________________________________________________________
3
monitored during the production process to improve machine operation reliability
and reduce unavailability. Therefore, conducting effective condition monitoring
brings significant benefits to industry [1, 2]. The field of condition monitoring has
attracted the increasing attention of engineers in the quest to improve the economic
efficiency and safety of machinery operation. However, condition monitoring
requires effective fault diagnosis, which is a labour-oriented practice to this day.
Without effective and efficient diagnosis, one is unable to make reliable prediction of
lead-time to failure.
The automation of this labour-oriented process of diagnosis by implementing
intelligent diagnosis strategies is a natural procedure, which helps experts and
technicians to be relieved from this relatively expensive task. This research has the
following significance:
(1)The diagnosis of machine faults conventionally requires human interpretation for
the large amounts of data collected from an operating machine. This research
automated and facilitated the procedure of machine fault diagnosis. Humans can
therefore be relieved from laborious work. Consequently, the productivity and life of
diagnosed machines can be improved, and profits for industry increased.
(2)Automated diagnosis renders safe and reliable asset management. Labour-oriented
and unpredictable faults are the difficulties that confront asset managers. Automated
diagnosis requires less labour and makes faults predictable, thus reducing the effort
required to assign staff to monitor machines. In addition, automated diagnosis can
greatly assist the task of prognosing condition, which is a necessary step towards
residual life estimation.
1.4 Scope of Research
This thesis mainly presents research into novel methods for vibration based condition
monitoring fault diagnosis of rotating machinery. It is acknowledged that other
techniques for condition monitoring and fault diagnosis such as lubrication,
temperature, and operation performance, are also necessary technologies in condition
monitoring, but lie outside the scope of this research.
CHAPTER 1. GENERAL INTRODUCTION
____________________________________________________________________
4
The focus of this work was based on rotating machinery faults and rolling element
bearings in particular.
The new methods in this research were developed to represent non-stationary
vibration signals in the time-frequency domain. The time-frequency analysis was
mainly studied in terms of making the most use of wavelets in the application to fault
diagnosis of rolling element bearings. The methods included improved Discrete
Wavelet Packet Analysis (DWPA), Matching Pursuit, and Basis Pursuit - a novel
time-frequency analysis. Initial evaluation of these techniques were conducted using
simulated signals.
The time-frequency analysis methods were used to extract fault related features.
These features were presented in a distinguishable manner for the purpose of direct
interpretation of rolling element bearing faults. Other factors, which were irrelevant
to bearing faults, were not pursued.
Noise removal was also investigated using the Basis Pursuit. The new method and
improved methods were tested using both the simulated signals and experimental
data.
Neural Network techniques were applied to automatically detect and diagnose
bearing faults including Outer Race Fault (ORF), Inner Race Fault (IRF), and
Rolling Element Fault (REF).
Feed Forward Neural Networks (FFNN) were designed and tested to classify the
bearing faults among various types of Neural Networks. Other types of Neural
Networks such as radial basis function, and Self Organised Maps Neural Networks
were not attempted in this research. The reasons for this selected are described in
Section 4.5.
1.5 Originality of Research
The originality and contribution of this research are as follows:
(1) The advancement of time-frequency analysis, as applied to the vibration analysis
of rolling element bearings. Specifically:-
• This thesis is the first application of Basis Pursuit to fault diagnosis.
CHAPTER 1. GENERAL INTRODUCTION
____________________________________________________________________
5
• The DWPA procedure was improved and benchmarked with Basis Pursuit;
• The Matching Pursuit technique was refined and benchmarked with Basis
Pursuit.
This research work resulted in the following publications:
• “Time Frequency Techniques for Fault Diagnosis of Rolling Element Bearings”,
in the Proceeding of 10th Asia-Pacific Vibration Conference (APVC, 2003): pp.
789-794. ISBN: 06464 42853;
• “Feature extraction of faulty bearings vibration via Basis Pursuit”, in the
Proceedings of 10th Asia-Pacific Vibration Conference, (APVC, 2003): pp. 795-
800. ISBN: 06464 42853;
• “Fault diagnosis of rolling element bearings using basis pursuit”, to be published
in the Journal of Mechanical Systems and Signal Processing, Volume 19 (2),
2005: pp. 341-356.
(2) Development of additional features for time-frequency analysis:-
• Statistical parameters derived from wavelet packets such as Mean, Variance,
Root Mean Square (RMS), Skewness, Kurtosis, and Crest Factor;
• Wavelet packet based Matched Filter parameter.
• Selected components derived coefficients of Matching Pursuit; and Basis Pursuit.
(3) The design and evaluation of Neural Network classifiers and their performance in
automatic diagnosis. These include:
• The design of two architectures for Feed Forward Neural Networks;
• An evaluation of the performance of the Neural Networks on a variety of features
derived from DWPA, Matching Pursuit and Basis Pursuit. All of this work have
been published in:
o “Bearing fault classification using wavelet features”, in the Proceedings
of Intelligence Maintenance System 2004 International Conference.
Arles, France: Section 1-D;
CHAPTER 1. GENERAL INTRODUCTION
____________________________________________________________________
6
o “Matching Pursuit features based Neural Network pattern recognition of
rolling bearing faults”, in the Proceedings of the International Conference
of Maintenance Societies, 2004. Sydney, Australia: paper 74:pp. 1-8;
o “Basis Pursuit feature based pattern recognition of rolling bearing faults”,
in the Proceeding of Intelligence Maintenance System 2004 International
Conference, 2004. Arles, France: Section 2-C.
(4) Presenting an integrated application of the various wavelet based techniques
1.6 Organisation of Thesis
This thesis starts with a brief introduction to fault diagnosis of rotating machinery.
The objectives, significance, and originality of this research is covered in Chapter 1.
Chapter 2, provides a comprehensive literature review on vibration feature extraction
and AI techniques for condition monitoring and fault diagnosis of rotating
machinery. The feature extraction of vibrations, was categorized as:
• time domain feature extraction techniques,
• frequency domain feature extraction techniques, and
• time-frequency feature extraction techniques.
The advantages and disadvantages of these techniques are presented. The review also
provides the application of a variety of AI techniques such as Neural Networks,
fuzzy logic, expert systems, and hybrid AI techniques.
Chapter 3, provides the analysis and interpretation of vibration analysis techniques
for better representation of vibration signals in terms of diagnostic results. It covers
the Basis Pursuit, the best basis DWPA and the Matching Pursuit techniques.
In Chapter 4, the automated diagnosis procedure based on spectrum and time-
frequency analysis techniques is presented. The design and evaluation of Neural
Network classifiers for the automatic diagnostic schema is presented.
Chapter 5 presents the simulation and experimental phases of this thesis.
In Chapter 6, the results of the application of these techniques to the simulated and
experimental data are presented and discussed.
CHAPTER 1. GENERAL INTRODUCTION
____________________________________________________________________
7
Chapter 7 presents the conclusions of this work. Some suggestions to extend the
work contained in this thesis are presented in chapter 8.
CHAPTER 2. LITERATURE REVIEW
____________________________________________________________________
8
CHAPTER 2. LITERATURE REVIEW
2.1 Introduction
The background of fault diagnosis of rotating machinery is introduced in this chapter.
A literature of techniques for vibration based fault diagnosis is reviewed. It includes
the research work done in the past and presented in publications such as books,
conference articles, journal papers and reports. The variety of methods used, are
discussed and analysed with critical comments. Based on the overall review of
techniques for diagnosing machinery faults, some conclusions are drawn from the
literature.
2.2 Background
2.2.1 Condition Monitoring and Fault Diagnosis
Machine availability is a major concern in industry. In general, machine life can be
described using the bathtub curve (as shown in Figure 2.1). The first phase of a
machine is the preparation phase, that is, a period a new machine experiences until it
commences normal operation. The second phase is the normal operation phase,
which is followed by the failure phase. The failure phase starts with incipient faults
and ends in the breakdown of the machine. It is necessary to monitor the condition of
a machine during the period of machine life. Condition monitoring is a field of
technical activity in which selected parameters associated with machinery operation
are observed for determining integrity [3]. Condition monitoring is essential for
maintenance management in industry, which usually involves five distinct phases
such as detection of fault, diagnosis of fault, prognosis of fault progression,
prescription for treatment of a problem, and post mortem.
Fault diagnosis is a critical component for condition monitoring and mainly
ascertains location, cause, and severity of a machine fault.
CHAPTER 2. LITERATURE REVIEW
____________________________________________________________________
9
Figure 2.1: Machine life bathtub curve
Condition monitoring includes breakdown maintenance, routine maintenance,
condition-based maintenance, and pro-active maintenance [2]. To date, various
condition-monitoring techniques have been developed based on theory from areas
such as dynamics, industrial noise and vibration, tribology, and non-destructive
techniques of moving structures and rotating machinery [1, 2]. In a recent ISO
working party (ISO 1991), it was identified that the main techniques for machine
condition monitoring are:
• Vibration measurements;
• Tribological measurements;
• Electrical measurements;
• Process and performance measurements; and
• Non-destructive testing.
These measurements are analysed as wear debris, vibrations, temperature,
performance, and expired life [4]. As a general principle, the degree of deterioration
is detected by the ‘level’ of a measurement and its change with time, while the cause
is generally indicated by its ‘shape’.
These five basic techniques can be used as a guideline for planning the condition
monitoring of industrial plant or machines, since it provides a check-list to ensure
that all possible techniques have been considered.
All these techniques require some form of transduction and statistical parameter or
advanced signal processing for diagnosis and life estimation of systems. Among the
condition monitoring techniques, vibration condition monitoring is popular for its
Failure
Preparation
Normal
Ris
k
Time
CHAPTER 2. LITERATURE REVIEW
____________________________________________________________________
10
versatility and its effectiveness. There are many mechanical problems associated
with vibrations. Common problems are: (a) imbalance of rotating parts; (b) eccentric
components; (c) misalignment of couplings and bearings; (d) bent shafts; (e)
component looseness; (f) worn or damaged gears; (g) worn drive belts and drive
chains; (h) defective anti-friction bearings; (i) torque variations; (j) electromagnetic
forces; (k) aerodynamic forces; (l) hydraulic forces; (m) resonance; and (n) rubbing
[5]. All of the above problems can cause vibration. Meanwhile, vibration in machines
causes periodic stresses in machine parts, which lead to fatigue failure. If the motion
due to vibration is severe enough, it can cause machine parts to come into unwanted
contact, causing wear or damage [6]. Vibration of machines is a parameter, which
often indirectly represents the health of machines and is generally capable of
detecting more kinds of machine faults when compared with the other techniques.
Vibration monitoring also has advantages as a non-destructive, clean, relatively
simple and cost effective technique [7].
Vibration monitoring of rolling element bearings are typically conducted using a case
mounted transducer: an accelerometer, velocity pickup, and sometimes a
displacement sensor. Acceleration signals, obtained from case mounted sensors,
emphasize high frequency sources, while displacement signals emphasize lower
frequency sources, with velocity signals falling between the extremes.
Figure 2.2: Overall levels of a bearing of continuing phases of failure
2 3 4 1
1 Filtered High Frequency 2 Acceleration 3 Velocity 4 Displacement
Phase 0 Phase I Phase II Phase III Phase IV
Am
plitu
de
Time
CHAPTER 2. LITERATURE REVIEW
____________________________________________________________________
11
Figure 2.2 [8] depicts the overall amplitude levels obtained from a bearing as it
progresses through continuing phases of failure.
Figure 2.3 illustrates spectral phenomenon when bearings have different severities of
faults [8]. Generally, bearings fault progression can be described in four stages from
normal condition to most serious damage. The spectra of bearings with faults of
different stages contains harmonics of rotating frequencies, defect frequencies,
natural resonance frequency, and high frequencies in the ultrasonic region.
Stage I exhibits very high frequency content in the Spike Energy region. This zone
lies in the ultrasonic region and requires a sensor specifically designed to detect these
components. Physical inspection of the bearing at this stage may not show any
identifiable defects.
Stage II begins to generate signals associated with natural resonance frequencies of
the bearing parts as bearing defects begin to "ring" the bearing components. A
notable increase in zones 3 and 4 is normally recorded at this stage. Visual inspection
will show defects at this stage.
The Stage III condition shows the fundamental bearing defect frequencies present.
Harmonics of defect frequencies may be present depending upon the quantity of
defects and their dispersal around the bearing races. The harmonic frequencies will
be modulated, or side banded, by the shaft speed. Zone 4 signals continue to grow
throughout this stage.
Stage IV is the last condition before catastrophic failure of the bearing. This stage is
associated with numerous modulated fundamental frequencies and harmonics
indicating that the defects are distributed around the bearing races. The internal
clearances are greater and allow the shaft to vibrate more freely with associated
increases in the shaft frequencies associated with balance or mis-alignment due to the
increased degradation of the bearing. During the later phases of stage IV, the bearing
fundamental frequencies will decline and be replaced with random noise floor or
"hay stack" at higher frequencies. Zone 4 signal levels will actually decrease with a
significant increase just prior to failure.
CHAPTER 2. LITERATURE REVIEW
____________________________________________________________________
12
Figure 2.3: Spectral characteristics of different stage bearing faults
2.2.2 Artificial Intelligence
In an era when information technology and inter-disciplinary approaches need to be
employed to assess the integrity of systems and machinery, condition monitoring has
begun to incorporate these technologies to assist with the interpretation and
diagnostics of signals particularly with integration of artificial intelligence
techniques. Artificial intelligence (AI) can be defined as a branch of computer
science that is concerned with automation of intelligent behaviour [9]. The goal of
Artificial Intelligence is the development of paradigms or algorithms that require
machines to perform cognitive tasks, at which humans are currently better. An AI
system must be capable of doing three things: (1) store knowledge; (2) apply the
knowledge stored to solve problems; and (3) acquire new knowledge through
experience. An AI system has three key components: representation, reasoning, and
learning [10]. AI techniques have been widely applied in the engineering area which
includes expert systems, fuzzy logic and Neural Networks.
Zone 1 Zone 2 Defect frequencies
Zone 3 Natural resonance frequency
Zone 4 High frequency
Zone 1
Zone 1
Zone 1
Normal
Stage I
Stage II
Stage III Zone 1
Stage IV
Zone 4 High frequency
Zone 4 High frequency
Zone 4 High frequency
Zone 4 High frequency
1xR
PM
2
xRP
M
3xR
PM
1xR
PM
2
xRP
M
3xR
PM
1xR
PM
2
xRP
M
3xR
PM
1xR
PM
2
xRP
M
3xR
PM
1xR
PM
2
xRP
M
3xR
PM
Zone 3 Natural resonance frequency
Zone 3 Natural resonance frequency
Zone 3 Natural resonance frequency
Zone 3 Natural resonance frequency
Zone 2 Defect frequencies
Zone 2 Defect frequencies
Zone 2 Defect frequencies
CHAPTER 2. LITERATURE REVIEW
____________________________________________________________________
13
Expert systems (knowledge based systems) is a branch of AI techniques and
originated in 1965. An expert system generally consists of four essential
components: a knowledge base, an inference engine, a knowledge-acquisition
module, and an explanatory interface [11]. Expert systems can capture and retain the
expertise of a skilled person. This is often necessary, particularly in the condition
monitoring and fault diagnosis field, because, for example, that expert is about to
retire or change jobs. Expert systems for condition monitoring and fault diagnosis
have been built up to address the problems such as lack of staff, lack of correct skill
levels, lack of time to perform tasks, and inconsistent performance of tasks [5] .
Fuzzy logic was first introduced by Zadeh [12] and is one kind of logic satisfied with
the imprecise nature of reality. Traditional logic allows only for “true” and “false”
states. However, knowledge under the logic with alternative values will be partial
because imprecision is ubiquitous and realistic. Factors and variables should also be
permitted between “true” and “false”. A fuzzy set is a set of elements within an
interval and with a membership function on the interval [13, 14]. A fuzzy set is the
cornerstone of a non-additive uncertainty theory, namely possibility theory, and is a
versatile tool for both linguistic and numerical modelling: known as fuzzy rule-based
systems. Several works now combine concepts on fuzzy sets with other scientific
disciplines. In the field of machine condition monitoring, fuzzy logic can be used to
represent machine faults more intuitively than is possible for conventional (precise)
diagnosis. It has the potential to represent the imprecise nature of machine diagnosis.
For example, a machine fault can be represented as seriously or lightly damaged
rather than damaged or not.
Pattern recognition is the method of assigning data into one of a number of pre-
specified classes based on the extraction and processing of significant features [15].
A conceptual representation of a pattern recognition problem is shown in Figure 2.4
[16]. It is a key component in identifying failure models which are induced from the
monitored system. The three approaches to pattern recognition are:
• Statistical pattern recognition (decision-theoretic);
• Syntactic (linguistic or structural); and
• Neural Network (black box).
CHAPTER 2. LITERATURE REVIEW
____________________________________________________________________
14
Figure 2.4: Conceptual representation of a pattern recognition problem
Neural Networks are an important technique in the field of artificial intelligence,
which has been motivated right from its inception by the recognition that the human
brain computes in an entirely different way from conventional digital computers
[10]. The brain is a highly complex, nonlinear, and parallel computer (information-
processing system). It has the capability to organize its structural constituents, known
as neurons, so as to perform certain computations many times faster than the fastest
digital computer in existence today. Neural Networks are widely adopted for their
learning ability, which could be usefully applied in machine diagnosis. For example,
an eligible Neural Network can record new classes of faults so that they can be
utilized when the same fault happens again. Without the learning ability, a fault
diagnosis system may classify a new class of faults into an existing class thus leading
to a wrong judgement.
Jang [17] depicted the architecture of a Neural Network (as shown in Figure 2.5),
which was typically organized in layers. Layers were made up of a number of
interconnected “nodes” which contained an “activation function”. Patterns were
presented to the network via the “input layer”, which communicated to one or more
“hidden layers” where the actual processing was done via a system of weighted
“connections”. The hidden layers were then linked to an “output layer” where the
answer was provided [17, 18].
Figure 2.5: Architecture of a Neural Network
Physical
variables
Data acquisition
Data pre-processing
Decision classification
Phase I Phase II
Phase III
CHAPTER 2. LITERATURE REVIEW
____________________________________________________________________
15
2.3 Literature on fault diagnosis of rotating machinery
Generally, vibration based fault diagnosis can be conducted conventionally in the
following phases: data collection, feature extraction, and fault detection and
identification. Intelligent diagnosis procedure is shown in Figure 2.6, which begins
with the act of data collection (obtaining signals using transducers from machinery to
be diagnosed) followed by feature extraction (extracting characteristics whose values
quantitatively represent faults). As an example of extracted features, in vibration
analysis, impact pulses due to defect are usually regarded as an important
characteristic (feature) for machinery fault diagnosis.
Figure 2.6: Fault diagnosis -an overview
Various mechanical rotating parts, e.g. rolling bearings, gears, pumps, chains, belts,
and electric rotators are very important diagnostic objects in industry. Among these
machinery parts, rolling bearings are most often used mechanical parts. There are
many problems which can be related to bearings and their vibrations. Installation
problems are relatively common and are often caused by improperly forcing the
bearing onto the shaft or in the housing. Misalignment of bearings is a common
result of defective bearing installation. There are four ways of misalignment
including Out-of-Line, shaft deflection, cocked or tilted outer race, and cocked or
tilted inner race [19]. The problems can cause bearing failure. Usually the failure is
accelerated by overloading, over speeding, or starving the bearings of lubricants.
Bearing defects may be categorized as ‘distributed’ or ‘local’. Distributed defects
include surface roughness, waviness, misaligned races and off-size rolling elements.
Distributed defects are caused by manufacturing error, improper installation or
abrasive wear. The variation in contact area between rolling elements and raceways
due to distributed defects results in an increased vibration level. Localized defects
include cracks, pits and spalls in the rolling surfaces. The presence of a defect causes
Data
collection
Feature
Extraction
Fault
detection and
identification
CHAPTER 2. LITERATURE REVIEW
____________________________________________________________________
16
a significant increase in the vibration levels, which are used as the features to be
sought for diagnosing different machinery faults.
Feature extraction techniques for diagnosing rotating machinery faults are
widespread and can range from statistical to model based techniques and comprises a
variety of signal processing algorithms, which include wavelet transforms. Existing
techniques for extracting vibration features, which are critical and significant for
reliable fault diagnosis, are reviewed in section 2.1. Fault detection and identification
is a subsequent step and recently incorporates artificial intelligence techniques.
Conventionally, faults are detected and identified by using experts’ interpretation of
extracted features, and sometimes even direct inspection of running machines. The
advances of fault identification techniques enhance efficacy and reliability of fault
diagnosis. Techniques for fault detection and identification using artificial
intelligence techniques are reviewed in section 2.3.1. With the advancement of
modern fault diagnosis techniques, researchers synthetically employ techniques to
achieve effectiveness of fault diagnosis. Several feature extraction, artificial
intelligence techniques, or their combinations may be used together for one fault
diagnostic task. Diagnosis using both techniques for feature extraction and artificial
intelligence techniques are reviewed in section 2.3.2.
2.3.1 Feature Extraction for Fault Diagnosis of Rotating Machinery
A variety of features have been sought to assist fault diagnosis in literature. There are
many potential advantages obtained when performing feature extraction in the
application of Condition Monitoring and Fault Diagnosis (CMFD). From the point of
view of computing complexity, modern machine CMFD can have a number of inputs
obtained from collected digital signals, which often creates an information overload
for the operators. The information overload makes it difficult to gain an overview of
the machine, and determine exactly what is happening. Using feature selection in
conjunction with classifiers allows the classifiers to act as a data generalization layer,
interpreting the raw data to provide information that can be digested easily, and used
to make meaningful assessments of the condition of a machine. The effective
features can allow machine condition to be ascertained more reliably than by human
operators, and allows further high-level integration with artificial intelligence
CHAPTER 2. LITERATURE REVIEW
____________________________________________________________________
17
techniques such as expert systems, which can provide wide-scale monitoring of the
machine as a whole, rather than the sub-assemblies that the classifiers monitor.
The feature extraction phase is critical in fault diagnosis practice. The purpose of
feature extraction is twofold; firstly, feature extraction is an attempt to reduce the
dimensionality of the data presented to the classifier or human inspection, without
diminishing the content presented in the data. Secondly, feature extraction is utilized
to turn raw data into the information that the classifier can use. The vibration data as
sampled will consist of several hundreds, or even thousands of data points.
To train a classifier to deal with raw data always requires a very large training time,
makes a network extremely complex, and sometimes leads to failure of the
classification. It also makes the generalisation of the network fairly poor, as so many
input factors will make it difficult for the classifier to determine useful relationships
between inputs, and consequently, to generalize effectively. This effect, known as the
curse of dimensionality, can be dealt with using feature extraction, which is simple
and effective.
To extract features from raw data, numerous vibration analysis techniques have been
applied to fault diagnosis of rotating machinery. Mathew and Alfredson [20] gave a
review of vibration monitoring techniques in time and frequency domains and their
results on rolling element bearings. McFadden, Smith [21-23] and Kim [24-27]
included classical non-parametric spectral analysis; principal component analysis;
joint time-frequency analysis; the discrete wavelet transform; and change detection
algorithm based on residual generation Lebold and McClintic [28] reviewed
statistical methods for extracting vibration features when diagnosing gearbox. They
made an attempt to define and unify statistical technique terms, establish the
preprocessing needed for each feature, and provide the details needed to produce
consistent results. They categorized features into five different groups based on their
preprocessing needs when diagnosing gearboxes. They were: 1) Raw signals (RAW)
which includes Root Mean Square (RMS), Kurtosis, Delta RMS, Crest Factor,
enveloping and demodulation, 2) Time synchronous averaged signal (TSA) which
includes FM0 and the Comblet, 3) Residual signal (RES) include NA4 and NA4*, 4)
Difference signal (DIF) which includes FM4, M6A, and M8A, and 5) Band-pass
mesh signal (BPM) which includes NB4.
CHAPTER 2. LITERATURE REVIEW
____________________________________________________________________
18
Tandon and Choudhury studied a review on vibration and acoustic measurement
techniques for the detection of defects in rolling element bearings [29]. In the time
domain, RMS, Kurtosis and a shock pulse method have been analysed. In the
frequency domain, how to apply Fast Fourier Transform (FFT) has been explained.
Power Cepstrum is defined as the logarithm of power spectrum. Adaptive Noise
Cancelling (ANC) technique, envelop detection or the high-frequency resonance
technique (HFRT) are important signal processing techniques. Chow [30] provided a
brief review of model-based approaches and signal processing approaches on motor
fault detection and diagnosis.
A review of feature extraction techniques are updated according to signals under
time domain, frequency domain, and the combination of time and frequency domain
(as shown in Figure 2.7). These categories are treated and reviewed separately in the
following sections.
Figure 2.7: An overview of feature extraction techniques
2.3.1.1 Time domain Feature Extraction Techniques
Vibration signals are initially obtained as a series of digital values representing
proximity, velocity, or acceleration in the time domain. The time waveforms can be
processed to achieve diagnostic objectives. Certain features such as statistical
parameters can be signified using time domain vibration analysis techniques. The
machine faults can be distinguished using the quantitative representation of time
domain features. This section includes research appearing in literature, and reviews
vibration techniques in the time domain for various types of rotating machinery and
categorises these techniques into the following groups (as shown in Figure 2.8).
Feature extraction
Time domain Frequency domain Time and frequency domain
CHAPTER 2. LITERATURE REVIEW
____________________________________________________________________
19
Figure 2.8: An overview of time domain feature extraction techniques
(1) Statistical parameters, which include Root Mean Square (RMS), Mean, Variance,
Skewness, Kurtosis, and Crest Factor
Statistics is an area, which can provide many ideas for vibration analysis in fault
diagnosis of machinery. Statistical analyses of vibration signals have proved to be
useful in detecting machinery faults. Tandon [29] showed that the probability density
function is correlated with bearing defects. The probability density of acceleration of
a bearing in good condition has a Gaussian distribution, whereas a damaged bearing
results in non-Gaussian distribution with dominant tails, because of a relative
increase in the number of high levels of acceleration. Andrade [31] proposed a
comparison of the Cumulative Density Function (CDF) of a target distribution with
the CDF of a reference distribution and used the likelihood to successfully detect
gear tooth fatigue crack. Mathew and Alfredson [20] also reported obtaining a near-
Gaussian distribution for some damaged bearings. Instead of studying the probability
density curves, it is often more informative to examine the statistical moments of the
data, defined as
∫+∞
∞−
== mnxPxM nx ,,3,2,1)( L (2.1)
Filter based methods Raw signals
Statistical parameters
Time domain
Stochastic methods and other advanced methods
Time Synchronous Averaged Signal (TSA) based methods
Time synchronous averaged (TSA) signal, Residual signal (RES), and Difference
Demodulation, Prony model, and Adaptive noise cancelling
Chaos, Blind deconvolution, Thresholding, and Autoregressive model based method
Root Mean Square (RMS), Mean, Variance, Skewness, Kurtosis, and Crest Factor
CHAPTER 2. LITERATURE REVIEW
____________________________________________________________________
20
Where P(x) is the probability density function of instantaneous amplitude x. The first
and second moments are well known, being the mean value and the Variance,
respectively. The third moment normalized with respect to the cube of standard
deviation is known as the coefficient of ‘Skewness’. Kurtosis is defined as the fourth
moment of the distribution and measures the relative peakedness or flatness of a
distribution as compared to a normal distribution. Kurtosis provides a measure of the
size of the tails of distribution and is used as an indicator of major peaks in a set of
data. As rotating machinery faults present themselves, Kurtosis should signal an error
due to the increased level of vibration. Kurtosis has been applied to diagnosing
bearing, and gearbox faults [32].
RMS and Delta RMS
The root mean square (RMS) value of a vibration signal is a time analysis feature,
which is the measure of the power content in the vibration signature. This feature is
good for tracking the overall noise level, but it will not provide any information on
which component is failing. It can be very effective in detecting a major out of
balance in rotating systems.
Delta RMS is the difference between the current RMS value and the previous. The
RMS are value and Crest Factor have been applied in diagnosing bearings and gears
[29]. This feature can be very effective when detecting an imbalance in rotating
machinery. The most basic approach to measuring defects in the time domain is to
use the RMS approach, which is often not sensitive enough to detect incipient faults
in particular.
Kurtosis
Kurtosis is defined as the fourth moment of the distribution and measures the relative
peakedness or flatness of a distribution as compared to a normal distribution.
Kurtosis provides a measure of the size of the tails of distribution and is used as an
indicator of major peaks in a set of data. As a gear wears this feature should signal an
error due to the increased level of vibration.
CHAPTER 2. LITERATURE REVIEW
____________________________________________________________________
21
Crest Factor
The RMS level may not show appreciable changes in the early stages of gear and
bearing damage. A better measure is to use “Crest Factor” which is defined as the
ratio of the peak level of the input signal to the RMS level. Therefore, peaks in the
time series signal will result in an increase in the Crest Factor value. Crest Factor
may reach between 2 and 6 in normal operations. A value above 6 is usually
associated with machinery problems. This feature is used to detect changes in the
signal pattern due to impulsive vibration sources such as tooth breakage on a gear or
a defect on the outer race of a bearing.
The RMS, peak value, Kurtosis and Crest Factor have been combined with a high
frequency resonance technique and an adaptive line enhancer to detect and localize
the damage in rolling bearing [33].
(2)Time synchronous averaging based methods, which include Time Synchronous
Averaged (TSA) signal, residual signal (RES), and difference signal (DIFS)
The TSA signals are the signals obtained by time synchronous averaging of the
initial data and reducing redundant noise. The repetitive signals after TSA can
indicate the information related to the faults, which need to be diagnosed. Time
synchronous averaged signal (TSA) includes FM0 and the Comblet, Residual signal
(RES) included NA4 and NA4*, Difference signal (DIF) includes FM4, M6A, and
M8A, Band-pass mesh signal (BPM) includes NB4. The TSA including FM0 and
Comblet [28] requires knowing the repetitive frequency of the desired signal such as
defect frequencies of rolling bearings, gears, and shafts. Synchronous averaged
signals were utilized to diagnose faults in rolling bearings and gears successfully [34-
36].
Residual signals (RES) [28] was used for diagnosing gear faults. RES consists of
time synchronous averaged signal with primary meshing and shaft components along
with their harmonics removed. RES may be system dependent. Difference signals
(DIF) were calculated by removing the regular meshing components from the time
CHAPTER 2. LITERATURE REVIEW
____________________________________________________________________
22
synchronous averaged signal. The DIF signals were used to diagnose gearbox faults
effectively [37].
FM0 is a relatively simple method used to detect major changes in the meshing
pattern. Major tooth faults typically result in an increase of the peak-to-peak signal
levels, but do not change the meshing frequency. FM0 is defined as the peak-to-peak
level of the TSA signal divided by the sum of the amplitude at the gear-mesh
frequency and its corresponding harmonics. The peak-to-peak level remains constant
for heavy wear, while the meshing frequency decreases, causing the FM0 parameter
to jump up. The above situations result in a large increase in the FM0 parameter.
However FM0 is not a good indicator for minor tooth damage. The equation for FM0
is:
( )∑=
= n
iifA
PPAFM
1
0 (2.1)
where PPA is the peak-to-peak amplitude of the time synchronous averaged
waveform and ( )ifA is the amplitude of the gear-mesh fundamental and harmonics
in the frequency domain.
NA4
NA4 was developed to detect the onset of damage and to continue to react to this
damage as it spreads and increases in magnitude. NA4 is determined by dividing the
fourth statistical moment of the residual signal by the current run time averaged
Variance of the residual signal, raised to the second power. The equation for NA4 is
( )
( )2
1 1
2
1
4
14
−
−=
∑ ∑
∑
= =
=
m
j
N
ijij
N
ii
rrm
rrNA (2.2)
where r is the residual signal, r is the mean value of the residual signal, N is the
total number of data points in the time record, and m is the current time record
number in the run ensemble.
CHAPTER 2. LITERATURE REVIEW
____________________________________________________________________
23
NA4*
NA4* (or ENA4) was developed as an enhanced version of NA4, and was expected to
be more robust when progressive damage occurs. This added robustness is
incorporated into NA4* by normalizing the fourth statistical moment with the
residual signal Variance for a gearbox in good condition instead of the running
Variance, which is used for NA4. The equation for NA4* follows:
( )
( )22
1
4
~*4M
rrNA
N
ii∑
=
−= (2.3)
where r is the residual signal, r is the mean value of residual signal, N is the total
number of data points in time record, and 2~
M is the Variance of the residual signal
for a gearbox in good condition.
FM4
FM4 was developed to detect changes in the vibration pattern resulting from damage
on a limited number of gear teeth. FM4 is calculated by applying the fourth
normalized statistical moment to this difference signal as given in the equation:
( )
( )2
1
2
1
4
4
−
−=
∑
∑
=
=N
ii
N
ii
dd
ddNFM (2.4)
where d is the difference signal, d is the mean value of difference signal, and N is the
total number of data points in the time record. A difference signal from a gear in
good condition will be primarily Gaussian noise therefore resulting in a normalized
Kurtosis value of 3. As a defect develops in a tooth, peaks will grow in the difference
signal that will result in the Kurtosis value to increase beyond 3.
CHAPTER 2. LITERATURE REVIEW
____________________________________________________________________
24
M6A and M8A
M6A and M8A were proposed to detect surface damage on machinery components.
Both of these features are applied to the difference signal. The theory behind M6A
and M8A is the same as that for FM4, except that M6A and M8A are expected to be
more sensitive to peaks in the difference signal. The equations for M6A and M8A are
as follows:
( )
( )3
1
2
1
62
6
−
−=
∑
∑
=
=N
ii
N
ii
dd
ddNAM (2.5)
( )
( )4
1
2
1
82
8
−
−=
∑
∑
=
=
N
ii
N
ii
dd
ddNAM (2.6)
where d is the difference signal, d is the mean value of difference signal, and N is
the total number of data points in the time record.
NB4
NB4 is similar to NA4 except that instead of using the residual signal, NB4 uses the
envelope of a band-passed segment of the time synchronous averaged signal. The
idea behind this method is that a few damaged gear teeth will cause transient load
fluctuations that are different from the normal tooth load fluctuations. The theory
suggests that these fluctuations will be manifested in the envelope of a signal which
is band-pass filtered about the dominant meshing frequency. The dominant meshing
frequency is either the primary meshing frequency or one of its harmonics whichever
appears to give the most robust group of sidebands. Researchers suggest that the
width of the band-pass filter depends on the location of the meshing frequency to
other meshing frequency harmonics, while others suggest using a bandwidth giving
the maximum amount of sidebands even if the sidebands interfere with those from
other harmonics.
CHAPTER 2. LITERATURE REVIEW
____________________________________________________________________
25
The reasoning of the latter method is to assume that the interference from other
sidebands is negligible and includes as many of the primary modulating sidebands as
plausible. The envelope of the band-passed signal is the magnitude of the complex
(i.e., analytic) signal obtained by applying the Hilbert transform to the band-passed
signal:
( ) ( )( ) ( )[ ]22 tAHtAtE += (2.7)
where E(t) is the envelope of the band-passed signal, A(t) is the band-passed signal,
and H[A(t)] is the Hilbert transform of the band-passed signal. The analytic signal is
A(t)+iH[A(t)].
NB4 is then determined by dividing the fourth statistical moment of this envelope
signal by the current run time averaged Variance of the envelope signal, raised to the
second power, with the equation following:
( )
( )2
1 1
2
1
4
14
−
−=
∑ ∑
∑
= =
=
m
j
N
ijij
N
ii
EEm
EENNB (2.8)
where E is the envelope of the band-passed signal, E is the mean value of the
envelope signal, N is the total number of data points in the time record, and m is the
current time record number in the run ensemble.
(3) Filter based methods including demodulation, Prony model, and adaptive noise
cancelling (ANC)
Filters are widely used in feature extraction techniques for removing noise and
isolating signals. Generally all these methods were referred as filter based methods.
Filter based methods include demodulation, prony model, and adaptive noise
cancelling (ANC).
Demodulation including phase and amplitude demodulation is an important signal
processing technique. The amplitude demodulation was also known as envelope, or
CHAPTER 2. LITERATURE REVIEW
____________________________________________________________________
26
resonance demodulation, or high frequency resonance demodulation techniques [38].
The amplitude demodulation separates low-level, low-frequency signals from
background noise, enabling them to be easily measured. In the application of gear
faults detection, the amplitude demodulation focused on the fault-induced high-order
modulation sidebands around the dominant gear meshing harmonic [39]. It has also
been successfully applied to diagnose bearing faults [40]. The phase demodulation
emphasised the band associated with the structural resonance excited by the fault-
induced impacts [38].
Generally the demodulation procedure starts with using conventional Infinite
Impulse Response (IIR) Filters such as Butterworth, Chebyshev, Bessel, and Elliptic
in pass band or stop band. Prony's model was used as an algorithm for finding an IIR
filter with a prescribed time domain impulse response.
Enveloping
Enveloping is used to monitor the high frequency response of the mechanical system
to periodic impacts such as gear or bearing faults. An impulse is produced each time
a loaded rolling element makes contact with a defect on another surface in the
bearing or as a faulty gear tooth makes contact with another tooth. This impulse has
an extremely short duration compared to the interval between the pulses. The energy
from the defect pulse will be distributed at a very low level over a wide range of
frequencies. This wide distribution of energy makes bearing defects difficult to
detect by conventional spectrum analysis when they are in the presence of vibrations
from gears and other machine components. The impact usually excites a resonance in
the system at a much higher frequency than the vibration generated by the other
components. This structural energy is usually concentrated into a narrow band that is
easier to detect than the widely distributed energy of the bearing defect frequencies.
With tooth wear and breakage, the sideband activity near critical frequencies such as
the output shaft frequency is expected to increase. The entire spectrum contains very
high periodic signals associated with the gear mesh frequencies.
The envelope or high frequency technique focuses on the structural resonance to
determine the health of a gear or the type of failure in a bearing. This technique
CHAPTER 2. LITERATURE REVIEW
____________________________________________________________________
27
consists of processing structural resonance energy with an envelope detector. The
structural resonance is obtained by band-pass filtering the data around the structural
resonance frequency. The band-pass filtered signal is then processed by an envelope
detector, which consists of half-wave (or full-wave) rectifier and a peak-hold and
smoothing section.
The centre frequency of the band-pass filter should be selected to coincide with the
structure resonance frequency being studied. The bandwidth of the filter should be at
least double the highest characteristic defect frequency. This will ensure that the
filter will pass the carrier frequency and at least on a pair of modulation sidebands. In
practice, the bandwidth should be somewhat greater to accommodate the first two
pairs of modulation sidebands around the carrier frequency.
The rectifier in the envelope detector turns the bipolar filtered signal into a unipolar
waveform. The peak-hold smoothing section will then remove the carrier frequency
by smoothing/filtering the fast transitions in the signal. The remaining signal will
then consist of the defect frequencies.
This feature produces several values of merit for analysis use. The primary value of
merit is the peak frequency and amplitude in the power spectral density of the
enveloped data. Other values of merit include the RMS and Kurtosis values of the
filtering section and the standard deviation of the output from the rectification and
smoothing block.
The envelope technique has been widely used in numerous applications and has
shown successful results in the early detection of bearing faults. Besides early
detection, this process can help distinguish the actual cause of bearing failure by
inspecting the actual bearing defect frequencies.
Envelope detection or high-frequency resonance technique (HFRT) has been shown
to be effective in fault diagnosis [23, 34] [40].
Amplitude demodulation used in gear fault diagnosis
During a normal gear roll, one tooth essentially pushes the next without sliding.
When teeth wear, sliding occurs. The energy that went into pushing before will now
go into pushing and sliding, thus resulting in a change of amplitude or amplitude
CHAPTER 2. LITERATURE REVIEW
____________________________________________________________________
28
modulation of the vibrations at the gear mesh frequency (GMF) and its harmonics.
Demodulation identifies periodicity in modulation of the carrier. The carriers used in
this processing were the GMF and 2*GMF. Demodulation techniques detect the
amplitude modulation components induced by gear wear in the region of a single
frequency, in this case the GMF or 2*GMF. This differs from enveloping which
detects the combined effects over a range of frequencies. To implement the
demodulation technique, the raw data is high-passed filtered at 85%*GMF and then
low-passed filtered at 115%*GMF. The power spectral density of the filtered signal
is searched to obtain the actual carrier frequency (GMF). The actual carrier is used to
amplitude demodulate the filtered carrier signal. The power spectral density of the
resulting signal is searched within +/- five percent of the output shaft frequency. The
values of merit extracted for this technique are the frequency of the peak and the
magnitude squared amplitude.
Resonance demodulation [39], [41] is similar to the commonly used narrow-band
demodulation. The former emphasizes the band associated with the structural
resonance excited by the fault-induced impacts, whereas the latter focuses on the
fault-induced high-order modulation sidebands around the dominant gear meshing
harmonic.
A Prony model based method [42] was applied to bearing faults diagnosis. The
method shows potential for analysing transient vibration signals created from faulty
low speed rolling element bearings. Spectral plots can be generated by applying the
procedure to very short data samples, as well as trending parameters based on these
spectral estimations and Prony parameters. An equation was derived to quantitatively
determine the fault status. It is shown that application of the Prony model based
method has the potential to be an effective as well as efficient machine condition
monitoring and diagnostic tool where short duration transient vibration signals are
being generated.
A recently developed filter-adaptive filter was embedded into an adaptive noise
cancelling (ANC) system and showed promise in diagnosing bearing faults [43].
Adaptive noise cancelling is an approach to reduce noise based on reference signals.
In conventional adaptive noise cancelling systems, the primary input signal is a
combined signal and noise c(n)=s(n)+r0(n), and the reference signal is a noise signal
CHAPTER 2. LITERATURE REVIEW
____________________________________________________________________
29
r1(n) through another channel from the same noise source. Asynchronous adaptive
noise cancelling technology was employed to detect self-aligning roller bearing
faults successfully [44]. Wang [38] detected gear faults using phase and amplitude
demodulation.
(4) Stochastic methods (including chaos) and others (blind deconvolution, blind
source separation, thresholding, and autoregressive model based method)
Advanced methods such as stochastic parameters have been used to analyse
vibrations in the time domain. Chaos, whose computation parameters are known as
the correlation dimension, is used to characterise several induced faults of varying
severity in a rolling element bearing [45],[46]. The correlation dimension can
provide some intrinsic information of an underlying dynamical system, and can be
used to classify different faults intelligently [47].
D. Logan [45] applied a new field of chaos to mechanical systems. His research
proposed the computation of chaotic parameters, known as the correlation
dimension, and used this to characterise several induced faults of varying severity in
a rolling element bearing. Further detailed investigations were then made into the
parameters governing the multi-stage process of determining the correlation
dimension. The correlation dimension was obtained from a computationally
straightforward algorithm and appeared as a single scalar index.
The key to chaos theory was being able to extract the nature of a strange attractor
(presumably assuming one existed) from, in their case, turbulent fluid flow. Strange
attractors were normally characterised by fractal dimensionality d, where d is less
than the number of degrees of freedom of the system F, d<F. The correlation
dimension could provide some intrinsic information of an underlying dynamical
system, and could be used to classify different faults intelligently [47].
Nirbito [48] proposed and tested the feasibility of blind deconvolution for the
enhancement of bearing signals corrupted by noise. Blind deconvolution
(equalization) is a technique used to recover the desired signals from a single
CHAPTER 2. LITERATURE REVIEW
____________________________________________________________________
30
received channel without any priory knowledge about the unknown channel. The
technique has been widely used in network communication. A major advantage of
blind deconvolution is that it does not require a training stage, which is essential in
conventional equalization.
Serviere [49] applied Blind Source Separation (BSS) to rotating machinery
diagnosis. BSS consisted of recovering signals from different physical sources from
several observed combinations independent of the propagation medium. BSS was
used as a promising tool for non-destructive machine condition monitoring by
vibration analysis, as it was intended to retrieve the signature of a single rotating
machine from combinations of several working machines. In this way, BSS could be
seen as a pre-processing step that improves the diagnosis. BSS methods generally
assumed observations that were either noise-free or corrupted with spatially distinct
white noises. In the latter case, principal component analysis (PCA) was applied as a
first step to filter out the noise and whiten the observations. The efficiency of the
whole separation procedure depends on the accuracy of the first step PCA. However,
in the real world, signals of rotating machine vibration might be severely corrupted
with spatially correlated noises and therefore the signal subspace would not be
correctly estimated with PCA. A ‘robust-to-noise’ technique was proposed for the
separation of rotating machine signals. The sources were assumed to be periodic and
could be modelled as the sum of sinusoids of harmonic frequencies. A new estimator
of the signal subspace and the whitening matrix was introduced which exploited the
model of sinusoidal sources and used spectral matrices of delayed observations to
eliminate the influence of the noise. After whitening, the second step of source
separation remained unchanged. Finally, performance of the algorithm was
investigated with artificial data and experimental rotating machine vibration data.
The pseudo-phase portrait was sensitive to some rotating machinery faults [46].
Threshold denoising (including hard threshold and soft threshold) were often used to
denoise vibration analysis. The threshold denoising methods were usually combined
with envelope or some other methods together when diagnosing machinery faults. A
soft-thresholding method and hard thresholding method have also been used in
diagnosing machine faults [50]. An autoregressive model-based method has also
been successfully applied in fault diagnosis [41].
CHAPTER 2. LITERATURE REVIEW
____________________________________________________________________
31
Andrade [31] compared Cumulative Density Function (CDF) of a target distribution
with the CDF of a reference distribution and used the likelihood to detect gear tooth
fatigue crack. The statistical distance between two CDFs was converted into a
similarity probability using the Kolmogorov-Smirnov(KS) probability distribution
function ksQ defined as
( ) ( ) 222
1
112 λλ j
j
jks eQ −
∞
=
−∑ −= (2.9)
A synchronous averaging signal was utilized to diagnose faults in rolling bearings
and gears [34, 51]. Calculation of a synchronous averaging signal ( )ty of a time
signal x(t) using a trigger signal having a frequency tf is equivalent to the
convolution ( ) ( ) ( )txtcty *= where ( )tc is a train of N impulses of amplitude 1/N,
spaced at intervals tt fT /1= , given by ( ) ( )∑−
=
+=1
0
1 N
ntnTt
Ntc δ
In the frequency domain, this is equivalent to the multiplication of the Fourier
transform X(f) of the signal by C(f), represented by
( ) ( ) ( )fXfCfY ⋅= (2.10)
Where C(f), the Fourier transform of c(t), is a comb filter function of the form
( ) ( )( )fT
fNT
NfC
1
1
sin
sin1
ππ=
(2.11)
Increasing the number of averages N narrows the teeth of the comb, and reduces the
amplitude of the side lobes between the teeth. For very large N, only frequencies at
exact multiples of the trigger frequencytf are passed. Thus, synchronous averaging
can be viewed in the frequency domain, for large numbers of averages, as the
complete removal of all components except those that occur at integer multiples of
the frequency tf .
The autoregressive model-based method has been applied in fault diagnosis [1, 41,
52-55].
CHAPTER 2. LITERATURE REVIEW
____________________________________________________________________
32
Statistical methods have been combined with AI techniques such as Neural Networks
in order to diagnose faults more efficiently [56]. An 8×12×1 artificial Neural
Network has been used successfully for on-line monitoring of ball bearing
conditions. Peak amplitude in the frequency domain, peak RMS, and the power
spectrum of vibration signals have been used as inputs of the Neural Network while
the outputs indicate the bearing states.
An empirical model-based fault diagnosis system was developed for induction
motors using recurrent dynamic Neural Networks and multiresolution signal
processing methods [57]. It was pointed out that, in practice, it is desirable to
perform multi step (MS) predictions recursively, by relating current estimated output
with previous estimated output and previous input. The recursive relation between
inputs and outputs in MS prediction was expressed using a Feed Forward Neural
Network (FFNN). IIR type feedback (GF) and the lack GF in IIR type network (TF)
were used as learning algorithms to train the Neural Network. Air-gap eccentricity
for motor and broken rotor bars for motor was successfully detected using the model-
based diagnosis system.
Other methods such as shock pulse, Matched Filter root mean square method and
threshold denoising (including hard threshold and soft threshold) show promise.
The shock pulse method [24] is a signal processing technique used to measure metal
impact and rolling noise such as those found in rolling element bearings and gears. In
this method, an accelerometer is involved that is excited to resonance by shock
impacts from bearing defects. The impact components of signals can be effectively
focused and localised. The peak value of the recorded shock pulse is measured in
order to obtain an indication of the condition of a bearing. The disadvantage of
tuning the transducer resonance to the resonant frequency of the structure is that the
absolute amplitude of the vibration is not known. To surmount this problem the
shock pulse value is normalised by subtracting the shock pulse value of an
undamaged bearing. A similar method is spike energy measurement [24], which can
be obtained using spike energy meters. Spike energy meters detect bursts of vibration
at very high frequencies. A high-pass filter is used to filter our vibration components
below the accelerometer resonant frequency.
CHAPTER 2. LITERATURE REVIEW
____________________________________________________________________
33
The adaptive filter [58] and Matched Filter root mean square method [20] are also
popular techniques for fault diagnosis. The ANC technique [43] was used to
diagnose bearing faults. Soft-thresholding denoising method [50] and hard
thresholding denoising method are used widely to denoise signals when diagnosing
machine faults.
2.3.1.2 Frequency and Time-frequency Feature Extraction Techniques
Tandon [29] presented a review of vibration and acoustic measurement methods for
the detection of defects in rolling element bearings. He considered the detection of
both localized and distributed categories of defect. Vibration measurement in both
time and frequency domains along with signal processing techniques such as the
high-frequency resonance technique have been covered. Other acoustic measurement
techniques such as sound pressure, sound intensity and acoustic emission have been
reviewed. Recent trends in research on the detection of defects in bearings, such as
the wavelet transform method and automated data processing, have also been
included. In the time domain, RMS, Kurtosis and shock pulse methods have been
analysed. In the frequency domain, how to apply Fast Fourier Transform (FFT) has
been explained. Power cepstrum is defined as the logarithm of power spectrum. The
ANC technique, envelop detection or the high-frequency resonance technique
(HFRT) is an important signal processing technique.
This section starts from the advent of modern Fast Fourier Transform and then
emphasizes time-frequency representation. A comprehensive review with most
frequency and time-frequency methods for diagnosing machinery faults is covered. It
is found that in the literature both frequency and time-frequency analysis techniques
are being attempted to extract efficient coefficients by increasing the order of
transformation of all kinds of parameters. For instance, power spectrum as a second
order spectrum was applied successfully even though spectrum has been used
widely. Various high order parameters have already showed their capability of
magnifying vibration features in fault diagnosis research. Theoretically, it can be
found that high order transformation increases the magnitude of characteristic
frequency or time-frequency parameters.
CHAPTER 2. LITERATURE REVIEW
____________________________________________________________________
34
Table 2.1 gives an overview of developed high order frequency techniques and time-
frequency techniques. The detailed definitions and explanations of those parameters
are given in the context, which follows Table 2.1:
Table 2.1: An overview of frequency techniques and time-frequency techniques
First order Second order Third order Fourth order
Power spectrum Spectrum
Instantaneous Power Spectrum(IPS)
Bicoherence spectrum
Correlation of spectrum, signal averaging
Cyclostationarity Bilinearity
Spectrogram Short Time Fourier Transform (STFT) Wigner distribution Wigner bi
spectra Wigner tri spectra
Continuous Wavelet Transform(CWT)
Scalogram
Discrete Wavelet Packet Analysis (DWPA)
Matching Pursuit
Time-Averaged Wavelet Spectrum (TAWS)
Time-Frequency-Scale domain (TFS)
As is well known, frequency-domain or spectral analysis of the vibration signal is
perhaps the most widely used approach of bearing defect detection. The Discrete
Fourier Transform (DFT) and FFT are the most conventional diagnosis techniques
which have been widely used [29, 59-63]. DFT provides information about any
sampled signal, returning complex frequency spectra that can in turn be used to
calculate both the phase and magnitude spectra of a signal. Each one of the frequency
bins created as a result of this process has a resolution of fs/N, where fs is the
sampling frequency. Applying a DFT directly to a time series causes large peaks to
be generated at both ends of the spectrum; this is caused by the assumption that the
CHAPTER 2. LITERATURE REVIEW
____________________________________________________________________
35
signal is periodic, and any discontinuities in the signal cause the large peaks to be
generated to compensate for this. In order to alleviate this problem, a window is
applied to the data, which ensures that the beginning and end of the time sequence
start at zero.
( ) ( )∑−
=
−=1
0
/2N
n
NknjenxkX π (2.12)
Calculating the DFT is a fairly computationally intensive job, however FFT opened
up the field of simple spectral analysis. This method and its variants are still found
commonly in use today.
Spectrum is used to derive a few methods such as singular spectrum, envelope
spectrum, and Power Spectrum. Singular spectrum analysis can reveal the
complexity of a signal. By means of singular spectrum analysis one can reduce the
noise of a signal [46].
The procedures for obtaining the spectrum of the envelope signal by a high-
frequency resonance technique are well established and applied in bearing fault
detection [64].
Du [65] proposed a method for the fault diagnosis of roller bearings based on the
empirical mode decomposition (EMD) method and the Hilbert spectrum. The local
Hilbert spectrum and local Hilbert marginal spectrum were introduced. The
orthogonal wavelet bases were used to translate vibration signals of a roller bearing
into time-scale representation. Then, an envelope signal could be obtained by
envelope spectrum analysis of wavelet coefficients of high scales. By applying the
EMD method and Hilbert transform to the envelope signal, they derived the local
Hilbert marginal spectrum from which the faults in a roller bearing could be
diagnosed and fault patterns identified. Practical vibration signals measured from
roller bearings with out-race faults or inner-race faults were analysed by the
proposed method. The results showed that the proposed method was superior to the
traditional envelope spectrum method in extracting the fault characteristics of roller
bearings.
The Power Spectrum whose amplitude is square of amplitude of spectrum is also an
effective method for diagnosing machinery faults [61, 66]. The power spectrum is a
CHAPTER 2. LITERATURE REVIEW
____________________________________________________________________
36
development of the DFT. The DFT returns a complex answer to a real valued
problem, and examining the individual components it is hard to make sense of what
is actually happening within a signal. The Power Spectrum calculates the average of
the magnitude of the DFT, and is defined:
( ) ( ) ( ){ }kXkXEkSxx*= (2.13)
where ( )kX * is the complex conjugate of X(k). This is probably the most common
spectral technique used. By examining the peaks, harmonics can be spotted in the
spectrum, and from this the frequencies of interest can be calculated, and then related
back to the faulty components. However, in many cases it can be difficult to
distinguish harmonic peaks within a spectrum, dependent upon the degree of noise
present in the raw data, and the number of other vibration sources in the location
where the data was sampled.
The Instantaneous Power Spectrum (IPS) is a time-frequency technique, which has
been successfully applied in the detection of the local faults in helical gears [67]. The
IPS was combined with Neural Networks to effectively diagnose machinery faults
[68].
The Power Cepstrum, which is logarithm of power spectrum, was applied to bearing
fault diagnosis. Cepstrum is the inverse Fourier transform of the logarithmic power
spectrum and was used in the diagnosis of bearing faults [29].
Cepstrum analysis is a nonlinear signal processing technique with a variety of
applications in areas such as speech and image processing. The real Cepstrum of a
signal x, sometimes called simply the Cepstrum, is calculated by determining the
natural logarithm of magnitude of the Fourier transform of x, then obtaining the
inverse Fourier transform of the resulting sequence. This process creates another
spectrum, which when read, gives an indication of the presence of harmonics in the
power spectrum. Peaks in the Cepstrum correspond to the existence of harmonics in
the power spectrum, with the frequency giving the period of the separation between
the harmonics in the frequency domain. The harmonics on the Cepstrum plot (known
as harmonics) would indicate the existence of several different harmonics which are
a multiple of the frequency of interest.
CHAPTER 2. LITERATURE REVIEW
____________________________________________________________________
37
The Cepstrum is used in vibration monitoring because it is often the case that the
power spectra taken from a machine can be extremely noisy. This is due to vibration
coming from external influences, such as other machines operating on the same area.
Thus, it can be very difficult to determine what parts of the signal constitute
harmonics and what is noise in the power spectrum, and Cepstrum analysis provides
a relatively efficient way to isolate the harmonic frequencies present in the signal.
The higher order spectrum is also called Bispectrum and is defined by
( ) ( ) ( ) ( )[ ]yxyxyx ffSfSfSEffB += *,, (2.14)
It can be seen that the Bispectrum is complex and that the Bispectral values depend
on two frequencies xf and yf . Rewriting the above equation in terms of amplitude
and phase quantities gives
( ) ( ) ( ) ( ) ( )yx ffjyxyxyx effSfSfSffB ,, βθ+=
(2.15)
where ( ) ( ) ( ) ( )yxyxyx ffffff +−+= ββββ θθθθ , and is called biphase, which is
applied to fault diagnosis of motor bearing [68].
Bicoherence
The Bicoherence Spectrum is a third-order spectrum used to measure the phase
coherence among three spectral components due to nonlinear wave coupling.
Bicoherence has been used to monitor bearing condition [69].
Cyclostationarity
Signal averaging is a well-known synchronised averaging method, which is the
expression of Cyclostationarity of the first order. The spectral correlation function
issued from the second-order Cyclostationarity is an efficient parameter for the early
diagnosis of faults in rotating machines. Moreover, it is shown that vibration signals
measured on gear systems display second-order Cyclostationarity. Application in
early diagnosis of spalling in gear teeth demonstrates the power of this new
parameter [70]. A comparison between Cyclostationarity and Bilinearity has been
presented in an application to early diagnosis for helicopter gearboxes [71].
CHAPTER 2. LITERATURE REVIEW
____________________________________________________________________
38
Generally, frequency domain parameters were more consistent in the detection of
damage than time domain parameters. However, sufficient evidence is produced to
show that it would be unreliable to depend exclusively on any one technique to
detect bearing damage [72].
In recent years, the time-frequency signal analysis has been studied and applied in
machinery fault diagnosis due to its capability of representing signals in both time
and frequency domains. This characteristic of time-frequency analysis technique
meets the requirements for analysing vibration signals, which are non stationary
signals. Preliminary time-frequency analysis, Windowed Fourier Transform [73] and
Short Time Fourier Transform (STFT) [74] were used to monitor condition of
machinery.
The Wigner Distribution [73, 75-77] and the Spectrogram [36, 78, 79] are the most
well-known quadratic time-frequency representations belonging to the Cohen class
which have been applied in diagnosing gear faults.
The Spectrogram is defined as
( ) ( ) ( ) ( )2
22,, ∫
∞+
∞−
−−== τττ τπ detwxftSTFTftSPEC fjxx (2.16)
The Wigner Distribution (WD) is defined as
( ) ∫+∞
∞−
−
−
+= τττ τπ detxtxftWD fjx
2*
22,
(2.17)
The Wigner-Ville distribution has been used in the diagnosis of faults of rotating
machines and particularly in the analysis of gearbox faults. However, it is
computationally expensive to perform. The basic nature of such signals causes
significant interfering cross-terms, which do not permit a straightforward
interpretation of the energy distribution. To counter this, a modified version was
proposed by Choi and Williams, which has become more widely used. It reduces the
effect of the artefacts by the use of a kernel to minimise the cross terms. The Choi-
Williams distribution is defined:
CHAPTER 2. LITERATURE REVIEW
____________________________________________________________________
39
( ) ( ) ( )[ ] τττστπ
ω τωστ duduxuxetP jtuCW
+
−= −−−∫∫ 22
1
4
1,
22 4/
22/3
(2.18)
where σ is constant.
A new signal processing technique, the Directional Choi-Williams Distribution
(dCWD), is proposed to account for complex-valued time-varying signals, which
represent the planar motion of rotating machinery at each instant of time [80].
Directional Wigner Distributions (DWDs) defined for the forward and backward
pass analytic signals [81] have been applied in analysing order of rotating machines.
Higher-order statistics for fault detection within mechanical systems based on the
observation that impulsive signals tend to increase the Kurtosis values. The use of
the third- and fourth-order Wigner moment spectra, called the Wigner bi- and tri-
spectra respectively [82].
The sliced Wigner fourth-order moment spectra for multiple signals had problems
with its application which were due to the existence of non-oscillating cross-terms
not smoothed by conventional methods. A new method was developed to smooth
non-oscillation cross-terms. The techniques developed are applied to the diagnosis of
valve system faults in an engine [83].
The Continuous Wavelet Transform (CWT) has been used in time-frequency
analysis of rotating machinery fault diagnosis [35, 38, 84, 85].
Continuous Wavelet Transforms (CWTs) [86] were widely recognized as effective
tools for vibration-based machine fault diagnosis, as CWTs could detect both
stationary and transitory signals. However, due to the problem of overlapping, a large
amount of redundant information existed in the results that were generated by CWTs.
The appearance of overlapping could smear the spectral features and make the results
very difficult to interpret for machine operators. Misinterpretation of results might
lead to false alarms or failures to detect anomalous signals. Moreover, as
conventional CWTs only used a single mother wavelet to generate daughter
wavelets, the distortion of the time waveform in the resultant coefficients was
inevitable. Obviously, this would significantly affect the accuracy in anomalous
signal detection. To minimize the effect of overlapping and to enhance the accuracy
CHAPTER 2. LITERATURE REVIEW
____________________________________________________________________
40
of fault detection, a novel wavelet transform, which was named exact wavelet
analysis, had been designed for use in vibration-based machine fault diagnosis. The
design of exact wavelet analysis was based on genetic algorithms. At each selected
time frame, the algorithms would generate an adaptive daughter wavelet to match the
inspected signal as exactly as possible. The optimization process of exact wavelet
analysis was different from other adaptive wavelets as it considered both the
optimization of wavelet coefficients and the satisfaction of the admissibility
conditions of wavelets. The results obtained from simulated and practical
experiments proved that exact wavelet analysis not only minimized the undesirable
effect of overlapping, but also helped operators to detect faults and distinguish the
causes of faults. With the help from exact wavelet analysis, sudden shutdowns of
production and services due to the fatal breakdown of machines could be avoided.
The scalogram – the squared modulus of the CWT [41] was applied in diagnosing
gears. Vibration signatures are passed into a harmonic wavelet transform algorithm
and the mean square wavelet map [87] when diagnosing rotorcraft planetary
geartrain system.
The Discrete Wavelet Transform (DWT) was also applied in [27, 88, 89]. It was
demonstrated that the Discrete Wavelet Packet Analysis (DWPA) [2], wavelet packet
transform [90-93], discrete wavelet analysis [94] all showed potential in fault
diagnosis. Wavelet transform analysis were extended using the Gabor [95-97]
dictionary in Matching Pursuit [84, 96, 97], and using both continuous and discrete
Morlet wavelets [50, 98].
An approach to diagnose gear faults based on continuous wavelet transform was
presented [99]. Continuous wavelet transform can provide a finer scale resolution
than the orthogonal wavelet transform. It is more suitable for extracting mechanical
fault information. The concept of time-averaged wavelet spectrum (TAWS) based on
Morlet continuous wavelet transform was proposed. Two fault diagnosis methods,
the Spectrum Comparison Method (SCM) and Feature Energy Method (FEM) based
on TAWS were established. The results of the application to gearbox gear fault
diagnosis showed that TAWS could effectively extract gear fault information. The
feature energy of the TAWS identifies gear fault advancement very well and is
conically proportional to gear fault advancement.
CHAPTER 2. LITERATURE REVIEW
____________________________________________________________________
41
An adaptive wavelet filter based on the Morlet wavelet was introduced by Lin [100].
The parameters in the Morlet wavelet function were optimised based on the Kurtosis
maximisation principle. The wavelet used was adaptive because the parameters were
not fixed. The adaptive wavelet filter was found to be very effective in detecting
symptoms from vibration signals of a gearbox with early fatigue tooth crack. Two
types of discrete wavelet transform (DWT), the decimated with DB4 wavelet and the
undecimated with harmonic wavelet, were also used to analyse the same signals for
comparison. No periodic impulses appeared on any scale in either DWT
decomposition.
Non orthogonal wavelets are more like vibration impulses [50]. Lin denied that non
orthogonal wavelets can be used in DWT. However, in practice, CWT has been
approximated by DWT [101]. Orthogonality is often not crucial in the post-
processing of signal coefficients. One may thus further enlarge the freedom of choice
by approximating the signal with non-orthogonal vectors, chosen from a large and
redundant dictionary [102] .
Most applications of wavelet bases exploit their ability to efficiently approximate
particular classes of functions with few non-zero wavelet coefficients. This is true
not only for data compression but also for noise removal and fast calculations. In
signal processing, orthogonal bases are of interest because they can efficiently
approximate certain types of signals with just a few vectors. Two examples of such
applications are image compression and the estimation of noisy signals.
From families of wavelet packet bases and local cosine bases, a fast dynamical
programming algorithm is used to select the “best” basis that minimizes a Schur
concave cost function. Pursuit algorithms generalize these adaptive approximations
by selecting the approximation vectors from redundant dictionaries of time-
frequency atoms, with no orthogonality constraint.
A time-frequency-scale domain (TFS) technique has been developed to diagnose
machine faults [103]. The method was proposed to detect transients in mechanical
systems by matching wavelets with associated signals, leading to a development of
joint time-frequency-scale distribution. The three variables, the time, frequency and
scale, maximised the chance for finding similar signal segments from a system under
CHAPTER 2. LITERATURE REVIEW
____________________________________________________________________
42
inspection. The sensitivity was shown to be very high due to closer matching and
better choice of wavelet shapes, which was essential for early fault detection and
failure prevention. Fundamental types of wavelets were introduced based on the
shapes of widely encountered system responses. A method of processing the three-
dimensional image was suggested for interpreting the time-frequency-scale wavelet
map, where the properties of the object patterns uncovered the features of a signal
source, so as to understand the defect and to indicate the condition of a diagnosed
system. The joint distribution was demonstrated to be useful in detecting transients
from different mechanical systems.
2.3.1.3 The Application of Vibration Techniques
In conclusion, statistical parameters have been applied to analysing time domain
signals. The FFT and high order spectrum analyses of signals have been applied in
the frequency domain. Wavelet transforms can show features of signals in the
frequency as well as the time domains. The resolution of frequency and time can be
adjusted. Wavelet transformation has several limitations. A drawback of the wavelet
transform is that it has overlapping or inter terms between neighbouring frequency
and time. The Wigner distribution is a proposed algorithm using differentiation to
reduce the overlapping [104]. In discrete wavelet transform or wavelet packet
analysis, the best basis of discrete wavelet analysis was being investigated to detect
component information required for specific applications. In fault diagnosis,
characteristic frequencies lie in low as well as high frequencies. This makes dyadic
Discrete Wavelet Transform unreliable for fault diagnosis of rotating machinery due
to its low frequency focus. However, Discrete Wavelet Packet Analysis(DWPA)
determines packet coefficients both in low frequencies and in high frequencies,
which makes it useful in fault diagnosis of rotating machineries.
2.3.2 Artificial Intelligence Techniques in Fault Identif ication
The call for automatic diagnosis of faults in machinery systems is gaining more and
more importance, due to the increasing complexity and riskiness of modern
machinery systems, and the growing demands for quality, cost efficiency,
availability, reliability and safety. Artificial intelligence applied in fault diagnosis
brings two benefits: firstly, it reduces the negative effects of conventional human
CHAPTER 2. LITERATURE REVIEW
____________________________________________________________________
43
interpretation such as time consumption; secondly, it improves the reliability of the
fault diagnosis because of the removal of human subjectivity during the diagnostic
procedure.
To achieve the goal of automatic diagnosis, fault detection and identification
sometimes employs artificial intelligent (AI) approaches for pattern classification.
Numerous attempts have been made to improve the accuracy and efficiency of fault
diagnosis of rotating machinery by employing AI techniques. Few have attempted to
summarise these techniques comprehensively. Zhong [105] introduced new
developments in the theory and application of intelligent condition monitoring and
diagnostics in China. He concluded that the trends in intelligent diagnosis are NN-
based fault classifiers, NN-based expert systems, NN-based prognosis, behaviour-
based intelligent diagnosis, remote distributed intelligent diagnosis networks and
intelligent multi-agent architecture for fault diagnosis. He provided a good overview
of intelligent fault diagnosis of machinery but was somewhat general.
Pham [106] theoretically analysed the applicability of artificial intelligence in
engineering problems and predominantly looked at knowledge-based systems, fuzzy
logic, inductive learning, Neural Networks and genetic algorithms in different
branches of engineering but not in machinery fault diagnosis. Tandon [29] mentioned
that automatic diagnosis was a trend in the fault diagnosis of rolling elements. Gao
[107] gave an up-to-date review on recent progress of soft computing methods-based
motor fault diagnosis systems. He summarized several motor fault diagnosis
techniques using Neural Networks, fuzzy logic, neural-fuzzy, and genetic algorithms
(GAs) and compared them with conventional techniques such as direct inspection.
Chow [30] also gave a brief review, which listed 14 papers from experts in the area
of motor fault detection and diagnosis. He grouped those papers into five main
categories: survey papers, model-based approaches, signal processing approaches,
emerging technology approaches, and experimentation.
This section presents a review of the application of Artificial Intelligence (AI)
techniques in fault diagnosis of rotating machinery. In the literature, it appears that
diagnostics techniques for different rotating machinery parts can have much in
common. Previous research has been conducted predominately on specific
components and faults in rotating machinery. This section provides a review of
CHAPTER 2. LITERATURE REVIEW
____________________________________________________________________
44
techniques concerned with rotating elements and shows the potential of generic
application of techniques over a range of rotating elements.
The second part of this section presents a summary of AI based intelligent fault
diagnosis techniques. These are grouped into the following categories: Neural
Networks, fuzzy sets, expert systems, and hybrid approaches (as shown in Figure
2.9). The limitations and strengths of these different techniques are addressed
respectively.
Figure 2.9: An overview of fault detection and identification techniques
2.3.2.1 Neural Networks Based Fault Diagnosis
In general, intelligent diagnosis is carried out when known inputs are fed into black
boxes, which subsequently produce outputs in accordance with machine faults.
Neural Networks (NN) are suitable for these tasks and have been widely researched
as an artificial intelligence tool for machinery fault diagnosis [1]. By employing such
a tool, maintenance personnel need not understand or operate the internal
mechanisms of a Neural Network. He or she will only be responsible for inputting
the appropriate data to a Neural Network. The Neural Network will then be trained
on this data so that it can diagnose faults.
Neural Networks have become widely adopted in rotating machinery for fault
diagnosis due to their learning ability. The learning ability of NN makes Neural
Networks capable of tackling a new problem by making use of existing information.
In addition, Neural Networks also have the capacity to model complex nonlinear
problems, which may approximate real world fault diagnosis.
In rotating machinery fault diagnosis, a Neural Network is usually trained on features
extracted from obtained signals. The Neural Network is usually viewed as a fault
classifier. To date, several types of Neural Networks [17] have been applied in
machinery fault diagnosis including Back Propagation for Feed Forward Networks
Fault detection and identification
Neural Networks Fuzzy set Expert systems Hybrid AI techniques
CHAPTER 2. LITERATURE REVIEW
____________________________________________________________________
45
(BPFF), Multi Layer Perceptrons (MLP), Back Propagation Multiplayer Perceptrons
(BPMP), Radial Basis Function networks (RBF), Self Organised Maps (SOM), and
Principal Component Analysis (PCA).
A Feed Forward Neural Network consists of several successive layers of sigmoid
neurones, where every neuron in a layer receives as inputs the outputs of all the
neurones of the preceding layer. The input to the network is the n inputs to the
neurones in the first layer. The outputs of the m neurones in the last layer are the
network outputs.
The most common application of Feed Forward neural nets is for finding a mapping
that best interpolates from a given set of sample points of a multivariate function. For
this purpose the neural net designer has a choice of two variants: the multi-layer
perceptron (MLP) and Radial basis functions.
Multi-layer perceptrons are usually used for pattern recognition. Each perceptron has
the function of outputting either “0” or “1”. This makes it suitable for classification.
Radial basis function networks use Gaussian transfer functions for their nodes. Also
the inputs are not aggregated by an affine transform but by the Euclidean distance
from a centre. Each neuron has a different centre. Therefore the output of the neuron
decreases exponentially as the input point moves away from the centre.
Principle Components is a set of variables that define a projection that encapsulates
the maximum amount of variation in a dataset and is orthogonal (and therefore
uncorrelated) to the previous principle component of the same dataset. Principle
component analysis (PCA) is commonly used as a cluster analysis tool. It is designed
to capture the Variance in a dataset in terms of principle components. In effect, one is
trying to reduce the dimensionality of the data to summarise the most important (i.e.
defining) parts whilst simultaneously filtering out noise. Normalisation, however,
can sometimes remove this noise and make the data less variate, which could affect
the ability of PCA to capture data structure.
Back propagation is a commonly used training algorithm in the training of a Neural
Network. All back propagation methods are to apply chain rules for ordered partial
derivatives, which can be used to calculate the sensitivity that a cost function has
with respect to the internal states and weights of a network. In other words, the term
CHAPTER 2. LITERATURE REVIEW
____________________________________________________________________
46
back propagation is used to imply a backward pass of error to each internal node
within the network, which is then used to calculate weight gradients for that node.
Learning progresses by alternately propagating forward the activations and
propagating backward the instantaneous errors.
Recurrent Networks have the same characteristics as the standard Feed Forward
networks, but with feedback connections (as shown in Figure 2.10). Because of these
feedback connections, cycles are present in the network. Therefore, training is
sometimes iterated for a long period of time before a response is produced.
Figure 2.10: A Recurrent Neural Network
Recurrent networks tend to be more difficult to train than feed forward networks, as a
result of the cycles, though there are still a fair number of algorithms which are
frequently used, including Jordan, Elman or time delay networks. Some problems are
particularly suited to the use of recurrent networks in favour of Feed Forward
networks, such as time series prediction.
A Self Organising Map (SOM) Neural Network defines a mapping from input signal
of arbitrary dimension to a one or two dimensional array of nodes. This array of
nodes corresponds to a discrete map. The biological basis of a SOM is that sensory
inputs, such as motor, visual, auditory etc., are mapped onto corresponding areas of
the cerebral cortex in an orderly fashion. The neurons are close together and interact
Output
Input
Biases
CHAPTER 2. LITERATURE REVIEW
____________________________________________________________________
47
via short synapse connections. Figure 2.11 shows a one dimensional SOM structure.
Figure 2.12 displays an SOM in two dimensions.
Figure 2.11: One Dimensional Self Organising Map
As can be seen from Figure 2.11, the SOM is essentially a two layer Neural Network
with full connections, only with connections between neurons in the output layer.
Methods for learning in such architecture are to be discussed in the following
section.
Figure 2.12: Two dimensional Self Organising Map
Two dimensional output layer
Input node
One dimensional output layer
Input layer
CHAPTER 2. LITERATURE REVIEW
____________________________________________________________________
48
SOMs represent a topology differently because the connections exist between the
input layer and output layer as well as between neurons in the output layer.
Connections do not exist between neurons in the output layer for other network
types. The Kohonen SOM can have the following attributes:
Competition: Neurons in the network compete to see which input is closest to an
input pattern.
Cooperation: The winning neuron determines a neighbourhood - a group of
networks close by providing the basis for the neighbouring neurons to cooperate,
which is to increase the weights together.
Adaptation: The excited neurons steadily adapt their values to be closer to that of
the input pattern through adjustments to their weights.
Neural Networks, which have been successfully applied in rotating machinery fault
detection and identification, are listed in Table 2.2. The table provides a preliminary
summary of the type of NN used and its application in diagnosing faults in common
machine components.
Table 2.2 Neural Networks applied in diagnosing rotating machinery faults
Neural Networks
pumps rolling bearings
gears gearboxes shafts fans motor
RNN [108] [108] RBF [1] BPFF [75, 109] [1] [110] [109] [111] [108] MLP [63,
112] [63, 112]
[112] [63, 113]
[114] [112]
Kohonen SOM
[115]
LVQ [63] [63] [63]
where, FFNN represents a Feed Forward Neural Network, and RNN represents a
Recurrent Neural Network.
Lu [66] processed the vibration signal of the rolling bearing in motors using fast
Fourier transformation and auto power spectrum estimation. The natural frequencies
of the rolling bearing could be obtained and were used for diagnosing faults through
CHAPTER 2. LITERATURE REVIEW
____________________________________________________________________
49
Neural Network training period and identification period by using back-propagation
feed forward multilayer Neural Networks. Lu[66] obtained positive results from
rolling element bearings in electrical machines, which showed the feasibility and
effectiveness of his technique. It was clearly shown that the frequency signature was
more sensitive for damage detection than other signatures.
Baillie and Mathew’s results [1] in diagnosing faults of rolling element bearings
indicated that back propagation Neural Networks generally outperformed the radial
basis functions.
In the research of Baillie and Mathew, they introduced the concept of fault diagnosis
using an observer bank of autoregressive time series models. The concept was
applied experimentally to diagnose a number of induced faults in a rolling element
bearing using the measured time series vibration signal. Three distinct techniques of
autoregressive modelling were compared for their performance and reliability under
conditions of various signal lengths. The results indicated that back propagation
Neural Networks generally outperformed the radial basis functions and the
traditional linear autoregressive models. This model-based technique for fault
diagnosis was found to require much shorter lengths of vibration data than traditional
pattern classification techniques used in the field of machine condition monitoring.
Three techniques of autoregressive modelling were: the Box-Jenkins linear
autoregressive models, back propagation Neural Networks, and radial basis function
networks.
A back-propagation Neural Network is a multi-layer feed-forward architecture
Neural Network. In Baillie and Mathew’s work, the inclusion of a hidden layer
allowed the network to perform a non-linear mapping of vectors from input to
output. The output was only one node - the disadvantage is that a systematic method
for selection of the most appropriate parameters did not exist. Thus the building and
training of Artificial Neural Networks (ANN) typically required a trial and error
approach (and some experience). When data length was as long as 500 vector
representations, the classification of faults was 100% correct. It needed half the
length of data for accurate classification of the faults.
CHAPTER 2. LITERATURE REVIEW
____________________________________________________________________
50
Radial Basis Functions were characterised by the localised response of hidden layer
nodes, making some applications highly suited to this approach. Another important
feature of radial basis functions was their extremely rapid training times compared to
back propagation networks. Researchers had repeatedly observed training times of an
order of magnitude quicker.
Radial basis functions generally failed to satisfactorily classify vibration signals,
compared to the performance of the linear autoregressive models and back
propagation Neural Network models. It was also observed that the operation of the
radial basis functions was much slower than either the linear autoregressive models
or the back propagation models. This was attributable to the larger number of hidden
layer nodes inside the network, requiring many floating point calculations to be
performed.
McCormick and Nandi [113] used multi-layer perceptron and radial basis function
Neural Networks to classify the condition of a small rotating machine. It was found
that both networks achieved similar success in fault classification. However, Radial
Basis Function (RBF) networks could be trained in a significantly shorter length of
time. Multi-layer perceptrons (MLP) required fewer neurons and were faster in recall
operation.
Meesad and Yen [63] applied MLP and Learning Vector Quantization (LVQ)
classifiers to diagnosing faults in gears, bearings, and shafts. It was proven that these
two Neural Networks were successful while they required off-line training and
iterative presentation of the training data. They were cumbersome when applied to
pattern classification problems that needed fast, on-line, real-time, incremental
learning.
A diagnostic algorithm using Kohonen’s network proposed by Tanaka [115-120]
could classify data obtained from a motor in a plant. The proposed diagnostic
algorithm was compared with Back Propagation -networks for both fault detection
and fault identification. In Tanaka’s application, it was proven that the performance
of the Kohonen Self Organised Map (SOM) was better than Back Propagation
networks. It was also concluded that the map size of the Kohonen’s network was a
factor, which affected its performance.
CHAPTER 2. LITERATURE REVIEW
____________________________________________________________________
51
A Kohonen Self-Organizing Map (KN) was developed for pattern recognition, and
had been extended to fault classification. However, the KN could not be applied to
classify faults from the system output if it contains other factors, such as system state
and sensor mounting errors. To overcome this problem, a constrained KN (CKN)
[121] was proposed. To eliminate the effect of the system state and the mounting
errors, it was proposed that the weight vectors of the CKN were constrained in the
parity space. The training algorithm of the CKN was derived, and its convergence
discussed. Application of the CKN to fault classification was presented, and its
performance was illustrated by an example involving a redundant sensor system with
six sensors.
The performance of Neural Networks is sometimes subjective. The selection of the
network structure may affect the network performance to a significant extent. Some
Neural Networks can have very slow convergence speeds. Training these Neural
Networks take a substantial length of time. Even worse, training can fall victim to the
law of diminishing returns. In particular, the slow operating speed of some Neural
Networks is unable to meet the requirement that machinery be diagnosed on line or
in certain timeframes. Furthermore, a Neural Network can output local optimisation
thereby not guaranteeing an optimal solution. This disadvantage of Neural Networks
can lead directly to errors in diagnosis. Yet another limitation of NN is the lack of
semantics [93].
2.3.2.2 Fuzzy Sets Based Fault Diagnosis
Fuzzy logic-based fault diagnosis methods have the advantages of embedded
linguistic knowledge and approximate reasoning capability. The Fuzzy logic
proposed by Zadeh [12] performs well at qualitative description of knowledge.
However, the design of such a system depends heavily on the intuitive experience
acquired from practicing operators thus resulting in subjectivity of diagnosed faults.
The fuzzy membership function and fuzzy rules cannot be guaranteed to be optimal
in any case. Furthermore, fuzzy logic systems lack the ability of self-learning, which
is compulsory in some highly demanding real-time fault diagnosis cases [107].
Rough set based intelligence diagnostic systems have been constructed and used in
CHAPTER 2. LITERATURE REVIEW
____________________________________________________________________
52
diagnosing valves in three-cylinder reciprocating pumps [90] and turbogenerators
[122, 123].
Miguel [124] described a failure detection and diagnosis method for rotating
machinery which combines fuzzy logic with vibration analysis techniques. Vibration
analysis provided diagnosis information that contained uncertainties from several
sources such as measurement conditions and disturbances. However, this information
was qualitatively meaningful and redundant enough so as to obtain a reliable fault
detection and diagnosis by using fuzzy knowledge processing. Fault sensibility of the
different symptoms was the key factor to establish a fault diagnosis by a fuzzy
processing of those symptoms.
2.3.2.3 Expert Systems Based Fault Diagnosis
Expert systems are computer programs embodying knowledge about a narrow
domain for solving problems related to that domain. An expert system usually
comprises two main elements, a knowledge base and an inference mechanism. The
knowledge base contains domain knowledge which may be expressed as any
combination of ‘IF-THEN’ rules, factual statements, frames, objects, procedures and
cases [106].
A fault diagnosis expert system is a system with a knowledge base storing the
accumulated experience of fault diagnostics experts. However, expert systems can
only diagnose faults stored in the knowledge base and are thus unable to tackle new
problems which have not been classified, i.e., a fault diagnosis expert system cannot
diagnose new machinery faults when compared to real human expertise.
Furthermore, building knowledge bases is labour oriented and time consuming. This
often makes other AI techniques as the preferred ones due to their flexibility and
efficiency in diagnosing rotating machinery faults.
Available expert systems for rotating machinery fault diagnosis are limited.
Amethyst [125] is an expert system to assist with vibration based condition
monitoring of rotating machinery. Vibration patterns are collected from machinery
such as pumps, fans, motors and generators during normal operating conditions. The
Amethyst software performs automatically the same analysis of these vibration
patterns as an experienced machinery expert. The result of this analysis, is a
CHAPTER 2. LITERATURE REVIEW
____________________________________________________________________
53
diagnosis as to whether the machine is working properly and what the problems are.
Amethyst has been used by approximately 400 companies world-wide, making it one
of most widely used expert systems in the manufacturing industry.
Another expert system [126] developed by El Adawi performs preventive
maintenance tasks and detects faults/failure during standard operating cycles. The
expert system was a combination of an intelligent inference engine matched with a
database of information. That system enabled the operator to spot instantaneously the
parameters of interest. Predictive maintenance enabled the operator to minimize the
shut down time of faulty equipment and hence increased productivity. Furthermore,
the system minimized probable human faults and reduced production costs. The
design of this system was based upon the data from measuring equipment linked with
the information system for a Syngas compressor and other rotating machinery such
as pumps, motors, and turbines, etc.
Georgin [127] carried out research based on two of EDF's diagnostic expert systems,
DIVA (for turbine-generators) and DIAPO (for primary cooling pumps), which
interactively deals with users with different knowledge levels.
Shao [128] proposed a new concept on the degree of credibility of parameter value
variations (DCPV factor) and developed an expert system for on-line monitoring and
diagnostics of rolling element bearings.
2.3.2.4 Other Artificial Intelligence Techniques Based Fault Diagnosis
More and more hybrid artificial intelligence techniques are applied to fault diagnosis
of rotating machinery. Neurofuzzy computing recently emerged as an alternative
technique to diagnosing rotating machinery faults [63, 118, 129-139]. Neurofuzzy
techniques work in the following way: A fuzzy set interpretation is incorporated into
the network design to handle imprecise information. Neural Network architecture is
used to automatically deduce fuzzy if-then rules based on a hybrid supervised
learning scheme. For instance, Altmann and Mathew successfully applied an
adaptive neural fuzzy inference system (ANFIS), using a zero-order sugeno fuzzy
model (shown in Figure 2.13) in rolling element bearings fault diagnosis [139].
CHAPTER 2. LITERATURE REVIEW
____________________________________________________________________
54
input inputmembership
function
outputmembership
function
weightedsum
output
rule output
x
y W2Z2
W2
W1 W1Z1
x y
x y
A1
A2
B1
B2
normalisation node
B2
Σ
Σ
W1Z1+W2Z2
W1+W2
W1Z1+W2Z2
W1+W2
Figure 2.13: A Zero-Order Sugeno Fuzzy Model
A fuzzy expert system [140, 141] for real-time process condition monitoring and
incident prevention was developed. Its reasoning strategy was based on dynamic
membership functions of fuzzy systems. The fuzzy expert system could codify the
expertise knowledge to handle incidents, perform process condition monitoring, and
provide operation support with a multimedia user interface. The prototype of this
system was successfully used in a chemical pulp mill for process condition
monitoring and incident prevention.
Artificial Neural Network (ANN) technology and fuzzy logic [118] were applied to
the field of on-line machine condition monitoring (CM) for complex electro-
mechanical systems. A method utilising a combination of human knowledge,
encoded using techniques borrowed from fuzzy logic, Kohonen Neural Networks and
statistical K-means clustering was constructed. The methodology was discussed by
means of a direct comparison between this new approach and a purely neural
approach. An analysis of other situations where this approach would be applicable
was also presented and it is discussed other current research work in the area of
hybrid AI technologies which should assist further with the alleviation of the
problems under consideration.
In mechanical equipment monitoring tasks, fuzzy logic theory was applied to
situations where accurate mathematical models were unavailable or too complex to
CHAPTER 2. LITERATURE REVIEW
____________________________________________________________________
55
be established, but some obscure, subjective and empirical knowledge about the
problem under investigation may still exist [132]. Such knowledge is usually
formalized as a set of fuzzy relationships (rules) on which the entire fuzzy system is
based upon. Sometimes, the fuzzy rules provided by human experts were only partial
and rarely complete; while a set of system input/output data were available. Under
such situations, it was desirable to extract fuzzy relationships from system data and
combine human knowledge and experience to form a complete and relevant set of
fuzzy rules. The application of the B-spline Neural Network to monitor centrifugal
pumps was described. A neuro-fuzzy approach was established for extracting a set of
fuzzy relationships from observation data, where a B-spline Neural Network was
employed to learn the internal mapping relations from a set of features/conditions of
the pump. A general procedure was set up using the basic structure and learning
mechanism of the network and finally, the network performance and results were
discussed.
Turbo machinery diagnosis [142] used fuzzy logic and certainty factor techniques for
uncertainty maintenance. For each observation of symptom provided by the user, the
system would apply a set of diagnostic rules stored in the knowledge base to
postulate currently possible causes. This system could deal with all the 88 possible
symptoms typically observed in the field.
A probabilistic Neural Network for automatic diagnosis was able to deal with the
ambiguous diagnosis problems. Kawabe [143] proposed the "partially-linearized
Neural Network (PNN)" by which failure types could be quickly distinguished on the
basis of the probability distributions of symptom parameters. The knowledge
acquisition method for PNN learning using rough sets was also proposed. These
methods were applied to the failure diagnosis of centrifugal pumps, which were used
in a chemical plant. The results of the failure diagnosis verified that the methods
were effective. The methods could also be applied to other diagnosis or pattern
recognition problems.
A more recent development, Support Vector Machines (SVMs) has been used in
diagnosing rolling bearings and compared with ANNs [144]. Features, which are
used in these two strategies, include statistical features and spectral feature. The
performance of these features as input are compared with selected data of features
CHAPTER 2. LITERATURE REVIEW
____________________________________________________________________
56
using genetic algorithms. The SVMs have a high success rate in this work. However,
training the SVMs took longer than training an equivalent ANN. Additionally,
dependent upon the number of support vectors used in the trained solution, the SVM
may also be significantly slower to operate than the ANN.
Jack combined ANN and genetic algorithms in the fault detection of rotating
machinery [145]. In any given scenario, there were many different possible features
that may be used as inputs for the ANN. The main problem in the use of ANNs was
the selection of the best inputs to the ANN, allowing the creation of compact, highly
accurate networks that require comparatively little preprocessing. Jack proposed a
genetic algorithm to tackle the problem in the application of ANN to detect faults in
rotating machinery and examined the use of a genetic algorithm (GA) to select the
most significant input features from a large set of possible features in machine
condition monitoring contexts. Using the GA, a subset of six input features was
selected from a set of 66, giving a classification accuracy of 99.8%, compared with
an accuracy of 87.2% using an ANN without feature selection and all 66 inputs.
From a larger set of 156 different features, the GA was able to select a set of six
features to give 100% recognition accuracy.
In pump condition monitoring, ANN, mostly, with Feed-Forward Multiple Layered
(FFML) network structure, and Back-Propagation (BP) training method was seen to
be an effective strategy. Wang [146] proposed a systematic approach for constructing
FFML network structure based on the related concepts of Genetic Algorithms (GA).
However, when applying this strategy, a potential risk was the selection of the
network structure, which was poorly defined and affected the network performance
to a large extent.
2.3.3 Artificial Intelligence & Wavelet Transform for Fau lt Diagnosis
Several researchers have used wavelet transforms and Neural Networks with success
in fault diagnosis. Engin [110] outlined a wavelet transform (WT) based artificial
Neural Network (ANN) input data pre-processing scheme and presented the results
of localized gear tooth defect recognition tests by employing this proposed
methodology. The methodology consisted of calculating Daubechies' 20-order
(DAUB-20) mean-square dilation WT of the data, and then selecting predominant
CHAPTER 2. LITERATURE REVIEW
____________________________________________________________________
57
wavelet coefficients distributed to certain levels of these WT as inputs to ANN for
pattern recognition. The test results showed that a fairly small sized backpropagation
network trained with a reasonably small number of training sets could detect and
classify various types or degrees of failures occurring on a spur gear pair
successfully.
Lopez [85] presented successful initial results applying continuous wavelet
transforms coupled with conventional Neural Networks to the development of a real-
time fault detection and classification systems. The development of the real-time
fault detection and identification technologies allowed a migration, in the respective
theatre of operation, from expensive scheduled based maintenance to the more
efficient, less costly alternative of condition based maintenance. The approach
produced results in a general methodology which was shown to work equally well on
fault-seeded, helicopter gearbox data and operational data from Navy shipboard
pumps. The family of wavelet basis functions were specifically engineered to allow
for real-time implementation. The wavelet basis functions had a time-scale
decomposition mathematically inspired from biological systems and provided a
clustering in feature space, which allowed for the development of simplified Neural
Network classifiers. Application to various classes of fault data (helicopter and
shipboard pump data) resulted in perfect detection, no false alarms with only modest
deferral rates.
Paya [112] proposed the purpose of condition monitoring and fault diagnostics was
to detect and distinguish faults occurring in machinery, in order to provide a
significant improvement in plant economy, reduce operational and maintenance costs
and improve the level of safety. The condition of a model drive-line, consisting of
various interconnected rotating parts, including an actual vehicle gearbox, two
bearing housings, and an electric motor, all connected via flexible couplings and
loaded by a disc brake, was investigated. This model drive-line was run in its normal
condition, and then single and multiple faults were introduced intentionally to the
gearbox, and to the one of the bearing housings. These single and multiple faults
studied on the drive-line were typical bearing and gear faults which may develop
during normal and continuous operation of this kind of rotating machinery. He
presented the investigation carried out in order to study both bearing and gear faults
CHAPTER 2. LITERATURE REVIEW
____________________________________________________________________
58
introduced first separately as a single fault and then together as multiple faults to the
drive-line. The real-time domain vibration signals obtained for the drive-line were
preprocessed by wavelet transforms for the Neural Network to perform fault
detection and identify the exact kinds of fault occurring in the model drive-line. It is
shown that by using multi-layer artificial Neural Networks on the sets of
preprocessed data by wavelet transforms, single and multiple faults were successfully
detected and classified into distinct groups.
Classical signal processing techniques when combined with pattern classification
analysis could provide an automated fault detection procedure for machinery
diagnostics [147]. Artificial Neural Networks had been established as a powerful
method of pattern recognition. The Neural Network-based fault detection approach
usually required pre-processing algorithms, which enhance the fault features,
reducing their number at the same time. Various time-invariant and time-variant
signal pre-processing algorithms were studied here. These included spectral analysis,
time domain averaging, envelope detection, Wigner-Ville distributions and wavelet
transforms. A Neural Network pattern classifier with pre-processing algorithms was
applied to experimental data in the form of vibration records taken from a controlled
tooth fault in a pair of meshing spur gears. The results showed that faults could be
detected and classified without errors.
Wavelet transforms were combined with fuzzy sets in the application to fault
diagnosis. Lou [148] dealt with a new scheme for the diagnosis of localised defects
in ball bearings based on the wavelet transform and neuro-fuzzy classification.
Vibration signals for normal bearings, bearings with inner race faults and ball faults
were acquired from a motor-driven experimental system. The wavelet transform was
used to process the accelerometer signals and to generate feature vectors. An
adaptive neural-fuzzy inference system (ANFIS) was trained and used as a diagnostic
classifier. For comparison purposes, the Euclidean vector distance method as well as
the vector correlation coefficient method was also investigated. The results
demonstrate that the developed diagnostic method could reliably separate different
fault conditions under the presence of load variations.
Altmann [2] proposed diagnosis of low speed rolling element using best basis
wavelet transformation and ANFIS. First, he analysed the mathematical model of
CHAPTER 2. LITERATURE REVIEW
____________________________________________________________________
59
rolling element vibration. This model provided both an understanding of the
fundamental cause and behaviour of the vibrations. It also implied that it was
necessary to explore the capabilities and limitations of current signal processing
techniques and to facilitate research into new signal processing techniques. A
traditional technique such as FFT requires knowledge of characteristic frequencies to
diagnose faults, thus demanding maintenance persons know the geometry of rolling
elements. Such requirement increases the complexity of utilising the traditional
techniques. Altmann extracted features using Spectral Peak Ratio (SPR) and Kurtosis
to overcome this disadvantage of traditional techniques. After removing outliners,
the Kurtosis and SPR were used as the input to ANFIS (adaptive neuro fuzzy
inference system) that was based on the Sugeno network. The output of ANFIS was
the membership value of different classification and the severity of different faults.
Although Altmann successfully tested the efficiency of applying DWPA and ANFIS
into fault diagnosis of rolling element bearings, there were still several aspects which
needed to be further investigated. Firstly, the Shannon Entropy was chosen as the
criteria (cost function) to select time-frequency localisation. Some other alternatives
such as the Gauss-Marchov “log energy”, Lp norm, threshold entropy, and the
“SURE” entropy could also be used to select the basis of the tree of frequency
localisation. It had not yet been determined which of these could produce the optimal
basis tree for rotating machinery. Secondly, Debauchies 16 was used as the mother
wavelet to represent the signal collected from rotating bearings. In the field of signal
processing, there were other choices such as the Morlet, Coif, and Meyer wavelets.
Attempts to use the Morlet wavelet and Matching Pursuit occurred previously [84,
96, 97]. Comparison between these mother wavelets needed to be conducted so that
different mother wavelets may be applied for different diagnosis tasks.
2.4 Conclusion of Review
The literature review shows that the following topics need to be developed further.
(1) Developing of more in-depth feature extraction tools and methodologies with
correlations to experts’ interpretation of these features to achieve accurate fault
diagnosis.
CHAPTER 2. LITERATURE REVIEW
____________________________________________________________________
60
(2) Development of advanced artificial intelligence techniques (instead of human
expert interpretation) for diagnosis that uses advanced or traditional extracted
features.
(3) Development of advanced integrated automated diagnosis techniques using a
combination of artificial intelligence techniques and advanced feature extraction
techniques.
In particular, the application of AI techniques to fault diagnosis of rotating
machinery has become prominent due to its potential in being able to reduce
dependency on human expertise.
CHAPTER 3. ADVANCED FEATURE EXTRACTION TECHNIQUES
____________________________________________________________________
61
CHAPTER 3. ADVANCED FEATURE EXTRACTION
TECHNIQUES
3.1 Introduction
This chapter introduces the theoretical foundation of the methodology for feature
extraction investigated in this research, predominantly based on advanced time-
frequency analysis techniques. The representation of time-frequency planes of
signals was further developed to better present features of signals. Further studies of
time-frequency analysis techniques were conducted in this research to better extract
defect related impulsive components or represent time-frequency planes of signals
with better resolution.
DWPA has been applied with some success, to fault diagnosis of rotating machinery.
However, the investigations to date has been inadequate. For instance, the choice of
wavelet functions and best basis has been limited. In this chapter, an improved
DWPA is presented from the aspects of best basis selection as well as wavelet
function selection. A variety of basis selection criteria are presented with an analysis
and a final decision. Waveforms and properties of different wavelet functions are
presented.
Matching Pursuit, a recent technique, has been applied to fault diagnosis of rolling
bearings. It is one of the adaptive approximation techniques with pursuit, where a
variety of dictionaries are available for approximation. However, not all dictionaries
are suitable for the application of rolling bearing fault diagnosis. In this chapter, the
principles of Matching Pursuit are elaborated. The available dictionaries are
introduced and then the dictionary suitable for vibration analysis is selected and
presented in detail.
Another adaptive approximation technique, Basis Pursuit is introduced and is the
first application in rolling bearing fault diagnosis. The selection of wavelet packet
dictionary and optimisation algorithms of Basis Pursuit are presented. A Basis
Pursuit denoising method is employed to analyse vibration of rolling bearings under
different fault conditions.
CHAPTER 3. ADVANCED FEATURE EXTRACTION TECHNIQUES
____________________________________________________________________
62
The computation complexity of Basis Pursuit is discussed and analysed in the
chapter. A comparison of these three time-frequency analysis techniques is also
presented in terms of computation complexity and time-frequency component
resolution.
3.2 Best Basis Discrete Wavelet Packet Analysis
Two problems are often confronted when utilising classical methods such as
demodulation. Firstly, the most appropriate frequency band to demodulate needs to
be chosen. If the chosen frequency band does not consist of resonance excited by
bearing damage, demodulation analysis can fail in defect detection. Impulse
characteristics can change with the development of bearing severity. The impulses
are narrow and prominent when defects are incipient but would broaden and would
usually be embedded in the signal when defects develop in the bearing. The other
intricacy is to catch the transitional nature of defect-generated impulses. The
frequency components can change with time. However demodulation as well as
spectrum can only show certain frequency components of the signals [2]. In order to
overcome these problems, wavelet transform techniques are utilised to present
signals in the time-frequency domain.
In this work, the best basis DWPA for the collected signal was calculated in the joint
time-frequency domain. Wavelet coefficients, which were mainly affected by the
bearing faults, was determined and extracted. The resonance frequency need not be
known when using DWPA. Furthermore, a basis that is well matched to the
characteristics of the signal was chosen to improve the representation of bearing
impacts. Features from the interested frequencies were extracted, while contributions
from noise and other components rather than the interested frequencies were
substantially reduced.
An improved method for extracting bearing fault impulses of vibration signals, using
DWPA best basis to extract the wavelet correlation coefficients at an optimal time-
frequency resolution is presented. The DWPA accomplishes this by forming a well-
balanced binary tree from which the best basis is selected.
To demonstrate the improved best basis DWPA, concepts of wavelet transform are
introduced first and followed by the introduction of the fundamentals of best basis
CHAPTER 3. ADVANCED FEATURE EXTRACTION TECHNIQUES
____________________________________________________________________
63
discrete wavelet packet analysis. Discrete wavelet transform is also used in
comparison with the discrete wavelet packet analysis for further illumination of the
DWPA.
Concepts and fundamentals of DWPA [2]
A wavelet ψ is a function of zero average:
( ) 0=∫+∞
∞−dttψ
(3.1)
Which is dilated with a scale parameter s, and translated by u:
( )
−=s
ut
stsu ψψ 1
, (3.2)
The wavelet transform of f at the scale s and position u is computed by
correlating f with a wavelet ψ
( ) ( ) dts
ut
stfsuWf ∫
∞+
∞−
−= *1, ψ (3.3)
A wavelet orthonormal basis decomposes the frequency axis in dyadic intervals
whose sizes have an exponential growth. This fixed dyadic construction by
decomposing the frequency in intervals whose bandwidths may vary. Each frequency
interval is covered by the time-frequency boxes of wavelet packet functions that are
uniformly translated in time in order to cover the whole plane.
Although the wavelet transform has been widely researched during the last decade,
the founding principles behind wavelets can be traced back as far as 1909 when
Alfred Haar discovered another orthonormal system of functions, such that for any
continuous function f(x), the series
10),2()(0
12
02
<≤−=∑∑∞
=
−
=+ xforkxxf j
j k
j
kj ψα (3.4)
converges to f(x) uniformly over the interval 10 <≤ x . Haar’s research led to the
simplest of the orthogonal wavelets, a set of rectangular basis functions. The Haar
basis function was of limited use due to it being discontinuous in nature. This
resulted in it being inefficient in modelling smooth signals, as many levels need to be
CHAPTER 3. ADVANCED FEATURE EXTRACTION TECHNIQUES
____________________________________________________________________
64
included to obtain an accurate representation. Since then major advances in the
development of wavelets have been presented, which will be introduced in a
following section “wavelet selection”.
The wavelet transform is made up of a scaling function (father wavelet) and an
analysing function (mother wavelet). The scaling function is a solution of the dilation
equation:
( ) ( )∑ −=k
k kxcx 2φφ (3.5)
Where the constant kc satisfies the condition:
∑ =k
kc 2 (3.6)
The dilation equation generates the family of wavelets:
( )∑ −k
k kxc 2φ (3.7)
Where j,k Z∈ .
The sequence ( ){ }Zkj,:xkj, ∈φ
is an orthonomal family in ( )RL2
Equations for the mother wavelet can be derived from the scale function:
( ) ( ) ( )∑ += +k
1 21-x kxckk φψ (3.8)
The mother wavelet generates the family of wavelets:
( ) ( )kxjjkj −= −− 22x 2/, ψψ (3.9)
Where Zkj, ∈ .
As with the scaling function, the sequence ( ){ }Zkj, :x, ∈kjψ is an orthonormal
family in ( )RL2
.
Fundamental concepts of Discrete Wavelet Packet Analysis
Discrete wavelet packets form a redundant dictionary of bases, from which the best
basis to represent a given signal can be selected. Wavelet packets are composed of
CHAPTER 3. ADVANCED FEATURE EXTRACTION TECHNIQUES
____________________________________________________________________
65
elementary functions called wavelet packets ( ){ }Znkjkxjn
j ∈− ,,:22 2/ ω , where j, k
and n represent the index of scale, position and degree of oscillation, respectively.
Hubbard loosely described wavelet packets as the product of a wavelet, a wiggle, and
an oscillating function. The wavelet reacts to abrupt fluctuations in the signal, while
smooth oscillations are altered for by the wiggle function. It is this combination of
the wavelet and wiggle functions that enables the size, frequency and position of
time-frequency atoms all to be varied independently.
As with the wavelet transform, wavelet packets can be represented by a filter bank
constructed from quadrature mirror filters. The construction of the wavelet packet
bases can be expressed as:
A wavelet packet base allows any dyadic tree structure. At each point in the tree we
have an option to send the signal through the low pass-high pass filter bank, or not.
Wavelet packets are generated by the following iterations:
( ) ( )( ) ( )∑
∑
−=−
−=−
−−
+
−−
m
jnkm
jn
m
jnkm
jn
mtgkt
mthkt
222
222
21
12
21
2
ωω
ωω
(3.10)
where h and g represent the respective high- and low-pass quadrature mirror filters,
and 0ω and 1ω correspond to the father wavelet (scaling function) and mother wavelet
(analysing function).
For a given signal the wavelet packet coefficients can be iteratively computed by the
following equations.
∑
∑
−−
+
−−
=
=
k
jknlk
jn
k
jknlk
jn
cgc
chc
,21
1,12
,211,2
(3.11)
DWPA can also be explained in comparison with DWT. DWT decomposes the
signal into ortho-normal “wavelets”, scaled and shifted versions of the “mother
wavelet”,ψ . A function ( )tf can be expressed by its wavelet expansion,
( ) ( ) 10,20
12
020 <≤−+= ∑∑
∞
=
−
=+ tforkttf
j k
j
j
kj ψαα (3.12)
CHAPTER 3. ADVANCED FEATURE EXTRACTION TECHNIQUES
____________________________________________________________________
66
The integer j describes the different levels of wavelets, and k covers the number of
wavelets in each level. DWPA can decompose signals into both low frequency
components and high frequency components in a manner shown in Figure 3.1.
For instance, an example tree of wavelet packet decomposition of a signal with
sampling frequency 16k Hz is shown in Figure 3.2. It can be seen that the children
nodes of the best basis tree are in an order which is different from the numerical
order of the frequency.
Figure 3.1: Filter bank representation of DWT and DWPA [149]
3.2.1 Introduction to Different Wavelet Functions
Any discussion of wavelets begins with Haar wavelet, the first and simplest. Haar
wavelet (3.13) is discontinuous (as shown in Figure 3.3), and resembles a step
function. Daubechies (dbN), Biorthogonal (biorNr. Nd), Coiflet (CoifN), Symlets
(symN), Morlet (morl), Mexican Hat (mexh), Meyer, Batter-Lemarie are some other
wavelets which can be chosen [150].
≤≤−≤≤
=otherwise
x
x
H
0
12/11
2/101
ψ (3.13)
CHAPTER 3. ADVANCED FEATURE EXTRACTION TECHNIQUES
____________________________________________________________________
67
Figure 3.2: An example tree of wavelet packet decomposition [91]
0 0.2 0.4 0.6 0.8 1 1.2 1.4-1.5
-1
-0.5
0
0.5
1
1.5 Haar
Figure 3.3: Harr Wavelet
To further introduce some concepts of Shannon and Meyer wavelets, the following
notations are firstly presented.
Hψ
x
CHAPTER 3. ADVANCED FEATURE EXTRACTION TECHNIQUES
____________________________________________________________________
68
Notations
Let φ be a scaling function and h the corresponding conjugate mirror filter. Let ψ be
the function whose Fourier transform is
( )
=2
ˆ2
ˆ2
1ˆ
ωφωωψ g, (3.14)
with
( ) ( )πωω ω += *ˆˆ heg i . (3.15)
Let us denote
( )
−=j
j
nj
nt
jt
2
2
2
1, ψψ . (3.16)
For any scale j2 ,{ }Ζ∈nnj ,ψ is an orthonormal basis of jW . For all scales, { }( ) 2,, Ζ∈njnjψ is
an orthonormal basis of ( )RL2 .
Shannon wavelet
The Shannon wavelet is constructed from the Shannon multiresolution
approximation, which approximates functions by their restriction to low frequency
intervals. It corresponds to [ ]ππφ ,1ˆ−= and ( ) [ ]( )ωω ππ 2/,2/12ˆ
−=h for [ ]ππω ,−= .
We derive from (3.14) that
( ) [ ] [ ]
∪−−∈=−
otherwise
ifei
0
2,,2ˆ2 ππππωωψω
(3.17)
and hence
( )
−
−−
−
−=
2
12
1sin
2
12
2
12sin
t
t
t
tt
π
π
π
πψ . (3.18)
This wavelet is ∞C but has slow asymptotic time decay. Since ( )ωψ is zero in the
neighbourhood of 0=ω , all its derivatives are zero at 0=ω .
CHAPTER 3. ADVANCED FEATURE EXTRACTION TECHNIQUES
____________________________________________________________________
69
Meyer Wavelets
A Meyer wavelet is a frequency band-limited function whose Fourier transform is
smooth, unlike the Fourier transform of the Shannon wavelet. This smoothness
provides a much faster asymptotic decay in time. These wavelets are constructed
with conjugate mirror filters that are and satisfy
( )
∪
−−∈
−∈=
ππππω
ππωω
,3
2
3
2,0
3,
32
ˆ
if
ifh . (3.19)
The only degree of freedom is the behaviour of in the transition bands. It must satisfy
the quadrature condition
( ) ( ) 2ˆˆ 22
=++ πωω hh , (3.20)
and to obtain nC junctions at 3
πω = and 3
2πω = , the n first derivatives must
vanish at these abscissa. One can construct such functions that are ∞C . The
waveform of a meyer wavelet filter is shown in Figure 3.4. Another wavelet, coiflet
waveform is shown in Figure 3.5.
-8 -6 -4 -2 0 2 4 6 8-1
-0.5
0
0.5
1
1.5 Meyer
Figure 3.4: Meyer wavelet function
( )ωh
ω
CHAPTER 3. ADVANCED FEATURE EXTRACTION TECHNIQUES
____________________________________________________________________
70
0 1 2 3 4 5-2
-1
0
1
2
3(a) Coiflets1
0 5 10 15-1
-0.5
0
0.5
1
1.5
2(b) Coiflets2
0 5 10 15 20-1
-0.5
0
0.5
1
1.5(c) Coiflets3
0 5 10 15 20 25-1
-0.5
0
0.5
1
1.5(d) Coiflets4
Figure 3.5: Coiflets wavelet function
Battle-Lemarié Wavelets
Polynomial spline wavelets introduced by Battle and Lemarié are computed from
spline multiresolution approximations. For splines of degree m, ( )ωh and its first m
derivatives are zero at πω = .
This wavelet has an exponential decay. Since it is a polynomial spline of degree m, it
is m-1 times continuously differentiable. Polynomial spline wavelets are less regular
than Meyer wavelets but have faster time asymptotic decay. For m odd, ψ is
symmetric about 1/2. For m even, it is antisymmetric about 1/2.
Daubechies Wavelets
Daubechies wavelets (see Figure 3.6) have a support of minimum size for any given
number p of vanishing moments.
( )ωh
( )ωh
( )ωh
( )ωh
ω ω
ω ω
CHAPTER 3. ADVANCED FEATURE EXTRACTION TECHNIQUES
____________________________________________________________________
71
0 1 2 3-1.5
-1
-0.5
0
0.5
1
1.5
2(a) Daubechies2
0 2 4 6 8-1
-0.5
0
0.5
1
1.5(b) Daubechies4
0 5 10 15-1.5
-1
-0.5
0
0.5
1(c) Daubechies8
0 5 10 15 20-1.5
-1
-0.5
0
0.5
1(d) Daubechies10
Figure 3.6: Daubechies function
Symmlets
Daubechies wavelets are very asymmetric because they are constructed by selecting
the minimum phase square root of( )ωieQ − . Filters corresponding to a minimum
phase square root have their energy optimally concentrated near the starting point of
their support. They are highly non-symmetric, which yields very asymmetric
wavelets.
To obtain a symmetric or antisymmetric wavelet, the filter h must be symmetric or
antisymmetric with respect to the centre of its support. The Symmlet filters (as
shown in Figure 3.7) of Daubechies are obtained by optimizing the choice of the
square root ( )ωieR − of ( )ωieQ to obtain an almost linear phase.
( )ωh
( )ωh
( )ωh
ω ω
ω ω
( )ωh
CHAPTER 3. ADVANCED FEATURE EXTRACTION TECHNIQUES
____________________________________________________________________
72
0 1 2 3-2
-1.5
-1
-0.5
0
0.5
1
1.5(a) symlet2
0 2 4 6 8-1.5
-1
-0.5
0
0.5
1
1.5
2(b) symlet4
0 5 10 15-1
-0.5
0
0.5
1
1.5(c) symlet8
0 5 10 15 20-1
-0.5
0
0.5
1
1.5(d) symlet10
Figure 3.7: Symlet function
There is trade off between symmetry and computation complexity of a wavelet
function. Among the orthogonal wavelets, Harr wavelets have computational
advantages due to its symmetry property. However, Harr wavelets can hardly be used
in the analysis of vibration because it has straight waveforms, which has nothing in
common with vibration waveforms.
For the wavelet families, their properties are summarized in Table 3.1. It can be
concluded that symlet family and coiflet family have most properties such as
compactly supported, near symmetry, orthogonal analysis. These two wavelet
families will be preferable. In particular, symlet function shows good matched
vibration characteristics, therefore, will be a wavelet suitable for the time-frequency
analysis of the bearing vibrations. In this vibration analysis, symlet functions are
considered for reasons such as orthogonality and their waveforms similar to
vibration. The symlet family includes symlet2, symlet4, symlet8, and symlet10.
Among these wavelet functions, symlet8 appears very much similar to vibration
waveforms so that it is eventually chosen in the feature extraction procedure.
( )ωh
( )ωh
ω ω
ω ω
( )ωh
( )ωh
CHAPTER 3. ADVANCED FEATURE EXTRACTION TECHNIQUES
____________________________________________________________________
73
Table 3.1 Summary of Wavelet Families and Associated Properties (Manual of
wavelet toolbox in Matlab)
Property morl Mexh meyr haar dbN symN coifN biorNr.Nd Crude • • Infinitely regular
• • •
Arbitrary regularity
• • • •
Compactly supported orthogonal
• • • •
Compactly supported biothogonal
•
Symmetry • • • • •
Asymmetry • Near symmetry
• •
Arbitrary number of vanishing moments
• • • •
Vanishing moments for φ
•
Existence of φ
• • • • • •
Orthogonal analysis
• • • • •
Biorthogonal analysis
• • • • • •
Exact reconstruction
• • • • • • •
FIR filters • • • • • Continuous transform
• • • • • • • •
Discrete transform
• • • • • •
Fast algorithm
• • • • •
Explicit expression
• • •
CHAPTER 3. ADVANCED FEATURE EXTRACTION TECHNIQUES
____________________________________________________________________
74
3.2.2 Selection of Best Basis
When analysing a signal using DWPA, the information concerned might lie in any
position on the decomposition binary tree. There exists a variety of selection criteria
to choose the best basis for different application. Additive cost functions such as
entropy based criteria are well suited for efficient searching of binary tree structures,
and provide an accurate information cost function of a given signal. For the cost
function E to be additive it must satisfy the criteria E (0) =0 and E(s) = ( )∑i
isE .
The following is a list of additive cost functions that could be used:
The Shannon entropy
( ) ( )∑−=i
iiaddshannon sssE 22 ln (3.21)
The Gauss-Markov “log energy” entropy
( ) ( )∑=i
iadd ssE 2
Markov-Gauss ln (3.22)
The concentration in Lp norm
( ) 21 ≤≤−= ∑ pssEi
p
iaddnorm (3.23)
The threshold entropy
( )elsewhere
pssE iiadd
0
1 threshold
=>=
(3.24)
The “SURE” entropy
( ) { } ( )∑+≤−=i
iiaddsure pspsthatsuchinsE 22,min#
(3.25)
An alternative to additive cost functions implementing non-additive cost functions
can determine a near-best basis for the tree. A non-additive cost function cannot
guarantee the selection of a best basis, thus the basis selected is referred to as a near-
best basis. A non-additive cost function is defined as any function that can be used to
provide a comparison of basis, but does not satisfy the additive requirements. One
CHAPTER 3. ADVANCED FEATURE EXTRACTION TECHNIQUES
____________________________________________________________________
75
such cost function is the Shannon entropy of a finite scheme, where the probabilistic
events are identified with the normalised energies.
Shannon entropy of a finite scheme
( ) ∑
−= 2
2
2
2
2
2
lni
i
i
inon
s
s
s
ssE (3.26)
Best tree searching algorithms
There are two primary approaches used in searching a tree, the top-down search and
the bottom-up search algorithms. The top-down search algorithm calculates a near-
best or sub-optimal basis from a library of bases. An optimal basis cannot be
guaranteed through this method, as the entire library is not searched. However, this
technique does offer substantial computational cost savings, as the best basis can be
simultaneously calculated as the tree is constructed. This means it is not necessary to
calculate the transform coefficients for the entire tree, as pruning of the tree is done
on the fly, with the search terminating as soon as the sum of the children’s cost
functions is greater than the parents. This process reduces the overall calculation and
storage requirements of the optimisation process. The following steps are used to
locate the sub-optimal tree structure:
If ( )∑≤child
opt childEparentE )( , and the parent is not the root node, then set
( ) )( parentEnodeEopt = and terminate the branch.
If ( ) ( )∑>child
optopt childEnodeE , then split and set ( ) ( )∑=child
optopt childEnodeE .
Determination of the optimal or best basis can be achieved through a bottom-up
search algorithm using an additive cost function. Unlike the top-down algorithm, the
bottom-up approach searches through the entire family of bases, to ensure that it
finds the best basis. To do this it starts at the bottom of the tree and compares the
children’s cost function with that of their parent. If the children’s cost function is
greater than their parents, then the children get pruned from the tree, and the parent
becomes a child node. This procedure can be written as:
CHAPTER 3. ADVANCED FEATURE EXTRACTION TECHNIQUES
____________________________________________________________________
76
If E (parent) ( )∑≤child
opt childE , and the parent is not the root node, then prune the tree
and set ( ) )( parentEnodeEopt = .
If ( ) ( )∑>child
optopt childEnodeE , then set ( ) ( )∑=child
optopt childEnodeE .
An example of the best tree DWPA decomposition of a signal is shown in Figure 3.8,
and Figure 3.9. The example shows the top-down search saves on the computation of
the coefficients and entropy of many nodes.
Figure 3.8: An example of the best tree DWPA decomposition of a signal (with
depth position index)
Figure 3.9: An example of the best tree DWPA decomposition of a signal (with
Entropy value index)
120.09
62.624 64.56
27.934 29.436
4.9837 13.292
3.0661 39.037
18.051 14.827
2.8831 10.86
4.8559 5.1133
17.875 12.115 2.6565 0.6131
1.8115 2.9245
(0,0)
(1,0) (1,1)
(2,0) (2,1)
(3,2) (3,3)
(2,2) (2,3)
(3,6) (3,7)
(4,14) (4,15)
(5,30) (5,31)
CHAPTER 3. ADVANCED FEATURE EXTRACTION TECHNIQUES
____________________________________________________________________
77
Discussion on the selection of best basis
The best basis DWPA was applied to signals from the same bearing. It was seen that
even different data segments of a signal often have different decomposition trees
even if same criterion was used in the selection of best basis. It is usually difficult to
find a unique best basis for signals collected in same situation. Therefore, without
further processing, best tree is mostly suitable for the analysis of individual signal
segments and generally needs human interpretation. In the case a generalization is
needed, a best basis selection is not suitable then unless a unique basis is found
before data is processed. In particular, for the usage of a Neural Network, a great
number of data sets are required for the training process. The Neural Network
requires a unique input vector (i.e. a basis), which cannot be obtained using this
principle of best basis selection. The selection of best basis cannot be directly applied
in automatic diagnosis.
3.3 Adaptive Approximation with Pursuit
3.3.1 Fundamentals of Adaptive Approximation with Pursuit
3.3.1.1 Concepts of Adaptive Approximation
With the advance of time-frequency analysis signal processing techniques, recently,
adaptive approximation techniques [102] have become popular for obtaining
economical and precise representations of signals. In adaptive approximation, the
goal is to find the representation of a signal x as a weighted sum of elements γφ from
an over complete dictionaryΓ
γγ
γφα∑Γ∈
=x (3.27)
Or an approximate decomposition
( )mm
i
Rxii
+=∑=
γγ φα1
(3.28)
Where r is an index of a set,Γ ; γα is the coefficient of the elementγφ ; m is the order
of decomposition; and( )mR is a residual.
CHAPTER 3. ADVANCED FEATURE EXTRACTION TECHNIQUES
____________________________________________________________________
78
Adaptive approximation is accurate in representing signals mostly due to its
employing a set “Dictionary”. There have been a variety of dictionaries developed
substantially in the past. The variety of dictionaries will be introduced in the
subsequent section, where the Discrete Wavelet Packet Dictionary is elaborated and
chosen to analyse vibration signals in fault diagnosis of rolling bearings. An
investigation of the reasons why the discrete wavelet packet dictionary is suitable for
this study is also provided. Principles of two adaptive approximation methods are
introduced and have been applied using this Discrete Wavelet Packet dictionary. The
adaptive approximation methods are Matching Pursuit and Basis Pursuit.
3.3.1.2 Concepts and Principles of Different Dictionaries
Atoms and dictionaries
A considerable focus of activity in the recent signal processing literature has been the
development of basis used such as atoms or wavelets in signal representations.
Terminology introduced by Mallat and Zhang [25] was use. A dictionary Γ is
defined as a collection of parameterized waveformsγφ ,{ }Γ∈γφ γ | . The waveforms
γφ are discrete-time signals of length n called atoms. Depending on the dictionary,
the parameter γ can have the interpretation of indexing frequency, in which case the
dictionary is a frequency or Fourier dictionary, of indexing time-scale jointly, in
which case the dictionary is a time-scale dictionary, or of indexing time-frequency
jointly, in which case the dictionary is a time-frequency dictionary. Usually
dictionaries are complete or overcomplete, in which case they contain exactly n
atoms, or more than n atoms. In certain cases, an overcomplete dictionary can also
have continuum dictionaries containing an infinity of atoms and an undercomplete
dictionary might contain fewer than n atoms.
Interesting dictionaries have been proposed over the last few years including trivial
dictionaries, frequency dictionaries, time scale dictionaries, and many time-
frequency dictionaries, which are introduced as follows.
The Dirac dictionary The Dirac dictionary is simply the collection of waveforms
that are zero except in one point:{ }1,,1,0 −∈ nKγ and γφ . This is an orthogonal basis
of nR the standard basis.
CHAPTER 3. ADVANCED FEATURE EXTRACTION TECHNIQUES
____________________________________________________________________
79
The Heaviside dictionary
The Heaviside dictionary is the collection of waveforms that jump at one particular
point: { }1,,1,0 −∈ nKγ ; ( ) { }γγφ ≥= tt 1 ; Atoms in this dictionary are not orthogonal, but
every signal s has a representation
( ) γγ
γγ φφ ∑−
=−−+=
1
1100
n
ssss (3.29)
Frequency dictionaries
A Fourier dictionary is a collection of sinusoidal waveforms γφ indexed by
( )νωγ ,= , where [ )πω 2,0= is an angular frequency variable and { }1,0∈ν indicates
phase type: sine or cosine. In detail,
( ) ( ),cos0, tωφ ω = ( ) ( )tωφ ω sin1, = . (3.30)
For the standard Fourier dictionary, γ runs through the set of all cosines with Fourier
frequencies 2/,,0,/2 nknkk K== πω , and all sines with Fourier
frequencies .12/,,1, −= nkk Kω This dictionary consists of n waveforms; it is in fact
a basis, and a very simple one: the atoms are all mutually orthogonal. An
overcomplete Fourier dictionary is obtained by sampling the frequencies more finely.
Let l be a whole number > 1 and let lΓ be the collection of all cosines with
2ln/,,0ln,/2 K== kkk πω , and all sines with frequencies 12ln/,,0, −= Kkkω .
This is an l fold over complete system. Both complete and overcomplete dictionaries
based on discrete cosine transforms and sine transforms are also used.
Time-scale dictionaries
Several types of wavelet dictionaries are available. Haar dictionary is considered
with “father wavelet” ( ]1,01=ϕ and “mother wavelet" ( ] [ )2/1,01,2/1 11 −=ψ . The
dictionary is a collection of translations and dilations of the basic mother wavelet,
together with translations of a father wavelet. It is indexed by ( )vba ,,=γ , where
( )∞∈ ,0a is a scale variable, [ ]nb ,0∈ indicates location, and { }1,0∈v indicates
gender. In detail,
CHAPTER 3. ADVANCED FEATURE EXTRACTION TECHNIQUES
____________________________________________________________________
80
( ) ( )( ) abtaba ⋅−=ψφ 1,, , ( ) ( )( ) abtaba ⋅−= ϕφ 0,, . (3.31)
For the standard Haar dictionary, γ runs through the discrete collection of mother
wavelets with dyadic scales ( ) 1log,,,/2 20 −== njjna jj L , and locations that are
integer multiples of the scale 12,,0, −=⋅= jjj kakb L , and the collection of father
wavelets at the coarse scale 0j . This dictionary consists of n waveforms; it is an
orthonormal basis. An overcomplete wavelet dictionary is obtained by sampling the
locations more finely: one location per sample point. This gives the so-called
stationary Haar dictionary, consisting of O (n log2 (n)) waveforms.
A variety of other wavelet bases are possible. The most important variations are
smooth wavelet bases, using splines or using wavelets defined recursively from two
scale filtering relations. Although the rules of construction are more complicated
(boundary conditions [28], orthogonality versus biorthogonality [10], etc.), these
have the same indexing structure as the standard Haar dictionary.
Time-frequency dictionaries
Much recent activity in the wavelet communities has focused on the study of time-
frequency phenomena. The standard example, the Gabor dictionary, is due to Gabor
(1946); in notation, ( )tδθτωγ ,,,= , where [ )πω ,0∈ is a frequency, τ is a location,
θ is a phase, and tδ is the duration, and
atoms ( ) ( ) ( ){ } ( )( )θτωδτφγ +−⋅−−= tttt cos/exp 22 are considered. Such atoms indeed
consist of frequencies near ω and essentially vanish far away fromτ .
For fixed tδ , discrete dictionaries can be built from time-frequency lattices,
ωω ∆= kk and ττ ∆= ll , and { }2/,0 πθ ∈ ; with τ∆ and ω∆ chosen sufficiently fine
dictionaries are complete.
The usage of a time-frequency dictionary provides adaptedness for the
representations of vibration signals. Depending on different dictionaries, a parameter
γ for γφ can have different interpretations. In the application to analysing vibrations
of rolling bearings, the employed dictionary is a time-frequency dictionary,
γ indexes time-frequency jointly. In a time-frequency dictionary, the waveforms may
CHAPTER 3. ADVANCED FEATURE EXTRACTION TECHNIQUES
____________________________________________________________________
81
be chosen from different types of wavelet functions, cosine functions with a great
range of different scales. For instance, the wavelet functions can be a group of
Daubechies functions or a group of Morlet functions. Atoms can be produced by
scaling the function with certain time and frequency dimensions using the same
function.
The wavelets dictionary is only the best known of available alternate dictionries
collections of parameterized waveforms. Wavelets, steerable wavelets, segmented
wavelets, Gabor dictionaries, multiscale Gabor dictionaries, wavelet packets, cosine
packets, chirplets, warplets, and a wide range of other dictionaries are now available.
Time-frequency atoms
Mallat and Zhang [97] proposed that a general family of time-frequency atoms (as
seen in Figure 3.10) can be generated by scaling, translating and modulating a single
window function. It can be considered that wavelet atoms ( )tγψ are wavelets, which
are dilated with a scale parameter s, translated by u, and demodulated byξ . It is
denoted that ( )ξγ ,,su= :
( ) ( ) tisu e
s
ut
stt ξ
ξγ ψψψ
−== 1,,
(3.32)
The index γ is an element of the set2RR ×=Γ +[97]. The factor
s
1normalizes
( )tγψ to 1. The energy of ( )tγψ is mostly concentrated in a neighbourhood of u,
whose size is proportional to s. Let ( )ωψ be Fourier transform of ( )tψ , the Fourier
transform of ( )tγψ is
( ) ( ) ( )( )uiess ξωγ ξωψωψ −−−= ˆˆ
(3.33)
The energy of ( )ωψ γˆ is concentrated in a neighbourhood of ξ , whose size is
proportional to 1/s. The time-frequency representation of a wavelet family is built by
relating the frequency parameter nξ to the scale ns with n
n s0ξξ = , where 0ξ is a
CHAPTER 3. ADVANCED FEATURE EXTRACTION TECHNIQUES
____________________________________________________________________
82
constant. The wavelet atom is capable of catching high-energy components, which
are located in subtle time and frequency intervals.
Figure 3.10: An example of a time-frequency atom plot [151]
Wavelet packet dictionary
Recently, Coifman and Meyer [97] developed the wavelet packet especially to meet
the computational demands of discrete-time signal processing. A wavelet packet
dictionary includes, as special cases, a standard orthogonal wavelet dictionary, the
Dirac dictionary, and a collection of oscillating waveforms spanning a range of
frequencies and durations. This dictionary has wavelet functions (oscillating
waveforms), which are advantageous when diagnosing rotating machines (as
discussed in the section of introduction of wavelet functions). The orthogonality of
this dictionary makes it possible for us to get a unique basis for a signal. The dyadic
dictionary is an efficient way for computation of an approximation of a vibration
signal. The scaling of wavelet packet atoms is more flexible than wavelet transform
and discrete wavelet packet analysis.
CHAPTER 3. ADVANCED FEATURE EXTRACTION TECHNIQUES
____________________________________________________________________
83
For signals of N samples, each vector of a wavepacket dictionary (a wavelet packet
atom γψ ) is indexed by ( )kpj ,,=γ , with jj kNpNj 20,20),(log0 2 ≤≤≤≤≤≤ − .
Such an atom has similar time-frequency localization properties to a discrete window
function, dilated byj2 , centred at
+21
2 pj , and modulated by a sinusoidal wave of
frequency
+−
21
22 kjπ .
A wavelet packet dictionary with a symlet8 wavelet function is suitable for vibration
analysis of rolling bearings due to the following reasons:
(1) it is one type of time-frequency dictionary and able to catch time-frequency
atoms with high energy level. This is desirable for extracting time-frequency high
energy vibration features when diagnosing rolling bearings faults;
(2) The wavelet packet dictionary has wavelet functions (oscillating waveforms),
which match vibration waveforms. This makes their coefficients best represent
vibration level. The more similar the wavelet functions are to the vibrations, the less
number of coefficients are needed to represent the signal. This also helps improve
feature extraction performance.
(3) The orthogonality of this dictionary makes it possible to get a unique basis for a
signal.
(4) The dyadic dictionary is an efficient way for computation of an approximation of
a vibration signal.
3.3.1.3 Computational complexity of dictionaries, Φ and TΦ .
Different dictionaries can impose drastically different computational burdens. The
nominal cost of storing and applying an arbitrary n-by-p matrix to a p-vector is a
constant times np. Certain dictionaries have fast implicit algorithms. The
parameters αΦ and sTΦ can be computed, for arbitrary vectors α ands , (a) without
ever storing the matrices Φ and TΦ and (b) using special properties of the matrices
to accelerate computations.
CHAPTER 3. ADVANCED FEATURE EXTRACTION TECHNIQUES
____________________________________________________________________
84
Wavelets give a dictionary with a fast implicit algorithm; if S8-symmlet is used, both
Φ and TΦ may be applied in O(n) time. For the stationary wavelet dictionary, O(n
log(n)) time is required. Cosine packets and wavelet packets also have fast implicit
algorithms. Here both Φ and TΦ can be applied in order O(n log(n)) time and order
O(n log(n)) space much better than the nominal ( )nnnp 22 log= one would expect
from naive use of the matrix definition.
3.3.2 Matching Pursuit
3.3.2.1 Concepts and Principles of Matching Pursuit
Notations
The space ( )RL2 [25] is the Hilbert space of complex valued functions such that
( ) +∞<= ∫+∞
∞−dttff
22 (3.34)
The inner product of ( ) ( )22, RLgf ∈ is defined by
( ) ( )dttgtfgf ∫+∞
∞−=, (3.35)
where ( )tg is the complex conjugate of ( )tg . The Fourier transform of ( )∈tf ( )RL2
is written ( )ωf and defined by
( ) ( ) dtetff ti
∫+∞
∞−
−= ωωˆ (3.36)
Matching Pursuit uses a specific criterion to search and decide the atoms and their
coefficients in the adaptive approximation. The Matching Pursuit (MP) initiates the
approximation with xR =)0( and builds up a sequence of sparse approximation
stepwise.
At the first step, the Matching Pursuit selects wavelet atoms 1γψ where
1, γψx is
maximum over the whole dictionary. Then the signal can begin with the
representation,
)1(
11, Rxx += γγ ψψ . (3.37)
CHAPTER 3. ADVANCED FEATURE EXTRACTION TECHNIQUES
____________________________________________________________________
85
At the stage k , the Matching Pursuit identifies the dictionary atom that best
correlates with the residual and then adds to the current approximation a scalar
multiple of the atom, so that ( ) ( ) )(11 , kkk RRRkk
+= −−γγ ψψ . That is,
)1( +kγψ is selected
in the dictionary if maximum over ( )k
kR γψ,1− .
After the k iterations, the signal can be represented as
( ) kk
i
i RRxii
+=∑=
−
1
1 , γγ ψψ , (3.38)
Where ( )i
iR γψ,1− is the coefficient of the atomiγψ . The decomposition is ceased
when certain residual requirement is fulfilled according to the application of the
Matching Pursuit.
3.3.2.2 Complexity
Each Matching Pursuit iteration requires ( )( )NNO 2log operations.
3.3.3 Basis Pursuit
3.3.3.1 Concepts and Principles of Basis Pursuit
Basis Pursuit [151] represents signals in over complete dictionaries by convex
optimization. It obtains the decomposition that minimizes the 1l norm of the
coefficients occurring in the representation.
1min α (3.39)
The powerful over complete dictionaries and optimization rules of Basis Pursuit, if
properly applied to analysing vibration signals of machines, can lead to effective
machinery fault diagnosis.
A vibration signal of an operating bearing ( )ntxx t <≤= 0: (viewed as a vector
in nR ), is a discrete-time signal. This signal is sampled with equal time intervals
(depending on its sampling frequency) and a sampling length n. It is important to
find the precise reconstruction of the signal using superposition of certain elementary
waveforms γφ . This is the main objective of Basis Pursuit algorithm and which
yields an adaptive time-frequency transform to extract vibration features. In
CHAPTER 3. ADVANCED FEATURE EXTRACTION TECHNIQUES
____________________________________________________________________
86
particular, using Basis Pursuit, impulsive component at defect frequencies can be
represented in the time-frequency domain with subtle resolution and sparsity.
According to the residual requirement of an adaptive approximation, an adaptive
approximation can be achieved after a certain number of iterations. Best matched
wavelet packet atoms are selected from the wave packet dictionary in these iterations
based on the optimization principle of Basis Pursuit.
Basis Pursuit can be used with noisy data by solving an optimization problem trading
off a quadratic misfit measure with a 1l norm of coefficients. Basis Pursuit can stably
suppress noise while preserving structure that is well expressed in the dictionary
under consideration.
Basis Pursuit is closely connected with linear programming. Recent advances in
large-scale linear programming associated with interior-point methods can be applied
to Basis Pursuit and can make it possible, with certain dictionaries, to nearly solve
the BP optimization problem in nearly linear time.
The optimization principle for Basis Pursuit leads to decompositions that can have
very different properties from the wavelet transform, DWPA, and Matching Pursuit.
In particular, they can be much sparser. Vibration analysis using Basis Pursuit can
effectively extract features, which are sparser than features DWPA and Matching
Pursuit can extract. As a result, unnecessary features can be invisible and disregarded
when conducting the interpretation of features. Furthermore, because Basis Pursuit is
based on convex optimization, it searches for solutions globally. A signal can be
stably super resolved, ie, the time and frequency of features can be sufficiently
localised.
3.3.3.2 Algorithms used in solving Basis Pursuit problems
Basis Pursuit is realised using an optimization principle. This optimization procedure
is a complicated linear problem, which can be successfully solved given the
significant amount of work done on the solution of linear programs in the past.
Algorithms for solving linear problems are the simplex and interior-point algorithms,
which have been applied to Basis Pursuit. These two algorithms in the application of
CHAPTER 3. ADVANCED FEATURE EXTRACTION TECHNIQUES
____________________________________________________________________
87
solving Basis Pursuit problems are called Basis Pursuit - simplex or Basis Pursuit –
interior, and are introduced as follows:
Basis Pursuit - simplex
In standard implementations of the simplex method for LP, one first finds an initial
basis B consisting of n linearly independent columns of A for which the
corresponding solution bB 1− is feasible (nonnegative). Then one iteratively improves
the current basis by swapping one term, at each step, on the basis for one term not in
the basis, using the swap that best improves the objective function. There always
exists a swap that improves or maintains the objective value, except at the optimal
solution. Moreover, LP researchers have shown how one can select terms to swap in
such a way as to guarantee convergence to an optimal solution (anticycling rules).
Hence the simplex algorithm is explicitly a process of BP: iterative improvement of a
basis until no improvement is possible, at which point the solution is achieved.
Translating this LP algorithm into BP terminology, one starts from any linearly
independent collection of n atoms from the dictionary. This is the current
decomposition. The current decomposition is iteratively improved by swapping
atoms in the current decomposition for new atoms, with the goal of improving the
objective function. By application of anticycling rules, there is a way to select swaps
that guarantees convergence to an optimal solution (assuming exact arithmetic).
BP-interior
The collection of feasible points { }0,: ≥= xbAxx is a convex polyhedron in mR (a
“simplex"). The simplex method, viewed geometrically, works by walking around
the boundary of this simplex, jumping from one vertex (extreme point) of the
polyhedron to an adjacent vertex at which the objective is better. Interior point
methods instead start from a point ( )0x well inside the interior of the simplex
( )( )00 >>x and go “through the interior" of the simplex. Since the solution of a LP is
always at an extreme point of the simplex, as the interior-point method converges,
the current iterate ( )kx approaches the boundary. One may abandon the basic interior
point iteration and invoke a “crossover" procedure that uses simplex iterations to find
the optimizing extreme point.
CHAPTER 3. ADVANCED FEATURE EXTRACTION TECHNIQUES
____________________________________________________________________
88
Translating this LP algorithm into Basis Pursuit terminology, one starts from a
solution to the overcomplete representation problem ( ) s=Φ 0α with ( )0α > 0. One
iteratively modifies the coefficients, maintaining feasibility ( ) sk =Φα , and applying
a transformation that effectively sparsifies the vector ( )kα . At some iteration, the
vector has n≤ significantly nonzero entries, and it “becomes clear” that those
correspond to the atoms appearing in the final solution. One forces all the other
coefficients to zero and “jumps” to the decomposition in terms of the n≤ selected
atoms. (More general interior-point algorithms start with ( ) 00 >α but don't require the
feasibility ( ) sk =Φα throughout; they achieve feasibility eventually.)
3.3.3.3 Denoising
For noisy data, Basis Pursuit can be adapted to assume the form:
zsy σ+= ; (3.40)
where ( iz ) is a standard white Gaussian noise, σ > 0 is a noise level, and s is the
clean signal. In this setting, s is unknown, while y is known. Usually an exact
decomposition of y is not required. Instead decompositions like (3-28) become
relevant.
Basis Pursuit denoising (BPDN) refers to the solution of
1
2
221
min αλαα
⋅+Φ−y (3.41)
The details of the denoising algorithm can be referred to [151]. Basis Pursuit
denoising of the vibration signals was conducted in this research to highlight the fault
features.
3.3.3.4 Factors affecting the application of Basis Pursuit
When applying Basis Pursuit to analyse vibration signals, problems such as lengthy
calculation times or converging trends can be experienced. These often lead to
insufficiency or failure of the Basis Pursuit analysis of the vibration signals. The
performance of Basis Pursuit analysis of the vibration signals is determined by the
size of the problem, the parameter selection, and the complexity of the signal. These
CHAPTER 3. ADVANCED FEATURE EXTRACTION TECHNIQUES
____________________________________________________________________
89
factors are vital for the success of the Basis Pursuit analysis of signals and are
explained as follows.
(1). The size of problem
The complexity of the Basis Pursuit analysis increases as the size of the problem
grows. The innermost computational step (a conjugate-gradient iteration) has a
complexity that scales with problem size like O(n) or O(n log(n)) depending on the
type of dictionary used. The choice of right dictionary is vital for the successful
application of Basis Pursuit analysis.
(2). Selection of parameters in the Basis Pursuit analysis
The complexity of the primal-dual logarithmic barrier interior-point implementation
depends on both the accuracy of the solution and the accuracy of the conjugate-
gradient solver. The accuracy of the solution is determined by the two parameters
FeaTol, PDGapTol [151] controlling the number of barrier iterations, and the
parameter CGAccuracy, which decides the accuracy of the conjugate-gradient solver
and consequently the number of conjugate-gradient iterations. To meet with the
accuracy required for accurate fault diagnosis, FeaTol, PDGapTol, and CGAccuracy
are set up at 210− for superresolution.
(3). The complexity of the signal
When the vibration signal is straightforward with only a few tonal components, Basis
Pursuit can achieve sparse representation with the algorithm converging quickly.
3.4 Summary
This chapter has presented the basic theory and comments on the suitability of the
time-frequency analysis techniques for fault diagnosis of rolling bearings. According
to the analysis of the above sections, symmlet 8 wavelet functions were chosen for
DWPA. A discrete wavelet packet dictionary with symmlet 8 functions was used in
Matching Pursuit and Basis Pursuit analysis. The wavelet packet dictionary was
chosen for its properties as follows:
(1) It is a time-frequency dictionary,
(2) It has wavelet functions,
CHAPTER 3. ADVANCED FEATURE EXTRACTION TECHNIQUES
____________________________________________________________________
90
(3) It is orthogonal,
(4) It is a dyadic dictionary.
A primal dual log barrier interior algorithm was chosen in the procedure of the
optimization of Basis Pursuit.
For highly non-stationary signals, the entropy minimization in DWPA produces a
mismatch between the “best” orthonormal basis and many local signal components.
On the contrary, the Matching Pursuit is a greedy algorithm that locally optimizes the
choice of the wavelet packet function, for each signal residue. Matching Pursuit
analysis can thus adapt to varying structures. This greedy strategy requires more
computations than the best basis DWPA, whose total complexity is O (N logN). The
globally optimization of Basis Pursuit requires more computational complexity but
yields good results for the decomposition of vibration signals.
Signals were decomposed with the same length and number of decomposition levels
using these three time-frequency analysis techniques in the application to diagnose
rolling element bearing faults.
CHAPTER 4. AUTOMATIC DIAGNOSIS SCHEMA
____________________________________________________________________
91
CHAPTER 4. AUTOMATIC DIAGNOSIS SCHEMA
4.1 Introduction
This chapter presents automatic fault diagnosis methods. Neural Networks were
applied to automate the interpretation of the feature extracted using Fourier spectrum
analysis, time-frequency spectrogram analysis, or advanced time-frequency analysis
techniques (improved DWPA, Matching Pursuit, and Basis Pursuit) as illustrated in
Chapter 3. Conventionally, human experts will distinguish features by inspection. In
spectrum analysis, certain specific frequency components can be discernible for
specialists to analyse the conditions of rolling bearings. In the time-frequency maps
of spectrogram, diagnostic personnel can detect the condition of a rolling bearing by
inspecting the spectrogram time-frequency map of the rolling bearing signal.
In DWPA, certain wavelet packets of bearing signals are filtered to signify the status
of the bearing. Although the above mentioned techniques can sometimes be used
with limited success in diagnosing rolling bearing faults, the above mentioned
features may not be easily distinguished by human experts. To overcome this
limitation of the above methods (reliance on experts), this chapter presents some
common rules for feature extraction. These feature patterns were incorporated with
Neural Network classifiers to form an automatic pattern recognition strategy.
Neural Networks were chosen as pattern classifiers for the reasons mentioned in
Chapter 2. Of the various types of Neural Networks, Feed Forward Neural Networks
are the most well developed and generally applied with success in bearing fault
classification.
In each section of this chapter, the automatic diagnostic schema was designed and
explained one by one with a flowchart drawn first followed by a detailed description.
The schemas presented follow the process of feature extraction and rolling bearing
conditions classification. The strategy for feature selection is elaborated first in each
section. The principle of Neural Networks is presented with specific details and two
architectures are designed for pattern recognition.
CHAPTER 4. AUTOMATIC DIAGNOSIS SCHEMA
____________________________________________________________________
92
4.2 Automatic Fault Diagnosis Using Spectrum Analysis
The automatic diagnosis procedure used in this study is shown in Figure 4.1, where
frequency features are extracted first. Frequency features were extracted using
Fourier spectra and the time frequency spectrogram separately. The spectrum or time
series signals of each frequency band after spectrogram were processed to reduce the
dimension of feature vectors. A Feed Forward Neural Network (FFNN) was
designed and trained on these features to classify bearing conditions.
Figure 4.1: Automatic fault diagnosis procedure using spectrum, spectrogram
with NN
A signal x(t) is a series of sampling digital values which represents the vibration of a
bearing to be diagnosed.
Fourier Spectrum (FFT)
The FFT is the most common technique to convert signals from its time domain into
the frequency domain.
( ) ( )∫+∞
∞−
−= dtetxF tjωω (4.1)
Fault Classification
Feature Extraction
Collected Data
Diagnosis results: ORF, IRF, REF, HB
Fourier spectrum or spectrogram
Averaging
FFNN
CHAPTER 4. AUTOMATIC DIAGNOSIS SCHEMA
____________________________________________________________________
93
FFT of the signals is further calculated by averaging the spectrum in equally divided
frequency intervals. Sixteen intervals were divided on the frequency range of each
data set and averaged to formulate a vector as the input to an FFNN.
Spectrogram
The spectrogram is defined as the square of STFT.
( ) ( ) ( ) ( )2
22,, ∫
∞+
∞−
−−== τττ τπ detwxftSTFTftSPEC fjxx (4.2)
The spectrogram is a arguably the most useful presentation of time frequency
information at this time [149]. In this study, the vibration signals were analysed into
16 frequency bands using the spectrogram. The time series in each band were
averaged and further formulated into a feature vector.
4.3 Automatic Fault Diagnosis Based on DWPA
The automatic diagnosis procedure adopted in this investigation is shown in Figure
4.2. Features were extracted from input signals and subsequently classified to assess
bearing condition. The feature extraction process utilized the DWPA as a pre-
processor. A variety of parameters such as Mean, Root Mean Square (RMS),
Variance, Energy value, Skewness, Kurtosis, Crest Factor, and Matched Filter were
derived from the DWPA based wavelet packets and further used as features. Feed
Forward Neural Networks (FFNN) were used to classify bearing conditions based on
features obtained. The training of the Neural Network was conducted by feeding in
features of a number of signals collected from bearings with known conditions
including Normal (N) (or Healthy Bearing (HB), Inner Race Fault (IRF), Outer Race
Fault (ORF), and Rolling Element Fault (REF).
Choosing wavelet functions
In the procedure outlined above, symlet8 functions were used in feature extraction
for reasons mentioned in the previous chapter. Figure 4.3 shows the symlet8 function
with certain scale and location where the left value in the brackets indicates the
decomposition level and the right value identifies the location of bands (packets) in
the decomposition level.
CHAPTER 4. AUTOMATIC DIAGNOSIS SCHEMA
____________________________________________________________________
94
Figure 4.2: Automatic fault diagnosis procedure using DWPA and NN
Best level decision
A signal can be decomposed into wavelet packets after DWPA. The selection of
wavelet packets is critical for the success of the Neural Network classifier. The best
basis algorithm cannot be used to find a unique tree for different data segments of a
signal, therefore not suitable for the requirement of a Neural Network classifier. To
find a unique way of selecting wavelet packets, it is recommended to just choose the
decomposition level and keep all the nodes up to that level. In this case, information
Reference signals (Signals from bearings with known faults)
BEARING CONDITION CLASSIFICATION
Test signals (Signals from a bearing to be diagnosed)
Diagnosis results: ORF, IRF, REF, HB
A Feed Forward Neural Network: Single output NN or multi output NN
Testing sets
Training sets
FEATURE EXTRACTION
DWPA DWPA
Signals of wavelet packets
Mean, or Variance, or Skewness, or Kurtosis, or Crest Factor, or Energy, or Matched Filter
Wavelet Selection
Tree Level Decision
Wavelet Selection
Tree Level Decision
Signals of wavelet packets
Mean, or Variance, or Skewness, or Kurtosis, or Crest Factor, or Energy, or Matched Filter
CHAPTER 4. AUTOMATIC DIAGNOSIS SCHEMA
____________________________________________________________________
95
will not be lossy and then good for finding a uniquely generating input feature
vectors for a Neural Network classifier.
A high decomposition level leads to signals filtered in narrow frequency bands and
accurate frequency localisation of high energy vibrations. However it requires
significant computation time and thus cannot fulfil the requirement for efficient
feature extraction calculation in an automatic diagnostic procedure. There is trade off
between the precision of filter bands and computation time. It is important to select
an appropriate decomposition level to analyse signals in accurate frequency bands
while conducting the DWPA in reasonable time.
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 10
2
4
6
8
10
12
14
16Some S8 Symmlets at various scales and locations
(4, 3)
(4, 8)
(4,11)
(6,12)
(6,26)
(6,34)
(6,42)
(7,51)
(7,77)
(7,101)
(8,31)
(8,81)
(8,102)
(8,166)
(8,202)
Figure 4.3: Symlet8 function with a few scales and locations
Features based on wavelet packets
Parameters are further calculated based on the wavelet packets after the DWPA of
signals. This procedure can dramatically reduce the dimension of the inputs to the
Feed Forward Neural Network (FFNN) classifier. Identifying a reasonable dimension
is critical for the performance of the automatic interpretation of a FFNN outputs and
can avoid the curse of dimension.
CHAPTER 4. AUTOMATIC DIAGNOSIS SCHEMA
____________________________________________________________________
96
4.4 Automatic Fault Diagnosis Using Pursuit
The Pursuit based automatic diagnosis procedure is shown in Figure 4.4. The feature
extraction utilizes Matching Pursuit or Basis Pursuit by choosing dictionary and
decomposition iteration decisions. Feed Forward Neural Networks with multi outputs
were used to classify bearing conditions based on the features.
Figure 4.4: Automatic fault diagnosis procedure using Basis Pursuit (or
Matching Pursuit) and NN
Dictionary selection
In this feature extraction, Discrete Wavelet Packet Dictionary with symlet8 functions
were adopted.
Bearing Condition Classification
Reference signals (Signals from bearings with known faults)
Basis Pursuit Or Matching Pursuit analysis
Dictionary Selection
Decomposition Level Decision
Feature selection Feature selection
Test signals (Signals from a bearing to be diagnosed)
Diagnosis results: ORF, IRF, REF, HB
A Neural Network
Testing sets Training sets
Basis Pursuit Or Matching Pursuit analysis
Dictionary Selection
Decomposition Level Decision
CHAPTER 4. AUTOMATIC DIAGNOSIS SCHEMA
____________________________________________________________________
97
Iteration decision
The number of iterations in the calculation of Basis Pursuit (Matching Pursuit) needs
to be decided according to the residual requirement and the calculation efficiency
requirements. In these analyses, the decomposition was conducted from level 1, to 4.
Feature selection
The feature pattern is further formed from the Basis Pursuit coefficients of signals by
selecting high energy components, or a certain number of maximum values of the
coefficients. In this application, the number of selected values for the feature pattern
is the power of two, including 16, 32, 64, and 128. A value under 16 is considered
insufficient to represent a signal. A value over 128 is considered burdensome for
calculation.
4.5 Design of the Feed Forward Neural Network classifier
4.5.1 Feed Forward Neural Networks
Of the variety of activation functions and learning rules, the ones used in this thesis
are presented in the following sections. The back propagation rule will be also
presented.
Activation Function
Most neurons in Neural Networks convert their inputs using a scalar-to-scalar
function called an activation function, yielding a neuron output. The most commonly
used activation functions are linear functions, threshold functions, sigmoid functions,
and bipolar sigmoid functions (refer to APPENDIX). Among these activation
functions, the sigmoid function was selected in the Neural Network classifiers used
in this study.
Sigmoid function (Figure 4.5)
xexg −+
=1
1)( (4.3)
Sigmoid function is suitable for this application due to two reasons:
CHAPTER 4. AUTOMATIC DIAGNOSIS SCHEMA
____________________________________________________________________
98
(1) This function has particular advantages for use in back propagation Neural
Networks because it is easy to differentiate, and thus can dramatically reduce the
computational burden for training.
(2) It is suitable for applications whose desired output values are between 0 and 1.
This makes it suitable for pattern recognition applications, and in this case,
particularly for bearing fault classification.
Figure 4.5: Sigmoid function [152]
(3) Activation functions should be chosen to suit the distribution of target values for
output neurons. The sigmoid function is well suited for target values that are binary
[0,1]. The sigmoid functions have additional advantages for continuous-valued
targets with a bounded range, provided that either the outputs or the targets is scaled
to the range of the output activation function.
Back propagation was devised by generalizing the Widrow-Hoff learning rule to
multiple-layer networks and nonlinear differentiable transfer functions. Input vectors
and the corresponding target vectors are used to train a network until it can
approximate a function, associate input vectors with specific output vectors, or
classify input vectors in an appropriate way. Networks with biases, a sigmoid layer,
and a linear output layer are capable of approximating any function with a finite
number of discontinuities.
Back propagation is a gradient descent algorithm, as is the Widrow-Hoff learning
rule, in which the network weights are moved along the negative of the gradient of
the performance function. The term backpropagation refers to the manner in which
the gradient is computed for nonlinear multilayer networks.
CHAPTER 4. AUTOMATIC DIAGNOSIS SCHEMA
____________________________________________________________________
99
A few aspects need to be considered when designing the structure of Neural
Network classifiers:
(1) Number of neural nodes
A larger number of neurons enable a Neural Network to approximate functions with
more complexity. The shortcoming, however, is that while more neurons (and,
consequently, more parameters) can fit the training data better, the system will likely
provide greater degrees of freedom in between the training points. This might cause
the test data to fit the function desired poorly. Very few parameters relative to the
training data distribution and even the training set would not be adequately matched.
The gradient descent algorithm sometimes may have a tendency to get trapped in
local minima of the error surface.
(2) Layers of Neural Networks
Single-layer networks are sometimes more capable of representing only linearly
separable functions or linearly separable decision domains. Two hidden layered
networks can represent an arbitrary decision boundary to arbitrary accuracy and
could approximate any smooth mapping to any accuracy with sigmoid activation
functions.
Many hidden layers lead to a dramatic increase in the number of local minima.
Gradient-based optimization algorithms often only find local minima, and sometimes
may miss global minima. Even if the training algorithm can find the global minima,
it is highly possible to be stuck in a local minimum after many time-consuming
iterations. The training will then have to be ceased and started over.
In general, one hidden layer is the first choice for any practical feed-forward network
design. A Neural Network using a single hidden layer with a large number of hidden
neurons does not perform well. A second hidden layer with fewer processing neurons
need to then be considered.
Considering the limitation and strength of different numbers of hidden layers in
Neural Networks, only the single and double hidden layered networks were designed
and tested in this study.
CHAPTER 4. AUTOMATIC DIAGNOSIS SCHEMA
____________________________________________________________________
100
4.5.2 Design of the Structure of the Feed Forward Neural Network Classifiers
In the automatic diagnosis procedure, the features of signals extracted using the time
frequency analysis methods are fed into a FFNN to train the network and classify the
condition of the bearing to test.
In this research, two FFNNs were designed for the application of the classification of
bearing conditions. These two FFNNs have different architectures, with different
number of hidden layers and different number of output nodes. The first NN (as
shown in Figure 4.6) has eight input nodes, connected to one hidden layer with ten
nodes. The hidden layer was subsequently input to the one output node. The decision
space of the output node is divided into four intervals, which are related to diagnostic
results respectively.
The second FFNN (as shown in Figure 4.7) has eight input nodes, two hidden layers
with ten nodes and four output nodes. The input layer includes eight nodes, which
are connected to two hidden layers. The output layer of the NN comprises of four
nodes, which represent the classes of the rolling bearing conditions: Normal, IRF,
ORF, and REF respectively.
Figure 4.6: A single output FFNN
Input layer Hidden
layer Output layer
ORF
IRF
REF
Normal
CHAPTER 4. AUTOMATIC DIAGNOSIS SCHEMA
____________________________________________________________________
101
Figure 4.7: A multi output FFNN
The Neural Networks are trained using a back propagation algorithm. The training
can cease according to the criteria of either Mean Square Error (MSE) reach to
certain value or that the epoch of training reaches certain value. In our application, a
target Mean Square Error of 510− and a maximum iteration number (epoch) of 300 is
setup. The training process would stop if any of these conditions were met. The
initial weights and biases of the network were generated automatically by the
program. During our training processes, generally the epoch value of 300 is the one
reached first. The Mean Square Error (MSE) at this time is used as a criterion for
appraising the training performance of the Neural Network and the classification rate
as the criterion for appraising each diagnosis procedure.
4.6 Summary
This chapter presented automatic fault diagnosis procedures. The feature extraction
techniques employed ranged from conventional Fourier Spectrum analysis, time-
frequency spectrogram, improved DWPA, improved Matching Pursuit, to a first
application of the Basis Pursuit. The feature pattern based on spectrum analysis was
formed by selecting a certain number of averaged values of the spectrum. The feature
pattern from the spectrogram was formed by averaging signals in different frequency
bands. The DWPA pattern consisted of Mean, Variance, Skewness, Kurtosis, RMS,
Input layer Hidden
layer Hidden layer Output
layer
ORF
IRF
REF
Normal
CHAPTER 4. AUTOMATIC DIAGNOSIS SCHEMA
____________________________________________________________________
102
Energy Value, Crest Factor, and Matched Filter derived from the wavelet packets.
The feature pattern for the Matching Pursuit was formed by selecting maximum
value of Matching Pursuit coefficients while that of the Basis Pursuit was formed by
selecting maximum Basis Pursuit coefficients.
Details of Feed Forward Neural Networks were also introduced. At the same time, a
single hidden layer single output and double hidden layers multi-output Neural
Networks were designed for diagnosing faults in rolling bearings.
CHAPTER 5. SIMULATION AND EXPERIMENT
____________________________________________________________________
103
CHAPTER 5. SIMULATION AND EXPERIMENT
5.1 Introduction
This chapter presents vibration signals simulated using mathematical models.
Impulsive signals, which have similar waveforms as vibration waveforms, were
simulated to test the performance of the studied techniques. In the test, the analytical
results of these simulated signals by the proposed methods were used in comparison
with known theoretical characteristics in the designed model. The comparison shows
the performance of the proposed feature extraction techniques.
Experimental work was conducted to collect vibration signals to test the performance
of the proposed methods for experimental diagnostic tasks.
The framework for preparing test data from rolling bearings is briefly presented in
this chapter and includes test rigs and electronic instruments.
Various fault severities were artificially introduced. The dimensions of these bearing
faults are presented in the third section, where the characteristic frequencies of the
bearings are also provided.
The fifth section analyses failure development of different severity of faults in rolling
bearings.
5.2 Simulated Signals
Two signals were simulated as presented in Equations 5.1 and 5.2. The simulated
signals are impulsive signals with digitisation at 5000 Hz. These two simulated
signals consist of four sinusoidal components with exponential amplitude.
( ) ( )[ ] ( ) ( )[ ] ( )( )[ ] ( ) ( )[ ] ( )7.45000cos3200800exp75000cos2700600exp
4.55000cos3000400exp65000cos2400200exp22
221
tttt
ttttty
ππππ
−−+−−+
−−+−−=,
(5.1)
( ) ( )[ ] ( ) ( )[ ] ( )( )[ ] ( ) ( )[ ] ( )165000cos60800exp145000cos90600exp
185000cos70400exp155000cos80200exp22
222
tttt
ttttty
ππππ
−−+−−+
−−+−−=
(5.2)
CHAPTER 5. SIMULATION AND EXPERIMENT
____________________________________________________________________
104
The former has outliners similar to a sinusoid signal while the latter has outliners
which appear more impulsive. These two signals can successfully present vibration
signals of faulty bearings. In these models, the amplitude and frequency values of the
main components are clearly distinguished. The first model has primary components
with frequencies: 833 Hz, 463 Hz, 357 Hz, and 531 Hz. These frequency components
have time periods: 0.0012 s, 0.0233 s, 0.0028 s, and 0.00188 s. The second model
has primary components with frequencies: 167 Hz, 139 Hz, 178.5 Hz, and 156 Hz.
And associated periods: 0.006 s, 0.0072 s, 0.0056 s, and 0.0064 s.
Gaussian noise with Signal to noise ratio (SNR) of 15 dB were added to these two
simulated signals. These signals with noise more readily simulated actual vibration
signals from bearings, which are often contaminated by environmental noise or other
extraneous vibration sources.
5.3 Test Rig and Experiment Procedure
The experimental procedure adopted in the thesis work program is presented in
Figure 5.1. Accelerometers were attached to the test rigs. Signals were then
transmitted to an amplifier and low pass filtered.
Figure 5.1: Experimental apparatus
Experiments were conducted using two bearing test rigs. The first one comprised an
AC motor and a shaft supported by two rolling element bearings one of which was
the test bearing (see Figure 5.2). Load was applied using a V-belt configuration.
Acceleration transducer
Low pass filter
Amplifier
Data Data recorder A/D converter
Test rig
CHAPTER 5. SIMULATION AND EXPERIMENT
____________________________________________________________________
105
Figure 5.2: Test rig with V-belt load
The other test rig had an AC motor and DC motor connected via a geared shaft
which was supported by bearings (as shown in Figure 5.3 ).
Figure 5.3: Test rig without load
For the first test rig, an accelerometer was used to measure the vibrations and was
located on the plummer block of the faulty bearing. The accelerometer, ENDEVCO
Model 256HX, is a lightweight low impedance (constant current) piezoelectric
accelerometer with integral electronics, designed specifically for vibration
measurements on small structures. Its frequency response lay in the region, 15Hz to
20 kHz. Signals were amplified using a PCB Model 482A20 amplifier and low
passed using a Krohn-hite model 3202 filter for both noise reduction and anti-
aliasing.
CHAPTER 5. SIMULATION AND EXPERIMENT
____________________________________________________________________
106
A DAQP-308 16-bit A/D PCMCIA data acquisition card was used to convert
analogue signals to digital signals and which was controlled using the DaqEZ
professional data acquisition software.
For the other test rig, the whole experimental procedure is shown in Figure 5.4. An
NI BNC 2120 A/D card was used in this experiment and the Labview software was
used to record the signals.
Figure 5.4: Experimental apparatus
5.4 Fault Simulation
Faults were simulated by using a wire cutting technique to induce a “crack” in the
races of the two SKF 6204 bearings to be tested. Figure 5.5 illustrates the simulated
faults, one crack in the inner race of a ball bearing and one crack in the outer race of
the second ball bearing. The inner race crack was 0.593 mm wide and 0.858 mm
deep. The outer race crack was 0.660 mm wide and 2.098 mm deep. The ball passing
inner race frequency (IRF) was calculated to be approximately five times the shaft
rotational frequency. The ball passing Outer Race Frequency (ORF) is approximately
three times the shaft rotational frequency. These bearings were embedded in the first
test rig and used to collect data using DAQ acquisition system.
CHAPTER 5. SIMULATION AND EXPERIMENT
____________________________________________________________________
107
Figure 5.5: Fault simulation of SKF 6205
A series of bearing faults were also simulated in the second test rig. KOYO 6201 RS
bearings were used (shown as Figure 5.6). Laser processing was used to create cracks
with depth, 0.5 mm in inner races and outer races, and a spot in the ball of the
bearing. The width of the outer race and inner race cracks and the diameters of the
rolling element spots are specified in Table 5.1. The dimensions and characteristic
frequencies of these bearings are shown in Table 5.2. The dimensions and
characteristic of the SKF 6204 bearings are listed in Table 5.3.
Figure 5.6: Fault simulation of KOYO 6201 RS
CHAPTER 5. SIMULATION AND EXPERIMENT
____________________________________________________________________
108
Table 5.1 Fault specification for the bearings KOYO 6201 RS.
Fault location
Outer race (width)
Inner race (width)
Ball (diameter)
type Crack Spot 0.1 0.1 0.1 0.2 0.2 0.2
Fault size (mm)
0.5 0.5 0.5
Table 5.2 Drive end bearing KOYO 6201 (Size in mm)
Inside Diameter
Outside Diameter
Thickness Ball Diameter Pitch Diameter
12 32 10 6 22
The number of balls: 7.
Defect frequencies: (multiple of running speed in Hz)
Inner Ring Outer Ring Cage Train Rolling Element 4.45 2.55 0.36 3.39
Table 5.3 Drive end bearing: 6204 SKF, deep groove ball bearing
Defect frequencies: (multiple of running speed in Hz)
Inner Race Outer Race Cage Train Rolling Element 4.95 3.05 0.38 3.98 Ball Diameter (mm) Pitch Diameter (mm) 7.938 33.5
Data with marginal faults were also obtained from the Case Western Reserve
University (CWRU) website. Their bearing faults were simulated using Electric
Discharge Machining (EDM). The CWRU signals were collected from SKF 6205
bearings (Table 5.4 and Table 5.5) with different severities of inner race and outer
race faults. The ball passing inner race and outer race frequencies for these bearings
are 5.4 and 4.7 times shaft rotational frequency respectively.
CHAPTER 5. SIMULATION AND EXPERIMENT
____________________________________________________________________
109
Table 5.4 Drive end bearing: 6205-2RS JEM SKF, deep groove ball bearing
(Size in inches)
Inside Diameter
Outside Diameter
Thickness Ball Diameter Pitch Diameter
0.9843 2.0472 0.5906 0.3126 1.537
Defect frequencies: (multiple of running speed in Hz)
Inner Ring Outer Ring Cage Train Rolling Element 5.4152 3.5848 0.39828 4.7135
Table 5.5 Fault Specifications for 6204-2RS JEM SKF (All dimension in inches)
Bearing Fault Location Diameter Diameter Diameter Drive End Inner Race .007 .014 .021 Drive End Outer Race .007 .014 .021 Drive End Ball .007 .014 .021
Defect period: (s;1/Hz)
Inner Race Outer Race Cage Train Rolling Element 0.005624 0.008496 0.07647 0.006462
5.5 Summary
The structure of the test rigs and the details of instruments used are described in this
chapter. Bearing faults are simulated and measurements made to identify the fault
characteristics of the bearings. The simulated signals were designed with impulses
similar to vibration waveforms generated by bearing defects. Noise was also added to
the simulated signals to further improve the simulation.
Data collected from these experiments was used in rigorous testing of the proposed
methods in this project such as the feature extraction performance of the proposed
techniques such as Discrete Wavelet Packet Analysis (DWPA), Matching Pursuit,
and Basis Pursuit. Furthermore, a significant amount of data collected from this
experimental phase was used in the training and testing of the automatic diagnostic
techniques.
CHAPTER 6. RESULTS AND DISCUSSION
____________________________________________________________________
110
CHAPTER 6. RESULTS AND DISCUSSION
6.1 Introduction
To evaluate the performance of the proposed techniques in practice, the following
aspects need to be considered:
• How can these methods be applied to real diagnostic tasks?
• What would the time-frequency maps look like?
• How would one interpret the time-frequency maps?
• How has the automatic procedures performed?
The time-frequency analysis methods (including the improved DWPA, Matching
Pursuit, and the Basis Pursuit presented in the third chapter) and the automatic
diagnosis using these three methods (presented in the fourth chapter) were applied to
analyzing the data collected in the experiments.
A mixture of software codes were utilised in generating the results in this thesis. The
Matlab Wavelet toolbox, neural network toolbox, and the Matlab code for Basis
Pursuit [151] were utilized to extract features from the vibration signals and for
classification of the bearing condition
Section 6.2 presents the analysis of simulated data and experimental data using these
three time-frequency analysis techniques. The features extracted from the simulated
signals using the time-frequency analysis techniques were compared with the known
characteristics in the simulation to check the accuracy of the time-frequency analysis
techniques. The experimental data were analysed to extract defect related time-
frequency features, which were interpreted as bearing faults.
Section 6.3 presents the result of the application of the automatic diagnostic
techniques with emphasis on the preparation of feature inputs to the Neural
Networks. These features were derived from spectra, spectrograms, DWPA,
Matching Pursuit, and Basis Pursuit. Each feature is embedded into a diagnostic
procedure. The Neural Networks were then implemented by being trained and tested
on real data from faulty rolling element bearings.
CHAPTER 6. RESULTS AND DISCUSSION
____________________________________________________________________
111
6.2 Analysis using DWPA, Matching Pursuit, and Basis Pursuit
Simulated signals and data from the experiments were used to evaluate the Basis
Pursuit technique by comparisons with the best basis discrete wavelet packet analysis
(DWPA) and the Matching Pursuit particularly with regard to their performance on
time-frequency feature extraction. Basis Pursuit was also applied to detect different
fault severities. Moreover, features, which were extracted from signals of faulty
bearings using Basis Pursuit, were reconstructed into a time series with clear defect
impacts and little noise contamination. The results are presented in the following
sections.
6.2.1 Time-Frequency Analysis of Simulated Signals
DWPA, Matching Pursuit, and Basis Pursuit were tested using simulated signals as
shown in Figures 6.1- 6.4. In Figure 6.2, it can be seen that there are mainly four
frequency components in the analysed signal, which have time periods around
0.0012 s, 0.0233 s, 0.0028 s, and 0.00188 s. These detected frequency components
exactly match the frequency components of the theoretical signal. Figure 6.3 and
Figure 6.4 can also generally be used to determine the frequency components.
However, as seen in the figures, when decomposed into same level, the Matching
Pursuit and the DWPA roughly presented the time-frequency components. The
resolution of the time-frequency components from the Basis Pursuit analysis is better
than those from the DWPA or Matching Pursuit analysis.
CHAPTER 6. RESULTS AND DISCUSSION
____________________________________________________________________
112
0 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08-3
-2
-1
0
1
2
3
Time(s)
Am
plitu
de
Figure 6.1: Simulated impulse signal 1y
Time(s)
Fre
quen
cy(H
z)
0 0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16 0.18 0.20
500
1000
1500
2000
2500
Figure 6.2: Basis Pursuit TF plane of 1y
CHAPTER 6. RESULTS AND DISCUSSION
____________________________________________________________________
113
Time(s)
Fre
quen
cy(H
z)
0 0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16 0.18 0.20
500
1000
1500
2000
2500
Figure 6.3: Best basis DWPA TF plane of 1y
Time(s)
Fre
quen
cy(H
z)
0 0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16 0.18 0.20
500
1000
1500
2000
2500
Figure 6.4: Matching Pursuit TF plane of 1y
CHAPTER 6. RESULTS AND DISCUSSION
____________________________________________________________________
114
To make the simulation more real, Gaussian noise were artificially added into the
pure signal (as shown in Figure 6.5). From Figure 6.6, it can be seen that the noise
influence has been diminished significantly and the frequency components are
distinguished clearly in the time-frequency plane. The Basis Pursuit TF plane
provides very good resolution as well as sparsity for the detection of the time-
frequency components. The Basis Pursuit TF plane is still a clean map with clearly
distinguishable four main time-frequency components. In the Basis Pursuit TF plane,
the noise diminishes dramatically. In the DWPA (Figure 6.7) and Matching Pursuit
TF planes (Figure 6.8), redundant time-frequency components appear and affect the
precise interpretation of the TF planes. The DWPA TF and Matching Pursuit TF
planes shows several frequency components from noise. Basis Pursuit also is able to
remove most of the noise that was introduced in the study. Besides, it also attained
the best resolution of the techniques investigated.
0 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08-3
-2
-1
0
1
2
3
Time(s)
Am
plitu
de
Figure 6.5: Simulated impulse signal 1y with noise
CHAPTER 6. RESULTS AND DISCUSSION
____________________________________________________________________
115
Time(s)
Fre
quen
cy(H
z)
0 0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16 0.18 0.20
500
1000
1500
2000
2500
Figure 6.6: Basis Pursuit TF plane of 1y with noise
Time(s)
Fre
quen
cy(H
z)
0 0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16 0.18 0.20
500
1000
1500
2000
2500
Figure 6.7: DWPA plane of 1y with noise
CHAPTER 6. RESULTS AND DISCUSSION
____________________________________________________________________
116
Time(s)
Fre
quen
cy(H
z)
0 0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16 0.18 0.20
500
1000
1500
2000
2500
Figure 6.8: Matching Pursuit Plane of 1y with noise
High amplitude impulsive vibrations were also simulated as shown in Figures 6.9-12.
These impulses are meant to be similar to the bearing defects signals. In the time-
frequency planes of Basis Pursuit analysis of this signal, significant components
clearly lead to the interpretation of four main frequencies. The four main frequency
components have time periods around 0.006 s, 0.0072 s, 0.0056 s, and 0.0064 s,
which match the theoretical components of the simulated signal. The time periods of
four main components were obtained by measuring the time intervals of
neighbouring components. As seen from these figures, the time-frequency
components show up better in the Basis Pursuit plane than in the DWPA and
Matching Pursuit planes.
CHAPTER 6. RESULTS AND DISCUSSION
____________________________________________________________________
117
0 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08-3
-2
-1
0
1
2
3
Time(s)
Am
plitu
de
Figure 6.9. Simulated impulse signal 2y
Time(s)
Fre
quen
cy(H
z)
0 0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16 0.18 0.20
500
1000
1500
2000
2500
Figure 6.10: Basis Pursuit TF plane of 2y
CHAPTER 6. RESULTS AND DISCUSSION
____________________________________________________________________
118
Time(s)
Fre
quen
cy(H
z)
0 0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16 0.18 0.20
500
1000
1500
2000
2500
Figure 6.11: DWPA TF plane of 2y
Time(s)
Fre
quen
cy(H
z)
0 0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16 0.18 0.20
500
1000
1500
2000
2500
Figure 6.12: Matching Pursuit TF plane of 2y
CHAPTER 6. RESULTS AND DISCUSSION
____________________________________________________________________
119
The simulated impulsive signal was also corrupted with Gaussian noise as seen in
Figure 6.13. Further analysis was conducted using the proposed DWPA, Matching
Pursuit, and Basis Pursuit. Figure 6.14, Figure 6.15, and Figure 6.16 present the TF
planes of Basis Pursuit, DWPA, and Matching Pursuit analysis of this signal
respectively. In Figure 6.15 and Figure 6.16, the predominant time-frequency
components are shown but are mixed in with noise elements in both the DWPA and
Matching Pursuit results. In the Matching Pursuit case, it is difficult to distinguish
pertinent time-frequency components from unrelated components due to the
similarity of the amplitude (shown in grey) of the both types of components.
However, the Basis Pursuit plane can diminish the colour of the unrelated
components and clear show impulsive related components. Basis Pursuit is so far
proving to be the most accurate of these techniques.
0 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08-3
-2
-1
0
1
2
3
Time(s)
Am
plitu
de
Figure 6.13: Simulated impulse signal 2y with noise
CHAPTER 6. RESULTS AND DISCUSSION
____________________________________________________________________
120
Time(s)
Fre
quen
cy(H
z)
0 0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16 0.18 0.20
500
1000
1500
2000
2500
Figure 6.14: Basis Pursuit TF Plane of signal 2y with noise
Time(s)
Fre
quen
cy(H
z)
0 0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16 0.18 0.20
500
1000
1500
2000
2500
Figure 6.15: DWPA TF plane of signal 2y with noise
CHAPTER 6. RESULTS AND DISCUSSION
____________________________________________________________________
121
Time(s)
Fre
quen
cy(H
z)
0 0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16 0.18 0.20
500
1000
1500
2000
2500
Figure 6.16: Matching Pursuit TF plane of signal 2y with noise
6.2.2 Time-Frequency Analysis of the Signals of Bearings with Marginal Faults
Figures 6.17- 6.20 are the time waveform of a normal bearing, its Basis Pursuit
plane, best basis DWPA plane, and Matching Pursuit plane. The time-frequency
planes of DWPA and Basis Pursuit analysis shows one predominant component at
about 1000Hz, presumably caused by a resonance. Figures 6.21- 6.24 are the time
waveform of a bearing with inner race fault, its Basis Pursuit, best basis DWPA, and
Matching Pursuit planes. The DWPA and Matching Pursuit planes do indicate that a
fault is present but the data is imprecise compared to the Basis Pursuit plane. Figures
6.25- 6.28 are the time waveform of a bearing with ORF, its Basis Pursuit plane, best
basis DWPA plane, and Matching Pursuit plane.
CHAPTER 6. RESULTS AND DISCUSSION
____________________________________________________________________
122
0 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08-3
-2
-1
0
1
2
3
Time(s)
Am
plitu
de
Figure 6.17: Vibration signal of a bearing with normal condition
Time(s)
Fre
quen
cy(H
z)
0 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.080
1000
2000
3000
4000
5000
6000
Figure 6.18: Basis Pursuit TF plane
(m
/s2 )
CHAPTER 6. RESULTS AND DISCUSSION
____________________________________________________________________
123
Time(s)
Fre
quen
cy(H
z)
0 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.080
1000
2000
3000
4000
5000
6000
Figure 6.19: Best basis DWPA TF plane
Time(s)
Fre
quen
cy(H
z)
0 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.080
1000
2000
3000
4000
5000
6000
Figure 6.20: Matching Pursuit TF Plane
CHAPTER 6. RESULTS AND DISCUSSION
____________________________________________________________________
124
0 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08-3
-2
-1
0
1
2
3
Time(s)
Am
plitu
de
Figure 6.21: Vibration signal of a bearing with IRF
Time(s)
Fre
quen
cy(H
z)
0 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.080
1000
2000
3000
4000
5000
6000
Figure 6.22: Basis Pursuit TF plane
(m
/s2 )
E F
A B C D
CHAPTER 6. RESULTS AND DISCUSSION
____________________________________________________________________
125
Time(s)
Fre
quen
cy(H
z)
0 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.080
1000
2000
3000
4000
5000
6000
Figure 6.23: Best basis DWPA TF plane
Time(s)
Fre
quen
cy(H
z)
0 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.080
1000
2000
3000
4000
5000
6000
Figure 6.24: Matching Pursuit TF Plane
CHAPTER 6. RESULTS AND DISCUSSION
____________________________________________________________________
126
0 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08-3
-2
-1
0
1
2
3
Time(s)
Am
plitu
de
Figure 6.25: Vibration signal of a bearing with ORF
Time(s)
Fre
quen
cy(H
z)
0 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.080
1000
2000
3000
4000
5000
6000
Figure 6.26: Basis Pursuit TF plane
(m
/s2 )
A B C D
E F
CHAPTER 6. RESULTS AND DISCUSSION
____________________________________________________________________
127
Time(s)
Fre
quen
cy(H
z)
0 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.080
1000
2000
3000
4000
5000
6000
Figure 6.27: Best basis DWPA TF plane
Time(s)
Fre
quen
cy(H
z)
0 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.080
1000
2000
3000
4000
5000
6000
Figure 6.28: Matching Pursuit TF plane
CHAPTER 6. RESULTS AND DISCUSSION
____________________________________________________________________
128
6.2.3 Basis Pursuit and Best Basis DWPA and Matching Pursuit – a
Comparison
A comparison was made between the application of these three techniques to
diagnosing bearing faults regarding their feature extraction performance, sparsity,
superresolution, and computational complexity.
6.2.3.1 Feature extraction performance of Basis Pursuit, Matching Pursuit, and best
basis DWPA
Generally, marginal faults in rolling bearings can go undetected using vibration
techniques such as best basis DWPA and Matching Pursuit. However, in this study
the Basis Pursuit technique proved to be effective in detecting marginal faults. The
application of DWPA, Matching Pursuit, and Basis Pursuit is shown in Figures 6.17-
6.28. The signals were obtained from a normal bearing, a bearing with an inner race
fault, and a bearing with an outer race fault respectively. The darkness indicates the
amplitude of the coefficients of wavelet atoms in the Time-frequency (TF) planes. It
is noted that the results generated using the best basis DWPA and Matching Pursuit
contained a large number of features (wavelet atoms with high amplitude), which
were extracted with relatively rough resolution and often irrelevant to the bearing
faults. These extracted features may cause confusion in a routine diagnostic survey.
The DWPA is mostly considered as a multi-band pass filter. Certain bands of interest
can be selected from the TF plane of DWPA. Matching Pursuit provides more
discernible features in the TF plane but the features are not separated as identifiable
as features extracted from the Basis Pursuit. The features in the TF plane of Basis
Pursuit appear with certain frequencies that related to the bearing fault. Figures 6.17-
6.20 also illustrates the analysis of the original vibration signal of a normal bearing
including the DWPA TF plane, the Matching Pursuit TF plane, and the Basis Pursuit
TF plane. It again shows that Basis Pursuit TF plane provides more effective
features, which can be interpreted more easily to diagnostic faults, when compared
with the DWPA TF and the Matching Pursuit TF planes.
In Figures 6.21- 6.24, these three methods are used to diagnose a marginal inner race
fault in a rolling bearing. The IRF of the bearing to be diagnosed is 162Hz. In the
Basis Pursuit TF plane, the time differences between two neighbouring significant
CHAPTER 6. RESULTS AND DISCUSSION
____________________________________________________________________
129
atoms such as A, B, C, and D (marked) are approximately 0.0062s. This result
corresponds with the IRF defect frequency. It can also be recognized that the
difference between the two frequency components such as E and F matches the IRF,
162 Hz. Therefore, using Basis Pursuit, features of the IRF can be effectively
extracted and IRF of the bearing can be clearly identified. The outer race fault
detection was more straightforward as seen in Figures 6.25- 6.28, where the DWPA
TF and Matching Pursuit TF planes show promise in displaying the features.
However, Basis Pursuit provides the best presentation of features among these three
time-frequency methods. The atoms such as A, B, C, and D are distributed with time
intervals around 0.0093s, which matches the ORF, 107 Hz. It can also be clearly seen
that the difference between two frequency components matches the ORF, 107 Hz.
6.2.3.2 Sparsity
The objective of finding an effective feature extraction technique is to obtain the
sparsest and most representative features possible. Basis Pursuit provides the fewest
significant atoms and has the best sparsity. Matching Pursuit and DWPA produce
more irrelevant atoms for the decomposition of vibration signals from rolling
bearings.
6.2.3.3 Computational complexity
DWPA is the fastest algorithm among these three algorithms. Matching Pursuit is
faster than Basis Pursuit but slower than DWPA. The Basis Pursuit method
computation time requires around five times the Matching Pursuit method to analyse
the same length of data. This disadvantage of Basis Pursuit is offset by the short
length of data needed in its application.
6.2.3.4 Superresolution
Matching Pursuit and Basis Pursuit perform better than best basis DWPA with better
resolution. The improvement in performance of Matching Pursuit and Basis Pursuit
in diagnosing bearing faults is due to employing a wavelet packet dictionary. The
usage of a wavelet packet dictionary brings in more wavelet atoms. There are up to
)(log)1( 2 NN + atoms to be chosen for decomposition of a signal where N is the length
CHAPTER 6. RESULTS AND DISCUSSION
____________________________________________________________________
130
of the signal. The optimisation rule in Basis Pursuit makes the Basis Pursuit analysis
more accurate.
6.2.4 Severity of Bearing Faults Analysed Using Basis Pursuit
The Basis Pursuit technique was evaluated using varying severity of faults in our test
bearings. The effect that fault severity has on the time-frequency characteristic of
faulty bearings vibration is presented in Figures 6.29- 6.33. The signals in Figures
6.29- 6.32 were collected from bearings with different severity of Electro-Discharge
Machining (EDM) inner race faults (shown in Table 6.1). In Figure 6.33, the signal
was collected from a SKF 6204 bearing with a crack in its inner race. In Figures
6.34- 6.36, the signals were collected from bearings, which had different severity of
outer race faults (shown in Table 6.1). In Figure 6.37, the signal was collected from a
bearing with a crack in its outer race. These two figures present the original vibration
signals of faulty bearings and the Basis Pursuit TF features. The severity of bearing
faults in the figures increases from Figures 6.29- 6.37.
The ORF and IRF of bearings SKF 6205 are 107 Hz and 162 Hz respectively. The
ORF and IRF of SKF6204 bearings are approximately 60 Hz and 100 Hz given the
different rotational shaft speed. These defect frequencies can be identified either
from time intervals or frequency intervals of features.
Generally, the time intervals of features matched the defect frequencies when the
fault was marginal. As faults developed, the frequency intervals of features (which
match defect frequencies) became more discernible. Both time intervals and
frequency intervals are shown clearly (Figures 6.29-6.37) when faults grew more
severe.
CHAPTER 6. RESULTS AND DISCUSSION
____________________________________________________________________
131
0 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08-3
-2
-1
0
1
2
3
Time(s)
Fre
quen
cy(H
z)
0 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.080
1000
2000
3000
4000
5000
6000
Figure 6.29: The time waveform and its Basis Pursuit TF plane obtained from
the bearing with 0.07 inch EDM IRF
0 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08-3
-2
-1
0
1
2
3
Time(s)
Fre
quen
cy(H
z)
0 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.080
1000
2000
3000
4000
5000
6000
Figure 6.30: The time waveform and its Basis Pursuit TF plane obtained from
the bearing with 0.14 inch EDM IRF
Acc
ele
ratio
n (
m/s2 )
Acc
ele
ratio
n (
m/s2 )
CHAPTER 6. RESULTS AND DISCUSSION
____________________________________________________________________
132
0 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08-3
-2
-1
0
1
2
3
Time(s)
Fre
quen
cy(H
z)
0 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.080
1000
2000
3000
4000
5000
6000
Figure 6.31: The time waveform and its Basis Pursuit TF plane obtained from
the bearing with 0.21 inch EDM IRF
0 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08-3
-2
-1
0
1
2
3
Time(s)
Fre
quen
cy(H
z)
0 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.080
1000
2000
3000
4000
5000
6000
Figure 6.32: The time waveform and its Basis Pursuit TF plane obtained from
the bearing with 0.28 inch EDM IRF
Acc
ele
ratio
n (
m/s2 )
Acc
ele
ratio
n (
m/s2 )
CHAPTER 6. RESULTS AND DISCUSSION
____________________________________________________________________
133
0 0.02 0.04 0.06 0.08 0.1 0.12-4
-2
0
2
4
Time(s)
Am
plitu
de
Time(s)
Fre
quen
cy(H
z)
0 0.02 0.04 0.06 0.08 0.1 0.120
1000
2000
3000
4000
Figure 6.33: The time waveform and its Basis Pursuit TF plane obtained from
the bearing with a crack in inner race
0 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08-4
-2
0
2
4
Time(s)
Fre
quen
cy(H
z)
0 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.080
1000
2000
3000
4000
5000
6000
Figure 6.34: The time waveform and its Basis Pursuit TF plane of the bearing
with ORF: 0.07inch
Acc
ele
ratio
n (
m/s2 )
(m
/s2 )
CHAPTER 6. RESULTS AND DISCUSSION
____________________________________________________________________
134
0 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08-4
-2
0
2
4
Time(s)
Fre
quen
cy(H
z)
0 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.080
1000
2000
3000
4000
5000
6000
Figure 6.35: The time waveform and its Basis Pursuit TF plane of the bearing
with ORF: 0.14 inch
0 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08-4
-2
0
2
4
Time(s)
Fre
quen
cy(H
z)
0 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.080
1000
2000
3000
4000
5000
6000
Figure 6.36: The time waveform and its Basis Pursuit TF plane of the bearing
with ORF: 0.21inch
Acc
ele
ratio
n (
m/s2 )
Acc
ele
ratio
n (
m/s2 )
CHAPTER 6. RESULTS AND DISCUSSION
____________________________________________________________________
135
0 0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16 0.18 0.2-4
-2
0
2
4
Time(s)
Am
plitu
de
Time(s)
Fre
quen
cy(H
z)
0 0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16 0.18 0.20
500
1000
1500
2000
2500
Figure 6.37: The time waveform and its Basis Pursuit TF plane of the bearing
with ORF: crack
6.2.5 Basis Pursuit Denoising
The preceding sections have shown that Basis Pursuit can be used to extract vibration
features of a faulty bearing efficiently. The atoms in a Basis Pursuit TF plane can be
used to reconstruct the signal for a more accurate representation of the fault as shown
in Figure 6.38. Figure 6.38 (a), (c), and (e) show the time waveforms from a normal
bearing, a bearing with inner race fault, a bearing with outer race fault. Figure 6.38
(b), (d), and (f) show signals whose noise has been significantly removed by using
Basis Pursuit. It appears after the denoising application that:
(1) there are hardly any impacts detectable in Figure 6.38 (b), the reconstructed
signals of normal bearings;
(2) regular impacts are clearly shown such as spikes A and B in Figure 6.38 (d) (the
reconstructed signals of the inner race faulty bearing), and spikes A and B in Figure
6.38 (f) (the reconstructed signals of the outer race faulty bearing) ;
(3) the measured impacts due to inner race faults that were merged in noise are now
visible.
(m
/s2 )
CHAPTER 6. RESULTS AND DISCUSSION
____________________________________________________________________
136
These differences are quantified as “Signal to Noise” ratios for both the original and
denoised signals as shown in Table 6.2. Signal to Noise ratios were calculated using
both the time and frequency domains and expressed in decibels. In general, the Basis
Pursuit denoised signals achieved signal to noise ratios improvements of about 40dB
which is very significant.
Table 6.1 Fault severity specifications in Figures 6.29- 6.37
0.1778mm (0.007 inch)
0.3556 mm (0.014 inch)
0.5334 mm (0.021 inch)
0.7112mm (0.028 inch)
Wire-cut crack
IRF Figure 6.29 Figure 6.30 Figure 6.31 Figure 6.32 Figure 6.33
ORF Figure 6.34 Figure 6.35 Figure 6.36 Figure 6.37
0 0.1 0.2 0.3-4
-2
0
2
4
0 0.1 0.2 0.3-1
-0.5
0
0.5
1
0 0.1 0.2 0.3-4
-2
0
2
4
0 0.1 0.2 0.3-1
-0.5
0
0.5
1
0 0.1 0.2 0.3-4
-2
0
2
4
Time(s)
Am
plitu
de
0 0.1 0.2 0.3-1
-0.5
0
0.5
1
Time(s)
Am
plitu
de
Figure 6.38: The time waveforms and the Basis Pursuit denoised signals: (a), (b)
normal, (c), (d) with IRF (e), (f) with ORF
A B
A B
(a) (b)
(c) (d)
(e) (f)
CHAPTER 6. RESULTS AND DISCUSSION
____________________________________________________________________
137
Table 6.2 Signal to Noise Ratio of the Original and the BP Denoised signals
Signals of bearings with faults SNR of Time Waveforms (dB)
SNR of BP denoised signals (dB)
Time 21.5 61.7 IRF
Frequency 29.0 69.8
Time 24.8 66.0 ORF
Frequency 29.6 72.1
6.3 Automatic fault diagnosis
The results of the evaluation of the various automatic fault diagnosis schema is
presented in the following sections.
6.3.1 Spectrum Based Automatic Fault Diagnosis
Data collected from a healthy bearing and bearings with IRF, ORF, and REF was
analysed using spectrum and spectrogram. An averaging procedure was used to
reduce the curse of dimension. Feature vectors with 16 elements were used for ANN
classification.
Four data sets obtained from the bearings under different conditions, which were
analysed using spectrum, and spectrogram, and then averaged, are shown in Figures
6.43- 6.51. The feature vectors of bearings under different conditions can be
distinguished from each other, using averaging spectrum and spectrogram.
The NNs were trained and tested based on the above features. Training procedures
were stopped after 300 epochs. The classification rate and estimation error are shown
in Table 6.3. The classification performance of the NN based on spectrum averaged
features and spectrogram averaged features are satisfactory. This is consistent with
the distinguished features extracted using these two techniques (as shown from
Figure 6.43 to Figure 6.51 ). The classification rate is low when features are
extracted by averaging wavelet packets, which is consistent with that the feature
vectors are similar and difficult to distinguish from each other among the features of
signals of bearings with different conditions.
Frequency features and time-frequency features can be successfully combined with
Neural Networks to classify different bearing conditions.
CHAPTER 6. RESULTS AND DISCUSSION
____________________________________________________________________
138
0 1000 2000 3000 4000 5000 60000
0.01
0.02
0.03
0.04
0.05
0.06
Frequency
Acc
eler
atio
n
Figure 6.39: Spectrum of the signal of a normal bearing
0 1000 2000 3000 4000 5000 60000
0.02
0.04
0.06
0.08
0.1
0.12
Frequency
Acc
eler
atio
n
Figure 6.40: Spectrum of the signal of a bearing with IRF
(m
/s2 )
(m
/s2 )
(Hz)
(Hz)
CHAPTER 6. RESULTS AND DISCUSSION
____________________________________________________________________
139
0 1000 2000 3000 4000 5000 60000
0.05
0.1
0.15
0.2
0.25
0.3
0.35
Frequency
Acc
eler
atio
n
Figure 6.41: Spectrum of the signal of a bearing with ORF
0 1000 2000 3000 4000 5000 60000
0.005
0.01
0.015
0.02
0.025
0.03
0.035
0.04
0.045
Frequency
Acc
eler
atio
n
Figure 6.42: Spectrum of the signal of a bearing with REF
(m
/s2 )
(m
/s2 )
(Hz)
(Hz)
CHAPTER 6. RESULTS AND DISCUSSION
____________________________________________________________________
140
0 1000 2000 3000 4000 5000 60000
0.001
0.002
0.003
0.004
0.005
0.006
0.007
0.008
0.009
0.01
Frequency(Hz)
Acc
eler
atio
n
Figure 6.43: Feature vectors based on Spectrum of the signal of a normal
bearing
0 1000 2000 3000 4000 5000 60000
0.005
0.01
0.015
0.02
0.025
0.03
0.035
0.04
Frequency(Hz)
Acc
eler
atio
n
Figure 6.44: Feature vector based on Spectrum of the signal of a bearing with
IRF
(m
/s2 )
(m/s
2 )
CHAPTER 6. RESULTS AND DISCUSSION
____________________________________________________________________
141
0 1000 2000 3000 4000 5000 60000
0.01
0.02
0.03
0.04
0.05
0.06
0.07
0.08
0.09
Frequency(Hz)
Acc
eler
atio
n
Figure 6.45: Feature vector based on spectrum of the signal of a bearing with
ORF
0 1000 2000 3000 4000 5000 60000
0.005
0.01
0.015
0.02
0.025
Frequency(Hz)
Acce
lera
tion
Figure 6.46: Feature vector based on Spectrum of the signal of a bearing with
REF
(m/s
2 ) (m
/s2 )
CHAPTER 6. RESULTS AND DISCUSSION
____________________________________________________________________
142
Time
Fre
quen
cy
0 100 200 300 400 500 600 700 800 900 1000
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Figure 6.47: Spectrogram of the signal of a normal bearing
Time
Fre
quen
cy
0 100 200 300 400 500 600 700 800 900 1000
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Figure 6.48: Spectrogram of the signal of a bearing with IRF
(s)
(Hz)
(s)
(Hz)
CHAPTER 6. RESULTS AND DISCUSSION
____________________________________________________________________
143
Time
Fre
quen
cy
0 100 200 300 400 500 600 700 800 900 1000
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Figure 6.49: Spectrogram of the signal of a bearing with ORF
Time
Fre
quen
cy
0 100 200 300 400 500 600 700 800 900 1000
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Figure 6.50: Spectrogram of the signal of a bearing with REF
(s)
(Hz)
(s)
(Hz)
CHAPTER 6. RESULTS AND DISCUSSION
____________________________________________________________________
144
0 2 4 6 8 10 12 14 160
0.1
0.2
0.3
0.4
0.5
0.6
0.7
The No. of a frequency band
Acc
eler
atio
n
Figure 6.51: Feature vector based on spectrogram of the signal of a normal
bearing
0 2 4 6 8 10 12 14 160
0.5
1
1.5
The No. of a frequency band
Acc
eler
atio
n
Figure 6.52: Spectrogram feature of the signal of a bearing with IRF
(m/s
2 ) (m
/s2 )
CHAPTER 6. RESULTS AND DISCUSSION
____________________________________________________________________
145
0 2 4 6 8 10 12 14 160
0.5
1
1.5
2
2.5
3
3.5
4
4.5
The No. of a frequency band
Acc
eler
atio
n
Figure 6.53: Spectrogram feature of the signal of a bearing with ORF
0 2 4 6 8 10 12 14 160
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
The No. of a frequency band
Acc
eler
atio
n
Figure 6.54: Spectrogram feature of the signal of a bearing with REF
(m/s
2 ) (m
/s2 )
CHAPTER 6. RESULTS AND DISCUSSION
____________________________________________________________________
146
Table 6.3 Classification performance of automatic diagnosis based on Spectra
and Spectrogram
Feature name
Fault type Training sets
Test sets
Correct classification
Misclas-sification
Classification rate
MSE
IRF 40 0 100% 0.00012 ORF 40 0 100% 0.00040 REF 40 0 100% 0.00025
Mean based on Spectrum
Normal
80 40
40 0 100% 0.00018 IRF 40 0 100% 0.00023 ORF 40 0 100% 0.00037 REF 40 0 100% 0.00014
Mean based on spectrogram Normal
80
40
40 0 100% 0.00023
6.3.2 DWPA Feature Based Automatic Fault Diagnosis
An example of the time waveforms collected and analysed using DWPA to level
three with the symlet function, is shown in Figures 6.55- 6.58. The vertical axis
represents the decomposition level of DWPA, which ranges from one to six. The
horizontal axis includes the wavelet packets in each decomposition level. Each of the
packets is framed by vertical dashed lines and horizontal lines and presented as time
series signals (filtered within the frequency band of the wavelet packet). As shown in
the figures, it can be noted that four signals were filtered and then the energy of each
signal was localised in narrow frequency bands after DWPA. These high energy
frequency bands contain defect information. Therefore parameters derived from
these frequency bands features can be used as features. The energy is focused on the
frequency bands of the first and second wavelet packets of the signal obtained from a
normal bearing (see Figure 6.55). The energy appeared focused on the frequency
bands of the fourth and fifth wavelet packets with high vibration levels (see Figures
6.56- 6.58) by using DWPA analysis of the signals of the bearings with IRF, ORF,
and REF.
Features such as Mean Value, Variance, Energy, Skewness, Kurtosis, Root Mean
Square (RMS), Crest factor, and Matched filter were further derived from the
reconstructed time series signals of the DWPA wavelet packets. These features were
used to represent the time frequency characteristic of the vibration signals.
These features extracted from data sets of bearings are shown in Figures 6.59- 6.66.
In Figure 6.59, the mean value derived from wavelet packets as a feature, provides
little value as a predictor for fault classification.
CHAPTER 6. RESULTS AND DISCUSSION
____________________________________________________________________
147
As shown in Figures 6.60- 6.64, the parameters, Variance, Energy value, and RMS,
are successful in picking up high energy vibration frequency bands, which are often
induced by defects in rolling element bearings. The defect led to significant increases
of these parameters in these graphs. The feature graph of each signal can be easily
distinguished. Even the feature trends of bearings are clearly identifiable. These
features provide accurate classification rate, as presented in Table 6.4 and Table 6.5.
Skewness, and Kurtosis are high order statistical parameters, which are accepted as
efficient parameters in analysing time series of vibration signals. The Crest Factor
also has the potential to be an effective feature in time series classification. Figures
6.61- 6.66 show the trends of these three parameters. It can be observed that the
feature graphs of each bearing signal are distinctive from each other. Kurtosis is
shown to be effective in the automatic classification of the bearing conditions(as
shown in Table 6.4 and Table 6.5). Skewness can be used to reach acceptable
classification rates providing the Neural Network classifiers are well designed.
The Matched Filter is a parameter which extracts frequency features. The Matched
Filter derived from wavelet packets aims at capturing frequency characteristic in
each wavelet packet frequency band. As shown in the Figure 6.59, the Matched Filter
is also able to capture defect caused high energy vibrations, which are presented as
high amplitude peaks. The Matched Filter feature graphs of bearings under the four
conditions are almost as distinctive as those obtained from the RMS, Energy Value,
and Variance parameters.
The FFNN classifiers were evaluated by the input of the above mentioned features.
Training procedures were halted after 300 epochs. The classification rate and
estimation error of the single output Neural Network and the multi output Neural
Network are shown in Table 6.4 and Table 6.5, respectively. The correct
classification rate of the single output Neural Network classifier using Mean Value
as the feature and input ranged from 37.5% for REF to 72.5% for Healthy bearings.
A hundred percent classification rate was achieved using RMS as feature-inputs to
the multi-output Neural Network for all the bearing conditions.
CHAPTER 6. RESULTS AND DISCUSSION
____________________________________________________________________
148
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1-7
-6
-5
-4
-3
-2
-1
0
1
Spl
it Le
vel
Frequency[Time]
Figure 6.55: The DWPA of the signal of a normal bearing
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1-7
-6
-5
-4
-3
-2
-1
0
1
Spl
it Le
vel
Frequency[Time]
Figure 6.56: The DWPA of the signal of a bearing with IRF
CHAPTER 6. RESULTS AND DISCUSSION
____________________________________________________________________
149
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1-7
-6
-5
-4
-3
-2
-1
0
1
Spl
it Le
vel
Frequency[Time]
Figure 6.57: The DWPA of the signal of a bearing with ORF
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1-7
-6
-5
-4
-3
-2
-1
0
1
Spl
it Le
vel
Frequency[Time]
Figure 6.58: The DWPA of the signal of a bearing with REF
CHAPTER 6. RESULTS AND DISCUSSION
____________________________________________________________________
150
1 2 3 4 5 6 7 8-0.005
0
0.005
0.01
0.015
0.02
0.025
0.03
0.035
The No. of a Wavelet Packet
Mea
n V
alue
REFIRFORFNormal
Figure 6.59: Features based on wavelet packets: Mean Value
1 2 3 4 5 6 7 80
0.05
0.1
0.15
0.2
0.25
The No. of a Wavelet Packet
Var
ianc
e
REFIRFORFNormal
Figure 6.60: Features based on wavelet packets: Variance
CHAPTER 6. RESULTS AND DISCUSSION
____________________________________________________________________
151
1 2 3 4 5 6 7 8-0.8
-0.6
-0.4
-0.2
0
0.2
0.4
0.6
The No. of a Wavelet Packet
Ske
wne
ss
REFIRFORFNormal
Figure 6.61: Features based on wavelet packets: Skewness
1 2 3 4 5 6 7 82
4
6
8
10
12
14
16
The No. of a Wavelet Packet
Kur
tosi
s
REFIRFORFNormal
Figure 6.62: Features based on wavelet packets: Kurtosis
CHAPTER 6. RESULTS AND DISCUSSION
____________________________________________________________________
152
1 2 3 4 5 6 7 80
0.01
0.02
0.03
0.04
0.05
0.06
0.07
0.08
0.09
0.1
The No. of a Wavelet Packet
Ene
rgy
Val
ue
REFIRFORFNormal
Figure 6.63: Features based on wavelet packets: Energy
1 2 3 4 5 6 7 80
0.05
0.1
0.15
0.2
0.25
0.3
0.35
0.4
0.45
0.5
The No. of a Wavelet Packet
Roo
t M
ean
Squ
are
REFIRFORFNormal
Figure 6.64: Features based on wavelet packets: Root Mean Square
CHAPTER 6. RESULTS AND DISCUSSION
____________________________________________________________________
153
1 2 3 4 5 6 7 84
6
8
10
12
14
16
The No. of a Wavelet Packet
Cre
st F
acto
r
REFIRFORFNormal
Figure 6.65: Features based on wavelet packets: Crest Factor
1 2 3 4 5 6 7 8-20
0
20
40
60
80
100
The No. of a Wavelet Packet
Mat
ched
filt
er
REFIRFORFNormal
Figure 6.66: Features based on wavelet packets: Matched Filter
CHAPTER 6. RESULTS AND DISCUSSION
____________________________________________________________________
154
Table 6.4 Performance of the single output FFNN using DWPA features
Feature name
Fault type
Training sets
Test sets
Correct classific-ation
Misclass-ification
Classifi-cation rate
MSE
IRF 27 13 67.5% 0.04254 ORF 17 23 37.5% 0.04186 REF 25 15 62.5% 0.08220
Mean
Normal
80
40
30 10 72.5% 0.06997 IRF 40 0 100% 0.00027 ORF 39 1 97.5% 0.00013 REF 40 0 100% 0.00007
Variance
Normal
80 40
40 0 100% 0.00002 IRF 39 0 100% 0.00002 ORF 40 0 97.5% 0.00001 REF 36 4 90% 0.00004
Energy
Normal
80 40
40 0 100% 0.00001 IRF 25 15 62.5% 0.03795 ORF 27 13 67.5% 0.04283 REF 17 23 42.5% 0.04376
Skewness
Normal
80 40
40 14 65% 0.33935 IRF 38 2 95% 0.02693 ORF 36 4 90% 0.03392 REF 35 5 87.5% 0.03249
Kurtosis
Normal
80 40
39 1 97.5% 0.00188 IRF 39 1 97.5% 0.00001 ORF 39 1 97.5% 0.00009 REF 38 2 95% 0.00008
RMS
Normal
80 40
40 0 100% 0.00009 IRF 36 4 90% 0.09938 ORF 38 2 95% 0.15510 REF 21 19 52.5% 0.08706
Crest Factor
Normal
80 40
39 1 97.5% 0.06382 IRF 29 11 72.5% 0.03544 ORF 39 1 97.5% 0.03311 REF 39 1 97.5% 0.01245
Matched Filter
Normal
80 40
40 0 100% 0.02567
CHAPTER 6. RESULTS AND DISCUSSION
____________________________________________________________________
155
Table 6.5 Performance of the multi output FFNN using DWPA features
Feature name
Fault type
Training sets
Test Sets
Correct classif-ication
Miscl-sificat-ion
Classifi-cation rate
MSE
IRF 31 9 77.5% 0.03071 ORF 40 0 100% 0.02867 REF 24 16 60% 0.03621
Mean
Normal
80
40
40 0 85% 0.02355 IRF 40 0 100% 0.00029 ORF 39 1 97.5% 0.00020 REF 38 2 95% 0.00035
Variance
Normal
80 40
40 0 100% 0.00033 IRF 40 0 100% 0.00022 ORF 40 0 100% 0.00013 REF 40 0 100% 0.00010
Energy
Normal
80 40
40 0 100% 0.92000 IRF 25 15 62.5% 0.03121 ORF 30 10 75% 0.05035 REF 33 7 82.5% 0.03737
Skewness
Normal
80 40
37 3 92.5% 0.06933 IRF 34 6 85% 0.00996 ORF 40 0 100% 0.01326 REF 35 5 87.5% 0.00531
Kurtosis
Normal
80 40
39 1 97.5% 0.00738 IRF 40 0 100% 0.00025 ORF 40 0 100% 0.00014 REF 40 0 100% 0.00012
RMS
Normal
80 40
40 0 100% 0.00028 IRF 39 1 97.5% 0.02227 ORF 40 0 100% 0.01573 REF 33 7 82.5% 0.02476
Crest Factor
Normal
80 40
35 5 87.5% 0.02019 IRF 35 5 87.5% 0.01746 ORF 39 1 97.5% 0.01708 REF 39 1 97.5% 0.02501
Matched Filter
Normal
80 40
40 0 100% 0.07307
6.3.3 Matching Pursuit Feature Based Automatic Fault Diagnosis
Data collected from a Healthy bearing, and bearings with IRF, ORF, and REF were
analysed using Matching Pursuit with four iterations. The calculation to this iteration
level was efficient and precise enough for accurate FFNN classification. Figures
6.67- 6.70 shows typical data that was collected from bearings in Healthy condition,
and bearings with IRF, ORF, and REF, respectively.
CHAPTER 6. RESULTS AND DISCUSSION
____________________________________________________________________
156
The time series signal of a bearing under normal condition appears similar as the
signal of a bearing under REF condition. It is also difficult to distinguish ORF
condition and IRF condition by observing the time series signals. The four signals in
Figures 6.67- 6.70 were further analysed using Matching Pursuit with four iterations
using a wavelet packet dictionary with a symlet function, is shown in Figure 6.71.
The signals were decomposed to the fourth iteration and then presented as time
frequency maps with relatively coarse resolution. The high energy time frequency
components appear “blocky”. Although these decompositions were not accurate
enough for direct interpretation, the features were identifiable from the time
frequency analysis of bearing signals under Healthy, IRF, ORF, and REF conditions.
Note that the data were filtered by the wavelet function in Matching Pursuit analysis
with energy being localised in narrow frequency bands. In Figure 6.71, the high
energy of the signals was concentrated primarily in certain frequency bands for the
different conditions:
• 0-2 kHz band for the Healthy condition –see Figure 6.67,
• 1-5 kHz band for the Inner Race Fault – see Figure 6.68,
• 3-4 kHz band for the Outer Race Fault –see Figure 6.69,
• Under 2 kHz, and 2-4 kHz for the Rolling Element Fault –see Figure 6.70.
An added observation in the above mentioned four figures is the variations in time
intervals for the different frequency bands of the different signals. These time
variations further enable one to differentiate the faults.
The Matching Pursuit coefficients of the signals of bearings under different
conditions are shown in Figures Figure 6.71 (a)-(d). These graphs correspond to the
time frequency maps in Figures 6.67- 6.70, respectively. It can be seen that the
Matching Pursuit coefficients of signals for the different conditions are clearly
distinguishable with the most activity in Figure 6.71 (c) – the signal for the inner race
fault.
Feature vectors were further derived from the above Matching Pursuit coefficients
and were formed by selecting the maximum values among these coefficients (as
CHAPTER 6. RESULTS AND DISCUSSION
____________________________________________________________________
157
shown in Figure 6.72 (a)-(d)). These feature vectors for the different conditions
appear:-
• Flat for Healthy condition - see Figure 6.72 (a),
• With several peaks for Inner Race Fault, Outer Race Fault, and Rolling Element
Fault – see Figure 6.72 (b) - (d)).
These derived feature vectors can be used to classify the different bearing conditions.
The FFNN was tested using the above derived features. In total, 120 data sets of each
bearing condition were analysed in the training and testing of the proposed
methodology. The training procedures ceased after 300 epochs. The classification
rate and estimation error are shown in Table 6.6. The maximum value number used
as the dimension of the input feature vectors were 16, 32, 64, and 128 respectively.
The classification rate ranged from 70 % (outer race fault condition) to 97.5% (inner
race fault condition) when using 16 maximum values of Matching Pursuit
coefficients as the inputs. The classification rate ranged from 55 % (rolling element
fault condition) to 97.5% (inner race fault condition) when using 128 maximum
values of Matching Pursuit coefficients as the inputs. It appears that the higher
maximum value numbers did not increase the classification accuracy.
Time(s)
Fre
quen
cy(H
z)
0 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.080
1000
2000
3000
4000
5000
6000
Figure 6.67: The Matching Pursuit of the vibration signal of a bearing under
condition: Normal
CHAPTER 6. RESULTS AND DISCUSSION
____________________________________________________________________
158
Time(s)
Fre
quen
cy(H
z)
0 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.080
1000
2000
3000
4000
5000
6000
Figure 6.68: The Matching Pursuit of the vibration signal of a bearing under
condition: ORF
Time(s)
Fre
quen
cy(H
z)
0 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.080
1000
2000
3000
4000
5000
6000
Figure 6.69: The Matching Pursuit of the vibration signal of bearing under
condition: IRF
CHAPTER 6. RESULTS AND DISCUSSION
____________________________________________________________________
159
Time(s)
Fre
quen
cy(H
z)
0 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.080
1000
2000
3000
4000
5000
6000
Figure 6.70: The Matching Pursuit of the vibration signal of a bearing under
condition: REF
0 1000 2000 3000 4000 5000 6000-1
-0.5
0
0.5
0 1000 2000 3000 4000 5000 6000-2
0
2
0 1000 2000 3000 4000 5000 6000-10
0
10
0 1000 2000 3000 4000 5000 6000-1
0
1
Figure 6.71: The Matching Pursuit (MP) coefficients of vibration signals of
bearings under conditions: (a) Normal (b) ORF (c) IRF (d) REF
(a)
(b)
(c)
(d)
Am
plitu
de(m
/s2 )
The No. of MP coefficients
CHAPTER 6. RESULTS AND DISCUSSION
____________________________________________________________________
160
0 20 40 60 80 100 120
2
4
6
0 20 40 60 80 100 120
2
4
6
0 20 40 60 80 100 120
2
4
6
0 20 40 60 80 100 120
2
4
6
Figure 6.72: The Matching Pursuit Features of Vibration Signals of bearings
under conditions: (a) Normal (b) ORF (c) IRF (d) REF
Table 6.6 Classification performance of different procedures using Matching
Pursuit
Features Fault type Training sets
Test sets
Correct classifi-cation
Misclass-ification
Classif-ication rate
MSE
IRF 39 1 97.5% 0.01657 ORF 28 12 70% 0.01626 REF 33 7 82.5% 0.01835
Maximum 16 value
Normal
80 40
31 9 77.5% 0.02855 IRF 24 16 60% 0.00962 ORF 36 4 90% 0.00887 REF 26 14 65% 0.01043
Maximum 32 value
Normal
80 40
35 5 87.5% 0.00981 IRF 24 16 60% 0.00012 ORF 35 5 87.5% 0.00219 REF 32 8 80% 0.00033
Maximum 64 value
Normal
80 40
35 5 87.5% 0.00224 IRF 39 1 97.5% 0.00012 ORF 39 1 97.5% 0.00013 REF 22 18 55% 0.00065
Maximum 128 value
Normal
80 40
35 5 87.5% 0.00019
(a)
(b)
(c)
(d)
Am
plitu
de(m
/s2 )
The No. of feature vector
CHAPTER 6. RESULTS AND DISCUSSION
____________________________________________________________________
161
6.3.4 Basis Pursuit Feature Based Automatic Fault Diagnosis
Case studies were conducted using data collected from healthy bearings, as well as
bearings with IRF, ORF and REF to test the proposed intelligent diagnostic
technique. The data were analysed using Basis Pursuit for FFNN classification.
Figures 6.73- 6.76 illustrates an example of four data sets, which were analysed
using the Basis Pursuit with a symlet wavelet packet dictionary but at a low level of
decomposition. In this study, it was found that decomposition to level 4 was deemed
to be sufficient for accurate feature extraction as discussed in Chapter 4, Section 4.4.
In the time-frequency maps, the colours range from white to black, with shades of
grey representing, the amplitude values of the time-frequency components. The
signals were decomposed to the fourth iteration and then presented as time-frequency
maps with relatively coarse resolution. The high energy time-frequency components
appear clearly in the time-frequency maps. Although these decompositions were not
accurate enough for direct interpretation, the time-frequency analysis of bearing
signals under Healthy, IRF, ORF and REF conditions were clearly identifiable. It can
be noted that the data were filtered, and the resultant energy was localised in narrow
frequency bands after Basis Pursuit analysis. Referring to these four figures, it can be
seen that the high energy of the signals was concentrated primarily in certain
frequency bands within varying conditions:
• 0-3 kHz band for Healthy condition (see Figure 6.73),
• 2-5 kHz band for Inner Race Fault (see Figure 6.74),
• 3-4 kHz band for Outer Race Fault (Figure 6.75),
• 3-4 kHz for Rolling Element Fault (Figure 6.76).
The above mentioned four figures provide additional information relating to the
difference of the amplitude values of high energy time-frequency components and
the variations in time intervals for the different frequency bands of the different
signals. These time variations further enable one to differentiate the faults.
Comparing Basis Pursuit analysis with Matching Pursuit analysis (as shown in
Figures 6.67- 6.70), Basis Pursuit provides greater accuracy of information on fault
related features. The high energy components from Basis Pursuit analysis in Figures
CHAPTER 6. RESULTS AND DISCUSSION
____________________________________________________________________
162
6.73-76 are both more focussed and distinguishable than those from Matching
Pursuit analysis in Figure 6.71.
Figure 6.77 demonstrate the Basis Pursuit coefficients corresponding with the time-
frequency maps in Figure 6.73, Figure 6.74, Figure 6.75, and Figure 6.76
respectively. It can be seen that the Basis Pursuit coefficients of signals under
different conditions are clearly distinguishable with the most activity in Figure
6.77(c), the signal for the inner race fault.
Feature vectors were further derived from the above-mentioned Basis Pursuit
coefficients and formed by selecting the maximum values among these coefficients
(as shown in Figure 6.78 (a)-(d)). Results indicate that the feature vectors of the
bearing signals under different conditions:
• appear flat in the Healthy condition (refer to Figure 6.78 (a)); and
• have some peak values for Inner Race Fault, Outer Race Fault and Rolling
Element Fault (see Figure 6.78 (b) - (d)).
These derived feature vectors appear identifiable for the classification of the bearing
conditions. The FFNN was tested based on the above derived features. In total, 120
data sets of each bearing condition were analysed to test the proposed methodology.
The training procedures were ceased after 300 epochs. The resultant classification
rate and estimation error are shown in Table 6.7. The maximum value number used
as the dimension of the input feature vectors were 16, 32, 64 and 128 respectively.
The classification rate ranged from 85% for the Rolling Element Fault condition to
100% for the other conditions when using 16 maximum values of Basis Pursuit
coefficients as the inputs. The classification rate ranged from 55% for the Rolling
Element Fault condition to 97.5% for the Outer Race Fault condition when using 128
maximum values of Basis Pursuit coefficients as the inputs. These features
performed poorly for the classification of bearings under the Rolling Element Fault
condition.
CHAPTER 6. RESULTS AND DISCUSSION
____________________________________________________________________
163
Time(s)
Fre
quen
cy(H
z)
0 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.080
1000
2000
3000
4000
5000
6000
Figure 6.73: The Basis Pursuit of the vibration signals of bearings under
condition: Normal
Time(s)
Fre
quen
cy(H
z)
0 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.080
1000
2000
3000
4000
5000
6000
Figure 6.74: The Basis Pursuit of the vibration signals of bearings under
condition: ORF
CHAPTER 6. RESULTS AND DISCUSSION
____________________________________________________________________
164
Time(s)
Fre
quen
cy(H
z)
0 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.080
1000
2000
3000
4000
5000
6000
Figure 6.75: The Basis Pursuit of the vibration signals of bearings under
condition: IRF
Time(s)
Fre
quen
cy(H
z)
0 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.080
1000
2000
3000
4000
5000
6000
Figure 6.76: The Basis Pursuit of the vibration signals of bearings under
condition: REF
CHAPTER 6. RESULTS AND DISCUSSION
____________________________________________________________________
165
Figure 6.77: The Basis Pursuit coefficients of vibration signals of bearings under
conditions: (a) Normal (b) ORF (c) IRF (d) REF
0 10 20 30 40 50 60
0.20.40.60.8
11.2
0 10 20 30 40 50 60
0.20.40.60.8
11.2
0 10 20 30 40 50 60
0.20.40.60.8
11.2
0 10 20 30 40 50 60
0.20.40.60.8
11.2
Figure 6.78: The Basis Pursuit features of vibration signals of bearings under
conditions: (a) Normal (b) ORF (c) IRF (d) REF
0 1000 2000 3000 4000 5000 6000-0.2
0
0.2
0 1000 2000 3000 4000 5000 6000-0.5
0
0.5
0 1000 2000 3000 4000 5000 6000-2
0
2
0 1000 2000 3000 4000 5000 6000-0.2
0
0.2
(a)
(b)
(c)
(d)
Am
plitu
de (m
/s2 )
The No. of coefficients
(a)
(b)
(c)
(d) The No. of feature vector
Am
plitu
de (m
/s2 )
CHAPTER 6. RESULTS AND DISCUSSION
____________________________________________________________________
166
Table 6.7 Classification performance of different procedures using Basis Pursuit
Features Fault type
Training sets
Test Sets
Correct classification
Misclass-ification
Classification rate
MSE
IRF 40 0 100% 0.00014 ORF 40 0 100% 0.00043 REF 34 6 85% 0.00028
Maximum 16 value
Normal
80 40
40 0 100% 0.00005 IRF 39 1 97.5% 0.00001 ORF 38 2 95% 0.00001 REF 37 3 92.5% 0.00007
Maximum 32 value
Normal
80 40
40 0 100% 0.00015 IRF 28 12 70% 0.00371 ORF 31 9 77.5% 0.00222 REF 35 5 87.5% 0.00346
Maximum 64 value
Normal
80 40
40 0 100% 0.00224 IRF 25 15 62.5% 0.00021 ORF 37 3 97.5% 0.00037 REF 25 15 55% 0.00088
Maximum 128 value
Normal
80 40
38 2 95% 0.00007
6.4 Discussion and Conclusion
The study on TF plane representation of vibration signals has shown that the Basis
Pursuit technique is able to represent vibration signals with sparsely separated
features. Best basis DWPA on the other hand produces more irrelevant atoms, and
which do not contribute to fault features and which can lead to inaccurate diagnosis.
The Matching Pursuit technique performs better than the best basis DWPA with
improved resolution. However, its performance is judged to be less accurate than
Basis Pursuit in fault diagnosis. Matching Pursuit and Basis Pursuit outperform
DWPA in fault diagnosis due to the use of the wavelet packet dictionary. This
dictionary introduces additional wavelet atoms to the analysis. In general, a
maximum of )(log)1( 2 NN + atoms are chosen for decomposition of a signal, where
N is the length of the signal. This procedure provides more flexibility and accuracy
for representing the vibration characteristics of rolling bearings. Furthermore, the
global optimisation rule in Basis Pursuit makes it more accurate than Matching
Pursuit. Basis Pursuit resolves vibration signals more finely with very sparse
significant atoms during decomposition. As a result, defect frequency component
features are efficiently extracted and displayed in the time-frequency plane. An
CHAPTER 6. RESULTS AND DISCUSSION
____________________________________________________________________
167
added feature of Basis Pursuit is that the method needs relatively short lengths of
data for effective analysis while traditional diagnostic techniques require larger
amounts of vibration time series data.
Basis Pursuit also increases the signal to noise ratio in identifying fault features. It
removes high frequency noise thereby accentuating the defect frequency
components. The limitation of Basis Pursuit is that its computation tends to take
longer in comparison with DWPA and Matching Pursuit using the same length of
data. However, this disadvantage is offset by the fact that Basis Pursuit can use
shorter data lengths than the other techniques.
The results of NN classifier using different features also show that the proposed
techniques for feature extraction are mostly effective. The features derived from
spectrum, spectrogram, DWPA, Matching Pursuit, and Basis Pursuit are identifiable
for rolling element bearing conditions.
6.5 Summary
This chapter provides a discussion of the research findings obtained from
investigation into the application of three time-frequency analysis techniques – the
improved DWPA, improved Matching Pursuit, and the novel Basis Pursuit.
Comparisons were made regarding a variety of characteristics of these three analysis
techniques. Advantages and disadvantages of these three techniques were described.
This chapter concluded with results obtained from evaluation of the classification
rates of the automatic diagnostic techniques incorporating these three time-frequency
analysis techniques.
CHAPTER 7. CONCLUSION
____________________________________________________________________
168
CHAPTER 7. CONCLUSION
7.1 Introduction
This thesis has presented novel time-frequency analysis techniques and ANN
classifiers, which were validated on vibration signals of bearings under different
conditions.
This research contributes new knowledge to the feature extraction of vibration
signals of rolling element bearings and the intelligent interpretation of the extracted
features. The major objectives of this research were to improve the reliability of fault
diagnostic techniques and reduce post processing tasks normally conducted by
human experts.
This research focused on wavelet based feature extraction and automation of fault
identification using these features. The development of wavelet based vibration
analysis improves the time-frequency resolution of defect-related features as well as
identifies defects related features.
An overall conclusion of the conducted work is provided in the following sections.
7.2 The Improved DWPA, Matching Pursuit, and Basis Pursuit
The literature demonstrated clearly that research into the further development of the
DWPA and Matching Pursuit techniques for fault diagnosis would be both needed
and justified. In this work, it was proven that these wavelet based methods are
capable of accurate feature extraction. Basis Pursuit has not yet found application in
the diagnosis of bearing defects. This research has addressed the matter and has
presented the first successful application of this technique.
The best basis DWPA can be used to filter high energy frequency bands without
choosing frequency bands. Best basis selection helps DWPA capture relevant
features as well as optimising computation effort.
The Matching Pursuit and Basis Pursuit can produce time frequency maps which
often distinguish defect related components clearly. In particular, Basis Pursuit can
generate better resolution time-frequency maps with the same number of
decomposition levels when compared with the Matching Pursuit and the DWPA.
CHAPTER 7. CONCLUSION
____________________________________________________________________
169
Basis Pursuit denoising improves the signal to noise ratio of vibration signals
significantly and enables defects to be more discernible in time waveforms.
These three techniques still have limitations when applied to vibration analysis. The
wavelet filters do not always identify the useful components in a time frequency
map. Some of these components can still present problems to diagnostics personnel
in their interpretation. Practically these techniques can sometimes have limitations in
highlighting low frequency components such as those generated due to shaft
misalignment or balancing because of the high sampling frequency data sampling
often used in standard condition monitoring programs.
7.3 Automatic Diagnosis Using Spectrum
Conventionally, the Fourier spectrum is used as the basis in most automatic fault
diagnostics tasks. This research used Fourier spectra initially and subsequently used
the time frequency spectrogram in developing an automatic fault diagnosis
technique. Frequency features and time-frequency feature extraction can be
successfully combined with Neural Networks to classify different bearing conditions.
The spectrum based features and spectrogram features are identifiable for vibrations
of bearings under different conditions. The designed Feed Forward Neural Network
(FFNN) performed well in classification of faults. The proper design of an FFNN is
important for successful diagnosis. An FFNN, with two hidden layers and four
output nodes, produced accurate diagnostic results with accurate classification and
low Mean Square Error (MSE).
7.4 Automatic Diagnosis Using DWPA
The statistical parameters derived from DWPA wavelet packets can be used as
features and fed into Neural Network classifiers to distinguish healthy and faulty
bearings.
The parameters Variance, Skewness, Kurtosis, Energy, RMS, Crest Factor and
Matched Filter generally produced faithful classification using the multi output
FFNN. In particular, Variance, RMS, Energy value and Matched Filter provided the
best classification result. The proposed DWPA features for NN classifiers worked
CHAPTER 7. CONCLUSION
____________________________________________________________________
170
well in diagnosing bearing faults and can be widely applied to a large range of both
rotating and reciprocating machinery.
Compared with the spectrum and spectrogram analysis based automatic diagnosis,
features using the mean value derived from DWPA produced poor classification rate
of bearing faults.
7.5 Automatic Diagnosis Using Matching Pursuit
Matching Pursuit was proposed as a feature extraction technique in the automatic
diagnosis of rolling element bearing faults. Matching Pursuit can effectively extract
features which are subsequently fed to a Feed Forward Neural Network to accurately
classify bearing conditions. In particular, accurate classification was obtained with
16 and 32 maximum values of Matching Pursuit coefficients as features.
7.6 Automatic Diagnosis Using Basis Pursuit
Basis Pursuit was shown to be more effective then Matching Pursuit in being able to
extract features and provided better classification when used with the FFNN
classifier for automatic fault diagnosis. In this study, the FFNN classifier accurately
categorised bearings’ conditions for Healthy, IRF, ORF and REF. The 16, 32, and 64
maximum values of Basis Pursuit coefficients (as features) produced accurate
classification.
CHAPTER 8. FUTURE RESEARCH
____________________________________________________________________
171
CHAPTER 8. FUTURE RESEARCH
8.1 Signal Processing Techniques for Feature Extraction
The time-frequency analysis techniques developed and applied in this thesis
candidature can be further refined for feature extraction of vibration signals from the
point of view of:
• Reduced computation time;
• Producing time-frequency maps with better resolution, and
• Reconstruction of time waveforms with less noise.
The computation of Matching Pursuit and Basis Pursuit analysis can be time
consuming. Further work on investigating the algorithm for optimizing the selection
of atoms can be done to improve the computation time and the resolution of
extracted features. The algorithm for solving Linear Problems can also be improved
for economic computation by using advances in computation theory.
The wavelet packet dictionary used in this study assisted in producing effective
bearing fault detection. The wavelet packet dictionary can be further developed to
better correlate with vibrations in Matching Pursuit and Basis Pursuit analyses.
8.2 Artificial Intelligence for Diagnosis
Neural Networks were employed in this study of automatic fault diagnosis. This
work can be extended by employing other artificial intelligence techniques for fault
diagnosis. Evolutionary algorithms, genetic algorithms, fuzzy logic, combined with
expert systems may be applied in various condition monitoring and fault diagnosis.
Evolutionary algorithms can assist in determining the best features during feature
extraction. Fuzzy logic can assist in human language input and output. The
development of support vector machines in the area of computational intelligence
can assist in more accurate and faster classification of bearing conditions. Features
selected using Matching Pursuit analysis and Basis Pursuit analysis can be used as
inputs to support vector machines, evolutionary algorithms, genetic algorithms, or
fuzzy logic, for automatic fault diagnosis.
CHAPTER 8. FUTURE RESEARCH
____________________________________________________________________
172
A variety of Neural Networks can be developed and applied in intelligent fault
diagnosis. These Neural Networks can be Self Organised Maps, Recurrent Neural
Networks, and Probabilistic Neural Networks. These Neural Networks can be used as
classifiers based on the features extracted using DWPA, Matching Pursuit, and Basis
Pursuit analysis. Feature selection can be further investigated based on DWPA,
Matching Pursuit, and Basis Pursuit and incorporated into these various Neural
Networks to achieve fault diagnosis of rolling element bearings.
8.3 Incipient Fault Detection using Time-Frequency Analysis Techniques
This work has been concerned with the development of the overall diagnostics
technique and has not particularly focussed on incipient faults. Further work may
investigate the sensitivity of the technique to detecting marginal faults.
This study has been concerned with studying rolling element bearing faults which
represent a fundamental component in rotating machinery. The work can be extended
to other rotational components such as pumps, fans and motors. The techniques used
in this study may also be validated on various operating regimes such as high and
low speed condition.
8.4 Automated Diagnosis of Process Monitoring and Material
Degradation.
The time frequency techniques and automated diagnosis procedure evaluated in this
study can also be used in other fields for diagnostics such as process monitoring and
material degradation studies particularly for structural health monitoring.
8.5 Automatic diagnosis of transmission systems
Power transmission systems such as geartrains have similarities to the operation of
rolling element bearings given their rotational and periodic behaviour. The feature
extraction technique proposed in this thesis can be directly applied to these classes of
machines. For example, the characteristic fault frequencies of a transmission system
can be extracted from the time-frequency plane derived from DWPA, Matching
Pursuit, and Basis Pursuit and analysed in exactly the same manner as that adopted in
this thesis.
CHAPTER 8. FUTURE RESEARCH
____________________________________________________________________
173
8.6 Commercializing an integral intelligent diagnostic toolbox
This research provides a solid and successful study of automatic fault diagnosis of
rolling bearings using time-frequency analysis and neural networks. The Matching
and Basis Pursuit techniques can be incorporated into a comprehensive toolbox for
fault diagnosis which needs to be further developed and commercialized to improve
the practice of condition monitoring and diagnostics.
REFERENCES
____________________________________________________________________
174
REFERENCES
1. Baillie, D. and J. Mathew, Diagnosing Rolling Element Bearing Faults with
Artificial Neural Networks. Acoustics Australia, 22(3): pp. 79-84. (1994).
2. Altmann, J., Application of Discrete Wavelet Packet Analysis for the
Detection and Diagnosis of Low Speed Rolling-Element Bearing Faults.
Monash University, Melbourne. (PhD thesis,1999).
3. Mathew, J., Standards in condition monitoring, in Proceedings of Condition
Monitoring 97, (keynote Paper), Xian, P R China: pp. 1-9. (1997).
4. Yardley, E.D., Condition Monitoring-Engineering the Practice, Professional
Engineering Publishing, Bury St Edmunds and London, UK: pp. 25-27.
(2002).
5. Rao, B.K.N., Handbook of condition monitoring. Oxford, UK, Elsevier
Advanced Technology: pp. 49-80. (1996).
6. Bently, D.E. and C.T. Hatch, Fundamentals of Rotating Machinery
Diagnostics. Bob Grissom ed. Canada, Bently pressurized bearing press: pp.
11-12. (2002).
7. Hale, V. and J. Mathew, High and low speed bearings, in Condition
Monitoring Frontiers, the second (CM)2 Forum, Melbourne, Australia: pp.
21-30. (1995).
8. Rolling Element Bearings. http://www.cmcpweb.com/appnotes/reb.htm.
9. Russell, S. and P. Norvig, Artificial intelligence: a modern approach,
Englewood Cliffs, N.J, Prentice Hall: pp. 23-24. (1995).
REFERENCES
____________________________________________________________________
175
10. Haykin, S., Neural networks: a comprehensive foundation. New York;
Macmillan; Toronto; Maxwell Macmillan Canada; New York: Maxwell
Macmillan International: pp. 16-18. (1994).
11. Forsyth, R., Expert systems: principles and case studies. London, Chapman
and Hall: pp. 8-9. (1984).
12. Zadeh, L.A., Fuzzy sets. Information and Control, 8: pp. 338-353. (1965).
13. Chen, C.H., Fuzzy logic and neural network handbook. Computer
engineering series. New York, McGraw-Hill: pp: (various pagings). vol. 1.
(1996).
14. Kruse, R., J. Gebhardt, and F. Klawonn, Foundations of fuzzy systems.
Chichester, West Sussex, England New York, Wiley & Sons: pp xii, 265.
(1994).
15. Baillie.D.C., Applications of Artificial Neural Networks for Bearing Fault
Diagnosis. Monash University, Melbourne, Australia. (PhD thesis, 1996).
16. Bow, Pattern recognition. Marcel Dekker, Inc., Now York: pp. 55-56. (1984).
17. Jang, J.-S.R., C.-T. Sun, and E. Mizutani, Neuro-fuzzy and soft computing : a
computational approach to learning and machine intelligence. MATLAB
curriculum series. Upper Saddle River, NJ, Prentice Hall: pp. xxvi, 614.
(1997).
18. Browne, A., Neural network analysis, architectures, and applications.
Philadelphia, Pa., Institute of Physics Pub.: pp. 7-9. (1997).
REFERENCES
____________________________________________________________________
176
19. El Hachemi Benbouzid, M., A review of induction motors signature analysis
as a medium for faults detection. Industrial Electronics, IEEE Transactions,
47(5): pp. 984-993. (2000).
20. Alfredson, R.J. and J. Mathew, Time domain methods for monitoring the
condition of rolling element bearings. Mechanical Engineering Transactions -
Institution of Engineers, Australia, ME 10(2): pp. 102-107. (1985).
21. McFadden, P.D. and J.D. Smith, Information from the vibration of rolling
bearings. In: Condition Monitoring '84, Proc. Int. Conf. on Condition
Monitoring, (Swansea, U.K.: Apr. 10-13, 1984), M.H. Jones (ed.), Swansea,
U.K., Pineridge Press, 1984, Section 2, Paper 8, pp.178-190. (ISBN 0-
906674-32-8). (1984).
22. McFadden, P.D. and J.D. Smith, Model for the vibration produced by a single
point defect in a rolling element bearing. Journal of Sound and Vibration,
96(1): pp. 69-82. (1984).
23. McFadden, P.D. and J.D. Smith, The vibration produced by multiple point
defects in a rolling element bearing. Journal of Sound and Vibration, 98(2):
pp. 263-273. (1985).
24. Kim, P.Y., A review of rolling element bearing health monitoring (II)
preliminary test results on current technologies. In: Proc. on Machinery
Vibration Monitoring & Analysis Meeting: pp. 127-137. (1984).
25. Kim, P.Y., Review of rolling element bearing health (III): preliminary test
results on eddy current proximity transducer technique. I Mech E Conference
Publications (Institution of Mechanical Engineers): pp. 119-125. (1984).
REFERENCES
____________________________________________________________________
177
26. Kim, P.Y. and I.R.G. Lowe, Review of rolling element bearing health
monitoring. Proceedings - Machinery Vibration Monitoring and Analysis
Seminar and Meeting: pp. 145-154. (1983).
27. Kim, Y.W., et al., Analysis and processing of shaft angular velocity signals in
rotating machinery for diagnostic applications, in Acoustics, Speech, and
Signal Processing, 1995. ICASSP-95., 1995 International Conference on,
Dept. of Mech. Eng., Ohio State Univ., Columbus, OH, USA, Theoretical or
Mathematical: pp. 2971-2974. vol.5. (1995).
28. Lebold, M., et al., Review of vibration analysis methods for gearbox
diagnostics and prognostics, in the 54th meeting of the Society for Machinery
Failure Prevention Technology, Virginia Beach,VA: pp. 623-634. (2000).
29. Tandon, N. and A. Choudhury, A review of vibration and acoustic
measurement methods for the detection of defects in rolling element bearings.
Tribology International, 32(8): pp. 469-480. (1999).
30. Chow, M.-Y., Guest editorial special section on motor fault detection and
diagnosis. Industrial Electronics, IEEE Transactions on, 47(5): pp. 982-983.
(2000).
31. Andrade, F.A., I. Esat, and M.N.M. Badi, A new approach to time-domain
vibration condition monitoring: gear tooth fatigue crack detection and
identification by the kolmogorov-smirnov test. Journal of Sound and
Vibration, 240(5): pp. 909-919. (2001).
32. Altmann, J. and J. Mathew, High Frequency Transient Analysis for the
Detection and Diagnosis of Faults in Low speed Rolling Element Bearings, in
the Asia Pacific Vibration Conference, Kyongju, Korea: pp. 730-735. (1997).
REFERENCES
____________________________________________________________________
178
33. Williams, T., et al., Rolling element bearing diagnostics in run-to-failure
lifetime testing. Mechanical Systems and Signal Processing, 15(5): pp. 979-
993. (2001).
34. Mcfadden, P.D. and M.M. Toozhy, Application of synchronous averaging to
vibration monitoring of rolling element bearings. Mechanical Systems and
Signal Processing, 14(6): pp. 891-906. (2000).
35. Samuel, P.D. and D.J. Pines, Vibration separation methodology for planetary
gear health monitoring. Proceedings of the SPIE, the International Society for
Optical Engineering, 3985: pp. 250-60. (2000).
36. Wang, W.J. and P.D. McFadden, Early detection of gear failure by vibration
analysis--ii. interpretation of the time-frequency distribution using image
processing techniques. Mechanical Systems and Signal Processing, 7(3): pp.
205-215. (1993).
37. McClintic, K., et al., Residual and difference feature analysis with transitional
gearbox data, in the 54th meeting of the society for machinery failure
prevention technology, Virginia Beach,VA: pp. 635-645. (2000).
38. Wang, W.Q., F. Ismail, and M.F. Golnaraghi, Assessment of gear damage
monitoring techniques using vibration measurements. Mechanical Systems
and Signal Processing, 15(5): pp. 905-22. (2001).
39. Wang, W., Early detection of gear tooth cracking using the resonance
demodulation technique. Mechanical Systems and Signal Processing, 15(5):
pp. 887-903. (2001).
40. Brown, D.N., Envelope analysis detects bearing faults before major damage
occurs. Pulp & Paper, 63(13): pp. 113-117. (1989).
REFERENCES
____________________________________________________________________
179
41. Wang, W. and A.K. Wong, Some new signal processing approaches for gear
fault diagnosis, in Signal Processing and Its Applications, 1999. ISSPA '99.
Proceedings of the Fifth International Symposium on, Aeronaut. & Maritime
Res. Lab., Defence Sci. & Technol. Organ., Melbourne, Vic., Australia, New
Development, Practical: pp. 587-590. vol.2. (1999).
42. Chen, Z. and C.K. Mechefske, Diagnosis of machinery fault status using
transient vibration signal parameters. JVC/Journal of Vibration and Control,
8(3): pp. 321-335.
43. Li, Z. and Y. Fu, Adaptive noise cancelling technique and bearing fault
diagnosis. Journal of Aerospace Power/Hangkong Dongli Xuebao, 5(3): pp.
199-203. (1990).
44. Shao, Y. and K. Nezu, Detection of self-aligning roller bearing fault by
asynchronous adaptive noise cancelling technology. JSME International
Journal, Series C, 42(1): pp. 33-43. (1999).
45. Logan, D., Using the correlation method for the detection of faults in rolling
element bearings. Monash University, Melbourne. (PhD thesis, 1996).
46. Wang, W.J., et al., The application of some non-linear methods in rotating
machinery fault diagnosis. Mechanical Systems and Signal Processing, 15(4):
pp. 697-705. (2001).
47. Mevel, L., L. Hermans, and H. Van Der Auweraer, Application of a
subspace-based fault detection method to industrial structures. Mechanical
Systems and Signal Processing, 13(6): pp. 823-838. (1999).
48. Nirbito, W., C.C.Tan, and J. Mathew, The enhancement of bearing signals
corrupted by noise using blind deconvolution-a feasibility study, in
REFERENCES
____________________________________________________________________
180
Proceedings of the 2nd Asia-Pacific Conference on Systems Integrating and
Maintenance, Nanjing, China: pp. 309-314. (2000).
49. Serviere, C. and P. Fabry, Blind source separation of noisy harmonic signals
for rotating machine diagnosis. Journal of Sound and Vibration, 272(1-2): pp.
317-339. (2004).
50. Lin, J. and L. Qu, Feature extraction based on morlet wavelet and its
application for mechanical fault diagnosis. Journal of Sound and Vibration,
234(1): pp. 135-148. (2000).
51. Mcfadden, P.D., Detection of gear faults by decomposition of matched
differences of vibration signals. Mechanical Systems and Signal Processing,
14(5): pp. 805-817. (2000).
52. Hambaba, A. and E. Huff, Multiresolution error detection on early fatigue
cracks in gears, in 2000 IEEE aerospace Conference Proceedings, Big Sky,
MT, USA, NASA Ames Res. Center Moffett Field CA USA, 2000 IEEE
Aerospace Conference.: pp. 566-573. (2000).
53. Li, X., et al., Fault prognosis for large rotating machinery using neural
network, Dept. of Mech. Eng. Xi'an Jiaotong Univ. China,Applications of
Artificial Intelligence in Engineering IX. Proceedings of the Ninth
International Conference. Comput. Mech. Publications Southampton UK: pp.
618-625. (1994).
54. Salami, M.J.E., A. Gani, and T. Pervez, Machine condition monitoring and
fault diagnosis using spectral analysis techniques, Fac. of Eng. Int. Islamic
Univ. Malaysia,First International Conference on Mechatronics.
Mechatronics - An Integrated Engineering for the New Millennium.
Conference Proceedings. Int. Islamic Univ. Malaysia Kuala Lumpar
Malaysia: pp. 2 vol. xiv+743. (2001).
REFERENCES
____________________________________________________________________
181
55. Zhengjia, H., et al., Wavelet Transform in Tandem with Autoregressive
Technique for Monitoring and Diagnosis of Machinery, in Condition
monitoring and diagnostic engineering management, Proceedings of
COMADEM 94. (September 26-29, 1994).
56. Liu, T.I. and J.M. Mengel, Detection of ball bearing conditions by an A.I.
approach. Sensors, Controls, and Quality Issues in Manufacturing; American
Society of Mechanical Engineers, Production Engineering Division
(Publication) PED, 55: pp. 13-21. (1991).
57. Kim, K. and A.G. Parlos, Model-based fault diagnosis of induction motors
using non-stationary signal segmentation. Mechanical Systems and Signal
Processing, 16(2-3): pp. 223-253. (2002).
58. Haykin, S.S., Adaptive filter theory. 3rd ed. Prentice-Hall information and
system sciences series. Upper Saddle River, N.J, Prentice Hall: pp. 8-10.
(1996).
59. Igarashi, T. and H. Hamada, Studies on the vibration and sound of defective
rolling bearings. (First report: vibration of ball bearings with one defect).
Bull. JSME, 25(204): pp. 994-1001. (1982).
60. Burgess, P.F.J., Antifriction bearing fault detection using envelope detection.
Transactions of the Institution of Professional Engineers New Zealand,
Electrical/Mechanical Chemical Engineering Section, 15(2/EMCh): pp. 77-
82. (1988).
61. Hong, S.Y., Forecasting minor machine failure by DDS spectrum analysis,
Dept. of Mech. Eng. Wright State Univ. Dayton OH USA, Intelligent
Manufacturing Systems 1994 (IMS`94). A Postprint Volume from the IFAC
Workshop. Pergamon Oxford UK: pp. ix+534. (1994).
REFERENCES
____________________________________________________________________
182
62. Xu, F., Z. Wang, and S. Fu, Research of a data acquisition system of
vibration signal for large-scale rotating machine. Qinghua Daxue
Xuebao/Journal of Tsinghua University, 38(6): pp. 111-114. (1998).
63. Meesad, P. and G.G. Yen, Pattern classification by a neurofuzzy network:
application to vibration monitoring. ISA Transactions, 39(3): pp. 293-308.
(2000).
64. McFadden, P.D. and J.D. Smith, Vibration monitoring of rolling element
bearings by the high-frequency resonance technique-a review. Tribology
International, 17(1): pp. 3-10. (1984).
65. Yu, D., J. Cheng, and Y. Yang, Application of EMD method and Hilbert
spectrum to the fault diagnosis of roller bearings. Mechanical Systems and
Signal Processing, 19(2): pp. 259-270. (2005).
66. Lu, Q. and D. Li, Neural network method for diagnosing faults rolling
bearing in electrical machines with frequency signatures. Qinghua Daxue
Xuebao/Journal of Tsinghua University, 38(4): pp. 94-97. (1998).
67. Baydar, N. and A. Ball, Detection of gear deterioration under varying load
conditions by using the instantaneous power spectrum. Mechanical Systems
and Signal Processing, 14(6): pp. 907-921. (2000).
68. Yang, D.-M., et al., Third-order spectral techniques for the diagnosis of motor
bearing condition usign artificial neural networks. Mechanical Systems and
Signal Processing, 16(2-3): pp. 391-411. (2002).
69. Li, C.J., J. Ma, and B. Hwang, Bearing condition monitoring by pattern
recognition based on bicoherence analysis of vibrations. Proceedings of the
REFERENCES
____________________________________________________________________
183
Institution of Mechanical Engineers, Part C: Journal of Mechanical
Engineering Science, 210(3): pp. 277-285. (1996).
70. Capdessus, C., M. Sidahmed, and J.L. Lacoume, Cyclostationary processes:
application in gear faults early diagnosis. Mechanical Systems and Signal
Processing, 14(3): pp. 371-385. (2000).
71. Bouillaut, L. and M. Sidahmed, Cyclostationary approach and bilinear
approach: Comparison, applications to early diagnosis for helicopter gearbox
and classification method based on hocs. Mechanical Systems and Signal
Processing, 15(5): pp. 923-943. (2001).
72. Mathew, J. and R.J. Alfredson, Condition monitoring of rolling element
bearings using vibration analysis. Journal of Vibration, Acoustics, Stress, and
Reliability in Design, 106(3): pp. 447-453. (1984).
73. Pan, M.-C., P. Sas, and H. van Brussel, Nonstationary time-frequency
analysis for machine condition monitoring, in Time-Frequency and Time-
Scale Analysis, 1996., Proceedings of the IEEE-SP International Symposium
on, Dept. of Mech. Eng., Katholieke Univ., Leuven, Belgium, Theoretical or
Mathematical,Experimental: pp. 477-480. (1996).
74. Klein, R., D. Ingman, and S. Braun, Non-stationary signals: phase-energy
approach-theory and simulations. Mechanical Systems and Signal Processing,
15(6): pp. 1061-1089. (2001).
75. In Soo, K. and K. Whan Woo, The development of reactor coolant pump
vibration monitoring and a diagnostic system in the nuclear power plant. ISA
Transactions, 39(3): pp. 309-16. (2000).
REFERENCES
____________________________________________________________________
184
76. Lee, C.K., et al., A study on RCP vibration monitoring and diagnostics in
NPP,Proceedings of 3rd Asia-Pacific Conference on Control and
Measurement (APCCM'98), Dunhuang, China, Korea Atomic Energy Res.
Inst. Taejon South Korea,Proceedings of the 3rd Asia-Pacific Conference on
Control and Measurement. China Aviation Ind. Press Beijing China: pp.
vi+425. (1998).
77. Baydar, N. and A. Ball, A comparative study of acoustic and vibration signals
in detection of gear failures using Wigner-Ville distribution. Mechanical
Systems and Signal Processing, 15(6): pp. 1091-1107. (2001).
78. Wang, W.J. and P.D. McFadden, Early detection of gear failure by vibration
analysis i. calculation of the time-frequency distribution. Mechanical Systems
and Signal Processing, 7(3): pp. 193-203. (1993).
79. Wen Yi, W. and M.J. Harrap, Condition monitoring of rolling element
bearings by using cone kernel time-frequency distribution. Proceedings of the
SPIE The International Society for Optical Engineering, 2101(1): pp. 290-8.
(1993).
80. Lee, S.u., D. Robb, and C. Besant, The directional choi-willams distribution
for the analysis of rotor-vibration signals. Mechanical Systems and Signal
Processing, 15(4): pp. 789-811.(2001).
81. Han, Y.-S. and C.-W. Lee, Directional Wigner distribution for order analysis
in rotating/reciprocating machines. Mechanical Systems and Signal
Processing, 13(5): pp. 723-738. (1999).
82. Lee, S.K. and P.R. White, Higher-order time-frequency analysis and its
application to fault detection in rotating machinery. Mechanical Systems &
Signal Processing, 11(4): pp. 637-650. (1997).
REFERENCES
____________________________________________________________________
185
83. Chin-Hsing Chen, Jiann-Der Lee, and Ming-Chi Lin, Classification of
underwater signals using wavelet transforms and neural networks.
Mathematical and Computer Modelling, 27(2): pp. 47-60. (1998).
84. Liu, B., S.F. Ling, and R. Gribonval, Bearing failure detection using
matching pursuit. NDT & E International, 35(4): pp. 255-262. (2002).
85. Lopez, J.E., R.R. Tenney, and J.C. Deckert, Fault detection and identification
using real-time wavelet feature extraction, in Proceedings of IEEE-SP
International Symposium on Time- Frequency and Time-Scale Analysis,
Philadelphia, PA, USA, IEEE New York NY USA: pp. 217-221. (1994).
86. Tse, P.W., W.-x. Yang, and H.Y. Tam, Machine fault diagnosis through an
effective exact wavelet analysis. Journal of Sound and Vibration, 277(4-
5):1005-1024. (2004).
87. Samuel, P. and D. Pines, Health monitoring/damage detection of a rotorcraft
planetary geartrain system using piezoelectric sensors. Proceedings of the
SPIE The International Society for Optical Engineering, 3041: pp. 44-53.
(1997).
88. Mori, K., et al., Prediction of spalling on a ball bearing by applying the
discrete wavelet transform to vibration signals. Wear, 195(1-2): pp. 162-168.
(1996).
89. Wickerhauser, M.V., Adapted wavelet analysis from theory to software.
Wellesley, MA, A.K. Peters: p. xii, 486. (1994).
90. Liu, S. and W. Shi, Rough set based intelligence diagnostic system for valves
in reciprocating pumps, in Systems, Man, and Cybernetics, 2001 IEEE
REFERENCES
____________________________________________________________________
186
International Conference on, Dept. of Mech. Eng., Daqing Pet. Inst.,
Heilongjiang, China, Practical: pp. 353-358. vol.1. (2001).
91. Nikolaou, N.G. and I.A. Antoniadis, Rolling element bearing fault diagnosis
using wavelet packets. NDT & E International, 35(3): pp. 197-205. (2002).
92. Shi, W., R. Wang, and W. Huang, Application of rough set theory to fault
diagnosis of check valves in reciprocating pumps,Canadian Conference on
Electrical and Computer Engineering 2001, Toronto, Ont., Canada, Dept. of
Astronaut. & Mech. Harbin Inst. of Technol. China,Canadian Conference on
Electrical and Computer Engineering 2001. Conference Proceedings (Cat.
No.01TH8555). IEEE Piscataway NJ USA: pp. 2 vol.1414. (2001).
93. Shulin, L. and S. Wengang, Rough set based intelligence diagnostic system
for valves in reciprocating pumps,Proceedings of IEEE International
Conference on Systems, Man & Cybernetics, Tucson, AZ, USA, Dept. of
Mech. Eng. Daqing Pet. Inst. Heilongjiang China,2001 IEEE International
Conference on Systems Man and Cybernetics. e-Systems and e-Man for
Cybernetics in Cyberspace (Cat.No.01CH37236). IEEE Piscataway NJ USA:
pp. 3494. vol. 5. (2001).
94. Goumas, S., et al., Intelligent on-line quality control of washing machines
using discrete wavelet analysis features and likelihood classification.
Engineering Applications of Artificial Intelligence, 14(5): pp. 655-666.
(2001).
95. Akan, A. and L.F. Chaparro, Evolutionary chirp representation of non-
stationary signals via Gabor transform. Signal Processing, 81(11): pp. 2429-
2436. (2001).
96. Ferrando, S.E., et al., Probablistic matching pursuit with Gabor dictionaries.
Signal Processing, 80. (2000).
REFERENCES
____________________________________________________________________
187
97. Mallat, S.G. and Z. Zhang, Matching pursuits with time-frequency
dictionaries. IEEE Transactions on Signal Processing, 41(12): pp. 3397-3415.
(1993).
98. Lin, J., Feature extraction of machine sound using wavelet and its application
in fault diagnosis. NDT & E International, 34(1): pp. 25-30. (2001).
99. Zheng, H., Z. Li, and X. Chen, Gear fault diagnosis based on continuous
wavelet transform. Mechanical Systems and Signal Processing, 16(2-3): pp.
447-457. (2002).
100. Lin, J. and M.J. Zuo, Gearbox fault diagnosis using adaptive wavelet filter.
Mechanical Systems and Signal Processing, 17(6): pp. 1259-1269. (2003).
101. Zhang, J. and Z. Bao, Initialization of orthogonal discrete wavelet transforms.
IEEE Transactions on Signal Processing, 48(5): pp. 1474-1477. (2000).
102. Mallat, S., a wavelet tour of signal processing. Paris, Courant Institute, New
York University: pp. 242-244. (1999).
103. Wang, W.J., Wavelets for detecting mechanical faults with high sensitivity.
Mechanical Systems and Signal Processing, 15(4): pp. 685-696. (2001).
104. Tse, P.W. and W.X. Yang, Shortcoming of Wavelet Transforms in Machine
fault diagnosis and the proposed solution, in the Proceedings of the 3rd Asia-
Pacific Conference on System Integrity and Maintenance (ACSIM, 2002): pp.
357-362. ISBN: 1 86435 589 1.
REFERENCES
____________________________________________________________________
188
105. Zhong, B., Developments in intelligent condition monitoring and diagnostics,
in System Integrity and Maintenance , the 2nd Asia-Pacific
Conference(ACSIM2000), Brisbane, Australia: pp. 1-7. (2000).
106. Pham, D.T. and P.T.N. Pham., Artificial intelligence in engineering.
International Journal of Machine Tools & Manufacture, (39): pp. 937-949.
(1999).
107. Gao, X.Z. and S.J. Ovaska, Soft computing methods in motor fault diagnosis.
Applied Soft Computing, 1(1): pp. 73-81. (2001).
108. Parlos, A.G., S.K. Menon, and A.F. Atiya, Adaptive state estimation using
dynamic recurrent neural networks, in Neural Networks, 1999. IJCNN '99.
International Joint Conference on, Dept. of Nucl. Eng., Texas A&M Univ.,
College Station, TX, USA, Application Theoretical or Mathematical: pp.
3361-3364 vol.5. (1999).
109. Priddy, K.L., M.D. Lothers, and R.E. Saeks, Neural networks and fault
diagnosis in rotating machinery, in Systems, Man and Cybernetics, 1993.
'Systems Engineering in the Service of Humans', Conference Proceedings.,
International Conference on, Accurate Autom. Corp., Chattanooga, TN, USA,
Practical: pp. 640-644 vol.2. (1993).
110. Engin, S.N. and K. Gulez, A wavelet transform-artificial neural networks
(WT-ANN) based rotating machinery fault diagnostics
methodology,Proceedings of the IEEE-EURASIP Workshop on Nonlinear
Signal and Image Processing (NSIP'99), Antalaya, Turkey, Dept. of Electr.
Eng. Yildiz Tech. Univ. Istanbul Turkey,Proceedings of the IEEE-EURASIP
Workshop on Nonlinear Signal and Image Processing (NSIP'99). Bogazici
Univ Instanbul Turkey: pp. 2 vol. xxiii+894. (1999).
REFERENCES
____________________________________________________________________
189
111. Zhao, L. and Z. Sheng, Combination of discrete cosine transform with neural
network in fault diagnosis for rotating machinery. Proceedings of the IEEE
International Conference on Industrial Technology: pp. 450-454. (1996).
112. Paya, B.A., I.I. Esat, and M.N.M. Badi, Artificial neural network based fault
diagnostics of rotating machinery using wavelet transforms as a preprocessor.
Mechanical Systems & Signal Processing, 11(5): pp. 751-765. (1997).
113. McCormick, A.C. and A.K. Nandi, A comparison of artificial neural
networks and other statistical methods for rotating machine condition
classification,IEE Colloquium on Modelling and Signal Processing for Fault
Diagnosis (Ref, Leicester, UK, Dept. of Electron. & Electr. Eng. Strathclyde
Univ. Glasgow UK,IEE Colloquium on Modelling and Signal Processing for
Fault Diagnosis (Ref. No.1996/260). IEE London UK: p. 100. (1996).
114. Lowes, S. and J.M. Shippen, A diagnostic system for industrial fans.
Measurement and Control, 30(1): pp. 9-13. (1997).
115. Tanaka, M., et al., Application of Kohonen's self-organizing network to the
diagnosis system for rotating machinery, Fac. of Eng. Hiroshima Univ.
Japan,1995 IEEE International Conference on Systems Man and Cybernetics.
Intelligent Systems for the 21st Century (Cat. No.95CH3576-7). IEEE New
York NY USA: pp. 5 vol. 4711. (1995).
116. Deschenes, C.J. and J. Noonan, Fuzzy Kohonen Network for the
Classification of Transients Using the Wavelet Transform for Feature
Extraction. Information Sciences, 87(4): pp. 247-266. (1995).
117. Hoffman, A.J. and N.T. van der Merwe, The application of neural networks
to vibrational diagnostics for multiple fault conditions. Computer Standards
& Interfaces, 24(2): pp. 139-149. (2002).
REFERENCES
____________________________________________________________________
190
118. Melvin, D.G. and J. Penman, Fusing human knowledge with neural networks
in machine condition monitoring systems. Proceedings of the SPIE The
International Society for Optical Engineering, 2492(1): pp. 276-83. (1995).
119. Suzuki, M., et al., Application of neural network to failure diagnosis.
Research Reports of Kogakuin University. no., 81: pp. 33-8. (1996).
120. Taylor, O. and J. Macintyre, Modified Kohonen network for data fusion and
novelty detection within condition monitoring,Proceedings of EuroFusion
98., Great Malvern, UK, Sch. of Comput. & Inf. Syst. Sunderland Univ. UK,
Proceedings of EuroFusio 98. International Data Fusion Conference. DERA
Malvern UK.: pp. vi+228. (1998).
121. Chan, C. W., et al., Fault detection of systems with redundant sensors using
constrained Kohonen networks. Automatica, 37(10): pp. 1671-1676. (2001).
122. Hu, T., B.C. Lu, and G.J. Chen, A Rotary Machinery Fault Diagnosis
Approach Based on Rough Set Theory, in the 3rd World Congress on
Intelligent Control and Automation, Hefei, China: pp. 589-685. (2000).
123. Jeffries, M., et al., A fuzzy approach to the condition monitoring of a
packaging plant. Journal of Materials Processing Technology, 109: pp. 83-89.
(2001).
124. De Miguel, L.J., J. Fernandez, and J.R. Peran, Applying fuzzy logic to
rotating machinery diagnosis, Dept. of Syst. Eng. & Control Valladolid Univ.
Spain. Methodologies for the Conception Design and Application of
Intelligent Systems. Proceedings of the 4th International Conference on Soft
Computing. World Scientific Singapore: pp. 2 vol. xlii+974. (1996).
125. Milne, R., Amethyst, Intelligent Applications Ltd. Kirkton Bus. Centre
Livingston UK,IEE Colloquium on 'Intelligent Fault Diagnosis - Part 1:
REFERENCES
____________________________________________________________________
191
Classification-Based Techniques' (Digest No.045). IEE London UK: p. 32.
(1992).
126. El Adawi, S., et al., Computer based expert system for rotating machinery
(preventive and predictive maintenance), Dept. of Mech. Power Zagazig
Univ. Egypt,Proceedings of the Second IASTED International Conference.
Computer Applications in Industry. ACTA Press Zurich Switzerland: pp. 2
vol.vii+585. (1992).
127. Georgin, E., et al., The importance of cases and domain models in
explanation, Proceedings of ISAP '94, Montpellier, France, Centre for Electr.
Power Eng. Strathclyde Univ. Glasgow UK,ISAP '94. International
Conference on Intelligent System Application to Power Systems. EC2
Nanterre Cedex France: p. 2 vol. 894. (1994).
128. Shao, Y. and K. Nezu, An online monitoring and diagnostic method of rolling
element bearing with AI. Transactions of the Society of Instrument and
Control Engineers, 32(8): pp. 1287-93. (1996).
129. Vilim, R.B., H.E. Garcia, and F.W. Chen, Machine condition monitoring
using neural networks and the likelihood function. Intelligent Engineering
Systems Through Artificial Neural Networks, 7: pp. 653-659. (1997).
130. Emmanouilidis, C., J. MacIntyre, and C. Cox, An integrated, soft computing
approach for machine condition diagnosis, Sch. of Comput. & Inf. Syst.
Sunderland Univ. UK,6th European Congress on Intelligent Techniques and
Soft Computing. EUFIT '98. Verlag Mainz Aachen Germany: pp. 3 vol.
xxvi+2010. (1998).
131. Hao, L. and X. Xu, The application of rough set neural network system in
fault diagnosis. Control Theory & Applications, 18(5): pp. 681-5. (2001).
REFERENCES
____________________________________________________________________
192
132. Kesheng, W. and L. Bing, Using B-spline neural network to extract fuzzy
rules for a centrifugal pump monitoring. Journal of Intelligent Manufacturing,
12(1): pp. 5-11. (2001).
133. Satoh, S., M.S. Shaikh, and Y. Dote, Fast fuzzy neural network for fault
diagnosis of rotational machine parts using general parameter learning and
adaptation,SMCia/01, Blacksburg, VA, USA, Dept. of Comput. Sci. & Syst.
Eng. Muroran Inst. of Technol. Japan,SMCia/01. Proceedings of the 2001
IEEE Mountain Workshop on Soft Computing in Industrial Applications
(Cat. No.01EX504). IEEE Piscataway NJ USA: pp. xv+134. (2001).
134. Satoh, S., M.S. Shaikh, and Y. Dote, Fault diagnosis for dynamical systems
using soft computing,Proceedings of IEEE International Conference on
Systems, Man & Cybernetics, Tucson, AZ, USA, Dept. of Comput. Sci. &
Syst. Eng. Muroran Inst. of Technol. Japan: pp. 3494. vol. 5. (2001).
135. Yakuwa, F., et al., Fault diagnosis for dynamical systems using soft
computing, in Fuzzy Systems, 2002. FUZZ-IEEE'02. Proceedings of the 2002
IEEE International Conference on, Muroran Institute of Technology: pp. 261-
266. (2002).
136. Taniguchi, S., D. Akhmetov, and Y. Dote, Fault detection of rotating machine
parts using novel fuzzy neural network. Proceedings of the IEEE
International Conference on Systems, Man and Cybernetics, 1: pp. I-365 - I-
369. (1999).
137. Taniguchi, S., et al., Nonlinear modeling and fault detection using fuzzy-
neural network, Dept. of Comput. Sci. & Syst. Eng. Muroran Inst. of
Technol. Japan,Proceedings of the ISCA 9th International Conference
Intelligent Systems. Int. Soc. Comput. & Their Appl. - ISCA Cary NC USA:
pp. iv+158. (2000).
REFERENCES
____________________________________________________________________
193
138. Hsiao, I.-L., et al., On-line fault diagnosis of rotor vibration by using signal-
based feature generation and neural fuzzy inference. Journal of the Chinese
Society of Mechanical Engineers, Transactions of the Chinese Institute of
Engineers, Series C/Chung-Kuo Chi Hsueh Kung Ch'eng Hsuebo Pao, 20(4):
pp. 345-352. (1999).
139. Altmann, J. and J. Mathew, Multiple Band-Pass Based Automatic
Classification of Low Speed Rolling-element Bearing Faults, in ACSIM.,
Nanjing China, Queensland University of Technology, Brisbane, Australia.
(2000).
140. Feng, E., H. Yang, and M. Rao, Fuzzy expert system for real-time process
condition monitoring and incident prevention. Expert Systems with
Applications, 15(3-4): pp. 383-390. (1998).
141. Liu, T., J. Shigonahalli, and N. Iyer, Detection of Roller Bearing Defects
using Expert systems and Fuzzy logic. Mechanical Systems and Signal
Processing, 10(5): pp. 595-614. (1996).
142. Siu, C., Q. Shen, and R. Milne, A fuzzy expert system for vibration cause
identification in rotating machines. IEEE. Vol (1): pp. 555-560. (1997).
143. Kawabe, Y., et al., Diagnosis method of centrifugal pumps by rough sets and
partially-linearized neural network.1997 IEEE International Conference on
Intelligent Processing Systems, Beijing, China, Mitubishi Chem. Corp.
Kitakyushu Japan. 1997 IEEE International Conference on Intelligent
Processing Systems (Cat. No.97TH8335). IEEE New York NY USA: p. 2
vol. xxviii+1893. (1997).
144. Jack, L.B. and A.K. Nandi, Fault detection using support vector machines and
artificial neural networks, augmented by genetic algorithm. Mechanical
Systems and Signal Processing, 16(2-3): pp. 373-390. (2002).
REFERENCES
____________________________________________________________________
194
145. Jack, L.B. and A.K. Nandi, Genetic algorithms for feature selection in
machine condition monitoring with vibration signals. IEE Proceedings:
Vision, Image and Signal Processing, 147(3): pp. 205-212. (2000).
146. Wang, K. and B. Lei, Genetic algorithms for constructing feed forward
multiple layered neural network in a centrifugal pump condition monitoring.
Intelligent Engineering Systems Through Artificial Neural Networks, 8: pp.
303-310. (1998).
147. Staszewski, W.J. and K. Worden, Classification of faults in gearboxes. Pre-
processing algorithms and neural networks. Neural Computing &
Applications, 5(3): pp. 160-83. (1997).
148. Lou, X. and K.A. Loparo, Bearing fault diagnosis based on wavelet transform
and fuzzy inference. Mechanical Systems and Signal Processing, 18(5): pp.
1077-1095. (2004).
149. Altmann, J. and J. Mathew, DWPA best basis demodulation for the detection
and diagnosis of faults in rolling element bearings, in Proceedings of the 1st
Australian Conference on Systems Integrity and Maintenance, Surfers
Paradise: pp. 331-339. (1997).
150. Changsheng, H., et al., System analysis and design based on MATLAB-
Wavelet Analysis. Xian, Xian Electronices Science & Technology
University: pp. 124-135. (1999).
151. Chen, S.S., D.L. Donoho, and M.A. Saunders, Atomic Decomposition by
Basis Pursuit. SIAM REVIEW, 43(1): pp. 129-159. (2001).
152. http://www.gc.ssr.upm.es/inves/neural/ann1/concepts/basis.htm
PUBLICATIONS
____________________________________________________________________
195
PUBLICATIONS
Accepted for Publications
1. Yang, H., J. Mathew, and L. Ma. Intelligent diagnosis of rotating machinery
faults- a review, Proceedings of the 3rd Asia-Pacific Conference on System Integrity
and Maintenance (ACSIM): pp. 385-392. ISBN: 1 86435 589 1. (2002).
2. Yang, H., J. Mathew, and L. Ma. Vibration Feature Extraction for Diagnosis of
Rotating Machinery Faults-A Literature Survey, Proceedings of the 10th Asia-Pacific
Vibration Conference, (APVC): pp. 801--807. ISBN: 06464 42853. (2003).
3. Yang, H., J. Mathew, and L. Ma. Time Frequency Techniques for Fault Diagnosis
of Rolling Element Bearings, Proceeding of 10th Asia-Pacific Vibration Conference
(APVC): pp. 789-794. ISBN: 06464 42853. (2003).
4. Yang, H., J. Mathew, and L. Ma. Feature extraction of faulty bearings vibration
via basis pursuit, Proceeding of the 10th Asia-Pacific Vibration Conference,
(APVC): pp. 789-794. ISBN: 06464 42853. (2003).
5. Yang, H., J. Mathew, and L. Ma. Bearing fault classification using wavelet
features, Proceeding of Intelligence Maintenance System (IMS) International
Conference. Arles, France: Section 1-D. (2004).
6. Yang, H., J. Mathew, L. Ma, and V. Kosse. Matching Pursuit features based
Neural Network pattern recognition of rolling bearing faults, Proceeding of
International Conference of Maintenance Societies. Sydney, Australia: Paper.74:
pp.1-8. (2004).
7. Yang, H., J. Mathew, and L. Ma. Basis Pursuit feature based pattern recognition of
rolling bearing faults, Proceeding of Intelligence Maintenance System (IMS)
International Conference. Arles, France: Section 2-C. (2004).
8. Hongyu Yang, Joseph Mathew, Lin Ma, Fault diagnosis of rolling element
bearings using basis pursuit, The Journal of Mechanical Systems and Signal
Processing, Volume 19 (2): pp. 341-356. (2005)
GLOASSARY
____________________________________________________________________
196
GLOSSARY
Discrete Wavelet Transformation
The Discrete Wavelet Transform (DWT), in particular, decomposes the signal into
ortho-normal “wavelets”, scaled and shifted versions of the “mother wavelet”, ψ. A
function f(t) can be expressed by its wavelet expansion, defined as follows.
10),2()(0
12
020 <≤−+= ∑∑
∞
=
−
=+ tforktatf j
j k
j
kj ψα
The integer j describes the different levels of wavelets, and k covers the number of
wavelets in each level.
Discrete Wavelet Packet Analysis (DWPA)
DWPA can decompose signals into both low frequency components and high
frequency components.
Power spectrum
Power spectrum is the square of amplitude of spectrum.
Cepstrum
Cepstrum is the logarithm of power spectrum.
Wigner distribution
( ) ∫+∞
∞−
−
−
+= τττ τπ detxtxftWD fjx
2*
22,
GLOASSARY
____________________________________________________________________
197
The Choi-Williams distribution is defined:
( ) ( ) ( )[ ] τττστπ
ω τωστ duduxuxetP jtuCW
+
−= −−−∫∫ 22
1
4
1,
22 4/
22/3
where σ is constant.
Scalogram – the squared modulus of the CWT.
Spectrogram is defined as the square of STFT (has been applied in radar, not found
in fault diagnosis yet)
( ) ( ) ( ) ( )2
22,, ∫
∞+
∞−
−−== τττ τπ detwxftSTFTftSPEC fjxx
Directional Choi-Williams distribution (dCWD) , is to account for complex-valued
time-varying signals.
Recurrent Networks have the same characteristics as the standard Feed Forward,
but with feedback connections.
A Self Organising Map (SOM) Neural Network defines a mapping from input
signal of arbitrary dimension to a one or two dimensional array of nodes.
Adaptive approximation
The goal of adaptive approximation is to find the representation of a signal x as a
weighted sum of elements γφ from an over complete dictionaryΓ
γγ
γ φα∑Γ∈
=x
Or an approximate decomposition
GLOASSARY
____________________________________________________________________
198
( )mm
i
Rxii
+=∑=
γγ φα1
Where γ is an index of a set,Γ ; γα is the coefficient of the elementγφ ; m is the order
of decomposition; and( )mR is a residual.
A dictionary Γ is defined as a collection of parameterized
waveforms γφ ,{ }Γ∈γφ γ | . The waveforms γφ are discrete-time signals of length n
called atoms.
Matching Pursuit uses a specific criterion to search and decide the atoms and their
coefficients in the adaptive approximation. The Matching Pursuit identifies the
dictionary atom that best correlates with the residual and then adds to the current
approximation a scalar multiple of the atom.
Basis Pursuit represents signals in over complete dictionaries by convex
optimization. It obtains the decomposition that minimizes the 1l norm of the
coefficients occurring in the representation.
APPENDIX
____________________________________________________________________
199
APPENDIX
Activation functions
(1) Linear function (see Figure 1)
kxxg =)( (1)
Where k is a constant multiplied by the input x to form a linear function.
g(x)
-1
0
1
-1 0 1x
Figure 1: Identity function
(2) Threshold or heaviside function (see Figure 2)
A threshold function or Heaviside function is limited to one of the two values:
<≥
=) if(
) if(
0
1)(
θθ
x
xxg (2)
This kind of function is often used in single layer networks. Figure 2 illuminates the
function in the case of 1=θ .
g(x)
0
1
-1 0 1 2 3x
Figure 2: Threshold function
(3) Bipolar sigmoid function (see Figure 3)
x
x
e
exg −
−
+−=
1
1)( (3)
APPENDIX
____________________________________________________________________
200
This function has similar properties with the sigmoid function. It works well for the
applications that yield output values in the range of [-1,1].
g(x)
-1
0
1
-6 -4 -2 0 2 4 6 x
Figure 3: Bipolar sigmoid function
Characteristics of Flaw Induced Bearing Vibration
Depending on which rolling surface has the irregularities the bearing will excite
vibration at the following well-known defect frequencies:
Cage rotating frequency:
−= αcos1
2
1
c
r
d
dRPMFTF
Where rd is the diameter of the rolling elements.
( )inoutc ddd +=2
1is the diameter of the cage.
outd is the diameter of the outer race
ind is the diameter of the inner race
α is the contact angle between the rolling elements and rolling surfaces and RPM/60
is the shaft rotating frequency (expressed in Hz).
Rotational frequency of the rolling elements (BSF):
−= αcos1
2
12
c
r
r
c
d
d
d
dRPMBSF
APPENDIX
____________________________________________________________________
201
Ball-pass frequency on the outer race (BPFO):
zd
dRPMBPFO
c
r
−= αcos1
2
1
where: z is the number of rolling elements.
Ball-pass frequency on the inner race (BPFI):
zd
dRPMBPFI
c
r
+= αcos1
2
1
Very often, especially when the load is variable, vibration at other frequencies are
excited in the bearing. These frequencies are the harmonics and sum and difference
combinations of the preceding frequencies.