Condition monitoring and fault diagnosis of induction machine using artificial intelligence methods and empirical mode decomposition

ACKNOWLEDGEMENT
I would like to convey my most sincere thanks to my project supervisors, Prof Xu
Jianxin and A/Prof Sanjib Kumar Panda for their encouragement and advices. Their
profound knowledge and experiences in the field of machine learning techniques and
machine system are my source of inspiration. I also appreciate the opportunity to work in
the field of Empirical Mode Decomposition, which is the most versatile and powerful
signal processing tool I have read.

I would like to thank Mr Woo Ying Chee and Mr Mukaya Chandra from the
Electrical Machine and Drives Laboratory, who helped in setting up the Machine Fault
Simulator, DAQ measurement systems and preparing workstations for MATLAB
simulations used in the project, and their help in booking meeting room for project
briefing.

Lastly, I would like to thank Dr N. Rehman and Dr D.P. Mandic for making the
Noise-Assisted Multi-variate Empirical Mode Decomposition MATLAB code publicly
available at and Dr G.
Rilling and Dr P. Flandrin for making the mono-variate Empirical Mode Decomposition
MATLAB code publicly available at />
i


TABLE OF CONTENTS
CHAPTER 1: INTRODUCTION ................................................................................................. 1
1.1

Objectives ............................................................................................................. 2

1.2

Thesis Organization.............................................................................................. 3

1.3

Fault Mode Statistics Survey................................................................................ 6

1.4

Literature Survey .................................................................................................. 8

1.4.1

Vibration Signature Analysis ........................................................................ 8

1.4.2

Motor Current Signature Analysis ................................................................ 9

CHAPTER 2: MECHANICS OF MACHINE FAULT MODES ...................................................... 14
2.1

Eccentricity......................................................................................................... 15

2.1.1

Static Eccentricity ....................................................................................... 15

2.1.2

Dynamic Eccentricity.................................................................................. 16

2.1.3

Mixed Eccentricity Motor Current Signature ............................................. 17

2.2

Unbalanced Rotor Fault ..................................................................................... 18

2.2.1
2.3

Unbalanced Rotor Fault Motor Current Signature...................................... 20

Bearing Faults .................................................................................................... 20

2.3.1

Bearing Faults Vibration Signatures ........................................................... 22

2.3.2

Bearing Faults Motor Current Signatures ................................................... 25

2.4

Bearing General Roughness ............................................................................... 26

2.5

Broken Rotor Bar Motor Current Signature ....................................................... 27

2.6

Shorted Stator Winding Fault Motor Current Signature .................................... 28

2.7

Healthy Machine Signature ................................................................................ 29

2.8

Dynamic Estimation of the Machine Slip .......................................................... 30

CHAPTER 3: MOTOR CURRENT SIGNATURE AND VIBRATION SIGNATURE ANALYSIS ......... 32
3.1

Motor Current signal Analysis or Vibration Analysis? ...................................... 32

3.2

Challenges of Motor Current Signature Analysis .............................................. 33

3.3

Discussions: Proposed Ensemble Spectrum Approach ...................................... 35

CHAPTER 4: ARTIFICIAL INTELLIGENCE TECHNIQUES FOR MACHINE FAULT DIAGNOSIS .... 37
4.1

k-Nearest Neighbour (k-NN) ............................................................................. 37

4.1.1
4.2

k-NN Algorithm .......................................................................................... 38

Self-Organizing Map (SOM) ............................................................................. 38
ii


4.2.1

Structure and Operation of SOM ................................................................ 39

4.2.2

SOM Algorithm .......................................................................................... 41

4.3

Support Vector Machine (SVM) ........................................................................ 42

4.3.1

Multi-Class SVM (M-SVM) ....................................................................... 44

4.3.2

M-SVM Algorithm ..................................................................................... 45

4.4

Empirical Mode Decomposition (EMD) ............................................................ 46

4.4.1

EMD Algorithm .......................................................................................... 47

4.4.2

Mode Mis-alignment ................................................................................... 48

4.4.3

Multi-variante EMD.................................................................................... 57

4.4.4

Mode Mixing .............................................................................................. 57

4.4.5

Noise-assisted Multi-variate EMD (N-A M-EMD) .................................... 57

CHAPTER 5: A STUDY ON AUTOMATIC DIAGNOSIS OF BEARING AND UNBALANCED ROTOR
FAULTS .............................................................................................................................. 59
5.1

Fault Diagnosis using Time-Domain Vibration Signatures ............................... 59

5.1.1

Similarity measures by Cross-Correlation Operator ................................... 60

5.2

Machine Fault Simulator .................................................................................... 66

5.3

Machine Signatures Collection .......................................................................... 69

5.4

Experimental Results.......................................................................................... 70

5.5

Discussions: Difficulty in choosing a suitable value for k ................................. 71

5.6

Discussions: Larger Training Samples ............................................................... 72

5.7

Visualization of Classification Results by k-NN ............................................... 72

5.8

Fault Diagnosis using Frequency-Domain Vibration Signatures ....................... 75

5.8.1

Discrete Wiener Filter ................................................................................. 76

5.8.2

Frequency Analysis of Vibration Signatures .............................................. 78

5.8.2.1 Frequency content of Healthy Machine .................................................... 78
5.8.2.2 Frequency content of Unbalanced Rotor fault .......................................... 78
5.8.2.3 Frequency content of Bearing fault ........................................................... 78
5.8.2.4 Discussion on vibration frequency analysis .............................................. 78
5.8.3

Feature Extraction of Frequency domain information ................................ 83

5.8.4

Cluster Analysis of Vibration Feature Vectors ........................................... 84

5.8.5

Further Feature Extraction .......................................................................... 85

5.8.6

Multi-class SVM (M-SVM) for Classifying Machine Fault Data .............. 86
iii

5.9

Discussions: Frequency-domain Analysis of Vibration Signatures ................... 88

CHAPTER 6: A STUDY ON MOTOR CURRENT SIGNATURE USING EMPIRICAL MODE
DECOMPOSITION ................................................................................................................ 89
6.1

Fourier Transform .............................................................................................. 89

6.2

Wavelet Transform ............................................................................................. 91

6.3

Hilbert-Huang Transform................................................................................... 92

6.3.1

Hilbert Spectrum ......................................................................................... 93

6.3.2

Marginal Hilbert Spectrum ......................................................................... 94

6.4

Discussion: EMD as a suitable Analysis Tool ................................................... 94

6.5

N-A M-EMD Experiment Results...................................................................... 95

6.5.1

Discussions: IMF Derived by EMD ........................................................... 97

6.5.2

Discussions: Filter-bank Property of EMD Algorithm ............................... 99

6.5.3

Discussions: Significance of IMF1, IMF2, IMF3, IMF4......................... 101

6.5.4

Discussions: Significance of IMF10, IMF11 ........................................... 103

6.5.5

Discussions: Significance of IMF5, IMF6, IMF7, IMF8, IMF9 ............. 104

6.6

Visualization of the Comparison results by SOM ............................................ 107

6.7

Discussions: Discovery of Unique Features by SOM ..................................... 107

CHAPTER 7: CONCLUSION ................................................................................................ 111
REFERENCES .................................................................................................................... 113
APPENDIX A: INTRINSIC MODE FUNCTIONS DERIVED BY N-A MEMD ALGORITHM FOR
MACHINE SPEED AT 20HZ ................................................................................................. 124
APPENDIX B: HILBERT SPECTRUM AND MARGINAL HILBERT SPECTRUM OF MACHINE
SIGNATURE (AT MACHINE SPEED OF 20HZ) INTRINSIC MODE FUNCTION 5 TO 9 .............. 131
APPENDIX C: PSEUDO CODE FOR 2-CLASS SVM LEARNING ............................................. 135
APPENDIX D: PSEUDO CODE FOR MULTI-CLASS SVM LEARNING .................................... 140
APPENDIX E: PSEUDO CODE FOR SOM LEARNING............................................................ 147
APPENDIX F: PSEUDO CODE FOR K-NN LEARNING............................................................ 160

iv


SUMMARY
Induction machine are used widely in industrial process e.g., steel mills, chemical
plants etc. it is therefore vital to condition monitor the health of the machine to prevent
unexpected and untimely failure. Their untimely downtime have significant economic
and social impact, such as, disruption to production process, spoilage to work-inprogress, costly plant process re-start etc. It is therefore useful to investigate automatic
machine fault detection and diagnosis techniques. This creates the motivation for this
study. Why condition monitoring? Incipient machine faults can be detected by continuous
monitoring [1]. As such, condition-based maintenance has become a new maintenance
methodology that has rapidly been adopted by the industry as the standard operating
procedure. In the past, it is essentially a routine periodic machine shutdown for servicing
and inspection. This method has proved to be inefficient. In condition-based
maintenance, the machine is carefully and continuously condition monitored for

symptoms of failure. Based on such continuous tracking of the machine health-states,
imminent failures is detected and planned shutdown made, only when necessary. This
way, machine downtime and maintenance costs are reduced, and asset security and
reliability increased, both achieving efficiency and profitability for the organization. With
this in view, this project investigates the various machine condition monitoring
techniques, with the objective to implement effective automatic fault detection and
diagnosis methods, to reveal developing incipient faults, so that timely intervention is
made to prevent sudden catastrophic failures.

v


LIST OF TABLES
Table 1.1: Summary of percentage of each of the failure mode. ........................................ 6
Table 1.2: Fault statistics on 8 surveyed articles. (*MC denotes Most Common fault) ..... 7
Table 1.3: Percentage of each of the failure mode derived from Table 1.2. ....................... 7

Table 5.1: Similarity Measure for v1, v2 and v3 ................................................................. 66
Table 5.2: Training and test sets for k-NN classification. ................................................ 69
Table 5.3: Fault classification confusion matrix. .............................................................. 70
Table 5.4: Classification result summary.......................................................................... 70
Table 5.5: Tabulated results of error rate (%) with various k-neighbor values. ............... 71
Table 5.6: Distribution of spectrum of machine vibration signatures. ............................. 83
Table 5.7: Fault classification confusion matrix of vibration signature. .......................... 87
Table 5.8: M-SVM classification of vibration signatures result summary. ...................... 88

Table 6.1: Similarity measures of same-indexed pair of machine current IMFs at 20Hz. 98
Table 6.2: Similarity measures of same-indexed pair of machine current IMFs at 30Hz. 98
Table 6.3: Similarity measures of same-indexed pair of machine current IMFs at 40Hz. 98
Table 6.4: Frequency band for HTY30 (IMF 5-9) machine current signature. .............. 100

vi


LIST OF FIGURES
Figure 1.1: Approaches of this project to investigate machine fault diagnosis. ............... 13
Figure 2.1: Perturbation force (Fc) created by unbalanced mass (m) rotating at Ω. ........ 19
Figure 2.2: Unbalanced rotor fault signatures at various machine speeds. ....................... 19
Figure 2.3: Bearing assembly. .......................................................................................... 21
Figure 2.4: Defective rolling elements (adopted from [93]). ............................................ 21
Figure 2.5: Raceway faults (adopted from [93]). .............................................................. 22
Figure 2.6: Rolling element pitch, diameter and contact angle of a bearing. ................... 23
Figure 2.7: Rolling element fault signatures at various machine speeds. ......................... 24
Figure 2.8: Inner raceway fault signatures at various machine speeds. ............................ 24
Figure 2.9: Outer raceway fault signatures at various machine speeds. ........................... 25
Figure 2.10: Healthy machine signatures at various machine speeds. .............................. 29

Figure 4.1: Connections between the input and output neurons of SOM. ........................ 39
Figure 4.2: Linear decaying learning rate versus learning steps of a SOM. ..................... 40
Figure 4.3: Support vectors, decision boundary and margin of 2-class SVM. ................. 43
Figure 4.4: Multi-class SVM using one-versus-all learning strategy. .............................. 45
Figure 4.5: Mono-variate EMD of BRG signature. .......................................................... 49
Figure 4.6: Mono-variate EMD of UBR signature. .......................................................... 50
Figure 4.7: Mono-variate EMD of HTY signature. .......................................................... 51
Figure 4.8: Extremum of two signals x1(t) and x2(t). ........................................................ 52
Figure 4.9: Mean of 3D ―tube‖ of complex signal signals zb(t)........................................ 53
Figure 4.10: Evolution of bi-variate complex signal for (|HTY40(t)|,|BRG40(t),t). ......... 53
Figure 4.11: Mean of complex signal signals zt(t). ........................................................... 54
Figure 4.12: Evolution of tri-variate complex signal for
(|HTY40(t)|,|BRG40(t)|,|BRB40(t)|).................................................................................. 55

Figure 4.13: Mode mixing. ............................................................................................... 56
Figure 4.14: N-A M-EMD on signal with Guassian white noise added. .......................... 58

Figure 5.1: Template matching using cross-correlation of machine signatures. .............. 60
Figure 5.2: Deterministic signals v1(t), v2(t) and v3(t). ...................................................... 61
Figure 5.3: Normalized cross-correlation sum coefficients of a pair signature. ............... 64
Figure 5.4: Machinery Fault Simulator (MFS) by SpectraQuest®, Inc............................. 66
Figure 5.5: Schematic of machine signature acquisition using DAQ by Dewetron®. ...... 67
Figure 5.6: Bearing fault simulation using MFS. ............................................................. 68
Figure 5.7: Unbalanced rotor fault simulation using MFS. .............................................. 68
Figure 5.8: Error rate (%) versus k-neighbor values. ........................................................ 71
Figure 5.9: Unbalanced rotor fault misclassified as healthy machine. ............................. 73
Figure 5.10: Healthy machine signatures correctly classified. ......................................... 73
Figure 5.11: Unbalanced rotor signature correctly classified. .......................................... 74
vii


Figure 5.12: Unbalanced rotor signature correctly classified. .......................................... 75
Figure 5.13: Wiener Filter................................................................................................. 77
Figure 5.14: Filtered machine signatures at fr=15Hz and 31Hz. ...................................... 77
Figure 5.15: Frequency content of HTY signatures at fr=15Hz........................................ 79
Figure 5.16: Frequency content of HTY signatures at fr=31Hz........................................ 79
Figure 5.17: Frequency content of UBR signatures at fr=16Hz........................................ 80
Figure 5.18: Frequency content of UBR signatures at fr=32Hz........................................ 80
Figure 5.19: Frequency content of BRG signatures at fr=15Hz........................................ 81
Figure 5.20: Frequency content of BRG signatures at fr=32Hz........................................ 82
Figure 5.21: 11-dimensional feature vector at fr=15Hz and 31Hz.................................... 84
Figure 5.22: Semantic map of vibration signatures from two SOM different simulations.
........................................................................................................................................... 85
Figure 5.23: 2-dimensional feature vector. ....................................................................... 86

Figure 5.24: M-SVM classification (Gaussian kernel hsvm=8.0, slack factor C=0.1) of
vibration signature. ........................................................................................................... 87

Figure 6.1: Fourier Series (a finite sum of a 10Hz square wave with n=3 and n=10). ..... 90
Figure 6.2: Different wavelet basis functions. .................................................................. 92
Figure 6.3: A 7-channel Motor Current Signature decomposition by N-A MEMD. ........ 96
Figure 6.4: EMD as filter-banks for HTY30 (IMF 5 – 9) machine current signature. ..... 99
Figure 6.5: IMF1, IMF2, IMF3, IMF4 of the machine signatures at 20Hz. ................... 101
Figure 6.6: IMF1, IMF2, IMF3, IMF4 of the machine signatures at 30Hz. ................... 102
Figure 6.7: IMF1, IMF2, IMF3, IMF4 of the machine signatures at 40Hz. ................... 102
Figure 6.8: IMF10 and IMF11 (residue) of the machine signatures at 20Hz. ................ 103
Figure 6.9: IMF10 and IMF11 (residue) of the machine signatures at 30Hz. ................ 103
Figure 6.10: IMF10 and IMF11 (residue) of the machine signatures at 40Hz. .............. 103
Figure 6.11: IMF5-9 of the HTY machine signatures at 20Hz. ...................................... 104
Figure 6.12: IMF5-9 of the BRG machine signatures at 20Hz. ...................................... 105
Figure 6.13: IMF5-9 of the BRB machine signatures at 20Hz. ...................................... 105
Figure 6.14: IMF5-9 of the UBR machine signatures at 20Hz. ...................................... 106
Figure 6.15: IMF5-9 of the SWF machine signatures at 20Hz. ...................................... 106
Figure 6.16: Feature map using fea_IMF vector at fs=20Hz. ......................................... 108
Figure 6.17: Feature map using fea_IMF vector at fs=30Hz. ......................................... 109
Figure 6.18: Feature map using fea_IMF vector at fs=40Hz. ......................................... 110

viii


LIST OF SYMBOLS &ACRONYMS
ai: non-zero Lagrange multipliers
BMU: Best Matching Unit of SOM
bo: SVM bias
BRG: Bearing fault

BRB: Broken rotor bar
C: slack factor of SVM
Db: is the diameter of the rolling element
Dc: is the rolling element pitch
di , dj: data label for OVA learning strategy for M-SVM
"EMD": Empirical Mode Decomposition
fr: rotor frequency
FR: rotor mechanical frequency
fs: fundamental supply frequency
g0: radial air-gap length in the case of a uniform air-gap
hsvm: Guassian kernel width
h: inverter harmonic order
HTY: healthy machine
IMF: Intrinsic Mode Function
isD: instantaneous values of direct-axis component of monitored stator current
isQ: instantaneous values of quadrature-axis component of the monitored stator current
k-NN: k-Nearest Neighbor
ix


Lm: three-phase magnetizing inductance
Lr: three-phase self-inductance of the rotor winding
Ls’: stator transient inductances
m: rotor unbalanced mass
MFS: machine fault simulator
MMF: magneto motive force
MCSA: Motor Current Signature Analysis
M-SVM: multi-class Support Vector Machine
N: is the number of rotor bars
NB: is the number of rolling elements

nd: eccentricity order number, (static eccentricity=0, dynamic eccentricity=1)
N-A MEMD: noise-assisted MEMD
OVA: one-versus-all learning strategy for M-SVM
P: number of pole-pairs
P0: average air-gap permeance
r: distance between the centre of rotation and the centre of gravity of the rotor
Rr: resistance of rotor phase winding
s: machine slip
SOM: Self-Organizing Map
SWF: Shorted stator winding fault
SVM: support vector machine
Tr: open-circuit rotor time constant given by Lr/Rr
UBR: Unbalanced rotor fault

x


w_som(i,j): SOM output neuron‘s weight vector
w_svmi: SVM weight term
wk[i]: Wiener filter weights at instant ith
xi: support vectors for a SVM
zb(t): complex signal in space (|x1(t)|, |x2(t)|,t)
zt(t): complex signal in space (|x1(t)|, |x2(t)|, |x3(t)|)
β: rolling element contact angle
γ: ―nudge-to-zero‖ constant for Wiener filter
η0: initial learning rate of SOM
θ: angular position of r
θr: angular position of the rotor with respect to the stator reference
μ: constant to adjust the rate of convergence of the weights for Wiener filter
ν: order of the stator time harmonics present in the power supply

ρ: degree of eccentricity
Φ: particular angular position along the stator inner surface
Φn: phase delay
ψrd: instantaneous values of direct-axis component of the rotor flux linkage
ψrq: instantaneous values of quadrature-axis component of the rotor flux linkage
Ω: rotor shaft rotational speed
ω1: angular stator frequency
ωsl: angular slip frequency
ωv: frequency of the kth vibration due to bearing defect

xi


LIST OF RELEVANT PUBLICATIONS

1. W.-Y. Chen, J.-X. Xu, S.K. Panda, ―A Study on Automatic Machine Condition
Monitoring and Fault Diagnosis for Bearing and Unbalanced Rotor Faults‖, IEEE
International Symposium on Industrial Electronics (ISIE‘2011), Poland, Gdansk, 2830 Jun 2011, Accepted for publication

2. W.-Y. Chen, J.-X. Xu, S.K. Panda, ―Application of Artificial Intelligence Techniques
to the Study of Machine Signatures‖, IEEE International Symposium on Industrial
Electronics (ISIE‘2012), China, Hangzhou, 28-31 May 2012, Manuscript submitted
for publication

xii


CHAPTER 1: INTRODUCTION
The field of machine condition monitoring and fault diagnosis is vast. A literature
survey; which is presented subsequently, has shown wide ranging diagnostic techniques.

Various machine operation quantities may be used for monitoring the health of a motor,
e.g., partial discharge, thermo-graphic monitoring of hot-spots, chemical content; such as,
oil degradation detection, wear debris detection, machine axial leakage flux, acoustic,
torque, machine power efficiency, machine vibration signal, and motor current signature
[2, 3]. Among these, the technique by analyzing machine stator current is known as
Motor Current Signature Analysis (MCSA) is the state-of-the-art technique [102]. It is a
popular research area where many algorithms have been proposed, but a single effective
method that is able to detect and diagnosis multiple classes of machine fault still elude
researchers.
The current harmonics that is present in the motor current is mainly created by the
machine asymmetries and vibrations due to machine faults. Hence, this project focuses on
two fault detection techniques, namely, vibration signature and MCSA. There are a
number of issues to address in the formulation of a reliable fault detection and diagnosis
scheme [4]:


definition of a single diagnostic procedure for any type of faults



insensitive to and independent of operating conditions



reliable fault detection for position, speed and torque controlled drives



reliable fault detection for drives in time-varying conditions



quantify a stated fault threshold independent of operating conditions

1


1.1

Objectives
With the above issues in mind, this project aims to accomplish two main

objectives, namely,
Objective 1:

To investigate and formulate an automatic machine condition monitoring
scheme to detect and diagnose the most common machine fault modes,
namely, bearing and unbalanced rotor fault, that is insensitive to machine
operating speed

Objective 2:

To investigate and study the use of MCSA to cover a wider range of
machine fault modes; apart from bearing and unbalanced rotor faults, to
include broken rotor bars and shorted winding faults as well, where
vibration analysis is difficult to diagnose, and to discover unique nonlinear
and non-stationary features for automatic fault classifications

In these studies, computational intelligence are applied. Of particular interests, are
the Self-Organizing Map (SOM), multi-class Support Vector Machine (M-SVM), kNearest Neighbor (k-NN) case-based learning and the Empirical Mode Decomposition

(EMD).

On the first objective, this project has formulated and implemented a simple and
effective data-based scheme, using time-domain vibration data, for the continuous
automatic condition monitoring and diagnosis of bearing and unbalanced rotor faults is
proposed. The key idea is to use a novel normalized cross-correlation sum operator as
2


similarity measure, and in combination with the use k-NN algorithm for the effective
automatic classification of machine faults. This technique is both noise-tolerant and shiftinvariant., It also has a low error rate and insensitive to machine operating speed, as
shown subsequently in this thesis. Further, the diagnosis of these two mechanical faults
using vibration frequency-domain information is also shown, where SOM is used to
discover cluster information on the extracted features in an unsupervised fashion, and an
M-SVM is next used to derive the clusters globally optimal separating hyperplanes for
the automatic classification of the fault modes.
On the second objective, this project use EMD technique to study the motor
current signatures harmonic contents of a healthy machine (HTY), a machine with
bearing fault (BRG), unbalanced rotor fault (UBR), broken rotor bar fault (BRB) and
shorted stator winding fault (SWF). In this project, new unique non-linear and nonstationary features are discovered for these fault modes at machine operating speed of
20Hz and 30Hz. However, it is also observed in this project that uniqueness of these
features is not obvious at higher speed of 40Hz. With the newly discovered unique
features at 20Hz and 30Hz, future works on automatic fault classifications by a single
effective fault detection and diagnosis scheme based on EMD technique is achievable.

1.2

Thesis Organization

This thesis consists of seven chapters.

3


Chapter 1: Introduction on the issues of formulating a reliable machine fault diagnostic
scheme, and the rationale for condition monitoring using MCSA and vibration
analysis, and sets the stage for stating the objectives of this research. Fault
statistics and literature survey are also carried out to compile the fault
statistics and identifies the most common failure modes. This allows research
effort to be directed at the most common failure modes. Fault diagnostic
technique literature survey is next conducted, to understand how various novel
diagnostic techniques are formulated and the difficulties encountered. This
identifies niche research area where this project adds values.

Chapter 2: Mechanics of machine fault elucidates the origin of different type of
machine faults, presents the various fault vibration signatures and the
expected motor current fault spectrum for MCSA.

Chapter 3: Motor Current Signature and Vibration Signature Analysis explain the
difficulties, challenges and issue of vibration analysis and MCSA techniques
and a new approach is proposed.

Chapter 4: Application of Artificial Intelligence (AI) techniques for fault diagnosis
presents the various AI techniques used in this project.

Chapter 5: A study on Automatic Diagnosis of Bearing and Unbalanced Rotor faults
presents the results of the data-based machine fault detection and diagnosis

4

scheme using time-domain vibration data. It explains how cross-correlation
sum operation in time-domain data series is a suitable similarity measure for
the vibration signatures for the purpose of automatic pattern classification
using k-NN classifier, and presents the fault classification error rate and
confusion matrix. It also presents feature extraction using vibration frequencydomain information, fault-class clusters study and discovery using
unsupervised learning by SOM, the clusters globally optimal separating
hyperplane derived from a M-SVM using one-versus-all learning strategy, and
the M-SVM classification error rate and confusion matrix.

Chapter 6: A study on

Motor Current Signature

using Empirical

Mode

Decomposition explains the disadvantages of the traditional analysis tool for
MCSA using Fourier-based and Wavelet transform, the rationale for using
EMD techniques as an effective tool for the analysis of machine current
signatures, with a view to discover new information that Fourier and Wavelet
transform may not be able to reveal.

Chapter 7: Conclusion

5


In the next, a fault mode statistics survey is carried out with a view to identify the

most commonly occurring machine fault modes. The survey article in 1985 [5] reveals
that bearing fault is the most common machine fault mode. The followings further survey
and present the situation in the 1990s.

1.3

Fault Mode Statistics Survey
In [1], a detail survey on fault statistics was done in 2008. In this comprehensive

survey, several sources; including the private communication between the author and an
original equipment manufacturer, referenced about 80 journal papers published in IEEE
and IEE on the subject over the past 26 years since 2008, were used. The table below
summaries the survey result.

Table 1.1: Summary of percentage of each of the failure mode.

Failure Modes
Bearing
Stator Related
Rotor Related
Others

%
52.5%
22.0%
13.0%
12.5%

A majority of the failure mode is due to bearing (52.5%). If bearing fault is
combined with stator related faults, this together accounted for more than 87.5% of the

total faults. Further fault information from referenced articles [6-13] is conducted. The
following table summarizes the findings.

6


Table 1.2: Fault statistics on 8 surveyed articles. (*MC denotes Most Common fault)

Referenced articles
Bearing
Stator Related
Rotor Related
Others

[6]
45%
-

[7]
50%
40%
10%
-

[8]
10%
-

[9]
45%

-

[10]
52%
25%
6%
17%

[11]
41%
37%
10%
12%

[12]
50%
40%
10%
-

[13]
40%
-

By taking the average across the rows of Table 1.2, the following table is derived.

Table 1.3: Percentage of each of the failure mode derived from Table 1.2.

Failure Modes
Bearing

Stator Related
Rotor Related
Others

%
46.1%
35.5%
9.2%
14.5%

Table 1.3 presents a similar failure mode statistics as in Table 1.1. Bearing fault
accounted for more than half the total faults, and the second most common fault is the
stator related faults. These two faults together accounts for more than at least 65% of the
total faults. This finding is consistent with that in Table 1.1. As such, from Table 1.3,
bearing faults accounts for about half of the fault modes, hence it is worthwhile to focus
research effort on bearing faults. If the fault coverage is extended to include rotor and
stator related faults e.g., unbalanced rotor, broken rotor bars and shorted stator windings,
about 85% of all fault modes is covered. With this information, objective 2, presented
above, is set.

7


1.4

Literature Survey
This section presents the literature survey to shed lights on the various techniques

used and progress made. The survey focuses on machine vibration signature and MCSA,
as these techniques are able to detect bearing, unbalanced rotor, broken rotor bars and

shorted stator windings fault modes [1].

1.4.1 Vibration Signature Analysis
Vibration signature analysis is the most commonly monitored operation parameter
for detection and diagnosis of mechanical fault modes e.g., bearing defects and
eccentricities [14]. Using Wavelet techniques to preprocess the vibration signal is
popular. The articles [15-17] presented such a study where the wavelet coefficients were
the feature vectors. It is interesting to note that, in Elsevier collection of articles, the use
of Morlet wavelet basis function is common, whereas in IEEE collection of articles,
Daubechies wavelet is popular. Higher order spectral analysis using Bispectral transform,
is used for noise suppression, detection of non-Gaussian data, and to detect nonlinearity
of the fault information in [18]. Envelope analysis is used in [19, 20] for feature
extraction. Article [21] showed that combining features extracted from Mel-frequency
Cepstral Coefficients (MFCC) and Kurtosis, are effective for diagnosing bearing faults.
Independent component analysis can be used to extract features from vibration signatures
for bearing fault diagnosis [22]. After feature extraction, AI techniques e.g., neural
network, SVM, SOM, are next used to predict fault modes. Excellent examples of works
done are [23-26].
8


However, the most interesting technique is the use of EMD for the analysis of
machine vibration signatures, where the basis function is derived based on empirical data
in terms of Intrinsic Mode Functions (IMFs). Articles [27-31] presented such a study.
The IMFs thus derived are the feature vectors for fault diagnosis.

1.4.2 Motor Current Signature Analysis
A survey on MCSA technique reveals that the approaches are numerous and
wide-ranging. The articles maybe broadly categorized into: reviews, model construction
for fault modes, feature extraction techniques, and the use of computational intelligence

for machine fault diagnosis.
Over the years, a series of reviews have been made by [1, 4, 7, 32-37]. They offer
a good overview on how progress has been made. A notable change is the progress from
the use of traditional Fourier transform e.g., Fast Fourier Transform (FFT), to analyze
motor current signatures, and the increasingly popular use of Wavelet transform e.g.,
Discrete Wavelet Packet transform, to identify fault spectrum and extract unique features
for fault diagnosis. FFT is the traditional tool for MCSA where by locating individual
fault spectrum, the machine fault is diagnosed. This approach is successful for broken
rotor bars and eccentricities faults [11, 38, 39].
In [40, 41], a good comparative study of the various techniques for MCSA for
broken rotor bar and air-gap eccentricities is presented. In [42], a good review is given on
the various diagnosis methods for stator voltage asymmetry and rotor broken bars.
However, these techniques are mainly Fourier based. It is worth to note that, as the fault
spectrum is functions of machine slip, it is particularly difficult to locate the fault
9


spectrum in low slip situation. Further, since motor current signature is non-linear and
non-stationary in nature [43], as such, Wavelet multi-resolution decomposition approach
is popular. In [44-52], Wavelet transform is used to decompose the motor current into
various approximate and detail levels wavelet coefficients, and features are extracted
from these coefficients for fault diagnosis. Diagnoses of bearing, broken rotor bar,
eccentricities faults were reported. However, careful selection of a wavelet basis function
is not trivial [40, 49], as wavelet decomposition is a convolution computation of machine
signature with wavelet basis function and hence a different choice of basis function
produces different results.
Beside wavelet technique, other high resolution frequency-domain techniques e.g.,
Eigen-analysis Multiple Signal Classification (MUSIC) spectrum Estimator, Welch, Burg
[53] are used. In [54], these techniques are applied for the detection of rotor cage faults.
In [8], a detail study using different auto-regressive parametric methods e.g., YuleWalker, and the possibility of using a lower sampling rate were explored. The article [55],

showed that a sliding window ROOT-MUSIC algorithm for bearing fault diagnosis is
possible, and in [56] a novel combination of maximum covariance method for frequency
tracking and Zoom-FFT technique, to selectively increase the frequency resolution of the
frequency range of interest for fault diagnosis were demonstrated. Other methods
incorporating temporal information of the motor current using higher order statistic, such
as, spectral kurtosis is used in [57].
With the popular use of inverter speed controller for machine, the effect of PWM
inverter harmonics on MCSA was investigated in [58, 59]. In [60], inverter input and
output current were studied with a view to detect the twice fundamental frequency

10


harmonics for diagnosis of rotor faults. It is shown that detection of these harmonics is
possible using inverter input current near zero frequency.
To extend the type of fault coverage, stator winding faults are investigated as well.
In [61], a novel diagnostic indicator for stator winding fault, that does not involve ground
fault, is formulated using positive and negative sequence line-voltage and line-current
information. The key idea is that various indicators were determined at various machine
speeds, and a kind of lookup table was created for diagnosis at different machine
operating speed. However, the use of line-voltage, made the method invasive, where
potential transformer (PT) is required. The article [62] presented a method using
Extended Park‘s Vector Approach (EPVA), where instead of observing the ovality of the
signature in the D-Q plane, the frequency-domain information revealed the presence of
fault for stator winding. This approach may be used for bearing fault as well [10].
Survey also revealed other innovative approaches. Instead of using steady-state
information, transient start-stop information may be used for diagnosis as well, as shown
in [63, 64]. In [6], monitoring instantaneous power factor and motor efficiency [65] are
possible approaches too. An interesting approach is presented in [66], where a noise
cancellation technique was used for diagnosing general roughness fault. The scheme

assumed that all frequencies that are not related to the bearing faults, e.g., supply
frequency, supply unbalance harmonics, the eccentricity harmonics, the slot harmonics,
saturation harmonics and interferences from environmental sources are regarded as noise
and are estimated by a Wiener Filter. All these noise components are then cancelled out
by their estimate in a real-time manner. The remaining components are hence related to

11


bearing fault, and the RMS value of this noise-cancelled signal is next calculated online
as fault index, with impending fault as an increase in fault index.
Model-based approach aims to construct a mathematical model of the machine
and thereby using the model to analysis and predict fault mode [67-74]. Finite element
analysis is popular for simulating and studying of fault mode; especially for broken rotor
bar. Winding-Function model is specially formulated for modeling air-gap eccentricity as
shown in [75, 76].
Data-based approach collects real machine fault data rather than using
sophisticated mathematical model to calculate them, and uses these data as examples for
fault diagnosis. These examples are collected from fault simulator. With the fault data
available, AI techniques e.g., SOM, neural network, fuzzy logic, M-SVM etc., are used to
automatically classify the faults [77-81]. Other modeling approaches are possible, such as,
[12] use Autoregressive (AR) Spectrum Estimation; a form of parametric spectrum
estimation technique to model a healthy motor signature, and deviation from this baseline
indicates a bearing general roughness fault. However, this method requires the use of
notch-filter to remove the dominant fundamental frequency and a series of filter banks to
remove the harmonics of other possible faults e.g., unbalance voltage source, cyclical
load torque, eccentricities, broken rotor bars, rotor slotting effects etc. This adds to the
complexity of this method. Recently, the use of Independent Component Analysis (ICA)
has achieved remarkable results, where the diagnostic procedure is independent of
machine operation speed for the diagnosis of bearing and broken rotor bars [82-84].

12


EMD is applied for the diagnosis of shorted stator winding fault and broken rotor
bars in [85, 86]. However, in each of the study, only one fault mode was covered. This
runs the risk of mis-diagnosing a fault, as another fault signatures not covered in the
study, may produce similar features. This project aims to widen the scope of motor
current signature study to cover more fault modes. Figure 1.1 illustrates the approaches of
this project to investigate the automatic fault diagnosis of AC synchronous machine. In
the next, the mechanics of machine fault mode is presented.
AC synchronous machine fault diagnosis

Vibration signatures

Motor current signatures

Unbalanced Rotor Bar fault

Broken Rotor Bar fault

Bearing fault

Unbalanced Rotor Bar fault
Bearing fault
Shorted Winding fault

Time-domain analysis
Normalized Cross-correlation

Time-domain analysis
Empirical Mode Decomposition

Wiener filter

Frequency-domain analysis
Fast Fourier Transform

Frequency-domain analysis
Hilbert Hwang Transform

Figure 1.1: Approaches of this project to investigate machine fault diagnosis.

13