Zhenyou Zhang

Data Mining Approaches for

Intelligent Condition-based
Maintenance
A Framework of Intelligent Fault Diagnosis and
Prognosis System (IFDPS)

Thesis for the degree of Philosophiae Doctor

Trondheim, May 2014

Norwegian University of Science and Technology

Faculty of Engineering Science and Technology
Department of Production and Quality Engineering
NTNU
Norwegian University of Science and Technology

Thesis for the degree of Philosophiae Doctor

Faculty of Engineering Science and Technology

Department of Production and Quality Engineering

© Zhenyou Zhang

ISBN 978-82-326-0074-8 (printed ver.)

ISBN 978-82-326-0075-5 (electronic ver.)
ISSN 1503-8181

Doctoral theses at NTNU, 2014:75

Printed by NTNU-trykk
Acknowledge

It gives me an immense pleasure to present the thesis report in its completed form.
First of all, I would like to thank my main supervisor Prof. Kesheng Wang for his
guidance, but more importantly for his moral support through this research work.
Without his timely advice and through knowledge in Data Mining and Condition-
based Maintenance, the research would not have been accomplished such a great
success. I am extremely thankful for his support.
I thank my co-supervisor Odd Myklebust for his support and help during these
years.
I thank my colleagues PhD Candidate Quan Yu, Associate Prof. Yiliu Liu, Dr.
Lijuan Dai, Dr. Rhythm S. Wadhwa and all of whom helped and supported me
during these years.
I thank Prof. Lilan Liu who came to NTNU as a visit scholar from Shanghai
University. I was very pleasure to work together with her during her half year
visiting.
I thank Dr. Guijuan Lin who came to NTNU as an exchange PhD student from
Tongji University. I was very pleased to work with her during three months.
I also thank two Master students, Deborah Cruciani and Roberta Cusanno, from
University of Bologna to work together in fault diagnosis and maintenance
scheduling optimization.
Last, but the most important, I am extremely grateful to my parents and my wife
for being there for me all these years, for their patience, support and
encouragement.

Trondheim, May 2014

Zhenyou Zhang

I
II
Abstract

Condition-based Maintenance (CBM) is a maintenance policy that take

maintenance action just when need arises with real-time condition monitoring.
Intelligent CBM means a CBM system is capable of understanding and making
maintenance decisions without human intervention. To achieve this objective, it is
needed to detect current conditions of mechanical and electrical systems and
predict the fault of the systems accurately. What’s more, the maintenance
scheduling need to be optimized to reduce the maintenance cost and improve the
reliability, availability and safety based on the results of fault detection and
prediction.

Data mining is a computational process of discovering patterns in large data sets

involving methods at the intersection of artificial intelligence, machine learning,
statistics, and database systems. The goal of the data mining is to extract useful
information from a data set and transform it into an understandable structure for
further use.

This thesis develops framework of Intelligent Fault Diagnosis and Prognosis

System (IFDPS) for CBM based on Data Mining Techniques. It mainly includes
two tasks: the one is to detect and predict the condition of the equipment and the
other is to optimize maintenance scheduling accordingly. It contains several phases:
sensor selection and its placement optimization, signal processing and feature
extraction, fault diagnosis, fault prognosis and predictive maintenance scheduling
optimization based on results of fault diagnosis and prognosis. This thesis applies
different data mining techniques containing Artificial Neural Network such as
Supervised Back-Propagation (SBP) and Self-Organizing Map (SOM), Swarm
Intelligence such as Particle Swarm Optimization (PSO), Bee Colony Algorithm
(BCA) and Ant Colony Optimization (ACO), and Association Rule (AI) in most of
these phases.

The outcomes of the thesis can be applied in mechanical and electrical system in
industries of manufacturing, wind and hydro power plants.

III
IV
 Table of Contents

Table of Contents

Acknowledge .............................................................................................................I
Abstract ................................................................................................................... III
Table of Contents ..................................................................................................... V
List of Figures ......................................................................................................... XI
List of Tables ........................................................................................................ XV
Abbreviations ......................................................................................................XVII
1 Introduction ....................................................................................................... 1
1.1 Motivation of Present Work ...................................................................... 1
1.2 Literature Review...................................................................................... 2
1.2.1 Review of Maintenance Strategies ........................................................ 2
1.2.1.1 CorrectiveMaintenance(CM).......................................................5
1.2.1.2 PreventiveMaintenance...............................................................6
1.2.1.3 PredictiveMaintenance(PM).......................................................7
1.2.2 Review of Sensor System and Sensor Placement Optimization ........... 9
1.2.2.1 SensorClassification......................................................................9
1.2.2.2 WirelessSensorNetworks(WSNs).............................................11
1.2.2.3 RadioͲfrequencyIdentification(RFID).........................................12
1.2.2.4 SensorPlacementOptimization..................................................14
1.2.3 Review of Fault Diagnosis and Prognosis .......................................... 15
1.2.4 Review of Maintenance Scheduling Optimization ............................. 16
1.3 Contributions........................................................................................... 17
1.4 List of Scientific Articles ........................................................................ 18
1.5 Outline of Thesis ..................................................................................... 20
2 Framework of Intelligent Fault Diagnosis and Prognosis Systems (IFDPS) for
CBM........................................................................................................................ 21
2.1 Introduction ............................................................................................. 21
2.2 Objectives and Benefits .......................................................................... 21
2.3 Structure of IFDPS .................................................................................. 22
2.3.1 Data Acquisition ................................................................................. 23

V
 Table of Contents

2.3.1.1 ClassificationofSensors..............................................................24
2.3.1.2 SensorPlacementOptimization..................................................24
2.3.2 Signal Preprocessing and Feature Extraction ...................................... 25
2.3.3 Fault Diagnosis and Identification ...................................................... 25
2.3.4 Fault Prognosis and Remaining Useful Life Evaluation ..................... 26
2.3.5 Maintenance Scheduling Optimization ............................................... 29
2.4 Summary ................................................................................................. 29
3 Data Mining Techniques for IFDPS ................................................................ 31
3.1 Introduction ............................................................................................. 31
3.2 Artificial Neural Networks (ANN) ......................................................... 32
3.2.1 Supervised Learning ANNs ................................................................ 32
3.2.1.1 ForwardPhase............................................................................33
3.2.1.2 BackwardPhase..........................................................................34
3.2.2 Self-Organizing Map (SOM) .............................................................. 35
3.3 Semi-supervised Learning Methods (Manifold Regularization) ............. 37
3.4 Association Rules.................................................................................... 40
3.4.1 Market-basket Analysis....................................................................... 40
3.4.2 Mining Association Rules Steps ......................................................... 42
3.4.3 The Apriori Algorithm ........................................................................ 42
3.4.4 Generating Association Rules from Frequent Itemset ........................ 43
3.4.5 Improving the Efficiency of the Apriori Algorithm ............................ 45
3.4.5.1 PartitionͲbasedApriori................................................................45
3.4.5.2 Sampling......................................................................................45
3.4.5.3 Hashing........................................................................................46
3.4.5.4 Transactionremoval...................................................................46
3.5 Swarm Intelligence ................................................................................. 46
3.5.1 Ant Colony Optimization (ACO) ........................................................ 47
3.5.2 Particle Swarm Optimization .............................................................. 49
3.5.2.1 BiologicalMetaphor....................................................................49
3.5.2.2 BasisAlgorithmofPSO................................................................50
3.5.2.3 TheParametersofPSO...............................................................51

VI
 Table of Contents

3.5.2.4 VariantsofPSO............................................................................52
3.5.3 Bee Colony Algorithm ........................................................................ 54
3.5.3.1 BiologicalMetaphor....................................................................54
3.5.3.2 AlgorithmofBCA.........................................................................56
3.6 Summary ................................................................................................. 57
4 Sensor Classification and Sensor Placement Optimization ............................. 59
4.1 Introduction ............................................................................................. 59
4.2 Classification of Sensors ......................................................................... 60
4.3 Wireless Sensor Networks ...................................................................... 67
4.4 RFID Sensor Networks ........................................................................... 71
4.4.1 RFID System ....................................................................................... 72
4.4.2 Embedded RFID Sensor Monitoring .................................................. 73
4.5 General Sensor Networks........................................................................ 74
4.6 Sensor Placement Optimization (SPO) ................................................... 74
4.6.1 Problem Description ........................................................................... 76
4.6.2 Application of PSO in Sensor Placement Optimization ..................... 77
4.6.2.1 TheProcessofPSOApplicationinSensorPlacement
Optimization...............................................................................................77
4.6.2.2 CaseStudyandItsResults...........................................................77
4.6.3 Application of BCA in Sensor Placement Optimization..................... 85
4.6.3.1 TheProcessofApplicationofBCAinSensorPlacement
Optimization...............................................................................................85
4.6.3.2 CaseStudyandItsResults...........................................................85
4.7 Summary ................................................................................................. 87
5 Signal Preprocessing and Feature Extraction .................................................. 89
5.1 Introduction ............................................................................................. 89
5.2 Signal Preprocessing ............................................................................... 90
5.3 Feature Extraction ................................................................................... 91
5.3.1 Feature Extraction in Time Domain.................................................... 91
5.3.2 Feature Extraction in Frequency Domain ........................................... 93
5.3.3 Feature Extraction in Time-Frequency Domain ................................. 95
5.3.3.1 ContinuousWaveletTransform(CWT).......................................96

VII
 Table of Contents

5.3.3.2 DiscreteWaveletTransform(DWT)............................................97
5.3.3.3 WaveletPacketDecomposition..................................................98
5.3.3.4 WaveletͲbasedFeatures...........................................................100
5.4 Feature Selection ................................................................................... 102
5.5 Summary ............................................................................................... 104
6 Fault Diagnosis based on Data Mining Techniques ...................................... 105
6.1 Introduction ........................................................................................... 105
6.2 Fault Diagnosis based on SBP .............................................................. 107
6.3 Fault Diagnosis based on SOM ............................................................. 108
6.4 Fault Diagnosis based on Semi-supervised Learning ........................... 109
6.5 Fault Diagnosis based on Association Rules ........................................ 113
6.6 Case Study 1: Fault Diagnosis Integration of WPD, PCA and BP
Network ............................................................................................................. 113
6.6.1 Experimental Setup ........................................................................... 114
6.6.2 Experimental Procedure .................................................................... 114
6.6.3 Features Extraction in Wavelet Domain ........................................... 115
6.6.4 Principal Component Analysis (PCA) .............................................. 116
6.6.5 Fault Diagnosis using BP Network ................................................... 117
6.6.6 Results and Discussion...................................................................... 118
6.7 Case Study 2: Fault Diagnosis Integration of WPD, FFT and BP Network
.............................................................................................................. 120
6.7.1 Feature Extraction ............................................................................. 121
6.7.2 Fast Fourier Transform to WPD Signals........................................... 123
6.7.3 Fault Diagnosis Procedure of Integrating WPD, FFT and BP Network .
.......................................................................................................... 124
6.7.4 Experiment and Results .................................................................... 125
6.7.5 Discussion ......................................................................................... 126
6.8 Case Study 3: Fault Diagnosis based on Self-organizing Map ............. 130
6.8.1 Experimental Setup ........................................................................... 130
6.8.2 Fault Types of Centrifugal Pump System ......................................... 131
6.8.3 Experiment and Results .................................................................... 133
6.9 Summary ............................................................................................... 135
7 Fault Prognosis based on Artificial Neural Network ..................................... 137

VIII
 Table of Contents

7.1 Introduction ........................................................................................... 137

7.2 Procedure of Fault Prognosis based on Artificial Neural Network ....... 138
7.3 Fault Prognosis based on Indicator Prediction by ANN for Wind Turbine
Monitoring ......................................................................................................... 140
7.3.1 SCADA Dataset Description ............................................................ 141
7.3.2 Modeling of SCADA Parameter Normal Behavior .......................... 142
7.3.2.1 ParameterSelection..................................................................142
7.3.2.2 TrainingANNModel..................................................................143
7.3.3 Prediction and Detection of Rear Bearing Fault ............................... 146
7.3.4 Discussion ......................................................................................... 147
7.4 Summary ............................................................................................... 149
8 Maintenance Scheduling Optimization based on Data Mining Techniques .. 151
8.1 Introduction ........................................................................................... 151
8.2 Predictive Maintenance Scheduling Optimization Based on Swarm
Intelligence ........................................................................................................ 153
8.3 Generating Unit Maintenance Scheduling (GMS) using PSO .............. 154
8.3.1 Fitness function and Constraints of GMS ......................................... 155
8.3.2 Improved PSO (IPSO) Algorithm ..................................................... 156
8.3.3 Case Study and Results ..................................................................... 157
8.4 Dynamic Condition-Based Maintenance Scheduling using BCA......... 159
8.4.1 Model of Condition based PM .......................................................... 159
8.4.1.1 ModelingofManufacturingSystem.........................................160
8.4.1.2 ModelingofEquipmentInspection...........................................160
8.4.1.3 DeteriorationModelforEachMachine....................................161
8.4.1.4 ModellingofCostFunction.......................................................162
8.4.1.5 ModellingofProfitfortheManufacturingSystem...................163
8.4.2 Numerical Examples ......................................................................... 164
8.5 Routing and Scheduling Optimization of Maintenance Flee (RSOM) for
Offshore Wind Farm ......................................................................................... 166
8.5.1 Mathematical Model of RSOM......................................................... 166
8.5.2 Application of Duo-ACO in RSOM Problem ................................... 169
8.5.3 Numerical Examples ......................................................................... 170

IX
 Table of Contents

8.6 Summary ............................................................................................... 176

9 Conclusion and Future Work ......................................................................... 179
9.1 Summary and Conclusions.................................................................... 179
9.2 Suggestions of Future Work ................................................................. 180
References ............................................................................................................. 181

X
 List of Figures

List of Figures

Fig. 1.1 Maintenance History ................................................................................... 3

Fig. 1.2 Maintenance Types...................................................................................... 5
Fig. 1.3 The Classification of Sensors. ..................................................................... 9
Fig. 1.4 Wireless Sensor Networks ......................................................................... 11
Fig. 1.5 Basic Network Topologies ........................................................................ 12
Fig. 2.1 Framework of IFDPS ................................................................................ 23
Fig. 2.2 Model-based and Data-driven Fault Diagnosis Techniques ...................... 26
Fig. 2.3 Remaining Useful Life Distribution for Each Condition .......................... 28
Fig. 3.1 A BP Neural Network with Single Hidden Layer ..................................... 33
Fig. 3.2 Kohonen Model of SOM ........................................................................... 36
Fig. 3.3 Different Forms of the Neighborhood in SOM Network around U c ........ 37
Fig. 3.4 Application of Association Rules in Market-basket Analysis ................... 41
Fig. 3.5 The process of Aprioi Algorithm .............................................................. 43
Fig. 3.6 Example Generation of Association Rules using Apriori Algorithm ........ 44
Fig. 3.7 Generation of Frequent Itemsets using Partition-based Apriori ................ 45
Fig. 3.8 The implementation steps of ACO ............................................................ 49
Fig. 3.9 Birds Flocking of PSO............................................................................... 50
Fig. 3.10 The flowchart of PSO algorithm ............................................................. 52
Fig. 3.11 The waggle dance .................................................................................... 55
Fig. 3.12 The Behavior of the Bees ........................................................................ 55
Fig. 4.1 Schematic Representation of a Measuring Device .................................... 60
Fig. 4.2 The Classification of Sensors. ................................................................... 64
Fig. 4.3 Principle of Ultrasonic Sensors ................................................................ 65
Fig. 4.4 Wireless Sensor Networks ......................................................................... 69
Fig. 4.5 Single-hop Versus Multi-hop Communication in Sensor Networks ......... 70
Fig. 4.6 Typical RFID System ................................................................................ 72
Fig. 4.7 RFID Tags Communication Methods ....................................................... 72
Fig. 4.8 General Structure of Embedded RFID Sensing System ............................ 74
Fig. 4.9 General Sensor Network Structure ............................................................ 75
Fig. 4.10 Structure of PSO Application in Sensor Placement Optimization .......... 77

XI
 List of Figures

Fig. 4.11 Initial Placement of Measuring Points on the Blower ............................. 78
Fig. 4.12 The Finite Element Model of Blower and Its First Four Modes ............. 79
Fig. 4.13 Fitness Changes with Change of Iteration PSO ( n 5 ) for Total
Displacement Mode ................................................................................................ 83
Fig. 4.14 Fitness Changes with Change of Iteration PSO ( n 5 ) for X Direction
Displacement Mode ................................................................................................ 84
Fig. 4.15 Fitness Changes with Change of Iteration PSO ( n 5 ) for Y Direction
Displacement Mode ................................................................................................ 84
Fig. 4.16 Fitness Changes with Change of Iteration PSO ( n 5 ) for Z Direction
Displacement Mode ................................................................................................ 84
Fig. 4.17 Structure of BCA Application in Sensor Placement Optimization ......... 85
Fig. 4.18 Fitness Changes with Change of Iteration BCA ( n 5 ) for Total
Displacement Mode ................................................................................................ 86
Fig. 4.19 Fitness Changes with Change of Iteration BCA ( n 5 ) for X Direction
Displacement Mode ................................................................................................ 86
Fig. 4.20 Fitness Changes with Change of Iteration BCA ( n 5 ) for Y Direction
Displacement Mode ................................................................................................ 86
Fig. 4.21 Fitness Changes with Change of Iteration BCA ( n 5 ) for Z Direction
Displacement Mode ................................................................................................ 87
Fig. 5.1 Vibration Signal in Time Domain ............................................................. 94
Fig. 5.2 Frequency Response Function of Vibration Signal in Fig. 5.1.................. 95
Fig. 5.3 3-layer Signal Decomposition by Discrete Wavelet Transform ................ 98
Fig. 5.4 Decomposed Signals by DWT .................................................................. 98
Fig. 5.5 Decomposed Signals by WPD................................................................... 99
Fig. 5.6. Wavelet Packet Coefficients and Their Relevant Standard Deviation ... 101
Fig. 6.1 Procedure of Fault Diagnosis BP Network.............................................. 108
Fig. 6.2 Procedure of SOM in Fault Diagnosis..................................................... 110
Fig. 6.3 Solution of Two-moon Problem without Unlabelled Dataset ................. 111
Fig. 6.4 Solution of Two-moon Problem with Unlabelled Dataset ...................... 112
Fig. 6.5 Procedure of Semi-supervised Learning in Fault Diagnosis ................... 112
Fig. 6.6 The Structure of Association Rule-based Fault Diagnosis ...................... 113
Fig. 6.7 Hardware of Experimental Setup ............................................................ 114
Fig. 6.8 Sensors Setup on Blower ......................................................................... 114
Fig. 6.9 Parts for Simulation Degradations........................................................... 115

XII
 List of Figures

Fig. 6.10 Raw Signals with Different Degradations ............................................. 115

Fig. 6.11 Tree structures of wavelet packet transform (4 levels).......................... 116
Fig. 6.12 Wavelet Packet Coefficients (WPC) and Their Relevant Standard
Deviation ............................................................................................................... 116
Fig. 6.13 The first four Principal Components ..................................................... 117
Fig. 6.14 Procedure of Fault Diagnosis Integrating BP Network, PCA and WPC118
Fig. 6.15 Errors of Condition 0 ............................................................................. 119
Fig. 6.16 Errors of Condition 0.3 .......................................................................... 119
Fig. 6.17 Errors of Condition 0.7 .......................................................................... 119
Fig. 6.18 Errors of Condition 1 ............................................................................. 120
Fig. 6.19 3-layer Structure of Wavelet Packet Decomposition ............................ 121
Fig. 6.20 Decomposed Signal of Condition 0 ....................................................... 121
Fig. 6.21 Decomposed Signal of Condition 0.3 .................................................... 122
Fig. 6.22 Decomposed Signal of Condition 0.7 .................................................... 122
Fig. 6.23 Decomposed Signal of Condition 1 ....................................................... 122
Fig. 6.24 FFT for Each Version Signal of Condition 0 ........................................ 123
Fig. 6.25 FFT for Each Version Signal of Condition 0.3 ..................................... 123
Fig. 6.26 FFT for Each Version Signal of Condition 0.7 ..................................... 124
Fig. 6.27 FFT for Each Version Signal of Condition 1 ........................................ 124
Fig. 6.28 Procedure of Diagnosis Integrating WPD, FFT and BP Network ......... 125
Fig. 6.29 Errors for Each Condition ..................................................................... 128
Fig. 6.30 Output with Different Number of Hidden Layer Nodes........................ 128
Fig. 6.31 BP Network Training Time with the Increasing of Training Data ........ 129
Fig. 6.32 BP Network Training Time with the Increasing of Hidden Layer Nodes
.............................................................................................................................. 129
Fig. 6.33 Vibration Measurement Points .............................................................. 130
Fig. 6.34 Visualization of SOM ............................................................................ 134
Fig. 6.35 Classification Result of SOM ................................................................ 134
Fig. 7.1 Procedure of Fault Prognosis................................................................... 139
Fig. 7.2 Neural Network Turbine Rear Bearing Temperature Model Training Data
.............................................................................................................................. 144
Fig. 7.3 Rear Bearing Model Testing Input Data.................................................. 145
Fig. 7.4 Rear Bearing Model Output in Normal Condition .................................. 146

XIII
 List of Figures

Fig. 7.5 Fault Detection Results of Rear Bearing ................................................. 147

Fig. 7.6 Rear Bearing Model Testing Input Data of New Turbine ....................... 148
Fig. 7.7 Rear Bearing Model Output in Normal Condition of New Turbine ........ 149
Fig. 8.1 Maintenance Scheduling Optimization Scheme ...................................... 154
Fig. 8.2 Fitness Value by the Change of the Number of Iteration ........................ 159
Fig. 8.3 Inspection Point Schematic Diagram ...................................................... 161
Fig. 8.4 Degradation Model for One Machine...................................................... 161
Fig. 8.5 Fitness Value by the Change of Iterations ............................................... 164
Fig. 8.6 The implementation steps of Duo-ACO .................................................. 171
Fig. 8.7 The offshore wind farm example ............................................................ 172
Fig. 8.8 Objective Value Changes with Iteration (8 turbines) .............................. 174
Fig. 8.9 Objective Value Changes with Iteration (28 turbines) ............................ 176

XIV
 List of Tables

List of Tables

Table 2.1 The Methods of Signal Pre-process and Signal Process. ........................ 25
Table 3.1 A Model of a Simple Transaction Database ........................................... 41
Table 4.1 Measurands of Sensors ........................................................................... 60
Table 4.2 Technological Aspects of Sensors .......................................................... 62
Table 4.3 Detection Means Used in Sensors........................................................... 62
Table 4.4 Sensor Conversion Phenomena............................................................... 62
Table 4.5 Fields of Application............................................................................... 63
Table 4.6 Sensor Materials ..................................................................................... 63
Table 4.7 Main Natural Frequencies of Blower ...................................................... 78
Table 4.8 Total Displacement Mode for Each Point Order .................................... 79
Table 4.9 X Directional Displacement Mode for Each Point Order ....................... 80
Table 4.10 Y Directional Displacement Mode for Each Point Order ..................... 80
Table 4.11 Z Directional Displacement Mode for Each Point Order...................... 81
Table 4.12 Optimal Sensor Placement for Different Number of Measuring Point
using Total Displacement Mode ............................................................................. 82
Table 4.13 Optimal Sensor Placement for Different Number of Measuring Point
using X Direction Displacement Mode ................................................................... 82
Table 4.14 Optimal Sensor Placement for Different Number of Measuring Point
using Y Direction Displacement Mode ................................................................... 83
Table 4.15 Optimal Sensor Placement for Different Number of Measuring Point
using Z Direction Displacement Mode ................................................................... 83
Table 5.1 The Methods of Signal Pre-process and Signal Process ......................... 89
Table 5.2 Comparing Different Time-Frequency Analysis Methods ..................... 96
Table 6.1 Variance for each component ............................................................... 117
Table 6.2 Part of Training Data ............................................................................ 126
Table 6.3 Test Data and the Results ...................................................................... 127
Table 6.4 Measurement Points and Their Corresponding Vibration Types .......... 131
Table 6.5 Parameters Calculated from Vibration Signals ..................................... 133
Table 7.1 Input and Outputs of ANN Model ........................................................ 143
Table 8.1 Weekly Peak Load in Percent of Annual Peak (%) .............................. 158
Table 8.2 Data of Generators ................................................................................ 158

XV
 List of Tables

Table 8.3 Result (Maintenance period) ................................................................. 159

Table 8.4 Machine Parameters .............................................................................. 165
Table 8.5 Results of PM and CM by BCA ........................................................... 166
Table 8.6 Parameters of Maintenance Vessels...................................................... 173
Table 8.7 Parameters of 8 Turbines ...................................................................... 173
Table 8.8 Mmaximum Working Hours for Each Day........................................... 173
Table 8.9 Results of Maintenance Routing with 8 Turbines ................................. 174
Table 8.10 Parameters of 28 Turbines .................................................................. 175
Table 8.11 Results of Maintenance Routing with 28 Turbines ............................. 176

XVI
 Abbreviations

Abbreviations

CBM Condition-based Maintenance

IFDPS Intelligent Fault Diagnosis and Prognosis System
DM Data Mining
RUL Remaining Useful Life
CM Corrective Maintenance
TPM Time-based Preventive Maintenance
RCM Reliability Centered-Maintenance
EPRI Electric Power Research Institute
FMEA Failure Mode and Effect Analysis
OM Opportunity Maintenance
DOM Design out Maintenance
PM Predictive Maintenance
AM Anticipatory Maintenance
RFID Radio-frequency Identification
GA Genetic Algorithm
SVM Support vector machine
HMM Hidden Markov Model
RBF Radial-basis-Function
FLS Fuzzy Logic System
ANN Artificial Neural Network
HCI Hybrid Computational Intelligence
FFT Fast Fourier Transform
PSO Particle Swarm Optimization
BCA Bee Colony Algorithm
ACO Ant Colony Optimization
STFT Short Time Fourier Transform
FEA Finite Element Analysis
SOM Self-organizing Mapping
BP Back-Propagation
SCADA Supervisory Control and Data Acquisition system

XVII
 Abbreviations

CI Computational Intelligence
AR Association Rules
DT Decision Trees
SI Swarm Intelligence
CFT Continues Fourier Transform
DFT Discrete Fourier Transform
WVD Wigner-ville Distribution
TSA Time Synchronous Averaging
WT Wavelet Transform
WP Wavelet Packet
SBP Supervised Back-Propagation
SSL Semi-supervised Learning
RKHS Reproducing Kernel Hilbert Spaces
TID Ttransaction Identifier
HBCA Honey Bee Colony Algorithm
MEMS Microelectromechanical System
RTD Resistance Temperature Detector
WSN Wireless Sensor Network
GPS Global Positioning System
ARR Analytical Redundancy Relation
PCA Principal Component Analysis
BTA Boosting Tree Algorithm
CWT Continuous Wavelet Transform
DWT Discrete Wavelet Transform
WPD Wavelet Packet Decomposition
CBR Case-based Reasoning
SDWPC Standard Deviation of Wavelet Packet Coefficients
PDF Probability Density Function
GMS Generating Unit Maintenance Scheduling
IPSO Improved PSO
RSOM Routing and Scheduling Optimization of Maintenance
OWT Offshore Wind Turbines

XVIII
 Chapter 1: Introduction

1 Introduction

1.1 Motivation of Present Work

With the rapid development of manufacturing, automobile, aeronautics and

aerospace industries, the equipment of those become more and more complex and
integrated, and thus an unanticipated breakdown of the equipment can cause more
losses in economy and human sources. To avoid the unanticipated failure of
equipment, the maintenance action should be performed before the machine
becoming failure. The number of maintenance actions should not exceed its
necessary, or it may increase the cost of maintenance and reduce the product life.
Therefore, the maintenance action should be performed just before the machine
failure. To reach this objective, condition monitoring has to be performed in
equipment and processes of manufacturing and operations to support the
maintenance decision. Therefore, the motivation for present work can be described
as following paragraphs.
Because of the complex, integration and associativity of the equipment, the right
maintenance policy of equipment has to be researched for reducing the loss and
increasing the life cycle of products. As mentioned above, for the key components
of equipment, the maintenance action should be performed just when it is
necessary before failure. Condition-based Maintenance (CBM) policy is based on
the condition of equipment and tries to maintain the correct equipment at right time.
To implement CBM policy, the healthy condition of equipment needs to be
assessed according to the real-time information from the sensors mounted on the
equipment. The present work establishes a framework called IFDPS for CBM to
reducing the maintenance cost and increase the life cycle of products.
To carry out CBM policy, equipment health must be assessed based on the
condition information of the equipment. Fault diagnosis, which means detecting,
isolating, and identifying an impending or incipient failure condition in which
affected component is still operational even though at a degraded mode, is a very
important technique to obtain the information. Normally, when a machine goes
down, most of downtime is used to identify the causes of the failure while only a
small part of that is used to repair or maintain the machine. Diagnostics can answer
this question: why the performance of the observed process, or equipment is
degrading, or in other word, what is the cause of the observed process or machinery
degradation [Djurdjanovic et al., 2003]. It is to see that diagnostics’ function is to
identify the components or causes of the failure happening or about to happen.
Therefore, the present work introduces many Data Mining (DM) methods to carry
out diagnostics to tell the staff which components should be repaired or maintained.
Meanwhile, to carry out CBM policy, the fault prognosis is very important as well
to support maintenance decision. The prognostics can answer the question: when
the observed process, or equipment is going to fail, or degrade to the point when its
performance becomes unacceptable [Djurdjanovic et al., 2003]. CBM policy can
make predictive maintenance scheduling based on the condition of machine.

1
 Chapter 1: Introduction

Condition, the state of a machine, is related to the Remaining Useful Life (RUL).
In the industrial and manufacturing arenas, fault prognosis can be used to estimate
the remaining useful life of a machine or a component once an impending failure
condition is detected, isolated and identified. It is obviously seen that fault
prognosis with fault diagnosis is a basis of predictive maintenance scheduling.
Therefore, the present work also proposes methods for fault prognosis to support
CBM policy.
To carry out CBM, condition monitoring is very important and obtaining parameter
information of machine is the bases for all the processes include diagnostics,
prognostics and predictive maintenance decision. Normally, sensors are used to
collect information of machines. There are two issues need to be considered for
sensors. The one is what kind of sensors should be chosen to collect the
information. The other is where the sensors should be set up on the machine to get
the information continuous or periodically. Actually, the present work focuses on
the second issue, i.e. the sensor placement optimization.
Data Mining (DM) techniques could be very useful for maintenance scheduling,
prognostics, diagnostics and sensor placement selection. Many companies, such as
BMW, ABB, Boeing and Statoil, have lots of history data. But the data has not
been used effectively in current time. DM techniques can be used to extract useful
information from the history data to support all process mentioned above.
Therefore, during the three years of PhD work, Intelligent Fault Diagnosis and
Prognosis System (IFDPS) for Condition-based Maintenance in Manufacturing
systems and processes is established. The framework IFDPS includes almost all
processes of sensor selection, sensor placement optimization, fault diagnosis, fault
diagnosis and prognosis, and maintenance scheduling optimization. It is hoped that
IFDPS can help the companies to carry out near-zero breakdown manufacturing
and further to carry out zero-defect manufacturing.

1.2 Literature Review

As mentioned above, the present work mainly based on the maintenance policy,
methods of diagnostics and prognostics, signal process and sensor strategy. In this
section, the state-of-the-arts for these topics are reviewed briefly.

1.2.1 Review of Maintenance Strategies

Maintenance is defined [EN 13306: 2001, 2001] as the combination of all technical,
administrative and managerial actions during the life cycle of an item intended to
retain it in, or restore it to, a state in which it can perform the required function (a
function or a combination of functions of an item which are considered necessary
to provide a given service). It is a set of organized activities that are carried out in
order to keep an item in its best operational condition with minimum cost acquired.
The maintenance actions could be either repair or replacement activities, which are
necessary for an item to reach its acceptable productivity condition, and these

2
 Chapter 1: Introduction

activities should be carried out with a minimum possible cost. In the period of pre-
World War II, people saw maintenance as an added cost to the plant which did not
increase the value of finished product, and thus, the maintenance at that era was
restricted to fixing the unit when it breaks because it was the cheapest option.
During and after World War II at the time when the advances of engineering and
scientific technology developed, people developed other types of maintenance,
which were much cheaper such as preventive maintenance and in addition, people
in this era classified maintenance as a function of the production system.
Nowadays, increased awareness of such issues as environment safety, quality of
product and services makes maintenance one of the most important functions that
contribute to the success of the industry and world-class companies are in
continuous need of a very well organized maintenance plan to compete world-wide.
The brief history of maintenance mentioned above can be seen in Fig. 1.1 [Shenoy
& Bhadbury, 1998].

Fig. 1.1 Maintenance History

It is very important for a manufacturing company to choose a right maintenance

policy because the affections of maintenance are not only on economy, reliability
and availability but also on personnel safety. The range of maintenance cost is from
15% for manufacturing companies and 40% for iron and steel industry of the whole
cost of manufactured parts and machines in 1990s [Keith Mobley, 2002] and even
more nowadays. The corresponding cost in United Stated is more than 200 billion
dollars every year [Chu et al., 1998]. This shows the significance of maintenance in
the viewpoint of economy. Unexpected failure causes tremendous losses in
economy and production, and may also cause hazard of staff and equipment in
manufacturing plant. Therefore, the maintenance actions are performed before the
failure is very important which is referring to the preventive maintenance and
predictive maintenance.
Maintenance objectives should be consistent with and subordinate to production
goals. The relation between maintenance objectives and production goals is

3
 Chapter 1: Introduction

reflected in the action of keeping production machines and facilities in the best
possible condition. Typically, the objectives of maintenance can be classified into
three groups [Boucly, 2001; Marquez, 2007; Wireman, 1990]:
x Technical objectives. These objectives are the operational imperatives from
the business sector of a company or plant. In general, operational
imperatives are linked to a satisfactory level of equipment availability and
people safety. A generally accepted method to measure the fulfillment of
this goal is the Overall Equipment Effectiveness (OEE), as described in
TPM method [Nakajima, 1988].
x Legal objectives/Mandatory regulations. Normally it is a maintenance
objective to fulfill all these existing regulations for electrical devices,
pressure equipment, vehicles, protection means, etc.
x Financial objectives. To satisfy the technical objective at the minimum cost.
From a long term perspective global equipment life cycle cost should be a
suitable measure for this.
Generally, the objectives can be list as bellowing:
1) Maximizing production or increasing facilities availability at the lowest
cost and at the highest quality and safety standards.
2) Reducing breakdowns and emergency shutdowns.
3) Optimizing resources utilization.
4) Reducing downtime.
5) Improving spares stock control.
6) Improving equipment efficiency and reducing scrap rate.
7) Minimizing energy usage.
8) Optimizing the useful life of equipment.
9) Providing reliable cost and budgetary control.
10) Identifying and implementing cost reductions.
A maintenance action may include a set of maintenance activities: inspection,
monitoring, routine maintenance, overhaul, rebuilding and repair. Inspection can be
performed by measuring, observing, testing or gauging the relevant features of an
item before, during or after other maintenance activity. Monitoring is a kind of
activities performed manually or automatically, continuously or periodically
intended to obtain the actual state of the equipment which can be used to evaluate
parameters changes of the equipment when the equipment is in operating state.
Routine maintenance is a kind of regular elementary maintenance activities, such
as cleaning, tightening of connections and checking lubrication, which usually do
not need special qualification authorization or tools. Overhaul is a comprehensive
set of examinations and actions performed at prescribed intervals of time or a
number of operations in order to maintain the required level of reliability,
availability and safety, and sometimes may require partial or complete dismantling
of the items. Rebuilding is performed when the equipment or components are
approaching their useful life or should be regularly replaced in order to provide the
equipment with a useful life that may be greater than the lifespan of the original
equipment. Repairing is a physical action to restore the required functions of faulty
equipment [Marquez, 2007]. A maintenance action could include some of one or

4
 Chapter 1: Introduction

more above activities. The maintenance may also need fault diagnosis and
prognosis for monitored equipment.
With a long history development, maintenance has been made great progress. At
the beginning, maintenance action is performed when the equipment become
failure. However, this kind of maintenance policy cannot meet the requirement of
the industry and many other types of maintenance are emerged during the several
decades as seen in Fig. 1.2 [EN 13306: 2001, 2001]. In many literatures, Condition-
based Maintenance (CBM) is also called predictive maintenance. This section
mainly reviews corrective maintenance and preventive maintenance briefly, and
review predictive maintenance in detail.

Fig. 1.2 Maintenance Types

1.2.1.1 Corrective Maintenance (CM)

Corrective Maintenance is similar to repair work, which is undertaken after a
breakdown or when obvious failure has been located. That is why it is also called
run-to failure maintenance, maintenance-on-failure or breakdown maintenance. In
CM, the plant item is allowed to failure before maintenance is performed and thus
that it is only suitable if the consequences of failures are small, such as light bulb.
It is only appropriate to apply CM policy if it does not matter whether the machine
fails, or how long the repair will take or how much it will cost. Sometimes a failure
is not predictable using any instrument or analysis, and only checking for failure
will detect the fault. Unfortunately the strategy is widely used in inappropriate
situations. At failure, the task of the repair team is to restore the machine to a state
in which it can perform the required function as quickly as possible [Holmberg et
al., 2010]. Therefore, CM at its best should be utilized only in non-critical areas
where capital costs are small, consequences of failure are slight, no safety risks are
immediate, and quick failure identification and rapid failure repair are possible.
Corrective maintenance is maintenance carried out after fault recognition and
intended to put the equipment into a state in which it can perform a required
function. It could be immediate or deferred [Marquez, 2007]. Immediate
maintenance means the maintenance is carried out without delay after a fault has
been detected to avoid unacceptable consequences, while deferred maintenance

5
 Chapter 1: Introduction

means the maintenance is not immediately carried out after fault detection but is
delayed according to given maintenance rules.
The CM policy has its advantages. Its planning is very simple because the
maintenance action is needed only when the failure happens and the plan is only to
consider the failure rate. The maintenance work is not scheduled until it is really
needed. However, it has major disadvantages [Holmberg et al., 2010]:
x Failure can, and probably will, occur at an inconvenient time, e.g., when the
plant is at full load, or while it is starting.
x A component fault may go unnoticed, leading to expensive consequential
damage, e.g., bearing seizure causes damage to a shaft.
x Dangerous and/or expensive failure consequences should be expected.
x No data are available regarding the past, present and possible future state of
the machine.
x A large breakdown crew may need to be available on standby. All the
required expertise should be either within the plant or easily accessed from
external resources, which is almost always costly, or a longer waiting time
should be expected.
x A large spares inventory is necessary to ensure quick repair.
x Failures exceeding the capacity of the repair team lead to “fire-fighting”.

1.2.1.2 Preventive Maintenance

Preventive Maintenance can be defined as maintenance carried out at
predetermined intervals or according to prescribed criteria and intended to reduce
the probability of failure or the degradation of the functioning of an item [EN
13306: 2001, 2001]. The preventive maintenance action can be condition-based or
predetermined maintenance. Predetermined maintenance carried out in accordance
with established intervals of time or number of units of use but without previous
condition investigation. Condition based maintenance is preventive maintenance
based on performance and/or parameter monitoring and the subsequent actions
which is also known as predictive maintenance which is reviewed in section 1.2.1.3.
Predetermined maintenance means that the maintenance is scheduled without the
occurrence of any monitoring activities [Niu et al., 2010; Zhang & Wang, 2013].
The scheduling can be based on the number of hours in use, the number of times an
item has been used the number of kilometers the items has been used, according to
prescribed dates.
While predetermined maintenance is not the optimum maintenance program, it
does have several advantages over that of a purely corrective program as follows
[Sullivan et al., 2010]:
x Cost effective in many capital-intensive processes.
x Flexibility allows for the adjustment of maintenance periodicity.
x Increased component life cycle.
x Energy savings.
x Reduced equipment or process failure.

6
 Chapter 1: Introduction

x Estimated 12% to 18% cost savings over corrective maintenance program.

However, there are still some disadvantages [Sullivan et al., 2010]:
x Catastrophic failures still likely to occur.
x Labor intensive.
x Includes performance of unneeded maintenance.
x Potential for incidental damage to components in conducting unneeded
maintenance.

1.2.1.3 Predictive Maintenance (PM)

Condition based maintenance is also known as predictive maintenance which
means a set of activities that detect changes in the physical condition of equipment
(signs of failure) in order to carry out the appropriate maintenance work for
maximizing the service life of equipment without increasing the risk of failure. It
depends on continuous or periodic condition monitoring equipment to detect the
signs of failure.
Condition-based Maintenance (CBM) is most popular policy in modern industries
[Carnero Moya, 2004; Dieulle et al., 2001; Han & Song, 2003]. CBM is
maintenance when need arises which is performed after one or more indicators
show that equipment is going to fail or that equipment performance is deteriorating.
It was introduced in 1975 in order to maximize the effectiveness of PM decision
making. CBM is a maintenance program that recommends maintenance actions
(decisions) based on the information collected through condition monitoring
process [Jardine et al., 2006]. In CBM, the lifetime (age) of the equipment is
monitored through its operating condition, which can be measured based on
various monitoring parameters, such as vibration, temperature, lubricating oil,
contaminants, and noise levels. The motivation of CBM is that 99 percent of
equipment failures are preceded by certain signs, conditions, or indications that a
failure is going to occur [Bloch & Geitner, 2012]. Therefore, CBM is needed for
better equipment health management, lower life cycle cost, catastrophic failure
avoidance etc. [Ahmad & Kamaruddin, 2012].
The heart of CBM is the condition monitoring process, where signals are
continuously monitored using certain types of sensor or other appropriate
indicators [Campos, 2009]. Thus, maintenance activities (e.g., repairs or
replacements) are performed only ‘when needed’ or just before failure [Andersen
& Rasmussen, 1999]. In general, the main goal of CBM is to perform a real-time
assessment of equipment conditions in order to make maintenance decisions,
consequently reducing unnecessary maintenance and related costs [Gupta &
Lawsirirat, 2006].
Monitoring is defined as: ‘An activity which is intended to observe the actual state
of an item’ [SS-EN 13306, 2001]. In other words, condition monitoring is a tool
used to indicate the condition of equipment in a system [Hameed et al., 2009]. In
general, the purpose of the condition monitoring process is twofold. First, it
collects the condition data (information) of the equipment. Second, it increases

7
 Chapter 1: Introduction

knowledge of the failure causes and effects and the deterioration patterns of
equipment [Ahmad & Kamaruddin, 2012].
The condition monitoring process can be carried out into two ways: on-line and
off-line. On-line processing is carried out during the running state of the equipment
(operating state), while off-line processing is performed when the equipment is not
running. In addition, condition monitoring can be performed either periodically or
continuously. Typically, periodical monitoring is carried out at certain intervals,
such as every hour or every working shift end, with the aid of portable indicators,
such as hand-held meters, acoustic emission units, and vibration pens. The
condition monitoring process also includes evaluations based on human senses to
measure or evaluate equipment conditions, such as degree of dirtiness and
abnormal color. As for continuous monitoring, as its name suggests, monitoring is
performed continuously and automatically based on special measurement devices,
such as vibration and acoustic sensors.
There are two main limitation of continuous monitoring exist: it is expensive
because many special devices are required and inaccurate information may be
obtained because the continuous flow of data creates increased noise. In contrast,
the main limitation of periodic monitoring is the possibility of missing some
important information of equipment failure between monitoring intervals [Jardine
et al., 2006]. Most equipment failures are preceded by certain signs, conditions, or
indications that such a failure was going to occur and many condition monitoring
techniques can be used to monitor equipment conditions [Bloch & Geitner, 2012].
PM has some advantages over other maintenance policies: 1) Improving
availability and reliability by reducing downtime; 2) Enhancing equipment life by
reducing wear from frequent rebuilding, minimizing potential for problems in
disassembly and reassembly and detecting problems as they occur; 3) Saving
maintenance costs by reducing repair costs, reducing overtime and reducing parts
inventory requirements; 4) Decreasing number of maintenance operations causes
decreasing of human error influence. However, there are still some challenges of
PM: 1) Initiating PM is costly because the cost of sufficient instruments could be
quite large especially if the goal is to monitor already installed equipment; 2) The
goal of PM is accurate maintenance, but it is difficult to achieve for the complexity
of equipment and environment; 3) Introducing PM will invoke a major change in
how maintenance is performed, and potentially to the whole maintenance
organization in a company. Organizational changes are in general difficult.
There are many kinds of techniques, such as sensors techniques, signal process
techniques, fault diagnosis techniques, fault prognosis techniques and maintenance
optimization techniques, can be used to support maintenance decision making. All
these techniques will be reviewed.

8
 Chapter 1: Introduction

1.2.2 Review of Sensor System and Sensor Placement Optimization

1.2.2.1 Sensor Classification

There are many kinds of data could be collected from the system because of the
complex and integrated manufacturing system and process [Vachtsevanos et al.,
2006]. Therefore, the selection of the suitable sensors is the key for effective
condition monitoring. A variety of sensors exist to effectively monitor and control
various process parameters (Fig. 1.3):
x Mechanical sensors such as acceleration, position, displacement, speed
sensors and strain gauges etc.;
x Performance sensors such as pressure, fluid and thermodynamic sensors etc.;
x Electrical measurement sensors such as eddy-Current proximity probes and
micro-electromechanical system sensors etc.;
x Fibre-optic sensors.

Accelerometers (Vibration Measurements)

Strain gauges
Mechanical Sensor Ultrasonic Sensor System
Systems Position, speed, acceleration, torque, strain
ĂĂ
Temperature Sensors / Thermography
Performance Sensors Pressure, Fluid and thermodynamic
Optical properties and biochemical elements
ĂĂĂĂ
Eddy-Current Proximity Probes
Electrical Measurement
Microelectromechanical System (MEMS) Sensors
Fiberoptic Sensors

Fig. 1.3 The Classification of Sensors.

Mechanical sensor systems have been studied extensively, and a large number of
such devices are currently in use to monitor system performance for operational
state assessment and tracking of fault indicators. A number of mechanical
quantities - position, speed, acceleration, torque, strain, temperature, etc. - are
commonly employed in dynamic systems. Most of devices for measuring these
quantities are available commercially, and their operation has been amply
described in textbooks and publications [Silva, 1989; Stuart & Allocca, 1984].
However, the most useful Mechanical sensors for condition monitoring are
accelerometers and strain gauge.
System performance and operational data are monitored routinely in all industrial
establishments, utility operations, transportation systems, etc. for process control,
performance evaluation, quality assurance, fault diagnosis and prognosis, and

9
 Chapter 1: Introduction

maintenance decision support purposes. A large number of sensor systems have

been developed and employed over the years. The list includes devices that are
intended to measure such critical properties as temperature; pressure; fluid,
thermodynamic, and optical properties; and biochemical elements, among many
others. Sensors based on classic measuring elements—inductive, capacitive,
ultrasound-have found extensive applications. More recently, biochemical sensors
have begun taking central stage, and their detection principles and requirements are
described in the technical literature [Guardia, 1995]. Characteristic chemical-sensor
properties with potential application to structural fault diagnosis include liquid and
solid electrolytic sensors, photochemical sensors, humidity sensors, and field-effect
and mass-sensitive devices. As an example of their principles of operation,
consider conductivity sensors. In these devices, the interaction of the gas with the
solid (semiconducting metal oxide or organic semiconductor) causes a change in
conductivity. A change in resistance also can be caused by a change in the
temperature of the sensor material [Vachtsevanos et al., 2006].
Electromechanical, electrical, and electronic systems constitute a major component
of industrial engine. They are the dominant element in such areas as transportation
systems, utilities, biomedical instrumentation, communications, computing, etc. A
number of sensor systems have been developed and applied in the recent past in an
attempt to interrogate critical components and systems for fault diagnosis and
prognosis. Transducing principles based on eddy-current response characteristics,
optical and infrared signal mentoring, microwaves, and others have been
investigated [Vachtsevanos et al., 2006; Zou et al., 2000].
Fiber optics has penetrated the telecommunications and other high technology
sectors in recent years. They find utility in the sensor field because of their
compact and flexible geometry, potential for fabrication into arrays of devices,
batch fabrication, etc. Fiberoptic sensors have been designed to measure strain,
temperature, displacement, chemical concentration, and acceleration, among other
material and environmental properties. Their main advantages include small size,
light weight, immunity to electromagnetic and radio frequency interference, high-
and low-temperature endurance, fast response, high sensitivity, and low cost. The
basic physics leads to a very stable, accurate, and linear temperature sensor over a
large temperature range. These sensors are also quite small and therefore ideal for
applications where restricted space or minimal measurement interference is a
consideration. The size also leads to a very small time response as compared with
other temperature measurement techniques [Ansari, 1998; Lienhart & Brunner,
2003; Vachtsevanos et al., 2006].
Normally, the output of sensor is electrical signal whatever the physical signals is.
The electrical signals need to be transferred to a database for analysis. The signals
can be transferred with cables or wireless network. Recently, a fast development
technology RFID can be used to transfer the signals. Transferring signals with
cables is a kind of traditional methods and is very effective. However, for some
companies such as wind generator plant, the monitoring equipment may be far
away from the plant and thus wireless network and RFID could be used to solve
this problem.

10
 Chapter 1: Introduction

1.2.2.2 Wireless Sensor Networks (WSNs)

Wireless sensor networks (WSNs) are important in applications where wires cannot
be run owing to cost, weight, or accessibility. Properly designed WSNs can be
installed and calibrated quickly and can be up and running in a very short time
frame [Lewis, 2004]. Typically, WSNs generally consist of a data-acquisition
network, a data-distribution network monitored and controlled by a management
center as shown in Fig. 1.4 [Lewis, 2004]. Too many of available technologies
make even the selection of components difficult, let alone the design of a consistent,
reliable, robust overall system.
The basic issue in WSNs is the transmission of messages to achieve a prescribed
message throughput and quality of service which can be specified in terms of
message delay, message due dates, bit error rates, packet loss, economic cost of
transmission, transmission power, etc. Depending on quality of service, the
installation environment, economic considerations, and the application, one of
several basic network topologies may be used. A communication network is
composed of nodes, each of which has computing power and can transmit and
receive messages over communication links, wireless or cabled. The basic network
includes fully connected, mesh, star, ring, tree, bus as shown in Fig. 1.5 [Lewis,
2004]. A single network may consist of several interconnected subnets of different
topologies. Networks are further classified as Local Area Networks (LAN), e.g.
inside one building, or Wide Area Networks (WAN), e.g. between buildings.

Fig. 1.4 Wireless Sensor Networks

11
 Chapter 1: Introduction

Fig. 1.5 Basic Network Topologies

Fully connected networks suffer from problems of NP-complexity [Garey &

Johnson, 1979]; as additional nodes are added, the number of links increases
exponentially. Therefore, for large networks, the routing problem is
computationally intractable even with the availability of large amounts of
computing power. Mesh networks are regularly distributed networks that generally
allow transmission only to a node’s nearest neighbors. The nodes in these networks
are generally identical, so that mesh nets are also referred to as peer-to-peer nets.
Mesh nets can be good models for large-scale networks of wireless sensors that are
distributed over a geographic region. Since there are generally multiple routing
paths between nodes, these nets are robust to failure of individual nodes or links.
An advantage of mesh nets is that, although all nodes may be identical and have the
same computing and transmission capabilities, certain nodes can be designated as
‘group leaders’ that take on additional functions. If a group leader is disabled,
another node can then take over these duties [Lewis, 2004]. Star topology means
that all nodes are connected to a single hub node. The hub requires greater message
handling, routing, and decision-making capabilities than the other nodes. If a
communication link is cut, it only affects one node. However, if the hub is
incapacitated the network is destroyed. Ring topology means all nodes perform the
same function and there is no leader node. Messages generally travel around the
ring in a single direction. However, if the ring is cut, all communication is lost. In
the bus topology, messages are broadcast on the bus to all nodes. Each node checks
the destination address in the message header, and processes the messages
addressed to it. The bus topology is passive in that each node simply listens for
messages and is not responsible for retransmitting any messages [Lewis, 2004].

1.2.2.3 Radio-frequency Identification (RFID)

Radio-frequency identification (RFID) is one of numerous technologies grouped
under the term of Automatic Identification (Auto ID), such as bar code, magnetic
inks, optical character recognition, voice recognition, touch memory, smart cards,
biometrics etc. Auto ID technologies are a new way of controlling information and

12
 Chapter 1: Introduction

material flow, especially suitable for large production networks [Ilie-zudor et al.,
2006]. RFID is the use of a wireless non-contact radio system to transfer data from
a tag attached to an object, for the purposes of identification and tracking
[http://en.wikipedia.org/wiki/Radio-frequency_identification]. In general terms, it
is a means of identifying a person or object using a radio frequency transmission.
The technology can be used to identify, track, sort or detect a wide variety of
objects [Lewis, 2004]. Recently, RFID become more and more interesting
technology in many fields such as agriculture, manufacturing and supply chain
management.

The history of RFID technology can be tracked back to the radio-based

identification system used by allied bombers during World War II [Garfinkel &
Holtzman, 2005]. Early identification Friend or For (IFF) systems were used to
distinguish Allied fighter and bomber by identifying the correct signals sent by
Allied aircrafts, from aircrafts sent by enemy at night. After the war, Harry
Stockman realized that it is possible to power a mobile transmitter completely from
the strength of a received radio signal, and then he introduced the concept of
passive RFID systems [Stockman, 1948]. In 1972, a patent application for
“inductively coupled transmitter-responder arrangement” was filed which is used
separate coils for receiving power and transmitting the return signal [Kriofsky &
Kaplan, 1975]. In 1979, a patent application for “identification device” (two
antennas was combined) was filed which is seen as a RFID landmark because it
emphasized the potentially small size of RFID device [Beigel, 1982]. The 1980s
became the decade for full implementation of RFID technology, though interests
developed somewhat differently in various parts of the world. The greatest interests
in the United States were for transportation, personnel access, and to a lesser extent,
for animals. In Europe, the greatest interests were for short-range systems for
animals, industrial and business applications, though toll roads in Italy, France,
Spain, Portugal, and Norway were equipped with RFID. The 1990s was a
significant decade for RFID since it saw the wide scale deployment of electronic
toll collection in the United States. The world's first open highway electronic
tolling system opened in Oklahoma in 1991 and then extended to the whole world.
Interest was also keen for RFID applications in Europe during the 1990s. Both
Microwave and inductive technologies were finding use for toll collection, access
control and a wide variety of other applications in commerce [Landt, 2001]. The
21st century opens with the smallest microwave tags built using, at a minimum,
two components: a single custom CMOS integrated circuit and an antenna. Tags
could now be built as sticky labels, easily attached to windshields and objects to be
managed [Landt, 2005]. It seems that there are still a great many developments of
RFID to look forward to as the history continues to teach that and RFID will be
presented in our daily life.
As mentioned above, most of applications of RFID are in logistics or Auto ID area.
However, from its principle, it is possible to apply this technology in signal
transmission in condition monitoring (vibration measuring). However, there are no
matured products of the RFID sensor for measuring vibration in production so far.
Generally, there are two kinds of RFID vibration sensor could be developed. The
one is combining the RFID tag and vibration sensor together to compose a new

13
 Chapter 1: Introduction

RFID vibration sensing tag. The other is to connect vibration sensor to RFID tag
and the RFID system only used to transmit vibration data to RFID reader and
further to host computer. This application can make the measuring vibration
become very flexible and effective [Wang & Zhang, 2012].

1.2.2.4 Sensor Placement Optimization

The basic problem for condition monitoring is to deduce the existence of a defect
in a structure from measurements taken at sensors distributed on the structure. The
correctness of defect diagnosis depends on the method of pattern recognition for
fault and effectiveness of signals from the sensors mounted on the machines. While
carrying out on-site condition monitoring for a machine, the inappropriate
distribution of sensors might result in weak incentives of certain order or modal,
and affect the accuracy of fault identification. The aim of optimizing the placement
of sensors is to obtain as much as possible of machine structural information with
as few as possible sensors, which benefit the company in the economy viewpoint.
Because of constraints of machine structure and environment, and consideration of
economy, only a small number of sensors are installed when a condition
monitoring system is established. It is very important to design the optimal position
of the sensor to mount in order to ensure the accuracy and correctness of
monitoring and fault judgment.
There are many literatures in optimal placement optimization of sensors in machine
level. The spatial controllability was used to find the optimal placement of
collocated actuator-sensor pairs for effective average vibration reduction over the
entire structure, and the maintaining modal controllability and observability were
used to select vibration modes for a thin plate [Halim & Moheimani, 2003].
Recently, intelligent optimization algorithm has developed well which is a method
to simulate the biological and physical process which can be used in sensor
placement optimization. Many researchers focus on Genetic Algorithm (GA)
application in sensor placement optimization and make up for a lot of shortage of
the traditional optimization algorithm [Li et al., 2000; Liu et al., 2008; Sun et al.,
2008]. But GA has to adopt binary coding and has complex operation process such
as mutation, genetic and crossover. PSO adopts real number coding to avoid the
complex operation, which is simple and easy to realize. So it is easy to apply in
sensor placement optimization. PSO and finite element analysis were combined
together to search the sensors optimal placement of a gearbox [Pan et al., 2010].
Binary PSO and Analytical redundancy Relations (ARRs) were combined to
optimize the sensor placement for fault diagnosis [Du et al., 2011]. The sensor
placement optimization is a very important aspect for many applications such as
modal test and parameter identification [Cherng, 2003; Papadimitriou, 2004;
Pennacchi & Vania, 2008], fault diagnosis [Bhushan & Rengaswamy, 2000;
Staszewski, 2002; Worden & Burrows, 2001] and process monitoring [Wang et al.,
2002]. The PhD work tries to apply Swarm Intelligence (SI) such as Particle
Swarm Optimization (PSO) and Bee Colony Algorithm (BCA), and finite element
analysis in sensor placement optimization in order to get enough information of
machine structure using a small number of sensors and ensure the accuracy and
correctness of condition monitoring.

14
 Chapter 1: Introduction

1.2.3 Review of Fault Diagnosis and Prognosis

During a system failure, only a small fraction of the downtime is spent to maintain
or repair the components that cause the fault. Up to 80% of that is spent to locate
the source of the fault [Kegg, 1984]. In case of complex installation such as
automotive manufacturing plant, one minute downtime may cause as high as
$20,000 cost [Spiewak et al., 2000]. Early fault diagnosis is crucial for avoiding
major malfunction and massive loss in economy and productivity. In diagnosing
rotating machinery, sound emissions or vibration signals are used to monitor the
performance of the machine and could be used to judge whether the machine is
failure or degrading. Many useful techniques for signal analysis have been applied.
These techniques can be classified into three types: time domain [Chen et al., 2008;
Wang et al., 2010], frequency domain such as Fast Fourier Transform [Corinthios,
1971; Liu et al., 2010; Rai & Mohanty, 2007] and time-frequency domain such as
the Short Time Fourier Transform [Portnoff, 1980], Hilbert-Huang Transform [Yu
et al., 2007], Wigner-ville distribution [Andria et al., 1994; Staszewski et al., 1997;
Wang et al., 2008] and Wavelet Transform [Dongyan Chen & Trivedi, 2005; Lin &
Qu, 2000; Prabhakar et al., 2002; Seker & Ayaz, 2003; Tse et al., 2004; Wu &
Chen, 2006; Wu & Kuo, 2009; Wu & Liu, 2009; Zheng et al., 2002].
Autoregressive model method can also be used to extract features of a machine or
component for fault diagnosis and prognosis [Li et al., 2009]. Wavelet transform is
the best of these tools because short time Fourier transform only provides a
constant time-frequency resolution, and Wigner-ville distribution produced
interface terms on the time-frequency domain in a critical condition [Wu & Chen,
2006]. It has particular advantages for characterizing signals at different
localization levels in time as well as signal processing, image processing, pattern
recognition, seismology and machine fault diagnosis.
After processing vibration signals and extracting the features, the more important
thing is identifying the fault and predicting the remaining useful life. There are
many methods could be used in this area. Support vector machine (SVM) learning
is a popular machine learning application due to its high accuracy and good
generalization capabilities [Saravanan et al., 2008]. Li et al. [Li et al., 2005]
proposed a hidden Markov model (HMM)-based fault diagnosis in speed-up and
speed-down process for rotary machinery. In the implementation of the system, one
PC was used for data sampling and another PC was used for data storage and
analysis. Wu and Chow [Wu & Chow, 2004] presented a self-organizing map
(SOM) based radial-basis-function (RBF) neural network method for induction
machine fault detection. The system was implemented by utilizing a PC and
additional data acquisition equipment. Many methods based on ANN have been
developed for online surveillance with knowledge discovery, novelty detection and
learning abilities [Kasabov, 2001; Markou & Singh, 2003; Marzi, 2004]. ANN,
Fuzzy Logic System (FLS), Genetic Algorithms (GA) and Hybrid Computational
Intelligence (HCI) systems were applied in fault diagnosis and a case of centrifugal
pump was utilized to show how the methods work [Wang, 2002]. Decision tree
method was used to identify fault in of mean shifts in bivariate processes in real
time [He et al., 2011]. Probability based Bayesian network methods was used to
identify vehicle fault which can be used to diagnose single-fault and multi-fault

15
 Chapter 1: Introduction

[Huang et al., 2008]. Lee, et al. [Lee et al., 2006] developed an intelligent
prognostics and e-maintenance system named “Watchdog Agent” with the method
of Statistical matching, and performance signature and Support Vector Machine
(SVM) based diagnostic tool.
There exist some literatures integrating these techniques for fault diagnosis and
prognosis. Momoh and Button integrated FFT and ANN to analyze and identify the
fault of aerospace DC arcing [Momoh & Button, 2003]. Fourier transform and
wavelet transform were integrated to detect and identify the fault of induction
motor using stator current information [Lee, 2011]. Wavelet analysis techniques
and ANN were integrated for fault diagnosis in induce motors [Lee, 2011],
automotive generator [Wu & Kuo, 2009] and gear box [Saravanan &
Ramachandran, 2010] and the results were pretty good. In the PhD work, some
techniques are integrated together to classify and predict fault and further to predict
the remaining useful life. These results can be used to support the maintenance
decision making and optimizing the scheduling.

1.2.4 Review of Maintenance Scheduling Optimization

As mentioned above, PM is a dynamic schedule according to the state of

equipment from continuous and/or periodic inspection. It utilizes the product
degradation information extracted and identified from on-line sensing techniques to
minimize the system downtime by balancing the risk of failure and achievable
profits. Mathematically, the maintenance scheduling problem is a multiple-
constraint, non-linear and stochastic optimization problem. This kind of problem
has been studied for several decades and many kinds of different methods have
been applied to solve it. Two methods for PM optimization had been developed
during 1980s. The first method [Perla, 1984; Walker, 1987] performs cost/benefit
analysis of each analyzed piece of manufacturing equipment. It is based on
identifying important equipment firstly, and then predicting its future performance
with and without changes in the regularly scheduled maintenance program. The
second approach is the Reliability-Centered Maintenance (RCM) [Crellin, 1986;
Hook et al., 1987; Vasudevan, 1985]. This methodology was adopted from the
commercial air transport industry. It is based on a series of orderly steps, including
identification of system/subsystem functions and failure modes, prioritization of
failures and failure modes (using a decision logic tree), and finally selection of PM
tasks that are both applicable (i.e. have the potential of reducing failure rate) and
effective (i.e. economically worth doing). In the last two decades, many kinds of
intelligent computational methods, such as the artificial neural network method,
simulated annealing method, expert system, fuzzy systems and evolutionary
optimization, have been applied to solve the maintenance scheduling problem and
obtained many very exciting results [Huang, 1998; Miranda et al., 1998; Satoh &
Nara, 1991; Sutoh et al., 1994; Yoshimoto et al., 1993]. And also, with the rapid
development of the evolutionary theory, genetic algorithms (GAs) had become a
very powerful optimization tool and obtained wide application in this area [Arroyo
& Conejo, 2002; Back et al., 1997; Huang et al., 1992; Lai, 1998; Lee & Yang,
1998; Wang & Handschin, 2000]. In recently years, several new intelligent

16
 Chapter 1: Introduction

computational methods such as Ant Colony Optimization (ACO) and Particle

Swarm Optimization (PSO) have been applied in preventive maintenance
scheduling [Benbouzid-Sitayeb et al., 2008; Pereira et al., 2010; Yare &
Venayagamoorthy, 2010].
All the above methods of maintenance scheduling are based on the specified time
periods other than based on the condition of the equipment or facilities. PM is a
good strategy which could be used to improve reliability and increase useful life of
the equipment and reduce the cost of maintenance according to the condition of
machine. When the condition of a system, such as its degradation level, can be
continuously monitored, PM policy can be implemented, according to which the
decision of maintaining the system is taken dynamically on the basis of the
observed condition of the system. Recently, genetic algorithms, Monte Carlo
method, Markov and semi-Markov methods are applied in PM [Amari et al., 2006;
Barata et al., 2001, 2002; BeĄrenguer et al., 2000; Grall et al., 2008; Marseguerra
et al., 2002]. Normally, to make a dynamic PM scheduling, there are main three
tasks as following and will be discussed in Chapter 8.
x Establishing a predictive maintenance model mathematically;
x Finding a suitable optimization method to optimize the predictive
maintenance model;
x Making a dynamic maintenance decision based on the predictive model and
optimization method.

1.3 Contributions

This section provides the overview of scientific contributions to the topic of

Condition-based Maintenance (CBM) especially in data mining approaches. CBM
is a technique that has not yet been implemented on a large scale in industry. Many
companies have installed various sensors on their equipment and used this gathered
information to determine the current health of the system. However, the gathered
information has seldom been used effectively for fault diagnosis and prognosis.
This PhD work tries to find an easy way to implement CBM technique.
The main contributions of this thesis are located in data mining approaches other
than the other techniques such as model based or statistical methods. The
framework contains many relevant aspects, i.e. sensor placement optimization,
signal processing and feature extraction, fault diagnosis and prognosis, and
maintenance scheduling optimization. The contributions of this thesis are described
as follows.
During the PhD work, the framework called Intelligent Fault Diagnosis and
Prognosis System (IFDPS) for Condition-based Maintenance (CBM) is established.
The thesis tried to apply data mining techniques in most of the aspects in this
framework.
Sensor placement optimization is a nontrivial problem of fault diagnosis and
prognosis for equipment. Optimal positions of sensors mounted on equipment can

17
 Chapter 1: Introduction

improve the effectiveness and reliability of condition monitoring and improve the
quality of data collection. This thesis proposes a method for sensor placement
optimization in machine level by combining Finite Element Analysis (FEA) and
Swarm Intelligence, i.e. Particle Swarm Optimization (PSO) and Bee Colony
Algorithm (BCA). The method can find the optimal positions of a number of
sensors.
Techniques of signal processing and feature extraction are crucial for obtaining key
performance information so that the system can diagnose and prognose effectively.
The thesis analyze the vibration signal through traditional methods such as fast
Fourier transform (FFT), short-time Fourier transform (STFT) and some modern
signal analysis techniques such as wavelet transform, etc. These techniques
together feature extraction method such as Principal Component Analysis (PCA)
For fault diagnosis, the thesis combines the methods of signal processing and
feature extraction mentioned above, and some data mining techniques such as
Artificial Neural Network (ANN) and Self-organizing Map (SOM). These methods
can detect and diagnose the fault effectively.
For fault prognosis, the thesis proposes a methodology to predict the indicator of
component fault based on the collected information by sensors and ANN other than
based on the traditional statistics methods. This methodology has already applied to
wind turbine fault prognosis and it works effectively. The method establishes ANN
model for the indicator in normal condition of wind turbine using the history
SCADA which is collected by wind farm operator but not use effectively. Then the
thresholds of different conditions can be set by using the history data with different
extent fault. Finally ANN model can be applied online to monitor the wind turbines
and gives staff earn warning of fault so that they can schedule the maintenance
actions in advance to reduce downtime, production loss and maintenance cost.
For different purpose, the different maintenance models are established. Based on
these models, the maintenance schedule can be optimized by Swarm Intelligence
(SI), i.e. Particle Swarm Optimization (PSO), Ant Colony Optimization (ACO) and
Bee Colony Algorithm (BCA).
The algorithms of PSO, ACO and BCA are improved or modified in order to be
applied to maintenance scheduling optimization.

1.4 List of Scientific Articles

¾ Zhang, Z., Wang, Y. and Wang, K., (2013). Fault Diagnosis and Prognosis
using Wavelet Packet Decomposition, Fourier Transform and Artificial Neural
Network. Journal of Intelligent Manufacturing, vol. 24 (6), pp. 1213-1227 (doi:
10.1007/s10845-012-0657-2).
¾ Zhang, Z., Wang, Y. and Wang, K., (2013). Intelligent Fault Diagnosis and
Prognosis Approach for Rotating Machinery integrating Wavelet Transform,
Principal Component Analysis, and Artificial Neural Networks. International
Journal of Advanced Manufacturing Technology, vol. 68 (1-4), pp. 763-773
(doi: 10.1007/s00170-013-4797-0).

18
 Chapter 1: Introduction

¾ Zhang, Z., and Wang, K., (2014). Wind turbine fault detection based on
SCADA data analysis using ANN. Advances in Manufacturing, 2(1), pp. 70-78
(doi: 10.1007/s40436-014-0061-6).
¾ Liu, Y., Zhang, Z. and Liu, Z., (2011). Customized Configuration for
Hierarchical Products: Component Clustering and Optimization with PSO. The
International Journal of Advanced Manufacturing Technology, 57. pp. 9-12.
¾ Zhang, Z. and Wang, K., (2013). Wind Turbine Fault Detection Based on
SCADA Data Analysis Using ANN. International Workshop of Advanced
Manufacturing and Automation (IWAMA2013), Nov. 27, pp. 323-335.
¾ Wang, K., Sharma, V. and Zhang, Z. (2013). SCADA Data Interpretation for
Condition-based Monitoring of Wind Turbines. International Workshop of
Advanced Manufacturing and Automation (IWAMA2013), Nov. 27, pp. 307-321.
¾ Zhang, Z. and Wang, K., (2012). Sensors Placement Optimization for
Condition Monitoring. Proceedings of International Workshop of Advanced
Manufacturing and Automation (IWAMA2012), June 20-21, pp. 69-76.
¾ Zhang, Z. and Wang, K., (2012). IFDPS-Intelligent Fault Diagnosis and
Prognosis System for Condition-based Maintenance. International Workshop of
Advanced Manufacturing and Automation (IWAMA2012), June 20-21, pp.77-84.
¾ Zhang, Z. and Wang, K. (2010). Application of Improved Discrete Particle
Swarm Optimization (IDPSO) in Generating Unit Maintenance Scheduling.
International Workshop of Advanced Manufacturing and Automation
(IWAMA2010), pp. 79-86.
¾ Zhang, Z. and Wang, K., (2012). Dynamic Condition-Based Maintenance
Scheduling Using Bee Colony Algorithm. Proceedings of International Asia
Conference on Industrial Engineering and Management Innovation (IEMI2012),
Oct. 27-29, pp.1607-1618.
¾ Zhang, Z. and Wang, K., (2011). Fault isolation using self-organizing map
(SOM) ANNs. IET International Conference of Wireless Mobile & Computing,
Nov. 14-16, pp. 425-431.
¾ Wang, K. and Zhang, Z., (2012). Application of Radio Frequency Identification
(RFID) to Manufacturing. SFI-Norman, SINTEF (ISBN 978-82-14-05388-3),
pp. 1-24.
¾ Wang, K. and Zhang, Z. (2011). Intelligent Fault Diagnosis and Prognosis
System (IFDAPS) for Condition-based Maintenance. Trondheim: SINTEF
A17147 (ISBN 978-82-14-05057-8) pp. 1-21.
¾ Wang, K., Sharma, V. and Zhang, Z., (2013). SCADA Data Mining for wind
turbine fault diagnosis and failure prognosis: Principles, Trends, Applications
and Research Areas, Trondheim: SINTEF (ISBN 978-82-14-05496-5), pp. 1-20.
¾ Zhang, Z. and Wang, K. (2012). Fault Diagnosis using Association Rules. In
Wang, K. and Wang, Y. edit: Data Mining for Zero-Defect Manufacturing, pp.
53-75.
¾ Cusanno, R., Zhang, Z. Regattieri, A. and Wang, K., (2012). Apply Particle
Swarm Optimization for Condition-based Maintenance Scheduling. In Wang, K.
and Wang, Y. edit: Data Mining for Zero-Defect Manufacturing, pp. 117-131.
¾ Crucian, D., Zhang, Z. and Wang, K., (2012). Fault Diagnosis and Prognosis
Using Self-organizing Map. In Wang, K. and Wang, Y. edit: Data Mining for
Zero-Defect Manufacturing, pp. 101-115.

19
 Chapter 1: Introduction

1.5 Outline of Thesis

The present thesis is structured in 9 Chapters. Chapter 2 describes the general

structure of framework of IFDPS very briefly. Chapter 3 introduces some
Computational Intelligence (CI) techniques, such as Artificial Neural Network
(ANN), Association Rules (AR), Decision Trees (DT), Particle Swarm
Optimization, Bee Colony Algorithm and Semi-Supervised Learning, which are
applied in the phases of the framework of Intelligent Fault Diagnosis and Prognosis
(IFDPS) for Condition-based Maintenance. Chapter 4 introduces the sensor
strategies very briefly and proposes the data mining methods application in sensor
placement optimization. Chapter 5 describes the techniques of signal processing
and feature extraction. Chapter 6 presents the data mining technology applications
in fault diagnosis. Chapter 7 proposes a fault prognosis method for fault prognosis
based on fault indicator prediction by data mining techniques. Chapter 8 presents
how to apply the Computational Intelligence methods in maintenance scheduling
optimization both for CBM and Preventive maintenance. Finally the conclusions
and future research directions are presented in Chapter 9. The problem of fault
diagnosis and prognosis and CBM are the very hot research points recently. The
present work is focus on the data mining methods for phases of IFDPS.

20
 Chapter 2: Framework of Intelligent Fault Diagnosis and Prognosis Systems
(IFDPS) for CBM

2 Framework of Intelligent Fault Diagnosis and Prognosis

Systems (IFDPS) for CBM

2.1 Introduction

Any operation or process done on machine or its components to enhance the

efficiency of machine before or after the breakdown is called maintenance
[Deshpande & Modak, 2002]. It contains all technical and administrative activities,
including management activities, which have the objective to sustain or recover
equipment state and thus enable it to perform at a required level. The maintenance
cost ranges between 15% (for manufacturing companies) and 40% (for iron and
steel industry) of the cost of the manufactured goods [Mobley, 1990]. In the United
States, this corresponds to more than 200 billion dollars every year. This shows the
importance of maintenance from an economical point of view. Usually, three
different types of maintenance are considered [Chu et al., 1998], that is: corrective
maintenance, preventive maintenance and predictive maintenance. Corrective
maintenance consists in repairing a system only after a breakdown occurred which
contains all maintenance performed in order to repair a failure [Wilson, 2002].
Corrective maintenance is probably the most commonly used approach, but it is
easy to see its limitations. When equipment fails, it often leads to downtime in
production and therefore this approach is often expensive. Preventive maintenance
consists in maintaining the system periodically to prevent breakdown. Statistics of
failures are used to define the period such as every 100 working hours, or every 10
000 km for a car. Because of uncertainty of products, the lifecycle can vary much
even if they are products in the same class. The periodical maintenance may carry
out far before or after the time of the product become failure. Therefore, the
predictive maintenance (also called Condition-based Maintenance) became a much
better alternative maintenance strategy, which consists in starting a maintenance
operation only when required by the state of the system, i.e. when a potential
failure is detected. With the predictive maintenance strategy, the maintenance
action can be done just before the product become failure, and thus it can prolong
the product life without breakdown.
Condition-based Maintenance (CBM) is the use of machinery run-time data to
determine the machinery condition and hence its current fault/failure condition,
which can be used to schedule, required repair and maintenance prior to breakdown
[Vachtsevanos et al., 2006]. To support CBM policy, a framework called
Intelligent Fault Diagnosis and Prognosis System (IFDPS) is established in KDL
for manufacturing systems and processes. This Chapter mainly introduces the
general idea of IFDPS and its functions very briefly.

2.2 Objectives and Benefits

The main objective of IFDPS is to establish a framework to show how to use the
signals, databases, analysis tools and maintenance decision-making techniques for

21
 Chapter 2: Framework of Intelligent Fault Diagnosis and Prognosis Systems
(IFDPS) for CBM

reaching near zero-breakdown in sustainable manufacturing. It is a part of a big

project called SFI Norman (NORMAN - Center for Research-based Innovation).
The final aim of IFDPS is to reach zero-defect manufacturing in which the first
step is to reach zero-breakdown manufacturing. Therefore, there are several
benefits of the framework of IFDPS.
x It can monitor plant floor assets, link the production and maintenance
operation system, acquire data, collect feedback from remote customer site,
and integrate it into upper level enterprise applications, discovery and
generate maintenance knowledge.
x It can monitor the degradation of manufacturing machine and process, and
predict the condition (remaining useful life) of the equipment.
x It can make predictive maintenance decision to prevent occurrence and
development of failures effectively, ensure the safety of equipment and
personnel, and reduce economic lost caused by failure.
x It can use fault diagnosis, performance assessment of level of degrading,
fault prognosis models to reach zero-breakdown performance and further
to reach zero-defect manufacturing, and improve the productivity of a
company.

2.3 Structure of IFDPS

Fig. 2.1 shows the general structure of IFDPS which presents from the machine
degrading, sensors, signal processing, fault diagnosis and prognosis, and
maintenance scheduling optimization. The main tasks performed by IFDPS are the
following:
x Continuous collection of data from different sensors mounted on the
machine includes the information of machine and environment.
x Continuous processing the data collected from sensors in order to get
useful information to evaluate the off-line and on-line health condition of
the machine and also to detect if some symptoms of degradation or
anomalies are present or could become present.
x According to the useful information mentioned above, the condition or the
fault can be identified. If there are any degradation become unaccepted, the
system can tell staff which components or machines and when should be
maintained or repaired.
x According to the condition of the component or machine, the remaining
useful life can be predicted.
x According to the condition of equipment and predicted remaining useful
life, the maintenance action plan can be scheduled by some intelligent
computational optimization algorithm.
The techniques of subtasks are presented in the following sections.

22
 Chapter 2: Framework of Intelligent Fault Diagnosis and Prognosis Systems
(IFDPS) for CBM

Fig. 2.1 Framework of IFDPS

2.3.1 Data Acquisition

Data acquisition is first phase of the IFDPS and is a basis of fault diagnosis and
prognosis and hence is foundation of intelligent Condition-based Maintenance
scheduling. The tasks of this phase are selecting a suitable sensors and optimal
sensor strategy. Sensors and sensing strategies constitute the foundational basis for
fault diagnosis and prognosis. Strategic issues arising with sensor suites employed
to collect data that eventually will lead to online realization of diagnostic and
prognostic algorithms are associated with the type, number, and location of sensors;
their size, weight, cost, dynamic range, and other characteristic properties; whether
they are of the wired or wireless variety; etc. [Vachtsevanos et al., 2006]. Data
collected by transducing devices rarely are useful in their raw form. Such data must
be processed appropriately so that useful information may be extracted that is a
reduced version of the original data but preserves as much as possible those
characteristic features or fault indicators that are indicative of the fault events we
are seeking to detect, isolate, and predict the time evolution of. Thus such data
must be preprocessed, that is, filtered, compressed, correlated, etc., in order to
remove artifacts and reduce noise levels and the volume of data to be processed
subsequently. Furthermore, the sensor providing the data must be validated; that is,
the sensors themselves are not subjected to fault conditions. Once the
preprocessing module confirms that the sensor data are “clean” and formatted
appropriately, features or signatures of normal or faulty conditions must be
extracted. This is the most significant step in the IFDPS architecture whose output
will set the stage for accurate and timely diagnosis of fault modes. The extracted-
feature vector will serve as one of the essential inputs to fault diagnostic algorithms.

23
 Chapter 2: Framework of Intelligent Fault Diagnosis and Prognosis Systems
(IFDPS) for CBM

Following will introduce the two aspects: sensor category and placement
optimization.

2.3.1.1 Classification of Sensors

There are many methods can be used to classify sensors such as measurends,
detection mass used in sensors, materials and applications [White, 1987]. However,
for fault diagnosis and prognosis, we only use the sensors can measure physical
measurends. For condition monitoring, it is generally agreed that two classes of
sensors are making significant inroads into system monitoring for fault diagnosis
and prognosis. The first one refers to classic or traditional transducers aimed at
monitoring mechanical, structural, performance and operational and
electrical/electronic properties that relate to failure mechanisms of mechanical,
structural, and electrical systems. In this category, there are many sensors that
measure fluid and thermodynamic, thermal, and mechanical properties of a variety
of systems or processes—gas turbines, ground vehicles, pumps, aerospace systems,
etc. The second important category refers to sensor systems that are placed almost
exclusively to interrogate and track system properties that are related directly to
their failure mechanisms. The most useful sensors in fault diagnosis and prognosis
are first category which is shown in Fig. 1.3. When the sensors need to be selected,
many aspects such as position, accuracy, ease of fitting and cost, need to be
considered. After the sensors are chosen, the most important this is where to install
the sensors.

2.3.1.2 Sensor Placement Optimization

Researches of sensor placement become very important issues for obtaining as
much as possible information of machines or components using as few as possible
sensors considering efficiency, effectiveness and economic issues. Traditionally,
the sensor are placed to meet control and performance monitoring objectives [Al-
Shehabi & Newman, 2001; Chen & Li, 2002; Faulds & King, 2000; Giraud &
Jouvencel, 1995]. It is instructive to take advantage of such sensors in a fault
diagnosis monitoring scheme because they can provide useful information relating
to fault behaviors of critical system variables. More recently, research on sensor
placement has focused on two different levels: the component level and the system
level. At the component level, attempts have been reported regarding placement at
the component’s range, for example, a bearing or an object in three dimensional
views [Faulds & King, 2000; Naimimohasses et al., 1995]. For complex, large-
scale systems consisting of multiple components/subsystems, a fault may propagate
through several components. With a large number of possible sensor locations,
selection of an optimal location, as well as the number and types of sensors, poses
a challenging problem that must be addressed at the system level.
IFDPS optimize the distribution of sensor placement in both component level and
system using some Swarm Intelligence (SI) such as Particle Swarm Optimization
(PSO) and Bee Colony Algorithm (BCA). For component level, the structure is
analyzed using Finite Element Method (FEM) which information can be used to
optimize the sensor placement using SI. For system level, the information transmit

24
 Chapter 2: Framework of Intelligent Fault Diagnosis and Prognosis Systems
(IFDPS) for CBM

flaw is analyzed which information can also be used to optimize the sensor
placement.

2.3.2 Signal Preprocessing and Feature Extraction

Generally, there are two steps to deal with the signals from sensors. The one is
signal preprocessing which is intended to enhance the signal characteristics that
eventually may facilitate the efficient extraction of useful information that is the
indictors of the condition of monitoring component or subsystem. The tools in this
category include filtering, amplification, data compression, data validation, and de-
noising which generally aim at improving the signal-to-noise ratio. And the other is
extracting features from preprocessed signals that are characteristic of an incipient
failure or fault. Generally, the features can be extracted from three domains: time
domain, frequency domain and time-frequency domain. All possible signal
preprocessing and feature extraction methods are shown in Table 2.1 and which
features could be selected depend on the real machines or system. All these kinds
of methods are selectable in IFDPS and which methods are applied can be decided
by real machine or system analysis. What’s more, in order to express the enough
information to express the condition, the methods in the table can be combined
together to be indicators of the condition.

Table 2.1 The Methods of Signal Pre-process and Signal Process.

Signal process
Signal
Preprocessing Wavelet
Time Domain Frequency Domain
Domain
Filter, Mean, RMS,
Amplification, Continues Fourier
Shape factor, Transform (CFT), Discrete
Signal Skewness, Kurtosis, Fourier Transform (DFT),
Conditioning, Wavelet
Crest factor, Fast Fourier Transform Transform
Extracting Weak (FFT),
Entropy Error, (WT) and
Signals, De-noising
Entropy estimation, Wigner-ville Distribution Wavelet Packet
Vibration Signal (WVD) and (WP)
Compression and Histogram Lower
and Short Time Fourier
Time Synchronous Transformation (STFT)
Averaging (TSA) Histogram upper

2.3.3 Fault Diagnosis and Identification

Fault diagnosis strategies have been developed in recent years and have found
extensive utility in a variety of application domains. The enabling technologies
typically fall into two major categories: model based and data-driven, as shown in
Fig. 2.2. Model-based techniques rely on an accurate dynamic model of the system
and are capable of detecting even unanticipated faults. They take advantage of the
actual system and model outputs to generate a ‘‘discrepancy’’ or residual, as it is
known, between the two outputs that is indicative of a potential fault condition. On

25
 Chapter 2: Framework of Intelligent Fault Diagnosis and Prognosis Systems
(IFDPS) for CBM

the other hand, data-driven techniques often address only anticipated fault
conditions, where a fault ‘‘model’’ now is a construct or a collection of constructs
such as neural networks, expert systems, etc. that must be trained first with known
prototype fault patterns (data) and then employed online to detect and determine
the faulty component’s identity.

Fig. 2.2 Model-based and Data-driven Fault Diagnosis Techniques

IFDPS focus on the data-driven techniques and hybrid techniques. If the historical
data can be obtained easily, the data-driven is very good to identify the fault and
evaluate the condition. When only part of historical can be obtained, the hybrid
techniques which combine the data-driven techniques and model-based techniques
can be used to evaluate the condition of machine effectively. The semi-supervised
learning method also can be used to evaluate condition and identify fault when
only part of historical data is available and it is very effective. All these techniques
are selectable according to the real manufacturing system analysis.

2.3.4 Fault Prognosis and Remaining Useful Life Evaluation

Generally speaking, current prognostic approaches can be classified into three

basic groups: physical model prognostics, stochastic model prognostics and data-
driven model prognostics.
Since physical models usually are based on a physical mechanism (failure
mechanism) or process (failure process), they are valid for all problems where the
process/mechanism leads to a failure. Sometimes, they are restricted to specific
types of components. Typical for physical models is that the input parameters have
a clear meaning and represent real (and often measurable) quantities or natural,
physical or material constants. Thus, the models provide a clear understanding
about the model input and the output, resulting in a so-called white-box model.
Therefore, physical models are appealing for those who wish to get better
understanding of the mechanisms and processes leading to failure [Loucks et al.,
2005]. Physical models are in particular useful for design improvement.

26
 Chapter 2: Framework of Intelligent Fault Diagnosis and Prognosis Systems
(IFDPS) for CBM

In a first step, a physical model must be established if a good model is not available.
This can be a challenging work and requires good knowledge about the problem
that is modelled. However, once a good model is available, it can be applied to all
comparable problems where good estimates or measurements of the model input
parameters are available. Since the processes in real world may be quite
complicated and may be affected by many mechanisms and effects, one usually has
not the possibility to take all of them into account. Thus, a physical model may be
restricted to include the main mechanisms and main effects only.
Physical models are often empirical, which means they are based on observation or
experiments. Physical models can basically be used for all kinds of predictions,
both long-term and short-term, depending on what they are designed for.
Most stochastic can be applied for many different problems. An advantage of
stochastic models applied to lifetime models are of general nature and prediction is
that both an estimate of the mean lifetime and various estimates of uncertainty can
be established, such as variance of the lifetime, confidence intervals for parameters
and predictions, etc. Parameter estimation in stochastic modelling is based on the
observation of the model output. Thus, observations of the model output, such as
observations of lifetime or degradation, are usually collected as basis for parameter
estimation. When possible, one should fit different stochastic models to the data
and choose the model that gives the best prediction. Many techniques exist to
choose the best model and to check the goodness of fit (e.g. p-value, confidence
intervals, comparison of maximum likelihood values and various graphical
methods such as probability plots). As alternative to data collection, expert
judgement can be used for parameter estimation. There exist different techniques
for expert judgement, e.g. [Cooke, 1992]. Stochastic models can basically be used
for both short-term and long-term predictions. However, for lifetime prediction,
they are mostly used to make medium and long-term predictions. Furthermore,
they are often used in system modelling or as input in other models (such as
maintenance and optimization models) where the main interest is in long- term
averages (such as failure rates). They can also successfully be applied for
comparing and explaining the lifetime influence of different designs or other
factors either by looking on the results from different samples or by incorporating
explanatory variables in the model.
However, in the absence of a reliable and accurate system model, and statistical
data, another approach to determine the remaining useful life is to monitor the
trajectory of a developing fault and predict the amount of time until the developing
fault reaches a predetermined level requiring action, which is the so called data-
driven prognostic method. Data-driven techniques utilize monitored operational
data related to system health. They can be beneficial when understanding of first
principles of system operation is not straightforward or when the system is so
complex that developing an accurate alternative model is prohibitively expensive.
An added value of data-driven techniques is their ability to transform high-
dimensional noisy data into lower dimensional information useful for decision-
making [Dragomir et al., 2007]. Furthermore, recent advances in sensor technology
and refined simulation capabilities enable us to continuously monitor the health of
operating components and manage the related large amount of reference data.

27
 Chapter 2: Framework of Intelligent Fault Diagnosis and Prognosis Systems
(IFDPS) for CBM

Many data-driven models can be classified as black-box models because the

relation of input and output variables and the model parameters is unclear in such
types of models. Parameter estimation in black-box models is often based on
learning and training. Thus, the models require data, and often data covering a time
period where a failure was observed, in order to make a prediction of the lifetime.
Learning can be based on data from a situation identified as normal. Then, all
situations that are different from the normal situation may be defined as abnormal
and (potentially) erroneous. Such an approach is appealing for diagnostic
applications, because the observation of failures is not required. However, this
approach is not sufficient for making predictions of the remaining lifetime.
Data-driven models are mostly suitable for making short-term predictions when the
component reaches the end of life and when a potential failure becomes apparent in
monitoring data. Since there are many models and methods in the field of AI, that
in addition often are quite different, it is difficult to make general statements about
models properties and the ways of parameter estimation. Many models can be
considered as black-box models. Some others however, as for example expert
systems, are white-box models where the internal model logic is based on expert
knowledge.
IFDPS evaluate the remaining useful life by data-driven techniques because of
physical or mathematical model absence. Traditional prognostic techniques are to
find the relationship between the remaining useful life and time of the machine or
component has been used. IFDPS try to find the relations between the remaining
useful life and the condition of machine or component. Fig. 2.3 shows an example
of this relationship. When the condition is identified, the remaining useful life can
be predicted with a standard deviation.

0.1

0.09

0.08

0.07

0.06

0.05

0.04

0.03

0.02

0.01

0
0 50 100 150 200 250 300 350 400 450 500

Fig. 2.3 Remaining Useful Life Distribution for Each Condition

After the condition of the component is determined, the remaining useful life can
be evaluated according to the condition. Most current RUL estimation methods are
based on the event data or condition monitoring data which want to find the
relationship between RUL and time the component used or RUL and feature values

28
 Chapter 2: Framework of Intelligent Fault Diagnosis and Prognosis Systems
(IFDPS) for CBM

[Lee et al., 2006; Si et al., 2011]. The method of Fig. 2.3 tries to find the
relationship between the RUL and the condition of a component that is evaluating
RUL by the condition and RUL distribution for each condition. The distributions of
RUL are obtained by the statistical methods. For example, if the condition of a
component is 0, the remaining useful is 350h with a certain standard deviation.
When the condition is 1.0, the RUL is much closed to 0 which means the
component has to be maintained or repaired. From Fig. 2.3 the RUL distribution
become narrow that means the RUL evaluation is more accuracy when the
condition closed to failure. Therefore the confidence value of RUL increases with
the condition deterioration.
IFDPS also propose another method to predict the RUL by establishing ANN
model for machines in normal condition and set thresholds in several different
levels. This method has applied in real industries such as wind power industry
which is described in Chapter 7.

2.3.5 Maintenance Scheduling Optimization

The functions of the maintenance are finding out fault status of maintenance object
and maintenance effect, and selecting right maintenance policy to achieve expected
maintenance effect. The purpose of it is making maintenance decision based on
current information to prevent occurrence and development of failure effectively,
ensure the security of equipment and personnel, and reduce economic lost caused
by failure. Maintenance scheduling optimization is a kind of NP problem and the
SI algorithms could be a very good technique to solve this kind of problem. IFDPS
apply Genetic Algorithm (GA), Particle Swarm Optimization (PSO), Ant Colony
Optimization (ACO) and Bee Colony Algorithm (BCA) and try to find the optimal
dynamic predictive maintenance scheduling. All these kind methods are selectable
in IFDPS to solve maintenance scheduling optimization problems.

2.4 Summary

IFDPS (Intelligent Fault Diagnosis and Prognosis System) for Condition-based

Maintenance is developed to monitor the manufacturing system and process, and to
classify and predict faults and states, and to evaluate remaining useful life. Based
on this system, the suitable maintenance actions can be made before any failure
happens to ensure the security of equipment and personnel, and reduce economic
lost caused by failure. In this framework, the suitable sensors should be selected to
monitor the manufacturing system firstly, and then the collected data from the
sensors is processed. For signal processing, the parameters of time domain,
frequency domain and time-frequency domain can be used to process the signals
and extract features as indicators of machines’ condition. The condition of
machines can be identified and predicted based on the extracted features from the
real time signals. The remaining useful life can be evaluated based on the condition
of machine and finally the maintenance decision can be made using some Swarm
Intelligence algorithms.

29
Chapter 2: Framework of Intelligent Fault Diagnosis and Prognosis Systems
(IFDPS) for CBM

30
 Chapter 3: Data Mining Techniques for IFDPS

3 Data Mining Techniques for IFDPS

3.1 Introduction

There are many aspects can be researched for the Framework of IFDPS. The
volumes of data from sensors and processing procedure become tremendous filling
the computers and networks. Sometimes, the data is too huge and too complicated
to analyze effectively and thus how to get the useful information from these data
becomes very significant point. This PhD work mainly focus on the application of
Data Mining (DM) techniques in all processes of IFDPS such as sensor placement
optimization, fault diagnosis, fault prognosis and maintenance scheduling
optimization. There are already some researches in these areas but most of these
researches focus on one process. DM technology has recently become a hot topic
for decision-makers because it provides valuable, hidden business and scientific
“Intelligence” from a large amount of historical data. It is a kind of methods to
extract information and knowledge from recorded data. This Chapter describes
some DM techniques used in the PhD work.
Data mining can be defined as the analysis of (often large) observational data sets
to find unsuspected relationships and to summarize the data in novel ways that are
both understandable and useful to the data owner [Hand et al., 2001]. It is the entire
process of applying computer-based methodology, including new techniques for
knowledge discovery, from data. It draws ideas and resources from multiple
disciplines, including machine learning, statistics, database research, high
performance computing and commerce. This explains the dynamic, multifaceted
and rapidly evolving nature of the data mining discipline. Generally, there are two
main goals of data mining: prediction and description. Prediction involves using
some variables or fields in the dataset to predict unknown or future values of other
variables of interest. Description focuses on finding patterns describing the data
that can be interpreted by humans. Therefore, the data mining activities can be
classified into two categories: predictive data mining which produces the model of
the system described by the given dataset, and the descriptive data mining which
produces new, nontrivial information based on the available dataset. The main
tasks of DM techniques are [Kantardzic, 2003]:
x Classification – discovery of a predictive learning function that classifies a data
item into one of several predefined classes.
x Regression – discovery of predictive learning function, which maps a data item
to a real-value prediction variable.
x Clustering – a common descriptive task in which one seeks to identify a finite
set of categories or clusters to describe the data.
x Summarization – an additional descriptive task that involves methods for
finding a compact description for a set of data.
x Dependency Modeling – finding a local model that describes significant
dependencies between variables or between the values of a feature in a data set
or in a part of a data set.

31
 Chapter 3: Data Mining Techniques for IFDPS

x Change and Deviation Detection – discovering the most significant changes in

the data set.
To carry out these tasks, many DM techniques are available so far and more
techniques will appear in the future. This Chapter will introduce some DM
techniques used in the IFDPS framework.

3.2 Artificial Neural Networks (ANN)

The pattern classification theory has become a key factor in fault diagnosis and
prognosis. Some classification methods for equipment performance monitoring use
the relationship between the type of fault and a set of patterns which is extract from
the collected signals without establishing explicit models. Currently, ANN is one
of the most popular methods in this domain. ANN is a model that emulates a
biological neural network [Wang, 2005]. The origin of ANN can be traced back to
a seminar paper by McCulloch and Pitts [McCulloch & Pitts, 1943] that
demonstrated a collection of connected processors, loosely modeled on the
organization of brain, could theoretically perform any logical or arithmetic
operation. Then, the development of ANN techniques is very fast which is
extensive to many categories containing Back-propagation (BP), Self-organization
Mapping (SOM) and Radial Basis Function (RBF), etc. The application of artificial
neural network models lies in the fact that they can be used to infer a function from
observations. This is particularly useful in applications where the complexity of the
data or task makes the design of such a function by hand impractical. This
attribution is very nontrivial in diagnostic problems. BP neural network is a main
type of ANN used to solve fault diagnosis and prognosis problems.
ANN can deal with complex non-linear problem without sophisticated and
specialized knowledge of the real systems. It is an effective classification
techniques and low operational response times needed after training. The
relationship between the condition of component and the features is not linear but
non-linear. BP neural network does not need to know the exact form of analytical
function on which the model should be built. This means neither the functional
type nor the number and position of the parameters in the model-function need to
know. It can deal with multi-input, multi-output, quantitative or qualitative,
complex system with very good abilities of data fusion, self-adaptation and parallel
processing. Therefore, it is very suitable to be selected as a method of fault
diagnosis and prognosis. There are many papers dealing with the use of ANN and
most of their contributions are ANN training efficiency and strategies for ANN
itself. ANN in IFDPS will be used to detect and predict the condition of machines
with other techniques such as wavelet analysis and Fourier transform. Two ANN
techniques of Supervised Back-Propagation (SBP) and Self-Organizing Map are
introduced in this subsection.

3.2.1 Supervised Learning ANNs

BP neural network which is the most widely used neural network model currently
was proposed by Rumelhart and McCelland in 1986 [Rumelhart et al., 1986]. It is a

32
 Chapter 3: Data Mining Techniques for IFDPS

multilayer feed-forward network usually containing the input layer, hidden layer,
and output layer (Fig. 3.1), which trained by an error back propagation algorithm.
The biggest advantage of ANNs trained by back propagation is that there isn’t need
to know the exact form of analytical function on which model should be built. So
it’s not necessary have neither the function type not even the number and position
of the parameters in the model function. Moreover, BP network can learn and store
a lot of input-output model mapping without mathematical equations which
describing this mapping. The learning method of BP is the steepest descent method
which is adjusting the weights and thresholds of the network to minimize the sum
of squared errors. The general procedure of BP network training can be
summarized as follows [Wang, 2005].
1) Initialize the weights to small random vales (-1, 1);
2) Select a training vector pair (input and the corresponding desired output)
from the training set and present the input vector to the inputs layer of the
ANN;
3) Calculate the actual outputs (forward phase);
4) Adjust the weights to reduce the difference according to the error between
actual output and target (backward phase);
5) Return to step 2 and repeat for each pattern p until the total error has
reached an acceptable level;
6) Stop.
The backward phase and forward phase are described separately in this section.

Fig. 3.1 A BP Neural Network with Single Hidden Layer

3.2.1.1 Forward Phase

Fig. 3.1 shows a BP network with signal hidden layer which is used to explain how
the BP network works. In the figure, there are m nodes of input layer, h nodes of
hidden layer and n nodes of output layer. Weight connections between input layer
ith (i 1, 2,", m) node and hidden layer j th (i 1, 2,", h) node are denoted as v ji ,
while Weight connections between hidden layer j th (i 1, 2,", h) node and output

33
 Chapter 3: Data Mining Techniques for IFDPS

layer k th (i 1, 2,", n) are denoted as wkj . xi represents the i th input value, y j

represents the output value of j th node of hidden layer, z k represents the k th output,
and tk represents the target value of the k th output. The following terms are now
defined:
m
Hj ¦v
i 1
ji xi i 1, 2 " , m; j 1, 2," h (3.1)

h
Ik ¦w kj yj k 1, 2" , n; j 1, 2," h
j 1
(3.2)
where H j is the combined or net input to hidden layer node j , while I k is the net
input to the node k of the output layer. The output for each node of hidden layer
and output layer can be obtained as following respectively:
yi f (H ) j 1, 2," , h (3.3)
zk f (Ik ) (3.4)
where f is an activation function. Finally, the output of node k of output layer can
be rewritten as:
ª h m
º
zk f « ¦ wkj f (¦ v ji xi ) » (3.5)
¬j 1 i 1 ¼

3.2.1.2 Backward Phase

After calculating outputs of all nodes of output layer, the backward phase can be
started to calculate according to Gradient Decent Learning [Wang, 2005]. The
update rule for output layer node can be obtained:
wE
'wkj K KG k yi K (tk  zk ) f '( I k ) yi (3.6)
wwkj

where K is a constant often called the learning rate. Then the update rule for hidden
layer nodes can be obtained as:
n n
'v ji KG j xi K xi f '( H j ) ¦ G k wkj K xi f '( H j ) ¦ [(tk  z k ) f '( I k ) wkj ] (3.7)
k 1 k 1

Then, all weights wkj and v ji can be updated according to Eq. (3.6) and Eq. (3.7)
respectively as following:
wkjnew wkjold  'wkj wkjold  K y j (tk  zk ) f '( I k )
(3.8)
n

ji  K xi f '( H j ) ¦ [(t k  z k ) f '( I k ) wkj ]

v new
ji ji  ' v ji
v old v old (3.9)
k 1

34
 Chapter 3: Data Mining Techniques for IFDPS

This is the process of forward phase and backward phase in BP network training.
Afterward, the whole training processes can be done according to the steps
descripted above. The objective of ANN training is to obtain suitable weights to
meet the inputs and the targets of training data. After the training of BP network,
for each set of test data or query data, there is a set of output calculated by the final
updated weights. BP network is a very useful model in real application especially
when the real physical model and mathematic model are unavailable. It acts as a
black box, which allows no physical interpretation of its internal parameters and
functions. This propriety is very important to apply BP network in condition
monitoring because most real mathematic models are unavailable. For a specific
application in fault diagnosis and prognosis, after training by features extracted
from processed historic data, the BP network can classify the fault and predict the
states of the monitored components or machine units.

3.2.2 Self-Organizing Map (SOM)

Machine learning is an approach of using data to synthesize programs. In a case
when the data are input/output pairs, it is called supervised learning as BP learning
mentioned above. In a case, where there are no output values and the learning task
is to gain some understanding of the process that generated the data, this type of
learning is said to be unsupervised [Kankar et al., 2011]. The concept of SOM was
introduced by Teuvi Kohonen in 1982 [Kohonen, 1982], and numerous versions,
generalizations, accelerated learning scheme, and application of SOM have been
developed since then. It is a type of Artificial Neural Network that is trained using
unsupervised learning mode to produce a low-dimensional, discretized
representation using the input datasets of the training samples. SOM is the closest
of all Artificial Neural Networks architectures and learning schemes to the
biological neuron network. Its network is composed by only one layer of neurons
arranged in two-dimensional plane with a well-defined topology.
The most important unsupervised ANNs learning algorithm is the Kohonen
competitive learning algorithm, and Fig. 3.2 shows a typically example of Kohonen
map. The neurons on the output layer (also called competitive layer) can find the
organization of relationship among input patterns. The output of each neuron isn’t
connected to all of the other neurons in the plane, but only to a small number that
are topologically close to it. The network map shows the natural relationship
between the patterns, that is, each input neuron is connected to every neuron on the
competitive layer which is organized as two-dimensional grids. The network is
presented by a set of training input patterns without target output patterns. At the
beginning one of the patterns is chosen randomly, and then each neuron in the
input layer of the SOM takes on the value of the corresponding entry in the input
pattern. In the competitive learning, only one neuron in the output layer is selected
after input occurs, regardless of how close the other neurons are from the best one.
This is so-called “Winner takes it all” method.

35
 Chapter 3: Data Mining Techniques for IFDPS

Zij

Fig. 3.2 Kohonen Model of SOM

In generally, the learning process of SOM network can be several steps [Wang,
2005].
1) Initializing the weight vector randomly Zij , the learning rate K and other
relative training parameters.
2) For each input vector, the responses of all neurons in the output layer are
calculated and the winning node U c is selected. The winning node means its
weight Zij best matches the input vector that is the Euclidean Distance is the
smallest among all nodes.
3) After the winning node is selected, identifying the neighborhood around U c ,
that is the set of competitive units close to the winning node. Fig. 3.3 shows the
two examples of a neighborhood around wining node: the one is rectangular
lattice and the other is the hexagonal lattice. The size of the neighborhood
begins with a large enough size and then decreases with the number of iterations
of the network.
4) Updating the weight vectors of node Uc and all nodes in the neighborhood
around it by the following functions:

­°Z j (t )  K (t ) ˜ f ( d c  d i ) ˜ ( x  Z j (t )) j  H (t )
Z j (t  1) ® (3.10)
°̄ Z j (t ) otherwise

t
H (t ) H 0 (1  ) (3.11)
T
T t
K (t ) (K max  K min )  K min (3.12)
T 1

36
 Chapter 3: Data Mining Techniques for IFDPS

where:
t : the current learning epoch;
x: input vector;
T : the total number of learning epoch;
H0 : the initial neighborhood size;
dc  di : the topological distance between the central neuron c and the current
neuron i ;
f : topology dependent function.
H (t ) : the actual neighborhood size in t epoch;
th

Z j (t ) : the weigh vectors of U c and its neighborhood in t epoch;

th

K (t ) :
the learning rate in t th epoch;
5) Updating the learning rate using Eq. (3.12).
6) Reduce the neighborhood function using Eq. (3.11).
7) Loop from 2) to 6) until no noticeable changes of the feature map.

Fig. 3.3 Different Forms of the Neighborhood in SOM Network around U c

SOM network has some advantages and some disadvantages. SOM permits to
cluster data where there is no prior knowledge of the results or of the clustering. It
is able to convert multi-dimensional data clusters into the form of a two-
dimensional grid preserving the topological relationship of the data. It may be used
where there is ample supply of “good normal” data containing some but little bad
or usual data. That is engine monitoring or alarm monitoring. The SOM has very
serious computational disadvantages, which affects the performance of large scale
application running on parallel computers. In order to find which neuron is to be
stimulated, the program has to check all of the neurons. This is a big restriction
when large SOM network are to be trained. Sometimes grid size may need to be
adjusting in response to number of clusters expected.

3.3 Semi-supervised Learning Methods (Manifold Regularization)

Semi-supervised learning (SSL) is halfway between supervised learning and

unsupervised learning. In addition to unlabeled data, the algorithm is provided with

37
 Chapter 3: Data Mining Techniques for IFDPS

some supervision information – but not necessarily for all examples. In this case,
the data set X ( xi )i[ n] can be divided into two parts: the points X l ( x1 ," , xl ) for
which labels Yl ( y1 ," , yl ) are provided, and the points X u ( xl 1 ," , xl u ) , the
labels of which are not known. SSL is very useful in real industry application
especially when the history data are huge but only a small part of them are labeled.
Semi-supervised learning methods fall into ¿ve categories: SSL with generative
models, SSL with low density separation, graph-based methods, co-training
methods, and self-training methods [Blum & Chawla, 2001; Yuan, 2012].
Recently graph-based methods with more applicable assumption have attracted
considerable attention. Speci¿cally, graph-based manifold regularization [Belkin et
al., 2006] exploits the geometric structure of the marginal distribution of the CM
data in the feature space. The incorporation of unlabeled data has demonstrated the
potential for improved accuracy in time series prediction [Wei & Keogh, 2006],
speech recognition [Jansen & Niyogi, 2005], calibration-eơort reduction problem
[Pan et al., 2001]. In this paper, we are looking for a general semi-supervised
classi¿cation framework for fault detection. The manifold regularization based
methods is a good option.
The Manifold regularization combines the ideas of spectral graph theory, manifold
learning and kernel methods in a coherent and natural way to incorporate both the
cluster assumption and the manifold assumption in Reproducing Kernel Hilbert
Spaces (RKHS) regularization framework. In this section, we address the manifold
regularization based SSL framework concisely following the description of [Belkin
et al., 2006]. More details refer to [Sindhwani et al., 2005].
As mentioned above, consider a set of l labelled samples {( xi , zi )}li 1 and a set of u
unlabelled samples { x j }lj ul 1 , where xi , x j  R d are the feature vectors collected
from the input space 0 according to the marginal distribution (0 , and zi  R is
the classi¿ed label determined by the conditional distribution ( ( z | x ) . Manifold
regularization introduces the regularized risk functional as an additional regularizer
that serves to impose this assumption on the learning problem. The learning
problem corresponds to solving the following optimization problem:
1 l
¦ i 1 . ( xi , zi , f )  J A f JI ³
2
f arg min K
’M ,’M (3.13)
f H K l M

which ¿nds the optimal function f in the RKHS space H K of functions f : 0 o R
corresponding to a Mercer kernel K : 0 u 0 o R , e.g. a linear or Gaussian kernel.
The ¿rst term of the regularized risk functional in Eq. (3.13) is de¿ned on the loss
function . measured the discrepancy between predicted value f ( xi ) and actual
label zi . The second term controls the complexity of f in terms of the RKHS norm,
with J A being the RKHS norm regularization parameter. The third term is speci¿c
to manifold regularization and is based on the assumption that the support of (0
forms a compact sub-manifold % . It controls the complexity of f in the intrinsic
geometry of (0 , with J I being the corresponding manifold regularization

38
 Chapter 3: Data Mining Techniques for IFDPS

parameter. The third term is approximated using the graph Laplacian de¿ned on all
l  u labelled and unlabelled examples without using the label information. Hence
the optimization problem can be reformulated as:
1 l JI
¦ . ( xi , zi , f )  J A f fˆ T Lfˆ
2
f arg min  (3.14)
f H K l i1 K
(l  u ) 2

where fˆ ( f ( x1 ), . . . , f ( xl u ))c and L is the Laplacian matrix of a graph that models the
underlying geometric structure.
From the extended Representer theorem [Belkin et al., 2006], the optimal function
can be expressed in the following form:
l u
f ( x) ¦ D K ( x , x)
i 1
i i
(3.15)

When lose function . in Eq. (3.14) is adopt to be the squared loss function
.( xi , zi , f )  zi  f (x) , the Laplacian Regularized Least Squares (LapRLS) algorithm
2

formulates the optimization problem:

1 l JI
¦  zi  f ( x )   J A fˆ T Lfˆ
2 2
f arg min f  (3.16)
f H K l i 1 K
(l  u ) 2

For the LapRLS, the optimal solution in Eq. (3.16) D (D1 ,},Dl u )c is given by:
J Il
D ( JK  J AlI  LK ) 1 Z (3.17)
(l  u ) 2

where K is the (l  u) u (l  u) Gram matrix over all labelled and unlabelled samples,
Z is an (l  u ) -dimensional label vector given by Z ( z1 , }, zl , 0, }, 0)c , and
J (1, } ,1, 0, } , 0) is an (l  u ) u (l  u ) diagonal matrix with the ¿rst l diagonal
entries being 1 and the rest being 0.
When lose function . in Eq. (3.14) is adopt to be the hinge loss function
. ( xi , zi , f ) 1  zi f ( x )  , the algorithm formulates the Laplacian Support Vector
Machines (LapSVM). Please refer to [Belkin et al., 2006] in details. The manifold
regularization algorithms SSL can be summarized as following [Belkin et al.,
2006]:
Input: l labelled examples {( xi , yi )}li 1 , u unlabelled examples { x j }lj ul 1 .

Output: Estimated function f : \ n o \ .

Step 1: Construct data adjacency graph with ( l  u ) nodes using, for example, k -
nearest neighbours or a graph kernel. Choose edge weights wij , for example,
2
 xi  x j

binary weights or heat kernel weights wij e 4t .

Step 2: Choose a kernel function K ( x, y ) . Compute the Gram matrix Kij K ( xi , x j ) .

39
 Chapter 3: Data Mining Techniques for IFDPS

Step 3: Compute graph Laplacian matrix: L D  W where D is a diagonal matrix

¦
l u
given by Dij j 1
Wij .

Step 4: Choose J A and J I .

Step 5: Compute D * using Eq. (3.17) for square loss (Laplacian RLS).

¦
l u
*
Step 6: Output function f (x) i 1
Di* K ( xi , x) .

3.4 Association Rules

Association rules are one of the major techniques of data mining and it is perhaps
the most common form of local-pattern discovery in unsupervised learning systems.
Association rules mining retrieve all possible interesting associations (patterns,
relationships or dependencies) in large sets of the data items which are stored in the
form of transactions that can be generated by an external process, or extracted from
relational database or data warehouse. Due to good scalability characteristics of the
association rules algorithm and the ever-growing size of the accumulated data,
association rules are an essential data mining tool for extracting useful knowledge
from database. The most important thing in this case would be a rule that is
interesting, that tells you something about your database that you have not already
known and probably weren’t able to explicitly articulate.

3.4.1 Market-basket Analysis

Market-basket analysis is one of the most intuitive applications of association rules
which strive to analyze customer buying patterns by finding associations between
items that customers put into their basket. For example, customers visit to a
grocery store or an online purchase who may buy milk and bread together and even
that some particular brands of milk are more often bought with certain brands of
bread. That means for each customer, there is a typical transaction. Retails
accumulate huge collections of transactions by recording business activities over
time. Then, the transactions database is analyzed to find sets of items, or itemsets
that appear together, such as bread and milk, in many transactions. These and other
more knowledge can be used to maximize the profits by helping to design
successful marketing campaigns and customizing store layout. A number of
association rules can be generated from the market basket database as shown in Fig.
3.4.

40
 Chapter 3: Data Mining Techniques for IFDPS

Fig. 3.4 Application of Association Rules in Market-basket Analysis

From the database of sales transactions, the important associations among items
such that the presence of some items in a transaction will imply the presence of
other items in the same transactions can be discovered. Let I ^i1 , i2 , " , im ` be a
set of literals which called items. Let D (database) be a set of transactions where
each transaction T is a set of items such that T Ž I . Note that the quantities of the
items bought in a transaction are not considered which means each item is a binary
variable indicating whether an item was bought or not. Each transaction is
associated with an identifier called a transaction identifier ( TID ). An example of
the model for such transaction database is given in Table 3.1.

Table 3.1 A Model of a Simple Transaction Database

TID Items
001 Apples, Celery, Diapers
002 Beer, Celery, Eggs
003 Apples, Beer, Celery, Eggs
004 Beer, Eggs
A transaction T is said to contain a set of items X if and only if X Ž T . A
transaction rules implies the form X Ÿ Y where X Ž I , Y Ž I , and X  Y I . The
rule X Ÿ Y holds in the transaction set D with confidence c if c% of the
transaction in D that contain X also contain Y . The rule X Ÿ Y has support s in
the transaction set D if s% of the transaction in D that contain X * Y . Here, two
important concepts are defined bellowing:
—Support, which indicates the frequency (probability) of the entire rule with the
respect to D . It is defined as a ratio of the number of transactions containing A and
B to the total number of transactions (the probability of both A and B co-
occurring in D ):
^T  D|A * B Ž T`
support  A Ÿ B  =P  A * B  = (3.18)
D

41
 Chapter 3: Data Mining Techniques for IFDPS

—Confidence, which indicates the strength of implication in the rules. It is defined

as ratio of the number of transactions containing A and B to the number of
transaction s containing A (Conditional probability of B given A ):
^T  D|A * B Ž T`
confidence  A Ÿ B  =P  A|B  = (3.19)
T  D|A Ž T

It is often desirable to pay attention to only those rules that may have a reasonably
large support. Such rules with high confidence and strong support are referred to as
strong rules. The task of mining association rules is essentially to discover strong
association rules in large databases.

3.4.2 Mining Association Rules Steps

The following four steps are used to generate association rules:
1) Prepare input data in the transaction format;
2) Select items of interest, i.e. itemsets;
3) Calculate support counts to evaluate whether the selected itemsets are
frequency which depend on whether the support s is above the
predetermined minimum threshold;
4) Generate the association rules for the database that have confidence c
above the predetermined minimum threshold using the large itemset.
The computational performance of an association-rule mining is determined by the
second and the third step above. Then, the large itemsets are identified; the
corresponding association rules can be derived in a straightforward manner.
Efficient counting of large itemsets is thus the focus of most mining algorithm, and
many efficient solutions have been designed to address previous criteria. Therefore,
the following discussion concentrates on these two steps.

3.4.3 The Apriori Algorithm

The simplest way to calculate frequency itemsets is to consider all possible
itemsets, compute their support, and check whether they are higher than the
predetermined minimum threshold. The number of test of this method grows
exponentially with the number of the items, and thus for large problems the
computations would take an unacceptable long time. This reasoning resulted in the
development of the Apriori algorithm.
The Apriori algorithm uses prior knowledge about an important property of
frequent itemsets. The Apriori property of an itemset says that all nonempty
subsets of a frequent itemset must also be frequent. In other words, if a given
itemset is not frequent, any superset of this itemset will also be not frequent,
because it cannot occur more frequently than the original itemset. The simplest
superset of an itemset is the itemset with one more added item. The Apriori
property is used to reduce the number of itemsets that must be searched to find
frequent itemsets. The association-rule mining algorithm, the Apriori algorithm,
performs the iterative search through itemsets, starting with 1-itemsets, through 2-
itemsets, 3-itemsets, etc. In general, it finds and processes k-itemsets based on the

42
 Chapter 3: Data Mining Techniques for IFDPS

exploration of ( k – 1 )-itemsets. Using the Apriori property, the Apriori algorithm is

shown in Fig. 3.5.
Based on the Apriori property, in each iteration, k -itemsets that do not satisfy the
minimum support are removed and only the remaining k -itemsets are used to
generate itemsets for the next, k  1 , iteration. This process substantially reduces
the number of itemsets that have to be checked if they are frequent. The only
unknown in implementing the Apriori algorithm is how to perform generation of
Ck , which is a set of k -itemsets based on Lk  1 . These k -itemsets are checked
against the minimum support to derive Lk . The Ck is generated in two steps:
1) For each frequent itemset ( FI ) from Lk  1 , find each item i that does not
belong to FI , but belongs to some other frequent ( k  1 )-itemset in Lk 1 .
Add i to FI to create a k -itemset. Remove duplicate k -itemsets after all
additions for all ( k  1 )-itemsets are finished.
2) If frequent ( k  1 )-itemsets from Lk  1 have ( k  2 )-items in common,
then create a k -itemset by adding the two different items to ( k  2 )
common items.

Fig. 3.5 The process of Aprioi Algorithm

3.4.4 Generating Association Rules from Frequent Itemset

The last step of the four that are used to generate single-dimensional association
rules is to generate association rules from frequent itemsets. The association-rule
mining algorithm requires the generation of strong rules, i.e., those that satisfy both
minimum confidence and minimum support. The minimum support level is

43
 Chapter 3: Data Mining Techniques for IFDPS

guaranteed by using frequent itemsets, and thus we need only to generate the rules
and prune those rules which do not satisfy the minimum confidence. The
confidence can be defined based on the corresponding support values as follows:
support (A * B)
confidence  A Ÿ B  =P  A|B  = (3.20)
support (A)

Where support ( A * B) is the number of transactions in D containing the itemset

A * B , and support ( A) is the number of transactions in D containing the itemset A .
Based on this formula, each frequent itemset FI is used to generate association
rules in two steps:
a) Generate all nonempty subsets of items, Y , of FI ;
support_count(FI)
b) For each Y, output the rules Y Ÿ ( FI  Y ) if the value of
support_count(Y)
is larger than minimum confidence threshold.
To demonstrate the Apriori algorithm in action, we generate association rules from
the transactional data given in Table 3.1. Fig. 3.6 shows the process of how to
select the frequent itemsets. Finally, the association rules can be derived as
bellowing from the generated frequency 3-itemset ^ B, C , E` with support=50%
which also have to satisfy the minimum confidence=60%.
B and C ֜ E with support = 50% and confidence = 2 / 2 = 100%
B and E ֜ C with support = 50% and confidence = 2 / 3 = 66.7%
C and E ֜ B with support = 50% and confidence = 2 / 2 = 100%
B ֜ C and E with support = 50% and confidence = 2 / 3 = 66.7%
C ֜ B and E with support = 50% and confidence = 2 / 3 = 66.7%
E ֜ B and C with support = 50% and confidence = 2 / 3 = 66.7%

TID Transactions Support Support

Itemset Itemset
001 Apples (A), Celery (C), Diapers (D) count count
002 Beer (B), Celery (C), Eggs (E) A 2 A 2
003 Apples (A), Beer (B), Celery (C), Eggs (E) B 3 B 3
004 Beer (B), Eggs (E) C 3 C 3
D 1 E 3
E 3

Support
Support Support Itemset
Itemset Itemset count
count count
Support A,B 1
Itemset A,B,C 1 A,C 2
count A,C 2
A,B,E 1 B,C 2
B,C,E 2 A,E 1
A,C,E 1 B,E 3 B,C 2
B,C,E 2 C,E 2 B,E 3
C,E 2

Fig. 3.6 Example Generation of Association Rules using Apriori Algorithm

44
 Chapter 3: Data Mining Techniques for IFDPS

3.4.5 Improving the Efficiency of the Apriori Algorithm

Since the amount of the processed data in mining frequent itemset tends to be huge,
it is significant to devise efficient algorithms to mine such data. Our basic Apriori
algorithm scans the database several times depending on the size of largest frequent
itemsets. Several refinements have been proposed that focus on reducing the
number of database scan, the number of candidate itemsets counted in each scan, or
both.

3.4.5.1 Partition-based Apriori

Partition-based Apriori is an algorithm that requires only two scans of the
transaction database. The database is divided into disjoint partitions, each small
enough to fit into available memory. In the first scan, the algorithm reads each
partition and computes local frequent itemsets on each partition. The frequent local
itemsets may or may not be frequent in transaction database D , but any itemset
that is potentially frequent in D must be frequent in at least one subset. Therefore,
local frequent itemsets from all subsets become candidate itemsets for D . The
collection of all local frequent itemsets is referred to as global itemsets with respect
D . In the second scan, the algorithm counts the support of all global frequent
itemsets toward the complete database D . Then comparing between the support
values and minimum support threshold and find out which of the global candidate
itemsetsare frequent itemsets. The process of partition-based Apriori algorithm to
select frequent itemsets is shown in Fig. 3.7.



Fig. 3.7 Generation of Frequent Itemsets using Partition-based Apriori

3.4.5.2 Sampling
As the database size increases, sampling appears to be an attractive approach to
data mining. Sampling generates association rules based on a sampled subset of
transactions in D . In this case, a randomly selected subset S of D is used to
search for the frequent itemsets. The generation of frequent itemsets from S is
more efficient (faster), but some of the rules that would have been generated from
D may be missing, and some rules generated from S may not be present in D , i.e.,
the “accuracy” of the rules may be lower. Usually the size of S is selected so that
the transactions can fit into the main memory, and thus only one scan of the data is
required (no paging). To reduce the possibility that we will miss some of the

45
 Chapter 3: Data Mining Techniques for IFDPS

frequent itemsets from D when generating frequent itemsets from S , we may use
a lower support threshold for S as compared with the support threshold for D .
This approach is especially valuable when the association rules are computed on a
very frequent basis.

3.4.5.3 Hashing
Hashing is used to reduce the size of the candidate k-itemsets, i.e., itemsets
generated from frequent itemsets from iteration k  1 , Ck , for k ! 1 . For instance,
when scanning D to generate L1 from the candidate 1-itemsets in C1 , we can at
the same time generate all 2-itemsets for each transaction, hash (map) them into
different buckets of the hash table structure, and increase the corresponding bucket
counts. A 2-itemset whose corresponding bucket count is below the support
threshold cannot be frequent, and thus we can remove it from the candidate set C 2 .
In this way, we reduce the number of candidate 2-itemsets that must be examined
to obtain L2 .

3.4.5.4 Transaction removal

Transaction removal removes transactions that do not contain frequent itemsets. In
general, if a transaction does not contain any frequent k -itemsets, it cannot contain
any frequent ( k  1 ) itemsets, and thus it can be removed from the computation of
any frequent t -itemsets, where t ! k .

3.5 Swarm Intelligence

Swarm intelligence (SI), which is an Artificial Intelligence (AI) discipline, is

concerned with the design of intelligent multi-agent systems by taking inspiration
from the collective behavior of social insects such as ants, termites, bees, and
wasps, as well as from other animal societies such as flocks of birds or schools of
fish. Colonies of social insects have fascinated researchers for many years, and the
mechanisms that govern their behavior remained unknown for a long time. Even
though the single members of these colonies are non-sophisticated individuals, they
are able to achieve complex tasks in cooperation. Coordinated colony behavior
emerges from relatively simple actions or interactions between the colonies’
individual members. Many aspects of the collaborative activities of social insects
are self-organized and work without a central control. For example, leafcutter ants
cut pieces from leaves, bring them back to their nest, and grow fungi used as food
for their larvae. Weaver ant workers build chains with their bodies in order to cross
gaps between two leaves. The edges of the two leaves are then pulled together, and
successively connected by silk that is emitted by a mature larva held by a worker.
Other examples include the capabilities of termites and wasps to build
sophisticated nests, or the ability of bees and ants to orient themselves in their
environment [Abraham et al., 2006]. The research scientists extract the term swarm
intelligence according to these collaborative activities of social insects. The term
swarm intelligence was first used by Beni in the context of cellular robotic systems

46
 Chapter 3: Data Mining Techniques for IFDPS

where simple agents organize themselves through nearest-neighbor interaction

[Beni, 1988]. Meanwhile, the term swarm intelligence is used for a much broader
research field [Bonabeau et al., 1999]. Swarm intelligence methods have been very
successful in the area of optimization especially to find an optimal solution for NP
problem, which is of great importance for industry and science.
The main algorithms of Swarm Intelligence currently are Ant Colony Optimization
(ACO), Particle Swarm Optimization (PSO) and Bee Colony Algorithm (BCA).
ACO deals with artificial system that is inspired from the foraging behavior of real
ants, which are used to solve discrete optimization problems [Dorigo et al., 1996].
The main idea is the indirect communication between the ants by means of
chemical pheromone trials, which enables them to find short paths between their
nest and food. PSO incorporates swarming behaviors observed in flocks of birds,
schools of fish, or swarms of bees, and even human social behavior, from which
the idea is emerged [Clerc & Kennedy, 2002; Kennedy & Eberhart, 2001;
Parsopoulos & Vrahatis, 2004]. PSO is a population-based optimization tool,
which could be implemented and applied easily to solve various functional
optimization problems, or the problems that can be transformed to functional
optimization problems. As an algorithm, the main strength of PSO is its fast
convergence, which compares favorably with many global optimization algorithms
like Genetic Algorithms (GA) [Goldberg, 1989], Simulated Annealing (SA) [Orosz
& Jacobson, 2002; Triki et al., 2005] and other global optimization algorithms. For
applying PSO successfully, one of the key issues is finding how to map the
problem solution into the PSO particle, which directly affects its feasibility and
performance. There are several advantages of Swarm Intelligence:
Flexibility - the swarm can quickly respond to internal perturbations and
external challenges.
Adaptability - The swarm can adapt to a changing environment.
Robustness - even if one or more individuals in the swarm fail, the swarm can
still complete its tasks.
Self-organization - Paths to solutions are emergent rather than predefined.
Decentralization - The swarm needs relatively little supervision or top-down
control. In other words, there is no central control in the swarm.
Scalability - The control mechanisms used are not dependent on the number of
agents in the swarm.
This section will mainly introduce the algorithms of ACO, PSO and BCA.

3.5.1 Ant Colony Optimization (ACO)

Ant colonies can accomplish complex tasks that far exceed the individual
capabilities of a single ant [Dorigo & Stützle, 2004]. The ACO model is applied
firstly to solve the Travelling Salesman Problem (TSP). The two main phases of
the algorithm constitute the solution construction and the pheromone update. For
TSP, m ants concurrently build a tour and select cities randomly at the beginning

47
 Chapter 3: Data Mining Techniques for IFDPS

of the tour construction. At each construction step, ant k decides which city to visit
next according to a random proportional rule. The probability with which ant k ,
currently at city i , chooses to go to city j is:

[W ij ]D [Kij ]E
pijk , if j  Nik (3.21)
¦
'

lNik
[W ij ]D [Kij ]E

where W ij is the pheromone deposited on arc (i, j ) , Kij 1/ dij , which represents the
visibility of city j towards city i which is inversely proportional to the distance
dij , D and E are two parameters which determine the relative influence of the
pheromone trail and the heuristic information, and N ik is the set of cities that ant k
has not visited yet [Dorigo & Stützle, 2004].
The pheromone trails are updated after tours constructing by evaporating at a
constant rate and accumulating with new deposits:
m
W ij m (1  U )W ij  ¦ 'W ijk ,  (i , k )  L (3.22)
k 1

where 0  U d 1 is the pheromone evaporation rate and 'W ijk is the amount of
pheromone that ant k deposits on the arcs it has visited, defined as follows:
­1 / C k if arc(i, j) belongs to T k
'W ijk ®
¯0 otherwise

where C k is the length of the tour T k built by ant k . By using this rule, the
probability increases that forthcoming ants will use this arc. A brief pseudo-code
and the implementation steps of ACO can be written as following pseudo-code and
Fig. 3.8.

Begin
Initialization
While stopping criterion not satisfied do
Deploy each ant in a starting city
For each ant
Repeat
Calculate probability of remaining cities selected to be next city
Choose next city according to probability using roulette wheel selection
algorithm
Until all cities are visited
Update pheromone
End for
Update the best route (beat solution)
End while
Record and output the beat route (solution)
End

48
 Chapter 3: Data Mining Techniques for IFDPS

Fig. 3.8 The implementation steps of ACO

3.5.2 Particle Swarm Optimization

3.5.2.1 Biological Metaphor

We can imagine such a scenario that a group of birds were searching of food
randomly. There is only one food in this region. All the birds don’t know where the
food is but know how far they away from the food. Then, what is the best strategy
to find the food?
The easiest way is to search the region around of nearest bird from the food. And
according to their own experience of flying to judge the position of food.

49
 Chapter 3: Data Mining Techniques for IFDPS

3.5.2.2 Basis Algorithm of PSO

The Particle Swarm Optimization (PSO) algorithm is a heuristic approach
motivated by the observation of social behavior of composed organisms such as
birds flocking (Fig. 3.9). A number of simple entities – the particles – are placed in
the search space of some problem or function, and each evaluates the objective
function at its current location. Each individual in the particle swarm is composed
of D dimensional vectors, where D is the dimensionality of the search space.

o o
( pi (t )  xi (t ))

o o
( pg  xi (t ))
o
xi (t ) o
vi (t  1) o

o
xi (t  1)
pi (t ) o
vi (t )
o
pg

Fig. 3.9 Birds Flocking of PSO

o
The current position xi can be considered as a set of coordinates describing a point
in space. If the current position is better than any that has been found so far, then
o
the coordinates are stored in the vector pi . The value of the best function result so
o
far is stored in a variable that can be called pg . The objective, of course, is to keep
o o
finding better positions and updating pi and pg . New points are chosen by adding
o o o
vi coordinates to xi , and the algorithm operates by adjusting vi , which can
effectively be seen as a step size. The steps of implementing PSO were shown as
follows:
Step 1: Initialize a population array of particles with random positions and
velocities on D dimensions in the search space.
Step 2: Loop
Step 3: For each particle, evaluate the desired optimization fitness function in
D variables.
o
Step 4: Compare particle’s fitness evaluation with that of its pi . If current
o o
value is better than that of pi , then set pi equal to the current
coordinates.

50
 Chapter 3: Data Mining Techniques for IFDPS

Step 5: Identify the particle in the neighborhood with the best success so far,
o
and assign it to the variable pg .
Step 6: Change the velocity and position of the particle according to the
following equations:
o o o o o o
vi (t  1) Z ˜ vi (t )  c1 ˜ r1 ( pi  vi (t ))  c2 ˜ r2 ( pg  vi (t )) (3.23)
o o o
xi (t  1) xi (t )  vi (t  1) (3.24)
where Z is the inertia weighting; c1 and c2 are acceleration coefficients,
positive constraint; r1 and r2 are the random numbers deferring
uniform distribution on [0, 1]; i represents i th iteration.
Step 7: If a criterion is met (usually a sufficiently good fitness or a maximum
number of iterations), exit loop.
The flowchart of PSO can be seen as Fig. 3.10. In PSO, every particle remembers
its own previous best value as well as the neighborhood best. therefore it has a
more effective memory capability than the GA. PSO is also more efficient in
maintaining the diversity of the swarm, since all the particles use some information
related to the most successful particle in order to improve themselves, whereas in
GA, the worse solutions at every generation are discarded and only the good ones
are saved for next generation. Therefore in GA the population does the evolution
around a set of best individuals in every generation. In addition, PSO is easier to
implement and there are fewer parameters to adjust compared with GA [Valle et al.,
2008].

3.5.2.3 The Parameters of PSO

The role of inertia weight Z in Eq. (3.23), is considered critical for the
convergence behavior of PSO. The inertia weight is employed to control the impact
of the previous history of velocities on the current one. Accordingly, the parameter
Z regulates the trade-off between the global (wide-ranging) and local (nearby)
exploration abilities of the swarm. A large inertia weight facilitates global
exploration, i.e. searching new areas, while a small one tends to facilitate local
exploration, i.e. fine-tuning the current search area. A suitable value for the inertia
weight Z usually provides balance between global and local exploration abilities
and consequently results in a reduction of the number of iterations required to
locate the optimum solution. Initially, the inertia weight is set as a constant.
However, some experiment results indicates that it is better to initially set the
inertia to a large value, in order to promote global exploration of the search space,
and gradually decrease it to get more refined solutions [Eberhart & Shi, 2000].
Thus, an initial value is set to maximum one Zmax (for example around 1.2) and
gradually reducing towards the minimum one Zmin (for example around 0.6) can be
considered as a good choice. A better method is to use some adaptive approaches
(example: fuzzy controller), in which the parameters can be adaptively fine-tuned

51
 Chapter 3: Data Mining Techniques for IFDPS

according to the problems under consideration [Liu & Abraham, 2005; Shi &
Eberhart, 2001].

Fig. 3.10 The flowchart of PSO algorithm

The parameters c1 and c2 , in Eq. (3.23), are not critical for the convergence of PSO.
However, proper fine-tuning may result in faster convergence and alleviation of
local minima. As default values, usually, c1 c2 2 are used, but some experiment
results indicate that c1 c2 1.49 might provide even better results. From Eq.(3.23),
it is better for local exploitation when c1 ! c2 , while it is better for global
exploration when c1  c2 . Recent work reports that it might be even better to choose
a larger cognitive parameter, c1 , than a social parameter, c2 , but with c1  c2 d 4
[Clerc & Kennedy, 2002]. Therefore, the parameter c1 can be changed from c1min to
c1max and the parameter c2 can be changed from c2 max to c2 min regularly in order to
make the algorithm promote global exploration in the beginning and get more
refined solutions (local exploitation) in the end.

3.5.2.4 Variants of PSO

There are many different variants of the PSO algorithm. Some of these variants
have been proposed to incorporate either the capabilities of other evolutionary
computation techniques, such as hybrid versions of PSO or the adaptation of PSO

52
 Chapter 3: Data Mining Techniques for IFDPS

parameters for a better performance (adaptive PSO). In other cases, the nature of
the problem to be solved requires the PSO to work under complex environments as
in the case of the multi-objective or constrained optimization problems or tracking
dynamic systems. There are also some discrete variants of PSO and other
variations to the original formulation that can be included to improve its
performance. This section will present some of them.
A. Binary PSO
Kennedy and Eberhart proposed a discrete binary version of PSO for binary
problems [Kennedy & Eberhart, 1997]. In their model a particle will decide on
"yes" or “no", "true" or "false", "include" or "not to include" etc. also this binary
values can be a representation of a real value in binary search space.
In the binary PSO, the particle’s personal best and global best is updated as in
continuous version. The major difference between binary PSO with continuous
version is that velocities of the particles are rather defined in terms of probabilities
that a bit will change to one. Using this definition a velocity must be restricted
within the range [0, 1]. So a map is introduced to map all real valued numbers of
velocity to the range [0, 1] [Kennedy & Eberhart, 1997]. The normalization
function used here is a sigmoid function as:
1
Sig (vij (t ))  vij ( t )
(3.25)
1 e
JG
where vij (t ) means the j th component of vector vi (t ) . The Eq. (3.23) is also used
to update the velocity vector of the particle. And the new position of the particle is
obtained using the following equation:
­ 1 rij  sig (vij (t  1))
xij (t  1) ® (3.26)
¯0 otherwise

where: rij is a uniform random number in the range [0, 1].

B. Hybrid PSO (FPSO)

A natural evolution of the particle swarm algorithm can be achieved by
incorporating methods that have already been tested in other evolutionary
computation techniques. Many authors have considered incorporating selection,
mutation and crossover, as well as the differential evolution (DE), into the PSO
algorithm. The main goal is to increase the diversity of the population by: 1) either
preventing the particles to move too close to each other and collide [Blackwell &
Bentley, 2002; Krink et al., 2002] or 2) to self-adapt parameters such as the
constriction factor, acceleration constants [Miranda & Fonseca, 2002], or inertia
weight [Lovbjerg & Krink, 2002]. As a result, hybrid versions of PSO have been
created and tested in different applications. The most common ones include hybrid
of genetic algorithm and PSO (GA-PSO), evolutionary PSO (EPSO) and
differential evolution PSO (DEPSO and C-PSO). All these variants of PSO can be
seen in paper [Valle et al., 2008] which described very detail.

53
 Chapter 3: Data Mining Techniques for IFDPS

There are also some other variants of PSO such as Fuzzy PSO [Shi & Eberhart,
2001; D. Tian & Li, 2009], Adaptive PSO [Valle et al., 2008], Gaussian PSO
[Krohling, 2004, 2005; Secrest & Lamont, 2002], Dissipative PSO (DPSO) [Biskas
et al., 2006; Xie et al., 2002], PSO With Passive Congregation (PSOPC) [He et al.,
2004], Stretching PSO (SPSO) [Kannan et al., 2004; Parsopoulos & Vrahatis,
2002], Cooperative PSO (CPSO) [Baskar & Suganthan, 2004; Bergh &
Engelbrecht, 2004; El-Abd & Kamel, 2006], and Comprehensive Learning PSO
(CLPSO) [Liang et al., 2006]. Each variant of PSO mentioned above have
improved its performance in one or more aspects. We can choose a suitable one
when we need apply PSO or its variants to find optimal solution.

3.5.3 Bee Colony Algorithm

Bee Colony Algorithm (BCA) is based on the waggle dance which was discovered
by the Austrian ethnologist and Nobel laureate Karl von Frisch in 1967 [Frisch,
1967]. In the recent years some people has used this knowledge to develop
algorithms to solve real problems. There are several different algorithms inspired
from the bee colony behaviors such as Artificial Bee Colony (ABC) [Karaboga &
Basturk, 2007; Karaboga, 2005], Bees Algorithm (BA) [Pham et al., 2006], Honey
Bee Colony Algorithm (HBCA) [Chong et al., 2006], and Bee Colony
Optimization (BCO) [Teodorovic & Dell’Orco, 2005]. This section will introduce
an algorithm called Bee Colony Algorithm (BCA) [Karaboga & Akay, 2009;
Karaboga & Basturk, 2007].

3.5.3.1 Biological Metaphor

BCA is inspired from the behavior of bee colony during their forage and thus this
biological behavior is introduced firstly before the BCA algorithm. A first
explanation of the behaviour of the bees, was given by the biologist Karl von
Frisch [Frisch, 1967]. The bees use a dance language inside the hive to
communicate the location of the food sources. Sources which are located closer to
the hive are solicited by a round dance. With this dance, the bees describe several
circles with a changing orientation. Food sources with a further distance to the hive
are communicated through a waggle dance which is the most important aspect for
our purpose.
A waggle dance consists (Fig. 3.11) of one to 100 or more circuits, each of which
has two phases: the waggle phase and the return phase. For this dance, the dancing
bee starts to bounce with its abdomen. A worker bee's waggle dance involves
running through a small figure-eight pattern: a waggle run (waggle phase) followed
by a turn to the right to circle back to the starting point (return phase), another
waggle run, followed by a turn and circle to the left, and so on in a regular
alternation between right and left turns after waggle runs.
The meaning of the direction and duration of waggle runs is the direction and
distance of the patch of flowers being advertised by the dancing bee. Flowers
located directly in line with the sun are represented by waggle runs in an upward
direction on the vertical combs, and any angle to the right or left of the sun is coded
by a corresponding angle to the right or left of the upward direction. The distance

54
 Chapter 3: Data Mining Techniques for IFDPS

between hive and recruitment target is indicated by the duration of the waggle runs.
The farther the target, the longer the waggle phase, with a rate of increase of about
75 milliseconds per 100 meters.

Fig. 3.11 The waggle dance

After unloading the collected food, a foraging bee returning to the beehive from a
food source (employed bee) decides whether to abandon the food source or not. If
the food source is abandoned, the bee observes the dances of other employed bees
and follows one of the possible ways adverted for other bees as a follower bee or
starts to search for an entirely new source as a scout bee. However, if the food
source is not abandoned, the employed bee decides whether to dance for the source
to recruit other bees or not and just keep on going to the same food source without
adverting it. Fig. 3.12 shows the decision model of bees’ behaviour.

EBB FS(B)
FS(New)
OB
SB
OB EBB

SB
Unload
OB nectar
Unload
Da
nectar nce
for a rea EB B
B
Dance area

SB
for New

a
r A re

EBA
fo e a
c
an
D

Unload
OB
nectar
HIVE

x Food Source (FS) OB EBA

x Employed Bee (EB)
x Onlooker Bee (OB) OB
EBA
x Scout Bee (SB) FS(A)

Fig. 3.12 The Behavior of the Bees

55
 Chapter 3: Data Mining Techniques for IFDPS

Some researchers have developed a model for foraging behaviour of a honeybee

colony based on reaction–diffusion equations as Karaboga did. His model, that
leads to the idea of collective intelligence of bee swarms consists of three essential
components: food sources, employed foragers, and unemployed foragers, and
defines two ways for the bee colony behaviour: recruitment to a food source and
abandonment of a source [Karaboga & Akay, 2009]. The explanation of the main
components of the model is:
x Food Sources: In order to select a food source, a forager bee evaluates
several properties related with the food source such as its closeness to the
hive, richness of the energy, taste of its nectar, and the easiness or
difficulty of extracting this energy. To simplify, the quality of a food
source can be represented by only one quantity although it depends on
various parameters mentioned above.
x Employed Bees: An employed bee is employed at a specific food source
which is currently exploiting, carrying information about this specific
source and sharing it with other bees waiting in the hive. The information
includes distance, direction and profitability of the food source.
x Unemployed Bees: A forager bee that looks for a food source to exploit is
called unemployed. It can be either a scout who searches the environment
randomly or an onlooker who tries to find a food source by means of the
information given by the employed bee. The mean number of scouts is
about 5–10%.

3.5.3.2 Algorithm of BCA

In BCA algorithm [Karaboga & Akay, 2009; Karaboga & Basturk, 2007], the
position of a food source represents a possible solution to the optimization problem
and the nectar amount of a food source corresponds to the quality of the associated
solution. The number of the employed bees and the onlooker bees is equal to the
number of solutions in the population. The process of the behavior of bees to
search food can be described as following Fig. 3.12: the first phase is called
employed bee phase. In this phase, every food source (FS) is visited by one
employed bee (EB) who then take nectar to hive, and do the waggle dance in the
dance area to express the quality of nectar. The second phase is onlooker bee (OB)
phase. The onlooker bees will chose the food source to visit according to the
waggle dance by the employed bees. The finally phase is scout bee (SB) phase. If
the reaming food source is not good, the scout bees will be set out to find new food
sources, take the nectar back to hive and dance in the dance area. The new food
sources with the old ones will be combined together to be visited by onlooker bees
and employed bees according to their qualities of nectar.
The algorithm of BCA can be described as following steps:
G
1) Initialize the positions of solutions xi , the colony size ( NP ), the maximum
cycle number ( maxCycle ), the number of parameters ( D ), and the number
of trials to improve a source ( limit ).
2) Evaluate the population using fitness function.

56
 Chapter 3: Data Mining Techniques for IFDPS

3) Repeat ( Cycle 1 )
G
4) Produce new solutions vi (food source positions) in the neighbourhood of
G
xi for the employed bees using Eq. (3.27) and evaluate these solutions
using fitness function.
vij xij  Iij ( xij  xkj ) (3.27)

where
Iij : Random number between [-1, 1].
i : {1, 2," , C} the i th food source.
j : the j th component of parameters.
k : {1, 2," , SN } randomly chosen index of parameter (dimension) which is
different from j .
G G
5) Apply the greedy selection process for employed bees between vi and xi .
G
6) Calculate the probability value pi for the solutions xi by means of their
fitness values using Eq. (3.28).
fiti / ¦ n
SN
pi 1
fit n (3.28)
G
7) Produce the new solutions vi (new positions) for the onlookers from the
G
solutions xi using Eq. (3.27), which selected depending on pi , and
evaluate them.
G
8) Apply the greedy selection process for the onlooker bees between vi and
G
xi .
9) Determine the abandoned solution (source), if exists, replace it with a new
G
randomly produced solution xi for the scout using the following equation.
xij xijmin  rand[0,1]( xijmax  xijmin ) (3.29)
10) Memorize the best food source position (solution) achieved so far.
11) Cycle Cycle  1
12) Exit if Cycle maxCycle or other criterion is met.
In the process of BCA algorithm, step 5) and step 6) constitute the employed bee
phase, step 7) and step 8) constitute the onlooker bee phase while step 9) is scout
bee phase. The problem of dynamic CBM scheduling is a kind of NP problem and
the BCA is a good method to find the optimal solution for this kind problem.

3.6 Summary

This Chapter introduced the basic concepts of Data Mining techniques and
algorithms. ANN includes supervised learning and unsupervised learning is mainly
applied in the case of the accurate physical model or mathematical model is
unavailable, but the huge history data are available. When there are huge history

57
 Chapter 3: Data Mining Techniques for IFDPS

data but only a small part of them are labeled, the Semi-supervised learning can be
very good method to build the model. Association rules are mainly used to find the
relations between the features. Swarm Intelligence, such as particle Swarm
Optimization and Bee Colony Algorithm, is mainly used to solve the optimization
problems, find the optimal solution for NP problems.
There are too many methods and algorithm of Data Mining techniques. This
Chapter only introduced some of them which will be applied in IFDPS system.

58
 Chapter 4: Sensor Classification and Sensor Placement Optimization

4 Sensor Classification and Sensor Placement Optimization

4.1 Introduction

A sensor is a converter that measures physical quantity and converts it into signal
which can be read by an observer or by instruments. It is device that detects
changes in the ambient conditions or in the state of another device or a system, and
conveys or records this information in a certain manner. Sensors and sensing
strategies constitute the foundational basis for fault diagnosis and prognosis
systems. Most of sensors are well-developed in market and the customers just need
to choose suitable sensors to collect data which can be used to monitor the
condition of components or machines. When choosing sensors for diagnostics and
prognostics, many parameters and features of the sensors must to be considered
which are the type, number, and location of sensors; their size, weight, cost,
dynamic range, and other characteristic properties; whether they are of the wired or
wireless variety; etc. The raw data collected from the sensors are rarely useful
because they may contain much noise or no explicit features. These data must be
processed appropriately so that useful information may be extracted that is a
reduced version of the original data but preserves as much as possible those
characteristic features or fault indicators that are indicative of the fault events we
are seeking to detect, isolate, and predict the time evolution of. Thus such data
must be preprocessed, that is, filtered, compressed, correlated, etc., in order to
remove artifacts and reduce noise levels and the volume of data to be processed
subsequently. Furthermore, the sensor providing the data must be validated; that is,
the sensors themselves are not subjected to fault conditions. Once the
preprocessing module confirms that the sensor data are ‘‘clean’’ and formatted
appropriately, features or signatures of normal or faulty conditions must be
extracted. This is most important in the framework of IFDPS because it is the input
of the processes of diagnostics and prognostics [Wachtsevanos et al. 2006]
Sensor suites are specific to the application domain, and they are intended to
monitor such typical state awareness variables as temperature, pressure, speed,
vibrations, etc. Some sensors are inserted specifically to measure quantities that are
directly related to fault modes identified as candidates for diagnosis. Among them
are strain gauges, ultrasonic sensors, proximity devices, acoustic emission sensors,
electrochemical fatigue sensors, interferometers, etc., whereas others are of the
multipurpose variety, such as temperature, speed, flow rate, etc., and are designed
to monitor process variables for control and/or performance assessment in addition
to diagnosis. More recently we have witnessed the introduction of wireless devices
in the area of condition monitoring.
For a normal word ‘sensor’ device, it actual have two components: sensor and
transducer. A sensor is defined as a device that is sensitive to light, temperature,
electrical impedance, or radiation level and transmits a signal to a measuring or
control device. On the other hand, a transducer is defined as a device that receives
energy from one system and retransmits it, often in a different form, to another
system. A measuring device passes through two stages while measuring a signal.

59
 Chapter 4: Sensor Classification and Sensor Placement Optimization

First, the measurand (a physical quantity such as acceleration, pressure, strain,

temperature etc.) is sensed by the sensor. Then, the measured signal is transduced
into a form that is particularly suitable for transmitting, signal conditioning, and
processing. For this reason, output of the transducer stage is often an electrical
signal that is then digitized. The sensor and transducer stages of a typical
measuring device are represented schematically in Fig. 4.1.

Fig. 4.1 Schematic Representation of a Measuring Device

The sensor strategies are mainly focused two issues: the one is which kind of
sensors is suitable to measure the signals, and the other is which place the sensors
should be set up. This Chapter will introduce these two issues.

4.2 Classification of Sensors

There are many kinds of sensors in the business market. White [White, 1987]
presented out a sensor classification scheme for categorizing sensors which are
recalled in the following tables. Table 4.1 shows most measurands for which
sensors may be needed under the headings: acoustic, biological, chemical, electric,
magnetic, mechanical, optical, radiation (particle), and thermal, etc. With a
particular measurand, one is primarily interested in sensor characteristics such as
sensitivity, selectivity, and speed of response which is shown in Table 4.2 called
technological aspects of sensors. Table 4.3 shows the detection means used in
sensors. Table 4.4 is intended to indicate the primary phenomena used to convert
the measurand into a form suitable for producing the sensor output. The application
fields are listed in Table 4.5. Most sensors contain a variety of materials (for
example, almost all contain some metal). The entries in Table 4.6 should be
understood to refer to the materials chiefly responsible for sensor operation.

Table 4.1 Measurands of Sensors

A. Measurands
Al. Acoustic
Al.1 Wave amplitude, phase, polarization, spectrum
A1.2 Wave velocity
A1.3 Other (specify)
A2. Biological
A2.1 Biomass (identities, concentrations, states)
A2.2 Other (specify)
A3. Chemical
A3.1 Components (identities, concentrations, states)
A3.2 Other (specify)
A4. Electric
A4.1 Charge, current

60
Chapter 4: Sensor Classification and Sensor Placement Optimization

A4.2 Potential, potential difference

A4.3 Electric field (amplitude, phase, polarization, spectrum)
A4.4 Conductivity
A4.5 Permittivity
A4.6 Other (specify)
A5. Magnetic
A5.1 Magnetic field (amplitude, phase, polarization, spectrum)
A5.2 Magnetic flux
AS.3 Permeability
AS.4 Other (specify)
A6. Mechanical
A6.1 Position (linear, angular)
A6.2 Velocity
A6.3 Acceleration
A6.4 Force
A6.5 Stress, pressure
A6.6 Strain
A6.7 Mass, density
A6.8 Moment, torque
A6.9 Speed of flow, rate of mass transport
A6.10 Shape, roughness, orientation
A6.11 Stiffness, compliance
A6.12 Viscosity
A6.13 Crystallinity, structural integrity
A6.14 Other (specify)
A7. Optical
A7.1 Wave amplitude, phase, polarization, spectrum
A7.2 Wave velocity
A7.3 Other (specify)
A8. Radiation
A8.1 Type
A8.2 Energy
A8.3 Intensity
A8.4 Other (specify)
A9. Thermal
A9.1 Temperature
A9.2 Flux
A9.3 Specific heat
A9.4 Thermal conductivity
A9.5 Other (specify)
A10. Other (specify)

61
Chapter 4: Sensor Classification and Sensor Placement Optimization

Table 4.2 Technological Aspects of Sensors

B. Technological Aspects of Sensors
B1 Sensitivity
B2 Measurand range
B3 Stability (short-term, long-term)
B4 Resolution
B5 Selectivity
B6 Speed of response
B7 Ambient conditions allowed
B8 Overload characteristics
B9 Operating life
B10 Output format
B11 Cost, size, weight

Table 4.3 Detection Means Used in Sensors

C. Detection Means Used in Sensors
C1 Biological
C2 Chemical
C3 Electric, Magnetic, or Electromagnetic Wave
C4 Heat, Temperature
C5 Mechanical Displacement or Wave
C6 Radioactivity, Radiation
C7 Other (specify)

Table 4.4 Sensor Conversion Phenomena

D. Sensor Conversion Phenomena
Dl. Biological
D1.1 Biochemical transformation
D1.2 Physical transformation
D1.3 Effect on test organism
Dl .4 Spectroscopy
D1.5 Other (specify)
D2. Chemical
D2.1 Chemical transformation
D2.2 Physical transformation
D2.3 Electrochemical process
D2.4 Spectroscopy
D2.5 Other (specify)
D3. Physical
D3.1 Thermoelectric
D3.2 Photoelectric
D3.3 Photomagnetic
D3.4 Magnetoelectric
D3.S Elastomagnetic
D3.6 Thermoelastic
D3.7 Elastoelectric
D3.8 Thermomagnetic
D3.9 Thermooptic
D3.10 Photoelastic
D3.11 Other (specify)

62
 Chapter 4: Sensor Classification and Sensor Placement Optimization

Table 4.5 Fields of Application

F. Fields of Application
F1 Agriculture
F2 Automotive
F3 Civil engineering, construction
F4 Distribution, commerce, finance
F5 Domestic appliances
F6 Energy, power
F7 Environment, meteorology, security
F8 Health, medicine
F9 Information, telecommunications
F10 Manufacturing
F11 Marine
F12 Military
F13 Scientific measurement
F14 Space
F15 Transportation (excluding automotive)
F16 Other (specify)

Table 4.6 Sensor Materials

E. Sensor Materials
El Inorganic
E2 Organic
E3 Conductor
E4 Insulator
E5 Semiconductor
E6 Liquid, gas or plasma
E7 Biological substance
E8 Other (specify)

The scheme shown in above tables can facilitate comparing sensors,

communicating with other workers about sensors, and keeping track of sensor
progress and availability. Categorizing might help one think about new sensing
principles that could be explored, and Table 4.2 might serve as a checklist to
consult when considering commercial sensors. Refer to the sensor application in
Condition Monitoring, the first step is to determine which measurands need to be
measured which is shown in Table 4.1, and then analyze the requirements of the
system to decide the technological aspects of the selected sensors which are shown
in Table 4.2. In this Chapter, only the sensors can be used to collect data for fault
diagnosis and prognosis are considered. Fig. 4.2 shows the most kinds of sensors
applied in condition monitoring for fault diagnosis and prognosis. Some of these
kinds of sensors are described following.
Mechanical sensor systems have been studied extensively, and a large number of
such devices are currently in use to monitor system performance for operational
state assessment and tracking of fault indicators. A number of mechanical
quantities—position, speed, acceleration, torque, strain, etc.—are commonly
employed in dynamic systems. The most widely used sensors in condition
monitoring for manufacturing machines are vibration sensors and strain gauges.

63
 Chapter 4: Sensor Classification and Sensor Placement Optimization

Accelerometers (Vibration Measurements)

Strain gauges
Mechanical Sensor Ultrasonic Sensor System
Systems (A6) Position, speed, acceleration, torque, strain
ĂĂ
Temperature Sensors / Thermography
Performance Sensors Pressure, Fluid and thermodynamic
(A1, A3, A7 and A9) Optical properties and biochemical elements
ĂĂĂĂ
Eddy-Current Proximity Probes
Electrical Measurement
(A4)
Microelectromechanical System (MEMS) Sensors
Fiber Optic Sensors

Fig. 4.2 The Classification of Sensors.

Recent years have seen an increased requirement for a greater understanding of the
causes of vibration and the dynamic response of failing structures and machines to
vibratory forces. An accurate, reliable, and robust vibration transducer therefore is
required to monitor online such critical components and structures. Piezoelectric
accelerometers offer a wide dynamic range and rank among the optimal choices for
vibration-monitoring apparatus. They exhibit such desirable properties as
[Wachtsevanos et al., 2006]:
x Usability over very wide frequency ranges;
x Excellent linearity over a very wide dynamic range;
x Electronically integrated acceleration signals to provide velocity and dis-
placement data.
x Vibration measurements in a wide range of environmental conditions while
still maintaining excellent accuracy
x Self-generating power supply
x No moving parts and hence extreme durability
x Extremely compact plus a high sensitivity-to-mass ratio
Piezoelectric accelerometers are used to measure all types of vibrations regardless
of their nature or source in the time or frequency domain as long as the
accelerometer has the correct frequency and dynamic ranges.
A strain-gauge sensor is based on a simple principle from basic electronics that the
resistance of a conductor is directly proportional to its length and resistivity and
inversely proportional to its cross-sectional area. Applied stress or strain causes the
metal transduction element to vary in length and cross-sectional area, thus causing
a change in resistance that can be measured as an electrical signal. Certain
substances, such as semiconductors, exhibit the piezoresistive effect, in which
application of strain greatly affects their resistivity. Strain gauges of this type have
a sensitivity approximately two orders greater than the former type. The transducer
usually is used within a Wheatstone bridge arrangement, with one, two, or all four
of the bridge arms being individual strain gauges, so that the output voltage change

64
 Chapter 4: Sensor Classification and Sensor Placement Optimization

is an indication of measurand (the strain) change. The output of the strain-gauge is

a simple voltage signal that can be connected to an oscilloscope to view the strain
output or to the data-acquisition system to take strain-gauge data.
Ultrasonic sensor systems are being considered for monitoring the health of critical
structures such as airplanes, bridges, and buildings. Ultrasonic methods are
particularly suitable for structural health monitoring both because ultrasonic waves
travel long distances and thus have the potential to monitor a large volume of
material and because ultrasonic methods have proven useful for nondestructive
inspection of such structures during maintenance. There are three main types of
ultrasonic waves that are suitable for structural health monitoring: guided waves,
bulk waves, and diffuse waves. Regardless of the type of wave, the strategy is to
monitor changes and then detect, localize, and characterize damage based on the
nature of the change. This strategy of looking for changes can enable detection
sensitivity to be similar to that of nondestructive inspection despite the limitation of
fixed sensors. The theory behind ultrasonic ranging is quite simple (as shown in
Fig. 4.3). Typically a short ultrasonic burst is transmitted from the transmitter.
When there is an object in the path of the ultrasonic pulse, some portion of the
transmitted ultrasonic wave is reflected and the ultrasonic receiver can detect such
echo. By measuring the elapsed time between the sending and the receiving of the
signal along with the knowledge of the speed of sound in the medium, the distance
between the receiver and the object can be calculated.

Fig. 4.3 Principle of Ultrasonic Sensors

System performance and operational data are monitored routinely in all industrial
establishments, utility operations, transportation systems, etc. for process control,
performance evaluation, quality assurance, and fault diagnosis purposes. A large
number of sensor systems have been developed and employed over the years. The
list includes devices that are intended to measure such critical properties as
temperature; pressure; fluid, thermodynamic, and optical properties; and
biochemical elements, among many others. Sensors based on classic measuring
elements—inductive, capacitive, ultrasound— have found extensive applications.
Temperature variations in many mechanical, electrical, and electronic systems are
excellent indicators of impending failure conditions. Temperatures in excess of
control limits should be monitored and used in conjunction with other

65
 Chapter 4: Sensor Classification and Sensor Placement Optimization

measurements to detect and isolate faults. Temperature sensing has found

numerous applications over the years in such areas as engineering, medicine,
environmental monitoring, etc. Therefor the temperature sensors play a very
important role in condition monitoring. A temperature sensor is a device that
gathers data concerning the temperature from a source and converts it to a form
that can be understood either by an observer or another device. Temperature
sensors come in many different forms and are used for a wide variety of purposes,
from simple home use to extremely accurate and precise scientific use. They play a
very important role almost everywhere that they are applied. The best known
example of a temperature sensor is the mercury-in-glass thermometer. Mercury
expands and contracts based on changes in temperature; when these volume
changes are quantified, temperature can be measured with a fair degree of
accuracy. The outside temperature is the source of the temperature measurements
and the position of the mercury in the glass tube is the observable quantification of
temperature that can be understood by observers. Typically, mercury-in-glass
thermometers are only used for non-scientific purposes because they are not
extremely accurate. In some cases, they can be used in high school or college
chemistry labs when a very accurate measurement of temperature is not important.
The most common temperature sensors in scientific area are resistance temperature
detectors (RTDs), whose principle of operation is variation of the resistance of a
platinum wire or film as a function of temperature. Platinum usually is employed
because of its stability with temperature and the fact that its resistance tends to be
almost linear with temperature. Such temperature devices have higher accuracy
than that of mercury-in-glass thermometer and thus are used widely in condition
monitoring when the temperature of the machines or environment needs to be
monitor accurately.
Electrical measurements are the methods, devices and calculations used to measure
electrical quantities. Measurement of electrical quantities may be done to measure
electrical parameters of a system. Using transducers, physical properties such as
temperature, pressure, flow, force, and many others can be converted into electrical
signals, which can then be conveniently measured and recorded. According to the
principle, a number of sensor systems based on the Electrical measurements have
been developed and applied in the recent past in an attempt to interrogate critical
components and systems for fault diagnosis and prognosis. Transducing principles
based on eddy-current response characteristics, optical and infrared signal
mentoring, microwaves, and others have been investigated.
Response characteristics of induced eddy currents in conducting media are
monitored for changes in their behavior owing to material anomalies, cracks, shaft
or mating-part displacements, etc. Eddy-current proximity probes are a mature
technology that has been used for protection and management of rotating
machinery. They are employed commonly in high-speed turbo machinery to
observe relative shaft motion directly, that is, inside bearing clearances of fluid-
film interfaces. Zou et al. describe the application of eddy-current proximity
sensing to the detection of a crack in a seal/rotor drive-shaft arrangement [Zou et
al., 2000].

66
 Chapter 4: Sensor Classification and Sensor Placement Optimization

Of interest also are sensor systems that can be produced inexpensively, singly or in
an array, while maintaining a high level of operational reliability.
Microelectromechanical systems (MEMS) and sensors based on fiber-optic
technologies are finding popularity because of their size, cost, and ability to
integrate multiple transducers in a single device. Micro-machined MEMS devices
in silicon or other materials are fabricated in a batch process with the potential for
integration with electronics, thus facilitating on-board signal processing and other
‘‘smart’’ functions. A number of MEMS transducer and sensor systems have been
manufactured in the laboratory or are available commercially, monitoring such
critical parameters as temperature, pressure, acceleration, etc. [Wachtsevanos et al.,
2006].
Fiber optics has penetrated the telecommunications and other high-technology
sectors in recent years. They find utility in the sensor field because of their
compact and flexible geometry, potential for fabrication into arrays of devices,
batch fabrication, etc. Fiber optic sensors have been designed to measure strain,
temperature, displacement, chemical concentration, and acceleration, among other
material and environmental properties. Their main advantages include small size,
light weight, immunity to electromagnetic and radio frequency interference
(EMI/RFI), high- and low-temperature endurance, fast response, high sensitivity,
and low cost. Fiber optic technologies are based on extrinsic Fabry-Perot
interferometry (EFPI), chemical change in the fiber cladding, optical signal
changes owing to fiber stress and deformation, etc.
There are also some other kinds of sensors available in the market and most of
them are very good to meet the monitoring requirement. We only need choose
suitable ones to collect data from the machines for monitoring.

4.3 Wireless Sensor Networks

A sensor network is a group of specialized sensors with a communications

infrastructure intended to monitor and record conditions at diverse locations.
Commonly monitored parameters are temperature, humidity, pressure, wind
direction and speed, illumination intensity, vibration intensity, sound intensity,
power-line voltage, chemical concentrations, pollutant levels and vital body
functions.
Sensor networks may consist of many different types of sensors such as seismic,
low sampling rate magnetic, thermal, visual, infrared, acoustic and radar, which are
able to monitor a wide variety of ambient conditions that include the following
[Estrin et al., 1999]:
x Temperature,
x Humidity,
x Vehicular movement,
x Lightning condition,
x Pressure,
x Soil makeup,

67
 Chapter 4: Sensor Classification and Sensor Placement Optimization

x Noise levels,
x The presence or absence of certain kinds of objects,
x Mechanical stress levels on attached objects, and
x The current characteristics such as speed, direction, and size of an object.
A sensor network consists of multiple detection stations called sensor nodes, each
of which is small, lightweight and portable. Every sensor node is equipped with a
transducer, microcomputer, transceiver and power source. The transducer generates
electrical signals based on sensed physical effects and phenomena. The
microcomputer processes and stores the sensor output. The transceiver, which can
be hard-wired or wireless, receives commands from a central computer and
transmits data to that computer. The power for each sensor node is derived from
the electric utility or from a battery.
Sensor networks can be deployed in the following two ways [Intanagonwiwat et
al., 2000]:
x Sensors can be positioned far from the actual phenomenon, i.e., something
known by sense perception. In this approach, large sensors that use some
complex techniques to distinguish the targets from environmental noise are
required.
x Several sensors that perform only sensing can be deployed. The positions
of the sensors and communications topology are carefully engineered (Fig.
1.5). They transmit time series of the sensed phenomenon to the central
nodes where computations are performed and data are fused.
The above described features ensure a wide range of applications for sensor
networks. Some of the application areas are health, military, and security. For
example, the physiological data about a patient can be monitored remotely by a
doctor. While this is more convenient for the patient, it also allows the doctor to
better understand the patient’s current condition. Sensor networks can also be used
to detect foreign chemical agents in the air and the water. They can help to identify
the type, concentration, and location of pollutants. In essence, sensor networks will
provide the end user with intelligence and a better understanding of the
environment [Akyildi et al., 2002]. Sensor networks can also be very helpful in
condition monitoring for manufacturing machines, wind turbines, transporters and
infrastructure because they may be distributed in different place. Potential
applications of sensor networks may include:
x Condition monitoring for factory or infrastructure;
x Industrial automation;
x Automated and smart homes;
x Video surveillance;
x Traffic monitoring;
x Medical device monitoring;
x Monitoring of weather conditions;
x Air traffic control;
x Military applications;
x Robot control.

68
 Chapter 4: Sensor Classification and Sensor Placement Optimization

While many sensors connect to controllers and processing stations directly (e.g.,
using local area networks), an increasing number of sensors communicate the
collected data wirelessly to a centralized processing station which are compose a
Wireless Sensor Network (WSN). This is important since many network
applications require hundreds or thousands of sensor nodes, often deployed in
remote and inaccessible areas. Therefore, a wireless sensor has not only a sensing
component, but also on-board processing, communication, and storage capabilities.
With these enhancements, a sensor node is often not only responsible for data
collection, but also for in-network analysis, correlation, and fusion of its own
sensor data and data from other sensor nodes. When many sensors cooperatively
monitor large physical environments, they form a WSN. Sensor nodes
communicate not only with each other but also with a base station (BS which could
be a gateway) using their wireless radios, allowing them to disseminate their sensor
data to remote processing, visualization, analysis, and storage systems. For
example, Fig. 4.4 shows two sensor fields monitoring two different geographic
regions and connecting to the Internet using their base stations [Dargie and
Poellabauer, 2010].

Fig. 4.4 Wireless Sensor Networks

The capabilities of sensor nodes in a WSN can vary widely, that is, simple sensor
nodes may monitor a single physical phenomenon, while more complex devices
may combine many different sensing techniques (e.g., acoustic, optical, magnetic).
They can also differ in their communication capabilities, for example, using
ultrasound, infrared, or radio frequency technologies with varying data rates and
latencies. While simple sensors may only collect and communicate information
about the observed environment, more powerful devices (i.e., devices with large
processing, energy, and storage capacities) may also perform extensive processing
and aggregation functions. Such devices often assume additional responsibilities in
a WSN, for example, they may form communication backbones that can be used by
other resource-constrained sensor devices to reach the base station. Finally, some
devices may have access to additional supporting technologies, for example,
Global Positioning System (GPS) receivers, allowing them to accurately determine
their position. However, such systems often consume too much energy to be
feasible for low-cost and low-power sensor nodes [Dargie and Poellabauer, 2010].

69
 Chapter 4: Sensor Classification and Sensor Placement Optimization

The well-known IEEE 802.11 family of standards was introduced in 1997 and is
the most common wireless networking technology for mobile systems. It uses
different frequency bands, for example, the 2.4-GHz band is used by IEEE 802.11b
and IEEE 802.11g, while the IEEE 802.11a protocol uses the 5-GHz frequency
band. IEEE 802.11 was frequently used in early wireless sensor networks and can
still be found in current networks when bandwidth demands are high (e.g., for
multimedia sensors). However, the high-energy overheads of IEEE 802.11-based
networks make this standard unsuitable for low-power sensor networks. Typical
data rate requirements in sensor networks are comparable to the bandwidths pro-
vided by dial-up modems, therefore the data rates provided by IEEE 802.11 are
typically much higher than needed. This has led to the development of a variety of
protocols that better satisfy the networks’ need for low power consumption and low
data rates. For example, the IEEE 802.15.4 protocol [Callaway et al., 2002] has
been designed specifically for short- range communications in low-power sensor
networks and is supported by most academic and commercial sensor nodes.
The network topologies can be seen as in Fig. 1.5 and the most widely used ones
are topologies of star and mesh. When the transmission ranges of the radios of all
sensor nodes are large enough and the sensors can transmit their data directly to the
base station, they can form a star topology as shown on the left in Fig. 4.5. In this
topology, each sensor node communicates directly with the base station using a
single hop. However, sensor networks often cover large geographic areas and radio
transmission power should be kept at a minimum in order to conserve energy;
consequently, multi-hop communication is the more common case for sensor net-
works (shown on the right in Fig. 4.5). In this mesh topology, sensor nodes must
not only capture and disseminate their own data, but also serve as relays for other
sensor nodes, that is, they must collaborate to propagate sensor data towards the
base station. This routing problem, that is, the task of finding a multi-hop path from
a sensor node to the base station, is one of the most important challenges and has
received immense attention from the research community. When a node serves as a
relay for multiple routes, it often has the opportunity to analyze and pre-process
sensor data in the network, which can lead to the elimination of redundant
information or aggregation of data that may be smaller than the original data. The
more detailed information about Wireless Sensor Networks can be found at the
reference of [Dargie & Poellabauer, 2010].

Fig. 4.5 Single-hop Versus Multi-hop Communication in Sensor Networks

70
 Chapter 4: Sensor Classification and Sensor Placement Optimization

4.4 RFID Sensor Networks

Radio Frequency IDentification (RFID) is one of numerous technologies grouped

under the term of Automatic Identification (Auto ID), such as bar code, magnetic
inks, optical character recognition, voice recognition, touch memory, smart cards,
biometrics etc. Auto ID technologies are a new way of controlling information and
material flow, especially suitable for large production networks [Ilie-zudor et al.,
2006]. RFID is the use of a wireless non-contact radio system to transfer data from
a tag attached to an object, for the purposes of identification and tracking. In
general terms, it is a means of identifying a person or object using a radio
frequency transmission. The technology can be used to identify, track, sort or
detect a wide variety of objects [Lewis, 2004]. Recently, RFID become more and
more interesting technology in many fields such as agriculture, manufacturing and
supply chain management.
The history of RFID technology can be tracked back to the radio-based
identification system used by allied bombers during World War II [Garfinkel &
Holtzman, 2005]. Early identification Friend or For (IFF) systems were used to
distinguish Allied fighter and bomber by identifying the correct signals sent by
Allied aircrafts, from aircrafts sent by enemy at night. After the war, Harry
Stockman realized that it is possible to power a mobile transmitter completely from
the strength of a received radio signal, and then he introduced the concept of
passive RFID systems [Stockman, 1948]. In 1972, a patent application for
“inductively coupled transmitter-responder arrangement” was filed which is used
separate coils for receiving power and transmitting the return signal [Kriofsky &
Kaplan, 1975]. In 1979, a patent application for “identification device” (two
antennas was combined) was filed which is seen as a RFID landmark because it
emphasized the potentially small size of RFID device [Beigel, 1982]. The 1980s
became the decade for full implementation of RFID technology, though interests
developed somewhat differently in various parts of the world. The greatest interests
in the United States were for transportation, personnel access, and to a lesser
extent, for animals. In Europe, the greatest interests were for short-range systems
for animals, industrial and business applications, though toll roads in Italy, France,
Spain, Portugal, and Norway were equipped with RFID. The 1990s were a
significant decade for RFID since it saw the wide scale deployment of electronic
toll collection in the United States. The world's first open highway electronic
tolling system opened in Oklahoma in 1991 and then extended to the whole world.
Interest was also keen for RFID applications in Europe during the 1990s. Both
Microwave and inductive technologies were finding use for toll collection, access
control and a wide variety of other applications in commerce [Landt, 2001]. The
21st century opens with the smallest microwave tags built using, at a minimum,
two components: a single custom CMOS integrated circuit and an antenna. Tags
could now be built as sticky labels, easily attached to windshields and objects to be
managed [Landt, 2005]. It seems that there are still a great many developments of
RFID to look forward to as the history continues to teach that and RFID will be
presented in our daily life.

71
 Chapter 4: Sensor Classification and Sensor Placement Optimization

4.4.1 RFID System

Typical RFID systems fundamentally consist of four elements: the RFID tags, the
RFID readers, the antennas and choice of radio characteristics, and the computer
network (if any) that is used to connect the readers (Fig. 4.6). Tags are attached to
objects and each of them has a certain amount of internal memory (E2PROM) in
which it stories information about the object, such as its unique ID number, or in
some cases more details including manufacture data and product composition.
When these tags pass through a field generated by a reader, they transmit this
information back to the reader, thereby identifying the object. Until recently, the
focus of RFID technology was mainly on tags and readers which were being used
in systems where relatively low volumes of data are involved. This is now
changing as RFID in the supply chain is expected to generate huge volumes of data,
which will have to be filtered and routed to the backend IT systems. To solve this
problem companies have developed special software packages (Middleware),
which act as buffers between the RFID front end and the IT backend [Wang &
Zhang, 2012].

Fig. 4.6 Typical RFID System

Fig. 4.7 RFID Tags Communication Methods

There are two main communication principles between RFID readers/antennas and
RFID Tags: inductive coupling and backscatter reflection which are used in near
field and far field respectively (Fig. 4.7). The principle of inductive coupling
means transferring energy from one circuit to another through mutual inductance.
Near field employs inductive coupling of the tag to the magnetic field circulating
around the reader antenna (like a transformer). In RFID systems using inductive
coupling, the reader antenna and the RFID tag antenna each have a coil which

72
 Chapter 4: Sensor Classification and Sensor Placement Optimization

together forms a magnetic field so that the tag draws energy from the field to
change the electrical load on the tag antenna. The change is picked up by the reader
and read as a unique serial number. Far field uses similar techniques to radar
(Backscatter reflection) by coupling with the electric field. RFID tags using
backscatter technology reflect radio waves at the same carrier frequency back to
the tag reader, using modulation to transmit the data.
The communication process between the reader and tag is managed and controlled
by one of several protocols, such as the ISO 15693 and ISO 18000-3 for HF or the
ISO 18000-6, and EPC for UHF. Basically what happens is that when the reader is
switched on, it starts emitting a signal at the selected frequency band (typically 860
- 915MHz for UHF or 13.56MHz for HF). Any corresponding tag in the vicinity of
the reader will detect the signal and use the energy from it to wake up and supply
operating power to its internal circuits. Once the Tag has decoded the signal as
valid, it replies to the reader, and indicates its presence by modulating (affecting)
the reader field.
The communication principle can be used to compose parts of wireless sensor
network.

4.4.2 Embedded RFID Sensor Monitoring

RFID sensor enabled tags, which can be used in such fields as project tracking,
environmental monitoring, automotive electronic system, telemedicine and
manufacturing processes controlling, etc., are bred as the result. Without doubts,
they will play important roles in more and more areas as the technology is
progressively growing. Roughly, the primary sensors in use today can be classified
according to their functions in many categories such as: temperature, pressure,
acceleration, inclination, humidity, light, gas sensor and chemical sensors
[Ruhanne et al., 2008] .
Fig. 4.8 shows the system architecture for a generic sensor tag and its interaction
with RFID systems as it passes through various stages of the manufacturing,
assembling and supply chain. The RFID tags can be combined to the sensor
devices (many different sensors) and transfer the sensing data to the RFID reader
and further to the database through radio waves. Typically for the supply, there are
a number of RFID portals and at each of these passive RFID tag is interrogated.
The data obtained could be used for improving the process and scheduling of
supply chain and production process [Wang & Zhang, 2012]. For the
manufacturing systems and processes, there are many sensors mounted on the
machines which can be combined with RFID tags. The collected data can be
transmitted to RFID reader and database, and the data with some processing
techniques can be used to monitor the condition of the machine and improve the
performance.

73
 Chapter 4: Sensor Classification and Sensor Placement Optimization

Fig. 4.8 General Structure of Embedded RFID Sensing System

4.5 General Sensor Networks

This section is a summary of sensor network techniques mentioned above. The

wired network, Wi-Fi wireless networks, Bluetooth and RFID can be integrated
together to collect data according to the requirements of real projects. The general
structure of the integrated sensor networks are shown in Fig. 4.9. In the real
application sites, suitable sensors are selected to collect data of the machines. The
collected data can be transmitted to database through wired network, wireless
network (Wi-Fi), RFID and Bluetooth. The customers may use one or more these
kinds of methods to transfer the data according to the requirements and considering
the cost of human resource, economy and so on.

4.6 Sensor Placement Optimization (SPO)

The basic problem for condition monitoring is to deduce the existence of a defect
in a structure from measurements taken at sensors distributed on the structure. The
correctness of defect diagnosis depends on the method of pattern recognition for
fault and effectiveness of signals from the sensors mounted on the machines. While
carrying out on-site condition monitoring for a machine, the inappropriate
distribution of sensors might result in weak incentives of certain order or modal,
and affect the accuracy of fault identification. The aim of optimizing the placement
of sensors is to obtain as much as possible of machine structural information with
as few as possible sensors, which benefit the company in the economy viewpoint.
Because of constraints of machine structure and environment, and consideration of
economy, only a small number of sensors are installed when a condition
monitoring system is established. It is very important to design the optimal position
of the sensor to mount in order to ensure the accuracy and correctness of
monitoring and fault judgement.

74
 Chapter 4: Sensor Classification and Sensor Placement Optimization

Fig. 4.9 General Sensor Network Structure

There are many literatures in optimal placement optimization of sensors in machine

level. The spatial controllability was used to find the optimal placement of
collocated actuator-sensor pairs for effective average vibration reduction over the
entire structure, and the maintaining modal controllability and observability were
used to select vibration modes for a thin plate [Halim & Reza Moheimani, 2003].
Recently, intelligent optimization algorithm has developed well which is a method
to simulate the biological and physical process which can be used in sensor
placement optimization. Many researchers focus on Genetic Algorithm (GA)
application in sensor placement optimization and make up for a lot of shortage of
the traditional optimization algorithm [Li et al., 2000; Liu et al., 2008; Sun et al.,
2008]. But GA has to adopt binary coding and has complex operation process such
as mutation, genetic and crossover. PSO adopts real number coding to avoid the
complex operation, which is simple and easy to realize. So it is easy to apply in
sensor placement optimization. PSO and finite element analysis were combined
together to search the sensors optimal placement of a gearbox [Pan et al., 2010].
Binary PSO and Analytical redundancy Relations (ARRs) were combined to
optimize the sensor placement for fault diagnosis [Du et al., 2011]. The sensor
placement optimization is a very important aspect for many applications such as
modal test and parameter identification [Cheng 2003; Papadimitriou 2004;
Pennacchi and Vania 2008], fault diagnosis [Bhushan and Rengaswamy, 2000;
Molter et al., 2010; Staszewski, 2002; Worden and Burrows, 2001] and process
monitoring [Wang et al., 2002]. This section tries to apply PSO and finite element
analysis in sensor placement optimization in order to get enough information of

75
 Chapter 4: Sensor Classification and Sensor Placement Optimization

machine structure using a small number of sensors and ensure the accuracy and
correctness of condition monitoring.

4.6.1 Problem Description

Modal analysis (finite element analysis) is a very important method for fault
diagnosis and condition monitoring. Faults of a machine, such as crack, axis
loosening and fatigue, usually accompany with the change of physical parameters,
such as natural frequency, modal damping, vibration mode and frequency response
function. The faults can be diagnosed according to these changes. The machine’s
vibration is supposed to be a n degree of freedom linear time-invariant system
which differential function can be written as [Wei and Pan 2010]:
xx x
M x(t )  C x(t )  Kx(t ) f (t ) (4.1)
where: M , C and K are the system mass, damping and stiffness matrix
x xx
respectively which are n u n matrix. x (t ) , x(t ) and x (t ) are n order response
vectors of system displacement, velocity and acceleration respectively. f (t )
represents n order excitation force vector. Then the frequency displacement
response function can be obtained by Fourier transform and set x(t ) xe jZt as:
x(Z ) H (Z ) F (Z ) (4.2)
where H (Z ) means the frequency displacement response function which is a
matrix. If the actuation is charged in i point of the machine, the frequency
response function of j point can be written as:
n I jr Iir
H ij (Z ) ¦ Z
r 1
2
M r  jZ Cr  K r
(4.3)

where M r , C r K r and I r represent modal mass, modal damping, modal stiffness

and each order vibration mode vector. Eq. (4.3) shows the relationship between
transfer function and the modal parameters and for a certain machine the value of
(Z 2 M r  jZCr  Kr ) is always the same because it only depends on the frequency
and damping ratio. Therefore, the value of frequency response function depends on
vibration mode vector of i and j points.
Let I [I1 , I2 ," , In ] be a displacement mode in which Ii is a N dimension vector
where N means the freedom degree of the machine structure. Let m be the number
of sensors (or number of measurement points) mounted on the machine while
o N  m be a non-measurement points. The fitness function can be as:
n n
f ¦ ¦ ¦I I
i 1 j 1 ro
ri rj (4.4)

76
 Chapter 4: Sensor Classification and Sensor Placement Optimization

where Iri means the r th component of j th vibration mode and r  o means all
calculation vectors are of non-measurement points. Compared Eq. (4.3) and
Eq.(4.4), it is only task to find the minimum value of Eq. (4.4) for the optimal
distribution of sensors. Therefore, it is chosen to be fitness function to find optimal
placement of sensors.

4.6.2 Application of PSO in Sensor Placement Optimization

4.6.2.1 The Process of PSO Application in Sensor Placement Optimization

The principle of PSO has been introduced in Section 3.5. This section induced how
to apply PSO to solve the Sensor placement optimization problems. Sensor
placement can be solved by simple mathematical calculation, but it will be time-
consuming. For example, if there are n possible measuring points and m sensors
are available to set up there will be n !/ (m !(n  m)!) times need to be calculated.
PSO is a good optimization algorithm which can easily solve this problem. Fig.
4.10 shows the structure to apply PSO in sensor placement optimization. First of all,
the machine structure is analyzed using finite element analysis, and at the same,
according to the shape and the application, all possible measurement points can be
determined. From result of above step, all vibration displacement modes can be
calculated. Then, input all of these data to PSO to find the optimal sensor
placement which can be sent to design and management center. According to the
result, the staff can improve the structure design, or make it is easy to monitor the
machine with high accuracy and correctness.

Evaluate fitness For all

particle with Eq. 4.4
Start
Finite Element
CAD Mode Is this fitness better
Analysis
than pbest ?
Calculation all NO
vibration Initialization YES
displacement of velocities Set this value as pbest Keep origin pbest
mode for each and
measurement Positions
Find all points and structure Loop for Is pbest better than
possible parameters all particles NO Loop until
gbest ? maximum iteration
measurement
points
Set pbest as gbest Keep origin gbest

Update velocity and position

with Eq. 3.21 and Eq. 3.22

Design and Management YES Is termination condition NO

Optimal sensor placement
Center satisfied ?

Fig. 4.10 Structure of PSO Application in Sensor Placement Optimization

4.6.2.2 Case Study and Its Results

In order to validate the effectiveness of the proposed method, a blower is chosen to
analyze. The 3D model is bolted based in the practical installation and possible 10
measurement points are chosen to analyze (Fig. 4.11). When it is analyzed, the
elastic modulus is set to E 210 Gpa, mass density to 7800 kg/m3 and Poisson ratio

77
 Chapter 4: Sensor Classification and Sensor Placement Optimization

to 0.3. The 3D solid model of the blower is built using the three-dimensional
software Solidworks and then import to ANSYS 13.0 to carry out finite element
analysis calculation and modal analysis. The blower is bolted to the floor in real
installation, and thus the boundary condition of baseboard of blower is set to fixed
constraint. This study calculates total 10 order natural frequency (Table 4.7) and its
10 vibration mode shapes of the blower are obtained. The finite element model and
its first four vibration modes are shown in Fig. 4.12. Fig. 4.12(a) to Fig. 4.12 (d)
shows from first to forth order of vibration shape mode respectively. In these
figures, the arrows mean the movement directions of that mode. The natural
frequency results (displacement) in total is shown in Table 4.8, and in three
different directions (X, Y and Z) are shown in Table 4.9-Table 4.11.

Table 4.7 Main Natural Frequencies of Blower

Order Frequency Order Frequency

1 58.282 6 185.62
2 116.77 7 214.41
3 121.98 8 229.1
4 164.62 9 243.49
5 165.52 10 250.3

Fig. 4.11 Initial Placement of Measuring Points on the Blower

78
 Chapter 4: Sensor Classification and Sensor Placement Optimization

Fig. 4.12 The Finite Element Model of Blower and Its First Four Modes

Table 4.8 Total Displacement Mode for Each Point Order

Measuring st 10th
1 order 2nd order 3rd order 4th order 5th order 6th order 7th order 8th order 9th other
Point order

1 0.17836 9.060e-2 7.433e-2 6.391e-2 4.649e-2 0.2028 0.32562 0.10588 4.435e-2 3.509e-2
2 0.16318 9.666e-2 6.106e-2 6.268e-2 4.406e-2 0.18447 0.30796 9.893e-2 4.048e-2 4.081e-2
3 0.17333 0.13224 8.327e-2 7.811e-2 4.548e-2 0.259 0.31935 0.11036 4.282e-2 2.617e-2
4 0.17537 0.21849 0.12898 0.10694 4.355e-2 0.38771 0.30514 0.1153 4.059e-2 3.384e-2
5 5.862e-3 1.840e-3 6.917e-4 1.681e-3 2.018e-3 1.222e-2 2.505e-2 9.816e-3 6.142e-4 4.418e-3
6 2.433e-2 2.246e-2 9.974e-3 1.343e-2 6.225e-3 5.518e-2 3.595e-2 3.119e-2 3.028e-3 3.148e-2
5.9661e- 2.6120e- 2.7633e- 7.6294e- 2.009e- 1.951e- 4.300e- 5.0898e- 2.5647e- 8.9125e-
7
10 10 10 11 11 10 10 11 11 11
8 0.2517 9.544e-3 8.621e-2 3.165e-2 3.646e-2 0.10458 0.14308 0.19168 9.056e-2 0.12752
9 0.31278 4.151e-2 9.815e-2 5.240e-2 4.757e-2 0.11371 0.1602 0.18728 4.534e-2 0.12261
10 0.3195 6.082e-2 0.10333 6.415e-2 5.369e-2 0.20737 0.15862 0.2448 4.327e-2 0.16902

79
 Chapter 4: Sensor Classification and Sensor Placement Optimization

Table 4.9 X Directional Displacement Mode for Each Point Order

Measuring 10th
1st order 2nd order 3rd order 4th order 5th order 6th order 7th order 8th order 9th other
Point order

1 -5.07e-3 9.585e-2 -4.30e-3 6.389e-2 3.341e-3 0.19631 -5.48e-2 0.10362 -7.03e-4 2.618e-2
2 -4.91e-3 8.861e-2 -4.54e-3 6.260e-2 3.204e-3 0.18686 -5.21e-2 9.645e-2 -6.69e-4 2.406e-2
3 -6.96e-3 0.1349 -5.64e-3 7.755e-2 4.552e-3 0.25141 -6.62e-2 0.10779 -1.02e-3 1.254e-2
4 -1.23e-2 0.21812 -8.48e-3 0.10665 7.873e-3 0.38268 -9.18e-2 0.10778 -1.79e-3 -2.65e-2
5 4.178e-5 9.337e-4 -1.40e-4 1.545e-3 6.528e-5 3.761e-3 -1.82e-3 2.943e-3 -3.67e-5 1.914e-3
6 -7.02e-3 6.974e-3 7.819e-4 7.756e-3 1.026e-3 3.040e-2 7.178e-3 -1.79e-4 2.035e-3 -3.20e-3
6.717e- -1.3584e- 1.3694e- -5.702e- -9.6312e- -3.1304e- 6.3539e- 1.2883e- 1.4565e- 3.8828e-
7
12 10 11 11 12 10 11 11 12 11
8 -9.822e-4 8.667e-3 2.214e-4 3.020e-2 -4.575e-3 3.122e-2 -3.36e-2 0.19303 3.795e-4 0.12611
9 2.46e-3 3.002e-2 -2.68e-3 4.980e-2 -2.146e-3 9.763e-2 -4.31e-2 0.18305 1.835e-5 0.11062
10 -2.6e-3 3.381e-2 -3.37e-3 5.157e-2 -2.602e-3 0.10477 -4.23e-2 0.18105 -4.48e-4 0.10807

Table 4.10 Y Directional Displacement Mode for Each Point Order

Measuring 1st 3rd 7th 9th 10th

2nd order 4th order 5th order 6th order 8th order
Point order order order other order
-5.21e- -6.76e- -9.81e-
1 -4.59e-3 4.28e-2 -3.77e-4 1.085e-2 -1.19e-2 -3.50e-3 -3.84e-3
2 2 3
-4.95e- 4.459e- -6.46e- -9.65e-
2 -7.17e-3 -4.24e-3 1.072e-2 -1.89e-2 -1.18e-2 -8.59e-3
2 2 2 3
-8.09e- 6.938e- -1.67e-
3 -6.84e-3 -1.05e-3 1.763e-2 -1.77e-2 -0.1073 -5.09e-3 -6.50e-3
2 2 2
-3.13e-
4 -0.1542 -1.28e-2 0.12835 -1.65e-3 3.355e-2 -3.35e-2 -0.2039 -8.85e-3 -1.23e-2
2
8.573e- 2.018e- -7.87e- -3.39e-
5 2.268e-4 1.432e-4 -2.52e-4 5.930e-4 1.417e-3 2.517e-4
4 4 4 6
-4.25e- -2.38e- -7.73e- -8.34e-
6 2.351e-3 -2.21e-3 1.412e-3 -3.25e-4 -1.11e-2 -8.52e-3
3 3 3 4
- - - -
3.1173e- 5.5385e- 7.3163e- 2.3211e- 2.3103e- 5.8361e-
7 9.490e- 6.281e- 3.944e- 0.605e-
11 12 13 11 12 13
11 12 11 13
8 6.86e-2 5.901e-4 2.12e-2 -3.92e-4 -1.79e-2 -2.00e-3 -3.3e-2 -1.30e-3 -2.1e-2 -1.94e-3
5.224e- 1.575e- -1.39e- -3.56e-
9 -6.94e-2 -1.43e-2 -4.16e-3 -3.77e-2 -2.21e-2 -1.24e-2
2 2 2 3
3.905e- 1.841e- -1.13e- -4.11e-
10 -1.87e-2 -2.56e-2 -3.92e-2 -6.34e-2 -6.66e-2 -3.80e-2
2 2 2 4

80
 Chapter 4: Sensor Classification and Sensor Placement Optimization

Table 4.11 Z Directional Displacement Mode for Each Point Order

Measuring st 4th 5th 8th 10th

1 order 2nd order 3rd order 6th order 7th order 9th other
Point order order order order
-4.53e- 2.239e- 2.347e-
1 0.16931 1.879e-2 -5.70e-2 8.68e-3 6.517e-2 0.31279 4.327e-2
2 2 2
-5.27e- -4.25e- 1.322e- 3.204e-
2 0.15121 -2.33e-2 -4.03e-2 -2.36e-3 0.29841 3.957e-2
3 2 2 2
4.932e- -4.21e- 2.198e- 2.328e-
3 0.15458 1.774e-2 -4.40e-2 5.924e-2 0.29066 4.027e-2
3 2 2 2
8.660e- -2.73e- 1.825e- 1.603e-
4 8.796e-2 1.260e-2 1.308e-2 4.474e-2 0.1988 2.641e-2
3 2 2 2
1.195e- 1.903e- -6.82e- -6.28e-
5 -6.27e-3 1.490e-3 -9.35e-4 7.195e-3 2.383e-2 8.225e-4
3 3 3 3
-1.11e- -5.78e- 2.910e- 3.007e-
6 1.839e-2 -2.14e-2 9.795e-3 -4.79e-2 3.553e-2 2.075e-3
2 3 2 2
- - - -
2.2494e- 8.1034e- 6.7551e- 3.0784e- 6.1781e- 2.7575e-
7 4.605e- 1.717e- 4.387e- 7.837e-
10 11 11 10 10 11
11 11 11 11
1.070e- -7.81e- -2.76e- -1.72e-
8 0.23909 3.971e-3 8.188e-2 -8.09e-3 -0.1400 3.71e-2
4 3 2 2
-1.60e- -4.64e- 2.965e- 5.089e-
9 0.31009 -1.83e-2 9.733e-2 -7.03e-2 -0.1563 -4.47e-2
2 2 2 2
-3.22e- -5.38e-
10 0.3156 -4.82e-2 0.1005 -0.1671 -0.1538 0.1491 -4.33e-2 0.12277
2 2

All parameters are presented in above figures and tables. Accordingly, the process
of PSO application in sensor placement optimization in Fig. 4.10 and the fitness
function Eq. (4.4), the optimal sensor placement of the blower can be obtained
using PSO. For the PSO algorithm, the number of particles is initialized as 10 and
n (1 ~ 10) sensors are assumed to place on blower measuring points. The weight Z
is set to 1.2-0.8 with decreasing linearly, and the acceleration coefficient c1 and c2
is set as 1.2. The vibration mode parameters in Table 4.8-Table 4.11 are input to
the PSO respectively which can be used to calculate the fitness value.
Table 4.12 shows the smallest fitness value and the corresponding sensor
placement for the different number of measuring points using the total
displacement mode for each measuring point (Table 4.8). From this table, the
amount of information on the blower increases with the increasing of measuring
points, because the fitness become smaller and smaller. The smallest fitness is very
big (8.824) when there only one sensor place on the blower while it became very
small even equal 0 when the number of sensors increasing to 8, 9 to 10. From this
table, the importance of measuring points can be obvious observed. The point 4 is
the most important while the point 7 is the least important. The amount information
also can be calculated from this table. Just take measuring point 6 as example, its
amount information can be calculated as fitness value in point 5 minus in point 6
(2.422-1.213=1.209).
Table 4.13, Table 4.14 and Table 4.15 present the smallest fitness values and the
corresponding sensor placement for the different number of measuring points using
displacement modes for each measuring point in X direction, Y direction and Z

81
 Chapter 4: Sensor Classification and Sensor Placement Optimization

direction respectively. With these tables, the same conclusions can be obtained as
the Table 4.12, and what’s more, when the same number of sensors is planned to
installed to the blower, the optimal places may different using different
displacement modes. When optimal sensor placement is applied in real machine, it
is very significant to know which direction is important for deformation referring
to failure of machine.
Fig. 4.13 to Fig. 4.16 show fitness values changes with the changes of iteration
PSO ( n 5 ) for total, X direction, Y direction and Z direction respectively. From
these figures, the optimal sensor placement can be obtained within 20 iterations of
PSO for using all kinds of displacement mode. Combining all these figures and
tables, PSO can successfully solve the optimal sensor placement problem.
As PSO has it important advantages in solving the optimization and NP problems,
it is employed to solve sensor placement optimization problem for improving
product design and fault diagnosis. Fitness is established for PSO application in
sensor placement optimization based on the analysis on placement guidelines of
vibration sensors. Generally, the proposed method combined the structure finite
element modeling and its modal analysis, and PSO the carry out the optimal sensor
placement distribution. The proposed method combining PSO and FEM analysis
can be applied in machine level and component level but not system level because
it need finite element mode and modal analysis of the structure. Therefore, the
future research will be on the method for optimal sensor distribution in system
level.

Table 4.12 Optimal Sensor Placement for Different Number of Measuring Point using Total
Displacement Mode
Measuring Point Sensor place Measuring Point Sensor place
Fitness Fitness
No. position No. position
1 4 8.824 6 1 2 3 4 9 10 1.213
2 4 10 6.793 7 1 2 3 4 8 9 10 0.059
3 3 4 10 5.183 8 1 2 3 4 6 8 9 10 0.004
12345689 4.1E-
4 3 4 9 10 3.786 9
10 18
123456789
5 1 3 4 9 10 2.422 10 0
10

Table 4.13 Optimal Sensor Placement for Different Number of Measuring Point using X
Direction Displacement Mode
Measuring Point Sensor place Measuring Point Sensor place
Fitness Fitness
No. position No. position
1 4 1.7706 6 1 2 3 4 9 10 0.1886
2 34 1.3236 7 1 2 3 4 8 9 10 0.0046
3 134 1.0166 8 1 2 3 4 6 8 9 10 0.0002
12345689 4.26E-
4 1 3 4 10 0.7351 9
10 19
123456789
5 1 2 3 4 10 0.4605 10 0
10

82
 Chapter 4: Sensor Classification and Sensor Placement Optimization

Table 4.14 Optimal Sensor Placement for Different Number of Measuring Point using Y
Direction Displacement Mode
Measuring Point Sensor place Measuring Point Sensor place
Fitness Fitness
No. position No. position
1 4 0.3971 6 1 2 3 4 9 10 0.0299
2 34 0.2888 7 1 2 3 4 8 9 10 0.0017
2.240e-
3 3 4 10 0.1860 8 1 2 3 4 6 8 9 10
5
12345689 4.17E-
4 3 4 9 10 0.1257 9
10 20
123456789
5 2 3 4 9 10 0.0729 10 0
10

Table 4.15 Optimal Sensor Placement for Different Number of Measuring Point using Z
Direction Displacement Mode
Measuring Point Sensor place Measuring Point Sensor place
Fitness Fitness
No. position No. position
1 10 2.7715 6 1 2 3 8 9 10 0.2538
2 9 10 2.0660 7 1 2 3 4 8 9 10 0.0478
3 1 9 10 1.4790 8 1 2 3 4 6 8 9 10 0.0032
12345689 2.287E-
4 1 3 9 10 0.9906 9
10 18
123456789
5 1 2 3 9 10 0.5705 10 0
10

3.5
Fitness Value

2.5

0 20 40 60 80 100 120 140 160 180 200

Iteration

Fig. 4.13 Fitness Changes with Change of Iteration PSO ( n 5 ) for Total Displacement
Mode

83
 Chapter 4: Sensor Classification and Sensor Placement Optimization

0.5

0.495

0.49

0.485
Fitness Value

0.48

0.475

0.47

0.465

0.46
0 20 40 60 80 100 120 140 160 180 200
Iteration

Fig. 4.14 Fitness Changes with Change of Iteration PSO ( n 5 ) for X Direction
Displacement Mode

0.16

0.15

0.14

0.13
Fitness Value

0.12

0.11

0.1

0.09

0.08

0 20 40 60 80 100 120 140 160 180 200

Iteration

Fig. 4.15 Fitness Changes with Change of Iteration PSO ( n 5 ) for Y Direction
Displacement Mode

0.95

0.9

0.85

0.8
Fitness Value

0.75

0.7

0.65

0.6

0.55
0 20 40 60 80 100 120 140 160 180 200
Iteration

Fig. 4.16 Fitness Changes with Change of Iteration PSO ( n 5 ) for Z Direction
Displacement Mode

84
 Chapter 4: Sensor Classification and Sensor Placement Optimization

4.6.3 Application of BCA in Sensor Placement Optimization

4.6.3.1 The Process of Application of BCA in Sensor Placement

Optimization
The principle of BCA was introduced in Section 3.5.2 and this part will present
how to apply BCA to solve the problem of sensor placement optimization. Fig.
4.17 shows the process of application of BCA to find the optimal sensor placement
for manufacturing machines and other equipment. The necessity of using
intelligent algorithm was described in Section 4.6.2.1. The 3D model of a machine
can be established using Solidworks and transfer it into FEM software ANSYS to
calculate the vibration displacement mode for all measuring point. The parameters
obtained from vibration displacement modes are input to BCA to find the optimal
sensor placement to get as much as information using as less as sensors. The results
can be used to improve the machine design, management and operations.

Fig. 4.17 Structure of BCA Application in Sensor Placement Optimization

4.6.3.2 Case Study and Its Results

The object of this case study is the same as Section 4.6.2.2 and the finite element
model analysis is the same as well. Therefore, the 3D model and the vibration
mode shapes of blower are shown in Fig. 4.11 and Fig. 4.12, and the parameters of
the blower are the same as shown in Table 4.7 to Table 4.11. All these parameters
are input to the BCA to find the optimal sensor placement. For the BCA algorithm,
the colony size is set to 10, the maximum cycle number is set to 50, and the
number of trial to improve a solution is set to 20.
The final results are the same as PSO application in Section 4.6.2.2 in Table 4.12 to
Table 4.15, and the explanations are also the same as that. In order to compare to
the PSO method, Fig. 4.18 to Fig. 4.21 show the the fitness value changing of
iteration of BCA in total, X direction, Y direction and Z direction displacement
mode for 5 sensors ( n 5 ) installing on the blower. From these four figures, the
optimal can be found within 10 iterations of BCA and the convergence is faster
than PSO compared to Fig. 4.13 to Fig. 4.16.

85
 Chapter 4: Sensor Classification and Sensor Placement Optimization

2.65

2.6

Fitness Value 2.55

2.5

2.45

0 5 10 15 20 25 30 35 40 45 50
Iteration

Fig. 4.18 Fitness Changes with Change of Iteration BCA ( n 5 ) for Total Displacement
Mode

0.5

0.495

0.49

0.485
Fitness Value

0.48

0.475

0.47

0.465

0.46
0 5 10 15 20 25 30 35 40 45 50
Iteration

Fig. 4.19 Fitness Changes with Change of Iteration BCA ( n 5 ) for X Direction
Displacement Mode

Best function values

0.16

0.15

0.14

0.13
Fitness Value

0.12

0.11

0.1

0.09

0.08

0 5 10 15 20 25 30 35 40 45 50
Iteration

Fig. 4.20 Fitness Changes with Change of Iteration BCA ( n 5 ) for Y Direction
Displacement Mode

86
 Chapter 4: Sensor Classification and Sensor Placement Optimization

Best function values

0.95

0.9

0.85

0.8
Fitness Value

0.75

0.7

0.65

0.6

0.55
0 5 10 15 20 25 30 35 40 45 50
Iteration

Fig. 4.21 Fitness Changes with Change of Iteration BCA ( n 5 ) for Z Direction
Displacement Mode

4.7 Summary

This Chapter introduced sensor classification scheme for categorizing and list some
criteria to categorize sensors. Most of sensors are very mature in the market. When
a machine needs to be monitored, the properties of signals and the parameters of
sensors can be firstly determined, and then the suitable sensors can be found from
the market. The more important thing in this Chapter is to define a sensor
placement optimization problem which is a NP problem, and introduce two Swarm
Intelligence algorithms: PSO and BCA to solve this problem. The Swarm
Intelligence algorithms are very good at solving the NP problems, and thus, they
are suitable to solve the sensor placement optimization problems. Finally, a case
study is descripted in this Chapter which shows that both BCA and PSO can find
the optimal sensor placement accurately and fast.
When a machine needs to be monitored, one always wants to use as few as possible
sensors to obtain as much as possible information of the machine. To find the
optimal sensor placement could be a basis of condition monitoring of
manufacturing machines which can reduce the number of sensors used and thus
reduce the cost.

87
Chapter 4: Sensor Classification and Sensor Placement Optimization

88
 Chapter 5: Signal Preprocessing and Feature Extraction

5 Signal Preprocessing and Feature Extraction

5.1 Introduction

The main challenge of Condition-based Maintenance is that how to find the

relations between the collected data/signals and the conditions of machines. For a
complex machine, there could be many sensor are installed for monitoring its
condition and thus too many signals and information are collected. However, the
collected data cannot indicate the machine condition automatically, and sometimes
it is very difficult to get the real machine condition because the mass of signals.
Data are rarely useful or usable in their row form, because they may contain too
much noise or too weak, and sometimes only because they are too large. Consider,
for example, vibration data sampled at 100 kHz. Such large amounts of data are
unmanageable unless they are processed and reduced to a form that can be
manipulated easily by fault diagnostic and prognostic algorithms. The objective in
processing the raw sensor data is to reflect the true and correct information of
machine from the signals.
Generally, there are three steps of the raw sensor signal processing: signal
preprocessing, feature extraction and feature selection. The aim of signal
preprocessing is to improve the general quality of the signal, or in other words,
improving the signal-to-noise ratio, for more accurate analysis and measurement,
which eventually may facilitate the efficient extraction of useful information, that
is, the indicators of the condition of a failing component or subsystem. The tools of
preprocessing include filtering, amplification, data compression, data validation,
and de-noising. The aim of feature extraction is to extract features or indicators
from the preprocessed data that are characteristic of an incipient failure or fault.
The main aim of feature selection is to determine a minimal feature subset from a
problem domain while retaining a suitably high accuracy in representing the
original features. Table 5.1 shows the techniques of these three phases. This
Chapter will introduce some of these techniques which are used in IFDPS.

Table 5.1 7KH0HWKRGVRI6LJQDO3UHSURFHVV)HDWXUH([WUDFWLRQDQG)HDWXUH6HOHFWLRQ

Feature Extraction
Signal Feature
Preprocessing Frequency Time-Frequency Selection
Time Domain
Domain Domain
Filter, Continues
Mean, RMS, Fourier Short Time
Amplification, Principal
Shape factor, Transform Fourier
Signal Component
Skewness, (CFD), Discrete Transformation
Conditioning, analysis,
Kurtosis, Fourier (STFT), Wavelet
Extracting Weak Transform Support Vector
Crest factor, Transform
Signals, De- (DFT), Machine,
Entropy Error, (WT), Wavelet
noising, Boosting Tree
Entropy Fast Fourier Packet (WP),
Vibration Signal Algorithm, etc.
estimation, etc. Transform etc.
Compression, etc. (FFT), etc.

89
 Chapter 5: Signal Preprocessing and Feature Extraction

5.2 Signal Preprocessing

There are many methods for signal preprocessing as shown in Table 5.1. As
mentioned above, the main aims of signal pre-processing are to improve the signal-
to noise ratio, enhance the signal characteristics, and facilitate the efficient
extraction of useful information from the signals. The electrical signals generated
by sensors are often not adequate for useful information extraction because they
may be very nosy, of low amplitude, biased and dependent on secondary
parameters such as temperature and humidity. What’s more, the quantities of
interested parameters may be not able to be measured directly but can only
measure their related quantities. Therefore, signal conditioning is required which
can be performed with hardware and/or software which can include: amplification,
filtering, converting, range matching, isolation and any other processes required to
make sensor output suitable for processing after conditioning [Gutierrez-Osuna et
al., 2003]. Denoising techniques aim at eliminating noise from measured data while
trying to preserve the important signal features (such as texture and edges) as much
as possible[Ramani et al., 2008]. It is very important step to enhancing data
reliability and improving the accuracy of signal analysis methods. Wavelet based
denoising methods have been successfully applied for signal analysis to improve
the signal-to-noise ratio[Benouaret et al., 2012; Patil & Chavan, 2012]. Soft-
thresholding [Donoho, 1995] and wavelet-shrinkage denoising [Zheng et al., 2000]
are two popular denoising methods. There are still some other denoising techniques:
adaptive threshold denoising for fault detection in power systems [Yang & Liao,
2001], acoustic emission signal denoising for fatigue cracks detection in rotor
heads [Menon et al., 2000], denoising using modulus maxima algorithm for
structure fault detection in fighter aircraft [Hu et al., 2000], signal decomposition
technique (wavelets, wavelet packets and matching pursuit method) based
denoising methods for improving signal-to-noise ratio of knee-joint vibration
signals [Krishnan & Rangayyan, 2000], and reducing background noise level using
The second order displaced power spectral density (SDPSD) function for localized
defects in roller bearings [Piñeyro et al., 2000]. The amount of data collected from
industrial systems tends to be voluminous and, in most cases, difficult to manage
because the increasing of sensors and sample rates. Therefore, data compression is
very important for condition monitoring system especially for those implemented
online or Internet-based systems. Transient analysis is mostly used to compress
data because it can significantly improve the performance of sensor arrays with
careful instrument design and sampling procedures which are: improving
selectivity, reducing acquisition time and increasing sensor lifetime. There are
main three classes of transient analysis methods: Sub-sampling method [Gutierrez-
Osuna et al., 1999; Kermani et al., 1998; Roussel et al., 1998; White et al., 1996],
parameter-extraction method [Eklöv et al., 1997; Gibson et al., 1997; Llobet et al.,
1997; D. M. Wilson & DeWeerth, 1995], and system-identification method [Eklöv
et al., 1997; Gutierrez-Osuna et al., 1999; Nakamoto et al., 2000]. The signal
preprocessing techniques for condition monitoring are mature, and more techniques
and more detail can be referred in the literatures [Gutierrez-Osuna et al., 2003;
Marwala, 2012; Vachtsevanos et al., 2006].

90
 Chapter 5: Signal Preprocessing and Feature Extraction

5.3 Feature Extraction

Feature and condition indicator extraction and selection play curial roles in
condition monitoring especially for accuracy and reliability of fault diagnosis and
prognosis. The function of condition monitoring mainly depends on a set of
features extracted from sensor data that can distinguish between fault categories of
interest, and detect and isolate a specific fault at early initiation stages. These
features should be fairly insensitive to noise and within fault class variations. It
should beware that not losing useful information in feature extraction stage. For
time series signals, such as vibration signals, voltage signals and current signals,
the features can be extracted from four domains: time domain, frequency domain,
time-frequency domain and wavelet domain.

5.3.1 Feature Extraction in Time Domain

Features in time domain is very traditional methods for extraction features, but is
very widely used in fault diagnosis and prognosis which mainly computer the
statistical parameters from signals. The following features are some of these
statistical parameters [Vachtsevanos et al., 2006; Wang & Zhang, 2010]:
Peak value,
1
Pv ª max  xi   min  xi  ¼º (5.1)
2¬
where xi (i 1, 2" , N ) is the amplitude at sampling point i and N is the number
of sampling points.
RMS value,
1 N
¦  xi 
2
RMS (5.2)
N i1

Standard deviation,
1 N
¦  xi  x 
2
SD (5.3)
N i1

Kurtosis value,
1
¦  xi  x 
N 4

Kv N i 1
(5.4)
 RMSValue 
4

Crest factor,
Peak Value
Crf (5.5)
RMS Value

91
 Chapter 5: Signal Preprocessing and Feature Extraction

Clearance factor,
PeakValue
Clf 2 (5.6)
§1 ·
¨ ¦ i 1 | xi
N
|¸
©N ¹
Impulse factor,
Peak Value
Imf (5.7)
1
¦
N
| xi |
N i 1

Shape factor,
RMS Value
Shf (5.8)
1
¦
N
| xi |
N i 1

Weibull negative log-likelihood value was used recently for feature extraction from
vibration signals. The Weibull negative log-likelihood value ( Wnl ) and the normal
negative log-likelihood value ( Nnl ) of the time domain vibration signals are used
as input features along with the other features defined above in this study. The
negative log-likelihood function is defined as:
N
ȁ ¦Log ¬ª f  x , T , T ¼º
i 1
i 1 2

where f  xi ,T1 ,T2  is the probability density function ( pdf ). For Weibull negative
log-likelihood function and normal negative log-likelihood function, the pdfs are
computed as follows:
Weibull pdf :

ª § x ·E º
f  xi , E ,K  EK  E | xi |E 1 exp «  ¨ i
¸ » (5.9)
«¬ © K ¹ »¼

Where E and K are the shape and the scale parameters respectively.
Normal pdf :

­ ª  x  P 2 º ½
1 ° ¼°
exp ® ¬
i
f  xi , P , V  ¾ (5.10)
V 2S ° 2V 2
°
¯ ¿
Where P and V are the mean and the standard deviation respectively.
There are still three time domain parameters, i.e. Activity, mobility and complexity,
can be used for feature extraction [Hjorth, 1970; Xinyang Li et al., 2011]:
Activity var( x(t )) (5.11)

92
 Chapter 5: Signal Preprocessing and Feature Extraction

x
Activity ( x)
Mobility (5.12)
Activity ( x(t ))

x
Mobility ( x)
Complexity (5.13)
Mobility ( x(t ))

The above three parameters are often referred as Hjorth parameters and have been
widely applied [Cecchin et al., 2010; Obermaier et al., 2001]. There are also other
parameters can be used for feature extraction: Time-Domain Morphology and
Gradient [Mazomenos et al., 2012], Correlation, Covariance and Convolution
[Vachtsevanos et al., 2006]. For details of these techniques, the corresponding
literatures can be found in above mentioned references.

5.3.2 Feature Extraction in Frequency Domain

In many situations, especially with rotating machinery [Eisenmann & Eisenmann,

1998], the frequency domain data of measured time signals, such as vibration,
carries a great deal of information useful in diagnosis [Vachtsevanos et al., 2006].
Frequency domain methods are difficult to use in that they contain more
information than is necessary for fault detection. There is no method to select the
frequency bandwidth of interest, and they are usually noisy in anti-resonance
regions [Ewins, 1995; Marwala, 2012]. However, frequency domain methods still
have some advantages: 1) the measured data comprise the effects of out-of-
frequency-bandwidth modes; 2) one measurement offers ample data; 3) modal
analysis is not necessary and consequently modal identification errors are
circumvented; and, 4) frequency domain data are appropriate to structures with
high damping and modal density [Marwala, 2012]. A fault on a component of a
machine might be indicated by the base rotational frequency, two times of this
frequency or n times of this frequency. This principle can be used in fault
diagnosis and prognosis. The main algorithm of frequency of a signal is Fourier
Transform which is introduces in this section.
The Fast Fourier Transform (FFT) is basically a computationally efficient
technique for calculating the Fourier transform which exploits the symmetrical
nature of the Fourier transform [Marwala, 2012]. The theory of FFT is retrieved
here from the literature [Ewins, 1995; Marwala, 2012]. If the FFT is applied to the
response, the following expression is obtained:
1 f
X (Z )
2S ³ f
x (t )e  iZ t dt (5.14)

Similarly, the transformed excitation can be written as:

1 f
F (Z )
2S ³
f
f (t )e  iZt dt (5.15)

93
 Chapter 5: Signal Preprocessing and Feature Extraction

Frequency Response Function (FRF) D (w) of the response at position i to the

ij
excitation at j is the ratio of the Fourier transform of the response to the transform
of the excitation:
X i (Z )
D ij (Z ) (5.16)
F j (Z )

The FRF matrix is related to the spatial properties by the following expression:
1
>D (Z)@ ª¬Z 2 > M @  jZ >C @  > K @º¼ (5.17)

Here D is the frequency response function, Z is the frequency, > M @ is the mass
matrix, >C @ is the damping matrix, > K @ is the stiffness matrix and j 1 . The
above transform is for the continue signals. For the discrete signals, the frequency
response function can be expressed as:
N
X (k ) ¦ x ( n )e
n 1
 j 2 S ( k 1)( n 1) / N
k 1, 2, " , N (5.18)

where N is the number of time series x ( n ) . Fig. 5.1 shows a vibration signal with
the sample rate 4096 in one second while Fig. 5.2 shows the corresponding
frequency response function. From Fig. 5.2, the base frequency of this vibration
signal is 46 Hz, the second order frequency is 92 Hz and the third order frequency
is 138 Hz. There are also some other features can be extracted from the FRF figure
to find the characteristics of the signals for fault diagnosis and prognosis. For
example, power spectral density (PSD):
1 1 2
\x X (k ) X ' ( k ) X (k ) (5.19)
N N
is for fault diagnosis and prognosis which is easier to see details in the frequency
response than using X (k ) [Vachtsevanos et al., 2006].

0.015

0.01

0.005
Amplitude (m/s 2)

-0.005

-0.01

-0.015
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Time (s)

Fig. 5.1 Vibration Signal in Time Domain

94
 Chapter 5: Signal Preprocessing and Feature Extraction

20

18

16

14

12
Amplitude (m/s 2)

10

0
0 200 400 600 800 1000 1200 1400 1600 1800 2000
Frequency (Hz)

Fig. 5.2 Frequency Response Function of Vibration Signal in Fig. 5.1

5.3.3 Feature Extraction in Time-Frequency Domain

Although FFT based methods are powerful tools for fault diagnosis and prognosis,
they are not suitable for non-stationary signals. For analysis in the time-frequency
domain, the Wigner-Ville distribution (WVD) and the short time Fourier transform
(STFT) are the most popular methods for non-stationary signal analysis. However,
WVD suffers from interference terms appearing in the decomposition, and STFT
cannot provide good time and frequency resolution simultaneously because it uses
constant resolution at all frequencies. Moreover, no orthogonal bases exist for
STFT that can be used to implement a fast and effective STFT algorithm
[Okamura, 2011; Vachtsevanos et al., 2006]. The methods for time-frequency
analysis are compared in Table 5.2 [Vachtsevanos et al., 2006]. This section mainly
introduces Wavelet transform for time-frequency analysis and feature extraction.
Wavelet transform is a time-frequency decomposition of a signal into a set of
“wavelet” basic function. Wavelet analysis has proved its great capabilities in
decomposing, denoising, and signal analysis which made the analysis of non-
stationary signals achievable as well as detecting transient feature components as
other methods were inept to perform since wavelet can concurrently impart time
and frequency structures. Wavelet Transform (WT) gives good time and poor
frequency resolution at high frequencies, and good frequency and poor time
resolution at low frequencies. Analysis with wavelets involves with breaking up a
signal into shifted and scaled versions of the original (or mother) wavelet, i.e., one
high frequency term from each level and one low frequency residual from the last
level of decomposition. There are three categories of this transformation:
Continuous Wavelet Transform (CWT), Discrete Wavelet Transform (DWT) and
WPD.

95
 Chapter 5: Signal Preprocessing and Feature Extraction

Table 5.2 Comparing Different Time-Frequency Analysis Methods

Methods Resolution Interference Speed

Good frequency and low time resolution for
Wavelet low-frequency components; low frequency
None Fast
Transform (WT) and high time resolution for high-frequency
components
Short time Slower
Depending on window function used, either
Fourier transform None than
good time or good frequency resolution
(STFT) CWT
Wigner-Ville Slower
distribution Good time and frequency resolution Severe than
(WVD) STFT
Choi-Williams
Less than Very
distribution Good time and frequency resolution
WVD slow
(CWD)
Cone-shaped
Less than Very
distribution Good time and frequency resolution
WVD slow
(CSD)

5.3.3.1 Continuous Wavelet Transform (CWT)

A CWT is used to divide a continuous-time function into wavelets. Unlike Fourier
transform, the continuous wavelet transform possesses the ability to construct a
time-frequency representation of a signal that offers very good time and frequency
localization [Soman & Ramachandran, 2005]. The continuous wavelet transform of
a time function x(t) is given by following equation:
f

³ x(t )\ (5.20)
*
CT (a, b) ( a ,b ) (t ) dt
f

where \ ( a ,b) (t ) is a continuous function in both the time domain and the frequency
*

domain called the mother wavelet and * represents operation of complex conjugate.
\ (*a,b) (t ) can be expressed as:

1 §t b· (5.21)
\ (*a , b ) (t ) \¨ ¸ where a , b  R , a z 0
a © a ¹
The main purpose of the mother wavelet is to provide a source function to generate
the daughter wavelets which are simply the translated and scaled versions of the
mother wavelet. As seen in Eq. (5.21), the transform signal CT ( a , b ) is defined on
a  b plane, which a and b are used to adjust the frequency and the time location
of the wavelet in Eq. (5.21). A small a produces a high-frequency wavelet when
high frequency resolution is needed and the reverse is also true. The WT’s superior
time-localization properties stem from the finite support of the analysis wavelet: as
b increases, the analysis wavelet transverses the length of the input signal, and a
increases or decreases in response to changes in the signal’s local time and
frequency content. Finite support implies that the effect of each term in the wavelet

96
 Chapter 5: Signal Preprocessing and Feature Extraction

representation is purely localized. This sets the WT apart from the Fourier
Transform, where the effects of adding higher frequency sine waves are spread
throughout the frequency axis.

5.3.3.2 Discrete Wavelet Transform (DWT)

In numerical analysis and functional analysis, DWT is a wavelet transform for
which the wavelet \ ( a,b) is discretely sampled. As with CWT, a key advantage it
has over Fourier transforms is temporal resolution: it captures both frequency and
location information (location in time). Usually, the DWT can be derived from
discretization of CWT. The most common discretization is dyadic method:
f

³ x(t )\ (5.22)
*
DT (a, b) ( j ,k ) (t )dt
f

1 § t  2j k ·
\ * (t ) \¨ j ¸ (5.23)
© 2 ¹
( j ,k )
2j

where a and b are replaced by 2 j and 2 j k respectively [Daubechies, 1988; Mallat,

1989]. An efficient way to implement this scheme using filters was developed by
Mallat [1989]. The original signal x (t ) passes through two complementary filters
and emerges as low frequency called approximations Aj (t ) and high frequency
called details Di (t ) as shown in Eq. (5.24). The decomposition process can be
iterated, with successive approximations being decomposed in turn, such that a
signal can be broken down into many lower-resolution components.
i j
f (t ) ¦ D (t )  A (t )
i 1
i j (5.24)

Where Di (t ) denotes the wavelet detail and Ai (t ) stands for the wavelet
approximation at the j th level. DWT analysis is more efficient still with the
identical accuracy [Goumas et al., 2001].
As discussed above, DWT can decompose the signal into two parts: low-frequency
A1 and high frequency D1 . In the process of decomposition, the lost information
belonging to the low frequency part is captured by the high frequency part. In the
next level of decomposition, this method will also decompose A1 into two parts:
low-frequency A2 and high frequency D2 . The lost information belonging to low
frequency A2 is capture by the high-frequency D2 , and thus, a deeper level
decomposition can be done. The 3-layer structure of signal based on DWT is
shown in Fig. 5.3 in which only approximation version is decomposed.

97
 Chapter 5: Signal Preprocessing and Feature Extraction

Fig. 5.3 3-layer Signal Decomposition by Discrete Wavelet Transform

For the case of signal with the maximum frequency 2048 Hz, D1 , D2 , D3 and A3
represent the frequency 1024~2048 Hz, 512~1024 Hz, 128~512 Hz and 0~128 Hz
respectively. The decomposed signals by DWT from vibration signal (Fig. 5.1) are
shown in Fig. 5.4.
D1 D2
0.1 0.2

0.05 0.1
Amplitude

Amplitude

0 0

-0.05 -0.1

-0.1 -0.2
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
Time (s) Time (s)
D3 A3
0.2 0.2

0.1 0.1
Amplitude

Amplitude

0 0

-0.1 -0.1

-0.2 -0.2
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
Time (s) Time (s)

Fig. 5.4 Decomposed Signals by DWT

5.3.3.3 Wavelet Packet Decomposition

The structure of wavelet packet Decomposition (WPD) is similar to DWT. The
typical 3 layers signal decomposition by WPD is shown in Fig. 5.5. Both have the
framework of multi-resolution analysis. The main difference in the two techniques
is the WPT can simultaneously break up detail ( Di ) and approximation ( Ai )
versions [Li et al., 2003] while DWT only breaks up as an approximation version.
Therefore, the WPD have the same frequency bandwidths in each resolution and
DWT does not have this property. The mode of decomposition does not increase or
lose the information within the original signals. Therefore, the signal with great
quantity of middle and high frequency signals can offer superior time-frequency
analysis. The WPT suits signal processing, especially non-stationary signals

98
 Chapter 5: Signal Preprocessing and Feature Extraction

because the same frequency bandwidths can provide good resolution regardless of
high and low frequencies.

Fig. 5.5 Decomposed Signals by WPD

Wavelet packets consist of a set of linearly combined usual wavelet functions

which inherit the attributes of their corresponding wavelet functions such as
orthonormality and time–frequency localization. A wavelet packet is a function
with three indices of integers, i , j and k which are the modulation, scale and
translation parameters, respectively [Shinde, 2004],
j
\ ij , k  t  2 2\ j
2 t  k 
i
i 1, 2, 3, " (5.25)

The wavelet functions \ j can be obtained from the following recursive relations:
f
\ 2 j t  2 ¦h  k \ i  2t  k  (5.26)
f

f
\ 2 j 1  t  2 ¦ g  k \ i  2t  k  (5.27)
f

The original signal f (t ) after j level of decomposition can be stated as:

2j
f t  ¦ f t  ,
i 1
j
i
(5.28)

while the wavelet packet component f ji (t ) can be stated by a linear combination of

wavelet packet functions \ ij , k (t ) in such a way:
f
f ji, k  t  ¦c  t \  t 
f
i
j,k
i
j,k
(5.29)

where the wavelet packet coefficients cij ,k  t  can be obtained from,

f
c ij , k  t  ³ f (t )\  t  dt
i
j ,k (5.30)
f

providing that the wavelet packet functions are orthogonal:

\ mj , k  t \ nj , k  t  0 if m z n (5.31)

99
 Chapter 5: Signal Preprocessing and Feature Extraction

Since different types of wavelet functions have different time–frequency structures,

a function with a time–frequency structure matching superlatively that of the
transient component must be used to effectively detect the transient component. In
general, the smooth wavelets are better for regular, stationary, periodic data and the
compact wavelets are better for non-stationary, transient data [Staszewski, 1997].
As a result, Daubechies 4 (Db4) wavelet function has been chosen for this case
after several trials as it is often chosen arbitrarily for signal analysis and synthesis
by experiments in many papers in the field (e.g., [Vafaei & Rahnejat, 2003]) there
is no computational logic behind the selection of Daubechies order.
The dilation equations may be used to generate orthogonal wavelets. The scaling
function ߮ሺ‫ݐ‬ሻ is a dilated (horizontally expanded) version of M (2t ) . The dilation
equation in general has the form [Vachtsevanos et al., 2006]:
M  t  c0M  2t   c1M  2t  1  c2M  2t  2   c3M  2t  3  (5.32)

The Daubechies D4 wavelet coefficients have values:

(1  3) (3  3) (3  3) (1  3)
c0 ,?c1 c2 c3 (5.33)
4 2 4 2 4 2 4 2
Thus, a particular family of wavelets is specified by a particular set of numbers,
called the wavelet filter coefficients. The above set of numbers c0 , c1 , c2 and c3 are
called the Db4 wavelet filter coefficients.
In general, for an even M number of wavelet filter coefficients ck (k 1, 2," , M  1) ,
the scaling function is defined by:
M 1
I t  ¦ c I  2t  k 
k 1
k
(5.34)

and the corresponding wavelet is derived as:

M 1
w t  ¦ ( 1)
k 1
k
ck I  2t  k  M  1 (5.35)

It is observed that the scaling function has a low-pass form, whereas the wavelet
function has a high-pass form. Thus, the wavelet function is essentially responsible
for extracting the detail (high-frequency components) of the original signal.

5.3.3.4 Wavelet-based Features

There are several types of features can be extracted from wavelet-based methods,
which can be categorized roughly into wavelet coefficients-based, wavelet energy-
based, singularity-based, and wavelet function-based methods. All these features
are retrieved from the literature [Vachtsevanos et al., 2006].
The wavelet coefficients ck can be used to extracted features for fault detection and
prognosis. The standard deviations of the coefficients are good features for feature
extraction. For example, for each of the signal, wavelet packet is applied up to the
fourth level thus giving 16 signal coefficient sets. The wavelet packet coefficients

100
 Chapter 5: Signal Preprocessing and Feature Extraction

and their corresponding standard deviations for medium-worn fault sampled at

second configuration are shown in Fig. 5.6. At the end, the standard deviation of
wavelet packet coefficients of pre-processed signals is used as features for fault
diagnosis and prognosis.

Fig. 5.6. Wavelet Packet Coefficients and Their Relevant Standard Deviation

Energy-based features: The most important advantage of wavelets versus other

methods is their ability to provide an image visualization of the energy of a signal,
making it easier to compare two signals and identify abnormalities or anomalies in
the faulty signal. Based on these observations, suitable features can be designed
either using image-processing techniques or simply exploiting the energy values
directly. For example, in a study of a helicopter ’s planetary gear system [Saxena &
Vachtsevanos, 2005], it was observed that the energy distribution among the five
planets became asymmetric as a fault (crack) appeared on the gear plate. This
observation led to a feature based on the increasing variance among energy values
associated with the five planets as time evolves. Similarly, other features can be
designed that characterize a visible property in a numerical form.
Singularity-Based Features: Singularities can be discerned from wavelet phase
maps and can be used as features for detecting discontinuities and impulses in a
signal. Singularity exponents, extracted from the envelope of vibration signals,
have been used to diagnose breakers’ faults [Yang & Liao, 2001]. Other
applications reported in the technical literature relate to detection of shaft center
orbits in rotating mechanical systems [Peng et al., 2002].
Some of the most interesting applications of wavelet-based features include fault
detection in gearboxes [Chen & Wang, 2002; Hambaba & Huff, 2000; Yen & Lin,
1999, 2000; Zheng et al., 2002], Fault detection in rolling element bearings
[Altmann & Mathew, 2001; Shibata et al., 2000], and fault diagnosis in analog
circuits[Aminian, 2001], among others.
Wavelet energy-based features often cannot detect early faults because slight
changes in the signal result in small energy changes. Coefficient-based features are
more suitable for early fault detection. Similarly, singularity- based methods are
not very robust to noise in the signal, and denoising must be carried out before
calculating any such features [Vachtsevanos et al., 2006].

101
 Chapter 5: Signal Preprocessing and Feature Extraction

5.4 Feature Selection

There might be too many features extracted from the signals and collected from
sensors which make extraction of useful and understandable information from
these features become difficult. Therefore the dimensionality of the features needs
to be reduced. Feature selection is primarily performed to select relevant and
informative features which can reduce the dimensionality of features effectively. It
can have the other motivations, including [Guyon & Elisseef, 2006]:
1) General data reduction, to limit storage requirements and increase
algorithm speed;
2) Feature set reduction, to save resources in the next round of data collection
or during utilization;
3) Performance improvement, to gain in predictive accuracy;
4) Data understanding, to gain knowledge about the process that generated
the data or simply visualize the data
Many data mining algorithms can be used to carry out feature selection: neural
network ensemble (NNE) [Hansen & Salamon, 1990], neural network (NN) [Liu,
2001; Siegelmann & Sontag, 1994], boosting regression tree (BRT) [Friedman,
2001, 2002; Smola & Scholkopf, 2003], support vector machine (SVM) [Schölkopf
et al., 1999; Steinwart & Christmann, 2008], random forest with regression (RF)
[Breiman, 2001], standard classification and regression tree (CART) [Speybroeck,
2012], k nearest neighbour neural network (kNN) [Shakhnarovich et al., 2005],
wrapper approach integrated with the genetic or the best-first search algorithm
[Espinosa et al., 2005; Tan et al., 2006] and principal component analysis (PCA)
[Jolliffe, 2002]. All these algorithms are widely used for feature selection. Zhang
and Kusiak applied all these algorithm for parameter selection in wind turbine
condition monitoring and compared these algorithm [Kusiak & Verma, 2011;
Kusiak & Zhang, 2010; Zhang & Kusiak, 2012].
PCA is an unsupervised learning approach for dimensionality reduction that uses
correlation coefficients of the parameters to combine and transform them into a
reduced dimensional space [Miranda et al., 2008]. The concept of Principal
Component Analysis (PCA) was invented in 1901 by Karl Pearson [Pearson, 1901].
It is a mathematical procedure that uses an orthogonal transform to convert a set of
observations of possibly correlated variables into a set of values of uncorrelated
variables called principal components. This transform is defined in such a way that
the first principal component has as high a variance as possible, which means
accounting for as much of the variability in the data as possible, and each
succeeding component in turn has the highest variance possible under the
constraint that it be uncorrelated with the preceding components. It can reduce data
dimension and eliminate multi-collinearity. Currently, PCA mostly used to reduce
the dimension while maintain the main information in data mining analysis and
making models. This section mainly introduces the principle of PCA.
PCA computes a new set of uncorrelated multivariate (vector) samples by a
transform of coordinate rotation from original correlated multivariate samples. A

102
 Chapter 5: Signal Preprocessing and Feature Extraction

matrix composed by n rows which means n samples are collected and m columns
which represent the number of features are expressed as bellowing:
ª x11 } x1m º
« # % # »
X « » (5.36)
«¬ xn1 " xnm »¼

PCA can obtain a new set of vector according to the following steps:
1) Calculate the correlation coefficient matrix
The correlation coefficient matrix is calculated according to the following equation:
(n  1) ˜ Cov(i, j )
R Cor (i, j )
n n
r 
ij mu n (5.37)
¦  x (k )  P  ¦  x (k )  P 
2 2
i i j j
k 1 k 1

where n is the number of samples. The dimension of the correlation matrix R is

m u n . Cov(i, j ) means the covariance which matrix ism u n can be expressed as:

1
Cov (i, j ) ( xi  Pi )( x j  P j ) i, j 1, 2,..., m (5.38)
( n  1)

where P i and P j are the averages of the i th and jth rows of matrix X respectively.

2) Calculate the eigenvectors and eigenvalues of the matrix R

The m eigenvalues Oi which have the constraint as O1 t O2 t " t Om and their
responding eigenvectors Vi are calculated from correlation matrix. Oi and Vi
satisfy the following equation:
AVi Oi ˜ Vi i 1, 2," , m (5.39)
where A is a m u n covariance matrix or correlation matrix and the vector Vi can
be expressed as Vi [V1i , V2i ," , Vmi ] .
3) Generates the new samples
A new set of uncorrelated multivariate (vector) samples are computed according to
the following equation:
X new VT ˜ X (5.40)
where X new is the new uncorrelated multivariate (vector) sample, and X is the
original correlated multivariate (vector) samples. Both of them are n u m matrices
whose row vectors represent a single channel sample. V is eigenvectors matrix
which is also called the weight matrix. Each column of V is one principle
component. X new is the principal component scores. Each row of X new is the scores
for one principal component. Each Oi is variance of the scores for one principal
component. Most of time, only first several components in X new are selected as
principal components according to the variance threshold. The case study of how to

103
 Chapter 5: Signal Preprocessing and Feature Extraction

apply PCA to reduce dimensionality of features will be presented in Chapter 6 with

the case study of fault diagnosis.

5.5 Summary

This Chapter introduced the techniques of signal preprocessing, feature extraction

and feature selection. Signal preprocessing and feature extraction are mainly for
time series such as vibration signals and electrical signals. The features can be
extracted in time domain, frequency domain and time-frequency domain in which
the features extracted based on wavelet transform has special advantages and
becomes very popular. The extracted features might be too many to manage for
condition monitoring because there are too many sensors and many features can be
extracted from one signal. Therefore, dimensionality reduction and feature
selection become important for ease of management for these features but not
reduce the useful information. All these processes are the preparation for fault
diagnosis and prognosis which are crucial parts of condition monitoring.

104
 Chapter 6: Fault Diagnosis based on Data Mining Techniques

6 Fault Diagnosis based on Data Mining Techniques

6.1 Introduction

Fault diagnosis has become the subject of numerous investigations over the past
two decades. Researchers in many disciplines, such as medicine, engineering, the
sciences, business, and finance, been developing methodologies to detect fault
(failure) or anomaly conditions, pinpoint or isolate which component or object in a
system or process is faulty, and decide on the potential impact of a failing or failed
component on the health of the system [Vachtsevanos et al., 2006]. Fault
diagnostic algorithms must have the ability to detect system performance,
degradation levels, and faults (failures) based on physical property changes through
detectable phenomena. Referring the fault diagnosis and condition monitoring, the
following concepts need to be defined and distinguished [Vachtsevanos et al.,
2006]:
x Fault diagnosis. Detecting, isolating, and identifying an impending or
incipient failure condition—the affected component (subsystem, system) is
still operational even though at a degraded mode.
x Failure diagnosis. Detecting, isolating, and identifying a component
(subsystem, system) that has ceased to operate.
x Fault (failure) detection. An abnormal operating condition is detected and
reported.
x Fault (failure) isolation. Determining which component (subsystem, sys-
tem) is failing or has failed.
x Fault (failure) identification. Estimating the nature and extent of the fault
(failure).
Therefore, the aim of fault diagnosis is to detect abnormal condition of machine
before the failure happens, and also identify which component of the machine will
become failure. To evaluate the techniques for fault diagnosis of a condition
monitoring system, several qualification factors can be used [Vachtsevanos et al.,
2006]:
x Isolability. A measure of the model’s ability to distinguish between certain
specific failure modes. Enabling technologies include incidence matrices
involving both deterministic (zero-threshold) and statistical (high-threshold)
isolability.
x Sensitivity. A qualitative measure characteristic of the size of failures. This
factor depends on the size of the respective elements in the system’s
matrices, noise properties, and the time to failure. Filtering typically is
used to improve sensitivity, but it is rather difficult to construct a straight-
forward framework.
x Robustness. This factor refers to the model’s ability to isolate a failure in
the presence of modeling errors. Improvements in robustness rely on
algebraic cancelation that desensitizes residuals according to certain
modeling errors.

105
 Chapter 6: Fault Diagnosis based on Data Mining Techniques

There are many techniques can be used for fault diagnosis. The development of
model-based fault diagnosis began in the early of the 1970s [Dirilten, 1972; Hayes,
1971]. This method of fault detection in dynamic systems has been receiving more
and more attention over the last two decades [Schubert et al., 2011; Soman et al.,
2012; Van den Kerkhof et al., 2012]. It has much to offer in addressing system-
based fault diagnosis issues for complex systems [De Kleer & Williams, 1987;
Isermann, 2005]. It is used to detect any discrepancy between the system outputs
and model outputs. It is assumed that this discrepancy signal is related to a fault.
This method is perfect when the mathematical model or physical model is accurate
and the system outputs are no noise. However, the same difference signal can
respond to model plant mismatches or noise in real measurements, which are
erroneously detected as a fault. What’s more, sometimes, it is impossible to model
nonlinear systems by analytical equations [Mendonqa, 2006]. Therefore, the
model-based fault diagnosis techniques are not very good for some cases such as
non-linear system which mathematical model is not available.
Case-based Reasoning (CBR) [Aamodt & Plaza, 1994; Reisbec & Schank, 1989]
offers a reasoning paradigm that is similar to the way people routinely solve
problems which is another method can be used for fault diagnosis. CBR began to
be applied in fault diagnosis in 1990s [Grant et al., 1996; Patterson & Hughes,
1997], and become very popular afterwards [Fu et al., 2011; Tsai, 2009]. The
cyclic process of CBR can be described as following. When a new problem
happens, one or more similar cases are retrieved from the case base. A solution
suggested by the matching cases then is reused and tested for success. Unless the
retrieved case is a close match, the solution probably will have to be revised,
producing a new case that can be retained. Currently, this cycle rarely occurs
without human intervention and most CBR systems are used mainly as case
retrieval and reuse systems [Watson & Marir, 2009]. The CBR designer is faced
with two major challenges: coding of cases to be stored into the case library or case
base and adaptation, that is, how to reason about new cases so as to maximize the
chances of success while minimizing the uncertainty about the outcomes or actions.
Additional issues may relate to the types of information to be coded in a case, the
type of database to be used, and questions relating to the programming language to
be adopted [Vachtsevanos et al., 2006].
It is obviously that the Model-based fault diagnosis techniques can detect and
identify any faults even for unanticipated ones. But these methods need accurate
mathematical model or physical model which is usually not available for complex
machines. Therefore, data-driven methods could be better solution for fault
diagnosis when the model is unavailable and the CBR does not work well.
In contrast to model-based approaches, data-driven fault diagnostic techniques rely
primarily on process and data which are from sensors specifically designed to
respond to fault signals, to model a relationship between fault features or fault
characteristic indicators and fault classes. Such “models’’ may be cast as expert
systems or artificial neural networks or a combination of these computational
intelligence tools. They require a sufficient database (both baseline and fault
conditions) to train and validate such diagnostic algorithms before their final online
implementation. They lack the insight that model-based techniques provide

106
 Chapter 6: Fault Diagnosis based on Data Mining Techniques

regarding the physics of failure mechanisms, but they do not require accurate
dynamic models of the physical system under study. They respond only to
anticipate fault conditions that have been identified and prioritized in advance in
terms of their severity and frequency of occurrence, whereas model-based methods
may be deployed to detect even unanticipated faults because they rely on a
discrepancy or residual between the actual system and model outputs
[Vachtsevanos et al., 2006]. In the past few years, many Computational
Intelligence (CI) techniques have been applied as tools for fault diagnosis [Sun et
al., 2012; Wang, 1996]. This Chapter mainly introduces data mining techniques
especially of CI techniques application in fault diagnosis. Some case studies will be
used to show how these techniques work in fault diagnosis.

6.2 Fault Diagnosis based on SBP

The pattern classification theory has become a key factor in fault diagnosis. Some
classification methods for equipment performance monitoring use the relationship
between the type of fault and a set of patterns which is extract from the collected
signals without establishing explicit models. Currently, ANN is one of the most
popular methods in this domain. The principle of ANN has been introduced in
Section 3.2 which included Back-propagation (BP), Self-organization Mapping
(SOM). The application of artificial neural network models lies in the fact that they
can be used to infer a function from observations. This is particularly useful in
applications where the complexity of the data or task makes the design of such a
function by hand impractical. This attribution is very nontrivial in diagnostic
problems. BP neural network is a main type of ANN used to solve fault diagnosis
and prognosis problems.
ANN can deal with complex non-linear problem without sophisticated and
specialized knowledge of the real systems. It is an effective classification
techniques and low operational response times needed after training. The
relationship between the condition of component and the features is not linear but
non-linear. BP neural network does not need to know the exact form of analytical
function on which the model should be built. This means neither the functional
type nor the number and position of the parameters in the model-function need to
know. It can deal with multi-input, multi-output, quantitative or qualitative,
complex system with very good abilities of data fusion, self-adaptation and parallel
processing. Therefore, it is very suitable to select as a method of fault diagnosis.
Fig. 6.1 shows the procedure of fault diagnosis based on BP network. There are
mainly three phases of this method. The first phase is training phase to establish an
ANN model for a specific type of fault. The training data could be history data or
collected data from sensors. The collected raw signals, such as vibration signals
and acoustics signals are very hard to be used to train ANN model, and thus need
extract features from these signals. The signals of vibration and acoustics may
contain noise electrically or mechanically, thus the signals need to be processed to
filter out the noise, improve signal-to-noise ratio and amplify the weak signals.
Then the extracted features can be used to training ANN to establish the model of

107
 Chapter 6: Fault Diagnosis based on Data Mining Techniques

the fault. Once the ANN model is established, it can be used to judge if the
machine has fault and identify which component will be failure. This phase called
test phase. The data here used to test ANN model must be the same kind of features
as the training data. Thus, the techniques to be used for signal processing and
feature extraction must be same as that of the training data. The last phase is
maintenance decision making based on the test results of ANN model. This phase
will be complete in Chapter 8.

Fig. 6.1 Procedure of Fault Diagnosis BP Network

6.3 Fault Diagnosis based on SOM

Unsupervised learning [Jain et al., 1999; Oja, 2002] is another method of data
classification and clustering in addition to the supervised methods (for example BP
network) in the field of data analysis. Supervised methods mostly deal with
training classifier for known symptoms, while unsupervised learning (clustering)
provides exploratory techniques for finding hidden patterns in the data. With huge
volumes of data being generated from different systems every day, what makes a
system intelligent is its ability to analyze the data for efficient decision-making
based on known or new cluster discovery. Unsupervised data clustering is an
intelligent tool for delving deep into the unknown and unexplored data. It is a tool
that brings out the hidden patterns and association between different variables in a
multivariate dataset. When the knowledge of the data is not well known and

108
 Chapter 6: Fault Diagnosis based on Data Mining Techniques

explored, the unsupervised learning method can be used to analyze the data. That
means in the field of fault diagnosis, it can be used to understand the data and
cluster the fault hidden in database.
Self-Organizing Mapping (SOM) is a competitive learning network, it uses self-
learning mode of non-supervision and non-direction, and its algorithm is simple
with function of sidewise association [Brando et al., 2007]. It is one of the most
popular unsupervised learning algorithms which can be used to explore useful
information from not well-known data, and thus can be applied in fault diagnosis
when the knowledge of the history data is not known. The principle of SOM is
introduced in Section 3.2.2. Fig. 6.2 shows the procedure of applying SOM in fault
diagnosis which mainly has three phases: training phase, test phase and decision
making phase. The whole procedure of SOM application in fault diagnosis is
similar with that of SBP application. In the training phase, the sensors are used to
collect the data from the monitored mechanical equipment. The data could be the
signals such as vibration and acoustics or the time series such as temperature. The
former signals need to be processed in order to extract useful features while the
later data can be used as features. The features, then, can be used to train SOM for
establish classifier model of different type of faults. Once the classifier is
established, the test phase can be done. The features used to test the classifier must
be the same type with the training data. The finally phase is maintenance decision
making based on the test results of SOM classifier.

6.4 Fault Diagnosis based on Semi-supervised Learning

Fault diagnosis for mechanical equipment is the essence of pattern recognition

problem over the condition monitoring data, in the process of which the balanced
fault and fault-free data and the features definition are the basis for data-driven
diagnosis model such as BP model. However, the collection of fault data is very
difficult because of its expensive costs and stochastic causes for offline system.
Mostly, the fault data with label (type of fault) is collected by test rig in the lab.
However, data-driven model such as BP network, cannot inherently support the
transplant model of fault diagnosis if the test rig is not enough similar with real
system. Typically for a long time running machine, there are lots of unlabeled
samples of condition monitoring data which may contain valuable information of
normal or abnormal conditions. Traditional supervised classifier cannot explore
these data, but it is very improvident to just throw them [Yuan, 2012].
Conventional fault diagnosis methods using supervised learning are good at solving
the problems of condition monitoring (CM) data with labels, but not well at
utilizing unlabelled CM data. Semi-supervised learning algorithm can be
implemented in fault identification by using labelled data and unlabelled data
collected from sensors. Manifold Regularization (MR) is one of the most popular
semi-supervised algorithms which principle is introduced in Section 3.3. MR has
the capacity of learning intrinsic geometric structure of complexity nonlinearity
fault samples and exploiting the intrinsic geometric distribution property embedded

109
 Chapter 6: Fault Diagnosis based on Data Mining Techniques

in the high-dimensional fault patterns. Thus, the well-trained model can be utilized
to further conditions based monitoring as well as fault diagnosis and prognostics.

Fig. 6.2 Procedure of SOM in Fault Diagnosis

To show the advantages of semi-supervised learning with additional unlabelled

dataset, a toy example called two-moon problem is presented.
Fig. 6.3 shows the solution of two-moon problem without unlabelled dataset. In
this figure, the number of labelled training dataset is 50 and the number of test
dataset is 200. From the figure, there are some dataset are misclassified. Fig. 6.4
shows the solution of two-moon problem with unlabelled dataset. The difference
between
Fig. 6.3 and Fig. 6.4 is that the latter figure utilizes 500 unlabelled dataset with the
50 labelled dataset to train the model. Comparing the two figures, Fig. 6.4 is much
better for two-moon problem which almost no test dataset is misclassified.

110
 Chapter 6: Fault Diagnosis based on Data Mining Techniques

The fault diagnosis process of semi-supervised learning is shown in Fig. 6.5. The
manifold regularization based on semi-supervised manifold learning for fault
diagnosis system can be described as follows:
1) Building up general condition monitoring system to collect the labelled and
unlabelled data from both local monitoring machines and the test rig;
2) Implementing feature extraction and feature selection from the labelled and
unlabelled examples according to the criteria which determines the features set
that represent the geometric structure well;
3) Constructing a data adjacency graph with labelled and unlabelled nodes using
graph kernel, which describes an intrinsic manifold, and regulating
classi¿cation decision boundary with manifold regularization algorithm, and
then classifying the online patterns in the features space with classi¿ed labels;
4) Obtaining diagnosis information by classi¿cation of the results, then
determining the failure causes, and putting the corresponding decision or
control measures back to local condition monitoring system.

1.5

0.5

-0.5

-1
-1.5 -1 -0.5 0 0.5 1 1.5 2 2.5

Fig. 6.3 Solution of Two-moon Problem without Unlabelled Dataset

111
 Chapter 6: Fault Diagnosis based on Data Mining Techniques

1.5

0.5

-0.5

-1
-1.5 -1 -0.5 0 0.5 1 1.5 2 2.5 3

Fig. 6.4 Solution of Two-moon Problem with Unlabelled Dataset

Fig. 6.5 Procedure of Semi-supervised Learning in Fault Diagnosis

112
 Chapter 6: Fault Diagnosis based on Data Mining Techniques

6.5 Fault Diagnosis based on Association Rules

Association rules mining is a kind of data mining techniques which can discover
significant association rules between items in database [Agrawal et al., 1993]. The
basic concept and process of association rules are introduced in Section 3.4. This
part will propose an Association Rule-based Fault Diagnosis which structure is
shown in Fig. 6.6.

Fig. 6.6 The Structure of Association Rule-based Fault Diagnosis

Whatever a machines, cars or Robots, after long time running, their performance
may become degradation or failure. Some suitable kinds of sensors should be
selected to monitor their conditions. The data should be pre-processed before
features extraction because the raw data from sensors may contain noise. After
extracting the features, all the data are stored in a database called “Raw Training
Database” which can be used to mine the association rules. For each kind of fault,
several rules can be mined from the training data. Then, select and combine all the
rules together as the whole association rules which can classify the fault or judge
the condition of monitored equipment. Finally, the features extracted from pre-
processed real time data can be used to diagnose the fault using the association
rules generated above. According to result from association rules, the maintenance
or control decision can be made correctly and efficiently.

6.6 Case Study 1: Fault Diagnosis Integration of WPD, PCA and BP

Network

To demonstrate how the BP network works in fault diagnosis for mechanical

machines, a lab setup is established in Knowledge Discovery Lab (KDL) in NTNU.
The first case will show fault diagnosis integrating WPD, PCA and BP network.
This case study is retrieved from [Zhang et al., 2013].

113
 Chapter 6: Fault Diagnosis based on Data Mining Techniques

6.6.1 Experimental Setup

Fig. 6.7 shows the hardware of the experimental setup which includes a blower,
three vibration sensors, power supply for sensors, connector, DAQ card and a
computer. In this setup, the blower is selected as our monitoring object and a kind
of vibration sensors (Kistler: Type 8702B100) are chosen to collect the signals
from the blower. Three sensors are mounted on the blower in three directions
which can collect the vibration signals in different directions (Fig. 6.8). The signals
are collected from the sensors and processed using some processing method like
filter, de-noising and compression. Then the features are extracted in wavelet
domain which can be used to train and query BP network. After training, the
system can judge the real states of monitored components using real time signals.

Fig. 6.7 Hardware of Experimental Setup

Fig. 6.8 Sensors Setup on Blower

6.6.2 Experimental Procedure

In the present study, four different degradation’s levels of unbalance are simulated
using three different parts (Fig. 6.9) which are mounted in the axis end of the
blower. The unbalance degradation (condition) contains 0, 0.3, 0.7 and 1 which
represents the performance states from perfect to absolutely failure (unbalance). In
the first case, power on the blower, collect and store signals with sample rate 1024
per second from the sensors without mounting any simulation part. Next, power off
the blower and mount first part in the axis end and then, power on the blower,

114
 Chapter 6: Fault Diagnosis based on Data Mining Techniques

collect and store the signals from sensors. Repeat this process until collect all the
degrading signals simulated by simulation parts. Fig. 6.10 shows the signals of the
second sensor from perfect state to absolutely failure.

Fig. 6.9 Parts for Simulation Degradations

Condition 0 Condition 0.3

1 1

0.5 0.5
Amplitude

Amplitude

0 0

-0.5 -0.5

-1 -1
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
Time (s) Time (s)
Condition 0.7 Condition 1
1 4

0.5 2
Amplitude

Amplitude

0 0

-0.5 -2

-1 -4
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
Time (s) Time (s)

Fig. 6.10 Raw Signals with Different Degradations

6.6.3 Features Extraction in Wavelet Domain

Wavelet packet method [Li et al., 2003] which is a generalization of wavelet
decomposition offers a richer range of possibilities for signal analysis. Contrary to
WT, the wavelet packets contain a complete set of decompositions and details at
every level and hence providing a higher resolution in the high frequency region,
i.e., the wavelet detail component at each level is further decomposed to obtain its
approximation and detail components. The principle of Wavelet Packet
Decomposition (WPD) was introduced in Section 5.3.3. In this experiment, the
structure of Wavelet Packet Decomposition (WPD) algorithm broke up to 4
resolution levels is shown in Fig. 6.11 In the figure, the node (4, 0) presents the
symbol for a subspace that stands for the 4th resolution and the 0th subspace. For

115
 Chapter 6: Fault Diagnosis based on Data Mining Techniques

this case, each node present the frequency bandwidth 64 Hz which means a node (4,
0) presents the signal character of the bandwidth between 0 Hz and 64 Hz.

Fig. 6.11 Tree structures of wavelet packet transform (4 levels)

For each signal, wavelet packet was applied up to the fourth level, thus giving 16
signal coefficients. The wavelet packet coefficients (Eq.(5.30)) and their
corresponding standard deviations for one signal are shown in Fig. 6.12. In the end,
the Standard Deviation of Wavelet Packet Coefficients (SDWPC) of processed
signals is selected as feature vector which is used to train ANN after PCA analysis.

1.57504 0.023591
5.45224 0.032451
0.630946 0.054542
1.9242 0.033564
0.443467 0.316338
0.395521 0.194326
0.302271 0.063202
0.15776 0.071707

Fig. 6.12 Wavelet Packet Coefficients (WPC) and Their Relevant Standard Deviation

There are three vibration sensors mounted on the blower and for signal from each
sensors, 16 parameters are extracted and thus overall 48 features for each time
signals. Therefore, Principal Component Analysis (PCA) is employed to reduce the
dimension of the features.

6.6.4 Principal Component Analysis (PCA)

PCA is a good option to reduce dimension of the features and its principle was
introduced in Section 3.5. In this experiment, 200 samples for each condition are
collected as training data and are analyzed by PCA. There are 48 variables
(SDWPC) in each sample. Now the original sample matrix’s dimension is 800 u 48 .
Then, these data are analyzed by PCA. The variance for each component is shown
in Table 6.1 (only first 8 values are shown) and the first four principal components
were displayed in Fig. 6.13. If the value of threshold is set toɂ ൌ ͳ, only first
principal component was selected as feature to train ANN. If the value of threshold
is set toɂ ൌ ͲǤͷ, only first two principal components were selected as features to
train ANN.

116
 Chapter 6: Fault Diagnosis based on Data Mining Techniques

Table 6.1 Variance for each component

Component No. 1 2 3 4 5 6 7 8 …

Variance 5625.841 0.681 0.424 0.133 0.053 0.005 0.002 0.001 …

160

140
Values of each component in each sample

120

100

80

60

40

20

-20
0 100 200 300 400 500 600 700 800
Samples

Fig. 6.13 The first four Principal Components

6.6.5 Fault Diagnosis using BP Network

The PCA new features from SDWPC of vibration signals are used to estimate the
fault status of components and machines. The input nodes of BP neural network
come from the test signal sensors. BP neural network made up of one input layer,
one output layer and one hidden layers of nodes. And it has been proved that such
three layers’ BP neural network model can approach any continuous functions at
any precision. The values of output are from 0 to 1 which represent from perfect
condition to complete failure of specific kinds of fault.
For convenience of handling the signal collection, signal processing and interface
things, the Labview are selected as program software in this case study. However,
the capability of mathematical calculation of Labview is not as good as Matlab.
Therefore, both kinds of software are combined. The procedure of fault diagnosis
and prognosis integrating BP Network, PCA and WPC is shown in Fig. 6.14 which
is modified from Fig. 6.1. The historic data is collected and processed which are
fist two steps. Then, the features in wavelet domain (SDWPC) are extracted from
the processed signals. These features are analyzed by the PCA which can generate
new features called principal component which can used to train ANN. After
training, the signals in real time are collected and used to query the BP network,
and then the condition of the monitored components can be obtained.

117
 Chapter 6: Fault Diagnosis based on Data Mining Techniques

Fig. 6.14 Procedure of Fault Diagnosis Integrating BP Network, PCA and WPC

6.6.6 Results and Discussion

In this case study, four conditions are defined for the monitored component which
are 0, 0.3, 0.7 and 1. They represent from perfect performance (condition 0) to
completely failure (condition 1) discretely. For each condition, 200 training signals
are collected and processed. The new feature vectors are generated using PCA
from SDWPC. These new features are put into BP network for training. Finally,
test signals are collected and processed like the training data. In this experiment,
for each condition, 20 samples are collected which used to test trained BP network
for verification.
For each testing data, the output of BP network and the nominal values which can
be called “error from nominal value” (average value of testing data for each
condition) are compared. The values of these errors are shown in Fig. 6.15-Fig.
6.18. There are two curve-lines for each figure. One represents only using the
features of SDWPC as inputs while the other represents using the new features
generated by PCA from the features of SDWPC as inputs to BP network.

118
 Chapter 6: Fault Diagnosis based on Data Mining Techniques

0.05

0.04
error from nominal value

0.03

0.02

0.01

0
0 20 40 60 80 100 120 140 160 180 200
No. of Training data for each degradation

Fig. 6.15 Errors of Condition 0

0.25

0.2
error from nominal value

0.15

0.1

0.05

0
0 20 40 60 80 100 120 140 160 180 200
No. of Training data for each degradation

Fig. 6.16 Errors of Condition 0.3

0.25

0.2
error from nominal value

0.15

0.1

0.05

0
0 20 40 60 80 100 120 140 160 180 200
No. of Training data for each degradation

Fig. 6.17 Errors of Condition 0.7

119
 Chapter 6: Fault Diagnosis based on Data Mining Techniques

-3
x 10
8

0
0 20 40 60 80 100 120 140 160 180 200

Fig. 6.18 Errors of Condition 1

All these four figures show the differences between the predicted values and
nominal values of four different conditions using the features of SDWPC and new
features generated by PCA from SDWPC. Fig. 6.15 shows the result of condition
0. The error is much smaller of the result using the new features generated by PCA
from SDWPC compared to using the features of SDWPC as inputs of ANN. Fig.
6.16 and Fig. 6.18 show the results of condition 0.3 and condition 1 respectively.
When the number of training sets is very small, the results using new features
generated by PCA from SDWPC are much better than using features of SDWPC in
these two figures. However, with the number of the training data increasing, the
results of using both features are almost the same in these two figures and both of
them are correct and precise. Fig. 6.17 shows the result of condition 0.7. In this
figure, in both kinds of features, the performance is very effective and corrective
whatever the number of training data is, but the result of using the new features
generated by PCA from SDWPC is much better than using features of SDWPC.
We can see from Fig. 6.16 and Fig. 6.17, when the condition is neither perfect nor
completely failure, the result of using SDWPC is not believable if the number of
training data is very small because the ‘error from nominal value’ is large. But it is
still believable of using new features generated by PCA from SDWPC to training
and testing ANN in these conditions. We can see from the four figures, the
precision is better of using new features generated by PCA from SDWPC than
using features of SDWPC in any condition and in any number of training data.

6.7 Case Study 2: Fault Diagnosis Integration of WPD, FFT and BP

Network

This case study is retrieved from [Zhang et al., 2012]. The experimental setup and
Experimental Procedure are the same as Section 6.6.1 Experimental Setup and
Section 6.6.2 Experimental Procedure. However, the data analysis and diagnostic
algorithm are different.

120
 Chapter 6: Fault Diagnosis based on Data Mining Techniques

6.7.1 Feature Extraction

The feature extraction algorithm is combining WPD and FFT. For WPD, not like
previous case, the vibration signals are decomposed to 3 levels and for each level,
the approximation part is not decomposed in order to reduce the dimension of the
parameters without omitting much information. The structure of wavelet packet
decomposition is shown in Fig. 6.19.

Fig. 6.19 3-layer Structure of Wavelet Packet Decomposition

For this case, the signal maximum frequency is 512 Hz, and thus D1, D2, D3 and
A3 represent the frequency 256~512 Hz, 128~256 Hz, 64~128 Hz and 0~64 Hz
respectively in Fig. 6.19. In this experiment, only these four parts are analyzed to
judge the degradation of the performance. The decomposed signals by WPD from
the different degrading signals are shown in Fig. 6.20-Fig. 6.23.
D1 D2
0.4 0.4

0.2 0.2
Amplitude

Amplitude

0 0

-0.2 -0.2

-0.4 -0.4
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
Time (s) Time (s)
D3 A3
0.3 0.2

0.2
0.1
Amplitude

Amplitude

0.1
0
0
-0.1
-0.1

-0.2 -0.2
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
Time (s) Time (s)

Fig. 6.20 Decomposed Signal of Condition 0

121
 Chapter 6: Fault Diagnosis based on Data Mining Techniques

D1 D2
0.1 0.2

Amplitude 0.05 0.1

Amplitude
0 0

-0.05 -0.1

-0.1 -0.2
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
Time (s) Time (s)
D3 A3
0.2 0.2

0.1 0.1
Amplitude

Amplitude
0 0

-0.1 -0.1

-0.2 -0.2
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
Time (s) Time (s)

Fig. 6.21 Decomposed Signal of Condition 0.3

D1 D2
0.1 0.2

0.05 0.1
Amplitude

Amplitude

0 0

-0.05 -0.1

-0.1 -0.2
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
Time (s) Time (s)
D3 A3
0.2 0.2

0.1 0.1
Amplitude

Amplitude

0 0

-0.1 -0.1

-0.2 -0.2
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
Time (s) Time (s)

Fig. 6.22 Decomposed Signal of Condition 0.7

D1 D2
0.2 0.4

0.1 0.2
Amplitude

Amplitude

0 0

-0.1 -0.2

-0.2 -0.4
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
Time (s) Time (s)
D3 A3
0.4 0.4

0.2 0.2
Amplitude

Amplitude

0 0

-0.2 -0.2

-0.4 -0.4
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
Time (s) Time (s)

Fig. 6.23 Decomposed Signal of Condition 1

122
 Chapter 6: Fault Diagnosis based on Data Mining Techniques

6.7.2 Fast Fourier Transform to WPD Signals

The principle of FFT was introduced in Section 5.3.2. In Section 6.7.1, the original
signal was decomposed as on approximation and details. Then, the decomposed
signals are transformed with FFT which are shown in Fig. 6.24-Fig. 6.27 which
present different conditions from condition 0 to condition 1. From the result of FFT,
some kinds of features can be chosen. In this paper, the peaks for each part are
selected as features to judge the condition of monitored equipment.
FD1 FD2
15 10

8
10
6
|Y(fft)|

|Y(fft)|
4
5
2

0 0
0 200 400 600 800 1000 0 200 400 600 800 1000
Frequency (Hz) Frequency (Hz)
FD3 FA3
15 20

15
10
|Y(fft)|

|Y(fft)|

10
5
5

0 0
0 200 400 600 800 1000 0 200 400 600 800 1000
Frequency (Hz) Frequency (Hz)

Fig. 6.24 FFT for Each Version Signal of Condition 0

FD1 FD2
6 8

6
4
|Y(fft)|

|Y(fft)|

4
2
2

0 0
0 200 400 600 800 1000 0 200 400 600 800 1000
Frequency (Hz) Frequency (Hz)
FD3 FA3
8 40

6 30
|Y(fft)|

|Y(fft)|

4 20

2 10

0 0
0 200 400 600 800 1000 0 200 400 600 800 1000
Frequency (Hz) Frequency (Hz)

Fig. 6.25 FFT for Each Version Signal of Condition 0.3

123
 Chapter 6: Fault Diagnosis based on Data Mining Techniques

FD1 FD2
6 15

|Y(fft)| 4 10

|Y(fft)|
2 5

0 0
0 200 400 600 800 1000 0 200 400 600 800 1000
Frequency (Hz) Frequency (Hz)
FD3 FA3
10 50

8 40

6 30
|Y(fft)|

|Y(fft)|
4 20

2 10

0 0
0 200 400 600 800 1000 0 200 400 600 800 1000
Frequency (Hz) Frequency (Hz)

Fig. 6.26 FFT for Each Version Signal of Condition 0.7

FD1 FD2
8 20

6 15
|Y(fft)|

|Y(fft)|

4 10

2 5

0 0
0 200 400 600 800 1000 0 200 400 600 800 1000
Frequency (Hz) Frequency (Hz)
FD3 FA3
15 10

8
10
6
|Y(fft)|

|Y(fft)|

4
5
2

0 0
0 200 400 600 800 1000 0 200 400 600 800 1000
Frequency (Hz) Frequency (Hz)

Fig. 6.27 FFT for Each Version Signal of Condition 1

6.7.3 Fault Diagnosis Procedure of Integrating WPD, FFT and BP Network

Fig. 6.28 shows the procedure of fault diagnosis integrating WPD, FFT and BP
Network which is modified from Fig. 6.1. The historic data is collected and
processed which are first two steps. Then, the processed signals can be
decomposed by WPD. Each part of decomposed signals can be transformed using
FFT and the peak value for each of them is selected as feature to train BP network.
After training, the signals in real time are collected and used to query the BP
network, and then the condition of the monitored components can be obtained.
Finally, the remaining useful life is evaluated for decision making of maintenance
according to the condition.

124
 Chapter 6: Fault Diagnosis based on Data Mining Techniques

Fig. 6.28 Procedure of Diagnosis Integrating WPD, FFT and BP Network

6.7.4 Experiment and Results

In this case study, four conditions for the monitored component are defined which
include 0, 0.3, 0.7 and 1 which represent from perfect performance to completely
failure discretely. For each condition, 200 training signals were collected and
processed. The training signals are pre-processed firstly and then decomposed by
WPD. For each part of decomposed signal, calculating the peak value in its
frequency domain transformed using FFT which called PFD1, PFD2, PFD3, and
PFA3. In this case, there are three sensors and thus there are 12 parameters are
input to input nodes of BP network and one output value which represents
condition of the monitored component (Called C). A part of training data is shown
in Table 6.2. After training, test data or query data obtaining from real system can
be used to test or query BP network. In this case, 20 sets of test data (Table 6.3) are
used to test BP network.
There is no mathematical method to select the best structure of the BP network, but
the three layers SBP structure was validated its powerful function to build a
complex model. The SBP structure in this experiment is set to three layer 12×20×1
networks. 12 means the number of input parameters (features in this experiment),
20 means the number of the hidden layer nodes and 1 means only one output in this
BP network structure (condition). Its maximum training epoch is set to 5000. For
each condition, 80 training sets are used to train ANN and 20 sets of features are
chosen to test it. Table 6.3 shows the results of the test data. As mentioned before,

125
 Chapter 6: Fault Diagnosis based on Data Mining Techniques

there are 20 sets of test data in which there are 5 sets of them for each condition.
The nominal condition is called NC while the output condition of test is called TC
in this table. From this table, the results are 100% correct in the above parameter
sets. However, the output is not exactly the same as the nominal condition and
there are deviations between them. The precision of the output is discussed next
section.

Table 6.2 Part of Training Data

Sensor 1 Sensor 2 Sensor 3

C
PFD1 PFD2 PFD3 PFA3 PFD1 PFD2 PFD3 PFA3 PFD1 PFD2 PFD3 PFA3
4.20 3.18 3.768 49.05 4.07 3.756 3.26 95.17 4.325 4.323 2.816 101.08 0
4.46 2.965 2.788 20.38 4.54 3.404 3.346 102.0 3.891 4.108 3.248 107.36 0
4.58 4.039 3.874 317.6 6.15 3.603 3.704 4170 4.663 3.55 5.447 1094.9 0.3
3.42 3.802 3.227 314.3 3.46 3.765 3.659 4220 4.261 3.42 4.659 1132.5 0.3
4.87 4.238 5.951 482.2 5.19 4.184 4.617 6975 3.523 3.845 2.723 1889.8 0.7
4.49 3.745 4.178 395.6 4.03 4.412 4.289 6828 4.781 3.705 3.022 1861.4 0.7
6.41 3.007 18.46 1933 4.94 3.053 5.048 2035 5.095 2.919 5.601 6189.9 1
4.54 4.304 18.43 1936 4.72 4.23 4.73 2103 4 4.506 6.062 6391.8 1
… … … … … … … … … … … … …

6.7.5 Discussion
In this section, three issues will be discussed. The first one is how many training
sets should be used in order to achieve enough accurate condition of the machine
from BP network. The second one is attempting to discuss the relationship between
the accuracy and the number of hidden layer nodes. The last issue is convergent
time of the BP network training.
To discuss the first issue, the numbers of training sets for each condition are
changed from 1 to 200. The number of hidden layer nodes is set to 20 and the
number of training epoch is set to 5000. For each testing data, compare the output
of ANN and the nominal value which is called “error from nominal value” which is
average value of testing data for each condition. The values of these errors are
shown in Fig. 6.29. We can see from this figure, the result is believable whatever
the condition of the component is when the number of training data is larger than
20. For condition 0 and condition 1, the result is still believable even if the number
of training data is smaller than 20. It is clear that the result will be believable if
there are only two conditions (0 and 1 or good and fault) even if the number of
training data is very small. But if there are more conditions, the number of training
data should be increased. Therefore, the number of conditions should be considered
in designing of how many training sets are used to trained BP neural networks.

126
 Table 6.3 Test Data and the Results

Sensor 1 Sensor 2 Sensor 3 Results

Deviation
PFD1 PFD2 PFD3 PFA3 PFD1 PFD2 PFD3 PFA3 PFD1 PFD2 PFD3 PFA3 NC TC
3.71 3.382 2.941 37.608 4.636 4.582 3.018 99.027 4.095 3.719 4.749 107.685 0 0.04 0.038
4.755 3.079 3.049 30.693 6.705 4.092 3.187 81.135 3.951 3.525 3.583 105.698 0 0.05 0.046
4.29 4.083 2.418 20.416 4.258 4.069 3.583 85.343 6.563 3.09 4.254 100.095 0 0.04 0.041
4.803 3.506 3.056 39.464 4.561 3.792 2.952 76.756 4.398 4.545 4.052 111.304 0 0.03 0.033
4.114 3.856 2.557 32.6 4.77 4.005 3.258 112.054 4.752 3.517 3.48 116.38 0 0.04 0.038
4.133 3.684 3.162 275.02 4.114 3.736 3.227 4135.45 4.701 3.452 4.835 1199.69 0.3 0.28 0.017
4.433 3.475 3.589 280.2 3.944 3.532 3.121 4174.94 4.215 3.701 3.25 1223.66 0.3 0.28 0.025
5.301 3.352 3.539 280.78 4.89 4.044 4.023 4280.03 5.667 4.422 4.008 1259.53 0.3 0.29 0.014

127
5.346 6.322 4.353 257.69 4.934 3.491 3.361 4175.62 5.982 5.922 3.044 1212.37 0.3 0.28 0.02
3.852 3.516 3.548 303.43 4.874 3.825 3.852 4193.24 3.817 3.952 3.428 1233.34 0.3 0.29 0.011
4.699 3.421 3.31 311.82 4.327 4.911 5.273 7101.38 4.158 3.558 3.183 2098.35 0.7 0.71 0.005
4.087 4.644 3.008 278.09 3.865 3.482 5.644 7211.46 5.392 4.981 3.51 2147.95 0.7 0.64 0.059
3.978 3.719 3.463 286.09 4.321 3.635 5.177 7122.77 4.196 3.682 3.883 2094.52 0.7 0.7 0.002
4.347 2.976 3.434 279.05 5.284 5.405 4.546 7157.75 3.978 4.449 3.333 2126.02 0.7 0.67 0.033
4.44 3.505 3.345 262.11 5.521 3.628 4.63 7080.4 3.991 3.587 4.037 2082.84 0.7 0.67 0.031
3.603 4.235 8.397 910.92 5.633 5.258 9.274 21669.1 3.839 3.875 5.985 6824.31 1 1 0.002
5.451 3.87 6.187 885.07 4.922 6.128 12.67 21416.9 5.687 3.643 8.407 6661.59 1 1 0
Chapter 6: Fault Diagnosis based on Data Mining Techniques

5.957 3.575 8.829 918.7 5.764 5.873 8.867 21244.5 4.995 4.161 4.444 6594.82 1 1 0.002
4.818 3.049 8.352 882.03 6.918 5.931 10.12 20684.6 5.035 3.511 8.835 6461 1 1 0.002
3.745 3.083 8.32 885.89 6.902 5.976 10.9 20605 5.183 3.216 8.642 6455.84 1 1 0.002
 Chapter 6: Fault Diagnosis based on Data Mining Techniques

0.25

Condition 0

0.2 Condition 0.3

Condition 0.7
Errors from nominal value

Condition 1
0.15

0.1

0.05

0
0 20 40 60 80 100 120 140 160 180 200
No. of training data

Fig. 6.29 Errors for Each Condition

Fig. 6.30 Output with Different Number of Hidden Layer Nodes

To discuss the second issue, the number of hidden layer nodes is changed from 5 to
135. The number of training data is set to 80 and the number of maximum training
epoch is set to 5000. For each training process, several test sets for every condition
are used to test the trained SBP networks. The results are shown in Fig. 6.30. From
the figure, with the increasing of the number of hidden layer nodes, the fluctuations
of the output for each condition are small. So the changing of the number of hidden
layer nodes does not affect the accuracy of the output. What’s more, there is no
mathematical method to prove what the number of it is best. Therefore, the number
of hidden layer nodes does not need to be considered much.

128
 Chapter 6: Fault Diagnosis based on Data Mining Techniques

To discuss the last issue, the number of hidden layer nodes is set as 20 and the
training epoch is set as 2000. Fig. 6.31 shows the BP network training time with
the number of training data increasing from 10 to 200. From the figure, training
time is not apparently increasing with the increasing of training data sets.
Therefore, when the BP network is need, we should use as many as possible data
sets to complete the training. Fig. 6.32 shows the training time changes with the
increasing of hidden layer nodes. The number of training data sets is set as 200 and
the training epoch is set as 2000. From this figure, the training time increase
gradually with the increasing of hidden layer nodes. Therefore, when a BP network
need to trained, the number of hidden layer nodes should be considered. However,
from the experience of previous work, the numbers of the hidden layer neurons
depends both on the input layer number and the output layer neuron number but the
numbers can not be too many [Meng & Meng, 2010].

35

30
Training time (s)

25

20

15

0 20 40 60 80 100 120 140 160 180 200

No. of training datasets

Fig. 6.31 BP Network Training Time with the Increasing of Training Data

45

40

35
Training time (s)

30

25

20

15
0 20 40 60 80 100 120 140
No. of hidden layer nodes

Fig. 6.32 BP Network Training Time with the Increasing of Hidden Layer Nodes

129
 Chapter 6: Fault Diagnosis based on Data Mining Techniques

6.8 Case Study 3: Fault Diagnosis based on Self-organizing Map

SOM is a type of Artificial Neural Network (ANN) which is trained by

unsupervised learning to map a high dimensional dataset into low dimensional
space. It is very suitable for classification and clustering. The principle of SOM has
been introduced in Section 3.2.2. This example shows how the SOM works in fault
diagnosis, i.e. fault classification which is retrieved from [Zhang & Wang, 2011].

6.8.1 Experimental Setup

The experimental set-up consists of a centrifugal pump designed for a pressure
increase of 6.6 bars at 90m3/h and at an operating speed of 3000 rpm. The drive
unit is a 3 phase induction motor with an output of 26 kW. The pump rig is rigidly
mounted in a relatively noise-free environment. It is designed to lift and circulate
water. Both the motor and the pump are equipped with ball bearings. Vibration
measurements were taken in axial direction at the free ends of both the motor and
the pump. Measurements were also taken at the vertical and the horizontal
directions on the bearing housing at both the pump and the motor drive and free
ends. Along the vertical direction on the pump casing, another measurement was
taken close to the impeller (Fig. 6.33). The following types of vibration
measurements were carried out on the pump rig:
x High Frequency Domain (HFD) parameter (5-60 kHz).
x Low Frequency (LF) spectrum (0- 400 Hz).
x High Frequency (HF) spectra (0-8 kHz).
For the respective measuring point, the frequency components in Table 6.4 will be
registered.

Fig. 6.33 Vibration Measurement Points

130
 Chapter 6: Fault Diagnosis based on Data Mining Techniques

Table 6.4 Measurement Points and Their Corresponding Vibration Types

Equipment Measurement Points Type of vibration measurement

Free-end axial (1)

Free-end horizontal (2) Low Frequency Spectrum
Drive-end horizontal (4)

High Frequency Domain

Pump Free-end vertical (3)
Low Frequency Spectrum
Free-end vertical (5)
High Frequency Spectrum

High Frequency Domain

Pump Casing (6)
High Frequency Spectrum

Drive-end horizontal (7)

Free-end horizontal (10) Low Frequency Spectrum
Free-end axial (11)
Motor
High Frequency Domain
Drive-end vertical (8)
Low Frequency Spectrum
Free-end vertical (9)
High Frequency Spectrum

6.8.2 Fault Types of Centrifugal Pump System

There are several types of faults in the centrifugal pump system have different
symptoms with different failure. The most important problems to be monitored are
introduced as the following:
Leakage from worn wearing-ring (L)
Inner leakage as a result of worn wearing-ring in a pump will have a result that a
large part of the delivered capacity, which the pump delivers, will go directly back
to the suction side of the pump. The efficiency will be lower, and the pump will no
longer be able to produce the same pressure at a given capacity.
This problem is simulated by exchanging one of the two rings with another which
had a clearance of 1.0 mm instead of the recommended clearance of 0.25 mm.
Clearance is measured as the maximum distance between the inner side of the ring
and the impeller in radial direction.
Unbalance on impeller (U)
The efficiency of the pump is decreased as a result of the unbalance on the
impeller, and this leads to high current consumption at speed and flow. This
symptom includes high and steady once per revolution component (1x) at both

131
 Chapter 6: Fault Diagnosis based on Data Mining Techniques

bearing. Approaching a phase difference of 90 digress between the 1x components

in the vertical and horizontal directions for both bearing, and approaching zero
phase difference between 1x vibration at the pump’s free-end and drive-end
bearings.
This problem is simulated by exchanging the impeller with another which has a
steel weight of 0.114 kg mounted on the suction side, at a radius of 100 mm. the
shape of the weight was designed to give minimum disturbance to the flow around
the impeller.
Unbalance on coupling (N)
The symptoms include the high and steady once per revolution component (1x) at
the bearing on both side of the coupling, approaching a phase difference of 90
degrees between 1x vibration at the pump’s free-end and drive-end bearings.
This problem is simulated by mounted a steel weight of 0.102 kg on the periphery
of the coupling at a radius of 80 mm.
Misalignment between Motor and Pump (M)
Misalignment manifests itself as coupling misalignment, and can therefore cause
deflection forces to be generated in the rotor, friction in seals and casings, and
bearing failure, etc. This can result in high current consumption at high flows and
rpm. Misalignment is characterized by steady 1x and 2x frequency amplitude
components ,approaching a phase difference of 90 degrees between 1x vibration
component at the pump’s free-end and drive-end bearings. It can be accompanied
by large axial vibration, up to about 50% of the radial level.
This problem is simulated by moving the motor in both vertical and horizontal
directions; as a result a combination of parallel and angular misalignment exists. A
laser alignment monitoring instrument was used to verify the amount of
misalignment present.
Bearing Damage (B)
The effect could be measured by high frequency domain parameter and the total
revolution-level for a high frequency.
This problem is simulated by exchanging the ball bearing at the pump’s free-end
with another bearing that has a small cavity on the inner ring.
Cavitation (C)
Cavitation is one of the most frequent occurring problems in centrifugal pump
system which is unfavorable operational state as a result of pressure losses along
the pipe at the suction side at high flow rates. The most important effects of
cavitation are increased hydraulic losses, noise and vibration, and massive wear of
surface.
In experiment, a flow valve is used to regulate the flow on the suction side to a
pressure below the pump’s Net Position Suction Head (NPSH).

132
 Chapter 6: Fault Diagnosis based on Data Mining Techniques

When a problem is going to occur or has occurred, a certain number of symptoms

or parameters will demonstrate themselves in a certain way, i.e., some frequency
component value will change significantly when there is a bearing problem, these
parameters are then called “features” for monitoring or detecting that problem. In
this system, there are 56 parameters can be collected which is shown in Table 6.5.
A part of data used to Train SOM network while the others are used to test SOM
classifier. In this case study, 100 training data sets are used to train SOM network
and 20 data sets are used to test these network.

Table 6.5 Parameters Calculated from Vibration Signals

SCFLOW, SCSPEED, CPFATOTL, CPFA1X, CPFA2X, CPFHTOTL,

CPFH1X, CPFH2X, CPFH36X, CPFHSYNC, CPFHSUBS, CPFHNONS,
CPFRAT1X, CPFVTOTL, CPFV1X, CPFV2X, CPFV36X, CPFVSYNC,
CPFVSUBS, CPFVNONS, CPFVTOTH, CPFVHFD, CPDHTOTL,
CPDH1X, CPDH2X, CPDH36X, CPDHSYNC, CPDHSUBS,
Parameters CPDHNONS, CPDRAT1X, CPDVTOTL, CPDV1X, CPDV2X,
CPDV36X, CPDVSYNC, CPDVSUBS, CPDVNONS, CPDVTOTH,
CPDVHFD, CMDH1X, CMDV1X, CMDRAT1X, CMFATOTL,
CMFA1X, CMFA2X, CPFVHPHA, CPDVHPHA, CPFDPHA,
CMDVHPHA, CPMAPHA, CPCATOTH, CPCASHFD, CPOWER,
CTORQUE, CDELTAP
SCFLOW - normalized flow, SCSPEED - normalized speed, CX…X-
normalized data, X…X - raw data, P - pump, M - Motor, F - free end, D -
drive end, H - horizontal direction, V - vertical direction, A -axial direction,
CAS - pump casing, 1X, 2X, …, 6X - first, second up to sixth fundamental
frequency, PHA - phase angle, TOTL-total low frequency, TOTH - total
Terminology high frequency, SYNC - synchronous frequency, SUBS - sub synchronous
frequency, NONS - non-synchronous frequency, FREQ - frequency, P1 -
inlet pressure, P2 - outlet pressure, PH - phase, DELTAP - difference
between P1 and P2, VOLTAGE - voltage, CURRENT - current, SPEED -
rotational speed, FLOW - flow rate, CPFRAT1 - difference between
CPFV1X and CPFH1X.

6.8.3 Experiment and Results

Fig. 6.34 shows the visualization of SOM for U-matrix and first 5 variables. Fig.
6.34(a) shows the U-matrix which is short for unified distance matrix which means
the Euclidean distance between the SOM node vectors of the neighboring neurons
is depicted in a gray scale image. From this figure, it is easy to see that the maps
are classified as several different groups. Fig. 6.34(a)-(f) displays the distributions
of the variables from Variable1 to Variable5. It is noticeable that the groups in
variables are not the same as ones in U-matrix map and are not the same
themselves. This is because the effects of each variable for different type of faults
are different. For each variable, the map is classified as several groups clearly
according to its values. Combining all the maps together, the general map like U-
matrix could be generated which presents the whole groups of clustering, and here
presents the types of faults.

133
 Chapter 6: Fault Diagnosis based on Data Mining Techniques

Fig. 6.35 shows the results of SOM classification of Centrifugal Pump System. The
labels “B”, “C”, “L”, “M”, “N”, “U” and “M” present the types of faults. In this
map, the neurons with the fault type means the inputs with the same fault are
mapped into this node. It is noticeable that there are probably more than one fault
types locating in the same node which means the input data represent
corresponding fault types. It is also noticeable that the neurons representing the
same fault type may be not in the same area, which means a type of fault may be
caused by different parameters. The numbers located in the neurons mean the
sequence of the test data sets. From the map, the fault type of the test data could be
classified very clearly.
U-matrix Variable1 Variable2
4.55 13.7
9.93

2.34 1.77
4.93

0.128 -10.1 -0.0673

d d
(a) (b) (c)
Variable3 Variable4 Variable5
22.4 18.9 18.6

9.96 1.36 4.61

-2.47 -16.2 -9.39

(d) d (e) d (f) d

Fig. 6.34 Visualization of SOM

4
BCL 5
BCL 8B9 10
BL 3
BL 7B BL BL BL

B M

BCL BL BL CN M

BL BL B B BL M

1
BL BL CU 6 2
BL

BL N

11
L U BL BL CU N

L U

L L M N 16
N 17
N

L L U U U N

L L L L 12
L U U U

13
L L U N

M O O U U

O 19
O O M

M M 18
O O 20
O M 15
M 14
M N N

Fig. 6.35 Classification Result of SOM

134
 Chapter 6: Fault Diagnosis based on Data Mining Techniques

6.9 Summary

This Chapter mainly describes how the data mining techniques work in fault
diagnosis of mechanical machines. These data mining techniques include: BP
network, SOM, Semi-supervised learning and association rules. Some case studies
are used to verify these techniques except semi-supervised learning because there
are no data for case study and a two-moon problem was used to show how it
works. From these case studies, the data mining techniques are suitable to diagnose
the faults of mechanical equipment. The discussion for each case study is presented
in following paragraphs.
Case Study 1 and Case Study 2 described two examples integrating BP network
with other two techniques. The former described case study with the method of
integrating BP network, PCA and WPD while the later described the case study of
methods integrating BP network, WPD and FFT. To verify the correctness and
effectiveness of these two methods, Blower Fault Diagnosis System was
established. These methods demonstrated high effectiveness in diagnosing machine
faults. They can classify the condition of the monitored components.
In former case study, PCA was applied to reduce the input dimension (number of
variables) of BP network without omitting the useful information. BP network
model may become over specified, i.e. more input variables than is strictly
necessary, due to including superfluous variables which are uninformative, weakly
informative, or redundant [May et al., 2011]. In this case, the total volume of the
modeling problem domain increases exponentially with the linearly increasing of
variable dimensionality which is called curse of dimensionality [Bellman, 1961].
This will cause many problems such as: computational burden increasing which is
a significant influence in determining speed of training and training difficulty due
to inclusion of redundant and irrelevant input variables. By reducing the
dimensionality of variables, PCA can solve these problems and improve the
effectiveness of BP network training. Therefore, the method provides a faster, more
effective and more precise solution for fault diagnosis and prognosis. The latter
case study applies FFT after WPD to extract features which does not too many
variables, and thus the PCA is not applied.
In these two cases, the minimum bandwidth 0‫׽‬64 Hz is chosen in WPD because
the fundamental frequency of the vibration signal is 47.5 Hz. In a real system, the
minimum bandwidth of WPD (which means how many levels should be
decomposed) should be selected according to the real fundamental frequency.
There is only one type of fault (unbalance) simulated. In the future, multi-fault
diagnosis should be a research topic. These two methods can also be applied to
decide many other faults such as wear, crack, and fatigue of bearings and gearbox
which faults can be reflected by vibration signals. To apply these methods, the
fundamental frequency has to be known firstly and thereafter the sample rate of
vibration signals, the level of wavelet decomposition, and the structure of BP
network can be determined properly. The degradation information could be very
useful for maintenance decision making, and thus, how to apply this degradation
information in maintenance decision making should be a research issue as well in
the future.

135
 Chapter 6: Fault Diagnosis based on Data Mining Techniques

Case Study 3 described a Self-organizing Map (SOM) classifier applying in fault

isolation for a machine (here is a centrifugal pump), which could be called as
pattern clustering as well. The Self-Organizing mapping describes a mapping from
a higher dimensional input space to a lower dimensional space. In the experiment,
SOM maps 56 dimensional variables into a two dimensional space (15×15). The
result of SOM method is very effective, clear and easy to understand. The results of
the experiment shows SOM is very suitable for solving this kind of problem like
fault isolation, fault classification and pattern recognition.
From this case, the way to continuously monitor the condition of machines,
components, systems and processes have been found. The first stage is determining
the number of neurons of a two-dimensional SOM lattice and the number of
clusters of the conditions and fault type according to the real machines or systems.
The second stage is to train SOM which is mapping the many of variables to the
predefined SOM neurons, that is, finding the groups of lattice which each of them
represent a kind of condition or fault. The third stage is finding the location of the
test data or real time data inside the lattice. Finally, the conditions or faults of
machines, components, systems or processes could be determined according to the
trained lattice and the location of the test data or real time data.
In this case, for each type of fault, only two conditions (normal or failure) are
considered. However, this is not enough and not fitting the real situation. In the
future, more conditions should be considered for each type of fault. Here, only
offline data are used to test trained SOM neurons, but in the future, the real time
data should be used for real time monitoring, control and maintenance. Finally, in
the future, SOM could be applied in these kinds of problems such as pattern
recognition combining with other machine learning method such as Support Vector
machine (SVM) and Supervised Back-propagation (SBP).

136
 Chapter 7: Fault Prognosis based on Artificial Neural Network

7 Fault Prognosis based on Artificial Neural Network

7.1 Introduction

Prognosis is the ability to predict accurately and precisely the Remaining Useful
Life (RUL) of a failing component or subsystem. The task of the prognostic
module is to monitor and track the time evolution (growth) of the fault. In the
industrial and manufacturing arenas, prognosis is interpreted to answer the question:
“what is the RUL of a machine or a component once an impending failure
condition is detected, isolated, and identified?” It is a basis of a Condition-Based
Maintenance (CBM) system and presents major challenges to CBM system
designer primarily because it entails large-grain uncertainty. Long-term prediction
of the fault evolution to the point that may result in a failure requires means to
represent and manage the inherent uncertainty. Moreover, accurate and precise
prognosis demands good probabilistic models of the fault growth and statistically
sufficient samples of failure data to assist in training, validating, and fine-tuning
prognostic algorithms. Fault prognosis has been approached through probabilistic,
artificial intelligence and other methodologies. Specific techniques include fuzzy-
adaptive Kalman predictor [Tian et al., 2011], Autoregressive Model [Xin et al.,
2012], fuzzy-filtered neural networks [Li et al., 2013] and Case-Based Reasoning
[Berenji, 2006]. However, there are still some challenge in this area [Vachtsevanos
et al., 2006]:
x How we can infer the actual crack dimension over time in the absence of
the techniques of measuring creak length directly?
x How do we predict accurately and precisely the temporal progression of
the fault?
x How do we prescribe the uncertainty bounds or confidence limits
associated with the prediction?
x Once we have predicted the time evolution of the fault and prescribed the
initial uncertainty bounds, how do we improve on such performance
metrics as prediction accuracy, confidence, and precision?
The techniques of fault prognosis can be classified into three categories: model-
based, probability-based and data-driven methodologies. The model-based
techniques can predict any fault of the machines or components if the accurate
physical model or mathematical model is available. The advantages of this
technique are very apparent: it can predict any type fault in any component in any
stage of faults of a machine. However, determining a complete dynamic model in
terms of differential equations that relate the inputs and outputs of the system being
considered may be impractical in some instances since the machine becomes more
and more complex and integration. Often, historical data from previous failures for
a given class of machinery can be used to establish probabilistic model [Hu et al.,
2011] based on statistic methods. These methods require less detailed information
than model-based techniques because the information needed for prognosis resides
in various probability density functions (PDFs), not in dynamic differential

137
 Chapter 7: Fault Prognosis based on Artificial Neural Network

equations. Advantages are that the required PDFs can be found from observed
statistical data and that the PDFs are sufficient to predict the quantities of interest
in prognosis. Moreover, these methods also generally give confidence limits about
the results, which are important in giving a feeling for the accuracy and precision
of the predictions. In many instances, one has historical fault/failure data in terms
of time plots of various signals leading up to failure, or statistical data sets. In such
cases, it is very difficult to determine any sort of model for prediction purposes. In
such situations, nonlinear network approximators can be used for prediction of
failure which provides desired outputs directly in term of data using well-
established formal algorithms. This is so-called data-driven technique fault
prognosis. This Chapter will describe the process of fault prognosis based on
neural network.

7.2 Procedure of Fault Prognosis based on Artificial Neural Network

As mentioned before, most of big companies have huge history data which is not
effectively used currently. This kind of data can be used to predict and identify the
fault of machines before the failure happens. Fig. 7.1 shows how the history data
can be used in fault prognosis by ANN. This figure and the following sections just
take the SCADA data of wind turbines as an instance of research objects. These
kinds of history data normally contain the performance parameters such as
temperatures, vibrations, speed and lubrication, etc., and alarm/fault/warning list of
all components of the machines. The first step of fault prognosis is to select right
parameters to be analyzed. For a specific fault/failure, there is normally one or
more performance parameters could be the indicator to determine if the
fault/failure happens. For instance the temperature of bearing can be the indicator
of the bearing defect. It is not too difficult to choose the right indicator for a
specific fault through data analysis or experience. Besides, the related performance
parameters with the indicator should be also selected through data analysis,
experience or some algorithms such as boosting tree algorithm [Kudo &
Matsumoto, 2004] and wrapper with genetic search [Kohavi & John, 1997]. Then
the ANN model in normal condition can be trained in which the selected
performance parameters could be the input while the indicator of the fault could be
the output of the ANN model. The trained ANN is so-called ANN model of normal
behavior. The second step is to establish ANN predictor for fault prognosis. In this
step, the history data with fault will be used. With the ANN normal behavior and
the selected performance parameter values, the theoretical values of the indicator
can be estimated and compared with real values from the history data. Through the
comparison and how early the customer wants to have early warning, close alarm
and emergency stop, the thresholds of these levels can be set. Fig. 7.5 could be an
example of these functions. Finally, the ANN model with these thresholds can be
the fault predictor of a machine.

138
 Chapter 7: Fault Prognosis based on Artificial Neural Network

Mechanical System and/or Electrical System

Existing History Data:

Online Condition Monitoring
Temperature, vibration, speed...
(Online Data)
Fault, Alarm, warning...

Parameter Selection

ANN Training using Data with Fault

Healthy Data
Parameter Values

Real Time
Performance

ANN Model for Normal Performance

Behavior parameters and
Indicator Values

Estimated Indicator Real Indicator Values

Values

Comparison

ANN Fault Predictor and Identifier

Fault Prediction and Identification, Early warning

Fig. 7.1 Procedure of Fault Prognosis

139
 Chapter 7: Fault Prognosis based on Artificial Neural Network

7.3 Fault Prognosis based on Indicator Prediction by ANN for Wind

Turbine Monitoring

Renewable energy sources are playing an important role in the global energy mix,
as a means of reducing the impact of energy production on climate change. Wind
energy is the most developed renewable energy technologies worldwide with more
than 282.48 GW installed capacity at the end of 2012 [GWEC, 2013]. Certain
forecasts indicate that the share of wind in Europe’s energy production will reach
up to 20% in the close future [Krohn et al., 2007]. Today, large wind turbines (2-
6MW) are becoming established as economically viable alternatives to traditional
fossil-fuelled power generation. In some countries, such as Denmark, Germany and
Spain, wind turbines have become a key part of the national power networks [Pinar
Pérez et al., 2013].
Condition monitoring of wind turbines is of increasing importance as the size and
remote locations of wind turbines used nowadays makes the technical availability
of the turbine very crucial. Unexpected faults, especially of large and crucial
components, can lead to excessive downtime and cost because of restricted turbine
accessibility especially for some remote controlled wind farms on mountain and
offshore wind farms. However, even smaller issues and faults of auxiliary
equipment like pumps or fans can also cause expensive turbine downtime due to
the same causes. From an operator’s point of view it is therefore worth increasing
the effort spent to monitor the turbine condition in order to reduce unscheduled
downtime and thus operational costs. The key part of wind turbine monitoring
system is to detect and predict fault (fault diagnosis and prognosis) of turbines as
early as possible so that the maintenance staff can manage and prepare the
maintenance action in advance.
Most wind turbines installed nowadays are integrated with SCDA system which
can monitor the main components. SCADA system typically monitors parameters
such as temperatures of bearings, lubricating oil, windings and vibration levels of
driven train [Becker & Poste, 2006]. This monitored data is collected and stored
via a SCADA system that archives the information in a convenient manner, usually
for all of the turbines in the wind farm. This data quickly accumulates to create
large and unmanageable volumes that can hinder attempts to deduce the health of a
turbine’s components. It would prove beneficial, from the perspective of utility
companies, if the data could be analyzed and interpreted automatically to support
the operators in identifying defects. One main function of SCADA data analysis is
fault detection and predict as early as possible to support the decision of
maintenance action and operation.
Model based methods require a comprehensive physical or mathematical model
which is normally unavailable. Success of data based methods is conditioned by
the significance of historical data and the mathematical method used to detect the
patterns in data. For wind turbine systems where an important amount of data is
stored regularly by SCADA system and process model is not available, the use of
data driven methods is preferred [Nassim, 2011].

140
 Chapter 7: Fault Prognosis based on Artificial Neural Network

This section describes Artificial Neural Network (ANN) that can be used to predict
and identify incipient faults in the main component of a turbine, such as main
bearing, gearbox and blades, through the analysis of this SCADA data. BP network
is one type of ANN which can solve the non-linear problems without sophisticated
and specialized knowledge of the real systems. It is suitable to be applied in fault
detection and predict and the principle of BP network was described in Section
3.2.1.The SCADA data sets are already collected and stored, and therefore, no new
installation of specific sensors or diagnostic equipment is required. The technique
developed normal behavior model by ANN and SCADA data analysis which can
calculate the theoretical value of related parameters and compare to the real
measurement of the same parameters. The parameters mentioned above can be
indicator of abnormal behavior of incipient component failure. In this way, only
interesting information is highlighted to the operator, therefore significantly
reducing the volume of data they are faced with. This section just take the main
shaft rear bearing monitoring as an instance to show how the technique works.

7.3.1 SCADA Dataset Description

An operational wind farm typically generates vast quantities of data which is well
known SCADA data.
x The SCADA data contain information about every aspect of a wind farm,
from power output and wind speed to any errors registered within the
system. Thus by keeping track of both wind speed and power output
parameters, the overall health of the turbine can be supervised.
x SCADA data may be effectively used to “tune” a wind farm, providing
early warning of possible failures and optimizing power output across
many turbines in all conditions
It is common for “condition monitoring” to be applied to a wind farm. However,
this involves the addition of extra instrumentation, involving wind farm down time,
extra cost and potential warranty implications. As distinct from condition
monitoring, performance monitoring using existing instrumentation to analyze
SCADA data of wind turbines is no extra instrumentation, no down time and no
cost. It has the advantage of using data already routinely gathered. By making use
of specially-designed software tools, a great deal of information may be gathered
and analyzed to provide a detailed look at the performance of the wind farm.
Typical parameters recorded by SCADA on wind turbines could be broadly
categorized into following types which could be used in fault detection and
diagnosis activity [Verma & Kusiak, 2012].
x Wind parameters, such as wind speed and wind deviations;
x Performance parameters, such as power output, rotor speed, and blade
pitch angle;
x Vibration parameters, such as tower acceleration and drive train
acceleration; and

141
 Chapter 7: Fault Prognosis based on Artificial Neural Network

x Temperature parameters, such as bearing temperature and gearbox

temperature.
Specifically the SCAD data recorded and used for condition monitoring from wind
turbines are as follows:
x Active power output (10 min max/min/average)
x Anemometer-measured wind speed (10 min max/min/average)
x Turbine speed (10 min max/min/average)
x Nacelle temperature (10 min max/min/average)
x Turbine rear bearing temperature (10 min max/min/average)
x Turbine rear vibration (10 min RMS max/min/average)
x Turbine front bearing temperature (10 min max/min/average)
x Turbine front vibration (10 min RMS max/min/average)
The technique presented utilizes only some types of the data mentioned above. The
parameters listed above are typical of data collected and stored by commercial
wind turbine SCADA systems. This means the approach developed in this section
can be widely applied by wind farm operators.

7.3.2 Modeling of SCADA Parameter Normal Behavior

A parameter of main shaft rear bearing in the SCDA data, i.e. turbine rear bearing
temperature, gives an indication of how hot of the bearing are running, and
therefore offer the possibility to detect rear bearing overheating. The
straightforward threshold check which has already been applied in real wind farm,
could be used to flag up temperature exceeding a certain limit, this might be too
late to avoid significant damage to the main shaft rear bearing. The desired
functionality should take into consideration any relevant aspects of turbine
operation. This approach would allow temperatures to be detected that are too high
in the context of the concurrent level of power generation, leading to a quicker and
more effective identification of abnormal behavior.

7.3.2.1 Parameter Selection

In order to establish the normal behavior ANN model of main shaft rear bearing
temperature, the variables that can affect this temperature must be taken into
consideration to build an accurate model. Wind turbines can only aerodynamically
capture a proportion of the energy in the incident wind [Hansen, 2007]. This
energy is converted by the rotor blades into mechanical power and is transmitted
directly to the generator by the main shaft because the turbine monitored in this
paper is direct-driven turbine without gearbox. Zaher et al. [Zaher et al., 2009]
established the normal behavior model of gearbox bearing temperature and cooling
oil temperature and Sanz-Bobi et al. [Garcia et al., 2006] built the same model in
addition to cooling oil thermal difference. The former model utilized active power
and nacelle temperature while the later also utilized the operation of fans used to
cool the gearbox. However, there is no gearbox in direct-driven wind turbine and
thus avoid to faults of comprehensive gearbox. The main shaft bearings are the key

142
 Chapter 7: Fault Prognosis based on Artificial Neural Network

components of this type of wind turbine. Therefore, one of the key components,
main shaft rear bearing, is main monitoring object in this section.
Accordingly, the parameters may affect the rear bearing temperature contain:
active power output, nacelle temperature, turbine speed and cooling fan status.
Unfortunately, the cooling fan status is not available in current SCADA data and
thus the parameters selected to establish ANN model for the parameter of main
shaft rear bearing temperature can be chosen as seen in Table 7.1.

Table 7.1 Input and Outputs of ANN Model

Model Output Input

Rear bearing temperature (t-1)
Active power output (t)
Rear Bearing Temperature
Nacelle temperature (t)
Turbine speed (t)

7.3.2.2 Training ANN Model

The models are trained using the parameters discussed in Table 7.1. In order to get
an accurate representation for the parameter under study, the range of the
parameters as inputs to ANN should be as varied as possible while still ensuring
the turbines are in normal operational condition in ANN training process. This was
achieved through selection the period of training data with many conditions as
possible: the starting and stopping of turbine, big changes of turbine speed, with
and without active power output. Therefore, three months from 01.01.2009-
01.04.2009 SCADA data are chosen as training data for ANN rear bearing normal
behavior model as seen in Fig. 7.2. This amounts to roughly 13,000 data points for
each input. The training process then attempts to capture the nonlinear relationship
between these parameters, i.e. the associated rear bearing and nacelle temperatures
for the turbine speed and corresponding power output. The number of training
cycles used, also known as epochs, was 1000. Determining the architecture for the
network is an iterative process and depends solely on the structure that yields the
best accuracy when tested. The final architecture used for rear bearing model was
5-10-1.
The trained model was tested in new data from a healthy wind turbine which had
not been used in training process of ANN. Fig. 7.3 shows the input data used to test
rear bearing ANN model from the same turbine of the date from 26.05.2009 to
26.07.2009. The test data shown here was also very varied. Fig. 7.4 shows the
output of rear bearing ANN model (EstimatedTemp), the real temperature of that
(BearTemp) and the difference between these values. The average difference
between actual and estimated value is 0.026 oC, and the root mean square error is
0.2, which is considered to be an acceptable level for successful fault detection and
prediction. This means that the output of ANN model can be used directly as a
comparison with the actual temperature to assess if a fault is present. If the
difference between the estimated value and the actual value increases for a

143
 Chapter 7: Fault Prognosis based on Artificial Neural Network

continuous number of instances, i.e. a prolonged period of time and not a minor
fluctuation, then this would flag as a fault.

Fig. 7.2 Neural Network Turbine Rear Bearing Temperature Model Training Data

144
 Chapter 7: Fault Prognosis based on Artificial Neural Network

Turbine Rear Bearing Temperature (t-1)

45
Temperature (Celsius Degree)

40

35

30

25
2009-05-31 2009-06-07 2009-06-14 2009-06-21 2009-06-28 2009-07-05 2009-07-12 2009-07-19 2009-07-26
(a)
Turbine Speed
20
Turbine Speed (RPM)

15

10

0
2009-05-31 2009-06-07 2009-06-14 2009-06-21 2009-06-28 2009-07-05 2009-07-12 2009-07-19 2009-07-26
(b)
Nacelle Temperature
35
Temperature (Celsius Degree)

30

25

20

15
2009-05-31 2009-06-07 2009-06-14 2009-06-21 2009-06-28 2009-07-05 2009-07-12 2009-07-19 2009-07-26
(c)
Active Power
4000
Active Power (KW))

3000

2000

1000

0
2009-05-31 2009-06-07 2009-06-14 2009-06-21 2009-06-28 2009-07-05 2009-07-12 2009-07-19 2009-07-26
(d)
Time (yyyy-mm-dd)

Fig. 7.3 Rear Bearing Model Testing Input Data

145
 Chapter 7: Fault Prognosis based on Artificial Neural Network

Real Temperature and Estimated Temperature

45
EstimatedTemp
BearTemp

40

35

30

25
2009-05-31 2009-06-07 2009-06-14 2009-06-21 2009-06-28 2009-07-05 2009-07-12 2009-07-19 2009-07-26

(a)

Difference Between Actual and Estimated Temperature

3

-1

-2

-3
09-05-31 09-06-07 09-06-14 09-06-21 09-06-28 09-07-05 09-07-12 09-07-19 09-07-26
(b)

Time (yyyy-mm-dd)

Fig. 7.4 Rear Bearing Model Output in Normal Condition

7.3.3 Prediction and Detection of Rear Bearing Fault

Once the normal behavior of rear bearing ANN model was trained, it can be used
to detect and predict the corresponding fault of rear bearing by comparing
estimated and actual temperature. Fig. 7.5(a) shows the evolution of rear bearing
temperature from the period of July 2010 to March 2011 which contains eight
months where it eventually fails. Fig. 7.5(b) shows the difference trend between the
estimated and actual temperature of rear bearing in this period. The first important
deviation from the model estimates occurred from the start of October 2010, i.e.
point ķ. The frequency of deviation and their duration increased in the following
months. From point ĸ, the deviation from the model estimates increased to 4 0C
and lasted to point Ĺ where the turbine was stopped because of overheating. Then,
the operator of wind farm tried to solve the problem two times in point Ĺ and
point ĺ, but not successful and finally the turbine was completely stopped because
of the same overheating. From this figure, the method can give the operator a
warning as early as three months in point ķ before the failure happens. With the
evolution of the failure, the deviation from model estimation increases and the
alarm can be given to operator when the deviation reaches the level of point ĸ.
Therefore, the alarm can be given as early as 10 days before the failure happens.
The results produced by ANN model for rear bearing fault detection and prediction
are very positive. They can provide an early warning of problems developing in the
bearing before the absolute temperature becomes apparently high. The results of
the fault detection can be used to help the operator to make the schedule of

146
 Chapter 7: Fault Prognosis based on Artificial Neural Network

maintenance actions before the failure happens to reduce the maintenance cost,
reduce the unanticipated downtime and improve the reliability of the wind turbine.

Actual and Estimated Temperature

60 EstimatedTemp
BearTemp

50

40

30

20

10

2010-08-01 2010-09-01 2010-10-01 2010-11-01 2010-12-01 2011-01-01 2011-02-01 2011-03-01

(a)

Difference Between Actual and Estimated Temperature

15

4
10
5
2 3
1

5
4

1.5
0

-5
2010-08-01 2010-09-01 2010-10-01 2010-11-01 2010-12-01 2011-01-01 2011-02-01 2011-03-01
(b)

Time (yyyy-mm-dd)

Fig. 7.5 Fault Detection Results of Rear Bearing

7.3.4 Discussion

This section mainly discusses whether the model established from the SCADA data
of one turbine as in Fig. 7.2 can apply in fault detection for other turbines. Then,
another turbine is selected to test for this purpose. However, there is no same fault
in this wind farm and thus the SCADA data from a new turbine in normal
condition. If the differences between actual and estimated temperatures are located
within 1.5 0C as in Fig. 7.4, the conclusion can be drawn that the ANN model for
rear bearing can be applied for fault detection and prediction for other turbines.
Fig. 7.6 shows the three months’ SCADA data of a new turbine in same wind farm
in normal condition. The data presented in this figure is also very varied: the
turbine speed is varied with starting and stopping, the active power changes from
500 KW to more than 3000 KW, and the temperatures are also varied. These three
months’ SCADA data supposed to be large varied of the parameters in normal
condition. Fig. 7.7 shows the results of ANN model using the SCADA data from
the new turbine. The estimated rear bearing temperature is very close to the actual
value. The maximum difference between estimated and actual temperatures is less
than 1.5 0C which is an early warning level as shown in Fig. 7.5. This means the

147
 Chapter 7: Fault Prognosis based on Artificial Neural Network

new turbine is in normal condition. Therefore, the ANN model of rear bearing
using SCADA data from one turbine can be applied for other turbines in same type.

Turbine Bearing Temperature (t-1)

40

35

30

25

2010-07-01 2010-08-01 2010-09-01 2010-10-01

(a)

Turbine Speed

15

10

0
2010-07-01 2010-08-01 2010-09-01 2010-10-01
(b)

Nacelle Temperature

30

20

10

0
2010-07-01 2010-08-01 2010-09-01 2010-10-01
(c)
Active Power

3000

2000

1000

0
2010-07-01 2010-08-01 2010-09-01 2010-10-01
(d)

Time (yyyy-mm-dd)

Fig. 7.6 Rear Bearing Model Testing Input Data of New Turbine

148
 Chapter 7: Fault Prognosis based on Artificial Neural Network

Actual and Estimated Temerature

50
EstimatedTemp
45 BearTemp

40

35

30

25

20

15
2010-07-01 2010-08-01 2010-09-01 2010-10-01

Difference between Actual and Estimated Temerature

2

1.5

0.5

-0.5

-1

-1.5

-2
2010-07-01 2010-08-01 2010-09-01 2010-10-01

Fig. 7.7 Rear Bearing Model Output in Normal Condition of New Turbine

7.4 Summary

This Chapter described ANN technique for early fault prediction and identification
for the main components of wind turbines especially for the bearings based on the
existing SADA data collected by commercial SCADA system. The result shows
that it can deal with large volume SCADA data and give the operators of wind
farm very early warning and close alarm to help them make the right maintenance
schedule and action decision in advance. In this way, the information presented to
the operator is dramatically reduced without omitting useful information. The
maintenance and operation cost can also be reduced by optimize the maintenance
plan, staff and preparation of tools according to the early warning and alarm. The
instance presented in this section only established a normal behaviour ANN model
for one component, i.e. main shaft rear bearing. In the future, more components
need to establish normal behaviour model. In this section, the ANN model
established by one turbine tested in new one only in normal condition, and in the
future, the test should be done in different conditions contain normal level, warning
level and alarm level.

149
Chapter 7: Fault Prognosis based on Artificial Neural Network

150
 Chapter 8: Maintenance Scheduling Optimization based on Data Mining
Techniques

8 Maintenance Scheduling Optimization based on Data

Mining Techniques

8.1 Introduction

The range of maintenance cost is from 15% for manufacturing companies and 40%
for iron and steel industry of the whole cost of manufactured parts and machines
[Mobley, 1990]. The corresponding cost in United Stated is more than 200 billion
dollars every year [Chu et al., 1998]. This shows the significance of maintenance in
the viewpoint of economy.
Generally, there are three different types of maintenance strategies. The first one is
called Corrective Maintenance (CM) which is similar to repair work, is undertaken
after a breakdown or when obvious failure has been located. However, CM at its
best should be utilized only in non-critical areas where capital costs are small,
consequences of failure are slight, no safety risks are immediate, and quick failure
identification and rapid failure repair are possible. The second one is called
preventive maintenance which is scheduled without the occurrence of any
monitoring activities. The scheduling can be based on the number of hours in use,
the number of times an item has been used, or the number of kilometers the items
has been used, according to prescribed dates. The preventive maintenance may
cause much more or much less maintenance activities, which may cause more
maintenance cost or hazard of personnel and equipment. The last one is called
Predictive Maintenance (PM) which is a set of activities that detect changes in the
physical condition of equipment (signs of failure) in order to carry out the
appropriate maintenance work for maximizing the service life of equipment
without increasing the risk of failure. PM is a dynamic schedule according to the
state of machines from continuous and/or periodic inspection. It utilizes the product
degradation information extracted and identified from on-line sensing techniques to
minimize the system downtime by balancing the risk of failure and achievable
profits.
PM has some advantages over other maintenance policies: 1) Improving
availability and reliability by reducing downtime; 2) Enhancing equipment life by
reducing wear from frequent rebuilding, minimizing potential for problems in
disassembly and reassembly and detecting problems as they occur; 3) Saving
maintenance costs by reducing repair costs, reducing overtime and reducing parts
inventory requirements; 4) Decreasing number of maintenance operations causes
decreasing of human error influence. However, there are still some challenges of
PM: 1) Initiating PM is costly because the cost of sufficient instruments could be
quite large especially if the goal is to monitor already installed equipment; 2) The
goal of PM is accurate maintenance, but it is difficult to achieve for the complexity
of equipment and environment; 3) Introducing PM will invoke a major change in
how maintenance is performed, and potentially to the whole maintenance
organization in a company. Organizational changes are in general difficult. The
objective of maintenance scheduling optimization is to optimize maintenance

151
 Chapter 8: Maintenance Scheduling Optimization based on Data Mining
Techniques

scheduling in order to maximize the whole profit, ensure safety and increase
availability.
Mathematically, the maintenance scheduling problem is a multiple-constraint, non-
linear and stochastic optimization problem. This kind of problem has been studied
for several decades and many kinds of different methods have been applied to
solve it. Two methods for PM optimization had been developed during 1980s. The
first method [Perla, 1984; Walker, 1987] performs cost/benefit analysis of each
analyzed piece of manufacturing equipment. It is based on identifying important
equipment firstly, and then predicting its future performance with and without
changes in the regularly scheduled maintenance program. The second approach is
the Reliability-Centered Maintenance (RCM) [Crellin, 1986; Hook et al., 1987;
Vasudevan, 1985]. This methodology was adopted from the commercial air
transport industry. It is based on a series of orderly steps, including identification
of system/subsystem functions and failure modes, prioritization of failures and
failure modes (using a decision logic tree), and finally selection of PM tasks that
are both applicable (i.e. have the potential of reducing failure rate) and effective
(i.e. economically worth doing). In the last two decades, many kinds of intelligent
computational methods, such as the artificial neural network method, simulated
annealing method, expert system, fuzzy systems and evolutionary optimization,
have been applied to solve the maintenance scheduling problem and obtained many
very exciting results [Huang, 1998; Miranda et al., 1998; Satoh & Nara, 1991;
Sutoh et al., 1994; Yoshimoto et al., 1993]. And also, with the rapid development
of the evolutionary theory, genetic algorithms (GAs) had become a very powerful
optimization tool and obtained wide application in this area [Arroyo & Conejo,
2002; Back et al., 1997; Huang et al., 1992; Lai, 1998; Lee & Yang, 1998; Y.
Wang & Handschin, 2000]. In recently years, several new intelligent computational
methods such as Ant Colony Optimization (ACO) and Particle Swarm
Optimization (PSO) have been applied in preventive maintenance scheduling
[Benbouzid-Sitayeb et al., 2008; Pereira et al., 2010; Yare & Venayagamoorthy,
2010].
All the above methods of maintenance scheduling are based on the specified time
periods other than based on the condition of the equipment or facilities. PM is a
good strategy which could be used to improve reliability and increase useful life of
the equipment and reduce the cost of maintenance according to the condition of
machine. When the condition of a system, such as its degradation level, can be
continuously monitored, PM policy can be implemented, according to which the
decision of maintaining the system is taken dynamically on the basis of the
observed condition of the system. Recently, genetic algorithms, Monte Carlo
method, Markov and semi-Markov methods are applied in PM [Amari et al., 2006;
Barata et al., 2001, 2002; BeĄrenguer et al., 2000; Grall et al., 2008; Marseguerra
et al., 2002]. However, there are very few literatures on applying the intelligent
computational methods in predictive maintenance based on the conditions
(degradation) of monitored machines.
This Chapter will build PM scheduling models and optimize it using Swarm
Intelligence algorithms.

152
 Chapter 8: Maintenance Scheduling Optimization based on Data Mining
Techniques

8.2 Predictive Maintenance Scheduling Optimization Based on

Swarm Intelligence

Swarm Intelligence (SI) is an innovative distributed intelligent paradigm for

solving optimization problems that originally took its inspiration from the
biological examples by swarming, flocking and herding phenomena in vertebrates.
Particle Swarm Optimization (PSO) incorporates swarming behaviors observed in
flocks of birds, schools of fish, or swarms of bees, and even human social behavior,
from which the idea is emerged. Ant Colony Optimization (ACO) deals with
artificial systems that are inspired from the foraging behavior of real ants, which
are used to solve discrete optimization problems. Bee Colony Algorithm (BCA) is
an optimization algorithm based on the intelligent foraging behavior of honey bee
swarm, proposed by Karaboga in 2005 [Karaboga, 2005]. These optimization
algorithms are metaheuristic which can solve the difficult optimization problems
even the problem is NP problem. They can easily to be applied in maintenance
scheduling optimization.
The maintenance scheduling mentioned in this Chapter does not refer to the
scheduling for one machine or one component in its life cycle but to number of
machines or components in specific time duration in order to reduce cost and
increase productivity or profit. Fig. 8.1 shows the scheme of maintenance
scheduling optimization. The results of fault diagnosis and prognosis are the key
information of the maintenance scheduling optimization. The objective of
maintenance optimization is to maximize or minimize the fitness function with
some constraints such as crew constraint and maintenance window. The following
sections will show how the swarm intelligence techniques work in maintenance
scheduling optimization through some case studies.

153
 Chapter 8: Maintenance Scheduling Optimization based on Data Mining
Techniques

Fig. 8.1 Maintenance Scheduling Optimization Scheme

8.3 Generating Unit Maintenance Scheduling (GMS) using PSO

Power generating companies must generate sufficient electrical power to cater for
the varying demands of consumers. Electricity cannot be easily and cheaply stored,
so it must be continuously generated based on the customers’ demand. With the
increasing demand of electricity, the generating unit maintenance scheduling (GMS)
of power system has become a complex, multi-object-constrained optimization
problem. Within the last three decades, several techniques have appeared in the
literature addressing such optimization problems under different scenarios
[Marwali & Shahidehpour, 2000; Negnevitsky & Kelareva, 1999]. The primary
goal of the GMS is the effective allocation of generating units for maintenance
while ensuring high system reliability, reducing production cost, prolonging
generator life time subject to some units and system constraints [Yare et al., 2008].
In order to obtain an approximate solution of a complex GMS, some new concepts
have been proposed in recent years. They include applications of probabilistic
approach [Billinton & Abdulwhab, 2003], simulated annealing [Satoh & Nara,
1991], decomposition technique [Yellen et al., 1992] and genetic algorithm (GA)
[Firmo & Legey, 2002]. A flexible GMS that considered uncertainties is proposed
with a fuzzy 0-1 integer programming technique adopted and applied to the Taiwan
power system. The application of GA to GMS has been compared with and
confirmed to be superior to other conventional algorithms such as heuristic
approaches and branch-and-bound (B&B) in the quality of solutions [Firmo &
Legey, 2002]. However, the application of particle swarm optimization (PSO) and

154
 Chapter 8: Maintenance Scheduling Optimization based on Data Mining
Techniques

their variants to GMS has not been fully explored in the literature. This section is
retrieved from [Zhang & Wang, 2010].

8.3.1 Fitness function and Constraints of GMS

Generally, there are two main categories of objective functions in GMS problems,
namely based on reliability and economic cost. The reliability criteria of levelling
reserve generation for the entire period of study is considered in this paper. As an
objective function of the GMS problems, we establish annual supply reserve ratio
levelling one of the deterministic index. Because algorithms for levelling supply
reserve ratio is easy to implement without considering probabilistic simulation
procedures operation cost, it is possible to establish an annual GMS problem (52-
week horizon). However, it has a weak point in not considering probabilistic
conditions such as generators' forced outage. Actually generation companies have
been utilizing minimizing annual supply reserve ratio more than probabilistic index
methods. For our research, we just focus on PSO algorithm accessing to GMS
problem, so it is enough to formulate the objective function as an annual supply
ratio levelling.
The problem studied here was solved by minimizing the annual supply reserve
ratio. The problem has a number of units and system constraints to be satisfied
which were described as follows:
x Load constraints – total capacity of the units running at any interval should
be not less than predicted load at that interval.
x Crew constraint – for each period, the capacity of maintenance units cannot
exceed the maximum available maintenance capacity considering crew in
this period.
x Start week of maintenance – Each unit has its maintenance periods, the
maintenance schedule cannot exceed these periods.
x Maintenance window constraints – defines the starting of maintenance at
the beginning of an interval and finish at the end of the same interval
which may contain one or several weeks. The maintenance cannot be
aborted or finished earlier than scheduled.
The objective function to be minimized is given by Eq. (8.1) subject to the
constraints given be Eq. (8.2)-(8.5).
2
T ª§ Ac  Lt · 1 T § Act  Lt · º
it Min¦ «¨ t ¸  ¦¨ ¸»
«© Lt ¹ T t 1 © Lt ¹ »¼
t 1 ¬

2
(8.1)
ª§ IC  SLt  Lt
T
· 1 T § IC  SLt  Lt ·º
Min¦ «¨ ¸  ¦¨ ¸»
«©
t 1 ¬ Lt ¹ T t 1© Lt ¹ ¼»

Subject to load constraint

155
 Chapter 8: Maintenance Scheduling Optimization based on Data Mining
Techniques

IC  SLt ! SLt (8.2)

Subject to crew constraint
N
SLt  CRt ,? SLt ¦P C
t 1
jt j
(8.3)

Subject to start week of maintenance

S min
j  S j  S max
j (8.4)

Subject to maintenance window

­° t d S j or t t S j  M j no maintenance
® (8.5)
°̄S j  t  S j  M j maintenance

Where:
T : Length of the maintenance planning scheduling (normally 52 weeks);
ACt : Available generation capacity at t th week;

Lt : Load demand at t th week;

IC : Total Installed Capacity;

SLt : Capacity loss in t th week because of maintenance;

CRt : Maximum available maintenance capacity at t th week considering crew

[MW];
S j : Starting week for maintenance scheduling of jth unit;

S min
j : Feasible minimum starting week for maintenance scheduling of jth unit;
S max
j : Feasible maximum starting week for maintenance scheduling of jth unit;
C j : Capacity of jth unit;

Pjt P{t | S j  t  S j  M j ? : Whether the jth unit maintenance in t th week.

8.3.2 Improved PSO (IPSO) Algorithm

PSO performs well in the early iterations, but they have problems approaching a
near-optimal solution. If a particle’s current position accords with the global best
and its inertia weight multiply previous velocity is close to zero, the particle will
only fall into a specific position. If their previous velocities are very close to zero,
all the particles will stop moving around the near-optimal solution, which may lead
to premature convergence of algorithm. All the particles have converged to the best
position discovered so far which may be not the optimal solution. So, an improved
PSO (IPSO) is proposed here.

156
 Chapter 8: Maintenance Scheduling Optimization based on Data Mining
Techniques

In IPSO, before updating the velocities and positions in every iteration, the
particles are ranked according to their fitness values in descending order. Select the
first part of particles (suppose mutation rate is D , fist part is (1  D ) and put them
into the next iteration directly. Regenerate the rest part of particles (D ) randomly.
In this case, we can regenerate the positions and velocities according to the
following equations instead of Eq. (3.23)-(3.24):

xid 
round rand u  S max ( j )  S min ( j )   S min ( j )  (8.6)

vid  t  vmax  round  rand u 2vmax  vid  t   [ vmax , vmax ] (8.7)

8.3.3 Case Study and Results

In order to investigate the performance of IPSO for the GMS problem, a test
system comprising 32 units over a planning period of 52 weeks was used. The case
study is described below and implemented in a MATLAB environment.
There are 32 generating units, annual peak load demand is 2,850 MW, and installed
capacity is 3,450 MW. The weekly peak loads in present of annual peak are shown
in Table 8.1. The specific data of the generators are shown in Table 8.2 which
include capacity (MW), maintenance period and load constraints. The value of
crew constraint is constant at 800 MW.
To implement PSO and IPSO, a population size of 150 particles was chosen to
provide sufficient diversity into the population taking into account the
dimensionality and complexity of the problem. This population size ensured that
the domain was examined in full but at the expense of an increase in execution
time. The other parameters of PSO and IPSO were: c1 = c2 = 2.0, Ȧ = 1.2 - 0.8
with linearly decreasing, total iteration = 300 and V  [-3, 3].
Annual supply reserve ratio values by the change of the number of iteration are
shown in Fig. 8.2. We compared simulation results between the PSO and IPSO
algorithms. We can see from this figure, the IPSO algorithm has a better
performance than PSO in GMS problems to find optimal solutions. The particles of
IPSO have a higher possibility to find optimal solution than those of PSO. The
optimal solutions of GMS problems using PSO and IPSO are shown in Table 8.3. It
contains global particles of PSO and IPSO respectively which have a best
maintenance period satisfying maintenance continuity and crew constraints, etc.

157
 Chapter 8: Maintenance Scheduling Optimization based on Data Mining
Techniques

Table 8.1 Weekly Peak Load in Percent of Annual Peak (%)

Week Load Week Load Week Load Week Load

1 86.2 14 75.0 27 75.5 40 72.4
2 90.0 15 72.1 28 81.6 41 74.3
3 87.8 16 80.0 29 80.1 42 74.4
4 83.4 17 75.4 30 88.0 43 80.0
5 88.0 18 83.7 31 72.2 44 88.1
6 84.1 19 87.0 32 77.6 45 88.5
7 83.2 20 88.0 33 80.0 46 90.9
8 80.6 21 85.6 34 72.9 47 94.0
9 74.0 22 81.1 35 72.6 48 89.0
10 73.7 23 90.0 36 70.5 49 94.2
11 71.5 24 88.7 37 78.0 50 97.0
12 72.7 25 89.6 38 69.5 51 100
13 70.4 26 86.1 39 72.4 52 95.2

Table 8.2 Data of Generators

Generat Capaci Maintena Generat Capaci Maintena

Maintena Maintena
or ty nce or ty nce
nce period nce period
(Unit) (MW) Window (Unit) (MW) Window
1 12 1-52 2 17 76 18-29 3
2 12 1-52 2 18 76 18-29 3
3 12 1-52 2 19 76 18-29 3
4 12 1-52 2 20 100 18-29 3
5 12 1-52 2 21 100 18-29 3
6 20 18-29 2 22 100 18-29 3
7 20 18-29 2 23 155 1-52 4
8 20 18-29 2 24 155 1-52 4
9 20 18-29 2 25 155 1-52 4
10 50 41-52 2 26 155 1-52 4
11 50 41-52 2 27 197 1-52 4
12 50 41-52 2 28 197 1-52 4
13 50 1-27 2 29 197 1-52 4
14 50 1-27 2 30 350 1-52 5
15 50 1-27 2 31 400 1-52 6
16 76 18-29 3 32 400 1-52 6

158
 Chapter 8: Maintenance Scheduling Optimization based on Data Mining
Techniques

Fig. 8.2 Fitness Value by the Change of the Number of Iteration

Table 8.3 Result (Maintenance period)

Maintenance period (week) Maintenance period (week)

Unit Unit
PSO IPSO PSO IPSO
1 35,36 14, 15 17 26,27,28 26,27,28
2 21,22 21,22 18 18,19,20 26,27,28
3 31,32 8,9 19 26,27,28 18,19,20
4 21,22 35,36 20 26,27,28 26,27,28
5 17,18 21,22 21 26,27,28 26,27,28
6 27,28 21, 22 22 21,22,23 18,19,20
7 21,22 27,28 23 14,15,16,17 34,35,36,37
8 27,28 21,22 24 14,15,16,17 29,30,31,32
9 18,19 27,28 25 34,35,36,37 14,15,16,17
10 41,42 50,51 26 14,15,16,17 6,7,8,9
11 41,42 50,51 27 10,11,12,13 10,11,12,13
12 41,42 41,42 28 40,41,42,43 14,15,16,17
13 6,7 14,15 29 38,39,40,41 37,38,39,40
14 14,15 3,4 30 29,30,31,32,33 9, 10, 11, 12, 13
38, 39, 40, 41, 42,
15 21,22 14, 15 31 34,35,36,37,3839
43
31, 32, 33, 34, 35,
16 18,19,20 21,22,23 32 8,9,10,11,12,13
36

8.4 Dynamic Condition-Based Maintenance Scheduling using BCA

8.4.1 Model of Condition based PM

In order to show the general idea of applying BCA in condition-based PM

scheduling, a manufacturing model is built assuming the features of the system that

159
 Chapter 8: Maintenance Scheduling Optimization based on Data Mining
Techniques

we analyze. There are several assumptions of the manufacturing system as

following:
1) The manufacturing system is subjected to deterioration.
2) Periodically the system is under inspection and each inspection reveals the
system deterioration state perfectly.
3) Machine inspection is planned at the beginning of each period.
4) The inspection time is very short and can be ignored compared to the whole
period.
5) Following an inspection based on the current state of machine ( Si ), one of the
following action is taken:
x 0 d Si  Sk : no maintenance is performed ( Sk is PM threshold);
x Sk d Si  Sn : PM is planed ( S n is CM threshold) but is not always performed ;
x Si t Sn : CM has to be performed.
6) Following a PM or CM, the machine is restored to an as-good-as-new
condition.
7) The duration of PM action is much less than that of CM action for a same
machine.

8.4.1.1 Modeling of Manufacturing System

The manufacturing system has a number of machines marked as M , and for each
machine, the productivity is Prodi . Therefore, the maximum productivity of this
system can be expressed as Eq. (8.8) while its real total productivity can be
expressed as Eq. (8.9).
M
Prod max ¦ Prod
i 1
i
(8.8)

M
Prod tot ¦ ( Prod
i 1
i x Zi ) (8.9)

where: Prodi is the productivity of i th machine. Zi is the coefficient of i th machine

productivity. The value of Zi is 1 if the i th machine is not under any kind of
maintenance, the value is 0 if it is under CM action, and the value is 0.5 if the
machine is under a PM action in a period.

8.4.1.2 Modeling of Equipment Inspection

The value of the state can belong to arrange from 0 to 1 which represent the perfect
state to the totally failure of the component. The state of machine is can be
discretized as S1 , S2 ,", Sn which S1 can be set equal 0 while S n can be set equal 1 or
a value very closed to 1 (for example 0.98 which is the CM threshold). In this
model, the beginning condition is considered for each interval. During each period,
degradation of each machine is independent and random distribution according to
Poisson distribution. At the start of each period, there is the inspection of a
machine and the obtaining of the value of the state S of a machine as shown in Fig.

160
 Chapter 8: Maintenance Scheduling Optimization based on Data Mining
Techniques

8.3. The states for all the machines can be used as parameters in predictive
maintenance scheduling.

Fig. 8.3 Inspection Point Schematic Diagram

8.4.1.3 Deterioration Model for Each Machine

Deterioration means a process where the important parameters of a system
gradually worse. If left unattended, the process will lead to deterioration failure.
Therefore, the deterioration has to be considered when a maintenance policy needs
to be employed. Fig. 8.4 shows the deterioration model of a machine. The state S
of a machine can be a value among [ S1 , Sn ] . In Fig. 8.4, Si (i 1, 2," , n) is the
predefined state of a machine, Sk is while S n is CM threshold. Pij is the transition
probability for the state from Si to S j in one period. The P M should be planned
when the state is between Sk and S n . If the state goes to S n , the CM action must be
performed which means Pn1 1 . The state transition matrix P can be expressed as
Eq. (8.10) with the constraint of Eq. (8.11).

Fig. 8.4 Degradation Model for One Machine

ª P11 P12 " P1n º

«P P22 " P2 n »»
P « 21 (8.10)
« # # % # »
« »
¬ Pn1 Pn 2 " Pnn ¼
n

¦P
j 1
ij 1 i 1," , n (8.11)

This model is very similar with the Markov model in lack of a random variable of
inspection time. With the Markov, the mean time between CM and mean time
between PM can be estimated [Amari et al., 2004]. But with the Markov model,
the accumulative error is very difficult to eliminate. The result is only the mean
time between CM and mean time between PM rather than the real plan or
scheduling of CM or PM . With that result, the maintenance action CM and PM
could be much more or less than it necessary because of uncertainty of mechanical
products. Therefore, the inspection action is performed in the beginning of every

161
 Chapter 8: Maintenance Scheduling Optimization based on Data Mining
Techniques

period as mentioned in section 8.4.1.2. What’s more, in this model, there is no any
CM or PM action when the state of the machine is in the range between S1 and
S k 1 . The PM plan is made when the state of the machine is in range between Sk
and S n 1 , and as mention above, the CM action is performed if and only if the state
of machine reach or exceed S n . To simplify the analysis, for the element values in
the state transition matrix in Eq. (8.10), from the S1 to S k 1 , only Pii and
Pi ,i 1 (i 1, 2,", k  1) have positive values and others are all zero, while from the Sk
to S n 1 , only Pii , Pi , i 1 and Pi1 (i k , k  1," , n  1) have positive values and the others
are all zero as well. The new equation can be expressed as Eq. (8.12).
ª P11 P12 0 0 0 0 0 " 0 0 º
« 0 P22 P23 0 0 0 0 " 0 0 »
« »
« 0 0 P33 P34 0 0 0 " 0 0 »
« »
« # # # % # # # " # # »
« Pk1 0 0 0 0 Pkk Pk , k 1 " 0 0 »
P « » (8.12)
« Pk 1,1 0 0 0 0 0 Pk 1, k 1 " 0 0 »
«P 0 0 0 0 0 0 " 0 0 »
« k  2,1 »
« # # # # # # # % # # »
« »
« Pn 1,1 0 0 0 0 0 0 " Pn 1, n 1 Pn 1, n »
«¬ 1 0 0 0 0 0 0 " 0 0 »¼

The ideal values of all the elements in Eq. (8.12) for the perfect deterioration model
are express from Eq. (8.13) and Eq. (8.14).
Pii 0 & Pi ,i 1 1, i 1" k (8.13)

Pii 0 & ( Pi ,i 1 1 or Pi1 1), i k , k  1" n  1 (8.14)

For the state of S n in Eq. (8.12), Pn1 1 and all values of other elements are 0
which mean that when the state reach S n , CM has to be performed. These values
could be a real situation of a manufacturing machine but it is difficult make the
values reality. To achieve this point, the values of states from S1 to S n should be
adjusted after a number of periods by statistics.

8.4.1.4 Modelling of Cost Function

There are many types of costs for each period which are analysed one by one as in
this section. All the costs calculated in this section are just for only one period.
Production Cost: it is due to the amount products produced by the manufacturing
system which means how much money it need cost to produce the amount of
products.
C prod Prodtot <C piece (8.15)

162
 Chapter 8: Maintenance Scheduling Optimization based on Data Mining
Techniques

where Cprod represents the production cost for a period while Cpiece represents the
cost for producing one piece.
Maintenance Cost: it is due to the performing PM and CM, which means the how
much money needed to perform the PM and CM.
M
CM ¦ (CM
i 1
i x C ci  PM i x C pi ) (8.16)

where CM i represents if the i th machine is under the CM (0 means no CM action

while 1 means under that action). PM i represents if the i th machine is under the
PM (0 means no PM action while 1 means under that action). Cci and C pi
represent the costs of one CM and PM action respectively for i th machine.
Total Cost: it is the total cost for one period.
Ctot C prod  CM  CI (8.17)

where Ctot is the total cost in one period while CI is the inspection cost. Because in
this model all machines are inspected for every period, the value of CI is fixed.

8.4.1.5 Modelling of Profit for the Manufacturing System

After above analysis, the total profit for one period can be calculated using Eq.
(8.18). This equation could be an objective function for optimization. The total
number of produced products in the period should be more than a minimum
number which can be describe as Eq. (8.19). Furthermore, the number of CM and
P M have a limitation because of the resources limitation, such as repairers and
tools limitation.
Profit Prodtot x Pr  Ctot Prodtot x Pr  (C prod  CM  CI )
M (8.18)
Prodtot x Pr  [ Prodtot x C piece  ¦ (CM i x Cci  PM i x C pi )  CI ]
i 1

Prod tot t Prod min (8.19)

M

¦ (CM
i 1
i  PM i ) d M max (8.20)

where P r is the price for one piece of product, Prod min is the minimum amount
products limitation of one period, and M max is a limitation of the maximum
maintenance action can be performed. In this model, to find the optimal dynamic
predictive maintenance plan for each period, Eq. (8.18) could be an objective
function and Eq. (8.19) and Eq. (8.20) could be two constraints. The aim is to make
PM maintenance scheduling to obtain maximum Profit with two constraints of Eq.
(8.19) and Eq. (8.20).

163
 Chapter 8: Maintenance Scheduling Optimization based on Data Mining
Techniques

8.4.2 Numerical Examples

In order to investigate the performance of BCA for the condition-based PM

scheduling problem, a test system comprising 30 machines is used. According to
the conditions at the start of period, a dynamic PM scheduling is made period by
period. The case study is described below and implemented in a MATLAB
environment. In this case, the number of machine is 30, and the machine
parameters are shown in Table 8.4. In the table, the Prodi means the one day
productivity for i th machine. The problem of this case could be described as:
making a PM and CM scheduling decision for 30 machines in a week according to
the initial state of each machine.
There is no mathematical method to select the best population size of the BCA.
However, there are some empirical parameters from experience. In this example,
the value of population is set to 20. The profit (fitness value) for one period (a
week) by the change of the number of iteration is shown in Fig. 8.5. The result of
PM decision and CM decision are shown in Table 8.5. In the table, the values of
PM are 0 or 1 which mean perform or not perform the PM action. The CM value is
the same mean as PM. The optimal fitness value of this numerical example is
3081390. The result shows that BCA can make the dynamic PM scheduling
optimization very effective and clear with PM model.
6
x 10
3.09

3.08

3.07
Fitness Value

3.06

3.05

3.04

3.03
0 500 1000 1500 2000 2500
Iteration

Fig. 8.5 Fitness Value by the Change of Iterations

164
Chapter 8: Maintenance Scheduling Optimization based on Data Mining
Techniques

Table 8.4 Machine Parameters

Machine Prod i C piece Pr CI Si Cpi Cci

M1 200 70 140 200 0.46 4000 10000

M2 200 70 140 200 0.51 4000 10000
M3 200 70 140 200 0.94 4000 10000
M4 200 70 140 200 0.55 4000 10000
M5 200 70 140 200 0.93 4000 10000
M6 200 70 140 100 0.96 6000 12000
M7 200 70 140 100 0.42 6000 12000
M8 200 70 140 100 0.66 6000 12000
M9 200 70 140 100 0.35 6000 12000
M10 200 70 140 100 0.35 6000 12000
M11 300 60 140 210 0.86 6500 16000
M12 300 60 140 210 0.95 5500 12000
M13 300 60 140 210 0.72 7800 16000
M14 300 60 140 210 0.82 8000 15000
M15 300 55 140 150 0.71 9000 17000
M16 300 55 140 150 0.64 6000 13000
M17 300 55 140 180 0.95 7000 15000
M18 300 55 140 180 0.75 8000 16000
M19 300 55 140 170 0.92 8000 16000
M20 300 55 140 170 0.66 8000 16000
M21 400 70 140 220 0.33 9000 20000
M22 400 70 140 220 0.5 10000 17000
M23 400 70 140 220 0.91 7500 20000
M24 400 70 140 220 0.6 10000 16000
M25 400 70 140 220 0.37 8400 20000
M26 400 70 140 220 0.97 10000 18000
M27 400 60 140 200 0.4 9000 18000
M28 400 60 140 200 0.46 9000 17000
M29 400 60 140 200 0.91 9000 18000
M30 400 60 140 200 0.7 9000 16000

165
 Chapter 8: Maintenance Scheduling Optimization based on Data Mining
Techniques

Table 8.5 Results of PM and CM by BCA

Machine PM CM Machine PM CM Machine PM CM

M1 0 0 M11 1 0 M21 0 0
M2 0 0 M12 0 1 M22 0 0
M3 1 0 M13 1 0 M23 1 0
M4 0 0 M14 1 0 M24 0 0
M5 1 0 M15 1 0 M25 0 0
M6 0 1 M16 0 0 M26 0 1
M7 0 0 M17 0 1 M27 0 0
M8 0 0 M18 1 0 M28 0 0
M9 0 0 M19 1 0 M29 1 0
M10 0 0 M20 0 0 M30 1 0

8.5 Routing and Scheduling Optimization of Maintenance Flee

(RSOM) for Offshore Wind Farm

Wind energy industry has experienced an extensive and worldwide growth during
the past years. Certain forecasts indicate that the share of wind in Europe’s energy
production will reach up to 20% in the close future [Krohn et al., 2007]. The
efficient operation of installed turbines has an increasing significance. Among
operational decisions, the planning and scheduling of maintenance tasks is decisive
regarding both turbine availability and operational costs. Considering the spread of
offshore installations and the fact that their operational costs including specialized
support resources for offshore operations, such as service vessels and personnel,
can be estimated to be five to ten times more expensive than that of the onshore
farms [Bussel & Zaaijer, 2001; Markard & Petersen, 2009], maintenance
scheduling will receive even more emphasis. Meanwhile, the support resources are
often restricted by the environmental conditions at the site, and certain operations
are allowed only in short weather windows. Missing the weather window may lead
to production interruption and economic loss.
The Chapter aims to investigate an operational decision problem, i.e. routing and
scheduling of a maintenance fleet for offshore wind farms which can be used to
avoid a time-consuming process of manually planning the scheduling and routing
with a presumably suboptimal outcome. Mathematical model of RSOM is retrieved
from a literature [Dai, 2014] and then a swarm intelligence, i.e. Ant Colony
Optimization (ACO) is modified as Duo-ACO to be applied to solve this problem.

8.5.1 Mathematical Model of RSOM

Let there are n offshore wind turbines (OWTs) indexed by i . Associate to the
delivery location of OWT i a node i , and to its pick up location a node n  i . Also
associate to the harbor, nodes 0 and 2n  1 . The definitions of the variables can be
given as following:

166
 Chapter 8: Maintenance Scheduling Optimization based on Data Mining
Techniques

Sets:
Z  : the set of delivery nodes, Z = ^1, 2,3," , n` .

Z  : the set of pick up nodes, Z  ^n  1, n  2," , 2n` .

Z Z * Z .

Z V Ž Z : the set of nodes that require the vessel present during the maintenance
operations.
N : the set of all the nodes; N Z * > 0, 2n  1@ .

V : the set of service vessels.

T : the set of days in the planning period; T ^1, 2,"` represents the length of the
period.
Constants
Tvij : the time (hours) for vessel v traversing arc (i, j ) .

Cv : the traveling cost of vessel v per hour.

Ti M : the time needed for performing the maintenance task on turbine i ;

T0M T2Mn1 0.

Li : the weight of spare parts and equipment for maintenance on turbine i .

Pi : the required personnel number for maintenance on turbine i .

TvdMAX : the maximum working hours on day d for vessel v , which is used as the
weather limitation for different vessels.
LMAX
v : the load capacity of vessel v .

PvMAX : the personnel capacity of vessel v .

Ti LATE : the latest day to perform the maintenance task on turbine i without incurring
a penalty cost.
CiPE : the penalty cost per day for the delaying maintenance task on turbine i
beyond Ti LATE .
Decision variables
­1, vessel v travels from node i to node j on maintenance day d
xvijd ®
¯ 0, otherwise

yi : the number of delayed days for maintenance task on turbine i .

tvid : the time at which vessel v visits turbine i on maintenance day d .

167
 Chapter 8: Maintenance Scheduling Optimization based on Data Mining
Techniques

k vid : the total load weight on vessel v just after it leaves node i on maintenance
day d .
qvid : the total personnel number on vessel v just after it leaves node i on
maintenance day d .
Objective function
­ ½
min ®¦¦ Cv tv (2 n 1) d  ¦ CiPE yi ¾ (8.21)
¯vV d T iZ  ¿
Constraints
¦¦¦x
j N vV d T
vijd 1, i  Z , (8.22)

¦x
i N
v 0 id 1, v  V , d  T , (8.23)

¦x
j N
vjid ¦x
j N
vijd , v  V , d  T , i  N , (8.24)

¦x
i N
vi ( 2 n 1) d 1, v  V , d  T , (8.25)

¦x
j N
vjid ¦x
j N
v ( n  i ) jd , v  V , d  T , i  Z  , (8.26)

¦¦ x
vV d T
vi ( i  n ) d 1, i  Z V , (8.27)

tv ( n  i ) d  tvid t Ti M , i  Z  , v  V , d  T , (8.28)

¦ ¦ ¦ (d < x
j N vV d T
vijd  yi ) d T jLATE , i  Z  , (8.29)

(tvid  Tij  tvjd ) xvijd d 0, i, j  N , v V , d  T , (8.30)

¦¦Lx i ijvd d LMAX

v , v  V , d  T , (8.31)
iZ  j N

(kvid  L j  kvjd ) xvijd 0, i  Z  , j  N , v  V , d  T , (8.32)

(kvid  kvjd ) xvijd 0, i  N \ Z  , j  N , v  V , d  T , (8.33)

(qvid  Pj  qvjd ) xvijd 0, i  Z  , j  N , v  V , d  T , (8.34)

(qvid  Pj  qvjd ) xvijd 0, i  Z  , j  N , v  V , d  T , (8.35)

0 d kvid d LMAX
v i  N , v V , d  T , (8.36)

0 d qvid d PvMAX i  N , v V , d  T , (8.37)

tv (2 n 1) d TvdMAX v  V , d  T , (8.38)

tv 0 d 0 v  V , d  T , (8.39)

168
 Chapter 8: Maintenance Scheduling Optimization based on Data Mining
Techniques

yi t 0, i  Z  (8.40)
Constraints -
1) Eq. (8.22) ensure that each OWT is visited only once for delivery and once
for pick up.
2) Eq. (8.23) and (8.25) ensure that each vessel leaves and returns the harbor
only once every day.
3) Eq. (8.24) and (8.26) ensure flow conservation at each node.
4) Eq. (8.27) means that if the vessel needs to present during the maintenance
operation on one OWT, it will only leave the OWT when the operation is
completed.
5) Eq. (8.28) is precedence constraints which force the pickup is not done
before completing the maintenance operation on the same OWT.
6) Eq. (8.29) is soft constraints which require that the maintenance task is
performed within the preferred time.
7) Eq. (8.30) keeps the travelling time compatibility of each vessel.
8) Eq. (8.31) ensures the service vessels are not overloaded.
9) Eq. (8.32) expresses the compatibility requirements between routes and
vessel loads.
10) Eq. (8.33) ensures that no extra load added when the vessels pick up from
OWTs.
11) Eq. (8.34) and (8.35) describe the compatibility requirements between
routes and personnel number on the vessels.
12) Eq. (8.36) and (8.37) guarantee that neither of load or personnel number
exceeding the vessel limitations.
13) Eq. (8.38) imposes a maximal working time of the service vessels on each
day.
14) Eq. (8.39) means the time is counted from the vessels leaving the harbor.
15) Eq. (8.40) set the delayed maintenance day to be non-negative.

8.5.2 Application of Duo-ACO in RSOM Problem

ACO is a meta-heuristic technique which is inspired by the foraging behavior of

some ant species [Marco Dorigo et al., 2006]. It is a very good algorithm for
solving optimization problem typically Travelling Salesman Problem (TSP). In
classical TSP problem, there are many cities and only one salesman. If there are
two sales man to travel all these cities and each city can and only can be traveled
once, how to solve this Duo-TSP problem? The RSOM problem may have two or
more vessels which is very similar with duo-TSP problem. This section describes
the principle of Duo-ACO.
The idea of Duo-ACO is evolved from basic ACO which is introduced in Section
3.5.1. Duo-ACO has two groups with the same number of ants and each group has
its own pheromone (group1, group2 and pheromone1, pheromone2 respectively).
The procedure of the algorithm can be written as:
Begin
Initialization

169
 Chapter 8: Maintenance Scheduling Optimization based on Data Mining
Techniques

While stopping criterion not satisfied do

Deploy each ant ( k ) in a starting city for group1
Deploy each ant ( k ) in a starting city for group2
(The ant with the same sequence ( k ) of two groups cannot in same city)
For each ant (same sequence ant ( k ) for both group)
Repeat
Calculate probability of remaining cities selected to be next city for
group1
Choose next city according to probability using roulette wheel selection
algorithm for group1

Calculate probability of remaining cities selected to be next city for

group2
Choose next city according to probability using roulette wheel selection
algorithm for group2
Until all cities are visited
Update pheromone1
Update pheromone2
End for
Update the best routes (route1 for group1 while route2 for group2)
End while
Record and output the beat routes (solutions)
End
The implement steps of duo-ACO are shown in Fig. 8.6. In each iteration, the two
ants with the same index ( k ) in two groups select nodes (cities in TSP problem)
alternatively according to their probabilities. Accordingly, the pheromones for two
groups can be updated respectively. After all ants passed all nodes, the iteration
number increases one by one until maximum iteration. Then the best routes of two
groups with the same index are recorded as the best solution. Referring to apply
Duo-ACO in RSOM problem, the solution of each group represents the route of
corresponding vessel. The dij in Eq. (3.21) is replaced by Reciprocal of Eq. (8.21)
in the case of RSOM problem.

8.5.3 Numerical Examples

To examine the effectiveness of Duo-ACO application in RSOM problem, several

case studies are presented in this section. Fig. 8.7 shows an example of offshore
wind farm with 64 wind turbines. The states of turbines can be “Replacement”,
“Repair” and “No service demand” according to the results of condition monitoring
system especially of fault diagnosis and prognosis. There are two vessels can be
used as maintenance flee and the parameters of them are shown in Table 8.6. The
parameters of turbines and the maximum working hours for each day are shown in
Table 8.7 and Table 8.8 respectively. The maximum working hours in Table 8.8
can be obtained from weather condition forecasting.

170
 Chapter 8: Maintenance Scheduling Optimization based on Data Mining
Techniques

Start

Initializetheparameters,
SetThealgorithmiterationnumber NC=0

Place all ants of two groups to the random positions,

Computeprobabilitiesofnextselectednode forall
unvisitednodesforant1(k)byEq.(3.21)

Choosingnextnode accordingtotheprobabilities
NC=NC+1 k=k+1
androulettewheelselectionprinciple

Allnodesarevisitedbyant1(k)and Yes
ant2(k)?

No
Computeprobabilitiesofnextselectednode forall
unvisitednodesforant2(k)byEq.(3.21)

Choosingnextnode accordingtotheprobabilities
androulettewheelselectionprinciple

NO Allnodesarevisitedbyant1(k)and
ant2(k)?

Yes

Updatepheromone1andpheromone2 byEq.(3.22)

No
Allantshavevisitedallnodes?

Yes
No
Reachmaximumiteration
(NC=NCmax)
orotherterminationcriterion?

Yes
End

Fig. 8.6 The implementation steps of Duo-ACO

171
 Chapter 8: Maintenance Scheduling Optimization based on Data Mining
Techniques

The process of the program was shown in Fig. 8.6. There are two groups of ants
and each of them represents a vessel. The routing of ant represents the routing and
scheduling of maintenance. From the experience, the number of ants in each group
should be approximately the number of nodes the ants be visited which are
offshore turbines in this case. Therefore, the parameters of Duo-ACO are set as:
number of ants of each group is 10, the maximum iteration is 300, important
coefficient of pheromone D and E are set as 1 and 5 respectively, and the
pheromone evaporation coefficient is 0.1.
The results of maintenance scheduling and routing with 8 offshore turbines are
shown in Table 8.9. Vessel1 and vessel2 visit and maintain 5 turbines and 3
turbines respectively. The routing number here is the same mean as Table 8.7. The
result is that the two vessels can visit and maintain these turbines within one day
(5.4915 and 8.7097 hours respectively) and the objective value of Eq. (8.21) is
3848.5.

Fig. 8.7 The offshore wind farm example

172
 Chapter 8: Maintenance Scheduling Optimization based on Data Mining
Techniques

Table 8.6 Parameters of Maintenance Vessels

Speed Load Capacity Personal Capacity Cost

Vessels
S
( , km/h) ( , kg)
L ( )
P C
( , €/h)
Vessel1 33 1500 12 225
Vessel2 20 26000 12 300

Table 8.7 Parameters of 8 Turbines

Time window Penalty cost Required Required Task duration

Unit Turbines Task type (day) (euro/day) load (kg) personnel (hours)
LATE
Ti CiPE Li Pi Ti M
1 T12 Repair 2 1600 200 2 3
2 T19 Replacement 1 3000 200 2 2
3 T30 Repair 3 1200 100 2 1
4 T36 Replacement 2 2000 800 5 3
5 T39 Replacement 3 2000 300 2 2
6 T42 Repair 4 1000 200 2 2
7 T52 Replacement 2 2000 800 4 3
8 T60 Repair 4 1000 500 3 4

Table 8.8 Mmaximum Working Hours for Each Day

Mmaximum Working Hours

Time (Day)
Vessel1 Vessel2
Day 1 6 10
Day 2 6 10
Day 3 8 12
Day 4 7 11
Day 5 7 11
Day 6 5 8
Day 7 6 10
Day 8 6 10
Day 9 6 10
Day 10 7 11

173
 Chapter 8: Maintenance Scheduling Optimization based on Data Mining
Techniques

Table 8.9 Results of Maintenance Routing with 8 Turbines

No. of
Needed Time Objective
Vessels Visited Routing
(hours of each day) Value
Turbine
Vessel1 5 0-1-2-6-3-5-3-2-6-5-1-0 5.4915
3848.5
Vessel2 3 0-7-8-4-7-4-8-0 8.7097

6500

6000

5500
Objective Value

5000

4500

4000
3848.5

3500
0 50 100 150 200 250 300
Iteration

Fig. 8.8 Objective Value Changes with Iteration (8 turbines)

In order to examine the Duo-ACO performance for a large number turbines’ wind
farm, a new offshore wind farm with 28 turbines are tested. The information of two
vessels is the same as shown in Table 8.6 and the maximum working hours for
each day is the same as Table 8.8. The conditions and parameters of 28 turbines are
shown in Table 8.10. The parameters of Duo-ACO changes because of the
increasing the number of wind turbine. The number of ants of each group is set as
30 and the maximum iteration is set as 1000. The results are shown in Table
8.11Table and Fig. 8.9. The vessel1 and vessel2 visit and repair, inspection or
replacement 19 turbines and 9 turbines respectively. Vessel1 need four days to visit
and maintain all these 19 turbines, and it needs 5.859, 5.7938, 6.9541, and 4.8947
hours for each day which are less than that of the maximum working hours of
vessel1 in Table 8.8. Vessel2 need 2 days to visit and maintain 9 turbines, and it
needs 7.4814 and 7.0311 hours for each day which are also less than that of
maximum working hours of vessels2 in Table 8.8. The objective value of fitness
function of Eq. (8.21) is 94641.6 as shown in Table 8.11.
These two numerical examples show how to apply Duo-ACO in scheduling and
routing of maintenance fleet for offshore wind farms which is a complex non-linear
problem. Example 1 shown the problem solution with 8 offshore turbines while
example 2 shows that of 28 offshore turbines and both of examples show the
effectively of Duo-ACO application of the scheduling and routing problems of
offshore wind farms.

174
 Chapter 8: Maintenance Scheduling Optimization based on Data Mining
Techniques

Table 8.10 Parameters of 28 Turbines

Time Task
Penalty cost Required Required
window duration
Unit Turbines Task type (euro/day) load (kg) personnel
(day) (hours)
CiPE Li Pi
Ti LATE Ti M
1 T3 Replacement 3 2000 800 3 3
2 T4 Repair 6 500 50 2 2
3 T6 Replacement 4 1500 800 3 3
4 T11 Inspection 12 0 20 1 1
5 T12 Repair 4 1600 200 2 3
6 T13 Replacement 2 2500 500 3 2
7 T14 Replacement 2 2000 500 3 2
8 T16 Repair 5 1000 300 3 2
9 T19 Replacement 1 3000 200 2 2
10 T21 Repair 7 1000 50 1 2
11 T23 Inspection 12 0 20 1 1
12 T25 Inspection 10 0 20 1 1
13 T27 Replacement 2 2500 500 2 3
14 T30 Repair 4 1200 100 2 1
15 T36 Replacement 3 2000 800 5 3
16 T38 Inspection 12 0 20 1 1
17 T39 Replacement 4 2000 300 2 2
18 T42 Repair 5 1000 200 2 2
19 T44 Inspection 10 0 20 1 1
20 T45 Repair 8 1000 500 2 2
21 T49 Replacement 1 3000 800 4 3
22 T52 Replacement 2 2000 800 4 3
23 T54 Repair 5 1000 50 1 1
24 T55 Replacement 3 2000 500 3 2
25 T58 Inspection 13 0 20 1 1
26 T60 Repair 6 1000 500 3 4
27 T61 Repair 7 1000 300 2 3
28 T62 Inspection 12 0 20 1 1

175
 Chapter 8: Maintenance Scheduling Optimization based on Data Mining
Techniques

Table 8.11 Results of Maintenance Routing with 28 Turbines

No. of Needed Time
Objective
Vessels Visited Routing (hours of
Value
Turbine each day)
5.859,
0-5-4-9-12-21-25-4-12-25-9-5-21-0-22-22-
5.7938,
Vessel1 19 0-24-23-28-27-26-19-23-28-19-24-27-26-
6.9541,
0-8-11-16-17-20-18-11-16-8-17-20-18-0
4.8947 9641.6

0-2-1-13-15-2-1-13-15-0-3-7-6-10-14-14- 7.4814,
Vessel2 9
7-6-10-3-0 7.0311

4
x 10
1.35

1.3

1.25

1.2
Objective Value

1.15

1.1

1.05

1
9641.6

0.95
0 100 200 300 400 500 600 700 800 900 1000
Iteration

Fig. 8.9 Objective Value Changes with Iteration (28 turbines)

There is also a drawback of the Duo-ACO to solve this problem. With the
increasing of turbines, the process to find the solution using Duo-ACO becomes
time-consuming. However, this problem is not so time sensitive which means the
key point is to find the optimal solution regardless how much time it using.
Therefore, the Duo-ACO is a suitable algorithm to solve this non-linear scheduling
and routing problem.

8.6 Summary

This Chapter mainly described the maintenance scheduling optimization based on

Swarm Intelligence such as PSO, BCA and ACO. For each algorithm, applications
in industry or numerical examples were described to indicate how the algorithm
works in maintenance scheduling.
For algorithm of PSO, the problem of maintenance scheduling of generating units
for reliable operation of a power system with 32 units were tested. In this case, the
annual supply reserve ratio was selected as fitness function with constraints of load,
crew and maintenance window. The maintenance schedule was not based on the

176
 Chapter 8: Maintenance Scheduling Optimization based on Data Mining
Techniques

condition of machines but based on a fixed period (a year) and so it can be seen as
preventive maintenance scheduling. PSO was improved with mutation rate D
called improved PSO (IPSO) to apply in generating units maintenance scheduling.
Both PSO and IPSO can find optimal maintenance schedule of generation units but
IPSO has better performance with faster convergence speed and better fitness value.
For application of BCA in predictive scheduling optimization, a model of dynamic
model of condition based maintenance was established. The dynamic predictive
maintenance model is based on the condition of machines other than fixed period
like preventive maintenance. The main effort of predictive maintenance (PM) is to
avoid unnecessary maintenance action tasks by taking maintenance action just in
case of detecting any evidence of abnormal performance in physical condition. A
PM program can significantly decline the maintenance cost by decreasing the
number of needless scheduled preventive activities. PM program allows the
maintenance function to do only the right things, at the correct time, minimizing
spare parts cost, system downtime and time spent on maintenance. Based on the
model and condition of each machine, a dynamic scheduling of PM and CM can be
done using BCA. The result obtained from the numerical example confirms the
trend of successful application using this algorithm in the field of PM, where a
dynamic approach has a fundamental importance. Although the desired results
have fully achieved, and the analysis has helped to highlight and solve many
critical issues, it is clear that more careful analysis should be done when analyzing
PM maintenance model. In this Chapter, only one single kind condition for each
machine. However, mostly, more than one parameters get together to determine the
state of a machine. Therefore, how to get the state of a machine using different
parameters could be a future research field. Furthermore, the case study in this
Chapter only consider the one period because the limitation of our resources. In the
future, the long history period should be considered and the methods for adjusting
the state Si (i 1, 2," , n) could be a good topic to research.
For application of ACO in maintenance scheduling, a model of scheduling and
routing of maintenance fleet for offshore wind farms was established. ACO was
varied with two groups of ants which called Duo-ACO. Through the numerical
examples, Duo-ACO can solve this problem effectively even if the number of
turbines increasing. The drawback of the methodology is that it is impossible to
know if the optimal solution found by Duo-ACO is the best one.

177
Chapter 8: Maintenance Scheduling Optimization based on Data Mining
Techniques

178
 Chapter 9: Conclusion and Future Work

9 Conclusion and Future Work

This chapter provides general overall comments and concluding remarks about the
work presented in this thesis and some suggestions for future work.

9.1 Summary and Conclusions

The goals of this thesis are to develop a framework of intelligent Condition based
Maintenance (CBM) and apply data mining techniques in its phases. CBM is a
sufficient maintenance strategy which take maintenance action just before the
failure based on the condition of equipment to increase the reliability and
availability of the equipment and meanwhile reduce maintenance and operation
cost. It can also improve the safety for both equipment and operation staff. There
are mainly two tasks of CBM: the one is fault diagnosis and prognosis for the
equipment and the other is based on which to optimize the maintenance scheduling.
Chapter 2 presented framework of Intelligent Fault Diagnosis and Prognosis
System (IFDPs) for CBM which showed phases of the CBM and data mining
techniques applied in the system.
Chapter 3 presented data mining techniques applying in IFDPS, including Artificial
Neural Network (ANN), Swarm Intelligence (SI) and Association Rules (AR). The
techniques of ANN and AR are supposed to be applied in fault diagnosis and
prognosis while the techniques of SI are supposed to be applied in sensor place
optimization and maintenance optimization.
Chapter 4 introduced the sensor classification and sensor placement optimization
techniques. The presented methods sensor placement optimization is combination
of Finite Element Analysis (FEA) and SI algorithm such as PSO and BCO are
suitable for component level and machine level of sensor placement optimization.
However, the system level sensor placement optimization need to be further
researched.
Chapter 5 presented methods of signal processing typically for vibration signals
and feature extraction. The vibration signals can be processed in time domain,
frequency domain, time-frequency domain and wavelet domain analysis which
many features (parameters) can be extracted. The parameters extracted from
signals may be too many to be classified or predicted using data mining techniques
and thus feature selection techniques need to be used to reduce the dimensionality
of the parameters. PCA is an unsupervised learning approach for dimensionality
reduction that uses correlation coefficients of the parameters to combine and
transform them into a reduced dimensional space. It transforms high
dimensionality features to lower dimensionality but not select the features from
original features directly. Therefore, the feature selection directly from original
features should be researched.

179
 Chapter 9: Conclusion and Future Work

Chapter 6 presented the methods of fault diagnosis, i.e. fault detection and
classification, based on data mining techniques such as BP network, SOM and
Association Rules. The conclusions have been presented at the end of this chapter.
When the history data is available but the physical model and mathematical model
are not available or not accurate, the data-driven techniques can be sufficient
applied in fault diagnosis.
Chapter 7 presented fault prognosis based on the indicator prediction of the fault
using BP network. The traditional methods of data-driven fault prognosis are based
on statistics of the history data [Lee et al., 2006]. ANN model is supposed to be
used for multi-component, multi-fault prognosis but the case study for wind turbine
fault prognosis in this chapter only one component and one fault was used. In the
future, the multi-component, multi-fault ANN model should be further researched.
Chapter 8 presented the maintenance optimization based on data mining techniques.
Three different models and Swarm Intelligence (variants of PSO, BCA and ACO)
were presented in this Chapter. Generating Unit Maintenance Scheduling is a
preventive maintenance optimization, while the following two examples are
predictive maintenance or so-called CBM, and both of which can use data mining
techniques to solve.

9.2 Suggestions of Future Work

The following are proposed for future work:

¾ Developing sensor placement optimization methods in system level, i.e. more
than two machines.
¾ Developing methods of feature selection directly from original features.
¾ Developing hybrid methods of model-based and data-driven for fault
diagnosis and prognosis to improve accuracy.

180
 Reference

References

Aamodt, A., & Plaza, E. (1994). Case-Based Reasoningௗ: Foundational

Issues , Methodological Variations , and System Approaches. AI
Communications, Vol. 7, No. 1, pp. 39–59.
Abraham, A., Grosan, C., & Ramos, V. (2006). Swarm intelligence in data
mining. Berlin/Heidelberg: Springer.
Agrawal, R., ImieliĔski, T., & Swami, A. (1993). Mining association rules
between sets of items in large databases. Proceedings of the 1993 ACM
SIGMOD international conference on Management of data -
SIGMOD ’93, pp. 207–216. New York, USA: ACM Press.
doi:10.1145/170035.170072.
Ahmad, R., & Kamaruddin, S. (2012). An overview of time-based and
condition-based maintenance in industrial application. Computers &
Industrial Engineering, Vol. 63, No. 1, pp. 135–149.
doi:10.1016/j.cie.2012.02.002.
Akyildiz, I. F., Su, W., Sankarasubramaniam, Y., & Cayirci, E. (2002).
Wireless sensor networks: a survey. Computer Networks, Vol. 38, No.
4, pp. 393–422. doi:10.1016/S1389-1286(01)00302-4.
Al-Shehabi, A. G., & Newman, B. (2001). Aeroelastic vehicle optimal
sensor placement for feedback control applications using mixed gain-
phase stability. Proceedings of the 2001 American Control Conference.
(Cat. No.01CH37148), Vol. 3, pp. 1848–1852. IEEE.
doi:10.1109/ACC.2001.946005.
Altmann, J., & Mathew, J. (2001). Multiple Band-Pass Autoregressive
Demodulation for Rolling-Element Bearing Fault Diagnostics.
Mechanical Systems and Signal Processing, Vol. 15, No. 5, pp. 963–
977. doi:10.1006/mssp.2001.1410.
Amari, S. V, Ph, D., Corporation, R. S., & Mclaughlin, L. (2004). Optimal
Design of a Condition-Based Maintenance Model. Annual Symposium
Reliability and Maintainability, pp. 528–533.
Amari, S. V, Ph, D., Corporation, R. S., & Mclaughlin, L. (2006). Cost-
Effective Condition-Based Maintenance Using Markov Decision
Processes. Annual of Reliability and Maintainability Symposium, Vol.
00, pp. 464–469.
Aminian, F. (2001). Fault Diagnostics of Analog Circuits using Bayesian
Neural Networks with Wavelet Transform as Preprocessor. Journal of

181
 Reference

Electronic Testing: Theory and Applications, Vol. 17, No. 1, pp. 29–36.
doi:10.1023/A:1011141724916.
Andersen, T. M., & Rasmussen, M. (1999). Decision support in a condition
based environment. Journal of Quality in Maintenance Engineering,
Vol. 5, No. 2, pp. 89–102. doi:10.1108/13552519910271793.
Andria, G., Savino, M., & Trotta, A. (1994). Application of Wigner-Ville
Distribution to Measurements on Transient Signals. IEEE Transactions
on Instrumentation and Measurement, Vol. 43, No. 2, pp. 187–193.
doi:10.1109/19.293418.
Ansari, F. (1998). Fiber Optic Sensors for Construction Materials and
Bridges. Lancaster: Taylor & Francis.
Arroyo, J. M., & Conejo, A. J. (2002). A parallel repair genetic algorithm to
solve the unit commitment problem. IEEE Transactions on Power
Systems, Vol. 17, No. 4, pp. 1216–1224.
doi:10.1109/TPWRS.2002.804953.
Back, T., Hammel, U., & Schwefel, H.-P. (1997). Evolutionary computation:
comments on the history and current state. IEEE Transactions on
Evolutionary Computation, Vol. 1, No. 1, pp. 3–17.
doi:10.1109/4235.585888.
Barata, J., GuedesSoares, C., Marseguerra, M., & Zio, E. (2001). Monte
Carlo simulation of deteriorating systems. Proceedings ESREL, pp.
879–886.
Barata, J., Soares, C. G., Marseguerra, M., & Zio, E. (2002). Simulation
modelling of repairable multi-component deteriorating systems for “on
condition” maintenance optimisation. Reliability Engineering & System
Safety, Vol. 76, No. 3, pp. 255–264.
Baskar, S., & Suganthan, P. N. (2004). A novel concurrent particle swarm
optimization. Proceedings of the 2004 Congress on Evolutionary
Computation (IEEE Cat. No.04TH8753), pp. 792–796. IEEE.
doi:10.1109/CEC.2004.1330940.
Beƍrenguer, C., Grall, A., & Castanier, B. (2000). Simulation and evaluation
of Condition-based maintenance policies for multi-component
continuous-state deteriorating systems. Proceedings of the Foresi.
Becker, E., & Poste, P. (2006). Keeping the Condition Monitoring of Wind
Turbine Gears. Wind Energy, Vol. 7, No. 2, pp. 26–32.
Beigel, M. (1982). Identification Device. US.
Belkin, M., Niyogi, P., & Sindhwani, V. (2006). Manifold Regularizationௗ:
A Geometric Framework for Learning from Labeled and Unlabeled

182
 Reference

Examples. Journal of Machine Learning Research, Vol. 7, pp. 2399–

2434.
Bellman, R. (1961). Adaptive control processes: a guided tour. New Jersey:
Princeton University Press.
Benbouzid-Sitayeb, F., Ammi, I., Varnier, C., & Zerhouni, N. (2008).
Applying Ant Colony Optimization for the Joint Production and
Preventive Maintenance Scheduling Problem in the Flowshop
Sequencing Problem. 2008 3rd International Conference on
Information and Communication Technologies: From Theory to
Applications, pp. 1–6. IEEE. doi:10.1109/ICTTA.2008.4530343.
Beni, G. (1988). The concept of cellular robotic system. Proceedings IEEE
International Symposium on Intelligent Control 1988, pp. 57–62. IEEE
Comput. Soc. Press. doi:10.1109/ISIC.1988.65405.
Benouaret, M., Sahour, A., & Harize, S. (2012). Real time implementation
of a signal denoising approach based on eight-bits DWT. AEU -
International Journal of Electronics and Communications, Vol. 66, No.
11, pp. 937–943. doi:10.1016/j.aeue.2012.04.001.
Berenji, H. R. (2006). Case-Based Reasoning for Fault Diagnosis and
Prognosis. 2006 IEEE International Conference on Fuzzy Systems, pp.
1316–1321. IEEE. doi:10.1109/FUZZY.2006.1681880.
Bergh, F. van den, & Engelbrecht, A. . (2004). A Cooperative Approach to
Particle Swarm Optimization. IEEE Transactions on Evolutionary
Computation, Vol. 8, No. 3, pp. 225–239.
doi:10.1109/TEVC.2004.826069.
Bhushan, M., & Rengaswamy, R. (2000). Design of sensor location based
on various fault diagnostic observability and reliability criteria.
Computers & Chemical Engineering, Vol. 24, pp. 735–741.
doi:http://dx.doi.org/10.1016/S0098-1354(00)00331-8.
Billinton, R., & Abdulwhab, A. (2003). Short-term generating unit
maintenance scheduling in a deregulated power system using a
probabilistic approach. IEE Proceedings - Generation, Transmission
and Distribution, Vol. 150, No. 4, pp. 463.
Biskas, P. N., Ziogos, N. P., Tellidou, A., Zoumas, C. E., Bakirtzis, A. G.,
Petridis, V., & Tsakoumis, A. (2006). Comparison of Two
Metaheuristics with Mathematical Programming Methods for the
Solution of OPF. Proceedings of the 13th International Conference on,
Intelligent Systems Application to Power Systems, Vol. 153, pp. 510–
515. IEEE. doi:10.1109/ISAP.2005.1599316.

183
 Reference

Blackwell, T. M., & Bentley, P. (2002). Don’t push me! Collision-avoiding

swarms. Proceedings of the 2002 Congress on Evolutionary
Computation. CEC’02 (Cat. No.02TH8600), Vol. 2, pp. 1691–1696.
IEEE. doi:10.1109/CEC.2002.1004497.
Bloch, H. P., & Geitner, F. K. (2012). Machinery failure analysis and
troubleshooting, 4th ed., pp. 1–760. Houston: Butterworth-Heinemann.
Blum, A., & Chawla, S. (2001). Learning from Labeled and Unlabeled Data
using Graph Mincuts. ICML ’01 Proceedings of the Eighteenth
International Conference on Machine Learning, pp. 19–26.
Bonabeau, E., Dorigo, M., & Theraulaz, G. (1999). Swarm Intelligence:
From Natural to Artificial Systems. New York: Oxford University
Press.
Boucly, F. (2001). Le Management de la maintenance: évolution et
mutation, pp. 1–307. Paris: Association Française de Normalisation
(AFNOR).
Brando, G., Dannier, A., Del Pizzo, A., & Rizzo, R. (2007). Quick
identification technique of fault conditions in cascaded H-Bridge
multilevel converters. 2007 International Aegean Conference on
Electrical Machines and Power Electronics, Vol. 3, pp. 491–497. IEEE.
doi:10.1109/ACEMP.2007.4510549.
Breiman, L. (2001). Random Forests. Machine Learning, Vol. 45, No. 1, pp.
5–32. doi:10.1023/A:1010933404324.
Bussel, G. J. W. van, & Zaaijer, M. B. (2001). Reliability , Availability and
Maintenance aspects of large-scale offshore wind farms , a concepts
study . In Marine Renewable Energies Conference, Vol. 113, pp. 119–
126. Newcastle.
Callaway, E., Gorday, P., Hester, L., Gutierrez, J. A., Naeve, M., Heile, B.,
& Bahl, V. (2002). Home networking with IEEE 802.15.4: a
developing standard for low-rate wireless personal area networks.
IEEE Communications Magazine, Vol. 40, No. 8, pp. 70–77.
doi:10.1109/MCOM.2002.1024418.
Campos, J. (2009). Development in the application of ICT in condition
monitoring and maintenance. Computers in Industry, Vol. 60, No. 1, pp.
1–20. doi:10.1016/j.compind.2008.09.007.
Carnero Moya, M. C. (2004). The control of the setting up of a predictive
maintenance programme using a system of indicators. Omega, Vol. 32,
No. 1, pp. 57–75. doi:10.1016/j.omega.2003.09.009.

184
 Reference

Cecchin, T., Ranta, R., Koessler, L., Caspary, O., Vespignani, H., &
Maillard, L. (2010). Seizure lateralization in scalp EEG using Hjorth
parameters. Clinical neurophysiologyࣟ: official journal of the
International Federation of Clinical Neurophysiology, Vol. 121, No. 3,
pp. 290–300. doi:10.1016/j.clinph.2009.10.033.
Chen, D., & Trivedi, K. S. (2005). Optimization for condition-based
maintenance with semi-Markov decision process. Reliability
Engineering & System Safety, Vol. 90, pp. 25–29.
doi:10.1016/j.ress.2004.11.001.
Chen, D., & Wang, W. J. (2002). Classification of Wavelet Map Patterns
Using Multilayer Neural Networks for Gear Fault Detection.
Mechanical Systems and Signal Processing, Vol. 16, No. 4, pp. 695–
704. doi:10.1006/mssp.2002.1488.
Chen, G., Liu, Y., Zhou, W., & Song, J. (2008). Research on intelligent fault
diagnosis based on time series analysis algorithm. The Journal of
China Universities of Posts and Telecommunications, Vol. 15, No. 1,
pp. 68–74.
Chen, S. Y., & Li, Y. F. (2002). A method of automatic sensor placement
for robot vision in inspection tasks. Proceedings 2002 IEEE
International Conference on Robotics and Automation (Cat.
No.02CH37292), Vol. 3, pp. 2545–2550. IEEE.
doi:10.1109/ROBOT.2002.1013614.
CHERNG, A.-P. (2003). OPTIMAL SENSOR PLACEMENT FOR
MODAL PARAMETER IDENTIFICATION USING SIGNAL
SUBSPACE CORRELATION TECHNIQUES. Mechanical Systems
and Signal Processing, Vol. 17, No. 2, pp. 361–378.
doi:10.1006/mssp.2001.1400.
Chong, C., Hean Low, M., Sivakumar, A., & Gay, K. (2006). A Bee Colony
Optimization Algorithm to Job Shop Scheduling. Proceedings of the
2006 Winter Simulation Conference, pp. 1954–1961. IEEE.
doi:10.1109/WSC.2006.322980.
Chu, C., Proth, J., & Wolff, P. (1998). Predictive maintenanceௗ: The One-
unit Replacement Model. International Journal of Production
Economics, Vol. 54, No. 3, pp. 285–295.
doi:http://dx.doi.org/10.1016/S0925-5273(98)00004-8.
Clerc, M., & Kennedy, J. (2002). The particle swarm - explosion, stability,
and convergence in a multidimensional complex space. IEEE
Transactions on Evolutionary Computation, Vol. 6, No. 1, pp. 58–73.
doi:10.1109/4235.985692.

185
 Reference

Cooke, R. M. (1992). Experts in Uncertainty: Opinion and Subjective

Probability in Science, p. 336. New York: Oxford University Press.
Corinthios, M. J. (1971). A fast Fourier transformation for high-speed signal
processing. IEEE Transaction on Computers, Vol. 20, No. 8, pp. 843–
846. doi:http://doi.ieeecomputersociety.org/10.1109/T-C.1971.223359.
Crellin, G. L. (1986). Use of reliability-centered maintenance for the
McGuire Nuclear Station feed water system, EPRI NP-4795.
Dai, L. (2014). Safe and efficient operation and maintenance of offshore
wind farms. NTNU.
Dargie, W., & Poellabauer, C. (2010). Fundamentals of Wireless Sensor
Networks: Theory and Practice, p. 311. West Sussex: JohnWiley &
Sons Ltd.
Daubechies, I. (1988). Ortho-normal bases of compactly supported
wavelets.pdf. Communications on Pure and Applied Mathematics, Vol.
41, pp. 909–996.
De Kleer, J., & Williams, B. (1987). Diagnosing Multiple Faults. Artificial
Intelligence, Vol. 32, No. 1, pp. 97–130. doi:10.1016/0004-
3702(87)90063-4.
Deshpande, V. ., & Modak, J. . (2002). Application of RCM for safety
considerations in a steel plant. Reliability Engineering & System Safety,
Vol. 78, No. 3, pp. 325–334. doi:10.1016/S0951-8320(02)00177-1.
Dieulle, L., BCrenguer, C., Grall, A., & Roussignol, M. (2001). Continuous
Time Predictive Maintenance Scheduling for a Deteriorating System.
Dirilten, H. (1972). On the Mathematical Models Characterizing Faulty
Four-Phase MOS Logic Arrays. IEEE Transactions on Computers, Vol.
C-21, No. 3, pp. 301–305. doi:10.1109/TC.1972.5008954.
Djurdjanovic, D., Lee, J., & Ni, J. (2003). Watchdog Agent — an
infotronics-based prognostics approach for product performance
degradation assessment and prediction, Vol. 17, pp. 109–125.
doi:10.1016/j.aei.2004.07.005.
Donoho, D. L. (1995). De-noising by soft-thresholding. IEEE Transactions
on Information Theory, Vol. 41, No. 3, pp. 613–627.
doi:10.1109/18.382009.
Dorigo, M., Birattari, M., & Stutzle, T. (2006). Ant colony optimization.
IEEE Computational Intelligence Magazine, Vol. 1, No. 4, pp. 28–39.
doi:10.1109/MCI.2006.329691.

186
 Reference

Dorigo, M., Maniezzo, V., & Colorni, A. (1996). Ant system: optimization
by a colony of cooperating agents. IEEE transactions on systems, man,
and cybernetics. Part B, Cyberneticsࣟ: a publication of the IEEE
Systems, Man, and Cybernetics Society, Vol. 26, No. 1, pp. 29–41.
doi:10.1109/3477.484436.
Dorigo, M., & Stützle, T. (2004). Ant colony optimization. Cambridge, Mass:
MIT Press.
Dragomir, O. E., Gouriveau, R., Zerhount, N., & Dragomir, F. (2007).
Framework for a distributed and hybrid prognostic system. In B.
Octavian (Ed.), 4th IFAC Conference on Management and Control of
Production and Logistics (2007), pp. 431–436. doi:10.3182/20070927-
4-RO-3905.00072.
Du, M., Cai, J., Liu, L., & Chen, P. (2011). ARRs based sensor placement
optimization for fault diagnosis. Procedia Engineering, Vol. 16, pp.
42–47. doi:10.1016/j.proeng.2011.08.1049.
Eberhart, R. C., & Shi, Y. (2000). Comparing inertia weights and
constriction factors in particle swarm optimization. Proceedings of the
2000 Congress on Evolutionary Computation. CEC00 (Cat.
No.00TH8512), Vol. 1, pp. 84–88. IEEE.
doi:10.1109/CEC.2000.870279.
Eisenmann, R. C. S., & Eisenmann, R. C. J. (1998). Machinery Malfunction
Diagnosis and Correction. Englewood Cliffs, NJ: Prentice-Hall.
Eklöv, T., Mårtensson, P., & Lundström, I. (1997). Enhanced selectivity of
MOSFET gas sensors by systematical analysis of transient parameters.
Analytica Chimica Acta, Vol. 353, No. 2-3, pp. 291–300.
doi:10.1016/S0003-2670(97)87788-4.
El-Abd, M., & Kamel, M. S. (2006). A hierarchical cooperative Particle
Swarm Optimizer. Proc. of Swarm Intelligence Symposium, pp. 43–47.
EN 13306: 2001 Maintenance Terminology, European Standard (2001).
CEN (European Committee for Standardization), Brussels.
Espinosa, J., Vandewalle, J., & Wertz, V. (2005). Fuzzy Logic,
Identification and Predictive Control. London: Springer-Verlag.
Estrin, D., Govindan, R., Heidemann, J., & Kumar, S. (1999). Next century
challenges. Proceedings of the 5th annual ACM/IEEE international
conference on Mobile computing and networking - MobiCom ’99, pp.
263–270. New York, New York, USA: ACM Press.
doi:10.1145/313451.313556.

187
 Reference

Ewins, D. J. (1995). Modal testing: theory and practice. Letchworth:

Research Studies Press.
Faulds, A. L., & King, B. B. (2000). Sensor Location in Feedback Control
of Partial Differential Equation Systems. Proceedings of the 2000
IEEE International Conference on Control Applications, pp. 536–541.
Firmo, H. T., & Legey, L. F. L. (2002). Generation expansion planning: an
iterative genetic algorithm approach. IEEE Transactions on Power
Systems, Vol. 17, No. 3, pp. 901–906.
doi:10.1109/TPWRS.2002.801036.
Friedman, J. H. (2001). Greedy Function Approximation: A Gradient
Boosting Machine. The Annals of Statistics, Vol. 29, No. 5, pp. 1189–
1232. doi:10.1214/aos/1013203451.
Friedman, J. H. (2002). Stochastic gradient boosting. Computational
Statistics & Data Analysis, Vol. 38, No. 4, pp. 367–378.
doi:10.1016/S0167-9473(01)00065-2.
Frisch, K. von. (1967). The dance language and orientation of bees.
Cambridge: Harvard University Press.
Fu, X., Zhang, Y., & Zhu, Y. (2011). Rolling bearing fault diagnosis
approach based on case-based reasoning. Journal of Xi’an Jiaotong
University, Vol. 45, No. 11, pp. 79–84.
Garcia, M. C., Sanz-Bobi, M. a., & del Pico, J. (2006). SIMAP: Intelligent
System for Predictive Maintenance Application to the Health
Condition Monitoring of a Wind Turbine Gearbox. Computers in
Industry, Vol. 57, No. 6, pp. 552–568.
doi:10.1016/j.compind.2006.02.011.
Garey, M. R., & Johnson, D. S. (1979). Computers and Intractability: a
Guide to the Theory of NP-completeness. San Francisco: W. H.
Freeman.
Garfinkel, S., & Holtzman, H. (2005). Understanding RFID Technology in:
Applications, Security and Privacy, pp. 15–36. Addison-Wesley.
Gibson, T. D., Prosser, O., Hulbert, J. N., Marshall, R. W., Corcoran, P.,
Lowery, P., Ruck-Keene, E. a., et al. (1997). Detection and
simultaneous identification of microorganisms from headspace samples
using an electronic nose. Sensors and Actuators B: Chemical, Vol. 44,
No. 1-3, pp. 413–422. doi:10.1016/S0925-4005(97)00235-9.
Giraud, C., & Jouvencel, B. (1995). Sensor selection: a geometrical
approach. Proceedings 1995 IEEE/RSJ International Conference on
Intelligent Robots and Systems. Human Robot Interaction and

188
 Reference

Cooperative Robots, Vol. 2, pp. 555–560. IEEE Comput. Soc. Press.

doi:10.1109/IROS.1995.526271.
Goldberg, D. E. (1989). Genetic Algorithms in search, optimization, and
machine learning. Addison-Wesley Publishing Corporation, Inc.
Goumas, S., Zervakis, M., Pouliezos, A., & Stavrakakis, G. S. (2001).
Intelligent on-line quality control of washing machines using discrete
wavelet analysis features and likelihood classification. Engineering
Applications of Artificial Intelligence, Vol. 14, No. 5, pp. 655–666.
doi:10.1016/S0952-1976(01)00028-8.
Grall, A., Beƍrenguer, C., & Chu, C. (2008). Optimal dynamic
inspection/replacement planning in condition-based maintenance.
Proceedings of the European Safety and Reliability Conference, pp.
381–388.
Grant, P. W., Harris, P. M., & Moseley, L. G. (1996). Fault diagnosis for
industrial printers using case-based reasoning. Engineering
Applications of Artificial Intelligence, Vol. 9, No. 2, pp. 163–173.
doi:10.1016/0952-1976(96)00009-7.
Guardia, M. (1995). Biochemical sensors: The state of the art. Mikrochimica
Acta, Vol. 120, No. 1-4, pp. 243–255. doi:10.1007/BF01244435.
Gupta, A., & Lawsirirat, C. (2006). Strategically optimum maintenance of
monitoring-enabled multi-component systems using continuous-time
jump deterioration models. Journal of Quality in Maintenance
Engineering, Vol. 12, No. 3, pp. 306–329.
doi:10.1108/13552510610685138.
Gutierrez-Osuna, R., Nagle, H. T., & Schiffman, S. S. (1999). Transient
response analysis of an electronic nose using multi-exponential models.
Sensors and Actuators B: Chemical, Vol. 61, No. 1-3, pp. 170–182.
doi:10.1016/S0925-4005(99)00290-7.
Gutierrez-Osuna, R., Nagle, T., Kermani, B., & Schiffman, S. (2003).
Signal Conditioning and Pre-Processing. Handbook of Machine
Olfaction: Electronic Nose Technology, pp. 1–50. Wiley-VCH.
Guyon, I., & Elisseef, A. (2006). An Introduction to Feature Extraction. In I.
Guyon, M. Nikravesh, S. Gunn, & L. Zadeh (Eds.), Feature Extraction,
pp. 1–25. Berlin/Heidelberg: Springer.
GWEC. (2013). Global Wind Statistics 2012, pp. 1–4.
Halim, D., & Moheimani, R. S. O. (2003). An optimization approach to
optimal placement of collocated piezoelectric actuators and sensors on

189
 Reference

a thin plate. Mechatronics, Vol. 13, No. 1, pp. 27–47.

doi:10.1016/S0957-4158(01)00079-4.
Hambaba, A., & Huff, E. (2000). Multiresolution error detection on early
fatigue cracks in gears. 2000 IEEE Aerospace Conference. Proceedings
(Cat. No.00TH8484), Vol. 6, pp. 367–372. IEEE.
doi:10.1109/AERO.2000.877912.
Hameed, Z., Hong, Y. S., Cho, Y. M., Ahn, S. H., & Song, C. K. (2009).
Condition monitoring and fault detection of wind turbines and related
algorithms: A review. Renewable and Sustainable Energy Reviews,
Vol. 13, No. 1, pp. 1–39. doi:10.1016/j.rser.2007.05.008.
Han, Y., & Song, Y. H. (2003). Condition monitoring techniques for
electrical equipment-a literature survey. IEEE Transactions on Power
Delivery, Vol. 18, No. 1, pp. 4–13. doi:10.1109/TPWRD.2002.801425.
Hand, D. J., Mannila, H., & Smyth, P. (2001). Principles of Data Mining.
London: MIT Press & IEEE.
Hansen, L. K., & Salamon, P. (1990). Neural network ensembles. IEEE
Transactions on Pattern Analysis and Machine Intelligence, Vol. 12,
No. 10, pp. 993–1001. doi:10.1109/34.58871.
Hansen, M. O. L. (2007). Aerodynamics of Wind Turbines, 2nd ed., p. 208.
London, UK: Earthscan.
Hayes, J. P. (1971). A Nand Model ror Fault Diagnosis in Combinational
Logic Networks. IEEE Transactions on Computers, Vol. C-20, No. 12,
pp. 1496–1506. doi:10.1109/T-C.1971.223162.
He, S., Wen, J. Y., Prempain, E., Wu, Q. H., Fitch, J., & Mann, S. (2004).
An improved particle swarm optimization for optimal power flow.
2004 International Conference on Power System Technology, 2004.
PowerCon 2004., Vol. 2, pp. 1633–1637. IEEE.
doi:10.1109/ICPST.2004.1460265.
He, S.-G., He, Z., & Wang, G. A. (2011). Online monitoring and fault
identification of mean shifts in bivariate processes using decision tree
learning techniques. Journal of Intelligent Manufacturing.
doi:10.1007/s10845-011-0533-5.
Hjorth, B. (1970). EEG analysis based on time domain properties.
Electroencephalography and Clinical Neurophysiology, Vol. 29, No. 3,
pp. 306–310. doi:10.1016/0013-4694(70)90143-4.
Holmberg, K., Adgar, A., Arnaiz, A., Jantunen, E., Mascolo, J., & Mekid, S.
(2010). E-maintenance, pp. 1–241. London: Springer.

190
 Reference

Hook, T. G., Hughes, E. A., Levline, R. E., Morgan, T. A., & Parker, L. M.
(1987). Application of reliability-centered maintenance to San Onofre
Units 2 & 3 auxiliary feed water system, EPRI NP-5430.
Hu, J., Zhang, L., Ma, L., & Liang, W. (2011). An integrated safety
prognosis model for complex system based on dynamic Bayesian
network and ant colony algorithm. Expert Systems with Applications,
Vol. 38, No. 3, pp. 1431–1446. doi:10.1016/j.eswa.2010.07.050.
Hu, S., Zhou, C., & Hu, W. (2000). A New Structure Fault Detection
Method Based on Wavelet Singularity. Journal of Applied Sciences,
Vol. 18, No. 3, pp. 198–201.
Huang, C. J., Lin, C. E., & Huang, C. L. (1992). Fuzzy approach for
generator maintenance scheduling. Electric Power Systems Research,
Vol. 24, No. 1, pp. 31–38. doi:10.1016/0378-7796(92)90042-Y.
Huang, S. (1998). A genetic-evolved fuzzy system for maintenance
scheduling of generating units. International Journal of Electrical
Power & Energy Systems, Vol. 20, No. 3, pp. 191–195.
doi:http://dx.doi.org/10.1016/S0142-0615(97)00080-X.
Huang, Y., Mcmurran, R., & Jones, D. R. P. (2008). Probability based
vehicle fault diagnosisௗ: Bayesian network method, pp. 301–311.
doi:10.1007/s10845-008-0083-7.
Ilie-zudor, E., Kemény, Z., Egri, P., & Monostori, L. (2006). The RFID
Technology and Its Current Applications. Proceeding of The Modern
Information Technology in the Innovation Process of the Industrial
Enterprise-MITIP, pp. 29–36.
Intanagonwiwat, C., Govindan, R., & Estrin, D. (2000). Directed Diffusionௗ:
A Scalable and Robust Communication Paradigm for Sensor Networks.
Proceedings of the ACM Mobi- Com’00, pp. 56–67. Boston.
Isermann, R. (2005). Model-based fault-detection and diagnosis – status and
applications. Annual Reviews in Control, Vol. 29, No. 1, pp. 71–85.
doi:10.1016/j.arcontrol.2004.12.002.
Jain, A. K., Murty, M. N., & Flynn, P. J. (1999). Data clustering: a review.
ACM Computing Surveys, Vol. 31, No. 3, pp. 264–323.
doi:10.1145/331499.331504.
Jansen, A., & Niyogi, P. (2005). A Geometric Perspective on Speech Sounds,
pp. 1–50.
Jardine, A. K. S., Lin, D., & Banjevic, D. (2006). A review on machinery
diagnostics and prognostics implementing condition-based

191
 Reference

maintenance. Mechanical Systems and Signal Processing, Vol. 20, No.

7, pp. 1483–1510. doi:10.1016/j.ymssp.2005.09.012.
Jolliffe, I. T. (2002). Principal Component Analysis, 1st ed. New York:
Springer Series in Statistics.
Kankar, P. K., Sharma, S. C., & Harsha, S. P. (2011). Rolling element
bearing fault diagnosis using wavelet transform. Neurocomputing, Vol.
74, No. 10, pp. 1638–1645. doi:10.1016/j.neucom.2011.01.021.
Kannan, S., Slochanal, S. M. R., Subbaraj, P., & Padhy, N. P. (2004).
Application of particle swarm optimization technique and its variants
to generation expansion planning problem. Electric Power Systems
Research, Vol. 70, No. 3, pp. 203–210. doi:10.1016/j.epsr.2003.12.009.
Kantardzic, M. (2003). Data Mining: Concepts, Models, Methods, and
Algorithm. New York: John Wiley & Sons, Inc.
Karaboga, D. (2005). An idea based on honey bee swarm for numerical
optimization.
Karaboga, D., & Akay, B. (2009). A comparative study of Artificial Bee
Colony algorithm. Applied Mathematics and Computation, Vol. 214,
No. 1, pp. 108–132. doi:10.1016/j.amc.2009.03.090.
Karaboga, D., & Basturk, B. (2007). A powerful and efficient algorithm for
numerical function optimization: artificial bee colony (ABC) algorithm.
Journal of Global Optimization, Vol. 39, No. 3, pp. 459–471.
doi:10.1007/s10898-007-9149-x.
Kasabov, N. (2001). Evolving fuzzy neural networks for
supervised/unsupervised online knowledge-based learning. IEEE
transactions on systems, man, and cybernetics. Part B, Cyberneticsࣟ: a
publication of the IEEE Systems, Man, and Cybernetics Society, Vol.
31, No. 6, pp. 902–18. doi:10.1109/3477.969494.
Kegg, R. . (1984). On-line machine and process diagnostics. Annals of the
CIRP, Vol. 32, No. 2, pp. 469–573.
Kennedy, J., & Eberhart, R. (2001). Swarm Intelligence. San Francisco:
Morgan Kaufmann Publishers, Inc.
Kennedy, J., & Eberhart, R. C. (1997). A discrete binary version of the
particle swarm algorithm. 1997 IEEE International Conference on
Systems, Man, and Cybernetics. Computational Cybernetics and
Simulation, Vol. 5, pp. 4104–4108. IEEE.
doi:10.1109/ICSMC.1997.637339.
Kermani, B. G., Schiffman, S. S., & Nagle, H. T. (1998). A novel method
for reducing the dimensionality in a sensor array. IEEE Transactions

192
 Reference

on Instrumentation and Measurement, Vol. 47, No. 3, pp. 728–741.

doi:10.1109/19.744338.
Kohavi, R., & John, G. H. (1997). Wrappers for feature subset selection.
Artificial Intelligence, Vol. 97, No. 1-2, pp. 273–324.
doi:10.1016/S0004-3702(97)00043-X.
Kohonen, T. (1982). Self-organized formation of topologically correct
feature maps. Biological Cybernetics, Vol. 43, pp. 59–66.
Krink, T., Vesterstrom, J. S., & Riget, J. (2002). Particle swarm
optimisation with spatial particle extension. Proceedings of the 2002
Congress on Evolutionary Computation. CEC’02 (Cat. No.02TH8600),
Vol. 2, pp. 1474–1479. IEEE. doi:10.1109/CEC.2002.1004460.
Kriofsky, T. ., & Kaplan, L. M. (1975). Inductively Coupled Transmitter-
Responder Arrangement. US.
Krishnan, S., & Rangayyan, R. M. (2000). Denoising knee joint vibration
signals using adaptive time-frequency representations. Engineering
Solutions for the Next Millennium. 1999 IEEE Canadian Conference
on Electrical and Computer Engineering (Cat. No.99TH8411), Vol. 3,
pp. 1495–1500. IEEE. doi:10.1109/CCECE.1999.804930.
Krohling, R. A. (2004). Gaussian swarm: a novel particle swarm
optimization algorithm. IEEE Conference on Cybernetics and
Intelligent Systems, 2004., Vol. 1, pp. 372–376. IEEE.
doi:10.1109/ICCIS.2004.1460443.
Krohling, R. A. (2005). Gaussian Particle Swarm with Jumps. 2005 IEEE
Congress on Evolutionary Computation, Vol. 2, pp. 1226–1231. IEEE.
doi:10.1109/CEC.2005.1554830.
Krohn, S., Morthorst, P.-E., & Awerbuch, S. (2007). The economics of wind
energy. Renewable and Sustainable Energy Reviews, Vol. 13, pp. 1–
156. Retrieved from
http://linkinghub.elsevier.com/retrieve/pii/S1364032108001299.
Kudo, T., & Matsumoto, Y. (2004). A Boosting Algorithm for
Classification of Semi-Structured Text. EMNLP, pp. 301–308.
Kusiak, A., & Verma, A. (2011). Prediction of Status Patterns of Wind
Turbines: A Data-Mining Approach. Journal of Solar Energy
Engineering, Vol. 133, No. 1, pp. 1–10. doi:10.1115/1.4003188.
Kusiak, A., & Zhang, Z. (2010). Analysis of Wind Turbine Vibrations
Based on SCADA Data. Journal of Solar Energy Engineering, Vol.
132, No. 3, pp. 1–12. doi:10.1115/1.4001461.

193
 Reference

Lai, L. L. (1998). Intelligent system application in power engineering:

evolutionary programming and neural networks. New York: Wiley,
John & Sons.
Landt, J. (2001). Shrouds of Time The history of RFID, pp. 1–11. Pittsburgh,
USA.
Landt, J. (2005). The history of RFID. IEEE Potentials, Vol. 24, No. 4, pp.
8–11. doi:10.1109/MP.2005.1549751.
Lee, I. (2011). Fault Diagnosis of Induction Motors Using Discrete Wavelet
Transform and Artificial Neural Network, pp. 510–514.
Lee, J., Ni, J., Djurdjanovic, D., Qiu, H., & Liao, H. (2006). Intelligent
prognostics tools and e-maintenance. Computers in Industry, Vol. 57,
No. 6, pp. 476–489. doi:10.1016/j.compind.2006.02.014.
Lee, K. Y., & Yang, F. F. (1998). Optimal reactive power planning using
evolutionary algorithms: a comparative study for evolutionary
programming, evolutionary strategy, genetic algorithm, and linear
programming. IEEE Transactions on Power Systems, Vol. 13, No. 1,
pp. 101–108. doi:10.1109/59.651620.
Lewis, F. L. (2004). Wireless Sensor Networks. In D. J. Cook & S. K. Das
(Eds.), Smart Environments: Technologies, Protocols, and
Applications, pp. 1–18. New York: John Wiley. doi:10.1007/b117506.
Lewis, S. (2004). A Basic Introduction to RFID technology and Its Use in
the Supply Chain, pp. 1–30.
Li, D., Wang, W., & Ismail, F. (2013). Enhanced fuzzy-filtered neural
networks for material fatigue prognosis. Applied Soft Computing, Vol.
13, No. 1, pp. 283–291. doi:10.1016/j.asoc.2012.08.031.
Li, G., Qin, Q., & Dong, C. (2000). Monitor the Optimal Placement Position
of the Sensors in System by Genetic Algorithm. Engineering
Mechanics, Vol. 17, No. 1, pp. 25–34.
Li, R., Sopon, P., & He, D. (2009). Fault features extraction for bearing
prognostics. Journal of Intelligent Manufacturing, Vol. 23, No. 2, pp.
313–321. doi:10.1007/s10845-009-0353-z.
Li, X., Ge, S. S., Pan, Y., Hong, K., Zhang, Z., & Hu, X. (2011). Feature
extraction based on common spatial analysis for time domain
parameters. 2011 8th International Conference on Ubiquitous Robots
and Ambient Intelligence (URAI), pp. 377–382. IEEE.
doi:10.1109/URAI.2011.6145846.
Li, X., Qu, L., Wen, G., & Li, C. (2003). Application of wavelet packet
analysis for fault detection in electro-mechanical systems based on

194
 Reference

torsional vibration measurement. Mechanical Systems and Signal

Processing, Vol. 17, No. 6, pp. 1219–1235.
doi:10.1006/mssp.2002.1517.
Li, Z., Wu, Z., He, Y., & Fulei, C. (2005). Hidden Markov model-based
fault diagnostics method in speed-up and speed-down process for
rotating machinery. Mechanical Systems and Signal Processing, Vol.
19, No. 2, pp. 329–339. doi:10.1016/j.ymssp.2004.01.001.
Liang, J. J., Qin, A. K., Suganthan, P. N., & Baskar, S. (2006).
Comprehensive learning particle swarm optimizer for global
optimization of multimodal functions. IEEE Transactions on
Evolutionary Computation, Vol. 10, No. 3, pp. 281–295.
doi:10.1109/TEVC.2005.857610.
Lienhart, W., & Brunner, F. K. (2003). MONITORING OF BRIDGE
DEFORMATIONS USING EMBEDDED FIBER OPTICAL
SENSORS. Proceedings of11th FIG Symposium on Deformation
Measurements, pp. 1–7. Santorini.
Lin, J., & Qu, L. (2000). Feature extraction based on Morlet wavelet and its
application for mechanical diagnosis. Journal of Sound and Vibration,
Vol. 234, No. 1, pp. 135–148. doi:10.1006/jsvi.2000.2864.
Liu, G. P. (2001). Nonlinear identification and control: a neural network
approach. International Journal of Adaptive Control and Signal
Processing, Vol. 21. London: Springer. doi:10.1002/acs.953.
Liu, H., & Abraham, A. (2005). Fuzzy adaptive turbulent particle swarm
optimization. Fifth International Conference on Hybrid Intelligent
Systems (HIS’05), p. 6 pp. IEEE. doi:10.1109/ICHIS.2005.49.
Liu, W., Gao, W., Sun, Y., & Xu, M. (2008). Optimal sensor placement for
spatial lattice structure based on genetic algorithms. Journal of Sound
and Vibration, Vol. 317, No. 1-2, pp. 175–189.
doi:10.1016/j.jsv.2008.03.026.
Liu, Y., Guo, L., Wang, Q., An, G., Guo, M., & Lian, H. (2010).
Application to induction motor faults diagnosis of the amplitude
recovery method combined with FFT. Mechanical Systems and Signal
Processing, Vol. 24, No. 8, pp. 2961–2971.
doi:10.1016/j.ymssp.2010.03.008.
Llobet, E., Brezmes, J., Vilanova, X., Sueiras, J. E., & Correig, X. (1997).
Qualitative and quantitative analysis of volatile organic compounds
using transient and steady-state responses of a thick-film tin oxide gas
sensor array. Sensors and Actuators B: Chemical, Vol. 41, No. 1-3, pp.
13–21. doi:10.1016/S0925-4005(97)80272-9.

195
 Reference

Loucks, D. P., van Beek, E., Stedinger, J. R., Dijkman, J. P. M., & Villars,
M. T. (2005). Water Resources Systems Planning and Management:
An Introduction to Methods, Models and Applications, p. 690. Paris:
UNESCO.
Lovbjerg, M., & Krink, T. (2002). Extending particle swarm optimisers with
self-organized criticality. Proceedings of the 2002 Congress on
Evolutionary Computation. CEC’02 (Cat. No.02TH8600), Vol. 2, pp.
1588–1593. IEEE. doi:10.1109/CEC.2002.1004479.
Mallat, S. G. (1989). A theory for multiresolution signal decomposition: the
wavelet representation. IEEE Transactions on Pattern Analysis and
Machine Intelligence, Vol. 11, No. 7, pp. 674–693.
doi:10.1109/34.192463.
Markard, J., & Petersen, R. (2009). The offshore trend: Structural changes
in the wind power sector. Energy Policy, Vol. 37, No. 9, pp. 3545–
3556. doi:10.1016/j.enpol.2009.04.015.
Markou, M., & Singh, S. (2003). Novelty detection: a review–part 2: neural
network based approaches. Signal Processing, Vol. 83, No. 12, pp.
2499–2521. doi:10.1016/j.sigpro.2003.07.019.
Marquez, A. C. (2007). The Maintenance Management Framework, p. 340.
Marseguerra, M., Zio, E., & Podofillini, L. (2002). Condition-based
maintenance optimization by means of genetic algorithms and Monte
Carlo simulation. Reliability Engineering & System Safety, Vol. 77, No.
2, pp. 151–165. doi:10.1016/S0951-8320(02)00043-1.
Marwala, T. (2012). Condition Monitoring Using Computational
Intelligence Methods. London: Springer London.
Marwali, M. K. C., & Shahidehpour, S. M. (2000). Coordination between
long-term and short-term generation scheduling with network
constraints. IEEE Transactions on Power Systems, Vol. 15, No. 3, pp.
1161–1167. doi:10.1109/59.871749.
Marzi, H. (2004). Real-time fault detection and isolation in industrial
machines using learning vector quantization. Proceedings of the
Institution of Mechanical Engineers, Part B: Journal of Engineering
Manufacture, Vol. 218, No. 8, pp. 949–959.
doi:10.1243/0954405041486109.
May, R., Dandy, G., & Maier, H. (2011). Review of Input Variable
Selection Methods for Artificial Neural Networks. In K. Suzuki (Ed.),
Artificial Neural Networks - Methodological Advances and Biomedical
Applications. InTech. doi:10.5772/16004.

196
 Reference

Mazomenos, E. B., Chen, T., Acharyya, A., Bhattacharya, A., Rosengarten,

J., & Maharatna, K. (2012). A Time-Domain Morphology and Gradient
based algorithm for ECG feature extraction. 2012 IEEE International
Conference on Industrial Technology, pp. 117–122. IEEE.
doi:10.1109/ICIT.2012.6209924.
McCulloch, W. S., & Pitts, W. (1943). A logical calculus of the ideas
immanent in nervous activity. The Bulletin of Mathematical Biophysics,
Vol. 5, No. 4, pp. 115–133. doi:10.1007/BF02478259.
Mendonqa, L. F. (2006). Fault Isolation Using Fuzzy Model-based
Observers. IFAC Fault Detection, Supervision and Safety of technical
processes, pp. 735–740.
Meng, X., & Meng, X. (2010). Nonlinear System Simulation Based on the
BP Neural Network. 2010 Third International Conference on
Intelligent Networks and Intelligent Systems, pp. 334–337. IEEE.
doi:10.1109/ICINIS.2010.159.
Menon, S., Schoess, J. N., Hamza, R., & Busch, D. (2000). Wavelet-Based
Acoustic Emission Detection Method with Adaptive Thresholding. In
V. V. Varadan, N. M. Wereley, R. O. Claus, Y. Bar-Cohen, S.-C. Liu,
T. T. Hyde, V. K. Varadan, et al. (Eds.), Proc. SPIE 3986, Smart
Structures and Materials 2000: Sensory Phenomena and Measurement
Instrumentation for Smart Structures and Materials, Vol. 3986, pp. 71–
77. doi:10.1117/12.388093.
Miranda, A. A., Borgne, Y.-A., & Bontempi, G. (2008). New Routes from
Minimal Approximation Error to Principal Components. Neural
Processing Letters, Vol. 27, No. 3, pp. 197–207.
Miranda, V., & Fonseca, N. (2002). New evolutionary particle swarm
algorithm (EPSO) applied to voltage/VAR control. Proc. 14th Power
Syst. Comput. Conf.
Miranda, V., Srinivasan, D., & Proença, L. (1998). Evolutionary
computation in power systems. International Journal of Electrical
Power & Energy Systems, Vol. 20, No. 2, pp. 89–98.
doi:10.1016/S0142-0615(97)00040-9.
Mobley, R. K. (1990). An Introduction to Predictive Maintenance. New
York: Van Nostrand Reinhold.
Mobley, R. K. (2002). An introduction to predictive maintenance, Second.,
pp. 1–438. New York: Elsevier Science.
Molter, A., da Silveira, O. a. A., Fonseca, J. S. O., & Bottega, V. (2010).
Simultaneous Piezoelectric Actuator and Sensor Placement
Optimization and Control Design of Manipulators with Flexible Links

197
 Reference

Using SDRE Method. Mathematical Problems in Engineering, Vol.

2010, pp. 1–23. doi:10.1155/2010/362437.
Momoh, J., & Button, R. (2003). Design and analysis of aerospace DC
arcing faults using fast fourier transformation and artificial neural
network. 2003 IEEE Power Engineering Society General Meeting
(IEEE Cat. No.03CH37491), pp. 788–793. IEEE.
doi:10.1109/PES.2003.1270407.
Naimimohasses, R., Barnett, D. M., Green, D. a, & Smith, P. R. (1995).
Sensor optimization using neural network sensitivity measures.
Measurement Science and Technology, Vol. 6, No. 9, pp. 1291–1300.
doi:10.1088/0957-0233/6/9/008.
Nakajima, S. (1988). Introduction to TPM: Total Productive Maintenance,
pp. 1–129. Portland: Productivity Press.
Nakamoto, T., Iguchi, A., & Moriizumi, T. (2000). Vapor supply method in
odor sensing system and analysis of transient sensor responses. Sensors
and Actuators B: Chemical, Vol. 71, No. 3, pp. 155–160.
doi:10.1016/S0925-4005(99)00186-0.
Nassim, L. (2011). Support Vector Machines for Fault Detection in Wind
Turbines. In B. Sergio (Ed.), Proceedings of the 18th IFAC World
Congress, 2011, pp. 7067–7072. Milano, Italy. doi:10.3182/20110828-
6-IT-1002.02560.
Negnevitsky, M., & Kelareva, G. (1999). Application of genetic algorithms
for maintenance scheduling in power systems. ICONIP’99. ANZIIS'99
& ANNES'99 & ACNN'99. 6th International Conference on Neural
Information Processing. Proceedings (Cat. No.99EX378), Vol. 2, pp.
447–452. IEEE. doi:10.1109/ICONIP.1999.845636.
Niu, G., Yang, B.-S., & Pecht, M. (2010). Development of an optimized
condition-based maintenance system by data fusion and reliability-
centered maintenance. Reliability Engineering & System Safety, Vol.
95, No. 7, pp. 786–796. doi:10.1016/j.ress.2010.02.016.
Obermaier, B., Guger, C., Neuper, C., & Pfurtscheller, G. (2001). Hidden
Markov models for online classification of single trial EEG data.
Pattern Recognition Letters, Vol. 22, No. 12, pp. 1299–1309.
doi:10.1016/S0167-8655(01)00075-7.
Oja, E. (2002). Unsupervised Learning in Neural Computation. Theoretical
Computer Science, Vol. 287, No. 1, pp. 187–207.
Okamura, S. (2011). The Short Time Fourier Transform and Local Signals.
Carnegie Mellon University.

198
 Reference

Orosz, J. E., & Jacobson, S. H. (2002). Analysis of Static Simulated

Annealing Algorithms 1. Journal of Optimization theory and
Applications, Vol. 115, No. 1, pp. 165–182.
doi:10.1023/A:1019633214895.
Pan, H., Wei, X., & Xu, X. (2010). Research of optimal placement of
gearbox sensor based on particle swarm optimization. 2010 8th IEEE
International Conference on Industrial Informatics, pp. 108–113. IEEE.
doi:10.1109/INDIN.2010.5549454.
Pan, J. J., Yang, Q., Chang, H., & Yeung, D. (2001). A Manifold
Regularization Approach to Calibration Reduction for Sensor-Network
Based Tracking. Proceedings of the Twenty-First National Conference
on Arti¿cial Intelligence, pp. 988–993. Boston.
Papadimitriou, C. (2004). Optimal sensor placement methodology for
parametric identification of structural systems. Journal of Sound and
Vibration, Vol. 278, No. 4-5, pp. 923–947.
doi:10.1016/j.jsv.2003.10.063.
Parsopoulos, K. E., & Vrahatis, M. N. (2002). Recent approaches to global
optimization problems through Particle Swarm Optimization. Natural
Computing, Vol. 1, pp. 235–306. doi:10.1023/A:1016568309421.
Parsopoulos, K. E., & Vrahatis, M. N. (2004). On the Computation of All
Global Minimizers Through Particle Swarm Optimization. IEEE
Transactions on Evolutionary Computation, Vol. 8, No. 3, pp. 211–224.
doi:10.1109/TEVC.2004.826076.
Patil, P. B., & Chavan, M. S. (2012). A wavelet based method for denoising
of biomedical signal. International Conference on Pattern Recognition,
Informatics and Medical Engineering (PRIME-2012), pp. 278–283.
IEEE. doi:10.1109/ICPRIME.2012.6208358.
Patterson, D. W. R., & Hughes, J. G. (1997). Case-based reasoning for fault
diagnosis. New Review of Applied Expert Systems and Emerging
Technologies, Vol. 3.
Pearson, K. (1901). On lines and planes of closest fit to systems of points in
space. Philosophical Magazine, Vol. 2, No. 6, pp. 559–572.
Peng, Z., He, Y., Chen, Z., & Chu, F. (2002). Identification of the Shaft
Orbit for Rotating Machines Using Wavelet Modulus Maxima.
Mechanical Systems and Signal Processing, Vol. 16, No. 4, pp. 623–
635. doi:10.1006/mssp.2002.1494.
Pennacchi, P., & Vania, A. (2008). Diagnostics of a crack in a load coupling
of a gas turbine using the machine model and the analysis of the shaft

199
 Reference

vibrations. Mechanical Systems and Signal Processing, Vol. 22, No. 5,

pp. 1157–1178. doi:10.1016/j.ymssp.2007.10.005.
Pereira, C. M. N. a., Lapa, C. M. F., Mol, A. C. a., & da Luz, A. F. (2010).
A Particle Swarm Optimization (PSO) approach for non-periodic
preventive maintenance scheduling programming. Progress in Nuclear
Energy, Vol. 52, No. 8, pp. 710–714.
doi:10.1016/j.pnucene.2010.04.009.
Perla, H. F. et al. (1984). A Guide for Developing Preventive Maintenance
Programs in Electric Power Plants, EPRI NP-3416.
Pham, D. T., Ghanbarzadeh, A., Koç, E., Otri, S., Rahim, S., & Zaidi, M.
(2006). The Bees Algorithm – A Novel Tool for Complex Optimisation
Problems. Proceeding of 2nd International Virtual Conference on
Intelligent Production Machines and Systems, pp. 454–459.
Pinar Pérez, J. M., García Márquez, F. P., Tobias, A., & Papaelias, M.
(2013). Wind turbine reliability analysis. Renewable and Sustainable
Energy Reviews, Vol. 23, pp. 463–472. doi:10.1016/j.rser.2013.03.018.
Piñeyro, J., Klempnow, A., & Lescano, V. (2000). Effectiveness of new
spectral tools in the anomaly detection of rolling element bearings.
Journal of Alloys and Compounds, Vol. 310, No. 1-2, pp. 276–279.
doi:10.1016/S0925-8388(00)00964-6.
Portnoff, M. R. (1980). Time-Frequency Representation of . Digital Signals.
IEEE Transactions on Acoustics, Speech and Signal Processing ASSP,
Vol. 28, No. 1, pp. 55–69.
Prabhakar, S., Mohanty, a. R., & Sekhar, a. . (2002). Application of discrete
wavelet transform for detection of ball bearing race faults. Tribology
International, Vol. 35, No. 12, pp. 793–800. doi:10.1016/S0301-
679X(02)00063-4.
Rai, V. K., & Mohanty, A. R. (2007). Bearing fault diagnosis using FFT of
intrinsic mode functions in Hilbert–Huang transform. Mechanical
Systems and Signal Processing, Vol. 21, No. 6, pp. 2607–2615.
doi:10.1016/j.ymssp.2006.12.004.
Ramani, S., Blu, T., & Unser, M. (2008). Blind optimization of algorithm
parameters for signal denoising by Monte-Carlo Sure. 2008 IEEE
International Conference on Acoustics, Speech and Signal Processing,
pp. 905–908. IEEE. doi:10.1109/ICASSP.2008.4517757.
Reisbec, C. K., & Schank, R. C. (1989). Inside Case-Based Reasoning.
Hillsdale, NJ: Lawrence Earlbaum Associates.

200
 Reference

Roussel, S., Forsberg, G., Steinmetz, V., Grenier, P., & Bellon-Maurel, V.
(1998). Optimisation of electronic nose measurements. Part I:
Methodology of output feature selection. Journal of Food Engineering,
Vol. 37, No. 2, pp. 207–222. doi:10.1016/S0260-8774(98)00081-8.
Ruhanne, A., Hanhikorpi, M., Bertuccelli, F., Colonna, A., Malik, W.,
Ranasinghe, D., López, T. S., et al. (2008). Sensor-enabled RFID Tag
Handbook, pp. 1–47.
Rumelhart, D. E., Hinton, G. E., & Williams, R. J. (1986). Learning Internal
Representations by Error Propagation. In D. E. Rumenhart & J. L.
McCelland (Eds.), Parallel Distributed Processing: Explorations in the
Microstructure of Cognition, pp. 318–362. Cambridge: MIT Press.
Saravanan, N., Kumar Siddabattuni, V. N. S., & Ramachandran, K. I.
(2008). A comparative study on classification of features by SVM and
PSVM extracted using Morlet wavelet for fault diagnosis of spur bevel
gear box. Expert Systems with Applications, Vol. 35, No. 3, pp. 1351–
1366. doi:10.1016/j.eswa.2007.08.026.
Saravanan, N., & Ramachandran, K. I. (2010). Incipient gear box fault
diagnosis using discrete wavelet transform ( DWT ) for feature
extraction and classification using artificial neural network ( ANN ).
Expert Systems With Applications, Vol. 37, No. 6, pp. 4168–4181.
doi:10.1016/j.eswa.2009.11.006.
Satoh, T., & Nara, K. (1991). Maintenance scheduling by using simulated
annealing method (for power plants). IEEE Transactions on Power
Systems, Vol. 6, No. 2, pp. 850–857. doi:10.1109/59.76735.
Saxena, a., & Vachtsevanos, G. (2005). A methodology for analyzing
vibration data from planetary gear systems using complex morlet
wavelets. Proceedings of the 2005, American Control Conference,
2005., Vol. 2, pp. 4730–4735. IEEE. doi:10.1109/ACC.2005.1470743.
Schölkopf, B., Burges, C. J. C., & Smola, A. J. (1999). Advances in Kernel
Methods: Support Vector Learning. Cambridge, MA: The MIT Press.
Schubert, U., Kruger, U., Arellano-Garcia, H., de Sá Feital, T., & Wozny, G.
(2011). Unified model-based fault diagnosis for three industrial
application studies. Control Engineering Practice, Vol. 19, No. 5, pp.
479–490. doi:10.1016/j.conengprac.2011.01.009.
Secrest, B. R., & Lamont, G. B. (2002). Visualizing particle swarm
optimization - Gaussian particle swarm optimization. Proceedings of
the 2003 IEEE Swarm Intelligence Symposium. SIS’03 (Cat.
No.03EX706), pp. 198–204. IEEE. doi:10.1109/SIS.2003.1202268.

201
 Reference

Seker, S., & Ayaz, E. (2003). Feature extraction related to bearing damage
in electric motors by wavelet analysis. Journal of the Franklin Institute,
Vol. 340, No. 2, pp. 125–134. doi:10.1016/S0016-0032(03)00015-2.
Shakhnarovich, G., Darrell, T., & Indyk, P. (2005). Nearest-Neighbor
Methods in Learning and Visionࣟ: Theory and Practice. Cambridge,
MA: MIT Press.
Shenoy, D. B., & Bhadbury, B. (1998). Maintenance resources management:
Adapting MRP. London: Taylor & Francis.
Shi, Y., & Eberhart, R. C. (2001). Fuzzy adaptive particle swarm
optimization. Proceedings of the 2001 Congress on Evolutionary
Computation (IEEE Cat. No.01TH8546), Vol. 1, pp. 101–106. IEEE.
doi:10.1109/CEC.2001.934377.
Shibata, K., Takahashi, A., & Shirai, T. (2000). Fault Diagnostics of
Rotating Machinery Through Visualization of Sound Signals.
Mechanical Systems and Signal Processing, Vol. 14, No. 2, pp. 229–
241. doi:10.1006/mssp.1999.1255.
Shinde, A. D. (2004). A Wavelet Packet Based Sifting Process and Its
Application for Structural Health Monitoring. Worcester Polytechnic
Institute.
Si, X., Wang, W., Hu, C., & Zhou, D. (2011). Remaining useful life
estimation – A review on the statistical data driven approaches.
European Journal of Operational Research, Vol. 213, No. 1, pp. 1–14.
doi:10.1016/j.ejor.2010.11.018.
Siegelmann, H. T., & Sontag, E. D. (1994). Analog computation via neural
networks. Theoretical Computer Science, Vol. 131, No. 2, pp. 331–360.
doi:10.1016/0304-3975(94)90178-3.
Silva, C. W. De. (1989). Control sensors and actuators. Englwood Cliff, NJ:
Prentice Hall.
Sindhwani, V., Niyogi, P., & Belkin, M. (2005). Beyond the point cloud:
from transductive to semi-supervised learning. Proceeding of
ICML ’05, pp. 824–831. New York.
Smola, A. J., & Scholkopf, B. (2003). A Tutorial on Support Vector
Regression. Statistics and Computing, Vol. 14, No. 3, pp. 199–222.
Soman, K. P., & Ramachandran, K. I. (2005). Insight into Wavelets from
Theory to Practice, 2nd ed., p. 404. India: PHI Learning Pvt. Ltd.
Soman, R. R., Davidson, E. M., McArthur, S. D. J., Fletcher, J. E., &
Ericsen, T. (2012). Model-based methodology using modified sneak

202
 Reference

circuit analysis for power electronic converter fault diagnosis. IET

Power Electronics, Vol. 5, No. 6, pp. 813.
Speybroeck, N. (2012). Classification and regression trees. International
journal of public health, Vol. 57, No. 1, pp. 243–6.
doi:10.1007/s00038-011-0315-z.
Spiewak, S. a., Duggirala, R., & Barnett, K. (2000). Predictive Monitoring
and Control of the Cold Extrusion Process. CIRP Annals -
Manufacturing Technology, Vol. 49, No. 1, pp. 383–386.
doi:10.1016/S0007-8506(07)62970-9.
SS-EN 13306 Maintenance Terminology (2001). Swedish Standards
Institute.
Staszewski, W. J. (1997). Vibration data compression with optimal wavelet
coefficients. Second International Conference on Genetic Algorithms
in Engineering Systems, Vol. 1997, pp. 186–190. IEE.
doi:10.1049/cp:19971178.
Staszewski, W. J. (2002). Intelligent signal processing for damage detection
in composite materials. Composites Science and Technology, Vol. 62,
No. 7-8, pp. 941–950. doi:10.1016/S0266-3538(02)00008-8.
Staszewski, W. J., Worden, K., & Tomlinson, G. R. (1997). Time-frequency
Analysis Gearbox Fault Detection Using the Wigner Ville Distribution
and Pattern Recognition. Mechanical Systems and Signal Processing,
Vol. 11, No. 5, pp. 673–692. doi:10.1006/mssp.1997.0102.
Steinwart, I., & Christmann, A. (2008). Support Vector Machines. New
York: Springer-Verlag.
Stockman, H. (1948). Communication by Means of Reflected Power.
Proceedings of the IRE, Vol. 36, No. 10, pp. 1196–1204.
doi:10.1109/JRPROC.1948.226245.
Stuart, A., & Allocca, A. (1984). Transducers: theory and Applications.
Englwood Cliff, NJ: Prentice-Hall.
Sullivan, G. P., Pugh, R., Melendez, A. P., & Hunt, W. D. (2010).
Operations & Maintenance Best Practices Release 3.0: A Guide to
Achieving Operational Efficiency, pp. 1–321.
Sun, H.-C., Huang, Y.-C., & Huang, C.-M. (2012). Fault Diagnosis of
Power Transformers Using Computational Intelligence: A Review.
Energy Procedia, Vol. 14, pp. 1226–1231.
doi:10.1016/j.egypro.2011.12.1080.
Sun, X., Li, H., & Ou, J. (2008). Improved Genetic Algorithm of Dynamic
Detection Measurement Point Optimization of Large Bridge and its

203
 Reference

Application. Architectural and Science Journal of Xian, Vol. 38, No. 5,

pp. 624–628.
Sutoh, T., Suzuki, H., & Nagai, N. (1994). Large scale generator
maintenance scheduling using simulated evolution. Proceedings of the
International Conference on Intelligent System Application to Power
Systems. Montpellier.
Tan, P.-N., Steinbach, M., & Kumar, V. (2006). Introduction to Data
Mining. Boston, MA: Pearson Education/Addison Wesley.
Teodorovic, D., & Dell’Orco, M. (2005). Bee colony optimization – a
cooperative learning approach to complex transportation problems.
Proceeding of 10th EWGT Meeting and 16th Mini-EURO Conference,
pp. 51–60.
Tian, D., & Li, N. (2009). Fuzzy Particle Swarm Optimization Algorithm.
2009 International Joint Conference on Artificial Intelligence, pp.
263–267. doi:10.1109/JCAI.2009.50.
Tian, X., Cao, Y. P., & Chen, S. (2011). Process fault prognosis using a
fuzzy-adaptive unscented Kalman predictor. International Journal of
Adaptive Control and Signal Processing, Vol. 25, No. 9, pp. 813–830.
doi:10.1002/acs.1243.
Triki, E., Collette, Y., & Siarry, P. (2005). A theoretical study on the
behavior of simulated annealing leading to a new cooling schedule.
European Journal of Operational Research, Vol. 166, No. 1, pp. 77–92.
doi:10.1016/j.ejor.2004.03.035.
Tsai, Y. T. (2009). Applying a case-based reasoning method for fault
diagnosis during maintenance. Proceedings of the Institution of
Mechanical Engineers, Part C: Journal of Mechanical Engineering
Science, Vol. 223, No. 10, pp. 2431–2441.
doi:10.1243/09544062JMES1588.
Tse, P. W., Yang, W., & Tam, H. Y. (2004). Machine fault diagnosis
through an effective exact wavelet analysis. Journal of Sound and
Vibration, Vol. 277, No. 4-5, pp. 1005–1024.
doi:10.1016/j.jsv.2003.09.031.
Vachtsevanos, G., Lewis, F., Roemer, M., Hess, A., & Wu, B. (2006).
Intelligent Fault Diagnosis and Prognosis for Engineering System, p.
454. New Jersey: John Wiley & Sons, Inc.
Vafaei, S., & Rahnejat, H. (2003). Indicated repeatable runout with wavelet
decomposition (IRR-WD) for effective determination of bearing-
induced vibration. Journal of Sound and Vibration, Vol. 260, No. 1, pp.
67–82. doi:10.1016/S0022-460X(02)00900-8.

204
 Reference

Valle, Y., Member, S., Venayagamoorthy, G. K., Member, S., & Harley, R.
G. (2008). Particle Swarm Optimizationௗ: Basic Concepts , Variants
and Applications in Power Systems. IEEE TRANSACTIONS ON
EVOLUTIONARY COMPUTATION, Vol. 12, No. 2, pp. 171–195.
Van den Kerkhof, P., Gins, G., Vanlaer, J., & Van Impe, J. F. M. (2012).
Dynamic model-based fault diagnosis for (bio)chemical batch
processes. Computers & Chemical Engineering, Vol. 40, pp. 12–21.
doi:10.1016/j.compchemeng.2012.01.013.
Vasudevan, R. (1985). Application of reliability-centered maintenance to
component cooling-water system at Turkey Point Units 3 and 4, EPRI
NP-4271.
Verma, A., & Kusiak, A. (2012). Fault Monitoring of Wind Turbine
Generator Brushes: A Data-Mining Approach. Journal of Solar Energy
Engineering, Vol. 134, No. 2, pp. 021001. doi:10.1115/1.4005624.
Walker, I. (1987). Development of a maintenance program. Proceedings of
the 14thInte-ram Conference. Toronto.
Wang, C., Kang, Y., Shen, P., Chang, Y., & Chung, Y. (2010). Applications
of fault diagnosis in rotating machinery by using time series analysis
with neural network. Expert Systems with Applications, Vol. 37, No. 2,
pp. 1696–1702. doi:10.1016/j.eswa.2009.06.089.
Wang, C., Zhang, Y., & Zhong, Z. (2008). Fault Diagnosis for Diesel Valve
Trains based on Time–frequency Images. Mechanical Systems and
Signal Processing, Vol. 22, No. 8, pp. 1981–1993.
doi:10.1016/j.ymssp.2008.01.016.
Wang, D. D. (1996). Computational intelligence based machine fault
diagnosis. Proceedings of the IEEE International Conference on
Industrial Technology (ICIT’96), pp. 465–469. IEEE.
doi:10.1109/ICIT.1996.601632.
Wang, H., Song, Z., & Wang, H. (2002). Statistical process monitoring
using improved PCA with optimized sensor locations. Journal of
Process Control, Vol. 12, pp. 735–744.
doi:http://dx.doi.org/10.1016/S0959-1524(01)00048-8.
Wang, K. (2002). Intelligent Condition Monitoring and Diagnosis Systems.
Amsterdam: IOS Press.
Wang, K. (2005). Applied Computational Intelligence in Intelligent
Manufacturing Systems. Australia: Advanced Knowledge International
Pty Ltd.

205
 Reference

Wang, K., & Zhang, Z. (2010). Intelligent Fault Diagnosis and Prognosis
systems (IFDPS) for Condition-based Maintenance, pp. 1–21.
Trondheim.
Wang, K., & Zhang, Z. (2012). Application of Radio Frequency
Identification (RFID) to Manufacturing, pp. 1–24. Trondheim.
Wang, Y., & Handschin, E. (2000). A new genetic algorithm for preventive
unit maintenance scheduling of power systems. International Journal
of Electrical Power & Energy Systems, Vol. 22, No. 5, pp. 343–348.
doi:10.1016/S0142-0615(99)00062-9.
Watson, I., & Marir, F. (2009). Case-based reasoning: A review. The
Knowledge Engineering Review, Vol. 9, No. 04, pp. 327.
doi:10.1017/S0269888900007098.
Wei, L., & Keogh, E. (2006). Semi-supervised time series classification.
Proceedings of the 12th ACM SIGKDD international conference on
Knowledge discovery and data mining - KDD ’06, p. 748. New York,
New York, USA: ACM Press. doi:10.1145/1150402.1150498.
Wei, X., & Pan, H. (2010). Particle Swarm Optimization and Intelligent
Fault Diagnosis. Beijing: National Defence Industry Press.
White, J., Kauer, J. S., Dickinson, T. a, & Walt, D. R. (1996). Rapid analyte
recognition in a device based on optical sensors and the olfactory
system. Analytical chemistry, Vol. 68, No. 13, pp. 2191–202.
doi:10.1021/ac9511197.
White, R. M. (1987). A sensor classification scheme. IEEE transactions on
ultrasonics, ferroelectrics, and frequency control, Vol. 34, No. 2, pp.
124–6.
Wilson, A. (2002). Asset Maintenance Management: A Guide to Developing
Strategy and Improving Performance. New York: Industrial Press, Inc.
Wilson, D. M., & DeWeerth, S. P. (1995). Odor discrimination using
steady-state and transient characteristics of tin-oxide sensors. Sensors
and Actuators B: Chemical, Vol. 28, No. 2, pp. 123–128.
doi:10.1016/0925-4005(95)80036-0.
Wireman, T. (1990). World Class Maintenance Management. New York:
Industrial Press.
Worden, K., & Burrows, A. P. (2001). Optimal sensor placement for fault
detection. Engineering Structures, Vol. 23, pp. 885–901.
doi:http://dx.doi.org/10.1016/j.bbr.2011.03.031.
Wu, J., & Chen, J.-C. (2006). Continuous wavelet transform technique for
fault signal diagnosis of internal combustion engines. NDT & E

206
 Reference

International, Vol. 39, No. 4, pp. 304–311.

doi:10.1016/j.ndteint.2005.09.002.
Wu, J., & Kuo, J.-M. (2009). An automotive generator fault diagnosis
system using discrete wavelet transform and artificial neural network.
Expert Systems with Applications, Vol. 36, No. 6, pp. 9776–9783.
doi:10.1016/j.eswa.2009.02.027.
Wu, J., & Liu, C.-H. (2009). An expert system for fault diagnosis in internal
combustion engines using wavelet packet transform and neural
network. Expert Systems with Applications, Vol. 36, No. 3, pp. 4278–
4286. doi:10.1016/j.eswa.2008.03.008.
Wu, S., & Chow, T. W. S. (2004). Induction Machine Fault Detection Using
SOM-Based RBF Neural Networks. IEEE Transactions on Industrial
Electronics, Vol. 51, No. 1, pp. 183–194.
doi:10.1109/TIE.2003.821897.
Xie, X., Zhang, W., & Yang, Z. (2002). Dissipative particle swarm
optimization. Proceedings of the 2002 Congress on Evolutionary
Computation. CEC’02 (Cat. No.02TH8600), Vol. 2, pp. 1456–1461.
IEEE. doi:10.1109/CEC.2002.1004457.
Xin, W., Xiao-yun, C., Ke-ju, X., Rong-min, Z., Qing-jun, P., & Lin, L.
(2012). Ice-coated AC transmission lines tension prognosis with
autoregressive model. 2012 3rd International Conference on System
Science, Engineering Design and Manufacturing Informatization, Vol.
2, pp. 168–170. IEEE. doi:10.1109/ICSSEM.2012.6340835.
Yang, H., & Liao, C. (2001). A de-noising scheme for enhancing wavelet-
based power quality monitoring system. IEEE Transactions on Power
Delivery, Vol. 16, No. 3, pp. 353–360. doi:10.1109/61.924810.
Yare, Y., & Venayagamoorthy, G. K. (2010). Optimal maintenance
scheduling of generators using multiple swarms-MDPSO framework.
Engineering Applications of Artificial Intelligence, Vol. 23, No. 6, pp.
895–910. doi:10.1016/j.engappai.2010.05.006.
Yare, Y., Venayagamoorthy, G. K., & Aliyu, U. O. (2008). Optimal
generator maintenance scheduling using a modified discrete PSO. IET
Generation, Transmission & Distribution, Vol. 2, No. 6, pp. 834.
doi:10.1049/iet-gtd:20080030.
Yellen, J., Al-Khamis, T. M., Vemuri, S., & Lemonidis, L. (1992). A
decomposition approach to unit maintenance scheduling. IEEE
Transactions on Power Systems, Vol. 7, No. 2, pp. 726–733.
doi:10.1109/59.141779.

207
 Reference

Yen, G. G., & Lin, K. (1999). Conditional health monitoring using vibration
signatures. Proceedings of the 38th IEEE Conference on Decision and
Control (Cat. No.99CH36304), Vol. 5, pp. 4493–4498. IEEE.
doi:10.1109/CDC.1999.833249.
Yen, G. G., & Lin, K. (2000). Wavelet Packet Feature Extraction for
Vibration Monitoring, Vol. 47, No. 3, pp. 650–667.
Yoshimoto, K., Yasuda, K., Yokoyama, R., & Cory, B. J. (1993).
Decentralized Hopfield neural network applied to maintenance
scheduling of generating units in power systems. Third International
Conference on Artificial Neural Networks, 1993.,, pp. 277–281.
Brighton.
Yu, D., Yang, Y., & Cheng, J. (2007). Application of Time–frequency
Entropy Method based on Hilbert–Huang Transform to Gear Fault
Diagnosis. Measurement, Vol. 40, No. 9-10, pp. 823–830.
doi:10.1016/j.measurement.2007.03.004.
Yuan, J. (2012). Manifold Assumption and Semi-Supervised Learning for
Fault Diagnosis. Data Mining for Zero-Defect Manufacturing, pp. 133–
148. Trondheim: Tapir Academic Press.
Zaher, A., McArthur, S. D. J., Infield, D. G., & Patel, Y. (2009). Online
wind turbine fault detection through automated SCADA data analysis.
Wind Energy, Vol. 12, No. 6, pp. 574–593. doi:10.1002/we.319.
Zhang, Z., & Kusiak, A. (2012). Monitoring Wind Turbine Vibration Based
on SCADA Data. Journal of Solar Energy Engineering, Vol. 134, No.
2, pp. 021004. doi:10.1115/1.4005753.
Zhang, Z., & Wang, K. (2010). Application of Improved Discrete Particle
Swarm Optimization (IDPSO) in Generating Unit Maintenance
Scheduling. In K. Wang, O. Myklebust, & D. Tu (Eds.), International
Workshop of Advanced Manufacturing and Automation (IWAMA2010),
pp. 79–86. Shanghai: Tapir Academic Press.
Zhang, Z., & Wang, K. (2011). Fault Isolation Using Self-organizing Map
(SOM) ANNs. IET International Conference of Wireless Mobile &
Computing, pp. 425–431. Shanghai: Institute Engineering and
Technology.
Zhang, Z., & Wang, K. (2013). Dynamic Condition-Based Maintenance
Scheduling Using Bee Colony Algorithm (BCA). In E. Qi, J. Shen, &
R. Dou (Eds.), Proceedings of International Asia Conference on
Industrial Engineering and Management Innovation (IEMI2012), pp.
1607–1618. Berlin, Heidelberg: Springer Berlin Heidelberg.
doi:10.1007/978-3-642-38445-5_169.

208
 Reference

Zhang, Z., Wang, Y., & Wang, K. (2012). Fault diagnosis and prognosis
using wavelet packet decomposition, Fourier transform and artificial
neural network. Journal of Intelligent Manufacturing.
doi:10.1007/s10845-012-0657-2.
Zhang, Z., Wang, Y., & Wang, K. (2013). Intelligent fault diagnosis and
prognosis approach for rotating machinery integrating wavelet
transform, principal component analysis, and artificial neural networks.
The International Journal of Advanced Manufacturing Technology,
Vol. 68, No. 1-4, pp. 763–773. doi:10.1007/s00170-013-4797-0.
Zheng, H., Li, Z., & Chen, X. (2002). Gear Fault Diagnosis Based on
Continuous Wavelet Transform. Mechanical Systems and Signal
Processing, Vol. 16, No. 2-3, pp. 447–457.
doi:10.1006/mssp.2002.1482.
Zheng, Y., Tay, D. B. H., & Li, L. (2000). Signal extraction and power
spectrum estimation using wavelet transform scale space filtering and
Bayes shrinkage. Signal Processing, Vol. 80, No. 8, pp. 1535–1549.
doi:10.1016/S0165-1684(00)00054-2.
Zou, M., Dayan, J., & Green, I. (2000). Dynamic simulation and monitoring
of a non-contacting flexibly mounted rotor mechanical face seal.
Proceedings of the Institution of Mechanical Engineers, Part C:
Journal of Mechanical Engineering Science, Vol. 214, No. 9, pp.
1195–1206. doi:10.1243/0954406001523632.

209