Machine fault diagnosis and condition prognosis using adaptive neuro fuzzy inference system and classification anf regression trees

Thesis for the Degree of Doctor of Philosophy Machine Fault Diagnosis and Condition Prognosis using Adaptive Neuro-Fuzzy Inference System and Classification and Regression Trees by Van

Thesis for the Degree of Doctor of Philosophy

Machine Fault Diagnosis and Condition

Prognosis using Adaptive Neuro-Fuzzy Inference System and Classification and Regression Trees

by Van Tung Tran Department of Mechanical Engineering

The Graduate School Pukyong National University

February 2009

Machine Fault Diagnosis and Condition

Prognosis using Adaptive Neuro-Fuzzy Inference System and Classification and Regression Trees

Advisor: Prof Bo-Suk Yang

by

Van Tung Tran

A thesis submitted in partial fulfillment of the requirements

for the degree of

Doctor of Philosophy

in the Department of Mechanical Engineering, The Graduate School,

Pukyong National University

February 2009

List of Figures v

List of Tables viii

List of Symbols ix

Abstract

I Introduction 1

1 Background 1

2 Motivation of This Research 6

3 Research Objectives 6

4 Tools and Approaches 7

5 Scientific Contribution of This Research 7

6 Organization of Thesis 8

References 9

II The State-of-The-Art of Machine Fault Diagnosis and Prognosis 11

1.Machine Fault Diagnosis 11

1.1 Model-based approaches 11

1.2 Knowledge-based approaches 13

1.3 Pattern recognition-based approaches 15

2 Machine Fault Prognosis 19

2.1 Statistical approaches 20

2.2 Model-based approaches 21

2.3 Data-driven based approaches 22

References 22

III Background Knowledge 36

1 Feature-Based Diagnosis and Prognosis: a Review 36

1.1 Feature extraction techniques 37

1.2 Feature selection techniques 39

2 Feature Representation 40

2.1 Features in time domain 40

2.1.1 Cumulants 40

2.1.2 Upper and lower bound histogram 44

2.1.3 Entropy estimation and error 45

2.1.4 Auto-regression coefficients 45

2.2 Feature in frequency domain 46

2.2.1 Fourier transform 46

2.2.2 Spectral analysis 47

2.2.3 Frequency parameter indices 48

3 Classification and Regression Trees (CART) 49

3.1 Introduction 49

3.2 Tree growing 50

3.2.1 Classification tree 50

3.2.2 Regression tree 52

3.3 Tree pruning 54

3.3.1 Classification tree 54

3.3.2 Regression tree 55

3.4 Cross-validation for selecting the best tree 56

4 Adaptive Neuro-Fuzzy Inference System (ANFIS) 57

4.1 Architecture of ANFIS 57

4.2 Learning algorithm of ANFIS 60

5 Conclusions 61

References 61

IV CART and ANFIS Based Fault Diagnosis for Induction Motors 67

1 Introduction 67

2 Induction Motor Faults 67

2.1 Bearing faults 70

2.2 Stator or armature faults 72

2.3 Broken rotor bar and end ring faults 74

2.4 Eccentricity related faults 75

3 The Proposed Fault Diagnosis System for Induction Motors 77

3.1 Experiment and data acquisition 79

3.2 Feature calculation 81

3.3 Feature selection and classification 83

4 Conclusion 90

References 91

V Machine Condition Prognosis 94

1 Introduction 94

2 Prediction Strategies 97

2.1 Recursive prediction strategy 97

2.2 DirRec prediction strategy 98

2.3 Direct prediction strategy 98

3 Time Delay Estimation 99

4 Determining Embedding Dimension 100

4.1 Cao’s method 100

4.2 False nearest neighbor method (FNN) 101

5 Proposed System for Machine Condition Prognosis 103

6 Experiment 105

7 Case Studies of Machine Condition Prognosis 108

7.1 Case study 1: CART and OS prediction 108

7.2 Case study 2: parallel CART and MS direct prediction 115

7.2.1 Parallel structure of CART 115

7.2.2 Results and discussions 116

7.3 Case study 3: ANFIS and MS direct prediction 124

8 Conclusions 130

References 132

VI Conclusions and Future Works 134

1 Conclusions 134

2 Future Works 135

Acknowledgements

List of Figures

Fig 1.1 System costs depending on type of maintenance strategy 3

Fig 1.2 Architecture of a CBM system 4

Fig 3.1 Histogram for bearing signal with different condition 44

Fig 3.2 Classification tree 51

Fig 3.3 Regression tree 53

Fig 3.4 Schematic of ANFIS architecture 58

Fig 4.1 View of a squirrel cage induction motor 68

Fig 4.2 Four types of rolling-element bearing misalignment 70

Fig 4.3 Bearing sizes marked 71

Fig 4.4 Proposed system for fault diagnosis 78

Fig 4.5 Test rig for experiment 79

Fig 4.6 Faults on the induction motors 80

Fig 4.7 Vibration and current signals of each fault condition 81

Fig 4.8 Decision tree of features obtained from vibration signal 84

Fig 4.9 Decision tree of features obtained from current signal 84

Fig 4.10 Topology of ANFIS architecture for vibration signals 85

Fig 4.11 The network RMS error convergence curve 86

Fig 4.12 Bell shaped membership functions for vibration signals 87

Fig 5.1 Hierarchy of prognostic approaches 96

Fig 5.2 Proposed system for machine fault prognosis 104

Fig 5.3 Low methane compressor: wet screw type 105

Fig 5.4 The entire of peak acceleration data of low methane compressor

106

Fig 5.5 The entire of envelope acceleration data of low methane compressor

107

Fig 5.6 The faults of main bearings of compressor 108

Fig 5.7 Training and validating results of peak acceleration data (the first 300 points) 109

Fig 5.8 Predicted results of peak acceleration data 110

Fig 5.9 Peak acceleration of low methane compressor 110

Fig 5.10 The values of E1 and E2 of peak acceleration data of low methane compressor 111

Fig 5.11 Training and validating results of peak acceleration data 112

Fig 5.12 Predicted results of peak acceleration data 112

Fig 5.13 Data trending of envelope acceleration of low methane compressor 113

Fig 5.14 The values of E1 and E2 of envelope acceleration data 114

Fig 5.15 Training and validating results of envelope acceleration data

114

Fig 5.16 Predicted results of envelope acceleration data 115

Fig 5.17 Architecture and input values for sub-model of parallel-structure of CART 116

Fig 5.18 Time delay estimation 117

Fig.5.19 The relationship between FNN percentage and embedding

dimension 118

Fig 5.20 Training and validating results of peak acceleration data 120

Fig 5.21 Training and validating results of envelop acceleration data 121

Fig 5.22 Predicted results of peak acceleration data 123

Fig 5.23 Predicted results of envelop acceleration data 124

Fig 5.24 Training and validating results of the ANFIS model for peak acceleration data 126

Fig 5.25 Training and validating results of the ANFIS model for envelope

acceleration data 126

Fig 5.26 RMSE convergent curve 127

Fig 5.27 The changes of MFs after learning 128

Fig 5.28 Predicted results of ANFIS model for peak acceleration data

129

Fig 5.29 Predicted results of the ANFIS model for envelope acceleration data 130

Fig 6.1 The general hybrid system 136

List of Tables

Table 3.1 Cumulants for bearing signal with different condition 42

Table 4.1 Comparison of detection technologies 69

Table 4.2 The description of faulty motors 80

Table 4.3 Feature parameters 82

Table 4.4 Descriptions of data sets 83

Table 4.5 The confusion matrix for CART-ANFIS of 800 epochs 88

Table 4.6 The value of statistical parameters 90

Table 5.1 Training set D for direct prediction strategy 99

Table 5.2 Information of the system 105

Table 5.3 The RMSEs of CART and parallel-structure CART 119

Table 5.4 The RMSEs of CART and ANFIS 130

E ⋅ The statistical expectation

E(t L) Impurity of the left branch node

E(t R) Impurity of the right branch node

f(t) Representation of a time signal

h i Columns of the histogram

h L Lower bound of histogram

h U Upper bound of histogram

L Learning sample

MSF Mean square frequency

n Order of the AR model

N Number of data points

R T Total within-node sum of squares

R(t L) Sum of squares of the left subset

R(t R) Sum of squares of the right subset

RVF Root variance frequency

t L Left branch nodes

t R Right branch nodes

T max Wholly expanded tree in the growing phase

T Set of current nodes of T

VF Variance frequency

V Mutually exclusive data sets

x abs Absolute value

x i Discrete time signals

α′ Complexity parameter for the geometric midpoint

∆E(s,t) Drop of impurity

ε t Residual

Chapter IV CART and ANFIS Based Fault Diagnosis for Induction Motors

f Fundamental supply frequency

f b Detectable broken bar frequency

f ec Frequency components of interest

f r Rotor rotational frequency

y i (d) The ith reconstructed vector

y n(i,d) (d) Nearest neighbor of y i (d)

Machine Fault Diagnosis and Condition Prognosis using Adaptive Neuro-Fuzzy Inference System and Classification and Regression

Trees

Van Tung Tran

Department of Mechanical Engineering, The Graduate School,

Pukyong National University

Abstract

Sustaining the productivity is a key strategy of manufacturers to exist on the drastic competition of global market In order to keep up the productivity, manufacturers need to reduce the manufacturing costs by using maintenance due

to its major part of the total costs of the manufacturing process Consequently, a good maintenance strategy plays a crucial role in the existence and development

of the organizations Additionally, in accompany with the fast development of technology, the equipment becomes more and more complex The traditional maintenance strategies such as corrective maintenance and prescheduled maintenance cannot guarantee the functional operation of equipments and are progressively replaced by intelligent maintenance strategies in which condition based maintenance is one of the delegates

Condition-based maintenance has been defined as maintenance actions which are based on actual conditions of equipments obtained from nondestructive inspections, operations and condition measurements This means that the equipment condition is accessed under operation for making conclusions whether that equipment will be failed and the effective maintenance actions are necessary

to avoid the consequences of that failure or not The use of condition-based maintenance systems ensures that the condition of equipment is always monitored and alarm limitations can be indicated if the condition exceeds predefined levels

In condition-based maintenance system, fault diagnosis and condition prognosis are crucial components which have been considerably received much attention from the community of researchers and maintainers Fault diagnosis is the ability

to detect fault, isolate the component which is failure, and decide on the potential impact of failed component on the health of the system; while condition prognosis

is defined as a capability to foretell the future states, predict the remaining useful life – the time left for the normal operation of machine before breakdowns occur

or machine condition reaches the critical failure value

In this study, classification and regression trees (CART) and adaptive neuro-fuzzy inference systems (ANFIS) will be developed as an effective intelligent system for performing machine fault diagnosis and condition prognosis CART is known as one of the illustrious techniques of the decision tree induction and used for the purpose of either classification or regression depending on the output variable which is categorical or numerical CART recursively partitions the entire data into binary descendant subsets which are as homogeneous as possible with respect to the response variables High effective computation and reliability are the remarkable advantages of this algorithm In the second technique, ANFIS

is an excellent integration of the adaptive capability of neural networks and the modeling human knowledge ability of fuzzy logic During the learning process, the parameters of fuzzy membership functions initially determined by experts are adapted to the relationship between the input and output That combination makes the ANFIS model more systematic and less dependent on the expert knowledge For implementing the fault diagnosis, CART and ANFIS are combined with another technique so-called feature-based technique This technique is one of the powerful techniques to represent the raw data as features which are

representatives of values indicating the machine condition By using features, the encountered problem in data transfer and data storage could be effortlessly solved Feature-based technique consists of data acquisition, data preprocessing, feature representation, feature extraction, feature selection and classifiers In the proposed system for fault diagnosis, CART is used as a feature selection tool to select pertinent features which can characterize the machine conditions from the whole feature set whilst ANFIS plays a role as a classifier In order to be evaluated, this system is applied to diagnose the faults of induction motor, which is an indispensable part in several industrial applications The high performance results indicate that this system offers a potential for machine fault diagnosis

Foretelling the future states of machine has become more and more significant

in modern industry It assists maintainers or system operators in monitoring, inspecting the machines’ operating conditions, and detecting the incipient faults

so that they could opportunely perform remedial actions to avoid the catastrophic failures Furthermore, it enables the scheduled maintenance to be more effective

In this study, the future machines’ operating conditions are predicted by using CART and ANFIS model in combination with time series techniques These time series techniques consist of methods which are utilized to determine the optimal observations and the steps ahead as the inputs and outputs of predictors, respectively The trending data of a low methane compressor is used to validate the proposed method The predicted results show that CART and ANFIS predictors are reliable and promising tools in machine condition prognosis

I Introduction

1 Background

A failure in equipment of production line results in not only the loss of productivity but also timely services to customer, and may even lead to safety and environmental problems which destroy the organization image This emphasizes the need of maintenance activities in manufacturing operations of organization Maintenance activities can sustain the reliability and availability of product equipment, improve the product quality, increase productivity, and undertake the safety requirements However, maintenance activities have been historically regarded as a necessary evil by the various management functions in an organization since maintenance costs form a large part of the total operating and production costs in capital-intensive industries According to [1], the maintenance cost for industrial companies in the USA has increased 10-15% per year since

1979 Depending on the specific industry, maintenance costs can represent between 10 and more than 40% of the costs of goods produced For example, maintenance costs as a percentage of total value-added costs could be 20-50% for mining, 15-25% for primary metal and 3-15% for processing and manufacturing industries [2] In a study of Knights and Oyanader [3], in open-pit mining operations in Chile, the world’s primary copper producer, maintenance costs were estimated at 44% of total mine production costs Similarly, 20-40% of these maintenance costs are related to the repair of major components Therefore, major system repair costs indirectly account for 9-18% of the total operating costs of an open-pit mining operation Also, for equipment manufacturers and distributors, system repair and maintenance costs are the major cause of liabilities through

warranty programs and repair and maintenance contracts, especially when taking into account that the use of rebuilt system is of growing interest to the industry, both on a rental or leasing basis The recent surveys of maintenance management effectiveness in US manufacturing industry indicate that one third of all maintenance costs is wasted as the result of unnecessary and improper maintenance activities Additionally, with the augment of mechanization and automation, many modern plants have installed flexible computer-controlled automatic and unmanned equipments, the maintenance costs have been increased substantially Consequently, an efficient and reasonable maintenance strategy is in need of implementing so that organization’s overall goals and objectives can be attained at minimal costs

Traditionally, maintenance strategies in industry are broadly classified as two categories, namely corrective maintenance and preventive maintenance [4, 5] Corrective maintenance, also known as breakdown maintenance, is carried out merely after the occurrence of an obvious functional failure, malfunction, or breakdown of equipment Its actions are able to restore the functional capabilities

of failed or malfunctioned equipment by either repairing or replacing the failed component Corrective maintenance is reactive approach to maintenance because the action is triggered by the unplanned event of an equipment failure Preventive maintenance involves scheduled maintenance and condition-based maintenance (CBM) [6] Scheduled maintenance is performed periodically at predetermined interval to prevent the functional failure by replacing critical components before the end of their expected useful lives CBM is a method used to reduce the uncertainty of maintenance activities, and is carried out according to the need indicated by equipment condition Unlike the strategies mentioned above, CBM does not normally involve in an intrusion into the equipment and actual preventive action is taken only when an incipient failure is believed to have been detected According to [7], the variation in costs with number of maintenance events is

depicted in Fig 1 With corrective maintenance, even though the maintenance cost is low, its total cost is still high due to the operational failures Conversely, preventative maintenance practice produces low operations costs; however more preventative actions produce greater maintenance costs Evidently, the condition-based plan is the most efficient, in terms of total operations costs

Fig 1 System costs depending on type of maintenance strategy [7]

A complete CBM system comprises a number of functional modules listed as follows [8-9]:

• Sensing and data acquisition

• Signal processing and feature extraction

• Production of alarms or alerts

• Failure or fault diagnosis and health assessment

• Prognosis: projection of health profiles to future health or estimation the remaining useful life (RUL)

• Decision reasoning: maintenance recommendations, or evaluation of asset readiness for a particular operational scenario

• Management and control of data flow or test sequences

• Management of historical data storage and historical data access

• System configuration management

• Human system interface

Passive/

Smart sensor

Data manipulation

processing Feature extraction Feature selection

Pre-Condition monitior

Thresshold Fuzzy logic

Health assessment

Component specific feature extraction Anomaly &

diagnosis

Health assessment

based prognosis Model- prognosis

Feature-Automatic decision reasoning

Data fusion Classifier Response generator

Computer interface

Human-Physical model

Mission Plants

Fig 2 Architecture of a CBM system [8]

Not all these modules are usually used in CBM system Normally, the CBM system consists of the modules as follows: sensing and data acquisition, signal processing, condition monitoring, fault diagnosis and health assessment, prognosis, decision support, and presentation as shown in Fig 2 According to [10], diagnosis is defined as an ability to detect fault or anomaly conditions, isolate which component in the system is faulty, and decide the potential impact

of a failing or failed component on the health of system In the industrial and manufacturing areas, prognosis is the capability to foretell the RUL of component The task of prognosis module is to monitor and track the time evolution of the fault Consequently, diagnosis and prognosis modules gradually become the major components of CBM and have much consideration to be developed

There are several benefits obtained from implementing the CBM system in general and diagnosis/prognosis in particular:

• Reduce maintenance costs

• Increase equipment reliability and availability

• Reduce equipment downtime

• Extend service life of the observed system

• Provide constant evaluation of the system condition

• Increase operational safety

• Reduce severity of faults

• Attempt to totally eliminate catastrophic failures

• Extend maintenance cycle

• Reduce technician training requirements

Due to these benefits and low maintenance cost, CBM system has been received great attention of researchers and practical maintainers This study addresses to develop an intelligent machine fault diagnosis and machine condition prognosis based on adaptive neuro-fuzzy inference systems (ANFIS) and classification and regression trees (CART) These techniques were initially introduced by Jang [11] and Breiman et al [12], respectively Both CART and ANFIS techniques and their extensive researches have been successfully used in several applications such as handwritten recognition, face detection, medical issues, fault diagnosis of components for the rotating machinery, etc However, these researches and published papers in the combination of these techniques for machine fault diagnosis are still not mentioned Especially, in machine prognosis aspect, the use of CART as well as ANFIS for predicting the RUL or the future states of machine condition are rarely appeared in literatures Therefore, the aim

of this research is to contribute the development of intelligent method based on CART and ANFIS to machine fault diagnosis and condition prognosis

2 Motivation of This Research

Industrial machines are vital for manufacturing operations They consist of many components that will be degraded the operating conditions and could be failed due to wear and fatigue during the operating process Sustaining their condition or avoiding their failures leading to the stoppage of machine is of vital importance to be taken into consideration Condition-based maintenance in which fault diagnosis and condition prognosis are the key components has been progressively received much interest for their potential benefits as mentioned in the previous section Hence, the development of intelligent system for machine fault diagnosis and condition prognosis based on ANFIS and CART is the motivation of this research

3 Research Objectives

From the necessity of the supplemented methods for machine diagnosis and condition prognosis as stated in the introduction section, this research objectives are to redevelop and combine the existed algorithms which are CART, ANFIS, and time-series techniques for obtaining better performances in classification and prediction process These objectives are detailed as follows:

• Developing preprocessing method of feature selection for reducing the computational complexity and obtaining the high accuracy in classification

• Developing the data-driven approaches for machine condition prognosis by using the forecasting model to both of short-term prediction and long-term prediction purposes

• Applying the redeveloped diagnosis system for machine fault diagnosis

• Implementing the developed prognosis system for foretelling the future operating conditions of machine

4 Tools and Approaches

In order to attain the objectives of this research, the following methods have been adopted:

• Combining the CART with ANFIS algorithm for classification in which the former is used as a feature selection tool to select the significant features from original feature set which is obtained from feature calculation procedure and the latter is employed as a classifier

• Applying the combined technique to diagnose the faults of induction motors

• Developing the prediction technique in which the CART and ANFIS predictors are integrated with time-series techniques for forecasting the future states of machine

• Applying the developed prediction system for prognosticating the future operating conditions of low methane compressor

5 Scientific Contributions of This Research

The development of intelligent fault diagnosis and condition prognosis systems using the CART and ANFIS algorithms is the main contribution of this research Other significant scientific contributions of this research are as follows:

• The ability to obtain the optimal features for fault classification using feature selection

• The developed system was successfully applied in real application to diagnose and detect the faults of induction motors based on vibration and current signals

• Several of prediction models are developed for machine condition prognosis

• The developed prognosis systems were used to short-term prediction and long-term prediction purposes for forecasting the operating condition of low methane compressor based on routine vibration data

6 Organization of This Thesis

This dissertation is organized as follows:

Chapter 1 introduces the necessity and motivation behind this research It also

briefly describes the main objectives and the contributions as well as the organization of this dissertation

Chapter 2 outlines the state-of-the-art approaches to the machine fault

diagnosis and condition prognosis

Chapter 3 reviews the literature on the use of feature extraction and feature

selection in fault diagnosis Furthermore, the basic theory of feature representation

as well as the fundamentals of CART and ANFIS is profoundly described

Chapter 4 presents the faults frequently occurred in induction motor and the

use of developed methods for diagnosing these faults

Chapter 5 addresses several problems encountered in machine condition

prognosis It firstly introduces the techniques involved in prognosis approaches In addition, the prediction strategies of time-series techniques are also represented in this chapter Furthermore, it describes the methods used for determining the embedding dimension and the time delay which are the observations and enable predicting steps, respectively Finally, it addresses the case study of proposed system in real condition prognosis

Chapter 6 states the conclusion based on the results obtained from previous

chapters This chapter also recommends some directions for further research in the future

on Mining Innovation, Santiago, 2004

[4] A.H.C Tsang, Condition-based maintenance: tools and decision making, Journal of Quality in Maintenance Engineering 1 (3) (1995) 3–17

[5] R.C.M Yam, P.W Tse, P Tu, Intelligent predictive decision support for condition-based maintenance, The International Journal of Advanced Manufacturing Technology 17 (5) (2001) 383-391

[6] M Bengtsson, Condition based maintenance system technology – where is development heading, Proceeding of the 17th European Maintenance Congress, Barcelona, 2004

[7] http://www.osacbm.org

[8] M Lebold, K Reichard, D Boylan, Utilizing DCOM in an open system architecture framework for machinery monitoring and diagnostics, Proceedings of IEEE Aerospace Conference Vol.3, 2003, pp 1227-1236 [9] M Lebold, and M Thurston, Open standards for condition-based maintenance and prognostic systems, Maintenance and Reliability Conference, 2001

[10] G Vachtsevanos, F Lewis, M Roemer, A Hess, B Wu, Intelligent fault diagnosis and prognosis for engineering systems, John Wiley & Son, Inc., New Jersey, 2006

[11] J.S.R Jang, ANFIS: Adaptive-network-based fuzzy inference system, IEEE

Transactions on System, Man and Cybernetics 23 (3) (1993) 665-685

[12] L Breiman, J.H Friedman, R.A Olshen, C.J Stone, Classification and regression trees, Chapman & Hall, 1984

II The State-of-The-Art of Machine Fault

Diagnosis and Prognosis

1 Machine Fault Diagnosis

Several methods have been proposed in order to solve the fault detection and fault diagnosis problems The most commonly employed solution approaches for fault diagnosis system include (a) model-based, (b) knowledge-based, and (c) pattern recognition-based approaches [1] Generally, analytical model-based methods can be designed in order to minimize the effect of unknown disturbance and perform the consistent sensitivity analysis Knowledge-based methods are used when there is a lot of experience but not enough details to develop accurate quantitative models Pattern recognition methods are applicable to a wide variety

of systems and exhibit real-time characteristics The applications of these methods

to machine fault diagnosis are reviewed as follows:

1.1 Model-based approaches

The model-based methods perform fault diagnosis relied on analytical redundancy in which the consistency between the measurements and expected behavior of the process is checked by analytical models These analytical models could be physical specific or explicit mathematical model of the monitored machine Based on this explicit model, residual generation methods such as Kalman filter, parameter estimation (or system identification), and parity relations are used to obtain signals - so called residuals which is indicative of fault presence

in the machine Finally, the residuals are evaluated to achieve fault detection, isolation and identification

Different approaches for fault detection using mathematical models have been developed [2-8] A variety of models have been examined including linear system models [9], graph models [10, 11], process models [12], component models [13], and behavioral models [14, 15] Furthermore, model-based methods have been successfully applied for diagnosing the faults of components of mechanical system such as gearboxes [16, 17], bearings [18-20], rotors [21, 22] and cutting tools [23] Bartelmus [24, 25] used mathematical model and computer simulation

as a tool for aiding signal processing and interpretation of gearbox diagnosis Hansen et al [26] proposed an approach to a more robust diagnosis of meshing gears based on the fusion of sensor-based and model-based information Vania and Pennacchi [27] developed some methods to measure the accuracy of the results obtained from model-based techniques aiming to identify faults of generator The information provided by these methods was shown to be very useful in having more precise fault identification along with evaluating the confidence of a diagnostic decision

Recently, model-based techniques for diagnosis have been examined in the artificial intelligence community under the title of model-based reasoning Furthermore, the use of artificial intelligent (AI) techniques or the combination of conventional techniques and AI techniques has greatly enhanced the efficiency of model-based approaches in fault diagnosis For example, Su and Chong [28] used neural network model as an analytical method in fault diagnosis of induction motor to avoid the ineffectiveness of traditional methods due to their non-adaptation with vibration signal Liang and Du [29] combined a model-based fault detection and diagnosis method with support vector machine technique and successfully used this integrated technique for diagnosing the fault of heating, ventilation and air conditioning systems

Generally, model-based approaches can be more effective if a correct and accurate model is built However, explicit mathematical models may not be

feasible for complex systems since it would be very difficult or even could not be established for such systems

1.2 Knowledge-based approaches

Knowledge-based system (KBS) or expert system (ES) for fault diagnosis is performed based upon the evaluation of on-line monitored data according to a rule set which is determined by expert knowledge This knowledge includes the locations of input and output process variables, patterns of abnormal process conditions, fault symptom, operational constraints, and performance criteria The operators and engineers’ intelligence related to the specific process systems can

be implemented into this approach Their knowledge can help to recognize the potential faults based upon previous experiences This approach can reduce the difficulties on exact numeric information and automates the human intelligence for process supervision

Although KBSs are expensive and time-consuming during evolvement, they are dedicated to applying them to machine fault diagnosis was reported in the literatures [30-35] In [34], an ES including adaptive order tracking technique and artificial neural networks (ANNs) for fault diagnosis in internal combustion engines is introduced This system consisted of two stages In the first stage, the engine sound emission signals were recorded and treated as the tracking of frequency-varying bandpass signals Then the sound energy diagram was utilized

to normalize the features and reduce computation quantity In the second stage, the probability neural network was used to train the signal features and engine fault conditions Similarly, an expert system using order tracking technique in combination with fuzzy logic inference was presented to fault diagnosis of scooter platform [35] The fuzzy-logic inference was used in this system for developing the diagnostic rules of the data base in the present fault diagnosis system

Data mining techniques are also applied to KBS as tools extracted diagnosis

knowledge from data base A decision tree, which is a method in data mining technology, was developed from the information gathered from the machine maintenance manual, electrical prints, and a human expert The information from the decision tree was then used to develop the rules that represent the knowledge base Yang et al [36] proposed an expert system namely VIBEX to aid plant operators in diagnosing the cause of abnormal vibration for rotating machinery Decision tree in their work was used as an acquisition of structured knowledge to obtain the diagnosing rules from decision table which is built by the cause-symptom matrix

Evolutionary algorithms (EAs) are also utilized for domain expert knowledge

in a computer program with an automated inference engine to perform reasoning for solving problem Three main reasoning methods for ES used in the area of machinery diagnostics are rule-based reasoning [37-39], case-based reasoning [40-41] and model-based reasoning [42] Another reasoning method, negative reasoning, was introduced to mechanical diagnosis by Hall et al [43] Unlike other reasoning methods, negative reasoning deals with negative information, of which absence or lack of symptoms is indicative of meaningful inferences Stanek

et al [44] compared case-based and model-based reasoning and then proposed a combination of these methods for a lower-cost solution of machine condition assessment and diagnosis

Other examples of diagnostic ESs were used for troubleshooting electrical faults in a shuttle car [45, 46] and hydraulic system [47, 48] in mining areas Petri nets, as a general purpose graphical tool for describing relations between conditions and events, have recently been applied to machine fault detection and diagnosis Propes [49] used a fuzzy Petri net to describe an operating mode transition and to detect a mode change event for fault detection of diagnostics of complex systems Yang [50] proposed a hybrid Petri-net modeling method coupled with fault tree analysis and Kalman filtering for early failure detection

and fault isolation Yang et al [51] introduced an approach for integrating case-based reasoning with Petri net for fault diagnosis of induction motors The integrated approach was shown to outperform the conventional case-based reasoning expert system

Compared with model-based approaches, knowledge-based approaches are particularly suitable for large industrial plants since those non-linear real plants are extremely difficult to be modeled and the linear approximation of the model results in large errors In addition, knowledge-based approaches are able to reduce the complexity when implementing the corresponding safety system and make it flexible and easy to understand and follow Combining knowledge-based fault diagnosis methods with real-time process variables monitoring will improve the efficiency and reliability of detecting fault behavior and overall effectiveness of the system

1.3 Pattern recognition-based approaches

Pattern recognition is a process of mapping the information obtained in the measurement space and/or features in the feature space to machine faults in the fault space Traditionally, pattern recognition is manually done by auxiliary graphical tools such as power spectrum graph, phase spectrum graph, cepstrum graph, autoregressive spectrum, spectrogram, wavelet phase graph, etc However, manual pattern recognition requires expertise in the specific area of diagnosis application Thus, highly trained and skilled personnel are necessary Therefore, automatic pattern recognition is exceedingly desirable This can be achieved by classification of signals based on the information and feature extracted from the signals

Pattern recognition approaches include AI techniques which have been increasingly applied to machine diagnosis and have shown superior performance over conventional approaches However, it is not easy to apply AI techniques in

practice due to the lack of efficient procedures for obtaining training data and specific knowledge, which are necessary for training the models Thus far, most

of the applications in the literatures just used experimental data for model training Popular AI techniques for machine diagnosis are ANNs, fuzzy logic systems, fuzzy-neural networks (FNNs), and neural-fuzzy systems A review of recent developments in applications of AI techniques for induction machine stator fault diagnostics was given in [52]

ANNs mimic the human brain structure which consists of simple arithmetic units connected to complex layer architecture They are capable of representing highly nonlinear functions and performing multi-input, multi-output mapping The ANN learns the unknown function by adjusting its weights with observations of input and output This process is usually called training process of an ANN There have been various neural network models applied for pattern recognition Feed forward neural network (FFNN) structure is the most widely used neural network structure in machine fault diagnosis [53-56] Multilayer perceptron with the back propagation (BP) training algorithm, which is a special FFNN, is also employed for pattern recognition and classification as well as machine fault diagnostics [57-59] However, the BP neural networks have two main limitations which are difficult to determine the network structure and the number of nodes; and slow convergence of the training process Cascade correlation neural network (CCNN) is alternative which does not require initial determination of the network structure and the number of nodes CCNN can be used in case that on-line training

is preferable CCNN was applied to bearing fault classification and showed that it can result in the minimum network structure for fault recognition with acceptable accuracy [60] Other neural network models including radial basis function neural networks, recurrent neural networks, and counter propagation neural networks were also applied in machine diagnostics [61-64]

The ANN models mentioned above usually use supervised learning algorithms

which require external input such as a priori knowledge about the target or desired output For example, a common practice of training a neural network model is to utilize a set of experimental data with known faults Conversely, unsupervised learning does not require external input The unsupervised neural network learns itself by using new available information The applications of unsupervised neural network for fault diagnosis were introduced in [65-67] In [65], faults of rotating machine were detected by using self-organizing map (SOM) and learning vector quantization Tallam et al [66] proposed some self-commissioning and on-line training algorithms for FFNN with particular application to electrical machine fault diagnostics Sohn et al [67] employed an auto associative neural network to separate effects of damages caused by the environmental and vibration variations

of the system Then a sequential probability ratio test was performed on the normalized features for damage classification

The combination of neural networks and other techniques would be a significant alternative to improve the performance of machine diagnosis For instance, Silva et al [68] used two neural networks, which are SOM and adaptive resonance theory (ART), combined with an ES based on Taylor’s tool life equation to classify tool wear state DePold and Gass [69] studied the applications

of neural networks and ESs in a modular intelligent and adaptive system in gas turbine diagnostics Yang et al [70] presented an approach for integrating case-based reasoning ES with an ART-Kohonen neural network to enhance fault diagnosis It showed that the proposed approach outperforms the self-organizing feature map-based system with respect to classification rate Garga et al [71] proposed a hybrid reasoning approach combining neural network, fuzzy logic and

ES to integrate domain knowledge and test operation data from the machine for machine diagnostics and prognostics

In condition monitoring practice, knowledge from domain specific experts is usually inaccurate and reasoning on knowledge is often imprecise Therefore,

measures of the uncertainties in knowledge and reasoning are required for ES to provide more robust problem solving Unremarkably used uncertainty measures are probability, fuzzy member functions in fuzzy logic theory and belief functions

in belief networks theory An example of applying fuzzy logic to machine fault classification was given in [72] to classify frequency spectra representing various rolling element bearing faults Du and Yeung [73] introduced an approach so- called fuzzy transition probability, which combines transition probability (Markov process) with the fuzzy set, to monitor progressive faults Fuzzy logic is also incorporated with other techniques such as neural networks and ES for fault diagnostic application For example, Zhang et al [74] developed an FNN for fault diagnosis of rotary machines to improve the recognition rate of pattern recognition, especially in the case that sample data are similar Lou and Loparo [75] employed an adaptive neural-fuzzy inference system as a diagnostic classifier for bearing fault diagnosis Liu et al [76] applied fuzzy logic and ESs to build a fuzzy ES for bearing fault detection Chang et al [77] built a system for decision-making support in a power plant using both rule-based ES and fuzzy logic Genetic algorithms (GAs), which are the most ordinarily used type of EA, have shown to have merits in applications to machine diagnostics Several examples of ANN incorporating GA and other EAs for machine fault classification and diagnostics are [78-80] A technique called support vector machine (SVM) is a new general machine learning tool based on the structural risk minimization principle It has received much consideration in recent times due to its high accuracy and good generalization capabilities The use of SVM and its extension for machine fault diagnosis were summarized in [81]

Adaptive neuro-fuzzy inference system (ANFIS) which is an integration of ANNs and fuzzy logic system has been widely used for automated detection and diagnosis of machine condition It takes advantage of good learning capability of ANN and human knowledge representation of fuzzy logic Successful

implementations of ANFIS in fault diagnosis were reported in [82-85] Shukri et

al [82] applied ANFIS for fault detection and diagnosis of three-phase induction motor Lou and Loparo [83] used combined method for the diagnosis of localized defects in ball bearings in which wavelet transform was used to process the accelerometer signals and to generate feature vectors; and ANFIS was trained and utilized as diagnostic classifier Tran et al [84] used the integration of CART and ANFIS for diagnosing the faults of induction motors In [85] and [86], multiple ANFIS algorithm and its combination with GAs were employed for detecting rotor bar breakage, air gap eccentricity faults of three-phase induction motor and rolling bearings, respectively Other applications of ANFIS for machine fault diagnosis were introduced in [87-90]

2 Machine Fault Prognosis

The literatures of prognosis are much smaller in comparison with those of fault diagnosis The most obvious and normally used prognosis is to use the given current and past machine condition to predict how much time is left before a failure occurs The time left before observing a failure is usually called remaining useful life (RUL) In order to predict the RUL, data of the fault propagation process and/or the data of the failure mechanism must be available The fault propagation process is usually tracked by a trending or forecasting model for certain condition variables There are two ways in describing the failure mechanism The first one assumes that failure only depends on the condition variables, which reflect the actual fault level, and the predetermined boundary The definition of failure is simply defined that the failure occurs when the fault reaches a predetermined level The second one builds a model for the failure mechanism using available historical data In this case, different definitions of failure can be defined as follows: (a) an event that the machine is operating at an

unsatisfactory level; or (b) it can be a functional failure when the machine cannot perform its intended function at all; or (c) it can be just a breakdown when the machine stops operating, etc The approaches to prognosis fall into three main categories: statistical approaches, model-based approaches, and data-driven based approaches

2.1 Statistical approaches

Statistical approaches, which are the simplest forms of prognosis techniques, collect statistical information from a large number of component samples to indicate the survival duration of component before failure occurs and uses these statistics to predict the RUL of individual components Yan et al [91] used a logistic regression model to calculate the probability of failure for given condition variables and an autoregressive moving average time series model to trend the condition variables for failure prediction Then a predetermined level of failure probability was used to estimate the RUL Phelps et al [92] proposed to track sensor-level test-failure probability vectors instead of the physical system or sensor parameters for prognosis A Kalman filter with an associated interacting multiple models was used to perform the tracking The proportional hazards models (PHM) and proportional intensity model (PIM), which are two statistical models in survival analysis, are useful tools for RUL estimation in combination with a trending model for the fault propagation process Banjevic and Jardine [93] discussed RUL estimation for a Markov failure time process which includes a joint model of PHM and Markov property for the covariate evolution as a special case Vlok et al [94] applied PIM with covariate extrapolation to estimate bearing residual life Hidden Markov model is also a powerful tool for RUL estimation [95, 96] Lin and Makis [97] introduced a partially observable continuous-discrete stochastic process model to describe the hidden evolution process of the machine state associated with the observation process Wang et al [98] proposed a

stochastic process with hazard rate for predicting of residual life The condition information was the expert judgment based on vibration analysis Wang [99] used the residual delay-time concept and stochastic filtering theory to derive the residual life distribution

2.2 Model-based approaches

Model-based prognostic approaches are applicable to where accurate mathematical models can be constructed from physical system These methods use residuals as features, which are the outcomes of consistency checks between the sensed measurements of system and the outputs of a mathematical model Ray and Tangirala [100] used a non-linear stochastic model of fatigue crack dynamics for real-time computation of the time-dependent damage rate and accumulation in mechanical structures Li et al [101, 102] introduced two defect propagation models via mechanistic modeling for RUL estimation of bearings Oppenheiner and Loparo [103] applied a physical model for predicting the machine condition

in combination with fault strengths to life model based on crack growth law to estimate RUL A general method was purposed by Chelidze and Cusumano [104] for tracking the evolution of hidden damage process in the situation that a slowly evolving damage process is coupled to a fast, directly observable dynamical system Some different approaches used model-based techniques for prognosis were proposed in [105-110] However, model-based techniques are merely applied for some specific components and each requires a different mathematical model Changes in structural dynamics and operating conditions can affect the mathematical model as it is impossible to model all real-life conditions Furthermore, it is difficult to establish the suitable model to mimic the real life

2.3 Data-driven based approaches

Data-driven techniques are also known as data mining techniques or machine learning techniques They utilize and require large amount of historical failure data to build a prognostic model that learns the system behavior Among these techniques, artificial intelligence was regularly used because of its flexibility in generating appropriate model Several of the existing approaches used ANNs to model the systems and estimate the RUL Zhang and Ganesan [111] used self-organizing neural networks for multivariable trending of the fault development to estimate the residual life of bearing system Wang and Vachtsevanos [112] proposed an architecture for prognosis applied to industrial chillers Their prognostic model included dynamic wavelet neural networks, reinforcement learning, and genetic algorithms This model was used to predict the failure growth of bearings based on the vibration signals SOM and back propagation neural networks (BPNN) methods using vibration signals to predict the RUL of ball bearing were applied by Huang et al in [113] Wang et al [114] utilized and compared the results of two predictors, namely recurrent neural networks and ANFIS, to forecast the damage propagation trend of rotating machinery In [115], Yam et al applied a recurrent neural network for predicting the machine condition trend Dong et al [116] employed a grey model and a BPNN to predict the machine condition Altogether, the data-driven techniques are the promising and effective techniques for machine condition prognosis

References

[1] V Venkatasubrsmanian, Towards integrated process supervision: current status and future directions, Proceedings of the IFAC International Conference on Computer Software Structures, Sweden, 1994, pp.1-13

[2] J Chen, R.J Patton, Robust model-based fault diagnosis for dynamic