support vector machines in fault diagnostics of electrical motors
-
Upload
nguyennguyet -
Category
Documents
-
view
224 -
download
2
Transcript of support vector machines in fault diagnostics of electrical motors
Helsinki University of Technology Control Engineering Laboratory Espoo 2002 Report 131
SUPPORT VECTOR MACHINES IN FAULT DIAGNOSTICS OF ELECTRICAL MOTORS Sanna Pöyhönen
TEKNILLINEN KORKEAKOULU TEKNISKA HÖGSKOLAN HELSINKI UNIVERSITY OF TECHNOLOGY TECHNISCHE UNIVERSITÄT HELSINKI UNIVERSITE DE TECHNOLOGIE D´HELSINKI
Helsinki University of Technology Control Engineering Laboratory Espoo September 2002 Report 131
SUPPORT VECTOR MACHINES IN FAULT DIAGNOSTICS OF ELECTRICAL MOTORS1 Sanna Pöyhönen Abstract: Continuous and trouble-free operation of electrical motors is an essential part of
modern power and production plants. Faults and failures of electrical machinery may cause
remarkable economical losses but also highly dangerous situations. In the industry, model-based
methods are still most common choice for condition monitoring of electrical machinery, but
during last decade also different kinds Artificial Intelligence (AI) based methods have established
a firm position.
Support Vector Machine (SVM) is a novel machine learning method introduced in early 90’s.
It has been successfully applied to numerous classification and pattern recognition problems,
and in many applications, SVM has shown to have better generalisation properties than
traditional classifiers. Also, efficiency of SVM based classification does not depend on the
number of features of classified entities, which makes it attractive for fault diagnostics
applications.
In this thesis, a SVM based fault classification scheme is designed for electrical machines.
SVM is used to classify power spectrum estimates of different variables of the motor based on
the motor condition. Also the fusion of outputs of 2-class SVM’s to find the global classification
decision is studied, and a so-called mixture matrix approach is found to be the most suitable
method. Further result is that there exist better fault indicators than stator line current, which is
currently widely used in fault diagnostics of electrical machines.
Keywords: fault diagnostics, electrical machine, support vector machine, multi-class classification
Helsinki University of Technology
Department of Automation and Systems Technology
Control Engineering Laboratory
1 The summary for a Licentiate Thesis work, 2002.
Distribution:
Helsinki University of Technology
Control Engineering Laboratory
P.O. Box 5400
FIN-02015 HUT, Finland
Tel. +358-9-451 5201
Fax. +358-9-451 5208
E-mail: [email protected]
http://www.control.hut.fi/
ISBN 951-22-6133-2
ISSN 0356-0872
Picaset Oy
Helsinki 2002
Preface This thesis was written in Control Engineering Laboratory, Helsinki University of Technology. I want to thank the head of the laboratory, Professor Heikki Koivo, for supervising the thesis, and providing relaxed and pleasant atmosphere to the laboratory. I am also deeply thankful to Professor Heikki Hyötyniemi for inspirational guidance throughout the work, and giving valuable insights to the world of machine learning and support vector machines. In addition to professors, I want to thank the whole personnel of Control Engineering Laboratory for being humorous and delightfully eccentric fellow workers. Further, I want to thank Antero Arkkio, Marian Negrea and Pedro Jover for successful co-operation in the research of fault diagnostics of electrical machines. The work has been financially supported by the National Technology Agency (TEKES), the Technology Promotion Foundation, the Finnish Cultural Foundation and the Neles Foundation, which are gratefully acknowledged. Finally, I want to thank my supportive family, and especially the Crazy Monkey for all the fun. Espoo, September 2002
Sanna Pöyhönen
List of Publications P1 Pöyhönen, S., Negrea, M., Arkkio, A., Hyötyniemi, H., Koivo, H.:”Support
Vector Classification for Fault Diagnostics of an Electrical Machine”, Proc. of 6th Int. Conf. on Signal Processing (ICSP’02), Vol.2, pp. 1719-1722, Beijing-China, August, 2002
P2 Pöyhönen, S., Negrea, M., Arkkio, A., Hyötyniemi, H., Koivo, H.: ”Fault
Diagnostics of an Electrical Machine with Multiple Support Vector Classifiers”, to be published in Proc. of The 17th IEEE Int. Symp. on Intelligent Control (ISIC'02), Vancouver, British Columbia, Canada, October 27-30, 2002
P3 Pöyhönen, S., Negrea, M., Arkkio, A., Hyötyniemi, H.: “Comparison of
Reconstruction Schemes of Multiple SVM’s Applied to Fault Classification of a Cage Induction Motor”, HUT, Control Eng. Lab., Report 130, 2002
P4 Pöyhönen, S., Negrea, M., Jover, P., Arkkio, A., Hyötyniemi, H.: ”Numerical
Magnetic Field Analysis and Signal Processing for Fault Diagnostics of Electrical Machines”, Conf. record of 15th Int. Conf. on Electrical Machines, paper 365 - on the CD, Bruges-Belgium, August, 2002
Author’s contribution to the publications: [P1]: Author has written the paper excluding Chapter III and some parts of the
introduction. Author has also constructed a SVM classification structure in MATLAB. 2nd and 3rd authors have provided the simulation data of an induction motor and written Chapter III. 4th and 5th authors have given valuable insights to the subject.
[P2]: Author has written the paper excluding Chapter IV and some parts of the
introduction. Author has also constructed a multi-classification structure and the noise filtering algorithm in MATLAB. 2nd and 3rd authors have provided the simulation data of an induction motor and written Chapter IV. 4th and 5th authors have given valuable insights to the subject.
[P3]: Author has written the paper excluding Chapter 4. Author has also constructed
all reconstruction schemes of SVM’s in MATLAB. 2nd and 3rd authors have provided the simulation data of a cage induction motor and written Chapter 4. 4th author has given valuable insights to the subject.
[P4]: Author has written half of the introduction, Chapter 3, Chapter 4, Chapter 5 B,
half of Chapter 6 and Chapter 7. Author has provided all classification results in MATLAB. 2nd and 4th authors have provided the simulation data of a cage induction motor and a slip-ring machine. 2nd author has written half of the introduction, Chapter 2 and Chapter 5 A. 3rd author has provided the experimental data from a cage induction motor. 5th author has given valuable insights to the subject.
Contents
Abstract Preface List of Publications Nomenclature Contents 1. Introduction…………………………………………………..……2
2. Fault diagnostics of electrical machines…………………….……4
2.1 Model-based methods…………………………………….…………...…….4
2.2 AI based methods……………………………………..………………..……6
2.1.1 Expert systems…………………………………………………..………..7
2.1.2 Fuzzy logic…………………………………………………………….….7
2.1.3 Neural networks…………………………………………….…………….9
2.1.4 Fuzzy-neural networks..………………….…………………….………..12
2.1.5 Genetic algorithms…………………………...…………………….……13
3. SVM based classification………………………..……………… 14
3.1 Introduction to SVM…………………………………………..…..……… 14
3.2 SVM theory……………………………………………………..……...……16
3.3 Multi-class classification……………………………………………….…20
4. Summary of publications…………………….….……………… 22
5. Conclusions……………………………………………….………24
References
2
1. Introduction
The fast growth of computation capacity has brought new possibilities to develop
fault diagnostics and condition monitoring methods for modern industrial plants.
Firstly, it has given possibilities to build sophisticated numerical models of diagnosed
systems in healthy operation and in different fault situations, and simulation times to
provide the important virtual measurement data have drastically decreased. Secondly,
enhanced Artificial Intelligence (AI) based methods have become common alongside
with traditional model-based methods. For example Neural Networks (NN) and fuzzy
logic have attracted a wide following in the area of fault diagnostics [Filippetti00].
Rotating electrical machines are widely used in the world’s industrial life, and there is
a strong demand for their reliable and safe operation. Faults and failures of critical
electro-mechanical parts can lead to excessive downtimes and generate costs of
millions of euros in reduced output, emergency maintenance and lost revenues. Thus,
finding efficient and reliable fault diagnostics methods especially for electrical
machines is extremely important. In the industry, model-based methods are still most
common choice for condition monitoring of electrical machinery, but during last
decade also different kinds of AI based methods have established a firm position.
Support Vector Machine (SVM) is a relatively new machine learning method based
on statistical learning theory presented by V. N. Vapnik [Vapnik98]. SVM based
classifier is claimed to have better generalisation properties than NN based classifiers.
In addition to this, SVM based classification is interesting, because its efficiency does
not depend on the number of features of classified entities. This property is very
useful in fault diagnostics, because the number of features to be chosen to be the base
of fault classification is thus not limited. The aim of this thesis is to build a SVM
based fault classification scheme for electrical machines.
3
In Section 2, several fault diagnostics methods are reviewed in general, and especially
in application to electrical machines. The section is divided to consideration of
traditional model-based methods and Artificial Intelligence (AI) based methods. In
Section 3, the base of SVM is explained, and applications where it has been utilised
are reviewed. In Section 4, the author’s publications are summarized. Finally, in
Section 5, some conclusions are made from the results.
4
2. Fault diagnostics of electrical machines
2.1 Model-based methods
Model-based fault diagnosis methods take advantage of mathematical models of the
diagnosed plant. The idea is to generate quantities that reflect inconsistencies between
nominal and faulty system operation. These quantities are called residuals, and they
are generated using analytical approaches, such as observers (e.g. [Liu97]), parameter
estimation (e.g. [Isermann93]) or parity equations (e.g. [Gertler92]).
Using observers, the underlying idea is to estimate the system outputs from the
available inputs and outputs of the system. The residual will then be a weighted
difference between the estimated and the actual outputs.
Parameter estimation approach makes use of the assumption that faults of a dynamic
system reflect to the physical parameters of the process (e.g. friction, mass velocity
resistance) and thus also to the model parameters. It detects faults through the
estimation or identification of model parameters. Differences between healthy and
faulty model parameters can be considered to be residuals.
According to [Gertler92], parity equations are mathematical relationships linking a
number of variables, arranged in such a way that all terms appear on the same side of
the equation. Parity equations can be statistical or dynamical, and the outputs of the
equations are residuals.
In Fig. 2.1, there is a diagram of the performance of a model-based fault diagnosis
method from [Gertler88], where model-based methods are reviewed. Residuals are
generated with physical measurements and with analytical model of the system. With
5
zero residuals the system is assumed to be in normal operation condition. It is
however obvious that in real systems, residuals are rarely exactly zero because of
noise. The deviation of residuals from zero is a combination of noise and faults. With
any significant noise present, some kind of statistical analysis is needed. A logical
pattern is generated showing which residuals can be considered normal and which
ones indicate faults. Such a pattern is called signature of the failure. The statistical
testing is made using the fact that the noise is random with zero mean while failures
are deterministic. In general, the statistical testing does the fault detection and the
final step of the procedure is the analysis of the logical patterns obtained from the
residuals, with the aim of isolation and identification the failure or failures that cause
them.
Figure 2.1. Performance of a model-based fault diagnosis
Traditional model-based methods have been widely utilised also in the fault
diagnostics of electrical machines. For example, in [Loparo00], multiple model
framework is used to develop monitoring, fault detection and diagnosis system in
rotating machines. Each fault to be identified is associated with a certain model
structure and parameters in the rotating machinery model. Fault diagnosis is based on
statistical testing of residuals of the bank of stochastic non-linear observers. The
residuals of the filters are monitored, and the conditional probability that each filter
model is the process model is computed, and the filter with the highest probability is
declared to match the current operating condition.
In [Wieser98], the sensitivity and robustness of the on-line model based Vienna
monitoring method is addressed. The proposed condition monitoring method
Measure- ments
Model Design
Model Statistical Testing
Decision Making
Residuals Signatures Inference
6
compares the outputs of a reference model that represents an ideal machine to a
measurement model. Observing the deviations of these two models makes it possible
to detect and even locate rotor faults. The method utilises a voltage and a current
model structure, which respond differently to the faulty rotor bar. Differences of the
model outputs are evaluated and clustered. The same researchers have studied the
method also in [Kral2000] and in [Wieser97].
In [Combastel98], a model-based method and fuzzy logic are combined to create a
fault isolation method for a DC motor current loop. The global model is divided into
multiple local models. Instead of using crisp thresholds to detect an abnormal state of
the system, fuzzy information processing is implemented, which allows the fusion of
both numerical and symbolic information in the decision-making procedure, and thus
improves the isolation ability.
2.2 AI based methods
Model-based fault diagnostics methods require precise mathematical model of the
process under consideration, and based on the model and process measurements they
monitor the health of the system. In real systems, this may become a problem, since
any unmodeled dynamics can effect on fault diagnostics process. Model-based
methods are still widely used in fault diagnostics, but different kinds of AI based
methods have also been developed to overcome problems with model-based methods.
AI methods have also been combined to more traditional methods e.g. in failure
isolation and identification tasks or in building the system models [Marcu97],
[Dexter97].
Artificial intelligence methods usually include neural networks, fuzzy logic, expert
systems and genetic algorithms. Three first ones are widely utilised in the field of
fault diagnostics, either alone or combined with some other method. Genetic
algorithms, however, are rarely used alone. In this chapter, each of these methods and
their fault diagnostics applications are shortly described. Also, in [Filippetti00], recent
developments of induction motor drives fault diagnostics using AI techniques are
presented.
7
2.2.1 Expert systems
An expert system is a program that contains knowledge of a process under
consideration. The whole knowledge of an expert system is called a knowledge base.
The knowledge base is built to offer expert knowledge to a non-expert and it is
formed utilising knowledge of a human expert of the process. Human expert is a
person, who has wide knowledge of the specific problem considered, and who knows
practical solutions to the problem. Building the knowledge base requires the
description of the considered problem and the problem solution. One should also pay
attention to the selection of questions that most explicitly relieve the problem.
In fault diagnostics, the human expert could be a person who operates the diagnosed
machine or process and who, thus, is very well aware of different kinds of faults
occurring in it. Building the knowledge base could be done interviewing the human
operator on faults occurring in the diagnosed machine and on their symptoms.
Intelligence of traditional expert systems is based on a knowledge base that includes
simple if-then rules. A certain state of the monitored system activates a certain rule. A
sort of safety factor can be included to express how certainly a process state is known,
and how certainly the decision can be made by a specific rule. It is easy to implement
small expert systems with simple rules, but a large number of rules is difficult to
maintain, and system operation may get unbearable slow. Concerning the operator’s
work, it may also be difficult to piece together too many rules, and they may not be
intuitively related to the real world.
An example of expert system in fault diagnostics is presented in [Padalkar91].
2.2.2. Fuzzy logic
Traditional expert systems can be highly enhanced with fuzzy logic. Expert systems
are usually suitable for problems, where a human expert can linguistically describe
the solution. Typical human knowledge is vague and inexact, and handling this kind
of information has often been a problem with traditional expert systems. For example,
8
the limit, when the temperature in a sauna is too high, is vague in human mind. Fuzzy
logic provides a systematic framework to process vague, qualitative knowledge. It is
speculated that in the future, most of the expert systems use fuzzy sets and fuzzy logic
instead of traditional crisp sets. In Fig. 2.2 an overview of a fuzzy inference system is
presented. Theory of fuzzy logic is presented e.g. in [Wang97].
Figure 2.2. Basic configuration of a fuzzy system.
One of the benefits of fuzzy logic is that rules in the knowledge base do not have to be
so detailed and exact as with traditional expert systems. With fuzzy logic, rules can be
generalized to cover a higher number of cases than without it. Another benefit of
fuzzy approach is that it provides an easy way to deal with contradictions in the
knowledge base. Considering fault diagnosis, fuzzy systems are useful, because fault
diagnosis often needs a knowledge-based treatment. In practise, it is very difficult to
obtain adequate representations of the complex and highly non-linear behaviour of
faulty plants using quantitative models. The use of fuzzy qualitative models can also
take account of the uncertainties associated with describing the system.
Fuzzy logic applications for fault diagnosis are reviewed in [Isermann98] and
[Dexter95]. Dexter divides fuzzy fault diagnostics applications in two classes: shallow
knowledge and deep knowledge applications. In the first class, implicit fuzzy models
are used. Tasks may be, for example, to analyse qualitative statements of the
differences between the actual values and those predicted by quantitative models, to
adapt the threshold for evaluating the residuals generated with an observer, or to
identify faults using fuzzy inference based diagnostic model. In the second class,
Fuzzy Rule Base
Fuzzy Inference Engine
Fuzzifier Defuzzifier
x in U
fuzzy sets in U
fuzzy sets in V
y in V
9
explicit fuzzy models are used. For example, fuzzy classifiers and fuzzy pattern
recognition may be used to detect and isolate faults.
Isermann emphasizes combining quantitative and qualitative knowledge. According
to him, fuzzy approaches in fault diagnostics are especially attractive for symptom
generation with fuzzy thresholds, linguistically described observed symptoms and the
approximate reasoning with multi-level fuzzy-rule-based systems for fault-symptom
tree structures.
Also, in fault diagnostics of electrical machines, fuzzy logic has become common
especially in the decision making part of the diagnostics scheme. For example, in
[Nejjari99], fuzzy logic is applied to induction motor’s condition monitoring and its
stator and phase conditions through the amplitude features of the stator currents. In
[Lasurt00], higher order statistical analysis (HOS) is used as a pre-processing
procedure applied to a machine vibration signal. A combination of data reduction,
parametrization and fuzzy logic procedures is then applied to the HOS signatures to
enable diagnosis of the machine fault.
2.2.3 Neural networks
Neural computing is a class of soft computing methods that imitate behaviour of
neural cells. Perceptron networks are general non-linear function approximators,
which are built from a network of artificial neurons connected by appropriate weights.
With neural networks it is possible to estimate a function without requiring a
mathematical description of how the output functionally depends on the input – neural
networks learn from examples.
Neural networks have gathered a plenty of interest during recent years. The most
commonly mentioned advantages of neural networks are their ability to model any
non-linear system (given suitable weighting factors and appropriate architecture), the
ability to learn, the highly parallel structure and the ability to deal with inconsistent or
noisy data.
10
In fault diagnostics, some of the difficulties of using mathematical models can be
overcome and fault diagnosis algorithms can be made more applicable to real systems
using neural networks. The neural network can be used to both generate residuals and
isolate a fault. In residual generation, the residual vector is determined in order to
characterize each fault. The second step, decision making - or classification -
processes the residual vector to determine the location and occurrence of the faults. In
Fig. 2.3, the general fault diagnosis scheme with neural networks from [Patton99] is
presented. It is possible to replace either residual generation or decision making part
with some other AI-method or model-based algorithm.
One of the main features of the neural networks is their ability to learn from
examples. Hence, neural networks can be trained, for example, to represent
relationships between measurement data of the system and certain fault conditions.
Neural networks are often used in situations, where it is possible to get plenty of
measurement data of the system. The large amount of numerical data from the system
is also an essential requirement for training the neural network. In some cases,
difficulties might occur in creating a reliable network, if there are not enough
measurements available.
Figure 2.3 General fault diagnostics scheme with neural networks [Patton99]
Another disadvantage of neural networks is that the net architecture with weighting
factors is difficult to figure out by human. This may be a problem in tuning the
system, or explaining the diagnosis results to a system operator.
11
In [Haykin99], theory of neural networks is studied thoroughly. Sorsa has studied
neural networks applications for fault diagnostics in his doctoral thesis [Sorsa95].
Also in [Patton99], various neural network based fault diagnosis methods are
presented.
In fault diagnostics of electrical machines, neural networks are often used in different
parts of the diagnostics scheme. In [Chow93], the general design considerations for
feedforward artificial neural networks to perform motor fault detection are presented.
An example of using neural networks for modelling an induction motor is presented
in [Filippetti95]. The faulted machine models used to formalize the knowledge base
of the diagnostic system are formed with neural networks.
Examples of using neural networks in classification of faults are presented in
[Yang00], [Alguindigue93], [Li00]. In these articles, neural network based classifiers
have been used to monitor rolling bearings of a motor. Also, in [Penman94], a neural
network is used as a learning and pattern recognition device, and it was able to
successfully associate input signal patterns with appropriate machine states.
In [Schoen95], an interesting neural network based clustering approach for fault
diagnostics of an electrical machine is presented. There neural networks are used to
learn on-line the spectral characteristics of a healthy motor operation. A special
frequency filter is used to pass only those harmonics, which are known to be of
importance in fault detection, to a neural net clustering algorithm. After a sufficient
training period, the neural network signals a potential failure condition, when a new
cluster is formed and persisted for some time.
12
2.2.4 Fuzzy-neural networks
The crisp numerical values obtained from the neural networks can be seen as a
drawback of the diagnostic system, because heuristic or qualitative information may
be needed, and often knowledge for the diagnostic system is available only in
qualitative form. The solution is to combine neural networks and fuzzy logic to create
fuzzy-neural networks. This approach has shown to be promising. By integrating
qualitative and quantitative knowledge through a neuro-fuzzy system, it is feasible to
combine learning ability of neural networks with the explicit knowledge
representation of fuzzy logic.
In fault diagnostics, combinations of fuzzy logic and neural networks offer benefits.
The black-box approach of pure neural networks does not allow utilisation of
qualitative knowledge of faults and their symptoms, whereas fuzzy logic –based fault
diagnostics systems are often static, i.e. they do not allow changes throughout the
experiments. With fuzzy-neural networks better understanding of the diagnosis
process of the system can be achieved, and, also, the fault detector can be adapted to
provide more accurate solutions under different operating conditions.
In [Altug99], ANFIS (Adaptive Neuro Fuzzy Inference System) -based fault
diagnostics system of an induction motor is compared with another adaptive neuro-
fuzzy system FALCON (Fuzzy Adaptive Learning Control Network). Altug & al.
have found out that both structures provide good fault diagnostics framework under
varying operation conditions. Also, in [Goode95], a neuro-fuzzy system is applied in
fault diagnostics of induction motors. The neuro-fuzzy fault detector is used to
monitor the condition of a motor bearing and stator winding insulation. After the
detector is trained, in addition to the accurate motor condition information, it also can
provide the heuristic reasoning behind the fault detection process and the actual motor
conditions due to integrated fuzzy inference system and neural network features.
13
2.2.5 Genetic algorithms
Genetic algorithms are stochastic optimisation techniques that were introduced by
Holland in 1970’s [Holland75]. They are based on the mechanisms of natural
selection and genetics. Encoding mechanism is used for representing variables of the
optimisation problem. Fitness function - or objective function - provides the
mechanism for evaluating each string and forming the fittest population. After
selection of the fittest strings to the population, crossover is used to combine strings.
After crossover, strings are subject to mutation to keep the solution space rich enough.
Genetic and evolutionary algorithms have proven to be a powerful search and
optimisation tools.
Genetic algorithms have been utilised in fault diagnostics usually in co-operation with
some other AI-method. For example in [Betta98] and [Gao00], a genetic algorithm is
used to design and to train a neural network that detects faults. In [Gao00], the neural
network is designed for motor fault detection. In [Jack00], a genetic algorithm is used
to isolate the features of input space providing the most significant information to a
neural network that detects faults in the system. Thus, the number of inputs to the
network is decreased, and the diagnostics process becomes faster and the
classification more accurate. In [Patton95], a genetic algorithm is used alone to solve
a robust fault detection problem, which is formulated to find a trade-off between
sensitivity to faults and robustness to model uncertainties. Patton reformulates all
objectives into a set of inequality constraints and applies genetic algorithm to find the
optimal solution to satisfy these constraints.
14
3. SVM based classification
3.1 Introduction to SVM
SVM is a relatively new computational learning method based on statistical learning
theory presented by V. N. Vapnik [Vapnik98]. In SVM, original input space is
mapped into a high dimensional dot product space called feature space, and in the
feature space the optimal hyperplane is determined to maximize the generalisation
ability of the classifier. The optimal hyperplane is found by exploiting optimisation
theory, and respecting insights provided by the statistical learning theory
[Cristianini00].
SVM:s have potential to handle very large feature spaces, because training of SVM is
carried out so that the dimension of classified vectors does not have influence on the
performance of SVM. That is why, it is noticed to be especially efficient in large
classification problems. Concerning fault classification, this is a benefit, because thus
the number of features does not have to be limited. Aggressive feature selection could
result in a loss of information.
Also, SVM based classifiers are claimed to have better generalization properties than
e.g. NN based classifiers, because in training the SVM classifier a so-called structural
misclassification risk is to be minimized, while traditional classifiers are trained so
that the empirical risk is minimized.
SVM has been successfully applied to different kinds of classification problems. For
example to:
15
- text categorization e.g. in [Joachims97]
- image recognition e.g. in [Pontil98]
- phoneme classification e.g. in [Salomon01]
- hand written digit recognition e.g. in [Boser92]
- medicine, breast cancer prognosis e.g. in [Freiss98]
- bioinformatics, protein fold recognition e.g. in [Ding01]
- gene expression e.g. in [Brown97]
SVM based classification has not been applied to fault diagnostics of electrical
machines until this research. Although, SVM has shown good performance in
different kinds of classification applications, its appropriateness even to fault
diagnosis in general has not been widely studied. The author found only three articles
concerning utilisation of SVM in fault diagnostics.
Saunders & al. [Saunders00] examine the possibility of using pattern recognition
techniques to determine correct repairs for faults from past production history. They
claim that pattern recognition algorithms in general are suitable for fault diagnostics,
because fault diagnosis problem can be seen to be similar for example with the
problem of text categorisation. Specifically, with SVM based pattern recognition the
authors obtain good results.
In [Rychetsky99], engine knock detection is carried out with classical neural networks
(multilayer perceptron, Adaboost), with SVM and with another large margin
classifier: Kernel Adatron. Rychetsky & al. found out that both large margin
classifiers outperform classical neural networks.
Feng & al. [Feng02] apply SVM’s to quality monitoring in robotized arc welding.
Through the feature extraction of the welding process, a SVM classifier is constructed
to establish the relationship between the feature of process parameters and the quality
of weld penetration. With the constructed method, the authors obtain good results in
identifying defects online in welding production.
16
3.2 SVM theory
Let n-dimensional input xi (i = 1,…,M) belong to Class I or Class II and associated
labels be yi = 1 for Class I and yi = –1 for Class II. For linearly separable data, we can
determine a hyperplane f(x) that separates the data:
1
( ) .n
j jj
f b w x b=
= + = +∑x w xi (1)
A separating hyperplane satisfies the constraints that define the separation of the data
samples, i.e. ( ) 1if x ≥ + , if yi = +1, and ( ) 1if x ≤ − , if yi = -1 [Cherkassky98, p. 357].
This results:
( ) ( ) 1, for 1,...,i i i iy f y b i M= + ≥ =x w xi . (2)
where w is an n-dimensional vector and b is a scalar. Notation w�xi corresponds to dot
product of vectors w and xi. The weighting vector w defines the direction of the
separating hyperplane f(x) and with w and b (bias) it is possible to define the
hyperplane’s distance from the origin.
The separating hyperplane that has the maximum distance between the hyperplane
and the nearest data, i.e. the maximum margin, is called the optimal separating
hyperplane. An example of optimal separating hyperplane of two datasets is presented
in Fig. 3.1. The optimal hyperplane is perpendicular to the shortest line between
border lines of two sets, and the plane and the shortest line intersect each other in the
halfway of the line. The geometrical margin γ is half of the sum of the distances
between arbitrary separating hyperplane and the nearest negative and positive datum
(x– and x+):
2 2 2
1 1(( ) ( )) (( ) ( )) .
2 2γ + − + −= ⋅ − ⋅ = ⋅ − ⋅w w
x x w x w xw w w
17
Without loss of generality we can search the optimal separating hyperplane among so-
called canonical hyperplanes, which fulfil w�x++ b = 1 and w�x- + b = -1
[Cristianini00, p. 94]:
2 2
1 1(( ) ( )) .
2γ + −= ⋅ − ⋅ =w x w x
w w
Figure 3.1. Optimal hyperplane
The optimal hyperplane maximizes the geometrical margin. Thus the optimal
hyperplane can be found by solving the following convex quadratic optimisation
problem:
2
i
1minimize
2subject to y ( ) 1 .i b+ ≥
w
w xi (3)
The same optimisation problem can also be formulated by minimizing the guaranteed
risk for classification problem (i.e. maximizing the generalisation ability). For this
approach, see e.g. [Cristianini00].
If the number of attributes of data examples is large, we can considerably simplify
calculations by converting the problem with Kuhn-Tucker conditions into the
equivalent Lagrange dual problem. Lagrange function for (3) is:
18
( )( )1
1( , , ) ( ) 1
2
M
i i i ii
L b y bα=
= − + − ∑w � � � �i i , (4)
where � = (α1,…, αM) is the Lagrange multiplier. The dual problem is:
m axim ize ( , , )
subject to 0, 1, ..., .i
L b
i Mα ≥ =w
(5)
By differentiating (4) with respect to w and b and imposing stationarity, we get:
1
1
( , , )
( , , ) .
M
i i ii
M
i ii
Lb y
Lb y
b
α
α
=
=
∂ = − =∂∂ = =∂
∑
∑
w � � �w
w � (6)
From (4), (5) and (6) we get the dual representation of the optimisation problem:
M
i=1 , 0
M
i=1
1maximize ( )=
2
subject to 0, 0, 1,..., .
M
i i k i k i ki k
i i i
W y y
y i M
α α α
α α
=
−
= ≥ =
∑ ∑
∑
� �i
(7)
The number of variables of the dual problem is the number of training data.
Let us assume that optimal solution for the dual problem is �* and b*. According to
the Karush-Kuhn-Tucker theorem, the equality condition in (2) holds for the training
input-output pair (xi,yi) only if the associated αi* is not 0. In this case the training
example xi is a support vector. Solving (7) for � = (α1,…,αM), we can obtain the
support vectors for Classes I and II. Then the optimal separating hyperplane is placed
at the equal distances from the support vectors for classes I and II, and b* is given by:
*1 2
1
1* ( )
2
M
k k k kk
b y α=
= − +∑ s x s xi i ,
where s1 and s2 are respectively, arbitrary support vectors for Class I and Class II. In
Fig.1, support vectors are bolded. Notice that support vectors are such training
19
samples that are on the margin of two datasets. The optimal separating hyperplane
would be the same, if only support vectors had been used as training data.
So far we have assumed that the training data is linearly separable. In the case where
the training data cannot be linearly separated, we introduce non-negative slack
variables ξi to (2), and add to the objective function given by (5), the sum of the slack
variables multiplied by the parameter C. This corresponds to adding the upper bound
C to �. In both cases, the decision functions are the same and are given by:
* *
1
( )M
i i ii
f y bα=
= +∑x x xi .
Then unknown data example x is classified as follows:
Class 1, if ( ) 0
Class 2, otherwise .
f >∈
xx
SVM is a non-linear kernel-based classifier, which maps the data to be classified onto
a space, where the data can be linearly classified. The space is called a feature space.
Using the non-linear vector function �(x) = (Φ1(x),…,Φl(x)) that maps the n-
dimensional input vector x into the l-dimensional feature space, the linear decision
function in dual form is given by
1
( ) ( ) ( )M
i i ii
f yα=
= ∑x � �i . (8)
Notice that in (8) as well in the optimisation problem (7), the data occur only in inner
products. In SVM, the actual mapping function, Φ, is not necessary to be known, but
one can calculate the optimal separating hyperplane with inner products of the
original data samples. If it is possible to find this kind of procedure to calculate inner
products of feature space in original data space, it is called a
kernel, ( , ) ( ) ( )K = Φ Φx z x zi . Then the learning in the feature space does not require
evaluating � or even knowing it, because all the original samples are handled only
20
with Gram matrices , 1(( ))Mi j i jG == x xi . Using a Kernel function, the decision function
will be:
*
support vectors
( ) ( , )i i if y Kα= ∑x x x .
However, all kernels do not correspond to inner products in some feature space. With
a so-called Mercer’s theorem it is possible to find out, whether a kernel K depicts an
inner product in that space where Φ is mapped [Cristianini00]. For example,
polynomials of degree q have inner product kernel ( )( , ) 1q
K = ⋅ +x z x z and radial
basis functions of the form 2
21
( ) ( exp )n
ii
i
sign ασ=
− = −
∑ x x� , where σ defines the
width, have the inner product kernel2
2( , ) expK
σ − = −
x zx z .
3.3 Multi-class classification
SVM’s are 2-class classifiers. They are designed to separate only two classes from
each other. However, in most of the real applications, multi-class classification is
required. For example, in fault classification of an electrical machine, there exist
several fault classes in addition to healthy operation.
The solution is to decompose a multi-class problem to several 2-class problems, train
classifiers to solve these problems, and then reconstruct the solution of the multi-class
problem from outputs of the classifiers. One of the simplest multi-class classification
structures is a so-called one-against-others approach. In this method, K pairwise
classifiers are built in the way that each classifier separates one class from all the
others. However, in many applications, this approach has been found to be inferior to
a pairwise coupling approach, where 1
( 1)2
K K − 2-class classifiers are built, each
separating one class from another ignoring all the other classes. Pairwise classifiers’
outputs are then fused to find the global solution to the K-class problem. In this
21
approach, a higher number of 2-class classifiers are needed than in the former case,
but using it, the total classification performance can usually be highly improved.
There exist numerous schemes to reconstruct the final classification solution from the
outputs of pairwise classifiers’ solutions. The simplest methods are based on majority
voting [Friedman96]. Pairwise classifiers give votes for classes and the class that gets
most of the votes is selected to be a final class decision for a sample considered.
An important problem occurs when applying majority voting. For a given sample x,
the voting scheme weights equally the outputs of all pairwise classifiers, without
considering their significance. Of course, the relevant classifiers concerning the
success of the classification are not known in advance. However, redundancy of some
pairwise classifiers may be considered with a so called mixture matrix. With this
approach, outputs of classifiers are linearly combined with the mixture matrix created,
for example, with least square estimation, to minimize the error between the correct
class decision and the linear combination of pairwise classifiers’ outputs. A mixture
matrix approach is proposed in [Mayoraz99], but it has been considered there in
scaling the outputs of one-against-others type of classifiers. In some applications, also
a nonlinear combination – e.g. in the form of a neural network – can improve the
performance of the classification structure.
Other reconstruction schemes suggested in literature are, for example, binary trees
[Schwenker00] and a fuzzy logic based method [Inoue01]. When applying binary
trees, a proper hierarchy of classifiers should be known before training the classifiers.
This requires a priori knowledge of the solution of the classification problem or
implementation of sophisticated clustering or vector quantisation algorithms. When
using the fuzzy logic approach, choosing and tuning of the membership functions is
an application dependent task, and may be quite time-consuming in some
applications.
In this thesis, reconstruction schemes are considered with pairwise SVM’s, but they
can also be used with any other pairwise classifiers.
22
4. Summary of publications
Faults of rotating machines are traditionally detected in frequency domain based on a
spectrum of currents, voltages or vibrations of the machines. Lately, most of the fault
diagnostics research of electrical machines has concentrated on monitoring spectrum
of the stator line current of the machine [Benbouzid00]. We followed this approach in
[P1], where a power spectrum of the stator line current of a 15 kW induction motor
was used as a medium of fault detection, and SVM:s were trained to distinguish
healthy spectrum from faulty spectra and faulty spectra from each other. Six different
faults were studied in addition to the healthy operation of the motor. Numerical
magnetic field analysis [Arkkio90] was used to provide virtual measurement data
from operation of the motor. Power spectra estimates of the stator current of the motor
were calculated with Welch’s method [Welch67]. Results were promising. Most of
the faults could be separated correctly from each other.
In [P2], we fused the outputs of pairwise SVM’s that were built in [P1] to get the
global classification decision. We used a simple majority voting approach, and also
the influence of noise was studied.
Without noise the classification structure performed well, but noise degraded the total
classification rate. With noise filtering, the fault detection rate only slightly increased.
This raised a question, whether the stator line current is the best choice for a fault
detection medium. It is widely used, because measuring it does not require access to
the motor, but perhaps there exist other variables that more clearly show the faults in
the motor. However, some faults could be easily detected from stator line current
regardless of the noise. With 3-class classification structure, detection rate of shorted
coil, shorted turn and healthy operation were adequate even in noisy situation.
23
It was noticed that also in noiseless case, the errors in 2-class classifiers’ outputs were
cumulated in reconstruction of the final n-class classification solution. The
malfunction of 2-class classifiers should be able to be taken into account while
reconstructing the final classification solution. In [P3], we studied different schemes
to reconstruct a multi-class classifier from one-to-one SVM based classifiers. A neural
network reconstruction resulted in the best multi-class classification results, but with a
much simpler reconstruction approach relying on a mixture matrix almost equal
classification performance was obtained. In this application, a linear combination is a
practical choice for the reconstruction scheme, because training and tuning a neural
network is an exhausting task, and the benefits of applying a nonlinear approach are
marginal. Majority voting approach with rough reconstruction was found to be
inferior to the other methods considered.
In [P4], SVM based classification was applied to fault diagnostics of a 35 kW cage
induction motor and a slip-ring generator. A mixture matrix reconstruction scheme
was used to combine 2-class SVM:s. Stator line current, circulating currents between
parallel branches and forces acting on the machine’s rotors were compared as fault
indicators. Circulating currents between parallel branches and forces on rotor were
found to be superior indicators of faults compared to the stator current.
24
5. Conclusions
Electrical machines play an important part in the world’s industry. Their fault
diagnostics and condition monitoring is an important research subject. In addition to
traditional model-based fault diagnostics, different kinds of AI based methods have
become popular in the area of fault diagnostics of electrical machinery. A wide
variety of neural network, fuzzy logic or genetic algorithm based applications can be
found from the literature. SVM is a modern machine learning method, and although it
has been successfully applied to numerous classification and pattern recognition
problems, its utilization in fault diagnostics is low. In fault diagnostics of electrical
machines, SVM had not been applied before this research.
In this thesis, SVM based fault classification approach was studied for different
electrical machines. Firstly, pairwise SVM’s were trained to discriminate between
healthy and faulty power spectrum estimates of a stator line current of a motor.
Secondly, five different fusion techniques of the SVM’s were studied to get the final
decision of the motor condition. A mixture matrix fusion was found to be the best
technique in this application. Finally, different variables of rotating machines were
compared as indicators of motors’ condition, and it was found that there exist better
fault indicators than stator line current, e.g. forces on the rotor and currents between
parallel branches.
SVM based classification showed to be an efficient and reliable way to do the
classification of faults, especially, when either circulating currents in parallel branches
or forces on the rotor are used as a fault indicator. Comparison of different fusion
techniques of pairwise SVM’s is also important knowledge concerning the research of
classification methods in general, and the mixture matrix approach had not been
earlier applied exactly in this form.
25
References
[Alguindigue93]: Alguindigue, I.E., Loskiewicz-Buczak, Uhrig, R., “Monitoring and
Diagnosis of Rolling Element Bearings Using Artificial Neural Networks”, IEEE
Transactions on Industrial Electronics, Vol. 40, No.2, pp. 209-217, April 1993.
[Altug99]: Altug, S., Mo-Yuen Chow, Trussell, H.J. “Fuzzy Inference Systems
Implemented on Neural Architectures for Motor Fault Detection and Diagnosis”,
IEEE Transactions on Electronics, Vol. 46 Issue: 6, pp.1069 –1079, Dec. 1999.
[Arkkio90]: Arkkio, A., “Finite Element Analysis of Cage Induction Motors Fed by�
Static Frequency Converters. IEEE Transactions on Magnetics, 26, 2, pp. 551--
554,1990.
[Benbouzid00]: Benbouzid, M., “A Review of Induction Motor Signature Analysis as
a Medium for Faults Detection”, IEEE Transactions on Industrial Electronics, Vol.
47, No. 5, pp. 984-993, October 2000.
[Boser92]: Boser, B.E., Guyon, I.M., Vapnik, V.N., “A Training Algorithm for
Optimal Margin Classifiers”, In Proc. Of the 5th Annual ACM Workshop on
Computational Learning Theory, pp. 142-152, ACM Press, 1992.
[Brown97]: Brown, M.P.S., Grundy, W., Lin, D., Cristianini, N., Sugnet, C., Furey,
T.S., Ares, M. Jr., Haussler, D.: “Knowledge-based Analysis of Microarray Gene
Expression Data Using Support Vector Machines”, Proceedings of the National
Academy of Sciences, Vol. 1, pp. 262-267, 1997.
26
[Betta98]: Betta, G., Liguori, C., Pietrosanto, A., ”An Advanced Neural-Network-
Based Instrument Fault Detection and Isolation Scheme”, IEEE Transactions on
Instrumentation and Measurement, Vol. 47, No.2, pp. 507-512, April 1998.
[Cherkassky98]: Cherkassky, V., Mulier, F., “Learning From Data”, John Wiley &
Sons, 1998.
[Chow93]: Chow, M. Y., Sharpe, R. N., Hung, J. C.: “On the Application and Design
of Artificial Neural Networks for Motor Fault Detection (Part I and Part II),” IEEE
Trans. Ind. Electron., vol. 40, pp. 181–197, April 1993.
[Combastel98]: Combastel, C., Gentil, S., Rognon, J.-P.: “A symbolic reasoning
approach to fault detection and isolation applied to electrical machines”, Proc. of the
IEEE Int. Conf. on Control Applications, Vol.1, pp. 475-479, 1998.
[Cristianini00]: Cristianini, N., Shawe-Taylor, J., “Support Vector Machines and
Other Kernel-Based Learning Methods”, Cambridge University Press, 2000.
[Dexter95]: Dexter, A.L., “Fuzzy Model Based Fault Diagnosis”, IEE Proc. Control
Theory Appl., Vol. 142, No. 6, pp. 545-550, November 1995.
[Dexter97]: Dexter, A.L., Benouarets, M., “Model-Based Fault Diagnosis Using
Fuzzy Matching”, Systems, Man and Cybernetics, Part A, IEEE Transactions on, Vol.
27 Issue: 5, pp. 673 –682, Sept. 1997.
[Ding01]: Ding, C. H.Q, Dubchak, I., “Multi-Class Protein Fold Recognition Using
Support Vector Machines and Neural Networks”, Bioinformatics, Vol.17, No.4,
pp.349-358, April 2001.
[Feng02]: Feng, Y., Lun, S.Y., Di, L., Zong, L.Y., “Application of Support Vector
Machines to Quality Monitoring in Robotized Arc Welding”, Int. Joint Conf. On
Neural Networks, Vol. 3, pp. 2321-2326, 2002.
27
[Filippetti95]: Filippetti, F., Franceschini, G., Tassoni, C.: “Neural Network Aided
On-Line Diagnostics of Induction Motor Rotor Faults,” IEEE Trans. Ind. Applicat.,
vol. 31, pp. 892–899, July/Aug. 1995.
[Filippetti00]: Filippetti, F, Franceschini, Tassoni C., Vas, P., ”Recent Developments
of Induction Motor Drives Fault Diagnostics Using AI Techniques”, IEEE
Transactions on Industrial Electronics, Vol.47, No.5, pp. 994-1004, October 2000.
[Freiss98]: Freiss, T., Cristianini, N., Campbell, C., “The Kernel-Adatron Algorithm:
A Fast and and Simple Learning Procedure for Support Vector Machines”, Proc. Of
the 15th Int. Conf. On Machine Learning, ICML’98, pp. 180-196, 1998.
[Friedman96]: Friedman, J., “Another Approach to Polychotomous Classification”,
Technical Report, Stanford University, Department of Statistics, 1996.
[Gao00]: Gao, X.Z., Ovaska, S.J., Dote, Y. “Motor Fault Detection Using Elman
Neural Networks with Genetic Algorithm-aided training”, IEEE Int. Conf. on
Systems, Man, and Cybernetics, Vol. 4, pp. 2386 –2392, 2000.
[Gertler88]: Gertler, J.J., “Survey of Model-Based Failure Detection and Isolation in
Complex Plants”, IEEE Control Systems Magazine, Vol.8, Issue 6, pp.3-11,
Dec.1988.
[Gertler92]: Gertler, J.J., Anderson, K.C., “An Evidential Reasoning Extension to
Quantitative Model-Based Failure Diagnosis”, IEEE Transactions on Systems, Man
and Cybernetics, Vol. 22, Issue 2, pp. 275 –289, March-April 1992.
[Goode95]: Goode, P.V., Mo-Yuen Chow “Using a Neural/Fuzzy System to Extract
Heuristic Knowledge of Incipient Faults in Induction Motors: Part II-Application”
IEEE Transactions on Industrial Electronics, Vol. 42, Issue 2, pp. 139 –146, April
1995.
28
[Haykin99]: Haykin, S., ”Neural Networks – A Comprehensive Foundation”, Prentice
Hall International, Inc., 1999.
[Holland75]: Holland, J.H., “Adaptation in Natural and Artificial Systems”, Univ. of
Michigan Press, Ann Arbor, Mich., 1975.
[Inoue01]: Inoue, T., Abe, S., “Fuzzy Support Vector Machines for Pattern
Classification”, Proc. of Int. Joint Conference on Neural Networks, IJCNN '01, Vol.
2, pp. 1449 -1454, 2001.
[Isermann93]: Isermann, R., “Fault Diagnosis of Machines via Parameter Estimation
and Knowledge Processing”, Automatica, Vol.29, No. 4, pp. 815-835, 1993.
[Isermann98]: Isermann, R., “On Fuzzy Logic Applications for Automatic Control,
Supervision, and Fault Diagnosis”, IEEE Transactions on Systems, Man and
Cybernetics – Part A: Systems and Humans, Vol. 28, No. 2, pp. 221-235, March
1998.
[Jack00]: Jack, L.B., Nandi, A.K., ”Genetic Algorithms for Feature Selection in
Machine Condition Monitoring with Vibrating Signals”, IEEE Proc. Vis. Image
Signal Processing, Vol.147, No.3, pp.205-212, June 2000.
[Joachims97]: Joachims, T., “Text Categorization with Support Vector Machines:
Learning with Many Relevant Features ”, University of Dortmund, Technical Report,
1997.
[Kral00]: Kral, C., Pirker, F., Pascoli, G. “Detection of Rotor Faults in Squirrel Cage
Induction Machines at Standstill for Batch Tests by Means of the Vienna Monitoring
Method”, Conf. Record of IEEE Conf. On Industry Applications, Vol. 1, pp. 499 –
504, 2000.
[Lasurt00]: Lasurt, I., Stronach, A.F., Penman, J., “A Fuzzy Logic Approach to the
Interpretation of Higher Order Spectra Applied to Fault Diagnosis in Electrical
29
machines”, 19th Int. Conf. Of the North American Fuzzy Information Processing
Society, NAFIPS, pp. 158 –162, 2000.
[Li00]: Li, B., Chow, M-Y, Tipsuwan, Y., Hung, J.C., “Neural-Network-Based Motor
Rolling Bearing Fault Diagnosis”, IEEE Transactions on Industrial Electronics, Vol.
47, No. 5, pp. 1060-1069, October 2000.
[Liu97]: Liu, B., Si, J. “Fault Isolation Filter Design for Linear Time-Invariant
Systems”, IEEE Transactions on Automatic Control, Vol.42, Issue 5, pp. 704 - 707,
May 1997.
[Loparo00]: Loparo, K.A., Adams, M.L., Lin, W., Abdel-Magied, M.F., Afshari, N.:
”Fault Detection and Diagnosis of Rotating Machinery”, IEEE Transactions on
Industrial Electronics, Vol. 47, Issue: 5, pp. 1005 –1014, Oct. 2000.
[Mayoraz99]: Mayoraz, E., Alpaydin E., “Support Vector Machines for Multi-Class
Classification”, Int. Workshop on Artificial Neural Networks (IWANN), Vol. 2, pp.
833-842, 1999.
[Marcu97]: Marcu, T., Mirea, L., “Robust Detection and Isolation of Process Faults
Using Neural Networks”, IEEE Control Systems Magazine, Vol. 17, Issue 5, pp. 72 –
79, Oct. 1997.
[Nejjari99]: Nejjari, H., Benbouzid, M.E.H., ”Application of Fuzzy Logic to
Induction Motors Condition Monitoring” IEEE Power Engineering Review, Vol. 19,
Issue 6, pp. 52 –54, June 1999.
[Padalkar91]: Padalkar, S., Karsai, G., Biegl, C., Sztipanovits, J., Okuda, K.,
Miyasaka, N. “Real-Time Fault Diagnostics”, IEEE Expert, Vol. 6, Issue: 3, pp.75-85,
June 1991.
30
[Patton95]: Patton, R.J. “Robust Fault Detection of Dynamic Systems via Genetic
Algorithms”, Conf. Publications, Genetic Algorithms in Engineering Systems:
Innovations and Applications, pp. 511-516, 12-14 September 1995.
[Patton99]: Patton, R.J., Lopez-Toribio, C.J., Uppal, F.J., ”Artificial Intelligence
Approaches to Fault Diagnosis”, Condition Monitoring: IEE Colloquium on
Machinery, External Structures and Health (Ref. No. 1999/034), pp.5/1-5/18,1999.
[Penman94] J. Penman, C. M. Yin, “Feasibility of Using Unsupervised Learning NN
for the Condition Monitoring of Electrical Machines,” Proc. IEE—Elect. Power
Applicat., vol. 141, no. 6, pp. 317–322, 1994.
[Pontil98]: Pontil, M., Verri, A., “Object Recognition with Support Vector
Machines”, IEEE Trans. On PAMI, Vol.20, pp. 637-646, 1998.
[Rychetsky99]: Rychetsky, M., Ortmann, S., Glesner, M., “Support Vector
Approaches for Engine Knock Detection”, Int. Joint Conf. On neural Networks, Vol.
2, pp. 969-974, 1999.
[Salomon01]: Salomon, J., ”Support Vector Machines for Phoneme Classification”,
M.Sc Thesis, University of Edinburgh, 2001.
[Schoen95]: Schoen, R.R., Lin, B.K., Habetler, T.G.; Schlag, J.H.; Farag, S. “An
Unsupervised, On-Line System for Induction Motor Fault Detection Using Stator
Current Monitoring”, IEEE Transactions on Industry Applications, Vol. 31 Issue: 6,
pp. 1280 –1286, Nov.-Dec. 1995.
[Saunders00]: Saunders, C., Gammerman, A., Brown, H., Donald, G., “Application of
Support Vector Machines to Fault Diagnosis and Automated Repair”, Proc. Of the
11th Int. Workshop on Principles of Diagnosis (DX’00), June 8-10, 2000.
[Sorsa95]: Sorsa, T., “Neural Network Approach to Fault Diagnosis”, Doctoral
Thesis, Tampere University of Technology Publications 153, 1995.
31
[Schwenker00]: Schwenker, F., “Hierarchical Support Vector Machines for Multi-
Class Pattern Recognition”, Proc. of 4th Int. Conference on Knowledge-Based
Intelligent Engineering Systems and Allied Technologies, Vol. 2, pp. 561 –565, 2000.
[Vapnik98]: Vapnik, V.N., “Statistical Learning Theory”, John Wiley & Sons, 1998.
[Wang97]: Wang, Li-Xin, “A Course in Fuzzy Systems and Control”, Prentice Hall
PTR, 1997.
[Welch67]: Welch, P.D, “The Use of Fast Fourier Transform for the Estimation of
Power Spectra: A Method Based on Time Averaging Over Short, Modified
Periodograms”, IEEE Transactions on Audio Electroacoustics, Vol. AU-15, pp. 70-
73, June 1967.
[Wieser97]: Wieser, R., Kral, C., Pirker, F., Schagginger, M. “Condition Monitoring
of Inverter Fed Induction Machines by Means of State Variable Observation”, Eighth
Int. Conf. on Electrical Machines and Drives, No. 444, pp. 336 –340, 1997.
[Wieser98]: Wieser, R., Kral, C., Pirker, F., Schagginger, M. “Robust Induction
Machine Cage Monitoring Technique for Highly Distorted Voltage and Current
Waveforms, the Vienna method“, Seventh Int. Conf. on Power Electronics and
Variable Speed Drives, No. 456, pp. 194 –199, 1998.
[Yang00]: Yang, D-M, Penman, J.: “Intelligent Detection of Induction Motor Bearing
Faults Using Current and Vibration Monitoring”, Proc. Of Int. Cong. On Condition
Monitoring and Diagnostics Engineering Management, Vol. 1, pp. 461-470,
December 2000.
HELSINKI UNIVERSITY OF TECHNOLOGY CONTROL ENGINEERING LABORATORY Editor: H. Koivo Report 119 Gadoura, I. A. Design of Intelligent Controllers for Switching-Mode Power Supplies. November 1999. Report 120 Ylöstalo, T., Salonen, K., Siika-aho, M., Suni, S., Hyötyniemi, H., Rauhala, H., Koivo, H. Paperikoneen kiertovesien konsentroitumisen vaikutus mikrobien kasvuun. September 2000. Report 121 Cavazzutti, M. Fuzzy Gain Scheduling of Multivariable Processes. September 2000. Report 122 Uykan, Z. Intelligent Control of DC/DC Switching Buck Converter. December 2000. Report 123 Jäntti, R. Power Control and Transmission Rate Management in Cellular Radio Systems - A snapshot analysis
approach. May 2001. Report 124 Uykan, Z. Clustering-Based Algorithms For Radial Basis Function and Sigmoid Perceptron Networks. June 2001. Report 125 Hyötyniemi, H. Multivariate Regression - Techniques and tools. July 2001. Report 126 Kaartinen, J. Data Acquisition and Analysis System for Mineral Flotation. October 2001. Report 127 Ylén, J.-P. Measuring, Modelling and Controlling the pH value and the Dynamic Chemical State. November 2001. Report 128 Gadoura, I. A., Suntio, T. Implementation of Optimal Output Characteristic for a Telecom Power Supply - Fuzzy-logic approach.
April 2002. Report 129 Elmusrati, M. S. Power Control and MIMO Beamforming in CDMA Mobile Communication Systems. August 2002. Report 130 Pöyhönen, S., Negrea, M., Arkkio, A., Hyötyniemi, H. Comparison of Reconstruction Schemes of Multiple SVM’s Applied to Fault Classification of a Cage
Induction Motor. August 2002. Report 131 Pöyhönen, S. Support Vector Machines in Fault Diagnostics of Electrical Motors. September 2002. ISBN 951-22-6133-2
ISSN 0356-0872
Picaset Oy, Helsinki 2002