Support vector machine based classification in condition monitoring of induction motors

Abstract

Continuous and trouble-free operation of induction 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 addition to analytical and knowledge-based models, application of data-based models has established a firm position in the induction motor fault diagnostics during the last decade. For example, pattern recognition with Neural Networks (NN) is widely studied.

Support Vector Machine (SVM) is a novel machine learning method introduced in early 90's. It is based on the statistical learning theory presented by V.N. Vapnik, and it has been successfully applied to numerous classification and pattern recognition problems such as text categorization, image recognition and bioinformatics. SVM based classifier is built to minimize the structural misclassification risk, whereas conventional classification techniques often apply minimization of the empirical risk. Therefore, SVM is claimed to lead enhanced generalisation properties. Further, application of SVM results in the global solution for a classification problem. Thirdly, SVM based classification is attractive, because its efficiency does not directly depend on the dimension of classified entities. This property is very useful in fault diagnostics, because the number of fault classification features does not have to be drastically limited. However, SVM has not yet been widely studied in the area of fault diagnostics. Specifically, in the condition monitoring of induction motor, it does not seem to have been considered before this research.

In this thesis, a SVM based classification scheme is designed for different tasks in induction motor fault diagnostics and for partial discharge analysis of insulation condition monitoring. Several variables are compared as fault indicators, and forces on rotor are found to be important in fault detection instead of motor current that is currently widely studied. The measurement of forces is difficult, but easily measurable vibrations are directly related to the forces. Hence, vibration monitoring is considered in more detail as the medium for the motor fault diagnostics.

SVM classifiers are essentially 2-class classifiers. In addition to the induction motor fault diagnostics, the results of this thesis cover various methods for coupling SVMs for carrying out a multi-class classification problem.