The design of classifiers for condition monitoring and fault diagnosis applications

Firstly, the thesis addresses the problem caused by the limited availability of data for some classes (particularly fault classes), for supervised neural networks. Two novel techniques are developed to deal with this problem. The first of these techniques, referred to here as the 'Equal-Weightage' (EW) algorithm, involves a modification to the standard multi-layer Perceptron (MLP) training algorithm. The second approach, referred to here as 'Duplicate Data' (DD) training, is used to alter the configuration of the data set. Each technique is explored both theoretically and empirically, and is shown to result in significantly improved classifier performance.;Secondly, a 'fusion' classifier framework is developed which systematically addresses the issue of 'unclassified' and 'misclassified' patterns, in order to improve the performance of a classification system. The complete blackboard-based framework involves majority voting, Dempster-Shafer (D-S), MLP and expert system component. The D-S component involves a novel approach to mass assignment in D-S theory: an efficient implementation of this approach is also developed. Overall, the framework is seen to provide substantial improvements in classifier performance.;The techniques developed in the thesis are principally applied to the cooling system of a diesel engine. However, the techniques are also demonstrated in a different domain (classifying electrocardiographs) and it is argued that the results will prove valuable in a wider range of application areas in future studies.