Adaptive learning and neural networks in fault diagnosis

Abstract

Neural networks provide a solution to overcome some of the drawbacks of quantitative fault diagnosis because they are capable of modelling systems by using training data off-line. The neural networks are particularly good for fault diagnosis of systems that have few a priori and imperfect and/or noisy data. Two basic learning methods and their application to fault diagnosis were studied: supervised and unsupervised learning methods. Two types of neural networks based on supervised learning were considered: multi-layered perceptron networks and radial basis function networks. Most neural network-based fault diagnosis systems require a priori fault classes that are used to train the networks in order to recognise faults during system operation. It may be extremely difficult or dangerous to acquire fault data from real systems. To solve this problem, unsupervised learning is recommended to be used. In this approach the neural network classifies the data and the network learns new faults and adapts them to similar faults already occurred. Two types of neural networks based on unsupervised learning are investigated: Kohonen and counterpropagation networks. As a result of research the radial basis function and counterpropagation network were selected and applied to a model of an autonomous mobile vehicle in order to diagnose fault in actuators, sensors and the system.