Fault diagnosis of engineering systems using neural networks: a practical approach

The work reported in this paper is motivated by the need to define guidelines for building neural network-based diagnostic architectures for engineering systems in the fluid power domain. These systems are necessarily quite complex and whilst some progress has been made it has been difficult establishing the fundamental principles needed to reliably specify a diagnostic architecture for a given system configuration. Experience shows that as systems become increasingly complex, the size of neural network (NN) and amount of training data required increases at a much greater rate. This causes a disproportionate increase in network training time. With this in mind the systems chosen as examples in this paper are deliberately of a very simple nature. This reduced the computing overhead and allowed a much larger number of experiments to be performed in the available time. However, it is assumed that there is no loss in generality by considering simple systems. That is, the principles developed for small scale systems are applicable to larger systems. This will generally be true as long as there are sufficient computing hardware resources available for network training