Application of artificial neural networks to condition monitoring
Predictive maintenance or condition-based maintenance offers significant advantages over the traditional methods of preventive or breakdown maintenance of electromechanical systems. Despite its benefits, predictive maintenance is difficult to implement. This has led to the development of various techniques which allow the early detection of many common fault conditions, through analysis of quantities such as spectral components of line currents, magnetic fields and frame vibrations. Associating the observed signal patterns with the condition of the machine depends to a great extent on the experience and knowledge of experts. The fact that human operators are very successful at these monitoring tasks suggests that one possible method for designing computer-based monitoring systems is to model the learning and decision-making abilities of a human operator. The philosophy pursued in this research is therefore to design a system that should be able to emulate as closely as possible the learning, pattern recognition and the sensor fusion abilities of human operators. This thesis is, therefore, concerned with the application of artificial neural networks to condition monitoring of electrical drives, with particular reference to induction motors. The neural networks studied were the multi-layered perceptron (MLP) and the Kohonen self-organising feature map (KFM). The learning paradigm of the former is supervised, while the latter is unsupervised. A comprehensive theoretical basis is provided for both neural networks employed, and their effectiveness is verified by suitable experiments. The ability of the neural networks to predict the condition of the machine for varying fault severity as well as the transferrability of a trained network to monitor other machines of similar characteristics were also of interest. The suitability of the neural network fault diagnosis system for inverter-fed machines was also studied.