Artificial neural networks in fault diagnosis of dynamical systems

Summary form only given. The lecture starts with the discussion of the methodology of Fault Detection and Isolation (FDI) for dynamic systems. Then recent model-based approaches to FDI analytical ones and those based on soft computing are surveyed. Taking into account many limitations of analytical methods, the main attention is focused on the use of neural networks in FDI for solving specific tasks such as fault isolation, but mainly fault detection. Two kinds of dynamic neural networks the MultiLayer Perceptron (MLP) and the Group Method of Data Handling (GMDH) are discussed for the purpose of modelling the diagnosed systems. Irrespective of the neural networks used, there is always the problem of neural model uncertainty, i.e. the model-reality mismatch. Therefore, the neural network-based fault diagnosis scheme should provide robustness to model uncertainty. It will be shown how to determine the structure and parameters of the GMDH network as well as how to estimate modelling uncertainty of the resulting neural model using a Bounded-Error Approach (BEA). Such an approach gives the possibility of formulating an algorithm that allows obtaining a neural network with relatively small modeling uncertainty. The presentation describes how to develop an adaptive threshold with the GMDH model using some knowledge regarding its uncertainty, and how to increase the robustness of GMDH-based fault diagnosis. To illustrate the effectiveness of the GMDH network in fault diagnosis, several powerful examples a sugar factory value actuator (DAMADICS benchmark problem) and a laboratory two-tank system are presented.