Artificial neural network models for fault detection and isolation of industrial processes

The paper focuses on using of artificial neural networks in model-based fault detection and isolation. Modelling of a system both at its normal operation conditions and in faulty states is considered and a comparative study of three different methods of system modelling that use a linear model, neural network nonlinear autoregressive with exogenous input model, and neural network Wiener model is presented. Application of these models is illustrated with an example of approximation of a dependence of the juice steam pressure in the stage two on the juice steam pressures in the stages one and three of a five stage sugar evaporator. Parameters of the linear model are estimated with the recursive pseudolinear regression method, whilst the backpropagation and truncated backpropagation through time algorithms are employed for training the neural network models. All the considered models are derived based on the experimental data recorded at the Lublin Sugar Factory.