Active Fault Detection for Neural Network Based Control of Non-linear Stochastic Systems

This paper deals with design of active fault detection of non-linear stochastic systems. As general solution of the problem is extremely difficult, a special case of active detector design for a given set of controllers for jump Markov non-linear Gaussian models is considered. The optimal active detector for a given set of controllers is intractable and therefore, the rolling horizon technique will be used to reduce computational costs. The system is modelled using a multi-layer perceptron neural network where structure and unknown parameters are obtained by means of an off-line training process based on the extended Kalman filter estimation method and structure optimization using pruning of the insignificant connections. The proposed active detector is compared with a passive one based on open-loop feedback strategy and the performance is illustrated in an example by simulation and Monte Carlo analysis.