Robust neural-network-based fault detection with sequential D-optimum bounded-error input design

A growing demand for technologically advanced systems has contributed to the increase of the awareness of systems safety and reliability. Such a situation requires the development of novel methods of robust fault diagnosis. The application of the analytical redundancy based methods for system fault detection causes that their effectiveness depends on model quality. In this paper, a new methodology for the improvement of the neural model with a D-optimum sequential experimental design technique combined with outer bounding ellipsoid algorithm is proposed. Moreover, a novel method of robust fault detection against neural model uncertainty and disturbances is developed. Such an approach is used for modelling and robust fault detection of the three-screw spindle oil pump.