A neural network based adaptive fault detection scheme

Abstract

An adaptive neural network augmented observer for fault detection in nonlinear systems is presented. The key feature of this fault detection scheme is the use of a sliding mode observer to characterize the unmodeled dynamics, and facilitate the training of the neural network. The scheme provides robust fault detection in the presence of modeling errors. The fault detection scheme is validated by simulating faults in a section of a thermal power plant model. Simulations show that the adaptive fault detection scheme learns the unmodeled dynamics, and is able to distinguish between faults, and modeling errors.