On a neural network-based fault detection algorithm

Abstract

In recent years, many techniques have been proposed for the detection and isolation of abrupt changes in dynamical systems. Two of these schemes, namely, robust observers and detection filter, are studied in this thesis. The sensitivity of these approaches to parameter ariations, or modeling errors or both, makes them unable to deect faults and causes false alarms in the detection logic. In this thesis, a potential solution is presented. It involves the use of neural networks along with the current observer-based scheme. With this approach, one can achieve a robust failure detection scheme with minimal sensitivity to parameter perturbation, system's non-linearities, and white noise. A four tank non-linear system is used for illustration.