Fault Detection for a Class of Unknown Non-Linear Systems via Associative Memory Networks

Abstract

This paper presents a novel approach for the detection of faults for a class of non-linear systems whose parameters are unknown non-linear functions of both the measurable operating point and the faults of the system. Neural networks are used to estimate the healthy model's parameters, based on the measurable operating points, when no fault occurs within the system (this procedure is called the training of a healthy system). For this purpose, a modified version of the recursive least-squares algorithm with normalized signals and an output-error dead zone are employed. After the training of the healthy system, this recursive algorithm remains on-line to estimate the system parameters which, together with trained neural networks, are used to recognize, and differentiate, parameter changes that are caused either by the variation in the measured operating points or by faults.