Learning methodology for failure detection and accommodation

A major goal of intelligent control systems is to achieve high performance with increased reliability, availability, and automation of maintenance procedures. In order to achieve fault tolerance in dynamical systems many algorithms have been developed during the past two decades. Fault diagnosis and accommodation methods have traditionally been based on linear modeling techniques, which restricts the type of practical failure situations that can be modeled. This article presents a learning methodology for failure detection and accommodation. The main idea behind this approach is to monitor the physical system for any off-nominal behavior in its dynamics using nonlinear modeling techniques. The principal design tool used is a generic function approximator with adjustable parameters, referred to as online approximator. Examples of such structures include traditional approximation models such as polynomials and splines as well as neural networks topologies such as sigmoidal multilayer networks and radial basis function networks. Stable learning methods are developed for monitoring the dynamical system. The nonlinear modeling nature and learning capability of the estimator allow the output of the online approximator to be used not only for detection but also for identification and accommodation of system failures. Simulation studies are used to illustrate the learning methodology and to gain intuition into the effect of modeling uncertainties on the performance of the fault diagnosis scheme