Fault diagnosis for a class of closed-loop systems via deterministic learning

Early detection of small faults in closed-loop systems is a challenging issue in the fault diagnosis literature. The effect of faults in closed-loop systems will be obscured by a robust feedback control, especially when the controller is coupled with nonlinear uncertainty. In this paper, an approach for rapid detection for small faults in a class of closed-loop uncertain systems is proposed based on the deterministic learning (DL) theory. First, two types of neural networks are constructed. One is embedded in the controller, another one is embedded in state estimator. Through DL, learning the unknown system dynamics (including fault dynamics) contained in the control signal and state signal can be achieved. Second, a bank of estimators are constructed by using the learned knowledge. The effect of feedback control can be compensated by using the learned knowledge, giving the diagnosis scheme an advantage to be sensitive to small faults. A rigorous analysis of the performance of the detection scheme is also given. Application to a practical system is presented. It is illustrated how a fault in a ocean surface vessel can be detected by using the developed methodology.