Neural Network based Active Fault Diagnosis with a Statistical Test

Abstract

The paper focuses on designing an active fault detector (AFD) for a nonlinear stochastic system subject to abrupt faults. The neural network (NN) based models of the monitored system and their prediction error uncertainties are identified using historical input-output data obtained from the system under fault-free and all considered faulty conditions. The fault detector is based on a multiple hypothesis CUSUM-like statistical test that uses the identified NN models. The quality of decisions provided by such a detector is improved by a closed loop input signal generator. The input signal generator is represented by another NN and it is designed using a reinforcement learning method. The proposed AFD is illustrated by means of a numerical example.