Fault diagnosis of nonlinear systems using recurrent neural networks

Abstract

In this work, we propose a fault detection and isolation (FDI) methodology that enables diagnose of single, multiple and simultaneous actuator and sensor faults regardless of the utilized model structure. The basis of the proposed methodology is modelling and inversion of nonlinear systems using recurrent neural network (RNN)s. To this end, a bank of RNNs is used to estimate system inputs and/or outputs and build predictive models using the obtained estimates and an RNN as the plant model. Then a bank of residuals is generated in a way each residual is sensitive to a subset of faults and insensitive to the rest and as a result of this, a unique fault signature is obtained for each fault scenario. RNNs can be replaced by other machine learning techniques such as partial least square (PLS)s, random forest, etc. One of the advantages of the proposed methodology is that it does not require the existence of plant fault history or first principles models unlike other existing results in the literature. Also, it enables isolation of actuator and simultaneous actuator and sensor faults in highly interconnected systems. The effectiveness of the proposed FDI scheme is shown via simulation examples.