Research and Application of Failure Detection Method Based on Closed-loop Systems

Abstract

When a neural network model is trained to predict system output, the prediction error can be used as residual to report fault. Most existing research uses system open-loop input/output data, while the trained model is used to detect fault when the system runs under closed-loop control. This paper analyses the drawback of the training data acquisition and proposes new data acquisition method, so that the model accuracy is greatly improved. In addition, detection of the sensor fault, which is involved in the closed-loop, is discussed and the simulation for detecting such sensor faults are conducted. The new scheme is assessed and validated by being applied to the automotive engine air path to detect some simulated faults. The simulation results show that the developed method is effective and the residual is more sensitive to the faults