A Hybrid Semi-parametric Model for Robust Fault Detection based on GMDH Neural Network

This paper presents a hybrid Semi-parametric model for nonlinear system fault detection. The proposed hybrid model integrates the analytical model and the group method of data handing-type neural network (GMDH neural network) model. The GMDH neural network model is used to correct the modeling error of the analytical model for system fault diagnosis. There are two improvements to the hybrid semi-parametric model. Firstly, the evaluation criteria of the middle layer neurons are improved to improve the structural stability of the GMDH neural networks, especially the comprehensive evaluation of neurons using root mean square error, maximum absolute error, and network complexity function. Secondly, the reasonable threshold parameter is designed to enhance the robustness of the fault detection. The simulation results on a reaction flywheel of a satellite demonstrate that the proposed model is effective.