Research On Fault Diagnosis Method Of Turbofan Engine Sensor Based On Dynamic System Identification

With the development of machine learning and artificial intelligence,data-driven modeling methods can also obtain accurate output of complex systems without prior knowledge of system processes.Based on the system identification theory,an aero-engine model was established,and the research on engine sensor fault diagnosis methods was carried out.The main works are as follows:Firstly,the identification modeling of DGEN380 was carried out,and the input the excitation signals were designed under three typical engine operating conditions.The DGEN380 was regarded as a linear system in a small neighborhood of three operating conditions.The appropriate linear model structure was selected using the minimum prediction error criterion FPE,and the linear system parameters were identified based on the recursive least squares algorithm RLS.Subsequently,a UD decomposition algorithm was introduced which is able to identify the structure and parameters of a linear system model synchronously,and this method was expanded into a MIMO linear system.Then,the DGEN380 engine was regarded as a nonlinear system,and a nonlinear model was established based on the nonlinear autoregressive moving average model NARMAX.A forward term-selection method based on the AIC saliency criterion was used to select the NARMAX model structure.Due to the complicated modeling process and difficulty in nonlinear MIMO system modeling by classical nonlinear identification method,the LOLIMOT network was used to construct a nonlinear aero-engine model,which was a kind of fuzzy neural network.By comparing the linear and the non-linear identification method,it was found that the accuracy of the nonlinear identification algorithm was higher than that of the linear identification algorithm,and the accuracy of the linear system model can be improved by increasing the order of the linear system model.In order to solve sensor fault diagnosis of linear system,an improved RLS identification algorithm was proposed.The sensor fault parameters were regarded as parameters to be identified,and the optimal estimation of the fault parameters was obtained by minimizing the secondary performance index of the system.An identification algorithm SF-RLS combining forgetting strategy and RLS algorithm was proposed to solve the sensor fault diagnosis of time-varying linear systems.For the sensor fault diagnosis of non-linear systems,a modified NARMAX identification structure was proposed.A method combining wavelet decomposition theory and feature extraction was proposed to identify different kinds of sensor faults.An on-line sensor faults diagnosis struture was designed based on system identification and wavelet decomposition theory,which was tested by a sinusoidal sensor fault signal.