Research On Fault Prediction Method For High Speed Railway Traction System Based On Bayesian Network

The scale of Chinese high-speed train network is now unthinkably massive.It imposes an increasing demand on the operation safety,and technically,it urgently requires effective fault prediction methods.Fault prediction of the key equipment in high-speed trains can not only reduce the daily maintenance cost,but also effectively prevent the occurrence of serious safety accidents.This thesis focuses on the traction system in CRH5 high-speed railway trains,and covers system modeling,fault injection,fault prediction,degradation modeling and remaining use life(RUL)prediction methods.The main contents are as follows.(1)A closed-loop traction system model based on bond graph theory has been built.On the basis of the open-loop bond graph model of CRH5 traction system developed previously by the project team members,the direct torque control module via bond graph modeling has been realized and then combined with the open-loop model to obtain the closed-loop bond graph model of CRH5 traction system.Common IGBT faults in the inverter and the broken rotor bar fault in the motor have been simulated and injected into the developed closed-loop bond graph model.The simulation results can verify the correctness of the model and the rationality of fault injection.(2)Fault probability prediction has been carried out by predicting the probability of Fault 2 when Fault 1 occurs.The Bayesian network structure for fault prediction is determined according to the closed-loop bond graph model,and the directed cycles are simplified to avoid causal conflict.Then,the Bayesian network is trained based on the fault sample data to predict the probability of the motor rotor resistance fault when IGBT fault in the traction inverter occurs.The simulation results verify the feasibility and correctness of the algorithm.(3)Degradation feature extraction of the key component and remaining useful life(RUL)prediction of traction system have been realized.The degradation process of the key component is divided into several health states according to the length of its remaining useful life(5 states are selected in this thesis).The system degradation features are extracted by combining data mining technology and restricted Boltzmann machine.The health status labels at every testing time are determined by clustering algorithm.Finally,the current health status and remaining useful life of the system are inferred by Bayesian network.The degradation processes of capacitance and resistance in the DC-link circuit are simulated on the semi-physical simulation platform of high-speed railway traction system in CRRC Zhuzhou Institute,and the feasibility and correctness of this method are verified with the help of the platform.