Research On Fault Diagnosis Of Traction Inverter Based On Neural Network

Traction converter system,as the primary power of high-speed EMU and high-power electric locomotive,directly affects the safety and stable operation of the train.According to the statistical analysis of the number of failures of the elements in the traction converter system in a certain time,IGBT,the basic element of the inverter,has a high failure rate.IGBT faults can be divided into open circuit faults and short circuit faults.Short circuit fault will be open circuit fault because of the over-current protection.If open circuit fault occurs in IGBT,it is not only difficult to determine the specific location of the faulty device,but also will cause other faults,which will further damage the whole system.Therefore,it is very important to detect and locate the IGBT device with open circuit fault in time.Firstly,the function and control mode of IGBT power devices in traction inverter are analyzed.The model of traction inverter system is built based on matlab/simulink platform to simulate normal and fault conditions.This paper analyzes 22 open circuit conditions of up to two IGBT,further selects six typical faults,analyzes the characteristics of three-phase current output of inverter,and provides theoretical basis for fault diagnosis of IGBT.Secondly,based on the above analysis,the magnetic field current and torque current are obtained as the original data of fault diagnosis by the simulation model.Then,the wavelet packet energy entropy,wavelet energy coefficient and DC component of the two current are obtained by wavelet transform and FFT transform,which constitute 20 dimension fault eigenvector.The eigenvector can represent the circuit characteristics of different IGBT open circuit.Finally,for 22 kinds of working conditions,samples are obtained under different motor speeds,and used for BP and RBF neural network training and testing.The advantages and disadvantages of different feature quantities and classifiers are compared from the two aspects of diagnostic accuracy and sample training time.The simulation results show that the 20-dimensional eigenvector can reflect 22 different conditions,and the accuracy of RBF diagnosis is close to 100%,which is obviously better than BP network.