| At present,high-speed railway is the first choice for people to travel a long distance.The large and complex motor traction system in high-speed trains is a high incidence of faults.A large number of components in the system interact with each other.And the chain reaction of and multiple single faults will result in compound faults,which bring difficulties to fault diagnosis.Therefore,taking the traction motor system in high-speed trains as the research object,this paper studies the diagnosis methods of the compound faults occurring in the current sensor and traction motor,which are two important components of motor traction system.The specific research contents are as follows:Firstly,considering that the features of rotor bar broken and gap eccentricity compound fault in traction motor are masked by the noise and the fundamental frequency of the power supply,a compound fault diagnosis method based on extended Park vector method is proposed.The current signal is decomposed and reconstructed by the improved EMD decomposition method through median filter to reduce the interference of noise.Then,by using the extended Park vector method,the three-phase current signal is converted into Park vector mode according to the coordinate transformation principle,so that the fundamental frequency of the power supply is converted into the dc component,which highlights the compound fault characteristics.By using fast Fourier transform method,the feature vector is extracted from the spectrum of Park vector module to diagnose the compound fault.Finally,the feature vector is used to train the decision tree classifier to estimate the severity of the compound fault.Based on the data of fault injection semi-physical simulation platform of high-speed railway traction system in Zhuzhou Electric Locomotive Research Institute,the feasibility of the diagnosis method is verified.Secondly,a diagnosis method based on the improved BP neural network for the compound fault of double current sensors bias on the output side of the inverter is proposed,considering that the time domain characteristics of the signal are not obvious when the degree of the compound fault is small,and the single sensor bias fault and double sensors bias compound fault are difficult to distinguish.At first,Savitzky-Golay filter method is used for smoothing the stator current to reduce the interference of noise.Then the wavelet packet transform is used to obtain the detailed information of each frequency band of the signal,from which the compound fault characteristic vector is extracted,which is used to train the BP neural network classifier to realize the diagnosis of different fault modes.The method of additional momentum factor and iterative step-size self-adjustment are adopted to optimize the network weight adjustment,which improve the rapidity and stability of BP neural network,as well as the accuracy of fault diagnosis.The data collected from Zhuzhou semi-physical simulation platform is used to verify the validity of the diagnosis method.Finally,a compound fault diagnosis method based on random forest and XGBoost is proposed to solve the motor stator turn-to-turn short circuit and gap eccentricity compound fault.The wavelet packet transform is used to extract the characteristic information of the total current of U phase and V phase to form the fault feature vector.Principal component analysis is used to reduce the dimension of the feature,and the optimized feature vector is used to train the random forest classifier,which improves the prediction accuracy and generalization performance of the random forest model.The feature sets of high importance screened by random forest are used to train XGBoost classifier,which further improves the accuracy of compound fault diagnosis.The diagnosis method is applied to the vehicle-mounted experiment of the traction motor in CRH2 high-speed train in Zhuzhou,which proves the feasibility of the method. |
