Study On The Intelligent Diagnosis Methods Of Composite Faults In Traction Systems Of High Speed Train

Traction transmission system is a key subsystem in high-speed train structure,which provides high-speed train power through electromechanical conversion.As a large and complex electromechanical coupling system,multiple faults in the traction system occur at the same time,which brings potential danger to the safe operation of high-speed trains.This paper focuses on the traction drive system of CRH2 high-speed trains.Based on the traction system fault data obtained on the hardware-in-the-loop simulation platform of Zhuzhou Electric Locomotive Research Institute,the diagnostic methods for the following three types of complex faults that are prone to occur in practical engineering are studied:(1)composite fault of current sensor gain and offset,(2)traction Motor rotor broken bars and inter-turn short circuit composite fault,(3)rectifier IGBT power tube open circuit and current sensor offset composite fault.The specific contents are as follows:With composite fault(1),the fault features in time-domain indicators are often fuzzy and coupled,such as mean,range,root mean square,and kurtosis.A combination model based on the combination of gradient boosting tree(GBDT)and logistic regression(LR)was designed considering the existing methods of manually selecting features with low efficiency and bad results.The original features are filtered and recombined through GBDT,and then classified using the LR classifier.Two independent combined models were used to diagnose the two types of single faults corresponding to the composite fault to complete the task of this type of composite fault diagnosis.With composite fault(2),the time domain characteristics of the fault are not obvious.The fast Fourier transform(FFT)is used to perform time-frequency domain conversion on the original signal,and the fault characteristic frequency ranges near the fundamental frequency overlap with each other.In order to obtain effective fault characteristics from the frequency spectrum,a four-layer deep convolutional neural network model(DCNN4)is used to extract the deep fault characteristics of the signal spectrum to achieve the purpose of feature decoupling.Then use the softmax multi-classifier at the fully connected layer to convert multi-label tasks into multi-class tasks to diagnose composite fault.The experimental results verify the diagnosis effect of the proposed FFT-DCNN4 model,which is significantly better than the diagnosis effect using time domain features and the ordinary convolutional neural network model.The composite fault(3)has a small sample data size,and ordinary artificial intelligence model training methods appear overfitting on this data set,which leads to poor generalization performance and low robustness of the fault diagnosis method.A composite fault diagnosis scheme based on transfer learning is proposed to solve the problem.This solution implements adaptive model migration by judging the maximum mean difference(MMD)between data samples.It can use other types of fault historical data similar to the target composite fault to compensate for the lack of target fault data to a certain extent.With appropriate similar conditions,the accuracy of compound fault diagnosis on a small sample data set is significantly improved.