Research On Braking Friction Fault Diagnosis Of High-speed Train Based On Deep Feature Extraction

With the rapid development of high-speed railway technology,more and more attention has been paid to the braking system as one of the train components.Accurate and effective fault diagnosis of braking friction is crucial to ensure passenger safety and improve operation stability.As the most important part of the braking system,the disc brake will have various faults at the friction interface in the complex working environment,resulting in large economic losses and even casualties.Therefore,it is particularly urgent to achieve accurate and effective fault diagnosis of braking friction.As a direct reflection of braking performance parameters,brake friction state information has the characteristics of nonlinear,easy to be covered by noise and difficult to extract features,which improves the difficulty of brake friction fault diagnosis.In view of the characteristics of the above brake friction fault signals,this paper proposes a high-speed train brake friction fault diagnosis method based on deep feature extraction.The main contents are as follows:Firstly,aiming at the problem that it is difficult to obtain the fault samples of high-speed trains,the braking fault is simulated on the self-built high-speed train braking test bench.According to the characteristics of the friction signal,the friction coefficient,tangential vibration acceleration and vibration noise are selected as the friction state signals and collected.Then,the braking friction fault diagnosis model based on convolutional neural network(CNN)feature extraction is established.The validity of the model is verified by three kinds of collected signal data,which lays a foundation for later research.Secondly,aiming at the nonlinear and non-stationary problems of brake friction interface signals,considering the strong correlation between local signals,a brake friction fault diagnosis method based on Smoothness Priors Approach(SPA)and Residual Network(Res Net)depth feature extraction is proposed.Firstly,SPA is used to decompose the collected signals,and then the decomposed three signal components are transformed into images by matrix method.Then Res Net depth is introduced to extract image features to realize brake friction fault diagnosis.The optimal fault diagnosis model is determined by experiments on each super parameter in the method,and the feasibility of the deep feature extraction method is verified on the braking data.Thirdly,to obtain a more robust model,a dual-input deep spatial-temporal fusion network is proposed by comprehensively considering the local spatial characteristics and the temporal characteristics of the braking fault data.Based on the spatial features extracted by Res Net,1DCNN and long short-term memory network are introduced to extract the time features of the signal,and 2DCNN is used to fuse the two features to realize the brake friction fault diagnosis under noise conditions.Experiments on the effectiveness and anti-noise performance of the proposed network verify the superiority of the network in the diagnosis of noise environment.Finally,aiming at the problem of difficult feature extraction caused by unbalanced train fault samples,considering the correlation characteristics between fault samples,a brake friction fault diagnosis model based on graph convolution neural network is proposed.The model first uses the fault signals and the Jensen-Shannon divergence between faults to construct the graph data,then uses 1DCNN to compress and preliminarily extract the graph node characteristics,and then uses Graph SAGE to aggregate the graph data node characteristics to realize the brake friction fault diagnosis under data imbalance.Through the experiment of unbalanced data in various proportions,the characteristics of each layer are analyzed and compared with the unbalanced data processing method,which shows the effectiveness of the method.In this thesis,the friction block in the disc brake of high-speed train is taken as the research object.Aiming at the problems that the fault data of the friction block are difficult to collect,the signal is easy to be noised and the data are unbalanced,considering the excellent characteristics of deep learning in data feature extraction,the simulation fault experiment of the friction block is carried out,and the corresponding deep learning network is proposed to realize the fault diagnosis of the friction block.Through experimental analysis and comparison with other methods,it is proved that the proposed method is superior in the fault diagnosis of the friction block of high-speed train,and has certain practical engineering application value.