| At present,the scale of China’s high-speed railway network is getting larger and larger,and the number of high-speed trains continues to grow,so the study on fault diagnosis and prediction of train data has become more valuable.Braking system is relatively easy to occur on high-speed train fault of one of the subsystems,and braking system structure is too complex to build simple linear mapping relation bet ween the states and faults,so the traditional fault prediction methods are difficult to comprehensively study the fault characteristics and can be easily affected by cumulative error.As a result,the medium-term and long-term forecast effect is not good,and it is difficult to meet the actual needs.This thesis carries out the research on the braking system fault prediction method based on high-speed train historical state data,which mainly includes the following contents:(1)Based on data,we divided braking system faults into eight fault labels.At the same time,statistical method is used to screen the historical status data,and some features that have correlation with braking system faults are obtained as the data basis for subsequent research.(2)A fault diagnosis method based on one-dimensional CNN is proposed.In addition to the status data at the current moment,the input of the model also go back the time window size and inputs the historical data within the time window.In this way,CNN can be used to extract features of historical states in time dimension to obtain more information.The result of the experiment shows that the accuracy in the braking system fault label recognition scene reaches 87.6%,which is 5.5% higher than that of the MLP.(3)Traditional fault prediction methods can only maintain high prediction accuracy in short-term prediction,and the long-term fault prediction effect is often affected by the accumulation of errors,and it is difficult to achieve good performance.To solve this problem,a fault prediction method based on Informer is proposed in this thesis.The generative Decoder model in this method can generate multi-step prediction results at one time,avoid error accumulation effect,and improve the performance of medium-term and long-term prediction.In this thesis,LSTM and other models are used to design comparative experiments,and the result proved that Informer has significant advantages in fault label prediction for 30 minutes and above.(4)A vehicle-ground cooperative fault prediction scheme is proposed.On the one hand,it uses its richer data resources on the vehicle side for short-term online fault prediction,and on the other hand,it trains and deploys offline models in the ground data center for medium and long-term fault prediction.This thesis constructs two online fault prediction models based on the output fusion strategy and feature fusion strategy on the basis of the auto-encoder.After experiments,these two models have better prediction effect and faster calculation speed in online scenarios.In addition,this thesis develops an offline model distributed deployment system,through which the trained model can be easily deployed and used,and the model can be distributed to the computer cluster to make full use of the computing resources of the cluster.This thesis compares the advantages and disadvantages of the network architecture with registration center and the network architecture without registration center through analysis and experiments.The conclusion is that the network architecture with registration center is more suitable for this scenario.There are 39 pictures,19 tables and 59 references. |
