Research On Fault Prediction Method Of Axle Box Bearing Of Highspeed Train Based On Machine Learning

With the rapid development of China's high-speed rail,the safety of train operation is particularly important.At present,China's high-speed trains mainly monitor the temperature status of axle box bearing,the key component of the running part of emu,and give early warning based on the set threshold,which has the following defects: poor adaptability and high false alarm rate;the judgment of the fault mainly depends on expert experience;the data accumulated by the train is not fully utilized..In terms of repair system,fixed-mileage maintenance and replacement are commonly used at present,which may lead to under-repair,over-repair or other situations.In view of the above problems,comprehensively considering the effect of the actual train running time,train speed,environmental temperature and other factors on the axle temperature and the relationship between them,the machine learning method is applied to carry out deep mining processing on the monitored data in this paper.Finally,a random forest method combined with comprehensive sampling is proposed,through the comparison of various improved methods and models,to realize the fault prediction of axle box bearing.The main work is as follows:(1)According to the characteristics of high dimension,large amount,time series and low data quality of train operation and maintenance data,the data preprocessing method is determined.Firstly,the relevant feature data is selected,cleaned and integrated Then these data are divided according to the time window.Finally,the features are selected,extracted and constructed in the divided time windows.Meanwhile,they are associated with the fault records to attach corresponding labels to each group of feature data.Finally,the feature data set is standardized and dimension-reduced to complete the production of the data set.(2)To solve the problem of data imbalance,improvement methods are proposed from data and algorithm.For data level,comprehensive sampling method is adopted.First,under-sampling is used for the normal data by active acquisition.Then,new fault data are generated by using the Borderline-SMOTE over-Sampling algorithm to achieve the balance of data set.For algorithm level,cost sensitive learning is realized by modifying the weights of different categories in the loss function.(3)The final fault prediction model of axle box bearing is proposed through comparative experiments.First,different machine learning models is set up by Sklearn,a machine learning toolkit based on Python language.Combining the two improved methods of data imbalance,a comparative experiment is established.Through the nested k-fold cross-validation method in the grid search,the different improved models are optimized and evaluated.Then,the random forest method combined with comprehensive sampling is selected as the fault prediction model of the axle box bearing.Finally,two untrained fault data are used to verify the selected improved model.The model successfully predicts the bearing fault 6-7 hours before the fault,which proves the validity of the model.The combined comprehensive sampling of random forest model can realize the fault prediction of axle box bearings of high-speed trains,which gives more time for the processing of fault reserve.This method can not only avoid the situation of train delay caused by emergency braking and forced low speed operation,improve the travel experience of passengers,but also guarantee the running safety of the train,and provide a reference for optimizing repairing system.