Fault Prediction For Train Braking System Based On Multivariate Time Series Feature Enhancement

With the development of the economic,the railway transportation is becoming more and more important.As a result,it has become a top priority to improve the transportation capacity of heavy-haul trains,and the brake system has become the key to ensuring the safe and stable operation of trains.Electro-pneumatic braking system is usually used in heavy-haul trains.The braking system is complex in structure and components,what's more,the components interact with each other and the sensor data of each component changes with time,which makes the fault mode diversified and difficult to excavate.Therefore,it is important to effectively warn the failure of the brake system,prevent the occurrence of faults,and reduce unnecessary personnel and property losses.At present,artificial experience and postmortem analysis are the main ways to predict the fault of heavy-haul train braking system,which lead to misjudgments in the case of large data.In addition,the braking system data shows the characteristics of multivariate time series,and the fault mode fluctuates with time.Therefore,the existing time series prediction method should be further strengthened in the mining of multivariate failure modes.This thesis proposes a braking system fault prediction algorithm named MTSFE based on multivariate time series feature enhancement and extreme random forest model.The main work of this thesis are summarized as follows:(1)First,the data of the actual brake control unit is cleaned.Then,the brake system monitoring data is represented in a multivariate time series to analyze the importance of each variable on the predicted output system status value.Finally,in order to reduce the horizontal dimensionality of real data and the complexity of the prediction model,an adaptive LASSO method is proposed to select the corresponding variables of important components.(2)The size of sliding window is designed for data with important variables selected to divide into data segments.Then,the time domain,frequency domain and wavelet packet features are extracted with statistical analysis and wavelet packet decomposition method through the time sequence characteristics and the correlation of these data segments using.Finally,the gray correlation analysis method is used to analyze the correlation degree between the features and the system state values according to the fault features extracted from data segments,and the features with higher correlation degree are selected as the optimal feature set of the prediction model.(3)Based on the proposed feature set of multivariate time series,the fault prediction model and algorithm MTSFE of the extreme random forest(ERF)model are proposed,and the model is optimized by continuous parameter optimization.In this thesis,the real data of a certain type of electro-pneumatic braking system are used as training data and testing data to train and predict the proposed model.Experiments in this thesis include the test of the influence of feature number and feature enhancement on prediction accuracy,and comparisons with commonly used prediction algorithms.Experiments show that the proposed MTSFE is superior to other common prediction algorithms in terms of multiple evaluation indicators.Finally,the MTSFE algorithm is implemented.As the background prediction engine of the brake fault prediction system,the MTSFE algorithm can effectively assist the field personnel in the brake system fault prediction.