Fault Prediction Of High-speed Train Bearings Based On A Hybridized Physical And Data-driven Temperature Model

As a key component connecting axle and frame,axle bearings of the high-speed train are essential for the safe operation of modern high-speed trains.The current on-board detection system,which compares the axle box temperature measured from the sensor to a predefined threshold value,introduces problems such as poor adaptability to various service conditions,undesired high rate of false alarms and missing alarms.To solve those problems,in the present study,a novel physical-data hybrid model is proposed for predicting the axle box temperature and identifying the bearing fault.The hybrid model includes a physical model reflecting the heat transfer mechanism between the axle box and the environment,and a data-driven model adopting the same inputs but representing the relationship by means of back-propagation(BP)neural network and long-short term memory neural network(LSTM).By training and testing the established model on operation dataset collected from complex working conditions,the fault prediction accuracy is increased by 21.2 percent compared to the current on-board system and by 8.7 percent compared to the best single model.The study shows that in addition to the enhanced predictive accuracy,the hybrid modeling strategy has the advantage of adapting to different operating environments and pinpointing whether the fault is in a bearing or a temperature sensor.The main contents are as follows:(1)The characteristics of collected original train operation data are large in amount but poor in quality.To exclude the influence caused by missing values or inconsistent sampling frequency,the data was first preprocessed by means of cleaning,compensation and filtering.After preprocessing,according to train operation characteristics,the data is subdivided into several segments using a technique called window segmentation method.Finally,the change trend and distribution of healthy bearing data and fault data are analyzed,and the correlation between variables is analyzed qualitatively and quantitatively.(2)To simulate the thermal behavior of the axle box in a normal state,heat generation and heat dissipation mechanism of axle box are analyzed.The development mechanism of axle box temperature is analyzed by analytical method,and the unknown parameters of physical model are calibrated by numerical method.Based on the healthy dataset,genetic algorithm and least square method are employed to calibrate parameters of the physical model.The thermal behavior of normal axle boxes can be accurately modeled and used as a reference for the identification of the abnormal friction heat.(3)To compensate for the low prediction accuracy of a single physical model,a data-driven model with stronger nonlinear fitting ability is established based on machine learning method.By analyzing and comparing the applicability of common machine learning methods,BP neural network and LSTM neural network models are finally established.Network structure and hyper-parameters of data driven models are determined by systematic trial-and-error procedures.The neural network model is trained on the healthy dataset to achieve higher precision prediction effect of axle box temperature.(4)The discrepancy between the predicted and real temperature at the axle box is analyzed,and the bearing fault discrimination strategy is formulated by using the fault discrimination method based on statistical process control.The paper compares results obtained from each individual model,proposes a hybrid mechanism,which comprehensively considers the prediction results of physical model and data-driven model to discriminate and predict axle bearing faults,and demonstrates how such a hybrid model can improve fault prediction accuracy through the instance.