Research On Bearing Temperature Deep Anomaly Detection Of High-speed Train Based Onprediction

Bearing of high-speed train is an important part of train bogie,which bears complex excitation and changeable working conditions.Bearing temperature,as one of the state variables to measure the bearing condition,can avoid bearing failure threatening the train safety by setting threshold.Therefore,the prediction and monitoring of the bearing temperature in operation,the establishment of a real-time bearing temperature model,with the aid of early warning decision-making,can effectively improve the safety of the bearing operation and early diagnosis of fault.However,the bearing condition monitoring big data collected by railways presents multidimensional characteristics,and lacks typical bearing fault type data.It is difficult to early warning of faults of key components such as bearings,and there are few studies on them that explore the key points of each bearing the linear and non-linear coupling relationship between the components and the condition monitoring data of each axis.To this end,this thesis carries out related research,the main research work is as follows:(1)A bearing temperature prediction method for high-speed train based on complex correlation is proposed.Firstly,train history data is analyzed,and Pearson correlation coefficient is used to analyze the correlation bearing with high linear correlation.The data with low linear correlation and correlation bearing data are input into Light GBM model,and the features that affect bearing temperature are filtered again to reduce dimension,so that the model can learn more valuable features.Secondly,based on the deep neural network model,the bidirectional Gated Recurrent Unit is realized the bearing temperature prediction model is established by the ring unit,and the prediction model is trained by a large number of normal bearing history data;finally,the method is verified.It is concluded that the short-term prediction strategy and the method of correlated measurement points can effectively improve the accuracy and stability of the model.(2)A method for detecting abnormal bearing temperature of high-speed train based on Multi-Task Learning is proposed.First,the associated bearing temperature under normal operating conditions is used as the model input to construct a bearing temperature prediction model.When the actual temperature is abnormal,the correlation between the predicted value and the actual value presents an abnormal change.Secondly,with the introduction of a MultiHead Attention mechanism,the constructed model can simultaneously predict the temperature of nine bearings in three categories: axle box,gear box,and motor on one axle.Finally,an isolated forest model is established,and the results prove that the isolated forest model has the ability to detect anomalies on multiple bearings at the same time.(3)A method of bearing temperature anomaly detection for high-speed train based on confidence interval is proposed.Combined with the idea of uncertainty,the Aleatoric Uncertainty,Epistemic Uncertainty and Mixture Uncertainty are added to the model.The point prediction model is transformed into the probability prediction model,and the distribution probability of temperature is obtained.Through probability calculation,the bearing temperature anomaly detection is directly assisted by the model.The experimental results show that the probabilistic prediction model can realize the end-to-end direct detection,eliminate the accuracy loss of secondary modeling by collecting the residual between the predicted value and the actual value after the model is constructed,and achieve higher confidence.(4)A set of high-speed train bearing temperature anomaly detection and early warning system is developed.The system integrates the above research results,and can detect and warn the abnormal bearing based on the state detection parameters of high-speed train.The application of the system shows that the system achieves the expected design function.