An Online Prediction Method for Axle Temperature of High-Speed Train Based on RNN-EWC

Abstract

AbstractThe axle is one of the most important parts of a high-speed train. If the axle temperature rises abnormally during the running of the train, it may cause a driving accident. At present, most of the predictions of train axle temperature adopt offline models, but this offline model has many limitations in real train operation scenarios, such as untimely and inaccurate predictions of sudden changes in axle temperature caused by emergencies. This paper proposes an online prediction model based on continuous learning method, namely RNN-EWC, which uses recurrent neural network combined with the elastic weight consolidation (EWC) in continuous learning method to predict train axle temperature in real time. When the model is updated after a new sample of axle temperature data arrives, the EWC method will calculate the more important weights in the network, retain important “knowledge” and alleviate the long-term dependence of the RNN network. This method enable the network to have continuous learning capabilities and improve accuracy and real-time of predictions. Finally, through the comparative analysis of experiments, RNN-EWC can obtain significant prediction effect.KeywordsHigh-speed trainAxle temperatureOnline predictionRNNContinuous learning