Spatial Graph Attention Network for High-Speed Train Axle Temperature Forecasting

Bearings are one of the most vulnerable components in the operation of high-speed trains, and they need to be detected in real time to ensure timely maintenance decisions. The shaft temperature change can effectively reflect the working state of the bearing, and it is of great significance to predict the trend of shaft temperature change in real time for the adjustment of the driving strategy of the train. This paper proposes a spatial graph attention network for train axle temperature forecasting, where spatial features obtained from GAT are input into GRU and LSTM models for combined use. The results from the experiment reveal that the fusion model enhances accuracy by over 15% compared to the single prediction model, particularly with the GAT and GRU combination showing the best prediction performance. In addition, the proposed fusion model performs well in the long sequence prediction task, which significantly improves the accuracy and reliability of the prediction. This method can effectively predict the trend of shaft temperature change in advance, and ensure reliable technical support for the safe running of high-speed trains.