SiamMTS: Self-Supervised Representation Learning for High-Speed Train Traction System State Prediction

Abstract

Accurate prediction of the high-speed train traction system state ensures safe train operation. Currently, common prediction methods involve dividing the traction system into multiple equipment and establishing separate models based on multisensor signals for each. While effective, this method requires repeated training for different prediction tasks, consuming significant computational resources and limiting the flexibility of model application. Therefore, developing a universal learning framework for predicting the state of various equipment within the traction system is crucial. To this end, this article proposes SiamMTS, a self-supervised representation learning (SSRL) framework based on a Siamese network architecture for multisensor time-series signals. SiamMTS performs self-supervised learning by minimizing the distance between different augmented views of the same sensor sequence in the feature space, thereby extracting a universal time-series representation for improved state prediction performance from the multisensor signals monitoring the traction system. Experimental results demonstrate that SiamMTS performs well when processing datasets from multiple high-speed train traction systems. The encoder obtained during its pretraining phase provides reasonable initialization parameters for downstream tasks, enabling effective prediction of the state of various equipment within the system. Compared with the Supervised model with the same encoder architecture, SiamMTS reduces the average root mean square error (RMSE), mean absolute error (MAE), and mean absolute percentage error (MAPE) by 22.27%, 23.13%, and 33.21%, respectively, at a prediction step of 10 and by 15.31%, 16.67%, and 20.53%, respectively, at a prediction step of 20. In addition, the total computation time of SiamMTS in predicting the state of four traction system equipment is 41.31% of that required by the Supervised model.