Real-time assessment for running safety of high-speed railway based on physical models and deep neural networks

This paper proposes a method for real-time assessment the running safety of high-speed railways based on the field test, train-track-bridge dynamic analysis, and deep neural networks (N-BEATS). Firstly, the sensors are employed to obtain the acceleration of the bridge-track system under the excitation of the high-speed train. Then, the train-track-bridge coupled model was established and verified by comparing the calculated results with the field test. The derailment coefficient is considered an index for assessing the running safety of HSR. Therefore, the derailment coefficient at different running speeds were calculated using the coupled model, and the maximum of sixteen derailment coefficients were selected based on the position of the bogie. In order to match the derailment coefficient, the rail acceleration time-history data was divided into 16 equal parts according to the bogie wheelbase, and feature extraction was performed. The feature values of rail acceleration and running speed are selected as the inputs of the N-BEATS model, and the derailment coefficient is the output. The covariance matrix adaptation evolution strategy was used to find the optimal hyperparameter combination. The prediction performance and metrics at different running speeds prove that the N-BEATS model has high accuracy, robustness, and generalization. Finally, the prediction results of the N-BEATS model were compared with the commonly used machine learning models. Overall, this paper proposes a method that can accurately assess running safety in real-time, providing a reference for high-speed railway operation and maintenance decision-making.