Identification of wheel-rail forces on high-speed railways based on physical model and hybrid recursive neural networks

Abstract

This paper aims to propose a practical framework for fast inversion and identification of wheel-rail forces of railway based on deep learning, with a view of achieving continuous measurement of the wheel-rail forces of the entire railway line in real time by means of the easily-measure vibration response of the train. For this purpose, a calibrated vehicle-track physical model is used to generate vehicle vibration data and wheel-rail forces, and a data-driven hybrid recursive convolutional and long short-term neural network (CNN-LSTM) and forward rolling prediction method based on recursive mechanisms is developed to facilitate the inversion of wheel-rail forces. Besides, a probabilistic point selection method is applied to compress the massive measured track irregularity data, subsequently, the representative samples are reconstructed to obtain reliable responses from the physical model, and then the variational mode decomposition (VMD) is introduced to extract feature components of the vibration signals for enhancing the capability of the deep learning model. Afterward, the identification performance of the presented framework for the wheel-rail force is examined and discussed by the different working conditions. The results indicate the hybrid recursive VMD-CNN-LSTM holds excellent robustness and generalization ability for the cases of running speeds and track states. Finally, the accuracy and efficiency of the identification framework based- vibration response is further elaborated through the 1 km measured track irregularity data. The correlation between the predicted value and the real wheel-rail forces reaches 0.99, with a time cost of only 4 s, which indicates that the presented framework has promising potential for the online real-time monitoring of wheel-rail force.