Wheel-rail force inversion via transfer learning-based residual LSTM neural network with temporal pattern attention mechanism

As urbanization progresses, metropolitan transit vehicles are encountering a growing frequency of curved pathways, which presents challenges pertaining to both the safety of the vehicles and the comfort of the passengers. There is no doubt that reliable acquisition of wheel-rail force is critical, since it has great significance for the safety and stability of vehicle operation. However, conventional wheel-rail force measurement methods are costly and difficult to use on in-service vehicles. A data-driven approach to inverting the wheel-rail force will overcome the above problems. In this work, a transfer learning-based residual long short-term memory neural network with temporal pattern attention mechanism (TPA-ResLSTM) is proposed to realize real-time monitoring of wheel-rail force, even in scenarios where the dataset is deficient in sufficient features. Initially, a learnable wheel-rail force inversion neural network model is developed based on the physical relationship that exists between the wheel-rail force and acceleration. Subsequently, a dynamic model for a B-type metro vehicle is utilized to simulate various scenarios, serving as a virtual source to provide data for the neural network. Afterward, the performance of the model is synthetically validated by the ablation study and field experimental data. Finally, the deep learning model is further improved by the transfer learning network, whose performance is comprehensively evaluated using limited data. The results show that the inversion model still has remarkable accuracy, in which the coefficient of determination is more than 0.9, under the case of limited training data. The proposed methodology diminishes the data requirements for the network while facilitating real-time monitoring and feedback regarding wheel-rail forces, thereby enhancing the realism of operational safety assessments for trains.