Real-time prediction of high-speed rail pantograph-catenary contact force using deep learning framework

In high-speed railway systems, the contact force is a crucial metric for describing the interaction performance between the pantograph and the catenary. However, measuring the contact force directly is challenging compared to other measurable quantities such as panhead acceleration and uplift. Leveraging advancements in artificial intelligence, this work proposes a deep learning framework to predict the contact force time history based on panhead acceleration and uplift data. Canonical recurrent neural network models, including vanilla RNN, LSTM and GRU are employed to model and predict the contact force. To enhance efficiency by reducing model parameters and extracting sequential feature maps, we propose a hybrid GRU model integrated with a one-dimensional convolutional neural network, called CNN-GRU. The database for this study is built using a validated and reliable finite element model of the pantograph-catenary system, incorporating variations in structural parameters, speed levels and cut-off frequencies. Four training strategies are designed to evaluate prediction performance across these variations. The results demonstrate the potential of using deep learning to predict contact force based on panhead acceleration/uplift, with some limitations in predicting across different cut-off frequencies. While vanilla RNN, LSTM and GRU models achieve acceptable prediction accuracy, the proposed CNN-GRU outperforms them, particularly in terms of prediction consistency across varying train speeds.