Abstract: A digital twin method for predicting the safety performance of train curve negotiation is proposed to overcome these challenges posed by multiple-degree-of-freedom coupling modelling and the uncertainty factors analysis in traditional dynamics simulations, and to become more accurate and real-time. A digital twin for the safety prediction of train curve negotiation is built, and the dynamic safety indicators are visualized when a train passes a curved rail. The robustness and efficiency of the deep learning algorithm of MQRNN are helpful to extract features, simulate and predict the safety indicators of lateral acceleration of the frame, lateral force of the wheel shaft, the vertical force of the wheel and rail, as well as derailment coefficient in real-time. The results show that compared with the LSTM method, the proposed MQRNN method reduces the maximum error to 0.017 and 0.09, respectively, and gives prediction results with a 90% confidence interval, demonstrating its superior anti-interference ability. The proposed method can serve as a foundation for further digital twin-based decision-making of the train curve negotiation.

Key words: train curve negotiation, digital twin, deep learning, real-time predicted