Dynamic response prediction method of long-span road-rail suspension bridge based on LSTM

For long-span road-rail suspension bridge, it is often difficult to establish a mapping relationship between loads and response due to the huge and mixed dynamic data output from the structural health monitoring system, thus hindering the adoption of reasonable maintenance strategies. A specific long-span road-rail suspension bridge was selected as the research subject. The dynamic response data from the health monitoring system were analyzed in detail, and the prediction performance of vibration responses (acceleration and deflection) under various loading factors (highway vehicles, railway vehicles, temperature, and wind speed) was investigated using the long short-term memory (LSTM) neural network method. The results indicate that the LSTM network can effectively predict the dynamic response of such bridges. Specifically, deflection can be predicted with only temperature load as input after slightly more than ten iterations, whereas vertical and lateral vibration acceleration require simultaneous inputs of temperature, wind, highway and railway loads and can be accurately predicted after approximately 400 iterations.