A Deep Transfer Learning-Based Approach for Predicting Dynamic Responses of Vehicle–Bridge Coupled Systems Using Enhanced Neural Network Model

Neural networks have emerged as promising tools for dynamic response prediction of vehicle–bridge coupled systems. This paper proposes a deep transfer learning-based approach for dynamic response prediction of vehicle–bridge coupled systems. The feasibility of transfer learning between the 2D and 3D models is first established. Subsequently, the prediction model trained with datasets obtained from the computationally efficient 2D model is transferred to build the prediction model trained with limited data generated by the sophisticated 3D model for vehicle–bridge coupled systems. To facilitate transfer learning, an enhanced neural network model, i.e. CNN-BiLSTM-Attention consisting of feature extraction block, self-attention block, and temporal prediction block is put forth. The proposed approach is illustrated on four typical and significantly different vehicle–bridge coupled systems with field measurement data available in high-speed railway, heavy-haul railway, and urban rail transit lines. Results show that the developed transfer learning CNN-BiLSTM-Attention model can deliver accurate response time-history predictions with maximum absolute relative differences generally less than 10% across the scenarios. The transfer learning model reduces prediction error metrics by 68.5–81.1% compared to the model trained from scratch. The proposed approach provides a generalizable deep learning framework for dynamic response prediction of vehicle–bridge coupled systems with low cost and high accuracy.