Transfer learning-based scoping prediction for train induced ground vibration using both simulated and experimental data

Rapid scoping prediction of train-induced ground vibrations along the route is crucial during the planning phase of a metro line. Recently, machine learning methods have emerged for rapid prediction of train-induced vibration. However, their accuracy is often hindered by the scarcity and high cost of experimental vibration data. Therefore, this paper introduces an unprecedented transfer learning-based approach to predict train-induced vibrations. This method leverages prior knowledge embedded in the numerical model to predict experimental vibration data, mitigating the reliance on extensive experimental data and improving prediction accuracy. Furthermore, a case study is conducted where models are trained, optimized, and validated using both simulated and experimental vibration data from the Beijing subway line. The findings indicate that the proposed transfer learning-based model outperforms models trained solely on experimental data, particularly under conditions where such data are scarce. This underscores the feasibility and advancement of the proposed approach to practical railway applications.