To improve the utilization of land resources, in recent years, there has been an increase in the design and construction of over-track buildings in metro depots in China. However, there are daily tasks such as testing, parking, and maintenance of metro trains in the depot, vibration induced by trains in those tasks may exert impact upon the residents living in the buildings over the metro depots. In this paper, detailed on-site measurements were carried out to investigate the source and transmission characteristics of train-induced vibrations in the metro depot. The main results indicated that (i) the magnitudes of train-induced vibration on the testing lines and throat areas are greater than other areas, and the issue of environmental vibrations induced by trains in the building near these areas need to be given more attention. (ii) Train-induced vibrations are more attenuated from the ground to the platform and less attenuated from the platform to the building floor, and most vibration acceleration levels below 30 Hz are even amplified during the transmission from the platform to the floors in the building, which could be because of the matching of the frequencies of train-induced vibrations and the natural frequencies of the superstructure.

To deal with the issues of high computational cost and prediction uncertainty of numerical models in train-induced ground-borne vibration prediction, a prediction method based on transfer learning is proposed in this study. In this method, the vehicle–track-coupled analytical model and three-dimensional finite element model are first used to calculate the train-induced ground vibration under various condition variables, and these data were used as training samples to pre-train the deep neural network models. Numerous train-induced ground vibration experiments were then conducted along the metro lines in Beijing, and those measured vibration data were used to fine-tune the pre-trained deep neural network model with the transfer learning strategy. A random variable obeying a Gaussian distribution is assumed over the predicted vibration acceleration levels to model the randomness of train-induced vibration, and the parameters of this distribution were determined by the statistical results of vibration monitoring data in the metro tunnels. The fully trained model could complete the prediction of train-induced ground vibration in seconds. Finally, a case study was carried out, by comparing the probabilistic prediction results with the statistical results of the field measurements, and the feasibility and the improvement of the proposed method were demonstrated.