Kelvin’s model is widely used to simulate the dynamic characteristic of a resilient mat under a slab track. To develop an effective calculation model for a resilient mat using a solid element, a three-parameter viscoelasticity model (3PVM) was employed. With the help of the user-defined material mechanical behavior, the proposed model was implemented in software ABAQUS. To validate the model, a laboratory test was performed on a slab track with a resilient mat. Then, a finite element model of the track-tunnel-soil system was built. The calculation results using the 3PVM was compared with those using Kelvin’s model and the test results. The results indicate that the 3PVM can better reflect the dynamic characteristics of resilient mat than Kelvin’s model, especially over 10 Hz. Compared with the test results, the 3PVM has an average error of 2.7 dB and a max error of 7.9 dB at 5 Hz.

The tunnel vibration level is usually employed as a vibration source intensity of the empirical prediction method. Currently, the analogy test and data base are two main means to determine the vibration source intensity. To improve the accuracy efficiency, the machine learning (ML) method was introduced to predict the tunnel vibration responses. To acquire model training samples, the measurements were performed in 80 different running tunnel sections of Beijing metro lines. Two types of method, back propagation neural network (BPNN) and generalised regression neural network (GRNN) were employed, which can make full use of characteristics of measured samples and reduce the data noise. The results indicate that the prediction efficiency is high and the mean square errors of the two ML methods are acceptable. Accordingly, both of the ML methods can be used as the reference of vibration source intensity in metro train-induced environmental impact evaluation. GRNN has relatively better predicting ability than BPNN.