Determination Of Metro Train-induced Vibration Source Intensity Based On Machine Learning Method

With the vigorous development of urban rail transit construction in China,the environmental vibration problems arising from metro operation are becoming more and more prominent.A reasonable vibration mitigation design has put forward higher requirements to the environmental vibration prediction.At present,the empirical chain formula is mainly used to predict environmental vibrations in the feasibility stage of a metro project,and the vibration source intensity is the most important basic value in the formula.In view of the limitations of the current vibration source intensity prediction,this thesis proposes to apply the machine learning method to the prediction of vibration source intensity.The main research contents and results are as follows.(1)Establishing the vehicle-track-tunnel-soil model.The multi-body dynamics software SIMPACK and the finite element software ABAQUS were used to establish the vehicle-track and the tunnel-soil models,respectively.The train-induced acceleration response at the tunnel wall were calculated,and the two-position calibration method was used to compare the numerical calculation results with the in-situ measurement results,to ensure the accuracy of the numerical model.(2)Analysis of factors affecting the vibration source intensity.Four variables were considered: wheel wear state,rail wear state,vehicle speed and soil dynamic elastic modulus.A total of 432 sets of data were calculated by the proposed numerical model and the calculated basic database was used for machine learning training.The influence of each variable on vibration source intensity was analyzed on the basis.(3)Machine learning model design.In the data pre-processing stage,the One-Hot Encoding method was used to consider the qualitative indicators;in the prediction evaluation stage of machine learning model,the K-fold cross validation and the LeaveOne-Out method were used to eliminate the influence of the training set data selection on model evaluation.Two machine learning methods,support vector machine and BP neural network,were selected for model training and optimisation,and the absolute errors of predicted and measured values are all within ±2 d B,and the average absolute errors are0.59 d B,which meets the specification requirements and can be applied to actual engineering prediction.(4)Transfer learning of BP neural network.Using the Transfer learning method,the pre-trained neural network model was optimised twice using the measured data to make it better adaptable to the measured data features.The method integrates the advantages of rich variation of numerical calculation data and representative characteristics of measured data,and provides supports for the application of machine learning to practical engineering prediction.