Data-Driven Dynamics Modelling And Simulation Research For Railway Vehicles

With the increasing advancement of the railway vehicle research, the vehicle dynamics related subjects based on the classical mechanics have got a completed and systematical development. Besides, the application of the information science, especially the big data related theories (e.g. the machine learning techniques), is undergoing an extraordinary growth in engineering fields, and the application of machine learning theories in the traditional vehicle dynamics has becoming a new field in the railway vehicle research. In recent decades, researchers firstly introduced the computer aided modeling to simulate the dynamic behaviors of vehicles, which has largely reduced the financial and labor cost from real experiments. To ensure the liability of the vehicle motion simulation, it is common to build a sophisticated finite element model, especially for the vehicle crash research. While the complex structure of mesh grids and the mesh distortion problem in big structure deformation simulation would undermine the computation efficiency of the finite element method. Another alternative is the multi-body dynamics method, which shows a low computation accuracy when involving the highly nonlinear characteristics in dynamics, though it has a satisfied calculation efficiency.To maintain the advantages of both the finite element method and the multi-body dynamics method, in this paper we proposed a data-driven simulation method for vehicle dynamics based on the machine learning theory. The main idea of this method is:taking the data from real experiments or finite element simulations as the training data where the mechanical characteristics is extracted to make a surrogate force element, and using the multi-body model of railway vehicles as the calculation unit, the co-simulation techniques can be utilized to carry out the vehicle simulation by embedding the surrogate element into the multi-body model. Based on that idea, this paper presented a theoretic framework of the data-driven modeling and simulation method to obtain a surrogate force element from training data, which includes the data collection and feature selection, the data compression and training data sampling, the data-driven model building (the Kriging model and Legendre polynomials regression model are introduced in this paper), the model evaluation and model selection.To assess the performance of our method, we carried out two case studies in simulating the vertical and longitudinal dynamics of railway vehicles, and compared the capability of two data-driven models, the Kriging model and the Legendre polynomials regression model, in the aspects of the convergence condition, the effect the size of training samples acts on the calculation accuracy, the time efficiency, the tolerability to different levels of track spectrum and the prediction accuracy of unknown cases. The calculation result shows that the Legendre polynomials regression model presented a better ability in expressing the mechanical characteristics of the nonlinear force element in vehicle dynamics simulation, which largely cut down the calculation time without compromising of the accuracy.