Research On Data Mining For Collision Characteristics Of Rail Vehicles

Considering the problems of high computational complexity,difficulty in precise determination of boundary conditions and material parameters,and the dependence of mesh quality in finite element analysis,a data mining framework for the study of collision characteristics of rail vehicles is proposed in this paper.The framework mainly includes an application of collision response prediction based on end-to-end data mining model and an application of structural optimization design based on data mining model.Based on the simulation results of finite element collision,a data mining model of collision characteristics about initial simulation conditions is constructed according to the specific application.In the application of response prediction,the construction of deep neural network(DNN)based data mining model,the evaluation and improvement of model accuracy,the regularization of model,the selection of model and cross-validation are discussed in detail.On this basis,aiming at the particular problem of rail vehicle collision characteristics,this paper also improves the loss function of non-equilibrium regression problem,model input features based on prior knowledge and aggregates model to improve generalization accuracy.In the application of structural crashworthiness optimization design,the optimization problems are defined firstly,then the general process of structural optimization design based on data mining model is discussed in detail which includes the experimental design method,the establishment of agent model,elitist genetic algorithm(EGA)optimization and simulation verification of optimization results.Based on the above framework,a data mining software platform for studying rail vehicle collision characteristics is developed in this paper.The software platform includes two application modules: rail vehicle collision response prediction and structural optimization design,as well as four basic modules for data reading,model preservation,model training and result visualization.In order to verify the feasibility of the data mining framework,this paper uses the selfdeveloped software platform to verify the cases of the two major applications.On the problem of response prediction,the ensemble model based on DNN achieves satisfactory precision in energy absorbing and anti-climbing device and single-vehicle collisions.On the optimization design of energy absorber,the DNN model is superior to the SVR model in both model capacity and generalization accuracy,and the optimization process combining DNN model and EGA algorithm can quickly converge to the global optimal solution.From the comparison of the simulation results of the energy absorber before and after optimization,the absolute approximate error of the DNN proxy model is less than 3%,and the crashworthiness of the energy absorber is improved.