基于近似模型的微型客车碰撞指标预测精度研究

Study on Prediction Accuracy of Collision Indicators for Mini-bus Based on Approximate Model

用近似模型来代替真实模型可以减少大量的计算时间,保证优化过程的可行性,但需要保证近似模型的精度满足要求。近似模型精度越高,优化的结果可靠性也就越高。基于微型客车正面40%重叠可变形壁障碰撞有限元模型,选取车身前部对碰撞性能影响较大的10个板料厚度作为变量,把B柱下端加速度、总质量、仪表板管梁处侵入量、前围板转向柱孔处侵入量及下前围板离合器踏板处侵入量作为响应值。利用拉丁超立方试验设计方法选取70个样本点,建立了设计变量与响应的近似模型,比较了响应面、径向基神经网络、Kriging、正交多项式4种近似模型响应的误差散点图、平均相对误差和决定系数。结果表明:响应面近似模型和径向基神经网络近似模型在B柱下端加速度峰值、前围板转向柱孔处及下前围板离合器踏板处侵入量的预测精度未在可接受范围内;正交多项式近似模型质量响应精度较高,但其余各响应的精度并不能满足要求,且3种近似模型预测精度受响应量与变量的线性关系的影响明显;而Kriging近似模型各响应的预测精度均满足要求,受线性关系影响较小,故选择Kriging近似模型代替原模型。最后用粒子群算法对Kriging近似模型进行优化,结果表明:Kriging近似模型预测结果与有限元结果拟合精度高,其优化结果达到预期目标。

In order to reduce a large amount of computing time and to guarantee the feasibility of optimization process,the approximate model can be used to replace the real model,but it needs to ensure the accuracy of the approximate model meet the requirements. The higher the approximate model＇ s accuracy,the higher the reliability for the optimization result. Based on a finite element model of the 40% frontal overlap deformable barrier impact of minibus,10 front car steel plate thicknesses which has much influence on collision performance are chosen as variables,the acceleration of B pillar lower end,the total mass,the intrusion volumes of dashbord beam,steering column hole on cowl panel and clutch pedal on cowl panel lower part are chosen as response values. Choosing 70 sample points by latin hypercube experimental design method,the approximate model of design variables and the response is built. The error scattergrams,average relative errors and decision coefficients of the responses of the response surface approximate model,radial basis neural network approximate model,Kriging approximate model and orthogonal polynomial approximate model arecompared. The result shows that（ 1） The prediction accuracy of response surface approximate model and radial basis neural network approximate model in peak acceleration of B pillar lower end and the intrusion volumes of steering column hole on cowl panel and clutch pedal on cowl panel lower part cannot meet the requirement.（ 2） The orthogonal polynomial approximate model has a high prediction accuracy in total mass,but other responses ＇ prediction accuracy cannot meet the requirement. The prediction accuracy of 3 approximate models is obviously influenced by the linear relationship between response and variable. The Kriging approximate model meets the requirement of each responses＇ prediction accuracy,and the linear relationship is less affected,so the kriging approximate model is chosen to replace original model. Finally,Kriging approximation model is optimized by using particle swarm optimization,the result shows that Kriging approximation model has high fitting precision as FEM,and the optimization result also reached the expected aim.