Multi-Objective Optimization Design of Vehicle Side Crashworthiness Based on Machine Learning Point-Adding Method

Abstract

Multi-objective optimization problems are often accompanied by complex black-box functions which not only increases the difficulty of solving, but also increases the solving time. In order to reduce the computational cost of solving such multi-objective problems, this paper proposes an ARBF-MLPA (Adaptive Radial Basis Function neural network combined with Machine Learning Point Adding) method, which uses an ABRF (Adaptive Radial Basis Function) neural network and OLHS (Optimized Latin Hypercube Sampling) to establish the first generation metamodel and uses the NSGA-II (Non-dominated Sorting Genetic Algorithm II) optimization algorithm to obtain the optimal front edge of Pareto. The ARBF-MLPA method is continuously used to select and add points to update the meta-model, then dynamically improve the accuracy of the meta-model until the optimal front edge converges. Then the ARBF-MLPA method and RBF-UDPA (Radial Basis Function neural network combined with Uniform Point Adding) method are compared using the test functions of three different frontier features. The performance evaluation indexes of Inverted Generation Distance (IGD), Hypervolume (HV) and Spacing Metric are superior to RBF-UDPA. Finally, ARBF-MLPA method combined with the NSGA-II optimization algorithm is applied in the multi-objective optimization design of vehicle-side crashworthiness. The model converges after 6 iterations. Comparing the results obtained by the ARBF-MLPA method with the finite element simulation results, the error is within 5%, which meets the error requirements. The optimized model reduces chest intrusion by 4.32%, peak collision force by 2.11% and reduces mass by 14.05%.