In order to improve the design method of vehicle occupant restraint systems, it is necessary to reduce the computational load of simulations, to improve the global search capability, and to examine and integrate analytical methods to understand the complex interaction between design variables and objective functions. Therefore, in this study, we integrated the following three artificial intelligence technologies and applied them to the design of a vehicle occupant restraint system: (1) construction of a highly accurate approximate model by machine learning, (2) improvement of global search capability by evolutionary multi-objective optimization and (3) visualization and knowledge acquisition of multidimensional information using multivariate analysis methods. First, we obtained the minimum number of actual calculation samples using a crash analysis model with the design of experiments, and then used these samples to construct a highly accurate approximate model using machine learning. Next, we used the approximate model to perform a global search in the design space by evolutionary multi-objective optimization to obtain a pareto solution set that takes into account the trade-off relationship between the objective functions. finally, multivariate analysis using cluster analysis and self-organizing maps was performed on the pareto solution set. As a result, a fast global search was realized by substituting the evaluation calculation of evolutionary multi-objective optimization with a highly accurate approximate model. The pareto solution set obtained therein was then partitioned into clusters by cluster analysis, and the partitioned clusters were analysed by self-organizing maps, which provided perceptual information on the factors governing the trade-offs between the objective functions and the interactions between the design variables, and were useful for design engineers’ insights.
cluster analysis, evolutionary computation, machine learning, multi-objective optimization, self-organizing maps, vehicle occupant restraint system
 Obayashi, S., Jeong, S. & Chiba, K., Multi-objective design exploration for aerodynamicconfigurations, AIAA-2005-4666, 2005.
 Natori, S. & Yu, Q., An Application of CAP (Computer-Aided Principle) to StructuralDesign for Vehicle Crash Safety, SAE Technical Paper 2007-01-0882, 2007.
 Deb, K., & Srinivasan, A., Innovization: innovative design principles throughoptimization, Proceedings of the 8th annual conference on Genetic and evolutionarycomputation, pp.1629—1636, 2006.
 Horii, H., Estimate modelling for assessing the safety performance of occupant restraintsystems, WIT Transactions on the Built Environment, Vol.134, pp.627—635, 2014.
 Rasmussen, C.E. & Williams, C.K.I., Gaussian Processes for Machine Learning, MITPress, 2006.
 Sasaki, D. & Obayashi, S., Efficient search for trade-offs by adaptive range multiobjectivegenetic algorithms, AIAA Journal of Aerospace Computing, Information andCommunication, 2, pp. 44–64, 2005.
 Horii, H., Multi-objective optimization of vehicle occupant restraint system byusing evolutionary algorithm with response surface model, International Journal ofComputational Methods and Experimental Measurements, 5(2), pp. 163–173, 2017.
 Murtagh, F. & Legendre, P., Ward’s hierarchical clustering method: clustering criterionand agglomerative algorithm, Journal of Classification, 31(3), pp. 274–295, 2014.
 Kohonen, T., Self-Organizing Maps, Springer-Verlag, 2001.