Accelerated Virtual Evaluation of Restraint System Performance for Vehicle Occupants with Varying Body Shapes
This thesis describes the application of machine learning in the context of development and assessment methods in the domain of safety systems for occupant safety. Supervised learning methods are applied to generate rapid predictions for the results of finite element simulations in the domain of the assessment of occupant safety systems. The key concept revolves around the question of how a metamodel could be designed without explicit characterization of the vehicle. The base machine learning architecture uses the result of finite element simulation as a reference to predict the result of a finite element simulation in the same vehicle configuration but different occupant anthropometrics. The input for the metamodel is of the type multivariate time-series. Different strategies to deal with this data type are compared. Starting from a detailed finite element model of the Honda Accord 2014, using an incremental and structured approach, a simplified version of the finite element model is derived. Each simplification step is reasoned and evaluated regarding the loss of the model quality in relation to the gained improvements in computation time. Databases featuring different designs of experiments and the anthropomorphic test device Hybrid III and the human body model VIRTHUMAN are generated. The influence of different database sizes and an adaptive querying strategy is studied. Transfer learning approaches as a method to deal with limited data are discussed as well. The homogeneous transfer between Hybrid III and VIRTHUMAN and the direct cross-domain prediction between them is investigated.
