Deep learning-based inverse prediction of side pole collision conditions of electric vehicle

Abstract

To improve safety of electric vehicles under side pole collisions, accident reconstruction and failure risk prediction on battery cell are essential. Accident reconstruction and analysis is complex due to structural nonlinearities and vehicle rotation during collision. Such task becomes more challenging due to the ill-posedness of this inverse problem. This study proposed a deep-learning based method to inversely predict the collision conditions when only the deformation of battery pack exterior structure is available. Battery cell deformation was also predicted to assess the accident severity. To build the dataset, a large number of finite element simulations were run at pack level. Compared to the comprehensive coverage of collision condition domain, the collision response domain inevitably exhibits poor filling, leading to non-uniqueness in inverse prediction. To address this, the model was trained with pre- and post-collision images of the side structure. A convolutional neural network integrated with residual network (ResNet) was applied to improve model performance. The amount of input feature information and the network structure were thoroughly discussed. The model also demonstrated good interpretability and robustness, maintaining stable performance with added noise. This proposed approach would become an effective tool for analyzing collision scenarios where limited information is available.