A feed-forwarded neural network-based variational Bayesian learning approach for forensic analysis of traffic accident

Abstract

In this work, a variational Bayesian learning-based computation algorithm is developed to “inversely” identify the deformation field of a crashed car and hence their residual strain fields based on its final damaged structural configuration (wreckage), which is important in three-dimensional traffic collision reconstruction and its forensic analysis. Different from our previous Generalized Bayesian Regularization Network (GBRN) algorithm (Xie et al. [2002] Computational Mechanics 69, 1191–1212), the present method is based on Variational Bayesian Learning theory coupled with a Feed-forward Neural Network architecture, and it provides a higher computation efficiency. This is because it requires less number of iterations and produces more accurate registration results, since the locality of nodes are greatly preserved during registration process. In this work, we have demonstrated that the developed machine learning algorithm has a unique capability to practically identify the deformation field of a real crashed car and to recover its initial pre-crash state based on residual damaged geometric configuration, and it shows great potential in forensic analysis of car crash and vehicle crashworthiness evaluation.