An artificial intelligence framework for vehicle crashworthiness design

Abstract

Numerical crash test simulations are crucial for vehicle safety design. In the automotive industry, frameworks based on finite element methods are most common as they are precise and reliable. A few of the setbacks are simulation time and computation resources required for simulation. This thesis presents an artificial intelligence framework that utilizes recurrent neural networks to reduce the time and computational resources required to predict axial crash tests on the LS-DYNA models of thin-walled UWR4-like aluminum extrusion profiles. In addition, the work provides an overview of several data preprocessing techniques aiming to improve framework training time; ensembling of neural networks for the framework is explored as an addition to data preprocessing to improve framework performance. The thesis includes a detailed description of the data used and the machine learning models utilized in the framework. Threedifferent sampling techniques are compared to reduce the time required to train the framework– two variants of random sampling and importance sampling; model ensembling is explored to improve accuracy on framework trained on data samples. Experiments show that the artificial intelligence framework reduces the time required to obtain one simulation of an axial crash test by the factor of 270, with a tradeoff of accuracy. Additional experiments on data preprocessing and model ensembling show that the training time of the framework could be reduced from 111hours to 37 minutes for a single sample or 3 hours for a models ensemble with an additional cost of accuracy.