Applicability of using machine learning to improve computational efficiency of combined pre-crash and crash simulations
Vehicle occupant safety is commonly evaluated in dynamic finite element simulations. Common finite element occupant models are human body models (HBM), active human body models (AHBM), and dummy models in pre-crash and crash simulations are time-consuming, and a simulation may cost dozens of hours or even days, which will affect economic efficiency. Machine learning offers a potentially time-efficient approach with high accuracy. This thesis project aimed to compare finite element (FE) simulations and machine learning (ML) prediction results to determine the applicability and efficiency of the ML method for pre-crash and crash simulations. The Hybrid III fast finite element model was always used as an occupant model in combined pre-crash (braking or steering) and crash simulations. Six parameters related to occupant kinematics and safety system design were varied. Pre-pretensioner force, activation time of pre-pretensioner, braking level, braking duration, steering level, and steering duration in 200 FE simulations. From these FE simulations head kinematics, rib strains and seatbelt forces were extracted as output. Using the commercially available ML software, LUNAR, corresponding curves were predicted with algorithms that LUNAR suggested. Then the prediction accuracy was evaluated by calculating mean square error (MSE), root mean square error (RMSE), mean absolute error (MAE) and mean absolute percentage error (MAPE) between actual FE- simulation results and ML-model predictions. The conclusion shows that ML can provide a method that saves more than 19% of the time to obtain highly accurate results for head node displacement, rib strain and belt force prediction in both braking and steering simulations, while for head node acceleration especially in in-crash phase, it has low accuracy.