Every year, 1.19 million people worldwide fall victim to road accidents. In an effort to increase safety, car manufacturers develop their vehicles to minimize the chance of injury in staged crash tests. For restraint systems, two settings are decisive for optimizing the crashwortiness: the active vent of the airbag and the seatbelt tightener.
However, hardware-based crash testing and finite element method simulations are too expensive to explore the entire design space of these settings.
As an alternative, researchers have been employing deep learning to predict crash outcomes based on the one-dimensional acceleration signal measured at the B-pillar of a vehicle during a crash test.
This thesis researches how multi-variate signals - accelerations, forces, and deflections - measured in different parts of the test dummy can be modeled to predict not only a specific variable of interest but a holistic crash safety metric.
A novel convolutional and attention-based architecture, called Crashformer is proposed, that achieves a 15% reduction in prediction error and a 400% increase in computational efficiency over existing methods.
The methods are presented as a recommendation framework showcasing how such methods can be used to suggest optimal restraint system settings to engineers.
The framework and models are evaluated on hardware-based crash tests. Additionally, different methods to increase the interpretability of the predictions are explored, and novel theoretical guarantees about the physical feasibility of predictions of PCA-encoded crash signals are derived.