Abstract:
Advances made in Advanced Driver Assistance Systems such as Antilock Braking Systems (ABS), have significantly
improved the safety of road vehicles. ABS enhances the braking performance and steerability of a vehicle under
severe braking conditions. However, ABS performance degrades on rough terrain. This is largely due to noisy
measurements, the type of ABS control algorithm used, and the excitation of complex dynamics such as higher
order tyre mode shapes that are neglected in the control strategy. This study proposes a model-free intelligent
control technique with no modelling constraints that can overcome these un-modelled dynamics and parametric
uncertainties. The Double Deep Q-learning Network algorithm with the Temporal Convolutional Network is
presented as the intelligent control algorithm. The model is initially trained with a simplified single wheel model.
The initial training data is transferred to and then enhanced by using a validated full-vehicle model including a
physics-based tyre model, a 3D rough road profile with added stochasticity. The performance of the newly
developed ABS controller is compared to a Bosch algorithm tuned for off-road use. Simulation results show a
generalizable and robust control algorithm that can prevent wheel lockup over rough terrain without
significantly deteriorating the vehicle’s stopping distance on smooth roads
