Vehicle yaw stability control with a two-layered learning MPC

This paper proposes a novel framework to address the model-vehicle mismatch (MVM) challenge of yaw stability control, which introduces the Gaussian Process-based Model Predictive Control into the classical two-layer structure. In the upper layer, the supervisor controller calculates the compensation torque required for vehicle turning. Gaussian Process regression is employed to compensate model mismatch and unmodelled dynamics. The chance constraints on sideslip angle are converted into double half-space constraints by the normal distribution quantile function. In the lower layer, the subordinate controller utilises the nonlinear slip rate model to distribute torque to the wheels in the form of braking torque. A new assignment rule of braking torque weight is proposed according to the yaw torque analysis. The stability of the two-layered learning MPC is ensured through proof and simulation. In the Simulink and Carsim co-simulation environment, the superiority of the proposed method is verified in the task of rapid double-lane change. Compared with the traditional MPC-based yaw stability controller, the proposed method has increased the stability tracking index and energy consumption index by around 13% and 8%.