Physics-Informed Neural Network Model for Trajectory Tracking Control of Unmanned Surface Vehicles

Rapid acquisition of an efficient and accurate system dynamics model is crucial for trajectory tracking control of unmanned surface vehicles (USVs) to meet diverse maritime mission requirements. To this end, this paper proposes a nonparametric modeling approach based on physics-informed neural networks (PINNs), where physical constraints derived from the physics model are incorporated into the training loss function to enhance physical consistency and generalization. An adaptive unscented Kalman filter (AUKF) is first employed to identify unknown parameters in the mathematical model of the USV for building the physics model through parameter identification, which are then used to guide the construction of the PINN model. Building on this, a novel integration framework of the nonparametric model with nonlinear model predictive control (NMPC) is developed to achieve accurate trajectory tracking control, without the need for the nominal model. The feasibility of the PINN-NMPC control method in this framework is validated by two simulation scenarios, including one with the measurement disturbances and another for the trajectory tracking task. The experimental results demonstrate that the proposed PINN-NMPC method has higher tracking accuracy and robustness compared with other integration approaches.