High-Angle-of-Attack Longitudinal Control of UAVs Using Physics-Informed Neural Networks and RBF-Compensated Nonlinear Dynamic Inversion

This study evaluates a longitudinal control method integrating a physics-informed neural network (PINN), nonlinear dynamic inversion (NDI), and radial basis function (RBF) compensation using an open-source F-16 dynamics model as a numerical benchmark. The pitch dynamics are embedded in the offline training loss as a physical constraint. During control operation, the trained PINN is evaluated at the current inputs with fixed weights and biases, and its predictions update the inverse model and control effectiveness terms. The RBF network provides online residual compensation. The PINN captures the trends of the pitching moment coefficient and its aerodynamic derivatives, providing model information for dynamic inversion. Compared with the implemented fixed-gain PID baseline, the model-informed controllers reduce angle-of-attack tracking errors and transient overshoot under the prescribed simulation conditions. RBF compensation further reduces the remaining tracking deviations. These results concern the prescribed numerical tracking test; they do not establish flight safety, loss-of-control prevention, emergency stall recovery, or performance on a civil aircraft. The stability result is conditional on a bounded operating region, a uniformly bounded residual, effective control authority, and the stated actuator realization conditions.