PhyRL-PINN: Full-scale adaptive physics-informed ship sideslip estimation via reinforcement-learning-guided parameter scheduling

Estimating a navigation-derived proxy of the ground-referenced ship sideslip angle under real-sea conditions requires combining physical structure with data-driven flexibility. PhyRL-PINN is a physics-informed residual-regularized neural estimator with reinforcement-learning-guided coefficient scheduling. A reduced-order lateral-dynamics prior is retained, while an actor–critic scheduler infers regularized equivalent structural coefficients from causal measurement histories. Within each 12-s window, six coefficient vectors are inferred at consecutive 2-s subinterval endpoints for endpoint-conditioned physical reconstruction, whose terminal state conditions the final estimate. Training and evaluation are performed offline on recorded measurements. The supervisory target, reconstructed from GNSS/INS speed over ground, course over ground, and heading, is a ground-referenced sideslip proxy rather than a direct water-relative hydrodynamic measurement. Using 59 voyages, 40,830 s of operation, and 163,379 synchronized samples, PhyRL-PINN achieves an RMSE of 1.74 × 1 0 − 3 rad and an R 2 of 0.9911 under seen-voyage blockwise interpolation, corresponding to RMSE reductions of approximately 78% and 60% relative to TimeMixer and MMAE, respectively. Under five-fold voyage-level grouped cross-validation, it retains an R 2 of 0.9341 on held-out voyages. Ablation results confirm the contributions of physics-residual regularization and state-dependent coefficient scheduling.