Improved Self-Adaptive Physics-Informed Neural Network for Ship Motion Response Model Parameter Identification

Abstract Accurate identification of ship maneuvering model parameters plays a critical role in ship motion control, navigation safety, and intelligent decision-making. However, the nonlinear and strongly coupled characteristics of ship dynamics render traditional identification methods difficult to apply under limited or noisy data conditions. To address these challenges, a self-adaptive physics-informed neural network (PINN) framework is proposed in this study, in which the governing equations of ship motion are coupled to formulate a multidimensional loss function and a scalable hyperparameter is embedded into the activation function to dynamically adjust the loss function landscape during training, thereby enhancing the prediction accuracy and convergence speed of ship maneuvering performance. The method is validated through parameter identification experiments using ± 20 ° zigzag maneuver simulation data of a Mariner-class vessel. The proposed adaptive-activation PINN method significantly reduces the identification errors of several key maneuvering parameters under limited sampling and noise-contaminated conditions, demonstrating higher accuracy and robustness. This study can provide dynamic support for intelligent navigation decision-making and optimal control, which is conducive to achieving high-precision control and decision optimization for autonomous ships in complex sea conditions.