Motion Trend Prediction of Unmanned Surface Vessels Based on Physics-Informed Neural Network

Due to the maneuverability and stealth capabilities of Unmanned Surface Vehicles (USVs), they have increasingly been deployed in maritime combat environments, significantly impacting traditional maritime defense systems. Accurately predicting opposing USVs' motion trends has become essential for defenders to formulate effective defensive strategies. The motion of USVs is influenced not only by their predefined trajectories but also by uncertainties arising from hidden intentions and other external factors. Consequently, this paper addresses the prediction of USV motion trends under uncertainty, with the following contributions: Firstly, the uncertainties of the USV motion is depicted by the Bayesian-LSTM network. By analyzing the prior physical features to determine probabilistic regions of USV presence, a refined posterior probability is computed. This method integrates a residual correction mechanism, continuously adjusting outputs to delineate the potential regions that USVs may reach and their associated probabilities. Secondly, a Physics-Informed Neural Network (PINN) is incorporated into the motion trend prediction model. By embedding physics-based loss functions into the neural network training process, the network's ability to learn and express physical principles is significantly enhanced, ensuring consistency between the model outputs and real-world physics. Finally, a residual-driven attention mechanism is introduced, dynamically adjusting the fusion weights of input data based on model residuals. This can substantially reduce the dependence on large-scale training datasets. Experimental results demonstrate that the proposed model outperforms baseline methods in terms of prediction accuracy and uncertainty quantification.