A decoupled physics-informed expert system for stochastic trajectory prediction via spatio-temporal-frequency fusion

The modeling and prediction of complex dynamical systems underpin a wide range of engineering applications, including aerospace guidance, autonomous navigation, and robotics. However, practitioners face a persistent cybernetic dilemma: reconciling deterministic physical knowledge with perturbed trajectories and measurements contaminated by multi-scale, non-Gaussian sensor noise and environmental disturbances. To address these gaps, we propose a decoupled physics-informed expert system centered on architectural separation and hierarchical fusion. The proposed framework structurally separates deterministic physical trajectory generation from stochastic deviation modeling. Specifically, a Physics-Informed Trajectory Generation Network (PITGN) establishes a physically consistent baseline trajectory using a velocity-level PINN constraint, while a Stochastic Time-Frequency Dynamics Network (STFD-Net) models the deviation needed to map the baseline trajectory to the disturbed trajectory by integrating a Learnable Spectral Wavelet Attention (LSWA) mechanism with a Neural SDE decoder. Extensive experiments on a high-fidelity 7-DOF aviation dataset demonstrate that our system achieves a mean absolute error of 5.05 m, significantly outperforming traditional filters and standard deep learning baselines. With an inference latency of 25 ms, this modular architecture demonstrates potential for practical deployment, and provides a structured framework for real-time trajectory prediction under uncertain navigation conditions.