Physics-guided neural network for underwater glider flight modeling

Abstract

Underwater glider (UG) is an energy-saving ocean exploration platform, and its powerful dynamic model is necessary for state estimation, planning, and motion control. This paper introduces a novel method for the UG flight modeling based on the physics-guided neural network (PNN), which combines the advantages of physics-driven and data-driven modeling. The theoretical model provides the baseline, and the radial basis function neural network (RBF) compensates for the modeling error. The sliding window method concentrates on the RBF to improve online learning functions. Towing tank test (Tank) and computational fluid dynamics (CFD) are conducted to calculate the hydrodynamic of the UG. Comparison experiments indicate that the PNN can improve the accuracy by 41% and 82% compared with the theoretical model (CFD-based) and data-driven method (RBF-based), respectively. The proposed PNN also applies to other underwater vehicles.