• Proposes a hybrid manoeuvring model, incorporating a physics-based control model and a data-driven model, within a fixed-gain non-linear state observer. • Presents five data-driven models trained with synthetic data that was gathered under simulated standard manoeuvres in accordance with the International Maritime Organisation guidelines. • The best-performing hybrid physics and neural observer has a relative improvement of 30% compared to a fixed-gain non-linear observer using the physics-based control model alone. Unlike conventional non-linear state observers that rely solely on physics-based models, this study proposes a hybrid state observer that combines a physics-based control model with a data-driven correction model to improve state estimation for maritime surface vessels. Five deep neural network architectures, trained on synthetic data akin to what can be obtained during sea trials, serve as correction models within the hybrid design. These architectures are evaluated through simulation scenarios featuring variable ocean currents and wind. The results show that the hybrid observer using a fully connected neural network with three layers of 64 neurons reduces estimation error by 30% compared to a fixed-gain non-linear observer based solely on a physics-based model. This improvement highlights the potential of hybrid observers to enhance state estimation accuracy, which is essential for the overall control system performance.
Frafjord et al. (Tue,) studied this question.