Dynamic model identification of unmanned surface vehicles using deep learning network

In this paper, a deep learning-based dynamic model identification method is proposed. The proposed method is designed to capture higher-order dynamic behaviors that result from the coupling of hydrodynamics and actuator dynamics. By adopting recent advancements in deep learning, our model addresses problems such as the regression problem in machine learning. Among various deep learning algorithms, long short-term memory (LSTM)-based recurrent neural network was used to deal with the hidden latent state of the USV dynamic model. The model validation was performed using free running test data of a USV. Analysis result shows that proposed model reduces surge speed prediction error by 76.9%, yaw rate prediction error by 60.7% and sway velocity prediction error by 27.9% over the conventional linear dynamic model.