Digital Twin Motion Modeling of an Autonomous Surface Vehicle Based on Meta-learning Predictor

This paper investigates the digital twin motion modeling of an autonomous surface vehicle (ASV) with fully unknown dynamics. Specifically, a general deep neural network (DNN) is trained offline by using meta-learning techniques with collected training data under different environment conditions to approximate the unknown dynamics. Then, a meta-learning predictor is designed to update the last layer weights of the DNN to adapt to the new environment. Finally, the digital twin motion model of the ASV is constructed based on the meta-learning predictor, which not only allows for deep portrayal and accurate characterization of the motion state of the ASV in the cyber space but also provides high-fidelity model information and disturbances information for motion control. The stability of the digital twin motion modeling method is analyzed via cascade theory. Simulation results show the effectiveness of the proposed digital twin motion modeling method of the ASV.