Dynamic system identification of underactuated ship dynamics based on Gaussian process regression

Abstract

At present, the problems of underactuated, nonlinearity, and poor real-time performance are common in the research of ship motion control methods. The modeling of ship dynamics is one of the key points in ship controller design This paper uses Gaussian process regression in machine learning to identify ship model, in which an empirical ship maneuvering is used to generate state information data sets for regression training, which reduces the computational cost by using only low data volume for training. This approach avoids the calculation of hydrodynamic derivatives in the traditional mechanism modeling process and simplifies to optimize a small number of hyperparameters of the kernel function in Gaussian regression. Finally, the accuracy and robustness of the regression model are tested by cross-validation. Simulation results show that Gaussian process regression can be accurately used to identify nonparametric dynamic systems of ships.