Non-parametric dynamics modeling for unmanned surface vehicle using spectral metric multi-output Gaussian processes learning

Abstract

High-precision dynamic models are crucial for various marine missions and the control and navigation of unmanned surface vehicles. However, the nonlinear, strong coupling, complex structure of these vehicles poses challenges in building such models. Traditional single-output models have limitations in data utilization and high computational complexity. Therefore, this paper proposes a spectral metric multi-output Gaussian process for continuous-time dynamic modeling of unmanned surface vehicles. Specifically, a spectral covariance matrix is constructed in this scheme to establish a spectral metric multi-output Gaussian processes learning model that captures correlations under different degrees of freedom. Meanwhile, an improved particle swarm optimization algorithm is used to dynamically tuning the correlation matrix for improved model accuracy, and utilizing a subspace index to enhance computational efficiency in multi-output Gaussian processes learning. Moreover, the scheme can conduct real-time simulation under environmental disturbances. Finally, the proposed scheme is applied to an unmanned surface vehicle, and the results demonstrate that the scheme is an effective modeling tool for multi degree-of-freedom coupled dynamics.