Dynamic System Identification of Underwater Gliders based on Multi-output Gaussian Process

Abstract

In this paper, a nonparametric system identification algorithm based on a multi-output Gaussian process for underwater gliders is proposed, which can predict the motion of UGs under the conditions of few training data, part measurable states, and high coupling degrees. The algorithm combines the nonlinear auto-regressive model with an external input structure and uses the conjugate gradient descent optimization algorithm to develop a nonparametric dynamic system identification scheme. The proposed scheme is implemented over data obtained from the simulated model of a UG ray-like manta of 5° and 10° Z-type steering data. The results show that the root means square errors of the prediction motion are less than 0.01500° compared with the real motion, and the multi-output Gaussian process can be accurately applied to the strong coupling, multi-degree-of-freedom (DOF) of the underwater gliders.