Incremental online non-parametric modeling of surface vehicle dynamics using adaptive spectral metric Gaussian processes learning.

High-precision online modeling is the basis for various marine control and navigation of surface vehicles. An incremental Gaussian processes learning based on subspace index and adaptive spectral metric kernels is proposed for incremental online non-parametric identification of surface vehicle dynamics. The scheme utilizes streaming data for online identification without special designs of training data and can consider environmental interference. The subspaces are updated online by similarity, so the proposed subspace index can be viewed as a sparse technique for streaming online identification data. In order to improve the online model accuracy, an adaptive version of spectral metric kernels is designed in parallel utilizing novel adaptive moment estimation. The feasibility of the proposed scheme is verified through random steering tests of a 4-degree-of-freedom surface vehicle. Compared with traditional non-parametric identification methods, the proposed online identification scheme is effective in prediction accuracy and model generalization. The results show that this online scheme can provide a model basis for intelligent surface vehicle automation. • An incremental online non-parametric modeling scheme is constructed to identify surface vehicle dynamics online. • A spectral metric kernel is designed for incremental Gaussian processes learning to improve online identification accuracy. • A subspace index method is proposed to improve the efficiency of the online identification scheme. [ABSTRACT FROM AUTHOR]