Parameter identification and real-time motion prediction for a water-jet unmanned surface vehicle based on online sparse least squares support vector machine algorithm

Abstract

A large amount of navigation state data and control command data would be generated during the operation of unmanned surface vehicle (USV). However, existing research rarely focuses on decoupling the mapping between actual navigation state data and control commands for constructing the maneuvering motion model of USV. This paper proposes an online learning method based on least squares support vector machine (LSSVM) for the USV’s mathematical modeling and online maneuvering prediction. A sliding window mechanism is introduced to update USV’s state variable data, maintaining the total number of samples within the time window constant, thereby enabling the traditional least squares support vector machine (LSSVM) method to acquire online recursion and identification capabilities. The incremental and decremental learning formulas for updating the inverse kernel function matrix are derived to improve algorithm’s real-time performance. Meanwhile, a novel leave-one-out cross-validation (LOOCV) pruning algorithm is proposed for sliding window data updating, which calculates LOOCV values for each sample and removes noise samples with lower modeling contribution. A cost-accuracy metrics method integrating both algorithm runtime and identification accuracy is designed to evaluate the performance of the identification algorithm. The feasibility and effectiveness of the developed method are validated through real-time motion prediction studies, utilizing actual steering and turning tests of a full-scale USV.