Nonparametric online modeling for differential-driven unmanned surface vehicle based on fuzzy C-means and PD-like event-triggered strategy

• The modified sand cat swarm optimization algorithm incorporating a nonlinear decay factor and Levy flight mechanism is proposed to optimize the hyperparameters of LS-SVM. • The FCM soft clustering method is adopted for data partitioning, and the separate models are built to better capture the complexity and diversity of ship maneuvering motion. • The PDLET strategy is proposed in this article, which could more effectively capture dynamic variations and update the model in a timely manner. • The data selection window mechanism is designed to avoid data redundancy while ensuring smooth transitions of the models. The modeling reliability of differential-driven unmanned surface vessel (USV) is greatly affected by the limited on-board computational resources, highly coupled nonlinear models, and internal and external variable factors. For balancing computational resources and prediction accuracy, this paper proposes an novel nonparametric online modeling method for dynamic model of differential-drive USV based on fuzzy C-means (FCM) and PD-like event-triggered (PDLET) strategy. Firstly, an improved sand cat swarm optimization (ISCSO) is proposed to optimize the hyperparameters of least squares-support vector machine (LS-SVM). Secondly, the FCM soft clustering algorithm is adopted to cluster the collected offline data, and then the optimized LS-SVM is adopted to establish offline models for different clusters. Thirdly, the velocity and acceleration errors are simultaneously considered to construct the PDLET strategy, which determines model update moments and better corrects the model. Further, the training data of each cluster is filtered by the designed data selection window to reduce prediction errors in model switching process. The offline experimental results show that the hyperparameters optimized by ISCSO have better predictive accuracy compared with the existing sand cat swarm optimization (SCSO), particle swarm optimization (PSO), and cuckoo search (CS). The online experimental results show that the proposed online method has better prediction accuracy and could significantly reduce modeling time and the number of model updates.