Ship dynamics models are the foundation to predict ship movement. Traditionally, mechanism-driven models have low accuracy and data-driven models have high data dependency. Thus, a novel mechanism-data dual-driven method is proposed to address the above problems. An MMG model is constructed to generate virtual data for the offline identification. An online model is established with reference to the offline model. The online model is trained by two state-of-the-art online algorithms, kernel adaptive projection and projection support vector machine. It has been verified that the proposed method can improve the prediction accuracy and reduce the data dependency.
In recent years, ship motion control technology has been widely used in the maritime autonomous surface ship (MASS). Before application, the control system should be tested to ensure sailing safety. Accurate ship dynamics model is the basis of testing platform (Liu et al.,2024). The accuracy of the traditional mechanism model is usually unsatisfactory due to the incomplete expression of the nonlinear hydrodynamic items. Moreover, it is difficult to consider the environmental disturbances and scale effect in the mechanism model (Shen et al., 2023). System identification is an efficient and low-cost modelling method. Besides, the method can be directly applied to full-scale ship to avoid the scale effect.
Lots of researches on ship dynamics model identification aim at obtaining parameters. The competitive methods include ridge regression method (Yoon et al., 2007), frequency domain spectrum analysis (Bhattacharyya and Haddara, 2006), backstepping (Casadoetal., 2007) and Bayesian approach (Xue et al., 2020). Because of the satisfactory generalization performance and the global optimality of solutions, support vector machine (SVM) and its extended version have been widely applied in the field of ship manoeuvring. Luo and Zou (2009) first applied SVM to identify the hydrodynamic coefficient of 3 degrees of freedom (DOF) Abkowitz model. Based on the reference parameters from SVM identification, Meng et al.(2022) further optimized them by a modified grey wolf optimizer to improve the model accuracy. Zhu et al. (2019) enhanced the global optimization capability of SVM by using artificial bee colony algorithm to optimize hyperparameters. Considering roll, Wang et al. (2013) obtain 4 DOF ship dynamics model by least square support vector machine (LS-SVM). Before identifying parameters by LS-SVM, Hu et al. (2021) used wavelet threshold denoising to preprocess train data, which can reduce parameter drift and improve generalization performance.
