Online identification of a ship maneuvering model using a fast noisy input Gaussian process

The design of maritime traffic simulators, model-based controllers, and maritime autonomous surface ships require accurate ship dynamic models. Using different nonparametric methods to identify a ship dynamic model using measurement data is a popular research topic. Among the proposed nonparametric models, the Gaussian process (GP) approach is a high-accuracy nonparametric modeling technique, and it can provide the uncertainty of the prediction. However, the high computational complexity of the GP restricts its further application, especially for online learning. Here, we present a novel online identification scheme, namely, the online fast noisy input Gaussian process (online-FNIGP) to identify ship response models. To avoid a high computational burden and the influence of noise, the fully independent training conditional (FITC) algorithm is introduced to the noisy input Gaussian process (NIGP), and it can significantly reduce the computational complexity. The developed scheme can incorporate new noisy measurements online and make fast predictions. The proposed method is verified by two examples: a simulation zigzag test with a parametric container ship and a random maneuver experiment with an unmanned surface vessel. For comparison with the proposed model, a regular Gaussian process is implemented to validate the developed approach. Both the simulation study results and the experimental study results indicate that the developed scheme is a powerful online identification tool for ship maneuvering systems.