System identification of ship dynamic model based on Gaussian process regression with input noise

Abstract

As a critical step designing the ship controller and the maritime traffic simulator, the system identification of a ship dynamic model from input-output data is a promising direction. However, the noise in the training data significantly reduces the accuracy in identification and prediction. Here, we present a robust nonparametric system identification technique for a ship maneuvering model based on Gaussian Process (GP) regression. To solve the problem caused by noise, the input noisy Gaussian Process (NIGP) model is employed, and it can automatically propagate the input uncertainty to the output in the learning model using the Taylor approximation. Polluted simulation datasets obtained from the numerical model of a parametric container ship are used for training. Zigzag and turning circle maneuvering motions are performed, and a comparison with GP and SVM is implemented to validate the proposed approach. The results indicate that the developed scheme for system identification of ships is accurate and robust with noisy input.