Non-parameterized ship maneuvering model of Deep Neural Networks based on real voyage data-driven

Abstract

While Deep Neural Network (DNN) models for ship maneuvering model are commonly constructed using experimental model ships or simulation data, this study focuses on training four Non-parameter Ship maneuvering Network models (NSM_Net_models) by using real ship voyage data, then evaluating their performance by comparison. To support the data-driven approach, real ship voyage data is collected from the Voyage Data Recorder (VDR) system of a training ship. Along with this, the input variables of the models are enhanced by incorporating the past movement history of ships. Furthermore, the impact of three different durations of past ship motion history on the models’ performance is compared and analyzed. The results suggest the feasibility and practical application value of the ship maneuvering motion model built by the deep network.