Neural networks as a tool for improving the mathematical model of ship motion

Abstract

Using neural networks opens up great opportunities for studying mathematical models of ship motion. Correction by a network of identified parameters of the selected model should be as adequate as possible to the results of standard full-scale tests defined by the IMO Resolution N 137 of 2002. A mathematical model in displacements is considered, containing 16 parameters that determine the hydrodynamic forces acting on the ship's hull and steering gear, and is the source of a data set for training the network by randomly varying the parameters and subsequent computer testing. The standard maneuver is a steady-state circulation with fixation of the maneuvering elements: diameter, linear velocity, drift angle and angular velocity of rotation. Improving the quality of the model has consisted of changing its parameters and minimizing the mean square errors of the values of the maneuvering elements obtained during testing. For these purposes, a neural network with 16 inputs (model parameters) and four outputs (maneuvering elements for steady-state circulation) has been built. The data set for training the network was obtained using a program developed by the authors and intended for calculating parameters and conducting maneuver tests. A tanker with a displacement of 30,000 tons was chosen as a test object. Various options for network architecture and tools for working with it have been considered; the Statistica Neural Nets (SNN) software environment and the ANN package in the SciLab environment have been used. Comparative assessments of the results of working with these tools have been given.