MACHINE LEARNING FOR MODELING UNDERWATER VEHICLE DYNAMICS: OVERVIEW AND INSIGHTS

Abstract

Accurate modeling of underwater vehicle dynamics is an essential component of various solutions designed to address a range of challenges involved in both the vehicle’s design and operation. Such models are usually parametric, including dynamic equations that simulate the vehicle’s response to various controls and environment conditions. They can be used to determine the vehicle’s capabilities, estimate the vehicle’s state in the absence of external communications, or to derive control signals to produce desired state responses. While a range of explicitly derived models have been commonly used in various applications, modeling the complex nonlinear dynamics using machine learning has recently attracted considerable interest. This topical review focuses on the integration of machine learning in underwater vehicle modeling, and covers two categories: artificial neural networks and non-parametric regression models. The first category includes recurrent neural networks and physics-informed neural network. They are trained to estimate model parameters, forces and moments from damping and disturbances, or to completely replace the dynamic model by outputting the expected state responses. The second category of the reviewed models covers support vector machines and Gaussian process models. These are non-parametric dynamic models and their training requirements are generally lower than ANN-based models. An overview of the theory behind each model is presented, along with examples of specific applications. The capabilities of each machine learning method are compared, and the challenges of their implementation for underwater vehicle dynamic modeling are discussed.