Data-driven enhancement to ship dynamic model for motion prediction

Abstract

Ensuring safety and efficiency is the most critical agenda for the maritime industry. It is essential to have a good understanding of ship dynamics and make accurate predictions of ship motion. This enables us to properly evaluate risks such as collision and excessive motion. The bedrock to do so is to develop advanced ship dynamic models for making accurate motion predictions.

Ships are complex systems composed of multi-scale and multi-disciplinary sub-systems, such as a hull, thrusters, engines, etc. They are exposed to external disturbances such as wind, waves, and ocean currents in real-world sea conditions. These complexity and non-linearity lead to strong uncertainties in dynamic models, making it difficult to achieve high performance. The maritime industry has faced increasingly rigorous cost requirements, leading to a demand for a more agile framework for delivering dynamic models.

To address this, this dissertation proposes a framework that makes a synergy of two modeling principles. Physics-based modeling has played a dominant role, emphasizing a mechanical understanding of ship motion. The resulting model is accompanied by physical interpretations and is trusted by experts for their knowledge foundation and reliability. The flip side of its physics-based inspiration is that it tends to perform poorly on the full-scale dataset of interest. Moreover, it requires much expertise and meticulous experiments, resulting in high costs in both time and resources. On the other hand, data-driven modeling aims to learn inductively hidden patterns from full-scale data. It excels at handling strong non-linearity and complexity. With the use of cutting-edge machine learning models that are robust to noise and disturbances, it is possible to effectively utilize onboard data for better ship dynamic models. However, due to their poor interpretability, replacing the strong aspects of physics-based models with them would not be a good idea.

Two modeling principles play different roles and cannot be replaced by each other. Instead, in the agile framework, they need to leverage each other's strengths for complementing each other's weaknesses. To this end, this dissertation develops a framework in which data-driven models enhance the performance of the physics-based model while receiving sufficient knowledge transfer and continuity from the physics-based model. New techniques are presented in four case studies for various pragmatic problems where full-scale data needs to rapidly benefit the performance of the physics-based model. The comprehensive experimental study addresses how the physics-based model and data benefit the enhanced performance in such a framework. A novel method is also presented to bridge the dynamics of similar ships and ease model development. This dissertation proposes specific enablers of such a framework that enhances physics-based models with data-driven methods and demonstrates its promising effectiveness through simulation and real-world validation experiments.