Data-driven and Machine-learning Optimized Hydrodynamic Models for High-Speed Planing Hull Unmanned Surface Vehicles
Abstract
Modeling control systems for high-speed planing hull unmanned surface vehicles (USV) under hydrodynamics simulations pose a challenge due to the complexity of the underlying physics and the presence of nonlinearities. The systems are commonly modeled as coupled ordinary differential equations (ODEs) embedded in PID controllers, which can be numerically solved using suitable ODE solvers in real-time. The solutions describe the USV control response to changes in the environment or external conditions. Control system engineers typically use predefined ODEs with a range of parameters. Those parameters are tweaked based on data collected through numerical simulations or sensors during real-time missions. The goal is to have the most robust control system under various environmental changes. In real-time missions, however, more than having predefined ODEs may be required. In the case of rare events or vehicle damages, those ODEs may not fully represent the control system under normal circumstances, resulting in unpredictable behavior of the USV.Moreover, engineers may be unable to perform real-time controller updates, and a backup controller may be needed to ensure safe missions. Neural ODE is a computational method that utilizes machine learning to form coupled ODEs based entirely on the data. Having predefined ODEs is optional, although the method can also be used to improve preexisting models. For example, one can use predefined ODEs for linear models and neural ODE to model nonlinearities. The novel contribution of this project will be the development of computational methods and strategies employing neural differential equations to develop a digital model that describes the motion ofa high-speed planing craft USV in six degrees of freedom and use techniques to mitigate performance bottlenecks of data-driven scientific machine-learning frameworks. Metrics and validation will include application assessments of the robustness of neural ODE architectures for the USV control system in the case of unforeseen or extreme conditions. These types of early applications are representative of the challenges that Navy autonomous systems will experience. For example, persistent systems may experience unexpected changes in system dynamics that require learning and verifying the safety and effectiveness of a new control policy using data collected during deployment to continue operating optimally. Navy autonomous systems may need to update autonomy policies, robust backup control systems, and goals in situ to adapt to their environment and complete tasking successfully. Understanding the viability of neural-ODE-based control systems as a primary or supporting controller to the robustness of the USV performance is critical. The proposed research will support the Navy#s need to explore state-of-the-art methods in data-driven simulation-based dynamical system modeling, improve the overall Naval confidence in machine learning methods, and prepare a graduate student to work at the US Naval UnderseaWarfare Center.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Jun 13, 2024
- Source ID
- N000142412332
Entities
People
- Alfa Heryudono
Organizations
- Office of Naval Research
- United States Navy
- University of Massachusetts