Multi-fidelity data-adaptive autonomous seakeeping

Abstract

Safe operation of a ship in heavy weather and high sea states requires accounting for the risk of extreme ship motion responses in stochastic ocean waves. Excessive ship motions can lead to hazardous and unsafe conditions such as pure loss of stability, surf-riding and broaching. Mitigation of these risks can be performed through selection of ship speeds and headings for a given seaway, and avoiding conditions likely to lead to severe motions. To address this challenge, a robust, fast, data-adaptive model is a prospective enabling capability for onboard autonomous seakeeping. In this study, data-adaptive Long Short-Term Memory (LSTM) neural networks are investigated as part of a multi-fidelity approach incorporating Large Amplitude Program (LAMP), and a reduced-order model known as SimpleCode. An assessment of this multi-fidelity approach focuses on prediction of ship motion responses in waves. LSTM networks are trained and tested with LAMP simulations as a target, and SimpleCode simulations and wave time-series as inputs. LSTM networks are shown to improve the fidelity of SimpleCode seakeeping predictions relative to LAMP, while retaining the computational efficiency of a reduced-order model. Potential areas of application include unmanned and reduced-crew vessels, operator guidance for manned systems, and weather-informed ship route planning.