Prediction of Extreme Vessel Responses Utilizing Artificial Intelligence

Abstract Since mankind put to sea, the seakeeping capabilities of its vessels has been the decisive factor on the save voyage of their seafarers. In modern times the analysis of vessel responses to seaway by numerical calculations has developed to be the industry standard alongside model tests. At the Institute of Ship Design and Ship Safety of the Hamburg University of Technology the Software E4-ROLLS is used to perform such kind of analyses. Originating in the 1980s this piece of software has been validated through plentiful research as well as accident investigations. While sufficiently accurate and fast in its predictions on the roll motion of oceangoing ships, such calculations are currently performed only for specific loading conditions and seaways due to its requirement of some preceding calculations and manual user interaction. The current research is set out to widen the scope onto the entire operational profile of ships and all necessary seaway situations. This is achieved by first generating a large number of loading conditions for the vessel in question using a parameterized description of its deadweight items and a Monte-Carlo-based approach. Second all necessary calculations for and with E4-ROLLS are automatized to produce a large quantity of accurate data using the generated loading conditions. The third step is to find a fitting mathematical model of this data. For this methods of machine learning, as a sub-category of artificial intelligence, are implemented and tested on the achievable prediction accuracy. Currently polynomial regression is utilized in a two-step process. In the first step the seakeeping capabilities of each loading condition are modelled by a multidimensional polynomial function optimizing the polynomial degrees for a low mean quadratic error. The second step is used to extract polynomial coefficients for any in-between loading conditions. The current implementation of this process results in a rather fast calculation, while the overall mean quadratic error remains below one meter of limiting significant wave height for an exemplary ultra large container vessel. Current work focuses on refining the polynomial model. However if the accuracy cannot be improved further, other methods may be employed in the future. With the ability to calculate a real-time prediction model of the seakeeping behaviour of a ship it is further planned to incorporate this as a warning system aboard ships assisting in avoiding dangerous vessel motion.