Motion-Based Wave Inference With Neural Networks: Transfer Learning From Numerical Simulation to Experimental Data
人工神经网络
推论
计算机科学
运动(物理)
学习迁移
人工智能
机器学习
作者
Gustavo A. Bisinotto,Pedro Cardozo de Mello,Fábio Gagliardi Cozman,Eduardo A. Tannuri
出处
期刊:Journal of offshore mechanics and Arctic engineering [ASME International] 日期:2024-01-30卷期号:146 (5)被引量:3
标识
DOI:10.1115/1.4064618
摘要
Abstract The directional wave spectrum, which describes the distribution of wave energy along frequencies and directions, can be estimated from the measured motions of a vessel subjected to a particular sea condition by resorting to the wave-buoy analogy. Several methods have been proposed to address the inverse estimation problem; recently, machine learning techniques have been assessed as further alternatives. However, it may be difficult to gather large datasets of in-service motion responses and the associated sea states to train effective data-driven models. In this work, an encoder–decoder neural network is trained with the synthetic responses of a station-keeping platform supply vessel (PSV) to estimate the directional wave spectrum. This estimation model is directly applied to perform wave inference from motion data of wave basin tests with a small-scale model of the same vessel. Furthermore, fine-tuning is also used to incorporate experimental data into the neural network model. Results show a satisfactory match between estimated and measured values, both with respect to the energy distribution and the integral spectrum parameters, indicating that the proposed approach can be employed to obtain data-driven wave inference models when there is little or no availability of measured motion records and the corresponding sea conditions.