This paper investigates a data-driven methodology for estimating sea state parameters from the motions of a spread-moored FPSO operating off the Brazilian coast. The approach is motivated by operational environments where dedicated wave measurement is unavailable, leveraging motion sensors already installed on-board. A systematic hyperparameter search over different deep learning models was conducted using in-service motion records collected under multiple loading conditions, combined with wave statistics from the ERA5 reanalysis model, used as training targets. Results show that the proposed neural network model achieved good performance, with median errors for significant wave height below 6% across several bimodal sea conditions, but also indicated some of its shortcomings and opportunities for improvement. A transfer learning strategy was also explored, in which networks pre-trained on simulated responses from synthetic sea states were fine-tuned using a limited subset of measured data. Although the fine-tuned models did not reach the accuracy of those trained directly on in-service data, results demonstrated the feasibility of exploiting simulations to reduce the requirements for extensive real-world measurements to derive effective sea state estimation models. • Data-driven sea state estimation using full-scale motion records of in-service FPSO. • Estimation model trained across multiple drafts achieves median H s errors below 6%. • Inference under crossed sea states by estimating two sets of wave statistics. • Inception-based neural networks selected via systematic statistical comparison. • Transfer learning from numerical simulations reduces reliance on in-service data.
Bisinotto et al. (Fri,) studied this question.