Sea state identification using machine learning—A comparative study based on in-service data from a container vessel. (September 2022)

• Record Type: Journal Article
• Authors: Mittendorf, Malte; Nielsen, Ulrik Dam; Bingham, Harry B.; Storhaug, Gaute
• Is Part Of: Marine structures. Volume 85(2022)
• Publication Date: 2022-09
• Subjects: Sea state estimation -- Wave buoy analogy -- Sensor data -- Wave radar -- Deep learning -- Multi-task learning; Naval architecture -- Periodicals; Offshore structures -- Periodicals; Architecture navale -- Périodiques; Structures offshore -- Périodiques; Naval architecture; Offshore structures; Periodicals; 620.4162
• DOI: 10.1016/j.marstruc.2022.103274
• Languages: English
• ISSNs: 0951-8339

Abstract

Abstract: This paper is concerned with a machine learning-based approach for sea state estimation using the wave buoy analogy. In-situ sensor data of an advancing medium-size container vessel has been utilized for the prediction of integral sea state parameters. The main novelty of this contribution is the rigorous comparison of time and frequency domain models in terms of accuracy, robustness and computational cost. The frequency domain model is trained on sequences of spectral ordinates derived from cross response spectra, while the time domain model is applied to 5-minute time series of ship responses. Multiple deep neural networks were trained and the sensitivity of individual sensor recordings, sample length, and frequency discretization on estimation accuracy was analysed. An Inception Architecture adapted for sequential data yields the highest out of sample performance in both considered domains. Additionally, multi-task learning was employed, as it is known for increased generalization capability and diminished uncertainty. Overall, it was found that the frequency domain method provides both superior performance and significantly less computational effort for training. Highlights: Measurement data of a container ship trading in the Northern Atlantic is used. The significant wave height, peak period and encounter direction are estimated. Several deep neural networks are compared in the time and frequency domain. Multi-task learning shows diminished uncertainty in case of the temporal models. The frequency domain method shows higher accuracy and computational efficiency.