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

Estimating waves via measured ship responses

“…The assessment of the hybrid approach is made with the same dataset as presented in the previous subsection focused on in-service data from a container ship. The training of the CNN is based on 80% of the total (filtered) dataset and comprises samples 32 , while the validation set makes up the remaining 20% of the data, which means that samples are left as unseen data. The reader is reminded that a sample in this context consists of a 25-minute set of MRU (motion response unit) time series and the corresponding ERA5 wave parameters in terms of { , , }.…”

“…In this paper, the sea state identification , to differentiate from wave spectrum estimation where the entire directional wave spectrum is obtained, builds on an implementation of an Inception model 69 . Specifically, the identification algorithm 32 makes use of response spectra, applied in a multi-task learning (MTL) setting 70 , for identification of significant wave height, peak period, and wave direction. In MTL, each output is considered as a separate task with its own dedicated branch of fully connected hidden layers and corresponding output layers.…”

“…Such an architecture is selected to favour physical interdependencies between the wave parameters and improve the model’s generalisation capability 71 . In the current work, the ship’s speed and draught were not fed into the model; draught was not measured, thus “unknown”, and the speed was relatively constant, varying in the range of 18-20 knots for the vast majority of the data 32 . In case of more varied operational conditions, however, it will be necessary to refine the neural network, for instance by considering a multi-modal convolutional neural network, which is characterised by several input branches and discussed in greater detail in the literature 32 .…”

“…Such a framework has recently been developed by establishing an algorithm 11 that outputs a directional wave spectrum which is conditioned on results from a machine learning method. More specifically, the hybrid framework takes estimated wave parameters from a convolutional neural network 32 and uses them as constraints on the directional wave spectrum subsequently computed by a physics-based method 8 , 13 . It is noteworthy that the specific physics-based method relies on Bayes’ theorem 37 , and, in the following, we will refer to it as the Bayesian method .…”

SAFENESS: A Semi-Supervised Transfer Learning Approach for Sea State Estimation Using Ship Motion Data

“…In effect, the ship is inherently outfitted with a system for SSE [11]. SSE methodologies based on ship motion responses are typically classified as either model-based or model-free [12], [13].…”