Directional wave spectrum estimation with ship motion responses using adversarial networks

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

“…Scholcz and Mak [22] extend the work of Düz et al [20] and present a deep learning methodology for the non-parametric estimation of the directional wave spectrum based on wave radar data using a convolutional encoder-decoder network applied to in-service time series data. Lastly, Han et al [23] provide an investigation for non-parametric SSE and establish an approach based on a generative adversarial network, in which the generator predicts the 2D spectrum relying on cross response spectra, and the discriminator classifies the validity of the prediction. This iterative approach is then compared to a traditional model-based approach and shows satisfactory accuracy in non-forward speed scenarios; emphasizing that simulated time series data has been considered exclusively.…”

“…It is noted that 25 min are chosen as the maximum sample length since recordings close to the 30 min threshold were frequently missing. In addition, the focus of the herein presented work is on using smaller time frames, noting that other studies use longer durations, say, from 30 min in [23] to 60 min in [21]. The synchronization step decreased the size of the dataset further to 5182 samples.…”

“…Thus, the number of discrete frequencies 𝛺 is part of a second sensitivity study, in which also the effect of the sample length on prediction accuracy will be investigated. For reference, 71 and 100 discrete spectral ordinates were extracted from individual cross response spectra in the work of [21,23], respectively. Lastly, the shape of the input matrix for the machine learning model in the spectral domain has a shape of 𝛺 × 𝑀, where 𝑀 takes a maximum value of 43.…”

“…Hence, it will be studied herein, whether the prediction of the forward speed enhances the prediction of the sea state parameters and could act simultaneously as a reliability indicator. Lastly, the tasks of the two investigated architectures are related or even complementary, however, adversarial training as in [23] is also feasible using multi-task learning.…”

“…In their work [52], it is deduced that the relative wave direction is most susceptible to instationarity, whereas 𝐻 𝑠 and 𝑇 𝑝 are seen as more stable. For minimizing this kind of uncertainty, it is the objective of this work to use shorter sample lengths as compared to 30 min in [23] and 60 min in [21]. However, the uncertainty regarding stable spectral analysis is a counteracting contribution, which will increase with shorter time series samples.…”