A frequency-enhanced interpretable patch decomposition network for seakeeping prediction under irregular-wave conditions

To address the challenge that the nonlinear and locally time-varying characteristics of ship motion responses under irregular-wave conditions make it difficult for existing forecasting models to sufficiently characterize key frequency components and historical segment contributions, this paper proposes a Frequency-Enhanced Interpretable Patch Decomposition Network (FEIPDN). First, a frequency attention mechanism is employed to enhance the dominant frequency components relevant to future response prediction. Second, the frequency-reweighted sequence is divided into local patches, and a one-dimensional deformable convolutional encoder is used to extract local dynamic features from different historical segments. Then, a patch decoder is adopted to establish the relationship between historical patches and future prediction segments, thereby achieving ship motion response forecasting. Finally, patch contribution decomposition is introduced to quantify the influence of different historical segments on the prediction results, improving the interpretability of the forecasting process. Experimental results on five Delft 523 irregular-wave conditions show that FEIPDN achieves the best prediction performance, with R 2 values ranging from 0.9231 to 0.9432; further validation on continuous full-scale voyage data from the SA Agulhas II yields an R 2 of 0.9189. These results demonstrate that FEIPDN improves prediction accuracy and generalization capability while enhancing the interpretability of the forecasting process.