A Small-Sample Transfer Learning Framework for Cross-Domain Ship Motion Prediction in Extreme Sea States

Accurate ship motion prediction is important for navigation safety, onboard decision support, and shipborne digital twins. Yet data-driven models trained on one vessel often degrade when applied to different hull forms, speed regimes, and sea states. This study proposes a small-sample transfer learning method for cross-domain ship motion prediction in extreme sea states. A CEEMDAN–SSA–Informer model is pre-trained on sea-trial data from the displacement-type training ship Yu Kun and adapted to a high-speed V-shaped target vessel under SS5–SS6 conditions. Only a 15-min target-domain segment is used to update the final model parameters. The method combines maximum mean discrepancy-guided layer freezing, domain-adaptive modal attention, and a peak-aware loss to preserve transferable representations and improve extreme-response prediction. Under SS6, the roll PRE decreases from 48.7% with direct transfer to 8.2% after adaptation, corresponding to an 83.2% reduction and indicating substantially improved prediction accuracy over high-amplitude roll-response samples. Ablation and sensitivity analyses show that the three transfer components improve both prediction accuracy and data efficiency. The proposed method provides a practical solution for adapting ship motion predictors across vessel and severe sea-state domains with limited target-domain data.