Spectrum-regularized deep learning with joint wave parameter prediction for sea state estimation from ship motions

Real-time in-situ sea state estimation is crucial for ship operations. However, existing data-driven methods often struggle with physical inconsistency, limited continuous parameter estimation, and poor robustness under sensor degradation. To address these challenges, this paper proposes a deep learning method that integrates spectrum-based self-consistency regularization to jointly estimate significant wave height (Hₛ), wave direction (Dir), and peak period (Tₚ) from ship motion data, while providing an auxiliary sea state class (Class) output. The proposed model predicts Hs, Tₚ, and the spectrum in parallel using a shared dual-domain encoder and three cross-attention branches, where spectral relations couple the outputs. During training, spectral algebraic relation serves as auxiliary self-consistency losses, with a delayed introduction strategy to balance spectral regularization and data-driven objectives. Six ablation experiments examine the effects of the main components and training choices. Based on privately owned high-fidelity simulated data, the obtained results show that compared with the data-driven baseline, the proposed method reduces Hₛ MAE by 15.53% and improves auxiliary classification accuracy by 5.61pp. It also shows lower sampling-rate sensitivity to downsampling, with 1.3% degradation at 1 Hz, compared with 30.2% for the baseline. Ablation results further show that spectral regularization mainly improves Hₛ estimation, delayed introduction improves the balance among the prediction targets, and soft labels improve the auxiliary classification output.

• Record URL:
  • https://doi.org/10.1016/j.oceaneng.2026.128186
• Record URL:
  • http://www.sciencedirect.com/science/article/pii/S0029801826040205
• Availability:
  • Find a library where document is available. Order URL: http://worldcat.org/issn/00298018
• Supplemental Notes:
  • © 2026 Elsevier Ltd. All rights are reserved, including those for text and data mining, AI training, and similar technologies. Abstract reprinted with permission of Elsevier.
• Authors:
  • Guo, Xinyu
  • Mao, Runze
  • 0000-0003-2392-492X
  • Han, Peihua
  • Xu, Jinyan
  • Zhang, Houxiang
• Publication Date: 2026-11-15

Language

• English

Media Info

• Media Type: Web
• Features: References;
• Pagination: 128186
• Serial:
  • Ocean Engineering
  • Volume: 368
  • Issue Number: 0
  • Publisher: Pergamon
  • ISSN: 0029-8018
  • EISSN: 1873-5258
  • Serial URL: http://www.sciencedirect.com/science/journal/00298018

Subject/Index Terms

• TRT Terms: Deep learning; Estimating; Predictive models; Properties of water waves; Sea states; Ship motion; Ship operations; Ships
• Subject Areas: Marine Transportation; Vehicles and Equipment;

Filing Info

• Accession Number: 02004876
• Record Type: Publication
• Files: TRIS
• Created Date: Oct 2 2026 4:54PM