As commercial ships grow larger and faster, severe wave impacts trigger rapid dynamic vibrations—known as springing and whipping—that dramatically increase bending stresses on the hull. Failing to predict these dynamic midship bending moments in real time can lead to catastrophic structural fatigue or catastrophic hull fractures.
Traditional hydrodynamic simulations and early machine-learning models struggle to track these sudden, non-stationary wave loads. Because dynamic loads fluctuate rapidly during slamming events, standard forecasting tools often compound prediction errors over time and fail to estimate structural safety margins.
To resolve this limitation, researchers developed a three-stage autoregressive neural network architecture designed to decouple complex load signals and compute continuous probabilistic confidence bounds.
The model first uses Variational Mode Decomposition (VMD) to break down raw time-series data into discrete sub-sequences, cutting complex signal noise (sample entropy) by 30% to 80%. A single-layer Bidirectional Gated Recurrent Unit (BiGRU) then models past and forward temporal sequence trends to generate deterministic load predictions. Finally, a Quantile Regression Neural Network (QRNN) builds tight probabilistic confidence intervals around these estimates, giving operators a clear view of structural risk limits.
Unlike previous models that rely on vessel speed, wave height, or motion sensors—inputs that introduce external measurement errors—the new system operates autoregressively on direct hull load history.
"Predicting hull stress under chaotic wave conditions is notoriously difficult because high-frequency whipping spikes mask the underlying load trend," explained Jun Ding, corresponding author of the study. "By breaking down complex wave frequencies and calculating probabilistic safety bounds, our system gives shipbuilders and crews clear, reliable data on real-time structural risk."
The team validated the framework using physical towing-tank experimental data from a 205,000-ton bulk carrier subjected to irregular severe sea conditions. The model achieved an average coefficient of determination (R²) of 0.9533 and maintained a 94.22% coverage rate at the 95% confidence level. Crucially, the algorithm successfully tracked sudden peak loads caused by bow slamming without losing precision.
Looking forward, the researchers plan to integrate the prediction engine into active hull monitoring hardware across commercial marine fleets.
"Our objective is to move from passive structural monitoring to predictive dynamic safety control," added Ding. "Integrating this framework into navigation systems will help captains adjust speed and heading before dynamic stress reaches critical safety thresholds."
Co-authors of the study include Peiqiao Zhu, Zhang Zhu, and Yiming Qiang of the China Ship Scientific Research Center and Taihu Laboratory of Deepsea Technological Science.
DOI Link:
https://doi.org/10.26599/OCEAN.2026.9470022
About Ocean :
Ocean (ISSN 3008-1203) is an international peer-reviewed journal that offers open access and serves as a multidisciplinary platform for the state-of-the-art research and practice in the domains of oceanscience, technology, and engineering. The journal is dedicated to publishing articles, reviews and perspectives in these areas, with the goal of promptly disseminating and promoting theoretical, numerical, site-based, and experimental advancements in the context of global sustainability.
Ocean
Hybrid deep learning method for predicting vertical bending moments at midship
8-Jun-2026