Enhancing naval hull stress monitoring with neural network-based data imputation

Modern naval vessels operate under complex sea conditions where cyclic wave-induced loads, slamming events and long-term exposure to corrosive environments may contribute significantly to fatigue damage, particularly at welded joints. While the deployment of Hull Monitoring Systems (HMS) provides a vessel-specific approach to fatigue life assessment by capturing continuous stress response data, the utility of such systems may be hindered by missing data. This study applies machine learning to impute missing HMS stress data from a naval ship using a fused dataset of HMS records and wave hindcast data. Deep learning models, including Artificial Neural Networks (ANNs), Recurrent Neural Networks (RNNs), and Long Short-Term Memory Networks (LSTMs), were developed in PyTorch, with feature selection informed by Random Forests, Gradient Boosted Trees and engineering judgement. A Bayesian Maximum Entropy Fusion (BMEF) approach was further employed to combine complementary model strengths. The highest prediction accuracy was achieved when stress data were available as inputs, with BMEF achieving a mean absolute percentage error of 5.06% and outperforming all other models. When stress data were absent, overall performance declined, but LSTMs provided the best results among the models tested. These findings highlight the value of both ensemble methods and temporal models in stress imputation and demonstrate that machine learning can significantly improve fatigue assessment, maintenance planning and life-of-type predictions for naval platforms.