Real-Time Wave Identification for Stationary Marine Units: A Deep Transfer Learning Approach With Adaptive Windowing

This study presents an artificial-intelligence-based framework for real-time wave parameter identification in stationary marine units, addressing long-standing challenges in sea state estimation (SSE). Traditional methods often suffer from significant time lag and limited accuracy under rapidly changing conditions. To overcome these, we integrate a dynamic adaptive predictor with a deep-transfer-learning-based SSE estimator. The predictor, based on hybrid signal decomposition and statistical evaluation, dynamically adjusts the observation window based on predicted future response fluctuations, thus minimizing latency and optimizing input segments for the deep learning model, responsible for SSE. Time-series responses are converted into scalogram images to extract rich spectral–temporal features, which are then used to fine-tune pretrained deep neural networks through transfer learning. This dual-stage design ensures that the deep learning model receives statistically stable and temporally relevant inputs, enhancing both responsiveness and accuracy. Results show that pretrained-image-based convolutional neural networks (CNNs), such as DarkNet-19, achieve over 97% validation accuracy within just a few training epochs, offering high efficiency and reduced training cost. In comparison, a 2-D CNN trained from scratch reaches 98.36% accuracy with longer training, while a 1-D CNN on raw signals achieves 83.42% . When integrated with the adaptive predictor, the proposed system estimates wave parameters with average errors as low as 2.94% for <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"><tex-math notation="LaTeX">${{H}_s}$</tex-math></inline-formula>, 3.20% for <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"><tex-math notation="LaTeX">${{T}_p}$</tex-math></inline-formula>, and 2.06% for <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"><tex-math notation="LaTeX">$\beta $</tex-math></inline-formula> in steady conditions and maintains robustness under random conditions. These findings highlight the framework’s real-time applicability and scalability for onboard SSE.