A joint Time–Frequency feature learning method for damage localization and severity assessment of offshore jacket structures

Under prolonged service in complex marine environments, the initiation and propagation of localized damage in offshore jacket platforms are recognized as critical threats to structural integrity. However, damage localization is prone to ambiguity, and effective identification of damage severity remains challenging when conventional approaches are employed. To address these limitations, a damage identification framework for jacket structures is proposed, which is grounded in joint time-frequency feature learning. First, a dual-branch identification model is constructed by integrating a Temporal Convolutional Network (TCN) with a Modal Spectral Representation Encoding Network (MSRE-Net). Specifically, temporal evolutionary features of structural response signals are extracted by the TCN branch, whereas modal frequency-domain energy distribution features are captured by the MSRE-Net branch. Subsequently, an adaptive weighting mechanism is introduced to enable dynamic synergy between time- and frequency-domain information, and key hyperparameters are optimized via the Lichtenberg algorithm, thereby enhancing localization accuracy and identification stability under complex operational conditions. Finally, a prototype-based distance metric and an ordinal constraint are formulated within the fused feature space, through which a continuous damage severity index is generated to achieve a gradient characterization of damage states ranging from incipient to critical. Concurrently, a discrimination accuracy of 90.73% is attained for damage severity assessment based on the continuous severity scores.