Unsupervised Screening and Lightweight Cross-Modal Classification for Damage Identification in Offshore Jacket Structures

Damage identification in offshore jacket structures requires state screening when damaged samples are scarce and local damage classification under measurement noise. We propose a framework that combines unsupervised screening with a lightweight bimodal classifier. An autoencoder learns undamaged response features from undamaged data alone and provides initial damage screening. The bimodal classifier combines acceleration time histories and mode-shape curvature difference (MSCD) images through bidirectional cross-modal attention and gated fusion. Beam simulations achieved 92.26 ± 1.41% accuracy at 15% Gaussian noise, exceeding direct concatenation by 3.63 percentage points. In scaled jacket tests, the autoencoder detected 98.58% of damaged samples. With 5% low-frequency colored perturbation, the bimodal classifier achieved 100% accuracy across seven structural states on a bootstrap-resampled test set. The classifier contains 3.44 million parameters and required approximately 12 ms per sample on the tested CPU.