Research on deep-sea mining risers damage identification based on Multi-Scale Squeeze-and-Excitation Residual Network

Deep-sea mining risers are critical components connecting the mining vessel to the buffer station. Long-term operation in complex marine environments makes them prone to crack initiation and propagation, potentially compromising structural integrity and operational safety. Early and accurate damage identification is therefore essential. However, research on damage localization and severity quantification in deep-sea mining risers remains limited. This study develops a Multi-Scale Squeeze-and-Excitation Residual Network (MS-SE-ResNet) framework with two independently trained subnetworks for damage localization and severity quantification. A cracked-riser model is established using Euler-Bernoulli beam theory and a local-flexibility crack model, with task-specific features constructed from the first six modal responses. Results show that, in a single run with random seed 42, the localization subnetwork achieved 98.68% exact node-level accuracy and a mean absolute position deviation of 0.0598 nodes. Across five random seeds, MS-SE-ResNet achieved the best mean severity-quantification performance among the evaluated ablation variants, with a mean R 2 of 0.9943 ± 0.0022. Noise tests conducted at different signal-to-noise ratios (SNRs) show that localization performance improves as the noise level decreases, with stronger cross-SNR generalization under lower-noise conditions. SHapley Additive exPlanations (SHAP) analysis shows that damage-severity predictions are primarily influenced by a limited number of damage-sensitive modal features.