A Domain-Adaptive Framework for Damage Detection of Offshore Wind Turbine Jacket Substructures using Machine Learning

Implementing effective Structural Health Monitoring (SHM) for offshore wind turbine jacket substructures is critical for minimizing Operations and Maintenance (O&M) costs, yet it remains challenging due to the scarcity of realworld fault data and the difficulty of detecting incipient damage. To address this, a high-fidelity coupled simulation framework integrating OpenFAST and Abaqus was established, generating a comprehensive database covering 17 structural health states, including complex cross-level damage scenarios. The study evaluates different feature engineering strategies across machine learning and deep learning models. Results indicate that Extreme Gradient Boosting (XGBoost), utilizing a hierarchical classification strategy and raw time-domain signals under wind-only loading conditions, achieved a robust 100 % accuracy, establishing a strong baseline. In contrast, deep learning approaches, including LSTM and ResNet, demonstrated the advantage of automated feature extraction, effectively capturing temporal dependencies and stiffness-induced spectral shifts respectively. However, under complex environmental loading, these architectures exhibited higher sensitivity to data scale and signal dominance (e.g., damage masking), highlighting the challenges of end-to-end learning in mixed-fault scenarios. Finally, to mitigate data scarcity for early-stage faults, a hybrid Semi-Supervised Domain Adaptation (SSDA) strategy integrating DANN with Maximum Mean Discrepancy (MMD) was investigated. This approach successfully facilitated knowledge transfer from severe to incipient damage scenarios, improving accuracy from 45 % to 56 %, demonstrating the preliminary feasibility of adversarial learning in enhancing the detectability of subtle structural degradation where labeled data is limited.

Keywords: Structural health monitoring, offshore wind turbine, jacket substructure, machine learning, deep learning, domain adaptation, semi-supervised learning.