Unsupervised anomaly detection in floating offshore wind turbines based on system responses

Floating offshore wind turbines (FOWTs) are emerging as a central technology for offshore wind development, yet their operational reliability is challenged by harsh marine environments and complex system dynamics. This study develops a simulation-based framework to benchmark unsupervised deep learning methods for fault detection in FOWTs. A synthetic dataset was generated using OpenFAST for a 22 MW reference turbine, comprising 1,110 samples across multiple IEC-compliant fault scenarios. All models were evaluated using system-level response signals only, enabling fault inference without component-specific sensors. Within this controlled setting, the proposed Two-phase Self-Refining Transformer (TSR-Former) demonstrated strong performance, yielding a sample-wise F1-score of 0.931 and a median detection latency of 8.85 s. The TSR-Former also maintained robustness under diverse conditions, including varying wind speeds, wind–wave misalignment, and severe additive sensor noise, achieving an F1-score of 0.78 at 10 dB SNR. While validated in a simulation environment, the framework provides a structured basis for evaluating unsupervised methods. Future research should incorporate experimental or field data to validate this approach and assess its robustness under real-world operational constraints. • Unsupervised FOWT fault detection is benchmarked on IEC-based simulations. • A novel Transformer model detects faults using response signals. • The model is robust against simulated sensor noise and complex sea states.