SWIFT: Spectral Wavelet Integrated Fault Transformer for Wind Turbine Anomaly Detection

Offshore wind turbine (OWT) condition monitoring is a critical challenge as wind fleets scale, driven by the high cost of maintenance, harsh marine environments, and the complex, non-stationary nature of turbine data. Existing anomaly detection methods - ranging from statistical control charts to deep prediction-based models - often struggle with high environmental noise, sparse fault events, and the wide range of temporal and frequency scales over which turbine failures manifest. In this paper, we propose SWIFT (Spectral Wavelet Integrated Fault Transformer), an end-to-end, domain-aware anomaly detection framework designed specifically for offshore wind turbine monitoring. SWIFT integrates three complementary components: i) an association discrepancy mechanism to detect anomalous temporal dependencies rather than relying solely on point-wise prediction errors, ii) a Stationary Wavelet Transform (SWT) front-end to disentangle fault-relevant frequency bands from non-stationary environmental noise, and iii) a novel Multi-Scale Attention (MSA) module with scale-specific priors to simultaneously capture transient faults and long-term degradation patterns. We evaluate SWIFT on CARE2Compare, a large-scale real-world dataset comprising 89 turbine-years from 36 turbines across three wind farms, and CART2, an experimental turbine built by the National Renewable Energy Laboratory. Experimental results demonstrate that SWIFT consistently outperforms the state-of-the-art Hawkeye model, as well as other popular approaches such as OCSVM and Isolation Forest. SWIFT achieves higher precision-recall stability and a maximum F-score of 0.5 and a ROC-AUC of 0.69 on CARE compared to Hawkeye's maximum F-score of 0.41 and ROC-AUC of 0.55. Ablation studies confirm that combining frequency-domain decomposition with multi-scale attention is essential for robust fault detection in offshore wind turbines. These findings highlight the importance of incorporating domain-specific, multi-resolution modeling strategies when analyzing high-dimensional, noisy time series in renewable energy systems.