Unsupervised Automated Identification Method for Abnormal States of Wind Turbine Gearboxes

• • The U-net autoencoder with log Mel-band energy features achieves 93.34% accuracy on both factory and wind farm test sets, outperforming MLSTM-AE (70.50%), DCAE (75.25%), and SDAE (86.47%) when using the same feature extraction, demonstrating a 6.87–22.84 percentage point improvement over these baselines. • • Without log Mel-band energy feature extraction, the same U-net model attains only 88.28% accuracy, a 5.06 percentage point drop, while other models degrade to 69.38–81.88%, proving that the feature extraction stage is critical for robust anomaly detection in raw vibration signals. • • The model requires only 7.7×10^6 parameters, which is 1–2 orders of magnitude lower than DCAE (3.071×10^8) and SDAE (3.100×10^6), enabling deployment on edge devices for online monitoring without substantial computational overhead. • • Validation on real wind farm data from Yangtouya, Shanxi, confirms the method's industrial applicability, achieving the same 93.34% accuracy as factory tests, and establishing a health baseline for full-lifecycle management without labeled fault data.