Abstract
With the growing deployment of wind energy, reliable fault diagnosis of wind turbines (WTs) is increasingly important for sustainable operation and maintenance. However, supervisory control and data acquisition (SCADA) system data are often characterized by severe class imbalance and complex correlations among multi-source features, which hinder accurate fault recognition. To address these challenges, this study proposes a geometric visibility augmentation-mechanism supervision graph (GVA-MSG) framework for imbalanced fault diagnosis of WTs. At the data level, GVA integrates visibility-based neighbor filtering, Voronoi-based boundary constraints, and density-adaptive directional filling into the synthetic minority oversampling technique (SMOTE) to generate minority samples within class manifolds and alleviate imbalance. At the model level, MSG captures temporal dependencies and introduces mechanism supervision based on the wind speed-power relationship. A mechanistic consistency loss aligns model decisions with physical principles, while a latent-space graph adaptively learns cross-channel feature relationships for fault recognition. Comparative, ablation, and generalization experiments are conducted on real SCADA datasets. GVA-MSG achieves a Macro-Accuracy, Macro-F1, and Macro-Gmean of 98.57 %, 96.42 %, and 97.75 %, respectively. Compared with the best-performing data generation baseline, these metrics improve by 1.29, 3.55, and 2.94 percentage points; compared with the best-performing diagnostic baseline, they improve by 0.17, 0.44, and 0.28 percentage points. These results demonstrate that GVA-MSG improves class balance, the quality of generated samples, and mechanism-supervised feature characterization. Moreover, the cross-wind-farm validation results indicate that GVA-MSG maintains stable performance under the tested distribution shifts, demonstrating promising transfer stability and application potential under the current validation conditions.