Distribution-aware federated learning for wind turbine fault diagnosis in edge-cloud under data imbalance

In the practical application of wind energy technology, the reliability of wind turbine systems is directly linked to the stable output of renewable electricity. Data imbalance in fault monitoring networks causes classification bias in diagnostic models, which increases the risk of misdiagnosis and ultimately undermines the stability of wind power. In the context of distributed monitoring, this issue is further exacerbated by the geographically dispersed layout of wind farms, which significantly amplifies data imbalance. To address these challenges, a global distribution-aware logits adjustment federated framework is proposed, along with a dynamic estimation mechanism for global fault distribution based on edge node logits. This mechanism accurately captures the global fault feature distribution and enhances attention to minority categories. Additionally, a feature compensation architecture for edge-cloud bidirectional collaboration is developed. By constructing a feature similarity matrix, knowledge distillation is performed at the feature level to transfer global feature knowledge to the edge, enabling local models to learn global common features while preserving the personalized feature extraction capabilities of edge nodes. On-site experiments at Kelmarsh and Xinjiang wind farms demonstrated that the proposed method achieved accuracies of 0.8617 and 0.9891, respectively, showcasing its effectiveness.