Privacy-Preserving Distributed Fault Diagnosis for Multiple Wind Farms Using a Federated Feature Fusion Method

While data-driven methods have gained prominence in wind turbine fault diagnosis, their effectiveness is increasingly constrained by the proliferation of data silos. This phenomenon primarily arises from the reluctance toward data sharing, a trend exacerbated by intensifying commercial competition and mounting privacy concerns. Such limitations fundamentally challenge centralized data processing paradigms in this field. Federated learning provides a viable technical solution to mitigate data silos while preserving data sovereignty. Nevertheless, two critical challenges that substantially hinder practical implementation remain unresolved: 1) pronounced data heterogeneity among geographically distributed wind farms, and 2) prohibitive communication overhead stemming from sluggish convergence rates in distributed optimization. To bridge these gaps, we propose the Federated Feature Fusion (Fed-FF), a novel cross-farm fault diagnosis framework that synergistically integrates feature-level federation with convergence acceleration. Furthermore, the theoretical guarantees of the proposed method are studied in this paper, where the convergence risk bounds for both convex and non-convex settings are derived. The proposed method is then validated against three real-world datasets collected in Yunnan Province, Jiangsu Province and Shanghai, China. The experimental results show that the proposed method achieves an average accuracy of 93.83%, outperforming several state-of-the-art fault diagnosis methods, while communication costs are reduced by up to 94.71%. Note to Practitioners—As competition intensifies and data privacy regulations tighten, wind farm operators are increasingly withholding operational data, creating persistent data silos. Compounding this challenge, the inherent heterogeneity of SCADA measurements across geographically dispersed farms exacerbates communication overhead in conventional federated learning. To overcome these limitations, we propose a novel approach that ensures data privacy while significantly accelerating convergence and reducing communication costs, thereby lowering operational expenses for wind farms. Supported by rigorous theoretical guarantees, our method enhances both reliability and practical applicability. As a distributed fault diagnosis framework, it facilitates efficient collaboration and statistical information sharing among multiple wind farms without exchanging raw data, enabling precise fault localization. Beyond wind farms, this method is also applicable to diverse domains, including healthcare, intelligent connected vehicles, photovoltaic power plants, and smart manufacturing, offering a versatile and privacy-preserving solution for distributed learning.