Frequency prototype-based source-free domain adaptation method and its application of wind turbine gearbox fault diagnosis

Abstract Source-free domain adaptation (SFDA) for wind turbine gearbox fault diagnosis enables cross-domain generalization without accessing source data, effectively addressing privacy, security, and deployment challenges in real-world wind farm scenarios. However, existing approaches remain limited by the sensitivity of time-domain features to operating conditions and the unstable of pseudo-label supervision, which together hinder reliable domain alignment and robust cross-domain adaptation. To overcome these limitations, a novel Frequency Prototype-based Source-Free Domain Adaptation (FP-SFDA) method is proposed for source-free cross-domain mechanical fault diagnosis. Firstly, the frequency-domain prototype extractor (FDPE) is proposed to capture domain-invariant representations from the spectral energy distribution of the source domain, enabling robust feature abstraction against condition variations. Furthermore, a frequency prototype induced domain alignment (FPIDA) mechanism is presented. Concretely, the prototype predictor is trained by aligning predicted and ground-truth frequency-domain prototypes to learn representative source features. Then, by feeding target features into this trained predictor, pseudo-source representations are generated, which serve as a reference for aligning the target domain and thus achieving effective source free domain adaptation.Experimental results on real wind turbine gearbox fault dataset and one public dataset demonstrate the effectiveness and stability of the proposed method in source-free wind turbine gearbox fault diagnosis compared with recent SFDA methods.