Incremental learning with prototype calibration and dynamic proxy for wind turbine fault diagnosis under time-varying operating conditions.

For supervisory control and data acquisition (SCADA) based intelligent fault diagnosis (IFD) of wind turbines, the domain-incremental learning (D-IL) paradigm enables the model to continuously adapt to new operational data while mitigating catastrophic forgetting (CF). However, for wind turbines under complex operating conditions, existing studies still have limitations: on the one hand, operating condition drift disrupts feature consistency, degrading the effectiveness of historical knowledge replay; on the other hand, class boundaries in the feature space become increasingly blurred, leading to a gradual deterioration in the ability to separate fault categories when their feature distributions shift. To address these issues, this paper proposes a D-IL paradigm with prototype calibration and dynamic proxy (PCDP). First, during IFD model training, dynamic proxies adjust the compactness of intra-class feature representations and the separability of inter-class representations, ensuring that feature changes under time-varying operating conditions are properly captured. Then, features extracted from graph structured data are aggregated into class level prototypes to represent the current operational status and support subsequent D-IL training. Finally, during the D-IL process, a prototype calibration network corrects the drift of class level prototypes used for prototype replay. The proposed PCDP is validated in two cases (main-shaft bearing faults and blade icing) and outperforms advanced approaches.