ABSTRACT The reliability of wind energy systems heavily relies on the condition of critical mechanical components, with rolling element bearings being particularly susceptible to degradation and premature failure. Accurately predicting the remaining useful life (RUL) of these components is crucial for implementing predictive maintenance strategies aimed at minimizing downtime and reducing maintenance costs. Yet, practical use of data‐driven prognostic models in industrial settings faces significant challenges, including limited access to run‐to‐failure data, varied operating conditions, and stringent data privacy requirements. This paper introduces a novel decentralized peer‐to‐peer federated learning (P2P‐FL) framework that addresses all three challenges simultaneously. The central methodological innovation is the “virtual client” (VC) paradigm, in which statistical health indicators (HIs) extracted from vibration signals—root mean square (RMS), standard deviation (STD), crest factor (CF), and peak‐to‐RMS ratio (P2RMS)—are treated as independent federated learning (FL) clients. This paradigm is positioned as a domain‐adapted instantiation of feature‐level federation designed for single‐asset industrial prognostics, where the conventional multi‐entity FL requirement is structurally infeasible. This abstraction enables FL from a single physical machine, eliminates the dependency on multiple physically distinct data owners, and embeds domain knowledge from bearing diagnostics directly into the learning architecture. Built upon this paradigm, a fully decentralized peer‐to‐peer (P2P) architecture replaces the conventional central aggregation server with direct P2P model parameter exchange, governed by an adaptive performance‐based aggregation mechanism. The framework operates under an honest‐but‐curious adversary model, providing data‐minimising architectural privacy (no raw HI data is ever transmitted) while acknowledging that formal differential privacy guarantees are identified as future work. To align prognostic evaluation with operational maintenance decisions, we further propose a dual regression–classification evaluation framework combining standard regression metrics with ROC analysis, precision–recall curves, and probability calibration. Experimental validation on vibration data collected from an operational 2.2 MW Suzlon wind turbine (WT) bearing undergoing natural inner race degradation demonstrates that the proposed P2P adaptive framework achieves R ² = 0.94, ROC‐AUC = 0.969, precision = 1.00 (zero false alarms), and a 52% root mean square error (RMSE) reduction over the centralized FL baseline. Additional validation on three FEMTO‐ST benchmark bearing trajectories confirms generalisability ( R ² = 0.89–0.93). These results are obtained while transmitting only model parameters—never raw operational data—at computational cost equivalent to standard FL, demonstrating that privacy preservation and state‐of‐the‐art prognostic accuracy are simultaneously achievable through careful architectural design.
Sidhom et al. (Mon,) studied this question.