Wind turbine operators face critical challenges in predicting bearing failures: data is limited, operating conditions vary widely, and sharing raw sensor data raises privacy concerns. A new decentralized federated learning framework addresses all three obstacles simultaneously, enabling accurate remaining useful life prediction from a single wind asset without compromising data security.
The innovation centers on a "virtual client" paradigm that treats statistical health indicators extracted from vibration signals—such as root mean square, standard deviation, and crest factor—as independent federated learning participants. Rather than requiring multiple wind turbines or multiple operators to collaborate, this approach allows a single 2.2 MW turbine to train prognostic models internally by partitioning its own diagnostic features as virtual learners. This abstraction preserves industrial applicability while embedding domain knowledge from bearing diagnostics directly into the architecture.
The framework operates on a peer-to-peer basis without a central aggregation server, with an adaptive mechanism governing model parameter exchange based on performance metrics. Crucially, only model parameters—never raw vibration or operational data—are transmitted, providing architectural privacy protection aligned with industrial data governance requirements.
Field validation on a Suzlon turbine experiencing natural bearing degradation yielded strong results: R² of 0.94, ROC-AUC of 0.969, and precision of 1.00, meaning zero false alarms that could trigger unnecessary maintenance. Root mean square error was 52% lower than centralized alternatives. Additional testing on three industry benchmark datasets confirmed generalization across different bearing failure modes, with R² values between 0.89 and 0.93.
The computational overhead matches standard federated learning, making deployment straightforward for existing wind operations. While formal differential privacy guarantees remain future work, the framework demonstrates that privacy preservation and prognostic accuracy are achievable through careful system design. For asset managers, this enables predictive maintenance strategies that reduce unplanned downtime while respecting operational confidentiality constraints.
