Federated Incremental Collaborative Fault Diagnosis Method for Dynamic Data Streams in Multiple Wind Farms.

Area of Science:

• Renewable Energy Systems
• Machine Learning
• Data Science

Background:

• Growing privacy concerns and data silos necessitate advanced solutions for wind turbine fault diagnosis.
• Existing federated learning methods struggle with dynamic data streams and the introduction of new fault classes, leading to impractical storage and computational demands.
• Current models suffer from significant memory degradation when diagnosing new faults in heterogeneous, resource-limited wind farm environments.

Purpose of the Study:

• To propose a federated incremental collaborative fault diagnosis method for dynamic data streams across multiple wind farms.
• To address the challenges of new fault class detection, model plasticity-stability balance, and global model adaptation in dynamic environments.
• To mitigate memory degradation and improve the performance of federated learning in wind turbine fault diagnosis.

Main Methods:

• A novel fault class detection method to identify the introduction of new fault classes.
• A plasticity-stability balance mechanism for local fault diagnosis models to combat the fading memory problem.
• A global model adaptive compensatory method to address memory degradation in the aggregated model due to data heterogeneity.

Main Results:

• The proposed method effectively mitigates fading memory issues in federated learning models for wind turbine fault diagnosis.
• Validation using real-world data from three Chinese wind farms demonstrates superior performance compared to state-of-the-art methods.
• The approach successfully handles dynamic data streams and the emergence of new fault classes.

Conclusions:

• The federated incremental collaborative fault diagnosis method offers a practical and effective solution for wind turbine maintenance in dynamic environments.
• The study highlights the importance of addressing model plasticity, stability, and heterogeneity for robust federated learning in industrial applications.
• This research contributes to enhancing the reliability and efficiency of wind energy by improving fault diagnosis capabilities.