Redundancy Removal and Knowledge Alignment-Based Personalized Federated Learning for Online Condition Monitoring.

Area of Science:

• Electrical Engineering
• Artificial Intelligence
• Cybersecurity

Background:

• Online monitoring of high-voltage electrical equipment for partial discharges (PDs) is crucial for safety and reliability.
• Existing systems face data security and privacy challenges during data transmission and storage.
• Federated learning (FL) offers a privacy-preserving approach for collaborative model training without raw data sharing.

Purpose of the Study:

• To develop an advanced FL framework for robust and secure online PD monitoring in switchgear.
• To improve the informativeness and representativeness of the global model in FL systems.
• To enhance client model personalization for local data while maintaining data privacy.

Main Methods:

• A novel FL framework employing a maximum diversity and minimum redundancy strategy for client model evaluation.
• Introduction of a spatial-logic alignment module with knowledge distillation for enhanced client model personalization.
• Implementation of a hybrid architecture with leaf clients, branch clients, and a central server, leveraging edge computing.

Main Results:

• The proposed framework generates a more informative and representative global model compared to traditional performance-based aggregation.
• Spatial-logic alignment and knowledge distillation significantly improve client model personalization.
• Experimental validation on multiple datasets demonstrates superior performance over state-of-the-art methods.

Conclusions:

• The novel FL framework effectively addresses data security and privacy concerns in PD monitoring.
• The diversity-based model evaluation and spatial-logic alignment enhance both global model accuracy and local personalization.
• The hybrid architecture and edge computing integration enable efficient, low-latency monitoring for high-voltage equipment safety.