Insulator Surface Defect Detection Method Based on Graph Feature Diffusion Distillation.

Area of Science:

• Electrical Engineering
• Computer Vision
• Materials Science

Background:

• Surface defects on power insulators pose risks to grid stability.
• Current defect detection methods struggle with limited defect samples, irregular shapes, and poor localization accuracy.

Purpose of the Study:

• To develop an automated defect detection method for power insulators that addresses sample scarcity and improves localization accuracy.
• To enhance the robustness and generalization capabilities of defect detection models.

Main Methods:

• Proposes a Graph Feature Diffusion Distillation (GFDD) method.
• Employs a dual-teachers architecture with graph feature consistency constraints to mitigate feature bias.
• Utilizes a cross-layer feature fusion module for dynamic multi-scale information aggregation.
• Incorporates a diffusion distillation mechanism and channel attention for enhanced global context modeling.

Main Results:

• Achieved 96.6% Pi.AUROC, 97.7% Im.AUROC, and 95.1% F1-score on a self-built dataset, outperforming existing methods by 2.4-3.2%.
• Demonstrated excellent generalization and robustness across multiple public datasets.
• GFDD provides a high-precision solution for automated inspection of insulator surface defects.

Conclusions:

• The GFDD method offers a significant advancement in automated power insulator defect detection.
• The approach effectively handles challenges like limited data and irregular defect morphologies.
• The method holds considerable engineering value for practical applications in power infrastructure maintenance.