Defect Detection in GIS X-Ray Images Based on Improved YOLOv10

Area of Science:

• Electrical Engineering
• Computer Vision
• Materials Science

Background:

• Internal defects in Gas-Insulated Switchgear (GIS) pose risks to power system reliability.
• Automated X-ray defect detection is challenging due to small, low-contrast defects and complex backgrounds.

Purpose of the Study:

• To develop an enhanced object-detection model for accurate GIS defect identification.
• To improve the detection of small and low-contrast defects in GIS X-ray images.

Main Methods:

• Utilized a lightweight YOLOv10n framework, enhanced with Normalized Wasserstein Distance (NWD) loss for small object localization.
• Integrated Monte Carlo (MCAttn) and Parallelized Patch-Aware (PPA) attention mechanisms for improved feature extraction.
• Employed a GFPN-inspired neck for effective multi-scale feature fusion.

Main Results:

• The enhanced model achieved a mean Average Precision (mAP) of 0.674 (IoU 0.5:0.95).
• Demonstrated a 5.0 percentage point improvement over the baseline YOLOv10n model.
• Outperformed other comparative models in detecting challenging defects.

Conclusions:

• The proposed model offers an effective solution for automated GIS defect detection.
• Significant potential to enhance power grid maintenance efficiency and safety.
• Qualitative results confirm superior detection of small and low-contrast defects.