A Preprocessing Method for Insulation Pull Rod Defect Dataset Based on the YOLOv5s Object Detection Network

Area of Science:

• Electrical Engineering
• Materials Science
• Computer Vision

Background:

• Gas-insulated switchgear (GIS) components, specifically insulation pull rods, are prone to micro-defects from production.
• Intelligent identification methods for insulation faults require large, balanced datasets, which are challenging to obtain due to limited defective samples and imbalanced defect types.
• Existing methods struggle with poor recognition performance when faced with imbalanced defect data.

Purpose of the Study:

• To propose an effective data preprocessing method for insulation pull rod defect feature datasets.
• To enhance the recognition performance of intelligent identification systems for insulation pull rod defects.
• To address the challenges of limited defective samples and imbalanced defect categories in actual production data.

Main Methods:

• Utilized the YOLOv5s algorithm for defect detection in insulation pull rod images, establishing a dataset with five defect categories.
• Introduced two preprocessing techniques: copy-paste augmentation within images and bounding box correction for hair-like impurities.
• Integrated copy-paste augmentation with Mosaic data augmentation and refined bounding box corrections for hair-like impurities.

Main Results:

• The proposed preprocessing methods effectively enhance small-sized defect targets (impurities and bubbles) while maintaining detection performance for other defect types.
• The combined approach of copy-paste augmentation, Mosaic augmentation, and bounding box correction significantly improved the overall model performance.
• Demonstrated a specific enhancement for small-sized defect targets, crucial for accurate fault identification.

Conclusions:

• The developed data preprocessing method is effective in improving the performance of defect detection models for insulation pull rods in GIS.
• The integration of specific augmentation and correction techniques addresses the issue of imbalanced defect data, leading to more robust identification systems.
• This approach offers a viable solution for enhancing the reliability of intelligent fault identification in electrical equipment.