Abstract
Insulation pull rods used in gas-insulated switchgear (GIS) inevitably contain the micro defects generated during production. The intelligent identification method, which requires large datasets with a balanced distribution of defect types, is regarded as the prevailing way to avoid insulation faults. However, the number of defective pull rods is limited, and the occurrence of different types of defects is highly imbalanced in actual production, leading to poor recognition performance. Thus, this work proposes a data preprocessing method for the insulation pull rod defect feature dataset. In this work, the YOLOv5s algorithm is used to detect defects in insulation pull rod images, creating a dataset with five defect categories. Two preprocessing methods for impurities and bubbles are introduced, including copy-paste within images and bounding box corrections for hair-like impurities. The results show that these two methods can specifically enhance small-sized defect targets while maintaining the detection performance for other types of targets. In contrast, the proposed method integrates copy-paste within images with Mosaic data augmentation and corrects bounding boxes for hair-like impurities significantly improving the model's performance.