Intelligent identification method of insulation pull rod defects based on intactness-aware Mosaic data augmentation and fusion of YOLOv5s

Abstract

The authors introduce the intactness-aware Mosaic data augmentation strategy, designed to tackle challenges such as low accuracy in detecting defects in insulation pull rods, limited timeliness in intelligent analysis, and the absence of a comprehensive database for information on insulation pull rod defects. The proposed strategy incorporates the YOLOv5s algorithm for detecting defects in insulation pull rods. Initially, the YOLOv5s network was constructed, and a dataset containing photos of insulation pull rods with white spots, fractures, impurities, and bubble flaws was compiled to capture images of defects. The research presented a data enhancement approach to improve the images and establish a dataset for insulation pull rod defects. The YOLOv5s algorithm was applied for both training and testing purposes. A comparative analysis was conducted to assess the detection performance of YOLOv5s against a conventional target detector for identifying defects in insulation pull rods. Furthermore, the utility of Mosaic's data augmentation technique, which incorporates intactness awareness, was evaluated to enhance the accuracy of identifying insulation pull rod defects. The research findings indicate that the YOLOv5s algorithm is employed for intelligent detection and precise localisation of flaws. The intactness-aware Mosaic data augmentation strategy significantly improves the accuracy of detecting faults in insulation pull rods. The YOLOv5s model used achieves a performance index [email protected]:0.95 of 0.563 on the test set, distinct from the training set data. With a threshold of 0.5, the [email protected] score is 0.904, indicating a substantial improvement in both detection efficiency and accuracy compared to conventional target detection methods. Innovative approaches for identifying defects in insulation pull rods are introduced.