Automatic detection of broken strands defects in distribution network conductors based on deep learning

This paper proposes an improved YOLOv5-based method for automatic detection of broken strand defects in distribution network conductors. To address challenges in complex environments and slender target localization, we introduce a lightweight neck network, channel attention mechanism, and an efficient feature extraction network (EFEN). Experiments on a self-built UAV inspection dataset of 1,364 images show that the improved model achieves 0.89 precision and 0.88 mAP, outperforming the original YOLOv5 by 6%. Field tests on over 1 km of power lines demonstrate an 88.4% detection rate with no false alarms.