Distribution line inspection method using multi-scale information augmentation and ensemble learning

This inspection of power distribution lines is evolving from manual to digitalized and intelligent methods with the advancement of Unmanned Aerial Vehicle (UAV) and artificial intelligence algorithms. However, existing universal or those focusing on specific defect detectors struggle to handle the complexity of real-world inspection tasks. To bridge this gap, we proposed a comprehensive distribution line inspection method based on You Only Look Once Exceeding (YOLOX), capable of accurately detecting 14 key defects across 6 critical components. Through large-scale UAV data analysis, we identified two major challenges in practical scenarios: Small Object Detection and Long-tailed Distribution. These insights guide the development of more robust and generalizable inspection methods. And to address these two challenges, we introduced a Multi-Scale Information Augmentation module to enhance the detection of small defects in high-resolution images, and a Dual-Branch Ensemble Classifier structure to mitigate the impact of imbalanced defect distribution. Extensive experiments demonstrated the practical effectiveness of our method, achieving a $Recall$ of 76.2%, $Precision$ of 59.7%, and $AveragePrecision$ at 50% Intersection over Union Threshold ($A{P}_{50}$) of 52.6%. Notably, our method delivered a 51.9% improvement in small-defect $Recall$ and a 24.3% increase in tail-category performance compared to the baseline YOLOX. A fully implemented inspection system built on our framework has been deployed to assist field engineers, significantly reducing manual workload and improving inspection efficiency. These results demonstrated the robustness and adaptability of our method, offering a valuable contribution to the intelligent operation and maintenance of modern power distribution networks.