Insulator mechanical damage dataset

Regular and rigorous inspection of outdoor insulators is essential for uninterrupted power grid operation. Recent advances in computer vision enabled replacing conventional subjective, costly, and inefficient visual insulator inspection with automated diagnosis from unmanned aerial vehicle (UAV) taken images. In this study, advanced computer vision algorithms, namely, family of YOLOv3 and YOLOv5 architectures, are trained and compared for classification of frequently encountered insulator mechanical faults from UAV images. Hence, a dataset of 8886 insulator images under normal, bird pecking, cracking, and missing cap conditions is collected to train and evaluate our classifiers. In model selection, YOLO models are compared using architectural complexity (number of parameters) and mean average precision at intersection over union (IOU) thresholds from 0.5 to 0.95 (mAP@0.5:0.95). According to model selection results, YOLOv5x and YOLOv5n are the best models in terms of mAP@0.5:0.95 and complexity. Models evaluation using test set reveals that YOLOv5x and YOLOv5n achieved remarkable performance of 95.8% and 90.9% in terms of mAP@0.5:0.95, respectively. Although YOLOv5x reported higher mAP@0.5:0.95, YOLOv5n requires ~41 times less memory and ~49 times less floating-point operations for image classification at the expense of ~5% reduction in mAP@0.5:0.95, which makes YOLOv5n an attractive option for resource-constrained hardware such as UAVs

对户外绝缘子开展定期且严格的检测，是保障电网不间断稳定运行的核心前提。近年来计算机视觉领域的技术进步，使得我们可以通过无人机（Unmanned Aerial Vehicle）拍摄的图像实现自动化故障诊断，替代传统主观性强、成本高昂且效率低下的人工绝缘子视觉检测方式。本研究针对无人机拍摄的绝缘子图像，对YOLOv3与YOLOv5系列等先进计算机视觉算法进行训练与对比，以实现绝缘子常见机械故障的分类识别。为此，本研究收集了包含8886张绝缘子图像的数据集，图像涵盖正常、鸟啄、开裂以及帽体缺失四种工况，用于训练与评估所搭建的分类器。在模型选型环节，本研究以模型架构复杂度（参数数量）以及交并比（Intersection over Union，IOU）阈值0.5至0.95下的平均精度均值（mAP@0.5:0.95）作为评估指标，对YOLO系列模型开展对比选型。模型选型结果显示，YOLOv5x与YOLOv5n分别在mAP@0.5:0.95性能与架构复杂度维度上表现最优。基于测试集的模型评估结果表明，YOLOv5x与YOLOv5n的mAP@0.5:0.95分别达到95.8%与90.9%，性能表现优异。尽管YOLOv5x的mAP@0.5:0.95更高，但YOLOv5n在图像分类任务中所需内存仅为前者的1/41左右，浮点运算量仅为前者的1/49左右，仅以约5%的mAP@0.5:0.95降幅为代价，这使其在无人机等资源受限的硬件场景中具备极具吸引力的应用价值。

搜集汇总

背景与挑战

背景概述

该数据集包含8886张无人机拍摄的绝缘子图像，涵盖正常状态及三种机械损伤类型（鸟啄、开裂、缺失帽），用于训练计算机视觉模型实现电力设备自动化检测。数据以多种格式存储，总容量28.38GB，已成功应用于YOLOv5等模型的性能验证。