随着我国特高压输电以及新型电力系统建设进入新时代，如何保证电力系统安全可靠运行并使用户得到优质的电能质量，成为电网检修和运维人员所需克服的主要困难。当前人工以及直升机巡检受到气候或地理条件限制，之后的电力巡检任务正逐步由无人机取代。在未来，如何利用卷积神经网络等人工智能方法对巡检图像，尤其是带有噪声和受背景影响的绝缘子图像进行快速准确地处理，以匹配无人机巡检的效率，是智能电网主要发展方向之一。

本文以输电杆塔处的绝缘子及其缺陷为主要检测目标，使用LabelImg处理无人机巡检过程中获得的图像，得到适合于网络训练的PASCAL VOC 2007格式数据集。结合无人机在线巡检要求，以YOLOv7算法为基础，分别提出了YOLOv7-ReLU-M高精度算法和YOLOv7-SAPD轻量化算法，能够及时发现包含自爆缺陷的绝缘子。主要研究内容如下：

由于无人机巡检获得的绝缘子图像含有噪声，并且会受到森林、城市等不同背景的干扰，提出了基于高斯模糊的绝缘子数据预处理方法，提高了网络训练效率，也更有利于提高实际应用中的预测精度。

在保证检测速度基本不变的前提下，为了提高巡检绝缘子的检测准确度和定位精度，提出了YOLOv7-ReLU-M目标检测网络。该网络以YOLOv7算法为基础，使用深层拉普拉斯金字塔和最大池化层深度融合了浅层特征与深层特征，并增大了网络检测部分的感受野，使其能更准确地检测绝缘子，尤其是自爆缺陷。另一方面，由于YOLOv7原始网络在训练数据时会占用大量内存，导致训练速度下降，本文提出了加权GsIoU Loss损失函数，通过平衡难易样本以及综合考虑预测框，大幅提升了算法的训练效率，并通过实验验证，也有利于提升检测准确度。对比实验及消融研究结果表明，经过上述改进的网络可大大提高对绝缘子和缺陷的检测准确度，而且其检测速度仍然符合在线巡检要求。

为了降低YOLOv7算法的运行要求，将YOLOv7的主干网络及部分特征提取网络与SAPD算法结合，提出YOLOv7-SAPD无锚框算法，降低了占用的运行内存，并在保证检测准确度的前提下，极大地提高了检测速度，使其更适用于城镇配电网的在线巡检。

As China's UHV transmission and the construction of new power systems enter a new era, how to make sure the securable and dependable movement of electrical power systems and users to obtain high-quality power quality have become the main difficulties for grid maintenance and operation and maintenance personnel to overcome. At present, manual and helicopter inspection is limited by climate or geographical conditions, so the inspection task in the mountains is gradually replaced by drones. In the future, how to use convolutional neural networks and other artificial intelligence methods to quickly and accurately process inspection images, especially insulator images with noise and background influence, to match the efficiency of UAV inspection is one of the main development directions of smart grid.

In this paper, the insulators and defects at the transmission tower are the main detection targets, and LabelImg is used to process the images obtained in the UAV inspection process to obtain a PASCAL VOC 2007 format data set suitable for network training. Combined with the requirements of UAV online inspection and based on YOLOv7 algorithm, the YOLOv7-ReLU-M high-precision algorithm and YOLOv7-SAPD lightweight algorithm are proposed respectively, which can timely detect insulators containing self-detonation defects. The primary investigation components are listed below:

Since the insulator image obtained by UAV inspection contains noise and will be interfered by different backgrounds such as forest and city, a pre-processing method of insulator data based on Gaussian fuzzy is proposed, which improves the efficiency of network training and is more conducive to improving the prediction accuracy in practical applications.

The YOLOv7-ReLU-M object detection network is proposed to improve the detection accuracy and positioning accuracy of inspection insulators on the premise of keeping the detection speed basically unchanged. Based on the YOLOv7 algorithm, the network uses deep Laplacian pyramid and maximum pooling layer to integrate shallow and deep features, and increases the receptive field of the detection part of the network, so that it can detect insulators more accurately, especially self-detonation defects. On the other hand, since the original YOLOv7 network occupied a large amount of memory when training data, resulting in a decrease in training speed, a weighted GsIoU Loss function is proposed, which greatly improve the training efficiency of the algorithm by balancing the difficulty samples and comprehensively considering the prediction box. And the experiment proves that it can improve the detection accuracy. The results of comparison experiment and ablation study show that the improved network can greatly improve the detection accuracy of insulators and defects, and its detection speed still meets the requirements of online inspection.

In order to reduce the operation requirements of YOLOv7 algorithm, the backbone network of YOLOv7 and part of the feature extraction network are combined with the SAPD algorithm, and the YOLOv7-SAPD anchor-free algorithm is proposed, which reduces the operating memory occupied and greatly improves the detection speed on the premise of ensuring the detection accuracy. It is more suitable for online inspection of urban distribution network.