绝缘子作为架空输电线路安全平稳工作至关重要的构件，定期对输电线路绝缘子巡检有着重大的意义。本文以航拍的输电线路绝缘子图像为主要研究对象，针对绝缘子缺陷在航拍图像中占比过小的问题，设计了基准模型为YOLOv7的级联目标检测网络，对基于深度学习的输电线路绝缘子的多种缺陷进行了检测研究。级联检测网络主要由两级目标检测网络组成，第一级目标检测网络识别了绝缘子数据集中的所有绝缘子，把原图的所有绝缘子数据都检测出来，并输入到第二级网络模型；第二级网络模型对绝缘子缺陷进行了目标检测，把绝缘子的各种缺陷情况都检测出来，并同时提供了缺陷的具体类别。本文中主要的研究内容如下：

（1）设计基准模型为YOLOv7的级联目标检测网络。首先，使用形状、颜色、对比度等数据增强算法对输入的图像做预处理，增加数据的多样性；其次，将处理后的图片输入级联检测网络的第一级检测模型YOLOv7-d6算法来识别图像中的绝缘子目标；最后，该级联检测网络中的第二级模型使用YOLOv7x算法，能够检测图像中的绝缘子目标，并对其缺陷类别进行检测。

（2）优化YOLOv7-d6算法模型。本文对第一级网络模型进行了改进，通过引入ECANet通道注意力模块和FReLU激活函数来优化YOLOv7-d6算法模型，旨在提高模型对缺陷信息的关注程度和深度信息获取能力。ECANet在每个基础卷积模块的批归一化层后添加通道注意力模块，并用FReLU替换Leaky Relu作为激活函数。在数据集上训练和测试级联目标检测框架后，第一级模型的mAP@0.5为96.6%，提高了3.76%，召回率recall达到96.7%。

（3）优化YOLOv7x算法模型。为了提高目标检测的精度，在YOLOv7x算法中引入CBAM注意力模块和小目标检测层，并采用DIoU_NMS代替NMS进行结果筛选。这样可以更好地学习特征，提高模型对小目标的检测能力，并减少误检和漏检的情况，从而提高整体检测精度。在对数据集进行训练和测试后，优化后的YOLOv7x算法模型在 mAP@0.5 指标上相较于YOLOv7x改进前提高了10.9%。

实验结果表明，本文提出的基准模型为YOLOv7的级联目标检测网络在绝缘子缺陷检测方面表现优异，能够满足各种缺陷检测需求，同时也能够高精度地检测出绝缘子中的小缺陷。

As an important component for the safe and stable operation of overhead transmission lines, insulator regular inspection of transmission lines is of great significance. In this thesis, aerial images of transmission line insulators are taken as the main research object. Aiming at the problem that insulator defects account for too small proportion in aerial images, a cascade target detection network with a reference model of YOLOv7 is designed to detect various defects of transmission line insulators based on deep learning. The cascade detection network is mainly composed of two levels of target detection network. The first level of target detection network identifies all insulators in the insulator data set, detects all insulator data in the original figure, and inputs it into the second-level network model. The second level network model carries out target detection for insulator defects, detects all kinds of defects of insulators, and provides specific types of defects. The main research contents in this thesis are as follows:

The cascade target detection network of YOLOv7 was designed based on the reference model. Firstly, data enhancement algorithms such as shape, color and contrast are used to preprocess the input image to increase the diversity of data. Secondly, the processed images are input into YOLOv7-d6 algorithm, the first-level detection model of cascade detection network, to identify insulator targets in the images. Finally, the second-level model in the cascade detection network uses the YOLOv7x algorithm to detect the insulator target in the image and detect its defect category.

Optimize the YOLOv7-d6 algorithm model. This thesis improved the first-level network model by introducing ECANet channel attention module and FReLU activation function to optimize the YOLOv7-d6 algorithm model, aiming to improve the model's attention to defect information and the ability to acquire in-depth information. ECANet adds the channel attention module after the batch normalization layer of each underlying convolutional module and replaces Leaky Relu with FReLU as the activation function. After the cascade target detection framework was trained and tested on the data set, the mAP@0.5 of the first-level model was 96.6%, an increase of 3.76%, and the recall rate reached 96.7%.

Optimize the YOLOv7x algorithm model. In order to improve the accuracy of target detection, CBAM attention module and small target detection layer are introduced into YOLOv7x algorithm, and DIoU_NMS is used to replace NMS for result screening. In this way, features can be learned better, the detection ability of the model to small targets can be improved, and the situation of false detection and missing detection can be reduced, so as to improve the overall detection accuracy. After training and testing the data set, the optimized YOLOv7x algorithm model increased by 10.9% in the mAP@0.5 index compared with that before the improvement of YOLOv7x.

The experimental results show that the cascade target detection network of YOLOv7, the reference model proposed in this thesis, has excellent performance in insulator defect detection, which can meet various defect detection requirements, and can also detect small defects in insulators with high precision.

[1]中国电力网. 中电联：预计2023年底并网太阳能发电装机达490GW [Z]. 2023- 01-29

[2]中国电力网. 中电建协发布《中国电力建设行业年度发展报告2022》[Z]. 2022-08-29

[3]刘传洋, 吴一全. 基于深度学习的输电线路视觉检测方法研究进展[J/OL].中国电机工程学报: 1-24 [2023-02-17].

[4]丘灵华, 朱铮涛. 基于深度学习的输电线路绝缘子缺陷检测研究[J]. 计算机应用研究, 2020, 37(S1): 358-360+365.

[5]冯万兴, 范鹏, 姚翔宇, 谷山强, 周自强, 周盛. 基于深度学习的输电线路绝缘子缺陷识别[J]. 水电能源科学, 2021, 39(01): 176-178+50.

[6]王素珍, 许浩, 邵明伟, 邓成禹. 基于改进YOLOv4-Tiny算法的绝缘子缺陷检测[J].国外电子测量技术, 2022, 41(09): 155-162.

[7]杨肖辉, 绳飞, 薛鹏, 谷峰颉, 米新. 基于深度学习的航拍电网绝缘子缺陷检测[J]. 信息技术, 2020, 44(04): 37-40+45.

[8]袁健. 无人机在高压输电线路巡检中的应用[J]. 电子技术, 2022, 51(06): 154-155.

[9]杨蔚, 周辉, 杨生兰等. 输电线路巡检绝缘子跟踪与测距技术[J].电子科技, 2017, 30(05): 124-127.

[10]Qing Yang, Zhong Yang, Tianyi Zhang, Guoxiong Hu. A random chemical reaction optimization algorithm based on dual containers strategy for multi‐rotor UAV path planning in transmission line inspection[J]. Concurrency and Computation: Practice and Experience,2019,31(12).

[11]Zhang Pengyu, Zhang Zhe, Hao Yanpeng, Zhou Zhiheng, Luo Bing, Wang Tingting. Multi-Scale Feature Enhanced Domain Adaptive Object Detection For Power Transmission Line Inspection[J]. IEEE ACCESS,2020,8.

[12]Qi Yue, Mu Shile, Wang Jun, Wang Liangliang. Intelligent Recognition of Transmission Line Inspection Image Based on Deep Learning[J]. Journal of Physics: Conference Series,2021,1757(1).

[13]皮俊, 邹怡. 基于图像处理的输电线路绝缘子缺陷检测[J]. 电气开关, 2020, 58(06): 62-64.

[14]周敏, 吴冰. 输电线路覆冰图像的Canny算子改进算法研究[J].计算机时代, 2019, No.320(02): 54-57.

[15]Huang Xinbo, Zhang Xiaoling, Zhang Ye, Zhao Long. A Method of Identifying Rust Status of Dampers Based on Image Processing[J]. IEEE Transactions on Instrumentation and Measurement,2020,69(8).

[16]尚方, 朱明泽, 刘生等. 基于颜色和距离的绝缘子航拍图像故障检测方法[J]. 黑龙江电力, 2020, 42(02): 100-103.

[17]Chi Zhang, Qing-wu Gong, Ting Wang, Koji Koyamada. Visual extraction system for insulators on power transmission lines from UAV photographs using support vector machine and color models[J]. Journal of Visualization,2020,23(6).

[18]王照, 胡日鹏, 葛馨远等. 基于傅立叶变换的电网线路污闪绝缘子检测研究[J]. 电子技术, 2020, 49(09): 58-61.

[19]In Hyuk Choi, Ja Bin Koo, u Am Son, Jun Sin Yi, Young Geun Yoon, Tae Keun Oh. Development of Equipment and Application of Machine Learning Techniques Using Frequency Response Data for Cap Damage Detection of Porcelain Insulators[J]. Applied Sciences,2020,10(8).

[20]Singh Laxman, Alam Altaf, Kumar K. Vijay,Kumar Devendra, Kumar Parvendra, Jaffery Zainul Abdin. Design of thermal imaging-based health condition monitoring and early fault detection technique for porcelain insulators using Machine learning[J]. Environmental Technology & Innovation,2021,24.

[21]于小彬. 基于图像处理与机器学习的复合绝缘子憎水性等级智能识别[D]. 南昌大学, 2021.

[22]马腾飞, 杨志勇, 谷长春. 基于工业互联网的玻璃绝缘子智能制造系统设计[J]. 物联网技术, 2021, 11(07): 55-58.

[23]张红民, 彭闯, 郭泓邑等. 一种基于SIFT-BOW模型的绝缘子图像识别方法[J]. 激光杂志, 2020, 41(11): 96-99.

[24]王格梓, 张军,李志峰等. 机器学习在瓷绝缘子超声检测中的应用[J]. 自动化应用, 2021(10): 10-14.

[25]杨晓旭, 温招洋. 深度学习在输电线路绝缘子故障检测中的研究与应用[J]. 中国新通信, 2018, 20(10): 208-210.

[26]杨罡, 张娜, 晋涛等. 基于YOLO算法的无人机航拍图片绝缘子缺失检测[J]. 科技资讯, 2018, 16(24): 30-31.

[27]Yanli Yang, Ying Wang, Hongyan Jiao. Insulator identification and self-shattering detection based on mask region with convolutional neural network[J]. Journal of Electronic Imaging,2019,28(5).

[28]Wang Shuqing, Liu Yifan, Qing Yihui, Wang Chenxi, Lan Tianze,Yao Ruotian. Detection of Insulator Defects With Improved ResNeSt and Region Proposal Network[J]. IEEE ACCESS,2020,8.

[29]何宁辉, 王世杰, 刘军福等. 基于深度学习的航拍图像绝缘子缺失检测方法研究[J].电力系统保护与控制, 2021, 49(12): 132-140.

[30]晋成龙, 沈云, 万鑫怡. 基于深度学习的航拍图片中绝缘子串轮廓提取方法[J]. 电子技术与软件工程, 2021, No.216(22): 136-138.

[31]Ma Bin, Fu Yongkang, Wang Chunpeng, Li Jian, Wang Yuli. A high‐performance insulators location scheme based on YOLOv4 deep learning network with GDIoU loss function[J]. IET Image Processing,2022,16(4).

[32]乔路丽, 蔺雨桐, 李静等. 基于深度学习的绝缘子缺失检测方法研究[J]. 电网与清洁能源, 2022, 38(10): 44-50.

[33]Premkumar Borugadda, R. Lakshmi,Surla Govindu. Classification of Cotton Leaf Diseases Using AlexNet and Machine Learning Models[J]. Current Journal of Applied Science and Technology,2021.

[34]Lin Tang, Chenqiang Gao,Xu Chen,Yue Zhao. Pose detection in complex classroom environment based on improved Faster R-CNN[J]. IET Image Processing,2019,13(3).

[35]Kelong Wang, Wei Zhou. Pedestrian and cyclist detection based on deep neural network fast R-CNN[J]. International Journal of Advanced Robotic Systems,2019,16(1).

[36]Pha Samphors, Zheng Hanbo, Sun Yonghui. Research on Detection Method of Insulator Image Based on Improved Faster R-CNN[J]. Journal of Physics: Conference Series,2022,2213(1).

[37]Lingyun Zhu, Chongliu Jia, Juan Zhang, Lu Wang, YueMou Jian. YOLO-C: a new surface defect detection model of steel plate based on YOLO optimisation[J]. International Journal of Wireless and Mobile Computing,2022,23(1).

[38]Zhou Hanbo, Zhao Yiming, Xiang Wei. Method for judging parking status based on yolov2 target detection algorithm[J]. Procedia Computer Science,2022,199.

[39]Liu Chuanyang, Wu Yiquan, Liu Jingjing, Sun Zuo. Improved YOLOv3 Network for Insulator Detection in Aerial Images with Diverse Background Interference[J]. Electronics,2021,10(7).

[40]Yi Li, Jinguo Li, Yongjie Zhai, Ping Meng. Detection of Self-explosive Insulators in Aerial Images Based on Improved YOLOv4[J]. Journal of Physics: Conference Series, 2022, 2320(1).

[41]Ding Jian, Cao Haonan, Ding Xulin, An Chenghui. High Accuracy Real-Time Insulator String Defect Detection Method Based on Improved YOLOv5[J]. Frontiers in Energy Research, 2022.

[42]Norkobil Saydirasulovich Saydirasulov, Abdusalomov Akmalbek, Jamil Muhammad Kafeel, Nasimov Rashid, Kozhamzharova Dinara, Cho YoungIm. A YOLOv6-Based Improved Fire Detection Approach for Smart City Environments[J]. Sensors, 2023.

[43]Everingham M, Eslami S A, Van Gool L, et al.The pascal visual object classes challenge: Aretrospective[J]. International journal of computer vision, 2015, 111(1): 98-136.

[44]王建明, 陈响育, 杨自忠等. 不同数据增强方法对模型识别精度的影响[J]. 计算机科学, 2022, 49(S1): 418-423.

[45]贺丹. 基于PyTorch的图像修复技术研究[J]. 电脑知识与技术, 2022, 18(09): 75-77.

[46]Tian Di, Han Yi, Wang Shu, Chen Xu, Guan Tian. Absolute size IoU loss for the bounding box regression of the object detection[J]. Neurocomputing, 2022, 500.

[47]徐聪. 基于深度学习的航拍图像中绝缘子缺陷检测的研究[D]. 华北电力大学, 2021.

[48]张超. P-R曲线与模型评估问题研究[J]. 现代信息科技, 2020, 4(04): 23-24+27.

[49]Tsung-Yi Lin, Michael Maire, Serge Belongie.Microsoft COCO: Common Objects in Context[J]. Computer Science. 2015, 3.

[50]Yisi Liu, Xiaojun Wang, Lei Wang, Dongliang Liu. A modified leaky ReLU scheme (MLRS) for topology optimization with multiple materials[J]. Applied Mathematics and Computation, 2019, 352.

[51]R. R. S. Ravi kumar, G. Apparao. Extreme Learning Machine with sigmoid activation function on large data[J]. International Journal of Recent Technology and Engineering (IJRTE), 2019, 8(2s11).

[52]王伟, 万晓刚. 结合注意力机制和特征融合的小目标检测方法[J]. 西安工程大学学报, 2022, 36(06): 115-123.

[53]Kim Yu Hwan, soo Jeong Yoo, Yun su Chung, Park Kil Houm. Age Estimation Method based on Comparative Convolutional Neural Network using Inception Module[J]. Journal of Korean Institute of Intelligent Systems, 2018, 28(3).

[54]Cui Zhigao, Wang Nian, Su Yanzhao, Zhang Wei, Lan Yunwei, Li Aihua. ECANet: enhanced context aggregation network for single image dehazing[J]. Signal, Image and Video Processing, 2022, 17(2).

[55]刘志远, 于晓军, 罗美玲, 林泽暄, 郝治国, 张宇博, 杨松浩. 基于CBAM-FCN的高压输电线路发展性故障识别方法[J]. 电网与清洁能源, 2022, 38(09): 25-33+44.

[56]周德政, 赖恵成, 李亚东, 孙晨皓, 汪烈军. 基于YOLOv3目标检测的深层特征增强算法[J]. 激光杂志, 2022, 43(11): 25-30.

[57]刘志远, 于晓军, 罗美玲, 林泽暄, 郝治国, 张宇博, 杨松浩. 基于CBAM-FCN的高压输电线路发展性故障识别方法[J]. 电网与清洁能源, 2022, 38(09): 25-33+44.

[58]周杨, 杨文柱, 申远. 基于深度卷积神经网络的人物检测改进算法[J]. 计算机应用与软件, 2022, 39(07): 215-221.

[59]Assefa Addis Abebe, Tian Wenhong, Acheampong Kingsley Nketia, Aftab Muhammad Umar, Ahmad Muhammad. Small-Scale and Occluded Pedestrian Detection Using Multi Mapping Feature Extraction Function and Modified Soft-NMS.[J]. Computational intelligence and neuroscience, 2022, 2022.

[60]丁楠楠, 胡旭晓, 吴跃成等. 基于深度学习的输电线路外力破坏检测方法研究[J]. 软件工程, 2022,25(1): 14-17.