利用无人机对高压输电线路巡检并基于计算机视觉技术对巡检数据中的故障目标进行自动、准确检测是输电线路巡检领域中的重要研究方向,同时也是一个极具挑战性的课题。因此，本课题探索了基于深度学习的输电线路多目标故障检测方法，该研究对提高输电线路效率及实现智能巡检具有重要意义。论文主要工作包括：
（1）针对复杂巡检环境中待检测目标存在多尺度特性以及部分遮挡造成传统算法难以准确检测问题，提出一种基于注意力机制与跨尺度特征融合的YOLOv5 检测算法。首先，搭建了YOLOv5 检测网络，为了抑制复杂背景干扰，在其基础上引入空间与通道卷积注意力模型，以增强待检测故障目标的显著度；然后，将原始YOLOv5 检测框架Neck 中的FPN+PAN 结构改为BiFPN 结构，从而使目标多尺度特征能够有效融合；其次，为了解决待检测目标特征表达能力不足而造成漏检和误检的问题，设计了多尺度与同尺度特征的自适应加权融合模块，以增强检测网络对被遮挡情况下故障目标的检测能力；最后，为了验证所提出算法的有效性，使用某巡检部门近4 年进行无人机巡检得到的数据来对算法进行验证。实验结果表明，所提算法能够对复杂环境中输电线路上的多尺度故障目标实现精确检测，其平均检测精度相比原网络的目标检测精度mAP 值提升了0.9%，目标检测精度mAP 值可达96.8%。
（2）针对复杂巡检环境下，输电线路小目标存在于多目标中及被遮挡造成难以检测，并同时实现模型的轻量化的问题，提出一种基于TransFormer 的轻量化YOLOv5 输电线路小目标检测算法。首先，搭建了YOLOv5 基础检测网络，在Backbone 中采用MobileNetV3 轻量化网络减少主干网络参数量，从而减少检测模型大小与检测时间；然后，在其Backbone 上引入一种更有效捕获位置信息和通道关系的注意力机制CA 模型，来抑制复杂背景干扰，增强待检测故障目标的显著度；最后，将YOLOv5 的Neck 末端中CSPnet 块替换为Transformer 块，形成一个Transformer 预测头（TPH），使模型可以在低分辨率特征图上降低较高的计算和存储成本。将所提算法与原YOLOv5 网络进行对比实验，实验证明所提算法能够对复杂环境中输电线路上的多尺度中故障小目标实现精确检测，其平均检测精度相比原网络的目标检测精度mAP 值提升了1%，目标检测精度mAP 值可达96.9%。

Automatic and accurate detection of fault targets in transmission line inspection data based on computer vision technology using UAVs for high-voltage transmission line inspection is an important research direction in the field of transmission line inspection, and also a challenging topic. Therefore, this topic explores the multi-target fault detection method for transmission lines based on deep learning, which is important for improving the efficiency of
transmission lines and realizing intelligent inspection. The main work of the paper includes:
(1)For the complex inspection environment in which the target to be detected has multiscale characteristics as well as partial occlusion that makes it difficult for traditional algorithms
to detect accurately, a YOLOv5 detection algorithm based on attention mechanism and crossscale feature fusion is proposed. Firstly, the YOLOv5 detection network is built, and in order to
suppress the complex background interference, the spatial and channel convolutional attention model is introduced on its basis to enhance the saliency of the faulty target to be detected; then,
the FPN+PAN structure in the original YOLOv5 detection framework Neck is changed to BiFPN structure, so that the target multi-scale features can be effectively fused; secondly, in order to solve the problem that the multi-scale features of the target to be Secondly, in order to solve the problem of missing and false detection caused by insufficient feature expression capability of the target to be detected, an adaptive weighted fusion module of multi-scale and same-scale features is designed to enhance the detection capability of the detection network for faulty targets under obscured conditions; finally, in order to verify the effectiveness of the proposed algorithm, the data obtained from the UAV inspection conducted by an inspection department in the past four years are used to validate the algorithm. The experimental results show that the proposed algorithm can achieve accurate detection of multi-scale fault targets on
transmission lines in complex environments, and its average detection accuracy improves by 0.9% compared with the target detection accuracy mAP value of the original network, and the target detection accuracy mAP value can reach 96.8%.
(2)For the complex inspection environment, transmission line small targets exist in multiple targets and are difficult to detect due to occlusion, and at the same time to realize the problem of lightweight model, a lightweight YOLOv5 transmission line small target detection algorithm based on TransFormer is proposed. First, the YOLOv5 base detection network is built, and the MobileNetV3 lightweight network is used in the Backbone to reduce the number of backbone network parameters, thus reducing the detection model size and detection time;
then, an attention mechanism CA model is introduced on its Backbone to capture location information and channel relationships more effectively to suppress complex background
interference and enhance the saliency of the target to be Finally, the CSPnet block in the Neck end of YOLOv5 is replaced with a Transformer block to form a Transformer Prediction Head (TPH), which allows the model to reduce higher computation and storage costs on lowresolution feature maps.The proposed algorithm is compared with the original YOLOv5 network for experiments, and it is demonstrated that the proposed algorithm can achieve accurate detection of small targets of faults in multiple scales on transmission lines in complex environments, and its average detection accuracy improves by 1% compared with the target
detection accuracy mAP value of the original network, and the target detection accuracy mAP value can reach 96.9%.

[1] 余贻鑫.智能电网实施的紧迫性和长期性[J].电力系统保护制,2019,47(17):1-5.

[2] 徐云鹏,毛强,李庭坚.输电线路机巡与人巡效果对比及协同巡检建议[J].南方电网技术,2016,10(02):44-47.

[3] 彭向阳,易琳,钱金菊,等.大型无人直升机电力线路巡检系统实用化[J].高电压技

术,2020,46(02):384-396.

[4] 黄山,吴振升,任志刚,等.电力智能巡检机器人研究综述[J].电测与仪表,2020,57(02):26-

38.

[5] 江秀臣,刘亚东,傅晓飞,等.输配电设备泛在电力物联网建设思路与发展趋势[J].高电

压技术,2019,45(05):1345-1351.

[6] “十四五”电力规划工作启动[J].变压器,2020,57(02):55.

[7] Tang Z,Abera G,Kishore M,et al.An applicable time reversal analysis of guided waves in

al-plate with ultrasonic immersion testing[J].Journal-Korean Physical Society,2019,

74(4):340-348.

[8] Colwell J E, Hansen, et al. Investigation of diurnal variability of water vapor in

Enceladus'plume by the Cassini ultraviolet imaging spectrograph[J]. Geophysical

Research Letters，2017，44(2):672-677.

[9] Zhao S N, SongY, Zhao Y F. Infrared target detection method based on the receptive field

and lateral inhibition of human visual system[J].Applied Optics, 2017,56(30):8555-8563.

[10] 陈豫宁,周路焱,崔文亮,等.以现代设备管理体系为依托的辖区外特高压跨区输电线路

管理经验探究[J].中国设备工程,2022(03):12-13.

[11] 翟瑞聪,张峰,许志海.架空输电线路机巡光电吊舱系统巡检数据智能整理技术研究[J].

科学技术创新,2017(20):19-20.

[12] 贾晓芬,吴雪茹,赵佰亭.绝缘子自爆缺陷的轻量化检测网络DE-YOLO[J/OL].电子测

量与仪器报:1-8.

[13] 廖圣龙,安居白.输电线路对绝缘子破损航拍图像检测仿真[J].计算机仿真,2016,33(04):

176-179+291.

[14] 翟永杰,王迪,赵振兵,程海燕.基于空域形态一致性特征的绝缘子串定位方法[J].中国

电机工程学报,2017,37(05):1568-1578.

[15] 方挺,韩家明.航拍图像中绝缘子缺陷的检测与定位[J].计算机科学,2016,43(S1):222-

225.

[16] 姚晓通,刘力,李致远.基于Canny 边缘特征点的接触网绝缘子识别方法[J].电瓷避雷

器,2020,21(01):142-148.

[17] 韩正新,乔耀华,孙阳,等.基于图像识别的无人机输电线路绝缘子故障检测方法研究

[J].现代电子技术,2017,40(22):179-181+186.

[18] 赵振兵,徐磊,戚银城,蔡银萍.基于Hough 检测和C-V 模型的航拍绝缘子自动协同分

割方法[J].仪器仪表学报,2016,37(02):395-403.

[19] 张晶晶,韩军,赵亚博等.形状感知的绝缘子识别与缺陷诊断[J].中国图象图形

学,2014,19(8):1194-1201.

[20] Zhai Y,Chen R,Yang Q,et al.Insulator ault detection based on spatial morphological

features of aerial images[J].IEEE Access,2018,6: 35316-35326.

[21] 周飞燕,金林鹏,董军.卷积神经网络研究综述[J].计算机学报,2017,40(06):1229-1251.

[22] 赵永强, 饶元, 董世鹏, 等. 深度学习目标检测方法综述[J]. 中国图象图形学

报,2020,25(04):629-654.

[23] 段仲静,李少波,胡建军,等.深度学习目标检测方法及其主流框架综述[J].激光与光电

子学进展,2020,57(12):59-74.

[24] GIRSHICK R, DONAHUE J, DARRELL T, et al. Richfeature hierarchies for accurate

object detection and semantic segmentation[C]//Proceedings of 2014 IEEE Conference on

Computer Vision and Pattern Recognition (CVPR). Columbus, OH, USA, 2014: 580–587.

[25] GIRSHICK R. Fast R-CNN[C]//Proceedings of 2015 IEEE International Conference on

Computer Vision (ICCV).Santiago,Chile,2015:1440–1448.

[26] 侯春萍,章衡光,张巍,等.输电线路绝缘子自爆缺陷识别方法[J].电力系统及其自动化

学报,2019,31(06):1-6.

[27] 刘思言,王博,高昆仑,等.基于R-FCN 的航拍巡检图像目标检测方法[J].电力系统自动

化,2019,43(13):162-168.

[28] 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.

[29] 白洁音，赵瑞，谷丰强，等．多目标检测和故障识别图像处理方法[J]．高电压技

术，2019，45(11)：3504-3511．

[30] 林刚,王波,彭辉,等.基于改进Faster-RCNN 的输电线巡检图像多目标检测及定位[J].

电力自动化设备,2019,39(05):213-218.

[31] Redmon J, Divvala S, Girshick R, et al. You onlylook once: unified, real.time object

detection[C]//Proceedings of 2016 IEEE Conference on Computer Vision and Pattern

Recognition (CVPR). Las Vegas, NV, USA, 2016:779-788.

[32] Liu Wei, ANGUELOV D, ERHAN D, et al. SSD: singleshot MultiBox detector[C]//Proce

edings of the 14thEuropean Conference on Computer Vision. Amsterdam,the Netherlands,

2016: 21–37.

[33] 颜宏文,陈金鑫.基于改进YOLOv3 的绝缘子串定位与状态识别方法[J].高电压技

术,2020,46(02):423-432.

[34] 王卓,王玉静,王庆岩,等,V.I.Mikulovich.基于协同深度学习的二阶段绝缘子故障检测

方法[J].电工技术学报,2021,36(17):3594-3604.

[35] 郝帅,马瑞泽,赵新生,等.基于卷积块注意模型的YOLOv3 输电线路故障检测方法

[J]．电网技术,2021,45(08):2979-2987.

[36] 徐文静，高云天，陈晨，等．基于YOLOv5 的绝缘子图像自动标注[J]．科学技术

创新，2021(14)：15-17．

[37] Shankar K,Lakshmanaprabu S K,D Gupta,et al.Optimal feature-based multi-kernel SVM a

pproach for thyroid disease classification[J].The Journal of Supercomputing,2020,76(28):

1-16.

[38] Ssr A, Skv B, Yk C. Review on recent development in infrared small target detection

algorithms[J]. Procedia Computer Science, 2020, 167(05):2496-2505.

[39] Redmon J, Farhadi A. YOLO9000: Better, Faster, Stronger[J]. IEEE, 2017:6517-6525.

[40] Redmon J, Farhadi A. YOLOv3:An incremental improvement [EB/OL]. [2018-09-

30].https://arxiv.org/abs/1804.02767.

[41] Jocher G, Stoken A, Borovec J.(2020,August13).ultralytics/yolov5:v3.0(Versionv3.0).Zen

odo.http://doi.org/10.5281/zenodo.3983579.

[42] Bochkovskiy A,Wang C Y,Liao H Y.YOLOv4:optimal speed and accuracy of object detect

ion[J/OL].arXiv:2004.109342020[cs.CV].(2020-04-23).https://arxiv.org/abs/2004.10934.

[43] Wang C Y，Liao H Y M，Wu Y H，et al．CSPNet：A new backbone that can enhance

learning capability of CNN[C]//2020 IEEE/CVF Conference on Computer Vision and

Pattern Recognition Workshops(CVPRW)．Seattle，USA：IEEE，2020：1571-1580．

[44] Goodfellow I J,Pouget-Abadie J,Mirza M,et al.Generative Adversarial

Networks[J].Advances in Neural Information Processing Systems,2014,3:2672-2680.

[45] Shaham T R, Dekel T, Michaeli T. SinGAN:Learning a Generative Model from a Single

Natural Image[J].2019.

[46] Hinz T,Fisher M,Wang O,et al.Improved techniques for training single-image

GANs[C]//Proceedings of the IEEE/CVF Winter Conference on Applications of Computer

Vision.2021:1300-1309.

[47] He K,Zhang X,Ren S,et al.Spatial pyramid pooling in deep convolutional networks for

visual recognition[J].IEEE Transactions on Pattern Analysis and Machine

Intelligence,2015,37(9):1904-1916.

[48] Lin T Y,Dollar P,Girshick R,et al.Feature pyramid networks for object detection[C]//2017

IEEE Conference on Computer Vision and Pattern Recognition(CVPR).Honolulu,HI,USA:

IEEE,2017:936-944.

[49] Liu S,Qi L,Qin H,et al.Path aggregation network for instance segmentation[C]//IEEE Con

ference on Computer Vision and Pattern Recognition(CVPR) .Salt LakeCity:IEEE,2017:8

759-8768.

[50] Hu J, Shen L, Sun G. Squeeze.and.excitation networks[C]//Proceedings of the IEEE Conf

erence on Computer Vision and Pattern Recognition(CVPR), 2018: 7132-7141.

[51] Woo S,Park J,Lee J Y, et al. Cbam:Convolutional block attention module[C]//Proceedings

of the European Conference on Computer Vision(ECCV), 2018: 3-19.

[52] Tan M,Pang R,Le Q V.EfficientDet:scalable and efficient object detection[C]//IEEE Confe

rence on Computer Vision and Pattern Recognition(CVPR),New York:IEEE,2020:10778-

10787.

[53]Liu S,Huang D,Wang Y.Learning spatial fusion for single-shot object detection[J].arXiv

preprint arXiv:1911.09516,2019.

[54]Howard A,Sandler M,Chu G,et al.Searching for MobileNetV3[C]//Proc-eedings of the

IEEE/CVF International Conference on Computer Vision.2019:1314-1324.

[55]Zhang R, Zhu F, Liu J, et al. Depth-wise separable convolutions and multi-level pooling

for an efficient spatial CNN-based steganalysis[J]. IEEE Transactions on Information

Forensics and Security, 2019, 50(15): 1138-1150.

[56]Dosovitskiy A,Beyer L,Kolesnikov A,et al.An image is worth 16x16 words:Transformers

for image recognition at scale[J].arXiv preprint arXiv:2010.11929,2020.

[57]Hou Q, Zhou D, Feng J. Coordinate attention for efficient mobile network

design[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern

Recognition. 2021: 13713-13722.

[58]Howard A G,Zhu M,Chen B,et al.Searching for MobileNets:Efficient convolutional neural

networks for mobile vision applications[J].arXiv:1704.04861,2017.

[59]Howard A, Zhmoginov A,Chen L C,et al.Inverted residuals and linea-r bottlenecks: Mobile

networks for classificationdetection and segmentation[J].2018.

[60]He J, Erfani S, Ma X, et al. Alpha -IoU: A Family of Power Intersection over Union Losses

for Bounding Box Regression[J]. Advances in Neural Information Processing Systems,

2021,34.

[61]Shi H, Yang N, Tang H, et al. ASGD: Stochastic Gradient Descent with Adaptive Batch

Size for Every Parameter[J]. Mathematics, 2022, 10(6): 863.