输电线路是电力系统的组成环节之一，由于长期运行在野外，各金具部件易出现锈蚀、断股、破损、异物等缺陷，这给输电线路的安全运行带来了极大隐患。利用无人机巡检数据，通过卷积神经网络对输电线路多种目标缺陷进行检测与识别，可为输电线路缺陷判别提供决策依据，对保障电力系统安全运行具有重要意义。然而现有机器学习算法在复杂巡检场景下输电线路关键部件识别与缺陷检测的精度仍较低，难以满足高精度巡检的需要。因此，本文利用无人机巡检图像构建了高质量输电线路样本数据集，针对复杂巡检场景下输电线路多类别缺陷，建立了基于改进Faster-RCNN（Faster Region based Convolutional Neural Network）算法的自然巡检场景下输电线路典型部件识别与缺陷检测优化模型。主要研究工作如下：

根据无人机巡检输电线路图像的基本特征，基于线性变换、均值滤波和HSI颜色模型对图像进行对比度增强、去噪处理和图像均衡化预处理。构建包含7类识别目标（均压环倾斜、均压环正常、防震锤锈蚀、防震锤正常、悬垂线夹缺垫片、悬垂线夹正常和鸟窝）的输电线路典型部件样本数据集。借助TensorFlow和LabelImg平台，对数据集进行人工标注，并通过图像增强方法扩增数据集，总计8400张图像。

利用自建输电线路典型部件样本数据集，在VGG16（Visual Geometry Group Network-16）主干特征提取网络下，对比分析SSD（Single Shot MultiBox Detector）和Faster-RCNN算法对输电线路典型部件缺陷识别的效率和不足。表明Faster-RCNN算法对输电线路典型部件缺陷识别的有效性（mAP（mean Average Precision，平均精度均值）为86.57%）明显高于SSD算法（mAP为57.16%）。然而，Faster-RCNN算法对复杂巡检场景下典型部件缺陷识别的准确率明显低于简单巡检场景（平均低30.43%）。

针对Faster-RCNN算法对复杂巡检场景下输电线路典型部件缺陷识别效果不佳的问题，通过优选VGG16网络，采用双线性插值方法，结合交替优化训练方式，将Faster-RCNN算法中的感兴趣区域池化改进为感兴趣区域校准。实验结果表明改进Faster-RCNN算法显著提高了复杂巡检场景下典型部件缺陷的检测精度，平均检测耗时0.307s，mAP为92.47%，可满足自然巡检场景下输电线路典型部件缺陷识别与检测要求。

本文提出的卷积神经网络优化模型对输电线路典型部件识别与缺陷检测具有良好的准确性、实时性和鲁棒性，可获得相比其他算法更好的识别效果，并且能够适应不同巡检场景下的检测与识别任务，可为推进智能电网和保障电力安全提供决策依据。

With the expansion, complexity and geographical span of China’s power grid, the safety and reliable operation of power grid has been widely concerned. As one of the important links of power system, transmission lines are exposed in the field for a long time during operation. Its components are prone to rust, damage, broken strands, foreign bodies and other defects. This brings great hidden dangers to the safe operation of transmission lines. Transmission line use environment includes complex terrain, inconvenient traffic conditions and harsh climate environment. The traditional manual inspection of transmission lines has many disadvantages, such as high labor intensity, low efficiency, low security of inspection personnel and limited inspection results by personnel skills. It is difficult to adapt to the development and safe operation of modern power grid. In recent years, with the development of unmanned air vehicle (UAV) technology and UAV assisted line inspection system, UAV assisted line inspection system with low cost, high efficiency and adaptability to complex environment has completely replaced the traditional inspection method. Through timely processing of UAV inspection images, the basic status of transmission lines can be obtained and equipment defects and hidden troubles can be found. However, the rapid growth of the power grid and large number of applications of helicopters and unmanned aerial patrol lines have resulted in a dramatic increase in the number of aerial images. The contradiction between the image detection of key components of transmission lines and the number of inspection personnel has become increasingly prominent. Therefore, combining UAV inspection data and deep learning technology is helpful to improve the timeliness and automation degree of transmission line defect inspection. This is great significant to ensure the safe operation of the power system. In this study, a sample detection dataset of transmission lines is constructed by using UAV inspection images; the efficiency and insufficiency of SSD algorithm and Faster-RCNN algorithm for multi-target defect recognition of transmission lines are compared and analyzed; and an optimization model for multi-objective defect recognition of transmission lines is established based on an improved Faster-RCNN. The main research contents are as follows:

By augmenting and labeling the UAV inspection images, a sample detection dataset of transmission lines is made for the experiment by using the UAV inspection image provided by a power grid company in Gansu Province, which contains seven categories of targets: normal_ring, error_ring, not_rush, is_rush, have_shim, no_shim and nest. At the same time, the dataset was preprocessed by the contrast enhancement, denoising and equalization using linear transformation, mean filter and enhancement algorithm based on the HSI model, respectively. Finally, the annotated and augmented dataset contains 8400 images.

Using the self-built transmission lines dataset, under the VGG16 backbone feature extraction network, experimental results show that the effectiveness of the Faster-RCNN algorithm for multi-target defect identification of transmission lines (average accuracy rate of 86.57%) is significantly higher than that of SSD algorithm (average accuracy rate of 57.16%). However, the accuracy of Faster-RCNN algorithm for multi-target defect recognition under complex surface background is significantly lower than that of simple surface background.

To improved the Faster-RCNN algorithm, adopted VGG16 as the feature extraction network and selected alternating optimization training method, and the ROI (Region of Interest) Pooling in convolutional neural network was improved to ROI Align regional feature aggregation. Results indicate that the improved Faster-RCNN for target recognition on transmission lines had the mean average Precision (mAP) was 92.47% and the average detection time was 0.307s, which can meet the multi-target defect identification and detection requirements under the background of complex underlying surfaces of transmission lines.

Capitalizing on these study findings, this study proposes a multi-objective defect recognition optimization model for transmission lines under natural conditions based on improved Faster-RCNN. The model has good accuracy, real-time and robustness, and can provide intelligent and effective decision-making basis for the identification of transmission lines defects in actual line inspections.

[1]江秀臣, 盛戈皞. 电力设备状态大数据分析的研究和应用[J]. 高电压技术, 2018, 44(4): 1041-1050.

[2] Hosseini M M, Umunnakwe A, Parvania M, et al. Intelligent damage classification and estimation in power distribution poles using unmanned aerial vehicles and convolutional neural networks[J]. IEEE Transactions on Smart Grid, 2020, 11, 3325-3333.

[3] 赵振兵, 崔雅萍. 基于深度学习的输电线路关键部件视觉检测方法的研究进展[J]. 电力科学与工程, 2018, 34(03):1-6.

[4] 李宁, 郑仟, 谢贵文, 陈炜. 基于无人机图像识别技术的输电线路缺陷检测[J]. 电子设计工程, 2019, 27(10): 102-106+112.

[5] 赵振兵, 齐鸿雨, 聂礼强. 基于深度学习的输电线路视觉检测研究综述[J]. 广东电力, 2019, 32(09): 11-23.

[6] Sui Y, Ning P, Niu P. A review of power inspection technology of mounted unmanned aerial vehicle for overhead transmission lines[J]. Power Grid Technology, 2021, 45(9): 3636-3648.

[7] 郑仟, 李宁. 输电线路无人机巡检智能管理系统的研究与应用[J]. 电子设计工程, 2019, 27(09): 74-77+82.

[8] Yang L, Fan J, Liu Y, et al. A review on state-of-the-art power line inspection techniques[J]. IEEE Transactions on Instrumentation and Measurement, 2020, 69(12): 9350-9365.

[9] Nguyen, V N, Jenssen R, Roverso D. Automatic autonomous vision-based power line inspection: A review of current status and the potential role of deep learning[J]. International Journal of Electrical Power & Energy Systems, 2018, 99, 107-120.

[10]郝睿. 直升机电力巡检作业发展方向的研究[J]. 科技创新与应用, 2016, 25, 186.

[11]李明明, 秦宇翔, 李志学. 无人机在输电线路巡检中的应用及发展前景[J]. 电子制作, 2014, 21, 61.

[12] Li J, Yan D, Luan K, et al. Deep learning-based bird’s nest detection on transmission lines using UAV imagery[J]. Applied Sciences, 2020, 10(18): 6147.

[13]王伟男, 杨朝红. 基于图像处理技术的目标识别方法综述[J]. 电脑与信息技术, 2019, 27(06): 9-15.

[14]Xu D, Wang L, Li F. Research summary of typical target detection algorithms for deep learning[J]. Computer Engineering and Applications, 2021, 57(8):10-215.

[15]钱金菊, 韩正伟, 易琳, 向林, 许志海. 图像处理技术在无人机电力线路巡检中的应用[J]. 电子技术与软件工程, 2017, 15, 72-73.

[16]Tao X, Zhang D, Wang Z, et al. Detection of power line insulator defects using aerial images analyzed with convolutional neural networks[J]. IEEE Transactions on Systems, Man, and Cybernetics: Systems, 2018, 50, 1486-1498.

[17]Krizhevsky A, Sutskever I, Hinton G E. ImageNet classification with deep convolutional neural networks[J]. Communications of theACM, 2017, 60(6): 84-90.

[18]Zhao B, Dai M, Li P, et al. Defect detection method for electric multiple units key components based on deep learning[J]. IEEE Access, 2020, 99, 1-1.

[19]石书山. 基于深度学习技术的输电线路缺陷智能分析系统研究与应用[J]. 通信电源技术, 2020, 37(06):74-75.

[20]方路平, 何杭江, 周国民. 目标检测算法研究综述[J]. 计算机工程与应用, 2018, 54(13): 11-18+33.

[21]段仲静, 李少波, 胡建军, 杨静, 王铮. 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.

[22]Jeong J, Park H, Kwak N. Enhancement of SSD by concatenating feature maps for object detection[J]. arXiv preprint arXiv:1705.09587, 2017.

[23]Ren S, He K, Girshick R, et al. Faster R-CNN: Towards real-time object detection with region proposal networks[J]. IEEE Transactions on Pattern Analysis & Machine Intelligence, 2016, 39(6): 1137-1149.

[24]刘志颖, 缪希仁, 陈静, 江灏. 电力架空线路巡检可见光图像智能处理研究综述[J]. 电网技术, 2020, 44(03): 1057-1069.

[25]谢娟英, 刘然. 基于深度学习的目标检测算法研究进展[J]. 陕西师范大学学报(自然科学版), 2019, 47(05): 1-9.

[26]赵敏. 基于深度学习的输电线路主要缺陷检测研究[D]. 四川: 西南科技大学, 2020.

[27]Zhang X, An J, Chen F. A Method of Insulator Fault Detection from Airborne Images[C]. Intelligent Systems (GCIS), 2010 Second WRI Global Congress on. IEEE, 2010, 2: 200-203.

[28]张涛. 高压输电线路目标缺陷识别算法设计与实现[D]. 扬州: 扬州大学, 2021.

[29]闫继爽. 基于神经网络的输电线路缺陷检测[D]. 大连: 大连理工大学, 2021.

[30]冼世平. 基于无人机自动巡检技术的输电线路缺陷识别方法的研究与应用[D]. 广州: 华南理工大学, 2020.

[31]Yang Z, Xu X, Wang K, et al. Multitarget detection of transmission lines based on DANet and YOLOv4[J]. Scientific Programming, 2021, 1, 1-12.

[32]Ni H, Wang M, Zhao L. An improved Faster R-CNN for defect recognition of key components of transmission line[J]. Mathematical Biosciences and Engineering: MBE, 2021, 18(4): 4679-4695.

[33]熊小萍, 许爽, 蒙登越, 韦香祥, 屠德然, 武文梁. 基于Faster R-CNN的输电线路缺陷识别模型研究[J]. 自动化与仪器仪表, 2020, 3, 1-6.

[34]张喆. 基于深度学习的航拍架空输电线路典型缺陷识别定位研究[D]. 保定: 华北电力大学, 2021.

[35]章毓晋. 图像处理和分析技术(第二版)[M], 北京: 高等教育山版社, 2008.

[36]Al-Amri S S, Kalyankai N V, Khamitkar S D. Linear and non-linear contrast enhancement image[J]. Amri, 2010, 10(2): 139-143.

[37]Bocketein I M. Color equalization method and its application to color image processing[J]. Journal of the Optical Society of America A, 1986, 3(5): 735-737.

[38]Pitas I, Kiniklis P. Multichannel techniques in color image enhancement and modeling[J]. IEEE Transactions on Image Processing, 1996, 5(1): 168-171.

[39]Chien C L, Tseng D C. Color image enhancement with exact Hsi Color Model[J]. International Journal of Innovative Computing Information & Control Ijicic, 2011, 7(12): 6691-6710.

[40]刘迪, 贾金露, 赵玉卿, 钱育蓉. 基于深度学习的图像去噪方法研究综述[J]. 计算机工程与应用, 2021, 57(07): 1-13.

[41]王延翔. 基于均值算法的混合噪声图像滤波算法的研究与实现[D]. 北京: 北京邮电大学, 2010.

[42]Umme S, Morium A, Mohammad S U. Image quality assessment through FSIM, SSIM, MSE and PSNR—A comparative study[J]. Journal of Computer and Communications, 2019, 7, 8-18

[43]Redmon J, Divvala S, Girshick R. You only look once: unified, real-time object detection[C]//IEEE Conference on Computer Vision and Pattern Recognition (CVPR). Boston, MA, USA: IEEE, 2016: 1-9.

[44]Ren S Q, He K M, Girshick R, et al. Faster R-CNN: Towards real-time object detection with region proposal networks[J]. IEEE transactions on pattern analysis and machine intelligence, 2017, 39(6): 1137-1149.

[45]Pang B, Nijkamp E, Wu Y N. Deep learning with TensorFlow: A review[J]. Journal of Educational and Behavioral Statistics, 2020, 45(2): 227-248.

[46]王君洋, 叶传奇, 朱炳旭. 基于SSD神经网络的图像目标物体检测[J]. 无线互联科技, 2021, 18(17): 41-42.

[47]Liu W, Anguelov D, Erhan D, et al. SSD: Single shot multibox detector[C]//Proc of the 14th European Conference on Computer Vision(ECCV 2016), 2016: 21-37.

[48]Huang H, Liang Q, Luo D, et al. Attention-enhanced one-stage algorithm for traffic sign detection and recognition[J]. Journal of Sensors, 2022, 1, 1-8

[49]吴之昊, 熊卫华, 任嘉锋. 基于轻量级SSD的电力设备锈蚀目标检测[J]. 计算机系统应用, 29(2): 262-267.

[50]蔡汉明, 赵振兴, 韩露, 曾祥永. 基于SSD网络模型的多目标检测算法[J]. 机电工程, 2017, 34(06): 685-688.

[51]Zeng Fu, Liu Y, Ye Y, et al. A detection method of Edge Coherent Mode based on improved SSD[J]. Fusion Engineering and Design, 2022, 179(12): 113141

[52]王功鹏, 段萌, 牛常勇. 基于卷积神经网络的随机梯度下降算法[J]. 计算机工程与设计, 2018, 39(02): 441-445+462.

[53]杨鹤标, 龚文彦. 基于卷积神经网络的反向传播算法改进[J]. 计算机工程与设计, 2019, 40(01): 126-130.

[54]Girshick R, Donahue J, Darrell T, et al. Region-based convolutional networhs for accurate obiect detection and segmentation[J]. IEEE transactions on pattern analysis and machine intelligence, 2015, 38(1):142-158.

[55]Girshick R. Fast R-CNN[J]. Computer Science, 2015.

[56]Karen S, Andrew Z. Very deep convolutional networks for large-scale image recognition [C]//San Dieo: International Conference on Learning Representations, 2015, 1-14.

[57]Szegedy C, Liu W, Jia Y, et al. Going deeper with convolutions[C]//Boston: IEEE Conference on Computer Vision and Pattern Recognition, 2015, 1-9.

[58]Zeiler M D, Fergus R.. Visualizing and understanding convolutional networks[M]. In: Fleet, D., Pajdla, T., Schiele, B., Tuytelaars, T. (eds) Computer Vision-ECCV 2014. ECCV 2014. Lecture Notes in Computer Science, vol 8689.

[59]闫建伟, 赵源, 张乐伟, 苏小东, 刘红芸, 张富贵, 樊卫国, 何林. 改进Faster-RCNN自然环境下识别刺梨果实[J]. 农业工程学报, 2019, 35(18): 143-150.

[60]Mi Z, Jiang X, Sun T, et al. GAN-generated image detection with self-attentio mechanism against gan generator defect[J]. IEEE Journal of Selected Topics in Signal Processing , 2020, 14(5): 969-981.

[61]Wang T, Zhang C, Ding R, et al. Mobile phone surface defect detection based on improved faster R-CNN[c]//International Conferenceon Pattern Recognition (ICPR), IEEE, Piscataway, NJ, 2021, 9371-9377.

[62]王森, 杨克俭. 基于双线性插值的图像缩放算法的研究与实现[J]. 自动化技术与应用, 2008, 7, 44-45+35.

[63]刘晓楠, 王正平, 贺云涛, 刘倩. 基于深度学习的小目标检测研究综述[J]. 战术导弹技术, 2019, 1, 100-107.