| 基于机器视觉的配网工程安全管控检测方法 |
| |
| 引用本文： | 马静,王庆杰,孟海磊,王栩成,董啸,赵文越,任敬飞. 基于机器视觉的配网工程安全管控检测方法[J]. 现代电力, 2022, 39(6): 685-693. DOI: 10.19725/j.cnki.1007-2322.2021.0330 |
| |
| 作者姓名： | 马静 王庆杰 孟海磊 王栩成 董啸 赵文越 任敬飞 |
| |
| 作者单位： | 1.华北电力大学电气与电子工程学院，北京市昌平区 102206 |
| |
| 基金项目： | 国家电网有限公司总部管理科技项目(5400-202116141A-0-0-00)。 |
| |
| 摘 要： | 针对配电网工程在施工现场受外界环境干扰因素多、现场监管难度大等问题，提出了一种基于改进的YOLOv5网络模型的配电网工程实时检测方法，并对配电网工程图像精确识别及缺陷检测进行了研究。首先，对配电网工程现场样本数据集进行标注，改进YOLOv5网络的特征提取网络，以加快多尺度融合并提高小目标物体检测的精度。在此基础上，改进损失函数、非极大值抑制模块，提高模型的识别精度与收敛速度。最后，经过Darknet深度学习模型对识别样本进行多次迭代训练，保存最优权重数据用于测试集的测试。算法通过 TensorBoard 可视化工具显示训练和测试结果。测试结果表明，每种配电网样本的平均识别准确率可达到95%以上，图片的识别速度可达到140 帧/s。同时，所改进算法检测准确率高，实时性强，满足工程现场实时使用需求。
|
| 关 键 词： | 配电网工程 YOLOv5 损失函数 目标检测 |
| 收稿时间： | 2021-11-22 |
A Machine Vision-Based Detection Method for Security Control of Distribution Network Engineering |
| |
| Affiliation: | 1.School of Electrical and Electronic Engineering, North China Electric Power University, Changping District, Beijing 102206, China2.State Grid Corporation of China, Xicheng District, Beijing 100031, China3.State Grid Shandong Electric Power Company, Jinan 250001, Shandong Province, China |
| |
| Abstract: | In view of such troubles as too much interference factors of external environment at the construction site and the worksite supervision difficulty and so on, an improved YOLOv5 network model-based realtime detection method for distribution network engineering was proposed, and the accurate image recognition as well as the defect detection of distribution network engineering were researched. Firstly, the on-site sample data set of distribution network engineering was labeled, and the feature extraction network for YOLOv5 network was improved to speed up the multi-scale fusion and to raise the detection accuracy of small target object. On this basis, the loss function and the non-maximum suppression module were improved to raise the recognition precision of the model and to accelerate the convergence speed. Secondly, by means of Darknet deep learning model the multiple iteration training was performed to the recognition samples and the optimal weight data was saved the for test set testing. Finally, by use of TensorBoard visual tool the training and test results could be displayed. Testing results show that the average recognition accuracy of each distribution network sample can reach more than 95%, and the speed of picture recognition can reach 140 fps. Meanwhile, The improved method possesses the advantages as high detection accuracy and strong real-time performance, so it can meet the need of on-site realtime use. |
| |
| Keywords: | |
|
| 点击此处可从《现代电力》浏览原始摘要信息 |
|
点击此处可从《现代电力》下载免费的PDF全文 |
|
