电工技术学报  2022, Vol. 37 Issue (9): 2241-2251    DOI: 10.19595/j.cnki.1000-6753.tces.210956
“电力装备可靠性与智能化”专题(特约主编:李奎教授 牛峰教授) |
基于改进EfficientDet的电网元件及缺陷识别方法
宋立业1, 刘帅1, 王凯2, 杨金丹1
1.辽宁工程技术大学电气与控制工程学院 葫芦岛 125000;
2.国网辽宁省电力有限公司葫芦岛供电公司 葫芦岛 125000
Identification Method of Power Grid Components and Defects Based on Improved EfficientDet
Song Liye1, Liu Shuai1, Wang Kai2, Yang Jindan1
1. College of Electrical and Control Engineering Liaoning Technical University Huludao 125000 China;
2. Huludao Power Supply Company of State Grid Liaoning Electric Power Company Limited Huludao 125000 China
全文: PDF (19998 KB)   HTML
输出: BibTeX | EndNote (RIS)      
摘要 针对现有的无人机电力巡检中的目标检测算法小目标识别精度低、检测的元件及缺陷类型较为单一、检测速度和精度无法同时满足的问题,提出一种改进的EfficientDet目标检测算法,该算法应用于无人机电力巡检图像的数据挖掘,对高压输电线路上的绝缘子、防震锤、均压环、屏蔽环、鸟巢同时进行目标检测及缺陷定位。首先通过Imgaug数据增强库对现有的1 468张国家电网某检修公司标准化无人机巡检数据集进行数据增强;然后在加强特征提取网络双向特征金字塔网络(BiFPN)特征融合时融入小一级尺度的特征层,提高了小目标检测能力,对主干特征提取网络EfficientNet的倒残差模块进行改进,引入坐标注意力机制(CA)提高了主干特征提取效率;最后进行对比训练实验,改进EfficientDet算法在元件检测及缺陷定位测试集上平均均值精度达到90.2%,较原始EfficientDet算法提高8.6%,亦优于其他先进目标检测算法,同时元件检测速率达到23.4f/s,缺陷定位达到17.2f/s,证明了该文方法可以满足电力巡检中准确性和快速性的要求。
服务
把本文推荐给朋友
加入我的书架
加入引用管理器
E-mail Alert
RSS
作者相关文章
宋立业
刘帅
王凯
杨金丹
关键词 电力巡检EfficientDet元件检测缺陷定位数据增强    
Abstract:An improved EfficientDet target detection algorithm was proposed to deal with the problem that the existing UAV electric power inspection target detection algorithm could easily produce low accuracy of small target, the detected components and defect types are relatively single, and could not satisfy with the detection speed and accuracy simultaneously. The algorithm was applied for data mining of power inspection images to detect insulators, dampers, grading-ring, shielding-ring and bird-house on the high voltage transmission lines and locate the corresponding defects at the same time. Firstly, this paper uses Imgaug data enhancement library to enhance the data of the existing 1 468 standardized UAV inspection datasets in the State Grid. Then, to improve the small target detection ability of the bidirectional feature pyramid networks and efficiency of the backbone network EfficientNet, a feature layer with a smaller scale was integrated weighted, the backdown residual module was improved and coordinate attention mechanism was introduced. Finally, a comparative training experiment is carried out, and the improved EfficientDet has the accuracy reaching up to 90.2% on the test set, which is 8.6% higher than the original EfficientDet and other advanced target detection algorithm. Meanwhile, the frame rate per second of component inspection and defect location reaches 23.4 and 17.2 respectively. It is proved that this method can meet the requirements of accuracy and efficiency in power inspection.
Key wordsPower inspection    EfficientDet    component detection    defect location    data enhancement   
收稿日期: 2021-06-26     
PACS: TM755  
基金资助:辽宁省教育厅科学技术研究创新团队项目(LT2019007)、辽宁省重点研发计划指导计划项目(2019JH8/10100050)和2019年辽宁省高等学校国(境)外培养项目(2019GJWZD002)资助
通讯作者: 刘 帅 男,1996年生,硕士研究生,研究方向为机器视觉与深度学习在电网巡检中的应用。E-mail:751411998@qq.com   
作者简介: 宋立业 男,1972年生,副教授,硕士生导师,研究方向为智能电网新技术,电力系统数字化监控技术。E-mail:372492761@qq.com
引用本文:   
宋立业, 刘帅, 王凯, 杨金丹. 基于改进EfficientDet的电网元件及缺陷识别方法[J]. 电工技术学报, 2022, 37(9): 2241-2251. Song Liye, Liu Shuai, Wang Kai, Yang Jindan. Identification Method of Power Grid Components and Defects Based on Improved EfficientDet. Transactions of China Electrotechnical Society, 2022, 37(9): 2241-2251.
链接本文:  
https://dgjsxb.ces-transaction.com/CN/10.19595/j.cnki.1000-6753.tces.210956          https://dgjsxb.ces-transaction.com/CN/Y2022/V37/I9/2241