电工技术学报  2021, Vol. 36 Issue (22): 4837-4848    DOI: 10.19595/j.cnki.1000-6753.tces.201595
高电压与绝缘 |
基于优化YOLOv4的主要电气设备智能检测及调参策略
律方成1,2, 牛雷雷1,2, 王胜辉1, 谢庆2, 王子豪2
1.新能源电力系统国家重点实验室(华北电力大学) 北京 102206;
2.河北省输变电设备安全防御重点实验室(华北电力大学) 保定 071003
Intelligent Detection and Parameter Adjustment Strategy of Major Electrical Equipment Based on Optimized YOLOv4
Lü Fangcheng1,2, Niu Leilei1,2, Wang Shenghui1, Xie Qing2, Wang Zihao2
1. State Key Laboratory of Alternate Electrical Power System with Renewable Energy Sources North China Electric Power University Beijing 102206 China;
2. Hebei Provincial Key Laboratory of Power Transmission Equipment Security Defense North China Electric Power University Baoding 071003 China
全文: PDF (7951 KB)   HTML
输出: BibTeX | EndNote (RIS)      
摘要 基于无人机和巡检机器人搭载的多光谱成像检测是高压设备非接触检测的发展趋势,而主要电气设备的识别是其绝缘状态智能诊断的基础。该文建立了绝缘子、均压环、防振锤、套管和导线训练与测试数据库;基于YOLOv4,改进了Mosaic数据扩充算法,使网络误差降低了0.7,识别准确度提高到84.3%;研究了基于边界框回归的交并比(IoU)算法对不同尺度检测目标的影响,提出了对大、小目标分别采用CIoU和GIoU的训练策略;研究了K-means和分层聚类算法对自建数据库的标注值宽高数据聚类效果及检测结果的影响;基于误差、识别准确度和训练速度,研究并优化了YOLOv4的网络参数,改进后的模型训练误差降低了3%,识别准确度提高了0.8%,较好地实现了主要电气设备的识别。该研究可用于多光谱成像电气设备运行状态的现场诊断。
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
关键词 电气设备识别YOLOv4数据扩充算法交并比K-means和层次聚类    
Abstract:Multi-spectral imaging method based on UAV and robot is the development trend of high voltage equipment inspection, and the identification of main electrical equipment is the basis of intelligent diagnosis of its insulation state. In this paper, the train and test database of insulator, griding ring,vibration damper, transformer bushing and conductor was established. Based on YOLOv4, the Mosaic data augmentation method was optimized , which reduces the network loss by 0.7 and improves the precision to 84.3%. The influence of IoU algorithm based on bounding box regression on detection target of different scales was studied, and the training strategies of CIoU and GIoU were proposed for large and small targets respectivly. The influence of K-means and hierarchical algorithm on the clustering result of label width and high data of self-built database and on network performance is studied. Based on the training loss, precision and speed, the network parameters of YOLOv4 is optimized. The average loss of the improved model was reduced by 3% and the mean average precision increased by 0.8%. This study can be used to assess the operation state of electrical equipment in-site with multi-spectrum image.
Key wordsElectric equipment detection    YOLOv4    data augmente algrithm    intersection ouer union    K-means and hierarchical cluster   
收稿日期: 2020-12-01     
PACS: TM852  
基金资助:国家重点研发计划资助项目(2018YFF01011900)
通讯作者: 牛雷雷 男,1989年生,博士研究生,研究方向为基于多光谱成像的电气设备外绝缘及其智能诊断技术。E-mail:niuleilei@ncepu.edu.cn   
作者简介: 律方成 男,1963年生,教授,博士生导师,研究方向为电气设备绝缘机理和状态检修、电气设备在线监测与故障诊断等。E-mail:lfc@ncepu.edu.cn
引用本文:   
律方成, 牛雷雷, 王胜辉, 谢庆, 王子豪. 基于优化YOLOv4的主要电气设备智能检测及调参策略[J]. 电工技术学报, 2021, 36(22): 4837-4848. Lü Fangcheng, Niu Leilei, Wang Shenghui, Xie Qing, Wang Zihao. Intelligent Detection and Parameter Adjustment Strategy of Major Electrical Equipment Based on Optimized YOLOv4. Transactions of China Electrotechnical Society, 2021, 36(22): 4837-4848.
链接本文:  
https://dgjsxb.ces-transaction.com/CN/10.19595/j.cnki.1000-6753.tces.201595          https://dgjsxb.ces-transaction.com/CN/Y2021/V36/I22/4837