摘要
作为输电线路中的重要金具部件,防振锤的缺陷将对输电线路构成严重威胁。针对由于防振锤缺陷样本数量稀少、背景复杂、区域形状尺寸不一造成的防振锤缺陷识别能力不足的问题,提出一种基于多尺度卷积注意力机制的防振锤缺陷检测方法。首先,通过统计不同缺陷的防振锤尺寸,设计适应不同类别的多尺度卷积注意力机制,使网络重点关注图像中的防振锤区域;其次,引入结构重参数化方法,以将网络中的多分支结构无损失地转换为单分支结构,在提高网络检测性能的同时维持检测速度在较高水平;最后,以渐进式特征金字塔网络结构(AFPN)为基础,融合更多的浅层网络,提高了网络检测防振锤小目标的能力。实际收集的防振锤缺陷数据集实验结果表明,设计的检测方法可显著提升防振锤缺陷检测的性能,检测精度mAP0.5达到了91.9%,在TITAN XP平台下检测速度达60.88帧/s,可为输电线路防振锤智能化巡检提供参考。
The presence of defective dampers in power transmission lines poses a significant risk to the secure and stable operation of the electrical grid.Advancing the intelligent development of damper inspections in transmission lines,a fast and accurate defect detection method holds paramount importance.Addressing the issue of insufficient damper defect recognition due to scarce defect samples,complex backgrounds,and varying regional dimensions,a novel damper defect detection network(RCA-YOLOv8)based on a multi-scale convolution attention mechanism was proposed.Firstly,the diverse sizes of dampers in images are analyzed and a multi-scale convolution attention mechanism composed of three sets of bar-shaped convolutions is constructed to precisely capture features of different-sized dampers.Subsequently,a structural reparameterization method is utilized to convert the multi-branch structure in the network into a single-branch structure,enabling to maintain consistent inference speed with the single-branch structure while benefiting from the detection performance improvement brought by the multi-branch structure.In addition,based on the YOLOv8 feature extraction structure,a Conv Block structure containing Conv-A structure and structural reparameterization method was constructed to propose multi-scale features of dampers.Moreover,more shallow network features are integrated by using the AFPNs structure,resolving feature conflicts between large,medium,and small targets in the images,thereby enabling accurate detection of small damper targets and further enhancing detection performance.In this model,Conv-A is more able to focus on the dampers area in the image,reducing background interference,and structural reparameterization greatly reduces computational costs.AFPNs solve the problem of feature conflicts between large and small dampers in the image,thus achieving a low computational cost and high detection accuracy model.For model experimentation,a dataset of damper defects in power transmission lines within substation scene
作者
张烨
李博涛
尚景浩
黄新波
翟鹏超
Zhang Ye;Li Botao;Shang Jinghao;Huang Xinbo;Zhai Pengchao(School of Electronics and Information,Xi’an Polytechnic University,Xi’an,710048,China;Xi'an Microelectronics Technology Institute,Xi’an,710000,China)
出处
《电工技术学报》
EI
CSCD
北大核心
2024年第11期3522-3537,共16页
Transactions of China Electrotechnical Society
基金
国家自然科学基金(52307182)
西安市科技计划(22GXFW0038)
陕西省科学技术协会青年人才托举计划(20220133)
金属成形技术与重型装备全国重点实验室开放课题(S2208100.W03)
西安工程大学博士科研启动基金(BS202125)资助项目。
关键词
防振锤
深度学习
注意力机制
实时缺陷检测
Damper
deep learning
attention mechanism
real-time defect detection
