基于轻量化YOLOv4的绝缘子故障检测  

Insulator Fault Detection Based on Light Weight YOLOv4

作　　者：何王金 王昭雷 张旭 侯亚欣 He Wangjin

机构地区：[1]深圳市浩瑞泰科技有限公司,广东深圳518063 [2]国网河北省电力有限公司超高压分公司,河北石家庄050070

出　　处：《工业控制计算机》2025年第4期41-43,共3页Industrial Control Computer

摘　　要：基于深度学习的绝缘子定位模型存在层次深、参数多、训练时对硬件设备要求高的问题,因此利用轻量型网络MobileNet对YOLOv4轻量化,以提高网络的检测速度;将空洞卷积引入SPP结构中,扩大了感受野,降低了多目标绝缘子漏检的概率;轻量化后YOLOv4网络会存在一定的精度损失,因此对其特征金字塔部分增加反馈机制,使其可以反复地提取特征;在减少参数量、易于训练的同时,兼顾了对绝缘子及其破损的检测能力,保证了网络检测的精确度。通过与YOLOv4、SSD算法进行对比,所采用的模型参数量降低了72.26%,mAP值提高了6.2%,帧速度达到31.1帧/s,提高了3.75帧/s。该算法具有对硬件设备要求低、通用性强的特点,可以满足绝缘子破损定位的实时性要求。The insulator positioning model based on deep learning has multiple hierarchies,parameters,and high hardware requirement,so this paper uses a lightweight network-MobileNet to light the YOLOv4 networks,and improves the detection speed of the network,put the empty convolution into the SPP structure,thereby expand the perceptual field,reducing the probability of multi-target insulator leaks.Lightweighted YOLOv4 networks will have accuracy loss,so the feedback mechanism is added to the feature pyramid section,so that it can be repeated extraction characteristics.Then the network has less parameters and better accuracy.Compared the algorithm used in this paper with the YOLOv4,SSD algorithm,the model parameters employed are reduced by 72.26%,while the mAP value has increased by 6.2%,the frame speed reaches 31.1 frame/s,and it has increased by 3.75 frames/s.The algorithm used in this paper has a lower hardware requirement,and more versatility.It can meet the requirements for real-time-positioning of damaged insulators.

关 键 词：绝缘子破损检测 YOLOv4 轻量化网络 空洞卷积 

分 类 号：TM216[一般工业技术—材料科学与工程]