联合剪枝和知识蒸馏的PCSA-YOLOs防振锤缺陷检测算法  

PCSA-YOLOs vibration damper defect detection algorithm combined with pruning and knowledge distillation

作　　者：张可 乐全明 黄文礼 黄承沛 任仰勋 ZHANG Ke;YUE Quan-ming;HUANG Wen-li;HUANG Cheng-pei;REN Yang-xun(State Grid Electric Power Research Institute,Nanjing 211000,China;State Grid Zhejiang Electric Power Corporation Hangzhou Power Supply Company,Hangzhou 310016,China;Electronic Information Engineering,Anhui University,Hefei 230601,China;University of Science and Technology of China,Hefei 230026,China)

机构地区：[1]国网电力科学研究院有限公司,江苏南京211000 [2]国网浙江省电力有限公司杭州供电公司,浙江杭州310016 [3]安徽大学电子信息工程学院,安徽合肥230601 [4]中国科学技术大学,安徽合肥230026

出　　处：《电工电能新技术》2022年第11期59-69,共11页Advanced Technology of Electrical Engineering and Energy

基　　金：国家电网公司总部科技项目(5500-202140127A-0-0-00、5500-202219275A-2-0-XG)。

摘　　要：准确检测输电线路中防振锤缺陷并及时更换部件是保证电力系统安全运行的重要措施。本文以包含FD、FDZ、FFH、FR四种常见型号防振锤的无人机输电线路图像为研究对象,针对复杂背景下的防振锤形态多样且目标较小的特点,设计了一种端到端的并联型通道-空间注意力模块(PCSA),PCSA-YOLO网络用于锈蚀、缺损和正常等3种防振锤的检测。在YOLOv4基础上,加入改进后的PCSA,关注复杂背景下防振锤小目标的关键区域,提高防振锤检测的精度。同时联合剪枝和知识蒸馏方法,对每个卷积层后正则化层的缩放因子施加L1正则化,然后根据稀疏后缩放因子的大小设定剪枝率来裁剪低于阈值的通道,达到压缩网络参数量的目的,并采用知识蒸馏策略以弥补网络剪枝造成的准确率下降,最终得到轻量型的防振锤检测网络模型PCSA-YOLOs。实验结果表明,在所构建的无人机输电线路数据集中,PCSA-YOLO的mAP@0.5可达94.0%,相较于YOLOv4提高了3.6%;轻量型PCSA-YOLOs的mAP@0.5可达92.7%,模型参数量为0.8 M,比YOLOv5s的mAP@0.5高出6%左右,模型参数量减少6 M左右,能满足智能巡检实时性的要求。It is important for the safe operation of power system to accurately detect the vibration damper defects and timely replace them in transmission lines.In this paper,the transmission line images of UAV including four common types of damper,FD,FDZ,FFH and FR,are taken as the research object.Aiming at the characteristics of various shapes and small targets of vibration dampers under complex background,an end-to-end PCSA-YOLO network is designed to detect corrosion,defect and normal vibration dampers.On the basis of YOLOv4,an improved parallel channel-space attention module PCSA is added to pay attention to the key areas of small targets of vibration damper in complex background,so as to improve the detection accuracy of vibration damper.Combining pruning and knowledge distillation,L1 regularization is applied to the scaling factor of BN layer after each convolution layer.According to the size of the scaling factor after thinning,the pruning rate is set to clip the channels below the threshold,so as to achieve the purpose of compressing the network parameters.Knowledge distillation strategy is used to compensate for the decrease of accuracy caused by network pruning,and finally a lightweight vibration damper detection network model PCSA-YOLOs is obtained.The experimental results show that the mAP@0.5 of PCSA-YOLO can reach 94.0%in the constructed UAV transmission line data set,which is 3.6%higher than YOLOv4.The mAP@0.5 of lightweight PCSA-YOLOs can reach 92.7%,and the model parameters are 0.8 M,which is about 6%higher than that of YOLOv5s.The model parameters are reduced by about 6 M,which can meet the real-time requirements of intelligent inspection.

关 键 词：防振锤 YOLOv4 缺陷检测 注意力机制 模型压缩 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]