融合SE-Attention图像识别模型的电力设备监测系统  

A power equipment monitoring system based on SE-Attention image recognition model

作　　者：何启远 潘德泰 李贵亮 林清 董芸州 李小敏 高振宇 HE Qiyuan;PAN Detai;LI Guiliang;LIN Qing;DONG Yunzhou;LI Xiaomin;GAO Zhenyu(Hainan Power Grid Co Ltd Information and Communication Branch,Haikou 570203,China;School of Mechanical Engineering,Sichuan University,Chengdu 610065,China)

机构地区：[1]海南电网有限责任公司信息通信分公司,海南海口570203 [2]四川大学机械工程学院,四川成都610065

出　　处：《西安工程大学学报》2021年第4期71-76,共6页Journal of Xi’an Polytechnic University

基　　金：海南省自然科学基金(20190704)。

摘　　要：针对目前电力设备监测主要依靠传统的经典神经网络自动化监测方式,准确率低且难以挖掘图像的深层信息,设计了基于SE-Attention(squeeze excitation attention)图像识别模型的电力设备监测系统。在卷积神经网络(convolutional neural network,CNN)基础上,结合SE-Net(squeeze excitation network)网络提取图像局部特征,采用深度学习中的行注意力机制、列注意力机制和通道注意力机制增加局部故障信息的权重,挖掘深层信息,提高识别电力设备故障的准确率。实验结果表明:与CNN和SE-Net检测方法相比,此检测方法在避雷器、断路器、电流互感器、电压互感器的识别准确率上分别有不同程度的提高。The current power equipment monitoring is mainly based on the traditional classical neural network automatic monitoring mode,which has such problems as low accuracy and difficulty to dig out the deep information of the image,a power equipment monitoring system based on SE-Attention(squeeze excitation attention)image recognition model was thus designed.Based on the convolutional neural network(CNN),combined with the SE-Net(squeeze extraction network)to extract the local features of the image,the row attention mechanism,column attention mechanism and channel attention mechanism in deep learning were adopted to increase the weight of local fault information,mine the deep information,and improve the accuracy of power equipment fault identification.The experimental results show that compared with CNN and SE-Net detection methods,this detection method has different improvement in the recognition accuracy of arrester,circuit breaker,current transformer and voltage transformer.

关 键 词：SE-Attention机制 电力设备监测 注意力机制 深层信息 识别故障 

分 类 号：TP277[自动化与计算机技术—检测技术与自动化装置]