基于卷积神经网络的电网变压器铭牌识别技术研究  

Research on Nameplate Identification Technology of Power Grid Transformer Based on Convolution Neural Network

作　　者：王元峰[1] 龙思璇 曾惜[1] 王宏远[1] 林家杰 陈华彬 WANG Yuan-fen;LONG Si-xuan;ZENG Xi;WANG Hong-yuan;LIN Jia-jie;CHEN Hua-bin(Chengbei Branch of Guiyang Power Supply Bureau of Guizhou Power Grid Co.,Ltd.,Guizhou Guizhou 550001;Guizhou Qianchi Information Co.,Ltd.,Guizhou Guiyang 550002)

机构地区：[1]贵州电网贵阳供电局城北分局,贵州贵阳550001 [2]贵州黔驰信息股份有限公司,贵州贵阳550002

出　　处：《数字技术与应用》2020年第7期113-115,共3页Digital Technology & Application

摘　　要：针对现有的CNN网络模型在电网变压器铭牌识别应用中容易发生过拟合,训练速度慢等问题,为提高变压器铭牌识别准确率,提升训练效率,基于传统CNN算法理论提出一种I_CNN算法。首先设计全局池化层来代替传统CNN网络中的全连接层,降低过拟合风险;然后引入一种改进的softmax分类器构建softmax分类层,有效提高训练效率;最后使用实地采集的变压器铭牌图片数据集上训练I_CNN网络模型,识别准确率达96.21%,并通过对比实验表明,本文提出的I_CNN算法具有较高的准确率和训练效率。In order to improve the accuracy of transformer nameplate recognition and training efficiency,an I_CNN algorithm is proposed based on the traditional CNN algorithm theory.Firstly,the global pooling layer is designed to replace the full connection layer in the traditional CNN network to reduce the risk of over fitting;then an improved softmax classifier is introduced to build the softmax classification layer to effectively improve the training efficiency;finally,the I_CNN network model is trained on the image data set of the transformer nameplate collected in the field,and the recognition accuracy is 96.21%,and the comparative experiment shows that this paper proposes The I_CNN algorithm has high accuracy and training efficiency.

关 键 词：卷积神经网络 图片识别 变压器 铭牌识别 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]