基于深度神经网络的电气元件符号识别算法  

Recognition Algorithm for Electrical Component Symbols Based on Deep Neural Network

作　　者：江再玉 石文娟 马晶 程瑛颖 JIANG Zaiyu;SHI Wenjuan;MA Jing;CHENG Yingying(Beijing China-Power Information Technology Company Limited,Beijing 100085,China;Marketing Service Center,State Grid Chongqing Electric Power Company,Chongqing 401121,China)

机构地区：[1]北京中电普华信息技术有限公司,北京100085 [2]国网重庆市电力公司营销服务中心,重庆401121

出　　处：《电力系统及其自动化学报》2022年第2期48-55,共8页Proceedings of the CSU-EPSA

基　　金：国家电网公司科技项目(5600-201927165A-0-0-00)。

摘　　要：在智能电网发展的新时期,提高业扩报装的工作效率以及智能化程度是一项重要任务,在这一过程中对于电气图纸中电气元件符号的识别尤其关键,已有方法在算法精度以及鲁棒性上都存在不足。为此,基于YO⁃LOv3提出了一种改进的电气符号识别算法,改进了模型超参数选取策略,构建了自下而上的特征融合网络以及基于图像冗余的图像预处理方法,有效地解决了传统方法精确度低的问题。平均准确率和召回率分别达到94.8%和96.5%,与传统的图像识别算法和基准方法相比都有明显的提升。In the new era of smart grid development,the improvement of the working efficiency of business expansion and the corresponding intelligence degree is an important task,during which the identification of electrical component symbols in electrical drawings is particularly critical.However,the existing methods have shortcomings in terms of algo⁃rithm accuracy and robustness.In this paper,based on YOLOv3,an improved electrical symbol recognition algorithm is proposed,which improves the selection strategy for the model’s hyperparameters and constructs a bottom-up feature fusion network and an image preprocessing method based on image redundancy.As a result,the problem of low accura⁃cy of the traditional methods is effectively solved.The average accuracy and recall rates reach 94.8%and 96.5%,re⁃spectively,which are greatly improved compared with those of the traditional image recognition algorithms and bench⁃mark methods.

关 键 词：深度神经网络 电气图纸 电气符号 目标检测 

分 类 号：TM930[电气工程—电力电子与电力传动]