摘要
为了解决高分辨率大尺寸电气厂站接线图图元符号检测精度不高、小目标图元漏检误检等问题,提出一种基于双层-分块检测网络的厂站接线图图符检测方法.该方法将电气厂站接线图按照电气逻辑切割后对断路器、隔离开关、电抗器、接地刀闸等11种典型图元进行识别.双层-分块检测网络由基于Area-YOLOv5网络的关键区域检测层和基于Obj-YOLOv5网络的具体图元识别层构成.首先,使用Area-YOLOv5网络检测出图纸的关键区域块,其关键区检测精度达到98.5%.其次,使用Obj-YOLOv5网络识别出具体图元符号,该网络采用融合了SE注意力机制和深度可分离卷积的LC_Block模块替换瓶颈部分中的普通卷积层,图符检测精度为0.963.所提方法以较高的精度实现了电气厂站接线图图元符号的识别检测.
To solve the problems of low detection precision of high-resolution large-size electrical plant station wiring diagram symbols and small target graphic symbols missing detection and false detection,a method for detecting plant station wiring diagram symbols based on double layer-block network is proposed.After cutting the electrical plant station wiring diagram according to the electrical logic,11 typical graphic symbols such as circuit breakers,isolating switches and reactors are identified.The double layer-block network consists of a key region detection layer based on Area-YOLOv5 network and a specific symbol recognition layer based on Obj-YOLOv5 network.First,the Area-YOLOv5 network is used to detect the key area blocks of the drawings with a key area detection accuracy of 98.5%.Secondly,the Obj-YOLOv5 network is used to identify the specific symbols.The network uses the lightweight CPU block(LC_Block),which combines the squeeze-and-excitation(SE)attention mechanism with the depthwise separable convolution to replace the ordinary convolution layer in the Neck.The detection accuracy of diagram symbols is 0.963.The proposed method realizes the identification of the electrical plant station wiring diagram with high precision.
作者
程鑫
褚雪汝
邓旭晖
杨凯
谭林林
陈中
曹卫国
Cheng Xin;Chu Xueru;Deng Xuhui;Yang Kai;Tan Linlin;Chen Zhong;Cao Weiguo(School of Software,Southeast University,Nanjing 211189,China;School of Cyberspace Security,Southeast University,Nanjing 211102,China;School of Electrical Engineering,Southeast University,Nanjing 210096,China)
出处
《东南大学学报（自然科学版）》
EI
CAS
CSCD
北大核心
2022年第6期1137-1144,共8页
Journal of Southeast University：Natural Science Edition
基金
国家电网公司总部科技资助项目(SGHEDK00JYJS2200012)。
