基于多光谱语义分割的GIS组件红外特征识别方法  

Infrared Characteristic Recognition Method of GIS Assembly Based on Multispectral Semantic Separation

作　　者：王炜斌 林熠 刘江 关向雨 WANG Wei-bin;LIN Yi;LIU Jiang;GUAN Xiang-yu(Electric Automatization College,Fuzhou University,Fuzhou 350108,China)

机构地区：[1]福州大学电气与自动化工程学院,福建福州350108

出　　处：《电气开关》2024年第2期37-42,47,共7页Electric Switchgear

基　　金：福建省自然科学基金资助项目(2020J01509)。

摘　　要：针对气体绝缘组合电器(Gas Insulated Switchgear,GIS)红外巡检中图像对比度低、易受道路等背景辐射干扰,各组件温升差异难以直接提取的不足,提出了基于多光谱语义分割和目标检测的GIS各组件红外特征智能识别算法。采用红外-可见光双光谱成像系统,构建了现场变电站GIS的红外和可见光多光谱数据集;基于MF-net多光谱语义分割框架,采用双分支编码器分别提取GIS外壳红外和可见光特征信息,进而在译码器上实现多光谱特征融合和GIS本体红外图像分割,从而排除了环境背景辐射对GIS组件识别和温升特性提取的影响;将分割后的GIS本体红外图像进行主母线、分支母线、互感器、断路器和隔离开关等组件标记,随后采用YOLOv4算法实现对GIS不同组件的识别。结果表明与未进行语义分割GIS红外组件识别模型相比,所提出模型可以达84.2%平均识别准确率,其中电流互感器精确度达93.68%、断路器精确度达92.68%、召回率达96.20%和F1值得分0.94,GIS组件温升识别能够去除道路等高辐射背景对测温结果干扰,对提高户外GIS设备温升红外热成像的结果可靠性和智能化具有应用价值。Inview of low image contrast of gas insulated switchgear in infrared scan test,easy to be radiated and disturbed by road and so on and temperature rise difference of various assembly difficult to be recovered,propose infrated characteristic intelligent recognition method of various assembly of GIS based on multispectral semantic separation and object detection.Adopt dual spectral imaging system finfrared visible light and set up the multispectral data collection of the infrared and visible light of the GIS for the field transformer substation.Based on MF-net multispectral semantic separation frame,adopt a double branched coder to separately recover infrared and visible light characteristic information of the GIS housing,thus achieving multispectral characteristic fusion and GIS body infrared image segmentation in the coder and removing the influence of environmental background radiation on GIS assembly recognition and the temperature rise characteristic extraction.After segmenting,take the infrared image of the GIS body of main bus,branch bus,transformer,circuit breaker and isolator assembly signed.Afther that,adopt YOLOV4 algorithm to achieve different assembly recognition of the GIS,The results show,compared with the GIS infrared assembly recognition model without semautic separation,the model proposed can come up to average 84.2%identification rate,which of current transformer accuracy comes up to 93.68%,circuit breaker 92.68%,recall rate 96.20%and F1 value score 0.94.The GIS assembly temperature rise recognition can remove high radiation back ground of the road to temperature measure result interference,which is of application value to increase esult reliability and intelligence of the infrared thermal imaging of the open GIS.

关 键 词：GIS YOLOv4 红外图像 可见光图像 语义分割 MFNet 

分 类 号：TM81[电气工程—高电压与绝缘技术]