摘要
针对气体绝缘组合电器(Gas Insulated Switchgear,GIS)红外热成像巡检缺乏对关键组件的红外特征实时识别功能、温变特性的提取缺乏针对性和易受干扰等不足,提出了一种基于知识蒸馏的轻量化改进GIS红外目标检测和温变特征提取算法。通过Mosaic和Mixup策略对变电站实际采集到的图片进行数据增强预处理,构建了3200条包含水平母线、垂直母线、互感器、断路器和隔离开关的GIS红外图谱数据集;采用YOLOX-s作为教师检测器,YOLOX-tiny作为学生检测器,使用教师网络模型指导学生网络模型进行知识蒸馏训练,进一步从整体上提升知识迁移的性能,实现对GIS不同组件的轻量化目标识别,在数据集上的测试结果表明明显优于其他目标检测算法,所提出的算法模型参数仅为19M,在上位机终端的识别速度可达29.5fps,同时保持了的82.3%的识别准确率;并采用灰阶差值算法进一步实现了对识别出的GIS组件的温度特征提取和分析。结果表明,本文所提出的算法可便于实现对智能红外终端的轻量化部署和集成,为提高GIS红外热成像巡检智能化水平和内部过热缺陷的诊断可靠性提供了一种有效方法。
In view of the gas insulated switchgear,the infrared imaging inspection is short of infrared featwre real-time indentifying function to key assembly and the temperature change characteristic extraction is short of directivity and easy to be disturbed etc.,A kind of light-weight modified GIS infrared object test and temperature change characteristic extraction algorithm based on knouledge distillation is advanced.Through Mosaic and Mixup tactics,the pictures actually collected by the transformer substation carried out intensifying pretreatment of data and setup 3200 including GIS infrared atlas data acquisition of horizontal bus,vertical bus,mutual inductor,circuit breaker and isolator,adopting TOLOX-s as student′s detector and YOLOX-tiny as student′s detector.Use the teacher′s network model to guide the student′s network model making knowledge distillation training make further improvement on knowledge transport performance as a whole so as to achieve the light weight object recognition to different assemldy of GIS.The test results for the data acquisition show they are obviously better than the other object detecting algorithm.The algorithm model parameter proposed is only 19M and the idetification speed on the upperseat machine terminal can be up to 29.5fps,at the same time keeping identifying accuracy of 82.3%,and adopt grey step D-value algorithm to further achieve the temperature characteristic extraction and analysis of the GIS assembly.The results show the algorthm proposed by the paper can be easy to achive the light weight disposition and integration to intelligent infrared terminal,which can provide a kind of effective method to increase diagnostic reliablility of GIS infrared thermal imaging inspection intelligence level and internal overheating defect.
作者
黄辉
张凯
程飞
HUANG Hui;ZHANG Kai;CHENG Fei(State Grid Hami Power Supply Company,Hami 839000,China)
出处
《电气开关》
2023年第2期54-59,共6页
Electric Switchgear
关键词
GIS
YOLOX
红外图像
温升提取
知识蒸馏
轻量级网络
GIS
YOLOX
infrared image
temperature rise extraction
knowledge distillation
light weight network
