基于红外图像实例分割的敞开式开关设备发热缺陷智能诊断方法  

Intelligent Diagnosis Method for Thermal Defects of Air Insulated Switchgear Based on Infrared Image Instance Segmentation

作　　者：叶剑涛 杨为 柯艳国 赵恒阳 胡迪 赵常威 YE Jiantao;YANG Wei;KE Yanguo;ZHAO Hengyang;HU Di;ZHAO Changwei(State Grid Anhui Electric Power Company Limited Research Institude,Hefei 230022,China)

机构地区：[1]国网安徽省电力有限公司电力科学研究院,合肥230022

出　　处：《高压电器》2021年第12期201-208,共8页High Voltage Apparatus

基　　金：国网安徽省电力有限公司科技项目资助(52120517000D)。

摘　　要：红外热成像是监测和诊断高压开关设备发热缺陷的方法之一,具有非接触、无损伤等优点。海量红外图像数据的处理对信息挖掘、目标识别和智能诊断提出了更高的要求,而现有方法在故障区域识别、关键特征提取和缺陷分类等方面仍存在不足。因此,文中提出了一种基于红外图像实例分割的敞开式开关设备发热缺陷智能诊断方法,利用Mask R-CNN进行感兴趣区域的自动提取与分割,构建轻量级卷积神经网络并引入到Mask R-CNN的最后一步,利用迁移学习进行模型训练,实现发热缺陷的自动识别。测试结果表明,文中方法对感兴趣区域识别的平均精确度为0.813,每秒帧数为5.196,在精度和计算速度上均高于目前的主流目标检测方法,同时在Mask R-CNN中引入的轻量级卷积神经网络使分类精度达到96.66%,特别适用于电力物联网下的智能终端设备。Infrared thermal imaging,with such advantages as no contact and no damage,is one of the methods to monitor and diagnose the thermal defect of high voltage switchgear.The processing of massive infrared image data has put forward higher requirement for information mining,target recognition and intelligent diagnosis.However,the existing methods are still inadequate in the aspects of fault area identification,key feature extraction and defect classification.Therefore,a kind of intelligent diagnosis method for thermal defects of air insulated switchgear based on infrared image instance segmentation is proposed in this paper.The Mask R-CNN is used for automatic extraction and segmentation of regions of interest,a lightweight convolutional neural network is constructed and brought to the last step of Mask R-CNN and model training is performed by transfer learning so to achieve automatic identification of thermal defects.The test results show that with this method the average accuracy of the recognized regions of interest is 0.813 and the number of frames per second is 5.196.The accuracy and calculation speed are higher than those of current mainstream target detection methods.Meanwhile,the lightweight convolutional neural network introduced in the Mask R-CNN keeps classification accuracy reach 96.66%and is especially suitable for the intelligent terminals in the power internet of things.

关 键 词：红外热成像 实例分割 敞开式开关设备 发热缺陷 迁移学习 

分 类 号：TM643[电气工程—电力系统及自动化] TP391.41[自动化与计算机技术—计算机应用技术] TN219[自动化与计算机技术—计算机科学与技术]