小样本变压器图像故障特征的分类方法设计  

Design of Classification Method for Fault Feature of Small Sample Transformer Images

作　　者：庄莉 宋立华[1] 徐凡 伍臣周 张晓东 ZHUANG Li;SONG Lihua;XU Fan;WU Chenzhou;ZHANG Xiaodong(Fujian Yirong Information Technology Co.,Ltd.,Fuzhou 350001,China;State Grid Information&Telecommunication Co.,Ltd.,Beijing 102211,China)

机构地区：[1]福建亿榕信息技术有限公司,福建福州350001 [2]国网信息通信产业集团有限公司,北京102211

出　　处：《自动化仪表》2025年第3期90-94,100,共6页Process Automation Instrumentation

基　　金：国网信通产业集团两级协同研发基金资助项目(SGITYLYRSCJS2310305)。

摘　　要：小样本变压器发生故障时的频率较低,判断故障类型难度较高。为了准确分类变压器故障特征,提出基于迁移学习的小样本变压器图像故障特征分类方法。利用变压器故障图像的高斯函数,计算变压器故障图像密度指数。通过分析和修正密度指标,得到迁移学习网络中变压器故障图像结构。利用维度方向拼接变压器图像特征。通过求解变压器故障图像特征提取的目标函数,更新变压器故障图像的权值并提取小样本变压器图像故障特征。将变压器故障图像划分为多个区域,提取各区域内图像颜色和纹理特征的差异性。利用变压器故障图像特征的关联损失函数,设计变压器故障特征分类算法,实现变压器图像故障特征分类。试验结果表明,该方法能准确分类变压器图像故障特征,并将特征分类的曲线下面积(AUC)值提高到0.95以上。该方法增强了分类的真实性。Small sample transformer has a low frequency when faults occur,and it is difficult to determine the type of faults.To accurately categorize transformer fault features,a migration learning-based fault feature classification method for small sample transformer images is proposed.Using the Gaussian function of transformer fault image,the density index of transformer fault image is calculated.By analyzing and correcting the density index,the transformer fault image structure in the migration learning network is obtained.The transformer image features are spliced using dimensional orientation.The weights of transformer fault image is updated by solving the objective function of transformer fault image feature extraction,and the small sample transformer image fault features are extracted.The transformer fault image is divided into multiple regions,and the variability of image color and texture features are extracted within each region.Using the correlation loss function of transformer fault image features,the transformer fault feature classification algorithm is designed to realize transformer image fault feature classification.The experimental results show that the method can accurately classify transformer image fault features and increase the area under the curve(AUC)value for feature classification to more than 0.95.The method enhances the realism of classification.

关 键 词：迁移学习 图像识别 故障特征 特征提取 小样本变压器 权值更新 

分 类 号：TH69[机械工程—机械制造及自动化]