面向电力缺陷场景的小样本图像生成方法  

A few-shot image generation method for power defect scenarios

作　　者：何宇浩 宋云海 何森 周震震 孙萌 陈毅 闫云凤[3,4] HE Yuhao;SONG Yunhai;HE Sen;ZHOU Zhenzhen;SUN Meng;CHEN Yi;YAN Yunfeng(CSG EHV Electric Power Research Institute,Hangzhou 510000,China;School of Mechanical Engineering,Zhejiang University,Hangzhou 310027,China;Hainan Institute of Zhejiang University,Sanya,Hainan 572000,China;School of Electrical Engineering,Zhejiang University,Hangzhou 310027,China)

机构地区：[1]南网超高压公司电力科研院,广州510000 [2]浙江大学机械工程学院,杭州310027 [3]浙江大学海南研究院,海南三亚572000 [4]浙江大学电气工程学院,杭州310027

出　　处：《浙江电力》2024年第1期126-132,共7页Zhejiang Electric Power

基　　金：浙江省科技计划项目(2022C01056);浙江省科技计划项目(LQ21F030017);CCF-联想蓝海科研基金。

摘　　要：由于电力缺陷数据稀缺,目前大多数缺陷检测方法都无法有效地对电力缺陷情况进行准确的检测。为此,使用小样本图像生成方法,基于改进的LoFGAN(局部融合生成对抗网络),设计基于上下文信息的小样本图像生成器,提高缺陷检测网络对细节特征的提取能力;引入基于LC-散度的正则化损失来优化图像生成模型在有限数据集上的训练效果。实验表明,小样本图像生成方法能够为电力场景缺陷情况生成有效且多样的缺陷数据,所提模型能够有效解决电力缺陷场景数据稀缺的问题。Due to the limited availability of power defect data,most current defect detection methods are unable to accurately detect power system anomalies.To overcome this challenge,a few-shot image generation method is em⁃ployed.Building upon the improved local-fusion generative adversarial network(LoFGAN),a context-aware fewshot image generator is designed to enhance the defect detection network’s capability to extract detailed features.A regularization loss based on LC-divergence is introduced to optimize the training effectiveness of the image genera⁃tion model on limited datasets.Experimental results reveal that the few-shot image generation method can generate effective and diverse defect data for power scenarios.The proposed model can address the issue of data unavailabil⁃ity in power defect scenarios.

关 键 词：小样本图像生成 电力缺陷 上下文信息 LC-散度 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术] TM50[自动化与计算机技术—计算机科学与技术]