基于多尺度特征和图像补全的变压器套管憎水性识别方法
Transformer Bushing Hydrophobicity Identification Method Based on Multi-Scale Features and Image Completion
刘旭光 1李婷婷 1李建萍1
作者信息
- 1. 国网冀北电力有限公司张家口供电公司,张家口 075000
- 折叠
摘要
提出基于多尺度特征和图像补全的变压器套管憎水性识别方法,及时、准确识别套管憎水性情况,避免电气故障发生,提高变压器运行的可靠性与电力系统的稳定性.利用BSCB算法补全变压器套管图像中的缺损区域,获取完整图像;利用LBP算法提取补全后图像的纹理特征向量,利用Canny算法对补全后图像实施分割处理,得到水珠与背景的分割图像,从分割图像中提取与套管憎水性密切相关的四个特征量来描述套管憎水性特征;构建概率神经网络模型,将获取的纹理特征向量与套管憎水性特征量作为输入项,输入至训练好的概率神经网络模型,其输出结果即是变压器套管憎水性最终的识别结果.实验表明,该方法可以精准识别变压器套管憎水性级别,当概率神经网络模型中的平滑因子设置为 0.7 时,识别效果最好.
Abstract
This paper proposes a method of transformer bushing hydrophobicity identification based on multi-scale features and image completion,which can identify the bushing hydrophobicity timely and accurately,avoid electrical faults,and improve the reliability of transformer operation and the stability of power system.BSCB algorithm is used to complete the defective area in the transformer bushing image to obtain the complete image.LBP algorithm was used to extract the texture feature vector of the completed image,and Canny algorithm was used to segment the completed image to obtain the image of water drop and background segmentation.Four features closely related to casing hydrophobicity were extracted from the segmented image to describe casing hydrophobicity characteristics.A probabilistic neural network model is constructed,and the obtained texture feature vector and casing hydrophobicity feature quantity are input to the trained probabilistic neural network model.The output result is the final recognition result of transformer casing hydrophobicity.Experimental results show that this method can accurately identify the hydrophobicity level of the transformer bushings,and the identification effect is best when the smoothing factor in the probabilistic neural network model is set to 0.7.
关键词
变压器套管/图像补全/憎水性识别/憎水性特征/纹理特征/概率神经网络Key words
transformer bushing/image completion/hydrophobic identification/hydrophobic characteristics/texture feature/probabilistic neural network引用本文复制引用
出版年
2024
NETL
NSTL
万方数据