基于改进Mask R-CNN的复合绝缘子憎水性状态评估方法

绳飞; 曹留; 刘玉龙; 黄杰; 黄雅倩; 朱彦卿

绳飞,曹留,刘玉龙,黄杰,黄雅倩,朱彦卿.基于改进Mask R-CNN的复合绝缘子憎水性状态评估方法[J].电测与仪表,2025,62(4):73-80.
SHENG Fei,CAO Liu,LIU Yulong,HUANG Jie,HUANG Yaqian,ZHU Yanqing.Evaluation method of hydrophobicity of composite insulators based on improved Mask R-CNN[J].Electrical Measurement & Instrumentation,2025,62(4):73-80.

基于改进Mask R-CNN的复合绝缘子憎水性状态评估方法

Evaluation method of hydrophobicity of composite insulators based on improved Mask R-CNN

DOI：10.19753/j.issn1001-1390.2025.04.009

中文关键词: 憎水性目标检测 Mask R-CNN 深度学习绝缘子

英文关键词:hydrophobicity, target detection, Mask R-CNN, deep learning, composite insulator

基金项目:国家电网科技项目(项目编号：SGTYHT/20-JS-221)

作者中文名	作者英文名	单位
绳飞	SHENG Fei	国网新疆电力有限公司乌鲁木齐供电公司
曹留	CAO Liu	国网新疆电力有限公司乌鲁木齐供电公司
刘玉龙	LIU Yulong	国网新疆电力有限公司乌鲁木齐供电公司
黄杰	HUANG Jie	湖南大学电气与信息工程学院
黄雅倩	HUANG Yaqian	湖南大学电气与信息工程学院
朱彦卿	ZHU Yanqing	湖南大学电气与信息工程学院

摘要点击次数: 197

中文摘要:

及时在线判别复合绝缘子憎水性状态是保证电力系统安全运行的一个重要举措。为提高复合绝缘子憎水性状态评估模型的泛化能力，解决现有分类模型评估憎水性程度变化不均的复合绝缘子时往往只关注憎水性不错的部分而误判的问题。文章提出将分类问题转变为目标检测问题，采用改进掩膜区域卷积神经网络（mask region-based convolutional neural network，Mask R-CNN)算法评估复合绝缘子憎水性等级。通过特征金字塔网络(feature pyramid network，FPN)确定图像中所有水滴的位置与大小，采用Mask R-CNN中特有的mask分支预测所有水珠憎水性等级，再计算出相应憎水性等级所占的面积，最终选择面积最大的作为图像的憎水性等级并输出分类结果。结合各等级憎水性图像的特点，引入改进非极大值抑制（soft non-maximum suppression，Soft-NMS）来减少高等级憎水性图像中水迹面积大且分布不规则时的目标漏检，并采用Giou-loss（generalized intersection over union loss）加快低等级图像中目标小而多时模型的收敛速度。最终通过对比实验，从mAP（mean average precision）、每秒帧率(frame per second，FPS)、准确率三项评判指标验证了基于改进Mask R-CNN的憎水图像识别算法的有效性与优越性。

英文摘要:

To ensure the safe operation of power system, it is necessary to discriminate the hydrophobicity level online of composite insulators in time. In order to improve the generalization ability of composite insulator hydrophobicity state evaluation model, and solve the problem that the existing classification models only focus on the parts with good hydrophobicity when evaluating the composite insulators with uneven hydrophobicity degree. In this paper, the classification problem is transformed into the target detection problem, and the improved mask region-based convolutional neural network（Mask R-CNN） algorithm is used to evaluate the hydrophobicity level of composite insulators. Firstly, the location and size of all water droplets in the image are determined by feature pyramid network（FPN) and the mask branch of Mask R-CNN is used to predict the hydrophobicity level of all water droplets. Then, the area of the corresponding hydrophobicity level is calculated. Finally, the hydrophobicity level with the largest area is selected as the classification results of the image. Especially, combined with the characteristics of each level of hydrophobicity image, we introduce the soft non-maximum suppression（Soft-NMS） structure to reduced the missed target detection because of the scale problem of water droplets and the irregular distribution in high level hydrophobicity images, and introduce generalized intersection over union loss（GIOU） to accelerate the convergence rate of the model with small and multi-temporal objects in low level images. Final validation through comparative experiments demonstrated the effectiveness and superiority of the enhanced Mask R-CNN-based hydrophobicity image recognition algorithm across three critical metrics: mean average precision (MAP), frames per second (FPS), and classification accuracy.

查看全文查看/发表评论下载PDF阅读器

关闭