摘要
针对变电站二次回路的定期检修与维护过于依赖人工规则、泛化能力差,导致故障检测识别精度不足的问题,提出一种基于深度学习的变电站二次图纸元件识别与结构化提取方法。首先,对原始变电站二次图纸进行图像分割,将其划分为多个子图;随后,基于YOLOv5框架构建元件识别模型,并引入多尺度锚框机制对子图进行检测。具体而言,通过计算不同尺度特征图上各锚框对应元件类别的概率分布,选取概率最大值所对应的标注框作为该元件的识别结果。最后,融合各子图的检测结果,提取全局内元件锚点间距离不超过最大连接距离的连接关系。实验结果表明:所提方法在测试集上的点云分类精确率达到96.0%,召回率为94.7%,F1分数为95.0%,且对不同元件识别结果的mAP50稳定在0.90以上。研究在保持较高召回率的同时有效提升了元件识别精度,为变电站智能化巡检与二次回路数字化管理提供了可靠的技术途径。
To address the issue of insufficient fault detection accuracy in the regular maintenance of substation secondary circuits,which stems from over-reliance on manual rules and poor generalization capabilities,this a deep learning-based method for identifying and structurally extracting components from substation secondary drawings is proposed.Firstly,the original substation secondary drawings undergo image segmentation,dividing them into multiple sub-graphs;then,an object recognition model is constructed based on the YOLOv5 framework,and employing a multi-scale anchor mechanism for sub-image detection.Specifically,by calculating the probability distribution of corresponding object categories across anchor boxes on feature maps of different scales,the object recognition result is determined as the anchor box with the maximum probability.Finally,detection results from all sub-images are fused to extract connection relationships where the distance between anchor points within the global object does not exceed the maximum connection distance.Experimental results demonstrate that this method achieves a point cloud classification accuracy of 96.0%,a recall of 94.7%,and an F 1 score of 95.0%on the test set.Moreover,the mAP50 for different component recognition results remains stable above 0.90.This research effectively improves component recognition accuracy while maintaining high recall,providing a reliable technical approach for intelligent substation inspection and the digital management of secondary circuits.
作者
刘军
季瑞
马慧超
王兆祺
LIU Jun;JI Rui;MA Huichao;WANG Zhaoqi(Economic and Technological Research Institute of State Grid Zhejiang Electric Power Company,Hangzhou 310020,China;Nanjing Yachen Zhonglian Electric Power Technology Co.,Ltd.,Nanjing 211100,China)
出处
《国外电子测量技术》
2025年第11期321-327,共7页
Foreign Electronic Measurement Technology
关键词
深度学习
变电站
二次图纸
元件识别
结构化提取
图像分割
deep learning
substation
secondary drawings
component identification
structured extraction
image segmentation
