基于注意力机制的变电站仪表智能读数方法  

Substation instrument intelligent reading method based on attention mechanism

作　　者：张东升 李雪松 范新健 崔昱 王啸宇 何畏 倪瑞璞 ZHANG Dongsheng;LI Xuesong;FAN Xinjian;CUI Yu;WANG Xiaoyu;HE Wei;NI Ruipu(Northeast Electric Power Design Institute Co.,Ltd.of China Power Engineering Consulting Group,Changchun 130021,China)

机构地区：[1]中国电力工程顾问集团东北电力设计院有限公司,吉林长春130021

出　　处：《现代电子技术》2023年第14期127-134,共8页Modern Electronics Technique

基　　金：国网吉林省电力有限公司智慧变电站建设关键技术科技基金项目(522371210003)。

摘　　要：针对变电站指针式仪表读数时存在的模型部署成本高、检测读数精度低的问题,文中提出一种基于注意力机制的变电站仪表智能读数方法。在仪表检测阶段,首先结合深度可分离卷积和空间域、通道域注意力机制的特性对YOLOv5s中的卷积运算进行改进,降低模型的部署成本;然后引入轻量级自注意力计算,构建检测目标与背景间的特征关系,以提升模型的检测性能。在仪表读数阶段,使用UNet实现指针与刻度的分割,再计算指针在刻度范围中的相对位置,最后根据预设的量程得到仪表的当前数值。实验表明,所提方法在检测阶段的模型参数量仅为3.667×10^(6),检测精度达到93.39%,在读数阶段误差保持在0.03以内,具有实际应用价值。In allusion to the problems of high model deployment cost and low reading detection accuracy when reading pointer instruments in substations,a substation instrument intelligent reading method based on attention mechanism is proposed.In the instrument detection stage,the convolution operation in YOLOv5s is improved by combining the characteristics of depth-separable convolution and spatial domain and channel domain attention mechanisms to reduce the deployment cost of the model.The lightweight self-attention computation is introduced to construct feature relationships between detection targets and backgrounds,so as to improve the detection performance of the model.In the reading phase of the instruments,the UNet is used to separate the pointer from the scale,and then calculate the relative position of the pointer within the scale range,so as to obtain the current value of the instrument according to the preset range.The experiment results show that the number of model parameters in the detection stage of the proposed method is only 3.667×10^(6),its detection accuracy can reach 93.39%,and the error in the reading stage is kept within 0.03.Therefore,the method has practical application value.

关 键 词：仪表检测 智能读数 注意力机制 YOLOv5s模型 图像预处理 注意力模块 对比实验 

分 类 号：TN919-34[电子电信—通信与信息系统] TP391[电子电信—信息与通信工程]