and Communication Branch of State Grid Shanxi Electric Power Company

【目的】变电站电力设备种类多设备之间连接关系复杂,普遍存在设备位置和图像对比较为单一,实际应用中只能获得有限数量的目标图像和标记以及传统方式分割带来的电力设备图像分割精度不够的问题,将卷积神经网络(ConvolutionNeuralNetwork,CNN)与Transform-mer结合组成新模型进行变电站电力设备的分割,提出了一种基于编解码器结构的新型SE-Transfomer(SubstationEquipmentTransformer)网络。【方法】为了获取局部上下文信息,编码器首先利用CNN提取空间特征图。同时,对特征图进行了精心改造,多尺度特征输入进行全局特征建模。该解码器利用Transformer提取全局深层特征,并进行逐步上采样来预测详细的分割图。SE-Transfomer在山西省梁家庄变电站数据集上进行广泛实验,其纵向结果的Dice系数(Dice)、召回率(Recall)、特异度(Specificity)和均方根误差(RMSE)分别为89.31%,90.52%,89.62%和11.32.【结果】结果表明,SE-Transfomer在变电站电力设备的扫描上获得了与以往最先进的分割方法相当或更高的结果。

【Purposes】 Owing to the varietry of electrical equipment and the complex connec-tion between them in transformer station, there are many common problems includeng relatively limited location and picture contrast of equipment, insufficient target images and markers in prac-tical applications, and inaccurate electrical equipment image segmentation brought by the tradi-tional way. In this paper, CNN (Convolutional Neural Network) is combined with Transformer to form a new model for segmentation of electrical equipment, and a new SE-Transfomer (Sub-station Equipment Transformer) network based on codec structure is proposed. 【Methods】 To obtain the local context information, the coder extracts the spatial feature map by using CNN at first. Meanwhile, the feature map is carefully modified with multi-scale feature inputs for global feature modeling. The decoder extracts global deep features using Transformer and performs stepwise up-sampling to predict the detailed segmentation map.SE-Transfomer is extensively ex-perimented on the dataset of Liangjiazhuang Transformer Station in Shanxi province, and its lon-gitudinal results of Dice, Recall, Specificity, and RMSE (Root Mean Square Error) are 89.31%, 90.52%, 89.62%, and 11.32, respectively. 【Findings】 The results indicate that SE-Transfomer obtains comparable or higher results than previous state-of-the-art segmentation methods on the scanning of electrical equipment in the transformer station.

国网山西省电力公司科技项目资助(52051C220003)

李洋,朱春山,张建亮,等.基于改进 Transformer的变电站复杂场景下电力设备分割[J].太原理工大学学报,2024,55(1):57-65.

LI Yang,ZHU Chunshan,ZHANG Jianliang,et al.Electrical equipment segmentation in complex substation sce-narios based on improved transformer[J].Journal of Taiyuan University of Technology,2024,55(1):57-65.