面向电力开关柜的轻量型GB-YOLOv5m状态检测方法

电力开关柜状态灯及仪表具有布局高密、异位同像的特点，从而对边端图像处理技术中的目标形貌、色度对比等基础特征检测能力以及轻量识别能力提出更高要求，为此该文提出一种Ghost-BiFPN-YOLOv5m(GB-YOLOv5m)方法。采用加权双向特征金字塔(BiFPN)结构，赋予特征层不同权重以传递更多有效特征信息；增加一个检测层尺度，提升网络对于小目标的检测精度，解决状态灯高密布局引起的小目标识别难问题；利用Ghost-Bottleneck结构替换原主干网络的Bottleneck复杂结构，实现模型的轻量化，为在边端部署模型提供有利条件；通过图像增强技术对有限样本进行状态灯和仪表传递特征的扩充，并通过迁移学习实现算法高速收敛。经10 kV开关柜实测，结果表明该算法对柜体状态灯及仪表共16类目标识别准确率高，均值平均精度(mAP)达97.3%，fps为37.533帧；相较于YOLOv5m算法，在模型大小缩小了37.04%的基础上，mAP提升了10.2%，说明所提方法对灯体与表体的检测能力大幅提升，且轻量识别效率提升明显，对于开关柜电力状态的实时核验与数字孪生信息交互，具有一定的现实意义。

Lightweight GB-YOLOv5m State Detection Method for Power Switchgear

Since status lights and instruments of power switchgear have the characteristics of high density and ectopic image, the detection ability of basic features such as target morphology, chromaticity comparison, and lightweight recognition ability in edge image processing technology presents higher requirements. Therefore, a Ghost-Bifpn-YOLOv5m(GB-YOLOv5m) method is proposed. Specifically, the Bi-directional Feature Pyramid Network (BiFPN) structure is adopted to give different weights to the feature layer to transmit more effective feature information. A detection layer scale is added to improve the detection accuracy of the network for small targets and tackle the complication of small target recognition caused by the high-density layout of status lights. The Ghost-Bottleneck structure is employed to replace the complex Bottleneck structure of the original backbone and realize the lightweight of the model, contributing to the formation of favorable conditions for deploying the model at the edge. Additionally, the transmission characteristics of status lights and instruments for limited samples are expanded with the image enhancement technology, and the high-speed convergence is realized through migration learning. The experimental results of 10 kV switchgear demonstrate that the algorithm has high recognition accuracy for 16 categories of cabinet status lights and instruments, with mean Average Precision (mAP) of 97.3% and fps of 37.533. Compared with the YOLOv5m algorithm, the model size is reduced by 37.04%, and mAP is increased by 10.2%, implying that the proposed method possesses a significantly enhanced detection ability for lamp bodies and table bodies, as well as remarkably improved lightweight recognition efficiency, which has certain practical significance for the real-time verification of the power state of the switchgear and the interaction of digital twins.