（1 Fuzhou Power Supply Company, State Grid Fujian Electric Power Co., Ltd, Fuzhou 350004, China;
2 College of Electrical Engineering and Automation, Fuzhou University, Fuzhou 350108, China）
Abstract: Due to the influence of interference factors such as stains, light, and shooting angles at the substation site, the instrument information in the images captured by the remote inspection system is weakened, and the pointer information is missing, resulting in a relatively low recognition accuracy of the instruments. Aiming at this problem, the target detection framework of YOLOv5 was adopted, and a cross-fusion feature pyramid network was designed, which enhanced the ability of the target detection network based on YOLOv5 to extract the position information of instruments. To address the problem of pointer fragmentation caused by blurred and tilted instrument images, the U-Net semantic segmentation network is adopted to identify the pointers in the instrument images, achieving the generation of instrument pointers in the interference environment. Experiments show that the proposed intelligent recognition algorithm for sulfur hexafluoride (SF6) instrument based on deep learning has demonstrated strong recognition capabilities in the complex environment of substations, with the accuracy rate of pointer recognition increasing from the original 63% to 96%.
Key words: substation; remote inspection; deep learning; SF6 instrument; intelligent recognition; cross-fusion; feature pyramid network
參考文獻
[1] 丁思奎，李健. 變電站巡檢機器人應用中存在的問題分析及解決方案[J]. 電工電氣，2016(2) ：57-58.
[2] 余福. 基于深度學習的變電站指針儀表讀數識別方法研究[D]. 吉林：東北電力大學，2022.
[3] 付世雄. 基于圖像處理的指針式儀表示數識別[D]. 哈爾濱：哈爾濱工業大學，2022.
[4] 錢玉寶，王紫涵，邱騰煌. 指針式儀表讀數識別的研究現狀與發展[J] . 電子測量技術，2024，47(8) ：110-119.
[5] 荊永菊，薄樹奎，郝曉玉，等. 基于圖像處理的指針式儀表讀數識別[J] . 信息技術與信息化，2024(12) ：27-31.
[6] LIN Ye, XU Zhezhuang, WU Yiying, et al.A multitask network for occluded meter reading with synthetic data generation technology[J].Advanced Engineering Informatics，2025，64 ：51-64.
[7] ZHOU Chuanhua, ZHOU Jiayi, YU Cai, et al.Multichannel sliced deep RCNN with residual network for text classification[J].Chinese Journal of Electronics，2020，29(5) ：880-886.
[8] 南曉虎，丁雷. 深度學習的典型目標檢測算法綜述[J].計算機應用研究，2020，37(S2) ：15-21.
[9] LIN Ye, XU Zhezhuang, YUAN Meng, et al.Pointer generation and main scale detection for occluded meter reading based on generative adversarial network[J].Measurement， 2024，234，114836-114850.
