融合知识迁移和改进YOLOv6的变电设备热像检测方法

PDF下载HTML

Title:

Thermd image detection method for substation equipment by incorporating knowledge migration and improved YOLOv6

作者:

赵振兵^1,2,3, 冯烁¹, 赵文清^2,4, 翟永杰⁴, 王洪涛^2,4

1. 华北电力大学 电子与通信工程系, 河北 保定 071003;
2. 华北电力大学 复杂能源系统智能计算教育部工程研究中心, 河北 保定 071003;
3. 华北电力大学 河北省电力物联网技术重点实验室, 河北 保定 071003;
4. 华北电力大学 控制与计算机工程学院, 河北 保定 071003

Author(s):

ZHAO Zhenbing^1,2,3, FENG Shuo¹, ZHAO Wenqing^2,4, ZHAI Yongjie⁴, WANG Hongtao^2,4

1. Department of Electronic and Communication Engineering, North China Electric Power University, Baoding 071003, China;
2. Engineering Research Center of Intelligent Computing for Complex Energy Systems, Ministry of Education, North China Electric Power University, Baoding 071003, China;
3. Hebei Key Laboratory of Power Internet of Things Technology, North China Electric Power University, Baoding 071003, China;
4. School of Control and Computer Engineering, North China Electric Power University, Baoding 071003, China

关键词:

变电设备; 热红外图像; 知识迁移; 样本生成; 目标检测; 扩散模型; 数据扩增; 深度学习

Keywords:

substation equipment; thermal infrared image; knowledge migration; sample generation; object detection; diffusion model; data augmentation; deep learning

分类号:

TP18

DOI:

10.11992/tis.202303030

摘要:

针对变电设备热像检测中存在复杂背景样本不足和设备定位困难的问题，提出了融合知识迁移和改进YOLOv6的变电设备热像检测方法。针对复杂背景样本不足问题，使用扩散模型从域外数据中提取背景知识来生成背景图像，将设备样本迁移到背景图像中生成人工热像；针对设备难以被精准定位的问题，将多头注意力机制和显示视觉中心模块融入YOLOv6模型，改善模型的特征提取能力。实验结果表明，该方法对设备的平均检测准确率达到86.4%，召回率达到89.4%，相较于基线模型分别提升了3.1%和1.5%，为变电设备热像检测提供了新的实现方法。

Abstract:

To address the problems of insufficient complex background samples and difficulty in device location in substation equipment thermal image detection, a fusion knowledge transfer and improved YOLOv6 detection method are proposed. The diffusion model was used to extract background knowledge from extraterritorial data for generating background images, solving the problem of insufficient complex background samples. The device samples were then migrated to the background images to generate artificial images. The multi-head self-attention mechanism and explicit visual center module were integrated into YOLOv6 to improve its feature extraction capability, solving the issue of difficulty in detecting devices. The experiment shows that the mAP and mAR of the proposed method reach 86.4% and 89.4%, indicating an improvement of 3.1% and 1.5% compared to the baseline model, respectively. This study provides a new implementation method for thermal image detection of substation equipment.