元学习优化的多重特征增强输变电缺陷检测网络  

Multi-Feature Enhancement Network for Power Transmission and Transformation Defect Detection Based on Meta-Learning Optimization

作　　者：毛万登 袁少光 姜亮 田杨阳 鲍华 MAO Wan-deng;YUAN Shao-guang;JIANG Liang;TIAN Yang-yang;BAO Hua(State Grid HENAN Electric Power Company,Zhengzhou 450000;College of Artificial Intelligence,Anhui University,Hefei 230601)

机构地区：[1]国网河南省电力公司,河南郑州450000 [2]安徽大学人工智能学院,安徽合肥230601

出　　处：《制造业自动化》2025年第2期59-67,共9页Manufacturing Automation

基　　金：安徽省自然科学基金面上项目(1908085MF217);安徽省教育厅重点项目(KJ2019A0022)。

摘　　要：对输变电设备进行缺陷检测已经成为维持电网稳定运行的重要一环。尽管深度学习方法输变电设备缺陷检测方面取得显著的进展,但依旧面临着缺陷样本稀少造成的小样本挑战性问题。为此,提出基于元学习的小样本输变电设备缺陷检测网络,以提高输变电设备的缺陷检测精度。该网络使用DarkNet-53作为检测框架的主干网络,并提出全局信息提取模块、通道注意力模块和跨阶段特征融合模块进行多重特征增强对网络主干进行改进,增强对数据的处理能力。通过将训练集拆分成支持集和查询集,依据元学习算法划分两阶段训练,通过元学习算法优化训练阶段的参数更新策略来解决小样本问题。实验表明,此方法的mAP@0.5可达到0.51,与现有的主流检测方法相比,此方法在输变电缺陷的检测上取得了显著的效果。Performing defect detection on power transmission and transformation equipment has become an important part of maintaining the stable operation of the power grid.Despite significant advancements in deep learning methods for defect detection in power equipment,there still exists challenges of a few-shot as the result of the scarcity of defect samples.To address this issue,a few-shot defect detection network for power transmission and transformation equipment is proposed based on meta-learning to improve the defect detection accuracy of power transmission and transformation equipment.The network utilizes DarkNet-53 as the backbone network of the detection framework and introduces multiple feature enhancement modules,including global information extraction,channel attention and cross-stage feature fusion to improve the backbone of the network and enhance the processing capability of data.By splitting the training set into support sets and query sets,a two-stage training process is conducted based on meta-learning algorithms.The meta-learning algorithm optimizes the parameter update strategy during the training stage to tackle the few-shot learning problem.The experimental results demonstrate that this method achieves a mean average precision(map)of 0.51 at IoU threshold 0.5,showing a significant improvement in the defect detection for power transmission and transformation equipment compared to the existing mainstream methods.

关 键 词：缺陷检测 元学习 特征提取 特征增强 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]