基于深度学习的线路走廊异物检测-数据融合与扩容  

Foreign Body Detection in Line Corridor Based on Deep Learning-Data Fusion and Augmentation

作　　者：高哲奇 赵必美 林克全 陈英达 GAO Zheqi;ZHAO Bimei;LIN Kequan;CHEN Yingda(Artificial Intelligence Technology Co.,Ltd.,China Southern Power Grid,Guangzhou 510000,Guangdong,China;China Southern Power Grid,Guangzhou 510000,Guangdong,China)

机构地区：[1]南方电网人工智能科技有限公司,广东广州510000 [2]中国南方电网有限责任公司,广东广州510000

出　　处：《电力大数据》2023年第6期28-35,共8页Power Systems and Big Data

摘　　要：本文旨在探讨数据增广技术在电力场景下的异物检测效果,重心是通过融合小样本数据来提升模型的训练数据量,从而提升检测算法的性能。在线路走廊中的异物种类繁多且形状复杂,但每一类别的数据量较小且难以获取大量样本进行训练,因此模型的性能提升受制于训练样本的数量。本文提出了一种方法,即通过融合多背景数据和多异物风筝样本来生成新的铁塔-风筝数据集。同时,本文还采用了基础的数据增强技术,以提升算法模型的泛化能力和鲁棒性。本文基于目标检测算法YOLOv5,通过对扩充后的数据集进行训练,实现了线路走廊的异物检测和识别。本文的研究成果将为电力场景下的异物检测问题提供新的解决思路和方法,并为相关情景的研究提供参考,具有重要的理论和实践意义。This paper aims to investigate the effectiveness of data augmentation techniques in foreign object detection within the electric power scenario.The focus is on enhancing the performance of detection algorithms by integrating small sample data to augment the training dataset.The variety and complexity of foreign object types in power line corridors pose a challenge as each category has a small amount of data,making it difficult to acquire a large number of samples for training.Consequently,the improvement of model performance is constrained by the quantity of training samples.This paper proposes a method that generates a new dataset of tower-kite images by fusing multiple background data and various foreign object kite samples.Additionally,basic data augmentation techniques are employed to enhance the generalization capability and robustness of the algorithm model.Based on the YOLOv5 object detection algorithm,this paper conducts training on the augmented dataset,achieving foreign object detection and recognition in power line corridors.The research outcomes of this paper offer novel approaches and methods for foreign object detection in electric power scenarios,serving as a reference for related studies and holding significant theoretical and practical implications.

关 键 词：线路走廊 异物检测 小样本数据 数据增强 YOLOv5 

分 类 号：TM752[电气工程—电力系统及自动化]