摘要
为了减少耐张线夹压接缺陷判别过程中人为因素对结果的干扰,提升评判效率,提出一种基于改进YOLOv4的输电线路耐张线夹缺陷检测方法。首先,通过“三跨”线路耐张线夹X光探伤工程,建立包含6类常见压接缺陷的图像样本集;其次,依次采用对比度拉伸、拉普拉斯算子增强、限制对比度自适应直方图均衡化及高斯滤波等图像处理方法对数据集进行预处理;然后,引入卷积块注意力模块和残差结构对原始YOLOv4模型的路径聚合网络层和空间金字塔池化层进行改进,增强模型对空间和通道维度的关注度,优化模型多感受野特征表达能力,提升算法对深层语义特征的提取效果;接着,利用预处理数据对YOLOv4及其改进模型进行训练与测试;最后,根据YOLOv4及其改进模型对各类压接缺陷的检测性能不尽相同这一特点,采用2次非极大值抑制搭建多网络融合缺陷检测模型,并完成模型的测试。先后的测试结果表明,改进YOLOv4、多网络融合模型对6类压接缺陷检测的平均精度均值分别为92.22%、93.08%,可实现对耐张线夹压接缺陷的有效检测。
In order to reduce the interference of human factors on the results in the process of identifying the crimping defects of the strain clamps and improve the evaluation efficiency,this paper proposes a method for detecting the defects of strain clamps of the transmission lines based on improved YOLOv4.Firstly,through the"three types of crossing span"line strain clamp X-ray flaw detection project,an image sample set containing six types of common crimping defects is established.Then,the image processing methods such as contrast stretching,Laplace operator enhancement,contrast limited adaptive histogram equalization(CLAHE)and Gaussian filtering are used to preprocess the data set.Secondly,the convolutional block attention module(CBAM)attention mechanism and residual structure are introduced to improve the path aggregation network(PANet)layer and spatial pyramid pooling(SPP)layer of the original YOLOv4 model,enhance the model s attention to space and channel dimensions,optimize multiple receptive field feature expression ability of the model,and improve extraction effect on deep semantic features of the algorithm.Subsequently,the preprocessed data is used to train and test the YOLOv4 and its improved model.Due to that the detection performance of various crimping defects is different using the YOLOv4 and its improved model,a multi-network integration defect detection model is built by using twice non-maximum suppression,and the model testing is completed.The successive experimental results show that the mAP values of the improved YOLOv4 and multi-network integration model for the detection of six types of crimping defects are 92.22%and 93.08%respectively,which can effectively detect the crimping defects of strain clamps.
作者
王郑
杜怀云
王金沛
彭冲
周大为
WANG Zheng;DU Huaiyun;WANG Jinpei;PENG Chong;ZHOU Dawei(Guiyang Bureau of CSG EHV Power Transmission Company,Guiyang,Guizhou 550081,China)
出处
《广东电力》
2023年第5期105-114,共10页
Guangdong Electric Power
基金
中国南方电网有限责任公司超高压输电公司(贵阳局)职工技术创新项目。
关键词
耐张线夹
压接缺陷
限制对比度自适应直方图均衡化
YOLOv4
卷积块注意力模块
多网络融合
strain clamp
crimping defect
contrast limited adaptive histogram equalization(CLAHE)
YOLOv4
convolutional block attention module(CBAM)
multi-network integration
作者简介
通讯作者:王郑(1989),男,湖北松滋人,工程师,研究方向为输电线路运行与维护,E-mail:455967200@qq.com;杜怀云(1981),男,贵州贵阳人,工程师,研究方向为输电线路运行与维护;王金沛(1984),男,贵州贵阳人,工程师,研究方向为输电线路运行与维护。
