基于低层特征融合多核卷积神经网络的管道缺陷漏磁图像识别方法  

Magnetic Flux Leakage Image Recognition Method of Pipeline Defects Based on Low-level Feature Fusion Multi-core Convolution Neural Network

作　　者：米亚鑫 郎宪明[1] 韩富丞 李雪川 MI Ya-xin;LANG Xian-ming;HAN Fu-cheng;LI Xue-chuan(School of Information and Control Engineering,Liaoning Petrochemical University,Fushun Liaoning 113001,China)

机构地区：[1]辽宁石油化工大学信息与控制工程学院,辽宁抚顺113001

出　　处：《当代化工》2023年第3期677-681,共5页Contemporary Chemical Industry

基　　金：国家自然科学基金项目(项目编号:61673199);国家自然科学基金项目(项目编号:62073158);中国博士后科学基金资助项目(项目编号:2020M660125);辽宁省博士科研启动基金计划项目(项目编号:2019-BS-158);辽宁省教育厅项目(项目编号:L2020017);辽宁石油化工大学引进人才科研启动基金资助项目(项目编号:2019XHHL-008)。

摘　　要：为了提高管道缺陷漏磁图像识别的准确率,提出了一种基于低层特征融合多核卷积神经网络的方法。首先构造一个由多卷积核Concat操作构成的Concatenation模块,由3个Concatenation模块和全连接层构成识别模型,该模型能得到不同尺度的信息。然后对原始图像进行下采样操作,与经过第一个Concatenation模块卷积操作后得到的特征映射进行融合,就能得到更多的空间、颜色、轮廓等信息,再经过两层Concatenation模块进行特征提取最后分类。结果表明:融合了原始图像的低层特征后的模型对管道漏磁缺陷的识别率相较于CNN提高了7.5%,相较于传统的机器学习算法准确率提升了11.5%。In order to improve the accuracy of magnetic flux leakage image recognition of pipeline defects,a method based on low-level feature fusion multi-core convolution neural network was proposed.Firstly,a Concatenation module composed of multiple convolution kernel Concat operations was constructed.The recognition model was composed of three Concatenation modules and full connection layer.The model could obtain information of different scales.Then,the original image was downsampled and fused with the feature map obtained after the convolution operation of the first Concatenation module.The model could get more space,color,outline and other information.Then feature extraction was carried out through two-layer Concatenation module,the last step was classification prediction.The experimental results showed that the accuracy of the model with low-level features was 7.5%higher than that of CNN,and 11.5%higher than that of the Machine Learning algorithm.

关 键 词：特征融合 漏磁 低层特征 卷积神经网络 

分 类 号：TE973.6[石油与天然气工程—石油机械设备]