基于ResNet18-ECA的管道焊缝交点定位方法  

A method for identifying pipeline weld intersections based on ResNet18-ECA

作　　者：肖放 杨理践[1] 田野 耿浩[1] XIAO Fang;YANG Lijian;TIAN Ye;GENG Hao(School of Information Science and Engineering,Shenyang University of Technology;PipeChina West Pipeline Co.,Ltd.)

机构地区：[1]沈阳工业大学信息科学与工程学院 [2]国家管网集团西部管道有限责任公司

出　　处：《管道保护》2025年第2期62-68,77,共8页

基　　金：国家自然科学基金青年资助项目“管道缺陷漏磁内检测正向问题解析方法研究”(62101356)。

摘　　要：环焊缝和螺旋焊缝交点的准确识别可作为目标管节精确定位的判断依据,因此,使用一种高效且智能的识别方法是非常有必要的。针对长距离管道焊缝交点的智能识别定位问题,提出了一种基于改进残差神经网络(residual neural net⁃work18,ResNet18)的智能识别定位方法。该方法从焊缝交点处漏磁信号的特征出发,通过分析焊缝交点处的漏磁场分布,结合人工判读的焊缝交点在漏磁信号中的通道信息和传感器编号,以及管道漏磁数据轴向分量、周向分量、径向分量在焊缝交点处的不同影响,在残差神经网络(ResNet18)中引入增强通道注意力机制(efficient channel attention,ECA)以建立ResNet18-ECA,并使用Huber损失函数、Leaky ReLU激活函数替换原来的MSE损失函数和ReLU激活函数,从而实现对焊缝交点的定位。选取焊缝交点数据集进行350轮训练,对比训练损失和均方根误差(root mean squared error,RMSE)的变化趋势。结果表明:使用改进的残差神经网络对焊缝交点定位有着很好的性能,在20个未经训练的漏磁数据样本中,焊缝交点传感器编号与采样点坐标预测模型对采样点位置的预测误差在±20 mm以内,对传感器编号的预测误差在±84 mm以内。该模型可有效识别焊缝交点,满足工程要求,同时可加快对焊缝交点的识别速度,在管道内检测数据处理领域具有良好的应用前景。The accurate identification of the intersection between girth and spiral welds serves as a foundation for precisely lo‐cating target pipeline segments,highlighting the necessity for an efficient and intelligent identification method.To address the challenge of intelligent identification and localization of weld intersections in long-distance pipelines,a method based on an improved Residual Neural Network 18(ResNet18)was proposed.Starting from the characteristics of magnetic flux leak-age signals at weld intersections,this method integrated an Efficient Channel Attention(ECA)mechanism into ResNet18 to create a ResNet18-ECA.It analyzed the distribution of flux leakage fields at weld intersections,combining channel information and sensor number from manual interpretations,as well as the varying effects of axial,circumferential,and radial components of pipeline magnetic flux leakage data.Additionally,the Huber loss function and Leaky ReLU activation function replaced the original MSE loss function and ReLU activation function to enhance the localization of weld intersections.Weld intersection data sets were selected for 350 training rounds,and the trends in training loss and root mean squared error(RMSE)were compared.The test results demonstrated that the improved ResNet18 performed effectively in locating weld intersections.In 20 untrained magnetic flux leakage data samples,the prediction error for the sampling point position,as determined by the weld intersection sensor number and sampling point coordinate prediction model,was within±20 mm,while the prediction error for the sensor number was within±84 mm.This model effectively identifies weld intersections,meets engineering requirements,and accelerates the identification process,demonstrating strong potential for application in in-line inspection data processing.

关 键 词：漏磁检测 残差神经网络 焊缝交点 注意力机制 

分 类 号：TE832[石油与天然气工程—油气储运工程]