基于深度学习的漏磁检测缺陷识别方法

传统漏磁信号缺陷量化缺少其他分量信息,人工特征的提取方式造成信息量有限。为此,提出一种基于深度学习的漏磁检测缺陷量化识别方法,并建立了漏磁检测缺陷识别模型,该模型包含深度卷积神经网络模块和回归模块。深度卷积神经网络模块利用卷积神经网络的多输入多输出互相关操作,完成漏磁缺陷信号3个分量(轴向、周向、径向)的数据融合,利用预训练的网络,迁移已有知识,实现缺陷信号的特征自动提取;回归模块中设计缺陷、长度和宽度联合损失函数,利用回归方式实现缺陷尺度的量化。采用有限元仿真和牵拉试验相结合的方式,建立漏磁信号缺陷量化数据集并划分为训练集和测试集,训练集用于模型训练,测试集进行方法验证。研究结果表明:90%置信度下,长度和宽度量化结果全部落在±10 mm的误差带上,深度量化结果全部落在±10%t (t为壁厚)的误差带上,满足工程检测要求,可有效完成管道漏磁缺陷识别。研究结果可为油气输送管道漏磁检测新技术的研究提供一定的参考。

Magnetic Flux Leakage Defect Detection Based on Deep Learning

Traditional defect quantization of magnetic flux leakage lacks some components,and manual-extracted information is limited.This paper introduces a quantitative identification method based on deep learning and establishes a defect detection model for magnetic flux leakage,which includes a deep convolution neural network module and a regression module.The deep convolutional neural network module uses the multi-input and multi-output cross-correlation of the convolutional neural network to complete fusion of three components(axial,circumferential,and radial)of magnetic flux leakage signals,and uses a pre-trained network to transfer known knowledge to automatically extract the characteristics of defect signals.The regression module designs the joint loss function of defect,length and width,and uses the regression method to quantify the defect scale.With combination of finite element simulation and traction test,a quantitative data set of magnetic flux leakage signal defects is established and divided into a training set and a testing set.The training set is used to train models,and the testing set is used to verify methods.The results show that,at 90%confidence,the quantitative results of length and width fall within a±10 mm error range,and the quantitative depth is within a±10%t error range.These results meet engineering inspection requirements and can effectively identify magnetic flux leakage defects of pipelines.These findings are helpful to the research on new technologies of magnetic flux leakage detection of oil and gas pipelines.