基于深度学习神经网络的油气管道漏磁检测缺陷诊断  

Defect Diagnosis of Magnetic Flux Leakage Testing for Oil and Gas Pipeline Based on Deep Learning Neural Network

作　　者：张力凡 魏航信[1] Zhang Lifan;Wei Hangxin(School of Mechanical Engineering,Xi'an Shiyou University,Xi'an 710000,China)

机构地区：[1]西安石油大学机械工程学院,西安710000

出　　处：《机电工程技术》2023年第1期260-266,共7页Mechanical & Electrical Engineering Technology

基　　金：国家自然基金青年科学基金资助项目(编号:51405385);陕西省科技厅科技攻关项目(编号:2014K07-20)。

摘　　要：管道缺陷尺寸预测时,由于管道漏磁数据量庞大以及常规预测方法精度不高,提出一种基于迁移学习的改进型卷积神经网络与贝叶斯优化算法,以预测管道缺陷尺寸。将已训练好的Lenet网络进行迁移,对迁移后的网络结构进行重构,建立了TB-CNN改进型网络。提出了改进型神经网络的贝叶斯优化算法,对迁移学习改进后的模型的训练过程进行超参数优化,得到最优的TB-CNN模型。在制作训练集时,提出了格拉姆角场(GAF)算法将一维的漏磁曲线数据转换为二维图像方法。通过仿真及实验,证明了该网络对长度和宽度为10~40mm范围、深度在2~10mm范围的缺陷预测均方误差根(RMSE)为0.0691,和其他网络相比,该方法对管道缺陷尺寸的预测精度更高。When predicting the size of pipeline defects,due to the huge amount of pipeline magnetic flux leakage data and the low precision of conventional prediction methods,an improved convolution neural network and Bayesian optimization algorithm based on transfer learning was proposed to predict the size of pipeline defects.The trained Lenet network was migrated,the migrated network structure was reconstructed,and an improved TB-CNN network was established.The Bayesian optimization algorithm of improved neural network was proposed,and the training process of the improved model of transfer learning was super-parametric optimized to obtain the optimal TB-CNN model.When making the training set,the Gram Angular Field(GAF) algorithm was proposed to convert the one-dimensional magnetic flux leakage curve data into twodimensional images.Through simulation and experiment,it is proved that the RMSE of this network is 0.0691 for the defects with length and width ranging from 10 mm to 40 mm and depth ranging from 2 mm to 10 mm.Compared with other networks,this method has higher prediction accuracy for pipeline defects.

关 键 词：管道缺陷 卷积神经网络 迁移学习 贝叶斯优化 格拉姆角场 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程] TE973[自动化与计算机技术—控制科学与工程]