【作者】 王伟；

【导师】 曹建树；

【作者基本信息】 北京石油化工学院 ， 机械工程， 2025， 硕士

【摘要】 油气管道的安全性随着油气资源的开发利用愈发重要,其完整性与管道系统的安全运行有着密切关系。打孔盗油行为破坏管道,遗留的孔洞缺陷隐蔽性高,包含诸多隐患。研究管道缺陷的相关参数对涡流阻抗信号的影响,可以实现对管道缺陷的定量识别表征。柔性阵列涡流检测是一种利用激励线圈形成的磁场在被检测的导体内产生电涡流,并通过多组检测线圈检测电涡流分布和大小变化的无损检测方法。此方法既可以检测导体内部的缺陷,也可以检测导体表面的缺陷,适用于管道缺陷的检测。本文为实现管道缺陷定量识别的目标,研究将柔性阵列涡流用于管道缺陷电磁和几何参数的无损检测过程。论文的主要内容和结果如下:(1)针对管道缺陷难以定量化识别的问题,建立基于柔性阵列涡流的多角度关联模型。通过COMSOL软件建立了管道缺陷有限元仿真模型,研究缺陷半径和缺陷深度变化对检测线圈阻抗信号产生的影响。结果显示线圈阻抗整体信号幅值随电导率的增加产生复杂波动,随着缺陷半径的增大而增大,随着缺陷深度的增大而减小。设计验证实验,研究实际缺陷半径与深度变化对线圈阻抗信号幅值的影响。对比发现实验结果与仿真结果一致,揭示阻抗信号与缺陷参数的非线性映射规律。(2)针对线圈阻抗信号幅值信息单一的问题,提出物理-数据双驱动混合反演框架。利用频域滤波和动态压缩优化阻抗信号幅值,采用变分模态分解和主成分分析得到关键时频特征,以均方根、峰度和频谱质心为主,反映阻抗信号的总体信息。构建加入关键时频特征作为物理约束的混合反演模型,设计物理信息神经网络的损失函数。发现物理信息神经网络与混合模型天然适配,通过物理约束和数据驱动的互补融合,促进多参数耦合反演。(3)针对管道缺陷参数预测难的问题,基于物理信息神经网络搭建并训练了预测模型。扩充用于训练和测试的数据集,提升模型的训练速度和鲁棒性。提出调节损失函数中用于平衡数据驱动和物理解析比重的超参数的方法,研究超参数变化对模型的影响。采用多种评价标准对模型训练结果进行评价。结果表明,模型的预测性能良好。更多还原

【Abstract】 The safety of oil and gas pipelines has become increasingly important with the development and utilization of oil and gas resources,and its integrity is closely related to the safe operation of the pipeline system.The act of drilling holes to steal oil damages pipelines,and the remaining hole defects are highly concealed and contain many hidden dangers.Studying the influence of relevant parameters of pipeline defects on the eddy current impedance signal can achieve quantitative identification and characterization of pipeline defects.Flexible array eddy current testing is a non-destructive testing method that uses the magnetic field formed by the excitation coil to generate eddy currents in the conductor under test,and detects the distribution and magnitude changes of the eddy currents through multiple sets of detection coils.This method can detect both internal and surface defects of conductors and is suitable for the detection of pipeline defects.In order to achieve the goal of quantitative identification of pipeline defects,this paper studies the application of flexible array eddy currents in the non-destructive testing process of electromagnetic and geometric parameters of pipeline defects.The main content and results of the thesis are as follows:(1)To address the problem of difficulty in quantitative identification of pipeline defects,a multi-angle correlation model based on flexible array eddy current is established.A finite element simulation model of pipeline defects is established using COMSOL software to study the influence of changes in defect radius and depth on the impedance signal of the detection coil.The results show that the overall signal amplitude of the coil impedance fluctuates complexly with the increase of electrical conductivity,increases with the increase of defect radius,and decreases with the increase of defect depth.A verification experiment is designed to study the influence of actual defect radius and depth changes on the amplitude of the coil impedance signal.The comparison reveals that the experimental results are consistent with the simulation results,revealing the nonlinear mapping relationship between the impedance signal and the defect parameters.(2)To address the problem of the single amplitude information of the coil impedance signal,a physical-data dual-driven hybrid inversion framework is proposed.Frequency domain filtering and dynamic compression are used to optimize the amplitude of the impedance signal,and variational mode decomposition and principal component analysis are used to obtain key time-frequency features,mainly root mean square,kurtosis,and spectral centroid,reflecting the overall information of the impedance signal.A hybrid inversion model incorporating key time-frequency features as physical constraints is constructed,and the loss function of the physical information neural network is designed.It is found that the physical information neural network is naturally compatible with the hybrid model,and the complementary fusion of physical constraints and data-driven promotes multi-parameter coupled inversion.(3)To address the problem of difficulty in predicting pipeline defect parameters,a prediction model based on physical information neural network is built and trained.The data set used for training and testing is expanded to improve the training speed and robustness of the model.A method for adjusting the hyperparameter used to balance the proportion of data-driven and physical analysis in the loss function is proposed,and the influence of hyperparameter changes on the model is studied.Multiple evaluation criteria are used to evaluate the training results of the model.The results show that the prediction performance of the model is good.更多还原