摘要：The pipelines＇ in-line inspection (ILI) is critical within the integrity management framework in the oil and gas industry. Furthermore, the reconstruction of defects＇ three-dimensional (3-D) profile using the magnetic flux leakage (MFL) signals acquired has great significance. However, most existing methods only focus on estimating defect sizes or shape parameters instead of the defect＇s 3-D profile. This study proposes an innovative approach for reconstructing the defect profile using a novel hybrid neural network to accurately and efficiently map three-axial MFL signals to the defects＇ 3-D profile. This paper utilizes the neural ordinary differential equation (ODE) as a module within the neural network architecture. The neural ODE is used to map the processed MFL signals to the spatial position of each point on the defective concave surface. Additionally, the model incorporates the Fourier integration kernel (FIK) to enhance computational efficiency. The proposed model is trained using finite element (FE) simulation data and then transferred to an experimental dataset, which addresses the challenge of limited availability of experimental data while maintaining accuracy. Furthermore, the proposed method also exhibits a high degree of accuracy in reconstructing the rotational angles of the defects. Therefore, the proposed method helps visualize defects in underground pipes via the analysis of MFL signals, facilitating operators in undertaking subsequent maintenance measures and providing a foundation for pipeline digital integrity management. ? 2025 The Author(s)

ISSN号：0951-8320

卷、期、页：卷258

发表日期：2025-06-01

期刊分区（SCI为中科院分区）：一区

收录情况：SCI（科学引文索引印刷版）,EI（工程索引）,SCIE（科学引文索引网络版）

发表期刊名称：Reliability Engineering and System Safety

参与作者：陈一诺,田志刚

通讯作者：魏昊天

第一作者：董绍华

论文类型：期刊论文

论文概要：陈一诺,田志刚,魏昊天,董绍华，Reconstruction of 3-D pipeline defect profile based on MFL signals and hybrid neural networks，Reliability Engineering and System Safety，2025，卷258

论文题目：Reconstruction of 3-D pipeline defect profile based on MFL signals and hybrid neural networks