An Iterative Stacking Method for Pipeline Defect Inversion With Complex MFL Signals

Abstract

Magnetic flux leakage (MFL) inspection in nondestructive testing (NDT) has been widely used in damaged pipeline defect inversion. The changeable environment and the complexity of MFL signal have brought severe challenges to the accurate estimation of defect sizes in inversion issue. This article proposes a novel pipeline defect inversion method (WT-STACK) based on stacking learning. This method consists of two parts. First, a multi-domain feature extraction with three-axis (axial, radial, and circumferential) signals is constructed. To avoid the feature information loss, signals are analyzed both in time and frequency domains. Second, to study the complex nonlinear relationship between the feature and defect size, an iterative stacking estimation network is developed with dynamic multi-domain features input. An adaptive learning is realized in the network, which enhances the generalization ability for different sample sets of defect inversion issue. Finally, the method is evaluated by experiments using MFL signals collected from experimental platform and simulation signals. Experimental results and comprehensive comparison analysis with other state-of-art methods validate the superiority of this method.