Three‐dimensional defect inversion from magnetic flux leakage signals using iterative neural network

Abstract

Defect inversion is of special interest to magnetic flux leakage (MFL) inspection in industry. This study proposes an iterative neural network to reconstruct three-dimensional defect profiles from three-axial MFL signals in pipeline inspection. A radial basis function neural network is utilised as the forward model to predict the MFL signals given a defect profile, and the defect profile gets updated based on a combination of gradient descent and simulated annealing in the iterative inversion procedure. Accuracy of the proposed inversion procedure is demonstrated in estimating the profile of different defects in steel pipes. Experimental result based on three-axial simulated MFL data also shows that the proposed inversion approach is robust even in presence of reasonable noise.