Convolutional Neural Network Models for Reconstruction and Sizing of Three-Dimensional Metal-Loss Defects in Oil and Gas Pipelines Based on Magnetic Flux Leakage Signals

In this study, a three-dimensional nonlinear finite element magnetostatic model is developed in the commercial package COMSOL Multiphysics (Version 5.6) to simulate magnetic flux leakage (MFL) signals for corroded pipelines. A large number of parametric finite element analysis cases covering wide ranges of locations and sizes of corrosions defects idealized as semi-ellipsoidal-shaped are used to produce the magnetic flux density signals. The white noises characterized by different signal-to-noise ratios are employed in the parametric analysis to represent the measurement errors in the MFL tool. The results of the parametric analyses are then used to train and validate a convolutional neural network (CNN) model to predict the location, depth, length and width of the corrosion defect simultaneously. The results indicate that the developed CNN model has a high predictive accuracy.