FEM of magnetic flux leakage signal for uncertainty estimation in crack depth classification using bayesian convolutional neural network and deep ensemble

This paper investigates the impact of uncertainties during dynamic magnetization due to relative motion between magnetic flux leakage (MFL) sensors and material under test. A 3D finite element method (FEM) based model for the MFL inspection system is developed to better understand the relationships between inline inspection (ILI) data and damage parameters in nonlinear ferromagnetic materials. In real-life field testing or inspection scenario, there exist lots of uncertainties associated with nondestructive evaluation (NDE), which will further affect damage condition-based decision making. Therefore, it is vital to quantitatively analyze the uncertainties in NDE. This paper investigates uncertainty quantification via numerical simulation by developing Bayesian based Convolutional Neural Network and Deep Ensemble methods, which are demonstrated in MFL based defect depth classification for NDE applications. Prediction accuracy and uncertainties are compared, which is valuable in better assessing the performance and quantifying uncertainties for generalized NDE problems.