Learning-Based Predictive Uncertainty Estimation of Magnetic Flux Leakage Data for Parametric Defect Classification

Magnetic flux leakage (MFL), one of the most popular electromagnetic-based Nondestructive Evaluation (NDE) methods, is an important inspection technique for pipeline safety to prevent long-term failures. In real-life field testing or inspection scenario, there exist lots of associated uncertainties that will affect damage condition-based decision-making, therefore, it is vital to address and quantify the involved uncertainty for ensuring the reliability of inspection. This paper investigates the effect of uncertainties in the dynamic magnetization process due to the relative motion of a magnetic flux leakage (MFL) sensor and the material under test in the axial and circumferential directions. During the inspection, the roughness of the surface of the measured material is a main source to affect sensor liftoff and is considered as one of the important uncertainty sources affecting the inspection results. Therefore, in this work, the uncertainties from sensing liftoff are investigated, which is propagated throughout the sensing system to affect the output data. Considering the complexity of describing the forward uncertainty propagation process, Deep Ensemble, a learning-based non-Bayesian uncertainty estimation method, is applied to address the input uncertainty from the response MFL data. For performance evaluation, a three-dimensional finite element method (FEM) based model is used to generate simulation data for MFL based defect depth classification, while experiment data are validated in MFL based defect size classification. Prediction accuracy and uncertainty with calibration are conducted which is valuable in assessing the prediction performance and quantifying uncertainties. Further, an autoencoding method is applied for tackling the lack of experimental data for training model, which is extended to address the bottleneck of insufficient experimental data in generalized NDE problems.