A novel multi-fidelity Gaussian process regression approach for defect characterization in motion-induced eddy current testing

Abstract

This study introduces a novel framework aimed at addressing the challenge of surface defect characterization in lab-scale tests. It utilizes a high-speed rotational disc setup to simulate the dynamics of rolling contact fatigue found in railway inspections through Motion-Induced Eddy Current Testing (MIECT). A key component of our approach was the integration of experimental data and finite element modeling, aimed at interpreting the relationship between defect dimensions, velocity, and their impact on magnetic sensor outputs. Our research focused on two main objectives: developing a forward model to predict the differential peak-to-peak amplitude (${\Delta V}_{pp}$) of sensor readings from defect size and velocity, and to perform inverse estimation of defect sizes from ${\Delta V}_{pp}$ across continuous velocity ranges. The key findings reveal that for the forward problem, the Radial Basis Function Multi-Fidelity Scaling (RBF-MFS) method outperforms other multi-fidelity and single-fidelity approaches. Moreover, the proposed Gaussian Process Regression with Multi-Fidelity Scaling and Feature Discretization (GPR-MFS-FD) method outperformed the state-of-the-art multi-fidelity method in the inverse estimation of defect geometries. This innovative method leverages high-fidelity experimental data together with low-fidelity physics simulations via multi-fidelity scaling and feature discretization to effectively manage velocity range inputs, reflecting real-world operational uncertainties in high-speed transport vehicles and infrastructures. Our integrated and novel data-driven approaches advance defect characterization, enhancing MIECT's application in surface defect detection and analysis, with potential extensions to other NDE applications.