Global Sensitivity Analysis of Eddy Current NDT System Based on SVR Model

The physics-based simulation model saves times and computational costs for the eddy current nondestructive testing (ECNDT) system to perform sensitivity analysis (SA) over experiments. However, the evaluation of SA by simulation model requires repeated calculations with a large number of uncertain parameters considered. The purpose of this paper is to reduce the cost of model-based SA for ECNDT systems by replacing accurate physics-based model with support vector regression (SVR) model to rapidly calculate Sobol indices. The accuracy of this model is evaluated based on the normalized root mean square error (NRMSE) at a fixed number of testing points. The proposed model is validated by the example of an eddy current NDT system with three uncertainty parameters selected. The numerical results show that the SVR model can be a good substitution of the physical simulation model for SA in ECNDT system.