Hyperband algorithm optimised CNN for accelerating MAPoD and sensitivity analysis in ECNDT

In this paper, a novel hyperparameter optimization algorithm (HOA) based convolutional neural network (CNN) is proposed to accelerate the study of the model-assisted probability of detection (MAPoD) and sensitivity analysis (SA) in eddy current nondestructive testing (ECNDT) system. MAPoD and SA are important methods to quantify the detection capability of NDT systems. However, a large number of model evaluations are required to propagate the uncertainties of the random input parameters, leading to extremely high computational costs, which makes it impossible to complete the MAPoD and SA studies of the ECNDT system within the specified time. To overcome this problem, in this work, CNNs are chosen to accelerate the MAPoD and SA studies, by partially replacing the accurate physical model. To optimize the selection of hyperparameters in CNN, Hyperband algorithm is proposed to replace manual trial and error and shows advantages over other HOA methods such as grid search, random search, and successive halving. Both the accuracy and efficiency of the proposed method are validated by 3D ECNDT benchmarks. Numerical results show that the proposed Hyperband optimized CNN algorithm has excellent performance over other HOA optimized ones with well-maintained accuracy for the study of uncertainty propagations in ECNDT..