Multimodal Data Analysis for Evaluating Hydrogen Diffusion in Steel

Abstract

Multimodal data analysis provides useful information that is not generally obtained from one of the analysis methods. In this study, time-course images of hydrogen distribution on a steel sample measured using electron stimulated desorption (ESD), scanning electron microscopy (SEM) images and electron backscatter diffraction (EBSD) images were fused to create a multimodal image data set. The fused multimodal images were analyzed by principal component analysis, least absolute shrinkage and selection operator (LASSO) and autoencoder. Each method is one of the most popular methods in each field, multivariate analysis, sparse modeling, and unsupervised learning based on artificial neural networks, respectively. The results of PCA, LASSO and autoencoder were consistent, and each method provides different aspects of the sample data information.