Sizing of metallic foam bubble flaws using direct current potential drop signals with the help of the neural network method

Abstract

To investigate the possibility of reconstructing the position and size of bubble flaws in metallic foam, a Neural Network approach is applied to predict the flaw profile from Direct Current Potential Drop (DCPD) signals. A feed-forward network, improved by Principal Component Analysis (PCA) is selected for the inverse analysis. Over 100 sets of DCPD signals due to flaws with different positions and sizes are calculated using a newly developed fast forward solver and are used for the inverse analysis. Satisfactory reconstruction results are obtained for these simulated signals.