Full wavefield processing by using FCN for delamination detection

Abstract

A novel full wavefield processing method by using fully convolutional neural networks is presented. The full wavefield of propagating Lamb waves in the fibre-reinforced composite plate was simulated by the parallel spectral element method. It resembles a full wavefield measurements acquired on a surface of the plate by the scanning laser Doppler vibrometer. The aim of the proposed technique is an identification of delamination location, size and shape. It is achieved by pixel-wise image segmentation by using the end-to-end approach. It is possible because of the large dataset of Lamb wave propagation patterns resulting from interaction with delaminations of random location, size and shape. It is demonstrated that the proposed method, tested on numerical data, is performing better than conventional adaptive wavenumber filtering method which was developed in previous work. Moreover, it enables better automation of delamination identification so that the damage map can be created without user intervention. The method was also tested on experimental data acquired on the surface of the specimen in which delamination was artificially created by a Teflon insert. The obtained results with the deep learning approach show its capability to predict the delamination in the numerically generated dataset with high accuracy compared to the conventional damage detection approach. Furthermore, the deep learning model shows the ability to generalize to a further experiential set.