Applicability of Data Augmentation Through Variational Autoencoder for Two-Dimensional Acoustic Emission Source Discrimination on Hollow Cylindrical Structures

Abstract

A data augmentation method is studied in this work to supplement previously proposed variational autoencoder (VAE) based acoustic emission (AE) source discrimination tool. The VAE model distinguishes the sources of collected AE signals through the multiple Lamb mode arrivals resulting from helical paths. It displayed the ability to discriminate pencil-lead-break (PLB) waveform dataset collected at a liquid nitrogen tank by source locations. VAE infers source-discriminative latent variable distribution conditioned on the observed waveforms and can be further applied to localization predictions for unseen waveforms. However, the prediction will be limited to source coordinates included in the training dataset. Therefore, this work studies the applicability of data augmentation using VAE to approximate waveform envelopes from sources coordinates that are not included in the PLB dataset, which are called target source in this study. Approximated waveforms of target source were introduced by interpolating the learned mode arrival characteristics to augment the original dataset. The augmented data set trained an updated latent variable distribution that accounted for the target source. Actual waveforms from target source were projected to updated latent space and validated the effectiveness of augment waveforms. This positive outcome provides confidence on pursuing data augmentation applications through VAE to assist AE localization.