Time-Domain Signal Synthesis with Style-Based Generative Adversarial Networks Applied to Guided Waves

Abstract

Data scarcity is a significant problem when it comes to designing machine learning systems for structural health monitoring applications, especially those based around data-hungry algorithms and methods, such as deep learning. Synthetic data generation could potentially alleviate this problem, lowering the number of measurements that need to be acquired in slow and often expensive conventional experiments. Such synthesis can be done by Generative Adversarial Networks, potentially creating unlimited synthetic samples recreating the original data distribution. While most of the research about these networks is centered around using them on image data, they have also been applied to audio waves - going as far as successfully synthesizing human speech. This suggests that these networks should also apply to synthesizing time-domain signals in various fields of structural health monitoring, guided waves in particular, as they are in many ways similar to audio wave signals. This work proposes an adaptation of style-based GAN architecture to time-domain signal generation, and presents its viability for guided waves synthesis, utilizing a database of signals collected in series of pitch-catch experiments on a composite plate.