Transfer Learning of Ultrasonic Guided Waves using Autoencoders: A Preliminary Study

In recent years, the use of scanning laser Doppler vibrometery and full wavefield acquisition has grown to aid the characterization of ultrasonic waves and the detection of structural detects. Yet, these methods require a considerable amount of time to acquire full wavefield data. Therefore, there is a significant need to reduce acquisition time. In this preliminary work, we present a transfer learning approach for reducing the number of sampled measurements necessary. Our method utilizes numerical simulations, combined with a small number of spatially sampled random measurements from an experimental structure, to reconstruct full wavefield data. Specifically, we use an autoencoder neural network to learn low-dimensional representations of wave propagation from numerical simulations. We then input a few experimental measurements into the neural network to reconstruct full wave field data. To demonstrate the ability of our framework, we show our initial success in three scenarios. We show reconstruction accuracies of 86% with one-fourth of the total measurements.