Machine learning on synthetic acoustic data using adversarial domain adaptation technique

Acousti Resonance Spectroscopy (ARS) measurements are attractive because they are highly sensitive to deviations in the structure’s material, geometry, and environmental conditions. However, retrieving the property of interest is challenging due to the complicated changes to the spectra as the property of interest changes and sensitivity of the spectra to changes in other properties simultaneously. Deep learning on ARS measurements has shown to efficient and robust to identify trends that correlate with the property of interest. In many applications, it is infeasible to collect enough experimental ARS data to train a model so strategies to augment experimental data remain an open issue. Synthetic data from simulations are typically easy to acquire. Therefore, learning from synthetic data is very desirable in this context. We introduce a novel approach leveraging the Finite Element Method (FEM) framework to generate synthetic spectra in pressure estimation tasks. The FEM simulations were performed to mimic various vessel configurations at different pressures, representing 100% of the dataset used to train of the CNN model. As a crucial step, we implemented a domain adaptation technique to mitigate discrepancies between real and synthetic spectra.