Into the latent space of capacitive sensors: interpolation and synthetic data generation using variational autoencoders

Abstract For many sensing applications, collecting a large experimental dataset could be a time-consuming and expensive task that can also hinder the implementation of Machine Learning models for analyzing sensor data. Therefore, this paper proposes the generation of synthetic signals through a Variational Autoencoder (VAE) to enlarge a spectra dataset acquired with a capacitive sensor based on a Dielectric Resonator. Trained with signals of several water/glycerine concentrations, this generative model learns the dataset characteristics and builds a representative latent space. Consequently, exploring this latent space is a critical task to control the generation of synthetic signals and interpolating concentrations unmeasured by the sensor. For this reason, this paper proposes a search method based on Bayesian Optimization that automatically explores the latent space. The results show excellent signal reconstruction quality, proving that the VAE architecture can successfully generate realistic synthetic signals from capacitive sensors. In addition, the proposed search method obtains a reasonable interpolation capability by finding latent encodings that generate signals related to the target glycerin concentrations. Moreover, this approach could be extended to other sensing technologies.