Variational Autoencoder for Calibration: A New Approach

TL;DR

The paper tackles sensor calibration for low-cost gas sensors by introducing a VAE whose latent space is trained to output calibration values, while preserving input reconstruction. The method uses a 4-input, 1D latent calibration output and a loss that combines reconstruction and calibration terms, evaluated on a multi-sensor MOS dataset with ground-truth reference measurements. Key contributions include the architectural design that binds calibration outputs to the latent space and a demonstration showing comparable calibration performance to existing methods (e.g., MAE for CO around 0.29) alongside faithful reconstructions. This work suggests that the latent space can encode transfer-function properties of sensors, offering a unified framework for calibration and drift adaptation that can be extended to additional datasets and loss formulations.

Abstract

In this paper we present a new implementation of a Variational Autoencoder (VAE) for the calibration of sensors. We propose that the VAE can be used to calibrate sensor data by training the latent space as a calibration output. We discuss this new approach and show a proof-of-concept using an existing multi-sensor gas dataset. We show the performance of the proposed calibration VAE and found that it was capable of performing as calibration model while performing as an autoencoder simultaneously. Additionally, these models have shown that they are capable of creating statistically similar outputs from both the calibration output as well as the reconstruction output to their respective truth data. We then discuss the methods of future testing and planned expansion of this work.

Figures (5)

• Figure 1: Proposed VAE Calibration Model Architecture where $X\approx X'$ and $Y\approx Y'$. The input to the model is represented by $X$ and $X'$ is the reconstruction, while $Y$ represents the truth value and $Y'$ represents the prediction.
• Figure 2: Correlation Coefficient Heatmap of Matched Test Dataset
• Figure 3: Plot showing the predictions of the CO VAE and truth data from CO(GT) sensor as seen in Table \ref{['tab:CO_features_metrics_summary']}.
• Figure 4: Plot showing the reconstruction of the input sensor data produced by the CO VAE as seen in Table \ref{['tab:CO_features_metrics_summary']}.
• Figure 5: Histogram of 4 separate components from the NO$_2$ model test. In the top row we see the truth data ($Y$) distribution from the "NO2(GT)" sensor and the model predictions ($Y'$) from the latent space. In the bottom row we see the "PT08.S4(NO2)" MOS input data (a component of $X$) distribution as well as the reconstruction data (a component of $X'$) distribution produced by the model. These symbols are in reference to Figure \ref{['fig:Vae_prop']}.