Variational Bayesian calibration of low-cost gas sensor systems in air quality monitoring

Abstract

Due to the impact of air quality on health, the use of low-cost gas sensor systems in air quality monitoring has increased. The deficiencies of low-cost gas sensors such as cross-sensitivities, interferences with environmental factors, and unit-to-unit variability have led to multidimensional calibration approaches with machine learning applied to field data. However, the trustworthiness of measurement results remains a concern, since sensor aging and concept drift are common problems. While most machine learning algorithms do not disclose the reliability and trustworthiness of measurement results, Bayesian models offer integrated sanity checking of predictions. Motivated by recent advancements in variational inference, this publication evaluates the potential of variational Bayesian linear regression and variational Bayesian neural networks for the calibration of low-cost gas sensors and sensor systems using laboratory as well as published field data. With the performed analysis, it can be shown that raw sensor measurements outside confidence regions, implicitly defined by the calibration input data, are detectable. In such situations, the uncertainty of the posterior predictive distribution increases, suggesting less trustworthy measurement results and necessity for maintenance.