Robust statistical calibration and characterization of portable low-cost air quality monitoring sensors to quantify real-time O3 and NO2 concentrations in diverse environments

Abstract

Abstract. Rising awareness of the health risks posed by elevated levels of ground-level O3 and NO2 have led to an increased demand for affordable and dense spatio-temporal air quality monitoring networks. Low-cost sensors used as a part of Internet of Things (IoT) platforms offer an attractive solution with greater mobility and lower maintenance costs, and can supplement compliance regulatory monitoring stations. These commodity low-cost sensors have reasonably high accuracy but require in-field calibration to improve precision. In this paper, we report the results of a deployment and calibration study on a network of seven air quality monitoring devices built using the Alphasense O3 (OX-B431) and NO2 (NO2-B43F) electrochemical gas sensors. The sensors were deployed at sites situated within two mega-cities with diverse geographical, meteorological and air quality parameters – Faridabad (Delhi National Capital Region) and Mumbai, India. The deployment was done in two phases over a period of three months. A unique feature of our deployment is a swap-out experiment wherein four of these sensors were relocated to different sites in the two deployment phases. Such a diverse deployment with sensors switching places gives us a unique opportunity to ablate the effect of seasonal, as well as geographical variations on calibration performance. We perform an extensive study of more than a dozen parametric as well as non-parametric calibration algorithms and find local calibration methods to offer the best performance. We propose a novel local calibration algorithm based on metric-learning that offers, across deployment sites and phases, an average R2 coefficient of 0.873 with respect to reference values for O3 calibration and 0.886 for NO2 calibration. This represents an upto 9\u2009% increase over R2 values offered by classical local calibration methods. In particular, our proposed model far outperforms the default calibration models offered by the gas sensor manufacturer. We also offer a critical analysis of the effect of various data preparation and model design choices on calibration performance. The key recommendations emerging out of this study include (1) incorporating ambient relative humidity and temperature as free parameters (or features) into all calibration models, (2) assessing the relative importance of various features with respect to the calibration task at hand, by using an appropriate feature weighing or metric learning technique, (3) the use of local (or even hyper-local) calibration techniques such as k-NN that seem to offer the best performance in high variability conditions such as those encountered in field deployments, (4) performing temporal smoothing over raw time series data, say by averaging sensor signals over small windows, but being careful to not do so too aggressively, and (5) making all efforts at ensuring that data with enough diversity is demonstrated to the calibration algorithm while training to ensure good generalization. These results offer insights into the strengths and limitations of these sensors, and offer an encouraging opportunity at using them to supplement and densify compliance regulatory monitoring networks.