An improved low power measurement of ambient NO<sub>2</sub> and O<sub>3</sub> combining electrochemical sensor clusters and machine learning

Abstract

Abstract. Low cost sensors (LCS) are an appealing solution to the problem of spatial resolution in air quality measurement, but they currently do not have the same analytical performance as regulatory reference methods. Individual sensors can be susceptible to analytical cross interferences, have random signal variability and experience drift over short, medium and long timescales. To overcome some of the performance limitations of individual sensors we use a clustering approach using the instantaneous median signal from six identical electrochemical sensors to minimise the randomised drifts and inter-sensor differences. We report here a low power analytical device (

Performance comparison between electrochemical and semiconductors sensors for the monitoring of O<sub>3</sub>

“…So for ozone air quality studies, given the linearity properties, the reproducibility of the measurements between the different sensors and the low dependence on temperature variations, we recommend the use of electrochemical sensors. However the high cross-sensitivity between ozone and NO2 for these sensors (Spinelle et al, 2013(Spinelle et al, , 2014(Spinelle et al, , 2015a, (Smith et al, 2018) requires simultaneous measurements of ozone and NO2 in the field using O3/NO2 35 sensors (Alphasense Ltd., 2019a) and NO2 (Alphasense Ltd., 2019b). This NO2 sensor uses an ozone filter to measure only NO2 gas.…”

“…The high linearity and time-limited drift of electrochemical sensors (Spinelle et al, 2014(Spinelle et al, , 2015a, led us to test the performance of these sensors for use with our future mobile platform. However, these electrochemical sensors have a high cross-sensitivity with ozone and NO2 (Spinelle et al, 2013(Spinelle et al, , 2014(Spinelle et al, , 2015aSmith et al, 2018). So we choice to compare these performances with metal oxide ozone sensors whose linearity is less good but which are not very sensitive to other gases (Spinelle et al, 2016).…”

“…The performance of several field calibration methods for low-cost metal oxide and electrochemical sensors are evaluated (Spinelle et al, 2015b(Spinelle et al, , 2017. One study shows that supervised learning techniques can be more effective than linear or multilinear regression methods to estimate the concentration of a gas (Smith et al, 2018).…”

Long-term behavior and stability of calibration models for NO and NO<sub>2</sub> low-cost sensors

“…Laboratory calibration against reference material often has major drawbacks, and calibration of LCSs is therefore predominantly done based on co-location with reference instruments operated in traditional air quality monitoring stations. Thus, the sensor output and other relevant environmental variables (e.g., temperature and relative humidity) are related to the true concentration values as represented by the reference measurement in parametric (e.g., Jiao et al, 2016;Kim et al, 2018;Malings et al, 2019) and nonparametric regression models (e.g., Cross et al, 2017;Hagan et al, 2018) or by using machine learning techniques (Bigi et al, 2018;Smith et al, 2019). The obtained (mathematical) relationship forms a calibration model that can be used for converting the raw sensor data into a concentration of the air pollutant to be measured.…”

First-Principles Algorithm for Air Quality Electrochemical Gas Sensors

“…To eliminate such effects, models of varying levels of complexities have been exploited, including univariable to multivariable linear regression methods that aim to describe the effects of T and RH through linear to higher-order terms [17][18][19][20][21] . More recent studies also speculated that sensor response to T and RH may be very complex, occurring through processes that can be nonlinear, and have applied more advanced algorithms such as the Random Forest Regression and others based on Machine Learning 19, [22][23][24] , Artificial Neural Networks [25][26][27] , or the High-Dimensional Model Representation approach designed by Aerodyne and Princeton 28 . Mapping schemes (that link sensor raw reading with concentration) afforded by these above approaches can be unamenable for interpretation, may include terms that are noncausal 29 , and often offer limited insights into the actual governing physical parameters.…”