On Memory-Based Precise Calibration of Cost-Efficient NO2 Sensor Using Artificial Intelligence and Global Response Correction

Abstract

Nitrogen dioxide (NO2) is a prevalent air pollutant, particularly abundant in densely populated urban regions. Given its harmful impact on health and the environment, precise real-time monitoring of NO2 concentration is crucial, particularly for devising and executing risk mitigation strategies. However, achieving precise measurements of NO2 is challenging due to the need for expensive and cumbersome equipment. This has spurred the development of more affordable alternatives, although their reliability is often uncertain. The aim of this study is to present a new method for accurately calibrating low-cost NO2 sensors. Our approach utilizes artificial intelligence techniques, particularly neural networks (NNs) acting as surrogates, trained to forecast sensor correction coefficients. These predictions rely on environmental variables (temperature, humidity, etc.), data from additional nitrogen dioxide sensors, and a short series of previous NO2 readings from the main sensor, all serving as inputs for the NN metamodel. As shown, integrating short-time-scale previous measurements significantly improves the quality of the calibration process, further bolstered by global response correction. Similar enhancements are achieved by considering environmental parameter differentials. Our calibration approach has been validated using a custom-built, cost-efficient monitoring platform and reference data collected over five months from high-performance public stations in Gdansk, Poland. The results demonstrate outstanding correction quality, with a correlation coefficient close to 0.93 compared to the reference data and an RMSE below 2.8 µg/m3. This establishes the calibrated sensor as a practical and cost-effective alternative to expensive traditional NO2 monitoring stations.