High-performance machine-learning-based calibration of low-cost nitrogen dioxide sensor using environmental parameter differentials and global data scaling.

Accurate tracking of harmful gas concentrations is essential to swiftly and effectively execute measures that mitigate the risks linked to air pollution, specifically in reducing its impact on living conditions, the environment, and the economy. One such prevalent pollutant in urban settings is nitrogen dioxide (NO), generated from the combustion of fossil fuels in car engines, commercial manufacturing, and food processing. Its elevated levels have adverse effects on the human respiratory system, exacerbating asthma and potentially causing various lung diseases. However, precise monitoring of NO requires intricate and costly equipment, prompting the need for more affordable yet dependable alternatives. This paper introduces a new method for reliably calibrating cost-effective NO sensors by integrating machine learning with neural network surrogates, global data scaling, and an expanded set of correction model inputs. These inputs encompass differentials of environmental parameters (such as temperature, humidity, atmospheric pressure), as well as readings from both primary and supplementary low-cost NO detectors. The methodology was showcased using a purpose-built platform housing NO and environmental sensors, electronic control units, drivers, and a wireless communication module for data transmission. Comparative experiments utilized NO data acquired during a five-month measurement campaign in Gdansk, Poland, from three independent high-precision reference stations, and low-cost sensor data gathered by the portable measurement platforms at the same locations. The numerical experiments have been carried out using several calibration scenarios using various sets of calibration input, as well as enabling/disabling the use of differentials, global data scaling, and NO readings from the primary sensor. The results validate the remarkable correction quality, exhibiting a correlation coefficient exceeding 0.9 concerning reference data, with a root mean squared error below 3.2 µg/m. This level of performance positions the calibrated sensor as a dependable and cost-effective alternative to expensive stationary equipment for NO monitoring.