Efficient field correction of low-cost particulate matter sensors using machine learning, mixed multiplicative/additive scaling and extended calibration inputs

Particulate matter (PM) stands out as a highly perilous form of atmospheric pollution, posing significant risks to human health by triggering or worsening numerous heart, brain, and lung ailments, and even increasing the likelihood of cancer and premature mortality. Therefore, ensuring accurate monitoring of PM levels holds paramount significance, particularly urban zones of dense population. Still, achieving precise readings of PM concentration demands the use of bulky and costly equipment, typically stationed at widely spaced reference sites. The rise in popularity of low-cost PM sensors as potential substitutes has been noted, although their reliability is hampered by manufacturing flaws, instability, and susceptibility to environmental variations. In this work, we introduce a novel approach to field calibration for cheap PM sensors. Our method integrates multiplicative and additive corrections, with coefficients determined by an artificial neural network (ANN) surrogate. The ANN model accounts for environmental parameters and the sensor's PM readings as inputs, with its architecture fine-tuned to ensure optimal generalization capability. Additionally, we consider an extended set of input parameters, including local temporal changes of environmental variables, and short sequences of low-sensor readings, to further enhance calibration reliability. We validate our technique using a non-stationary measurement equipment alongside reference data acquired by government-approved reference stations in Gdansk, Poland. The obtained values of coefficients of determination reach as high as 0.89 for PM₁, 0.87 for PM_2.5, and 0.77 for PM₁₀, respectively, while the root mean square error (RMSE) is merely 3.0, 3.9, and 4.9 µg/m³. Such a performance positions the calibrated low-cost sensor as a potential alternative to stationary measurement equipment.