Influence of Concept Drift on Metrological Performance of Low-Cost NO2 Sensors

Low-cost air quality monitoring systems (LCAQMSs) and machine learning (ML) techniques are enabling a new paradigm in air quality monitoring networks. Nevertheless, compliance with data quality objective (DQO) is still an open point. The assessment of various calibration models proposed in the literature has ever neglected the concept drift, i.e., the differences in data distributions associated with input and target variables of the streaming data coming from dynamic nonstationary environments. The influence of the concept drift is investigated on the maintenance of the (calibrated) low-cost instrumentation. The data from mid-term co-location campaigns are first used to train a multiple linear regression (MLR) as calibration model. Then, an original methodology based on the two-sample Kolmogorov–Smirnov test (TSKS test) is proposed for automatically detecting the presence of the concept drift. The time evolution of the relative expanded uncertainty (REU) is considered as well, to highlight the negative influence of the concept shift on the metrological performance of the LCAQMS, proving the need for the estimation of a new calibration model during the maintenance of the instrumentation in order to match the DQOs. A quantitative analysis is carried out on the distances among the training and test distributions about the input and outputs of the calibration model by correlating them with the time evolution of uncertainty. The scheme of an add-on block based on the proposed approach is designed for the continuous monitoring of the metrological performance exhibited by the calibration model. This management of the concept drift is expected to allow the long-awaited achievement of DQOs.