STATION CLUSTERING FOR DETECTING DATA INSTABILITY IN AN AIR QUALITY MONITORING NETWORK

Abstract

This paper proposes and validates an approach for automated detection of data instability in an atmospheric air quality monitoring network using data mining methods. Unlike traditional threshold-based checks, the approach focuses on behavioral features describing the “measurement stream quality” (completeness, missingness rate, signal variability, and sensor “sticking” indicators) computed from hourly aggregates. The clustering objects are “station–sensor” pairs, which enables localization of issues both at the station level and at the level of individual measurement channels. K-Means clustering is applied with prior feature scaling; the optimal number of clusters is selected using the elbow method and the silhouette coefficient. For cluster interpretation, a projection onto two principal components is used, reflecting a data availability/incompleteness index and a signal dynamics index (variability versus “sticking”). Experiments on real-world data reveal stable degradation profiles of measurements and allow identification of reliable stations and problematic channels (in particular, sensors with a high missingness rate or near-zero hourly variation). The practical value of the study lies in the ability to integrate the proposed method into environmental information-and-analytical systems as a data quality control module, and to further use the results for selecting reference sensors, calibration, and building predictive models.