As key components of low-cost sensor systems in air quality monitoring, electrochemical gas sensors have recently received a lot of interest but suffer from unit-to-unit variability and different drift components such as aging and concept drift, depending on the calibration approach. Magnitudes of drift can vary across sensors of the same type, and uniform recalibration intervals might lead to insufficient performance for some sensors. This publication evaluates the opportunity to perform predictive maintenance solely by the use of calibration data, thereby detecting the optimal moment for recalibration and improving recalibration intervals and measurement results. Specifically, the idea is to define confidence regions around the calibration data and to monitor the relative position of incoming sensor signals during operation. The emphasis lies on four algorithms from unsupervised anomaly detection-namely, robust covariance, local outlier factor, one-class support vector machine, and isolation forest. Moreover, the behavior of unit-to-unit variability and various drift components on the performance of the algorithms is discussed by analyzing published field experiments and by performing Monte Carlo simulations based on sensing and aging models. Although unsupervised anomaly detection on calibration data can disclose the reliability of measurement results, simulation results suggest that this does not translate to every sensor system due to unfavorable arrangements of baseline drifts paired with sensitivity drift.
Keywords: air quality monitoring; anomaly detection; gas sensor; low-cost sensors; machine learning; predictive maintenance.