Concept Drift Mitigation in Low-Cost Air Quality Monitoring Networks.

Area of Science:

• Environmental Science
• Sensor Technology
• Machine Learning Applications

Background:

• Future air quality monitoring networks will rely on low-cost sensors calibrated with machine learning.
• Concept drift is a major challenge, causing data quality degradation in operational machine learning models.
• Accurate calibration is essential to meet European Directive's Relative Expanded Uncertainty (REU) limits.

Purpose of the Study:

• To address the calibration model update for low-cost NO2 sensors after concept drift detection.
• To identify optimal data for model updating to ensure compliance with REU limits.
• To evaluate the effectiveness of different calibration models and ensemble strategies in mitigating concept drift.

Main Methods:

• Investigated general/global and importance weighting calibration models for concept drift mitigation.
• Applied a stacking ensemble strategy combining both calibration models.
• Evaluated model performance based on extending the temporal validity of calibration models.

Main Results:

• Independently, neither the general/global nor the importance weighting models were adequate.
• The stacking ensemble strategy significantly extended the temporal validity of the calibration model by at least three weeks for all tested sensors.
• This approach maximized the utility of data gathered during the initial co-location process.

Conclusions:

• A stacking ensemble strategy effectively mitigates concept drift in low-cost NO2 sensor calibration.
• This method enhances the reliability and longevity of sensor calibration models.
• The findings support the integration of advanced machine learning techniques for robust air quality monitoring.