Calibration and Inter-Unit Consistency Assessment of an Electrochemical Sensor System Using Machine Learning.

Area of Science:

• Environmental Science
• Sensor Technology
• Data Science

Background:

• Low-cost electrochemical sensors offer a scalable solution for atmospheric pollutant monitoring.
• Challenges including sensor drift, cross-sensitivity, and unit inconsistency impact data reliability.
• Accurate calibration is crucial for the effective deployment of these sensors.

Purpose of the Study:

• To evaluate three distinct calibration methods for low-cost electrochemical air quality sensors.
• To compare manufacturer-provided equations against machine learning (ML) approaches using raw voltage signals.
• To assess the performance enhancement of ML algorithms by leveraging sensor cross-sensitivity.

Main Methods:

• Experimental calibration of CO, NO, NO2, and O3 sensors across three urban sites in Greece.
• Utilized high-end instrumentation for reference concentration data.
• Implemented and compared three calibration strategies: manufacturer equations, ML with converted data, and ML with raw voltage signals.
• Employed the Random Forest ML algorithm for performance evaluation.

Main Results:

• Directly applying ML to voltage signals reduced variability between identical sensors compared to manufacturer equations.
• Calibration efficiency for CO, NO, NO2, and O3 sensors improved when using voltage signals.
• Integrating all sensor voltage signals into the ML model leveraged cross-sensitivity, further enhancing calibration accuracy.
• The Random Forest algorithm demonstrated robust performance for urban air quality sensor calibration.

Conclusions:

• Machine learning applied to raw voltage signals is a superior method for calibrating low-cost electrochemical air quality sensors.
• This approach enhances sensor reliability and accuracy, making them more suitable for widespread urban monitoring.
• The Random Forest algorithm shows significant promise for developing effective calibration models for similar sensor networks.