Calibration Assessment of Low-Cost Carbon Dioxide Sensors Using the Extremely Randomized Trees Algorithm

Area of Science:

• Environmental Science
• Sensor Technology
• Data Science

Background:

• Accurate carbon dioxide (CO2) monitoring is crucial for assessing indoor and outdoor air quality.
• Low-cost CO2 sensors often suffer from reduced accuracy, impacting data reliability.
• Existing calibration methods may not fully address the limitations of low-cost sensor technologies.

Purpose of the Study:

• To develop and validate a machine learning approach for enhancing the accuracy of low-cost CO2 sensors.
• To evaluate the proposed method's effectiveness across different sensor types (MOS, NDIR) and environments (indoor, outdoor).
• To investigate factors influencing calibration performance, such as sensor exposure time and data variety.

Main Methods:

• Utilized an extremely randomized trees algorithm for sensor calibration.
• Collected experimental data from metal oxide semiconductor (MOS) and non-dispersive infrared (NDIR) CO2 sensors in an indoor setting.
• Analyzed a third-party dataset comprising geographically distributed low-cost NDIR sensors for outdoor validation.

Main Results:

• Achieved significant reductions in mean absolute error (MAE) for NDIR sensors, up to 90% indoors and 98% outdoors.
• Demonstrated excellent linearity (r^2 value of 0.994) with MOS sensors in the indoor experiment.
• Identified sensor exposure time and data variety as important factors for effective calibration.

Conclusions:

• Machine learning, specifically the extremely randomized trees algorithm, effectively enhances the accuracy of low-cost CO2 sensors.
• The proposed calibration method improves data quality for both indoor and outdoor air quality monitoring applications.
• The approach is versatile, applicable to various low-cost gas sensor technologies and environmental conditions.