Related Experiment Video
Updated: Jun 13, 2025
06:50
O-cresol Concentration Online Measurement Based On Near Infrared Spectroscopy Via Partial Least Square Regression
Published on: November 8, 2019
6.6K
Low-Cost CO2 NDIR Sensors: Performance Evaluation and Calibration Using Machine Learning Techniques
Ravish Dubey1, Arina Telles2, James Nikkel2
1School of the Environment, Yale University, New Haven, CT 06511, USA.
Sensors (Basel, Switzerland)
|September 14, 2024
Summary
This study compared low-cost carbon dioxide (CO2) sensors, finding that higher-priced models offered better accuracy. Machine learning calibration significantly improved sensor performance, especially for affordable options.
Area of Science:
- Environmental monitoring
- Sensor technology
- Data science
Background:
- Accurate carbon dioxide (CO2) monitoring is crucial for environmental and climate studies.
- Low-cost CO2 sensors offer potential for widespread deployment but require performance validation.
- Existing research often lacks comprehensive intercomparison across diverse sensor price points.
Purpose of the Study:
- To evaluate the performance of various low-cost CO2 sensors against a reference instrument.
- To investigate the efficacy of machine learning techniques for calibrating these sensors.
- To assess the potential of machine learning to enhance the accuracy of affordable CO2 sensing technologies.
Main Methods:
- Comparative analysis of three CO2 sensors (Senseair Sunrise AB, Senseair K30, Vaisala GMP 343) against a Los Gatos precision greenhouse gas analyzer.
- Application of machine learning models, including linear regression, gradient boosting regression, and random forest regression, for sensor calibration.
- Development and evaluation of a stack ensemble model combining multiple machine learning approaches.
Main Results:
- Significant performance variations were observed among the tested CO2 sensors, with higher-cost Vaisala sensors showing superior accuracy.
- Lower-cost Senseair Sunrise sensors provided reasonable accuracy, while the K30 model exhibited higher noise and variability.
- Machine learning calibration, particularly a stack ensemble model, improved sensor accuracy by approximately 65%.
Conclusions:
- Machine learning offers a powerful tool for enhancing the accuracy of low-cost CO2 sensors, bridging the gap between affordability and reliability.
- The study provides valuable data for selecting and calibrating CO2 sensors across different price tiers for environmental applications.
- Further research into advanced machine learning algorithms can unlock greater potential for low-cost sensor networks.
