Low-Cost CO Sensor Calibration Using One Dimensional Convolutional Neural Network.

Area of Science:

• Environmental Science
• Sensor Technology
• Artificial Intelligence

Background:

• Cost-effective sensors and the Internet of Things (IoT) enable high-resolution urban pollution monitoring.
• Low-cost sensors often exhibit poor accuracy, necessitating effective calibration methods.

Purpose of the Study:

• To propose and evaluate a One Dimensional Convolutional Neural Network (1DCNN) for calibrating low-cost carbon monoxide sensors.
• To benchmark 1DCNN performance against other Machine Learning (ML) calibration techniques using real-world data.

Main Methods:

• Utilized three large datasets from field-deployed low-cost sensors co-located with reference sensors.
• Applied 1DCNN and other ML techniques (including gradient boosting regression) for sensor calibration.
• Investigated the impact of incorporating temperature, relative humidity, and cross-sensitivity data.

Main Results:

• 1DCNN demonstrated consistent performance across all tested datasets.
• Gradient boosting regression also showed promising results for gas sensor calibration.
• Including temperature and relative humidity significantly improved calibration accuracy.
• Exploiting cross-sensitivity to other pollutants further enhanced accuracy.

Conclusions:

• 1DCNN is a robust method for calibrating low-cost carbon monoxide sensors.
• Environmental factors like temperature and humidity are crucial for improving sensor accuracy.
• Deploying sensors in suites or arrays to capture covariate factors is recommended for comprehensive pollution monitoring.