Modeling Errors in Hazardous Gas Sensors: A Comparative Study of Machine Learning Approaches Using Environmental and Pollutant Variables

This study addresses the measurement errors commonly found in low-cost carbon monoxide (CO) sensors used in practical applications. A multi-source variable modeling approach is proposed to analyze how environmental factors and co-occurring pollutants contribute to sensor reading deviations. Seven feature input groups are designed, incorporating primary variables, pollutant-related variables, environmental factors, and their combinations. Six representative regression models are evaluated using a consistent train-test split and a time-aware cross-validation scheme to ensure rigorous and temporally informed performance evaluation. The results show that pollutant-related features contribute more significantly to error modeling, while environmental variables such as temperature and humidity also offer supplementary value in certain models. Model sensitivity varies: XGBoost demonstrates robust performance, whereas some nonlinear models tend to overfit under high-dimensional combinations. Based on the findings, several recommendations are made to improve sensor system design, including enhancing key variable acquisition, reducing redundant signal interference, and improving system stability. The proposed approach offers practical insights for sensor data modeling and acquisition strategy development.