We aim to investigate a diffusive non-dispersive infrared (NDIR) CO₂ sensor, with a focus on improving its measurement accuracy and stability. Accurate measurement of CO₂ volume fraction is crucial for environmental monitoring and biomedical research. While NDIR-based CO₂ sensors offer advantages such as fast response times and high accuracy, they are susceptible to measurement deviations due to ambient temperature variations. We aim to address this issue by proposing a temperature compensation method using unit surface fitting to enhance the sensor’s performance under varying temperature conditions.

We design and implement a diffusive NDIR CO₂ sensor, which utilizes the absorption characteristics of CO₂ in the infrared spectrum. The sensor employs a dual-channel pyroelectric detector and a micro-electromechanical system (MEMS) infrared light source, with a measurement channel centered at (4.26±0.05) μm and a reference channel at (4.00±0.08) μm. To mitigate interference from water vapor, a sapphire window and a waterproof breathable membrane are used in the sensor probe structure. The proposed temperature compensation method is based on unit surface fitting. The volume fraction of CO₂ is modeled as a function of signal absorption and temperature, with the fitting process dividing the volume fraction-temperature surface into smaller units. Each unit is fitted using a cubic polynomial, and the volume fraction is calculated by searching for the intersection of lines representing constant volume fraction across different temperatures.

The results demonstrate that the proposed unit surface fitting method significantly improves the accuracy of CO₂ volume fraction measurements across varying temperatures. In experiments conducted at 37 ℃, the maximum relative error is 0.99% for a CO₂ volume fraction of 10000×10^-6 (Fig. 7). When measuring a non-characteristic volume fraction of 10000×10^-6 across different temperatures, the maximum relative error is 1.36% at 15 ℃ (Fig. 7). The method outperforms other compensation algorithms, such as the improved sparrow search algorithm optimized backpropagation (ISSABP) neural network and linear interpolation, with a mean absolute error (MAE) of 107.85, mean relative error (MRE) of 0.0036, and root mean square error (RMSE) of 186.18. The coefficient of determination (R2) is 0.999993 (Table 2), which indicates a high level of accuracy and reliability.

The proposed unit surface fitting method effectively compensates for temperature-induced errors in diffusive NDIR CO₂ sensors. This method enhances the sensor’s accuracy and reliability across a wide range of temperatures and volume fractions, with maximum relative errors well below the industry standard of 2%. Our study offers a novel and practical solution for improving the performance of open-path CO₂ sensors, thus contributing to advancements in environmental monitoring and biomedical applications.