A fast calibration algorithm for Non-Dispersive Infrared single channel carbon dioxide sensor based on deep learning

Abstract

As people pay more attention to environmental monitoring, Carbon dioxide (CO2) sensors are widely used. However, most of the infrared CO2 single-channel sensors are accompanied by low calibration efficiency and low accuracy. In order to save costs while improving calibration efficiency and accuracy, we proposed a fast calibration algorithm for Non-Dispersive Infrared (NDIR) single-channel carbon dioxide sensor based on deep learning. Firstly, we establish N network models which consist of N sensors by collecting m data points from different temperatures and concentrations. Secondly, we collect six data points from a new sensor which are measured at three temperatures and two concentrations. Thirdly, we choose multiple approximate models from N network models based on the matching of the data points. At last, we regard these models as the estimation model of the new sensor to calibrate the sensor concentration. This method eliminates the individual differences of a single model to a certain extent and achieves the purpose of rapid calibration. After comparing three kinds of neural networks and conducting relevant experiments, we chose BP neural network as the model, and set the number of selected models to three. The results show that the floating up and down by industry-standard 5% plus or minus 50 ppm calculation, the qualified rate of our method is up to 91.542% between 0 °C to 45 °C, and the qualified rate even reaches 99.063% between 20 °C to 35 °C. Compared with similar products, the qualified rate of our method in the calibration of carbon dioxide increases by 12.315% and 22.732% respectively.