Local dynamic neural network for quantitative analysis of mixed gases

Abstract

The Gas sensor array is commonly used in combination with quantitative analysis method for detecting mixed gases. Artificial neural network (ANN) is usually employed to achieve a quantitative analysis of the mixed gases. However, the current ANN models typically require a large number of test samples to obtain low relative errors. In this work, we fabricated a gas sensor array composed of four gas sensors (i.e., In2O3: NO2, Pd-ZnO: NH3, Au-SnO2: CH4, Pd-LaFeO3: CO2) and proposed a local dynamic neural network (LDNN) model for quantitative analysis of four mixed gases. By constructing and extracting features through a pre-trained autoencoder network, only a small sample size (25 local points) is input into the LDNN for training to realize the concentration prediction of four gases. The results show that the MAEs (mean absolute error) of the predicted concentrations of NO2, NH3, CH4, and CO2 are 0.01 ppm, 0.04 ppm, 0.13 ppm, and 42.67 ppm, while the MREs (mean relative error) are 0.19%, 0.85%, 1.17%, and 1.06%, respectively. Moreover, the MREs of the predicted concentrations for four gases are less than 2% within the 95% confidence interval. This work provides an effective quantitative analysis method with small sample size, simple structure and high-precision for the detection of four mixed gases.