Simultaneous inversion of methyl thiol, methane and water vapor concentration from wavelength modulation spectroscopy using neural network

Methyl thiol(CH₃SH) is a colorless, flammable, smelly toxic gas. A large amount of methyl thiol gas comes from organic chemicals and agricultural production, and its olfactory threshold is very low (2ppb), which not only causes pollution to the environment, but also seriously affects people's quality of life and physical health. At the same time, methyl thiol also belongs to the human body of the endogenous trace volatile organic compounds (VOCs), and periodontitis disease is closely related. Therefore, the detection of methyl thiol is very important for the assessment of the environmental atmosphere and exhalation analysis. In this paper, the concentration of methyl thiol in the atmosphere was retrieved, simultaneously with methane and water vapor, by using artificial neural network (ANN) method based on wavelength modulation spectroscopy (WMS). A mid-infrared spectral band of 3392.4~3394.7 nm was selected to cover the absorption of the three gases, which can be available by tuning an interband cascade laser in our laboratory. In order to determine the concentrations of the three gases simultaneously from the measured WMS signal, the error back propagation (BP) ANN classification algorithm model was established by a series of simulation WMS signals. The second harmonic (2f) signal detection with WMS for different concentration ratios of the three gases under atmospheric conditions was simulated by means of the Hitran and PNNL databases. Of these, 450 groups were used as training samples and 60 sets of data as testing samples. Feature extraction of peak and valley and principal component analysis (PCA) methods were used to reduce the dimension of spectral signals. Then BPANN was trained, and best validation performance of network learning is 0.000241 at epoch 46. The average relative errors are 0.0047 and 0.000742, respectively when validating with testing samples by using the two dimensional reduction methods. The results prove the feasibility of these two methods and BPANN learning algorithm. Finally, the mixed spectral data of WMS-2f signal obtained from the actual experiment were fed into the established BPANN model. However, there are large deviations between the output results of the BPANN model and the actual concentration of the three gases. The relative deviation of methane, water vapor and methyl thiol are 88.92%, 1.69% and 59.94%, respectively. Because there is the background signal, including etalon signal, noise and residual amplitude modulation, superposed in the actual WMS-2f signal, which makes the poor consistency between the experimental data and the ideal simulation data. Overall, the measurement model can be used for qualitative analysis of the actual mixed gases detection. The accuracy of quantitative analysis will be improved by BPANN model trained by more realistic simulation signals in our future work, which can be used for gas recognition and concentration measurement in the actual detection of mixed gases.