Study On Multi-component Gases Quantitative Analysis Based On Ftir

Quantitative analysis of toxic and hazard gases is very useful for fire rescuing, fire monitoring, gas leaking diagnosis and researching on material characteristics. The current gas monitoring technologies has both advantages and disadvantages on consideration of applied scope, sensitivity, reliability, interference immunity, service life, convenience, and economy. With high sensitivity, wide applied scope, fast response, long life and convenience, Fourier transform infrared spectroscopy (FTIR) was chosen as the ideal technology in the field of gas monitoring.The experimental system was established by FTIR, gas compounding equipment and other devices. Eight different kinds of classic toxic gases with low concentration, including CO, CO₂, NO, NO₂, SO₂, HCl, HBr and HCN, were chosen for quantitative analysis. Partial least square (PLS) regression model was established and corrected after appropriate spectrum region choosing, data pre-treating, standards selecting and parameters setting. The correlated coefficient between actual concentrations and predicting concentrations of each component is higher than 0.99, and the root-mean standard error of calibration (RMSEC) is lower than 15 ppm. The predicting performance of model is validate by the standards in validation which proved that the error of the concentration prediction is lower than±2% F.S. and the root-mean standard error of prediction (RMSEP) is lower than 20 ppm.To improve the application and verify the interference immunity of the mode, computing range of the spectrum data ware extended to the whole range of wavelength. To study the relationship between the spectrum data and gas concentration data, three new methods, including Polynomial partial least square(PPLS), partial least square with BP artificial neural net (PLS-BP) and least square with support vector regression (LS-SVR), were created. Compared with two classical methods, PLS and classical least square (CLS), the characteristics, prediction performance and complexion of each method were discussed. According to the value of RMSEC, the performance of nonlinear methods, PPLS, PLS-BP and LS-SVR, are better than linear method, CLS and PLS. Above all, PLS-BP method has the best performance. The value of RMSEC of each component is less than 2. According to the performance index R², the prediction performance of nonlinear methods, PPLS, PLS-BP and LS-SVR, are much better than linear method, CLS and PLS. The value of R2 of each component is higher than 0.9.