Research On The Application Of Bayes Neural Network And FTIR In Quantitative Analysis Of Multi-gases

Air pollution not only brings great harm to human health and the environment, but also destroys the ecological balance of the natural world. Therefore, for the earth's sustainable development, some monitoring methods to the analysis and prediction of polluting sources must be taken in order to expand the appropriate control work. Conventional gas detection technologies have some limitations and shortcomings in applied scope, sensitivity, service life and reliability, but the Fourier transform infrared spectroscopy technology (FTIR) has advantages in the above aspects, so it becomes an ideal gas detection method in most fields.The application and development of FTIR is described, and the current research situations both at home and abroad in infrared quantitative analysis are analyzed. On these bases, the multi-gases distribution and spectral acquisition system are designed. Then three types of common toxic gases CO, NO and NO2 are chosen to analyze in the system. According to the dynamic gas flow distribution, different proportion of gas mixtures are measured by controlling the precision flow meter and the spectral data of the mixtures are acquired by FTIR spectrometer. Because of the complex nonlinear relationship in Multi- gases analysis under low concentrations, the artificial neural network modeling method is selected. After studying the advantages and disadvantages of neural network and its improved algorithms, and considering the large amount of data and high precision in quantitative analysis, principal component analysis and Bayesian regularization method are used to improve the neural network model.First, the reasonable spectrum region is selected. Then the principal component analysis method is used to reduce the high dimensional data. Finally, the Bayesian regularization method is used to improve the error function of the neural network. The model is created and optimized in MATLAB neural network development environment. After verifying the model with the prediction sample sets, the results show that the quantitative fit accuracy can reach 0.974, and the root-mean-square error of prediction is less than 20ppm.Compared with the unreduced dimension network and the LM network which is widely used currently, the results show that the principal component-bayesian neural network can achieve faster modeling speed and better prediction accuracy in the application of infrared quantitative analysis.