Research On Pretreatment And Neural Network Pattern Recognition Of Pollutant Gas On FTIR

With the aggravation of air pollution, environmental problems are increasinglyconcerned. It plays a enormous role in the environmental protection to predict pollutant gasquickly and accurately. Fourier transform infrared （FTIR） spectroscopy technology has beenmore widely used in the field of gas detection because of its rapid measurement , highresolution and high measuring accuracy.The paper starts from the application and the advantage of FTIR spectrometer to analyzethe research status of the pretreatment of spectral data and pattern recognition at home andabroad. On these bases, the infrared spectral acquisition system is designed. And then threetypes of mixed gas CO, NO₂and NO are chosen to be analyzed in the system. Three gases aremixed at a different ratio by the mass flow controller （MFC）. The spectral data of themixtures is collected by FTIR spectrometer. Because of noise influence and the complexnonlinear relationship of IR data, wavelet de-noising and neural network are chosen andimproved, and both methods can effectively predict the spectral components.At first, IR spectral data collected is pretreated using a modified wavelet de-noising toeliminate high frequency noise. Then, principal component analysis （PCA） is selected todescend dimension for sample band. After descending dimension, the 99% information of theoriginal sample data could be completely covered by only three principal components, thustime and computation could be reduced. Four typical neural network and optimized neuralnetwork are used to build model for pattern recognition and sample data with the prediction isused to validate and analysis the model. The results show that the classification accuracy ofoptimized neural network can reach 97.78%, and mean square error （MSE） can reach 0.0222.The classification accuracy of optimized neural network model is better than the four typicalneural network model.