The Information Extraction And Quantitative Expression Of Near Infrared Spectrum In Condensed Systems

This thesis has put the emphases on the information extraction and quantitative expression during the model building of near infrared spectrum using the spectral data sets of the dimethyl benzene'isomer. The research work may be important to the application and accuracy development of the NIR model.In this thesis several kinds of wavelength selection methods, including Interval Partial Least-Squares Regression and Moving Window Partial Least-Square Regression, have been studied and compared. Both their merits and demerits have been pointed out. Specially, the method based on the correlation with concentration was carefully studied. The similarity between the PCs and the correlation coefficient was pointed out, and this will help us to better understand the meaning of PCs. The subtractive spectra technology was used to partly overcome the multicollinearity of the NIR spectrum data sets by wavelength selection.The Cross-Correlation algorithm was introduced as spectral pretreatment in NIR field. The algorithm gave a substantial improvement of the performance when the NIR spectra with complex backgrounds and big noise. What is more, the Cross-Correlation algorithm was also applied to calibration transfer of the NIR spectra, and a good result was gotten.The information theory was also introduced for analysis spectral data structure in the NIR field. The component entropy was used to evaluate the degree of overlapping of NIR spectra, and it gave a powerful solution for this problem. A new algorithm, named weighted Shannon's entropy, stemming from the Shannon's information, is used to assess the criteria for NIR spectral data optimization. The analysis results show that the quantitative analysis gives an accurate result and the algorithm works very well for assessing a right criterion for the NIR data pretreatment. Key words:...