The Effect Of Calibration Set Of PLS Regression On OP/FT-IR Spectra

This thesis includes a review of the application and improvement of PLS regression on the Open-path Fourier transform infrared spectra (OP/FT-IR). In order to improve the accuracy of PLS regression, we focused on the calibration set. We analyzed quantitative the errors of PLS regression and proposed a strategy of building the calibration set. Moreover, this method was improved and optimized. The primary content of the thesis is summarized as follows.1. Open-path Fourier transform infrared spectra were measured around animal farms. By analyzing those spectra for the concentrations of NH3, a systematic investigation was carried out on the quantitative errors. 63 spectra were chosen as the targets in which the concentrations of NH3 vary significantly, from 0 to 1400 ppm-m. In the quantitative analysis of the 63 spectra, three reference spectra were used. The concentrations of NH3 in the reference are 40, 700 and 1300 ppm-m, respectively. The results indicated that the quantitative method based on single reference is practically applicable when the accuracy requirement is not high (e.g. less than 5%). However, as the accuracy requirement increases, the applicability of the single reference spectrum is gradually limited. When the single reference spectrum is ineffective, the quantification was implemented by an iterative process in which a reference spectrum of high resolution (0.125 cm-1) was used. The results also showed an advantage of FT-IR spectrometry, i.e. multiple reference spectra at low resolution could be obtained without actual measurements if a single reference spectrum at high resolution was available.2. We developed a strategy for accuracy improvement by using multiple calibration sets (MCS) in PLS regression. For PLS with MCS, the large concentration range was divided into several segments evenly, and for each segment an independent calibration set was built; these calibration sets compose the multiple calibration sets. It was found the PLS regression with MCS is more flexible than single calibration set in processing the spectra measured around animal farms where the concentration of NH3 varies significantly from nearly zero to about 1,000 ppm-m. The relative error was reduced from about 6% to below 1.5%. Besides, we also carried out some research about the optimal conditions of PLS regression with MCS, such as the number of the calibration sets, the number of factors of PLS, the effect of baseline and the effect of the concentration range. 3. We also optimized the wavenumber in the PLS regression with MCS-4. The PLS regression with full spectrum contains much interference information, which will have negative influence on the ability of the developed model. To resolve this problem, we selected three pairs of wavenumber regions which were interfered by different factors such as the absorbance of water, the nonlinearity between the absorbance and the concentration and the intensity of the absorbance, and compared the concentrations obtained over these regions, respectively. Finally, the optimized wavenumber region was 970-955 cm-1. Moreover, at those wavenumber of water but not ammonia didn't affect the PLS model. However, those wavenumber of water and ammonia affected the PLS model and should be removed before building the model.