Preliminary Study On Model Optimization Of Near-Infrared Spectroscopy Analysis

Near-infrared spectroscopy analysis technique is efficient, rapid, low cost,noninvasive, not destroying environment. It is not only suitable for laboratory analysis,but also for in-field fast and real-time on-line analysis. In the dissertation, an in-depthstudy is carried out by synthesizing multi-disciplinary knowledge on the followingtopics: the abstraction of weak spectral signal, the optimal measurement condition, theoptimization of multivariate regression method, the enhancement of calibration modelrobustness and the physical interpretation of measurement results. The main research content of the dissertation involves: Firstly, the abstraction of interesting spectral signal and overcome ofmulti-collinear are studied with multivariate regression method. Improvement ofmodel robustness by elimination of outliers is also analyzed. In both real andsynthetic example, PLS (Partial Least Square) regression method can abstractpredictive information for predictor variable from spectrum effectively and canachieve a robust model. Furthermore, removing outliers can reduce predictionerror by 79.2 percent in four components glucose aqueous solution. Secondly, application of Orthogonal PLS (O-PLS) in near-infrared spectrum hasbeen investigated. O-PLS results in reduced model complexity with preservedprediction ability. The result shows that, the non-correlated systematic variation inspectrum is removed, and the number of resulting PLS components is always reducedto one, which makes interpretation and optimization of the resulting PLS modeleasier. Thirdly, effect of net analyte signal and figure of merit for multivariate model onprediction error is analyzed. The optimal measurement conditions, which includeadding the number of valid wavelength and adapting optimal optical length, isproposed in multivariate calibration. In addition, the wavelength selection approachon the basis of the net analyte signal's relative error indicator (EI) is studiedthoroughly. And improved calculation of EI is put forward in order to avoid relying onabsorbance coefficient, which makes EI suitable for the wavelength selection ofpractical near-infrared spectroscopy analysis. Lastly, different calibration model, outlier detection, orthogonal signal correctionand net analyte signal are studied in process of optimizing milk measurement modelbased on near-infrared spectroscopy. It is verified that, PLS model is more effective inabstracting the fat information, the removal of outliers can reduce prediction error by21.8 percent, PLS and O-PLS model based on the wavelength selected by regressioncoefficient of O-PLS can increase the precision by 48.1 and 55.6 percent separately,and the models based on the optimal wavelength selected by EI can increase theprecision by 55.8 and 56.8%.