Novel Approaches To Mutivariate Calibration And Classification And Calibration Transfer

Many contemporary chemical manufacturing processes and quality of their products are involved in many factors and indices, which are significantly important to the optimization of chemical manufacturing processes and quality of their products. It is theorial importance and high valuable to apply that extracting real information and filtering noise from complex data to build robust mathematical models, which is benifical for better understanding of manufacturing process and better on-line quality control of final product, because of complexity of manufacturing processes, high colinerarity of different factors or indices and existence of interference or noise in obtained data.Near infrared spectroscopy (NIRS) with wide application in on-line quality and process control was selected in the thesis as research object to propose several kinds of multivariate calibration methods, including signal processing, noise filtering, robust model building and baseline correction.In this dissertation, firstly, a comparison was made between two versions of Net Analyte Signal/Processing (NAS/NAP) algorithms and two different Orthogonal Signal Correction (OSC) algorithms to reveal the relationship between these four methods and that between them and PLS model. Although the NAP and OSC algorithms preprocessed spectra differently, we showed by comparison of the external prediction errors (RMSEP) of these algorithms based on real and synthetic data that these two types of algorithms had the same predictive performance. Under some conditions, the overlap and noise, slightly differentiated these algorithms:the performance of Lorber's NAP algorithm and Fearn's OSC algorithm tracked closely and that Olivieri's NAP algorithm tracked with the performance of a modified form of Fearn's OSC. All these four signal processing methods couldn't improve predictive performance from PLS model, but simplify it. Therefore, a novel signal processing method, Piecewise Net Analyte Processing (PNAP), was developed to enhance the predictive performance from multivariate calibration model based on NIR spectrum through local removal of unrelated information. And through comparison of predictive performance, PNAP was obviously superior to traditional OSC and NAP, even though the spectra contained noise and overlapping of different components. Like Olivieri's NAP and modified version of Fearn's OSC, it was also shown piecewise implementations of Olivieri's NAP and the modified version of Fearn's OSC to filter a set of spectra had the same predictive performance.Two novel algorithms which employed the idea of stacked generalization or stacked regression, Stacked Partial Least-Squares (SPLS) and Stacked Moving-Window Partial Least-Squares (SMWPLS) was reported to remove local redundant information through unevenly distributed weights. The new algorithms established parallel, conventional PLS models based on all intervals of a set of spectra to take advantage of the information from the whole spectrum by incorporating parallel models in a way to emphasize intervals highly related to the target property. It was theoretically and experimentally illustrated that the predictive ability of these two stacked methods was never poorer than that of a PLS model based only on the best interval. These two stacking algorithms generated more parsimonious regression models with better predictive power than conventional PLS, and performed best when the spectral information is neither isolated to a single, small region, nor spread uniformly over the response. Additionally, the work about these two stacked methods did not only demonstrate the improvement, but also demonstrate that stacked regressions had the potential capability of predicting property information from an outlier spectrum in the prediction set.In this dissertation, we also showed the capability of stacked methods to maintain the predictive performance from calibration model. Unlike transfer methods requiring measurement of transfer standard on the primary and secondary instruments, SPLS regression can be used to generate parsimonious regression models with good predictive power on both primary and secondary instruments, without calibration transfer in most cases. The predictive performance from SPLS for predicting samples between different instruments was demonstrated through lake sediment NIR data set. Conventional calibration transfer techniques could also take advantage of stacked PLS regression to minimize local instrumental differences between two instruments. Dual Domain Regression Analysis (DDRA) based on the same idea of data fusion might also contribute to the improvement in predictive performance from conventional calibration model, when calibration transfer methods were implemented in frequency domain. Data fusion in different domains could enhance the predictive performance from transferred model, but fusion using SPLS was much better.Classifiers in combination and fused classifiers removing redundant information locally, like that in calibration, might generate more accurate classification than any single classifier. In this dissertation, we developed a few new stacked classifiers, including Stacked Partial Least-Squares Discriminant Analysis (SPLSDA), a classifier based on SPLS and two approaches to Stacked Linear Discriminant Analysis (SLDA), a classifier that combines stacking with linear discriminant analysis. It was shown in this dissertation that improvement in classification performance obtained after application of stacked PLSDA and stacked LDA as compared with that obtained using PLSDA and LDA classifiers. A stacked PLSDA classifier developed by weighting and combining local classifiers built on separate regions of the data revealed the contributions of those regions of the data to the classification, and often required fewer latent variables for same classifications than a conventional PLSDA classifier applied to the whole data set. The stacking weights generated in stacked PLSDA and stacked LDA performed a kind of variable selection while compensating for information loss that might occur in conventional variable selection techniques.In this dissertation, we used Wavelet Orthogonal Signal Correction (WOSC) for multivariate classification through removal unrelated information in frequency domain. This new classification tool combined a wavelet prism decomposition of a spectral response (local and multi-scale property in frequency scale) and orthogonal signal correction (global filtering uncorrelated classification information) to significantly improve the classification performance either in term of reduction of classification errors and in reduction of model complexity. We showed that a discriminant analysis based on WOSC removed irrelevant classification information effectively and performed favorably as compared to a wavelength-domain filtering approach, such as that used in Orthogonal Partial Least-Squares Discriminant Analysis (OPLS-DA).