Development and application of pattern recognition and calibration methods for multivariate analytical chemical data

This dissertation deals with the application of both standard and novel multivariate statistical analyses to problems in analytical chemistry.; A fundamental problem is the detection of multivariate outliers. Based on new theoretical insights, two novel methods for detecting multivariate outliers were developed. Both the resampling by half means method (RMH) and the smallest half volume (SHV) method are simple to use, conceptually clear, and provide results superior to the current best-performing technique, the minimum covariance determinant.; Strategies for the classification of copy toners were developed. Reflection-absorption infrared microscopy (RA-IR), elemental analysis by x-ray dispersive scanning electron microscopy (SEM), and pyrolysis gas chromatography/mass spectrometry (py-GC/MS) were used to analyze toner samples. Multivariate discriminant analysis correctly classified 96.28% of the 430 RA-IR toner spectra. Principal component and cluster analysis of SEM data for 166 samples established 13 statistically different subgroups. Specific manufacturers were identified for 40.96% of toners for which there were both SEM and RA-IR data available. Py-GC/MS on poly(styrene:acrylate) based toners identified 8 important peaks and a small group containing 5 statistically different subgroups. For 57 toners for which both SEM and Py-GC/MS data was available, 31 could be differentiated. These results demonstrate that RA-IR, SEM, and pyrolysis GC/MS are all useful tools for characterization of copied or laser printed documents.; An improved version of principal component regression (iPCR) was developed. Statistical confidence/prediction intervals were derived from classical regression theory. iPCR successfully modeled the carboxyhemoglobin percentage in forensic blood samples from UV/VIS spectra. The results compared favorably to CO Oximeter analysis. Reduction of the blood prior to analysis is not required.; Improvements to the neural network (NN) modeling process were proposed and evaluated. Methods were developed to rapidly initialize NN weights, enhance the Levenberg-Marquardt optimization algorithm, rapidly and effectively prune the NN structure, and interpret the NN structure. The effectiveness of the methods was demonstrated on calibration problems using Raman and NIR spectra.