Introduction to multivariate Data Analysis

Multivariate data analysis is a universal tool for the evaluation of process spectroscopic data. In process analytics, huge amounts of information are produced i) by the many variables contained in one spectrum often exceeding 1000 (wavenumbers, wavelength, shifts…), and ii) the high number of spectra that is generated within the measurement period. Multivariate data analysis, often also called Chemometrics, help to extract the relevant information which is needed to examine and even control processes. Therefore, calibration models must be precise and robust, and moreover, must cover a wide range of variation of factors posing an influence on the process. In this pre-conference course an introduction to both, explorative data analysis by PCA (Principal Component Analysis), and regression analysis by the most frequently used method, i.e. PLSR (Partial Least Squares Regression) is given. In principal, all optical spectroscopic methods are suited for multivariate evaluation. It will be demonstrated that under certain preconditions, even process NMR spectra can be predicted by PLSR models. At first, basic principles of multivariate data analysis will be provided. This includes a short introduction into the concept of model building and interpretation of results. Detailed aspects of data pretreatment and calibration & validation strategies for chemometric models will be provided with own data from NIR, Raman and NMR spectroscopy. In a first example, the development of an online compatible method for the quantification of methanol in biodiesel by PLSR is presented. This also includes the classification of biodiesel feedstocks by PCA and statistical tools which allow for the estimation of a full uncertainty budget. The Raman spectroscopic prediction of Hydroformylation reaction in a miniplant is used to discuss shortcomings and pitfalls which may occur with the transfer of off-line models to the real processes. Design of experiment and strategies for suitable lab-scale experiments are presented as a possible way to overcome problems. In a third application the prediction of the reactants of an esterification reaction based on process NMR data is demonstrated.