SPECTROSCOPIC PREDICTION OF NONLINEAR PROPERTIES BY PRINCIPAL COMPONENT REGRESSION

The Principal Component Regression method is widely used in analytical spectroscopy, even for prediction of nonlinear or almost nonlinear pr operties. This study analyzes the error introduced by the nonlinearity , as compared to common factors such as experimental noise level and s pectral characteristics (e.g. overlapping). It has been found that non linearities are responsible for the major contribution to the final pr ediction errors. A simple algorithm to handle such nonlinearities and to significantly improve PCR prediction results is proposed and evalua ted. It is based on a spectral transformation that partially compensat es for the nonlinearity. The transformation is simple and is unique fo r the whole spectrum. Numerous simulations show that this algorithm co nsiderably improves linear PCR predictions and is comparable to more c omplicated common nonlinear calibrations. Its performance is best when the nonlinear functional form (connecting concentrations to the predi cted properties) is known, however, an algorithm to handle unknown fun ctions is also provided. Moreover, this method may be applied for inve stigating the functional forms and a simple example for this mode of t he algorithm is given. The algorithm is exemplified by its application to experimental data: It is applied to improve PCR prediction of elec trical conductivity from spectral information, in a system of acetic a cid and acetone solutions.