Classification and prediction of quality and process parameters of thick juice and beet sugar by fluorescence spectroscopy and chemometrics

Full spectrum fluorescence spectroscopy in combination with multivariate statistical methods (chemometrics) is used to withdraw information from white sugar solutions and thick juice samples from 6 different sugar factories within the same region in Northern Europe with respect to classification and prediction of quality and process parameters. The weekly samples are collected during the 1993 campaign. Classification of the sugar samples according to factory was possible by explorative data analysis (soft independent modelling of class analogy, SIMCA) with 6% misclassification. The factory classification was not so distinct, probably due to sampling problems, when based on thick juice samples diluted 976 times, where circa 30% of the samples were misclassified. These results demonstrate that the fluorescence spectra recorded reflect the quality of the raw material superimposed by the influence of the process, resulting in a characteristic spectral pattern of each factory. By multivariate calibration (partial least squares, PLS) the quality parameters ash, color, a -amino-N, and SO2 content of sugar samples were predicted from the fluorescence emission spectra. The test set correlation coefficients (R) between laboratory analysed values and the values predicted by the calibration model based on the fluorescence spectra are 0.91, 0.94, 0.98, and 0.85, respectively. In an analogue way the parameters ash and color were predicted in thick juice samples with correlation coefficients of 0.89 and 0.90. When the pH value of the sugar samples was corrected to pH = 7.0, the misclassification percentage of the factories increased to circa 10%, and the correlation coefficients to ash, color, a -amino-N, and SO2 content became 0.88, 0.93, 0.98, and 0.84, indicating that some information is lost. Five pairs of excitation-emission wavelengths were selected by a chemometric algorithm (principal variables, PV), giving reasonable correlations results compared to full spectrum models.