Abstract:

Near infrared spectroscopy was used to predict the main chemical ingredients of fermented grains of Chinese
liquor by modeling. The established models were optimized for improved prediction performance. Latent variables (LVs)
were extracted by partial least squares (PLS) and used as the input variables of least squares support vector machine (LSSVM)
for the establishment of NIR quantitative models to predict the alcohol, starch, moisture contents and acidity of
fermented grains. Furthermore, a comparison with the PLS models built with waveband selection using uninformative
variable elimination (UVE) was carried out. The results showed that compared with the PLS models, quantitative correlation
coefficients (R2), root mean square errors of prediction (RMSEP), and relative percent differences (RPD) of alcohol, starch,
moisture and acidity showed better performances in the LS-SVM models, respectively. The accuracy of the LS-SVM models
in predicting unknown samples was significantly higher than that of the PLS models. In summary, the accuracy, stability and
prediction performance of the LS-SVM models were better than those of the PLS ones. This study can provide a new way
for quantitative analysis of fermented grains of Chinese liquor.

Key words: fermented grains of Chinese liquor, least squares support vector machines (LS-SVM), latent variables (LVs), partial least squares (PLS), waveband selection