Baijiu Base Liquor Model Based on Variable Selection Neural Network Band Screening

This study addresses the need for near-infrared spectroscopy analysis of the base liquor of strong-flavor Chinese baijiu, and proposes a variable selection method for variable-length neural networks (VSNN). This method eliminates complex preprocessing steps such as smoothing of the original spectra and baseline correction. The method consists of three vector mechanisms: spectral sensitivity (SV), gradient sensitivity (GV), and weighted gradient sensitivity (WGV). The experiment uses nested five-fold cross-validation to divide the training and test datasets, and the feature retention threshold of VSNN is determined through internal cross-validation optimization. Subsequently, quantitative prediction models for base liquor components and qualitative classification models for grade determination are constructed to verify the effectiveness of the method. The test set results of the quantitative model show the best R² performance in three types of strong-flavor Chinese baijiu base liquor. For Category 1, the WGV vector method shows superior performance, with R² values ranging from 0.8195 to 0.9339; for Category 2, the GV vector method performs best, with R² values ranging from 0.8426 to 0.9022; for the third category, the WGV vector method achieves the best effect, with R² values ranging from 0.8192 to 0.9685. With the support of the random forest classifier, the GV and WGV vector methods achieve a prediction accuracy of 0.9403 in the qualitative model. This study confirms that VSNN can accurately identify the core spectral features related to the quality of base liquor, providing an efficient and convenient technical approach for the rapid quantitative and qualitative analysis of Chinese baijiu base liquor.