Raw Nongxiangxin Baijiu of different grades were collected during the distillation process, and their near infrared spectroscopy (NIR) data and gas chromatography-mass spectrometry (GC-MS) data were acquired. After preprocessing the NIR data through 5-point 2-fold convolutional smoothing, spectral feature wavelengths were selected using the competitive adaptive reweighted sampling (CARS) algorithm; combining Spearman’s rank correlation coefficient, maximum information coefficient (MIC) and random forest (RF) variable importance, the key flavor components (KC) identified by GC-MS affecting the grading of raw Baijiu were determined. Then, extreme gradient boosting tree (XGBoost) was applied to establish three grade identification models for raw Baijui based on NIR, GC-MS and their fused data. The results showed that the prediction accuracy of the model based on the spectral feature variables selected by CARS was 89.66%, the prediction accuracy of the model based on KC after feature selection was 94.83%, and the classification accuracy of the model based on the fused data of CARS + KC reached as high as 98.28%. This study shows that the fusion of effective feature information from GC-MS and NIR data can enable more accurate and stable grade identification of raw Nongxiangxin Baijiu than either analytical technique alone, which provides a new idea and theoretical basis for the grade identification and quality control of raw Baijiu.

The chemical composition of base baijiu is complex and diverse. A classification model for base baijiu of different sensory grades was established based on the gas chromatography-mass spectrometric (GC-MS) data for their volatile composition. In order to improve the accuracy and generalization capacity of the classification model, a method combining genetic algorithm (GA) and bootstrap aggregating (Bagging) was proposed to optimize the support vector machine (SVM) classifier. Using Spearman’s correlation analysis, 36 key substances were selected, and 12 kernel principal components were extracted as input to the model by kernel principal component analysis, which together accounted for 96.06% of the total variance. The radial basis kernel function support vector machine with the best performance was selected, and the parallel computing Bagging ensemble algorithm with strong adaptability to data diversity was used to construct a Bagging-SVM classifier for base baijiu classification. Finally, GA was used to optimize the parameters (C, γ, and N) of the Bagging-SVM classifier to construct a GA-Bagging-SVM model. The results showed that the accuracy, precision, recall rate, and F1-Score of the GA-Bagging-SVM model were 96.77%, 96.90%, 96.77%, and 96.78%, respectively, which were 6.45%, 5.61%, 6.45%, and 6.42% higher than those of the SVM model, and 3.22%, 2.29%, 3.22%, and 3.15% higher than those of the Bagging-SVM model, respectively. This method can be used as an optimization method for the quality evaluation model for base baijiu.