A Machine Learning Method for the Quality Detection of Base Liquor and Commercial Liquor Using Multidimensional Signals from an Electronic Nose

Abstract

Chinese liquor is a world-famous beverage with a long history. Base liquor, a product of liquor brewing, significantly affects the flavor and quality of commercial liquor. In this study, a machine learning method consisting of a deep residual network (ResNet)18 backbone with a light gradient boosting machine (LightGBM) classifier (ResNet-GBM) is proposed for the quality identification of base liquor and commercial liquor using multidimensional signals from an electronic nose (E-Nose). Ablation experiments are conducted to analyze the contribution of the framework’s components. Five evaluation metrics (accuracy, sensitivity, precision, F1 score, and kappa score) are used to verify the performance of the proposed method, and six other frameworks (support vector machine (SVM), random forest (RF), k-nearest neighbor (KNN), extreme gradient boosting (XGBoost), multidimensional scaling-support vector machine (MDS-SVM), and back-propagation neural network (BPNN)) on three datasets (base liquor, commercial liquor, and mixed base and commercial liquor datasets). The experimental results demonstrate that the proposed ResNet-GBM model achieves the best performance for identifying base liquor and commercial liquors with different qualities. The proposed framework has the highest F1 score for the identification of commercial liquor in the mixed dataset due to the contribution of similar microconstituents from the base liquor. The proposed method can be used for the quality control of Chinese liquor and promotes the practical application of E-nose devices.