Rapid Discrimination Of Chinese Liquor Using Spectroscopy

Chinese liquor has been one of the most popular alcoholic beverages in China for centuries. The sales and the revenues keep growing in the past years. Liquor prices on the market today range from a few Yuan to tens of thousands Yuan per bottle. Even produced by the same distillery, liquors are still dramatically distinct in price, flavor types, and grades, of which some are easy to tell apart, while others are very difficult even for those who drink regularly. Thus some lawbreakers seek access to exorbitant profits by spirit adulterating and counterfeiting. Such illegal acts violate the interests of both consumers and manufacturers. The discrimination, involving different brands, flavor types, ages and alcohol degrees, of Chinese liquors and the detection of adulteration and counterfeit capture increasing attention in Chinese liquor industry. Sensory analysis based on the trained expert panel test is a conventional way for these purposes. However, it is high-cost, inconvenient and artificially affected. Hence there is an urgent need for rapid and reliable detection methods based on modern instruments.In this study, Chinese liquor samples of 22 kinds,10 brands (Maotai, Jiannanchun, Wuliangye, Sanhua, Daohuaxiang, Langjiu, Baiyunbian, Fenjiu, Hongxin and Luzhou Laojiao) and 6 flavor types (Luzhou-flavor, Maotai-flavor, Maotai-Luzhou-flavor, Fen-flavor, rice-flavor and other flavor) were analyzed by visable and near infrared spectroscopy (VIS/NIR), attenuated total reflectance-fourier transform infrared spectroscopy (ATR-FTIR) and ultraviolet spectroscopy (UV), respectively, and then modeled by 3 classification methods including supporting vector machine (SVM), soft independent modeling of class analogy (SIMCA), and linear discriminate analysis based on principal component analysis(PCA-LDA).Before chemometric analysis, several pre-treatments were performed:Cubic-order Savitzky-Golay smoothing filtering, the first and second derivatives. Different wave bands were investigated for each modeling method. Before classification, principal component analysis (PCA) is used as a display method to visualize the data structure. Parameters for each model including the principle component numbers, SVM classification type (c-SVC, nu-SVC) and the kernel type (linear, polynomial, radial basis function and sigmoid) were firstly optimized to create model. Models discrimination ability of VIS/NIR, ATR-FTIR, and UV was compared respectively. The main results were as follows:1. In the method of VIS/NIR, the average correct rates of PCA-LDA were 98.94% in the training set and 95.70% in the test set, and the average correct rates of SIMCA were 97.5% in the training set and 96.67% in the test set. In the SVM the percent correctly classified were 95% in the training set and 98.33% in the test set. All the three models could be used to realize Chinese liquor discrimination. The correction rates of the PCA-LDA model for brands, flavor styles, ages and alcohol degree discrimination were higher than 95%.2. In the method of ATR-FTIR, the discrimination result of SIMCA was less than 10%. Only few samples were correctly classified in this model. The average correct rates of PCA-LDA were 92.92% in the training set and 63.53% in the test set. And the percent correctly classified in the model based on SVM were 73.09% in the training set and 63.53% in the test set. It is obviously that PCA-LDA is the best fit model in the method of ATR-FTIR. However, correct rates of some classes were rather low, which were less than 30%. Thus the discrimination ability of the models based on ATR-FTIR is rather low on the whole.3. In the method of UV, the average correct rates of SIMCA were 87.4%in the training set and 72.08% in the test set. In the SVM the percent correctly classified were 99.38% in the training set and 99.68% in the test set. And the average correct rates of PCA-LDA were 99.83% in the training set and 100% in the test set. In the PCA-LDA, only one sample of the Baiyunbian aged-5 years class was wrongly classified into the Baiyunbian Jiaocang class. The discrimination ability of the PCA-LDA model for brands, flavor styles, ages and alcohol degree discrimination is strong.4. PCA-LDA was the best fit modeling method in discrimination methods based on VIS/NIR (98.94% in the training set and 95.70%), ATR-FTIR (92.92% in the training set and 63.53% in the test set) and UV (99.83% in the training set and 100%). The highest correct average was obtained in the method of UV using PCA-LDA.