Classification Of Chinese Liquors With Different Aroma Type, Geographic Oirgin And Quality Grade By Mass Spectrometry And Chemometrics

Compared with other spirits, Chinese liquor has higher ethanol content. Benefiting fromits unique flavor, it is really popular in China. Because of differences in geographic condition,raw material and manufacturing practice, the aroma profiles of Chinese liquors are quitedifferent. In the study, the method using headspace solid phase microextraction massspectrometry (HS SPME MS) combined with chemometrics was successfully applied toclassify the liquors with different aroma type, geographic origin and quality grade.The details are as follows:(1) The ions abundance of Chinese liquor samples from different geographic origin wascollected by HS SPME MS technique, without pre-treatment or chromatographic separation.By combination of partial least squares discriminant analysis (PLS DA) and stepwise lineardiscriminant analysis (SLDA) methods,36characteristic ions were finally selected and then aback-propagation (BP) neural network using the36ions as inputs and different geographicorigin as outputs was built, whose prediction accuracy was up to84.4%. The optimalparameter of BP neural network was tansig, trainbfg and five neurons in hidden layer. Theions abundance was also used to develop models by partial least squares regression (PLS),principal component regression (PCR) analysis and support vector machine (SVM). TheSVM model was used to validate BP network mutually to guarantee prediction accuracy,which was developed by ions selected by different methods, PLS regression coefficients andPCR regression coefficients. The optimal SVM model was achieved with ions selected byPCR regression coefficients, whose prediction accuracy for the validation set was up to87.5%.87.5%. The optimal parameter c, σ was32and1.414in the SVM model.(2) The ions abundance of Xijiu liquor samples with different aroma type (strong aromatype and soy sauce aroma type) was collected by HS SPME MS technique. Seven ions wereselected by PLS regression analysis and used to develop BP neural network and SVM model,whose prediction accuracy were both up to100%. The SVM model was used to validate BPnetwork mutually. The optimal parameter of BP neural network was tansig, trainlm and threeneurons in hidden layer. The optimal parameter c, σ was1.414and1in the SVM model.(3) The ions abundance of Yanghe liquor samples with different quality grade wascollected by HS SPME MS technique. By combination of PLS DA and SLDA,14characteristic ions were finally selected and developed a BP neural network, whose predictionaccuracy was up to100%. The optimal parameter of BP neural network was tansig, trainbfgand four neurons in hidden layer. The ions abundance was also used to develop models byPLS, PCR and SVM analysis. The SVM model was built to validate BP network mutually.The optimal SVM model was achieved with ions selected by PLS regression coefficients, whose prediction accuracy was up to96.3%. The optimal parameter c, σ was5.66and32inthe SVM model.(4) The ions abundance of Niulanshan liquor samples with different quality grade wascollected by HS SPME MS technique. By combination of PLS DA and SLDA,8characteristic ions were finally selected and developed a BP neural network, whose predictionaccuracy was up to93.3%. The optimal parameter of BP neural network was tansig, trainlmand two neurons in hidden layer. The ions abundance was also used to develop models byPLS, PCR and SVM analysis. The SVM model was built to validate BP network mutually.The optimal SVM model was achieved with ions selected by PCR regression coefficients,whose prediction accuracy was up to86.7%. The optimal parameter c, σ was2and11.314inthe SVM model.