Study On Quality Identification Method Of Xiaoqu Mild-Flavor Liquor Based On Spectroscopy And Chemometrics

The production of Xiaoqu mild-flavor liquor is gradually moving towards mechanization and intelligence,and this transformation has necessitated the adjustment of management and control models,which in turn has brought about new requirements for the exploration of the fermentation production mechanism of Xiaoqu liquor.Faced with industrial upgrading,the shortcomings of relying on manual experience and traditional detection methods are gradually becoming apparent,and it is no longer adequate for the needs of mechanized production and intelligent management.Three-dimensional fluorescence has the advantages of no preprocessing,fast detection speed,and high sensitivity,and has been widely used in the detection of water bodies,food and other fields.It has great application potential in liquor detection,such as liquor flavor and vintage identification.This study analyzed the production parameters of mechanized Xiaoqu liquor production,sorted out the correlation between parameters,explored the significance of various parameters on production results,and established a spectral recognition model for the produced liquor at different fermentation days and a discrimination model for liquor of different quality grades.This provides new detection means for liquor detection and technical support for intelligent management of liquor enterprises.(1)To analyze the production variables,principal component analysis(PCA-X)was performed on the collected process data to obtain the score distribution of the first two principal component variables.Three rounds of elimination were carried out to remove 95% of the samples outside the confidence interval,ensuring that the fluctuation range of the selected sample variables was within a reasonable range.The main component load of each parameter point was analyzed to obtain the correlation size between different parameter points.Partial least squares regression analysis was performed on the production variables and production result variables,and the variable importance in projection(VIP)of each parameter was calculated to obtain the importance ranking of process data.The results showed that fermentation time has the greatest impact on production results.(2)The study analyzed the variations in substance content in the original liquor with different fermentation times.It was discovered that as the fermentation time was extended,the content of ethyl acetate and ethyl lactate,which are the main characteristic substances in mechanized Xiaoqu liquor,increased continuously,while the content of fusel oil decreased steadily.The highest liquor yield was obtained in the batch with a fermentation time of 30 days,and further extension of the fermentation time resulted in a decrease in yield.The threedimensional fluorescence characteristics of the highest content ratio of n-propanol and isoamyl alcohol in ethyl acetate,ethyl lactate and fusel oil were respectively analyzed in a 65% ethanolwater solution.The feature emission wavelengths that corresponded to the original liquor threedimensional fluorescence spectra were chosen as the modeling data,and a BP neural network was utilized to fit the regression model of liquor fermentation time.The R value of the whole set model is 0.806,and the mean square error was 62.894,which indicated that the prediction deviation is less than 6 days.(3)The three-dimensional fluorescence spectrum of Xiaoqu liquor mainly exhibited two fluorescence peaks,and the absolute and relative intensities of these peaks varied across different quality levels of liquor.To extract the feature wavelengths of different quality levels of liquor,four spectra of λex290,310,340,and 360 nm were selected.A total of 157 sample spectra were used as the training set to develop the model.Various data dimensionality reduction methods were tested,and principal component analysis(PCA)proved to be the most effective in enhancing the modeling effect.The cumulative explanation rate of the first three principal components reached 90%.The two discriminant models used were support vector machine(SVM)and linear discriminant analysis(LDA),with LDA being the most accurate for the first quality level and SVM being the most accurate for the second quality level.The prediction accuracy of the test set of the two models was 80.77% and 84.62%,respectively.