Methods For Quality And Wine Age Determination Of Chinese Rice Wine Based On Near Infrared Spectroscopy

Chinese rice wine, the symbol of Chinese civilization, contains abundant amino acid, sugar, vitamins, and so on. Thus, it is honored as "liquid cake". In recent years, Chinese rice wine industry has been developing in a high speed. However, the quality status of Chinese rice wine is not satisfied. For the commercial Chinese rice wine, the quality of some products are not up to the requirements of the national standards, the labels of some are not right, or the wine age of some are wrongly labeled. Near infrared (NIR) spectroscopy, which utilizes transmission or reflectance spectra to analyze physical structure or chemical components of substance, is a rapid, nondestructive, and green analytical technique in wine quality determination.In this research, Fourier transform near infrared (FT-NIR) spectroscopy and visible near infrared spectroscopy together with chemometrics were applied to determine Chinese rice wine quality and to discriminate wine age, respectively.The results and conclusions for FT-NIR spectroscopy analysis part were listed as follows:(1) The influences of spectra collection parameters on FT-NIR analysis results were analyzed. Through the comparison analysis, the parameters for Chinese rice wine spectra collection were chosen. The energy of signal energy was 5V, the reference spectrum was collected in a 1 mm path-length rectangular quartz cuvette with air as reference, and resolution and the number of scan was 16cm^-1 and 32, respectively.(2) The principles for sample outlier analysis, which were based on Dixon test, as well as leverage and studentized residual test, were studied. Firstly, Dixon test, as well as leverage and studentized residual test were applied to the whole sample set. If Mahalanobis distance, leverage value or studentized residual value of some sample was noticeably different from those of the others, the sample was considered as an outlier in the first-round analysis. And they would be removed from the sample set. In the second-round analysis, the samples considered as outliers were reclaimed one by one in order to know whether they provided any useful information or they must be removed. The results demonstrated that studentized residual value had the highest influence on the sample outlier analysis. It was concluded that if some sample had a noticeably different studentized residual value; it should be removed without further analysis.(3) The performances of models for alcoholic degree, sugar content, pH and total acid established by different wavelength ranges (full wavelength range, short wavelength range, long wavelength range, and two absorption band 1300～1650 nm, 2200～2400), different spectra preprocessing methods, and different mathematical methods were compared in this research. The results showed that the best models for alcoholic degree, sugar content, pH and total acid were developed by partial least squares regression. (PLSR) together with full wavelength range of raw spectra, with correlation coefficient of calibration (r_cal) was 0.969, 0.992, 0.969 and 0.979, correlation coefficient of validation (r_val) was 0.966, 0.986, 0.955 and 0.970, root mean standard error of calibration (RMSEC) was 0.106 %(V/V), 0.049 %, 0.014 and 0.058 g/L, root mean standard error of validation (RMSEP) was 0.112 %(V/V), 0.061%, 0.017 and 0.068 g/L, respectively.(4) The performance of wine age discrimination models developed by Discriminant Analysis (DA), Soft Independent Modeling of Class Analogy (SIMCA), Discriminant Partial Least Squares (DPLS) together with different wavelength range and different preprocessing methods were compared. The results demonstrated that DA together with long wavelength range of raw spectra gave the best wine age discrimination result. The percentage of sample correctly classified was 100 % for the calibration and validation sample set, respectively. The percentage of sample correctly classified for the calibration and validation models developed by SIMCA together with 5-point smoothed spectra were 99.4 % and 98.8 %, respectively. And those for the model established by DPLS were 98.2 % and 93.8 %, respectively.The results and conclusions for the visible near infrared spectroscopy analysis part were listed as follows:(1) Fiber spectroscopic systems for square bottled and round bottled Chinese rice wine samples were set up. For square bottled Chinese rice wine samples, transmission spectra of Chinese rice wine were collected by a line-light fiber spectroscopic system which consisted of a light source (tungsten halogen lamp, 20 W), a sample holder, two adjustable collimating lens, a fiber spectrometer, two pieces of optics fiber, and a computer. The spectrometer was equipped with a 2048-element linear silicon CCD array detector. The wavelength range was 600～1200 nm. For round bottled Chinese rice wine samples, reflectance transmission spectra were collected by a circle-light fiber spectrometer system which was composed of light source, a sample holder, two adjustable collimating lens, a fiber spectrometer, a piece of optics fiber, and a computer. The two systems established foundation for the practicalities.(2) The PLSR and multi linear regression (MLR) models for alcoholic degree, sugar content, pH and total acid of square bottled Chinese rice wine samples were established.①The alcoholic degree, sugar content and total acid models developed by PLSR together with raw spectra was the best. For the three models, r_cal was 0.933, 0.902 and 0.859, RMSEC was 0.127 %(V/V), 0.218 % and 0.106 g/L, r_val was 0.922, 0.908 and 0.840, and RMSEP was 0.147 %(V/V), 0.215% and 0.112 g/L, respectively. The best model of pH was developed by weighted partial least squares regression(WPLSR), r_cal and r_val was 0.834 and 0.767, and RMSEC and RMSEP was 0.024 and 0.028, respectively.②For MLR models, 7 wave bands of raw spectra gave the best results for alcoholic degree analysis, r_cal and r_val was 0.925 and 0.808, and RMSEC and RMSEP was 0.135 %(V/V) and 0.217 %(V/V), respectively; 7 wave bands of 13-point smoothed spectra gave the best results for sugar content analysis, r_cal and r_val was 0.908 and 0.819, and RMSEC and RMSEP was 0.210 % and 0.233 %, respectively; 12 wave bands of standard normal variate (SNV) spectra gave the best results for pH analysis, r_cal and r_val was 0.856 and 0.787, and RMSEC and RMSEP was 0.023 and 0.027, respectively, which were better than those for PLSR model; 7 wave bands of 5-point smoothed spectra gave the best results for total analysis, r_cal and r_val was 0.839 and 0.723, and RMSEC and RMSEP was 0.113 g/L and 0.148 g/L, respectively. In conclusion, the performance of PLSR models was better. However, the PLSR models were relatively complicated. MLR models were simple and rapid, which could be used for on-line analysis.The PLSR and MLR models for round bottled Chinese rice wine samples were also established. The performance of alcoholic degree, pH and total acid models developed by PLSR together with raw spectra were the best, r_cal was 0.901, 0.838 and 0.864, RMSEC was 0.178 %(V/V), 0.036 and 0.160 g/L, r_val was 0.889, 0.833 and 0.848, RMSEP was 0.187 %(V/V), 0.037 and 0.176 g/L, respectively. The performance of sugar content model developed by PLS together with SNV spectra was the best, r_cal and r_val was 0.806 and 0.794, and RMSEC and RMSEP was 0.155 % and 0.154 %, respectively. The performance of MLR models (except sugar content) was worse than that of PLR models. The results for alcoholic degree (MLR together with 7 wave bands of SNV spectra) were: r_cal=0.723, RMSEC=0.284 %(V/V), r_val=0.559, RMSEP=0.341%(V/V); the results for sugar content (6 wave bands of SNV spectra) were: r_cal=0.858, RMSEC=0.134 %, r_val=0.823, RMSEP=0.143 %; the results for pH (6 wave bands of raw spectra) were: r_cal=0.827, RMSEC=0.037, r_val=0.803, RMSEP=0.040; and the results for total acid (7 wave bands of first derivative spectra) were: r_cal=0.840, RMSEC=0.172 g/L, r_val=0.724, RMSEP=0.216 g/L.(3) The performance of wine age discrimination models for square and round bottled Chinese rice wine samples developed by DA, S1MCA, DPLS together with different preprocessing methods were studied. The best models for square bottled Chinese rice wine samples were developed by DA together with raw spectra, as well as SIMCA together with 25-point smoothed spectra. The percentage of samples correctly classified was 100 % for the 1-, 2-, 3-, 4- and 5-year-old sample groups, respectively. The best models for round bottled Chinese rice wine samples were developed by DA together with raw spectra, as well as SIMCA together with 21-point smoothed spectra, the 1-, 2-, 3-, 4- and 5-year-old sample groups were all correctly classified.