- Email: [email protected]
Identification of Jiangxi wines by three-dimensional fluorescence fingerprints
Spectrochimica Acta Part A: Molecular and Biomolecular Spectroscopy 96 (2012) 605–610
Contents lists available at SciVerse ScienceDirect
Spectrochimica Acta Part A: Molecular and Biomolecular Spectroscopy journal homepage: www.elsevier.com/locate/saa
Identification of Jiangxi wines by three-dimensional fluorescence fingerprints Yiqun Wan a,b,⇑, Fengqin Pan b, Mingyue Shen a,b a b
State Key Laboratory of Food Science and Technology, Nanchang University, Nanchang, Jiangxi 330047, China Center of Analysis and Testing, Nanchang University, Nanchang, Jiangxi 330047, China
h i g h l i g h t s
g r a p h i c a l a b s t r a c t
" Three-dimensional fluorescence
fingerprints were established to distinguish wine samples. " A pattern recognition approach could identify wine samples from different manufacturers. " The proposed method could provide the criterion for the quality control of wines.
a r t i c l e
i n f o
Article history: Received 2 May 2012 Received in revised form 1 July 2012 Accepted 9 July 2012 Available online 17 July 2012 Keywords: Three-dimensional fluorescence spectrum Fingerprints Wines Principal component analysis Hierarchical cluster analysis Back-propagation network
a b s t r a c t A new assay of identifying wines was developed based on fingerprints of three-dimensional fluorescence spectra, and 30 samples from different manufacturers were analyzed. The techniques of principal component analysis (PCA) and hierarchical cluster analysis (HCA) were used to differentiate and evaluate the character parameters of wines’ three-dimensional fluorescence spectra. At the same time, the back-propagation network (BPN) was applied to predict the attribution of unknown samples. The results of PCA and HCA showed that there was definite different information among the wine samples from different manufacturers. It was promising that the method could be applied to distinguish wine samples produced by different manufacturers. The proposed method could provide the criterion for the quality control of wines. Ó 2012 Elsevier B.V. All rights reserved.
Introduction Nowadays, objective and authentic food information are the major concern of many consumers. In general, wine is a widely consumed beverage in the world with thousands years of tradition. Determination of its authenticity is one of the most important aspects in food quality control and safety. Wine is a complex medium composed of at least several hundred compounds of various chemical natures and concentrations, and the chemical compounds responsible for wine aroma and quality are mainly esters [1], ⇑ Corresponding author at: State Key Laboratory of Food Science and Technology, Nanchang University, Nanchang, Jiangxi 330047, China. Tel./fax: +86 791 88321370. E-mail address: [email protected] (Y. Wan). 1386-1425/$ - see front matter Ó 2012 Elsevier B.V. All rights reserved. http://dx.doi.org/10.1016/j.saa.2012.07.030
phenols [2,3], carbonyl compounds [4], and acids [5]. Such compounds come from fermentation processes and wine ageing, and they determine the brand- or technology-specific aroma tones [6]. Thus the quality control of wine is closely related to the analysis of these constituents. Commonly used instrumental methods for the quality control of foods were gas chromatography (GC) [1,2,4,6] and liquid chromatography (LC) [3,5,7–14], for example, a method based on liquid chromatographic separation with ultraviolet detection (HPLC–UV) has been proved to be an useful analytical tool in quality control in the brewing industry for the determination of phenolic acids [3], and individual and free anthocyanins were also determined by HPLC in young red wines to establish the most appropriate conditions for ensuring color stability during the shelf life [7].
606
Y. Wan et al. / Spectrochimica Acta Part A: Molecular and Biomolecular Spectroscopy 96 (2012) 605–610
Since there are tens of constituents in wine samples, which are slightly different according to their production technology and geographical origins [1–6], we cannot select only several specific components as essential criteria. In light of this, characteristic fingerprint is developed for the quality control in the brewing industry, and many successful studies have shown that it is possible to distinguish wines or grape variety on the basis of fingerprinting [15–19]. For example, the anthocyanin and flavonol profiles of the skin and the flesh of three red grapes varieties were obtained by HPLC and were used to differentiate similar grape varieties [19]. As the characteristic fingerprints are complex multivariate data sets, chemometrics techniques, such as hierarchical cluster analysis (HCA) [20,21], principal component analysis (PCA) [22] and artificial neural network (ANN) [23,24], etc., should be taken into consideration for the discrimination of wine samples. Chemometrics methods can highlight the chemical differences between samples and reduce variation due to physical effects, and the theory and methodology for these chemometrics methods are described in detail elsewhere [20–24]. So far, the main techniques applied to the fingerprinting of wine are still chromatographic or spectrometric analytical methods, including gas chromatography [15,16], high performance liquid chromatography [19], mass spectrometry [17,18], etc. These analytical methods might be more accurate but it is more time consuming and expensive. Therefore, a new approach to discriminate wine samples from different manufacturers is essentially required for the quality control in the brewing industry. Threedimensional fluorescence spectra can be an excellent candidate for the discrimination of wine samples from different manufacturers because it could reflect the slight differences of total chemical information between them and it is fast and nondestructive [25– 27]. However, few studies have been reported on the applications of three-dimensional fluorescence spectra to the discrimination of wine samples until now. To our knowledge, it is the first example that three-dimensional fluorescence fingerprints were used to distinguish the wines from different manufacturers and batches. For establishing a physical–chemical method for identifying different wine samples objectively, three-dimensional fluorescence spectrum technique was adopted to construct the fingerprints of wines, and 30 samples from different manufacturers and batches were analyzed in this paper. The techniques of principal component analysis (PCA) and hierarchical cluster analysis (HCA) were applied to investigate the character parameters of wines’ three-dimensional fluorescence spectra, thus the 30 wine samples were classified into five groups rationally. Finally, optimized back-propagation network (BPN) was employed to train the samples which had been classified by PCA, meanwhile, the established network was adopted to predict the attribution of unknown samples. Materials and methods Wine samples Thirty wine samples were obtained from five different manufactures: samples A1–A6 were six different batches produced in 2009 from Zhanggong wine industry Limited Liability Company in Jiangxi province (A), samples B7–B12 were six different batches produced between 2009 and 2010 from Site wine Limited Liability Company in Jiangxi province(B), samples C13–C18 were six different batches produced between 2009 and 2010 from Linchuangong Limited Liability Company in Jiangxi province(C), samples D19– D24 were six different batches produced between 2009 and 2010 from the Lidu wine industry Limited Liability Company in Jiangxi province(D), and samples E25–E30 were six different batches produced between 2009 and 2010 from Qibaoshan wine industry Limited Liability Company in Jiangxi province(E) respectively.
Analytical procedures The fluorescence spectra of 30 wine samples were directly measured on Hitachi F-4500 Fluorospectrophotometer (Hitachi, Japan) under the same conditions. Three-dimensional excitation –emission spectra of wines were scanned from 250 to 500 nm in the emission wavelengths and from 210 to 400 nm in the excitation wavelengths, the slits of excitation and emission were both set at 5 nm, both regular steps were 5 nm. The scanning speed was set at 40 nm/s. Surface fitting, three-dimensional fluorescence gram, characteristic parameter and data analysis were derived from the instrument software to input to Matlab 6.5. Results and discussion Three-dimensional fluorescence spectrometric analysis of wines from the different manufactures According to analytical procedures, three-dimensional fluorescence fingerprints of 30 wine samples were obtained, and Fig. 1 showed the representative three-dimensional fluorescence spectra of wine samples from five manufacturers. From these fingerprints, it was clear that there were significant differences in the fluorescence spectra among the five manufacturers. It implied that the production technology and raw materials of different manufacturers had certain difference. On the other hand, the spectra of samples from different batches of a given manufacturer were qualitatively similar, although the production process of the same manufacturer had relatively better control. Thus, the three-dimensional fluorescence spectra has certain correlation with the quality of wine samples and could be applied to distinguish the samples of wines from the different manufactures. Data standardization of characteristic parameters of wines’ fingerprints The fluorescence intensities of different wine samples recorded from 250 to 500 nm in the emission domain and from 210 to 400 nm in the excitation domain were extracted from the threedimensional fluorescence spectra to form a data matrix. The data matrix was then subjected to statistical analysis by matlab 6.5 and 11 variables such as mean value of fluorescence intensity, variance, abscissa of gravity center, etc. (shown in Table 1) were obtained for classification and identification [28,29]. For reducing the effects of various operational factors on the statistical characteristic parameters, and improving the samples comparability, characteristic parameters were dealt with data standardization before HCA and PCA, that is, each characteristic parameter of the samples was datalized to be zero mean and one variance according to:
x0i ¼ ðxi xÞ=s where xi is the ith characteristic parameter of each sample, x is the mean of all the ith characteristic parameter of each sample and s is the variance of all the ith characteristic parameter of each sample. Hierarchical cluster analysis In this work, HCA was applied in order to discover the possible similarities or affinities amongst the wine samples. We took the Euclidean distance as metric and Ward’s method as the agglomeration rule, then a data matrix (30 11) was subjected to a HCA of samples. The dendrogram was obtained and shown in Fig. 2. It was clear that 30 samples were approximately divided into 5 groups, and the wines of the five different manufacturers were accumulated into five groups separately, which was agreed with the fac-
Y. Wan et al. / Spectrochimica Acta Part A: Molecular and Biomolecular Spectroscopy 96 (2012) 605–610
607
Fig. 1. Three-dimensional fluorescence fingerprint spectra of wine samples from the five different manufacturers (A, B, C, D and E).
tual results. Obviously, the Hierarchical Cluster Analysis method was a good way for classification of wine samples. Principal component analysis In this study, we used PCA method to examine the similarity among the five different manufacturers of wine samples. A total 82.48% of data variance was explained by relation between principal component 1 vs. principal component 2 (PC1 vs. PC2). Principal component 1 was responsible for 64.20% while component 2 was responsible for 18.28% of the variance. It could be clearly seen from the projection plot (Fig. 3a) that the wine samples studied were distinctly separated into five groups. In fact, PC1 shows a distinct discrimination between the sets A, C and D. Set D forms a clearly separated cluster with positive scores on PC1, while the other sets have negative PC1 scores, thereinto, the sets B, C and E have similar scores, the scores of the sets B and E are higher than those of set C, and set A has the lowest scores. On the other hand, sets B and E were distinguished from the other wines by PC2, sets B and E have positive scores, and set E has higher scores than set B, which could distinguish the wine samples between manufacturers B and E. The discrimination of the sets A, C and D on PC2 which all have nega-
tive PC2 scores is much less clear. The classical way to see the variables responsible for the formation of the five clusters in Fig. 3a is to inspect the correlation in Fig. 3b. Thus, in Fig. 3b, the loading plot of variables shows that ordination of gravity center and abscissa of gravity center loadings are particularly related to the separation of the set D, the sets A and C are distinguished from the other wines mainly owing to the x, y correlation coefficient vector with negative loadings on both PCs, conversely, the specific locations of the sets B and E are due to their high percentages of first order origin distance in comparison with the other wine samples. Finally, on the whole, it was clear from the PC1 vs. PC2 projection plane that wine samples from the different manufacturers clustered respectively, and each batch of products presented different quality characteristics completely according to their material origin and production technology, which revealed that the five plots could provide reasonable relationships among the 30 samples. The construction and optimization of BPN From the viewpoints of reducing constructional complexity and unacceptably high computational loads associated with training neural networks, there is an advantage in minimizing the number
608
Table 1 Characteristic parameter abstraction result of three dimension fluorescence spectrum. Mean
Variance
Abscissa of gravity center
Ordinate of gravity center
X direction marginal distribution
Y direction marginal distribution
First order mixed Center distance
First order origin distance
X,y correlation coefficient
Denseelliptic area
Denseelliptic Long axi slope
A1 A2 A3 A4 A5 A6 B7 B8 B9 B10 B11 B12 C13 C14 C15 C16 C17 C18 D19 D20 D21 D22 D23 D24 E25 E26 E27 E28 E29 E30
40.239 38.868 39.813 38.57 35.413 36.733 74.227 75.004 74.688 82.428 82.225 78.824 62.645 62.123 61.742 62.467 62.784 61.975 123.69 126.45 126.06 138.32 134.32 132.43 75.731 77.528 78.209 65.6 65.837 68.274
45.4466 48.1397 46.6092 44.8673 39.3781 45.7267 76.7625 77.4899 77.2722 81.6499 81.3325 75.1013 55.0776 54.3463 53.678 54.4609 54.3821 53.6718 101.173 104.016 104.348 118.771 114.011 113.35 78.7387 85.2574 85.6609 69.4143 69.5648 75.6188
375.691 375.869 374.732 375.344 380.325 377.144 365.405 365.229 365.616 366.435 367.32 367.661 383.98 383.974 384.407 384.517 384.595 384.764 384.171 383.531 383.248 382.346 382.894 383.113 363.076 362.647 362.834 369.728 369.158 368.645
285.383 286.702 284.138 285.296 286.085 286.937 270.677 270.961 270.155 271.222 272.369 270.396 281.51 282.276 282.003 281.8 283.182 282.425 294.72 292.343 293.578 292.805 291.663 290.785 274.61 277.621 277.27 278.685 277.009 280.194
1.39 108 1.35 108 1.37 108 1.33 108 1.24 108 1.28 108 2.50 108 2.52 108 2.52 108 2.78 108 2.78 108 2.67 108 2.22 108 2.20 108 2.19 108 2.21 108 2.23 108 2.20 108 4.38 108 4.47 108 4.45 108 4.87 108 4.74 108 4.68 108 2.53 108 2.59 108 2.62 108 2.24 108 2.24 108 2.32 108
1.06 108 1.03 108 1.04 108 1.01 108 9.34 107 9.71 107 1.85 108 1.87 108 1.86 108 2.06 108 2.06 108 1.96 108 1.63 108 1.62 108 1.60 108 1.62 108 1.64 108 1.61 108 3.36 108 3.41 108 3.41 108 3.73 108 3.61 108 3.55 108 1.92 108 1.98 108 2.00 108 1.68 108 1.68 108 1.76 108
3.96 1010 3.85 1010 3.90 1010 3.80 1010 3.54 1010 3.65 1010 6.72 1010 6.79 1010 6.75 1010 7.50 1010 7.53 1010 7.17 1010 6.21 1010 6.17 1010 6.13 1010 6.20 1010 6.27 1010 6.17 1010 1.28 1011 1.30 1011 1.30 1011 1.41 1011 1.37 1011 1.35 1011 6.94 1010 7.17 1010 7.23 1010 6.19 1010 6.16 1010 6.46 1010
3.99 1010 3.87 1010 3.92 1010 3.82 1010 3.57 1010 3.67 1010 6.81 1010 6.89 1010 6.85 1010 7.60 1010 7.63 1010 7.27 1010 6.27 1010 6.24 1010 6.20 1010 6.27 1010 6.34 1010 6.24 1010 1.30 1011 1.32 1011 1.32 1011 1.44 1011 1.40 1011 1.37 1011 8.49 1011 8.76 1011 8.83 1011 7.59 1011 7.57 1011 7.91 1011
0.7204 0.7213 0.7209 0.7208 0.7205 0.7210 0.7136 0.7137 0.7136 0.7142 0.7146 0.7138 0.7152 0.7153 0.715 0.7149 0.7152 0.7149 0.7111 0.7111 0.7113 0.711 0.7105 0.7102 0.7164 0.717 0.7172 0.7154 0.7152 0.716
4.80 1011 4.65 1011 4.71 1011 4.59 1011 4.28 1011 4.41 1011 8.29 1011 8.38 1011 8.33 1011 9.24 1011 9.27 1011 8.85 1011 7.62 1011 7.58 1011 7.54 1011 7.62 1011 7.70 1011 7.59 1011 1.59 1012 1.61 1012 1.61 1012 1.76 1012 1.70 1012 1.68 1012 8.49 1011 8.76 1011 8.83 1011 7.59 1011 7.57 1011 7.91 1011
1.0491 1.057 1.0462 1.0504 1.0296 1.0522 1.0040 1.0067 0.9998 1.0035 1.0059 0.9921 0.9875 0.9921 0.9885 0.9867 0.9947 0.9894 1.0672 1.0550 1.0644 1.0629 1.0525 1.0456 1.0451 1.0674 1.064 1.0332 1.0253 1.0486
Y. Wan et al. / Spectrochimica Acta Part A: Molecular and Biomolecular Spectroscopy 96 (2012) 605–610
Sample
609
Y. Wan et al. / Spectrochimica Acta Part A: Molecular and Biomolecular Spectroscopy 96 (2012) 605–610 Table 2 The statistics of prediction results. Samples No. (origin)
FAa
FBa
F Ca
FDa
FEa
Accuracy (%)
1 2 3 4 5
197 1 0 0 0
1 199 0 0 0
0 0 200 1 0
1 0 0 199 0
1 0 0 0 200
98.5 99.5 100 99.5 100
(A) (B) (C) (D) (E)
a FA, FB, FC, FD and FE are the frequencies of predicted results belong to A, B, C, D and E, respectively.
Fig. 2. Clustering analysis of the 30 wine samples from the five different manufacturers.
tegic placement of characteristic points to enable the required sensitivity in the output of the target classification. PCA analysis has been applied to compress the real higher dimensional space into a meaningful lower dimensional space. This approach has been well applied in other applications [23,24] in conjunction with GAs to optimize system parameters against performance criteria and to enhance feature determination in data. PCA has also been widely used as a data reduction technique, for example, in process of fault diagnosis [30] and analysis of bench-marking data [31]. In this application, it is shown that reduction can be achieved successfully when this novel approach was applied to predict results. To construct reasonable structure of network, over-fitting must be avoided. Therefore, the following important parameters should be optimized. In order to avoid over-fitting, the following condition must be reached,
N=TW > 1 where N represents the number of the samples, and TW represents the total number of the weight. When N is constant, the formula above can be met by reducing value of TW which can be estimated as follows:
TW ¼ ðI þ 1Þ H þ ðH þ 1Þ O where I represents the node of input layer, H represents the node of hidden layer and O represents that of output layer. The decrease of I could bring TW decrease to meet the demand referred above. In our study, PCA was adopted once again to process the 30 11 data matrix to reduce the number of variables. The score matrix of the first three principle components, as the input data matrix, was used to construct the network. According to Jiang et al.’s study [32], the node of hidden layer selected in this study was four. To estimate the stability of the optimized network, a statistical work was carried out to testify the accuracy. Under the same condition, the performance of this network model was assessed by 200 times of predicton for wine sample of each manufacturer. The statistical work was conducted and the rate of accuracy was calculated as shown in Table 2. It could be clearly seen from the results that the accuracy was up to 98.5%, which manifested that the optimized network could be predicted with good accuracy. Thus, the quality control could be achieved by determining the three-dimensional fluorescence fingerprints of unknown wine samples from five different manufacturers and then using the established BPN model to predict the brand of unknown wine samples. Fig. 3. Scores plot (a) and loading plot (b) of the 30 wine samples projected on the space of PC1 vs. PC2. (e) wine samples from manufacturer A (marked as set A); (s) wine samples from manufacturer B (marked as set B); (+) wine samples from manufacturer C (marked as set C); (h) wine samples from manufacturer D(marked as set D); (D) wine samples from manufacturer E (marked as set E).
of input nodes. This corresponds with minimizing the number of characteristic parameters. To achieve satisfactory performance whilst reducing the number of parametric inputs requires the stra-
Conclusions In this study, the fingerprints of wine samples from the five different manufacturers were established by three-dimensional excitation-emission matrix fluorescence. A multi-parametric approach which included both hierarchical cluster analysis (HCA) and principal component analysis (PCA) could preferably distinguish wine samples from different manufacturers. PCA was also employed to handle the data of wines’ characteristic parameters to reduce the
610
Y. Wan et al. / Spectrochimica Acta Part A: Molecular and Biomolecular Spectroscopy 96 (2012) 605–610
number of variables. The processed data was repeatedly trained to construct reasonable structure. The multiple predicted results manifested that the optimized network was characteristic of high accuracy and good stability.The BPN technology could be used to comprehensively predict the properties of unknown samples. Acknowledgements The authors gratefully acknowledge the financial support of this study by the Natural Science Foundation of China (20965005), the Research Program of State Key Laboratory of Food Science and Technology of Nanchang University (SKLF-MB201002; SKLF-TS200918). References [1] J.J. Rodríguez-Bencomo, J.E. Conde, M.A. Rodríguez-Delgado, F. GarcíaMontelongo, J.P. Pérez-Trujillo, J. Chromatogr., A 963 (2002) 213–223. [2] L. Minuti, R. Pellegrino, J. Chromatogr., A 1185 (2008) 23–30. [3] A. Alonso García, B. Cancho Grande, J. Simal Gándara, J. Chromatogr., A 1054 (2004) 175–180. [4] M.J. Gómez-Míguez, J.F. Cacho, V. Ferreira, I.M. Vicario, F.J. Heredia, Food Chem. 100 (2007) 1464–1473. [5] V. Pereira, M. Pontes, J.S. Câmara, J.C. Marques, J. Chromatogr., A 1189 (2008) 435–443. [6] G. Antalick, M.C. Perello, G. de Revel, Food Chem. 121 (2010) 1236–1245. [7] C. Pérez-Lamela, M.S. García-Falcón, J. Simal-Gándara, I. Orriols-Fernández, Food Chem. 101 (2007) 601–606. [8] A.S. Rodrigues, M.R. Pérez-Gregorio, M.S. García-Falcón, J. Simal-Gándara, Food Res. Int. 42 (2009) 1331–1336. [9] A.S. Rodrigues, M.R. Pérez-Gregorio, M.S. García-Falcón, J. Simal-Gándara, D.P.F. Almeida, Food Control 21 (2010) 878–884. [10] R.M. Pérez-Gregorio, M.S. García-Falcón, J. Simal-Gándara, A.S. Rodrigues, D.P.F. Almeida, J. Food Compos. Anal. 23 (2010) 592–598.
[11] A.S. Rodrigues, M.R. Pérez-Gregorio, M.S. García-Falcón, J. Simal-Gándara, D.P.F. Almeida, Food Chem. 124 (2011) 303–308. [12] M.R. Pérez-Gregorio, M.S. García-Falcón, J. Simal-Gándara, Food Chem. 124 (2011) 652–658. [13] M.R. Pérez-Gregorio, J. Regueiro, C. González-Barreiro, R. Rial-Otero, J. SimalGándara, Food Control 22 (2011) 1108–1113. [14] M.R. Pérez-Gregorio, C. González-Barreiro, R. Rial-Otero, J. Simal-Gándara, Food Control 22 (2011) 2052–2058. [15] J. Laaks, T. Letzel, T.C. Schmidt, M.A. Jochmann, Anal. Bioanal. Chem. 403 (2012) 2429–2436. [16] C. Dall’Asta, M. Cirlini, E. Morini, G. Galaverna, J. Chromatogr., A 1218 (2011) 7557–7565. [17] Z. Spacil, M. Shariatgorji, N. Amini, P. Solich, L.L. Ilag, Rapid Commun. Mass Spectrom. 23 (2009) 1834–1840. [18] E. Villagra, L.S. Santos, B.G. Vaz, M.N. Eberlin, V.F. Laurie, Food Chem. 131 (2012) 692–697. [19] M. Figueiredo-González, E. Martínez-Carballo, B. Cancho-Grande, J.L. Santiago, M.C. Martí nez, J. Simal-Gándara, Food Chem. 130 (2012) 9–19. [20] J.A.S. Almeida, L.M.S. Barbosa, A.A.C.C. Pais, S.J. Formosinho, Chemom. Intell. Lab 87 (2007) 208–217. [21] M.Y. Shen, M.Y. Xie, S.P. Nie, Y.Q. Wan, J.H. Xie, Anal. Lett. 44 (2011) 863–873. [22] Y.N. Ni, Y.H. Lai, S. Brandes, S. Kokot, Anal. Chim. Acta 647 (2009) 149–158. [23] J. Basak, R.L. De, S.K. Pal, Pattern Recogn. Lett. 19 (1998) 997–1006. [24] P. Tongpadungrod, T.D.L. Rhys, P.N. Brett, Sens. Actuators, A 105 (2003) 47–54. [25] E. Csomo’s, K. He’berger, L. Simon-Sarkadi, J. Agric. Food Chem. 50 (2002) 3768–3774. [26] P. Etie’vant, P. Schlich, R. Cantagrel, A. Bertrand, J.C. Bouvier, J. Sci. Food Agric. 46 (1989) 421–438. [27] K. He0 berger, E. Csomo0 s, L. Simon-Sarkadi, J. Agric. Food Chem. 51 (2003) 8055–8060. [28] K.H.J. Siddiqui, D. Eastwood, Part of the SPIE conference on environmental monitoring and remediation technologies II SPIE 3853 (1999) 60-71. [29] F.Q. Jiang, Y.D. Tao, Y. Shao, J. Ethnopharmacol. 111 (2007) 265–270. [30] J.F. MacGregor, T. Kourti, Control Eng. Pract. 3 (1995) 403–414. [31] R.M. Alonso-Salces, S. Guyot, C. Herrero, L.A. Berrueta, J.F. Drilleau, B. Gallo, F. Vicente, Food Chem. 91 (2005) 91–98. [32] H. Jiang, L.B. Zeng, J.M. Hu, Comp. Eng. Appl. (2001) 122–125.
Contents lists available at SciVerse ScienceDirect
Spectrochimica Acta Part A: Molecular and Biomolecular Spectroscopy journal homepage: www.elsevier.com/locate/saa
Identification of Jiangxi wines by three-dimensional fluorescence fingerprints Yiqun Wan a,b,⇑, Fengqin Pan b, Mingyue Shen a,b a b
State Key Laboratory of Food Science and Technology, Nanchang University, Nanchang, Jiangxi 330047, China Center of Analysis and Testing, Nanchang University, Nanchang, Jiangxi 330047, China
h i g h l i g h t s
g r a p h i c a l a b s t r a c t
" Three-dimensional fluorescence
fingerprints were established to distinguish wine samples. " A pattern recognition approach could identify wine samples from different manufacturers. " The proposed method could provide the criterion for the quality control of wines.
a r t i c l e
i n f o
Article history: Received 2 May 2012 Received in revised form 1 July 2012 Accepted 9 July 2012 Available online 17 July 2012 Keywords: Three-dimensional fluorescence spectrum Fingerprints Wines Principal component analysis Hierarchical cluster analysis Back-propagation network
a b s t r a c t A new assay of identifying wines was developed based on fingerprints of three-dimensional fluorescence spectra, and 30 samples from different manufacturers were analyzed. The techniques of principal component analysis (PCA) and hierarchical cluster analysis (HCA) were used to differentiate and evaluate the character parameters of wines’ three-dimensional fluorescence spectra. At the same time, the back-propagation network (BPN) was applied to predict the attribution of unknown samples. The results of PCA and HCA showed that there was definite different information among the wine samples from different manufacturers. It was promising that the method could be applied to distinguish wine samples produced by different manufacturers. The proposed method could provide the criterion for the quality control of wines. Ó 2012 Elsevier B.V. All rights reserved.
Introduction Nowadays, objective and authentic food information are the major concern of many consumers. In general, wine is a widely consumed beverage in the world with thousands years of tradition. Determination of its authenticity is one of the most important aspects in food quality control and safety. Wine is a complex medium composed of at least several hundred compounds of various chemical natures and concentrations, and the chemical compounds responsible for wine aroma and quality are mainly esters [1], ⇑ Corresponding author at: State Key Laboratory of Food Science and Technology, Nanchang University, Nanchang, Jiangxi 330047, China. Tel./fax: +86 791 88321370. E-mail address: [email protected] (Y. Wan). 1386-1425/$ - see front matter Ó 2012 Elsevier B.V. All rights reserved. http://dx.doi.org/10.1016/j.saa.2012.07.030
phenols [2,3], carbonyl compounds [4], and acids [5]. Such compounds come from fermentation processes and wine ageing, and they determine the brand- or technology-specific aroma tones [6]. Thus the quality control of wine is closely related to the analysis of these constituents. Commonly used instrumental methods for the quality control of foods were gas chromatography (GC) [1,2,4,6] and liquid chromatography (LC) [3,5,7–14], for example, a method based on liquid chromatographic separation with ultraviolet detection (HPLC–UV) has been proved to be an useful analytical tool in quality control in the brewing industry for the determination of phenolic acids [3], and individual and free anthocyanins were also determined by HPLC in young red wines to establish the most appropriate conditions for ensuring color stability during the shelf life [7].
606
Y. Wan et al. / Spectrochimica Acta Part A: Molecular and Biomolecular Spectroscopy 96 (2012) 605–610
Since there are tens of constituents in wine samples, which are slightly different according to their production technology and geographical origins [1–6], we cannot select only several specific components as essential criteria. In light of this, characteristic fingerprint is developed for the quality control in the brewing industry, and many successful studies have shown that it is possible to distinguish wines or grape variety on the basis of fingerprinting [15–19]. For example, the anthocyanin and flavonol profiles of the skin and the flesh of three red grapes varieties were obtained by HPLC and were used to differentiate similar grape varieties [19]. As the characteristic fingerprints are complex multivariate data sets, chemometrics techniques, such as hierarchical cluster analysis (HCA) [20,21], principal component analysis (PCA) [22] and artificial neural network (ANN) [23,24], etc., should be taken into consideration for the discrimination of wine samples. Chemometrics methods can highlight the chemical differences between samples and reduce variation due to physical effects, and the theory and methodology for these chemometrics methods are described in detail elsewhere [20–24]. So far, the main techniques applied to the fingerprinting of wine are still chromatographic or spectrometric analytical methods, including gas chromatography [15,16], high performance liquid chromatography [19], mass spectrometry [17,18], etc. These analytical methods might be more accurate but it is more time consuming and expensive. Therefore, a new approach to discriminate wine samples from different manufacturers is essentially required for the quality control in the brewing industry. Threedimensional fluorescence spectra can be an excellent candidate for the discrimination of wine samples from different manufacturers because it could reflect the slight differences of total chemical information between them and it is fast and nondestructive [25– 27]. However, few studies have been reported on the applications of three-dimensional fluorescence spectra to the discrimination of wine samples until now. To our knowledge, it is the first example that three-dimensional fluorescence fingerprints were used to distinguish the wines from different manufacturers and batches. For establishing a physical–chemical method for identifying different wine samples objectively, three-dimensional fluorescence spectrum technique was adopted to construct the fingerprints of wines, and 30 samples from different manufacturers and batches were analyzed in this paper. The techniques of principal component analysis (PCA) and hierarchical cluster analysis (HCA) were applied to investigate the character parameters of wines’ three-dimensional fluorescence spectra, thus the 30 wine samples were classified into five groups rationally. Finally, optimized back-propagation network (BPN) was employed to train the samples which had been classified by PCA, meanwhile, the established network was adopted to predict the attribution of unknown samples. Materials and methods Wine samples Thirty wine samples were obtained from five different manufactures: samples A1–A6 were six different batches produced in 2009 from Zhanggong wine industry Limited Liability Company in Jiangxi province (A), samples B7–B12 were six different batches produced between 2009 and 2010 from Site wine Limited Liability Company in Jiangxi province(B), samples C13–C18 were six different batches produced between 2009 and 2010 from Linchuangong Limited Liability Company in Jiangxi province(C), samples D19– D24 were six different batches produced between 2009 and 2010 from the Lidu wine industry Limited Liability Company in Jiangxi province(D), and samples E25–E30 were six different batches produced between 2009 and 2010 from Qibaoshan wine industry Limited Liability Company in Jiangxi province(E) respectively.
Analytical procedures The fluorescence spectra of 30 wine samples were directly measured on Hitachi F-4500 Fluorospectrophotometer (Hitachi, Japan) under the same conditions. Three-dimensional excitation –emission spectra of wines were scanned from 250 to 500 nm in the emission wavelengths and from 210 to 400 nm in the excitation wavelengths, the slits of excitation and emission were both set at 5 nm, both regular steps were 5 nm. The scanning speed was set at 40 nm/s. Surface fitting, three-dimensional fluorescence gram, characteristic parameter and data analysis were derived from the instrument software to input to Matlab 6.5. Results and discussion Three-dimensional fluorescence spectrometric analysis of wines from the different manufactures According to analytical procedures, three-dimensional fluorescence fingerprints of 30 wine samples were obtained, and Fig. 1 showed the representative three-dimensional fluorescence spectra of wine samples from five manufacturers. From these fingerprints, it was clear that there were significant differences in the fluorescence spectra among the five manufacturers. It implied that the production technology and raw materials of different manufacturers had certain difference. On the other hand, the spectra of samples from different batches of a given manufacturer were qualitatively similar, although the production process of the same manufacturer had relatively better control. Thus, the three-dimensional fluorescence spectra has certain correlation with the quality of wine samples and could be applied to distinguish the samples of wines from the different manufactures. Data standardization of characteristic parameters of wines’ fingerprints The fluorescence intensities of different wine samples recorded from 250 to 500 nm in the emission domain and from 210 to 400 nm in the excitation domain were extracted from the threedimensional fluorescence spectra to form a data matrix. The data matrix was then subjected to statistical analysis by matlab 6.5 and 11 variables such as mean value of fluorescence intensity, variance, abscissa of gravity center, etc. (shown in Table 1) were obtained for classification and identification [28,29]. For reducing the effects of various operational factors on the statistical characteristic parameters, and improving the samples comparability, characteristic parameters were dealt with data standardization before HCA and PCA, that is, each characteristic parameter of the samples was datalized to be zero mean and one variance according to:
x0i ¼ ðxi xÞ=s where xi is the ith characteristic parameter of each sample, x is the mean of all the ith characteristic parameter of each sample and s is the variance of all the ith characteristic parameter of each sample. Hierarchical cluster analysis In this work, HCA was applied in order to discover the possible similarities or affinities amongst the wine samples. We took the Euclidean distance as metric and Ward’s method as the agglomeration rule, then a data matrix (30 11) was subjected to a HCA of samples. The dendrogram was obtained and shown in Fig. 2. It was clear that 30 samples were approximately divided into 5 groups, and the wines of the five different manufacturers were accumulated into five groups separately, which was agreed with the fac-
Y. Wan et al. / Spectrochimica Acta Part A: Molecular and Biomolecular Spectroscopy 96 (2012) 605–610
607
Fig. 1. Three-dimensional fluorescence fingerprint spectra of wine samples from the five different manufacturers (A, B, C, D and E).
tual results. Obviously, the Hierarchical Cluster Analysis method was a good way for classification of wine samples. Principal component analysis In this study, we used PCA method to examine the similarity among the five different manufacturers of wine samples. A total 82.48% of data variance was explained by relation between principal component 1 vs. principal component 2 (PC1 vs. PC2). Principal component 1 was responsible for 64.20% while component 2 was responsible for 18.28% of the variance. It could be clearly seen from the projection plot (Fig. 3a) that the wine samples studied were distinctly separated into five groups. In fact, PC1 shows a distinct discrimination between the sets A, C and D. Set D forms a clearly separated cluster with positive scores on PC1, while the other sets have negative PC1 scores, thereinto, the sets B, C and E have similar scores, the scores of the sets B and E are higher than those of set C, and set A has the lowest scores. On the other hand, sets B and E were distinguished from the other wines by PC2, sets B and E have positive scores, and set E has higher scores than set B, which could distinguish the wine samples between manufacturers B and E. The discrimination of the sets A, C and D on PC2 which all have nega-
tive PC2 scores is much less clear. The classical way to see the variables responsible for the formation of the five clusters in Fig. 3a is to inspect the correlation in Fig. 3b. Thus, in Fig. 3b, the loading plot of variables shows that ordination of gravity center and abscissa of gravity center loadings are particularly related to the separation of the set D, the sets A and C are distinguished from the other wines mainly owing to the x, y correlation coefficient vector with negative loadings on both PCs, conversely, the specific locations of the sets B and E are due to their high percentages of first order origin distance in comparison with the other wine samples. Finally, on the whole, it was clear from the PC1 vs. PC2 projection plane that wine samples from the different manufacturers clustered respectively, and each batch of products presented different quality characteristics completely according to their material origin and production technology, which revealed that the five plots could provide reasonable relationships among the 30 samples. The construction and optimization of BPN From the viewpoints of reducing constructional complexity and unacceptably high computational loads associated with training neural networks, there is an advantage in minimizing the number
608
Table 1 Characteristic parameter abstraction result of three dimension fluorescence spectrum. Mean
Variance
Abscissa of gravity center
Ordinate of gravity center
X direction marginal distribution
Y direction marginal distribution
First order mixed Center distance
First order origin distance
X,y correlation coefficient
Denseelliptic area
Denseelliptic Long axi slope
A1 A2 A3 A4 A5 A6 B7 B8 B9 B10 B11 B12 C13 C14 C15 C16 C17 C18 D19 D20 D21 D22 D23 D24 E25 E26 E27 E28 E29 E30
40.239 38.868 39.813 38.57 35.413 36.733 74.227 75.004 74.688 82.428 82.225 78.824 62.645 62.123 61.742 62.467 62.784 61.975 123.69 126.45 126.06 138.32 134.32 132.43 75.731 77.528 78.209 65.6 65.837 68.274
45.4466 48.1397 46.6092 44.8673 39.3781 45.7267 76.7625 77.4899 77.2722 81.6499 81.3325 75.1013 55.0776 54.3463 53.678 54.4609 54.3821 53.6718 101.173 104.016 104.348 118.771 114.011 113.35 78.7387 85.2574 85.6609 69.4143 69.5648 75.6188
375.691 375.869 374.732 375.344 380.325 377.144 365.405 365.229 365.616 366.435 367.32 367.661 383.98 383.974 384.407 384.517 384.595 384.764 384.171 383.531 383.248 382.346 382.894 383.113 363.076 362.647 362.834 369.728 369.158 368.645
285.383 286.702 284.138 285.296 286.085 286.937 270.677 270.961 270.155 271.222 272.369 270.396 281.51 282.276 282.003 281.8 283.182 282.425 294.72 292.343 293.578 292.805 291.663 290.785 274.61 277.621 277.27 278.685 277.009 280.194
1.39 108 1.35 108 1.37 108 1.33 108 1.24 108 1.28 108 2.50 108 2.52 108 2.52 108 2.78 108 2.78 108 2.67 108 2.22 108 2.20 108 2.19 108 2.21 108 2.23 108 2.20 108 4.38 108 4.47 108 4.45 108 4.87 108 4.74 108 4.68 108 2.53 108 2.59 108 2.62 108 2.24 108 2.24 108 2.32 108
1.06 108 1.03 108 1.04 108 1.01 108 9.34 107 9.71 107 1.85 108 1.87 108 1.86 108 2.06 108 2.06 108 1.96 108 1.63 108 1.62 108 1.60 108 1.62 108 1.64 108 1.61 108 3.36 108 3.41 108 3.41 108 3.73 108 3.61 108 3.55 108 1.92 108 1.98 108 2.00 108 1.68 108 1.68 108 1.76 108
3.96 1010 3.85 1010 3.90 1010 3.80 1010 3.54 1010 3.65 1010 6.72 1010 6.79 1010 6.75 1010 7.50 1010 7.53 1010 7.17 1010 6.21 1010 6.17 1010 6.13 1010 6.20 1010 6.27 1010 6.17 1010 1.28 1011 1.30 1011 1.30 1011 1.41 1011 1.37 1011 1.35 1011 6.94 1010 7.17 1010 7.23 1010 6.19 1010 6.16 1010 6.46 1010
3.99 1010 3.87 1010 3.92 1010 3.82 1010 3.57 1010 3.67 1010 6.81 1010 6.89 1010 6.85 1010 7.60 1010 7.63 1010 7.27 1010 6.27 1010 6.24 1010 6.20 1010 6.27 1010 6.34 1010 6.24 1010 1.30 1011 1.32 1011 1.32 1011 1.44 1011 1.40 1011 1.37 1011 8.49 1011 8.76 1011 8.83 1011 7.59 1011 7.57 1011 7.91 1011
0.7204 0.7213 0.7209 0.7208 0.7205 0.7210 0.7136 0.7137 0.7136 0.7142 0.7146 0.7138 0.7152 0.7153 0.715 0.7149 0.7152 0.7149 0.7111 0.7111 0.7113 0.711 0.7105 0.7102 0.7164 0.717 0.7172 0.7154 0.7152 0.716
4.80 1011 4.65 1011 4.71 1011 4.59 1011 4.28 1011 4.41 1011 8.29 1011 8.38 1011 8.33 1011 9.24 1011 9.27 1011 8.85 1011 7.62 1011 7.58 1011 7.54 1011 7.62 1011 7.70 1011 7.59 1011 1.59 1012 1.61 1012 1.61 1012 1.76 1012 1.70 1012 1.68 1012 8.49 1011 8.76 1011 8.83 1011 7.59 1011 7.57 1011 7.91 1011
1.0491 1.057 1.0462 1.0504 1.0296 1.0522 1.0040 1.0067 0.9998 1.0035 1.0059 0.9921 0.9875 0.9921 0.9885 0.9867 0.9947 0.9894 1.0672 1.0550 1.0644 1.0629 1.0525 1.0456 1.0451 1.0674 1.064 1.0332 1.0253 1.0486
Y. Wan et al. / Spectrochimica Acta Part A: Molecular and Biomolecular Spectroscopy 96 (2012) 605–610
Sample
609
Y. Wan et al. / Spectrochimica Acta Part A: Molecular and Biomolecular Spectroscopy 96 (2012) 605–610 Table 2 The statistics of prediction results. Samples No. (origin)
FAa
FBa
F Ca
FDa
FEa
Accuracy (%)
1 2 3 4 5
197 1 0 0 0
1 199 0 0 0
0 0 200 1 0
1 0 0 199 0
1 0 0 0 200
98.5 99.5 100 99.5 100
(A) (B) (C) (D) (E)
a FA, FB, FC, FD and FE are the frequencies of predicted results belong to A, B, C, D and E, respectively.
Fig. 2. Clustering analysis of the 30 wine samples from the five different manufacturers.
tegic placement of characteristic points to enable the required sensitivity in the output of the target classification. PCA analysis has been applied to compress the real higher dimensional space into a meaningful lower dimensional space. This approach has been well applied in other applications [23,24] in conjunction with GAs to optimize system parameters against performance criteria and to enhance feature determination in data. PCA has also been widely used as a data reduction technique, for example, in process of fault diagnosis [30] and analysis of bench-marking data [31]. In this application, it is shown that reduction can be achieved successfully when this novel approach was applied to predict results. To construct reasonable structure of network, over-fitting must be avoided. Therefore, the following important parameters should be optimized. In order to avoid over-fitting, the following condition must be reached,
N=TW > 1 where N represents the number of the samples, and TW represents the total number of the weight. When N is constant, the formula above can be met by reducing value of TW which can be estimated as follows:
TW ¼ ðI þ 1Þ H þ ðH þ 1Þ O where I represents the node of input layer, H represents the node of hidden layer and O represents that of output layer. The decrease of I could bring TW decrease to meet the demand referred above. In our study, PCA was adopted once again to process the 30 11 data matrix to reduce the number of variables. The score matrix of the first three principle components, as the input data matrix, was used to construct the network. According to Jiang et al.’s study [32], the node of hidden layer selected in this study was four. To estimate the stability of the optimized network, a statistical work was carried out to testify the accuracy. Under the same condition, the performance of this network model was assessed by 200 times of predicton for wine sample of each manufacturer. The statistical work was conducted and the rate of accuracy was calculated as shown in Table 2. It could be clearly seen from the results that the accuracy was up to 98.5%, which manifested that the optimized network could be predicted with good accuracy. Thus, the quality control could be achieved by determining the three-dimensional fluorescence fingerprints of unknown wine samples from five different manufacturers and then using the established BPN model to predict the brand of unknown wine samples. Fig. 3. Scores plot (a) and loading plot (b) of the 30 wine samples projected on the space of PC1 vs. PC2. (e) wine samples from manufacturer A (marked as set A); (s) wine samples from manufacturer B (marked as set B); (+) wine samples from manufacturer C (marked as set C); (h) wine samples from manufacturer D(marked as set D); (D) wine samples from manufacturer E (marked as set E).
of input nodes. This corresponds with minimizing the number of characteristic parameters. To achieve satisfactory performance whilst reducing the number of parametric inputs requires the stra-
Conclusions In this study, the fingerprints of wine samples from the five different manufacturers were established by three-dimensional excitation-emission matrix fluorescence. A multi-parametric approach which included both hierarchical cluster analysis (HCA) and principal component analysis (PCA) could preferably distinguish wine samples from different manufacturers. PCA was also employed to handle the data of wines’ characteristic parameters to reduce the
610
Y. Wan et al. / Spectrochimica Acta Part A: Molecular and Biomolecular Spectroscopy 96 (2012) 605–610
number of variables. The processed data was repeatedly trained to construct reasonable structure. The multiple predicted results manifested that the optimized network was characteristic of high accuracy and good stability.The BPN technology could be used to comprehensively predict the properties of unknown samples. Acknowledgements The authors gratefully acknowledge the financial support of this study by the Natural Science Foundation of China (20965005), the Research Program of State Key Laboratory of Food Science and Technology of Nanchang University (SKLF-MB201002; SKLF-TS200918). References [1] J.J. Rodríguez-Bencomo, J.E. Conde, M.A. Rodríguez-Delgado, F. GarcíaMontelongo, J.P. Pérez-Trujillo, J. Chromatogr., A 963 (2002) 213–223. [2] L. Minuti, R. Pellegrino, J. Chromatogr., A 1185 (2008) 23–30. [3] A. Alonso García, B. Cancho Grande, J. Simal Gándara, J. Chromatogr., A 1054 (2004) 175–180. [4] M.J. Gómez-Míguez, J.F. Cacho, V. Ferreira, I.M. Vicario, F.J. Heredia, Food Chem. 100 (2007) 1464–1473. [5] V. Pereira, M. Pontes, J.S. Câmara, J.C. Marques, J. Chromatogr., A 1189 (2008) 435–443. [6] G. Antalick, M.C. Perello, G. de Revel, Food Chem. 121 (2010) 1236–1245. [7] C. Pérez-Lamela, M.S. García-Falcón, J. Simal-Gándara, I. Orriols-Fernández, Food Chem. 101 (2007) 601–606. [8] A.S. Rodrigues, M.R. Pérez-Gregorio, M.S. García-Falcón, J. Simal-Gándara, Food Res. Int. 42 (2009) 1331–1336. [9] A.S. Rodrigues, M.R. Pérez-Gregorio, M.S. García-Falcón, J. Simal-Gándara, D.P.F. Almeida, Food Control 21 (2010) 878–884. [10] R.M. Pérez-Gregorio, M.S. García-Falcón, J. Simal-Gándara, A.S. Rodrigues, D.P.F. Almeida, J. Food Compos. Anal. 23 (2010) 592–598.
[11] A.S. Rodrigues, M.R. Pérez-Gregorio, M.S. García-Falcón, J. Simal-Gándara, D.P.F. Almeida, Food Chem. 124 (2011) 303–308. [12] M.R. Pérez-Gregorio, M.S. García-Falcón, J. Simal-Gándara, Food Chem. 124 (2011) 652–658. [13] M.R. Pérez-Gregorio, J. Regueiro, C. González-Barreiro, R. Rial-Otero, J. SimalGándara, Food Control 22 (2011) 1108–1113. [14] M.R. Pérez-Gregorio, C. González-Barreiro, R. Rial-Otero, J. Simal-Gándara, Food Control 22 (2011) 2052–2058. [15] J. Laaks, T. Letzel, T.C. Schmidt, M.A. Jochmann, Anal. Bioanal. Chem. 403 (2012) 2429–2436. [16] C. Dall’Asta, M. Cirlini, E. Morini, G. Galaverna, J. Chromatogr., A 1218 (2011) 7557–7565. [17] Z. Spacil, M. Shariatgorji, N. Amini, P. Solich, L.L. Ilag, Rapid Commun. Mass Spectrom. 23 (2009) 1834–1840. [18] E. Villagra, L.S. Santos, B.G. Vaz, M.N. Eberlin, V.F. Laurie, Food Chem. 131 (2012) 692–697. [19] M. Figueiredo-González, E. Martínez-Carballo, B. Cancho-Grande, J.L. Santiago, M.C. Martí nez, J. Simal-Gándara, Food Chem. 130 (2012) 9–19. [20] J.A.S. Almeida, L.M.S. Barbosa, A.A.C.C. Pais, S.J. Formosinho, Chemom. Intell. Lab 87 (2007) 208–217. [21] M.Y. Shen, M.Y. Xie, S.P. Nie, Y.Q. Wan, J.H. Xie, Anal. Lett. 44 (2011) 863–873. [22] Y.N. Ni, Y.H. Lai, S. Brandes, S. Kokot, Anal. Chim. Acta 647 (2009) 149–158. [23] J. Basak, R.L. De, S.K. Pal, Pattern Recogn. Lett. 19 (1998) 997–1006. [24] P. Tongpadungrod, T.D.L. Rhys, P.N. Brett, Sens. Actuators, A 105 (2003) 47–54. [25] E. Csomo’s, K. He’berger, L. Simon-Sarkadi, J. Agric. Food Chem. 50 (2002) 3768–3774. [26] P. Etie’vant, P. Schlich, R. Cantagrel, A. Bertrand, J.C. Bouvier, J. Sci. Food Agric. 46 (1989) 421–438. [27] K. He0 berger, E. Csomo0 s, L. Simon-Sarkadi, J. Agric. Food Chem. 51 (2003) 8055–8060. [28] K.H.J. Siddiqui, D. Eastwood, Part of the SPIE conference on environmental monitoring and remediation technologies II SPIE 3853 (1999) 60-71. [29] F.Q. Jiang, Y.D. Tao, Y. Shao, J. Ethnopharmacol. 111 (2007) 265–270. [30] J.F. MacGregor, T. Kourti, Control Eng. Pract. 3 (1995) 403–414. [31] R.M. Alonso-Salces, S. Guyot, C. Herrero, L.A. Berrueta, J.F. Drilleau, B. Gallo, F. Vicente, Food Chem. 91 (2005) 91–98. [32] H. Jiang, L.B. Zeng, J.M. Hu, Comp. Eng. Appl. (2001) 122–125.