Fresenius J Anal Chem (1997) 359: 143—149 ( Springer-Verlag 1997

OR I G I N A L P AP E R

Li-Xian Sun · Klaus Danzer · Gabriela Thiel

Classification of wine samples by means of artificial neural networks

and discrimination analytical methods

Received: 12 July 1996/Revised: 9 October 1996/Accepted: 12 October 1996

Abstract The three-layer artificial neural network origins of wine by reliable chemical analysis techniques
(ANN) model with back-propagation (BP) of error was in combination with modern chemometrical methods
used to classify wine samples in six different regions instead of traditional wine experts. Concerning the
based on the measurements of trace amounts of analysis methods of constituents in wine, there are
B, V, Mn, Zn, Fe, Al, Cu, Sr, Ba, Rb, Na, P, Ca, Mg, many reports such as GC for the analysis of aroma
K using an inductively coupled plasma optical emis- compounds in wine [1], HPLC for the determination
sion spectrometer (ICP-OES). The ANN architecture of phenolic acids [2], spectrophotometry for Fe [3],
and parameters were optimized. The results obtained Mn [4] and preservatives containing benzoic acid,
with ANN were compared with those obtained by sorbic acid, dehydroacetic acid and ethyl parahyd-
cluster analysis, principal component analysis, the robenzoate [5], ion-selective electrode (ISE) methods
Bayes discrimination method and the Fisher discrim- for the simultaneous determination of sulfite and phos-
ination method. A satisfactory prediction result (100%) phate [6] and ethanol [7], capillary zone electrophor-
by an artificial neural network using the jackknife esis (CZE) for phenolic compounds [8], flame atomic
leave-one-out procedure was obtained for the classi- absorption spectroscopy analysis (FAAS) and graphite
fication of wine samples containing six categories. furnace atomic absorption spectroscopy analysis
(GFAAS) for Fe, Cu, Ag, Co, Si, Zn, Pb [9—11], GC-MS
(mass spectrometry) for trifluoroacetylated glycosides
[12], GC-FTIR (Fourier transform infrared)-MS for
Introduction spirits [13], NMR [14] and so on. There is also
a monograph which describes various methods of wine
Wine is one of the most widely consumed beverages in analysis [15]. The inductively coupled plasma optical
the world and has very obvious commercial value as emission spectrometer (ICP-OES) provides a powerful
well as social importance. Therefore, the evaluation of means for fast analysis of a number of elements of the
the quality of wine plays a very important role for both periodic table. There is a problem, however, with how
manufacture and sale. Historically, the quality or geo- to treat the large quantity of data. Fortunately,
graphical origin of wine was determined only through chemometrical methods are able to solve such multi-
tasting by wine experts. However, recently more ad- variate data analysis problems [16, 17]. In the present
vanced instruments have become available. Therefore, paper, cluster analysis, principal component analysis,
it seems reasonable to determine the geographical discrimination analysis methods and artificial neural
networks are applied to wine classification based on
ICP-OES measurement data. Principal component
analysis [18—22], cluster analysis [23—27] and discrim-
Dedicated to Professor Dr. Gerhard Werner on the occasion of his
65th birthday
ination analysis methods [28—31] have found wide ap-
plications in various fields. Recently artificial neural
Li-Xian Sun networks (ANN) [32—43] have become powerful
On leave from the Department of Chemistry, Hunan Normal chemometric tools for modeling complex analytical
University, Changsha, 410081, People’s Republic of China systems and solving those complex non-linear prob-
K. Danzer ( ) · G. Thiel lems in which the relationship between the cause and
Institute of Inorganic and Analytical Chemistry, Friedrich Schiller effect cannot be precisely defined or described by chem-
University Jena, Lessingstrasse 8, D-07743 Jena, Germany ical reasoning. In this paper, trace amounts of
144

B, V, Mn, Zn, Fe, Al, Cu, Sr, Ba, Rb, Na, P, Ca, Mg, K in
wine samples were measured by ICP-OES. Both cluster
analysis and principal component analysis were used to
investigate the structure of the data and the three-layer
ANN model was used to classify the wine samples. The
results obtained by ANN were compared with those
obtained by conventional Bayes discrimination analy-
sis and Fisher discrimination methods.

Chemometric fundamentals

The principles of principal component analysis, cluster

analysis, Bayes discrimination analysis and Fisher dis-
crimination analysis are described in [44—47].
Details of ANN have been extensively explained in
[34, 35, 40, 42, 43]. In the present paper the weights
between hidden node j and input node i are adjusted in
the following way:
*w (q)"ed o #a*w (q!1) (1)
+* 1+ 1* +*
where the meaning of the symbols is the same as in
[31, 32, 37]. In this paper, the output of input node i for
the training pattern p, o in Eq. (1), was equal to the
1*
input signal to node i, which was in fact the corre-
sponding element in vector a , and these input data
1 Fig. 1 Architecture of a three-layer feed-forward ANN model
were all normalized by the Min-Max procedure [35]. p"1, 2, 2 , 170
An equation similar to Eq. (1) was used to adjust the
connection weights between output node k and hidden
transformation for the network’s expected output
node j.
codes c was used to make them lie between 0.20 and
A feed-forward ANN model with three layers of 1,
0.80 for smooth convergency:
nodes was constructed as shown in Fig. 1. Fifteen
element concentration values (mg/L) were measured at c "0.2#0.6(a !a )/(a !a )
seventeen different wavelengths for each wine sample +, +, ,,.*/ ,,.!9 ,,.*/
p (see Experimental section) and the 17-dimensional (j"1, 2, 2 , 170; k"1, 2, 2 , 17) (2)
concentration vector a (a , a , 2 , a ) served as the In order to achieve an optimum learning rate, the
1 11 12 117
input pattern; so the input layer contained 17 input adaptive learning rate method of Vogl [48] was used. If
nodes plus 1 bias node. The output layer represented E (q))E (q!1), then e(q#1) was set equal to 0.70:
the six different geographical origins. The code 1 1
e(q) to reduce the learning rate; if E (q)'E (q!1), then
(1 0 0 0 0 0) represents wine samples from the first origin 1 1
e(q#1) was set equal to 1.05: e(q) to slightly speed up
FS, FR and FM (see Table 1); (0 1 0 0 0 0) those from the learning rate. This method allowed the network to
GW, GR and GS; (0 0 1 0 0 0) those from BG, BS and rapidly stride across the detrimental search space. In
BR; (0 0 0 1 0 0) those from DR, DK and DG; (0 0 0 0 1 0) order to maintain a certain convergency speed of the
those from NR and NM and (0 0 0 0 0 1) those from OS, network with the decreasing of the objective function E ,
OM and OR, outputting through six nodes. 1
a minimum learning rate of 0.10 was set. Earlier inves-
During the network’s learning process, a series of tigations [32, 33, 37, 49] indicated that this rule worked
input patterns (i.e. 17-dimensional concentration vec- better than employing a fixed learning rate during the
tors) with their corresponding expected output patterns whole training process, i.e., the network’s convergency
(i.e. the true class of wine samples) were presented to the speed was increased and the prediction accuracy for
network to learn, with the connection weights between classification or calibration was also improved.
nodes of different layers adjusted in an iterative fashion
by the error-back-propagation algorithm [39, 40].
When the desired precision level for the discrepancies Experimental
between the network’s actual and expected outputs was
achieved, the learning process stopped. The seventeen samples from six different geographical origins are
Because the concentration range of the trace ele- listed in Table 1 along with their production time and corresponding
ments in wine samples varies, the following Min-Max abbreviation labels. The Maxim I (Fisons Instruments) and
 145

Table 1 List of origins of wine samples and labels by three different methods with two different instruments

Origin Plasmaquant 110 Plasmaquant 110 Maxim I Plasmaquant 110

Decomposition Direct method Direct method Standard addition

method method

1. Winzervereinigung Silvaner (FS), 1993 FS(1—3)*DEP! FS(4—6)DIP" FS(7—9)DIM# FS(10)SP$

Freyburg/Unstrut eG. Riesling (FR), 1993 FR(1—3)DEP FR(4—6)DIP FR(7—9)DIM FR(10)SP
Müller-Thurgau (FM),
1993 FM(1—3)DEP FM(4—6)DIP FM(7—9)DIM FM(10)SP
2. Weingut Weißburgunder (GW),
Forschungsanstalt 1993 GW(1—3)DEP GW(4—6)DIP GW(7—9)DIM GW(10)SP
Geisenheim Riesling (GR), 1993 GR(1—3)DEP GR(4—6)DIP GR(7—9)DIM GR(10)SP
Silvaner (GS), 1993 GS(1—3)DEP GS(4—6)DIP GS(7—9)DIM GS(10)SP
3. Staatliches Weinbauinstitut Gewürztraminer (BG),
Freiburg i. Br., Versuchs- und 1993 BG(1—3)DEP BG(4—6)DIP BG(7—9)DIM BG(10)SP
Lehrgut Blankenhornsberg Silvaner (BS), 1993 BS(1—3)DEP BS(4—6)DIP BS(7—9)DIM BS(10)SP
Riesling (BR) 1993 BR(1—3)DEP BR(4—6)DIP BR(7—9)DIM BR(10)SP
4. Staatliches Weinbauinstitut Riesling (DR), 1993 DR(1—3)DEP DR(4—6)DIP DR(7—9)DIM DR(10)SP
Freiburg i. Br., Versuchsrebgut Kerner (DK), 1993 DK(1—3)DEP DK(4—6)DIP DK(7—9)DIM DK(10)SP
Durbach Gewürztraminer (DG),
1993 DG(1—3)DEP DG(4—6)DIP DG(7—9)DIM DG(10)SP
5. Chemisches Untersuchungsamt Müller-Thurgau (NM),
Mainz, Staatliche Lehr- und 1994 NM(1—3)DEP NM(4—6)DIP NM(7—9)DIM NM(10)SP
Versuchsanstalt Oppenheim Riesling (NR), 1994 NR(1—3)DEP NR(4—6)DIP NR(7—9)DIM NR(10)SP
(Gemarkung Nierstein)
6. Chemisches Untersuchungsamt Silvaner (OS), 1994 OS(1—3)DEP OS(4—6)DIP OS(7—9)DIM OS(10)SP
Mainz, Staatliche Lehr- und Müller-Thurgau (OM),
Versuchsanstalt Oppenheim 1994 OM(1—3)DEP OM(4—6)DIP OM(7—9)DIM OM(10)SP
(Gemarkung Oppenheim) Riesling (OR), 1994 OR(1—3)DEP OR(4—6)DIP OR(7—9)DIM OR(10)SP

Note* each sample was measured 10 times and the number(s) in parentheses represent(s) the sequential No. of measurements with certain
methods and instruments (see (a—d)
! FS(1—3)DEP means that the wine sample FS is processed with the Decomposition method and then measured three times (i.e., from 1st—3rd)
by Plasmaquant 110
" FS(4—6)DIP means that the wine sample FS is processed with the Direct method and then measured three times (i.e., from 4th—6th) by
Plasmaquant 110
# FS(7—9)DIM means that the wine sample FS is processed with the Direct method and then measured three times (i.e., from 7th—9th) by
Maxim I
$ FS(10)SP means that the wine sample FS is processed with the Standard addition method and then measured only once (i.e. 10th) by
Plasmaquant 110

Plasmaquant 110 (Carl Zeiss Jena GmbH) ICP spectrometers were STATISTICA 5.0 [51] was used to carry out the cluster analysis
used to measure the concentrations of the fifteen trace elements of of the wine samples. The other programs were written in MATLAB
seventeen samples at 17 different wavelengths. The details of the [52] and run on IBM 486.
analytical program and the wavelengths used can be found in Table
2. To increase the reliability of the analytical results, each sample
was preprocessed using different methods (i.e., direct method, de-
composition method, and standard addition method) as shown in
Table 1 and was measured by both ICP-OES instruments. For Results and discussion
samples in a complex matrix the standard addition method is prefer-
red [50]. In the case of the direct method, standard solutions Principal component analysis
containing ethanol and different amounts of the elements to be
determined were used for calibration. In order to be able to compen-
sate for intensity differences caused by viscosity influences of the In Fig. 2 the first principal component PC1 of wine
organic components during sample nebulization, the organic com- samples is plotted against the second principal com-
pounds of the wine, in particular the alcohol, must be removed from ponent PC2, but the separations among different cate-
the samples. The wine samples were therefore decomposed using gories of wine samples from this PC1-PC2 scatter
perhydrol. This method, however, is very time-consuming. For more
details of these methods, see [50]. Finally, 170 samples were utilized point plot are not obvious. Therefore, the simple
for both ICP-OES measurements and data analysis because each PC1-PC2 scatter point plot appears to be inadequate
original wine sample was measured 10 times (see Table 1). to distinguish the wine samples.
146

Table 2 Measurement
parameters of the instruments Parameter Maxim I Plasmaquant 110
(Firma Fisons Instruments) (Firma Carl Zeiss Jena
GmbH)

Delay time 30 s sample transport 30 s sample transport

10 s signal stabilization 30 s output stabilization
Underground measurement 1 3
Line measurement 3 3
Integration time (s) 3 3
Coolant gas flow (L/min) 13 12
Auxiliary gas flow (L/min) 1.2 1
Carrier gas flow 0.48 (L/min) 200 kPa
Pump rate (mL/min) 3 1.6
Nebulizer type Meinhard-nebulizer cross-flow-nebulizer
Wavelength (nm) used:
B 249.6 249.6
V 292.4 292.4
Mn 257.6 257.6
Zn 206.2, 213.8 206.2, 213.8
Fe 259.9 259.9
Al 308.2, 396.1 308.2, 396.1
Cu 324.7 324.7
Sr 407.7 430.5
Ba 455.3 493.4
Rb 780.0 780.0
Na 589.0 588.9
P 214.9 214.9
Ca 317.9 317.9
Mg 279.0 279.0
K 766.4 766.4

Samples in each group in Fig. 3 are listed in Table 3.

From this table and figure, one can conclude that
samples from Durbach have obvious features which
form a separate group; samples from Freyburg-
Silvaner also tend to cluster into a single group; other
samples from the other five categories (i.e., exception
for Durbach’s samples) show some overlap. It appears
to be difficult to directly classify all wine samples by
only using the Ward’s cluster technique.

Bayes discrimination analysis

The wine samples were classified using the Bayes step-

wise discrimination analysis method as described
above. The leave-one-out method was used to estimate
the discrimination rates of different methods. The
results (Table 4) obtained show that an average
Fig. 2 Plot of the first principal component with respect to the
second principal component, using the principal component analy- discrimination rate for training is 99.4% and the
sis based on SVD for wine samples. corresponding prediction rate is 98.8%. However, two
# samples from Freyburg (i.e., all FS, FR and FM); samples (NR(10)SP and OR(3)DEP) were misclassified.
@ samples from Geisenheim (i.e., all GR, GW and GS);
* samples from Blankenhornsberg (i.e., all BG, BS and BR);
] samples from Durbach (i.e., all DR, DK and DG);
s samples from Gemarkung Nierstein (i.e., all NM and NR); Fisher discrimination analysis
d samples from Gemarkung Oppenheim (i.e., all OS, OM and OR)
Similarly to the Bayes discrimination, the leave-one-
Cluster analysis out method was also adopted for model establishment
and prediction using Fisher discrimination analysis.
Cluster analysis based on the Ward’s method is pre- The first canonical variable CV1 of the wine samples is
sented in Fig. 3, which shows that one can get 6 groups plotted with respect to the second canonical variable
if the tree diagram is cut at a linkage distance of 40. CV2 in Fig. 4, which shows that the separations among
 147

Fig. 3 Tree diagram of cluster

analysis based on Ward’s
method with Euclidean
distances for wine samples

Table 3 Results obtained by using the cluster analysis based on Ward’s method

1st Group 2nd Group 3rd Group 4th Group 5th Group 6th Group

DG(1—3)DEP OM(1—3)DEP BS(1—3)DEP BG(1—3)DEP OR(1—3)DEP FS(1—3)DEP

DG(4—6)DIP OM(4—6)DIP BS(4—6)DIP BG(4—6)DIP OR(4—6)DIP FS(4—6)DIP
DG(7—9)DIM OM(7—9)DIM BS(7—9)DIM BG(7—9)DIM OR(7—9)DIM FS(7—9)DIM
DG(10)SP OM(10)SP BS(10)SP BG(10)SP OR(10)SP FS(10)SP
DK(1—3)DEP NR(1—3)DEP BR(1—3)DEP NM(1—3)DEP OS(1—3)DEP
DK(4—6)DIP NR(4—6)DIP BR(4—6)DIP NM(4—6)DIP OS(4—6)DIP
DK(7—9)DIM NR(7—9)DIM BR(7—9)DIM NM(7—9)DIM OS(7—9)DIM
DK(10)SP NR(10)SP BR(10)SP NM(10)SP OS(10)SP
DR(1—3)DEP GS(1—3)DEP GR(4—6)DIP
DR(4—6)DIP GS(4—6)DIP GR(7—9)DIM
DR(7—9)DIM GS(7—9)DIM GR(10)SP
DR(10)SP GS(10)SP FR(1—3)DEP
GR(1—3)DEP FR(4—6)DIP
FR(7—9)DIM
FR(10)SP
GW(1—3)DEP
GW(4—6)DIP
GW(7—9)DIM
GW(10)SP
FM(1—3)DEP
FM(4—6)DIP
FM(7—9)DIM
FM(10)SP

Table 4 Comparison of the

results from different methods No. of Bayes discrimination Fisher discrimination ANN
based on the leave-one-out class
procedure Recognition Prediction Recognition Prediction Recognition Prediction
(%) (%) (%) (%) (%) (%)

1 100 100 100 96.8 100 100

2 100 100 100 96.8 100 100
3 100 100 100 100 100 100
4 100 100 100 100 100 100
5 95.0 95.0 94.8 85.0 100 100
6 100 96.8 76.9 70.0 100 100
148

Table 5 Effect of the gain (h)!

h 0.1 0.5 0.8 1.1 2.0

Discrimination rate (%) 99.4 100 100 100 99.4

! number of hidden nodes (N)"20, a"0.9

Table 6 Effect of the number of hidden nodes (N)!

N 5 8 10 15 20 30

Discrimination rate (%) 93.5 99.4 100 100 100 94.1

! h"0.8, a"0.9

Table 7 Effect of the number of hidden nodes (N)!

Fig. 4 Graph of the first canonical variable versus the second N 5 10 15 17 20 21 25

canonical variable of wine by Fisher discrimination analysis. The
meaning of the symbols is the same as in Fig. 2 Discrimination rate (%) 64.1 88.8 89.4 94.1 94.7 93.5 91.8

! h"1, a"0.9
different categories of wine samples from the CV1-CV2
scatter point plot are much more obvious than in Fig. 2
or Fig. 3. Therefore, it is possible to classify the wine shows that ANN classifies wine samples from six differ-
samples using this method. The results (Table 4) ob- ent regions with the most satisfactory results.
tained show that an average discrimination rate for
training is 95.3% and the corresponding prediction
rate is 91.8%. Conclusions

Principal component analysis and cluster analysis may

Artificial neural network be used to explore the rough structure of wine data, but
neither method can completely recover the entire in-
Since the gain-in node’s input-output transformation formation included in the data set. ANN shows the
function determines the slope of the sigmoid curve, it same discrimination rate of 100% for both training and
affects the network’s prediction accuracy for the test set prediction with the jackknife leave-one-out procedure
(shown in Table 5). It was found that the network using due to its adaptability; whereas the Bayes stepwise
the leave-one-out procedure gave satisfactory perfor- discrimination method gives only discrimination rates
mance by using a gain value between 0.5 and 1.1 when of 99.4% and 98.8% for training and prediction, re-
the number of hidden nodes is 20. In order to obtain an spectively, and the Fisher discrimination method pro-
optimum ANN model, various network architectures vides only discrimination rates of 95.3% and 91.8% for
were tested as shown in Table 6. The number of hidden training and prediction, respectively. From the point of
nodes (N) was an adjustable parameter that was view of time, ANN takes the longest computing time
optimized by reducing the number used until the net- among all methods. The classification of wine samples
work’s prediction performance deteriorated or the best from different geographical origins by using artificial
prediction accuracy was found. It was found that 10 to neural networks to analyze the measurements of trace
30 hidden nodes plus 1 bias node gave the best predic- elements in the samples using ICP-OES instruments is
tion performance for the wine samples. feasible. Both the Bayes stepwise discrimination analy-
With a gain of 0.8 and 10 to 30 hidden nodes, the sis method and the Fisher discrimination method can
ANN achieves good performance with a prediction rate also be applied with slightly smaller discrimination
of 100%. It must be pointed out that the node’s input- rates, but shorter computation time as compared with
output transformation function has a considerable ef- ANN.
fect on the performance of ANN. If one adopts the
input-output transformation functions: f(x) Acknowledgement Dr. Li-Xian Sun is grateful for a postdoctoral
"1/(1#e(~9@h)) for the output layer and f(x) fellowship from the Alexander von Humboldt Stiftung. The authors
would like to thank Dr. C.-J. Nötzold, Dr. V. Jüngel and K. Weniger
"2/(1#e(~29@h))!1 for the hidden layer, the best (Carl Zeiss Jena GmbH) for the possibility to carry out measure-
prediction rate of the ANN is only about 94.7% with ments on the Plasmaquant 110 ICP spectrometer and for valuable
a gain of 1.0 and 20 hidden nodes (Table 7). Table 4 discussions. We would like to thank Prof. Dr. H. Eschnauer for
 149

supporting this work. We are grateful to Mrs. Schubert in Winzer- 25. Yu RQ, Sun LX, Liang YZ (1995) Classification of materials. In:
vereinigung Freyburg/Unstrut eG., Prof. Dr. J. Dietrich in Kaliva JH (ed) Adaptation of Simulated Annealing to Chemical
Forschungsanstalt Geisenheim, Dr. R. Amann in Staatliches Optimization Problems. Elsevier, Amsterdam, pp 155—179
Weinbauinstitut Freiburg i. Br. and Mrs. Kreisel in Chemisches 26. Sun LX, Danzer K (1996) Journal of Chemometrics (in press)
Untersuchungsamt Mainz for the preparation of wine samples. 27. Shaffer RE, Small GW (1996) Chemometrics Intelligent Labor-
atory Systems 32:95—109
28. Martin MJ, Pablos F, Gonzalez AG (1996) Anal Chim Acta
320:191—197
References 29. Todeschini R, Marengo E (1992) Chemometrics and Intelligent
Laboratory Systems 16:25—35
1. Villen J, Senorans FJ, Reglero G, Herraiz M (1995) J Agric Food 30. Sun LX, Song XH, Xie YL, Liang YZ, Dong YF, Zhao XG,
Chem 43:717—722 Yu RQ (1995) Chin J Hunan Normal University 12:18—23
2. Buiarelli F, Cartoni G, Coccioli F, Levetsovitou Z (1995) 31. Sun LX, Song XH, Yu RQ (1994) Chinese Acta Scientiae
J Chromatogr A 695:229—235 Circumstantiae 14:501—505
3. Capitan Garcia F, Checa R, Avidad R, Capitan-Valley LF 32. Sun LX, Xu F, Tan J, Liang YZ, Yu RQ (1994) Chinese Acta
(1995) Talanta 42:711—715 Pharmaceutica Sinica 29:767—772
4. Zhou JY (1995) Lihua Jianyan, Huaxue Fence 31:42—43 33. Sun LX, Song XH, Yu RQ (1994) China Environmental Science
5. Sasaki, Hideto (1994) Tottori Daigaku Kogakuba Kenkyu 14:259—262
Hokoko 25:169—177 34. Zupan J, Gasteiger J (1991) Anal Chim Acta 248:1—30
6. Yao T, Satomura M, Nakahara T (1994) Talanta, 41:2113—2119 35. Long JR, Gregoriou VG, Gemperline PJ (1990) Anal Chem
7. Kokovkin VV, Smolyakav BS (1995) J Anal Chem 50:519—523 62:1791—1797
8. Gil MI, Garcia Viguera C, Bridle P, Tomas Barberan FA (1995) 36. Wythoff BJ (1993) Chemometrics Intelligent Laboratory Sys-
Z Lebensm Unters Forsch 200:278—281 tems 18:115—155
9. Escobal A, Iriondo C, Laborra C, Ulibarrena E (1995) At Spec- 37. Song XH, Chen Z, Yu RQ (1992) Chemometrics Intelligent
trosc 16:162—164 Laboratory Systems, 16:213—219
10. Soares ME, Bastos ML, Ferreira MA (1995) At Spectrosc 38. Borggaard C, Thodberg HH (1992) Anal Chem 64:545—551
16:256—260 39. Rumelhart DE, Hinton GE, Williams RJ (1986) In: Rumelhart
11. Jorhem L, Sundström B (1995) At Spectrosc 16:226—228 DE, McClelland JL (eds) Parallel distributed processing: explo-
12. Chassange D, Crouzet J, Baumes RL, Lepoutre JP, Bayonove rations in the microstructure of cognition, vol 1. MIT Press,
CL (1995) J Chromatogr A 694:441—451 Cambridge, MA
13. Laber H, Schulz J, Sponholz WR, Bremser W (1995) Fresenius 40. Pao YH (1989) Adaptive pattern recognition and neural net-
J Anal Chem 351:530—535 works. Addison-Wesley
14. Vogels JTWE, Tas AC, van den Berg F, van der Greef J (1993) 41. Ajay L (1994) Chemometrics and Intelligent Laboratory Sys-
Chemometrics Intelligent Laboratory Systems 21:249—258 tems 24:19—30
15. Linskens HF, Jackson JF (1993) Wine analysis. In: Modern 42. Smits JRM, Melssen WJ, Buydens LMC, Kateman G (1994)
methods of plant analysis. New Series, Volume 6 Chemometrics Intelligent Laboratory Systems 22:165—189
16. Danzer K, Venth K (1993) Fresenius J Anal Chem 346:520—524 43. Melssen WJ, Smits JRM, Buydens LMC, Kateman G (1994)
17. Danzer K, Wagner M (1994) Fresenius J Anal Chem Chemometrics Intelligent Laboratory Systems 23:267—291
350:339—343 44. Massart DL, Kaufman L (1988) The interpretation of analytical
18. Einax J, Geiß S, Danzer K (1987) Z Chem 27:332—333 chemical data by the use of cluster analysis. Wiley, New York
19. Einax J, Danzer K (1988) Staub-Reinhalt Luft 49:53—57 45. Sun SG, Pan EP (1990) Practical discrimination analysis.
20. Einax J, Oswald K, Danzer K (1990) Fresenius J Anal Chem Science, Beijing
336:394—399 46. Shatzoff M, Fienberge S, Tsak R (1968) Technometrics
21. Massart DL, Vandeginste BGM, Deming SN, Michotte Y, 10:769—775
Kaufman L (1988) Chemometrics: a textbook, 1st ed. Elsevier, 47. Furnival GM, Wilson RW (1974) Technometrics 16:499—508
Amsterdam, Chapter 21, 22 48. Vogl TP (1988) Biological Cybernetics 59:257—263
22. Alsberg BK, Kvalheim OM (1994) Chemometrics Intelligent 49. Song XH, Yu RQ (1992) Chemometrics Intelligent Laboratory
Laboratory System 24:31—42 Systems 19:101—109
23. Sun LX, Xie YL, Song XH, Wang JH, Yu RQ (1994) Computers 50. Thiel G (1995) PhD Thesis, Friedrich Schiller University Jena
Chem 18:103—108 51. STATISTICA for Windows Tulsa OK, USA
24. Sun LX, Xu F, Liang YZ, Xie YL, Yu RQ (1995) Chemometrics 52. MATLAB for Windows 4.2b, The Mathworks Inc, South
Intelligent Laboratory Systems 25:51—60 Natick 5.0, Statoft Inc, MA, USA