Application of an electronic nose to discrimination of coffees

Abstract

An investigation has been carried out into the response of an array of twelve tin oxide sensors to the headspace of coffee packs. Discriminant and classification function analyses are performed on the array response to each of three commercial coffees (covering two different blends and two roasts) as well as one coffee which has been subjected to a range of six roasting times. Multivariate functions are calculated from the entire data set (90 samples) or alternatively using half of it, to permit cross-validation. A success rate of 89.9% is achieved with the former procedure in classifying the three commercial coffee odours directly from the response (change in sensor conductances) of the array. This value falls to 81.1% when half of the data set is used for cross-validation. Preprocessing the array data, by normalizing the response of each sensor over the array, is found to increase the success rate (to 95.5%) on the entire data set only. The effect on coffee odour of a set of six roasting times (zero to 11.5 min) is also investigated and found to be considerable, some sensors registering an increase in conductance by a factor of three. A 100% group classification is achieved with zero and long roasting times, the overall success rate being 88.1%. The main conclusion is that tin oxide gas sensors can be used to discriminate between both the blend and roasting level of coffee, confirming their potential application in an electronic instrument for on-line quantitative process control in the food industry.

Full text

Sensors and Actuators B, 6 ( 1992) 71-75 71
Application of an electronic nose to the discrimination of coffees
J. W. Gardner, H. V. Shurmer and T. T. Tan
Department of Engineering, University of Warwick, Coventry CV4 7AL (UK)
Abstract
An investigation has been carried out into the response of an array of twelve tin oxide sensors to the headspace of
cotT,ee packs. Discriminant and classification function analyses are performed on the array response to each of three
commerdial coffees (covering two different blends and two roasts) as well as one coffee which has been subjected to
a range of six roasting times. Multivariate functions are calculated from the entire data set (90 samples) or
alternatively using half of it, to permit cross-validation. A success rate of 89.9% is achieved with the former
procedure in classifying the three commercial coffee odours directly from the response (change in sensor conduc-
tances) of the array. This value falls to 81.1% when half of the data set is used for cross-validation. Preprocessing
the array data, by normalizing the response of each sensor over the array, is found to increase the success rate (to
95.5%) on the entire data set only. The effect on coffee odour of a set of six roasting times (zero to 11.5 min) is also
investigated and found to be considerable, some sensors registering an increase in conductance by a factor of three.
A 100% group classification is achieved with zero and long roasting times, the overall success rate being 88.1%. The
main conclusion is that tin oxide gas sensors can be used to discriminate between both the blend and roasting level
of coffee, confirming their potential application in an electronic instrument for on-line quantitative process control
in the food industry.
1. Introduction
A prime requirement of the food industry is a
sensitive method of assessing volatiles, for identifi-
cation, authentication, process control and
product blending or formulation. The aim of the
present work was to demonstrate the feasibility of
using an electronic nose for a range of food appli-
cations [ 1 - 31. Coffee quality is assessed by expert
coffee tasters largely on the basis of its aroma and
flavour, and the highest-quality beans command a
considerable premium. Coffee volatiles are numer-
ous and varied in their aroma quality, potency and
concentration. Most of the volatiles are derived
from initially non-volatile components of the raw
bean, which break down and react during roast-
ing, forming a complex mixture. Pyrolysis, other
reactions and interactions of components, such as
sugars, amino acids, organic acids and phenolic
compounds, result in the formation of the charac-
teristic aroma and flavour of coffee. The final
composition of volatiles depends on a number of
factors, including species/variety of bean, climatic
and soil- conditions during growth and storage,
after both harvesting and roasting, time and tem-
perature of roasting, as well as the roasting equip-
ment used. Green coffee beans are generally
0925-4005/92/%5.00
regarded as having no agreeable aroma or flavour,
but they do possess a large number of volatiles,
most of which increase in concentration during
coffee roasting, although some tend to decrease
through degradation. Green and roasted beans
contain an appreciable amount of aliphatic hydro-
carbons, probably derived from the oxidation of
green bean lipids during storage or transport prior
to roasting. Green beans, on average, contain
about 13% lipid material, three-quarters of this
being triglyceride. There are also appreciable
amounts of diterpene, triterpene and sterol esters,
free diterpenes, triterpenes, sterols and phos-
phatides. Roasting has little effect on the volatiles
produced in this way and causes negligible change
in their overall concentration. In green coffee
beans, it has been reported that 21 hydrocarbons,
10 carbonyls, 10 esters, five sulphur compounds,
two alcohols, four heterocycles and methylindoline
may be detected, including furans, thiophenes, sul-
phides, aldehydes and ketones [4].
Roasting temperature, time, method of roasting
and cooling all affect the volatile composition. The
degree of roast affects the formation and degrada-
tion of the different volatiles and the final coffee
quality. A. study of the effect of roasting time on
the concentration of volatiles has shown that some
@ 1992 - Elsevier Sequoia. All rights reserved

12
reach the peak of their concentration during a
commercial roast, whilst others do not. Some
volatiles decrease significantly when roasting is
prolonged, whereas others increase, including phe-
nols, pyrroles and furfuryl alcohol. Aldehydes,
proponone and total phenols also increase
throughout roasting, slowly at first, then at a
greater rate, without reaching a peak. A more
comprehensive report on coffee volatiles, their re-
actions and their aroma is outside the scope of this
paper, but the details are given by Clarke and
Macrae [4].
2. Pattern-recognition techniques
The procedure employed to analyse the data
was based upon DFA (discriminant function anal-
ysis) and MANOVA (multivariate analysis of
variance) techniques. The experimental data were
initially screened to check for missing values, out-
liers and proximity to normal probability density
functions. Several transformations were assessed
to ensure that any analytical requirements were
met [ 51 as well as to optimize the output from the
sensor array 161.
The main aim of discriminant function analysis
is to predict group membership from a set of
dependent variables or predictors [ 5- lo]. DFA is
concerned with the problem of separating out
different groups on the basis of available informa-
tion, whereas MANOVA determines whether
group membership produces reliable differences on
a combination of predictors. If it does, then the
combination of variables can be used in DFA to’
predict group membership. On the other hand, the
main aim of MANOVA is to determine whether
group membership is associated with reliable
differences in combined dependent variable scores,
whereas in DFA it is to determine whether predic-
tors can be combined to predict group member-
ship reliably.
The general arrangement of the pattern-recogni-
tion system is shown in Fig. 1. The output from
the n-dimensional array of semiconducting sensors
is fed into a preprocessor that defines the optimal
response parameter, given the nature of the sen-
sors and pattern-recognition technique employed.
Features that characterize the odour patterns may
be identified from supervised pattern recognition,
stored in a data base and then extracted for com-
INPUT FEATuuE PATTERN OUTPUT
PIlocESOR - EXlmR -CLASSIFIER 13
SENsoR3
Fig. I General arrangement of pattern-recognition system.
parison with new data to predict or classify un-
known outputs in terms of a coffee type or flavour
attribute from organoleptic tests.
3. Experimental studies
Three commerical types of coffee (blend) were
supplied by Lyons Tetley Ltd. for testing. These
coffee types were as follows, with the letters (a) to
(i) used as labels: (a) a medium roasted coffee of
blend type 1; (b) a dark roasted coffee of blend
type 1; (c) a dark roasted coffee of blend type 2.
Additional coffee types (roasts) were prepared by
the Campden Food and Drinks Research Associa-
tion (CFDRA) from a single coffee blend with a
series of bean roasting times. These were un-
roasted coffee (d); and coffee roasted for 6 min
(e); 7.5 min (f); 9.5 min (g); 10.5 min (h); and
11.5 min (i). An array of 12 different commercial
tin oxide gas sensors with partially overlapping
sensitivities was used to evaluate the coffee
headspace [ 31.
4. Data analysis
4.1. MANOVA
The data were initially screened and found to be
reasonably well behaved, i.e., few outliers and
approximately normal probability distribution
functions. The analysis of the data was carried out
using a commercial statistical software package,
SPSS [9]. MANOVA was used to analyse the data,
confirming that group membership produced reli-
able differences on a combination of predictors,
with the sensor outputs S1 to & as the dependent
variables and the groups of coffee (a)-(c) the
independent variable (one independent variable
with three levels).

4.2. Discriminant function analysis
Discriminant function analysis was carried out
on the sensor array response obtained for the three
commercial coffees (30 samples of coffee (a), 30
samples of coffee (b) and 30 samples of coffee (c))
and the set of roasted coffees (7 samples of coffee
at each roasting time, (d)-(i)).
The raw array output (change in conductance)
and a preprocessed signal, which consisted of the
ratio of change in sensor conductance to the sum
of squares over the array, was analysed for the
three commercial coffees (a)-(c). The latter nor-
malization procedure has been reported as the
optimal response parameter with which to charac-
terize sensor performance [6]. The data were used
in the following manner: (i) all the samples in each
group were used to derive the classification co-
efficients and then used to classify themselves (su-
pervised learning); (ii) half the samples in each
group were used to derive the classification func-
tion and half to cross-validate the classification.
Experimental results are likely to contain some
degree of sensor drift and sample variation within
the data sets. For minimization of errors and, in
particular, guarding against the baseline drift, two
steps were implemented:
(1) Samples from different coffee types were
measured alternately, e.g., eight samples of coffee
type (a) and then eight samples of coffee type (b),
etc. This reduced the chance of falsely discriminat-
ing between groups due to baseline drift.
(2) For the cross-validation analysis, the sam-
ples were randomly chosen, 15 being used for
classification and 15 for cross-validation. The
roles of the classifying samples and the cross-
validation samples were then reversed and the
average of these two analyses was taken as the
final result.
The first method of analysis was then performed
on the raw data. For the commercial coffees, the
grouped data for 30 samples were used to derive
the classification coefficients in the analysis, where
each coffee had a distinct classification function.
Data for each sample were inserted into each
classification equation to develop a classification
score for each coffee. Each sample was assigned to
the group for which it had the highest classifica-
tion score, a process sometimes called jack-knife
classification. The results obtained with all 90 sam-
ples included in the analysis gave an overall
chance of 89.9% success in classifying the samples,
.5
Fist ckwini7ant function. Z
I
Fig. 2. Results of discriminant function analysis and classification of
three commercial coffees (90 samples without cross-validation, some
points are superimposed). Each group centroid is indicated by the
letter c.
coffee type (c) having the lowest success rate for
correct classification. The first two discriminant
functions were calculated and the classified scores
plotted in Fig. 2. When comparing the group
centroids (c), the first discriminant function, Z1,
best separates out the coffee types labelled (a) and
(b) while the second discriminant function, Z,,
best separates out coffee types (a) and (c). Some
dispersion in the individual coffee type (c) scores
caused an overlap with group (a) and thus a lower
degree of discrimination. In this supervised analy-
sis, the success rate for coffee classification im-
proved from a value of 89.9% to 95.5% by
preprocessing the data via the suggested normal-
ization procedure. These results agree with those
from a previous study on the choice of the prepro-
cessing expression for tin oxide gas sensors [6].
The normalization procedure helps reduce the
effect of sample variation.
In the second method of analysis, the classifica-
tion equation was based on only 15 out of 30
samples per coffee. The 15 samples of each coffee
used to derive the classification equation gave a
slight decrease in overall success in classifying
coffee groups (see Fig. 3 for a plot of the individ-
ual scores and group centroids). The success rate is
now lower at 88.9% for classifying samples from
which the equation was derived. This gives success
rates of 93.3%, 93.3% and 80.00/o for coffee groups

74
-5 0 5
Fist discrihhant function. 2
Fig. 3. Results of discriminant function analysis of 45 samples ( +) of
three commercial coffees (with 50% cross-validation shown by the
other symbols with some points superimposed). Each group centroid
is indicated by the letter c.
Fig. 4. Results of discriminant function analysis and classification of
roasted coffee samples (d) to (i). Some points are superimposed and
each group centroid is indicated by the letter x.
(a), (b) and (c), respectively, for correctly classify-
ing the group of samples. When the remaining 15
samples of each group were used for cross-valida-
tion, the success rates became 86.7%, 80.0% and
73.7% for groups (a), (b) and (c), respectively.
When these remaining samples were used as the
calibration samples, they were themselves classified
with success rates of lOO%, 93.3% and 86.7%,
respectively. Cross-validation of these results gave
corresponding figures of 86.7%, 100% and 73.3%.
The average success rates obtained from all these
results were 88.9% for self-classification and 8 1.1%
using cross-validation.
classification was only found to occur for either
group (e) or ( f). Similarly, coffee types (g) and (h)
are closer to each other than to any other coffee,
100% and 7 1.4% being correctly classified for
coffee types (g) and (h), respectively, with mis-
classification only be@veen the two mentioned
groups. Coffee type (i) is well separated from any
other coffees and. ‘Gas classified correctly fbr all
samples.
5. Results and discussion
The coffee types labelled (d) to (i), representing Reference 4 discusses the measured-changes in
different roasting times, were prepared by CD- the concentrations of some key voliitile compo-
FRA and are also analysed. However, the data set nents of coffee over the range of roasting times
could not be fully screened due to there being only used in our experiments, indicating, in particular,
seven samples in each group. The results obtained a rise in the level of furfuryl alcohol and phenol
are nevertheless worth noting, as the percentage of with roasting time. As tin oxide gas sensors are
grouped cases correctly classified was 88.1%. This particularly sensitive to combustible materials, this
compares very favourably with classification by should lead to a change in the array output.
chance, which is only about 16.7%. The group Results for six samples of each of the roasting
centroids, as shown in Fig. 4, tend to move from times (d) to (i) showing the responses (fractional
left to right with increasing roasting time. Coffee changes in conductance) for all twelve sensors are
type (d) is for zero roast, which is well away from indicated in Fig. 5. This confirms the expected
the other coffee types (100% classified correctly). conductivity increase with roasting time for sen-
Coffee types (e) and (f) are closer to each other sors 1 to 11, but sensor 12 appears #to be uninflu-
than to any other coffee (85.7% success and 71.4% enced. Scrutiny of these restilts shows that the
success for groups (e) and (f), respectively). Mis- sensitivity to blend is weaker than it is to roast.

RESPONSE
A
3.6
3
2.6
2
1.6 AS1 TIME
1 n
0.6
0 1 2 3 4 6 6 7 6 0 10 1112
SENSOR NUMBER
Fig. 5. Effect of roasting time on the response (fractional change in
conductance) of twelve Figaro gas sensors to coffee headspace.
6. Conclusions
The feasibility of using an electronic nose to
classify coffee aromas has been demonstrated in
this investigation. The success in classifying three
commercial coffees was greater than 80.0% and the
success in classifying samples prepared by CD-
FRA was greater than 88.1%. Improvement in the
experimental procedures, equipment and design
can be expected to enhance further the discrimi-
nating power of the electronic nose for coffees. A
mass-flow system that automatically acquires the
data with much higher reproducibility and consis-
tency (instead of a static rig using a syringe injec-
tion of headspace) should improve the quality of
the data acquired. Reducing the error in the data
itself then could further improve the success rate
of classifying coffee blends and roasts and, in
addition, there is scope for further advancement of
sensor technology, particularly in the selectivity
and stability of the sensors. The development of
integrated tin oxide sensors and conducting poly-
mer devices are priority items at Warwick and are
expected to facilitate improvements in resolution
[ll, 121.
Acknowledgements
This study was carried out in collaboration with
the Campden Food and Drinks Association as
part of a programme commissioned by the Min-
istry of Agriculture Fisheries and Food. We also
acknowledge the contributions of our colleagues,
Philip Bartlett and George Dodd, to the Warwick
Electronic Nose project.
References
I J. W. Gardner, P. N. Barlett, G. H. Dodd and H. V. Shurmer, in
D. Schild (ed.), Chemosensory Information Processing, Springer,
Berlin, 1990, pp. 13 I- 173.
2 H. V. Shurmer, J. W. Gardner and H. T. Chan, The application
of discriminating techniques to alcohols and tobaccos using tin
oxide sensors, Sensors and Actuators, 18 (1989) 359-369.
3 H. V. Shunner, J. W. Gardner and P. Corcoran, Intelligent
vapour discrimination using 12 element array, Sensors and Acfua-
lors, BI (1989) 256-260.
4 R. J. Clark and R. Macrae (eds.), Coffee: Vol. 1, Chemistry,
Elsevier Applied Science, Barking, UK, 1985, pp. 236-254.
5 B. G. Tabachnick and L. S. Fidell, Using Multivariate Statistics,
Harper and Row, New York, Ch. 4, 1983.
6 J. W. Gardner, Detection of vapours and odours from a multi-
sensor array using pattern recognition Part I. Principal compo-
nent and cluster analysis, Sensors and Actuators B, 4 (1991)
1099115.
1 W. R. Dillon and M. Goldstein, Multivariate Analysis, Methods
and Applications, Wiley, New York, 1984.
8 B. F. Manly, Multivariate Statistical Methods, A Primer, Chap-
man and Hall, London, 1986.
9 SPSS-X User’s Guide, SPSS Inc., USA, 1988.
10 J. Krzanowski, Principles of Multivariate Analysis: A User’s Per-
spective, Clarendon Press, Oxford, 1988.
11 J. W. Gardner, H. V. Shunner and P. Corcoran, Integrated tin
oxide odour sensors, Sensors and Actuators B, 4 ( 1991) 1 l7- 121.
12 H. V. Shurmer, P. Corcoran and J. W. Gardner, Integrated arrays
of gas sensors using conducting polymers with molecular sieves,
Sensors and Actuators B, 4 (1991) 29-33.