The electronic nose applied to dairy products : a review

Sensors and Actuators B 94 (2003) 1–12
The electronic nose applied to dairy products: a review
S. Ampuero, J.O. Bosset?
Swiss Federal Dairy Research Station, FAM, Schwarzenburgstrasse 161, Liebefeld, CH-3003 Bern, Switzerland
Received 20 December 2002; received in revised form 22 February 2003; accepted 3 March 2003
Abstract
The state-of-the-art and current trends in the development of “aroma” analysis with electronic noses are reviewed with special reference to
applications to dairy-products. Some of the reported problems with electronic noses have recently been reduced, e.g. the correction/reduction
of signal drift, the in?uence of humidity and temperature. New promising and reproducible sensor manufacturing techniques are being
implemented, e.g. electro-spray for QMB sensor production. The development of more selective and sensitive sensors, especially of QMB
and conducting polymer (CP) type, should improve their applicability. Interesting novel sampling techniques, such as SPME or SBSE,
offer more possibilities for the analysis of semi-volatile compounds which are generally more odoriferous. However, standard calibration
procedures and reference materials are not yet available. Although they are normally less powerful than human noses, electronic noses offer
some signi?cant advantages in the analysis of volatiles, for example, in instrumental classi?cations based on hedonic or sensory analyses
and in potentially automated on-line monitoring of volatiles. Several groups have explored the application of different electronic noses in
the investigation of various aspects of dairy products. The present review includes as examples the evaluation of Swiss and Cheddar cheese
aroma, the assessment of the ripening of Pecorino Toscano cheese (ewe’s), the detection of mould in Parmesan cheese, the classi?cation of
milk by trademark, by fat level and by preservation process, the classi?cation and the quanti?cation of off-?avours in milk, the evaluation
of Maillard reactions during heating processes in block-milk, as well as the identi?cation of single strains of disinfectant–resistant bacteria
in mixed cultures in milk.
© 2003 Elsevier Science B.V. All rights reserved.
Keywords: Gas sensor; Electronic nose; Arti?cial nose; Dairy product; Group classi?cation; Volatile compound analysis
1. Introduction
cals or biomolecules generally determined by spectroscopy
(e.g. FTIR, NIR, UV-Vis, etc.) [2] and this despite the ex-
Since the ?rst applications of solid state gas sensors in
treme importance of aroma as an indicator of quality and
arrays, some twenty years ago, “electronic noses” have un-
product conformity. This was mainly due to the lack of
dergone a great deal of development. Around a thousand
reliable odour assessing instruments and the practical im-
articles on this subject have been published over the last 4
possibility of employing sensory panels to the continuous
years, mainly in relation to the food and beverage indus-
monitoring of aroma. Electronic noses have the potential to
try [1], but also concerning environmental, agricultural, and
ful?l this task. Compared to sensory panels the main advan-
medical topics, in the automotive industry, etc. However, the
tage of electronic noses is that once calibrated they can per-
number of studies dedicated to dairy products is still very
form odour assessment on a continuous basis with a minimal
limited, probably due to the complexity of their matrices.
cost. Furthermore, once established this technique does not
The aim of the present paper is to review recent exploratory
require trained personnel like a sensory panel does, is not
studies of electronic noses applied to dairy products, in or-
subject to individual breakdown or variation of sensitivity
der to perceive the prospects and trends in this ?eld.
[3], is not overloaded under normal operation and takes com-
Traditionally in the food industry, monitoring of products
paratively very little time.
in terms of quality and control of production processes (e.g.
Before the advent of electronic noses the only possible
mixing, heating, drying, cooking, baking, extruding, fer-
instrumental analysis of “aroma” (the mixture of volatiles
menting, etc.) are performed via physicochemical measure-
present in the headspace of a product) was the identi?ca-
ments, i.e. pH-value, colour, concentration of given chemi-
tion/quanti?cation of individual chemical compounds, af-
ter a separation step (e.g. GC–MS, GC–FID, etc.). How-
? Corresponding author. Tel.: +41-31-3238167.
ever, the relationship between this sequential analysis and
E-mail address: jacques-olivier.bosset@fam.admin.ch (J.O. Bosset).
the perception of the global aroma of a product is not easily
0925-4005/03/$ – see front matter © 2003 Elsevier Science B.V. All rights reserved.
doi:10.1016/S0925-4005(03)00321-6

---

2
S. Ampuero, J.O. Bosset / Sensors and Actuators B 94 (2003) 1–12
established since the rules governing the combination of in-
factory epithelium which is an area of approximately 5 cm2
dividual chemical compounds in the generation of odours
located in the upper nasal cavity. There, the interactions of
are not yet fully understood [4–6].
odorants with the appropriate chemosensory receptors, ol-
It should be kept in mind that instrumental analyses,
factory neurons (?107 belonging to ?103 different classes
whether classical such as GC–MS, etc. or by electronic nose
[6]) produce electrical stimuli which are transmitted to the
are performed not only on odorous volatiles but also on
brain [3,6–9]. A pattern recognition process assisted by the
non-odorous compounds occurring in the headspace. This
memory then takes place using all the data in order to iden-
can be interesting when analysing hazardous non-odorous
tify, classify, or perform an hedonic analysis [9]. Evidence
compounds (e.g. carcinogens, toxins, solvents) but also im-
exists showing that a single olfactory neuron responds to
plies that instrumentally performed classi?cations/analyses
several odorants and that each odorant is sensed by mul-
might not be based on aroma relevant molecules. Further-
tiple olfactory neurons [10]. In the same way, electronic
more, hedonic assessment can not be performed by any in-
noses base the analysis on the cross-reactivity of an array of
strument. Classi?cation models have to be de?ned based on
semi-selective sensors. Hence, products with similar aroma
the results of sensory panels prior to performing analyses
generally result in similar sensor response patterns (sim-
with odour signi?cance.
ilar “?ngerprints”) whereas products with different aroma
show differences in their patterns (different “?ngerprints”).
The sampling step is carried out either by taking an aliquot
2. The electronic nose concept
of the sample headspace, with a syringe, and injecting it
into the detector, or by carrying the headspace with a gas
The name “electronic nose” comes from a certain parallel
stream into the detector. Sometimes the carrier gas is bub-
of the measurement concept of the instrument and that of
bled through the sample to strip out compounds. The inter-
the mammalian olfactory system. In the latter, upon being
action of volatiles with the array of sensors provokes a series
sniffed through the nose, or through the retro-nasal pathway
of signals which are then processed by the computer via a
when a product is tasted, volatile compounds reach the ol-
pattern recognition program.
Table 1
Detection threshold levels of human olfactory systems and electronic noses
Volatile compound
Reported human threshold (ppm)
Electronic nose threshold (ppm)
Type of electronic nose
Reference
Ethyl acetatea
7–17b
5–25
Fox 3000 (12 MOS)
[27]
Butyric acida
0.4–10b
<1
Fox 3000 (12 MOS)
[27]
Diacetyla
(4–15) × 10?3 b
(50–100) × 10?3
Fox 3000 (12 MOS)
[27]
n-Hexanala
(10–50) × 10?3
(10–50) × 10?3
Fox 3000 (12 MOS)
[27]
Methionala
(2–50) × 10?3
(10–50) × 10?3
Fox 3000 (12 MOS)
[27]
Furanola
(20–40) × 10?6 b
(50–100) × 10?6
Fox 3000 (12 MOS)
[27]
n-Nonanec
0.2–7
<0.2
20 CP composite
[13]
n-Octanec
3–9
0.6
20 CP composite
[13]
n-Heptanec
7–13
<2
20 CP composite
[13]
n-Hexanec
13–30
<10
20 CP composite
[13]
n-Pentanec
20–50
40
20 CP composite
[13]
1-Pentanolc
0.13–1.3
<0.06
20 CP composite
[13]
1-Butanolc
0.2–1.3
0.3
20 CP composite
[13]
1-Butanold
0.7
–
Aromascan (32 CP)
[1]
1-Butanold
–
Fox 3000 (12 MOS)
[1]
1-Butanold
+
6 Taguchi (SnO2)
[1]
1-Propanolc
0.9–1.9
1.3
20 CP composite
[13]
Ethanolc
5–500
2
20 CP composite
[13]
Methanolc
13–600
3
20 CP composite
[13]
Acetoned
141
–
Aromascan (32 CP)
[1]
Acetoned
+
Fox 3000 (12 MOS)
[1]
Acetoned
+
6 Taguchi (SnO2)
[1]
Ethanethiold
0.1 × 10?3
–
Aromascan (32 CP)
[1]
Ethanethiold
–
Fox 3000 (12 MOS)
[1]
Ethanethiold
–
6 Taguchi (SnO2)
[1]
+: Detected at the same concentration as submitted to human noses; –: not detected (when response <3× back ground noise) at the same concentration
as submitted to human noses.
a Concentration in water.
b Orthonasal analysis.
c Concentration in air.
d Concentration in vapour in equilibrium with a liquid phase at 22.5–25 ?C.

---

S. Ampuero, J.O. Bosset / Sensors and Actuators B 94 (2003) 1–12
3
A special type of system is slowly appearing in the mar-
contains a heating element. Oxygen from the air is dissolved
ket, the so called portable [4,11,12]. These are small instru-
in the semiconductors’ lattice, setting its electrical resis-
ments where the sensors array is con?ned to a chip. The
tance to a background level (stable when at equilibrium).
analysis proceeds by placing the instrument near the sam-
During the measurement, the volatile molecules (mainly
ple. Portables can be useful in simple and well determined
non-polar) are adsorbed at the surface of the semiconductor
cases, and when interference from the surroundings are mi-
where they react (oxidation/reduction) with the dissolved
nor or constant.
oxygen species causing a further modi?cation of the resis-
Just like the human olfactory system, electronic noses
tance (or conductivity) of the device. This last change is
do not need to be specially designed to detect a particular
taken as the response of the system to that particular sample
volatile. In fact, they can learn new patterns and associate
(Fig. 1) [10].
them with new odours via training and data storage func-
The sensitivity and selectivity of MOS sensors are deter-
tions as humans do. However, training of electronic noses
mined by the choice of the semiconductor material. Mod-
based on sensory panel classi?cations is required in order to
i?cations are induced by doping the semiconductor with
obtain odour-meaningful classi?cations. Often the sensitiv-
noble metal catalysts (e.g. Pt, Pd, Al, Au), by modulating
ity of electronic noses is similar to that of human noses but
the operational temperature (e.g. 200–500 ?C) or by intro-
humans are specially gifted in sensing speci?c compounds
ducing thermal gradients/cycles. Changing the particle size
(e.g. thiols, biogenic compounds, pyrazines, thiazoles, some
and the thickness of the semiconductor ?lm has also been
aldehydes [13]). The biological sensitivity can go down to
tried with the same aim, as well as sensor coating with a gas
ppt levels with a response time in the order of milliseconds
permeable membrane with varying thickness for enhance-
whereas instruments barely go under ppb levels with a re-
ment of the selectivity [17]. Doped sensors show greater
sponse time in the order of seconds (Table 1) [2,14].
sensitivity to oxygenated volatile organic compounds (e.g.
alcohols, ketones, etc) than to aliphatic, aromatic or chlori-
nated compounds [10]. Doping with Pt and Pd increases the
3. Overview of gas sensors: technology and
sensitivity of SnO2 sensors to gases such as benzene and
characteristics
toluene
[10].
The non-selectivity of solid state sensors (metal oxide
Due to the logarithmic dependence of the sensor response
sensors, MOS) was considered a severe drawback of this
on the concentration of volatiles, loss of sensitivity arises
technology intended as analytical tool. Back in the early
(towards low-volatile aroma compounds) in the presence of
1980s the idea of assembling arrays of such sensors with
highly concentrated detectable species such as ethanol [17].
different sensitivities and selectivities was put into practice.
Schaller and co-workers [22–24] have reported large back-
Thus, although both the qualitative and quantitative infor-
ground drift of CP and MOS sensors and MOS sensor poi-
mation obtained from each sensor was highly ambiguous,
soning when attempting to analyse cheese samples of Em-
their combination resulted in some sort of “?ngerprint” of
mental type. The poisoning of sensors was probably due to
the sample. And with the help of statistical programs the
the volatile fatty acids from the cheese. The recent models in
classi?cation of samples into groups could be achieved.
the market seem to be able to correct for drift and they usu-
Once the concept of assembling arrays of non-selective
ally include a temperature and humidity monitoring/control
sensors had been developed, various detection principles
device. Higher operating temperatures apparently make it
were tested, some of them almost accidentally as in the
possible to cope with poisoning as they allow for a better
case of MOSFET [15]. A few of them have given consis-
sensor regeneration after each analysis.
tent results and can be found on the market. Links to pro-
ducers, as well as to university groups performing R&D
3.2. CP
in this ?eld, can be found among others at the web ad-
dress: http://www.nose-network.org/review/. Other types of
Conducting organic polymer sensors (also called in-
devices have also been tried-out such as electrochemical sen-
trinsically conducting polymer (ICP)) are made of semi-
sors, optical ?bres coated with dye-impregnated polymers,
conducting materials, aromatic or heteroaromatic (e.g.
biosensors, etc. Several good papers [3,8,10,14,16–21] pro-
polypyrrole, polyaniline, polythiophene), deposited onto
vide interested readers with a more extensive insight into dif-
a substrate and between two gold-plated electrodes [25].
ferent gas sensor technologies. A brief description of some
Upon interaction with volatile molecules a reversible change
of the commercially available sensors follows.
of the devices’ electrical conductivity is observed.
Although mainly sensitive to polar volatile compounds,
3.1. MOS
their selectivity and sensitivity can be modi?ed by the use
of different functional groups, polymer structure and doping
Metal oxide sensors consist of a metal-oxide semi-
ions [26]. Thus, composites of polymer with thermoplastic
conducting ?lm (e.g. SnO2, TiO2, ZnO, ZrO2) coated onto a
binders or glass ?bres (e.g. polypyrrole with polyimide,
ceramic substrate (e.g. alumina). Most often the device also
polypyrrole with SnO2, or with copper and palladium

---

4
S. Ampuero, J.O. Bosset / Sensors and Actuators B 94 (2003) 1–12
Fig. 1. Working principle of a MOS sensor. R on the y-axis represents the sensor electrical resistance or conductivity. (a) The MOS sensor in presence
of air and at a given temperature. Oxygen dissolves in the sensor lattice setting its electrical resistance/conductivity to a background level. (c) Volatile
compounds, in this case methane, get in contact with the sensor. Upon adsorption/absorption of volatiles on the sensor oxidation/reduction reactions take
place changing the electrical resistance/conductivity of the sensor. The difference between steps (a) and (b) is usually taken as the response of the sensor
to the sample. (c) The sensor is regenerated to the background level under a ?ux of air and is ready to analyse the next sample.
inclusions) show large responses to non-polar volatiles [17].
samples and carrier gas, the system of Neotronics, although
In addition, biomaterials such as enzymes, antibodies, and
easier to operate, showed a reduced sensitivity.
cells may readily be incorporated into polymer structures
[10].
3.3. TSM
A variant of this type of sensors is based on electrically
insulating polymers loaded with carbon black as an electri-
Thickness-shear mode (or QCM quartz crystal microbal-
cally conducting ?ller. When exposed to volatile compounds
ance), BAW bulk acoustic wave, and SAW surface acoustic
the volume of the insulating polymer increases, enlarging
wave sensors, consist of a piezoelectric quartz crystal, with
the distance between the conducting carbon black particles.
gold electrodes, coated with a membrane which, depend-
This results in an increase in the electrical resistance [18].
ing on its af?nity, selectively adsorbs the volatile molecules
Generally polymers readily absorb water vapour and, as a
present. Adsorption of volatile compounds onto the sensing
result, the concentration of available binding sites for other
membrane increases the mass of the device resulting in a
volatiles decreases drastically. This is the reason for the re-
change in its resonance frequency. Selectivity and sensitiv-
duced sensitivity of CP gas sensors at high humidity levels.
ity of this type of sensor depend on the composition of the
Some authors have suggested the implementation of “?lters”
coating membrane (e.g. most frequently polymers but also
to retain undesirable compounds such as ethanol or water
biomolecules or metals) and on the operating frequency. The
prior to analysis [14,27], or during analysis in the case of
difference between SAW and QMB is the mode of oscil-
QMB sensors [28]. The other big drawback of this technol-
lations, at the surface and in the bulk respectively, deter-
ogy is the poor reproducibility in manufacturing polymer
mining the available range of oscillations: SAW operate at
sensors which is a continuing problem. However, CP-based
50–1000 MHz while QMB at 5–30 MHz. SAW devices are
sensors show linear responses and higher selectivities com-
more sensitive but also more unstable and require a high-tech
pared to MOS sensors. In contrast with MOS sensors, no poi-
control set-up.
soning effect with sulphur-containing compounds or weak
Different functional groups can be used as a coating mem-
acids has been observed. They show faster responses and
brane, offering the possibility of tailoring the sensor for the
base-line recoveries, and do not need high operating temper-
detection of speci?c volatiles. This represents a very inter-
atures. A comparative study of sensors done by Harper [29]
esting issue for TSM as well as for CP sensors. Indeed, com-
showed that an Alpha MOS instrument was the least sen-
bined approaches of computer (molecular) modelling and
sitive to water vapour compared to two CP-based sensors,
combinatorial synthesis are undertaken to obtain af?nity se-
one from AromaScan and the second one from Neotronics.
lective sensors [30]. Another approach is the production of
Whereas the system of AromaScan was dif?cult to oper-
molecularly imprinted polymers as very selective traps for
ate due to the complex control of the relative humidity of
speci?c volatiles [31], much in the way of a key and lock

---

S. Ampuero, J.O. Bosset / Sensors and Actuators B 94 (2003) 1–12
5
system that would work even in a noisy background. As an
cases could be complementary to electronic noses. Typi-
example, classi?cation of enantiomers has been reported by
cally, electronic tongues measure attributes such as saltiness,
the use of polymers with chiral functions both with CP and
sweetness, bitterness, sourness and metallic taste [37]. So
QCM sensors [9,17].
far, some interest has arisen for such sensors, for instance
The reproducibility in sensor manufacture is a recurrent
in the pharmaceutical ?eld testing the capacity of masking
problem given the fact that the life-time of sensors is rela-
bitterness in medications which is not readily done with a
tively short (e.g. 6–12 months for MOS and CP). An example
sensory panel for obvious reasons.
of the development of new techniques of coating deposition
is electrospray, instead of spin coating, for the deposition
of uniform ?lms with controlled thickness on QMB sensors
4. Data treatment
[32]. Besides reproducible sensors, there is also a need for
good calibration techniques so as to be able to correlate data
An essential step in the analysis with an electronic nose
obtained with different sensors [29,33,35].
of any kind is pattern recognition. In fact, together with the
Finally, in an attempt to broaden the applicability of gas
progress in electronics, which made possible the develop-
sensors, some companies offer hybrid sensor arrays combin-
ment of sensors, it is the high performance attained by sta-
ing MOS and CP, MOS and MOSFET, MOS and MS, etc.
tistical programs which made possible the introduction of
electronic noses. Although the best performing programs are
3.4. MS
sophisticated and, therefore, require the operation of skilled
personnel, most companies have implemented user-friendly
Mass detection-based electronic noses. Although not pre-
software for data treatment in commercially available elec-
cisely being gas sensors they can be used together with
tronic noses.
chemometric programs to obtain a ?ngerprint of the “aroma”
There exists linear multivariate analysis such as principal
of a product and to proceed to classi?cations. Electronic
component analysis (PCA), discriminant factor analysis or
noses based on mass detection typically use a quadrupole
discriminant function analysis (DFA), non-linear methods
mass spectrometer as a sensor array. Upon injection an MS
such as arti?cial neural networks (ANN) [38,39]. A classi?-
pattern of the unresolved volatiles mixture is created. In
cation can be supervised (e.g. DFA) or non-supervised (e.g.
other words, each mass to charge ratio (m/z) acts as a sensor
PCA), in other words based on predetermined groups or not.
that detects any molecule or fragment with that particular
A parametric classi?cation is very seldom possible. Usually
m/z. In this way, an MS-based electronic nose has potentially
the inclusion into a given group is determined by the Euclid-
hundreds of sensors. Particular fragment ions (m/z) can be
ian or the Mahalanobis distance. The latter takes into account
excluded from the data analysis to remove the in?uence of
the actual shape of the group whereas the former assumes
certain components such as water, ethanol, etc. In the same
that the data points belonging to the group are evenly dis-
way, fragment ions (m/z) aroma-relevant to the case under
tributed in a sphere around the centre of the group (Fig. 2).
study can be selectively chosen to be included in the data
Only a short description of some of the most frequently used
processing; provided that the aroma-relevant compounds are
pattern recognition methods is given here. Readers are re-
known. In this context, the creation of a data base of elec-
ferred to specialised literature for more information.
tronic nose MS spectra as suggested by some authors can
PCA is a linear combinatorial method which reduces the
be very useful, these spectra being different from those ob-
complexity of the data-set, from the initial n-dimensional
tained for individual compounds.
space (n sensors) to a few dimensions. The inherent struc-
A big advantage of this system over all the others is that it
ture of the data-set is preserved while its resulting variance
uses a very well-known technology [35]. The reproducibil-
ity, stability and sensitivity of mass spectrometers have long
been well established. An additional advantage is that dis-
crimination between groups provides with “relevant masses”
[37]. This information may be correlated to corresponding
chemical structures and further studied in combination with
other techniques such as GC–MS. All these features make
this system particularly interesting in the ?eld of R&D. On
the other hand, because of the bench-top type of MS elec-
tronic noses they are not foreseen as portables for in-?eld
applications in contrast to other types of sensors [17,36].
Among related sensor techniques are electronic tongues
which have recently appeared in the market. They are men-
Fig. 2. Representation of Euclidian (solid line) and Mahalanobis (dotted
line) distances. There are two outlayers according to Euclidian distance
tioned here because they also work by classi?cations based
and none according to Mahalanobis distance. A and B are equally far
on “?ngerprints”. They measure organic and inorganic com-
from the group according to Euclidean distance, whereas A is closer to
pounds in liquids (e.g. beverages, foods, etc.) and in some
the group than B according to Mahalanobis distance.

---

6
S. Ampuero, J.O. Bosset / Sensors and Actuators B 94 (2003) 1–12
is maximised. In other words, data points will be scaled
Aroma compounds are typically small hydrophobic or-
along new dimensions, linear combinations of the initial di-
ganic molecules, with a relatively low molecular mass, from
mensions. The magnitudes of the coef?cients, in the result-
30 to 300 amu, and often with a single polar group. Volatil-
ing linear combinations, give an indication of the relative
ity is reduced with higher molecular mass or higher molec-
importance of the initial dimensions in the data structure.
ular polarity. Although little is known about the process
PCA is performed with no information on the classi?cation
underlying odour sensing, there is some evidence that the
of samples. It is based solely on the variance of the data-set.
size and the shape of molecules are more relevant with
On the other hand, DFA is based on a priori data
regard to odour recognition than the chemical function
classi?cation. The linear combinations maximise the con-
or the position of the chemical function in the molecule
tribution of those dimensions that generate the largest dif-
[24].
ference between predetermined groups. With this method,
The concentration of a given compound, i, in the
different classi?cations on the same data-set are possible,
headspace of a sample is given by the partition coef?cient
following different properties (e.g. freshness, fruitiness,
Ki:
etc.). Particular care should be taken however to avoid
Ci(gas)
over-?tting of errors, i.e. classi?cations based on noise
Ki = C
rather than on real differences. The resulting DFA classi?-
i(matrix)
cation is highly dependant on the data-set used for training,
where Ci(gas) and Ci(matrix) are the concentrations of com-
i.e. it is important to verify the size of the training data-set
pound i in the gas phase and in the sample, respectively.
in terms of the complexity of the groups.
By displacing the equilibrium, i.e. by trapping the gaseous
ANN methods are very powerful and are inspired by the
compound i with a polymer for instance, more of this com-
way the mammalian brain processes information. Super-
pound must volatilise to restore the equilibrium. This ex-
vised classi?cations are common with ANN. The non-linear
traction (dynamic headspace sampling) performed prior to
character of this method makes it interesting specially for
analysis will potentially help the detection of compounds
non-linear technologies such as MOS [1]. Future develop-
with a low volatility. In any case, the concentration of most
ments will probably include adaptive neural networking,
of the volatiles will be increased depending on the af?n-
hybrid intelligent models based on Fuzzy-NN or Genetic
ity of the trap for the different volatiles and on the differ-
algorithms-NN, specially aiming at automated industrial ap-
ent equilibria taking place [5,43–45]. Among commercially
plications [40].
available devices for this purpose are: solid phase microex-
In all cases, to avoid classi?cation errors, the ratio of data
traction (SPME), stir bar sorptive extraction (SBSE), etc.
points (analyses) to variables (sensors), employed in the pat-
The SPME is a ?bre of fused silica coated with 1–3 poly-
tern recognition, should be at least three, but preferably six
mers. The ?bre is carried into a needle and is exposed only
[36]. Thus, replication of samples ?ve to eight times is usual
for physical–chemical sorption/desorption of volatiles dur-
in order to obtain a large number of data points as well as
ing sampling and during measurement. The twister (SBSE)
to ascertain the repeatability of the measurement. Since in-
is a magnetic bar coated with polymers, which can be held
dividual analyses are performed within a few minutes, or
in the headspace for sampling. Its loading capacity is much
even seconds with any of the systems currently available,
higher than that of SPME. The coatings can be chosen de-
and also because automatic sampling devices are generally
pending on the polar groups or the size of the targeted com-
implemented, the total analysis time remains competitive
pounds. Among available polymers used for coatings are:
compared to other analytical techniques such as GC–MS.
polydimethylsiloxane (PDMS), polyacrylate (PA), carboxen
Model validation with data points not used for the genera-
(CAR), etc. Other devices use porous traps such as Tenax.
tion of the model is also recommended [36], as well as ran-
More information is available in the specialised literature
domisation of sample analyses to avoid systematic errors.
and is not given here as this is beyond the scope of this ar-
Some authors have tried data pre-treatment to reduce noise,
ticle.
and data correction to target a given classi?cation by the use
The effect of the matrix on the release of aroma is well
of reference standards chosen among the key compounds
known. Steinhart and co-workers [5,46] showed that there is
under study [41,42]. Although so far absolute calibration is
a decrease in the concentration in the headspace of volatile
still lacking, calibration with reference compounds allows
2,3-ethyl-5-methalpyrazine (roast smell) with an increase
for correction of drift as well as for instrument to instrument
in the fat level in a coffee matrix. In general, most or-
matching [29,34,38].
ganic ?avour compounds are readily adsorbed and solu-
bilised in lipids, depending on their lipophylic character.
Proteins present in the matrix may in?uence the volatility
5. Additional factors affecting the analysis
of ?avour compounds via weak Van der Waals interactions
or by the formation of amides, esters and salts. Polysaccha-
Very often the importance of the sampling step is ignored.
rides can hinder the volatility of certain compounds whereas
Nevertheless the quality of the analysis can be greatly im-
other carbohydrates such as mono and disaccharides may
proved by adopting an appropriate sampling technique.
cause a salting-out effect [5,46].

---

S. Ampuero, J.O. Bosset / Sensors and Actuators B 94 (2003) 1–12
7
Other parameters that in?uence the volatility of com-
ture program between 150 and 180 ?C. The measurement of
pounds are: temperature (most samples release volatiles bet-
milk volatiles lasted not more than 7 min. Eighty-four sam-
ter at higher temperatures), equilibration time and to a lesser
ples of milk and 73 samples of chocolate milk were used,
extent pressure, pH-value (molecules can pass from a polar
taken directly from production lines over a 7-month-period.
state to a non-polar state and vice versa by changing the pH
Twenty samples of each set were used for model validation
[47]), ionic concentration (sometimes the addition of a salt
solely. Mass intensities were normalised by the intensity
provokes a “salting-out effect”, increasing the volatility of
of chlorobenzene (m/z 112). Partial least-squares modelling
certain compounds), surface area (grinding of a solid, mix-
predicted sample shelf-life with respect to sensory results
ing of a liquid). When analysing with electronic noses, the
with an accuracy of ±0.62 and ±0.88 days and a correla-
aim is very often the classi?cation of samples into differ-
tion coef?cient of 0.9801 and 0.9832 for milk and chocolate
ent groups. In such a case it is important to maximise dif-
milk, respectively. A signi?cant increase in concentration
ferences even if samples are slightly denatured during the
with ageing was detected by GC–MS for certain volatiles
analysis (by temperature, changes in pH, etc.) as long as the
such as: dimethyl sulphide, 2-heptanone, ethyl acetate,
same treatment is used for all samples.
pentanal, pyrrolidine, hexanoic acid, 2-methyl-butanal, fur-
Finally, changes in atmospheric temperature and humidity
furaldehyde, etc. [51].
may in?uence not only the sensors response but also the
concentration of volatiles in the gas phase.
6.2. Classi?cation of off-?avours in milk
Oxidation off-?avours in milk originate mostly from bac-
6. Applications to the dairy industry
terial metabolism, enzymatic activity, photo-oxidation, heat,
and oxidation catalysed by chemicals such as sanitizers for
A list of recently reported applications of electronic noses
production lines or pro-oxidant metals (copper, iron and
to dairy products is given below.
nickel) [51]. Exposing milk to light induces two major ef-
fects: The ?rst 2–3 days a burnt oxidised ?avour develops
6.1. Ageing of milk and shelf-life prediction
probably due to the degradation of sulphur-containing amino
acids from the whey into methional (relatively unstable),
The headspace of milk typically presents a complex
mercaptans, sulphides and disulphides (e.g. dimethyl disul-
mixture of organic volatiles (e.g. acetone at overwhelm-
phide from methionine) [52]. After the second day a per-
ing concentration, hexanal, 2-butanone, toluene, limonene,
sistent metallic, cardboard-like off-?avour occurs, attributed
heptanal, styrene, chloroform, etc.) at varying concen-
to the autoxidation of unsaturated fatty acids (?-oxidation)
trations and with a high percentage of relative humidity.
by the formation of free radicals induced by light. On the
Furthermore, the matrix is highly heterogeneous containing
other hand, heat provokes a typical boiled off-?avour prob-
different levels of lipids, proteins and carbohydrates.
ably resulting from the formation of sulphur compounds.
Correct classi?cation of groups of different ages was ob-
Great efforts have been devoted to the optimisation of UHT
tained for UHT [48,49] and pasteurised milk [48]. Groups
milk processing in order to avoid this effect. In general,
of UHT milk aged 1–8 days, and 1–3 days of pasteurised
low molecular weight aldehydes, ketones and fatty acids
milk were classi?ed by PCA performed on normalised data.
are responsible for most off-?avours observed in foods and
A home-made MOS sensors array was used (?ve SnO2 thin
beverages of which hexanal is the major by-product of the
?lm sensors, of which four were doped with Ni, Os, Pt and
degradation of linoleic acid (major polyunsaturated fatty
Pd). 10 ml of milk were placed in 20 ml vials, four to ?ve
acid in milk) upon exposure to light.
measurements per sample. Samples were incubated 15 min
Marsili has published several papers on the discrimina-
at 30 ?C during analysis, the headspace was carried into the
tion of off-?avours in milk [42,45,50–52]. The system used
injector with a ?ow of N2 (where the total gas ?ow was
was a prototype SPME–MS–MVA system above described.
100 sccm N2/100 sccm dry air), sensor temperature was set
PCA correctly classi?ed the set of samples by the origin of
at 250 ?C. Sensors showed a response and recovery time of
off-?avours: sanitizer-contaminated, copper-contaminated,
2–3 min.
spoiled by bacteria and fresh 2%-fat milk samples. Prior
Shelf-life prediction of 2%-fat pasteurised milk and
to analysis, samples were conditioned at 19 ?C for 16 h
whole-fat chocolate milk was obtained with a home-
except for fresh milk control samples, and then heated at
assembled SPME–MS–MVA system (solid phase micro-
45–50 ?C during the 12–15 min of ?bre exposure (75 ?m
extraction,
mass
spectrometry,
multivariate
analysis)
Carboxen/PDMS SPME ?bre). Masses corresponding to
[50,51]. 3 ml of milk sample and 5 ?l internal standard
volatiles normally present in non-defective milk were dis-
(10 ?g/ml chlorobenzene) were placed in 6 ml vials. A
regarded (e.g. acetone, 2-butanone) and intensity of masses
75 ?m Carboxen/PDMS SPME ?bre was used. Samples
from target compounds were ampli?ed (e.g. m/z 127, 142
were incubated 20 min at 50 ?C during ?bre exposure, oth-
and 94, this last one corresponding to dimethyl disulphide).
erwise they were stored at 7.2 ?C. The injector temperature
The intensities of all masses were normalised by the in-
was set at 275 ?C and the transfer line followed a tempera-
tensity of the internal standard (4-methyl-2-pentanone, m/z

---

8
S. Ampuero, J.O. Bosset / Sensors and Actuators B 94 (2003) 1–12
100). A fairly steady increase in pentanal and hexanal was
concentrating stages, from the initial raw material (milk
observed with increasing exposure time to 200 ft3 of ?uo-
and sugar mixtures with 20 wt.% dry matter) to the ?nal
rescent light (typical light exposure of milk in supermarket
block-milk product (97.8 wt.% dry matter). 60.0 g of sample
dairy cases) [52]. Typically, hexanal and dimethyl disul-
were mixed with an equal amount of MilliQ water just be-
phide were found to be good indicators of light damaged
fore analysis. No control of the relative humidity in the elec-
milk, whereas pentanal, isopentanal, hexanal, heptanal, oc-
tronic nose was undertaken. The response values were taken
tanal, nonanal, and 1-octen-3-one usually indicated copper
after 3 min of signal recording. The results correlate well
induced oxidation in milk [45].
with GC–MS and sensory analysis. PCA