You are on page 1of 8

FLAVOUR AND FRAGRANCE JOURNAL

Flavour Fra`r[ J[ 1999^ 04] 104Ð111

Detection of perfumes in diesel fuels with semiconductor


and mass-spectrometry based electronic noses
Roger Feldhoff,1 Claude-Alain Saby2 and Philippe Bernadet1
1
Elf-Aquitaine, Groupement de Recherches de Lacq, B.P. 34, 64170 Lacq, France
2
Elf Antar France, Centre de Recherches d’Elf Solaize, B.P. 22, 69360 Saint-Symphorien d’Ozon, France

Received 5 June 0888


Revised 0 February 1999
Accepted 0 February 1999

ABSTRACT] The gas!phase of two diesel fuels scented with perfume mixtures were examined with a FOX 3999
electronic nose "Alpha!M[O[S[\ France# and a Smart Nose GA "LDZ Laboratory\ Switzerland#[ The _rst system
uses chemical gas sensors\ the second a mass spectrometer[
The data of both instruments were evaluated by principal component analysis "PCA# and linear discriminant
analysis "LDA#[ The LDA results obtained allowed the detection of a much lower perfume concentration with
the mass spectrometry data than with the chemical sensor data[ Copyright Þ 1999 John Wiley + Sons\ Ltd[

KEY WORDS] diesel fuel^ gas!phase^ headspace^ volatile compounds^ detection limit^ perfume^ electronic nose^
chemical sensors^ mass spectrometry^ principal component analysis^ discriminant function analysis

Introduction gas phase is normally analysed by gas!chromatography


"GC# or gas chromatography coupled with mass
The aim of this article is to assess the usefulness of spectrometry "GC!MS#[
electronic noses as a complementary tool to odour Electronic noses could equally be used to measure
panels[ Electronic noses can certainly not replace the changes of the gas phase of chemical products[ The
human nose because they do not respond to the same advantages of electronic noses over classic GC are their
molecules[ On the other hand\ they are much faster\ simpler use and higher speed[ The greatest disadvantage
more repeatable and less expensive than expert panels[ is that chemical compounds can not be analytically
Potential applications of these devices can be product identi_ed[
or quality control in the chemical industry\ in particular Electronic noses have been proposed for quality
for health damaging products[ control of foodstu}4Ð6 or "alcoholic# beverages[7Ð09 They
The _rst study done in our laboratory describes were also applied to industrial food product control[00
the distinction of diesel fuels coming from di}erent Other papers give a more general overview about
re_neries[0 Chemical sensors and two mass spectrometer electronic nose technology and applications[01\02 Further
based electronic noses were used and compared[ The studies\ one of them carried out in our laboratory\
results obtained with both techniques were similar[ applied an electronic nose to perfumes[03\04 There are
Strong correlations between the mass spectral data and two articles which describe the application of an elec!
the chemical sensor responses were observed[ The mass tronic nose to fuels[05\06 The _rst one discusses the quali!
spectrometers showed a better reproducibility and were tative distinction of several automotive fuels "with
much more practical for this qualitative distinction of di}erent octane values#[05 The second one concentrates
diesel fuels[ on forensic and law enforcement applications[06
The goal of this study is to _nd out in which con!
centrations a perfume mixture can be detected in a
complex matrix like a diesel fuel[ The results obtained
with chemical sensors and mass spectrometers shall be
compared[ Experimental Conditions
The liquid phase of oil and oil derivatives are mostly
analysed by supercritical ~uid chromatography\1 mass Since chemical sensors and mass spectrometry are of a
spectrometry2 or ~uorescent indicator absorption[3 The very di}erent nature it was not possible to use the same
working conditions for both machines[ However\ to
 Correspondence to] R[ Feldho}\ Elf!Aquitaine\ Groupement de Recherches
compare the analytical potential of these techniques
de Lacq\ B[P[ 23\ 53069 Lacq\ France[ it seems to be more appropriate to optimise working

Copyright Þ 1999 John Wiley + Sons\ Ltd[


105 R[ FELDHOFF\ C[!A[ SABY AND P[ BERNADET

conditions for both instruments independently[ This is because they have a higher long!term stability and are
what was done in this study[ less sensitive to humidity ~uctuations than their con!
ducting polymer!based homologues[ The sensor signals
Chemical Sensors were recorded for 599 s with an acquisition point each
second[
The chemical sensor based instrument used was a FOX
3999 electronic nose "Alpha!M[O[S[\ Toulouse\
France#[ It was equipped with 07 semiconductor Mass Spectrometer
sensors[ The samples were passed by a HS499 auto! The second instrument used\ a Smart Nose GA 199\
sampler "CTC Analytics AG# with a capacity of 49 was based on a mass!spectrometer[ The corresponding
samples[ To provide a constant ~ux of dry air through
measurements were carried out by LDZ Laboratory\
the electronic nose\ a humidi_er "air condition unit CH!1963 Marin!Epagnier\ Switzerland[ The system|s
ACU\ model 0886\ Alpha!M[O[S[# was used[ The air detection unit consisted of a Balzers quadrupole mass
condition unit was supplied with synthetic air "nitro! spectrometer with 089 channels "from 09 to 088 amu#[
gen:oxygen R79#[ Its air ~ux was set to 299 ml:min\ the
The samples were 19 drops "about 079 mg# of diesel fuel
relative humidity was regulated between 16[1) and
_lled in 09 ml vials[ The injection unit was a CTC
18[4)\ the temperature varied between 38[6>C and Headspace Autosampler PAL[ The samples were incu!
49[1>C[ The dry air vector gas was passed via a 0 l bu}er bated at 49>C for 4 min "with agitation#[ The syringe
bottle before entering the electronic nose[ The FOX
and injector temperatures were 099>C and 019>C respec!
3999 instrument was set to 149 ml:min vector gas con!
tively[ After each injection of 1[4 ml of sample head!
sumption[ A third exit tube of the bu}er bottle allowed space\ the syringe was purged for 6 min and the injector
the super~uous air to escape[ The whole setup was for 09 min with nitrogen "9[4 bar#[ The acquisition time
mounted in an air!controlled laboratory with an ambi!
per sample was 09 min[ For data examination three
ent temperature of 10>C[ Careful instrument main! scans per sample were averaged[ Each diesel type was
tenance was necessary to obtain reproducible results
measured four times[
"for more details see reference ð0Ł#[
The samples "09 ml gas vials _lled with 099 ml of diesel
fuel# were passed in arbitrary order[ The vials were Samples
heated to 24>C by the autosampler "incubation time 29
Most car drivers _nd the smell of diesel fuel unpleasant[
min with agitation#[ The temperature of the injection
This circumstance leads many petrol companies to add
syringe was 39>C[ The syringe was _lled twice before
perfumes to their products[ The perfumes mask the
taking 399 ml of the sample headspace and injecting it
odour of heavily smelling volatile compounds[
into the FOX instrument[ Syringe _ll speed and injec!
The samples used were two di}erent diesel fuels] F91
tion speed were 099 ml:s and 0499 ml:s\ respectively[
"a winter diesel from the {elf|!re_nery at Feyzin# and
After each injection the syringe was baked out for 3 min
G00 "a summer diesel from the re_nery at Grandpuits#[
at 49>C[ It was then ~ushed with nitrogen "N 49# for
These two diesel fuels were scented in di}erent con!
5 ]29 min[ The delay time between two injections was 0 h
centrations with either an {elf| perfume mixture or with
29 min[ This comparatively long interval was necessary
a complex additive[ The additive is a mixture of a deter!
because the return of some of the gas sensors to their gent\ an anti!foam\ an anti!corrosion protection and 7
initial values was extremely slow[ The volatile com! g:099 g of the perfume[ The perfume concentration and
pounds contained in the headspace obviously caused a coding of these samples are listed in Tables 1 and 2
long desorption process[ This is probably related to the
Even at the lowest additive concentration used
viscous nature of diesel fuel with long!chained com! "sample F91A9979# a signi_cant odour di}erence was
pounds "09Ð19 carbon atoms#[ observed to the non!scented sample[ This was con_rmed
The 07 sensors of the FOX electronic nose are listed by an odour panel consisting of eight persons[ The
in Table 0[ We only used "doped# metal!oxide sensors
scented sample was described as more pleasant by the
panellists[ Odour descriptions corresponding to bad
Table 1. Metal-oxide sensors of the three chambers (A, smells obtained a lower weight[
B, C) of the FOX 4000 electronic nose The odour di}erences between the same two samples
Position Chamber A Chamber B Chamber C were equally con_rmed by gas!chromatography com!
bined with "human nose# sni.ng "GC!OLF#[ The odour
0 T29:0 P29:0 SY:LG
1 P09:0 P39:1 SY:G
pro_le obtained by headspace injection showed clear
2 P09:1 P29:1 SY:AA di}erences "not shown#[ On the other hand\ the pure
3 P39:0 P69:0 SY:Gh GC data "not shown# did not allow for any qualitative
4 T69:1 T39:0 SY:gCTl
5 PA1 TA1 SY:gCT
distinction between them[ There might be small di}er!
ences in some peaks which could not be resolved due

Copyright Þ 1999 John Wiley + Sons\ Ltd[ Flavour Fra`r[ J[ 1999^ 04] 104Ð111
DIESEL FUELS SCENTED WITH PERFUME MIXTURES 106

Table 2. F02 Diesel samples with different perfume and additive concentrations
Final Samples scented with the additive Samples scented with the pure perfume
perfume
concentration ðml:lŁ Code Quantity added ðml:lŁ Code Quantity added ðml:lŁ

9 F91 9 F91 9
79 F91A9979 0999 F91P9979 79
059 F91A9059 1999 F91P9059 059
219 F91A9219 3999 F91P9219 219
539 F91A9539 7999 F91P9539 539
0179 F91A0179 05 999 F91P0179 0179
09 999 F91A0) 014 999 F91P0) 09 999

Table 3. G11 diesel samples with different perfume and additive concentrations
Final Samples scented with the additive Samples scented with the pure perfume
perfume
concentration ðml:lŁ Code Quantity added ðml:lŁ Code Quantity added ðml:lŁ
9 G00 9 G00 9
79 G00A9979 0999 G00P9979 79
059 G00A9059 1999 G00P9059 059
219 G00A9219 3999 G00P9219 219
539 G00A9539 7999 G00P9539 539
859 G00A9859 01 999 G00P9859 859
0179 G00A0179 05 999 G00P0179 0179
0819 G00A0819 13 999 G00P0819 0819
1459 G00A1459 21 999 G00P1459 1459

to the low quantitative reproducibility of the manual group\ the samples with the pure perfume seem to form
injection method used[ The perfume consisted of many a separate cloud that touches a second cloud containing
di}erent compounds and its total concentration was the pure diesel fuel and the diesel with additive[ The
only 79 ppm in this sample[ The GC!peaks were too samples A0179 are all found at the right border of the
small to be detected or they were hidden by the numer! additive cluster but there is still no distinct separation
ous GC!peaks from the diesel itself[ from the main cloud at this perfume concentration[
Figure 1 shows the corresponding results for G00[
Again PCs 0 and 2 were used[ The relative weights of
Results and Discussion the _rst three principal components are PC0 83[3)\
PC1 1[0) and PC2 0[1)[ Apart from samples A0179\
Chemical Sensor Results with PCA A0819 and A1459 all other samples form a big cloud[
PC0 and PC2 both increase with additive concentration[
For data examination\ we used DRmax:R9\ which is the In PC2\ one might equally suspect a slight concentration
maximum change of the sensor|s electrical resistance in~uence of the perfume "see samples P1459 at the upper
divided by the initial resistance[ This value gives the border of the main cloud#[ PC1\ again\ did not contain
most stable results with respect to sensor baseline vari! much useful information for the distinction of the
ations[ di}erent samples[
A principal component score plot "PCs 0\ 2# for the In Figure 1\ two populations of data points can be
measurements done with diesel F91 is shown in Figure observed] a diagonal line with positive gradient\ starting
0[ PCs 0 and 2 were chosen for this representation from value 9[009 on the x!axis\ can be imagined to split
because PC1 did not contribute much to the separation the data into two groups[ This can best be seen on the
of the di}erently perfumed samples[ The importance of samples with higher additive concentrations[ The reason
the _rst three principal components are 78[8\ 2[3 and for this phenomenon was a spontaneous baseline shift
0[7)\ respectively[ According to the small importance of the chemical sensors during the experiment[ This
of PC2 with respect to PC0\ the data!distribution is baseline shift probably occurred due to the insu.cient
nearly one!dimensional[ PC0 expresses the di}erences temperature control of the instrument[
in additive concentration whereas PC2 contains infor!
mation about the diesel|s pure perfume content[ Only Chemical Sensor Results with LDA
three groups can be separated] samples with 0) of the
additive\ samples with 0) of the pure perfume and a How can the data model be optimized< One method to
rest group with lower perfume contents[ Within this rest improve the separation is linear discriminant analysis

Copyright Þ 1999 John Wiley + Sons\ Ltd[ Flavour Fra`r[ J[ 1999^ 04] 104Ð111
107 R[ FELDHOFF\ C[!A[ SABY AND P[ BERNADET

Figure 1. Principal component score plot for the (perfumed) diesel F02 as obtained with the FOX 4000 instrument (all
sensors used). F02 pure, P F02 with perfume, A F02 with (perfume containing) additive

Figure 2. Principal component score plot for the (perfumed) diesel G11 as obtained with the FOX 4000 instrument (all
sensors used). G11 pure, P G11 with perfume, A G11 with (perfume containing) additive

"LDA#[07\08\19 However\ the stability of LDA models the additive and PC2 about the perfume[ PC0 shows
strongly depends on a careful feature selection[0 very similar patterns in both _gures\ with strong con!
To select the most important sensors\ we looked at tributions from B1 and C1!3[ The similarities for PC2
the sensors contributing to the PCA models[ These are are less distinct[
shown in Figures 2 and 3 for F91 and G00\ respectively[ We then developed a LDA model only built on the
Since PC1 did not contribute to the separation of the sensors important for PCs 0 and 2[ For this\ we chose
di}erent samples\ we concentrated on PCs 0 and 2[ We those sensors which in both PCA models have a load
assume that PC0 mainly contains information about higher than 0[4[ According to this threshold value\ the

Copyright Þ 1999 John Wiley + Sons\ Ltd[ Flavour Fra`r[ J[ 1999^ 04] 104Ð111
DIESEL FUELS SCENTED WITH PERFUME MIXTURES 108

Figure 3. Loads of the first three principal component vectors for (perfumed) diesel F02 as obtained with the FOX
4000 instrument

Figure 4. Loads of the first three principal component vectors for (perfumed) diesel G11 as obtained with the FOX
4000 instrument

sensors retained from PC0 are A5\ B0\ B1 and C1!C3[ The corresponding discriminant function score plot
The main contributors to PC2 are sensors A0\ A3!5\ for F91 is shown in Figure 4[ The inner!class variances
B0!2 and C1\ C2[ The union of these two groups\ thus\ are smaller than obtained with PCA[ However\ the sep!
contains sensors A0\ A3!5\ B0!2 and C1!3[ LDA models aration of single groups is no better than before[ There
based on these sensors were computed for both\ F91 are again two clouds\ one contains only samples with
and G00[ In both cases there were software warnings perfume and the other only samples with additive[ The
that the matrix was close to singular[ This shows that pure diesel is found just in the middle between these
the observed di}erences between samples are quite small two clouds[ DF0 "70[6) of information# describes the
and are close to the instrument|s detection limits[ additive content and DF1 "04[9) of information# the

Copyright Þ 1999 John Wiley + Sons\ Ltd[ Flavour Fra`r[ J[ 1999^ 04] 104Ð111
119 R[ FELDHOFF\ C[!A[ SABY AND P[ BERNADET

Figure 5. Linear discriminant score plot of (perfumed) diesel data as obtained with the FOX 4000 instrument (sensors
1, 4–9, 14–16 used). F02 pure, P F02 with perfume, A F02 with (perfume containing) additive

perfume content[ In the case of the samples with additive Also in the case of G00 "Figure 5#\ the inner!group
there is a tendency towards a concentration dependence variance is much smaller with DFA than with PCA[
for sample concentrations above 219 ppm[ The lowest The sample separation is much better now[ Not only
additive concentration "well separated# could be sug! the samples with high additive concentrations but also
gested to be about 0179 ppm[ For the pure perfume it with high perfume content can now be distinguished
is not possible to de_ne a precise detection limit[ It from the weaker perfumed samples[ Starting from a
should be considerably higher than 1999 ppm but still main cloud containing weaker perfumed samples\ the
below 09 999 ppm "0)#[ additive concentration increases with DF0 "importance

Figure 6. Linear discriminant score plot of (perfumed) diesel data as obtained with the FOX 4000 instrument (sensors
1, 4–9, 14–16 used). G11 pure, P G11 with perfume, A G11 with (perfume containing) additive

Copyright Þ 1999 John Wiley + Sons\ Ltd[ Flavour Fra`r[ J[ 1999^ 04] 104Ð111
DIESEL FUELS SCENTED WITH PERFUME MIXTURES 110

41[4)#\ the pure perfume concentration with DF1 reasonable results and give an idea about the possi!
"importance 39[1)#[ The lowest concentrations "well bilities and the limits of chemical sensor based electronic
separated# are 0179 ppm in the case of the additive and noses[ Because of the closeness of the discussed
between 0179 ppm and 0819 ppm in the case of the measurements to the instruments detection limits we did
perfume[ not continue to use this kind of electronic noses for
It is remarkable that the behaviour of the two diesels this application[ Another reason was that "spontaneous#
is quite di}erent[ In the case of F91\ the samples with instrument drifts were observed[
low perfume and low additive concentrations form two
di}erent clouds[ In the case of G00 there is a "linear#
Mass Spectrometry Results
concentration dependence in both cases[ The detection
limits for the additive are about the same with F91 and The F91 samples were also investigated with a Smart
G00 "about 0179 ppm#[ In the case of the pure perfume\ Nose "mass spectrometer based# electronic nose[ The
the detection limit is much higher with F91 "signi_cantly whole dataset "09 to 088 amu# was normalized on mass
more than 1999 ppm# than with G00 "0179Ð0819 ppm#[ 39[ A PCA was then performed on masses from 55 to
This observation is underlined by the fact that the 062\ which were the channels with the most important
importance of DF1 is much bigger in the case of G00 information "see also reference ð0Ł#[ The result of the
"39[1)# than in the case of F91 "04[9)#[ A reason for PCA "not shown# was poor[ No sample discrimination
this di}erence in behaviour might be that F91 is a winter was possible[
diesel "W# and G00 is a summer diesel "S#[ Another\ Then a feature selection by subsequent elimination
maybe more probably reason\ may be the di}erent of less important masses was performed[ The retained
chemical composition of the two fuels due to their masses were 62\ 79\ 76\ 77\ 82\ 83\ 090\ 097\ 098\ 011
di}erent origins[ In an earlier study we found that the and 025[ After range!scaling\ a LDA was performed[
sample origin plays a more important role on the chemi! The results are shown in Figure 6[ The sample with 0)
cal composition than the fuel type "see reference ð0Ł#[ perfume was left out to obtain a more convenient axis
Signi_cant di}erences in nitrogen content between the scale[ The distribution is nearly one!dimensional with
samples from Feyzin "079Ð149 ppm# and Grandpuits relative importances of 86[1) for DF0 and only 1[2)
"about 49 ppm# were found[ for DF1[ There is a separation of all of the groups now[
Normally\ LDA models must be veri_ed "e[g[ by There seem to be two di}erent phenomena] the _rst is
leave!one!out cross!correlation#[ In this case this was expressed by DF0 and increases linearly with perfume
not done because the computed model "{matrix close to concentration[ The second is expressed by DF1 and
singular|# did not allow reliable predictions[ Another obtains saturation at a perfume concentration of about
reason was that the data base was not su.ciently large 059 ppm[
enough[ However\ the obtained LDA maps show Presently it is not known how well the long term

Figure 7. Linear discriminant score plot for (perfumed) diesel data as obtained with the Smart Nose instrument (masses
73, 80, 87, 88, 93, 94, 101, 108, 109, 122, 136 used). F02 pure, P F02 with perfume

Copyright Þ 1999 John Wiley + Sons\ Ltd[ Flavour Fra`r[ J[ 1999^ 04] 104Ð111
111 R[ FELDHOFF\ C[!A[ SABY AND P[ BERNADET

Table 4. Comparison of the parameters used and the performance of the two systems with respect to the
identification of perfumed diesel samples
System parameter or property FOX 3999 Smart Nose
Detection principle Chemical sensors Mass spectrometer
Injection technology Syringe injection Syringe injection
Sample volume 9[0 ml 19 drops\ ½9[1 ml
Incubation temperature 24>C 49>C
Incubation time 29 min 4 min
Headspace injection volume 099 ml 1499 ml
Time delay between two injections 0[4 h 09 min
Analytical information used 07 sensor signals Mass spectra "09Ð088 amu#
Measurement repeatability of individual samples − ¦
Detection limits found with F91 for perfume:additive Ł1999 ml:l: ½ 0179 ml:l ½219 ml:l: not tested
Detection limits found with G00 for perfume:additive 0179Ð0819 ml:l:½0179 ml:l not tested

drift of mass spectrometer based electronic noses can Acknowled`ements * We wish to thank Dr Francžois Hartmann\ Dr
Laurent Germanaud\ Dr Franck Eydoux "Centre de Recherches d|Elf
be mastered[ It is\ therefore\ not clear whether both of Solaize# and Prof Dr P[ McLeod "Ecole Pratiques des Hautes Etudes\
these parameters would be stable over a longer time Massy# for support and productive discussions[ We further thank Dr
period[ If one only uses DF0 "86[1)#\ the perfume Zesiger "LDZ Laboratory\ Switzerland# for his cooperation[ Thanks
for their kind help and professional work are equally due to the
detection limit may be ½219 ppm[ The sample with members of the odour panel of the Groupement de Recherches de
219 ppm can be distinguished from the pure diesel by Lacq[
Mahalanobis Distance Analysis[

References
Conclusion
0[ Feldho} R\ Saby C!A\ Bernadet P[ Analyst 0888^ 013] 0056Ð
The measurements carried out show that perfume detec! 0062[
tion in diesel fuels is di.cult with both types of elec! 1[ American Society for Testing + Materials\ Philadelphia[ 0882a
and 0882b "ASTM 4075!80#[
tronic noses] the most important system parameters 2[ American Society for Testing + Materials\ Philadelphia[ 0881b
used and the observations made are summarized in "ASTM D1314!77#[
Table 3[ The measurement repeatability of the mass 3[ American Society for Testing + Materials\ Philadelphia[ 0881a
"ASTM D0208!78#[
spectrometer is much better than that of the chemical 4[ Winquist F\ Hornsten EG\ Sundgren H\ Lundstrom I[ Meas[ Sci[
sensors[ It is also obvious that the mass spectral data Technol[ 0882^ 3] 0382Ð0499[
lead to a much lower "pure# perfume detection limit 5[ Bazzo S\ Loubet F\ Tan TT\ Hewett!Jones JD\ Engelen!Cornax
CEM\ Quadt JFA[ Semin[ Food Anal[ 0887^ 2"0#] 04Ð14[
"½299 ppm# than the chemical sensor data "Ł1999 6[ Simon JE\ Hetzroni A\ Bordelon B\ Miles GE\ Charles DJ[ J[
ppm#[ Food Science 0885^ 50"4#] 856Ð858[
With sensor selection and LDA it was possible to 7[ Pearce TC\ Gardner JW\ Friel S\ Bartlett PN\ Blair N[ Analyst
0882^ 007] 260Ð266[
make a dependence of the chemical sensor responses 8[ di Natale C\ Davide FAM\ Amico A\ Nelli P\ Gropelli S\
from the perfume concentration visible[ However\ the Sberveglieri G[ Sens[ Actuators B 0885^ 22"0Ð2#] 72[
measurement repeatability of the sensors was not 09[ Nanto H\ Kondo K\ Habara M\ Douguchii Y\ Waite RI\
Nakazumi H[ Sens[ Actuators B 0885^ 24Ð25] 072Ð075[
su.cient to build a stable and reliable LDA model[ 00[ Strassburger KJ[ Sem[ Food Anal[ 0887^ 2] 4Ð02[
The measurement repeatability of the mass spec! 01[ Gardner JW\ Bartlett PN[ Sens[ Actuators B 0883^ 07Ð08] 100Ð
trometer was much better[ This is the main reason why 119[
02[ Gardner JW[ An Introduction to Electronic Nose Technolo`y[
the perfume detection limit obtained with mass spec! Neotronics Scienti_c Ltd] UK\ 0885[
trometry was signi_cantly lower than that obtained with 03[ Moy L\ Tan T\ Gardner JW[ Perfumer + Flavorist 0883^ 08] 00Ð
chemical sensor measurements[ After selection of 00 05[
04[ Carrasco T\ Saby C\ Bernadet P[ Flavour Fra`r[ J[ 0887^ accepted[
masses\ a LDA model was computed[ Unfortunately\ 05[ Ho}heins BS\ Lauf RJ[ Analysis 0881^ 19] 190Ð196[
there were not enough data to do a _nal evaluation of 06[ Barshick SA\ Griest WH\ Vass AA[ SPIE 0886^ 1830] 52Ð63[
the stability of this LDA model[ 07[ Dillon WR\ Goldstein M[ Multivariate Analysis\ Methods and
Applications[ Wiley] New York\ 0873[
It is important to mention that this technique can only 08[ Coomans D\ Jonckheer M\ Massart DL\ Broeckaert I\ Blockx P[
be used to monitor di}erences between two samples "e[g[ Anal[ Chim[ Acta 0867^ 092] 398Ð304[
a pure and a perfumed one# but not to give analytical 19[ Massart DL\ Vandeginste BGM\ Buydens LM\ de Jong S\ Lewi
PJ\ Smeyersverbeke J[ Data handlin` in science and technolo`y
information about the perfume content of an unknown 19A\ Handbook of chemometrics and qualimetrics] Part B\
sample[ Elsevier\ 0886[

Copyright Þ 1999 John Wiley + Sons\ Ltd[ Flavour Fra`r[ J[ 1999^ 04] 104Ð111

You might also like