Analytica Chimica Acta 571 (2006) 167174

Combination of mid- and near-infraredetermination of the quality pro

b, M

nos Ab D Muno

April006

Abstract

The comb empof beers, suc mpletypes of bee icatewas used, w lusteand validation data sets. The calibration set was composed of 15 samples, thus leaving 28 for validation. A critical evaluation of the predictioncapability of multivariate methods established from the combination of NIR and MIR spectra was made. Partial least squares (PLS) and artificialneural networks (ANN) were evaluated for the treatment of data obtained in each individual technique and the combination of both. Differentparameters of each methodology were optimized. A slightly better predictive performance was obtained for NIRMIR combined spectra, and in allthe cases ANN performs better than PLS, which may be interpreted from the existence of some non-linearity in the data. The root-mean-sqare-errorof prediction0.076% w/w 2006 Else

Keywords: Dtechniques; N

1. Introdu

Beer isobtained bypresence ofcarbohydrabrewing in

Originain the beersrespectiveltent is a keythe beer clausually exp

CorresponE-mail ad

1 Present aQumica InorPabellon 2, C

0003-2670/$doi:10.1016/j(RMSEP) values obtained for the combined NIRMIR spectra for the determination of real extract, original extract and ethanol were, 0.14% w/w and 0.091% v/v.vier B.V. All rights reserved.

etermination; Original and real extract; Alcohol content; Beers; Artificial neuronal networks; Partial least squares; Combination of spectroscopicear-infrared; Middle-infrared; Attenuated total reflectance

ction

one of the two most common alcoholic beveragefermentation of cereals germinated in water in theyeast [1]. The main components of beers are water,

tes and ethanol, which are commonly used in thedustry for quality control of the final product.l and real extract correspond to the amount of sugarsbefore and after producing the fermentation process,

y and are normally expressed in % w/w. Ethanol con-economic and organoleptic parameter affecting both,ssification (in term of taxes) and its taste [1,2]. It isressed in % v/v.

ding author. Tel.: +34 96 354 4838; fax: +34 96 354 4838.dress: miguel.delaguardia@uv.es (M. de la Guardia).ddress: Laboratorio de Analisis de Trazas, Departamento deganica, Analtica y Qumica Fsica, Universidad de Buenos Aires,iudad Universitaria, (1428) Buenos Aires, Argentina.

The official methods of the Analytical division of EuropeanBrewery Convention for the determination of the aforemen-tioned parameters are based on the distillation of beer and themeasurement of the distillate and remaining solution density[3]. The density values of both solutions are introduced in semi-empirical tables for obtaining the percentage of extracts and theethanol content. These methods involve a high sample handlingand are time consuming and therefore inefficient in terns of timeand cost. So, there is a need for developing fast alternative meth-ods for routine quality control programs. Among those, the mostcommon strategy implemented in the industry is the use of auto-analyzers, which reduces the sample handling, but requires about30 min for obtaining a triplicate measurement of each parameter.

In two previous contributions we have reviewed the determi-nation of real extract, original extract and ethanol in beers bymeans of infrared (MIR) [4] and near-infrared (NIR) [5] spec-troscopy. In these works, we have also presented the methodswe have developed for the simultaneous determination of theseimportant parameters, as until that time most work conducted in

see front matter 2006 Elsevier B.V. All rights reserved..aca.2006.04.070Fernando A. Inon a,1, Salvador Garriguesa JENCK S.A., Av. Alvarez Thomas, 228 Bue

epartamento de Qumica Analtica. Universidad de Valencia, Edicio Jero`nimReceived 24 December 2005; received in revised form 21

Available online 5 May 2

ination of infrared (MIR) and near-infrared (NIR) spectroscopy has beenh as original and real extract and alcohol content. A population of 43 sar, was evaluated. For each technique, spectra were obtained in triplhereas attenuated total reflectance measurements were used in MIR. Cd spectroscopy for theperties of beersiguel de la Guardia b,

ires C1427CCPz C/Dr. Moliner, 50, 46100 Burjassot, Valencia, Spain2006; accepted 26 April 2006

loyed for the determination of important quality parameterss obtained from the Spanish market and including different. In the case of NIR a 1 mm pathlength quartz flow cellr hierarchical analysis was employed to select calibration

168 F.A. Inon et al. / Analytica Chimica Acta 571 (2006) 167174

this area only addressed the determination of only one parameterat time.

One queof the specttion capabicomparingcharacteriztwo-dimenNIR and Mtivariate cudependentMIR havepropertiesRaman andtion perforfrom theseto the usepurposes.

Regardiknowledgetaneous deconductedneural netwpredictionproduct or tmineral cosion spectrwhich the ntroscopy foproperties.of a methochemometrters.

For incmodel, a hselecting thtion proceselsewhere [

As theinclude theto evaluateficial neurapredictioncompared i(RMSEP).

2. Experim

For NIRBrucker Msoftware fr1.00 mm pa(Mullheim

For MIRSeries 750for WindowUSA) and e

Sampler horizontal ATR from Graseby Specac (Orpington, UK)with a 45 crystal ZnSe through top-plate.

m tter wtly dmpletherm

Minpum tub

ple

otalere

atchreme

rests ar

nd inples

the Aent

R an

ginaroo

g thell a

we

25 s1 (da329 cs 35bac

R pSAmpl

rvalseasu

pecth, saThed 60serv

ationroun

as a

spelling

an

pleswherlysistion that remains unanswered is if the combinationra acquired by NIR and MIR may improve the predic-lity of the multivariate models. There are many worksone technique versus other, some combining both foration or identification purposes, including those usingsional correlation analysis (i.e., [611]). CombinedIR spectral data has been used together with mul-rve resolution (MCR) for monitoring temperature-transitions of proteins [12,13]). Also, Raman andbeen used in combination for prediction of yarn

[14]. The authors concluded that concatenating theinfrared spectra does not enhance the PLS predic-

mance, not even after wavelength selection. Apartcontributions no works have been found devotedof multitechnique combination for quantitative

ng the chemometric treatment, to the best of ourno attempts to use neural networks for the simul-

termination of quality properties of beer have beenso far. Work published in the literature related toorks and beers [1519] address the optimization or

of industrial process rather quality control of the finalhe characterization of beer samples according to theirntent by inductively coupled plasma atomic emis-ometry (ICP-AES) [20]. No works has been found ineural network was coupled to any vibrational spec-r the determination of the aforementioned qualityTherefore, this study was devoted to the evaluation

dology based upon NIR and MIR measurements andic data treatment for the evaluation of beer parame-

reasing the application range of the multivariateeterogeneous sample population was chosen fore calibration and the validation datasets. The selec-s of each sample set was carried out as reported4,5].

combination of data from different techniques mayaddition of non-linearity to the system we decided

both, partial-least squares (PLS)-NIRMIR and arti-l networks (ANN)-NIRMIR methodologies, for theof ethanol, real and original extracts. Models weren terms of the root-mean-square-error of prediction

ental

measurement, it was used a Fourier transform NIRPA spectrometer controlled by Opus for Windowsom Brucker Optik (Bremen, Germany), and with athlength (50l volume) quartz flow cell from Hellma

, Germany) mounted on a home-made adaptor.measurement, it was employed a Nicolet Magnamodel FTIR spectrometer controlled by Omnics software from Nicolet Instrument (Madison, WI,quipped with Specaclamp IN-Compartment Contact

Roomome

nificanFor saand aGilsonstalticTeflon

3. Sam

A tcans, w

each bmeasu

of intesamplebe fou

Samfillinginstrum

4. MI

OritrolledforminATR cspectraaging4 cmof 0.6require

Thethe ATMA, Uthose eat inteafter mof IR s

Botmode.850 anwas obto variband aCO2, w

Theby refi

5. NIR

Samroom

the anaemperature was monitored using a mercury ther-ith a precision of 0.5 C and it did not vary sig-uring acquisition of the spectra of all the samples.preparation, a magnetic stirrer Select, (Singapur)ostatic bath Grant (Cambridge, UK) were used. Aipuls 2 from Gilson (Villiers-le-Bel, France) peri-p was used to fill the flow cell through 0.5 mm i.d.ing.

s

of 43 beer samples, contained in sealed aluminiumobtained from the Spanish market. Duplicate cans ofwere always collected, one was used for NIR and MIRnts and the other was used to measure the propertiesby reference procedures. The characteristics of thee indicated in Table 1, and a detailed description can

[5].were degassed by stirring and then filtered beforeTR cell or pumped through the flow cell of the NIR

.

alysis

l samples were placed in the same temperature con-m where the spectrometer was located before per-e analysis. Three sub samples were poured into thend the FTIR spectra were taken as follows. Samplere scanned between 4000 and 600 cm1, by aver-cans per spectrum with a nominal resolution ofta spacing of 1.93 cm1) and using a mirror velocitym s1. The acquisition of each averaged spectrums [4].kground and blank spectra were acquired by fillinglate cell with Millipore Q-purified water (Bedford,) and using the same instrumental conditions thanoyed for samples. Background spectra were scannedof seven samples and blank spectra were collectedrement of each sample, to ensure that memory effectra was negligible after cleaning the ATR crystal.mple and blank spectra were collected in absorbanceregions between 4000 and 3050 cm1 and between0 cm1 were eliminated prior the calculations as ited that variations in these regions cannot be ascribeds in sample composition. Furthermore, the absorptiond 2382 and 2314 cm1, which is due to atmosphericlso cut off from raw spectra.ctra of the three sub samples of each beer were takenthe ATR cell once again.

alysis

were placed in the same temperature controllede the spectrometer was located before to carry outs. During the acquisition of triplicate FT-NIR spectra

F.A. Inon et al. / Analytica Chimica Acta 571 (2006) 167174 169

Table 1Composition of samples employed in this study

Sample numb xtrac

12345a67a8a9

101112a131415161718a19a202122a232425a26a27282930a3132a333435a3637a3839a4041a4243

a Samples e

using the flothe cell usirefilled afte

Sampleaveraging 22 cm1. Th35 s. The mestimated t

The bacthe cell wiand using thfor sampleseven sampment of eacer Classification Original extract (% w/w) Real e100% malt 12.12 4.01Especial 12.95 4.57Normal 10.32 3.44Normal 10.6 3.54Normal 11.35 3.63100% malt 11.51 3.76Normal 10.7 3.63Especial 13 4.81Normal 10.9 3.7Normal 10.33 3.47With soda 8.72 2.9Normal 10.46 3.45Normal 10.32 3.45German type 12.47 4.15Normal 10.49 3.53

Normal 10.41 3.41Normal 10.78 3.58Alcohol free 4.25 2.17With soda 9.11 3.02Normal 10.44 3.43Normal 10.4 3.5Normal 10.6 3.57Normal 10.82 3.52Normal 10.7 3.4100% malt 12.07 4.45Normal 10.47 3.75Normal 10.84 3.53Normal 10.54 3.48Normal 10.52 3.64Normal 10.67 3.44Normal 10.77 3.62Normal 10.34 3.6Normal 10.7 3.51Normal 10.57 3.36Normal 10.69 3.42Normal 11.09 3.78Normal 10.51 3.47Normal 10.89 3.76Normal 10.73 3.72Normal 10.64 3.63Normal 10.81 3.72Normal 10.87 3.8Normal 11 3.96

mployed for calibration in all cases.

w cell, the sample was continuously pumped throughng a peristaltic pump (flow rate =1.5 mL min1) andr each measurement.spectra were scanned between 800 and 2857 nm, by5 scans per spectrum with a nominal resolution ofe acquisition of each averaged spectrum requiredaximum standard deviation of the response (sR) waso be 0.0005 a.u.kground and blank spectra were acquired by fillingth Millipore Q-purified water (Bedford, MA, USA)e same instrumental conditions than those employed

s. Background spectra were scanned at intervals ofles and blank spectra were acquired after measure-h sample, to avoid cross-contamination.

Both, samode. Theto 1850 nm1850 to 21of water in

6. Data an

Spectraexported in(South Natfunctions wate the simiassess the nt (% w/w) Alcohol (% v/v) Alcohol (% w/w)5.34 4.195.55 4.354.4 3.524.6 3.625.03 3.955.06 3.994.62 3.625.43 4.234.76 3.744.47 3.513.71 2.924.53 3.594.47 3.515.47 4.294.53 3.56

4.55 3.574.7 3.691.31 2.173.94 3.114.57 3.594.5 3.554.67 3.644.76 3.744.76 3.745.53 4.344.38 3.444.77 3.754.59 3.614.49 3.534.71 3.714.66 3.664.39 3.444.74 3.734.67 3.664.74 3.724.81 3.784.42 3.474.79 3.764.73 3.724.62 3.634.74 3.724.83 3.84.62 3.62

mple and blank, spectra were collected in absorbanceavailable spectral-analytical windows were from 800and from 2050 to 2400 nm. The spectral region from

00 nm was not available due to the high absorbancethis region.

alysis

collected from the Opus and Omnic software weretext format and analyzed using The Mathworks Inc.ick, MA, USA). Firstly, laboratory-written Matlabere used for hierarchical cluster analysis to evalu-

larity of samples in terms of their NIR spectra and toumber of characteristic subsets in which the available

170 F.A. Inon et al. / Analytica Chimica Acta 571 (2006) 167174

samples could be divided. Multivariate calibration calculationswere made, for PLS, with the MVC1 toolbox [21] and, for ANN,through a hmizing theof the Chem

The follpredictive pscope wasthe actual d

PRESS,square erroerror of crocoherencediction capmean-squamodel wasare availab

In orderthe referen(i) the absoand referenences (sxystandard erstated by Mcoefficientdata not othe data pogives somethe estimat

To buildnumber ofon the crite

In the cafor buildinor nor of areducing thof extractedber of inpuFor each anparametersbolic tangecases, the n

7. Results

7.1. Comb

Before cof correctiothe average(for ATRMeach indivirected specATRMIRthe wavenufurther corwavenumb

Comber; (bmbers

umbh samd ma

one

ompnd erestn 85afo

s: (iumbultivugm6 cohyd

m1re m

withl andwater, the beer spectra shown are due to the absorptione beer constituent plus the reduction of the amount ofIn the case of the formers, positive bands were observed,as in the later case, negative peaks were found. From there, the following general assignments can be made in theof interest: the bands located between 2200 and 2400 nmto the first set of CH combination bands, and for OH, the

nation bands occur between 2050 and 2200 nm, region ina broad band is observed [25,26].. 1 shows the combined spectra of three different typesrs. The 100% malt is the sample with the highest extractand it can be seen that the band in the sugar region ishest one. On the other hand, the sample without alcohollowest extract value, and accordingly, it has the lowest

tion in the aforementioned spectral region. In the selectedegion the main differences between the spectrum of theithout alcohol and that of a normal beer correspond tods located at 1100, 1050 and 875 cm1. These bands areome-made graphical user interface (GUI) for maxi-potential of the neural network calibration routinesoAC Toolbox [22].

owing figures for the models fit to the data and theower were used throughout the text. In all cases, the

to evaluate the average deviation of the model fromata.as the sum of squares prediction error, root-mean-r of calibration (RMSEC), and root-mean-squaress-validation (RMSECV) were used to evaluate theof the calibration models and to determine the pre-abilities, of each model it was employed, the root-re-error of prediction (RMSEP) established when theapplied to new data for which the reference values

le.to compare the spectroscopy methodology against

ce ones, different quality indicators were also given:lute mean difference (dxy) between predicted valuesce data, (ii) the standard deviation of mean differ-), the quality coefficient (QC) and (iii) the pooledror of prediction for validation samples (sreg). Asassart et al., the QC is to be preferred over correlationof the regression between predicted and referencenly because it gives a better idea of the spread ofints around the fitted straight line but also because itindication on the percentage error to be expected fored concentration [23].and select PLS models, the selection of the optimumfactors, which minimizes the RMSECV, was basedrion of Haaland and Thomas [24].se of ANN, there are many factors to be considered

g calibration models. These are (i) the requirementprevious principal component analysis (PCA) for

e number of variables, and in this case, the numbercomponent; (ii) the topology of the network (num-

t and output nodes), (iii) type of transfer function.alyte, an extensive evaluation of the aforementionedwas made. Regarding transfer function, only hyper-nt and linear function were considered. In all theumber of hidden layers was fixed to one.

and discussion

ined NIRMIR spectra of beer

ombining spectra acquired by NIR and MIR, a seriesns were carried out. For correcting additive artefacts,of absorbance values between 2007 and 2056 cm1IR) and between 2220 and 2222 nm (for NIR) of

dual spectrum was respectively subtracted. These cor-tra were utilized for further calculations. In case of, as the penetration depth is inversely proportional tomber, the absorbance at a given wavenumber was

rected by multiplying the absorbance value by theer value and rationing the result by the maximum

Fig. 1.malt bewavenu

waven

of eacmenteof NIR

A cbe fouof intebetwee

Thereason

waven

best mThus aand 31

The1052 cbeers aregionethanousingof somwater.whereliteraturegionare duecombiwhich

Figof beevalue,the highas theabsorpMIR rbeer wthe banined NIRMIR spectra of main types of beer samples. (a) 100%) normal beer and (c) beer without alcohol. Wavelengths andwere plotted in the same axis for clarity purposes.

er in the spectrum. After that, the mean spectrumple was calculated from acquired replicates. Aug-

trix were built allocating the MIR spectra at the rights.

lete description of main spectral features in the IR canlsewhere [4,5]. In the present study, the main bandsare located in the range 22202354 nm for NIR and0 and 1201 cm1 for MIR.

rementioned regions were selected due to two main) to keep low the total number of wavelengths anders and (ii) because in our previous contributions theariate models were built from these spectral regions.ented matrix was composed by 43 rows (samples)lumns (182 for NIR and 134 for MIR).roxyl groups (OH) of sugars absorb in MIR aroundcorresponding to carbohydrates. NIR spectra of

ainly related to the absorption of water in this spectralsome features of the other two main constituents:carbohydrates. As background signal was acquired

F.A. Inon et al. / Analytica Chimica Acta 571 (2006) 167174 171

ascribed to the presence of ethanol. Similar differences can beascribed in the NIR spectra.

7.2. Select

Calibratobtained frtering methorder to evand to seleing spectrasignificanta principalthe regionthe linkagedatasets waleast one sis compriseselected fortotal numbsamples weber of samthan the nuwithin a gi

Followition sets co

It musta previoustheir spectrples of theconstructio

7.3. Deterethanol usi

A calibrMIR data,lengths depthe wavelenmized for e

Table 2for the threthe real exbased on f0.095% w/extract, theand QC vafor ethanolof 0.103%

The aforvalues obtaRMSEP anand (iii) ethand 1.9%,of NIR, thew/w and 1.be seen, res

Table 2Prediction capabilities of PLS-NIRMIR for real and original extract and ethanoldetermination

Real extract Original extract Ethanol

4 3 4Va 0.208 0.211 0.103a 0.095 0.192 0.121P 2.6% 1.8% 2.6%

0.029 0.026 0.0390.092 0.194 0.1172.60% 1.81% 2.69%0.05 0.05 0.050.0005 0.0005 0.00050.04 0.03 0.03

ectral range used was that indicated in Fig. 1. sreg is the standard errorction.es given in % w/w for real and original extract and % v/v for ethanol.

itional details see the text.

lightly better (or similar) than those obtained on using atechnique.. 2 shd forcan

tors;IRdelNIRan

ed.le 2ned

es thrd deon ov/v,

Real and original extract and ethanol for net analyte sensitivity (sk)

or combined NIRMIR PLS model. Wavelengths and wavenumberstted in the same axis for clarity purposes.ion of the calibration set

ion and validation sets were based on the resultsom hierarchical cluster analysis [5]. In brief, a clus-od was used before chemometric data treatment inaluate possible classes among samples considered

ct properly a representative calibration. For calculat-l distance, Euclidian distance of the scores of the mostprincipal component (PC) was used (PC selected aftercomponent analysis (PCA) applied to NIR data infrom 2220 to 2370 nm). Ward method was used atstage. The selection of the calibration and validations made using the obtained dendrogram, selecting atample of each cluster for calibration. If the clusterd of more than one sample, the number of samplescalibration was approximately the root square of the

er of samples included in the cluster; the remainingre integrated in the validation data set. So, the num-ples assigned to the validation was equal or highermber of those employed for calibration. The samplesven cluster were selected randomly.ng the aforementioned rules, calibration and valida-mprised of 15 and 28 samples, respectively.be noticed that a calibration set selection based onidentification of the different class of samples, froma, offers guaranties about the consideration of sam-some type than those to be analysed in the model

n.

mination of real extract, original extract andng PLS and combined spectra

ation model was built in terms of combined NIR andconsidering the range of wavenumbers and wave-icted in Fig. 1. No efforts were conducted to reducegth/wavenumbers interval as their were already opti-ach individual spectroscopy.shows the calibration and validation results obtainede properties analysed using the combined spectra. Fortract determination the optimum PLS method wasour extracted factors, obtaining a RMSEP value ofw and a QC value of 2.6%. In the case of originaloptimum number of factors was 3, yielding a RMSEPlues of 0.192% w/w and 1.8%, respectively. Finally,4 factors were extracted, obtaining a RMSEP valuev/v and a QC value of 2.7%.ementioned reported values can be compared to thoseined using only NIR or MIR data [4,5]. For MIR,d QC values for (i) real extract, (ii) original extractanol was: (i) 0.103% w/w and 2.9%, (ii) 0.2% w/w(iii) 0.12% v/v and 2.6%, respectively. In the casese values were: (i) 0.12% w/w and 3.0%, (ii) 0.18%5%, (iii) 0.10% v/v and 1.9%, respectively. As it canults obtained when using the combined spectra tend

FactorsRMSECRMSEPRRMSEd(xy)as(xy)aQCsc

a

sRsreg

a

Note: Spof predi

a ValuFor add

to be ssingle

Figmethotion. Itall vecNIRMthe moon theof MIRbalanc

Tabmentioincludstandadeviati0.03%tively.

Fig. 2.vector fwere ploows the net sensitivity vector associated to the PLSreal extract, original extract and ethanol determina-be appreciated that there is no direct match betweenall parameters are not directly associated to the samespectral feature. Moreover, in the case of ethanol,tends to be more focused in the MIR spectra that, whereas for the other properties the contribution

d NIR spectra to the net analyte signal seems to be

includes the figures of merit obtained under all afore-conditions. The standard error of prediction (that

e uncertainty from the model [27,28], considering theviation of reference concentration (sc) and standardf the response (sR) were 0.04% w/w 0.03% w/w, andfor original extract, real extract and ethanol respec-

172 F.A. Inon et al. / Analytica Chimica Acta 571 (2006) 167174

7.4. Determination of real extract, original extract andethanol using neural networks

The neufeed-forwaAn excellefound elsewthe bias andof each neuing dataset

The outof a ANNnumber ofimpossibleinal variabneurons frrange (850computer.pre-treatmethe number

As PCAably emploa surface ring an AN(PC). Optiperformingwithout autive resultsto remembables beforANN traintion, and tvariables toscale eachrange of thples were rscaling [22samples.

The resprincipal cshown thatthe highestvariance, ththe next onof the rema

The opt(ii) originaand 4 and ((0.090% v/2 neurons.de dimensi

In eachcheeked wilinear onescan be catcout plottingdata at the

of the data may indicate that the data set is mainly linear whendata are activated in their linear portion or indicates a non-linear

our when data are activated in their strongly non-linear. Thus, it is possible to obtain information on the degree

-linearity of a given data set.. 3 shows for real extract and ethanol the activation ofdes by validation samples. It can be seen that in threer) of the activation nodes the activation occurs in the

portion, whereas in the case of ethanol it occurs in theear portion. Therefore, for each analyte the number ofand non-linear neurons has been evaluated in terms ofSEP value. Table 3 summarizes different characteristicsoptimum ANN model built for all the properties. Forination of real extract the optimum ANN method was

on two extracted factors and two neurons (one linear andn-linear). For original extract, nine factors were kept in

timized model and the topology of the network consistede linear and one non-linear neuron in the hidden layer.hanoon twed fale 3aforode

le 2.or oaximons (tivel

quon om oof

ing tuirene thl conedned st pro

on cadeterm

in hia

a

P

actorsean les giv

itionaral network chosen for this work is a double-layerrd type with the error back-propagation learning rule.nt review addressing all details of this ANN can behere [22]. The basis of ANN optimization is to selectweight values acting in the selected transfer-functionron for minimising the RMSEP values of a monitor-

.put nodes were fixed to one, and the optimizationfor each property has been made. Due to the highsensors (wavelengths and wavenumbers) it is almostto construct and optimize the ANN from the orig-les. As a matter of fact, evaluate the number ofom 1 to 10 in the hidden-layer for selected MIR1200 cm1) took 27 h in a Pentium IV 3.0 Gz HTTherefore a compression technique, in particular ant based on PCA, has been adopted in order to reduceof variables entering into the ANN.is a linear compression method and ANN is prefer-

yed for non-linear systems, it is important to constructesponse curve including also high PCs when build-N from scores of extracted principal componentsonally, the original data can be autoscaled before

PCA. We have tested the use of PCA with andtoscaled spectra, and in all the cases better predic-have been achieved using raw spectra. It is importanter that it is not necessary to mean-center input vari-e training since the biases act as offsets in the model.ing is not based on variancecovariance maximiza-herefore it is not necessary to scale the different

unit variance. The only constraint for ANNs is toinput variable so that training starts within the activee non-linear transfer functions. In this work, sam-ange-scaled with a linear mapping called minmax], determined for the training set and applied to all

ponse surface was explored up to 12 first extractedomponents (PCs) from calibration dataset. Resultsjust nine were needed for the model which requiresnumber of PCs (see below). Regarding the explainede fist component explain the 99.86% of the variance,

e explain an additional 0.11%, while the contributionining ones is bellow the 0.01%.

imum number of PCs and neurons for (i) real extract,l extract and (iii) ethanol were: (i) 2 and 2, (ii) 9iii) 5 and 7. In the latter case, similar RMSEP valuesv against 0.089% v/v) has been found with 3 PCs andSo, this bounder condition were preferred to reduceon of the model.case the behaviour of non-lineal neurons has beenth the idea of replace as much non-linear neurons by. Doing so, linear and non-linear behaviours of datahed by specific neurons. This evaluation is carriedhidden nodes activation of calibration or validation

end of training [22]. The activation of hidden nodes

behaviportionof non

Figthe no(of foulinearnon-linlinearthe RMof thedetermbasedone no

the opin threFor etbasedextract

Taball theANN min Tab1.9% fThe mminatirespecment.

Oneprecisispectruspectraobtainto acqcombithe fulcombicombithis las

Table 3Predictiethanol

Dataset

FactorsNeuronsRMSECRMSEPRRMSEd(xy)as(xy)aQCNote: Fand H m

a ValuFor addl determination, on the other hand, the model waso non-linear neurons in the hidden layer and threectors.also includes the figures of merit obtained under

ementioned conditions. Prediction capabilities usingl for all the properties are better than those presentedFor example, RMSEP values are 2.1%, 1.3% and

riginal extract, real extract, and ethanol respectively.um percentage error to be expected for new deter-QC) of these properties were 2.1%, 1.4% and 2.0%,y, which are better than those found for the PLS treat-

estion which remains open is how to estimate thef the method from available data. The combinedf each sample was obtained by averaging triplicateeach technique. Therefore, the usual procedure forriplicate combined spectra should be, in this case,three times three spectra in each technique and thenem. An alternative approach could be to perform

mbination individual spectrum, obtaining thus ninespectra, and to predict the concentration from eachpectra using the ANN method developed. Followingcedure, the repeatability of the determination, estab-

pabilities of ANN-NIRMIR for real and original extract andination

Real extract Original extract Ethanol

2 9 3dden layer 2 (1L, 1H) 4 (1H, 3L) 2 (2H)

0.141 0.059 0.0920.076 0.145 0.0912.1% 1.3% 1.9%0.007 0.014 0.0040.077 0.147 0.0922.12% 1.39% 2.01%

are the extracted principal components after performing PCA. Linear and hyperbolic tangent transfer function.en in % w/w for real and original extract and % v/v for ethanol.

l details see the text.

F.A. Inon et al. / Analytica Chimica Acta 571 (2006) 167174 173

lished fromconcentratifor originalbe appreciareference vtions are alare average

8. Conclu

The detein beers canof MIR anusing PLSobtained wFig. 3. Activation nodes for (A) real extract and (B) ethanol when considerin

the pooled standard deviation of the nine predictedons, were, 0.01% w/w, 0.02% w/w, and 0.01% w/w,extract, real extract and ethanol respectively. As canted, the precision is of the order of the uncertainty ofalues. The values calculated for QC in these condi-most equal to those obtained when individual spectrad before being combined.

sions

rmination of real extract, original extract and ethanolbe successfully carried out through the combination

d NIR techniques. Calibration model can be builtor ANN, although better predictive capabilities areith the latter method. Regarding the advantages of

combiningpredictionimprovemewith de Grbe justifiedportions.

Acknowled

The finai InvestigacGrupos 03de Vale`nciaAires (X01edge.g only hyperbolic tangent transfer functions.

or not the techniques, it can be argued that althougherrors are slightly lower in the combined method, thent is not significant. In this aspect we are in agreementoot et al. [14]. The use of ANN instead of PLS canbecause the activation of the nodes lies in non-linear

gments

ncial support of the Direccio General dUniversitatsio de la Generalitat Valenciana (Project GV04B/247,-118 and invited professor FAI grant), Universitat

(Project UV-AE-20050203), University of Buenos3) and ANPCyT (PICT-2003 17932)) is acknowl-

174 F.A. Inon et al. / Analytica Chimica Acta 571 (2006) 167174

References

[1] A.H. Varnam, J.P. Sutherland, Bebidas: Tecnologa, Qumica y Micro-biologa, Editorial Acribia, S.A., Zaragoza, 1996.

[2] Council Directive 92/84/EEC: OJ L 316, 1992. 31.10.1992 Bull. 10-1992.

[3] European Brewery Convention. Analytica EBC, 4th ed., Brauerei undGertranke Rundschau, Zurich, 1987.

[4] F.A. Inon, R. Llario, S. Garrigues, M. de la Guardia, Anal. Bioanal.Chem. 382 (2005) 1549.

[5] R. Llario, F.A. Inon, S. Garrigues, M. de la Guardia, Talanta 69 (2006)469.

[6] N. Afonina, L. Bhattacharyya, J.P. Hennessey, Analyst 129 (2004) 1091.[7] M.S. Lin, M. Al-Holy, H. Al-Qadiri, D.H. Kang, A.G. Cavinato, Y.Q.

Huang, B.A. Rasco, J. Agric. Food Chem. 52 (2004) 5769.[8] L.E. Rodriguez-Saona, F.M. Khambaty, F.S. Fry, E.M. Calvey, J. Agric.

Food Chem. 49 (2001) 574.[9] Y.L. Liu, R.K. Cho, K. Sakurai, T. Miura, Y. Ozaki, Appl. Spectrosc.

48 (1994) 1249.[10] F.E. Barton, D.S. Himmelsbach, Appl. Spectrosc. 47 (1993) 1920.[11] F.E. Barton, D.S. Himmelsbach, J.H. Duckworth, M.J. Smith, Appl.

Spectrosc. 46 (1992) 420.[12] S. Navea, A. de Juan, R. Tauler, Anal. Chem. 75 (2003) 5592.[13] L. Pillonel, W. Luginbuhl, D. Picque, E. Schaller, R. Tabacchi, J.O.

Bosset, Eur. Food Res. Technol. 216 (2003) 174.

[14] P.J. de Groot, H. Swierenga, G.J. Postma, W.J. Melssen, L.M.C. Buy-dens, Appl. Spectrosc. 57 (2003) 642.

[15] T. Becker, T. Enders, A. Delgado, Bioprocess. Biosyst. Eng. 24 (2002)347.

[16] T. Kurz, M. Fellner, T. Becker, A. Delgado, J. Inst. Brew. 107 (2001)241.

[17] I.C. Trelea, M. Titica, S. Landaud, E. Latrille, G. Corrieu, A. Cheruy,Math. Comput. Simul. 56 (2001) 405.

[18] L.A. Garcia, F. Argueso, A. Garcia, M. Diaz, J. Indust. Microbiol. 15(1995) 401.

[19] R. Simutis, I. Havlik, A. Lubbert, J. Biotechnol. 27 (1993) 203.[20] A. Alcazar, F. Pablos, M.J. Martin, A.G. Gonzalez, Talanta 57 (2002)

45.[21] A.C. Olivieri, H.C. Goicoechea, F.A. Inon, Chemom. Intell. Lab. Syst.

73 (2004) 189.[22] F. Despagne, D.L. Massart, Analyst 123 (1998) 157R.[23] D.L. Massart, B.G.M. Vandeginste, L.M.C. Buydens, S. de Jong, P.J.

Lewi, J. Verbeke, Handbook of Chemometrics and Qualimetrics (Part Aand B), Elsevier Science B.V, Amsterdam, 1997.

[24] D.M. Haaland, E.V. Thomas, Anal. Chem. 60 (1988) 1193.[25] L.G. Weyer, Appl. Spectrosc. Rev. 21 (1985) 1.[26] J.J. Workman, Appl. Spectrosc. Rev. 31 (1996) 251.[27] K. Faber, B.R. Kowalski, J. Chemom. 11 (1997) 181.[28] A.C. Olivieri, J. Chemom. 16 (2002) 207.

Combination of mid- and near-infrared spectroscopy for the determination of the quality properties of beersIntroductionExperimentalSamplesMIR analysisNIR analysisData analysisResults and discussionCombined NIR-MIR spectra of beerSelection of the calibration setDetermination of real extract, original extract and ethanol using PLS and combined spectraDetermination of real extract, original extract and ethanol using neural networks

ConclusionsAcknowledgmentsReferences