Automatic Sensor System for the Continuous Analysis of the Evolution of Wine

Transcription

1 Automatic Sensor System for the Continuous Analysis of the Evolution of Wine Jesús Lozano, 1,2 * José Pedro Santos, 2 José Ignacio Suárez, 1 Mariano Cabellos, 3 Teresa Arroyo, 3 and Carmen Horrillo 2 Abstract: An in situ and on-line electronic nose (e-nose) was developed and installed in a winery for the continuous measurement of wine evolution. The system has a novel sampling method that uses a carrier gas to extract aroma compounds directly from the headspace of the wine storage tank, and the volatile compounds are then automatically carried to the sensor cell. The system uses a tin oxide multisensor prepared by radio frequency sputtering onto an alumina substrate and treated with with chromium and indium. The whole system is fully automated and controlled by a computer and can be supervised remotely via the internet. Linear techniques such as principal component analysis and nonlinear ones such as artificial neural networks were used for pattern recognition, and partial least squares analysis was used for predicting GC-MS analysis. Results of two different wines show that the e-nose system can detect the evolution of aroma compounds for nine months. Correlation coefficients near to 1 were obtained in the prediction of the volatile organic compounds ethyl butyrate, isobutyric acid, isobutyl acetate, hexyl acetate, and ethyl octanoate. This system can be trained for monitoring wine preservation and evolution, detecting off-odors, and warning the wine expert to correct problems as soon as possible in order to prevent spoilage and improve wine quality. Key words: electronic nose, gas sensors, pattern recognition, wine evolution, aroma, GC-MS 1 Dept. Ingeniería Eléctrica, Electrónica y Automática, Universidad de Extremadura, Av. Elvas s/n, Badajoz, Spain; 2 Grupo de I+D en Sensores de Gases, Gridsen, Consejo Superior de Investigaciones Científicas, ITEFI, (CSIC), C/Serrano, 144, Madrid, Spain; and 3 Dept. Agroalimentación, Instituto Madrileño de Investigación y Desarrollo Rural, Agrario y Alimentario (IMIDRA), Km 38.2 N-II, Alcalá de Henares, Spain. *Corresponding author (jesuslozano@unex.es; tel: ; fax: ) Acknowledgments: The authors thank the Spanish Science and Technology Ministry for supporting the projects TIC C02-00 and RTA C02-00 and the Regional Government of Extremadura, through the European Regional Development Funds (GR10097). Manuscript submitted Aug 2014, revised Oct 2014, accepted Oct 2014 Publication costs of this article defrayed in part by page fees. Copyright 2015 by the American Society for Enology and Viticulture. All rights reserved. doi: /ajev The monitoring of quality and safety in the food industry is important because of issues of quality of life and health, as well as possible applications to wine flavor management. Organoleptic analysis, based on both analytical methods and trained inspectors who use odor evaluation, is the most common method to define quality and safety in foods. In many cases, monitoring and determining the constituents of a gas sample typically involves collecting and analyzing samples using traditional analytical instruments such as gas and liquid chromatography, mass spectrometry, nuclear magnetic resonance, and spectrophotometry. Though such methods are highly reliable and suitable for this purpose, they have certain drawbacks such as high cost and size, long processability, and low in situ and on-line usefulness. In addition, in some cases the sample preparation is time consuming and thus online, real-time analysis cannot be easily performed. Moreover, many applications, such as the detection of volatile organic compounds (VOC) or smells generated from food in the agrofood industry, need smaller, more portable, cheaper, and even disposable sensor-based systems designed to analyze such complex mixtures. Potentially, the most significant advance in objective sensory analysis may come from development of instruments such as electronic noses (e-noses). Among analytical methods, e-noses have the advantage of high portability for in situ and on-line measurements with lower costs and good reliability (Chatonnet and Dubourdieu 1999, Devarajan et al. 2011). In recent years, there have been a number of efforts to develop arrays of non-specific sensors coupled with pattern recognition methods. Applications of these methods include aroma identification and discrimination (Lozano et al. 2006, Gardner 2011); recognition of adulteration (Penza et al. 2004), varietal origin (Zoecklein et al. 2011, Devarajan 2011, Di Natale et al. 1996), oak toasting levels (Chatonnet 1999, Chatonnet and Dubordieu 1999), regional differences (Cynkar et al. 2010), and aging (Lozano et al. 2008); off-odors detection (Ragazzo-Sanchez et al. 2005); sensory properties prediction (Lozano et al. 2007a); and an objective method to establish wine quality (Adami et al. 2006). More recently, a combination of e-nose, e-tongue, and e-eye has been proposed to monitor wine aging (Apetrei et al. 2012). The authors have already demonstrated that it is possible to correlate the e- nose results with classic GC-MS analysis of wine (Santos et al. 2004, Lozano et al. 2007a). In all the above applications, measurements were performed off-line. The main goal of this work was to design an artificial olfactory system capable of analyzing wine in situ without taking tank samples. During design, comparative samples were taken and analyzed with traditional techniques. The evolution of two different wines was studied for nine months with new as well as traditional techniques. The main advantages of the proposed system, compared with traditional techniques of 148

2 Electronic Nose for Wine Cellars 149 wine analysis, are the low cost, ease of use, and noninvasive analysis. The proposed system is protected by patent no. WO A1. Materials and Methods The e-nose consists of four elements: an aroma extraction or air flow system which carries the volatile compounds from the samples to the next step, an array of chemical sensors which transform the aroma into electrical signals, an instrumentation system to measure the signal of the different sensors, and the control and automation system. The control and automation system includes pattern recognition designed to identify and classify the aroma of the measured samples into several previously learned classes. The e-nose was designed for detecting wine evolution in tanks and was installed in an experimental wine cellar as shown in Figure 1. Sampling method. When studying the volatile compounds responsible for wine aroma, choosing a suitable extraction procedure to qualitatively and quantitatively capture the original wine aroma can be problematic. Though several sampling methods can be used for wine aroma extraction for e-nose measurements (Lozano et al. 2007b), it can be difficult to obtain a representative extract of the wine which has not been degraded between sampling and analysis. In this work, a novel sampling method was developed to automatically extract and carry the aroma directly from the static headspace of wine tanks to the sensor cell. The main advantage of this system is that it is not necessary to subject the wine to storage prior to analysis. A general schematic of the sampling system is shown in Figure 2. Using the control program and several electric valves, the gas sample to be analyzed can be selected from several inputs: tank 1 or 2, where wine is stored; the calibration sample; or blank gas. The carrier gas used was % pure nitrogen. All measurements were carried out at a constant gas flow of 200 ml/min. Gas line tubes were stainless steel lined with fused silica in order to minimize gas adsorption in the line. In order to get aroma samples of wine directly from the tanks, headspace with effluent transfer was adapted for the use in wine tanks (Lozano et al. 2007b). The tanks were modified to allow the carrier gas to extract the wine aroma from the headspace, thereby not modifying the wine. The sampling method scheme is shown in Figure 3. Input and output gas lines were installed above the level of the wine in the tanks. When the carrier gas flows through the tanks, it carries the wine aroma to the sensors cells. At all times both tubes and tanks are filled with nitrogen to avoid wine spoilage or oxidization. Sensors. The multisensor included 16 sensor elements distributed in a circle onto an alumina substrate. Sensors were treated with different amounts of chromium and indium by changing the deposition time during the sputtering process. In this treatment, a trace impurity is added to the semiconductor to alter the electrical properties and to change the response to certain gases. Deposition conditions and composition of the multisensor are detailed elsewhere (Lozano et al. 2008). The multisensor was thermally treated in air at 520ºC for 4 hr to control the material morphology (stoichiometry and grain size of the tin oxide and treatment distribution) and to stabilize the semiconductor electrical Figure 2 Sampling system based on headspace with effluent transfer: (1) nitrogen generator, (2) mass flow controller, (3) calibration flask, (4) wine tanks, and (5) sensor chamber. Figure 1 View of the system installed in the wine cellar. The tanks contain wines made from either or grapes. Figure 3 Wine storage tanks modified for aroma sampling.

3 150 Lozano et al. resistance. Annealing was fundamental in order to obtain a good detection (Rickerby et al. 1997). The prepared array of sensors is sensitive to most of the compounds present in wine, including VOCs. The array responds to the range of wine aroma compounds. The sensors present cross selectivity with a broad spectrum in responses due to their composition. An image of the multisensor used for the measurements is shown in Figure 4. The array was placed in a 24 cm 3 stainless steel cell with a heater and a thermocouple. The operating temperature of sensors was controlled to 250ºC with a PID temperature controller. The same sensors were used in all experiments and no replacement of sensors was done. The lifetime of each sensor Figure 4 Sensor array. Figure 5 Instrumentation and control system scheme. is over one year. The sensors were calibrated once a week with a blank solution (12% (v/v) ethanol in deionized water) in order to reduce the effects of sensor drift (Gutierrez-Osuna 2002, Lozano et al. 2005). Instrumentation and control system. The instrumentation and control system was designed specifically for this application, based on the design of both home-developed electronic circuits and commercial instrumentation. The resistance of the sensors was measured with a digital multimeter coupled to a 40-channel multiplexor, which was connected to the personal computer through a general purpose interface bus. The control of carrier gas (using mass flow controllers and electrovalves) was performed using a multifunction peripheral component interconnect card with a homemade amplification board. Outputs were generated with the data acquisition board and powered by power operational amplifiers in case of analog signals and by relays and power transistors for digital outputs. The temperature of the sensors was controlled with a power supply and a commercial PID controller. Pressure of the pneumatic tips, ambient temperature, and the proper functioning of the equipment was monitored at all times. The whole system was fully automatic and could be controlled and supervised remotely via the internet with a program developed in Testpoint (Capital Equipment Corporation, Billerica, MA). A general scheme of the system is shown in Figure 5. Data processing. Sensor signals were acquired and stored in the computer. As illustrated in Figure 6, the process of data analysis can be split into four sequential stages: feature extraction, dimensionality reduction, classification and prediction, and decision making. The first computational stage extracts descriptive parameters from the sensor array response, which includes compensating for sensor drift and preparing the feature vector for further processing. The dimensionality reduction stage projects this initial feature vector onto a lower dimensional space in order to avoid problems associated with high-dimensional and sparse datasets. The resulting low-dimensional feature vector is then used to solve a given prediction problem, typically classification, regression, or clustering. Classification tasks address the problem of identifying an unknown sample as one from a set of previously learned aromas. In regression tasks, the goal is to predict a set of properties (e.g., concentration of an analyte, quality impact, organoleptic meaning) from the responses of another technique (in this case, gas sensors). In the final step, Figure 6 Building blocks of the pattern analysis system for an electronic nose.

4 Electronic Nose for Wine Cellars 151 the recognized class is selected from the previously learned classes or the value of prediction is performed. The selection of models and parameter settings, and the estimation of the true error rates for a trained model, are carried out by means of validation techniques (Gardner et al. 2002). By teaching a computer (or hardware) to recognize those patterns, the e-nose consequentially was trained to classify the wine aroma belonging to the different classes of learned aromas or patterns. An important characteristic of the e-nose is that it efficiently recognizes patterns. Several methods (including statistical methods) are used to determine the clusters of data representing different classes of odors. Some of these methods are based on different forms of artificial neural networks (ANN) in order to classify and quantify aromatic compounds and gas mixtures. Developing efficient pattern recognition algorithms is therefore one of the most important issues in the field of e-noses. One common method for pattern recognition is principal component analysis (PCA) (Gutierrez- Osuna 2002). PCA is a powerful, linear, unsupervised, and nonparametric pattern recognition technique that has been used to reduce the dimensionality of the pattern space in order to lead to better visualization of data clustering. For instance, if 16 sensors were used for our measurements, one measurement can be represented as a point in a 16-dimensional space, and therefore some of the sensors likely will respond in a similar (but not identical) manner. This means that the number of dimensions in the data set can be reduced without any loss of information. This method consists of expressing the response vectors in terms of linear combinations of orthogonal vectors along a new set of coordinate axes, and is sometimes referred to as vector decomposition, and thus helps to display multivariate data in two or three dimensions. A loading plot of a PCA shows to what degree the different sensors contribute to the principal components. In this plot, sensors with similar contributions (i.e., that contain similar information) will be close together. Sensors that are close to the origin have comparably small variance, and therefore probably contain little information. A popular supervised method to handle e-nose data is the ANN, which bears a similarity to the function of the human brain. In principle, ANN is constituted of many (~50 to 100) artificial neurons. The artificial neurons are organized in different layers, normally three, which together form a network. An artificial neuron is a simple processing element, which, similar to biological neurons, uses signals from several inputs to produce one output. A linear combination is taken of all the inputs, giving a single value. This value is then used in a transfer function, which could have arbitrary shape. Two different topologies of ANN were used in this work for pattern recognition: the classical feed-forward neural networks with backpropagation learning algorithm (BP) and, alternative to BP, probabilistic neural networks (PNN) combined with radial basis functions. PNN may require more neurons than standard feed-forward backpropagation networks, but they often are designed faster than the time required to train standard feed-forward networks and usually perform better when many training vectors are available. The learning of an ANN is performed by changing the parameters in the linear combination. By feeding data from known odors into the network, the parameters can be adapted to recognize the sensor signals from these odors (Lozano 2006). In order to adapt the parameters, the training data must be used many times. This is very similar to training of odor recognition for humans. It is unlikely for humans to remember an odor after a one-time exposure, whereas odors experienced often in youth are frequently recognized long thereafter. Of important note is that an ANN, similar to the human nose, cannot identify odors that it has not yet experienced. When confronted with the sensor signals from a new odor, the ANN can only indicate to which of the known odors the signals are most similar, or that it does not recognize the odor. A human can easily indicate if an unknown odor is pleasant or unpleasant, while an e-nose is unable to perform such subjective judgment. A further challenge for e-nose instruments is regression (Duda et al. 2001, Geladi and Kowalski 1986). The goal of regression is to establish a predictive model from a set of independent variables (e.g., gas sensor responses) to another set of continuous dependent variables. Partial least squares regression (PLS) combines the properties of multiple linear regression and PCA to produce a technique that can accept collinear data while also separating the statistical variability within the same class in order to make linear combinations in the dependent concentration matrix. PLS is used in chemometrics due to its ability to handle collinear data and reduce the number of required calibration samples (Geladi and Kowalski 1986). A stopping point for the sequential expansion is determined through cross-validation. Published works on multicomponent analysis using gas sensor arrays and PLS may be found in Hierlemann 1995, Sundgren et al. 1991, Wang et al. 1995, and Carey and Yee A regression model aimed at estimating sensory panel indicators by e-nose was built by PLS. Models were cross-validated by the leave-one-out method. Preprocessing, modeling, and validation were performed with Unscrambler 9.2 (Camo, Asa, Norway; unscrambler.html). Parameters used to evaluate model prediction ability were: root mean square prediction error and correlation coefficient between real and predicted Y variables. All variables were normalized prior to the analysis. Wine samples. Samples of two varietal wines, the white variety (20 Brix) and the red variety (25 Brix), manufactured with grapes of the majority varieties in the origin denomination (OD), Vinos de Madrid, were used for testing the discrimination capability of the system presented in this paper. Grapes, grown in the same vineyard, were harvested at commercial maturity. The harvested grapes were transported in plastic boxes (30 kg) to the cellar where they were processed following classic techniques for processing white and red wines. The grapes were destemmed and crushed, and 50 mg/l of SO 2 (antioxidant and antimicrobial agent) was added. The alcoholic fermentation was carried out in steel tanks (1000 L), keeping the fermentation temperature less than 30 C and 20 C in red and white vinification, respectively. The end of fermentation was defined as reducing sugar

5 152 Lozano et al. concentration below 2 g/l. At this time, the wine was stored in smaller volume tanks (100 L) for the e-nose measurements. GC-MS Analysis. The aromatic chemical composition of the wines was determined by custom GC-MS method, and 19 odorants were analyzed. The chemical compounds analyzed were acids (butyric acid, decanoic acid, hexanoic acid, isobutyric acid, isovaleric acid, and octanoic acid), alcohols (1-hexanol and 2-phenylethanol), esters (hexyl acetate, ethyl butyrate, ethyl decanoate, ethyl hexanoate, ethyl isovalerate, ethyl lactate, ethyl octanoate, isoamyl acetate, isobutyl acetate, diethyl succinate, and phenyl ethyl acetate) and phenols (4-vinyl-guaiacol). In this analysis, wine (50 ml) was extracted with etherhexane (Merck, Kenilworth, NJ) (1:1), three times (4, 2, 2 ml each). The extract was dried over anhydrous sodium sulfate and concentrated in a micro-kuderna-danish concentrator under a stream of pure N 2 to 500 µl. The extract was spiked with the internal standard solution (3-octanol) and injected into Hewlett Packard (HP; Palo Alto, CA) 6890 fitted to an HP Mass Selective 5973, which is equipped with the Chemstation software (Agilent, Avondale, PA). The oven initial temperature was 55ºC, held for 2 min, and then raised at 160ºC with a rate of 1.5ºC/min. The carrier gas was helium with a 1 ml/min flow. 1 µl of extract was injected. The column (30 m x 0.25 mm and 0.25 µm film thickness) was a HP-FFAP (Agilent). The operationing conditions of the HP Mass Spectrometer were as follows: ion source temperature, 200ºC; interface temperature, 200ºC; electron impact energy, 70eV; voltage, 200 V; mass range m/z, 25 to 500; sec/scan and delay time, 1.5 min. Calibration graphs were prepared for all the compounds by the analysis of synthetic samples containing known amounts of odorants. Compounds were identified by their retention time and of their mass spectrum (SCAN) compared to reference patterns. Results and Discussion As noted above, the designed system was installed in an experimental winery in Madrid (Figure 1) for the on-line and in situ monitoring of wine evolution directly in tanks. The tanks were filled with wine made from and grapes sourced from the Madrid OD. Measurements began in January and finished nine months later in September. Study of wine evolution by GC-MS. Several wine samples were taken from each of the tanks once a month for GC-MS analysis and tasting by a sensory panel. Chemical analysis and panel tasting showed that the wines experienced a loss of quality due to oxidation and an increase of volatile acidity in the first two months after fermentation (stages 1 to 3) because a lack of nitrogen in the sparging gas stream for several days at the three month treatment stage led to air exposure of the wine and SO 2 levels were not corrected. As a consequence, an accelerated oxidation of wines and a corresponding increase in volatile acidity occurred. At four months, SO 2 and total acid were corrected with the addition of SO 2 and tartaric acid respectively, in order to prevent further wine spoilage (between stages 3 and 4). Between five and six months, the wine was substituted with the same wine stored in a higher volume tank due to high spoilage. In the last stage, an increase of volatile acidity Table 1 Aromatic compounds in wine obtained in GC-MS analysis (mg/l). Jan 16 Stage 1 Feb 16 Stage 2 March 18 Stage 3 April 23 Stage 4 May 25 Stage 5 July 14 Stage 6 Sept 01 Stage 7 Alcohols 1-Hexanol Phenylethanol Sum Acetates Isobutyl acetate Isoamyl acetate Hexyl acetate Phenyl ethyl acetate Sum Esters Ethyl butyrate Ethyl hexanoate Ethyl octanoate Ethyl decanoate Ethyl lactate Diethyl succinate Sum Acids (4-5) Isobutyric acid Butyric acid Isovaleric acid Sum Acids (6-10) Hexanoic acid Octanoic acid Decanoic acid Sum Phenols 4-Vinyl-guaiacol

6 Electronic Nose for Wine Cellars 153 was observed reaching 0.6 g/l and 0.8 g/l for the and wines, respectively. A decrease of aromatic compounds (mainly alcohols, acetates, and esters) happened in the last months. This aspect was more remarkable for the wine. Tables 1 and 2 summarized the evolution of aromatic compound concentration obtained with GC-MS for the and wines, respectively. A typical chromatogram for wine is shown in Figure 7 as an example of GC-MS profile. Sampling dates and main incidences are collected in Table 3. Study of wine evolution by e-nose. The e-nose system was continuously in use for nine months after grape juice fermentation was completed. The evolution of the wine was confirmed by chemical and sensory analysis. Two e-nose measurements of each tank and a third e-nose measurement for calibration were performed daily. The data of the sensors were stored in disks and processed remotely via an internet connection. All data were normalized before analysis, as described in Materials and Methods. More than 500 measurements were made in each tank. In order to obtain a better visualization of wine evolution, the e-nose analysis was compared to the GC-MS analysis to determine the correlation between sampling methods at seven stages. Responses of the individual sensors were defined with regard to the minimum resistance to 12% (v/v) of ethanol for all the measurements. PCA was employed for data visualization and reduction of dimensionality. Further, analysis based on ANN was employed for discrimination and classification purposes (Duda et al. 2001). All data obtained were normalized before analysis. Figures 8 and 9 show the PCA score plot for the measurements of each tank. The clusters corresponding to the different time points of wine sampling are clearly separated. The spoiling and the correction of the wine can be detected between stages 2 and 4 for Figure 7 Chromatogram of minor compounds in wine: (1) isobutyl acetate, (2) ethyl butyrate, (3) isoamyl acetate, (4) ethyl hexanoate, (5) hexyl acetate, (6) ethyl lactate, (7) 1-hexanol, (8) 3-octanol (internal standard), (9) ethyl octanoate, (10) isobutyric acid, (11) butyric acid, (12) ethyl decanoate, (13) isovaleric acid, (14) diethyl succinate, (15) phenyl ethyl acetate, (16) hexanoic acid, (17) 2-phenylethanol, (18) octanoic acid, (19) 4-vinyl-guaiacol, (20) decanoic acid. Abundance is expressed in arbitrary units. Table 2 Aromatic compounds in wine obtained in GC-MS analysis (mg/l). Jan 16 Stage 1 Feb 16 Stage 2 March 18 Stage 3 April 23 Stage 4 May 25 Stage 5 July 14 Stage 6 Sept 01 Stage 7 Alcohols 1-Hexanol Phenylethanol Sum Acetates Isobutyl acetate Isoamyl acetate Hexyl acetate Phenyl ethyl acetate Sum Esters Ethyl butyrate Ethyl hexanoate Ethyl octanoate Ethyl decanoate Ethyl lactate Diethyl succinate Sum Acids (4-5) Isobutyric acid Butyric acid Isovaleric acid Sum Acids (6-10) Hexanoic acid Octanoic acid Decanoic acid Sum Phenols 4-Vinyl-guaiacol

7 154 Lozano et al. the (Figure 9). Substitution of a different unspoiled wine sample between measurements 4 and 5 is also detected by the e-nose: cluster 5 is close to cluster 2, corresponding to the wine at an earlier evolution stage prior to the onset of spoilage (Figure 8). Table 3 Sampling data and process remarks. Date Sample Remarks January 1 Measurements start March 2 Increase of volatile acidity and ph April 3 Wine oxidation May 4 Acidity corrected June 5 wine substituted July 6 Decrease of aromatic compounds September 7 Increase of volatile acidity In the classification tasks, seven classes were learned, corresponding to different stages of the wine evolution for the nine months of study. Leave-one-out was used as a validation method. 100% of classification success (percentage of cases correctly classified in validation) was obtained in both cases with BP and PNN. Prediction of GC-MS parameters. The aforementioned 19 compounds analyzed in GC-MS profiles were used as predictor variables. A model was created in order to predict these responses from sensor measurements. In this way, concentration of chemical compounds in wine determined by GC MS was correlated with e-nose response PLS regression analysis. Table 4 shows the correlation coefficients and errors (RMSEC and SEC) obtained for the correlation of each compound. As an example, Figures 10 and 11 show the plots estimated by Figure 8 PCA score plot of measurements of wine (stored in tank A). Table 4 Coefficients and errors obtained in correlation. Compound Correlation Coefficient RMSEC SEC Isobutyl acetate Ethyl butirate Isoamyl acetate Ethyl hexanoate Hexyl acetate Ethyl lactate Hexanol Ethyl octanoate Isobutyric acid Butyric acid Ethyl decanoate Isovaleric acid Diethyl succinate Phenyl ethyl acetate Hexanoic acid Phenylethanol Octanoic acid vinyl-guaiacol Decanoic acid Figure 9 PCA score plot of measurements of wine (stored in tank B). Figure 10 Concentration of hexyl acetate determined by GC-MS estimated by PLS model ( predicted value ) vs. real values.

8 Electronic Nose for Wine Cellars 155 Figure 11 Concentration of ethyl octanoate determined by GC-MS estimated by PLS model ( predicted value ) vs. real values. PLS model vs. real values of the concentration of hexyl acetate and ethyl octanoate. This model cannot quantitatively predict the concentrations of chemical compounds as the tendency is not clear in many compounds, e.g., hexanol, isovalerianic acid, diethyl succinate, and vinylguaiacol. A semiquantitative prediction could be performed with this model for some of the chemical compounds analyzed. This model could be improved by analyzing more samples with a broader range of responses. Conclusion A novel system based on an array of thin film sensors and a modified headspace sampling method was used to monitor changes in two wines over a nine-month time period. The advantages of this system are the in situ and on-line possibilities of performing continuous analysis without physically taking wine samples. Results showed that the system is capable of detecting changes in wine aroma compounds over nine months and useful in detecting wine spoilage. A correlation based in PLS was performed to predict GC-MS analysis of the wine using sensor response without taking samples. Future measurements of different wine samples will improve this model in order to quantitatively predict the concentrations of chemical compounds. Literature Cited Adami, A., L. Lorenzelli, V. Guarnieri, L. Francioso, A. Forleo, G. Agnusdei, A.M. Taurino, M. Zen, and P. Siciliano A WO 3 -based gas sensor array with linear temperature gradient for wine quality monitoring. Sensor. Actuat. B-Chem. 117: Apetrei, I.M., M.L. Rodríguez-Méndez, C. Apetrei, I. Nevares, M. del Alamo, and J.A. de Saja Monitoring of evolution during red wine aging in oak barrels and alternative method by means of an electronic panel test. Food Res. Int. 45: Carey, W.P., and S.S. Yee Calibration of nonlinear solid-state sensor arrays using multivariate regression techniques. Sensor. Actuat. B-Chem. 9: Chatonnet, P. Discrimination and control of toasting intensity and quality of oak wood barrels Am. J. Enol. Vitic. 50: Chatonnet, P., and D. Dubourdieu Using electronic odor sensors to discriminate among oak barrel toasting levels. J. Agric. Food Chem. 47: Cynkar, W., R. Dambergs, P. Smith, and D. Cozzolino Classification of Tempranillo wines according to geographic origin: Combination of mass spectrometry based electronic nose and chemometrics. Anal. Chim. Acta 660: Devarajan, Y.S., B.W. Zoecklein, K. Mallikarjunan, and D.M. Gardner Electronic nose evaluation of the effects of canopy side on Cabernet franc (Vitis vinifera L.) grape and wine volatiles. Am. J. Enol. Vitic. 62: Di Natale, C., F.A.M. Davide, A. D Amico, P. Nelli, S. Groppelli, and G. Sberveglieri An electronic nose for the recognition of the vineyard of a red wine. Sensor. Actuat. B-Chem. 33: Duda, R.O., P.E. Hart, and D.G. Stork Pattern Classification. Wiley-Interscience, New York. Gardner, D.M., B.W., Zoecklein, and K. Mallikarjunan Electronic nose analysis of Cabernet Sauvignon (Vitis vinifera L.) grape and wine volatile differences during cold soak and postfermentation. Am. J. Enol. Vitic. 62: Gardner, J., T. Pearce, T. Nagle, and S.S. Schiffmann Handbook of Machine Olfaction. Wiley/VCH, New York/Weinheim. Geladi, P., and B.R. Kowalski Partial least-squares regression: A tutorial. Anal. Chim. Acta 185:1-17. Gutierrez-Osuna, R Pattern analysis for machine olfaction: A review. IEEE Sens. J. 2: Hierlemann, A., U. Weimar, G. Kraus, M. Schweizer-Berberich, and W. Göpel Polymer-based sensor arrays and multicomponent analysis for the detection of hazardous organic vapors in the environment. Sensor. Actuat. B-Chem. 26: Lozano, J., J.P. Santos, and M.C. Horrillo Classification of white wine aromas with an electronic nose. Talanta 67: Lozano, J., J.P. Santos, M. Aleixandre, I. Sayago, J. Gutierrez, and M.C. Horrillo Identification of typical wine aromas by means of an electronic nose. IEEE Sens. J. 6: Lozano, J., J.P. Santos, T. Arroyo, M. Aznar, J.M. Cabellos, M. Gil, and M.C Horrillo. 2007a. Correlating e-nose responses to wine sensorial descriptors and gas chromatography mass spectrometry profiles using partial least squares regression analysis. Sensor. Actuat. B-Chem. 127: Lozano, J., J.P. Santos, J. Gutiérrez, and M.C. Horrillo. 2007b. Comparative study of sampling systems combined with gas sensors for wine discrimination. Sensor. Actuat. B-Chem. 126: Lozano, J., T. Arroyo, J.P. Santos, J.M. Cabellos, and M.C. Horrillo Electronic nose for wine ageing detection. Sensor. Actuat. B-Chem. 133: Penza, M., and G. Cassano Recognition of adulteration of Italian wines by thin-film multisensor array and artificial neural networks. Anal. Chim. Acta 509: Ragazzo-Sanchez, J.A., P. Chalier, and C. Ghommidh Coupling gas chromatography and electronic nose for dehydration and desalcoholization of alcoholized beverages: Application to off-flavour detection in wine. Sensor. Actuat. B-Chem. 106: Rickerby, D.G., M.C. Horrillo, J.P. Santos, and P. Serrini Microstructural characterization of nanograin tin oxide gas sensors. Nanostruct. Mater. 9: Santos, J.P., et al A comparative study of sensor array and GC- MS: Application to Madrid wines characterization. Sensor. Actuat. B-Chem. 102: Sundgren, H., F. Winquist, I. Lukkari, and I. Lundstrom Artificial neural networks and gas sensor arrays: Quantification of individual components in a gas mixture. Meas. Sci. Technol. 2: Wang, X., W.P. Carey, and S.S. Yee Monolithic thin-film metaloxide gas-sensor arrays with application to monitoring of organic vapors. Sensor. Actuat. B-Chem. 28: Zoecklein, B.W., Y.S. Devarajan, K. Mallikarjunan, and D.M. Gardner Monitoring effects of ethanol spray on Cabernet franc and Merlot grapes and wine volatiles using electronic nose systems. Am. J. Enol. Vitic. 62: