Analysis of CO and CH4 gas mixtures by using a micromachined sensor array

Analysis of CO and CH4 gas mixtures by using a micromachined sensor array

Sensors and Actuators B 78 (2001) 40±48 Analysis of CO and CH4 gas mixtures by using a micromachined sensor array S. Caponea,*, P. Sicilianob, N. BaÃ...

545KB Sizes 2 Downloads 86 Views

Sensors and Actuators B 78 (2001) 40±48

Analysis of CO and CH4 gas mixtures by using a micromachined sensor array S. Caponea,*, P. Sicilianob, N. BaÃrsanc, U. Weimarc, L. Vasanellib a

Dipartimento Ingegneria dell'Innovazione, UniversitaÁ di Lecce, Via per Arnesano, 73100 Lecce, Italy Istituto per lo Studio di Nuovi Materiali per l'Elettronica I.M.E.-C.N.R., Via per Arnesano, 73100 Lecce, Italy c Institute of Physical and Theoretical Chemistry, University of TuÈbingen, Auf der Morgenstelle 8, 72076 TuÈbingen, Germany b

Abstract An array of highly sensitive and mechanically stable gas sensors based on different sol±gel fabricated Pd-doped SnO2 nanocrystalline thick ®lms has been developed for the analysis of ternary mixtures in the concentration ranges of 0±100 ppm CO, 0±4000 ppm CH4 and 0± 50% relative humidity. The selectivity of the sensors has been modulated by varying the percentage of Pd content and the contacts geometry, while the use of micromachined hotplates as substrates for the sensors allowed a reduction of heater power consumption and a fast and accurate temperature control. Principal component analysis (PCA) as pattern recognition and principal component regression (PCR) as multicomponent analysis method have been used to analyze these mixtures qualitatively and quantitatively obtaining good results. # 2001 Elsevier Science B.V. All rights reserved. Keywords: Gas sensors array; Micro-hotplates; Pattern recognition; Multicomponent analysis

1. Introduction There is a growing need for a small and portable air pollution monitoring device that is inexpensive, reliable and able to detect and identify a wide range of air contaminants (CO, NO2, CH4, SO2, etc.), estimating their concentration either singly or in mixtures. Therefore, a considerable amount of scienti®c research has been directed towards the development of high-sensitive and high-performance gas-sensing devices to perform a qualitative and quantitative analysis of environments characterized by complex mixtures of gases and vapors. However, the feasibility of these electronic devices has been demonstrated only recently thanks mostly to considerable advances in gas-sensing area with solid-state gas sensors. Among such sensors, those based on conductance changes of semiconducting metal oxides, such as SnO2, result to be the most suitable devices to ful®ll a compromise between some principal advantages (low costs, high sensitivity, short response time, simple measurement electronics) and some well-known disadvantages (lack of selectivity, poor long-term stability, humidity dependence). Since high discrimination capability is required for environmental pollution monitoring, several approaches have * Corresponding author. Tel.: ‡39-832-320244; fax: ‡39-832-325299. E-mail address: [email protected] (S. Capone).

been studied for improving the sensing properties of semiconductor-oxide gas sensors and for modulating their selectivity, the most common consisting in the addition of noble metal sensitizers and/or active catalysts, in the control of preparation conditions and structuring of selective micromachined membranes, in the use of different operation temperatures and special measurements techniques (thermo-pulse operation, frequency-dependent analysis, etc.) [2±7]. However, the most attractive and promising approach to overcome the intrinsic non-selectivity of the sensors consists in the development of an electronic nose [15]. An electronic nose is composed of a chemical sensing system (e.g. an array of different sensing elements with partly overlapping sensitivity) and an automated pattern recognition system. Each chemical compound presented to the sensor array produces a signature or pattern characteristic of the gas. By presenting many different chemicals to the sensor array, a database of patterns is built up. This database of labeled signatures is used to train the pattern recognition system. The goal of this training process is to con®gure the recognition system to classify and quantify each chemical compounds in the mixture. The quantity and complexity of the data collected by sensors array is a powerful source of information but pattern recognition and multivariate analysis methods are commonly required to extract the maximum useful information from the sensor output signals, allowing a rapid data interpretation and

0925-4005/01/$ ± see front matter # 2001 Elsevier Science B.V. All rights reserved. PII: S 0 9 2 5 - 4 0 0 5 ( 0 1 ) 0 0 7 8 9 - 4

S. Capone et al. / Sensors and Actuators B 78 (2001) 40±48

characterization of a particular ambient atmosphere. To this purpose a variety of traditional supervised and unsupervised chemometric analysis methods like principal component analysis (PCA), cluster analysis (CA), multiple linear regression (MLR), principal component regression (PCR), partial least squares (PLS) is available together with nonlinear methods based on arti®cial neural network (ANN) [11±14]. This work reports the development of an array of eight different microsensors based on Pd-doped SnO2 nanocrystalline sensing layers deposited by the sol±gel technology in the form of thick ®lms on the top of Si-micromachined structures. This kind of structures consists of a Pt integrated heater and a suspended dielectric membrane, that isolates the high temperature area from the silicon bulk, allowing a reduction of power consumption, a good temperature uniformity and a low thermal inertia [1]. Combining the high sensitivity of thick ®lm SnO2 sensors with the fast response time of Si-micromachined substrates, we joined the advantages of both technologies. The application of the multisensor system to the analysis of binary gas mixtures of carbon monoxide and methane (CO/CH4) in air at different relative humidity levels (0, 30, 50% R.H.) has been considered. We selected CO and CH4 as target gases because are both toxic and hazardous gases present in a domestic and/or industrial environment. Moreover, we chose Pd as dopant in SnO2 sensors to promote the gases detection mechanisms based on surface chemical reactions between the adsorbed oxygen species and the considered gases. In fact, it is known that Pd (as Pt) is a catalytically active metal that improves the sensitivity to reducing gases, such as carbon monoxide and methane, through chemical or electronic interaction [2,3]. The sensing elements of the array were diversi®ed by varying the Pd-doping content (0.2, 2 wt.% of Pd) and the geometry of the Pt contact electrodes in comb-like and gap con®guration with differently spaced ®ngers. The idea was that to modulate the selectivity of each sensors of the array both by different Pd-doping and by simply changing the geometric arrangement of the electrodes since it leads to a different Schottky barrier modulation at the SnO2 semiconductor/metal contact interface and so to a different overall resistance modulation [10]. PCA as pattern recognition (PARC) technique, and PCR as multivariate analysis method have been applied to the input matrix that organizes in a vectorial way the response outputs of the sensors with the aim to identify and quantify the different binary CO/CH4 mixtures. We obtained a good classi®cation of the different mixtures and an actual quanti®cation of individual compounds in mixture. 2. Experimental A sensor array consisting of eight Pd doped SnO2 based sensors were used for this experimental work. The sensors

41

Table 1 Parameters of the sensor array Sensors label

Sensing layer

Pd content Electrodes (wt.%) geometry

A B C D E F G H

SnO2±Pd SnO2±Pd SnO2±Pd SnO2±Pd SnO2±Pd SnO2±Pd SnO2±Pd SnO2±Pd

0.2 0.2 0.2 2 2 2 0.2 0.2

Distance between a pair of electrodes (mm)

Interdigitated 5 Interdigitated 10 Interdigitated 100 Interdigitated 10 Interdigitated 50 Interdigitated 100 Gap 10 Gap 100

were fabricated by the thick-®lm technology starting from a powder prepared by the sol±gel route and doped with 0.2 and 2 wt.% of Pd by impregnation with the corresponding chlorides. The sensitive layers were obtained by mixing the calcinated powders with an organic solvent, obtaining a printable paste. A drop of this paste was transferred on the top of Si micromachined hotplates equipped on the backside with a meander Pt resistive type heater and on the frontside with differently spaced Pt contacts in interdigitated and gap con®guration. Subsequently, a short ®nal annealing at high temperature removed the solvent and bound the sensing layer to the substrate. Table 1 shows some parameters of sensor array. A detailed description of the preparation of the micromachined dielectric membranes used as substrates can be found in [1], while the details related to the preparation and deposition of sensing layers in [2,4]. Fig. 1 shows schematically a top view (a) and a cross-section of a single micromachined structure drop-coated by the SnO2±Pd sensing layer (b). The sensors, mounted on standard TO-4 sockets, were placed in a te¯on test chamber, capable to contain up to eight sensors simultaneously. Methane and carbon monoxide were introduced through stainless steel pipes in the test cell, both alone and in mixtures, by combining different concentrations of each gas. The well-de®ned test gases mixtures were produced from commercially available certi®ed bottles of pure gases in synthetic air, which are further diluted with synthetic air and/or mixed among them in de®ned proportions to ®nal composition of the different mixtures, by a mass ¯ow controller system, consisting of several computerdriven mass ¯ow controllers (Tylan MFCs) and electrovalves. The concentration ranges of test gases used for the mixtures were of 0±4000 ppm for methane and 0± 100 ppm for carbon monoxide. Table 2 shows all the combinations between the concentrations of CO and CH4 used for the measurements. Since water vapor is a common component of a real ambient atmosphere, its in¯uence on the response of SnO2-based sensors to test gases has been considered by performing all the measurements both in dry and in humid air (0, 30 and 50% relative humidity). The humidi®cation of the mixtures was obtained by the saturation method with a

42

S. Capone et al. / Sensors and Actuators B 78 (2001) 40±48

Fig. 1. Scheme of a single micromachined structure drop-coated by a sensing layer.

glass bubbler; the desired humidity level was ®xed by mixing the saturated water vapors with dry air in de®ned proportions. The total ¯ow rate was always 200 ml min 1. All the sensors operated at a constant temperature of 3508C. Previous experiments had shown that at this temperature the Pd-doped SnO2 sensors of the array show an high sensitivity to both carbon monoxide and methane, that usually need high temperatures to be detected. The electrical resistance of the sensor array elements was measured by a multimeter (Keithley DMM 199) equipped with a multiplexer to switch between the elements. A standard test cycle (30 min gas

Table 2 Combinations between selected CO and CH4 concentrations in the CO/ CH4 mixtures used in the gas sensing tests

exposure±30 min air recovering) was followed for each mixture, and it was repeated three times in order to prove the repeatability of the sensor array response to a same gaseous ambient and in order to collect enough data for the subsequent signal processing and pattern recognition analysis. Any preference in the choice of the sequence of the measurements related to all different classes of mixtures was adopted, but a random sequence among them was selected, so to avoid in such a way memory effects of the previous measurement cycle. The sensor response is taken as the ratio R0/Rgas between the resistance of the sensor in the reference gas (air) and the resistance in the presence of the test gas and/ or binary CO/CH4 mixture. 3. Results and discussion The problem related to this work, that is to analyze qualitatively an ambient atmosphere characterized by binary mixtures of CO and CH4 in air by discriminating mixtures of different composition and to estimate the concentration of each compound in mixture, is rather complicated, because chemical reactions between compounds or intermediates might occur in the mixture or at the surface causing enhancement/interference effects. Humidity is a further interfering compound that in¯uences the interaction of SnO2 gas sensors with reducing gases. Moreover, the interfering mechanisms might be different for CO and CH4, in fact, it seems that water vapor competes with methane, and not to carbon

S. Capone et al. / Sensors and Actuators B 78 (2001) 40±48

monoxide, in reacting on the SnO2 surface with the same oxygen sites. However, the effect of water vapor is controversial and it depends on the history of the sample [8,9]. Fig. 2 shows the radar plots of the sensor array response to only three selected mixtures among the all analyzed: 100 ppm CO, 3000 ppm CH4 and the mixture 100 ppm CO/3000 ppm CH4, in condition of dry air (0% R.H.) (a); 30% R.H. (b) and 50% R.H. (c). These star-like icons that plot the gas response of each sensor as a radial vector are

Fig. 2. Radar plots of the sensor array response to three selected mixture among the all analyzed: 100 ppm CO, 3000 ppm CH4, and the mixture 100 ppm CO/3000 ppm CH4 in dry air (a), 30% R.H. (b), 50% R.H. (c); (ppm CO/ppm CH4) are the labels for the CO/CH4 mixtures.

43

often used for an easy visualization and simple interpretation of the main features of the pattern recognition results. One can ®rst observe the qualitative difference between the response to pure gases and to mixture. The pattern shape of 100 ppm CO is different from that of 3000 ppm CH4, while the pattern of the mixture is more similar in shape to that of methane compared to that of carbon monoxide, indicating an higher contribute of methane to the sensor array response. Variations in the shape of the plots characterize the discrimination power of the array (selectivity), while the size represents the array's sensitivity to that particular mixture. The sensors resulted very sensitive both to CO and CH4, with an higher response to CH4 (3000 ppm)

Fig. 3. Radar plots of the sensor array response to each selected mixtures in the three situation of dry air, 30% R.H., 50% R.H.; 100 ppm CO (a), 3000 ppm CH4 (b), 100 ppm CO/3000 ppm CH4 (c).

44

S. Capone et al. / Sensors and Actuators B 78 (2001) 40±48

Fig. 4. PCA results in the plane of the first two principal components for the data set related to dry air.

respect to CO (100 ppm) in dry air, vice versa in humid air. The star related to the mixture is always external to those related to the pure gases, indicating generally an enhancement of the response of the SnO2/Pd sensors when carbon monoxide and methane are present simultaneously in the same atmosphere. The radar plots, reported in Fig. 3, compare the sensor array response to each of three selected mixtures in the different humidity condition considered in the experimental tests. Despite the utility of these radar plots to display sensor response in a easy graphical format, a more complete data processing procedure should be able to analyze the data as a whole, pointing out the differences between the different data sets. Therefore, PCA was performed on the whole data set of all measurements, with the aim to study the ability of array to discriminate the different classes of CO/CH4 mixtures. PCA is a powerful linear pattern recognition technique widely used in gas-sensing area to extract the main relations in the data matrix containing the sensor array responses. It projects the data in a new space of lower dimensionality, where the new axis (principal components) are linear combinations of the original axis (sensors) and are calculated from the latter in such a way as explain as much of the total variation in the data as possible. They are uncorrelated, orthogonal and ranked so that each one carries more information than any of the following one. A common way for determining the principal components of a data set is by calculating the eigenvectors of the data correlation matrix. The corresponding eigenvalues give an indication of the particular contribution to the systematic variance of the respective principal component. In the PCs coordinate system a graphical representation of the data allows easy interpretation of the data structure and rapid visualization of different data classes.

Figs. 4±6 show the results of the PCA performed on the data set related to the measurements carried out in dry air, 30% R.H. and 50% R.H., respectively; the projections of each data set in the plane of the ®rst two principal have been reported. Autoscaling (i.e. make the response of each sensor have a mean zero and normalize it to the standard deviation over all the samples) was used, even if it could introduce some noise by increasing the fractional error on low signals. Analyzing the score plot related to the data set in dry air (Fig. 4), we can ®rst observe that over 97% of the variance within data is contained in the ®rst two principal components. Clusters of different CO/CH4 mixtures are overlapped, so it is clearly dif®cult to discriminate between them, but looking at the PCA plot more carefully, one can observe a modest geometric structure of the data, being the clusters characterized by the same concentration of methane stretched along parallel directions. Analyzing the PCA plot related to the more realistic situations in presence of humidity (30% and 50% R.H.) (Figs. 5 and 6, respectively) one can analogously observe, ®rst of all, that over the 99 and 100% of variance of data is explained by the ®rst two principal components. Moreover, a more evident geometric arrangement of the clusters that are oriented along preferential directions, has to be highlighted. In fact, we can identify areas in the PC1±PC2 projection plane, in which the series of clusters with same concentration of methane and different concentrations of carbon monoxide are contained, distributed along the direction in which CO concentration increases. In a similar way it is possible to draw plane-sections in which the series of clusters with a same concentration of carbon monoxide and different concentrations of methane are contained, distributed along the direction in which CH4 concentration

S. Capone et al. / Sensors and Actuators B 78 (2001) 40±48

45

Fig. 5. PCA results in the plane of the first two principal components for the data set related to 30% R.H.

increases. The directions along which CO and CH4 concentration level increases are tilted in an acute angle. Thus, the classes of CO/CH4 mixtures are not randomly placed in the representative plane of the ®rst two principal components, but they are organized according to a grid of preferential lines. Even if some similar clusters are very closed to each other or superimposed, such kind of orderly structure in the data is an useful tool to discriminate between different mixtures of the two contaminants. It should be also noticed that this order in the data is higher in the data related to humid air than that related to dry air, where really the

arrangements of the clusters are generally confused. This result could be probably attributed, as shown in the radar plot of Fig. 3 to the increase in the response to CO and the decrease in the response to CH4 with humidity. The correlation circles of the PCA performed on each data set (R:H: ˆ 0, 30, 50%) showed high correlation between the sensors A and B and between sensors D and E, hence, the PCA was performed also putting out from the analysis the sensors A and D, but not a particular difference from the previous results as regards the discrimination power of the array was observed.

Fig. 6. PCA results in the plane of the first two principal components for the data set related to 50% R.H.

46

S. Capone et al. / Sensors and Actuators B 78 (2001) 40±48

Fig. 7. Score plot in the 3-D vectorial space of the first three principal components of the PCA performed on the whole data set.

PCA was also performed on the whole data matrix collecting all the data sets related to the three humidity conditions considered in this work (dry air, 30% and 50% R.H.). A better knowledge of the con®guration of all gas mixtures in the 3-D vectorial space of the ®rst three principal components was achieved (Fig. 7). It is interesting to observe a well-de®ned separation between the data in dry air and the data in humid air. For each level of humidity two directions, along which the clouds of clusters of pure CO and pure CH4 are stretched, can be individuate. They form an acute angle and the clusters of the CO/CH4 mixtures are spread in this sector of circle. The two data set in humid air (30% and 50% R.H.) are not so well distinct, but one can still observe that the corresponding sectors are rotated by an acute angle of few degrees. So, we can deduce that, even when the data are analyzed on a large scale, PCA reveals always the trend to an orderly and geometrically organized data structure. PCA allowed us to describe qualitatively complex mixtures of two important pollutants (CO and CH4), by de®ning the distribution of the data in the three-dimensional space of the ®rst three PCs, but a well ®tting mathematical model for calibration is required to predict the concentration of each component (dependent variables) in binary gas mixtures from the sensors responses (independent variables). The quality of a model depends on the quality of the sensors signals and on the design of the experiments to evaluate the model. Most of multi-components calibration methods

Fig. 8. PCR results for the data set in dry air. Predicted concentration of both components are plotted in function of the true ones.

S. Capone et al. / Sensors and Actuators B 78 (2001) 40±48

assume that sensors respond linearly to gas and use linear relationship between the response n  m matrix X (n samples, m sensors) and the n  p, Y matrix, of p unknown concentrations present in n measurements:

47

Multiple linear regression (MLR) method calculates the best square ®t describing the data points for each gas component and each sensor using this approach, whose main disadvantage is the collinearity of matrix X due to the cross-sensitivity of the sensors.

By taking into account this problem we chose to perform a PCR analysis on our experimental data to characterize quantitatively the CO/CH4 mixtures. PCR is a simple multivariate method based on PCA and MLR. It ®rst performs PCA on X, then ®ts an MLR model, using the PCs instead of the raw data as predictors. Since the PCs are uncorrelated and orthogonal, there is no collinearity problem. The sensor array was ®rst calibrated by PCR with all the samples, then cross validation has been applied as validation criterion. Cross validation uses the same samples both for calibration and testing. The method consists in leaving out some samples from the calibration data, using them for prediction and calibrating the model by the remaining data points. This process is repeated with another subset of the calibration data, and so on until every sample has been kept out once. PCR was performed according to this validation method on the data set related to the measurements in dry and humid air. Figs. 8±10 show the principal results of the PCR analysis, comparing predicted and true concentration of CO and CH4 in dry air, 30% and 50% R.H., respectively. It can be seen that the calculated CO and CH4 concentrations follow the real values with good correlation. Table 3 reports the correlation coef®cient of the regression linear ®t and the root mean square error of prediction (RMSEP) for all the regression plots and for both CO and methane. RMSEP is a

Fig. 9. PCR results for the data set in 30% R.H. Predicted concentration of both components are plotted in function of the true ones.

Fig. 10. PCR results for the data set in 50% R.H. Predicted concentration of both components are plotted in function of the true ones.

Y ˆ bX ‡ e where b is the calibration matrix of regression coefficients and e is an error matrix. The first approach to the problem is to use a data set (Xcal, Ycal) as calibration set where the analyte concentrations for n samples are known and a validated or testing set (Xval, Yval) where the analyte concentrations for n samples have to be predicted. If Xcal and Ycal are known, bcal can be determined by linear algebra bcal ˆ …X Tcal Xcal † 1 X Tcal Y cal Once this bcal is determined, a Xval data set can be used to ®nd the concentrations of the mixtures Y val ˆ Xval bcal

48

S. Capone et al. / Sensors and Actuators B 78 (2001) 40±48

Table 3 Some parameters results of the PCR analysis 0% R.H.

Correlation factor RMSEP (ppm)

References

30% R.H.

50% R.H.

CO

CO

CO

CH4

0.90 14

0.98 0.98 293 7

CH4

0.96 0.98 408 6

CH4 0.97 332

measurement of the average difference between predicted and measured response values. RMSEP can be thus interpreted as the average prediction error, expressed in the same units as the dependent variable. 4. Conclusions An approach for the simultaneous identi®cation and quanti®cation of carbon monoxide and methane in humid air, based on the performance of an array of eight optimized SnO2/Pd sensors in combination with PCA and principal components regression methods has been presented. It has been demonstrated that qualitative and quantitative gas mixture analysis with this sensor array is feasible with suf®cient accuracy under different humidity conditions. In particular, we found that data are spacely distributed according to a well-de®ned geometric structure, that is an useful tool for subsequent discrimination processes between the different classes of CO/CH4 mixtures, based on other pattern recognition methods like discriminant analysis (DA) or CA. PCR gave discrete results for the prediction of the real level of concentration of each component (CO and CH4) in mixture, especially for medium-high concentrations of the two gases. We have also proved the compatibility between the silicon micromachining technology for manufacturing micromachined substrates and the thick ®lm technology, that produces high sensitive sensing elements. Then the use of micro-hotplate as substrate offers the advantage to minimize the heat losses of the sensors and makes possible to develop a low-cost and low-power consumption device to identify and quantify CO and CH4 in gas mixture for some practical application, in which CO and CH4 has not to overcome some safety standard levels. Finally, the in¯uence of electronic behavior of the platinum contact electrode/SnO2 interface on the selectivity of the sensor form a topic for further study. Moreover, the effect of contact geometry and of humidity on the response to these two gases needs a more carefully investigation.

[1] D. Briand, B. van der Schoot, N.F. de Rooij, A. Krauss, U. Weimar, N. BarsaÃn, W. GoÈpel, High temperature micro-hotplates for drop coated gas sensors, in: Proceedings of the Conference of Eurosensors XIII, The Hague, The Netherlands, 12±15 September 1999, pp. 703±704. [2] A. DieÂguez, J.L. Alay, A. VilaÁ, A. Cabot, A. Romano-RodrõÂguez, J.R. Morante, J. Kappler, N. BarsaÃn, U. Weimar, W. GoÈpel, Influence on the gas sensor performances of the metal chemical states introduced by impregnation of SnO2 sol±gel nanocrystal, in: Proceedings of the Conference of Eurosensors XIII, The Hague, The Netherlands, 12±15 September 1999, pp. 109±112. [3] A. Heilig, N. BarsaÃn, U. Weimar, W. GoÈpel, Selectivity enhancement of SnO2 gas sensors: simultaneous monitoring of resistances and temperatures, Sens. Actuators B 58 (1999) 302±309. [4] A. DieÂguez, A. Romano-RodrõÂguez, J.R. Morante, J. Kappler, N. BarsaÃn, W. GoÈpel, Nanoparticle engineering for gas sensor optimisation: improved sol±gel fabricated nanocrystalline SnO2 thick film gas sensor for NO2 detection by calcinations, catalytic metal introduction and grinding treatments, Sens. Actuators B 60 (1999) 125±137. [5] J. Kappler, N. Barsan, U. Weimar, A. DieÂgez, J.L. Alay, A. RomanoRodriguez, J.R. Morante, W. GoÈpel, Correlation between XPS, Raman and TEM Measurements and the Gas Sensitivity of Pt and Pd Doped SnO2 Based Gas Sensors, Conf. Proc. 9. Tagung zur FestkoÈrperanalytik, Chemnitz (D) (6/1997); Fres. Journal. Analyt. Chem. 361 (1998) 110±114. [6] A. Heilig, N. BarsaÃn, U. Weimar, M. Schweizer-Berberich, J.W. Gardner, W. GoÈpel, Gas identification by modulating temperatures of SnO2-based thick film sensors, Sens. Actuators B 43 (1997) 45±51. [7] U. Weimar, W. GoÈpel, AC measurements on tin oxide sensors to improve selectivities and sensitivities, Sens. Actuators B 26 (1995) 13. [8] R. Ionescu, A. Vancu, C. Moise, A. Tomescu, Role of water vapor in the interaction of SnO2 gas sensors with CO and CH4, Sens. Actuators B 61 (1999) 39±42. [9] D.S. Vlachos, P.D. Skafidas, J.N. Avaritsiotis, The effect of humidity on tin-oxide thick film gas sensors in the presence of reducing and combustible gases, Sens. Actuators B 24/25 (1995) 491±494. [10] Uma Jain, A.H. Harker, A.M. Stoneham, D.E. Williams, Effect of electrode geometry on sensor response, Sens. Actuators B 2 (1990) 111±114. [11] K.D. Schierbaum, J. Geiger, U. Weimar, W. GoÈpel, Specific palladium and platinum doping for SnO2-based thin film sensor arrays, Sens. Actuators B 13/14 (1993) 143±147. [12] U. Hoefer, H. BoÈttner, A. Felske, G. KuÈhner, K. Steiner, G. Sulz, Thin-film SnO2 sensor array controlled by variation of contact potential: a suitable tool for chemometric gas mixture analysis in the TLV range, Sens. Actuators B 44 (1997) 429±433. [13] W.P. Carey, K.R. Beebe, E. Sanchez, P. Geladi, B.R. Kowalski, Chemometric analysis of multisensor array, Sens. Actuators B 9 (1986) 223±234. [14] S. Vaihinger, W. GoÈpel, Multi-Component Analysis in Chemical Sensing, Sensors-A Comprehensive Survey, Vol. 2, Part II, VCH, Weinheim, 1991, pp. 192±237. [15] J.W. Gardner, P. Barlett, A brief history of electronic nose, Sens. Actuators B 18/19 (1994) 211±220.