Discrimination of carbon dioxide and fluorocarbon using semiconductor gas sensors

Discrimination of carbon dioxide and fluorocarbon using semiconductor gas sensors

Sensors and Actuators B 67 Ž2000. 258–264 www.elsevier.nlrlocatersensorb Discrimination of carbon dioxide and fluorocarbon using semiconductor gas se...

250KB Sizes 0 Downloads 93 Views

Sensors and Actuators B 67 Ž2000. 258–264 www.elsevier.nlrlocatersensorb

Discrimination of carbon dioxide and fluorocarbon using semiconductor gas sensors F. Sarry ) , M. Lumbreras LICMr CLOES, UniÕersite´ de Metz et Supelec, ´ 2 rue E. Belin, 57070 Metz, France Received 11 October 1999; received in revised form 15 April 2000; accepted 17 April 2000

Abstract In this paper, we present the discrimination of two antagonist gases, using an array composed of five commercial gas sensors Žtwo TGS 832, two TGS 813, one TGS 800.. From the characterisation of the sensor array under forane R134a, carbon dioxide and mixtures of these two gases in dry synthetic air, we have extracted two main parameters, transient and steady-state conductance values. In general, only the steady-state conductance value is taken. However, we have characterised the sensor transient behaviour by using the measurement of the slope of the conductance when there is a step change in the gas concentration. Thus, discriminant factorial analysis ŽDFA. is carried out to show qualitatively that the selectivity is improving when taking into account simultaneously the two parameters. This makes possible the determination of the nature of several gases present in the studied atmosphere. q 2000 Elsevier Science S.A. All rights reserved. Keywords: Tin dioxide gas sensor; Carbon dioxide; Forane R134a; Gas mixture; Pattern recognition

1. Introduction Gas sensors exhibit a series of unpleasant characteristics such as cross-sensitivity, drift and humidity effects, ageing, poisoning, . . . w1x. It is yet well known that it is necessary to use them in an array to achieve a discrimination of different analytes w2x. Each chemical substance contributes differently to the output according to the sensor. So the sensor’s output represents a synthesis of the chemical feature of the environment. In a first time, we have carefully characterized our array to have a good look-up table and to determine the pertinent parameters to discriminate the gases w3x. In a second time, we use a pattern recognition method to extract information contained in the sensor output set. In this study, we shall test the ability of a commercial semiconductor gas sensor array to classify carbon dioxide,

)

Corresponding author. Tel.: q33-3-87-75-9600; fax: q33-3-87-759601. E-mail addresses: [email protected] ŽF. Sarry., [email protected] ŽM. Lumbreras..

forane R134a and their mixtures. The difficulty consists on the fact that forane R134a is a reducing gas and carbon dioxide an oxidising gas. So they have an antagonist effect on the sensitive layer. We will present the results obtained by discriminant function analysis on the two chosen criteria Žsteady-state and slope conductance values..

2. Analysed gas and experimental set-up We decided to test two main gases: forane R134a, carbon dioxide and a mixture of them. These two gases take part on the green house effect, which implies global warming w4x of the earth. Moreover, carbon dioxide may also cause headache, breathing troubles, . . . Its limit exposition level is fixed at 5000 ppm w5x. Concerning forane R134a, which is the new refrigerant gas and replaced CFCs, its effects on health are yet not well known, that’s why we may find sometimes that it is not dangerous w6x and sometimes that it is harmful w4x. Its limit exposition level is fixed at 1000 ppm w7x. The sensor array consisted of five commercially available Taguchi gas sensors ŽTable 1. that were operated as suggested by Figaro Engineering w8x ŽFig. 1.. The relation-

0925-4005r00r$ - see front matter q 2000 Elsevier Science S.A. All rights reserved. PII: S 0 9 2 5 - 4 0 0 5 Ž 0 0 . 0 0 5 1 8 - 9

F. Sarry, M. Lumbrerasr Sensors and Actuators B 67 (2000) 258–264

259

Table 1 Array components, main applications Model

Main applications

Number of sensors in the array

TGS 800

Air quality control Žcigarette smoke, gasoline exhaust, . . . . Combustible gas detection Halocarbon gas detection Žforane R134a.

1

TGS 813 TGS 832

2 2

Fig. 2. Experimental set-up.

as an interfering gas. In a following step, we will introduce the influence of humidity. ship between the sensor resistance and the monitored voltage is expressed by the following equation: 3. Results and discussion RS s RL

ž

VC Vout

y1

/

Ž 1.

Each measurement follows the same rules w9x. The sensors were preheated during a period of 2 weeks to stabilize the sensitive layer, and then the different voltages were applied and maintained constant during all the experiment time. Before each measurement, a constant dry synthetic air flow is fed to the sensors for 1 h to obtain the baseline. Each measurement was made during a period of 1 h to reach the steady-state value, and a period of 1 h under dry synthetic air flow between two measurements was necessary to regenerate the sensor surface. The gas line w10x was composed of three gas bottles containing dry synthetic air, carbon dioxide and forane R134a Ž1,1,1-2 tetrafluoroethane.. The gas sensors were mounted in a gas chamber connected to this gas line. The output voltage of the sensor was read by one ArD channel of the data acquisition card manufactured by National Instruments. The mass flow control system is schematically shown in Fig. 2. We have chosen to work with dry synthetic air as a carrier gas to be sure to separate the antagonist effects of the two target gases and also because wet air is considered

Fig. 1. Electronic circuit to monitor the sensor response.

3.1. Measurement Time responses presented by all the three types of gas sensors are practically similar for the two pure gases w11x. However, TGS 813 presents no response to CO 2 and for TGS 832 and 800, the response is very weak. Fig. 3Ža. and Žb. presents the time responses of the three types of

Fig. 3. Ža. and Žb. Conductance modification for different sensors under two mixtures.

260

F. Sarry, M. Lumbrerasr Sensors and Actuators B 67 (2000) 258–264

Fig. 4. Ža. and Žb. Comparison of the steady-state conductance value for the three types of sensor under different mixtures.

sensors for a mixture: a mixture of R134a with varying concentration from 0 to 1000 ppm in CO 2rsynthetic air base Ž1000 ppm., and the contrary, a mixture of CO 2 in R134arsynthetic air base. In the two cases of mixture, TGS 800 and TGS 813 have the same response, whereas there are two different responses for TGS 832. In the first case, the response is similar to the response to a mixture containing only R134a in synthetic air. In the second case,

we obtain a higher response. The CO 2 effect is here added to the R134a effect with the characteristically CO 2 peak yet presented w9x. As it was shown in Ref. w3x, the adsorption on the sensitive layer is a very slow process. For this reason, the steady-state conductance value is at best obtained only after half an hour, which is a too long period. Taking into account the steady-state behaviour implies a too long time to obtain the identification of the gases in the studied atmosphere so the slope of the sensor conductance taken during the five 5 min constitutes a characteristic parameter of the gas nature and concentration w11x. For the pure target gases, the steady-state conductance and the slope of the conductance have a foreseeable behaviour: increasing with the concentration of the reducing gas ŽR134a. and decreasing with the concentration of the oxidising gas ŽCO 2 .. In Fig. 4Ža. and Žb., we present the sensor steady-state conductance for the two different mixtures. In the case of varying CO 2 concentration in a base of 1000 ppm of R134arsynthetic air, we note an increase of the value with the concentration of the oxidising gas. Compared to the pure target gas response, this behaviour is completely opposite. This can be explained by a predominant effect of R134a reducing gas. The behaviour of the conductance slope is similar. So as to improve the data, it was necessary to determine the pre-treatment algorithms to be applied to the sensor signals before using them as the input of the pattern recognition method. In a previous paper w1x, we have presented the influence of short-term and long-term drift effect on the sensor response. For example, the baseline response, before each measurement, decreases exponentially along the time ŽFig. 5.. To take into account these

Fig. 5. Short-term drift of the sensor response.

F. Sarry, M. Lumbrerasr Sensors and Actuators B 67 (2000) 258–264

261

Fig. 6. Loading of the array components.

effects, we have studied several pre-treatments w12,16–20x like steady-state resistance or conductance, relative steady-state resistance or conductance, fractional relative steady-state resistance or conductance and finally the slope of the resistance or conductance. Because of the several problems presented by the gas sensors, i.e. lack of selectivity, sensitivity and drift problems, we have decided to use pattern recognition methods: principal component analysis ŽPCA. and discriminant factorial analysis ŽDFA. to extract the main information and we have tested all these parameters by PCA and DFA to select the most representative ones for the five sensors of our array. 3.2. Principal component analysis Many authors w10,12x use PCA to analyse the response of tin oxide chemical sensors. The PCA method belongs to

Fig. 7. Scores plot for the three studied species on the slope and the steady-state conductance values using PCA method.

the class of the multivariable statistical methods w13x. This pattern recognition technique is a powerful supervised method usually employed to tin oxide gas sensor arrays w14,15x. To apply this method, all the characterization measurements are grouped into a database corresponding to the sensor array response matrix X. In the response matrix of the array, each column was associated to a sensor while each row was related to a different experiment. This matrix is expected to contain collinear variables, inducing the existence of some dominating types of variability that carry most of the information. The main objective of PCA w12,15x consists of expressing this information by a lower number of variables called principal components. The principal components are chosen to contain the maximum data variance and to be orthogonal. The percentage of the data variance contained in each principal component is given by the corresponding eigenvalue.

Fig. 8. DFA on the steady-state conductance value for the sensor type TGS 832.

262

F. Sarry, M. Lumbrerasr Sensors and Actuators B 67 (2000) 258–264

Fig. 9. DFA on the fractional conductance value for the sensor type TGS 832.

First, the PCA was applied to all the sensors of the same type and for the three types of sensors under forane R134a to control their response proximity. It can be clearly seen in Fig. 6 how loadings for sensors of the same type are very close. This means that it is possible to change the sensors without having to re-characterize them if the database is sufficient enough. In a second time, PCA was applied to the sensor array matrix Žtwo TGS 832, two TGS 813, one TGS 800. corresponding to the three gas experiments Žpure R134a, pure CO 2 , mixture.. The first two principal components contained more than 88% of the total variance that contained the most relevant information to classify the species. In Fig. 7, the scores for Axes 1 and 2 are shown. Three different classes can be differentiated but are overlapped. Thus, it is not good enough to conclude the capability of the array to discriminate correctly the three gases. However, the steady-state conductance study gave the best results compared to the other pre-treatments. 3.3. Discriminant factorial analysis In this part, the DFA, which also belongs to the class of multivariable statistical methods, was used. In this method, the data are separated in k a priori defined groups, to show if the variables are sufficient or not to allow a well a posteriori classification of the data in

Fig. 10. DFA on the steady-state conductance values given by the sensor array.

Fig. 11. DFA on the slope of the conductance values given by the sensor array.

their a priori groups. For this aim, the discriminant procedure consists of maximizing the differences between all the groups and minimizing these differences inside each group. The goal of DFA is to find factorial axes corresponding to a linear combination of the components that best clusters, not all the set of experiments, but the set of experiment families. This clustering will be easier as the data families are distant from each other and the experiments belonging to the same family are close. For DFA analysis, we have chosen the three families of experiments yet differentiated by PCA: forane R134a, carbon dioxide and carbon dioxiderforane R134a mixtures for each type of sensor and then for the five sensor array. Fig. 8 shows the result of the DFA using the three a priori families ŽR134a, CO 2 , mixtures. on the steady-state conductance value for the sensor type TGS 832. We can note that the different gases are well separated. Such a result was also observed for the other sensor types. A similar result was obtained with the slope conductance value. Fig. 9 shows the result of the DFA on the fractional steady-state conductance value for the sensor type TGS 832. We can note that the pre-treatment has not given a better result on the discrimination. Moreover, there is now a great proximity between the data representing CO 2 and R134a after the pre-treatment. Such a result was also observed for the other sensor types. Other parameters as relative steady-state resistance or conductance, fractional

Fig. 12. DFA on the slope and the steady-state conductance values.

F. Sarry, M. Lumbrerasr Sensors and Actuators B 67 (2000) 258–264

263

array. The mixture has shown different effects corresponding to the first introduced gas. The use of PCA has pointed out the information redundancy of sensors of the same type. PCA has differentiated the data in three families, but with too important overlappings. The use of DFA has confirmed the choice of the two principal criteria, i.e. the slope of the conductance and the steady-state conductance values. The discrimination was improved with the use of DFA when associating the two parameters and confirmed with the classification testing. To conclude, our array associated with the look-up table and DFA is efficient enough to discriminate the gases in our specific atmosphere composed of two antagonist gases. Fig. 13. Results of the classification of four unknown gases.

relative steady-state resistance or conductance, steady-state resistance and finally the slope of the resistance or conductance have not given a good discrimination w11x. For all these other parameters, the slope of the conductance has given the best values. So to obtain a best identification, we decided to use the steady-state conductance or the slope of the conductance taken during the first 5-min values. In Figs. 10 and 11, we note that the discrimination is not very clear. When using in the same time these two parameters, the method simultaneously applied to the five sensor responses gives now a good discrimination corresponding to the three tested gases. It is possible to separate without any error all the data ŽFig. 12.. Moreover, the data corresponding to a mixture of forane R134a in carbon dioxide or the contrary are yet mixed in the corresponding cluster. So DFA applied to the two selected parameters improves the array selectivity compared with the PCA results. 3.4. Classification testing To evaluate the discriminating capacity of our smart array, we have tested the classification of four data. These data were extracted randomly from the whole set. A first classification was obtained from the reduced set Žthe whole set minus the four data. with the DFA method and have given eigenvectors. These vectors were then used to place the four data on the two principal axes. Fig. 13 shows the result. We can note that the data associated to forane, carbon dioxide and mixtures without taking into account the gas order introduction have been placed in the corresponding clusters. Such a result confirms the capacity of our array to discriminate between the different gases present in our specific atmosphere.

4. Conclusion We have presented in this paper the sensor behaviour of carbon dioxide, forane R134a and their mixtures with different gas introduction order on a three-sensor type

References w1x F. Sarry, M. Lumbreras, Drift problems and contaminant effects on SnO 2 gas sensors, in: Eurosensors XII, Southampton, September 13–16, IOP Publishing, 1998, pp. 685–688. w2x J. Gardner, P. Bartlett, Performance definition and standardisation of electronic noses, Sens. Actuators, B 33 Ž1996. 60–67. w3x F. Sarry, M. Lumbreras, Gas composition determination in air conditioned system using a sensor array: characterization of three different TGS sensors, Sens. Actuators, B 59 Ž1999. 94–99. w4x M. Duminil, Criteres ` objectifs de choix d’un frigorigene, ` Revue pour le Froid, no. 808, Decembre 1994r2. ´ w5x Security data, Elf Atochem, Product: carbon dioxide, no. CEE ŽEINECS. 2046969. w6x G. Lorentzen, J. Pettersen, A new efficient and environmentally benign system for car air-conditioning, Int. J. Refrig. 16 Ž1. Ž1993. 1. w7x Security data, Elf Atochem, Product: FORANE 134a, no. CEE ŽEINECS. 212-377-0. w8x Figaro gas sensor technical reference, 1992, Figaro Engineering, PO Box 357, 1000 Skokie Blvd., Room 575, Wilmette, IL 60091, USA. w9x F. Sarry, M. Lumbreras, Evaluation of commercially available fluorocarbon gas sensor for monitoring air pollutants, Sens. Actuators, B 47 Ž1998. 113–117. w10x H. Shurmer, J. Gardner, Odour discrimination with an electronic nose, Sens. Actuators, B 8 Ž1992. 1–11. w11x F. Sarry, PhD Thesis, Universite´ de Metz, France, 1998. w12x J. Gardner, Detection of vapours and odours from a multisensor array using pattern recognition: Part 1. Principal component and cluster analysis, Sens. Actuators, B 4 Ž1991. 109–115. w13x D.F. Morrison, Multivariable Statistical Method, McGraw-Hill, 1967. w14x H. Nagle, R. Gutierez-Osuna, S. Schiffman, The how and why of electronic noses, IEEE Spectrum 35-9 Ž1998. 22–34. w15x E. Llobet, J. Brezmes, X. Vilanova, J.E. Sueiras, X. Correig, Qualitative and quantitative analysis of volatile organic compounds using transient and steady-state responses of a thick film tin oxide gas sensor array, Sens. Actuators, B 41 Ž1997. 13–21. w16x K. Ihokura, J. Watson, The Stannic Oxide Gas Sensor: Principles and Applications, CRC Press, Boca Raton, USA, 1994. w17x J. Gardner, P. Barlett, Intelligent ChemSADs for artificial odor-sensing of coffees and lagers beers, Taste Sensor Ž1994. . w18x J. Gardner, E. Hines, H. Tang, Detection of vapours and odours from a multisensor array using pattern-recognition techniques: Part 2. Artificial neural networks, Sens. Actuators, B 9 Ž1992. 9–15. w19x S. Moore, J. Gardner, E. Hines, W. Gopel, U. Weimar, A modified ¨ multilayer perceptron model for gas mixture analysis, Sens. Actuators, B 15–16 Ž1993. 344–348. w20x J. Gardner, E. Hines, M. Wilkinson, Application of artificial neural networks to an electronic olfactory system, Meas. Sci. Technol. 1 Ž1990. 446–451.

264

F. Sarry, M. Lumbrerasr Sensors and Actuators B 67 (2000) 258–264

Biographies Frederic ´ ´ Sarry graduated in Telecommunication Engineering from the University of Metz in 1993. He was awarded a PhD degree in the Laboratory of Interfaces, Components and Microelectronics in the same university in 1998. His main areas of interest are in chemical sensors, chemical vapour discrimination using sensor arrays and pattern recognition methods.

Martine Lumbreras graduated in Electrical Engineering with specialization in Solids Electronics at the University of Montpellier in 1969. She was awarded a PhD degree in 1979 at the same university. She joined the University of Metz in 1979, and she was awarded a Doctor of Sciences Degree in 1987 in this university. She is a Professor at the University of Metz since 1991, and she has created a sensor research group in 1994 as part of the Laboratory of Interfaces, Components and Microelectronics.