- Email: [email protected]
A logarithmic multi-parameter model using gas sensor main and cross sensitivities to estimate gas concentrations in a gas mixture for SnO2 gas sensors
Sensors and Actuators B 123 (2007) 1064–1070
A logarithmic multi-parameter model using gas sensor main and cross sensitivities to estimate gas concentrations in a gas mixture for SnO2 gas sensors A. Chaiyboun a,b,∗,1 , R. Traute c , T. Haas b , O. Kiesewetter c , T. Doll a,b a
b
Institute of Physics, University of Mainz, Staudinger Weg 7, 55128 Mainz, Germany Department of Solid State Electronics, Technical University of Ilmenau, P.O. Box 100565, 98684 Ilmenau, Germany c UST Umweltsensortechnik GmbH, Dieselstraße 2, 98716 Geschwenda, Germany Received 3 August 2006; received in revised form 9 November 2006; accepted 13 November 2006 Available online 8 December 2006
Abstract In a metal-oxide semiconductor gas sensor, the sensitivity of the metal-oxide resistance to concentrations of reducing gases in the surrounding atmosphere is known to be related to adsorption and desorption of gas on the redox reactions between the gas and oxygen. Changes in the electric conductance due to these reactions were measured for tin dioxide semiconductor gas sensors. In this study, we propose a model of gas sensor responding behaviour using a relationship between sensor conductance and gas concentrations in a mixture. A least-squares method fit of measured data was applied to determining the values of coefficients. The proposed method uses main and cross sensitivities the describing the response of a gas sensor. Applying two-gas sensors which show different characteristics, the gas concentrations in a gas mixture can be evaluated. The proposed method has been applied to the estimation of gas concentrations in a mixture of hydrogen–methane, carbon monoxide–methane, propane–methane, ethanol–ammonia and propane–ammonia. The concentrations determined from the response curves were accurate within a 5% error. The results indicate that the proposed model is feasible for recognition of calculated estimations in a gas mixture. This paper shows a significant result through utilization of the proposed model of gas sensor response. © 2006 Elsevier B.V. All rights reserved. Keywords: Main sensitivity; Cross sensitivity; Gas concentration; SnO2 gas sensor; Gas mixture; Modelling
1. Introduction Metal-oxide semiconductor gas sensors are suitable for the detection of oxidizing and reducing gases, since they react to their presence with a measurable change of their electrical conductivity [1–3]. Among the used semiconductor layers heated SnO2 -layers are best tested and furthest common [2]. However, a single semiconductor gas sensor is only conditionally suitable for the selective proof of certain gasses, since it exhibits cross sensitivities opposite to practically all oxidizing and reducing gases. These sensors do not have usually high gas selectivity, and its recognition characteristics are largely depending on changes in the environment, such as temperature and humidity
∗ 1
Corresponding author. Tel.: +49 6131 3922406; fax: +49 6131 3922276. E-mail address: [email protected] (A. Chaiyboun). Since June 2006 at the University of Mainz.
0925-4005/$ – see front matter © 2006 Elsevier B.V. All rights reserved. doi:10.1016/j.snb.2006.11.012
[3]. Furthermore, metal-oxide gas sensors are non-linear systems which thereby cause considerable measurement errors. To the correction of these measuring errors, heuristic–mathematical or physical–chemical models can be used. The coefficients of these models must be determined with often significant effort by calibration and adaptation of the models to the calibrating data. For the determination of each parameter, value at least a measured value under exactly defined conditions must be acquired. The reaction of gas sensors to the measured gasses can be described by different models, which mostly indicate a power-law for the interrelation between partial pressure and sensor conductance. For the metal-oxide gas sensor, Clifford and Tuma [4,5] attempted to derive from experimental results an empirical formula which described the sensor resistance as a function of several gas concentrations. The model of Clifford was developed for application to sensors of the type Taguchi (TGS) [6]. Madou and Morrison [7] employed extensive theoretical considerations concerning the influence of oxygen and reducing gases on the sensor
A. Chaiyboun et al. / Sensors and Actuators B 123 (2007) 1064–1070
conductivity. At operating temperature, the resistance of semiconductor gas sensors follows a power-law dependence on the gas concentration in the environment air [4,7,8]. Hirobayashi et al. [9] proposed a logarithmic model for detecting the individual components of gas mixtures, demonstrating good approximation results for measured gas concentrations from 100 to 1000 ppm. Latterly, artificial neural network [10–13] and other pattern recognition methods [14] are employed with some success. Faglia et al. [15] used a neural network, to which information about the current humidity is supplied by separate measurement. Using an array of semiconductor gas sensors, they attempted to eliminate the humidity influence on the measurements. However, these methods require many gas concentration measurements for each object gas, which requires significant time and effort. Usually, gas sensor responses are obtained experimentally for each pure gas forming a mixture gas, and mixture gas concentrations are determined accurately using the sensor response equation. The object of the present study is to test and expand the formulation and verification of a logarithmic model in detail and describe an accurate estimation procedure for five kinds of gas mixtures using our proposed model. The model is particularly suitable for describing the low concentration range from 0.00001 to 10 ppm. This is important for application scenarios with low gas concentrations (e.g. ozone or carbon monoxide). Our logarithmic model delivers also good accuracy for the complete gas concentration range from 0 to 1000 ppm. The results show that accurate estimation is feasible using this model through detailed exploration of sensor output resistance characteristics. 2. A model of sensor characteristic curve A SnO2 gas sensor is reliable for detecting combustible gases. The sensors GGS1470 and GGS4470, herein after referred to as sensor 1 and sensor 2, respectively, are reliable and stable, but have no gas selectivity and show cross sensitivities to several gases; i.e. they sense different gases at random. Both sensors GGS XXX0 are commercial SnO2 -thick film gas sensors of the company UST. They consist of an Al2 O3 -carrier substrate with a structured platinum layer on the front and back side for the contacts and the heater. As an electrode structure a standard structure (SS) with two wide electrodes or an interdigital structure (IDS) is used. The sensitive layer (SnO2 ) is deposited via the screen printing technique on the Pt-contacts. The used sensors differ in their preparation of the sensitive layer, which leads to different main and cross sensitivities to the gases to be detected; for example, the sensor GGG1470 has main sensitivity to CH4 and the sensor GGS4470 to NH3 . The gas sensors used in the present study were sensor 1 and sensor 2. In many applications, these two types of sensors are widely used due to their advantages such as small size, light weight and high sensitivity. The sensor 1 is cross sensitive to hydrogen (H2 ), propane (C3 H8 ) and carbon monoxide (CO); this sensor also has a main sensitivity to methane (CH4 ) and it is especially suitable for the leakage detection of combustible gases. The sensor 2 is mainly sensitive to ammonia (NH3 ) and cross sensitive to
1065
Table 1 Lower explosion limit (LEL) for the used gases
LEL (vol%)
C2 H5 OH
C3 H8
CH4
CO
H2
NH3
3.3
2
5
10
17
4
ethanol (C2 H5 OH), propane (C3 H8 ) and other hazardous gases, and is often used in gas leakage alarms. The following combinations of testing mixture gases were chosen: (1) CH4 + H2 , (2) CH4 + C3 H8 , (3) CH4 + CO, (4) NH3 + C3 H8 , and (5) NH3 + C2 H5 OH. In Table 1, the lower explosion limit (LEL) is shown for each gas we used. By different limit ranges of each gas, one could have obtained better results, but the analysis and modelling of the small concentration range in the close proximity to 0 ppm were the centre of attention. Typically, semiconductor gas sensors react to different gases. Therefore, as a sensor signal the relative conductivity S(c) is used [16,17]. It corresponds to the resistance value of the sensor under gas influence related to the sensor resistance in pure atmosphere. S(c) =
Rg (c) G0 = , R0 Gg (c)
(1)
where G0 is the baseline conductance (i.e. in the presence of clean air) and Gg is the steady-state conductance of the sensor in the presence of a given gas or gas mixture. As previously mentioned, the function (1) applies only at constant temperature T of the sensor surface. It is well known that the electrical resistance Rg () of a gas sensor decreases as the gas concentration c (ppm) increases. Therefore, the sensor relative conductivity can be expressed as: S(c) = a − b ln(c + 0.5),
(2)
where a and b are the coefficients and the constant 0.5 is used for the defining of the sensor response at the gas concentration of 0 ppm (i.e. in the presence of clean air). The Eq. (2) shows that the sensor resistance or relative conductivity is proportional to the logarithm of the concentration c, and is satisfied when the temperature and relative humidity of the environment are fixed. The condition for maximum value of sensor resistance is the sensor response at 0 ppm measurement in clean air and that of minimum value of sensor resistance is the lower limit of sensor response at high gas concentrations. Gas sensor relative conductivity is shown in Fig. 1(a and b) as a function of the gas concentration for each object gas. The gas sensors 1 and 2 were measured to individual gas. For sensor 1 each object gas was varied from 1 to 1000 ppm (measured four times at 1, 10, 100, 1000 ppm) and for sensor 2 each gas was varied from 1 to 1000 ppm (measured seven times at 1, 3, 10, 30, 100, 300 and 1000 ppm). We took three measurements for each gas concentration, the mean value of which was used in the analysis. The solid line approximations connecting the experimental data were obtained from Eq. (2), using a leastsquares method.
1066
A. Chaiyboun et al. / Sensors and Actuators B 123 (2007) 1064–1070
gas sensor responses determined from fits for individual gas component in Fig. 1(a). Eq. (3) is a function for the gas sensor response, in which the cross sensitivity of the sensor is integrated (in this case the cross sensitivity to hydrogen); and we can regard this function as an approximation of the first order. The concentrations of cross gases are calculated subject to the main sensitivity of sensor 1; in our case subject to the main CH4 -sensitivity. The measured data of the gas mixture in Fig. 2 were fitted by Eq. (3) as a first approximation and the results show that the estimated curves are satisfactory, but with a large average error of 14% at low concentration from 1 to 10 ppm, and at high concentrations the error tends to decrease. The maximum error is 13.86% and 7.92% for sensor 1 and sensor 2, respectively. Fig. 3(a–c) shows the sensor response of sensor 1 to mixture gasses of methane and hydrogen, methane and propane, and methane and carbon monoxide, respectively. Methane (the main gas) was varied from 10 to 1000 ppm (measured three times at 10, 100, 1000 ppm) and hydrogen was varied from 1 to 1000 ppm (measured four times at 1, 10, 100 and 1000 ppm). Following a similar procedure to that for the methane and hydrogen, gas mixtures of methane and propane and of methane and carbon monoxide were measured at the same concentrations. At lower cross gas concentration, there is a greater sensitivity to methane concentration, as shown in Fig. 3(a–c), and the sensor relative conductivity decreased with increasing methane concentration as expected. For sensor 2, output characteristics of compounds of ammonia and propane and of ammonia and ethanol are shown in Fig. 4. The gas concentrations of cross gases of propane and ethanol were varied from 0.00001 to 1000 ppm and were measured five times (at 0.00001, 1, 10, 100 and 1000 ppm) at a constant ammonia concentration of 100 ppm.
Fig. 1. Measured values and approximated algorithmic fitting for each object gas using (a) sensor 1 and (b) sensor 2. S(c) the sensor relative conductivity.
3. Characteristic equation of two-gas system In this study, we investigated the sensor response characteristics of these two sensors to determine the specific gas concentration from various combinations of two-gas mixtures. The sensor relative conductivity can be defined as a function which describes the sensor response by various gas concentrations. Accordingly, the gas sensor relative conductivity for two gases can be expressed as: S(c) = a0 − b0 ln a0 − (a1 − b1 ln(ch + 0.5)) × cm + exp − 0.5 , b0 (3) where cm and ch are the gas concentrations of methane and hydrogen, respectively, and ak and bk are the coefficients for
Fig. 2. Measured data fitted with Eq. (3) for the gas mixture of methane and hydrogen. S(c) the sensor relative conductivity, cm , ch are the gas concentrations of CH4 and H2 , respectively.
A. Chaiyboun et al. / Sensors and Actuators B 123 (2007) 1064–1070
1067
Fig. 4. Measured values and approximated fitting with Eq. (4) for gas mixture of ammonia and propane and of ammonia and ethanol. ca the gas concentration of ammonia.
These results reveal that maintaining a unique-characteristic curve is difficult. Therefore, we attempted to extend Eq. (3) to be applicable to the concentrations of two different gases and to obtain optimized estimations. Using Eq. (3) the sensor signals can be described in the presence of a gas mixture, whereby with the help of additional product terms a possible interdependency between the individual gas components is considered. Eq. (3) verifies that the sensor relative conductivity is linearly proportional to the logarithm of the gas concentration. This holds for the low and high gas concentration of the main gas sensitivity but only for high concentrations of cross gas sensitivities (from 10 to 1000 ppm). The reason for that could be high sensor sensitivity to the cross gases at low gas concentrations, which affects the linearity of sensor. However, in the modelling subject region, results can generally be estimated to have a small error. In other words, the mixture gas sensor characteristics can be approximated with small errors. Therefore, product terms are added to Eq. (3) for the correction of estimations. For the gas mixture of methane and hydrogen, the mixture gas characteristic equation is then as follows: S(cm )
a0 − (a1 − b1 ln(ch + 0.5)) = a0 − b0 ln cm + exp b0 β [a0 −(a1 −b1 ln(ch +0.5))] − 0.5+α exp − 0.5 , b0 (4)
Fig. 3. Measured data fitted with Eq. (4) for gas mixtures of (a) methane and hydrogen, (b) methane and propane and (c) methane and carbon monoxide. S(c) the sensor relative conductivity, cm , ch , cp , cc are the gas concentrations of CH4 , H2 , C3 H8 and CO, respectively.
where α and β are the coefficients and the term α(exp(β(·)) describes the correction of Eq. (3). We assume that the sensor does not change its main sensitivity to the gas methane and the coefficients a0 and b0 remain constant. Using Eq. (4) the coefficients α and β can be determined by the least-squares method at constant methane gas concentra-
1068
A. Chaiyboun et al. / Sensors and Actuators B 123 (2007) 1064–1070
Fig. 5. Measured values and approximated non-linear fitting for the determining of the coefficients α and β using sensor 1. S(c) the sensor relative conductivity, cm the gas concentration of CH4 . Table 2 Coefficients of approximate equation (Eq. (4)) for gas mixtures for sensor 1
α β
CH4 + H2
CH4 + C3 H8
CH4 + CO
−5483 −0.136
−646.89 −0.283
34.73 0.739
Table 3 Coefficients of approximate equation (Eq. (3)) for each gas for the sensor 1 (single gas measurements) CH4 a0 b0
H2 1.047 0.104
a1 b1
C3 H8 0.584 0.067
CO
a2 b2
0.872 0.104
a3 b3
1.156 0.093
tions cm , in our case at 100 ppm, and the plots fitted well to the measured data as shown in Fig. 5. The solid line connecting the experimental data in Figs. 3 and 4 was obtained by approximation from Eq. (4) by least-squares fit. All measured values are closely consistent with the approximated curve in the range of concentrations from 1 to 1000 ppm. In the same way, the coefficients α and β can be determined for gas mixtures of methane and propane and of methane and carbon monoxide for sensor 1. Values for the coefficients of each function, as calculated by the least-squares method, are shown in Table 2, and the coefficients for each gas for sensor 1 are shown in Table 3. Table 4 shows the Table 4 Coefficients of approximate equations (Eqs. (3) and (4)) for sensor 2 (single gas and gas mixture measurements) NH3 a0 b0
0.775 0.095
C2 H5 OH
C3 H8
a1 b1
a2 b2
0.783 0.085
ca + cp 1.018 0.101
α1 β1
ca + ce 34.16 0.169
α2 β2
38.8 0.25
Fig. 6. Correlation between measured and estimated data for gas mixtures of (a) CH4 and H2 , (b) CH4 and C3 H8 and (c) CH4 and CO for the gas sensor 1. S(c)me the measured data, S(c)es the estimated data.
A. Chaiyboun et al. / Sensors and Actuators B 123 (2007) 1064–1070
1069
coefficients for single gas and gas mixture measurements for sensor 2. 4. Discussion We investigated the error of these estimation curves for different gas mixture combinations. The distribution of normalized errors between the measured concentration and the estimated concentration can be defined by: S(c)me − S(c)es S(c)error = × 100, (5) S(c)me where S(c)me denotes the measured data of the sensor relative conductivity and S(c)es denotes the estimated data. Fig. 6 shows the relationship between measured and estimated values for the gas mixtures of methane and hydrogen, methane and propane, and methane and carbon monoxide (sensor 1). Values are shown by the (䊉) signs and the approximation is shown by the solid line. The regression has a slope very close to 1, and for the best consistence the points must fall on this line. The calculated correlation coefficients are 0.9981, 0.9973 and 0.9965 for gas mixtures of methane and hydrogen, methane and propane, and methane and carbon monoxide, respectively, indicating a very high correlation. The maximum error for sensor 1 was 5.23%. Fig. 7 shows the relationship between the estimated values and the measured values using sensor 2 for gas mixtures of ammoniac and ethanol and of ammoniac and propane. In this case, the values of the calculated correlation coefficients are 0.99981 and 0.99940, also indicating a high correlation. The maximum error is 4.13% for sensor 2. Furthermore, we investigated the models of Clifford and Tuma [4] and Hirobayashi et al. [9] for two-gas mixtures. Fig. 8 shows as an example the fit results of the measured data of the
Fig. 7. Relationship between measured and estimated values for gas mixtures of NH3 and C2 H5 OH and of NH3 and C3 H8 for sensor 2, S(c)me the measured data, S(c)es the estimated data.
Fig. 8. Comparison of the fit with the models of Clifford, Hirobayashi and our model (Eq. (4)) for the gas mixture of CH4 (cm = 10 ppm) and H2 for sensor 1. S(c) the sensor relative conductivity, cm , ch the gas concentration of CH4 and H2 , respectively.
gas mixture of methane and hydrogen with these models. One recognises that for the models of Clifford and Hirobayashi the difference between measured values and fitted curves is very large. These differences are substantially smaller with the use of our model and thus the fit is clearly better (see Fig. 8).
Fig. 9. Average error of the fit by the use of the models of Clifford and Tuma [4], Hirobayashi et al. [9] and the logarithmic model for measured data of sensors 1 and 2.
1070
A. Chaiyboun et al. / Sensors and Actuators B 123 (2007) 1064–1070
The gas sensors 1 and 2 were simulated also with the models by Clifford and Tuma [4] and Hirobayashi et al. [9]. Fig. 9 shows the average error of the approximation with the models of Clifford, Hirobayashi and our model for sensors 1 and 2. The results show that the models of Clifford and Hirobayashi delivered relatively good approximations for the gas concentrations from 100 to 1000 ppm. However, at low gas concentration from 0.00001 to 10 ppm the error tends to increase up to 16%, whereas our logarithmic model delivered good accuracy for the complete gas concentration range from 0.00001 to 1000 ppm with an average error of less than of 5%. 5. Conclusions Semiconductor gas sensors are used to monitor gaseous indoor air pollution for estimation and verification of the gas concentrations of two-gas mixtures. A sensor response equation has been developed using a simple logarithmic model with two coefficients. These coefficients change with the type of sensor material, the type of reducing gas and the temperature of the sensor. The characteristics of two mixture gases (methane and hydrogen, methane and propane, methane and carbon monoxide, ammoniac and ethanol, and ammoniac and propane) were investigated for different combinations of gases in a mixture. Using the sensor response curves, we accurately evaluated the concentrations of mixture gases, given the characteristic for each of the component gases. Sensor relative conductivity was shown to be proportional to the logarithm of the concentration of each mixture gas, and was found to become non-linear when sensor cross sensitivities were low. Therefore, for gas sensors used, the main sensitivity was corrected with product terms of the cross sensitivity. Using the sensor main sensitivity-corrected equation, we found a strong correlation between estimated data and measured data. Results show that the estimated curves are almost identical to the measured data, with an average relative error of less than 5%. The choice and form of the model can affect the approximation results, so it is to be observed, e.g. that very low gas concentrations were not well estimated by the use of other models compared with our model. The proposed model has been shown to approximate the sensor response characteristics of mixture gases using a simple function with high accuracy, and can be applied to determine the concentration of mixtures of several gases. Acknowledgements This work was supported by the German Federal Ministry for education and research BMBF, grant 16 SV 1538 “IESSICA” and Umweltsensortechnik Geschwenda Ltd., which are gratefully acknowledged. References [1] P.T. Moseley, J.O.W. Norris, D.E. Williams, Techniques and Mechanisms in Gas Sensing, Adam Hilger, Bristol, UK, 1991. [2] P.T. Moseley, A.J. Crocker, Sensor Materials, Inst. of Physics Publ., Bristol, UK, 1996.
[3] S. Hirobayashi, H. Kimura, T. Oyabu, Dynamic model to estimate the dependence of gas sensor characteristics on temperature and humidity in environment, Sens. Actuators B 60 (1999) 78–82. [4] P.K. Clifford, D.T. Tuma, Characteristics of semiconductor gas sensors. I. Steady state gas response, Sens. Actuators B 3 (1982/1983) 233–254. [5] P.K. Clifford, D.T. Tuma, Characteristics of semiconductor gas sensors. II. Transient response to temperature change, Sens. Actuators B 3 (1983) 255–281. [6] Figaro Engineering Inc., Figaro Gas Sensors, 8-Series, Products Catalogue 1, 1998. [7] M.J. Madou, S.R. Morrison, Chemical Sensing with Solid State Devices, Academic Press, Boston, USA, 1989. [8] S. Str¨assler, A. Reis, Simple models for n-type metal oxide gas sensors, Sens. Actuators B 4 (1983) 491–495. [9] S. Hirobayashi, M.A. Kadir, T. Yoshizawa, T. Yamabuchi, Verification of a logarithmic model for estimation of gas concentrations in a mixture for a tin oxide gas sensor response, Sens. Actuators B 92 (2003) 269–278. [10] M.A. Mart´ın, J.P. Santos, J.A. Agapito, Application of artificial neural networks to calculate the partial gas concentrations in a mixture, Sens. Actuators B 77 (2001) 468–471. [11] H. Sundgren, F. Winquist, I. Lukkari, I. Lundstrom, Artificial neural networks and gas sensor arrays: quantification of individual components in a gas mixture, Meas. Sci. Technol. 2 (1991) 464–469. [12] H.E. Endres, W. G¨ottler, H.D. Jander, S.M. Drost, H. Sandmaier, G. Sberveglieri, G. Faglia, C. Perego, Improvement in signal evaluation methods for semiconductor gas sensors, Sens. Actuators B 27 (1995) 267–270. [13] M. Pardo, G. Faglia, G. Sberveglieri, M. Corte, F. Masulli, M. Riani, A time delay neural network for estimation of gas concentrations in a mixture, Sens. Actuators B 65 (2000) 267–269. [14] D.S. Lee, J.K. Jung, J.W. Lim, J.S. Huh, D.D. Lee, Recognition of volatile organic compounds using SnO2 sensor array and pattern recognition analysis, Sens. Actuators B 77 (2001) 228–236. [15] G. Faglia, F. Bicelli, G. Sberveglieri, P. Maffezzoni, P. Gubian, Identification and quantification of methane and ethyl alcohol in an environment at variable humidity by an hybrid array, Sens. Actuators B 44 (1997) 517–520. [16] Th. Becker, S. Ahlers, Chr. Bosch-v.Braunm¨uhl, G. M¨uller, O. Kiesewetter, Gas sensing properties of thin- and thick-film tin-oxide materials, Sens. Actuators B 77 (2001) 55–61. [17] R. Traute, F. Haase, O. Kiesewetter, G. M¨uller, T. Doll, Expert system for gas sensor array configuration and scenario management, in: Eurosensors XVII—The 17th European Conference on Solid-State Transducers, Guimar˜aes, Portugal, 2003.
Biographies A. Chaiyboun studied in Syria and Germany. He received his diploma in electrical engineering from the Technical University Ilmenau in 2001. Since 2003 he has been studying as a PhD student at the institute of solid state electronics at the Technical University Ilmenau. His field of interest focuses in chemical sensors, simulation and modelling of micromachined sensors. T. Haas received his diploma in electrical engineering in 2001 from Ilmenau Technical University in Germany. He is now working towards his PhD thesis at the Institute of Solid State Electronics in the field of computational design of photonic crystal devices ranging from novel sensors, PCF and fibre-chip interfaces. His research interests also include Linux based clusters. T. Doll received his diploma in physics from Munich University and his PhD and habilitation in micro systems/electrical engineering from the Bundeswehr University in Munich in 1995 and 1999, respectively. He was a visiting scientist at Caltech, 1998–2000 and afterwards a professor of solid state electronics at the Technical University of Ilmenau, there director of the Centre for Micro- and Nanotechnologies and a professor of micro structure physics at Mainz University since 2004 and director of the adlantis research centre in Dortmund since 2006. His research covers chemo- and solid state sensors, silicon technology, simulation and modelling of micromachined sensors, the development of chemical and high-temperature sensors, polymer transistors, photonic band gap and MEMS systems.
A logarithmic multi-parameter model using gas sensor main and cross sensitivities to estimate gas concentrations in a gas mixture for SnO2 gas sensors A. Chaiyboun a,b,∗,1 , R. Traute c , T. Haas b , O. Kiesewetter c , T. Doll a,b a
b
Institute of Physics, University of Mainz, Staudinger Weg 7, 55128 Mainz, Germany Department of Solid State Electronics, Technical University of Ilmenau, P.O. Box 100565, 98684 Ilmenau, Germany c UST Umweltsensortechnik GmbH, Dieselstraße 2, 98716 Geschwenda, Germany Received 3 August 2006; received in revised form 9 November 2006; accepted 13 November 2006 Available online 8 December 2006
Abstract In a metal-oxide semiconductor gas sensor, the sensitivity of the metal-oxide resistance to concentrations of reducing gases in the surrounding atmosphere is known to be related to adsorption and desorption of gas on the redox reactions between the gas and oxygen. Changes in the electric conductance due to these reactions were measured for tin dioxide semiconductor gas sensors. In this study, we propose a model of gas sensor responding behaviour using a relationship between sensor conductance and gas concentrations in a mixture. A least-squares method fit of measured data was applied to determining the values of coefficients. The proposed method uses main and cross sensitivities the describing the response of a gas sensor. Applying two-gas sensors which show different characteristics, the gas concentrations in a gas mixture can be evaluated. The proposed method has been applied to the estimation of gas concentrations in a mixture of hydrogen–methane, carbon monoxide–methane, propane–methane, ethanol–ammonia and propane–ammonia. The concentrations determined from the response curves were accurate within a 5% error. The results indicate that the proposed model is feasible for recognition of calculated estimations in a gas mixture. This paper shows a significant result through utilization of the proposed model of gas sensor response. © 2006 Elsevier B.V. All rights reserved. Keywords: Main sensitivity; Cross sensitivity; Gas concentration; SnO2 gas sensor; Gas mixture; Modelling
1. Introduction Metal-oxide semiconductor gas sensors are suitable for the detection of oxidizing and reducing gases, since they react to their presence with a measurable change of their electrical conductivity [1–3]. Among the used semiconductor layers heated SnO2 -layers are best tested and furthest common [2]. However, a single semiconductor gas sensor is only conditionally suitable for the selective proof of certain gasses, since it exhibits cross sensitivities opposite to practically all oxidizing and reducing gases. These sensors do not have usually high gas selectivity, and its recognition characteristics are largely depending on changes in the environment, such as temperature and humidity
∗ 1
Corresponding author. Tel.: +49 6131 3922406; fax: +49 6131 3922276. E-mail address: [email protected] (A. Chaiyboun). Since June 2006 at the University of Mainz.
0925-4005/$ – see front matter © 2006 Elsevier B.V. All rights reserved. doi:10.1016/j.snb.2006.11.012
[3]. Furthermore, metal-oxide gas sensors are non-linear systems which thereby cause considerable measurement errors. To the correction of these measuring errors, heuristic–mathematical or physical–chemical models can be used. The coefficients of these models must be determined with often significant effort by calibration and adaptation of the models to the calibrating data. For the determination of each parameter, value at least a measured value under exactly defined conditions must be acquired. The reaction of gas sensors to the measured gasses can be described by different models, which mostly indicate a power-law for the interrelation between partial pressure and sensor conductance. For the metal-oxide gas sensor, Clifford and Tuma [4,5] attempted to derive from experimental results an empirical formula which described the sensor resistance as a function of several gas concentrations. The model of Clifford was developed for application to sensors of the type Taguchi (TGS) [6]. Madou and Morrison [7] employed extensive theoretical considerations concerning the influence of oxygen and reducing gases on the sensor
A. Chaiyboun et al. / Sensors and Actuators B 123 (2007) 1064–1070
conductivity. At operating temperature, the resistance of semiconductor gas sensors follows a power-law dependence on the gas concentration in the environment air [4,7,8]. Hirobayashi et al. [9] proposed a logarithmic model for detecting the individual components of gas mixtures, demonstrating good approximation results for measured gas concentrations from 100 to 1000 ppm. Latterly, artificial neural network [10–13] and other pattern recognition methods [14] are employed with some success. Faglia et al. [15] used a neural network, to which information about the current humidity is supplied by separate measurement. Using an array of semiconductor gas sensors, they attempted to eliminate the humidity influence on the measurements. However, these methods require many gas concentration measurements for each object gas, which requires significant time and effort. Usually, gas sensor responses are obtained experimentally for each pure gas forming a mixture gas, and mixture gas concentrations are determined accurately using the sensor response equation. The object of the present study is to test and expand the formulation and verification of a logarithmic model in detail and describe an accurate estimation procedure for five kinds of gas mixtures using our proposed model. The model is particularly suitable for describing the low concentration range from 0.00001 to 10 ppm. This is important for application scenarios with low gas concentrations (e.g. ozone or carbon monoxide). Our logarithmic model delivers also good accuracy for the complete gas concentration range from 0 to 1000 ppm. The results show that accurate estimation is feasible using this model through detailed exploration of sensor output resistance characteristics. 2. A model of sensor characteristic curve A SnO2 gas sensor is reliable for detecting combustible gases. The sensors GGS1470 and GGS4470, herein after referred to as sensor 1 and sensor 2, respectively, are reliable and stable, but have no gas selectivity and show cross sensitivities to several gases; i.e. they sense different gases at random. Both sensors GGS XXX0 are commercial SnO2 -thick film gas sensors of the company UST. They consist of an Al2 O3 -carrier substrate with a structured platinum layer on the front and back side for the contacts and the heater. As an electrode structure a standard structure (SS) with two wide electrodes or an interdigital structure (IDS) is used. The sensitive layer (SnO2 ) is deposited via the screen printing technique on the Pt-contacts. The used sensors differ in their preparation of the sensitive layer, which leads to different main and cross sensitivities to the gases to be detected; for example, the sensor GGG1470 has main sensitivity to CH4 and the sensor GGS4470 to NH3 . The gas sensors used in the present study were sensor 1 and sensor 2. In many applications, these two types of sensors are widely used due to their advantages such as small size, light weight and high sensitivity. The sensor 1 is cross sensitive to hydrogen (H2 ), propane (C3 H8 ) and carbon monoxide (CO); this sensor also has a main sensitivity to methane (CH4 ) and it is especially suitable for the leakage detection of combustible gases. The sensor 2 is mainly sensitive to ammonia (NH3 ) and cross sensitive to
1065
Table 1 Lower explosion limit (LEL) for the used gases
LEL (vol%)
C2 H5 OH
C3 H8
CH4
CO
H2
NH3
3.3
2
5
10
17
4
ethanol (C2 H5 OH), propane (C3 H8 ) and other hazardous gases, and is often used in gas leakage alarms. The following combinations of testing mixture gases were chosen: (1) CH4 + H2 , (2) CH4 + C3 H8 , (3) CH4 + CO, (4) NH3 + C3 H8 , and (5) NH3 + C2 H5 OH. In Table 1, the lower explosion limit (LEL) is shown for each gas we used. By different limit ranges of each gas, one could have obtained better results, but the analysis and modelling of the small concentration range in the close proximity to 0 ppm were the centre of attention. Typically, semiconductor gas sensors react to different gases. Therefore, as a sensor signal the relative conductivity S(c) is used [16,17]. It corresponds to the resistance value of the sensor under gas influence related to the sensor resistance in pure atmosphere. S(c) =
Rg (c) G0 = , R0 Gg (c)
(1)
where G0 is the baseline conductance (i.e. in the presence of clean air) and Gg is the steady-state conductance of the sensor in the presence of a given gas or gas mixture. As previously mentioned, the function (1) applies only at constant temperature T of the sensor surface. It is well known that the electrical resistance Rg () of a gas sensor decreases as the gas concentration c (ppm) increases. Therefore, the sensor relative conductivity can be expressed as: S(c) = a − b ln(c + 0.5),
(2)
where a and b are the coefficients and the constant 0.5 is used for the defining of the sensor response at the gas concentration of 0 ppm (i.e. in the presence of clean air). The Eq. (2) shows that the sensor resistance or relative conductivity is proportional to the logarithm of the concentration c, and is satisfied when the temperature and relative humidity of the environment are fixed. The condition for maximum value of sensor resistance is the sensor response at 0 ppm measurement in clean air and that of minimum value of sensor resistance is the lower limit of sensor response at high gas concentrations. Gas sensor relative conductivity is shown in Fig. 1(a and b) as a function of the gas concentration for each object gas. The gas sensors 1 and 2 were measured to individual gas. For sensor 1 each object gas was varied from 1 to 1000 ppm (measured four times at 1, 10, 100, 1000 ppm) and for sensor 2 each gas was varied from 1 to 1000 ppm (measured seven times at 1, 3, 10, 30, 100, 300 and 1000 ppm). We took three measurements for each gas concentration, the mean value of which was used in the analysis. The solid line approximations connecting the experimental data were obtained from Eq. (2), using a leastsquares method.
1066
A. Chaiyboun et al. / Sensors and Actuators B 123 (2007) 1064–1070
gas sensor responses determined from fits for individual gas component in Fig. 1(a). Eq. (3) is a function for the gas sensor response, in which the cross sensitivity of the sensor is integrated (in this case the cross sensitivity to hydrogen); and we can regard this function as an approximation of the first order. The concentrations of cross gases are calculated subject to the main sensitivity of sensor 1; in our case subject to the main CH4 -sensitivity. The measured data of the gas mixture in Fig. 2 were fitted by Eq. (3) as a first approximation and the results show that the estimated curves are satisfactory, but with a large average error of 14% at low concentration from 1 to 10 ppm, and at high concentrations the error tends to decrease. The maximum error is 13.86% and 7.92% for sensor 1 and sensor 2, respectively. Fig. 3(a–c) shows the sensor response of sensor 1 to mixture gasses of methane and hydrogen, methane and propane, and methane and carbon monoxide, respectively. Methane (the main gas) was varied from 10 to 1000 ppm (measured three times at 10, 100, 1000 ppm) and hydrogen was varied from 1 to 1000 ppm (measured four times at 1, 10, 100 and 1000 ppm). Following a similar procedure to that for the methane and hydrogen, gas mixtures of methane and propane and of methane and carbon monoxide were measured at the same concentrations. At lower cross gas concentration, there is a greater sensitivity to methane concentration, as shown in Fig. 3(a–c), and the sensor relative conductivity decreased with increasing methane concentration as expected. For sensor 2, output characteristics of compounds of ammonia and propane and of ammonia and ethanol are shown in Fig. 4. The gas concentrations of cross gases of propane and ethanol were varied from 0.00001 to 1000 ppm and were measured five times (at 0.00001, 1, 10, 100 and 1000 ppm) at a constant ammonia concentration of 100 ppm.
Fig. 1. Measured values and approximated algorithmic fitting for each object gas using (a) sensor 1 and (b) sensor 2. S(c) the sensor relative conductivity.
3. Characteristic equation of two-gas system In this study, we investigated the sensor response characteristics of these two sensors to determine the specific gas concentration from various combinations of two-gas mixtures. The sensor relative conductivity can be defined as a function which describes the sensor response by various gas concentrations. Accordingly, the gas sensor relative conductivity for two gases can be expressed as: S(c) = a0 − b0 ln a0 − (a1 − b1 ln(ch + 0.5)) × cm + exp − 0.5 , b0 (3) where cm and ch are the gas concentrations of methane and hydrogen, respectively, and ak and bk are the coefficients for
Fig. 2. Measured data fitted with Eq. (3) for the gas mixture of methane and hydrogen. S(c) the sensor relative conductivity, cm , ch are the gas concentrations of CH4 and H2 , respectively.
A. Chaiyboun et al. / Sensors and Actuators B 123 (2007) 1064–1070
1067
Fig. 4. Measured values and approximated fitting with Eq. (4) for gas mixture of ammonia and propane and of ammonia and ethanol. ca the gas concentration of ammonia.
These results reveal that maintaining a unique-characteristic curve is difficult. Therefore, we attempted to extend Eq. (3) to be applicable to the concentrations of two different gases and to obtain optimized estimations. Using Eq. (3) the sensor signals can be described in the presence of a gas mixture, whereby with the help of additional product terms a possible interdependency between the individual gas components is considered. Eq. (3) verifies that the sensor relative conductivity is linearly proportional to the logarithm of the gas concentration. This holds for the low and high gas concentration of the main gas sensitivity but only for high concentrations of cross gas sensitivities (from 10 to 1000 ppm). The reason for that could be high sensor sensitivity to the cross gases at low gas concentrations, which affects the linearity of sensor. However, in the modelling subject region, results can generally be estimated to have a small error. In other words, the mixture gas sensor characteristics can be approximated with small errors. Therefore, product terms are added to Eq. (3) for the correction of estimations. For the gas mixture of methane and hydrogen, the mixture gas characteristic equation is then as follows: S(cm )
a0 − (a1 − b1 ln(ch + 0.5)) = a0 − b0 ln cm + exp b0 β [a0 −(a1 −b1 ln(ch +0.5))] − 0.5+α exp − 0.5 , b0 (4)
Fig. 3. Measured data fitted with Eq. (4) for gas mixtures of (a) methane and hydrogen, (b) methane and propane and (c) methane and carbon monoxide. S(c) the sensor relative conductivity, cm , ch , cp , cc are the gas concentrations of CH4 , H2 , C3 H8 and CO, respectively.
where α and β are the coefficients and the term α(exp(β(·)) describes the correction of Eq. (3). We assume that the sensor does not change its main sensitivity to the gas methane and the coefficients a0 and b0 remain constant. Using Eq. (4) the coefficients α and β can be determined by the least-squares method at constant methane gas concentra-
1068
A. Chaiyboun et al. / Sensors and Actuators B 123 (2007) 1064–1070
Fig. 5. Measured values and approximated non-linear fitting for the determining of the coefficients α and β using sensor 1. S(c) the sensor relative conductivity, cm the gas concentration of CH4 . Table 2 Coefficients of approximate equation (Eq. (4)) for gas mixtures for sensor 1
α β
CH4 + H2
CH4 + C3 H8
CH4 + CO
−5483 −0.136
−646.89 −0.283
34.73 0.739
Table 3 Coefficients of approximate equation (Eq. (3)) for each gas for the sensor 1 (single gas measurements) CH4 a0 b0
H2 1.047 0.104
a1 b1
C3 H8 0.584 0.067
CO
a2 b2
0.872 0.104
a3 b3
1.156 0.093
tions cm , in our case at 100 ppm, and the plots fitted well to the measured data as shown in Fig. 5. The solid line connecting the experimental data in Figs. 3 and 4 was obtained by approximation from Eq. (4) by least-squares fit. All measured values are closely consistent with the approximated curve in the range of concentrations from 1 to 1000 ppm. In the same way, the coefficients α and β can be determined for gas mixtures of methane and propane and of methane and carbon monoxide for sensor 1. Values for the coefficients of each function, as calculated by the least-squares method, are shown in Table 2, and the coefficients for each gas for sensor 1 are shown in Table 3. Table 4 shows the Table 4 Coefficients of approximate equations (Eqs. (3) and (4)) for sensor 2 (single gas and gas mixture measurements) NH3 a0 b0
0.775 0.095
C2 H5 OH
C3 H8
a1 b1
a2 b2
0.783 0.085
ca + cp 1.018 0.101
α1 β1
ca + ce 34.16 0.169
α2 β2
38.8 0.25
Fig. 6. Correlation between measured and estimated data for gas mixtures of (a) CH4 and H2 , (b) CH4 and C3 H8 and (c) CH4 and CO for the gas sensor 1. S(c)me the measured data, S(c)es the estimated data.
A. Chaiyboun et al. / Sensors and Actuators B 123 (2007) 1064–1070
1069
coefficients for single gas and gas mixture measurements for sensor 2. 4. Discussion We investigated the error of these estimation curves for different gas mixture combinations. The distribution of normalized errors between the measured concentration and the estimated concentration can be defined by: S(c)me − S(c)es S(c)error = × 100, (5) S(c)me where S(c)me denotes the measured data of the sensor relative conductivity and S(c)es denotes the estimated data. Fig. 6 shows the relationship between measured and estimated values for the gas mixtures of methane and hydrogen, methane and propane, and methane and carbon monoxide (sensor 1). Values are shown by the (䊉) signs and the approximation is shown by the solid line. The regression has a slope very close to 1, and for the best consistence the points must fall on this line. The calculated correlation coefficients are 0.9981, 0.9973 and 0.9965 for gas mixtures of methane and hydrogen, methane and propane, and methane and carbon monoxide, respectively, indicating a very high correlation. The maximum error for sensor 1 was 5.23%. Fig. 7 shows the relationship between the estimated values and the measured values using sensor 2 for gas mixtures of ammoniac and ethanol and of ammoniac and propane. In this case, the values of the calculated correlation coefficients are 0.99981 and 0.99940, also indicating a high correlation. The maximum error is 4.13% for sensor 2. Furthermore, we investigated the models of Clifford and Tuma [4] and Hirobayashi et al. [9] for two-gas mixtures. Fig. 8 shows as an example the fit results of the measured data of the
Fig. 7. Relationship between measured and estimated values for gas mixtures of NH3 and C2 H5 OH and of NH3 and C3 H8 for sensor 2, S(c)me the measured data, S(c)es the estimated data.
Fig. 8. Comparison of the fit with the models of Clifford, Hirobayashi and our model (Eq. (4)) for the gas mixture of CH4 (cm = 10 ppm) and H2 for sensor 1. S(c) the sensor relative conductivity, cm , ch the gas concentration of CH4 and H2 , respectively.
gas mixture of methane and hydrogen with these models. One recognises that for the models of Clifford and Hirobayashi the difference between measured values and fitted curves is very large. These differences are substantially smaller with the use of our model and thus the fit is clearly better (see Fig. 8).
Fig. 9. Average error of the fit by the use of the models of Clifford and Tuma [4], Hirobayashi et al. [9] and the logarithmic model for measured data of sensors 1 and 2.
1070
A. Chaiyboun et al. / Sensors and Actuators B 123 (2007) 1064–1070
The gas sensors 1 and 2 were simulated also with the models by Clifford and Tuma [4] and Hirobayashi et al. [9]. Fig. 9 shows the average error of the approximation with the models of Clifford, Hirobayashi and our model for sensors 1 and 2. The results show that the models of Clifford and Hirobayashi delivered relatively good approximations for the gas concentrations from 100 to 1000 ppm. However, at low gas concentration from 0.00001 to 10 ppm the error tends to increase up to 16%, whereas our logarithmic model delivered good accuracy for the complete gas concentration range from 0.00001 to 1000 ppm with an average error of less than of 5%. 5. Conclusions Semiconductor gas sensors are used to monitor gaseous indoor air pollution for estimation and verification of the gas concentrations of two-gas mixtures. A sensor response equation has been developed using a simple logarithmic model with two coefficients. These coefficients change with the type of sensor material, the type of reducing gas and the temperature of the sensor. The characteristics of two mixture gases (methane and hydrogen, methane and propane, methane and carbon monoxide, ammoniac and ethanol, and ammoniac and propane) were investigated for different combinations of gases in a mixture. Using the sensor response curves, we accurately evaluated the concentrations of mixture gases, given the characteristic for each of the component gases. Sensor relative conductivity was shown to be proportional to the logarithm of the concentration of each mixture gas, and was found to become non-linear when sensor cross sensitivities were low. Therefore, for gas sensors used, the main sensitivity was corrected with product terms of the cross sensitivity. Using the sensor main sensitivity-corrected equation, we found a strong correlation between estimated data and measured data. Results show that the estimated curves are almost identical to the measured data, with an average relative error of less than 5%. The choice and form of the model can affect the approximation results, so it is to be observed, e.g. that very low gas concentrations were not well estimated by the use of other models compared with our model. The proposed model has been shown to approximate the sensor response characteristics of mixture gases using a simple function with high accuracy, and can be applied to determine the concentration of mixtures of several gases. Acknowledgements This work was supported by the German Federal Ministry for education and research BMBF, grant 16 SV 1538 “IESSICA” and Umweltsensortechnik Geschwenda Ltd., which are gratefully acknowledged. References [1] P.T. Moseley, J.O.W. Norris, D.E. Williams, Techniques and Mechanisms in Gas Sensing, Adam Hilger, Bristol, UK, 1991. [2] P.T. Moseley, A.J. Crocker, Sensor Materials, Inst. of Physics Publ., Bristol, UK, 1996.
[3] S. Hirobayashi, H. Kimura, T. Oyabu, Dynamic model to estimate the dependence of gas sensor characteristics on temperature and humidity in environment, Sens. Actuators B 60 (1999) 78–82. [4] P.K. Clifford, D.T. Tuma, Characteristics of semiconductor gas sensors. I. Steady state gas response, Sens. Actuators B 3 (1982/1983) 233–254. [5] P.K. Clifford, D.T. Tuma, Characteristics of semiconductor gas sensors. II. Transient response to temperature change, Sens. Actuators B 3 (1983) 255–281. [6] Figaro Engineering Inc., Figaro Gas Sensors, 8-Series, Products Catalogue 1, 1998. [7] M.J. Madou, S.R. Morrison, Chemical Sensing with Solid State Devices, Academic Press, Boston, USA, 1989. [8] S. Str¨assler, A. Reis, Simple models for n-type metal oxide gas sensors, Sens. Actuators B 4 (1983) 491–495. [9] S. Hirobayashi, M.A. Kadir, T. Yoshizawa, T. Yamabuchi, Verification of a logarithmic model for estimation of gas concentrations in a mixture for a tin oxide gas sensor response, Sens. Actuators B 92 (2003) 269–278. [10] M.A. Mart´ın, J.P. Santos, J.A. Agapito, Application of artificial neural networks to calculate the partial gas concentrations in a mixture, Sens. Actuators B 77 (2001) 468–471. [11] H. Sundgren, F. Winquist, I. Lukkari, I. Lundstrom, Artificial neural networks and gas sensor arrays: quantification of individual components in a gas mixture, Meas. Sci. Technol. 2 (1991) 464–469. [12] H.E. Endres, W. G¨ottler, H.D. Jander, S.M. Drost, H. Sandmaier, G. Sberveglieri, G. Faglia, C. Perego, Improvement in signal evaluation methods for semiconductor gas sensors, Sens. Actuators B 27 (1995) 267–270. [13] M. Pardo, G. Faglia, G. Sberveglieri, M. Corte, F. Masulli, M. Riani, A time delay neural network for estimation of gas concentrations in a mixture, Sens. Actuators B 65 (2000) 267–269. [14] D.S. Lee, J.K. Jung, J.W. Lim, J.S. Huh, D.D. Lee, Recognition of volatile organic compounds using SnO2 sensor array and pattern recognition analysis, Sens. Actuators B 77 (2001) 228–236. [15] G. Faglia, F. Bicelli, G. Sberveglieri, P. Maffezzoni, P. Gubian, Identification and quantification of methane and ethyl alcohol in an environment at variable humidity by an hybrid array, Sens. Actuators B 44 (1997) 517–520. [16] Th. Becker, S. Ahlers, Chr. Bosch-v.Braunm¨uhl, G. M¨uller, O. Kiesewetter, Gas sensing properties of thin- and thick-film tin-oxide materials, Sens. Actuators B 77 (2001) 55–61. [17] R. Traute, F. Haase, O. Kiesewetter, G. M¨uller, T. Doll, Expert system for gas sensor array configuration and scenario management, in: Eurosensors XVII—The 17th European Conference on Solid-State Transducers, Guimar˜aes, Portugal, 2003.
Biographies A. Chaiyboun studied in Syria and Germany. He received his diploma in electrical engineering from the Technical University Ilmenau in 2001. Since 2003 he has been studying as a PhD student at the institute of solid state electronics at the Technical University Ilmenau. His field of interest focuses in chemical sensors, simulation and modelling of micromachined sensors. T. Haas received his diploma in electrical engineering in 2001 from Ilmenau Technical University in Germany. He is now working towards his PhD thesis at the Institute of Solid State Electronics in the field of computational design of photonic crystal devices ranging from novel sensors, PCF and fibre-chip interfaces. His research interests also include Linux based clusters. T. Doll received his diploma in physics from Munich University and his PhD and habilitation in micro systems/electrical engineering from the Bundeswehr University in Munich in 1995 and 1999, respectively. He was a visiting scientist at Caltech, 1998–2000 and afterwards a professor of solid state electronics at the Technical University of Ilmenau, there director of the Centre for Micro- and Nanotechnologies and a professor of micro structure physics at Mainz University since 2004 and director of the adlantis research centre in Dortmund since 2006. His research covers chemo- and solid state sensors, silicon technology, simulation and modelling of micromachined sensors, the development of chemical and high-temperature sensors, polymer transistors, photonic band gap and MEMS systems.