Sensors and Actuators B 45 (1997) 123 – 130
Simultaneous quantification of carbon monoxide and methane in humid air using a sensor array and an artificial neural network G. Huyberechts a,*, P. Szeco´wka b, J. Roggen a, B.W. Licznerski b a
b
IMEC, MAP/MS — Chemical sensors, Kapeldreef 75, B-3001 Leu6en, Belgium Technical Uni6ersity of Wroclaw, Wybrezeze Wyspianskiego 27, PL-50 -370 Wroclaw, Poland
Received 11 February 1997; received in revised form 8 September 1997; accepted 12 September 1997
Abstract The simultaneous quantification of carbon monoxide and methane in humid air is presented. The response of a three-sensor array, including an undoped and a platinum doped tin dioxide sensor showing non-ideal selectivity and a humidity sensor is fed into an optimised feed forward back propagation artificial neural network in order to obtain the carbon monoxide and methane concentration as network output. The gaseous environments under study were ternary mixtures in the concentration ranges of 0–0.5% methane, 0–1000 ppm carbon monoxide and 0 – 60% relative humidity at 20°C. The network structure, network output with respect to a priori known test concentrations and the influence of the size of the training data set is discussed. © 1997 Elsevier Science S.A. Keywords: Gas sensors; Sensor array; Artificial neural network; Carbon monoxide sensor; Methane sensor; Tin dioxide
1. Introduction Tin dioxide based gas sensors are commercially available for about three decades now [1] but still allow for further research [2– 5] towards better sensitivity, selectivity and stability. A variety of manufacturing techniques have been used to fabricate the sensors, including thin and thick film processing as complementary to the more classical manufacturing techniques. A variety of material modifications have resulted in sensors with modified selectivity, sensitivities, resistance to cross-sensitivities, etc. Various fabrication techniques, e.g. micro machining of miniaturised heater elements, or measuring methods, e.g. temperature programmed signal generation, have been used in order to either minimise the required heater power or in order to create measuring schemes that allow for intelligent extraction of information from the raw sensor signal. Nevertheless, a major problem with tin dioxide based sensors remains their lack of selectivity. It is known that, e.g. variations in humidity levels, result in more or less drastic changes in the conductivity of tin dioxide * Corresponding author. Tel: +32 16 281461; fax: +32 16 281501; e-mail:
[email protected]. 0925-4005/97/$17.00 © 1997 Elsevier Science S.A. All rights reserved. PII S 0 9 2 5 - 4 0 0 5 ( 9 7 ) 0 0 2 8 3 - 9
based sensors, thereby eventually masking the effect of toxic and hazardous gases on the sensor response. Various attempts have been reported to minimise the influence of humidity on the sensor response in order to reduce the number of false or unrecognised alarms from hazardous gas warning devices. One approach to the realisation of sensor systems showing less humidity dependence is of course the development of modified sensor materials and the selection of optimised operation conditions (temperature selection, temperature cycling, filters, etc.)—approaches that are also adopted by our laboratories. However, one can argue that on theoretical grounds the humidity effects can never be completely ruled out, and that by changing the operational conditions in order to decrease the humidity effect non-optimal detection conditions for the target gases can occur. On the other hand sensor array signal processing using artificial neural networks [6,7] (and other numerical techniques) has been used successfully in the identification and characterisation of complex gas mixtures [8–10]. Notwithstanding the large amount of research reports on the use of artificial neural networks in combination with sensor arrays (e.g. in ‘electronic noses’),
124
G. Huyberechts et al. / Sensors and Actuators B 45 (1997) 123–130
Fig. 3. Sensor response of the platinum doped tin dioxide sensor (: 0, : 20, : 40, ": 60% R.H., no carbon monoxide present).
Fig. 1. Screen printed gas sensor (width at base: 10.16 mm, total height: 23.6 mm, width tip: 1.25 mm).
gives, when combined with a trained artificial neural network as developed, a quantitative output for both methane and carbon monoxide irrespective their mutual presence and the presence of varying humidity levels.
relatively small amount of work is reported on the actual quantification of individual compounds in complex mixtures [11–13]. We report on an artificial neural network approach for processing signals arising from a three-sensor array. As target gases both carbon monoxide and methane were selected. Both gases are frequently considered as hazardous gases present in a domestic environment. The combination of two metal oxide based sensors and a humidity sensor allows for the formation of a microsystem for domestic applications that
A sensor array consisting of three sensors has been used for the reported research. Two different types of semiconducting metal oxide sensors have been realised using a screen printing technology and were used in combination with a commercially available humidity sensor (RHT-05, Rotronic).
Fig. 2. Sensor response of the undoped tin dioxide sensor (: 0, : 20, : 40, ": 60% R.H., no methane present).
Fig. 4. Overview of the responses of the tin dioxide based sensors to all mixtures.
2. Experimental
2.1. Sensor array description
G. Huyberechts et al. / Sensors and Actuators B 45 (1997) 123–130 Table 1 Training parameters of the optimised network Itertions
0 – 10 000
Learning ratio Input 0.9000 layer Hidden 0.3000 layer c1 Hidden 0.2500 layer c2 Output 0.1500 layer Momentum Input 0.6000 layer Other 0.8000 layers
10 000
30 000
70 000
–30 000
–70 000
–100 000
0.9000
0.9000
0.9000
0.2400
0.1540
0.0630
0.2000
0.1280
0.0520
0.1200
0.0768
0.0310
0.6000
0.6000
0.6000
0.6400
0.4100
0.1680
125
bon monoxide. It has been shown earlier that from a two-sensor array, comprising of the relevant sensor and a humidity sensor, quantitative information can be obtained about the target gas in binary mixtures [15].
2.2. Experimental set-up
The tin dioxide based sensors were fabricated at IMEC. The sensor lay out is a tip sensor laser machined out of a 96% alumina substrate with a platinum heater element, gold contact electrodes and appropriate dielectric layers for electrical insulation between the conductive layers, as shown in Fig. 1. The heater element is located at the upper edge of the sensor tip underneath the sensor element, allowing for high sensor operating temperatures. The contact pads on the bottom side, for applying power to the heating resistor and for the measurement of the sensor element resistance, are solderable and at 0.1¦ intervals, so that direct sensor insertion in standard connectors is possible. All thick film inks, except for the gas sensitive layer, are commercially available. The gas sensitive layers are screen printed, based on the in-house developed inks IMEC 116 and IMEC 117, respectively, based on pure tin dioxide and platinum doped tin dioxide [14]. During the tests the sensors are heated using a constant heater voltage. The average operating temperature of the pure tin dioxide sensor is 210°C and for the platinum doped sensor 450°C. Previous experiments had shown that these temperatures allowed for a maximum discrimination between methane and carbon monoxide responses. As can be expected the pure tin dioxide sensor shows a pronounced carbon monoxide sensitivity, whilst the conductivity of the platinum doped sensor is strongly influenced by changing methane concentrations. In Fig. 2, the response of the pure tin dioxide sensor versus carbon monoxide concentration is shown at various humidity levels and in absence of methane. The sensor responses as used are output voltages of interface circuits and are directly proportional to the sensor conductivity in the case of the tin dioxide based sensors. Fig. 3 depicts the response of the platinum doped sensor towards methane under different humidity levels, but in absence of car-
All three sensors were placed in a small test chamber (volume ca. 100 ml). The atmosphere over the sensors was changed by installing various combinations of methane and carbon monoxide concentrations at different humidity levels at 20°C ambient temperature. The total volumetric flow was kept constant throughout the experiments. Further details on the experimental set up have been published earlier [14]. The methane concentration was changed from 0 to 5000 ppm in steps of 250 ppm, and carbon monoxide from 0 to 1000 ppm in steps of 200 ppm. It should be noted that the concentration ranges used in this study were required for a specific industrial/domotic application. The requirements for the system were not identical to those required for classical household atmosphere monitoring, which is indicated both by the relatively high concentrations for carbon monoxide and the relatively low methane concentrations used in this study. However the concept as developed here is directly applicable to other domotic applications after considering a modified concentration range for training the artificial neural network. Humidity levels changed from 0 to 60% relative humidity at 20°C in steps of 20%, i.e. in the region where commonly the most pronounced dependence of tin dioxide based sensor responses on humidity levels is observed.
2.3. Artificial neural networks Artificial neural network approaches are classified amongst artificial intelligence systems, because of their ability to ‘learn’ and ‘generalise’ information. In this case, an artificial neural network should translate the physical responses of three sensors towards two output values: the methane concentration and the carbon monoxide concentration. The artificial neural network should use its trained knowledge about the effect of humidity and the presence of carbon monoxide and methane (either alone or in combination) on the response of a methane and on the response of a carbon monoxide sensor to deduce the real methane and carbon monoxide concentrations despite changing humidity levels. The ‘learning and generalisation’-ability, the implicit non-linear response of the sensors to the test environment and the absence of a priori models for the prediction of the sensor response in such moderately complex mixtures motivate the selection of an artificial neural
126
G. Huyberechts et al. / Sensors and Actuators B 45 (1997) 123–130
Fig. 5. Network performance: comparison between predicted and known concentrations at 0% R.H. (a: methane concentration, b: carbon monoxide concentration).
network approach to solve the stated problem. In order to appreciate the complexity of the system, and the power of the artificial neural network approach, the responses of the two tin dioxide sensors to the various test atmospheres are given in Fig. 4. The information contained in this figure, combined with the response of the humidity sensor under the same conditions, is used to extract the quantitative information on both methane and carbon monoxide concentrations. The numerical experiments were performed with NeuralWorks II (NeuralWare) — a specialised software tool providing simulation and development of several kinds of artificial neural networks. The research focused on heteroassociative networks and supervised training with an error backpropagation method. The sensor interface output values acted as input values for the network. The real methane and carbon monoxide concentrations, known a priori, were used as desired output patterns.
As depicted in Fig. 4, the three sensors were exposed to a total number of 462 combinations of methane and carbon monoxide concentrations and humidity levels. In general, one half of the available data set was used for learning and the whole data set for testing. The learning data were additionally perturbed by applying a uniform or, more often, a Gaussian distributed noise, in order to increase the generalisation abilities of the network. The selection of the whole data set for testing, contrary to common practice, is motivated by the fact that further in the study, the influence of the size of the learning set is modified. The selection of the whole data set as test set has the advantage that not only the generalisation aspect of the artificial neural network is tested, but that also the ‘calibration’ data set is included in the evaluation. Moreover, by selecting the total data set for testing, the influence of the selection of the appropriate size of the testing data set, as well as the influence of the actual data are circumvented.
Fig. 6. Network performance: comparison between predicted and known concentrations at 20% R.H. (a: methane concentration, b: carbon monoxide concentration).
G. Huyberechts et al. / Sensors and Actuators B 45 (1997) 123–130
127
Fig. 7. Network performance: comparison between predicted and known concentrations at 40% R.H. (a: methane concentration, b: carbon monoxide concentration).
Fig. 8. Network performance: comparison between predicted and known concentrations at 60% R.H. (a: methane concentration, b: carbon monoxide concentration).
More than 60 different network structures and learning strategies were evaluated, all based on feed forward back propagation paradigms. In order to compare the performance of the various network structures and learning strategies, the following objective function was defined: n
The network structure and learning strategies were optimised to minimise the value of the objective function, representing a minimisation of the absolute error in a least-squares-sum sense. Since all patterns are used in testing, n, the number of patterns, is constant for all computations described below.
2.4. Network structure and learning strategy
o= % (ydes − ynet)2 i=1
where, ynet is the current network output pattern, ydes is the desired output pattern, n is the number of patterns. Table 2 Size of training data set Net Net Net Net
c18 c28 c38 c48
243 124 65 33
The artificial neural network structure that shows the best quantitative results is a quite simple feed forward network with three neurons in the input layer (sensor response), two hidden layers wit twenty two neurons each, and three neurons in the output layer (i.e. concentration of methane, carbon monoxide and the relative humidity). The relatively large number of neurons in the hidden layers, although still acceptable small for practical implementation, e.g. in microcontrollers, probably originates from the sensor response as a func-
128
G. Huyberechts et al. / Sensors and Actuators B 45 (1997) 123–130
weights in a random manner) is applied to the presented network, we may assume that most likely a global minimum has been detected.
3. Results and discussion
Fig. 9. Influence of the size of the training data set on the network output (: net c 18., : net c 28, : net c 38, ": net c 48).
tion of methane concentration, as shown in Fig. 3. The optimal network was developed using simple summation of signals and a sigmoid transfer function in the neurons. Half of the measured data set was used as a learning set. The learning data were evenly distributed over the measurement data space. The values of carbon monoxide concentrations and humidity levels were normalised in the range [0, 1] prior to the numerical processing. The ‘cumulative update of weights’ method was used, i.e. the weights of the neuron connections are updated after every presentation of the whole learning set (epoch size) rather than after the presentation of a single pattern. The learning parameter values varied during the learning process as shown in Table 1. As indicated by the observation that only worse or comparable results were obtained when ‘weight jogging’ (i.e. changing the network
Fig. 10. Influence of the size of the training data set on the objective functions (: carbon monoxide., : methane, : relative humidity).
The performance of the optimised network is shown in Figs. 5–8. It is obvious that the response of the system including the sensors and the developed artificial neural is quite stable with little influence of the humidity level and the presence of the other target gas. The largest discrepancies between the network response and the known concentrations are observed for carbon monoxide with 0% relative humidity where the intrinsic sensitivity of the sensor itself is relatively low, and also for methane concentrations around 0.025%, where the sensor output behaves in a quite irregular way. Typically relative errors in the order of 5% or less are observed. Further accuracy improvement could be envisaged but this will result in loosing generalisation abilities and purely mimicking the input patterns. It has been shown that the combination of a small sensor array combined with an artificial neural network approach is a powerful tool to increase the information obtained from raw sensor signals. It should however be noted that the obtained results are based on a relatively large training data set, This results, even with a computerised test bench and flow control and gas mixing system, in an appreciable amount of calibration work. In the case of small sized sensor arrays the collection of the training data (rather than the computational time for the actual learning) might prove to be the most time consuming (and costly) step in the system as developed. Hence the influence of the size of the training data set was studied. Table 2 summarises the size of the training data set, with net c18 the optimised network as discussed before. The network structure and learning strategies were kept the same as for net c 18. For comparison of the results, as shown in Fig. 9, the influence of the size of the training data set on the carbon monoxide concentration at 0% relative humidity output of the network is shown. This is to be considered as a worst case situation since even with the optimised net and the largest learning data set this combination showed the poorest results. The overall results in terms of the objective function are shown in Fig. 10. Based on this graph and a pre-set uncertainty factor for each of the outputs one can deduce a required minimum size for the training data set (given that the data are evenly distributed over the measurement space as was the case in this study).
G. Huyberechts et al. / Sensors and Actuators B 45 (1997) 123–130
4. Conclusions An approach for the simultaneous quantification of carbon monoxide and methane in humid air, based on a three-sensor array and an artificial neural network has been presented. It is shown that very good quantitative results can be obtained with a relative error of 5% or better in most cases over the concentration range studied. It is shown that the size of the training data set influences the obtainable agreement between the network output and the a priori known concentrations. This however allows also the determination of the size of an acceptable learning data set. for obtaining a pre-set uncertainty in the obtained results during the calibration stage of similar arrays. Because of the fact that the actual collection of the data for the training set will be the most time consuming and costly step in this approach, the importance of the selection of the training data set both in size and in distribution over the concentration space can not be underestimated. Factorial designs will find their applications, if used with caution. Also the influence of unknown gases presented to the sensor array (leading to incorrect conclusions about the presence and concentration of the target gases) forms a topic for further study.
129
[7] J. Zupan, J. Gasteiger, Neural Networks for Chemists, VCH, Weinheim, 1993. [8] T. Nakamoto, K. Fukunishi, T. Moriizumi, Identification capability of odor sensor using quartz resonator array and neural network pattern recognition, Sensors and Actuators B 1 (1990) 473 – 476. [9] J.W. Gardner, F.L. Hines, M.W. Wilkinson, Application of artificial neural networks to an electronic olfactory system, Meas. Sci. Technol. 1 (1990) 446 – 451. [10] S.M. Chang, E. Tamiya, Y. Iwasaki, I. Karube, M. Suzuki, H. Muramatsu, Detection of odorants using an array of piezoelectric crystals and neural network pattern recognition, Anal. Chim. Acta 249 (1991) 323 – 329. [11] H. Sundgren, F. Winquist, I. Lukkar, 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. Sundgren, I. Lundstrom, H. Vollmer, Chemical sensor arrays and abductive networks, Sensors and Actuators B 9 (1992) 127 – 131. [13] V. Sommer, P. Tobias, D. Kohl, Methane and butane concentrations in a mixture with air determined by microcalorimetric sensors and neural networks, Sensors and Actuators B 12 (1993) 147 – 150. [14] P. Van Geloven, Tin oxide gas sensors for the simultaneous determination of gas concentrations in mixtures, PhD thesis, K.U. Leuven, 1992. [15] G. Huyberechts, P. Szeco´wka, J. Roggen, B.W. Licznerski, Artificial neural network approach for humidity influenced methane sensor, in: Proc. Electrical Properties of Metal/Nonmetal Microsystems: Physics, Technology and Applications, Polanica Zdro´j, 11 – 14 September 1995
Acknowledgements This research is performed with the aid of ‘het Ministerie van de Vlaamse Gemeenschap, Administratie Economie (Dossier 208/94-03-31/01-02-03)’ under the project ‘Samenwerkingsprogramma ter ondersteuning van KMO’s in de ontwikkeling van hoog-technologische produkten’.
References [1] K. Ihokura, J. Watson, The Stannic Oxide Gas Sensor: Principles and Applications, CRC Press, Boca Raton, FL, 1994. [2] P. Van Geloven, J. Moons, M. Honore´, J. Roggen, A comparison between sputtered, metallo-organic and screen printed metal oxide gas sensors for methane, Silicates Ind. 3/4 (1990) 81 – 85. [3] G. Faglia, P. Nelli, G. Sberveglieri, Frequency effect on highly sensitive NO2 sensors based on RGTO SnO2 (Al) thin films, Sensors and Actuators B 18/19 (1994) 497–499. [4] M. Honore´, S. Lenaerts, J. Desmet, G. Huyberechts, J. Roggen, Synthesis and characterisation of tin dioxide powders for the realization of thick film gas sensors, Sensors and actuators B 18/19 (1994) 621 – 624. [5] W. Fliegel, G. Behr, J. Wermer, G. Krabbes, Preparation, development of microstructure, electrical and gas sensitive properties of pure and doped SnO2 powders, Sensors and Actuators A 18 (1994) 474 – 477. [6] R. Beale, T. Jackson, Neural Computing: An Introduction, Institute of Physics, Bristol, 1992.
Biographies Guido Huyberechts obtained a PhD in Chemistry in 1988 at the Katholieke Universiteit Leuven. He joined the chemical sensor research group at IMEC in 1987 with activities in the field of gas sensors. Main interests are in the field of gas-solid interactions and the application and characterisation of ceramics, semiconducting metaloxides and organic materials in gas sensors and microsystems for chemical analysis for environmental and medical applications.
P. Szeco´wka graduated in Electronics/Automotion Engineering from the Technical University of Wroclaw in 1993. In 1993 he started PhD studies in the Institute of Cybernetical Engineering and in 1994 he joined the Solid State Electronics Group in the Institute of Electronic Technology. In 1995, he was at stage at IMEC researching artificial neural networks. His interests are in the field of artificial intelligence, signal and data processing, identifcation, microsystems.
J. Roggen obtained a PhD in physics from K.U. Leuven in 1977 and was with Philips company from
G. Huyberechts et al. / Sensors and Actuators B 45 (1997) 123–130
130
1978 to 1984. He moved to IMEC as head of the interconnection and packaging group and of the microsystems group since 1991. B.W. Licznerski PhD, DSc is full professor at the Technical University of Wroclaw and head of the Solid State Electronics Group in the Institute of Electronic
.
Technology. He is Member of the Alexander von Humboldt Foundation, the Committee of Electronics and Telecommunication at the Polish Academy of Science and chairman of the Material Science Commission of the Wroclaw Chapter at the Polish Academy of Science. His research interests are materials science, solid state electronics, microelectronics, sensors and microsystems.