Sensors and Actuators B 71 (2000) 9098

Explosive gas recognition system using thick lm

sensor array and neural network
Dae-Sik Leea, Ho-Yong Junga, Jun-Woo Lima, Minho Leeb, Sang-Woo Banb,
Jeung-Soo Huhc,1, Duk-Dong Leea,*
a
School of Electronic and Electrical Engineering, Kyungpook National University, Taegu 702-701, South Korea
b
Sensor Technology Research Center, Kyungpook National University, Taegu 702-701, South Korea
c
Department of Metallurgical Engineering, Kyungpook National University, Taegu 702-701, South Korea
Received 25 October 1999; accepted 30 June 2000

Abstract

A sensor array with nine discrete sensors integrated on a substrate was developed for recognizing the species and quantity of explosive
gases such as methane, propane, and butane. The sensor array consisted of nine oxide semiconductor gas-sensing materials with SnO2 as the
base material plus a heating element based on a meandered platinum layer all deposited on the sensor. The sensors on the sensor array were
designed to produce a uniform thermal distribution and show a high and broad sensitivity and reproductivity to low concentrations through
the use of nano-sized sensing materials with high surface areas and different additives. Using the sensitivity signals of the array along with
an articial neural network, a gas recognition system was then implemented for the classication and identication of explosive gases. The
characteristics of the multi-dimensional sensor signals obtained from the nine sensors were analyzed using the principal component analysis
(PCA) technique, and a gas pattern recognizer was implemented using a multi-layer neural network with an error back propagation learning
algorithm. The simulation and experimental results demonstrate that the proposed gas recognition system is effective in identifying
explosive gases. For real time processing, a DSP board (TMS320C31) was then used to implement the proposed gas recognition system in
conjunction with a neural network. # 2000 Elsevier Science B.V. All rights reserved.

Keywords: Tin oxide; Sensor array; Explosive gases; Neural network

1. Introduction consisting of various sensing materials such as metal oxides,

Recently, as the usage of LNG and LPG gas has increased,
the frequency of accidental explosions due to leakage has
sharply increased. The ability to detect and precisely mea-
sure gas leakage is crucial in preventing the occurrence of
such accidents. Accordingly, the development of sensors and
systems that can selectively detect and determine the quan-
tity of specic gases below the explosion limit, in situ, is
urgently needed. Until now, there have been various reports
on the use of different kinds of metal oxide sensors for
detecting explosive hydrocarbon gases, which have offered
many improvements from the perspective of stability [1].
Furthermore, to overcome the poor selectivity and long-term
drift of a sensor signal, studies have also been conducted
to develop an electronic nose with sensor array systems
* pattern recognition was then implemented using a DSP
Corresponding author. Tel.: 82-53-939-1074; fax: 82-53-939-1073.
E-mail addresses: mholee@knu.ac.kr (M. Lee), jshuh@knu.ac.kr (J.-S.
Huh), ddlee@ee.kyungpook.ac.kr (D.-D. Lee).
1
Tel.: 82-53-950-5562.
electrochemical sensors, conducting polymers, eld effect
transistor devices, surface acoustic wave devices, and hybrid
sensor arrays [211].
For continuous use, without relearning, an electronic nose
requires high reliability, stability, and sensitivity from the
individual sensors that make up the nose. However, sensor
responses tend to drift signicantly when sensors are used
over a long period of time resulting in poor selectivity and
redundant pattern recognition. Therefore, the fabrication of
more reliable and stable sensors, or the ability to adjust the
pattern recognition routine according to variations in the
sensor response is needed. Accordingly, nano-sized tin oxide
materials were synthesized using the coprecipitation method
and nine kinds of highly sensitive and stable thick lm
sensors were fabricated on a substrate. Using the sensitivity
signals of the arrays and an articial neural network, a gas
(TMS320C31) board with the aim of classifying and quan-
tifying specic explosive gases, including butane, propane,
and methane below their explosion limit value.

0925-4005/00/$ see front matter # 2000 Elsevier Science B.V. All rights reserved.
PII: S 0 9 2 5 - 4 0 0 5 ( 0 0 ) 0 0 6 1 4 - 6
 D.-S. Lee et al. / Sensors and Actuators B 71 (2000) 9098
2. Experimental
2.1. Preparation of base material
The priority for the reliable recognition of leaked explo-
sive gases is the development of stable individual sensors
that exhibit a high sensitivity to gases. Therefore, sensing
materials that can satisfy the above conditions need to be
fabricated and processed. It has been previously reported
that reducing particles to a nano-size through the addition of
Ca along with the spillover effect caused by adding the Pt
catalyst enhances the sensitivity of sensing materials to
gases even at low concentrations [12].
For preparing the raw materials, a solution containing
SnCl4, (CH3CO2)2Ca
xH2O and, H2PtCl6
6H2O was copre-
cipitated by controlling the pH value while dropping a
NH4OH solution to a white precipitated powder. The powder
was then washed with an NH4NO3 solution for 3 days, dried
in an oven at 1008C for 24 h, and calcined in an electric
furnace at 6008C for 1 h. The Ca and Pt contents were both
set at 0.1 wt.% relative to the amount of SnCl4. Using this
powder as a base material, nine kinds of sensing materials
with different additives such as Pd, Au, Pt, CuO, La2O3,
Sc2O3, TiO2, WO3, and ZnO within a range of 0.15 wt.%
were evenly mixed by hand-milling for 1 h to modify the
selectivity spectrum of the sensors. A process chart for the
base material is shown in Fig. 1. The selection of the
additives for the sensor array was based on previously
reported results [1216].
2.2. Fabrication of sensor array and measurement
The sensor array consisting of nine metal-oxide semi-
conductor gas sensors along with their dopants is shown in
Table 1. The coprecipitated SnO2/Ca,Pt (0.1 wt.%) powder
was used as the base material. The sensing lms were all
prepared using the silk printing technique. After being
simultaneously deposited by the silk printing method, the
sensing lms were then annealed at 8008C for 2 h in air. The
following presents a brief outline of the results produced by
Fig. 1. Process chart for the base material.
91
Table 1
Dopants of the sensors utilized in the array
Sensor no. Dopants
1 Pd (3 wt.%)
2 Au (0.1 wt.%)
3 Pt Pd (2 wt.%)
4 CuO (1 wt.%)
5 La2O3 (5 wt.%)
6 Sc2O3 (3 wt.%)
7 TiO2 (1 wt.%)
8 WO3 (10 wt.%)
9 ZnO (1 wt.%)
the sensors on the array listed in Table 1. All the sensing
materials were deposited on an alumina substrate
(55 m  14 m) along with a heating element and tempera-
ture sensor based on a meandered silk-printed platinum layer
on the backside of the substrate. Platinum pads and inter-
digitated platinum contacts were deposited on the front-face
of the substrate.
As shown in Fig. 2, the sensor array was connected with
the long bonding pins for electrical conduction and thermal
isolation from the socket and peripheral circuit. The char-
acteristics of the sensor array were then tested in a testing
chamber after injection of explosive gases of butane, pro-
pane, and methane at 4008C. The sensor signals were
transferred to a test system and monitored using a personal
computer in a testing chamber, as shown in Fig. 3.
The microstructure of the basic sensing-lm was inves-
tigated using transmission electron microscope (TEM),
X-ray diffraction (XRD), and Brunauer-Emmett-Teller
(BET) adsorption, and a voltage detecting method was used
to calculate the sensitivity of the sensor [17]. The sensitivity
was dened as Rair Rgas =Rair  100 (%), where Rgas and
Rair are the electrical resistances in explosive gas and clean
air, respectively.
3. Results and discussions
3.1. Analysis of thick films
To selectively determine the quantity of explosive gases
below their low explosion limit (methane: 5.3 vol.%, pro-
pane: 3.3 vol.%, butane: 1.9 vol.%), stable sensors with a
high sensitivity to low concentrations are required. To
improve the sensitivity of a sensor, the particle size, specic
surface area, and crystalline structure are all recognized as
signicant parameters [1820]. Accordingly, the micro-
structure of various base materials was investigated for its
adaptability.
The XRD patterns of calcined SnO2/Ca,Pt (0.1 wt.%)
powders at different annealing temperatures are shown in
Fig. 4. This gure indicates higher peak patterns with
increased temperatures. All peak patterns observed followed
92 D.-S. Lee et al. / Sensors and Actuators B 71 (2000) 9098

Fig. 2. Structure of sensor array. (a) Electrode (front); (b) heater (backside); (c) photograph of sensor array.

the rutile structure (JCPDS 21-1250) of the commercial analyses, as shown in Table 2. After measuring the full width
powder (99.9%, Aldrich Co.). at half maximum (FWHM) of the XRD pattern peaks, the
TEM photographs of this powder with and without Ca particle size was then calculated using Scherrer's formula
after being calcined at 6008C for 1 h are shown in Fig. 5. The
Ca-added powder exhibited restrained particle growth sizes.
Xu et al. reported that a reduced particle size produces a
signicant increase in sensitivity [19,20]. Accordingly, the
Ca-added powder would be more sensitive in detecting low-
level gas concentrations. Based on the above papers, the
particle sizes of commercial SnO2, precipitated SnO2, and
SnO2/Ca,Pt powders were measured using XRD and TEM

Fig. 4. XRD patterns of SnO2/Ca,Pt (0.1 wt.%) powders at varying

Fig. 3. Schematic diagram for measuring the properties of sensor array. temperatures.
 D.-S. Lee et al. / Sensors and Actuators B 71 (2000) 9098 93

Fig. 6. BET analysis of SnO2/Ca,Pt (0.1 wt.%) powder.

Using the BET method, the nitrogen sorption curve for the
base material was investigated, as shown in Fig. 6. The
hysterisis recorded in the graph indicated a higher number of
pores in the sensing materials, which produces more effec-
tive gas diffusion and a higher sensitivity. The specic
surface areas calculated from this curve are shown in
Table 3. As the annealing temperature increased, the specic
surface area decreased. The coprecipitation of SnO2 with Ca
and Pt produced an increased specic surface area.

3.2. Explosive gas-sensing characteristics of SnO2-based

gas sensors

Fig. 5. TEM photographs of (a) pure SnO2; (b) SnO2/Ca,Pt (0.1 wt.%)
powders with heat-treatment at 8008C for 1 h.
[21]. The differences between the values of the XRD and
TEM analyses were due to the pretreatment process for the
TEM photographs. The particle size of the precipitated SnO2
powder was smaller than that of the commercial powder. The
particle size of the coprecipitated SnO2/Ca,Pt powder was
the smallest at 7 nm after being calcined at 6008C for 1 h.
Accordingly, it is believed that a sensor with nano-sized
powders will be effective in detecting low-level gas con-
centrations.
The response time of a sensor to a gas is directly related to
gas detection and classication. Thus, the time response
property of the sensor array to objective gases is very
important. The time response curve of the sensor array to
4000 ppm propane at 4008C is shown in Fig. 7. It shows that
the gas reaction began to saturate within 35 s and de-
sorption of the gas was also completed within 30 s. The
detection and recognition of a gas was thus completely
nished within 23 min, in situ. The sensors also exhibited
a variety of sensitivity curves according to the additives used
to obtain gas selectivity.
Different sensing patterns for the sensors are critical for
conferring selectivity to a sensor array. Fig. 8 shows the

Table 2
Particle size of SnO2 and SnO2/Ca,Pt powders for the heating temperature

Calcining temperature (8C) Pure SnO2 (nm) Commercial SnO2 (nm) 0.1 wt.% SnO2/Ca,Pt (nm)

XRD TEM XRD TEM XRD TEM

Room temperature 5 5.5 15.8 16.5 2.5 2

500 5.6 5
600 9 10 34.8 35 7.2 7.5
700 15 15.5 10.8 10
800 20 20 12.1 12.5
900 24 24.5 17.6 18
1000 28 29 26.5 27
94 D.-S. Lee et al. / Sensors and Actuators B 71 (2000) 9098

Table 3
Specific surface area of SnO2 and SnO2/Ca,Pt powders for the heating
temperature

Condition SnO2 SnO2/Ca,Pt

(m2/g) (0.1 wt.%)

Calcination at 6008C, 1 h 25.5 35

Annealing at 8008C, 1 h (after calcination) 19.8 26

sensitivity of the sensor array to explosive gases such as

butane, propane, and methane below their LELs, specically
at 2000, 4000, and 6000 ppm, respectively. All sensors
exhibited a variety of sensitivity ranging from 20 to
100% and showed a monotonous increase of sensitivity with
increased gas concentrations. This gure also shows the
effects of additives in the base material, that is, sensing
materials with the addition of TiO2 showed the highest
sensitivity to butane, whereas the addition of TiO2 or
ZnO showed the highest sensitivity to propane. The highest
sensitivity to methane was shown in sensing materials with
the addition of La2O3. Accordingly, it is believed that the use
of these patterns will produce distinguishable selectivity in a
sensor array.
The sensor responses tend to drift signicantly when
Fig. 8. Sensitivity of nine-sensor array to three explosive gases. (a)
sensors are used over a long period of time resulting in Methane; (b) propane; (c) butane.
poor selectivity and redundant pattern recognition. There-
fore, the fabrication of more reliable and stable sensors, or
the ability to adjust the pattern recognition routine according long-term stability in air and in 2000 ppm butane for 12
to variations in the sensor response is needed. For example, months. But, the sensitivity decreased gradually and slowly.
Fig. 9 shows the long-term stability of a sensor with nano- Thus, by employing the conventional pattern recognition
sized tin oxide materials, especially SnO2/Ca,Pt routine, we worked on the adjustment to variations in the
(0.1 wt.% Pt,Pd (2 wt.%) sensor. The sensor shows sensor response.

Fig. 7. Time response of sensors with 4000 ppm propane gas.

 D.-S. Lee et al. / Sensors and Actuators B 71 (2000) 9098
Fig. 9. Long-term stability of SnO2/Ca,Pt 0:1 wt:% Pt; Pd (2 wt.%)
sensor.
The possibility of classifying the species and quantity of a
specic gas can be achieved by using a principal component
analysis (PCA) that can map multi-dimensional data onto
two- or three- dimensional axes with the minimum loss of
information [22]. After the data shown in Fig. 8 was
remeasured several times, it was transposed onto two axes,
as shown in Fig. 10. The 1st principal axis was set to the
direction in the largest dispersion of datum and 2nd principal
axis to the direction in the second largest dispersion of datum
in a multicomponent analysis. This demonstrates the ability
to cluster points roughly in types according to their species
and quantity. However, it was difcult to clearly classify the
6000 ppm propane and 2000 ppm butane as they are located
close to each other. Therefore, to precisely classify clusters
with similar characteristics and overcome statistical var-
iances, a gas pattern recognizer using a multi-layer neural
Fig. 10. PCA results of sensing data obtained from sensor array.
95
network was employed along with an error back propagation
learning algorithm, which is known to be effective in
classifying and identifying complex nonlinear patterns [22].
3.3. Implementation of gas pattern recognizer and gas
recognition system
Using the sensitivity signals from the sensor array as
multi-dimensional input patterns, a gas pattern recognizer
using a multi-layer neural network with an error back
propagation learning algorithm was then implemented
[2224]. The neural network consisted of an input layer
with nine nodes which received the data from the sensors on
the sensor array, a hidden layer with eight neurons, and an
output layer with nine nodes as shown in Fig. 11. The nine
nodes in the output layer indicated the recognition results for
three concentration levels of three kinds of explosive gases,
butane, propane, and methane. Fig. 11 shows the recognition
result in the case of an injection of 2000 ppm propane gas. In
this study, the experimental data were measured in triplicate
with 2000, 4000, and 6000 ppm concentrations of methane,
propane, and butane. The learning process was repeated
until the learning error reached 0.0001.
The implemented gas recognition system is shown in
Fig. 12. An analog multiplexer, lter, and signal transformer
for level shifting were all included in a data acquisition
board for transferring the multi-signal to a digital signal
processor (DSP) board. The learned connection-weights
were then saved in the DSP board memory. For real time
processing, the recognized gas species and concentrations
can then be displayed on a liquid crystal display (LCD)
96 D.-S. Lee et al. / Sensors and Actuators B 71 (2000) 9098

Fig. 11. Multi-layer neural network for gas recognition system.

board. A photograph of the fabricated explosive gas-recog-

Fig. 12. Architecture of proposed gas recognition system for identification
of explosive gases.
nition system using a DSP board is shown in Fig. 13.
The specic recognition results for the proposed explo-
sive gas recognition system with 2000 ppm butane,
4000 ppm propane, and 6000 ppm methane are shown in
Fig. 14. After learning, the proposed gas recognition system
demonstrated an almost 100% recognition rate. To verify
this recognition rate, new data was generated with a 5% error
in the original data and the resulting recognition rate was
examined. It also showed a high recognition rate of about
99%.

Fig. 13. Picture of implemented explosive gas recognition system.

 D.-S. Lee et al. / Sensors and Actuators B 71 (2000) 9098
Fig. 14. Recognition results of implemented gas recognition system for explosive gases: (a) 2000 ppm butane; (b) 4000 ppm propane; (c) 6000 ppm methane.
4. Conclusion
To recognize the species and quantity of explosive gases
under a low explosion limit, the sensor array with nine SnO2/
Ca,Pt-based metal oxide sensors integrated in an alumina
substrate were fabricated. The SnO2/Ca,Pt base material was
prepared using coprecipitation, and it exhibited a particle
size of about 7 nm, a high specic surface area of 35 m2/g, a
rutile structure after being calcined at 6008C for 1 h, and
long-term stability. To give good selectivity to the sensors,
several additives including Pt, Pd, Au, CuO, La2O3, Sc2O3,
TiO2, WO3, and ZnO were mixed with the base material.
The resulting sensor arrays showed high and selective
sensitivities and long-term stability to low gas concentra-
tions at 4008C.
Using the sensing signals from the sensor array along with
a multi-layer neural network using an error back propagation
learning algorithm, an intelligent gas recognition system
with a DSP board was implemented in order to produce a
real time recognition system for explosive gases such as
butane, propane, and methane. Furthermore, the fabricated
gas recognition system could precisely classify and identify
the species and concentration of each gas in real time
processing.
In future, the implementation of a sensor array and system
that can recognize continuously varying concentrations and
97
specic gases in a mixture of gases will be attempted.
Finally, a real time small-sized system that can consider
environmental conditions like peripheral temperature and
relative humidity will be designed to detect possible lea-
kages of explosive gases and recognize the species and
quantity of specic gases in gas pipes or workplaces.
References
[1] B. Yea, R. Konishi, T. Osaki, K. Sugahara, The discrimination of
many kinds of odor species using fuzzy reasoning and neural
networks, Sens. Actuators A 45 (1994) 159165.
[2] J.W. Gardner, P.N. Bartlett (Ed.), Sensors and sensory systems for an
electronic nose, NATO advanced research workshop, Reykjavik, 1991.
[3] J.W. Gardner, A brief history of electronic noses, Sens. Actuators B
18/19 (1994) 211220.
[4] J. Getino, J. Gutierrez, L. Ares, J.I. Robla, M.C. Horrillo, Integrated
sensor array for gas analysis in combustion atmospheres, Sens.
Actuators B 33 (1996) 128133.
[5] M.C. Horrillo, J. Getino, L. Ares, F.J. Gutierrez, Measurements of
VOCs with a semiconductor electronic nose, J. Electrochem. Soc.
145 (7) (1998) 24862489.
[6] U. Weimar, K.D. Schierbaum, W. Gopel, Pattern recognition methods
for gas mixture analysis: application to sensor arrays based on SnO2,
Sens. Actuators B 1 (1990) 9396.
[7] S. Somov, G. Reinhardt, W. Gopel, Gas analysis with arrays of solid
state electrochemical sensors; implications to monitor HCs and NOx
exhausts, Sens. Actuators B 35/36 (1996) 409418.
98 D.-S. Lee et al. / Sensors and Actuators B 71 (2000) 9098
[8] K.C. Persaud, S.M. Khaffaf, J.S. Payne, D.H. Lee, H.G. Byun, Sensor
array techniques of mimicking the mammalian olfactory system,
Sens. Actuators B 35/36 (1996) 267273.
[9] C.-F. Mandenius, T. Eklov, I. Lundstrom, Sensor fusion with on-line
gas emission multisensor arrays and standard process measuring
devices in Baker's yeast manufacturing process, Biotechnol. Bioeng.
55 (2) (1997) 427438.
[10] V.I. Anistimkin, R.G. Krystal, A.V. Medved, E. Verona, V.E.
Zemlyakov, Integrated array of gas sensors, Electron. Lett. 34 (13)
(1998) 13601361.
[11] J. Mitrovics, U. Weimar, W. Gopel, Linearisation in multicomponent
analysis based on a hybrid sensor array with 19 sensor elements,
Transducer'95, Stockholm, Sweden, 1995, pp. 707710.
[12] Y.H. Hong, D.D. Lee, Gas sensing characteristics of SnO2(Ca)/Pt
thick film using Pt electrode for hydrocarbon gases, J. Korean Sens.
Soc. 4 (2) (1995) 3144.
[13] D.D. Lee, Hydrocarbon gas sensors, Chemical Sensor Technology,
Vol. 5, Kodansha Ltd., 1994, pp. 7999.
[14] D.D. Lee, W.Y. Chung, Gas-sensing characteristics of SnO2x thin
film with added Pt fabricated by the dipping method, Sens. Actuators
B 1 (1989) 301305.
[15] W.Y. Chung, D.D. Lee, Effects of added TiO2 on the characteristics
of SnO2 based thick film gas sensors, Thin Solid Films 444 (1992)
304310.
[16] J.H. Sung, Y.S. Lee, B.H. Kang, J.W. Lim, D.D. Lee, Sensing
characteristics of tin oxide/gold sensor prepared by coprecipitation
method, in: Proceedings of the seventh International conference of
Chemical Sensors, Beijing, China, 1998, pp. 819821.
[17] T. Seiyama, Denki Kagaku 40 (1972) 244250.
[18] N. Yamazoe, N. Miura, Some basic aspects of semiconductor gas
sensors, Chemical Sensor Technology, Vol. 4, Kodansha Ltd., Tokyo,
Japan, 1992, pp. 1942.
[19] C. Xu, J. Kurokawa, N. Miura, N. Yamazoe, Stabilization of SnO2
untrafine praticles by additives, J. Mater. Sci. 27 (1992) 963971.
[20] C. Xu, J. Tamaki, N. Miura, N. Yamazoe, Relationship between gas
sensitivity and microstructure of porous SnO2, Denki Kagaku 58 (12)
(1990) 11431148.
[21] B.D. Cullity, Elements X-ray Diffraction, 2nd Edition, Addison
Wesley, London, 1978.
[22] A. Hierlemann, M. Schweizer, U. Weimar, W. Gopel, Pattern
recognition and multicomponent analysis, Sensors update, VCH,
Weinheim, Germany, 1996.
[23] P.D. Wasserman, Neural Computing Theory and Practice, VNR, NY,
USA, 1989.
[24] S. Haykin, Neural Networks a Comprehensive Foundation, Macmil-
lan, NY, USA, 1994.