1370 IEEE TRANSACTIONS ON INSTRUMENTATION AND MEASUREMENT, VOL. 50, NO.

5, OCTOBER 2001
Identification of Forane R134a in an Air-Conditioned
Atmosphere With a TGS Sensor Array
Claude Delpha, Maryam Siadat, and Martine Lumbreras
AbstractFor electronic nose applications based on tin oxide
gas sensors, relative humidity is a very influential parameter that
may cause false gas detection. In our application, we want to dis-
tinguish a refrigerant gas (Forane R134a) in an air-conditioned at-
mosphere using a TGS gas sensor array. First, this paper gives a
brief overviewfor accurate test equipment for characterizing these
types of sensors under closely controlled humidity and temper-
ature conditions. Next, the humidity influence is experimentally
shown, either for air or for Forane R134a gas concentrations in air.
Then, by using two complementary pattern recognition methods,
principal component analysis (PCA) followed by discriminant fac-
torial analysis (DFA), we show the ability to discriminate accu-
rately the target gas whatever the humidity rate. We also prove
that it is possible to identify unknown test cases with the resultant
decisive law.
Index TermsData analysis, discrimination and identification,
electronic nose, pattern recognition, tin oxide gas sensor array.
I. INTRODUCTION
T
ODAY, there is an increasing interest in electronic nose ap-
plications [1], [2]. In fact, the concept of electronic nose
is based on a gas sensor array with different selectivity patterns,
a signal processing and a data acquisition unit in association
with a pattern recognition system. The final aim of our work is
the conception of an environmental electronic nose, based on an
array of TGS-type metal oxide sensors, for the main detection of
a refrigerant gas Forane R134a, carbon dioxide (CO ) and dif-
ferent volatile organic compounds (VOCs) in an air-conditioned
atmosphere. These sensors are very sensitive, but their selec-
tivity and response are greatly affected by many pollutants like
the gas relative humidity ratio and temperature [3], [4]. Several
investigations have been done to solve the humidity problem by
incorporating additives into the oxide [5] or by modulating the
sensor operating temperature [6]. Unfortunately, this problem
was not completely solved by these methods. Another type of
investigation consists of taking the humidity ratio into account
in the pattern recognition analysis method. For our application,
this last method was used: the humid air is then considered as a
pollutant gas. Therefore, the gas sensor array has been charac-
terized under several humidity-controlled atmospheres, and the
results have been included in the sensor database for the pattern
recognition.
This paper concerns the Forane R134a identification in a
humid air atmosphere whatever the relative humidity rate. This
Manuscript received July 23, 1999; revised June 1, 2001.
The authors are with the Laboratoire Interfaces Composants et Microlec-
tronique, LICM/CLOES/SUPELEC, Universit de Metz, 57070 Metz, France
(e-mail: delpha@ese-metz.fr; lumbre@ese-metz.fr).
Publisher Item Identifier S 0018-9456(01)08736-8.
new generation refrigerant gas, Forane 134a, is studied because
it can cause greenhouse effects and also health problems in case
of leakage in a closed space. In this paper, we first describe the
sensor array characterization results, in terms of steady-state
conductance, for a humid air atmosphere, then for a mixture
of Forane R134a in dry or humid air. Then, by using the
principal component analysis (PCA) method, we show how the
experimental results can be correlated, and then separated in
groups. We then insert these groups into our database, and train
it with the discriminant factorial analysis (DFA) method. We
prove that it is possible to discriminate accurately the gas type
and afterwards identify unknown test data even if the relative
humidity rate varies.
II. EXPERIMENTAL
For the experiments, we have designed a systemable to create
closely controlled conditions of humidity and temperature [7].
It is a dynamic flow system (100 ml/min): the carrier gas (syn-
thetic air) is brought into a humidity generator based on the bub-
bler principle and then mixed with the main target gas (Forane
R134a) via gas lines controlled by mass flowcontrollers (Model
5850TR, Brooks Instruments). Then the mixture is introduced
in the test chamber that is placed in a gas-controlled tempera-
ture atmosphere (33 C). This chamber was specially designed
to be circular in stainless steel materials and also optimized to
provide the shorter gas exchange and laminar gas flow to the
enclosed gas sensors.
For our application, we are using three pairs of TGS-type gas
sensors provided by Figaro Engineering Inc. (TGS 832, a refrig-
erant gas sensor; TGS 800, a general air contaminant sensor; and
TGS 813, a hydrocarbon sensor) inside the test chamber. These
types of sensors were chosen because of their high sensitivity to
organic, pollutant, and combustible gases. In our application, all
the sensor output responses are collected via a data acquisition
board (LabPC , National Instruments), treated, and then ana-
lyzed by pattern recognition methods using statistical and data
analysis software (SPSS 8.0, SPSS Inc.).
III. RESULTS AND DISCUSSION
A. Characterization
To create our database, it is necessary to take into account
the simultaneous effect of the humidity and the studied gas for
obtaining a more accurate identification of the gas in the real
atmosphere. Therefore, we have characterized the sensors in
humid air or in a mixture of Forane R134a in dry or wet air.
00189456/01$10.00 2001 IEEE
DELPHA et al.: FORANE R134a IN AN AIR-CONDITIONED ATMOSPHERE 1371
Fig. 1. Typical sensor time-dependent behavior for reducing gases.
Fig. 2. Sensors steady-state response in humid air.
We have studied the behavior of the sensor array for five dif-
ferent relative humidity rates (18%, 35%, 52%, 68%, and 85%)
at 33 C. For the measurement in humid air, a dry synthetic air
flow is set for 1 h before submitting each humidity rate to the
sensors for 1 h. For each sensor, we have noted a similar be-
havior to a reducing gas exposure (Fig. 1), stabilized in the best
cases after 20 min [8]. Therefore, the time-dependent response
as well as the steady-state conductance increases as a function of
the relative humidity rate for the three types of sensors (Fig. 2).
As for humid air, the sensors have been characterized in a
mixture of Forane R134a in dry air. In this case, we have studied
the sensor array behavior in the Forane R134a concentration
range 200 ppm1000 ppm with steps of 200 ppm. Before each
Forane R134a-dry air flow, a 100 ml/min synthetic dry air flow
is set to stabilize the sensor sensitive layer [9] for 1 h. The sensor
array response to Forane R134a, which is a reducing gas, in-
creases along with the concentration (see Fig. 3) in the same
way as the previous results in humid air.
Then, we have studied the sensor behavior under a mixture of
Forane R134a in wet air. Therefore, for a Forane R134a/humid
air atmosphere, we present the sensor array conductance time-
dependent response (see Fig. 4). We can note, for all the sensors,
that the reducing effects due to humid air and Forane R134a are
cumulative. For the studied relative humidity rates, the sensors
offer similar responses. In fact, the steady-state conductance
values increase along with the gas concentrations and the rel-
ative humidity rates (see Fig. 5), but we note that the Forane
R134a response is masked for the lowest concentrations (200
ppm400 ppm) at the highest relative humidity value (85%).
Fig. 3. Sensors steady-state response in dry Forane R134a.
Fig. 4. TGS 800 dynamic behavior in humid Forane R134a.
Fig. 5. TGS 800 steady-state response in humid Forane R134a.
All these characterization results (160 measurements) have been
grouped in a database and exploited by using first PCA and then
DFA to prove that this base is efficient to identify the Forane
R134a even if the relative humidity rate varies.
B. Principal Component Analysis (PCA)
This pattern recognition technique is a powerful unsuper-
vised method [10] usually employed for the tin oxide gas sensor
array. To apply this method, steady-state conductance values
obtained from the characterization measurements are grouped
in a database. Then, linear combinations of the orthogonal re-
sponse vectors are expressed, and principal components are cal-
1372 IEEE TRANSACTIONS ON INSTRUMENTATION AND MEASUREMENT, VOL. 50, NO. 5, OCTOBER 2001
TABLE I
EIGENVALUES OF THE RESPONSE DATA BASE
Fig. 6. PCA results for the TGS sensor array steady-state conductance.
culated. Then, eigenvalues representing the data variance per-
centage contained in each principal component are given. The
aim of this technique consists of removing any redundancy and
reducing the dimensionality of the database. In Table I, we have
summarized the eigenvalues and the data variance percentages
for all the principal components. Because the two first compo-
nents allow us to take into account more than 95% of the data
base variance, we limit our representation to these two compo-
nents. Therefore, for these two axes, we present in Fig. 6 the
PCA results for all the studied TGS sensors steady-state con-
ductance. We note that the results can be separated into three
parts corresponding to each measurement gas type: humid air,
dry Forane R134a, and wet Forane R134a. However, the bound-
aries between the groups are not well defined, and we note
minor overlappings. Therefore, with these three groups defined
by PCA, we nowstudy the measurement database with the DFA
method to see if the steady-state conductance values are suffi-
cient variables to accurately identify the gas type. If not, we have
to define a new variable to improve the target gas type identifi-
cation.
C. Discriminant Factorial Analysis (DFA)
As for PCA, this technique is a factorial method [10] widely
used for electronic nose applications [11]. In fact, for this
method an a priori group is given for all the measurements in
the database. Then, the discriminant procedure, which consists
of maximizing the differences between all the groups and
minimizing these differences inside each group, is done. A
Fig. 7. DFA results for the TGS sensor array steady-state conductance.
TABLE II
CLASSIFICATION RESULTS FOR THE DFA WITH THE STEADY-STATE
CONDUCTANCE DATA BASE
linear combination of the variables which best characterizes
the differences among the groups is produced, and the obtained
discriminant function can be used to classify new cases. To
apply this method to our database, the three main groups de-
fined by PCA are applied, then two discriminant functions are
obtained. In Fig. 7, we present the DFA results for our sensor
array with two main axes representing 100% of the database.
With this representation, we show that the steady-state conduc-
tance values can be considered sufficient to generate a quite
good discriminant function for the gas type discrimination.
In fact, we have some reclassifications: only 98.1% of the
measurements in the database have been classified in their a
priori group. As shown in Table II, a few errors are produced
by classifying several dry or wet Forane R134a measurements
in the humid air group, but these measurements correspond to
the lowest gas concentrations. These results have then been
validated by the cross-validation method, and we have found
that only 97.5% of the cross-validated measurements have been
classified correctly.
In order to test the created decisive law, we have used a test
data set composed of 15 unknown cases (see Table III) corre-
sponding to different relative humidity ratios and Forane R134a
gas concentrations. In fact, we are able to accurately identify
all the unknown cases except the humid air measurements (five
cases) that were mainly identified as wet Forane R134a (see
Fig. 7). Therefore, the decisive law found with this database has
DELPHA et al.: FORANE R134a IN AN AIR-CONDITIONED ATMOSPHERE 1373
TABLE III
TEST DATA SET CORRESPONDING MEASUREMENTS AND IDENTIFICATION RESULTS
Fig. 8. Sensor conductance dynamic slope responses in humid air.
to be improved by associating a new variable to the steady-state
conductance.
A new variable known as the conductance dynamic slope
of the time-dependent response measured in the first five min-
utes of gas exposure is chosen (see Fig. 1). For the different
types of measurements, we found that this variable increases
significantly with the relative humidity ratio (see Fig. 8), or the
gas concentration (see Fig. 9). Therefore, for each sensor of
the array, these values were associated to the previous steady-
state conductance database. This new database is then trained
by using the same groups (Humid Air, Dry R134a, and Wet
R134a) and a new discriminant function is generated. As shown
in Fig. 10, the three groups are more separated compared to
Fig. 7. In fact, the classifying results obtained for this anal-
ysis give us only one reclassified data which corresponds to a
low concentration of dry Forane R134a classified in the humid
air group. Therefore, 99.4% of the complete new database has
been classified in the a priori group. The generated discriminant
function has been checked by the cross-validation method and
the same percentage was found, showing that this function is
correct and then can be used to classify unknown cases. Thus,
we have tried to identify the same test data set (see Table III)
Fig. 9. Sensor conductance dynamic slope response in dry Forane R134a.
Fig. 10. DFA results by using two kinds of variables for each sensor: the
dynamic slope and the steady-state conductance value.
with this new decisive law. We found that only one unknown
case was misidentified (see Fig. 10): only a humid air measure-
ment (85% RH) was classified as dry Forane R134a. Therefore,
1374 IEEE TRANSACTIONS ON INSTRUMENTATION AND MEASUREMENT, VOL. 50, NO. 5, OCTOBER 2001
the created lawcan be used to accurately identify unknown cases
whatever the gas relative humidity rate.
IV. CONCLUSION
We have characterized a TGS sensor array in a humid (085%
RH) atmosphere composed of a mixture of Forane R134a in syn-
thetic air (01000 ppm). The sensor steady-state conductance
values are arranged in a database and then treated by PCA and
DFA methods. For these two methods, we have shown that the
steady-state value is a variable that allows us to provide a quite
good separation of the gas types. However, this variable is not
really sufficient to create an accurate discriminant function to be
used to perfectly identify unknown cases. As a matter of fact,
another variable has been added for each sensor of the array:
the dynamic slope. This variable allows us to improve the gas
discrimination and to create a decisive law able to accurately
classify new unknown cases. Therefore, by using mathematical
treatment and statistical analysis, we are able to compensate for
the sensor variations and inaccuracies due to the presence of hu-
midity in an air-conditioned atmosphere.
REFERENCES
[1] W. Gpel, Chemical imaging: Concepts and visions for electronic and
bio-electronic noses, Sens. Actuators B, vol. 52, pp. 125142, 1998.
[2] J. W. Gardner and P. N. Bartlett, Performance definition and standard-
ization of electronic noses, Sens. Actuators B, vol. 33, pp. 6067, 1996.
[3] C. Delpha, M. Siadat, and M. Lumbreras, Environmental temperature
and humidity variation effects on the response of a TGS sensor array,
in Proc. ISOEN, Tbingen, Germany, Sept. 2022, 1999, pp. 156159.
[4] K. Ihokura and J. Watson, The Stannic Oxide Gas Sensor: Principles
and Applications. Boca Raton, FL: CRC, 1994.
[5] D. S. Vlachos, P. D. Skafidas, and J. N. Avaritsiotis, The effect of hu-
midity on tin oxide thick filmgas sensors in the presence of reducing and
combustible gases, Sens. Actuators B, vol. 2425, pp. 491494, 1995.
[6] P. Van Geloven, M. Honore, S. Leppavuori, and T. Rantala, The in-
fluence of relative humidity on the response of tin oxide gas sensors to
carbon monoxide, Sens. Actuators B, vol. 4, pp. 185188, 1991.
[7] C. Delpha, M. Siadat, and M. Lumbreras, Humidity effect on a com-
mercially available refrigerant gas sensor TGS 832, Proc. IEEE, vol.
3539, pp. 172179, 1998.
[8] , Humidity dependence of a TGS gas sensor array in an air-con-
ditioned atmosphere, Sens. Actuators B, vol. 59, no. 23, pp. 255259,
1999.
[9] C. Delpha, F. Sarry, M. Siadat, and M. Lumbreras, Evaluation of a
sensor array in an environmental controlled gas atmosphere, in Proc.
Sensor, vol. 2, Nrnberg, Germany, May 1820, 1999, pp. 513518.
[10] D. F. Morisson, Multivariate Statistical Methods, 2nd ed. Singapore:
McGraw-Hill, 1988, p. 415.
[11] H. T. Nagle, R. Gutierez-Osuna, and S. S. Schiffman, The howand why
of electronic noses, IEEE Spectrum, vol. 35, no. 9, pp. 2234, 1998.
Claude Delpha received the M.S. degree in biomedical electronics engineering
and the Instrumentation and Microelectronics postgraduate degree in the field
of gas sensing and signal processing techniques from the University of Nancy,
Nancy, France, in 1995 and 1996, respectively, and the Ph.D. degree in the field
of gas sensing and signal processing techniques from the Laboratory of Inter-
faces, Components, and Microelectronics, the University of Metz, Metz, France,
in 2000.
He is currently an Associate Professor at the University of Paris XI, Paris,
France. His main areas of interest are in electronics and instrumentation, semi-
conductor chemical sensors, chemical vapor discrimination using sensor arrays,
gas sensors humidity and temperature dependence, electronic nose, and pattern
recognition methods.
Maryam Siadat received the Dipl.Eng. degree in electronics in 1983 and the
Ph.D. degree in biomedical electronics engineering in 1989, both fromthe Poly-
technic Institute of Lorraine (ENSEM/INPL), Nancy, France.
She is currently an Associate Professor at the University of Metz, Metz,
France. Her research interests are in gas detection, signal and data processing,
sensor characterization, and numerical electronics circuits development.
Martine Lumbreras received the M.S. degree in electrical engineering with
specialization in solids electronics in 1969, and the Ph.D. degree in 1979, both
from the University of Montpellier, Montpellier, France, and the Dr. Sci. degree
in 1987 from the University of Metz, Metz, France.
Since 1991, she has been a Professor at the University of Metz. In 1994, she
created a sensor research group, which is part of the universitys Laboratory of
Interfaces, Components, and Microelectronics.