Sensors and Features Selection for Robust Gas Concentration
Evaluation
D. Ahmadou, E. Losson, M. Siadat and M. Lumbreras
Université de Lorraine, LCOMS, EA 7306, Metz, F-57000, France
Keywords: Gas Sensor Properties, Feature Comparison, Derivative Signal, Exposure and Purge times, Drift.
Abstract: This paper seeks to highlight the importance of the knowledge of metal oxide gas sensor behaviour before
conceiving an electronic nose for a dedicated application. Therefore, a depth study of sensor response
properties is needed for the selection of the more appropriate sensors via optimized measurement conditions
and extracted features. Especially for continuous gas evaluation, the most important aspects to consider are
the measurement time and the drift of the gas sensors. In this work, for fast recognition of pine oil vapour
dilutions, the performance of two features are shown: the maximum of the derivative curve (Peak), an
unusual feature which needs a very short gas exposure time, and the sensor amplitude voltage (Vs-V0)
obtained at the end of the gas exposition phase. The performance of the new feature Peak, validated by
Principal Component Analysis results, leads us to work with the shortest gas exposition and sensor
regeneration times, and allows us to choose the best sensors according to our application.
1 INTRODUCTION
Nowadays, electronic noses gain interest as general
purpose detectors of vapours in many fields of
application because these mobile and intelligent
instruments, easy to build, offer the possibility of
direct measurement (Falasconi et al., 2005; Cho et
al., 2008; Zhang and Wang, 2007). These systems
are largely used to detect, identify or quantify
complex atmospheres (Boilot et al., 2002; Branca et
al., 2003; Martin Negri and Reich, 2001). They
employ an array of gas sensors with different
selectivities, more often resistive metal oxide
sensors (Gutierrez-Osuna, 2002). The indisputable
advantages of these sensors are their high sensitivity,
robustness, and commercial availability. But two
main limitations must be taken into account to
provide fast and reliable gas identification: the delay
of the sensor response time and the gas sensor drifts.
So, the key requests of electronic noses, working in
continuous checking, are the conception of an
accurate sampling unit (Roussel et al., 1999) with
optimization of the recognition speed.
Considering the electronic nose as a “black box”
and referring only to the mathematical computing
results after recognition analysis cannot permit
robust real-time measurements. Therefore, the entire
knowledge of the gas sensor behaviour is very
important to select, for a given application, the best
measurement conditions, the best extracted features
and the best sensors by considering their
characteristics. This selection must be valid for the
entire chosen application.
For this purpose, reliable informative features
must first be selected to characterize the sensor time-
response. This feature selection should take into
account the behaviour of the gas sensors for all the
studied atmospheres. A lot of features have been
mentioned and compared in the literature (Llobet et
al., 2002; Distante et al., 2002; Paulsson, 2000;
Zhang et al., 2007). Representative features can be
extracted either from the transient phase (initial
slope, FFT and wavelet descriptors, integral,…) or
from the steady-state phase (absolute, relative,
fractional or log sensor conductance values) of the
sensor time-responses. In the case of steady-state
response, obtaining robust features needs generally a
long gas exposition time, not suitable for fast
recognition system.
We have particularly investigated a novel
transient parameter, deduced from the derivative
curve of the sensor time-response: the height of its
maximum (Peak), occurred before 100 seconds after
the gas exposition. The second studied feature is the
traditional relative change (Vs-V0), representing the
237
Ahmadou D., Losson E., Siadat M. and Lumbreras M..
Sensors and Features Selection for Robust Gas Concentration Evaluation.
DOI: 10.5220/0004670002370243
In Proceedings of the 3rd International Conference on Sensor Networks (SENSORNETS-2014), pages 237-243
ISBN: 978-989-758-001-7
Copyright
c
2014 SCITEPRESS (Science and Technology Publications, Lda.)
sensor response amplitude.
In this work, our electronic nose application
concerns the quantification of pine Essential Oil
(EO) vapours diluted in pure air. At first, the
analysis of the two features (Vs-V0) and Peak will
be used to optimize the measurement conditions in
order to obtain the fastest quantification. For this
purpose, discussions will be done about the
robustness of the selected features using the
optimized measurement protocol. After this first
step, sensors can be characterized by comparing our
two features: (Vs-V0) and Peak. The performance of
these features will be discussed along with the EO
concentrations and the sensor types. Finally, the
choice of experimental and calculation conditions,
validated by PCA, will allow us to identify the more
adequate sensors for our quantitative application.
2 MATERIALS AND METHODS
The study presented in this work concerns an
application for the estimation of EO vapour dilutions
by using a metal oxide gas sensor (MOX) array. The
global aim is to develop an electronic nose based
system to regulate the EO diffusion in a closed and
conditioned box.
2.1 Equipment Description
A test bench is mounted to generate various EO
concentrations in order to characterize and to
optimize the commercial MOX array of our
electronic nose. Figure 1 presents the functional
diagram of this experimental system.
The EO generation is made by bubbling
synthetic air flow in a bottle containing 1cm
3
of
liquid essential oil. To produce a desired EO diluted
atmosphere at a constant total flow rate, the created
odorant atmosphere is combined with pure air, and
then introduced into the gas sensor cell. So, various
concentrations are obtained by varying the flow rate
of the EO line to be combined with the pure air flow
rate. These EO concentrations (dilutions) are then
expressed as a percentage of the bubbling flow rate
in liquid oil over the total flow rate (100ml/min).
Pine oil at very low percentages (1, 2, 3, and 4%)
is utilized in this study. These concentrations
correspond to a pleasant odour (human panel) for
aromatherapy uses (Sambemana, Siadat and
Lumbreras, 2010). Gas chromatography
measurements were made on the EO pine samples
before the beginning and during the experiment
phase in order to control the stability of the EO
sample composition (molecules and their
concentrations).
The gas sensor cell contains 9 sensors
(TGS2620, TGS880, TGS822, TGS816, SPAQ1,
SPMW0, SP31, MQ3, MQ138) from Figaro, FIS
and Hanwei companies. Sensor responses are
digitalized and collected using a fast and high
resolution data acquisition board. The whole system
will be optimized for an accurate and rapid EO
concentration evaluation. In the functional diagram
(figure 1), we present also a sensor time-response in
terms of sensor voltage response versus time. The
signal shows first a voltage increase with an
inflexion point, corresponding to the gas exposition.
The second part corresponds to the sensor
regeneration.
Figure 1: Functional diagram of the gas sensor
characterization system.
2.2 Feature Determination
After each gas exposition, a sensor regeneration
must be undertaken to recover the conductance basis
value of the sensor. In previous studies, we used a
cycle composed of 5 minutes gas exposition time
followed by 20 minutes regeneration time. This
cycle allowed to obtain sensor response stabilization
during the exposition phase for all the sensors and
all the EO concentrations, and also a good
regeneration at the end of the purge phase.
We have tested many characteristic parameters
corresponding to transient and steady-state phases
(Szczurek and Maciejewska, 2012; Gualdron et al.,
2004), and then selected for this study two features:
one extracted from the sensor time-response, and the
second from the derivative curve of this response.
We have compared the performance of these two
features to discriminate the EO concentrations in
order to choose the best sensors acting with the
shortest measurement cycle, necessary for a real
time application.
2.2.1 Derivative Feature
To have a rapid evaluation of the gas concentration,
SENSORNETS2014-InternationalConferenceonSensorNetworks
238
it is necessary to consider the transient phase of the
sensor time-response (Ionescu, Vancu and Tomescu,
2000; Martinelli et al., 2003; Pardo and Sberveglieri,
2007). More we wait for a complete stabilization,
more the exposition time is long, and longer will be
the regeneration time. In our application, for several
studied cases (towards the sensors and/or the gas
concentrations) the time-response needs more than 5
minutes to reach 90% of the stabilization level.
The most studied transient feature is the initial
slope of the time-response signal (Sysoev et al.,
2007; Delpha, Siadat and Lumbreras, 2001). But the
difficulty is to determine the starting and the end
points of the linear transient phase. It is impossible
to fix a general rule for the calculation of this slope
because these points vary along with the gas
concentration and the sensor types.
So, we have decided to differentiate all the signal
time-responses in order to determine the maximum
of the derivative curve corresponding to the
inflexion point of the sensor time-response. To
reduce noise in the derivative signal, it was needed
the use of an adapted filtering. Several approaches
were tested as Butterworth low pass filtering,
Savitzky-Golay (S-G) derivative and smoothing
filter, and polynomial fitting (Savitzky and Golay,
1967).
(a)
(b)
Figure 2: Raw and filtered time-response signals of a gas
sensor (a) and their respective derivative curves (b) : Peak
apparition in the derivative curve.
The best results were obtained with S-G filter.
For each sensor, filter parameters (window width
and filter order) were adjusted whatever the used
concentration. Figure 2b underlines a notable
maximum of the derivative curve, obtained after
using an adequate filtering. This peak appears
generally in the 75 first seconds, and the height
value depends on the applied gas concentration and
the studied sensor.
In Figure 3 we present the derivative curves of
the 9 sensors for all the used concentrations. On this
figure the four concentrations are represented using
different colours. For the gas sensors (except MQ3
sensor), the peak height varies clearly with the
concentration. For MQ3 sensor, the superposition of
3% and 4% curves will be explained later.
Figure 3: Derivative curves (dV/dt) of each sensor versus
exposition time (s) along with the four EO concentrations.
2.2.2 Traditional Features
In most of electronic nose applications the
stabilization value of the sensor conductance is used.
To compare the Peak feature with this traditional
feature, we have determined the (Vs-V0) parameter
where Vs is the sensor response value at the end of
the exposure time and V0 the value of the initial
sensor level before the introduction of the EO
vapours.
Vs and V0 values are respectively calculated by
averaging five recorded data at the end and the
beginning of the sensor time-response signal, in
order to reduce the noise effects. The duration of V0
level is short (about 5 to 10 seconds according to the
sensor type) so 5 recorded data are used to average
the V0 value. Concerning Vs, this chosen average
gives satisfactory noise reducing.
In Figure 4 the time responses of all the sensors
for all the concentrations are drawn. We note that we
only obtain a good separation along with the
concentration for a few sensors (TGS2620, TGS880,
TGS816). The other sensors show high sensitivity to
the EO atmospheres than the three first cited sensors
with early sensor saturation. So, we see on the
corresponding graphs that the saturation occurs from
3% and even from 2% for the MQ3 sensor.
SensorsandFeaturesSelectionforRobustGasConcentrationEvaluation
239
Figure 4: Response signals (V) of each sensor versus
exposition time (s) along with the four EO concentrations.
2.2.3 Discussion
The choice of the sensors is predominant for a
reliable discrimination with electronic nose systems.
We have seen (Figure 4) that the saturation of the
sensor time-response occurs unfortunately for many
sensors, because of their high sensitivity to the
concerned effluent. So, for these sensors the
traditional parameter (Vs-V0) cannot well indicate
the concentration variation.
Concerning the “transient” parameter, Peak,
deduced from the derivative curve, the results
(Figure 3) show a better efficiency to discriminate
the concentration. In fact, this value is obtained
during the transient phase of the sensor time-
response (<75 seconds), then it is less influenced by
the saturation (excepted for MQ3 sensor).
So, these observations lead us to optimize our
detection system by reducing as much as possible
the gas exposition time. Consequently this reduction
might implicate the regeneration time reduction,
taking into account that these two phase times are
not linearly related.
This optimization is advantageous in two ways:
to reduce the measurement time and to improve the
efficiency of the traditional (Vs-V0) feature. This
approach will allow us to select the best sensors for
our real-time application.
3 MEASUREMENT
OPTIMIZATION
In this section we develop the optimization of the
measurement protocol, particularly important for
real time applications. After discussion about the
choice of the gas exposure and purge times, we
insist on the disparity between the sensor
behaviours. The study of these disparities permits us
to select the best sensors according to the optimized
measurement procedure and application.
3.1 Protocol Optimization
We know that measurement cycle has to be
composed of the gas exposure phase followed by the
sensor regeneration phase. In the considered
application, we need to determine the EO
concentration as quickly as possible, so one of our
goal was to reduce the times corresponding to the
measurement and regeneration phases with respect
of a good sensor regeneration.
So, several Exposure-Regeneration times were
tested. These experiments show us first that, even if
the exposure time becomes extremely short (for
example 60 seconds), the regeneration time remains
still very long (about 300s) to obtain a satisfactory
sensor layer cleaning. We have also noted that these
times are strongly related to the sensor type and of
course, for each sensor they depend on the used gas
concentration.
For each value of the studied exposition time,
several values of the regeneration time were applied
to control the sensor recovery. For an exposition
time less than 75s, the sensor time response does not
reach either the stabilization value, either the
inflexion point. So, it is impossible to determine a
reliable value of Peak (maximum of the derivative
curve). In contrary, an exposition time of 75 seconds
is convenient for all the sensors and most of the pine
EO concentrations. We have tested several
regeneration times for this exposition time. Figure 5
presents a set of cycles in the cases (a: 75s-150s) and
(b: 75s- 350s). In the case (a), all the graphs show an
important drift of the sensor initial values. The
sensor regenerations are not sufficient. In the case
(b), the regeneration is practically obtained for most
of the sensors. Other protocol (100s-500s) has given
practically the same results than the protocol (75s-
350s). This last cycle protocol is adopted for our
next investigation. This choice takes into account the
importance of a rapid and accurate measurement.
3.2 Sensor Selection
After adopting the measurement protocol, we looked
into the matter of the gas sensor selection. As we can
see on the Figure 5b, several sensors (TGS816,
TGS2620, SPAQ1, SPMW0 sensors) show a good
recovery into their initial conductance value after the
SENSORNETS2014-InternationalConferenceonSensorNetworks
240
(a) (b)
Figure 5: Set of repetitive Exposure-Regeneration cycles for all the sensors and 2% pine oil in the case of (a): 75s-150s and
(b) 75s-350s exposure and regeneration times.
regeneration phase. The other sensors present weak
or important drift, generally because they give a high
response to the EO atmospheres.
As we had characterized nine sensors, we
compared the recovery process of each sensor for all
the used EO concentrations. For this comparison we
have determined the Peak and the (Vs-V0) features.
The mean value and the corresponding standard
deviation are calculated from all the measurements
(8 repetitions), for each sensor and each EO
concentration. These values are plotted on the
Figure 6 for three representative sensors. We note
that the TGS2620 is the more appropriate for pine
EO concentrations discrimination: the values of
Peak and (Vs-V0) features show a very sensible rise
along the EO concentration with weak standard
deviations. But we can surprisingly see the
inefficiency of the SP31 sensor for this application.
Because of its high sensitivity to pine atmosphere,
the saturation occurs after 1% EO, represented by
abnormal evolution of the (Vs-V0) and Peak values
versus EO concentration. For the SPAQ1 sensor the
behaviour is intermediate, with a good variation of
Peak and a rather less efficient variation of (Vs-V0),
essentially higher than 3% EO concentration.
This comparison study leads us to detect three
qualities of sensors among our sensor array: very
good, good and non-adapted sensors for the
concerned protocol and application.
Very good: TGS 2620, TGS 880, SPMW0
Good: TGS 816, TGS822, SPAQ1, MQ138
Non-adapted: SP31, MQ3
(a) (b)
Figure 6: Feature evolutions of 3 gas sensors versus pine
EO concentrations (1, 2, 3, 4%); (a) Vs-V0, (b) Peak.
3.3 PCA Results
The measurements made for all the concentration
range (1, 2, 3, 4%) were analysed by Principal
Component Analysis (PCA) using as explicative
variables one of the two selected features (Peak or
(Vs-V0) of the nine sensors) separately. So, nine
principal components are obtained by linear
combinations of the original variables and
participate decreasingly to the construction of the
model. Figure 7 shows on the first two principal
components (PC1 and PC2) the loadings plots of
each of the two variable sets. A loading plot present
the correlation between the concerned variables, so
SensorsandFeaturesSelectionforRobustGasConcentrationEvaluation
241
(a) (b)
Figure 7: PCA loading plots; (a) Vs-V0 with PC1 explaining 88.9% of the variation and PC2 10.6%; (b) Peak with PC1
explaining 91.6% of the variation and PC2 5.5%.
all the representative points are positioned inside a
unit length circle, called “circle of correlation.
In our case, each variable characterises one of
the nine gas sensors. So, the loading plot, given by
PCA, provides a map of how the sensors relate to
each other. In this map, the more sensor projections
are closed together, the more they present similar
properties. Furthermore, the distance to the origin of
PC1 and PC2 also conveys information: the further
away from the plot origin a variable is located, the
stronger impact that variable has on the model with
respect of the EO concentration separation. The
more a variable is close to the origin of the plane,
the less important it is (Berna, Anderson and
Trowell, 2009; Jolliffe, 2002). In the same way,
since the PC1 explains the most important part of the
variation than PC2, this impact is stronger when the
variable is near to the unit length of PC1.
In Figure 7(a), where (Vs-V0) feature of each
sensor is used as representative variable, we can note
that SP31 and MQ3 sensors are situated far from the
unit length of the PC1. They are then less adapted
than the other sensors. This observation confirms the
previous result about the efficiency of these two
sensors. Other sensors of the array are positively
correlated and satisfy the condition of strong impact.
Considering Figure 7(b) where Peak is used as
representative feature, we can observe that the SP31
sensor becomes more efficient and joints other group
of sensor with high impact. But MQ3 sensor is
definitively less adapted for this study.
These PCA results confirm our sensor behaviour
study (section 3.2).
4 CONCLUSIONS
We have shown through this work that a deep
evaluation of the sensor behaviour according to the
studied atmosphere is required for reliable electronic
nose application such as gas quantification. Two
features extracted from the transient and the steady-
state phases of the sensor response signal (Peak: the
maximum of the derivative signal of sensor
response, and (Vs-V0): the response amplitude
voltage) were studied and compared. The
performance of the unusual Peak feature is
highlighted to provide fast and continuous
measurement. The capacity of this feature to
quantify pine oil vapour diffused in pure air has
permitted the optimization of the measurement time
conditions and also the selection of the best sensors.
In fact we have shown important disparities on the
stability and the performance of the chosen features
along with the sensor types. Loading plots obtained
with PCA confirm these results.
REFERENCES
Berna, A. Z., Anderson A. R., Trowell S. C., 2009. Bio-
Benchmarking of Electronic Nose Sensors. PLoS
ONE, 4(7): e6406;
[Online]; doi: 10.1371/journal.pone.0006406.
Boilot, P., Hines, E. L., Gongora, M. A., Folland, R. S.,
2002. Electronic noses inter-comparison, data fusion
and sensor selection in discrimination of standard fruit
solutions. In Sensors and Actuators B, 88, 80-88.
SENSORNETS2014-InternationalConferenceonSensorNetworks
242
Branca, A., Simonian, P., Ferrante, M., Novas, E., Martin
Negri R., 2003. Electronic based discrimination of a
perfumery compound in a fragrance. In Sensors and
Actuators B, 92, 222-227.
Cho, J. H., Kim, Y. W., Na, K. J., Jeon G. J., 2008.
Wireless electronic nose system for real-time
quantitative analysis of gas mixtures using micro-gas
sensor array and neuro-fuzzy network. In Sensors and
Actuators B, 134, 104-111.
Delpha, C., Siadat, M., Lumbreras, M., 2001. An
electronic nose using time reduced modelling
parameters for a reliable discrimination of Forane
134a. In Sensors and Actuators B, 77, 517-524.
Distante, C., Leo, M., Siciliano, P., Persaud, K.C., 2002.
On the study of feature extraction methods for an
electronic nose. In Sensors and Actuators B, 87, 274-
288.
Falasconi, M., Pardo,M., Sberveglieri, G., Ricco, I.,
Bresciani, A., 2005. The novel EOS
835
electronic nose
and data analysis for evaluating coffee ripening. In
Sensors and Actuators B, 110, 73-80.
Gualdron O., Llobet E., Brezmes J., Vilanova X., Correig,
X., 2004. Fast variable selection for gas sensing
applications. In Sensors, 2004. Proceeding of IEEE, 2,
892-895.
Gutierrez-Osuna R., 2002. Pattern analysis for machine
olfaction: A review. In IEEE Sensor Journal, 2, 189-
202.
Ionescu, R., Vancu, A., Tomescu, A., 2000. Time-
dependent humidity calibration for drift corrections in
electronic noses equipped with SnO2 gas sensors. In
Sensors and Actuators B, 69, 283-286.
Jolliffe, I.T., 2002. Principal Component Analysis, Second
edition. New York: Springer.
Llobet, E., Ionescu, R., Al-Khalifa, S., Brezmes, J.,
Vilanova, X., Correig, X., Barsan, N., Gardner, J.W.,
2002. Multicomponent gas mixture analysis using a
single tin oxide sensor and dynamic pattern
recognition. In Sensors Journal, IEEE, 1 (3), 207-213.
Martin Negri R. and Reich S., 2001. Identification of
pollutant gases and its concentrations with a
multisensory array. In Sensors and Actuators B, 75,
172-178.
Martinelli, E., Falconia, C., D’Amico A., Di Natale, C.,
2003. Feature Extraction of chemical sensors in phase
space. In Sensors and Actuators B, 95, 132–139.
Pardo, M. and Sberveglieri. G., 2007. Comparing the
performance of different features in sensor arrays. In
Sensors and Actuators B, 123, 437–443.
Paulsson, N., Larsson, E., Winquist, F., 2000. Extraction
and selection of parameters for evaluation of breath
alcohol measurement with an electronic nose. In
Sensors and Actuators B, 84, 187-197.
Roussel, S., Forberg, G., Grenier P., Bellon-Maurel, V.,
1999. Optimisation of electronic nose measurements.
partII : Influence of experimental parameters. In
Journal of Food Engineering, 39, 9-15.
Sambemana, H., Siadat, M., Lumbreras, M., 2010. Gas
sensing evaluation for the quantification of natural oil
diffusion. In Chemical Engineering Transaction, 23,
177-183.
Savitzky, A. and Golay, M.J.E., 1967. Smoothing and
Differentiation of Data by Simplified Least
SquaresProcedures. In Anal. Chem., 36 (8), 1627–
1639.
Sysoev, V. V., Goschnick, J., Schneider, T., Strelcov, E.,
Kolmakov, A., 2007. A gradient microarray electronic
based on percolating SnO2 nanowire sensing
elements. In Nano letters, 7 (10), 3182-3188.
Szczurek, A. and Maciejewska, A., 2012. Gas Sensor
Array with Broad Applicability, Sensor Array, Prof.
Wuqiang Yang (Ed.), ISBN: 978-953-51-0613-5,
InTech. Available from: http://www.intechopen.com/
books/sensor-array/gas-sensor-array-with-broad-
applicability.
Zhang, S., Xie, C., Zeng, D., Zhang, Q., Li, H., Bi, Z.,
2007. A feature extraction method and a sampling
system for fast recognition of flammable liquids with a
portable E-nose. In Sensors and Actuators B, 124,
437-443.
Zhang, H. and Wang, J., 2007. Detection of age and insect
damage incurred by wheat, with an electronic nose. In
Journal of Stored Products Research, 43, 489-495.
SensorsandFeaturesSelectionforRobustGasConcentrationEvaluation
243