Escolar Documentos
Profissional Documentos
Cultura Documentos
Facultat de Física
Departament d’Electrònica
Autora: Director:
Marta Padilla Ferran Dr. Santiago Marco Colás
Son muchas las personas a las que quiero agradecer por su influencia positiva en mi
vida durante estos largos años de tesis. No puedo nombrarlas a todas aquí y les pido
disculpas.
En primer lugar muchas gracias al Dr. Santiago Marco, director de mi tesis, que me
ha guiado, enseñado y ayudado para que, a pesar de las dificultades, este proyecto
saliera (finalmente!) adelante.
Muchas gracias también a los seniors, los doctores del grupo; Toni, Iván, Agus, Ale,
Manolo y más recientemente Dani y Eduard, algunos de los cuales ya no están en
el grupo aunque seguiremos en contacto. No me olvido de Xavi, uno de los seniors,
muchas gracias a ti también. Mis compañeros que (ya!) son doctores: Jordi y Rafa,
mis colegas: Sergi, Erola, Luis, Ana, Víctor, Didier, Miguel, Miquel, Fran, Idoya, y
los que se fueron: Benja, Ali, y Aina. Muchas gracias a todos por vuestra ayuda y
ánimos!!.
Aparte de la gente fantástica que he conocido durante mi tesis, agradezco los proyec-
tos que la hicieron posible: BioPatAna, Neurochem, Lotus, y también a la General-
itat por los fondos para poder disfrutar de mi estancia en Bélgica.
Y por último, aunque no menos importante, agradezco a mis padres su apoyo in-
condicional, al igual que a mi hermana Bárbara, a mi familia; Chelo, Toni, Joaqui, a
mis principales amigos; Ceci, Bel, Lucía, Ani, Patricia, Paco ..... a todos muchísimas
gracias por vuestros ánimos y por estar conmigo en los tiempos menos buenos!.
iii
Marta Padilla Ferran
Todo es mejorable.
Santi.
iv
Abstract
Information about the environment is essential for the survival of every living being,
since it determines the way to react with respect to external inputs. In the case of
humans, this information is collected through the five senses; sight, hearing, touch,
smell and taste. Sight, hearing and smell senses are considered specially interesting
because of their ability to get external information without direct interaction with
the sources. Sight is probably the most developed sense in humans, followed closely
by hearing. On the contrary, smell has always been considered the most primitive
and less important of those three senses, and was an almost total mystery until a
few years ago. Recently, in 2004, Dr. Richard Axel and Dr. Linda Buck shared the
Nobel Prize in medicine for their discoveries of odorant receptors and the organization
of the olfactory system, since then, studies on the smell and human olfactory system
have received a new interest.
With the aim of extending our senses abilities of obtaining information about our
surroundings, some devices have been developed which are inspired by the biologi-
cal mechanisms of the senses. Examples of such devices are well known and of very
extended use, like video cameras used for security issues in which image recognition
can be implemented (sight), or sound recording with speech recognition (hearing).
Relating smell, an instrument named electronic nose was proposed in 1982 by Per-
saud and Dodd [1] to differentiate odours. Electronic noses were very promising for
many qualitative and quantitative applications, since they were expected to provide
characteristics such as being of small size, low cost, fast and easy to use. These
features are specially interesting for on-field applications, compared to other well-
established instruments for gas/volatiles analysis which are big, heavy, expensive
and difficult to use, though they provide better chemical resolution.
Despite the many potential advantages of the use of electronic noses, nowadays,
more than 25 years after the first device, this instrument is not massively present
on the market. The main reason lies in the sensing area of the instrument, which
exhibits poor selectivity and bad stability. The chemical gas sensors used in elec-
tronic noses present problems like cross sensitivities, time instability, dependence
on previous gas exposures, etc. Therefore, instruments based on these sensors are
not robust and do not give enough reproducible results. The nature of the prob-
lems that influence chemical gas sensors is mainly technological, but affect sensors
of all state of the art technologies, though to different degrees. These deficiencies
can be mostly overcome as more research is made on improving fabrication process
or developing new technologies. However, while gas sensors technologies are being
improved, statistical signal processing can help to mathematically compensate, or
at least to reduce, the effect those mentioned issues have on the sensors responses
before pattern recognition is carried out.
The aim of this thesis is to explore the robustness of some sensor operation meth-
ods, and to propose the use of statistical signal processing techniques to correct
or compensate sensors responses affected by specific problems, such as sensor drift
v
Marta Padilla Ferran
and failure of one or more sensors in the array. This document is organized in five
chapters. In the first one, the concepts of machine olfaction and electronic noses
are revised, as well as the main problems that the instrument presents. Specially,
evidence of specific issues studied in this thesis and state of the art on solutions by
signal processing techniques are presented in this chapter. Concepts and definitions
of robustness are given in chapter two. In the third chapter, the aim of this disser-
tation is detailed. Then, chapter fourth explains the work made and presents the
papers which shows many, but not all, of the results obtained during these years of
work. Finally, conclusions are given in chapter five.
vi
Contents
1 Introduction 1
1.1 Electronic noses . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.1.1 Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.1.2 State-of-the-Art and future trends . . . . . . . . . . . . . . . 5
1.2 The electronic nose instrument . . . . . . . . . . . . . . . . . . . . . 6
1.2.1 Instrument general requirements . . . . . . . . . . . . . . . . 6
1.2.2 The sampling unit and delivery system . . . . . . . . . . . . 8
1.2.3 The sensors chamber . . . . . . . . . . . . . . . . . . . . . . 9
1.3 Gas sensors technologies . . . . . . . . . . . . . . . . . . . . . . . . 11
1.3.1 Metal Oxide Semiconductor sensors (MOX) . . . . . . . . . 13
1.3.2 Conductive Polymer Chemoresistors (CP) . . . . . . . . . . 16
1.4 Pattern recognition system . . . . . . . . . . . . . . . . . . . . . . . 17
1.5 Operation Modes . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
1.5.1 Sampling Transient Analysis . . . . . . . . . . . . . . . . . . 19
1.5.2 Temperature-Modulation Analysis . . . . . . . . . . . . . . . 22
1.6 Drift and sensor faults in a gas sensors array . . . . . . . . . . . . . 24
1.6.1 Characteristics of drift and sensor poisoning . . . . . . . . . 24
1.6.2 Factors involved in measurements instability . . . . . . . . . 25
1.6.3 Effects of drift and poisoning on the sensors response . . . . 26
1.6.4 Evidence of drift and faulty sensors in electronic nose appli-
cations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
1.6.5 Drift correction and fault detection by signal processing tech-
niques . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
2 Robustness 41
2.1 Robustness . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
2.1.1 Definition of Robustness and Ruggedness . . . . . . . . . . . 43
2.1.2 Robustness and Ruggedness tests . . . . . . . . . . . . . . . 44
2.1.3 Robustness of calibration models . . . . . . . . . . . . . . . 45
2.2 Signal processing methods for improving robustness . . . . . . . . . 47
2.2.1 Orthogonal Projection (OP) methods . . . . . . . . . . . . . 47
2.2.2 Multivariate Curve Resolution (MCR) methods . . . . . . . 50
2.2.3 Multiway (N-way) methods . . . . . . . . . . . . . . . . . . 53
2.2.4 Calibration Transfer . . . . . . . . . . . . . . . . . . . . . . 55
2.2.5 Fault detection and isolation (FDI) methods . . . . . . . . . 58
vii
Contents Marta Padilla Ferran
3 Thesis objectives 61
5 Conclusions 113
References 129
viii
Chapter 1
Introduction
In humans, the sense of smell gives place not only to instinctive responses but
also very variable and individual subjective ones. Initially, artificial olfaction (AO)
systems were thought to give a human odour impression, thus to mimic olfaction.
However, nowadays it has been shown that the sense of smell is too complex to be
imitated by an artificial instrument with the actual technology. This reason, and
problems derived from sensors technologies have made that electronic noses have
not reached the initially expected success.
In this chapter, artificial olfaction systems, or electronic noses, are described. Its
history is briefly revised, and also its future trends, state of the art and main ap-
plications. The terms artificial olfaction (AO) systems and electronic noses (e-nose)
are used indistinctly along the text. Then, the sensors technologies used in this
thesis are described, as well as strategies to extract maximum information from the
sensors-analyte interactions. Finally, a deeper insight is made on the main problems
that affect the chemical gas sensors; characteristics, evidence and state of the art in
counteraction by signal processing methods.
1
Introduction Marta Padilla Ferran
recognition unit to process sensors outputs and make a final decision, as the brain
does. Such instrument,electronic nose was proposed in 1982 by Persaud and Dodd
[1] to differentiate several gas compounds. This first concept of e-nose was based
on an array of three gas sensors with partial selectivity together with a processing
of the sensors response with a neural network, which discriminated among different
compounds the sensor array was exposed to. In that work, the authors introduced
for the first time the idea of a device that could mimic the discrimination ability of
the mammalian sense of smell.
The term electronic nose became popular some years later, when researchers ex-
pected to develop a device capable of recognizing odours and characterize them
with attribute descriptors such fruity, grassy, earthy, malty, etc. , just like the hu-
man sense of smell does. These human perception of odours would be of high interest
in the industry of aroma, perfumes, wines, food, etc.
Electronic noses were also designed to be an alternative to the use of other instru-
ments for gas analysis or detection, for example in an on-field application where
some gas compounds have to be differentiated. On the one hand, the objective of
gas analysers is to accurately measure and analyse the content of any pure compo-
nents in a given atmosphere. They can be costly, complex, big, heavy, difficult to
use, and deliver an outcome after a long time, but, on the contrary, results provide
a lot of information about the sample composition. A gas chromatographer/mass
Spectrometer (GC/MS) is a gas analyser indispensable and irreplaceable in many
applications. On the other hand, gas detectors are designed to fast detect a target
gas and then fire an alarm, they must be cheap and very fast in response. Ex-
amples of these devices are CO detectors in houses, also essential in those specific
applications. Electronic noses intend to be small size devices, low cost, low power
consumption, easy to use, portable and fast both for qualitative and quantitative
applications, requiring minimal or even absent sample preparation like prior separa-
tion in simpler compounds. Since these instruments were also expected to contribute
with some particular abilities coming from the emulation of the human olfaction,
they would offer an alternative technology for important applications that are not
covered by the mentioned (and others) well known gas instruments. The electronic
nose would complement those instruments, even sometimes its capabilities would
overlap with those of the other devices.
Besides, odour evaluation has traditionally been performed by human sensory panels,
whose responses are influenced by human problems like fatigue, illness, mood, etc.
The electronic nose was even expected to be, in the future, a complete substitute
for reference methods that use human sensory panels, since there would be a high
decrease in the cost and arbitrariness in the results of the panels.
During the nineties, given all these expected characteristics, the electronic nose was
the most promising among new analytical methodologies for objective odour evalu-
ation. A formal definition of the instrument was given by Gardner and Bartlett [2]:
an instrument, which comprises an array of electronic chemical sensors with par-
2
Electronic noses
However, a few years later, research on the capabilities of the electronic nose made
that some authors started to question the name given to the device; since an elec-
tronic nose is obviously electronic but not a nose [3]. And recently, another definition
has been given: an attempt to mimic the principles of smelling that gives another
view on the whole scene of volatiles compared to its biological inspiration [4].
After more than 25 years of research, it has become evident that, given the complex-
ity of the biological olfactory system and sense of smell, an electronic nose cannot
mimic olfaction. However, the term electronic nose is still in use, although the only
common characteristic between the instrument and the olfactory system is the par-
tial specificity of the sensors in the array and the presence of a signal processing unit.
Further, there is another point in which biological olfaction and electronic nose are
different. It is the fact that the mammalian olfaction is a general purpose instru-
ment, while an electronic nose has to be specially designed for a given application
due to technological limitations.
Today, once accepted that the initial expectations about mimicking the sense of
smell were not realistic, a change in the point of view of the problem has been given;
the actual approach is to design an instrument inspired by nature, not a replication
of it. The term used for this philosophy is biomimetics or bioinspiration.
Apart from the underestimation of the problem of the imitation of a biological sys-
tem, the problems derived from the sensor technologies have also contributed to the
small presence of the e-nose in industry. There exist many gas sensing technologies,
but all of them suffer from general problems that affect the reproducibility and relia-
bility of the final instrument, and thus make these devices not suitable for industry.
A great effort is being made for improving sensor technologies to overcome these
problems. Nevertheless, also signal processing techniques can help to mathemati-
cally compensate the degradation of the sensors responses and thus, to improve the
performance of the e-noses.
In summary, without the hope of that the instrument can mimic biological olfaction,
today the adopted approach is to design e-nose or AO instruments inspired from
biology and targeted to specific applications. In those applications these instruments
should fulfil the needs that other devices cannot cover, like low cost, portability, low
power consumption, simplicity of use, etc. Effort is being made on the improvement
of these characteristics and others related to the improvement of robustness and
performance.
3
Introduction Marta Padilla Ferran
1.1.1 Applications
For example, there are many references that study the quality of food products that
are often characterised by its smell (and taste) such as coffee [5], tea [6], olive oil
[7] and wine [8]. In these and other food products, the aim of other studies is to
differentiate the origin of the samples [9], to detect fermentation [10] and roasting
[11], which would affect the taste and smell of the product, to detect spoilage [12],
the presence of harmful substances [13], the presence of fungi [14], changes during
storage and transport [15] and defects in the production and packaging process [16].
Other applications are related to the security field, where hazardous substances [17]
and explosives are to be detected at very low concentrations [18, 19]. Environmental
monitoring applications include the characterization of the pollutant level of ambient
outdoor [20] and indoor air [21]. Specifically the detection of toxic [22] and exhaust
gases [23], leakages of certain gases [24] or the control of the ventilation of an indoor
atmosphere, determination of the odours emissions from animal production facilities,
malodorous produced through industrial factories and from decomposition process
of wastes in compost plants [25, 26]. Also the detection of microorganisms and
residues of insecticides in the headspace of drinkable water have been reported [27].
In disease diagnostics, volatiles have been detected from the skin, the sputum [28],
the urine, the stool, or the breath that have given indications of certain allergies
like asthma [29] and illnesses such as lung cancer [30], brain cancer [31], tuberculosis
[28], several infections in nose, throat, ear [32], urinary tract, vagina and skin, also
diabetes, halitosis [33], sinusitis [34], renal insufficiency, etc. However, given the
variability of human beings and the influence of lifestyle (diet, exercise, smoking,
etc. ), more research have to be done in health applications.
Very good reviews on the applications of AO systems can be found in [35, 36], and
more recently [4, 37], namely, applications on food control [38–40], dairy products
[41], medical applications [42] and disease diagnostics [43–45].
4
Electronic noses
Many of the above mentioned applications are carried out in research laboratories
and others require highly optimised systems for specific operations, but real-life
applications are still difficult to cover. In such applications, the high number of
interferents may hide the target compound and thus it is needed an increase in the
sensitivity and selectivity in the system. For this, several sensor technologies are
combined in the same instrument to provide complementary information, these are
hybrid e-noses. However, these new devices obtain better selectivity range at the
expense of an increasing of the set-up complexity and costs.
For instance, some new approaches include modules like mass spectrometers [46]
or gas chromatographers [47, 48]. Furthermore, new instrumental technologies are
sometimes considered as e-nose instruments, like ion mobility spectrometers (IMS)
and fast gas chromatographers [4]. Nevertheless, the design of a commercial e-nose
device depends on the target application or applications, since an ideal all-purpose
instrument is not possible nowadays due to technology limitations.
Actually, research is focused on new sensing materials, which provide new useful
information sometimes specific for a given type of target analytes, such as bio and
cell-based biosensors which show high selectivity especially for certain organic com-
pounds [49, 50], for example biological olfactory receptors [51]. Research is also
made about other issues aimed to improve the device performance such as ways
to increase the number of sensors in a system, micromachining to reduce power
consumption [52–55], optimisation modelling, etc. (fig. 1.1).
Rock et al. [4] give a complete overview of the current state of the art as well as
5
Introduction Marta Padilla Ferran
future trends in electronic noses. They also present a table containing commercial
electronic noses on the market in 2008.
Next, every individual part of the e-nose instrument is described in order to em-
phasise the factors that are important for obtaining robust measurements. In 1.3 a
more detailed description about the sensors technologies used in this thesis is given,
with emphasis in the transduction mechanisms and characteristics that lead to their
particular advantages and disadvantages.
As already mentioned, the most evident requirements for an electronic nose are a
low price, high portability and quick response. The sensors are responsible for such
requirements, since they are the heart of the device. Nevertheless, the remaining
parts of the instrument, such as sensors chamber and flow system, must not de-
grade the performance of the sensors, for this reason the whole instrument must be
carefully designed in all its parts (see sections 1.2.2 and 1.2.3).
Not all the desired properties of the sensors: high sensitivity to target volatiles,
low sensitivity to environmental variables (like temperature and humidity), fast
response and recovery times, robustness, time stability, reproducibility, small size
and low power consumption, can be met in the same device. Therefore the ideal
sensors characteristics must be evaluated and prioritised, sometimes establishing
compromises among several of them, according to the application the instrument is
going to be used for.
From the definition and philosophy of the electronic nose, it is obviously desired
that sensors response and recovery time are short. Short recovery time would mean
(almost) perfect reversibility of the physiochemical detection or sensing process, and
thus long sensor life and the possibility of making many measurements in a short
time. High sensitivity or low detection limits are also required in many applica-
tions. Usually, gas sensors reach ppm to ppb detection limits. Linearity is another
interesting property, it is usually accomplished for low concentrations of analyte,
but as concentrations increases non-linearity becomes more evident until a point of
saturation. But it has to be taken into account that high analyte concentration
levels also influences sensors response and recovery times, which may become much
longer and may be not reversible. Large dynamic range, so that the sensors are
able to respond from very low to very high concentrations with similar sensitivity
and no saturation, is also highly desired. Other nice property include the possibility
to obtain high selectivity or specificity to a target analyte for given applications.
6
The electronic nose instrument
This property is also convenient for general applications, since it has been found
that the performance of sensor array can be improved by using a combination of
broad selective and high selective sensors [56–58]. Finally, low cross-sensitivity to
interferents, and long-term stability of the sensor and sensing material are desired
as well [59].
Some of these properties, such us sensor sensitivity, selectivity, speed of response, and
reversibility, are determined by the thermodynamics and kinetics of sensor material-
analyte interactions. Desired requirements on these characteristics can impose con-
tradictory constraints on the sensor design, since the accomplishment of a given
requirement may imply difficulties to fulfil other. For instance, high sensitivity and
selectivity on the one hand are typically associated with strong interactions, whereas
perfect reversibility on the other hand requires weak interactions [59]. Consequently,
a trade-off among these properties is necessary. However, selectivity and sensitivity
can be in turn improved by the signal processing of some of the sensors individual
features, as will be explained in sections 1.3 and 1.5.
Gas sensor devices often present lack of reproducibility in their responses. This
makes difficult to establish a standard for measurements methodology and also de-
creases the robustness of the instrument. For this reason, when a faulty sensor has
to be replaced by a new one, a recalibration has to be performed, although both
sensors are supposed to be identical. This loss of calibration is mostly due to the
variability of the response among sensors of the same type. The fabrication pro-
cesses are the responsible of this unwanted property and it is common to different
transducing technologies. The only solution is a new calibration of the system.
Another important problem affecting chemical gas sensors, of all types but to differ-
ent degrees, is drift. Drift is a random temporal variation of the individual sensor
response when the sensor array is exposed to the same analyte under identical con-
ditions. It is due to several causes related to the degradation of the sensor’s sensing
material and other long-term effects. Environmental agents such as changes in hu-
midity, temperatures and flow rates, are also included by some authors into the
causes of drift, although they are better described as cross-interferences.
Drift and related problems lead to periodical re-calibrations of the device that are
costly and time consuming and which have to be minimised. An attempt to com-
pensate the degrading sensor response, especially due to drift and sensor poisoning,
by means of signal processing tools is one of the main objective of this work and
will be explained in more detail in section 1.6.
7
Introduction Marta Padilla Ferran
For electronic noses that need odour handling and delivery systems, several tech-
niques are available: sample flow, static, direct exposure to the gas and pre-
concentrator systems [60]. The main one is the sample flow, in which the sensors are
placed in the gas flow, allowing the rapid exchange of gases and hence shortening
the time of measure.
In the sample flow system, a flow control system is used to deliver the sample to the
sensor chamber. In general, the flow system consists of a pneumatic pump or a gas
pressure source with a mass flow controller (MFC) to create a constant gas flow rate.
Pipes are used to drive the gas towards the sensor array in the chamber. A MFC
precisely controls the flow rate independently of its pressure load, therefore this is
a suitable method for measurements of gaseous samples coming from gas cylinders.
The pipes can be made of Teflon or stainless steel to avoid chemical reactions between
them and the volatiles. Furthermore, the pipes and even the sensor chamber can be
heated to avoid the condensation of the volatiles on the walls, which would change
the concentration of the volatiles that arrives to the sensors.
The sample flow systems include methods such as headspace sampling, diffusion,
permeation tubes, bubblers, and sampling bags [60]. Gaseous samples can enter the
flow system by diffusion or be kept inside a tedlar bag for transportation from a
given place. Also for transportation of a gaseous sample, a technique based on the
pre-concentration of the volatiles in a sorbent trap containing tenax, hidrocarbon,
etc. which absorbs them, can be used. Later, these volatiles are desorbed by heating
the pre-concentrator, which requires an additional hardware. The advantage of this
method is that a great amount of volatiles can be captured in the pre-concentrator,
and thus enhance the selectivity of the electronic nose device.
On the other hand, an electronic nose can test solid or liquid samples from which
volatiles are evaporated. In these cases they are usually kept into a closed bottle or
vial, which has a mechanism in the cover to allow extraction of the emitted volatiles.
The headspace is the space just above the liquid or solid sample in the vial, which
gives the name to this sampling method. Inside the vial, when the equilibrium
between the sample phases is reached, the volatiles filling the headspace can be
extracted. The time this equilibrium is reached depends on the sample and on the
temperature through the vapour pressure of the volatile compounds, which must be
kept under the saturated point to avoid condensation into liquid drops. Therefore,
8
The electronic nose instrument
To drive the generated volatiles from the headspace to the sensors chamber, a system
based on two needles, which perforate the septum of every vial, is generally used
in automatic samplers. One of them connects the vial to the sensors chamber and,
through the other one, a carrier gas such as dry or external air, enters into the vial
to carry the volatiles towards the sensors chamber at a constant (controlled) flow
rate.
It has to be noted that the volatile concentration at the outlet of the vial gradually
changes its profile by dilution into the carrier gas, and this sometimes influences the
waveform of the sensor response. The volatile pulse supplied to the sensor must be
sufficiently short so that the concentration profile keeps constant, and sufficiently
large so that the sensors are exposed to the analyte for a sufficient residence time
to absorb it. Therefore, to avoid the influence of the concentration profile on the
sensor response, parameters such as temperature and flow rate must be carefully
selected [60].
Moreover, the sensors response must not be influenced by the flow system. Therefore,
all gas switching processes must be fast in comparison to the analyte diffusion and
reaction dynamics, and all the gas dead volumes in the flow system have to be
minimized. Also, the time span between the entrance of the gas sample into the
flow circuit and the moment at which the gas reaches the sensor, should be as short
as possible [59].
The design of the sensor chamber is an important issue that influences the sensors
response, although its importance have often been underestimated. Several studies
on the of the airflow patterns through human nasal cavity have been carried out
without practical application or transfer to for electronic noses (yet) due to its com-
plex anatomy [61, 62]. However, other studies have shown that an adequate design
of this part of the instrument would result in an improvement of its performance,
by reducing important sensors drawbacks such as variability among responses of
sensors of the main type or measurements repeatability.
It can be clearly seen that the chamber geometry greatly affects the sample concen-
tration inside through parameters like the chamber volume, shape and the sensors
and inlet and outlet position [63, 64]. These parameters play an important role in
the development of the fluid flow and the gas concentration fields inside the cham-
ber. Thus, they strongly influence the sensors response which become dependant of
9
Introduction Marta Padilla Ferran
its location inside the chamber. Indeed, an incorrect chamber design may lead to
high sensitivity to slight changes in the flow rates that would directly influence sen-
sors response characteristics such as stability, measurement reproducibility, response
time and amplitude [65].
The design of the chamber must guarantee that the same volatile concentration
arrives to every sensor’s surface by a suitable geometry and sensors arrangement
inside the chamber. At least this should not vary with input flow. Its geometry must
assure that there are no stagnant volumes or recirculating zones within the operating
flow rate range, and that the flow is homogeneous, so that the sensors have identical
exposure conditions. Also, a minimal chamber volume allows for rapid refresh rate,
minimize any effect due to the sensor location and keeps the volatile concentration
constant during the response time of the sensors [60]. Moreover, the condensation
or absorption in the walls must be avoided by the choice of a suitable material
[66]. In an ideal sensor chamber, at every sensor position it would be possible to
accurately reproduce the shape of the input concentration signal and obtain the
same concentration profile in repeated measurements. This would therefore result
in an improvement of measurements repeatability and a decrease in the variability
in the response of sensors of the same type at different positions, since the system
would be less dependant on flow rates.
Several studies about the sensors chamber characteristics have been made in order to
optimise the sensors response. These studies are based in dynamic flow simulations
[63, 66] or comparison of different designs of sensors chambers [66, 67]. In these
works, several sensors chambers are designed to fulfil the desired characteristics as
much as possible. We can find a radially symmetric chamber [65], diffusion grids
[63, 66], a planar chamber with a narrow flow channel [67] or a circular chamber
with a parallel arrangement of the sensors at 0o with respect to the flow direction
[66]. Also, Gmur et al. [68] performed measurements and finite-element model cal-
culations of the heat dissipation and gas fluidics, in order to design an optimum
packaging that optimises the power consumption and performance of a microar-
ray of tin oxide sensors. Figure 1.2 shows a computational fluid dynamics (CFD)
simulation of concentration of an introduced analyte after one second. The lighter
shading represents higher concentration of the analyte. It is interesting to see how
the chamber design produces different analyte concentration arriving to the sensors
surface depending on their position.
Regarding the material, the sensors chamber is usually made of stainless steel but
polytetrafluoroethylene (PTFE) is also recommended. It must be a material chem-
ically inert and non-adsorbent to avoid memory effects. Stainless steel has good
thermal conductivity, is tough, inert and non-absorbent but hard to machine. Be-
sides, PTFE is reasonably tough, flexible, useful up to a maximum temperature
of 260o C, has poor thermal conductivity and is inert, but it is microporous, so it
is absorbent and must be properly desorpted from any odorous materials between
samples [66].
10
Gas sensors technologies
Figure 1.2: Analyte concentration; a) rectangular chamber with sensors at 90o to flow
direction; b) semicircle chamber with sensors at 90o ; c) semicircle chamber with sensors at
0o ; d) semicircle chamber with sensors at 0o with diffused flow [66].
Inside the sensor chamber, a number of non-specific sensors with broad and par-
tially overlapping selectivity are placed building an array of sensors. When the
array is exposed to a given compound or mixture, the response of all the sensors
as a whole forms a characteristic pattern, an electronic fingerprint, from which the
given mixture or compound can be differentiated among other samples by the signal
processing unit (fig. 1.3).
The arrays of gas sensors can be formed by sensors of the same technology with
different characteristics or by sensors of different technologies. It is known that the
combination of several technologies in a device gives complementary information
about the sample, which therefore results in better characterization of the volatiles.
However, the combination of several types of sensors makes the device be more
complex and, consequently, more expensive. Furthermore, there is a limit to the
useful size of the array, since by increasing the dimensions of the array the noise
may be amplified, e. g. by sensitivity towards unimportant information, instead of
obtaining new information about the gaseous composition [4].
11
Introduction Marta Padilla Ferran
such as mass spectrometers, ion mobility spectrometers, infrared analysers, fast gas
chromatographers or optical sensors, complement the classical chemical sensors and
sometimes they are also named e-noses.
The chemical gas sensors technologies used in this thesis, metal oxide (MOX) and
conductive polymer (CP) sensors, are revised regarding their transduction mecha-
nism and characteristics to derive their particular advantages, disadvantages and at-
tempted technologies improvements. Excellent revisions on other technologies such
as acoustic wave devices (quartz crystal microbalances (QCM) and surface acoustic
wave (SAW), fig. 1.4), sensors based on field effect transistors (FET) (metal ox-
ide FET (MOSFET)(fig. 1.5(a)), ion selective FET (ISFET), etc. ), electrochemical
sensors, biosensors, potentiometric sensors, optical based sensors (surface plasmon
resonance (SPR), colorimetric, optical fiber (fig. 1.5(b)), etc. ), etc. can be found in
literature in AO systems reviews [4, 59] and books [70, 71], or specific reviews on
sensor technologies [72–85].
12
Gas sensors technologies
These type of sensors are probably the most widely used in electronic nose applica-
tions due to their low price and reasonably good characteristics related to lifetime,
sensitivity. Furthermore, they are available commercially in a variety of different
types with broadly different specificity. However, they exhibit a number of disad-
vantages that will be referred later in this section.
The use of a metal oxide semiconductor as a gas sensor was first proposed in 1962
by Seiyama [87], and hundreds of works to applications in gas sensing with these
sensors have appeared since then. The sensing mechanism involved in MOx sensors
is well known for simple gas molecules and some theories have been developped [36].
One of the first empirical mathematical expressions was given by the power law by
Clifford and Tuma [88]. This formula reveals that MOXs sensors response is not
linear with the gas concentration:
R
= (1 + Kgas · Cgas )−β (1.1)
Ro
where R is the sensor resistance, Ro is the sensor resistance in air, Cgas is the
concentration of the target analyte, β is the power law characteristic of the particular
sensor and the proportionality constant Kgas depends of the analyte [88, 89].
Normally, MOX sensors are composed of a substrate and a heating element, coated
with a sensing material between two electrodes (fig. 1.8). Their response is based
on a change on the conductivity of the sensing material when it is exposed to com-
pounds which induce an oxidizing or reducing reaction on the sensing layer. The
sensing material consist of a semiconductor of n-type or p-type. The first group,
13
Introduction Marta Padilla Ferran
n-type semiconductors such as SnO2 , ZnO, Fe2 O3 , TiO2 and WO3 , respond mainly
to reducing compounds such as H2 , CH4 , CO, C2 H5 , or H2 S by increasing their
conductivity. On the contrary, p-type semiconductors such as CuO, NiO or CoO,
respond mainly to oxidising compounds such as O2 , NO2 , and Cl2 . An advantage
of MOX sensors is that the characteristics of these semiconductors can be tuned in
several ways, such as changing the sensing film thickness, doping the semiconductor
with Pd or Pt metal catalysts and impurities, changing the operational temperature
or via the grain size of the material, in order to modify the sensitivity and selectivity
of the final sensor.
MOx SnO2 sensors respond faster than other types of sensors [84]. The MOXs’ re-
sponse and recovery times depend also on the the target gas and its concentration,
and on temperature, since high temperatures increase the speed of physic and chem-
ical reactions. The sensor sensitivity to certain analytes can be improved at high
temperatures, greater than 300o C, to increase the reactivity of the semiconductor
surfaces. But it results in high power consumption, which is one of the disadvantages
of these sensors, since it reduces the instrument applicability to portable systems.
The change in the sensor sensitivity with temperature can be used to extract more
information from a sensor by a thermal cycling technique. This extra information
have been reported to improve either the sensors selectivity and sensitivity [90].
Examples of the sensitivity dependence on temperature for several types of MOX
sensors and analytes, can be seen in figures 1.6 and 1.7 [91, 92].
Regarding disadvantages, apart from a high power consumption, MOX sensors suffer
from poisoning due to irreversible binding of compounds that contain sulphur or
weak acids [93, 94] to the sensor oxide [80] and ethanol can also blind the sensor
from other volatiles [83, 95] increasing their detection threshold. Slow base-line
recovery when exposed to high molecular weight compounds has also been reported.
Another potentially serious problem is the sensitivity of these devices to humidity,
which have been addressed by combining MOX sensors with a humidity sensor or by
increasing the sensor temperature of operation. Finally, other two drawbacks have to
be added; the high working temperature that makes MOX sensors inappropriate in
environments containing flammable chemicals and the poor level of sensor to sensor
reproducibility [96].
14
Gas sensors technologies
Figure 1.6: Sensitivity-temperature profile for SnO2 , Pt-, Pd- and Ag-doped SnO2 sensors
[91].
Figure 1.7: conductance-temperature response of a SnO2 gas sensor in (a) air, (b)
methane, (c) ethane, (d) propane, (e) n-butane, (f ) isobutene, (g) ethylene, (h) propy-
lene, and (i) carbon monoxide [92].
More information can be obtained from previously mentioned reviews and also [98–
100].
15
Introduction Marta Padilla Ferran
16
Pattern recognition system
composite causes an increase in the electrical resistance because the the number
of conducting pathways for charge carriers is reduced. Some works have reported
linear responses to concentration for various analytes and also good repeatability
after several exposures [83].
The use of CP sensors possess several advantages. First of all, a great range of
polymers is available in the market, and can be easily tailored towards the detection
of specific groups of analytes by the use of different functional groups or monomers
[93]. Another nice feature of these sensors is that they operate at room tempera-
ture, which leads to simpler required system electronics and low power consumption.
They are also relatively resistant to poisoning, show broad selectivity and reversible
responses. The reported response times vary considerably from seconds to minutes
[83], generally their present faster response and recovery times for polar compounds
and larger responses to non-polar species. Their fabrication method is very flexible
but has a considerable influence over the surface morphology and other critical pa-
rameters that prevents the sensors to be batch-to-batch reproducible [102], although
miniaturization and mass production are possible [84, 93]. On the other hand, al-
though conducting polymers show a good response to a wide range of volatiles, they
also present some difficulties in resolving certain types of analytes. This last fact is
due to the relatively low diversity in affinity of the polymers toward a diverse set
of analytes [84]. Other problem is the high sensitivity to humidity of the polymer
response, which results in the need of sensor operation in a controlled and condi-
tioned, environment. CP can also drift with time and their lifetime is quite short,
typically 9-18 months, due to oxidation of the polymer material [83].
The signal processing unit in the electronic nose analyses the pattern formed by the
sensors response in order to extract relevant information. There are many techniques
that can be used in this stage, coming from a variety of fields that range from
chemometrics, machine learning to pattern recognition, only some used in this thesis
are explained.
Excellent reviews on signal processing tools used in the field of AO systems can be
found in [103, 104].
17
Introduction Marta Padilla Ferran
Figure 1.9: Signal processing steps for pattern analysis in a smart sensor array [103].
Diverse designs of individual parts of the AO instrument have been briefly described
above. These different designs give place to systems that vary in cost and instrument
complexity. In turn, cost and complexity must be optimised in a compromise with
the desired performance for a given application. According to the variety of designs
of possible e-nose instruments, we can distinguish different operation modes of these
systems.
AO systems can consist on one or few more gas sensors directly exposed to the
(static) ambient atmosphere, this is the case of detectors and some e-noses [105].
In these systems, the sensors are continuously providing an output without inter-
mediate cleaning stages with clean air or a reference gas. Detectors are very simple
and cheap since they do not need any sensors chamber, sampling unit and delivery
system and also require simple signal processing (fig. 1.10). They are often used
for detecting a given gaseous analyte, such as carbon monoxide or methane in an
indoor environment.
Other AO systems include a delivery system to drive the target gas towards a sensors
chamber at a constant flow. Some instruments even allow automatic dilution of the
sample prior to exposure to the sensors, if high concentrations are detected [106]. In
AO instruments with delivery system, the sensors respond faster and provide better
performance than detectors, thus they are used for more sophisticated applications.
Often, these e-noses include a system to clean the sensors between measurements,
which requires more complexity since a source of clean air is needed. The inclusion of
cleaning cycles provides certain advantages, such as slow down sensors degradation,
avoid memory effects and increase information content for data analysis (see section
1.5.1). Figure 1.10 shows two example of commercial portable e-noses. GDA2 (gas
detector array) model by Airsense Analytics [106] is a multiple technologies sensor
array that comprises an Ion Mobility Spectrometer (IMS), a Photo Ionization De-
tector (PID), two semiconductor gas sensors (SC) and an electrochemical cell (EC).
It is designed for fast detection of chemical hazardous gases and it is already used
by fire brigades, police, military, civil protection and chemical companies. Besides,
18
Operation Modes
SNet sensor modules by JLM innovation [105] consist of four MOX sensors and are
designed to work in a network for applications in safety and security, environmental
monitoring, industrial control and robotics.
(a) GDA 2 by AirSense Ana- (b) SNet sensor module by JLM Innovation
lytics [106] [105]
E-nose instruments may also include a sampling unit, instead of just an inlet to-
wards the sensors chamber. This sampling unit can consist on several inlets to be
connected to vials and a system to automatically drive the volatiles from individual
vials (fig. 1.11(b)), or only one inlet and an automatic system to drive and con-
nect the vials to the sensors chamber inlet (fig. 1.11(a)). An example of the latter
is the Nordic Sensors Technologies NST-3320, where also a temperature control of
the vials in the carousel is provided. These instruments are thought to be used in
laboratories because of their size and complexity.
As already mentioned, it is known that the shape of the transient response of every
sensor gives much information about the kinetics of the reactions between the target
19
Introduction Marta Padilla Ferran
volatile and the sensing layer at its surface. These dynamic processes are unique for
each sensor-analyte pair, therefore derived parameters are very useful and indeed
lead to improved selectivity, reduced acquisition time, and even increased sensor
lifetime [59, 108, 109]. However, traditionally, only one measure point per sensor
and per sample is kept. This point corresponds either to the maximum point or to
the mean value among several points around the maximum point of the transient
signal of the sensor operating at given conditions, i. e. static or steady-state values.
The transient response approach consist of the analysis of the evolution of the sensors
response to a sudden change in the target sample, a step or a pulse in the concen-
tration of the analyte (fig. 1.12). The potential information content in a transient
signal is considerably higher than that from a steady-state signal, since, from the
shape of these transient signals f (t) many parameters can be extracted to be used
as input to multicomponent-analysis or pattern-recognition algorithms.
Figure 1.12: Transient response of one gas sensor in a measurement cycle [36].
20
Operation Modes
When a sharp change in the analyte concentration is given, the mechanisms deter-
mining the sensors transient response are mainly two; diffusion within the sensitive
layer and surface or bulk reaction kinetics of the sensitive material, provided the
diffusion processes in the measurement chamber are significantly faster [59].
Apart from the fact that by considering the transient signal f (t) the dimension of
the feature space is increased and thus, in general, the performance of the following
prediction algorithms is much improved, a number of other advantages have been
reported. First of all, since the recorded transient signal corresponds to unique reac-
tions between sensor and analyte, the selectivity of the sensor is improved. Also, the
acquisition and calibration times can be reduced if the initial sensor transients con-
tain sufficient discriminatory information, although the steady-state is not reached
[110]. As a consequence, the sensors also require less time to recover their baseline,
a process that can be very slow when the target analytes are present at high con-
centrations. Moreover, by reducing the duration of the analyte pulse, irreversible
bindings are minimized and thus the lifetime of the sensors can also be increased.
A short acquisition time is also useful in the case dynamic headspace analysis, since
the volatiles in the headspace may be depleted faster than they can be released from
the sample. In this case, steady-state response may not even be attainable [59].
Another advantage of the transient parameters is that they have been reported to
exhibit better repeatability than static descriptors [109].
Besides, even the entire transient response may be processed with suitable classi-
fication or regression models which include multiway techniques [116] (see section
2.2.3).
Spite of the many advantages of processing the transient responses, this technique
also presents some weakness, such as dependence in the sampling procedures and
the head-space generation for different products, in the fluid dynamics of the odour
delivery system and it is also known that the transient signals may also be influenced
by previous measurements (memory effect) and by drift [107, 109].
21
Introduction Marta Padilla Ferran
It is known that the operating temperature of the MOX sensors influences strongly
its selectivity, since the kinetics of the reaction between the sensing material and
the different volatiles and the stability of the oxygen adsorbed in its surface, are a
function of the temperature. Consequently, much discriminatory and quantitative
information can be obtained by the modulation of the operating temperature, which
can provide gas-specific temporal signatures. This dependence was first modelled
by Clifford and Tuma [117] and Sears et al. [118] proposed for the first time the use
of thermal modulation to improve the selectivity of a commercial MOX sensor.
22
Operation Modes
one cycle several peaks in sensitivity may appear. By this technique one sensor can
give place to several pseudo-sensors, since the sensor presents different sensitivities
acting as several sensors. Therefore, a multivariate dynamic signature of the sample
can be obtained, where not only the magnitude of the conductance but also the
shape of the dynamic response can be unique to each analyte.
Some issues regarding temperature modulation have been studied; the waveform of
the modulation, its frequency, the harmonic content of the induced waveform or the
preferred range of temperatures. All studies conclude that temperature cycling is a
very promising approach to improve the device performance [119].
Regarding the best operating temperatures, studies have shown that, in the low-
temperature ranges, the sensor response does not carry much information, since
most reactions occur at the surface level. In general, at high temperatures, bulk
reactions are increasingly involved, so the response patterns become more complex
and more characteristic of the target gases [59]. However, it has been shown that the
optimum temperature range is dependent on the gas species to be detected [91, 92].
23
Introduction Marta Padilla Ferran
As a result of drift, models built with learned sensor response patterns may become
obsolete over a relatively short time (on the order of few weeks), so periodical re-
calibrations may be required (section 1.2.1). For this reason drift is considered
to be one of the most serious limitations to the extensive use of the electronic
noses. Moreover, individual sensors in the array may fail and this introduce errors
in measurements that may give misclassification results.
The effects of drift on the sensor response, specifically from an array of sensors in an
e-nose system, and the advances to compensate it by signal processing, are discussed
in this section along with the subject of the presence of a failure in a sensor in the
array, specially due to poisoning. Methods for fault detection and isolation (FDI)
are discussed in section 2.2.5.
To evaluate the effectiveness of every drift reduction method several strategies can be
followed. For instance, the prediction capabilities of the classification/quantification
model with and without drift reduction can be compared. The change in the pre-
diction ability must be significant, otherwise the method should not be used since
the model complexity is increased. For classification applications, another measure
of the effectiveness of the method is the relative change of the Mahalanobis distance
between the different clusters, before and after drift reduction [128], or the evalu-
ation of the Fisher ratio, again before and after drift correction. In quantification
applications, the relative change in the root mean squared error in prediction (RM-
SEP) value can be used as a performance measure of the drift reduction method
[36].
In the evaluation of the drift correction method, care must be taken into the model
validation. If samples from the whole time period are taken for the model building
and validation is based on techniques that do not consider the temporal ordering of
the samples, like leave-one-out (LOO), random sub-sampling, bootstrap, and k-fold,
this would lead to overoptimistic results. The reason for this is that the evolution
of the data would be captured into the model, which is not a realistic situation.
Instead, the model must be built with the first measurements, and then apply it to
posterior samples for testing. This strategy must be followed to validate both, the
calibration model and the drift correction method.
Drift is not a problem that affects exclusively gas sensors, it also influences long-
time measurements of other type of sensors; humidity sensors[129], flow sensors
[130], fiber-optic sensors [131], hall sensors [132], magnetic sensors [133], etc. Also,
other analytical multivariate techniques are affected by drift and sample dispersion,
24
Drift and sensor faults in a gas sensors array
Drift, poisoning and other perturbations may result in a change in the data statistics.
Techniques to deal with time varying statistics have also recently appeared in other
fields of data processing. This problem is generally referred to as population drift
[134]. Some examples arise in medical applications, where there may be changes
in the patient characteristics thus the patient population may drift, radar target
recognition or object recognition in images, where the object changes its position
and ambient light conditions may change [135], email spam filtering systems which
needs to be adapted because the types of spam changes over time, etc.
Drift is a common problem for all chemical sensors, although different causes for
drift dominate for different sensor types. In the field of chemical gas sensors, drift is
considered strictly to be due to two main causes related to the sensor itself; ageing
and poisoning. Sensor ageing consist of the degradation of the sensor’s sensing layer,
and includes the re-organization of the sensing material, for example clustering of
metal particles, and changes in the number of reaction sites. These re-organization
may happen spontaneously, but it may be speeded up if the sensor operates in
reactive environments or at high temperatures [136].
Much effort is being made to find sensor materials which interact reversibly with
the gas, such that the molecules that have reacted on the sensor will leave it as
soon as the gas leaves the sensor surface. But this is a difficult task, instead and
as mentioned previously, different operation modes of the sensors are used, such
as short time exposition to the gas analyte and study of the sensors transient or
temperature modulation of the sensors.
Some variables that influence the sensor measurements over time that are related
to other sources than ageing and poisoning, such as environmental or measurement
system, can be considered and treated as drift from the point of view of signal
processing. Environmental sources include changes in temperature, pressure, or
humidity. Besides, a bad design of the measurement system may induce temperature
variations in the measured head space or on the sensors, humidity variations in the
sample or other physical and chemical processes, also fluctuations in flow rate or
even analyte condensation in the manifold, for example. Avoiding fluctuations of
these important parameters is essential, this can be addressed by careful control of
25
Introduction Marta Padilla Ferran
sample and sensor parameters with a proper system design (see sections 1.2.2 and
1.2.3) optimised for each application, or by measuring the variables that are known
to fluctuate to be later compensated for by software.
Additional factors that change the measurements characteristics along the time,
produce effects that may be confounded with drift. For instance memory effects
which, although it is a temporary effect that may last only for minutes or hours,
for long time experiments it cannot be distinguished from drift. Memory effects are
present when the response of the sensor depends on the volatiles it has recently been
exposed to, giving an additional effect to the sensor response. This phenomenon is
due to the fact that remnants of previous gases may still be present either in the
sampling system, or on the sensor surface itself. Memory effects can be considered
as an additional source of variance due to uncontrolled previous exposures. The best
way to avoid this problem is, when possible, to improve the measurement procedure,
e. g. by limiting the size of the sampling system, or to introduce cleaning cycles such
as short pulses of clean air and/or high temperature annealing between the samples
when possible [136].
Other effects that may be confounded with drift are hysteresis or systematic er-
rors due to fixed sampling sequences, short-term effects such as system warm-up or
thermal trends, noise or even the degradation of the samples themselves. Also, the
observed sometimes called short-term drift or warm-up effect, by which the sensors
response increases or decreases for the first minutes or hours of operation. This phe-
nomenon is due to the fact that the sensors need some time before they give a stable
response after start-up of a measurement series, and thus its nature is different from
ordinary drift [59, 136].
To reduce the effects of drift in the electronic nose by signal processing, different
mathematical models have been used to compensate for the changes in the sensor
behaviour, and therefore to maintain the gas identification capability. These meth-
ods have been applied for different situations and thus they have made different
assumptions of how the drift appears.
When mainly the baseline of the sensor is changed by drift, the drift is additive.
Instead, if drift mainly changes the sensitivity of the sensor, it is multiplicative, thus
the response is increased or decreased by some factor (fig. 1.14). However, in general,
the response of a sensor suffers changes in its baseline, sensitivity, selectivity and
magnitude along the time, therefore the data series may present gradual (sometimes
irregular) decreases or increases or even jumps, (usually of small size), also noise
can be superimposed on all the sensor signals and he speed of the response can be
also altered.
26
Drift and sensor faults in a gas sensors array
Figure 1.14: Ilustration of additive and multiplicative drifts on a sensor transient response
in a sequence of measurements [137].
Furthermore, the characteristic pattern of a given target gas may not be preserved
over time. The pattern considers relative relationship between different sensors
rather than their absolute output in a sample. If these relative relationships stay
constant over time, a simple normalization could correct the drift effect. However,
drift may also modify the characteristic patterns of the analytes, since it influences
the different sensors in different ways.
In a multi-sensor system, an exploratory tool as PCA shows ([128], see paper 4.1)
that drift may tend to move data mainly in one direction, which is similar for all
similar gas mixtures (fig. 1.15). This direction usually presents a big component
along a PCA model space and a smaller one in the residual space, thus violating
Hotelling’s T 2 and, also squared prediction error (SPE) limits respectively. To dis-
tinguish a phenomenon affecting all sensors, such as drift, from a fault in one or
several individual sensors, it has to be taken into account that drift and a process
fault would be registered by a band of correlated sensors [138]. Therefore, a sensor
fault would move the samples away from the PCA model space, surpassing mainly
the SPE limit [139].
The reason why the drift tends to go in only (mainly) one direction for each cluster,
is that, if sensors are similar, and they are exposed to the same analytes, they tend
to drift in a similar way [136]. However, different gas mixtures may drift in different
directions. This has to be taken into account when choosing a reference gas for
compensating drift, thus this gas has to be very similar to the test gases in the
application. In a situation where different sensor types are used, drift should be
described for each sensor type.
27
Introduction Marta Padilla Ferran
Figure 1.15: PCA scores plot for data from 39 sensors, measuring on 9 different gas
mixtures [128]. The direction of drift is similar for all mixtures.
The effect of drift in array of gas sensors have been observed since the beginning of
the development of electronic nose devices in long and not so long time experiments.
Given the broad variety of applications of the electronic noses, the effect of drift in
the gas sensors have been seen in many real life measurements for different situations,
like food applications, environment monitoring, etc.
For instance, among the food applications, Haugen et al. [140] showed graphically
the effect of drift on two experiments performed with a commercial device hybrid
gas-sensor array (NST 3220 by Nordic Sensor Technologies), which contained two
blocks of five MOS-FET sensors set at different temperatures and one block with
five Taguchi MOX sensors coupled in series. Indeed, they observed short-term drift
within a measurement series lasting several hours, and distinguished it from long
term drift during several measurement series. The first measurements were obtained
from a fish storage experiment that lasted 5 days and the second one was performed
on two types of milk samples of different qualities during 2 days. In this case, the ob-
jective was to discriminate fresh pasteurised milk from slightly oxidised pasteurised
milk. Many other experiments related to food quality have been carried out where
drift have had an important influence on the sensors. For instance, in an experi-
ment carried out by Pardo et al. [141] for the classification of several measurements
of extra-virgin olive oils in 14 different geographical provenances, they observed a
large mixture of the different classes which they attributed to drift. They used seven
commercial sensors from Alpha M. O. S. and five from Figaro, during the time of
performing 242 measurements. By testing several approaches they could reduce the
misclassification error from 25% to 10%. Drift was also present in a 50 days ex-
periment by Dutta et al. [6], where five tea samples with different qualities were
analysed using a MOx sensor based electronic nose, to discriminate between their
flavours. A PCA scores plot was plotted to show the effect of drift on the samples.
28
Drift and sensor faults in a gas sensors array
Schaller and co-workers [142, 143] have reported large background drift of CP and
MOS sensors, and also MOS sensor poisoning when attempting to analyse cheese
samples of Emmental type. Concita et al. [13] also observed drift on one of the
six thin film tin oxide sensors while detecting microbial contamination in processed
tomatoes. The drifting sensor was finally discarded for that application. In other
experiments, drift was taken into account and could be compensated by a reference
gas between measurements; measurements of samples of milk during 20 days with a
commercial electronic nose (Fox 4000, ALPHA MOS) based on MOx sensors [144],
on the monitoring of changes in apple fruit volatiles during 15 days by another com-
mercial electronic nose (Libra Nose by Technobiochip) based on QMB sensors [145],
also on the monitoring of changes in tomato aroma profiles of two different cultivar
during about 20 days [146], etc.
Kuske et al. [147] carried out work, the objective of which was to detect the presence
of moulds growing in building materials by using an e-nose with metal oxide sensors
and two different classifiers. In the experiment, two datasets were collected, the
second one five months after the first one with a time duration of 3 months. It
was found that drift had an important effect of both classifiers, which results were
reduced from 4 up to 9 percent points in the final classification rate with respect to
the one obtained with data from first data subset. The authors also present a PCA
score plot where drift can be perceived visually, since both data sets show the same
cluster shape but appear displaced a short distance from each other.
Drift has also been observed in environment monitoring applications. For example,
Tsujita et al. [148] run an experiment consisting on monitoring NOx values in Tokyo
over several weeks, with a distributed sensor system of four individual tin oxide sen-
sors. In that experiment, the drift of the individual sensor signals made necessary
one re-calibration every 10 days. Also, Nake et al. [149] measured volatile emissions
from wastewater plants with two commercial devices based on CP and MOx gas
sensors. They sampled air at different locations of the wastewater plant on different
days. However, the conductive polymer signals formed groups correlated to the day
of measurement, thus showing a substantial influence of ambient humidity (which
can be also considered as drift). Romain et al. [150, 151] have also reported strong
drift for in site monitoring of off-odours in experiments during several years. Inter-
esting plots show the evolution of the sensors response along the time (fig. 1.16),
and also their response when they are corrected. Other in field application carried
out by the same authors, included measurements from a compost plant [152] where
drift was also present. Perera et al. [24] observed drift in the monitoring of the state
of an air compressor during a long time, performed to detect oil vapour leakages.
The method they developed to detect these leakages was designed to cope with drift
effect. And also De Vito et al. [153, 154] had to deal with drift in an experiment
during 13 months, in which the concentration of urban environment pollution gases,
such as benzene, CO, NO2 and NOx, was monitored using a solid state gas sensor
array device. They observed large variation of the sensors response due to seasonal
meteorological effects and drift, which is shown graphically by plotting for example
absolute and relative errors on benzene concentration versus time in weeks [154].
29
Introduction Marta Padilla Ferran
Figure 1.16: Time evolution of six TGS sensors response during 3 years [150].
On the other hand, during long measurement series, some of the sensors of an array
may fail due to a variety of reasons, such as instrumental, or just the sensing material
degradation (gradual or abrupt) produced by time or poisoning. Apart from drift
during their long experiment in monitoring odours on a landfill site, Romain et
al. [150] also reported the need to replace some sensors. In experiments where sensors
were exposed to certain substances, such as sulphur compounds (e. g. H2 S or SOx),
hexamethyldixyloxane [155] or some acids (e.g. acetic) in the samples, e. g. cheeses
[142, 143], wines, vinegars, cruciferous vegetables or lampant virgin olive oils [96],
sensor poisoning has become a serious problem. Thus, in these situations, failing
sensors must be detected and identified in the array, in order to be intermediately
replaced. However, only few works in literature can be found which intend to detect
and correct a fault in an array of gas sensors. In an experiment carried out by
Pardo et al. [156] five semiconductor gas sensors were exposed to CO-NO2 mixtures.
In order to detect a possible malfunction of one of the sensor of the array during
continuous operation, the outputs of the sensors were estimated from the signals and
compared to the real values. Finally, Tomic et al. [157] studied calibration transfer
30
Drift and sensor faults in a gas sensors array
techniques for their application in sensor replacement problems, (see section 2.2.4).
In this section, a brief review of some techniques used in AO systems for drift
counteraction methods is made.
Robust features
A good starting point in the design of a technique for prediction which could work
under drift conditions, could be the identification of features that are (approxima-
tively) stable over time. As it has already been commented, the modulation of the
sensor operation temperature has been reported to generate features that are sig-
nificantly more robust against drift than isothermal features. However, this is only
possible in the case of MOx sensors (section 1.3), other extracted features such as
transient responses have shown to be drift sensitive (see paper 4.1), although they
provide more information for prediction than isothermal features.
Other approaches to find robust features used the pattern of responses rather than
individual sensor responses. For example, robust features were found by Wilson et
al. [114] by thresholding the values of the sensors response into a binary output.
They used these binary features for the discriminate between different chemicals
(acetone, ethanol, hexane, isopropanol, methanol and carbon monoxide and mix-
tures of two) by an electronic nose with an array of eight tin-oxide Figaro sensors
at eight different temperatures. To binarise the sensors signals, the sensors response
were arranged in ascending order and the output from each sensor was set to zero
if it was smaller than the median output, and set to one if it was larger than the
median. Therefore, the resulting output from the threshold function was a pattern
of zeros and ones and was more robust than the absolute sensor values since it is
relative and not as sensitive to noise. This technique removes much information,
specially the one related to concentration levels, and therefore it is useful only for
classification purposes.
Bednarczyk et al. [158] used a similar technique, but the sensors signals were mod-
ified pattern by pattern. For every sample, the sensors with the highest (winner),
and smallest (loser) outputs were identified and the slope between the winner, loser
and its two nearest neighbours was calculated. Since Bednarczyk used ten tin oxide
sensors, the method provided a total of six outputs from each sample giving a spe-
cific pattern that did not change much over time. In the classification task two steps
were involved; a first one using the winner and loser sensors which identify the family
of chemicals the sample belongs to, and a finer one with the calculated slopes which
further discriminates the chemical from others within the same family. However, the
31
Introduction Marta Padilla Ferran
authors mention that the system does not discriminate sufficiently among chemicals
within a large range of concentrations, they must be further processed to eliminate
this effect. And finally, although this approach seemed interesting, because the pre-
processing is made sample by sample, the system is still sensible to drift since the
characteristic patterns of the target analytes can also be altered (section 1.6.3).
Reference gas based methods are used to estimate drift. Usually, measurements
on one or several reference gases are made at the beginning of the measurement
series and with some intervals as long as the sensors are used. The change in the
sensor responses to the reference gas is taken as a measure of the change in the
sensors response for all other measurements (fig. 1.17). Furthermore, the use of
a reference gas also allows to distinguish between sensor drift and changes in the
samples themselves.
In the use of these methods, it is assumed that there is a strong correlation among
the reference gas, the drift and the other samples through the sensors response.
Therefore, the chosen reference gas or gases must be a similar compound of those
being measured, also close to the real samples in sensor response space, since the
drift in the reference gas should reflect the drift for all other samples. It must also
be stable over time, without degradation or varying concentration, available and
easy to measure. Authors have followed different approaches to choose a suitable
reference gas; some of them have used the same reference gas (often water) for all
applications, while others choose the reference gas depending on the application. In
other cases, replicates of the samples can be used as pseudo-reference gases, as long
as the samples are stable over time or reliable standardized samples are available
[159].
Examples on the use of gas reference methods include Fryder et al. [160], who cor-
Figure 1.17: Illustration of drift correction by a method using a reference gas measurement
before the target analyte measurement [137].
32
Drift and sensor faults in a gas sensors array
Haugen et al. [140] also performed re-calibrations within a single measurement se-
quence and between measurement sequences (fig. 1.18). At every time scale, the
temporal variations in the calibration gas was modelled, and a multiplicative correc-
tion factor was then applied to the new target samples. In fact, he reduced both the
additive and multiplicative effects of drift, again sensor by sensor. First, they calcu-
lated the ratio between the responses at time t and at the initial time for a reference
gas xRef
i
Gas
(t)/xRef
i
Gas
(0) for each sensor i, and used it to compensate the responses
of the samples. Then, they fitted a linear trend to this ratio (at + b) to find the
correction factor for the samples located between the reference gas measurements
(eq. 1.2):
xcorr
i (t) = xi (t) · (at + b)i (1.2)
This method is relatively simple and provided very good results. However, although
the technique was tested on complex mixtures, volatiles from food samples, the ex-
periments did not last more than 10 days. Further, many calibration measurements
have to be made and this is not possible in many situations.
Finally, Arthursson et al. [128] devised a technique based on PCA and PLS, called
component correction (CC). In this technique, a reference gas is used to estimate
the main direction of drift, done by computing the first principal component of its
samples. Then, this or these component/s, pref , are removed from the multivariate
array response x by:
The vector pref contain the contribution of the original variables to the identified
direction of drift, thus pref can be used also to see the most drifting parameters. On
the other hand, the removal of this direction preserves all the other directions and
keeps the important variances that separate different clusters and concentrations,
unless, of course, the information is found in the same direction as the drift. Results
were satisfactory (see fig. 1.15) but this method depends strongly on the reference
gas. If the analytes to be measure are similar, drift of all of them would show similar
trends but this is not necessary true if the analytes are different (section 1.6.3).
33
Introduction Marta Padilla Ferran
Examples of nonlinear models that follow an adaptive strategy are the use of a single
Kohonen self-organizing map (SOM) for all the classes, [161, 162] or a separate SOM
per class [163] (fig. 1.19). In the last work, the authors presented a new mSOM neu-
ral network methodology designed to improve the classification of different odours
using the output signals of a an electronic nose in an experiment during four weeks,
therefore under drift conditions.Self Organising Maps (SOMs) transformed incom-
ing signal patterns into a two-dimensional discrete map, imposing neighbourhood
constraints on the output units. However, the dynamic behaviour of the signal pat-
terns with time can make clusters to change its position in the map and thus these
maps may become useless. To deal with this problem, the authors provide a sys-
tem of multiple self organizing maps (mSOM) with as many maps as odour classes
for a classification task, that enable a self-adjusting process for each neuron in the
local maps. Zuppa et al. [163] showed that mSOM provided small error rates in
the classification of six compounds in an experiment that lasted 4 weeks. However,
for long time experiments, techniques that track the evolution of data with several
classes need that all the classes are alternatively measured, otherwise it may lose
the tracking of some of the classes.
34
Drift and sensor faults in a gas sensors array
Figure 1.19: Illustration of drift correction by an adaptive clustering method, which tracks
the evolution of the data due to the effect of drift [137].
assuming that sensors dynamics is more stable than the sensors response. Models
such as ARX, ARMA and Box-Jenkins were used by Holmberg et al. [164, 165]. They
concluded that Box-Jenkins was the more flexible to model the data because it does
not contain an auto-regressive part (unlike ARX model), and also includes a noise
model. They also found that the model was sensible to measurement pauses during
the experiment, in which five analytes were classified along 45 days. In another
work, Holmberg et al. [165] used a recursive least squares algorithm to update the
model. This method proved to be suitable even for the periods of measurement
pauses, thanks the ability of the updated model to adapt itself to the changes.
Later, for classification, a separate dynamical model was built for each analyte
class, and unknown analytes were assigned to the class whose model displays the
lowest prediction error. Finally, Perera et al. [24] have developed a fault detection
method based on recursive dynamic PCA [166, 167] that can operate under drift
conditions. The proposed technique is adaptive, therefore it is very useful for the
target application, in which a continuous monitoring is made to detect oil vapour
leakages in air compressors. Results are very good, but in this work, the component
of drift is simulated and linear.
Other appraches [168, 169] have used independent component analysis (ICA) to find
statistically independent components of the sensor array response, which could be
associated to drift and discriminatory information. For instance, Kermit and Tomic
[169] showed that ICA could separate three sources of information in the sensor
response: analyte identity, analyte concentration, and drift effects. In contrast, PCA
was only able to separate concentration information (first principal component), but
analyte identity and drift kept mixed together in the second and third principal
components. However, in this case ICA was applied in an off-line way, without
testing new incoming samples. In the experiment by Di Natale et al. [168], they
used the validation method of leave one out, which is not suitable for validation in
drift conditions (section 1.6).
Also, artificial neural networks (ANN) have been used to determine the concentra-
tion of a target analyte under different ambient conditions. This ANN took actual
35
Introduction Marta Padilla Ferran
values of the sensor’s resistance, temperature and humidity as inputs and resulted in
a very accurate estimation of the analyte concentration [170]. In long experiments
other factors that are not considered in this work may cause drift such as ageing,
which is difficult to measure.
Wavelets decomposition method has also been used as a drift correction method,
since by using Discrete Wavelet Transform (DWT) it is possible to have information
of the frequency content and time localization of drift components on a time signal
[171, 172]. Both papers show good results but signals were collected during few days
[171] or drift was simulated [172]. Both approaches are interesting and should be
tested on real drift conditions.
A new technique based on common principal component analysis (CPCA) was pro-
posed by Ziyatdinov et al. [174] in order to find and remove a direction of variance
common to all classes present in a dataset. They tested this method on measure-
ments collected during 7 months by 17 conductive polymer sensors and succeed in
improving the time stability of the system. Results are presented in a figure showing
the evolution of the classification rates of the species of analytes along the time of
measurements. However, they use 1000 samples in the smallest training set, which
is not feasible in many situations.
36
Drift and sensor faults in a gas sensors array
Figure 1.20: Illustration of an ideal situation where the directions of drift and analyte
information are orthogonal in a PCA model space [59].
Sisk et al. [176] studied the optimal calibration protocol. A condition-based cali-
bration, in which calibration was only performed when the classification performance
decreased below a predetermined threshold value, was observed to be superior to a
time-based calibration or to interval-based calibration protocols.
Romain and Nicolas [150] compare three simple drift counteraction techniques for
two arrays of six metal-oxide Figaro sensors continuously measuring during 7 years a
real-life atmosphere, which was the odour generated by an urban waste composting
facility. In that experiment, the temperature and humidity of the sensors cham-
bers were controlled and a calibration gas was available. The tested solutions to
compensate the drift effect were: signal pre-processing, univariate sensor correction
37
Introduction Marta Padilla Ferran
[140] and multivariate array correction [128]. In the first method, they preprocessed
every sensor signal by subtracting the sensor response to pure reference air. This
solution requires cycling operation between reference air and target gas. However, it
is not convenient for on-site applications since heavy gas cylinders should be carried
in the field or, as an alternative, the reference air could be generated by filtering
the ambient atmosphere, but this produce a reference without much purity which
would increase the data dispersion. Univariate sensor correction was performed by
using a multiplicative correction factor obtained from measurements of a reference
gas at regular intervals within the target measurement, and finally, the Arthursson’s
CC method was performed, which also required a calibration gas. In that work they
found that the univariate approach was the most adequate for their problem, since
they had to replace sensors in the array within the measurements series.
Sisk and Lewis [176] also made a comparison between some analytical and cali-
bration methods for an array of 15 carbon black-polymer composite sensors in an
experiment during four months. In that experiment, eight different analytes at low
concentrations where exposed to the sensors. The most difficult separation tasks,
e. g. the discrimination between hydrocarbons, were little improved by modelling
the dynamics of sensor drift, either through a linear regression or Fourier transform
decomposition of the individual responses of each sensor. Finally, a simple linear
sensor-by-sensor calibration scheme proved effective for classification performance
of binary separation tasks.
Zuppa et al. [163] tested the performance of their proposed mSOM (with several
internal parameters), mSOM with different filter types (Saviztky-Golay, Discrete
Wavelet Transform), and compared them with a previous version of SOM by Dis-
tante et al. [177] and other techniques such RBF and Fuzzy ARTMAP. They found
that their proposed mSOM with a filter was superior to the other tested techniques
and derived some conclusions about the internal parameters of the optimal mSOM
technique.
Summary
In the field of gas sensor systems, drift correction is a very important task and should
be included into the general scheme of pattern recognition. Several and different
signal processing techniques have been developed to handle this serious problem;
univariate [140, 176], multivariate [128, 162], linear [128] and non-linear [161, 162].
These techniques have been used following different strategies which depend on the
specific application; adaptive methods, where the prediction model is updated with
recently classified incoming samples, methods based on some measurements of a
reference gas distributed along the time, or methods which model a given trend in
the training subset that is extracted from new samples.
The adaptive methods strategy is useful in the case of an instrument working con-
38
Drift and sensor faults in a gas sensors array
tinuously and where target samples are known not to vary too much, as in process
monitoring tasks in a plant. However, their success rely on correct prediction; mis-
classification errors may cause the model to lose track of the class patterns and,
moreover, all analytes need to be sampled frequently to prevent their patterns from
drifting too far. Besides, the use of a reference gas increase the number of measure-
ments and assumes that the reference gas is available. Finally, by modelling a given
trend, the fact that the drift trends may change along the time is ignored. Indeed,
this would give place to local effects, which the model is unable to detect and thus
to compensate for. All strategies are totally useless in cases where abrupt changes
occur, except when measures of a reference gas are made in that period. In that
cases only new re-calibrations can keep the instrument in use.
Anyway, no technique is valid for all situations and more research is needed to find
more methods that could be adjusted to specific situations.
39
Introduction Marta Padilla Ferran
40
Chapter 2
Robustness
In this chapter, an overview on some signal processing techniques that can be used
to improve the robustness of enose instruments is presented. First, section 2.1 intro-
duces the notion of robustness for analytical methods in a general form, and then,
an exploration of useful methods for dealing with two specific problems affecting
array of gas sensors; drift and failure of individual sensors in an array, is made in
section 2.1.
These problems are related to the individual methods and technologies involved in e-
nose systems, but some of them, such as sensitivities to environmental parameters,
time stability or sensors failure, are also present in other analytical instruments,
which are highly used in industry. Indeed, industry demands these instruments to
fulfil certain requirements, within which robustness is one of the most important
ones. Every used instrumentation technology in industry is therefore rigorously
validated through a proper study of the instrument and methodology robustness.
A robustness test consists on a deep study of the behaviour of a given instrument
to variations in problematic variables, specific of every type of instrument, taking
into account not only the system but also the calibration method and algorithms
involved in the prediction step. Thus, those measurements technologies present in
41
Robustness Marta Padilla Ferran
industry consider the measurement of system robustness a very important issue, the
concepts and methods of which are being continuously developing.
On the other hand, AO systems are relatively young, and general standards and
concepts for the study of their robustness have not been established yet. However,
some of the concepts derived in other instrumentation areas, such as the variety
of methods in chromatography, spectroscopy, etc. could be adapted and transferred
to the field of artificial olfaction. For this reason, a general overview on the no-
tions of robustness and robustness test in such instrumentation fields is presented
in the following sections (sections 2.1.1 and 2.1.2). In particular, the robustness of
calibration models is an important task for signal processing, thus this subject is
discussed (section 2.1.3).
Further, as drift causes measurements performed at the calibration phase and after
a certain period of time to differ considerably, this situation can be seen as having
two instruments of the same kind; the first one corresponds to the calibration phase
and the second one is considered to be out of calibration. Hence, in order to avoid
to perform a full recalibration on the second device, techniques based on calibration
transfer methods can be applied to make this second device work as expected. The
aim of calibration transfer techniques is to reduce efforts in calibration of similar in-
struments or of new replaced sensors, avoiding full re-calibration of the instruments,
since otherwise re-calibrations can become a very time consuming and expensive
task. Section 2.2.4 gives a short presentation of these techniques.
Another particular problem under study in this work is related with individual sen-
42
Robustness
sors failing in an array. If that situation occurs, resulting measurements could lead
to false results and big errors at any prediction stage. Consequently, it is very impor-
tant to detect a failing sensor as soon as possible in order to be removed and replaced
when it is possible, and meanwhile the system must work with a reconstructed re-
sponse of the failing sensor calculated from remaining sensors. The issues of sensor
fault detection and isolation (FDI) and fault diagnosis, which involve faulty sensor
detection, identification and correction of its response, have been largely developed
for statistical process control. Among the techniques used for such issue in system
control, statistical methods are very interesting since they do not need to make a
model of the behaviour of the instrument, on the contrary, the model is made on its
measurement data. This topic is discussed in section 2.2.5.
2.1 Robustness
Industry is increasingly requiring for instruments able to give fast measurement with
suitable precision, accuracy and long-term stability under variable (industrial) condi-
tions, which include variations of parameters, instruments, addition of uncontrolled
factors not studied during calibration, etc. At the same time, industry is subjected
to strict regulatory requirements (specially in pharmaceutical industry) and costs.
This means that every used measurement method is required to be strictly validated
through a proper study of the robustness of the instrument and methodology. As a
result, the measurement of system robustness and model transferability have become
two key points in the development of new measurement technologies.
However, there is not an official definition for the word robustness valid for all type
of measurements, although it has always been considered an important property of
a measurement method. Many similar definitions can be found in literature, most
of them related to specific problems in specific applications, and thus sometimes dif-
ferent definitions can be found within one domain. Nevertheless, some international
organizations, such as International Organization for Standardization (ISO) [178] in
the International Conference on Harmonization of Technical Requirements for Reg-
istration of Pharmaceuticals for Human Use (ICH) [179] and US Pharmacopoeia
(USP) [180], gives some guidelines to characterize the quality of a measurement
method.
43
Robustness Marta Padilla Ferran
For instance, USP and ICH definition of robustness are equal: the robustness of
an analytical procedure is a measure of its capacity to remain unaffected by small,
but deliberate variations in method parameters and provides an indication of its
reliability during normal usage. But there exist another term used for robustness or
related properties, ruggedness. Also for this term, several definitions can be found
in literature, but also many times it is considered as a synonym of robustness. An
example is ICH, which considers ruggedness and robustness as synonyms, while USP
defines it as the degree of reproducibility of test results obtained by the analysis of the
same sample under a variety of normal test conditions, such as different laboratories,
different analysts, different instruments, different assay temperatures, different days,
etc. [181]. Therefore, the word ruggedness is mostly used when dealing with
transferability problems, while robustness is computed from intra-laboratory trials
(different temperature, concentration, etc. ) [182].
Robustness tests are widely applied to study the potential sources of variability in
method results. The ICH guideline defines robustness testing as: an experimen-
tal study in which one evaluates the influence of small changes in the operating or
environmental conditions on measured or calculated responses. And state that one
consequence of the evaluation of robustness should be that a series of system suit-
ability parameters (e.g. resolution tests) is established to ensure that the validity
of the analytical procedure is maintained whenever used [181]. The methodology of
robustness testing in chemometrics and statistics is in agreement with these ICH
gidelines [182].
In summary, robustness tests identify the parameters that can be varied without
affecting performance of an analytical method, providing reasonable operating tol-
erances that may be expected during routine use. Moreover, robustness tests also
determine which parameters need to be tightly controlled, and include these critical
variables in the procedural description of the analytical method. Hence, robustness
tests demonstrate that an analytical method is reliable during normal usage. Be-
sides, ruggedness tests measure the reproducibility of results taking into account
normal operating variations, and ensures that an analytical method will perform
suitably each time it is used.
44
Robustness
There are a number of factors that may be varied to evaluate robustness of an analyt-
ical method. In relation with a method based on an electronic nose, some examples
of typical parameters that may need to be studied are; the sampling method (sample
preparation, sampling time), influence of temperature, flow rates and humidity as
these parameters change the sensors responses. The range used to evaluate these
parameters should reflect the potential operating conditions. An interesting robust-
ness test performed on a electronic nose system is the study of the performance
and stability of the instrument when the operating conditions change. This is the
case of an array of metal oxide gas sensors when the sensors are heated by short
pulses instead of continuously. There is an considerable reduction in the power con-
sumption of the instrument when using pulsed heating sensors in comparison with
isothermal sensors, however the performance of the system may be degraded, as well
as the metrological characteristics of individual sensors, due to thermal effects that
degrades the sensors when repetitively switching temperature [183].
Recently, Zeaiter et al. [182] proposed a definition of robustness for multivariate cali-
bration in infra-red spectroscopy (IR) techniques that could also be adopted by the
45
Robustness Marta Padilla Ferran
The term robustness also appears in the field of statistics. The word robust was in-
troduced in statistics to describe procedures that give good results even though there
might be violations in the assumptions upon which these procedures are based [184].
The term robustness in robust statistics establish the resistance of estimators to the
occurrence of outliers [185]. Also, the model robustness refers to their insensitivity to
deviations related to the non-conformity with statistical hypothesis (linearity, nor-
mality, independence of the variables). In multivariate calibration, such a measure
of robustness is equivalent to the maximal percentage of outliers that a calibration
method can handle [182, 186, 187].
In multivariate calibration, the used indices used for robustness assessment are based
on minimizing the objective function of prediction error [182]. Indeed, a calibration
model is mostly computed using the cross validation technique, so calibration models
with less error of prediction are normally chosen and less attention is paid to their
robustness. Later, models are commonly tested by the root mean square error of
prediction (RMSEP) index. Sometimes, there are also included other indices as a
measurement of robustness, such as the calibration model dependence on training
samples, on the training set size or results from permutation tests, all based on
resulting prediction error.
On the other hand, there exist some ways of improving the robustness of calibration
models. Several statistical techniques that are often used to enhance prediction-
model performance, can also be used for this purpose. Zeaiter et al. [189] presented
an overview of these techniques in the field of near infra-red (NIR) spectroscopy. In
that work, she considered multivariate calibration methods to be linear functions
relating n predicted variables Y to n data samples X by the regression vector b:
Y =X ·b+e (2.1)
where e is the matrix of residuals supposedly due to random noise of zero mean. She
then stated that the robustness problem is due to variations in the measurement
conditions caused by variations in influence factors that affect the measurement
by adding a perturbation δx. This perturbation is represented in the prediction
46
Signal processing methods for improving robustness
δ ŷ = δxT · b (2.2)
which gives:
and therefore he conclude that, to minimize δ ŷ, one or more of the three terms in the
right hand of the equation 2.3 must be minimized. She showed how some prepro-
cessing methods commonly used in NIR techniques minimized some of these terms;
geometric spectral preprocessing methods (SNV, RNV, De-trend, MSC, Smoothing
and Differentiation), which correct for baseline shift, curvi-linearity and noise-related
terms, particularly reduce the kδxk term, smoothing reduces the second term kbk
giving an advantage over other geometric preprocessing methods but with a careful
selection of the width of the smoothing window. Also, ridge regression and regular-
ization methods in general, decrease the term kbk. Finally, Orthogonal projection
methods (OSC, DOSC, OPLS, IIR, and EPO) minimized the third term, |cos(δx, b)|,
representing the angle between the systematic variations and the model. These
methods seek to reduce |cos(δx, b)| to zero, to ensure the relative independence of
the model from variations in the x data. Variable-selection methods (Wrappers, Fil-
ters, and Embedded) affect the robustness via kδxk and |cos(δx, b)|. The term kδxk
could increase if the selected wavelengths are useless (or no information is given
by some gas sensors in the array). The term |cos(δx, b)| depends on the initially
selected wavelengths (gas sensors in our case); if the number of useless wavelengths
(sensors) increase, |cos(δx, b)| is increased. Hence, variable selection should be used
carefully in seeking to enhance the robustness of the calibration model.
This section revises four areas of knowledge in signal processing, highly applied in
other technologies, in which some developed algorithms can be very useful for the
improvement of the robustness of AO systems.
In the last decade, there has been a great effort to improve robustness in the areas
of chromatography and spectroscopy. In these fields, several signal processing tech-
niques have emerged to minimize these problems and improve the robustness of the
47
Robustness Marta Padilla Ferran
calibration methods (see section 2.1). Specifically, the family of orthogonal projec-
tion (OP) techniques which include net analyte signal (NAS), are commonly used
since the first OP method, orthogonal signal correction (OSC), appeared in 1998
[190]. OP family include other approaches for OSC by Fearn [191] and by Sjoblom
[192], direct OSC (DOSC) [193], direct orthogonalization (DO) [194], orthogonal
projection to latent structures (OPLS [195] and O2PLS [196]), OPLS discriminant
analysis (OPLS-DA) [197], Kernel OPLS (K-OPLS) [198], projected OSC (POSC)
[195], piecewise OSC (POSC) [199] and an improved version [200], external parame-
ter orthogonalization (EPO) [201], error removal by orthogonal subtraction (EROS)
[202], orthogonal discriminant projection (ODP) [203] and dynamic orthogonal pro-
jection (DOP) [204], etc. Other tecniques explored in this areas are net analyte
signal (NAS) (also called net analyte projection (NAP)) [205] and derived algo-
rithms [206–208] and independent interference reduction (IIR) [209]. In the field of
AO, the component correction method (CC) [128] can be considered as belonging
to the OP family.
In general, these methods are based on extracting from the multivariate measure-
ment space D, the subspace not related to components of interest Y . For this,
different approaches have been developed. For instance, in order to extract the X
data structure to be related to Y , these approaches can be based on partial least
squares (PLS).Several works comparing some of these techniques can be found in
literature [189, 206, 210, 211].
To list and compare the orthogonal projection methods methods, Zeaiter et al. [189]
consider the measurement variables space D as the sum of three orthogonal sub-
spaces: G space containing mainly effects due to variations in variable Y , N mainly
containing effects due to structured systematic variations, and E containing effects
due to variations of random noise (eq. 2.4). Thus, the data X, which constitute a
cloud of n sample points into D, can be split into two subsets; X + in subspace G,
X − in subspace N , and residuals in E. Therefore, X + is the useful part of X related
to G, X − is the useless part of X related to N . Mathematically, this is represented
by equation 2.4.
48
Signal processing methods for improving robustness
X ∗ = X · I − P − P −T (2.5)
Hence, according to Zeaiter [189], the different orthogonal projection methods are
based on the way P − is identified. Wold et al. OSC [190] and Wise and Gallagher
[212] approaches are OP methods that identify P − directly from X. These methods
starts by determining the structure of variance of X data by PCA. PCA scores are
then orthogonalized to Y giving T − . And finally, the matrix of latent structures
P − corresponding to T − is calculated using PLS or PCR. Besides, techniques that
identify P − from X − first compute X − , and then a PCA on X − to get P − . Also,
there are two ways of obtaining X − ; deducing it from X by intrinsic methods and by
using a set of dedicated experiments which incorporate systematic variations, these
are extrinsic methods.
At the same time, intrinsic methods include two approaches; a direct approach,
where X − is identified in the prediction space containing the common variations of
X and Y , and an indirect approach, where X − is identified from the structure of
X. In the direct approach, it was expected to avoid the problem of overfitting that
present the methods following the indirect approach. It finds a subspace common to
X and Y relative to X + . Then PCA is performed on this subspace to determine the
orthonormal base P + . OSC by Fearn, OPLS, O2PLS, DO and an improved version
of POSC follow this strategy. In contrast, the indirect approach is the strategy
used in projected orthogonal signal correction (POSC) [195] which is similar to
the direct orthogonal signal correction (DOSC) [193] and net analyte signal (NAS)
preprocessing [205].
On the other hand, extrinsic methods require the use of a special matrix Z − , which
contains spectra acquired at different levels of variations of known influence fac-
tors.Then, P − is identified using PCA on Z − . Different methods exist using this
direct approach, such as independent interference reduction (IIR) or external pa-
rameter orthogonalization (EPO). Most of these methods are able to remove only
the effect of known influence factors (temperature, scale, instruments, etc. ), or a
combination of them, this is the case of EPO, EROS or IIR. However, DOP is able
to remove variations related to unknown influence parameters and components in
the data that might affect the model. Thus, extrinsic methods have the advantage
of not requiring the Y information to reduce the complexity of the model, done
by removing the effects of known or unknown influence factors [213]. Component
correction (CC) technique, used in e-nose applications, can be considered also to
belong to these extrinsic methods.
Despite the elimination of variability non related to the information of interest, many
49
Robustness Marta Padilla Ferran
authors have reported that most of the intrinsic algorithms are not able to reduce
the error of prediction compared to ordinary PLS regression (only slight improve-
ment in specific applications), even though the number of latent variables (LVs) is
considerably decreased. Furthermore, these methods present the disadvantage of a
risk of overfitting the number of OSC components, which need to be adjusted, and
also errors in measurements of information in Y are not considered. The main ad-
vantage of applying the above mentioned OSC and NAS techniques lies therefore in
the possible easier interpretation and understanding from the analysis of corrected
data, and the possibility of an analysis of the orthogonal information removed, which
might also be useful. These techniques can also be used to remove specific variation,
if that variation can be described by a Y vector or matrix. Nevertheless, we think
that parsimonious models obtained by applying intrinsic correction methods can be
expected to be more robust, even though prediction error is not reduced compared
to ordinary PLS regression on raw data.
In contrast, the use of most extrinsic methods have resulted in an improvement not
only in an easier interpretation of the calibration model, but also better prediction
ability have been reported [213]. Although these methods do not need the informa-
tion Y , they require other type of information in the form of more measurements
and experiments, not always easy to perform.
Hence, given the variety of OP approaches, in order to extract the correct informa-
tion most related to the property to be predicted, the data preprocessing technique
must be selected and combined, based on the nature of the data to be analysed and
the relevant features in the data.
The term blind calibration has been mostly applied in the field of sensor networks
[215] and communications [216]. As in AO, individual sensors in a network may
be prone to calibration errors, and these errors are one of the major obstacles to
their practical use. Calibrating every sensor by hand is infeasible if the networks
contain tens of devices and, consequently, automatic methods for jointly calibrating
sensor networks in the field, without dependence on controlled stimuli data, is of
50
Signal processing methods for improving robustness
The objective of both calibration free and blind calibration methods is similar to the
so-called blind source separation (BSS) problem. BSS deals with situations that are
closely related to the cocktail-party problem, well known in signal processing, which
involve the recovery of the original speech signals of different speakers from the mixed
sound. In signal processing, the set of algorithms named independent component
analysis (ICA) are usual for solving such BSS problems. These methods have also
been used in other disciplines, including analytical chemistry [220] and electronic
noses [168, 169]. ICA algorithms uses high order statistics for de-mixing the original
data, assuming that the data matrix contains linearly mixed signals coming from
mutual statistically independent (non-Gaussian) and thus uncorrelated sources and
without any prior information about the sources or reference materials. Like ICA,
most calibration-free resolution techniques are based on the assumption that the
instrumental response in a mixture is an additive linear combination of the signals
from individual species, the pure components [217].
Mathematically, MCR methods de-mix a data matrix X, containing the raw mea-
surements of mixtures, into the contributions related to each of the pure components
in the system by decomposing the initial mixture data matrix X into the product of
two chemically meaningful data matrices C and S T . Each of these matrices contains
51
Robustness Marta Padilla Ferran
the pure response profiles of the n mixture components associated with the row and
the column directions of the initial data matrix, respectively (fig.). In matrix nota-
tion, the expression for all resolution methods is based on a standard bilinear model
of the form:
X = CS T + E (2.6)
where X(r × c) is the original data matrix, C(r × n) and S T (n × c) are the ma-
trices that contain pure-component profiles related to the data variation in the row
direction and in the column direction, respectively, and E(r × c) is the error ma-
trix, i. e. the residual variation of the data set that is not related to any chemical
contribution. The variables r and c represent the number of rows and the number of
columns of the original data matrix, respectively, and n is the number of chemical
components in the mixture or process. Matrices C and S T often refer to concen-
tration profiles and spectra, but they can refer to other factors in other problems
[222].
On the other hand, MCR methods include two steps; a first step finds an initial
estimates for matrix C or S T and the second one, often ALS, refines the previous
estimation step and incorporate the required constraints. Initial estimates can be
obtained by techniques of local rank analysis or by procedures for pure variable se-
lection [218]. Local rank analysis performs repeated PCA analyses in small parts of
the data set in order to know how the number and distribution of components evolve
along the data set. Evolving Factor Analysis (EFA) [223] and Fixed-Size Window
Evolving Analysis (FSWEFA) [224] are some examples. Furthermore, these tech-
niques are particularly relevant in the study of processes, where the concentration
profiles of the different components evolve smoothly and, often, following a sequen-
tial pattern. On the other hand, methods of pure variable selection find the most
representative row or column profiles for the different components in the data set
to determine the number of components. Simple-To-Use Interactive Self-Modelling
Mixture Analysis (SIMPLISMA) [225] is the most commonly used technique of this
52
Signal processing methods for improving robustness
type [218].
In AO systems MCR has never been applied until now. In this thesis, we present
an application of MCR-ALS where, given a dataset of measurements by an array of
gas sensors, with pure and mixture samples, information about the concentration
of the pure analytes and temperature sensitivities are obtained in matrix C and S
respectively.
According to the different operation modes of the AO instrument, every single sen-
sor’s response to temperature modulation or transient to a sudden exposition to the
target analyte, can be obtained (see section 1.5). The inclusion of any of these extra
information about all sensors in a data set, gives place to an increase on the di-
mension of the feature space, which therefore implies richer information. Such data
set can be organised in the form of a cube or a three way matrix (fig. 2.1), e. g. a
tensor where every edge or mode; row, column and tube (fig. 2.2)), refers to either
the samples, the sensors and every sensor’s transient signal or response to temper-
ature modulation. Even a four way matrix with dimensions samples × sensors ×
temperature modulation response × transient response can be considered.
The structure of a multiway data matrix can be preserved and exploited by using
multiway signal processing techniques. These methods are often used for preprocess-
ing. They provide more parsimonious models than classical methods, and extracted
features are meaningful and easier to interpret. A deep study on multiway methods
53
Robustness Marta Padilla Ferran
can be found in [116]. Also, De Juan and Tauler make a comparison of three-way
resolution methods for non-trilinear chemical data sets in [234].
The most common multiway techniques are parallel factor analysis (PARAFAC) and
Tucker, and for both extensions and modifications have been proposed (PARAFAC2,
Tucker3, etc. ).
For three way datasets PARAFAC is the alternative to PCA, however PARAFAC
requires the dataset to be trilinear and this is not always the case. If the multi-
way matrix is trilinear, it can be decomposed into three two-way matrix, each one
referring to a mode and having all the same number of factors. PARAFAC2 is a
modification of PARAFAC but it lets the second mode to have its own set of load-
ings. On the contrary, Tucker3 method does not require trilinearity but the model
building and final interpretation become much more complex. Tucker3 models de-
compose the multiway dataset into three matrices and another multiway matrix,
named core matrix G, containing the non trilinear part of the original dataset and
allowing different interactions among the modes. The decomposition of the three-
way matrix X, according to some of the multiway methods, is given by:
F
X
P CA : xij = aif bjf + eij (2.7)
f =1
F
X
P ARAF AC : xijk = aif bjf ckf + eijk (2.8)
f =1
X F X
F X H
T ucker3 : xijk = aid bjf ckh gdf h + eijk (2.9)
d=1 f =1 h=1
where xijk represents the ijkth element in the three-way data set, aif , bjf , ckf are the
elements in the decomposition matrices A, B, and C (see fig. 2.3 where the different
methods are represented graphically), eijk is the residual term and gdf h is the df hth
element of a core array (in Tucker model) of size D × F × H. D, F and G are the
selected number of factors in each decomposition matrix.
54
Signal processing methods for improving robustness
Skov and Bro [235] applied PARAFAC2 on an e-nose dataset with dimensions
samples × transient time × sensors. This method was preferred to classical
PARAFAC because of its ability to handle shifted sensors transient profiles. The au-
thors compared this technique with PCA and conclude that, although PARAFAC2
offers some advantages compared to PCA, the structure of the model required a
more careful interpretation of the model parameters.
When classical methods are to be used on the three-way matrix, which is the usual
approach in AO, this is unfolded often in a row-wise way (fig. 2.4), and therefore the
relationship among the three modes is destroyed and interesting information lost.
Processes and applications that require continuous monitoring run over long periods
of time, during which drift is not the only problem the instruments may present, they
also need maintenance, parts of them have to be repaired or have to be replaced, etc.
This means that calibration has to be repeated, implying the repetition of all the
tasks involved in it; experimental design, new data acquisition, data evaluation and
preprocessing, model building and validation, and finally application of the selected
model to predict responses of unknown samples. However, good calibrations require
many reference samples, depending on the instrument and the model complexity,
55
Robustness Marta Padilla Ferran
which must represent all possible situations that may arise during later operation
of the instrument. Other situation may arise when, after considerable effort placed
into constructing a robust model, the samples to be predicted are measured on a
different instrument or under differing environmental factors from those used to build
the model. Further, many identical or almost identical instruments can operate in
many locations in an industry. In this case, one often has a master instrument and
a number of field instruments, and thus, along their life time, regular re-calibrations
have to be run on each of the instruments.
Therefore, there exist many situations in which calibrations have to be repeated very
often, which is a time consuming and an expensive task and also a source of errors.
The aim of the field named calibration transfer is to minimize the cost and time
expense of re-calibration by transferring the information contained in the model of
a first instrument to a model for a second instrument using as few measurements as
possible. This is a problem for many different instrument manufacturers, and this
topic is expected to expand as calibration spreads into more complicated industrial
applications of continuous monitoring.
The transfer of calibration is made from one instrument to a slightly different (almost
identical) instrument. Being that, for identical instruments calibration models are
equal, and completely different instruments have no information in common, and
thus all possible situations have to be measured also on the second instrument.
When the instruments are similar, it is possible to assume that measurements in
similar environments change in a similar way for the second instrument and thus,
with a few new known measurements in both instruments, a new model can be built
for the second instrument [136].
Different methodologies can be followed in order to counteract for the data shift
between two instants measurements in the same instrument or measurements made
in two instruments. Feudale et al. [199] makes an interesting overview of these
methodologies in the area of near-infrared (NIR) spectra.
56
Signal processing methods for improving robustness
If these methods are applied in an adaptive way, pre-defined and well-known ref-
erence samples are needed. Adaptive methods continuously update the model at
certain intervals or use procedures which automatically identify changes and ap-
ply counteraction. Therefore, reference samples are needed in order to identify the
undesired changes. These reference samples must be carefully chosen; they must
be representative enough to describe the differences between the instruments, and
should ideally represent the entire experimental domain. Unfortunately, they tend
to be hard to obtain because of different sample conditions, samples degradation, en-
vironmental changes, etc. Methods that need such reference samples are often called
standardization methods and many techniques coming from the field of chemometrics
can be used, but, when transfer standards are not available, preprocessing methods
can be employed.
57
Robustness Marta Padilla Ferran
FDI methods can follow two different approaches; model-driven and data-driven.
The model-driven family of FDI techniques is most commonly based on models
that describe the physical and chemical principles underlying the process. However,
the development of these models requires a lot of process expert knowledge which
is not always available. Furthermore, since they are specifically developed for the
design of the processing plant, they usually focus on the optimal steady-state of
the process and are thus not suitable for the description of any transient states.
Besides, depending on whether the system model is represented as either a state
space or an input-output model, FDI can be classified into observer based, system
based and also adaptive observer based, this latter combines the best features of the
two previous approaches. In general model based FDI methods can only be applied
to low dimensional systems [240, 244].
On the other hand, data-driven based FDI can deal with high dimensional data,
and uses data dimension reduction techniques to highlight the important informa-
tion in the data. Yet, there are still some challenges in its use in FDI, as in cases
where the system is time-varying and highly non-linear for example. At the same
time, data-driven FDI techniques can be divided in other three groups; signal based,
multi-variable statistics based, and knowledge based. The common feature of these
methods is that they all use raw system data and sometimes process knowledge to
carry out the required FDI. Signal-based FDI methods use signal processing meth-
ods such as correlation functions, signal model identification, signal parity checks,
spectral analysis using fast Fourier transform and wavelet transform. Knowledge
based methods include artificial intelligence algorithms, including expert systems
58
Signal processing methods for improving robustness
[245].
In multi-variable statistics based FDI methods, techniques like PCA can be used
to reduce the high dimensionality of the data in a few features containing most of
the information of the original data. PCA divides the whole sensor space into a
principal component (PC) subspace and a residual subspace, and then performs the
FDI using Q-test and Hoteling T 2 test. This is done by checking whether or not the
new incoming data lies in the PC subspace through the mentioned statistics given
by the squared prediction error (SPE) or Q (eq. 2.10), and T 2 (eq. 2.11):
SP E = kx̃k2 (2.10)
T 2 = xT P Λ−1 P T x (2.11)
where x is the new incoming data, x̃ is the residual of x, P is the matrix of loadings
(PCs) and Λ is the diagonal matrix of eigenvalues. To identify the faulty sensor, the
PCA residuals have to be analysed, even methods for structuring its residuals have
been proposed in order to improve the accuracy of the identification task [245].
PCA is a powerful and robust technique that have been used in many different ap-
plications. It is able to detect process or sensor faults in large multivariate data sets.
However, PCA assumes that the data is (multivariate) Gaussian distributed, which
limits its application in complex industrial systems that exhibit a time-varying,
non-Gaussian and non-linear characteristics. Alternatively, independent component
analysis (ICA) does not assume the data to be Gaussian distributed and other meth-
ods can deal with non-linear characteristics.
59
Robustness Marta Padilla Ferran
60
Chapter 3
Thesis objectives
The aim of this thesis is to investigate signal processing methods that may improve
the robustness of AO systems. In particular, the main focus is the time stability
and sensor fault detection, identification and correction.
To improve time stability, we take a deeper look to techniques able to find restricted
subspaces that reject drift while keeping the predictive power of the models. We
will investigate the family of Orthogonal Projection methods, specially Orthogonal
Signal Correction, in their capacity to reduce drift but also unwanted variance. Its
abilities will be compared to state of the art techniques in AO like Component
Correction.
Sensor poissoning detection has received poor attention by the AO community. This
thesis proposes automatic methods to detect and correct faults in chemical sensor
arrays based on the inner correlation of the sensor signals.
61
Thesis objectives Marta Padilla Ferran
62
Chapter 4
Some of the results obtained during the course of this thesis are reported in four
journal and two conference papers, where the reader can find specific details con-
cerning the experiments, methodologies and employed data processing techniques.
A short introduction detailing the main contribution of each paper to the improve-
ment of the robustness of AO systems is presented next, before the compilation of
the journal papers.
As already mentioned, chemical sensors are prone to drift, which causes statistical
models to degrade in time. This is shown in this paper, where an electronic nose is
used to recognize building materials infected by moulds on two measurement series
separated by about four months.
A first goal in this work, is to perform several tests in order to ascertain several
questions relative to the detection of moulds by an array of ten commercial MOX
sensors. Studied issues includes the influence of every type of building material and
the the specie of mould on the final recognition of the substrate as infected or not,
and also whether or not an unknown specie of mould could be detected as well as
the presence of moulds could be detected when growing over an unknown building
material.
63
Journal and main Conference Papers Marta Padilla Ferran
The second objective is to test the time stability of the instrument in this application.
For this, a similar data set was collected four months after the first data set, from
which a classification model was built. Results from this second data set reveals
the degradation of the classification model caused by drift. Although the overall
classification rate is not too low compared to the one obtained in the first data set,
some individual rates degrade strongly depending on the mould specie and type of
the substrate being considered. This also illustrates the fact that drift influences
individual sensors in different ways, as was noted in section 1.6.3, since the evolution
of data along the time depends on the class they belong to.
As seen in section 1.5.1, the use of the full transient signal of the sensors response
increases the predictive capabilities of an electronic nose. When transient signals
from all sensors in the array are arranged in a three way matrix, data can be pro-
cessed by multiway techniques, which are reported to provide more parsimonious
and easier interpretation of the models. Extracted features can be used to build
posterior models for prediction.
In this paper, the multiway method parallel factor analysis (PARAFAC) is applied in
an application for quality control of a processed food product, fried potato chips with
different concentration of a flavouring agent. The measurements were performed
with a commercial electronic nose containing 13 metal oxide sensors. Later, only
one of the obtained features from PARAFAC is used to build an inverse least squares
(ILS) model, from which the concentration of aroma of new incoming samples can
be calculated. Remaining loadings from PARAFAC give information about the
influence of every individual sensor on the model and identify the time window
when the the sensors are exposed to the volatiles.
It has to be noted that, before the application of PARAFAC, the trilinearity property
required by PARAFAC must be tested. Furthermore, a number of factors for the
PARAFAC model have to selected. For this, several models of the calibration set are
made with different number of factors. Then, once the number of factors is selected,
a robustness test is made by repeating the procedure several times to confirm this
choice.
64
Introduction to papers
With the aim of exploring the ability of OP techniques to diminish the influence
of drift, in this paper it is proposed the use of one of the OP methods, orthogonal
signal correction (OSC) algorithm, to compensate drift on an experiment lasting
10 months. This experiment consists of measurements of three type of analytes at
several concentration levels by using an array of 17 conducting polymer gas sensors.
In addition to OSC, Arthursson’s component correction (CC) method, explained
in section 1.6.5, is tested on the same dataset in order to be compared with OSC
and with prediction when no preprocessing stage to counteract drift is employed.
Results are presented in the form of classification rates on the three type of analytes,
however it has to be noted that also information related to the concentration values
are used in the design of both CC and OSC techniques. Besides, it is also important
to note that the whole transient signals of the sensors have been used, since they
provide additional information for class discrimination, as mentioned in section 1.6.
Again, PCA score plots are very interesting, as they show firstly the effect on short
and long term drift on the data, and then the reduced class clusters obtained once
the preprocessing methods are applied.
65
Journal and main Conference Papers Marta Padilla Ferran
First, MCR-ALS is applied on synthetic data and then on real data. The analysis
results in a resolution of the gas mixture that gives C and S T profiles. Satisfying
results are obtained from synthetic data, however in real data non-linear effects are
observed.
Further, the analysis is made on measures from a single sensor, and later on data
containing responses from all sensors in the three way matrix (samples × sensors
× temperature modulation wavef orm), which is unfolded into a two way matrix
to be analysed.
Since in an array of sensors one or more may fail, it is important to detect the
fault on time, identify the failing sensor and correct its response until the sensor can
be replaced by anew one (section 2.2.5). In this work, a possible and subtle type
of fault typically affecting chemical gas sensors, is simulated. The poisoning fault
causes a change of the sensitivity profile of the affected sensor. Poisoning is difficult
to detect compared with other type of faults, because the sensor keeps working under
apparently normal conditions.
In this work, we proposed the use of simple statistical methods such as PCA, PLS
and PCA with structured residuals, to perform the three above mentioned tasks of
detection, identification and correction of one and two faulty sensors in an array of
17 conductive polymer gas sensors. All three techniques are compared in every step
by a receiver characteristic (ROC) curve for detection ability evaluation, percent
66
Full papers
of accuracy in the identified faulty sensor and classification rates before and after
correction of the faulty sensor’s response.
Also in this work, we test the ability of a signal processing method to detect, identify
and correct one or two poisoned sensors in an array, but with an additional difficulty
in which the system is subject to irregular drift. Therefore, since the method has
to deal with drift, the tasks of detection, identification and correction of the faulty
sensor become more challenging.
In this work, we propose the use of multiscale-PCA for the detection, identification
and correction of a faulty sensor under drift conditions. The multiscale step consists
on the prior decomposition of every sensor response along the time (sample mea-
sures) by a discrete wavelet transform (DWT). DWT allows to isolate the slower
varying part of the signal related to drift, from the part related to the information.
Then, the subset of coefficients corresponding to this information content from all
sensors are grouped in a matrix, to which the same PCA method that in section 4.1
is applied to perform the detection, identification and correction tasks.
67
Sensors and Actuators B 119 (2006) 33–40
Abstract
This work explores the detection of moulds growing in different building materials by using a metal oxide sensor array. Four moulds species
have been considered. Pattern classification provides detection rates on the order of 80–85% for different species. Drift degrades only slightly these
values subsequent test 4 months later.
© 2005 Elsevier B.V. All rights reserved.
0925-4005/$ – see front matter © 2005 Elsevier B.V. All rights reserved.
doi:10.1016/j.snb.2005.02.059
34 M. Kuske et al. / Sensors and Actuators B 119 (2006) 33–40
2. Materials and methods Agar, for 1 week. After this period, moulds were inoculated into
the building materials. Contaminated samples were incubated
An electronic nose instrument was designed and constructed in airtight glass jars (500 ml) closed with a Teflon cover. Each
at the Université de Liège. It consists of a sampling unit, a sen- jar contained two pieces of the same material of dimensions
sor array, and a signal processing system (Figs. 1 and 2). The 5 cm × 6 cm contaminated with one fungal species and 40 ml of
gas sensor array contains 12 metal oxide sensors in two sep- demineralised water in its bottom, for keeping the high degree
arate chambers placed in parallel. Each sensor chamber has of humidity that moulds need to grow. The samples were placed
a volume of 210 cm3 . The sensors were selected in order to on top of small glass vials to avoid direct contact with water.
obtain different responses for volatile compounds produced by The jars were kept at room temperature in the darkness during
moulds. Six of them were manufactured by Capteur; CAP01, the whole duration of the experiments.
CAP03, CAP06, CAP07, CAP23, CAP25, and six from Figaro; For each fungal species, two sequential experiments were
TGS2620, TGS2180, TGS825, TGS822, TGS2600, TGS2602. carried out, resulting in two datasets comprising the growth of
Four different materials typical of Belgian houses were used the moulds during 4 months. Clean and contaminated materials
as substrates for mould growing: plasterboard, particleboard (or (corresponding to different species) were randomly presented
chipboard), oriented strain board and wallpaper. Some combina- to the gas sensor array during this time. Contaminated samples
tions of them were also analysed; plasterboard with wallpaper were presented in order according to the species of moulds, in
and glue, particleboard with wallpaper and glue and oriented this way firsts sets of A. versicolor followed by P. chrysogenum
strain board with wallpaper and glue. The four fungal species were firstly measured, and then seconds sets of both species
were separately incubated on culture medium, Malt Extract in the same order. After that first sets of C. sphaerospermum
followed by P. aurantiogriseum were presented to the electronic
nose and finally seconds sets of the two species were measured
in the same order. The contents of all datasets are summarized
in Table 1.
The electronic nose has the following operation; clean dry
air is fed to a mass flow controller, then to a humidifier and then
to a jar with samples to sweep the volatiles accumulated in the
headspace. Two three-way valves commute clean air or the jar
headspace to the gas sensors chambers (see Fig. 1). A sample
cycle included a passage of clean humidified air (reference air),
followed by a passage of air from the samples during 3 min (time
of signal stabilisation) at 200 ml/min of constant air flow.
Table 1
Number of clean and contaminated samples for each substrate material and specie of mould in the two datasets
Materials Samples Type of mould
First set Second set First set Second set First set Second set First set Second set
vapour saturated carrier gas. In this way every measurement was For this reason only the Mahalanobis results are reported in this
constructed with 12 sensors signals, assembling a 12 dimen- paper.
sional vector. However two sensors (CAP23 and TGS2180)
were removed because of deficient behavior observed by visual 4. Results
inspection of the sensor patterns, therefore the formed pattern
vector has 10 dimensions. A PCA loadings plot on first two principal components
As already mentioned, two datasets were collected for this (Fig. 3), for the ensemble of all moulds, shows one group of five
study for each species. First dataset was used for all the data Figaro sensors and one Capteur sensor on the left hand side and
processing algorithm development, and the second one for a two groups of two Capteur sensors on the right. PCA loadings
posterior evaluation (including the evaluation of drift), since it plots for every specie of moulds are almost the same. Signals
was collected 4 months after the first one. provided by the five Figaro sensors (series TGS) are very similar
For validation of the classifier results on first dataset the fol- and anti-correlated to Capteur sensors signals on the first princi-
lowing strategy has been followed. This dataset (for each mould pal component (except for CAP01). CAP01 signal is interesting
species or all together) has been divided in two equally sized ran- because it varies inversely with the rest of Capteur sensors on
dom partitions. One part has been used for classifier design and PC1 and with all Figaro sensors on PC2. CAP03 and CAP25
optimization while the second half is used for classifier testing.
This process is repeated ten times and the average classification
rate is retained. Also the standard deviation of the classification
rate is estimated.
Later, the robustness of the electronic nose for mould detec-
tion was tested by including a wide variety of moulds and
building materials. Classification of each specie of mould is con-
sidered, as well as the capability of a classifier to generalize to
unknown mould species (not present in the training set). Finally
also the classification of samples with an unknown substrate was
explored.
Our primary objective was the binary classification of moulds
and no-moulds disregarding the particular fungal species and the
substrate material. For this purpose a Mahalanobis classifier was
applied on the raw data. The obtained results are on the same
order than the ones seen in the previous paper [19]. Although the
same classifiers as in [19] were considered; namely k-NN, fuzzy-
kNN with different initial membership functions on dimensional
reduced space and on raw data (a review of Fuzzy k-NN can be
found in our previous paper [19]) we found that the application
Fig. 3. PCA projection, loadings plot for all moulds together. Samples from
of complex classifiers on our small database did not provide any
first set of all moulds. PC1 and PC2 explains 82.27% and 10.88% of variance
advantage; the obtained classification rates were not in general respectively. The points corresponds to the five Figaro sensors and the plus
better than the ones obtained using the Mahalanobis classifier. symbol represent the Capteur sensors.
36 M. Kuske et al. / Sensors and Actuators B 119 (2006) 33–40
Fig. 4. PCA projection. Samples from first set of all moulds. PC1 and PC2 Fig. 6. Same PCA projection as in Fig. 4 regarding the four different species of
explains 82.27% and 10.88% of variance respectively. moulds on first dataset of every mould.
are not very significant for PC2, the rest of sensors gives infor-
mation to both first two principal components. This loading plot Fig. 4. Wallpaper and particleboard are mixed between them
also shows that Figaro sensors are more correlated among them but separated from oriented strain board and plasterboard. Due
than Capteur sensors, at least for this application. to the important background variability the detection of moulds
A PCA projection, a scores plot in the first and second becomes a difficult problem.
principal components, seems to indicate that the classes (con- In Fig. 6 the distribution of species of moulds on the PCA
taminated/clean) are no linearly separable, non-gaussian and projection is observed, the four types of moulds are completely
multi-modal (Fig. 4). Multi-modality clearly arises from the mixed so we can think that the classifications rates of each specie
presence of different substrate materials. In fact, visual explo- will be similar. This can be seen in Table 2, where the obtained
ration of the results shows that the influence of the substrate is classification rates for every type of mould are shown. All results
much larger than the influence of the presence of moulds. Fig. 5 are on the same order, with small differences. From these results
shows the distribution of the four substrates on the same PCA we can also think that the detection of a specie of mould not
projection than Fig. 4. The distribution of materials seems to present in the training set is possible.
be more separable than the contaminated and clean classes in To study the stability of the classifier a second experiment
for each specie of mould was performed in the same conditions
as the first one resulting in a second dataset. This experiment
started 4 months after the first one and was carried out dur-
ing more than 3 months. The Mahalanobis classifiers built with
firsts datasets were used to classify the samples from the sec-
ond datasets. Table 2 shows the obtained classification rates for
every type of mould and all moulds together. These results show
Table 2
Classification rates for every specie of mould and all moulds together, on first
and second datasets
Type of mould First set (%) Second set (%)
In ‘first set’ Mahalanobis model was built with half of the first dataset and
Fig. 5. Same PCA projection as in Fig. 4. Samples from first set of all mould. validated with the other half of first dataset, this was repeated 10 times mak-
Each color represents different substrate; plasterboard, particleboard, oriented ing random subsampling. In ‘second set’ the model was constructed with the
strain board, wallpaper. completed first dataset and validated with the second dataset.
M. Kuske et al. / Sensors and Actuators B 119 (2006) 33–40 37
Table 3
Classification of a specie of mould not considered in the training samples
Train samples (three species of moulds) Validation samples C.R. (%)
Table 5
Classification of moulds according to growing substrates
Type of material First set (%) Second set (%)
Wallpaper 74 ± 6 72
Plasterboard 83 ± 4 71
Chipboard 89 ± 3 82
Fig. 7. PCA projection from first set of all moulds. Points are samples from Oriented strain board 85 ± 4 72
second set of all moulds, circles are samples from first set of all mould. PC1 and
PC2 explains 82.27% and 10.88% of variance respectively. Samples from all species of moulds.
38 M. Kuske et al. / Sensors and Actuators B 119 (2006) 33–40
Fig. 8. Effect of sensors drift on a PCA projection of the two datasets on first set PCA model for every specie of mould. (a) A. versicolor PC1 81.16%, PC2 11.98%
variance explained. (b) C. sphaerospermum PC1 86.19%, PC2 8.66%. (c) P. aurantiogriseum PC1 86.11%, PC2 8.32%. (d) P. chrysogenum PC1 81.61%, PC2
11.51%.
5. Conclusions effect on the sensors depends on the specie of mould that is being
classified; A. versicolor has a classification rate of 77% while C.
In this work an electronic nose has been able to detect four sphaerospermum has 55%. However the obtained result for the
species of moulds growing on several building materials with set made from all moulds was 81%, so, in general, the sensors
a Mahalanobis classifier, obtaining an accuracy between 80 drift causes a small degradation without any further recalibration
and 85%. The use of other more complexes classifiers gave no of the system.
much better results, possibly this is due to the small size of our However, due to the modest values of classification rate we
database. consider that such a system could only be used for screening
The system is able to detect unknown species of mould with purposes followed for a more detailed analysis using alternative
accuracy between 73% and 86%, depending on the specie of analytical techniques.
mould. It can also detect contaminated samples of unknown
building materials, depending also on the type of material. The
obtained minimum accuracy was 67% for wallpaper samples and Acknowledgements
the maximum was 86% for chipboard. This accuracy is worse
than the one for the detection of unknown moulds because, as Spanish team has been funded by Cycit DPI2001-3213-C02-
seen in Fig. 5, the influence of the substrate materials is bigger 01 and Generalitat de Catalunya ACI2003-13. Belgian team has
than the influence of the different species of moulds. been funded by Region Wallone. One of the authors R.R. also
The stability of the designed classifier was tested with an acknowledges a PhD grant from the Spanish Ministry of Science
experiment made in the same conditions but 4 months later, and Technology. The assistance of CERTECH is also acknowl-
in which a second set of data was collected. The observed drift edged.
M. Kuske et al. / Sensors and Actuators B 119 (2006) 33–40 39
Santiago Marco Colás was born in Berga in 1965. Degree (1988) and Doc- “Tor Vergata” working on intelligent signal processing. Since 1995 he’s
tor in Physics (1993) at the University of Barcelona doing research on pres- an Associate Professor at the UB. His experience is centered in microsys-
sure sensors for biomedical uses which took place at the Centre Nacional tem modelisation, and in signal processing-based intelligent instrumenta-
de Microelectrònica. In 1994 he was post-doc at the University of Rome tion.
Sensors and Actuators B 116 (2006) 145–150
Abstract
When enose signals are analysed, the signal processing phase plays an important role in the quality of the end results. With the aim of getting
more reliable information, the proposal of incorporating the whole transitories of each gas sensor simultaneously recorded, to build a three-way
dataset seems to be a good option. But, anyway, this strategy must be accompanied by suitable signal processing/feature extraction of the data in
order to achieve stable solutions.
In this work the possibilities of the use of parallel factor analysis (PARAFAC) as a data compression technique suitable to deal with trilinear 3D
data arrays are shown. To exemplify its performance, a quantitative case focused on food analysis has been selected. The results obtained point out
the suitability of the technique to achieve a good predictive ability by using a simple inverse least squares (ILS) calibration onto a set of synthetic
samples.
© 2006 Elsevier B.V. All rights reserved.
0925-4005/$ – see front matter © 2006 Elsevier B.V. All rights reserved.
doi:10.1016/j.snb.2006.03.011
146 M. Padilla et al. / Sensors and Actuators B 116 (2006) 145–150
psychometrics [11,12] PARAFAC is gaining interest because sample into the cell and further purging of the system), and all
it is a processing technique that simultaneously determines the fluidics and valves necessary to properly deliver the different
the pure contributions to the dataset, optimizing each factor gas flows. All this system is controlled by embedded electronics,
as a time, in trilinear systems. Widely explored in different which in addition is in charge of the first signal treatment.
fields, such spectrofluorimetry, spectrophotometry, chromatog-
raphy and so on, it can deal with the increased complexity of 2.2. Experimental
process control data in science and industry [13,14].
Towards other related decomposition methods, such Tucker3 Thus, starting up from raw fried potato chips and a little
[15] and unfolded PCA [16,17], PARAFAC belongs to the amount of flavouring agent submitted by a Spanish food industry,
same family of multi- or bilinear methods of decomposition a set of 30 synthetic samples was prepared in different ses-
of multi-way data in a set of loading and score matrices. This sions. These samples were obtained by weighting of 1.00 g of
process allows to describe the initial variability decreasing its fried potato chips and further addition of different amounts of
dimensionality, thus leading to a compressed form. In this pro- flavouring agent, in order to span the percent of additive desired:
cess, PARAFAC uses less degree of freedom than Tucker3 or between 1% and 8% in weight. After proper weighting of the
unfolded PCA methods. This intrinsic characteristic leads to flavouring agent, the samples were tight closed in a glass vial
simpler models with less fit, but avoiding the incorporation of and properly stored.
non-significative effects such noise and redundant information The measurement of the headspace of the vials was
to the model [18]. done using a Nordic Sensor Technologies System, NST3320
This paper presents an illustration of the use of PARAFAC (Linköping, Sweden) with its accessories. This instrument con-
for feature extraction on a multi-way dataset obtained from an tains a gas detector array of 23 sensors, divided in 10 MOSFET
electronic nose. To check its performance, a quantization case catalytic field-effect sensors (not used in this work) and 13 MOX
for quality control of processed food products is studied. sensors. This unit is capable to deliver the headspace of the
different vials into the sensors chamber, after properly thermo-
2. Case statization of the sample, and to perform appropriate cleaning
operations of the measurement chamber.
2.1. Presentation In this work, the measures were done keeping a con-
stant temperature for all the samples of 40 ◦ C. After half an
The strong competition in the snacks industry has leaded to hour of thermal stabilization, the data were acquired follow-
the companies to diversify their offer as well as possible. In this ing a baseline–injection–recovery–flush–stabilization cycle of
way, the companies are offering variants of traditional products 15–10–10–120–110 s each, employing a sampling rate of 1 Hz.
with different tastes. Thus, at the present time, fried potato chips The data thus obtained, corresponding to an increase in the con-
being prepared with many tastes (‘ham’, ‘spicy’, ‘barbecue’, ductance value of the sensors, was properly stored for each of
etc.) the 13 MOX sensors and folded for each sample, in order to
The new products are often achieved through the addition generate a three-way matrix X of dimensions (30 × 13 × 265)
of alimentary flavours with different formulations, which incor- (samples × sensors × time), containing the described sinthetic
porate salt in a certain number of cases, and are part of the samples and its complete transitories. In a second step this
know-how of these companies and/or their suppliers. The addi- database was split in two, to obtain two differentiated subsets
tion process of these flavours is of utmost importance in order to to be further processed. Both subsets were prepared spanning
obtain a product of quality, with enough homogeneity between homogeneously the same concentration intervals, and recorded
batches. In spite of the importance of this phase, does not exist in different sessions. The number of samples of each of the two
a direct control of the addition of the alimentary flavour. At the subsets (calibration and validation) was of 20 and 10 samples,
present time, the process is only indirectly controlled in line by respectively.
properly weighting of the additive added, without any real con- In Fig. 1 can be seen the plot of the raw signal for sample 1
trol of the taste of the final product. Although off line analysis of the calibration subset. Data have been accordingly resized to
are done, through salt control in the laboratory of selected pro- avoid a confussing view. A transitory signal is obtained, present-
duction samples, this is only an indirect measure of the flavour ing different responses for each of the 13 MOX sensors, with
addition, and valid only if the flavouring agent contains salt. different behaviours along the measurement time.
Electronic olfaction appears as a possible alternative to quan-
tify the addition of this flavouring agent, starting up from the 3. Results
unspecific measure of the detached volatiles. This kind of mea-
surement presents the main advantage of its direct relationship To test the applicabilty of the PARAFAC method as a feature
with the magnitude to be determined (intensity of flavour/odour) extraction method from three-way datasets, the first step done
and provide some possibilities to its continuous implementation, was to check the inner trilinearity of the data. With this aim,
even tough it will come joined with appropriate modifications the calibration subset was properly row-wise (20 × (265 × 13)),
of actual instrumentation. These systems are often composed of column-wise ((20 × 265) × 13) and tube-wise ((20 × 13) × 265)
a suitable sampling system, a measurement cell (containing the unfolded to a two way array. If there exists trilinearity on the data,
sensor array), a pumping device (responsible of introducing the the number of estimated factors for both expanded arrays should
M. Padilla et al. / Sensors and Actuators B 116 (2006) 145–150 147
Fig. 1. Plot of enose signals of each sensor corresponding to sample 1 of cali- Fig. 2. Determination of the number of factors. Core consistency test results of
bration set. each of the three replicates calculated for each factor.
be nearly the same [19]. After SVD calculation on the centered To do this test, starting up from centered data, an exploration
data, they appeared to be the same (five) for each augmented of the different possible PARAFAC models was done, varying
array (see Table 1), thus confirming the trilinearity of the dataset. the number of factors and repeating the procedure by triplicate,
Taking the calibration subset as a basis, and before PARAFAC in order to check the stability of the results. After using the
compression were applied, several signal preprocessing tech- core consistency diagnostic [23] as an indicator of the success
niques were tested. At last, among others (smoothing, deriva- of the different models prepared (see Fig. 2), it was found that
tives), an offset correction based in the use of the average of the number of factors which better describe the data should be
the first 15 min of the transitories of each sensor was applied to three.
remove constant effects from run to run. Following this first step, Once this three factor model was selected, and after plotting
standard nomal variate (SNV) [20–22] was also applied to cor- the different loadings (scores and second and third mode ones),
rect the dispersion of the data. Both pretreatments were applied interesting tendences are observed across the scores (loadings
to the third mode (time), and were the ones which offered more of the first mode), as it can be seen in Fig. 3. There appeared
promising results. Because processing of the datasets along one three factors which shown increasing behaviours as the sample
or more modes can break its inner trilinearity condition, it must concentration increased. These increases describe, but, different
be ensured through a new check after the preprocessing step. In profiles, depending on the sample involved.
this case, the trilinearity of the dataset was re-evaluated render- Second loading matrix plot show the contribution of each
ing results that can be assumed as sufficiently positive. Once the sensor to the global signal. From this plot is possible to deter-
pretreatment was applied and to each variable slice, and veri- mine which are the sensors which are contributing with more
fied the trilinearity of the dataset, three-way data were centered information to each factor. As it can also be seen, these differ-
across the first mode (samples) before the testing of PARAFAC ences become milder as the number of factors increases. In the
models. case of the third loading plot, it shows important differences
Table 1
Trilinearity check
Eigenvalues of augmented data matrices
Before correction After correction
Fig. 3. Scores and loadings plot for the three factors PARAFAC model for each
of the three ways: samples (A), sensors (B) and recording time (C).
Table 2 for this dataset, rendering a root mean square error of prediction
ILS quantitative results of PARAFAC processed data (RMSEP) of 0.008.
Calibration Validation After obtaining this results, which seem to be fairly good,
Reference (g) Calculated (g) Reference (g) Calculated (g)
a residuals plot was done. As it can be seen in Fig. 3, the
distribution presents an homogeneous pattern across the cali-
0.000 0.006 0.000 0.009 bration range. The overall predictive results obtained (RMSEC
0.041 0.044 0.042 0.034
0.046 0.052 0.047 0.049
and RMSEP) agree with the absolute error (A.E. = ±0.01g) in
0.053 0.060 0.050 0.049 the determination of weights.
0.056 0.067 0.055 0.058
0.061 0.060 0.060 0.044
4. Conclusions
0.070 0.074 0.065 0.059
0.072 0.063 0.071 0.062
0.079 0.061 0.030 0.028 In this work it has been studied the use of PARAFAC as a
0.006 −0.001 0.036 0.047 feature extraction technique to perform a data reduction from
0.012 0.007 a specific three-way dataset. Previously to this extraction, a
0.015 0.020
suitable data pretreatment have been applied to obtain properly
0.019 0.023
0.024 0.030 quantitative results for our experimental case are desired. Even
0.029 0.033 though SNV is known as a processing method in other fields
0.038 0.035 prone to introduce nonlinearities in the datasets, it has revealed
0.039 0.039 itself to be the most suitable one, jointly applied with offset cor-
0.042 0.044
rection. Its application in this case has been effective, keeping
0.047 0.039
0.057 0.052 to a certain extent the trilinearity of the processed dataset.
Taking the quantitative case results as indicators of the perfor-
RMSEC 0.007 RMSEP 0.008
mance of this use of PARAFAC, the predictive errors provided
acceptable RMSE values, taking in mind the sensitivity of tech-
nique, and the amounts range. In this sense, PARAFAC performs
thus confirming the precision of the method. To check the accu- a suitable data compression task, incorporating the most of the
racy, a method comparison regression line (y = a + bx) between information of this complex data system. Although this work is
the calculated and the reference values was calculated. The focused in a quantitative preliminary approach, further work in
this direction can be done on the application of PARAFAC as
I 2
i=1 (X̂i − Xi )
a feature extraction method for its use for pattern recognition
RMSE = (1) purposes and on its application to more complex samples.
I
obtained results for the calibration set leaded to a linear equation:
y = 0.004 + 0.902x. To define the suitability of this comparison, References
the confidence intervals for this line were checked to ensure the
[1] E. Oja, Subspace Methods of Pattern Recognition, Research Studies
solution did not show any kind of systematic deviation. Confi- Press, Letchworth, Hertfordshire UK, 1983.
dence intervals obtained in this case for each parameter were for [2] E. Oja, J. Karhunen, Recursive construction of Karhunen–Loeve expan-
a ±0.007 and b ±0.147. Both intervals confirm that compared sions for pattern recognition purposes, in: Proceedings of the Fifth
methods (weighting of flavouring agent and ILS calibration on International Conference on Pattern Recognition, December, Springer-
Verlag, NY, 1980, pp. 1215–1218.
PARAFAC scores from enose data) leaded to similar results
[3] A. Pardo, S. Marco, C. Calaza, A. Ortega, A. Perera, T. Sundic, J.
(Fig. 4). Samitier, Methods for sensors selection in pattern recognition, Proc.
To check the predictive ability of the model built, a validation ISOEN (2000).
step using the second subset was performed. In this way, the [4] T. Eklov, P. Martensson, I. Lundstrom, Selection of variables for inter-
first necessary point was to obtain the PARAFAC scores of this preting multivariate gas sensor data, Anal. Chim. Acta 381 (2–3) (1999)
221–232.
second subset. This was done, once it was properly centered
[5] T. Sundic, S. Marco, A. Perera, A. Pardo, S. Hahn, N. Barsan, U.
and corrected, through its projection onto the PARAFAC model. Weimar, Fuzzy inference system for sensor array calibration: prediction
This scores estimation was done calculating the pseudoinverse of CO and CH4 levels in variable humidity conditions, Chemomet-
of Khatri–Rhao product of both loadings matrices (C and B), rics and Intelligent Laboratory Systems, October 2002. Corrected Proof,
multiplied by the vectorized form of the slice corresponding to Available online 11 October 2002.
[6] E. Llobet, J. Brezmes, X. Vilanova, X. Correig, J.E. Sueiras, Qualitative
the first object to be processed [25] (Eq. (2)).
and quantitative analysis of volatile organic compounds using transient
a = (C ⊗ B)+ vecXnew (2) and steady-state responses of a thick- lm tin oxide gas sensor array,
Sens. Actuators B, Chem. 41 (1997) 13–21.
Once the scores of the validation set were obtained, its flavouring [7] Y. Hiranaka, T. Abe, H. Murata, Gas dependant response in the temper-
agent amount values were obtained using the former regression ature transient of WGXGYZ gas sensors, Sens. Actuators B, Chem. 9
(3) (1992) 177.
coefficients obtained from the ILS calibration. The results from [8] J. Samitier, J.M. Lopez-Villegas, S. Marco, L. Camara, A. Pardo, O.
both datasets, calibration and validation, are shown in Table 2. Ruiz, J.R. Morante, A new method to analyse signal transients in chem-
As it can be seen, properly validation results were also obtained ical sensors, Sens. Actuators B, Chem. (1994).
150 M. Padilla et al. / Sensors and Actuators B 116 (2006) 145–150
[9] R. Gutierrez-Osuna, H.T. Nagle, S.S. Schiffman,.Transient response I. Montoliu was born in Barcelona in 1969. Degree
analysis of an electronic nose using multi-exponential models, Sens. (1995) and PhD in Chemistry (2001) at the Uni-
Actuators B, Chem. 61 (December) (1999) 170–182. versitat Autònoma de Barcelona, doing research in
[10] A. Perera, T. Sundic, A. Pardo, S. Marco, Feature extraction by sen- development and application of chemometrical meth-
sor waveform variance analysis, in: Proceedings of Ninth International ods to agroalimentary analysis. In 2000 joins the
Symposium on Olfaction and Electronic Nose, Roma, 2002. official distributor of Dionex products in Spain, as
[11] R.A. Harshman, Foundations of the PARAFAC procedure: model and chromatography applications and software special-
conditions for an ‘explanatory’ multi-mode factor analysis, UCLA Work- ist. Since 2004 is staying as Post Doctoral Research
ing Papers in Phonetics 16 (1970) 1. Associate Universitat de Barcelona. His areas of inter-
[12] J.D. Carroll, I. Chang, Analysis of individual differences in multidi- est are Chemometrics and its application to signals
mensional scaling via an N-way generalization of and Eckart-Young of different analytical instrumentation.
decomposition, Psychometrika 35 (1970) 283.
[13] P. Geladi, Analysis of multi-way (multi-mode) data, Chemom. Intell. A. Pardo was born in 1967 in Barcelona. Received
Lab. Syst. 7 (1989) 11. his deegree in physics 1991 and his PhD in 2000 from
[14] A.K. Smilde, Three-way analyses. Problems and prospects, Chemom. the University of Barcelona. His research is focused
Intell. Lab. Syst. 5 (1992) 143. in system identification with applications in gas sen-
[15] P.M. Kroonenburg, Three-mode principal component analysis, in: Theory sor systems, signal processing for gas sensors and
and Applications, DSWO Press, Leiden, 1983. pattern recognition.
[16] S. Wold, K. Esbensen, P. Geladi, Principal components analysis,
Chemom. Intell. Lab. Syst. 2 (1987) 37–52.
[17] B.M. Wise, N.B. Gallagher, S.W. Butler, D. White, G.G. Barna, A
comparison of principal components analysis, multi-way principal com-
ponents analysis, trilinear descomposition and parallel factor analysis for
fault detection in semiconductor etch process, J. Chemom. 13 (1999) A. Perera received his BS (1996) and his PhD (2003)
379–396. in physics from University of Barcelona and his BS
[18] Rasmus Bro. PARAFAC: tutorial and aplications. Chemom, Intell. Lab. in Electrical Engineering (2001) from the same uni-
Syst. 38 (1997) 149–171. versity. He was with Texas A&M University in the
[19] A. de Juan, R. Tauler, Comparison of three-way resolution methods for period from 2003 to 2005 as PostDoctoral Research
non-trilinear chemical data sets, J. Chemom. 15 (2001) 749–772. Associate. His research interests include applied pat-
[20] R.J. Barnes, M.S. Dhanoa, S.J. Lister, Standard normal variate transfor- tern recognition focusing chemical sensors, machine
mation and detrending of near-infrared diffuse reflectance spectra, Appl. olfaction and bioinformatics.
Spectrosc. 43 (1989) 772–777.
[21] R.J. Barnes, M.S. Dhanoa, S.J. Lister, Correction to the description of
standard normal variate (SNV) and de-trend (DT) transformations in
practical spectroscopy with applications in food and beverage analysis,
S. Marco was born in Berga in 1965. Degree (1988)
second ed., J. Near Infrared Spectrosc. 1 (1993) 185–186.
and Doctor in Physics (1993) at the University of
[22] M.S. Dhanoa, S.J. Lister, R. Sanderson, R.J. Barnes, The link between
Barcelona doing research on pressure sensors for
multiplicative scatter correction (MSC) and standard normal variate
biomedical uses which took place at the Centre
(SNV) transformations of NIR spectra, J. Near Infrared Spectrosc. 2
Nacional de Microelectrònica. In 1994 he was post-
(1994) 43–47.
doc at the University of Rome “Tor Vergata” working
[23] R. Bro, H.A.L. Kiers, A new efficient method for determining the
on intelligent signal processing. Since 1995 he’s an
number of components in PARAFAC models, J. Chemom. 17 (2003)
Associate Professor at the UB. His experience is
274–286.
centered in microsystem modelisation, and in signal
[24] B.G.M. Vandeginste, D.L. Massart, L.M.C. Buydens, S. De Jong, P.J.
processing-based intelligent instrumentation.
Lewi, J. Smeyers-Verbeke, Handbook of Chemometrics and Qualimetrics
(Part B), Elsevier, Amsterdam, 1998.
[25] A. Smilde, R. Bro, P. Geladi, Multi-Way Analysis, Application in the
Chemical Sciences, Wiley, Chichester, 2004.
Biographies
a r t i c l e i n f o a b s t r a c t
Article history: Drift is an important issue that impairs the reliability of gas sensing systems. Sensor aging, memory effects
Received 6 July 2009 and environmental disturbances produce shifts in sensor responses that make initial statistical models for
Received in revised form 7 October 2009 gas or odor recognition useless after a relatively short period (typically few weeks). Frequent recalibrations
Accepted 9 October 2009
are needed to preserve system accuracy. However, when recalibrations involve numerous samples they
Available online 7 November 2009
become expensive and laborious. An interesting and lower cost alternative is drift counteraction by signal
Keywords:
processing techniques. Orthogonal Signal Correction (OSC) is proposed for drift compensation in chemical
Gas sensor array sensor arrays. The performance of OSC is also compared with Component Correction (CC). A simple
Drift classification algorithm has been employed for assessing the performance of the algorithms on a dataset
Orthogonal Signal Correction composed by measurements of three analytes using an array of seventeen conductive polymer gas sensors
Component Correction over a ten month period.
Cross-validation © 2009 Elsevier B.V. All rights reserved.
Electronic nose
Data shift
1. Introduction methods to improve the stability over time of sensor arrays from a
signal and data processing perspective.
Gas sensor arrays are potentially economic and faster alternatives Gas sensor drift consists of a random temporal variation of the
for gas analysis and aroma evaluation compared to conventional sensor response when it is exposed to the same analytes under
analytical instruments such as gas chromatographs. Many different identical conditions. This problem is generally considered to be due to
technologies based on different principles for gas sensing are sensors aging [5], but it has also been attributed to a variety of sources,
available. These include Metal Oxide (MOX), Quartz Microbalances like environmental factors such as humidity variations [2], system
(QMBs), Conductive Polymers (CPs), Surface Acoustic Waves (SAWs), sampling non-specific adsorption [3,4], variations on flow rate,
and electrochemical sensors among others. While solid-state sensors thermo-mechanical degradation and poisoning among others [6,7].
offer very big advantages in terms of miniaturization, response times All of these factors can modify both the baseline and the sensitivity of
and power consumption, those sensors show poor selectivity. Since the sensors in the array in different ways, depending on sensor
the seminal paper of Persaud and Dodd [1], it has been known that technology.
improved selectivity may be achieved by combining different sensors Therefore, for operation over long periods, the ability of the
with partial and overlapping sensitivities and pattern recognition instrument to recognize analytes is degraded, since statistical models
systems. built in the calibration phase become useless after a short period of
While a large number of successful applications for gas sensor time, in some cases weeks or few months. After that time, the
arrays have been published, these are largely laboratory based and instrument must be completely re-calibrated, which is a time-
practical applications of chemical sensor arrays in real commercial consuming, laborious and expensive task, to ensure that the
applications are limited. This may be attributed to a number of predictions remain valid. The working hypothesis in this work is
reasons, such as insufficient sensitivity (or limit of detection too high), that system stability may be improved using proper multivariate data
lack of selectivity, and other problems. However, from the authors' processing techniques.
point of view, the main reason is the lack of stability over time and the Several methods have been reported in the literature in order to
cost of recalibration. This paper addresses the application of different improve stability over time by modifications in sensor technology and
design [8–10] or by the use of different sensor operation modes
⁎ Corresponding author. Departament d'Electrònica, Universitat de Barcelona, Marti i
[11,12]. On the other hand, signal processing methods for drift
Franquès 1, 08028 Barcelona, Spain. counteraction are based on different approaches: univariate [13,14] or
E-mail address: mpadilla@el.ub.es (M. Padilla). multivariate [15–17], linear [18,19] or non-linear [15–17], adaptive
0169-7439/$ – see front matter © 2009 Elsevier B.V. All rights reserved.
doi:10.1016/j.chemolab.2009.10.002
M. Padilla et al. / Chemometrics and Intelligent Laboratory Systems 100 (2010) 28–35 29
[15,16] or not, based in reference samples [14] or based in component Component Analysis (PCA) decomposition of a reference class data
removal [18,19]. Univariate corrections include basic baseline correc- subset that is later used to correct undesired variance from the rest of
tions [13], or more complex per-sensor correction by means of a the dataset. This method assumes that the reference class is
calibration sample [14]. These methods, however, do not take profit of representative of the entire population. Therefore variation found in
correlated drift effects. For this purpose, other multivariate methods the reference class will also be present in the rest of the dataset. This
have been proposed. Non-linear multivariate adaptive algorithms like means that, if the variance in the reference class is due to drift effects,
Self Organizing Maps (SOMs) [15,16], multiple SOM (mSOM) [17] and drift will be also removed from the complete dataset.
system identification theory [15,18] have also been proposed in the
past. However, fully unsupervised adaptive SOM methods have 2.2. Orthogonal Signal Correction (OSC)
problems in case of overlapping classes, since the reference pattern
updating may follow the wrong class. Techniques based on system Wold et al. firstly introduced Orthogonal Signal Correction (OSC)
identification theory predict sensors outputs by means of dynamical for its use on NIR spectra correction [27]. The main idea is to remove
models for every sensor from the rest of the sensors' response, variance not correlated to variables in a vector (or matrix) Y, which
assuming the sensors behave very similarly. Other adaptive contains some extra data information. This is done by constraining the
approaches include the use of sliding window wavelet decomposition deflation of non-relevant information of X, so that only information
based feature extraction for drift detection and compensation, orthogonal to Y should be removed. The inclusion of the condition of
sometimes post-processed by Principal Component Analysis or orthogonality to Y ensures that the signal correction removes as little
Fuzzy Adaptive Resonance Theory networks (ARTMAP) based algo- information as possible.
rithms [19–21]. Data processing by wavelets decomposes data into After Wold's paper a number of OSC-like algorithms that tried to
multiple time scales. Since drift is a slow process, it is better captured improve the original OSC method were published [28–34]. Compar-
by the coarsest scales, while noise and other events appear in the isons among them can be also found in literature [35,36]. In this work,
finest scales. Provided that the separation of drift from real responses the version applied is based on the Wise implementation of the
is feasible, this procedure improves posterior classification or algorithm [37]. This first searches for a direction of maximum variance
regression tasks. Linear methods like Component Correction (CC) of the data X using PCA. The scores vector t corresponding to this first
based on Principal Component Analysis (PCA) [22,23], or Canonical principal component is then orthogonalized with respect to the
Correlation Analysis (CCA) and Partial Least Squares (PLS) [24] have information matrix Y, in order to obtain a new scores vector nt not
also been reported to provide good results. In particular, Component correlated with Y that captures the highest possible amount of
Correction has received considerable attention in the community. variance of X. A Partial Least Squares (PLS) step between initial data X
However, Component Correction methods assume that all chemical and nt, with a suitable number of latent variables (LV), gives scores T1
classes behave in the same way in the presence of drift and this is not and loadings P1 vectors that contain the information not related to Y.
always the case (as it will be seen in the present study). For the rest of the paper, we will refer to this step as the inner PLS. The
Our primary goal is to demonstrate that Orthogonal Signal Correction number of inner PLS latent variables, it is calculated in this
(OSC) is suitable for drift compensation. OSC is a linear technique that implementation from an specification of the variance explained in
removes components orthogonal to a variable Y indicative of the data the X-block. We refer to this as the OSC tolerance and it will be given
structure, which is correlated to the data. This method is a very common in percentage values. In a final step T1 is again orthogonalized with
tool used in spectroscopy to correct spectra by removing systematic non- respect to Y and P1 is updated. The final T1 and P1 correspond to first
relevant information, such as baseline variation, but it is not so OSC factor and its products are then removed from original data. To
commonly used in the field of chemical sensors, although we find obtain a second OSC factor the complete treatment is then applied on
some examples in literature [25,26] where it is used for removing local corrected data, therefore for n OSC factors corrected data XOSC is
variance. This technique will be compared to Artursson's Component given by:
Correction (CC) method. The effectiveness of the proposed algorithms
n
will be evaluated on an experimental dataset composed by measure- XOSC = X− ∑ Ti Pi
′
ð1Þ
ments of several analytes using an array of conductive polymer sensors. i=1
The time duration of experiments has been 10 months.
A secondary goal of this paper is to study the impact of the training
set size on the ability of the methods for drift compensation. To be of 3. Experimental
practical interest, drift counteraction should be effective with only a
reduced set of training samples, spanning a limited time duration. 3.1. The dataset
A key point in this paper is the validation methodology. Very
common validation techniques, like k-fold cross-validation, random The dataset was from Osmetech plc (Cheshire, UK). Three different
subsampling, bootstrap and leave-one-out, which abound on the analytes (ammonia, propanoic acid and n-butanol), at different
literature, totally neglect the influence of drift. In those techniques concentrations levels, were periodically measured over 10 months
validation and training samples are interleaved along the time axis. with an array of 17 conductive polymer sensors, the total number of
Then future evolution of the sensor responses is modeled by the samples in the dataset being 3415. The concentration and number of
classifier and consequently provides overoptimistic results. Of course, samples for each analyte are shown in Table 1. We consider each
this is not representative at all of real operation conditions. The authors analyte as a class. Additionally, we define a group as a particular
would like to emphasize that those validation techniques should never analyte at a given concentration. Hence 3 classes and 8 groups are
be used in conditions where drift is present. Instead, in this work present in the dataset.
validation samples are always in the future of training samples. For every sample, the full sensor response to a sampling transient
is recorded. That is, the sensor array is initially exposed to clean air.
2. Theory Subsequently, the analyte at the desired concentration is introduced
in the sensor chamber for 185 s. Finally, clean air is introduced again.
2.1. Component correction Every transient signal lasts for 200 s at a sampling frequency of 1 Hz.
All waveforms are baseline corrected so that starting baseline in all
Artursson proposed in 2001 the Component Correction method samples is common at time 0 s. Fig. 1 shows an example of waveform
(CC) [22]. It is a signal processing technique based on a Principal for one sensor and three classes. To apply the proposed drift
30 M. Padilla et al. / Chemometrics and Intelligent Laboratory Systems 100 (2010) 28–35
In OSC, two internal parameters must be optimized; the number of drift, there is a long-term drift effect, which displaces and changes the
OSC components and the number of internal PLS latent variables. variance within the data set structure.
Here, instead of the number of inner PLS latent variables, we use the Fig. 5 shows the maximum of the transient for sensor one for the
equivalent amount of captured variance in the X-block (or tolerance) total duration of the experiment. In this figure, the effect of drift and
of the inner PLS relation. This is the way original Wise's algorithm is intra-class variability in the measurements is clearly seen, and also a
implemented [37]. clear correlation between the different traces can be observed. The
On the other hand, in the PCA-CC method the number of principal irregular shape of these curves may be due to sensors aging and also
components and a reference group have to be selected as well. The the environmental changes that happened during the 10 months of
criterion used to select the reference group was based on visual measurements, since conductive polymer sensors are strongly
inspection of the score plots in the calibration set. This group should perturbed by temperature and humidity.
have an inner covariance structure similar to most of the remaining
groups in the dataset. The reference group is modeled with a Principal 4.1. Fitting PCA-CC and OSC parameters
Components decomposition that captures a high amount of the group
inner variance in the calibration set. In fact, removing many components In Fig. 6, the Fisher ratio is plotted for different numbers of
in PCA-CC is risky, since some of those components may have infor- removed components and tolerances for OSC and PCA-CC. Fisher ratio
mation useful for posterior classifications. This is the case when drift measures the ratio between mean values of the distances of groups
direction is parallel to discriminant direction. Because of these factors among themselves and the groups' compactness:
only three components were removed by PCA-CC.
N N
FR = ∑ jcij j = ∑ mxi
4. Results and discussion i;j = 1;i≠j i=1
An exploratory analysis by PCA displays the initial distribution of where cij = |ci−cj| is the distance between the mean centers of groups
the classes. Fig. 4 shows the PCA scores plot of the calibration set i and j, and mxi is the mean of the distances of the Nk elements of
(solid symbols). In the figure, it can be seen that some classes group i to its center ci:
apparently mix, there is scattering due to drift and also additional Nk
intra-class variability. The direction of the main dispersion compo- 1 Nk 1
mxi = ∑ jxk −ci j; with ci = ∑ jx j
nent in all classes is quasi parallel, except ammonia 0.05%. Some Nk k = 1 Nk k = 1 k
groups, ammonia 0.01% and 0.02% and both concentration levels of n-
butanol, present dispersion over mainly only one direction. Main Therefore, high values are desired since they would mean small
dispersion directions are pointed out by an arrow for every type of groups and high separation among them. The proposed figure of merit
gas. A thick arrow shows the direction of the displacement of the helps to avoid over fitting, however it does not guarantee the best
eighth validation set (non-solid symbols) with respect to the choice for drift rejection in the long term, since the cross-validation
calibration set. Furthermore, in this data subset sensors are very samples are very close in time to the training samples.
correlated since only ten principal components capture more than Fig. 6 shows that the Fisher ratio is higher for OSC than for CC,
99% of the total variance, and the first two PCs capture about 90%. suggesting that the former will outperform the later in the final
A posterior data subset of 342 samples, measured between day 97 classification task.
and day 118, has also been projected on the same PCA subspace. For the OSC, we observe a continuous increase in the Fisher ratio
Scores of this projection are shown in the same figure in non-solid with the number of extracted components at multiple levels of the
symbols. The distribution of the groups in this latter subset is similar tolerance value. It may be surprising that this component removal does
to the one in the first subset; only ammonia 0.05% shows an important not saturate fast, but it has to be considered that this preprocessing is
deviation in its main dispersion directions. However, the whole subset carried out in an input space featuring high dimensionality, and the
is completely shifted from the location of the first data subset. graph only explores removal of up to 16 components. On the other
Therefore, additional to variability and scattering due to local noise/ hand, the figure of merit shows more sensitivity to the training
samples (larger error bars) when the number of OSC components
increases. This is probably due to an overfitting to the actual samples
Fig. 4. PCA scores of the calibration set (solid symbols) and eighth test set (measurements
between day 97 and day 118, in non-solid symbols). Fig. 5. Responses of sensor 1 to 3 analytes along the time.
32 M. Padilla et al. / Chemometrics and Intelligent Laboratory Systems 100 (2010) 28–35
Fig. 6. Fisher ratio vs. number of OSC and PCA-CC removed components and internal
OSC tolerances. PCA-CC is computed with two different groups of reference; ammonia
0.05% (group 3) and n-butanol 1% (group 8). Error bars represents sensitivity to training Fig. 7. First data removed component by PCA-CC and OSC.
samples.
Fig. 10. Classification rates along the time for drift correction by PCA-CC and OSC
Fig. 9. PCA scores of the calibration set (solid symbols) and eighth test set (non-solid
methods and non-corrected data (No Corr). Calibration set contains 308 samples.
symbols) corrected by PCA-CC (3 removed components, group of reference 8).
for each class, with decreasing concentrations towards the center of suitable method for drift correction. However, if the data is very close
the data distribution. Corrected test set is displaced from the training in time to training data OSC-8 is the most effective one. Up to 100 days
set, presents higher dispersion than training data and shows a small both OSC methods outperform PCA-CC.
change on the orientation of the branch corresponding to every The main reason of the different performance of both methods is
chemical specie. All these three effects are due to drift that OSC has the way they calculate the components of variance to be removed. The
not modeled. main component obtained by PCA-CC has an orientation given by the
Drift corrected data by means of PCA-CC (3 components, Fig. 9) direction of the largest variance of a reference group and, if more
also shows a distribution along three main directions, with different components are to be calculated, these ones are orthogonal to it. This
orientation according to the chemical species. It presents a very small makes that the removed components may be parallel to important
cluster for the reference group in the training set and other class of the directions for posterior classification, since it is not possible to control
same corresponding chemical specie. Analytes whose initial training their direction. On the other hand, PCA-CC only takes samples from
set distribution (Fig. 4) do not have parallel components to the ones of one group of reference to build a model, unlike OSC methods, which
the reference group, like ammonia 0.05%, show larger variance. Due to takes samples from all groups. PCA-CC reference group must be
drift, corrected validation data is also slightly displaced from training representative of all groups; therefore the more similar the groups are
data and it causes a change in the orientation of some class directions. the better this method performs. As a result PCA-CC depends strongly
The final eight groups are in general larger than in OSC plots and in the selection of the reference.
poorer in separation. Removed components in OSC are calculated considering informa-
It can also be noticed that the sparsest clusters in Figs. 8 and 9 are tion from all groups. The information regarding sample distribution is
the ones corresponding to ammonia. It has already been observed in in matrix Y, which contains the sample memberships to each of the
Fig. 4, that ammonia is the class whose covariance structure has 8 groups. Y is a matrix with dimensions number of samples x number of
changed the most from first subset to seventh, due to the effect of groups and binary elements; a ‘1’ at row i and column j means that
long-term drift that both signal processing methods are not able to sample i belongs to group j. The condition of orthogonality assures
correct. that extracted components of variance are not parallel to important
directions. Furthermore, OSC components are not forced to be
4.3. Time stability orthogonal among them. These facts mean that OSC gives very good
results for samples very close in time to the training set (little affected
In Fig. 10 the performance of PCA-CC and OSC drift counteraction by drift), outperforming PCA-CC until around day 100. Later it
techniques are plotted in the form of classification rates (CR) for degrades smoothly along the time resulting in a better and more
corrected and non-corrected data. Dataset dimensionality has stable behavior than non-corrected data, and even more stable than
been reduced to 10 PCs, capturing about 99% of variance, previous PCA-CC.
to the k-NN classifier. CR is calculated over nine test sets ordered in For both methods, selection of parameters is a critical step. More
time and models are built with 308 samples from calibration set. components or more strict tolerance (in the case of OSC) would result
Results show that both methods systematically improve the classifi- in a better-fitted model to training samples but in a loss of ability of
cation rate. Obviously, the correction is not perfect, and the CR still generalization to correct posterior data affected by drift. This fact can
decreases with time. be observed on comparing OSC-1 and OSC-8 results: OSC-8 outper-
OSC-8 (8 components) method outperforms OSC-1 (1 component) forms OSC-1 only on correcting samples from first validation set.
and PCA-CC providing higher CR values along the first month of In fact, in both cases the main difficulty resides in estimating from
measurements (first validation set). After first month, OSC-1 a short period (calibration set duration) the directions of drift. For the
clearly outperforms the other two methods until nearly day 100. method to be effective, the system should exhibit at least stability in
OSC-8 outperforms PCA-CC until around day 100. Later, all methods the statistical properties of the variance structure. If the structure of
become comparable, although PCA-CC presents the more unstable the noise/drift variance changes in time beyond the calibration phase,
behavior. Regarding sensitivity to training set samples, in general obviously the methods have no capability to re-adapt.
PCA-CC shows the smallest value and OSC-1 presents a lower error bar On the other hand, OSC classification results show higher
than OSC-8. This sensitivity was also observed in the Fisher ratio uncertainty than PCA-CC. We attribute this increased variance to the
figure (Fig. 6). We can conclude that OSC-1 is in general the most fact that it takes samples randomly from the eight groups in the
34 M. Padilla et al. / Chemometrics and Intelligent Laboratory Systems 100 (2010) 28–35
[12] M. Roth, R. Hartinger, R. Faul, H.E. Endres, Drift reduction of organic coated gas- Proceedings of the 7th International Symp. On Olfaction and Electronic Nose,
sensors by temperature modulation, Sensors and Actuators. B, Chemical 36 (1996) Brighton, UK, July 20–24, 2000, Institute of Physics Publishing, 2000, p. 147.
358. [25] O. Tomic, H. Ulmer, J.E. Haugen, Standardization methods for handling instrument
[13] E.L. Hines, E. Llobet, J.W. Gardner, Electronic noses: a review of signal processing related signal shift in gas-sensor array measurement data, Analytica Chimica Acta
techniques, IEEE Proceedings—Circuits Devices and Systems 146 (1999) 297. 472 (2002) 99.
[14] J.E. Haugen, O. Tomic, K. Kvaal, A calibration method for handling the temporal [26] J.E. Carlson, P.E. Martinsson, Exploring interaction effects in two-component gas
drift of solid state gas-sensors, Analytica Chimica Acta 407 (2000) 23. mixtures using orthogonal signal correction of ultrasound pulses, Journal of the
[15] M. Holmberg, F. Winquist, I. Lundstrom, F. Davide, C. DiNatale, A. Damico, Drift Acoustical Society of America 117 (2005) 2961.
counteraction for an electronic nose, Sensors and Actuators. B, Chemical 36 (1996) [27] S. Wold, H. Antti, F. Lindgren, J. Ohman, Orthogonal signal correction of near-
528. infrared spectra, Chemometrics and Intelligent Laboratory Systems 44 (1998) 175.
[16] S. Marco, A. Ortega, A. Pardo, J. Samitier, Gas identification with tin oxide sensor [28] T. Fearn, On orthogonal signal correction, Chemometrics and Intelligent
array and self-organizing maps: adaptive correction of sensor drifts, IEEE Laboratory Systems 50 (2000) 47.
Transactions on Instrumentation and Measurement 47 (1998) 316. [29] J. Sjoblom, O. Svensson, M. Josefson, H. Kullberg, S. Wold, An evaluation of
[17] M. Zuppa, C. Distante, P. Siciliano, K.C. Persaud, Drift counteraction with multiple orthogonal signal correction applied to calibration transfer of near infrared
self-organising maps for an electronic nose, Sensors and Actuators. B, Chemical 98 spectra, Chemometrics and Intelligent Laboratory Systems 44 (1998) 229.
(2004) 305. [30] J.A. Westerhuis, S. de Jong, A.K. Smilde, Direct orthogonal signal correction,
[18] M. Holmberg, F.A.M. Davide, C. DiNatale, A. Damico, F. Winquist, I. Lundstrom, Chemometrics and Intelligent Laboratory Systems 56 (2001) 13.
Drift counteraction in odour recognition applications: lifelong calibration method, [31] Li, A.J. Morris, E.B. Martin, Orthogonal signal correction: algorithmic aspects and
Sensors and Actuators. B, Chemical 42 (1997) 185. properties, Journal of Chemometrics 16 (2002) 556.
[19] H. Ding, H.F. Ge, J.H. Liu, High performance of gas identification by wavelet [32] R.N. Feudale, H. Tan, S.D. Brown, Piecewise orthogonal signal correction,
transform-based fast feature extraction from temperature modulated semicon- Chemometrics and Intelligent Laboratory Systems 63 (2002) 129–138.
ductor gas sensors, Sensors and Actuators. B, Chemical 107 (2005) 749. [33] J. Trygg, O2-PLS for qualitative and quantitative analysis in multivariate
[20] H. Ding, J.H. Liu, Z.R. Shen, Drift reduction of gas sensor by wavelet and principal calibration, Journal of Chemometrics 16 (2002) 283.
component analysis, Sensors and Actuators. B, Chemical 96 (2003) 354. [34] J. Trygg, S. Wold, Orthogonal projections to latent structures (O-PLS), Journal of
[21] E. Llobet, J. Brezmes, R. Ionescu, X. Vilanova, S. Al-Khalifa, J.W. Gardner, N. Barsan, Chemometrics 16 (2002) 119.
X. Correig, Wavelet transform and fuzzy ARTMAP-based pattern recognition for [35] O. Svensson, T. Kourti, J.F. MacGregor, An investigation of orthogonal signal correction
fast gas identification using a micro-hotplate gas sensor, Sensors and Actuators. B, algorithms and their characteristics, Journal of Chemometrics 16 (2002) 176.
Chemical 83 (2002) 238. [36] H.C. Goicoechea, A.C. Olivieri, A comparison of orthogonal signal correction and
[22] T. Artursson, T. Eklov, I. Lundstrom, P. Martensson, M. Sjostrom, M. Holmberg, net analyte preprocessing methods. Theoretical and experimental study, Chemo-
Drift correction for gas sensors using multivariate methods, Journal of Chemo- metrics and Intelligent Laboratory Systems 56 (2001) 73.
metrics 14 (2000) 711. [37] B.M. Wise, N.B. Gallagher, http://www.eigenvector.com/MATLAB/OSC.html.
[23] O. Tomic, T. Eklov, K. Kvaal, J.E. Haugen, Recalibration of a gas-sensor array system [38] J. Samitier, J.M. Lopez-Villegas, S. Marco, L. Camara, A. Pardo, O. Ruiz, J.R. Morante,
related to sensor replacemet, Analytica Chimica Acta 512 (2004) 199. A new method to analyse signal transients in chemical sensors, Sensors and
[24] R. Gutierrez-Osuna, Drift reduction for metal-oxide sensor arrays using canonical Actuators. B, Chemical 18 (1994) 308.
correlation regression and partial least squares, Electronic Noses and Olfaction:
Sensors and Actuators B 145 (2010) 464–473
a r t i c l e i n f o a b s t r a c t
Article history: Metal oxide (MOX) gas sensors have been widely used for years. Temperature modulation of gas sensors is
Received 25 July 2009 as an alternative to increase their sensitivity and selectivity to different gas species. In order to enhance the
Received in revised form extraction of useful information from this kind of signals, data processing techniques are needed. In this
10 December 2009
work, the use of self-modelling curve resolution techniques, in particular multivariate curve resolution-
Accepted 17 December 2009
alternating least squares (MCR-ALS), is presented for the analysis of these signals. First, the performance
Available online 4 January 2010
of MCR in a synthetic dataset generated from temperature-modulated gas sensor response models has
been evaluated, showing good results both in the resolution of gas mixtures and in the determination of
Keywords:
Temperature modulation
concentration/sensitivity profiles. Secondly, experimental confirmation of previously obtained conclu-
Multivariate curve resolution sions is attempted using temperature-modulated MOX sensors together with MCR-ALS for the analysis
MCR-ALS of carbon monoxide (CO) and methane (CH4 ) gas mixtures in dry air. Results allow confirming the pos-
Metal oxide sensors sibility of using the proposed approach as a quantitative technique for gas mixtures analysis, and also
reveal some limitations.
© 2009 Elsevier B.V. All rights reserved.
0925-4005/$ – see front matter © 2009 Elsevier B.V. All rights reserved.
doi:10.1016/j.snb.2009.12.051
I. Montoliu et al. / Sensors and Actuators B 145 (2010) 464–473 465
simple calibrations by comparison to a reference standard. How- sors to different binary gas mixtures. In a second part of the work,
ever, these results have become of limited utility in quantitative the exploration of the possibilities of the proposed approach for
analysis, mainly due to cross-sensitivity effects. quantitative purposes is also investigated.
Other authors have proposed the use of relative intensities of
real and imaginary components of the higher harmonics on the
Fast Fourier Transform (FFT) of gas sensor’s signals to distinguish 2. Experimental
between analytes (increasing selectivity) and this has been taken
as a basis for quantitative purposes. The use of 3D plots [10] of 2.1. Data description
concentrations vs. individual harmonics, with its further Artifi-
cial Neural Networks (ANN) + Fuzzy Inference processing [11,12], Let us consider that we have N different temperature-
have also provided partially successful qualitative and quantitative modulated MOX sensors. Their signals are acquired during the
results. total time of the experiment texp , with a sampling time of ts . The
In general, chemometric techniques [13] for mixture analysis sensor signals do produce then in general an array of dimen-
look for the transformation of multivariate data into relevant infor- sions (N, texp /ts ). Temperature modulation has a period of tcycle
mation, allowing the determination of relationships among the (ts tcycle texp ). To simplify the measurement conditions we
different individual system contributions to the global signal. assume that texp is an exact multiple of tcycle and tcycle is an exact
Multivariate curve resolution (MCR) is an algorithm for blind multiple of ts . We may consider that at the end of every modula-
source separation (BSS) [14] specifically adapted to chemical prob- tion cycle, each sensor outputs a measurement vector of tcycle /ts
lems. In this framework, the observed signals (sensor responses) are samples. Along the total duration of the experiment every sensors
considered to be a mixture of source signals (concentrations of pure produces texp /tcycle vectors. In this way, and for the whole dura-
components). BSS separation techniques try to recover the time tion of the experiment we may consider that each sensor produces
evolution of the source signals (concentrations of pure analytes) a data matrix of dimensions (texp /tcycle , tcycle /ts ). Since we have
without any knowledge of the mixing matrix. In other words, with- N sensors, data can be organized as a cube (or three-way array)
out any prior calibration step. BSS is typically attempted in signal of dimensions (N, texp /tcycle , tcycle /ts ) (see Fig. 1). In the follow-
processing by independent component analysis (ICA). ICA has been ing, we will refer to those dimensions as sensors, concentration
applied to chemical sensor signals by a number of authors [15,16]. (its evolution along the time) and sensitivities (due to temperature
However, for chemical problems a number of techniques have modulation).
been proposed that are able to perform this task (e.g. SIMPLISMA, Temperature modulation is accomplished by modulating the
MCR-ALS). Instead of imposing independence, other constraints as voltage applied to the sensor heater. In this way, sensor oper-
positiveness and others are used. The aim of self-modelling mul- ation temperature is basically a low-pass version of the power
tivariate curve resolution (MCR) methods is to generate a bilinear signal due to thermal inertia. MOX devices can exhibit diverse ther-
model that best describes the individual contributions of a reduced mal time constants depending on their internal structure. While
number of components to the global signal experimentally mea- fast MOX devices fabricated using Microsystem technologies can
sured. In other words, MCR methods seek for a bilinear model achieve time constants therm around 10 ms, older devices based in
which leads to the best fit of the original experimental data, in ceramic substrates can have slower thermal dynamics with time
a least squares sense. The most obvious difference with princi- constants on the order of 10 s. In any case, it is usually incorrect to
pal component analysis (PCA) is that the extracted components assume that sensors temperature follow instantaneously the power
are not required to be orthogonal. MCR are powerful methods signal applied to the sensor. In order to have a significant variation
because they are flexible and have low requirements. It is not of the temperature is necessary that tcycle therm .
necessary to have an exhaustive previous knowledge of the behav- Since the characteristic measurement time for every sensor is
ior of the system, although additional information can be used if given by tcycle . This time should be such that the evolution of the
it is available. The only initial assumption is the linear additiv- concentration within the gas chamber does not change dramati-
ity of the individual responses of the components of the analyzed cally. In other words, we will assume that chemical concentration
mixture. Unfortunately, metal oxide sensors are generally consid- is basically constant during the cycle time.
ered as non-linear sensing devices. However, in some particular
cases where the concentration range is limited and the signal is
expressed in conductance (G), sensor array signals can be approxi- 2.2. Synthetic dataset
mately described by a linear additive model, individually weighted
for each of the individual contributions. Typically, the concentra- A synthetic dataset was generated, intending to simulate the sig-
tion of these individual contributions or components can be used as nals that can be obtained from the measure of a ternary mixture of
a weighting element. In this work, we will explore to which extent gases in air. While concentration of one of them was kept constant
the non-linearity of the sensor responses allows the use of bilinear along all the measurement period, the other gases were incorpo-
additive models. rated and removed consecutively from the measurement chamber
In this work, MCR methods are proposed to determine the pure with a certain time overlap. The conductance of gas sensors was
component response (sensitivity) profiles as well as the concen- normalized to the conductance in pure air. To do so, two datasets
tration profiles from a set of temperature-modulated MOX sensors for each gas sample were generated. Thus, a first one was obtained,
signals. We have already mentioned that this approach has consid- containing the theoretical signal when the sensor is exposed to
erable interest since it can extract concentration profiles (signals) pure air, and a second one, giving its expected response when it
and sensitivities without prior calibration steps. The estimation is exposed to the ternary gas mixture.
of the sensitivity profiles has additional interest because it gives In order to generate these datasets, the dynamic Clifford–Tuma
additional insight regarding the sensor response and, moreover, model [17,18] was taken as a basis, with suitable parameters to
it may help to set-up the optimum operational temperatures in a design sensors based in n-type semiconductors. Thus, a model pro-
particular scenario. viding responses with different sensitivities to air, CO, CH4 and
Two cases have been investigated: a synthetical dataset corre- ethanol gases was obtained. It must be noted that all these gases
sponding to a ternary gas mixture, and an experimental dataset produce an increase in the conductance of n-type sensors. The sim-
corresponding to the exposure of four commercial MOX gas sen- ulation considered 11 different thermally modulated MOX sensors.
466 I. Montoliu et al. / Sensors and Actuators B 145 (2010) 464–473
Fig. 1. Procedure from original data to build data storage and preprocessing. (A) General scheme of sampling of the concentration during the measurement time cycle by
temperature modulation. Expected evolution of the signal as the gas sample enters to and exits from the gas cell. (B) Folding of the vectorized data from each sensor, rendering
as many rows as sampling data points have been recorded during each cycle (tcycle ) and as many columns as (texp /tcycle ), where texp corresponds to the total number of data
points recorded.
While it is not the objective to describe here a detailed model of a 3% multiplicative random gaussian noise was assumed on the
the sensors behavior, but only one capturing the complexity of the final conductance of each sensor. Signal conditioning was assumed
sensor signals, and for the sake of completeness, we present some to be obtained using a resistor half-bridge excited with 5 V and
hints on the implementation of the sensors models. a loading resistance of 1 K. Voltage was sampled with an 8-bit
Gas cell was considered to be 100 ml in volume, and contin- A/D converter using a 5 V reference introducing the corresponding
uously fed with a gas flow of 300 ml/min. Sensor chamber was quantization noise.
considered to act as a low pass filter for the gas concentrations Sampling rate was also fixed to 10 Hz, thus resulting to an overall
with time constant 20 s. of 100 points/cycle. The synthetic experiment had a total duration
The heating function, was numerically implemented using a of texp = 15 min, thus generating a total recording vector of 9000
voltage sawtooth function constrained between 0.2 and 0.9 V, dur- points for each sensor.
ing a period of tcycle = 10 s, producing a temperature excursion from Following the general procedure presented in Section 2.1,
200 to 450 ◦ C in static conditions. Sensors were considered to have a synthetically obtained vectorized data were properly folded to
thermal time constant of therm = 0.5 s. On the other hand, chemical generate data matrices of 100 × 90 dimension, where the two direc-
reactions at the sensor surface were considered to have a temper- tions or ways of measurement of the data were sensitivity and
ature dependent kinetics, being slow at low temperatures and fast concentration for both datasets: air (A1 , A2 , . . . ,A11 ) and ternary
at high temperatures. The overall contribution to the j sensor con- gas mixtures (R1 , R2 , . . . ,R11 ). Each pair of matrices (air/gas mix-
ductance (Gj ) by the different analytes was combined according to tures) were ratiored for each sensor (Xi = Ai /Ri ), giving a new set of
the expression 1. data matrices (X1 , X2 , . . . ,X11 ).
Gj = G0,j (T ) 1+ Si,j (T ) × ciˇi (1) 2.3. Experimental dataset
i
Experimental data was obtained using a set of four metal oxide
where G0,j is the static sensor conductance in air, Si,j is the sensitivity (MOX) sensors (SB11A, SB15, SB31 and SB42A) provided by FIS Inc.
to gas i. ci is the concentration of gas i, T is temperature and ˇi is a (Hyogo Japan), built in a custom designed gas cell. In accordance
gas depending parameter with values around 0.5. with their specifications, they are capable to detect hydrocarbons
In order to simulate realistic noise sources, the following (FIS SB11A, FIS SB15), solvents (FIS SB31) and Hydrofluorocarbons
assumptions were also made. Random gaussian noise with an (FIS SB42A). Data acquisition was performed using a DataTaker
amplitude of 2 ◦ C was assumed for the sensors temperature, and DT800 (dataTaker Pty Ltd., Australia) equipment, using a 2 Hz
I. Montoliu et al. / Sensors and Actuators B 145 (2010) 464–473 467
Table 1
Concentration of experimental gas mixtures.
Standard 1 0 3600
Standard 2 240 3000
Standard 3 360 2400
Standard 4 480 1800
Standard 5 600 0
4. Results and discussion sensitivities are always positive. The results obtained for sensor 5
are shown in Fig. 4. They confirmed the presence of three species in
4.1. Simulated data the measurement interval, with variable concentrations depending
on time. However, pure concentration profiles initially obtained in
With the purpose of resolving the pure sensitivity and concen- this way did not appear to be properly resolved. Although two of
tration profiles of one single sensor during the measurement cycle, the compounds were sequentially added, they emerged at nearly
its conductance response was studied by MCR-ALS. Thus, the matrix the same sampling time, and consequently with a high degree of
containing data from sensor 5 was selected. The 3D plot of this sen- overlap. On the other hand, the study of the sensitivity profiles
sor’s response is shown in Fig. 3a. Its singular value decomposition shows maxima in different positions for two of the gases, being
(SVD) indicated the presence of 2–3 major components. the component 2 the one showing the major sensitivity at a higher
As stated before, MCR-ALS algorithm initialization needs an temperature. For component 3, the corresponding sensitivity pro-
initial estimation either of the sensitivity, S, or concentration C file shows a decreasing profile in the coincidence region with the
profiles. To generate these initial profiles, EFA or SIMPLISMA can sensitivities of the other two factors, thus suggesting a certain non-
be used. Due to the time evolving characteristics of this dataset in linear effect. However, the main trends observed in the true profiles
particular, these pure profiles were generated by EFA. agreed with the sequential addition of the gases and with the dif-
MCR-ALS algorithm was then initialized applying nonnegativ- ferent sensitivities of the simulated sensors to each of the species.
ity constraints to sensitivity and concentration profiles, assuming MCR-ALS procedure was extended to the simultaneous analy-
that the concentration profiles cannot take negative values, and sis of the response signals from 11 sensors, as it is described in
that all the gases produce increases in the sensor conductance, so the method section. To do it so, a new row-wise (RW) augmented
I. Montoliu et al. / Sensors and Actuators B 145 (2010) 464–473 469
Fig. 4. (A) Pure sensitivity and (B) concentration profiles obtained from MCR-ALS
resolution of data from sensor 5.
Fig. 5. (A) Pure sensitivity and (B) concentration profiles obtained by MCR-ALS reso-
data matrix was properly set following a row wise augmentation lution of data form sensors 1–11. (C) Theoretical time/concentration profile of gases
(Eq. (6)) procedure. This process leaded to an augmented matrix of into sensors chamber. Both profiles have been displaced along the gas concentration
axis to allow a better visualization.
dimensions (90 × 1100). As it can be seen in Table 2, the determina-
tion of the number of factors was not simple, leaving this selection
to 2–3 factors. ysis previously done, choosing the concentration profiles resolved
MCR-ALS was applied to this augmented data matrix, assuming after the analysis by MCR-ALS of the sensor 5 data. Additionally
again nonnegativity constraints to both modes and normalization in this case, equality constraints were applied to one of the con-
of the data. Initial estimates were taken in this case from the anal- centration profiles. As it can be seen in Fig. 5a, the study of the
pure sensitivity profiles showed important differences among the
Table 2
3 different gases. For instance, it is observed a reversed behavior
Singular values for each row-wise augmented dataset. of the sensitivity of the different sensors for components 1 and 2
for each of the 11 sensors. On the other hand, the sensitivity profile
Factor Synthetic data. Real data. Real data. augmented.
augmented. augmented. (4 sensors × 5
corresponding to component 3, which corresponds to the gas kept
(11 sensors) (4 sensors) mixtures) at a constant concentration during all the experiment, shows an
increasing response with the sensor number. Both observed order-
Singular values
1 1625.18 1455.64 2748.56 ings are a direct consequence of the synthetic nature of the data
2 152.25 231.44 693.43 and must not be taken into consideration.
3 61.28 128.12 264.85 The incorporation and simultaneous analysis of the information
4 36.55 79.29 148.08 contained in multiple sensors, following the data augmentation
5 32.38 53.01 124.53
6 28.23 13.56 58.25
strategy proposed in Eq. (6), allowed an improved resolution of the
7 23.89 9.86 46.08 concentration profiles of the different gas species. As it can be seen
8 21.58 7.20 37.77 in Fig. 5b and c, now MCR-ALS pure and theoretical concentration
9 20.18 4.83 29.67 profiles were agreeing. This improvement in resolution should be
10 17.82 2.69 21.74
related to the different relative sensitivities to gas species provided
470 I. Montoliu et al. / Sensors and Actuators B 145 (2010) 464–473
Fig. 6. (A) Pure sensitivity and (B) concentration profiles obtained from MCR-ALS
resolution of experimental data from sensor SB42A, for CO (solid line) and CH4 Fig. 7. Multisensors case. (A) Sensitivity and (B) concentration profiles, obtained
(dashed line). from MCR-ALS resolution of experimental data from sensors (1) SB11A, (2) SB42A,
(3) SB15 and (4) SB31, for CO (solid line) and CH4 (dashed line).
Fig. 9. Data flow scheme for quantitative estimation of the analytes’ concentration.
case. Even so, the maxima of the different sensitivity profiles were
Fig. 8. Multisensors + multiset case. (A) Pure sensitivity and (B) concentration pro- nearly to be superimposed in the same heater voltage region, with
files obtained from MCR-ALS simultaneous resolution of experimental data from different intensities. Although these profiles were now improved,
sensors (1) SB11A, (2) SB42A, (3) SB15 and (4) SB31, exposed to five standard samples
they still were not presenting the expected shape.
with different concentrations in CO (solid line) and CH4 (dashed line).
To improve the resolution and to investigate the possibility of
performing quantitative estimations, two more standard mixtures
malization of the sensitivity profiles and nonnegativity constraints were simultaneously analyzed under the same experimental con-
to both modes were applied. ditions, according to the concentrations specified in Table 1. In
Pure profiles finally resolved by MCR-ALS are given in Fig. 7. addition to these mixture standards, two pure samples of both gases
Comparing these profiles with the previous ones using a single sen- were also recorded. Experimental data matrices obtained in these
sor, it can be observed now how the simultaneous analysis of the measurements were ordered following a CW + RW procedure, as it
four sensors provide a much better resolution of the gas concen- is expressed in Eq. (7), now building a new multisensor + multiset
tration profiles and also that they were closer to the expected ones, augmented data matrix to be resolved by ALS.
specially for CH4 . In particular, CH4 concentration profile agreed Multiset experimental dataset resulted to be well described
well with what was expected from experimental conditions. In the using two components (see Table 2, 4 sensors × 5 mixtures aug-
case of CO, the obtained concentration profile did not completely mented matrix). As in the previous case, initial estimations for both
agree with the expected one, since discordances appeared in the components were calculated by detection of purest variables. Non-
overlapping region (time ≈ 19–34 min). Even though, as in the pre- negativity constraint to the concentration profiles was added. Time
vious cases, it is possible to identify both gases from its deviations intervals of resolved concentration profiles, where one of the gas
of the baseline (starting times). species is known to be not present, were also constrained accord-
On the other hand, sensors sensitivities presented the same ingly. MCR-ALS solution provided improved estimations of the pure
shapes as in the previous case, but now with different relative inten- profiles for both concentration and sensitivity modes.
sities for each of the four SB sensors. Results show that sensor 2 When the resolved profiles were compared, it was found that
is more sensitive to CO than to CH4 while, in sensor 3 sensitiv- the different contributions from both individual gases, described
ities are reversed. Resolution for the selectivity profiles towards by the concentration and sensitivity profiles, were now in agree-
each gas due to the inclusion of sensors of different characteristics ment with the expected ones. These shapes can be seen in Fig. 8. In
improved significantly. However, this improvement of resolution the lower part of the figure in can be seen how the concentration
was not enough to provide sensitivity profiles in complete agree- profiles are showing distortions in the rising part of the peaks and
ment with the foreseen ordering of sensitivities along the heater on its overlapping regions (especially in the last mixture important
voltage cycle. Again, they still were not presenting the expected distortions of both concentration profiles are observed). This means
behavior for CO, appearing in a reversed order as in the previous that the augmentation of the data matrix with additional record-
472 I. Montoliu et al. / Sensors and Actuators B 145 (2010) 464–473
5. Conclusions
[10] S. Nakata, S. Akakabe, M. Nakasuji, K. Yoshikawa, Gas sensing based on a non- Roma Tauler obtained his Ph.D. in Chemistry at the Uni-
linear response: discrimination between hydrocarbons and quantification of versity of Barcelona in 1984. He is research professor
individual components in a gas mixture, Anal. Chem. 68 (1996) 2067–2072. at the Institute of Environmental Assessment and Water
[11] A. Heilig, N. Barsan, U. Weimar, M. Schweizer-Berberich, J.W. Gardner, W. Göpel, Research (IDÆA), CSIC, in Barcelona (Spain), since July
Gas identification by modulating temperatures of SnO2 -based thick film sen- 2003. Previously he was associate professor of Analytical
sors, Sens. Actuators B 43 (1997) 45–51. Chemistry at the Analytical Chemistry Department of the
[12] B. Yea, T. Osaki, K. Sugahara, R. Konishi, The concentration–estimation of University of Barcelona during 1987–2003. He performed
inflammable gases with a semiconductor gas sensor utilizing neural networks postdoct stays at Institut fu¨r Anorg. u. Anal. Chemie,
and fuzzy inference, Sens. Actuators B 41 (1997) 121–129. Univ. of Innsbruck, Innsbruck (Austria) in 1985 and 1989
[13] D.L. Massart, B.G.M. Vandeginste, L.M.C. Buydens, S. De Jong, P.J. Lewi, J. and a year’s sabbatical as a research scientist at the Cen-
Smeyers-Verbeke, Handbook of Chemometrics and Qualimetrics, Part B, Else- ter for Process Analytical Chemistry (CPAC), Department
vier, Amsterdam, 1997, p. 243. of Chemistry, University of Washington, Seattle, USA, in
[14] A. Cichoki, R. Zdunek, A.H. Phan, S.I. Amari, Non-negative Matrix Factoriza- 1992. At present, he is the Editor in Chief of the journal
tions: Applications to Exploratory Multi-way Data Analysis and Blind Source Chemometrics and Intelligent Laboratory Systems He has published more than 200
Separation, Wiley, Chichester, 2009. research papers, most of them in the field of chemometrics and its applications,
[15] C. Di Natale, E. Martinelli, A. D’Amico, Counteraction of environmental dis- and in particular in the development and applications of new multivariate reso-
turbances of electronic nose data by independent component analysis, Sens. lution methods. In the recent years he has focused more on the investigation of
Actuators B 82 (2002) 158–165. environmental problems.
[16] M. Kermitt, O. Tomic, Independent component analysis applied on gas sensor
array measurement data, IEEE Sensor J. 3 (2003) 218–228.
[17] P.K. Clifford, D.T. Tuma, Characteristics of semiconductor gas sensors I. Steady Marta Padilla was born in Sta. Cruz de Tenerife, Spain, on
state gas response, Sens. Actuators 3 (1982–1983) 233–254. 1973. She received a degree in Physics from the University
[18] P.K. Clifford, D.T. Tuma, Characteristics of semiconductor gas sensors II. Tran- of La Laguna in 1999 and a degree in Electronic Engineer-
sient response to temperature change, Sens. Actuators 3 (1982–1983) 255–281. ing from University of Barcelona in 2003. In the same year
[19] R. Tauler, D. Barceló, Multivariate curve resolution applied to liquid she became a Ph.D. student in Electronic Engineering in
chromatography—diode array detection, Trends Anal. Chem. 12 (1993) the University of Barcelona.
319–327.
[20] A. De Juan, R. Tauler, Chemometrics applied to unravel multicomponent pro-
cesses and mixtures: revisiting latest trends in multivariate resolution, Anal.
Chim. Acta 500 (1/2) (2003) 195–210.
[21] H.R. Keller, D.L. Massart, Evolving factor analysis, Chemom. Intell. Lab. Syst. 12
(1992) 209–224.
[22] W. Windig, J. Guilment, Interactive self-modeling mixture analysis, Anal. Chem.
63 (1991) 1425–1432.
Antonio Pardo received his diploma in Physics 1991 and
[23] W. Windig, Self-modeling mixture analysis of spectral data with continuous
his Ph.D. in 2000 from the University of Barcelona. During
concentration profiles, Chemom. Intell. Lab. Syst. 16 (1992) 1–16.
his Ph.D. studies he worked in system identification with
[24] E. Peré-Trepat, A. Hildebrandt, D. Barceló, S. Lacorte, R. Tauler, Fast chro-
applications in gas sensor systems. Since 2005 he is asso-
matography of complex biocide mixtures using diode array detection and
ciate professor at the University of Barcelona. His research
multivariate curve resolution, Chemom. Intell. Lab. Syst. 74 (2004) 293–303.
interest focused on signal processing for gas sensors and
[25] I.T. Jolliffe, Principal Component Analysis, 2nd ed., Springer, New York, 2002.
pattern recognition as well as on hardware and software
[26] G.H. Golub, C.F. Van Loan, Matrix Computations, 2nd ed., Johns Hopkins Uni-
developments for electronic nose instrumentation.
versity Press, Baltimore, 1991.
[27] R. Tauler, A. Smilde, B.R. Kowalski, Selectivity, local rank, three-way data analy-
sis and ambiguity in multivariate curve resolution, J. Chemom. 9 (1995) 31–58.
[28] R. Tauler, Calculation of maximum and minimum band boundaries of feasi-
ble solutions for species profiles obtained by multivariate curve resolution, J.
Chemom. 15 (2001) 627–646.
Santiago Marco is associate professor (Profesor Titu-
lar) at the Departamento d’Electronica of Universitat de
Biographies Barcelona since 1995. He received the degree in Physics
from the Universitat de Barcelona in 1988. From 1990 to
1993, he was regular visitor of the Centro Nacional de
Microelectrònica, Bellaterra, Spain. In 1993, he received
Ivan Montoliu was born in Barcelona in 1969. Degree his Ph.D. (honor award) degree from the Departamento
(1995) and Ph.D. in Chemistry (2001) at the Universitat de Física Aplicada i Electrònica, Universitat de Barcelona,
Autònoma de Barcelona, doing research in development for the development of a novel silicon sensor for in vivo
and application of chemometrical methods to agroali- measurements of the blood pressure. In 1994, he was
mentary analysis. In 2000 joins the official distributor of a post doc researcher at the Department of Electronic
Dionex products in Spain, as chromatography applications Engineering, Universita di Roma ‘Tor Vergata’, working in
and software specialist. From 2004 to 2007 he had a post data processing for artificial olfaction. In 2004 he had a
doc Position at the Department of Electronics at the Uni- Sabbatical Year at EADS Corporate Research in Munich working in Ion Mobility Spec-
versity of Barcelona. From 2008 he is with the Nestle trometry. He has published about 70 papers in scientific journals and books, as well
Research Center in Laussanne, Switzerland. His areas of as more than 130 conference papers. His current research interest is the applica-
interest are Chemometrics and its application to signals tion of signal and data processing to smart chemical instrumentation. Since 2008 he
of different analytical instrumentation. leads the Artificial Olfaction Lab at the Institut for BioEngineering of Catalonia.
Poisoning fault diagnosis in chemical gas sensor
arrays using multivariate statistical signal processing
and structured residuals generation
M. Padilla*, A. Perera†, I. Montoliu†, A. Chaudryº, K. Persaud‡ and S. Marco†
† Sistemas d’Instrumentació i Communicacions, Departament d’Electrònica, Universitat de Barcelona, Spain
º Crewe Business Park, Crewe, Cheshire, UK
‡ School of Chemical Engineering and Analytical Science, The University of Manchester,UK
Abstract – Chemical gas sensors are a cheaper and faster improvements, although signal processing can help to detect
alternative for gas analysis than conventional analytic and compensate the output signal of the affected sensor.
instruments. .However they are prone to degradation because of Poisoning consists on a degradation of the sensing
sensor poisoning and drift. Statistical methods like Principal properties of a sensor when it is exposed to a gas compound
Component Analysis (PCA) and Partial Least Squares (PLS) that irreversibly interacts with its sensing material. It is not an
have been proved to be very useful in the task of fault diagnosis
easily detectable problem because the sensors keep their
of malfunctioning sensors. In this work we test the effectiveness
of several techniques based on PCA and PLS on faults caused by responses within normal levels, although with an altered
sensor poisoning. These techniques will be evaluated on a dataset profile of sensitivities to the different analytes. Poisoning is a
composed by the signals of 17 conductive polymers gas sensors surface phenomenon and is closely related to physical
measuring three analytes at several concentration levels. These properties of the sensing material and the detecting gas, so for
techniques will be evaluated concerning their capabilities to heterogeneous arrays it can affect only some of the sensors in
detect the fault, identify the faulty sensor and correct their an array, which have to be immediately detected and
signal. identified in order to correct their responses until they can be
replaced by new ones.
Keywords – fault diagnosis, gas sensor array, poisoning. Poisoning may affect the sensors by changing its baseline
resistance [1], its sensitivity and recovery times [2]. The
degree of poisoning depends on the exposure temperature, the
I. INTRODUCTION time and the concentration of the poisonous substance [3], for
example HMDS[3], SO2 [2], NO2 [1] or O3 [2] have been
The use of chemical gas sensors in an array format with a reported as poisoning gases for different technologies.
pattern recognition system improves the degree of selectivity Event or fault detection on gas sensor arrays has been
of the individual gas sensors. In a sensor array each sensor previously tackled by other authors. Perera et al. [4] presented
has a partial specificity; the sensor array output produce a an adaptive method based on a recursive dynamic principal
'compound/ odour' fingerprint that can be identified by the component analysis (RDPCA) algorithm. This algorithm was
pattern recognition system. A sensor array is the base of the applied on real signals arising by oil vapor leakages in an air
instrument known as 'electronic nose' which is inspired by the compressor. Pardo et al. [5] used the correlation among the
mechanisms involved in human olfaction. Electronic noses responses of five gas sensors detect a possible malfunctioning
have been used for the analysis of volatile organic compounds of one of the sensors during continuous operation. The
in applications that range from the food or medical industry to outputs of the sensors were estimated as a function of the
environmental monitoring and process control. Nevertheless outputs of the remaining ones by Artificial Neural Networks.
problems like poisoning or drift degrade the ability of the On the other hand, the employed techniques of fault
instrument to recognize odours or gases. When using an array diagnosis can be based on statistical methods. Fault detection
of sensors one or more of them can fail due to a great variety using Principal Component Analysis (PCA) or Partial Least
of faults like sensor failure, degradation, external Squares (PLS) has already been studied in applications like
interferences, bias, noise, occasional outliers (noise spikes). process monitoring, quality control, or sensor and process
While these problems depend on the technology and fault diagnosis [6]. Also modifications of PCA have been
fabrication process and may be overcome by technology
−1
where fj is an estimate of the fault magnitude in the direction d = e T Re e (8)
ξj. The reconstructed vector xj is obtained minimizing the SPE
of sensor j response for a sample xj (SPEj). The minimization where Re is the covariance matrix of e*. This index d follows
of SPEj gives the value for fj: a Chi-square distribution when no fault is present.
This method allows the detection and identification of one
~ ~ ~ 2 or more faulty sensor in a single step; without having to
SPE j = ~
2
x j , dSPE j / df j = 0 ⇒ f j = ξ j T x / ξ j (4) remove a first faulty sensor detected and having to build a
new model with the rest of the sensors. Then it is possible to
~
x ,~
where ~ x j and ξ j are the projection of the sample x, xj, and know if there are one ore two malfunctioning sensors
depending on the design of the structure of the transformed
ξj respectively on the residual subspace. residuals.
Therefore the corrected component xj is the one that has To determine a suitable matrix W, a structure or incidence
the minimum size of its projection on the residual space. Then matrix has to be first designed. In this structure each fault has
the relation SPEj/SPE (j=1,..,N) is calculated for the detected a characteristic response, a fault code. Then, in the incidence
faulty sample. If sensor j is faulty its reconstruction will give matrix, the rows are the residual structures while the columns
a small SPEj, then SPEj/SPE will be close to zero, since SPE are the fault codes. A 1 element means that the concerned
increases when a fault occurs. If sensor j is not faulty its residual does respond to the concerned fault while a 0 means
reconstruction will not reduce SPEj, then SPEj/SPE will be it does not. It is possible to design a structure for a single fault
close to 1. Therefore SPEj/SPE will present the minimum or for multiple faults such as double or triple. In this case the
value for the direction that corresponds to the faulty sensor. incident matrix has to be carefully designed. More detailed
The technique for an optimum choice of the number of explanations and conditions about structured residuals can be
principal components for every PCA model is also explained found in [13].
by Dunia and Qin and can be found in [11]. To calculate W relations between its components, vectors
To detect two faulty sensors the first faulty sensor detected wi , and faults, bj, have to be established. The traditional
has to be removed and a new model has to be rebuilt with the structured residuals approach [12] chooses the vectors wi of
rest of the sensors. Then the same technique of detection and matrix W to be insensitive to certain faults and sensitive to
identification of a second faulty sensor has to be applied. others, but this wi is not unique and doesn't maximize the
response to those faults. Qin and Li propose [10] an algorithm
C. PCA structured residuals with maximum sensitivity (SRAM) that makes the residuals respond with maximum
approach (PCA-MS) sensitivity to the considered faults being insensitive to the
remaining ones. Mathematically it is equivalent to:
PCA with maximum sensitivity (PCA-MS) is a technique
that directly uses the residuals of PCA. A method based on n
structured residual was introduced by Gertler et. al. in 1990 max ∑ ( wiT b j / b j ) 2 with wiT bi = 0 , wi = 1 (9)
j ≠i
[12], and Qin and Li [10] proposed an extension of this j =1
method. The aim is to structure the residuals of a linear
method like PCA in such a way that one residual present most where wi is the ith row of matrix W and ri an ith element of r.
sensitivity to one subset of faults while being insensitive to Geometrically this means that wi and bi are orthogonal and the
other faults. This is done by means of a transformation angle between wi and other faults directions bj (i≠j) is
matrix, W which converts the primary residuals e of PCA into minimum. Finally, vector wi is given by the eigenvector that
structured residuals r with: corresponds to the largest eigenvalue of:
r = We (6) T T 2
Bi0 Bi0 wi = λwi where Bi0 = ( I − bi bi / b j ) B 0 (10)
The primary residuals e for a sample x with a fault e,
according to eq. (1), are given by: Later, some modifications have been introduced, for
example Li and Shah [14] use all the eigenvectors in eq. (10)
Bx = e = Bx* + Bξi f i = e* + bi f i (7) to generate a more powerful structured residual, which is a
vector, and Xu and Kwan [15] propose other algorithm which
really maximizes the sensitivity to other faults, not only
where bi is a column of B and the primary residuals e* are maximizing the sensitivity of the squares of fault responses as
obtained applying the model matrix B to the training data x*. SRAM does. This technique is called Max-min method and it
is more powerful than SRAM because it maximizes the systems. These are commonly classified in four types;
sensitivity of the response to the weakest of the faults of precision degradation, bias, drift and complete failure. The
interest. fault we consider in this work is characteristic of chemical
The obtained structured residuals are now used to build sensors which interacts directly with the measured substance.
scalar fault identification indices which will indicate which The poisoning of the sensor sensing element would lead to a
sensor fails. In [10] Qin and Li investigate several types of change in its sensitivity that can be not easily to detect, since
indices but here we only consider one of them, given by ri 2. If the sensor behaviour is apparently normal.
the identification index is Ii= ri 2/ ηi , being ηi a confident limit
determined by statistical techniques, it will be less than one in 12
Transient response of sensor 1 per class
Conductance (1/Ohm)
8
estimated by minimizing the magnitude of the error of the
corrected sample. 6
4 ammonia 0.01%
TABLE I. 10
Figure 2 shows the sensor responses during a sampling Fig.3 Examples of patterns corresponding to samples of 8 different
process. The sensor conductance increases and when the gas classes.
is removed it returns to the baseline. From this transient, only PCA scores
the maximum value is retained for further processing. At the 5 ammonia 0.01%
considered point of the transient signals the response of every ammonia 0.02%
ammonia 0.05%
n-buthanol 1%
2
applied on the complete dataset after being autoscaled. The
scores are shown in figure 4. In this dataset sensors are very 1
variance, and the first two PCs capture 96%. In the PCA
scores plot, the scores in different colors and shapes
-1
mixture of some classes, also several classes are scattered due 91.2% PC1
to the effect of intra-class variability and drift. Fig.4 PCA scores of the dataset.
10
90
IV. RESULTS 80
70
tp (%)
50
sensors. 0
0 10 20 30 40 50 60 70 80 90 100
with 342 samples and 50% of the samples with induced faults Fig.6. ROC representation for the detection index that presents the best ROC
in one and two sensors failing at the same time. The training curve for PCA, PCA-MS and PLS approaches.
set contains also 342 samples.
0
PCA PLS PCA-MS ACKNOWLEDGMENTS
Fig.7. Percent of accuracy in one and two faulty sensors identification for
PCA, PLS and PCA-MS approaches. We want to acknowledge the partial funding of this
work by the Spanish project MCYT TEC2004-07853-C02-01
and GOSPEL IST 507610 Fp6.
C. Faulty sensor correction
REFERENCES
After correcting the faulty sensors a reduction of
dimensionality by PCA, from dimension 17 to 5, has been [1] B. Ruhland, Th.. Becker and G. Müller, "Gas-kinetic interactions of
done. Later, a classifier (kNN with 3 nearest neighbors) has nitrous oxides with SnO2 surfaces", Sensors and Actuators B, vol 50,
been applied in order to compare the correction ability of the pp. 85–94, 1998.
[2] L. Mazet, C. Varenne, A. Pauly, J. Brunet and J.P. Germain, "H2, CO
algorithms. From figure 8 it can be seen that all methods and high vacuum regeneration of ozone poisoned pseudo-Schottky Pd–
present similar performance in correcting one faulty sensor InP based gas sensor", Sensors and Actuators B, vol 103, pp. 190–199,
signal, when two faulty sensors have to be corrected all the 2004.
methods degrades and PCA-MS is the one that degrades the [3] M. Matsumiya, W. Shin, F. Qiu, N. Izu, I. Matsubara and N.
Murayama, "Poisoning of platinum thin film catalyst by (HMDS) for
most. In any case the correction of the faulty signals clearly thermoelectric hydrogen gas sensor", Sensors and Actuators B, vol. 96,
improves classification rates. pp. 516–522 , 2003.
[4] A. Perera, N. Papamichail, N. Barsan, U. Weimar, S. Marco, “On-Line
Normalized classification rates Novelty Detection by Recursive Dynamic PCA and Gas Sensor Arrays
1
Under Drift Conditions”, IEEE Sensors Journal, Vol. 6, pp.770-783,
0.99 2006.
[5] M. Pardo, G. Faglia, G. Sberveglieri, M. Corte, F. Masulli and M.
0.98
Riani, "Monitoring reliability of sensors in an array by neural
0.97 networks", Sensors and Actuators B, vol. 67, pp. 128-133, 2000
0.96
[6] J. F. MacGregor and T. Kourti “Statistical process control of
multivariate processes“, Control Eng. Prac. Vol. 3, pp. 403-414, 1995.
CR
0.95
[7] T. Kourti, P. Nomikos and J. MacGregor “Analysis, monitoring and
0.94 fault diagnosis of batch processes using multiblock and multiway PLS”
non corrected Journal of Process Control, Vol. 5, (1995), pp 277-284
0.93
PCA [8] W. Ku, R. H. Storer and C. Georgakis, “Disturbance detection and
PLS
0.92
PCA-MS isolation by dynamic principal component analysis” Chemometrics and
0.91 Intelligent Laboratory Systems, vol. 30, pp. 179-196, 1995.
no fault 1 fault 2 faults
[9] R. Dunia, S. Joe Qin, T.F. Edgar and T.J. McAvoy, “Use of principal
component analysis for sensor fault identification “ Comp. Chem. Eng.
Fig.8. Normalized classification rates for the corrected responses by the three vol. 20, pp. 713-S718, 1996.
methods, having one and two faulty sensors and compared with non corrected [10] S. J. Qin and W. Li, “Detection, identification and reconstruction of
and no fault cases. faulty sensors with maximized sensitivity”, AIChE Journal, vol..45,
no.9, pp. 1963-1976, 1999.
[11] S. J. Qin and R. Dunia “Determining the number of principal
components for best reconstruction” Journal of Process Control, vol.
V. CONCLUSIONS 10, pp. 245-250, 2000.
[12] J. Gertler and D. Singer, “A new Structural Framework for Parity
Equation Based Failure Detection and Isolation”, Automatica, vol. 26,
In this work the performance of PCA, PLS and PCA-MS pp. 381, 1990.
for sensor poisoning fault diagnosis have been compared. [13] J. Gertler,, W. Li, Y. Huang, and T. McAvoy, “Isolation-enhanced
These techniques have been applied on a dataset consisting on Principal Component Analysis”, AIChE Journal, Vol. 45, pp. 323-334,
the measurements of three analytes at several concentration 1999.
[14] W. Li and S. Shah, “Structured residual vector-based approach to
levels (8 classes) by 17 conducting polymer sensors. Sensor sensor fault detection and isolation”, Journal of Process Control, vol.
poisoning fault has been simulated by changing the faulty 12, pp. 429-443, 2002.
sensor response to one of the 8 classes by its response to any [15] R. Xu and C. Kwan, "Robust isolation of sensor failures", Asian
other class. Journal of Control, vol. 5,pp. 12-23, 2003.
Methods based in PCA are more sensitive than PLS for
medium detection indexes and the contrary for higher
detection indexes, since different regions can be seen in the
ROC plot. In faulty sensor identification all methods show
Fault detection, identification, and reconstruction of faulty chemical
gas sensors under drift conditions, using Principal Component
Analysis and Multiscale-PCA.
M.Padilla, A.Perera, I.Montoliu, A.Chaudry, K.Persaud, S.Marco Member, IEEE
Abstract—Statistical methods like Principal Components can help to detect and compensate the output signal of the
Analysis (PCA) or Partial Least Squares (PLS) and multiscale affected sensor or sensors.
approaches, have been reported to be very useful in the task Poisoning consists of a degradation of the sensing prop-
of fault diagnosis of malfunctioning sensors for several types
of faults. In this work, we compare the performance of PCA erties of a sensor when it is exposed to a gas compound
and Multiscale-PCA on a fault based on a change of sensor that irreversibly interacts with its sensing material. The
sensitivity. This type of fault affects chemical gas sensors and it sensor response gradually degrades over a period of time
is one of the effects of the sensor poisoning. These two methods but seem to be operating normally. Therefore it is not an
will be applied on a dataset composed by the signals of 17 con- easily detectable problem, because the response of the faulty
ductive polymer gas sensors, measuring three analytes at several
concentration levels during 10 months. Therefore, additionally sensor is kept inside its range with an apparently normal
to performance’s comparison, both method’s stability along the behavior, although with an altered profile of sensitivities to
time will be tested. The comparison between both techniques the different analytes. Poisoning is a surface phenomenon
will be made regarding three aspects; detection, identification closely related to physical properties of the sensing material
of the faulty sensors and correction of faulty sensors response. and the detecting gas, so it can affect only some of the
sensors in an array, which have to be immediately detected
I. I NTRODUCTION and identified in order to correct their responses until they
can be replaced by new ones. Effects of sensor poisoning
Chemical gas sensors are a good alternative to conven- are gradually and irreversible changes in its baseline resis-
tional analytic instruments for gas analysis in terms of tance [26], [22], its sensitivity and recovery times [16]. The
promptness of response and low cost. However, gas sensors degree of poisoning depend on the exposure temperature,
suffer from several drawbacks that prevent them from being the time and the concentration of the poisonous substance
robust, for example lack of selectivity, high sensibility to [19][15], for example hexamethyldisiloxane (HMDS) [15],
humidity and temperature, drift and poisoning. The use of Sulfur dioxide (SO2 ) [19], [25], nitrogen dioxide (NO2 ) [16]
chemical gas sensors in an array format with a pattern or ozone (O3 ) [16], depending on the type of gas sensor. To
recognition system, improves the degree of selectivity of overcome poisoning new poison-resistant materials have to
the individual gas sensors. In the sensor array, each sensor be developed, however alternatives have been proposed like
has a partial specificity (non-specific) and its output produce the application of special treatments on the sensors [16], [19],
a compound/odor fingerprint that can be identified by a [25], [3].
pattern recognition system. A sensor array is the base of The detection and identification of a fault, identification
the instrument known as electronic nose, which is inspired of a faulty sensor and correction of the fault (known as
by the mechanisms involved in human olfaction. fault diagnosis) are issues very discussed in literature. In
In an array of gas sensors one or more sensors can fail due literature, a great variety of types of abnormal events have
to a great variety of faults like sensor failure, degradation, been considered and classified as faults, like external events
external interferences, bias, noise, or specific faults related or faults produced by the process or the equipment. However,
to chemical sensors, like drift or poisoning among others. less attention has been paid to faults related to the nature of
Although these problems can be overcome by technology the sensors.
improvements and careful system design, signal processing On the other hand, the employed techniques of fault
diagnosis can be based on statistical methods which have
S. Marco and M. Padilla, Departament d’Electrònica, Universitat de the advantage of extracting information from the process
Barcelona, Martı́ i Franquès 1, 08028 Barcelona, Spain (phone: +34 93
402 90 70 email: smarco@el.ub.es, mpadilla@el.ub.es ). data without taking into account the nature of the system.
Institute for Bioengineering of Catalonia, Baldiri i Reixach, 4-6, 08028 Techniques like Principal Components Analysis (PCA) ex-
Barcelona, Spain ploit the property of correlations among variables in order to
A. Perera, Centre de Recerca en Enginyeria Biomèdica, Universitat
Politécnica de Catalunya, Pau Gargallo 5, 08028 Barcelona, Spain extract this information. Fault detection using PCA or Partial
I. Montoliu, Nestlé Research Center, Metabonomics Biomarkers, CH- Least Squares (PLS) has already been studied in applications
1000 Lausanne 26, Switzerland like process monitoring, quality control, or sensor and pro-
A. Chaudry, Crewe Business Park, Crewe, Cheshire, UK.
K. Persaud, School of Chemical Engineering and Analytical Science, The cess fault diagnosis[23]. Also modifications of PCA have
University of Manchester,UK. been developed to analyze non steady data; non-linear [28],
multiblock [8], [12], multiway [8], [27], dynamic or recursive TABLE I: Sample distribution among analytes and concen-
PCA[9] and combinations [10]. More recently, the wavelets tration levels.
decomposition of the sensor signals before the application of
statistical techniques have reported very good performance Concentration
Analyte Samples levels
in various tasks such as sensor validation [11] or process 447 0.01%
monitoring [1]. Dunia and Qin [5], [4] study the sensor Ammonia 452 0.02%
and process fault detection, identification and reconstruction 307 0.05%
Propanoic 458 0.01%
based on PCA, in another posterior work Qin compared acid 457 0.02%
PCA with Multiscale-PCA (MS-PCA) [18]. In that work MS- 423 0.05%
PCA outperformed the conventional PCA based approach in n-Butanol 446 0.1%
425 1.0%
detecting and identifying the faults.
Event or fault detection on gas sensor arrays has been pre-
viously tackled by other authors. Perera et al.[21] presented
an adaptive method based on a recursive dynamic principal
component analysis (RDPCA) algorithm. This algorithm was 10
on a dataset composed by measurements of three compounds Fig. 1: Transient response of sensor 1 to three analytes. The
at different concentrations, using a 17 sensor array during 10 vertical dashed line shows the time instant where the three
months. The comparison between the algorithms at every signals are maximum.
step of time is made regarding three aspects: detection,
identification of the faulty sensors and correction of faulty
sensors response. to the effect of drift and intra-class variability. The irregular
variation, drift effect, along the time is shown in fig. 3 on the
II. DATA SET DESCRIPTION AND METHODOLOGY
signal of a single sensor (sensor 1) with the samples of only
A. The dataset three classes for better visualization. Drift has similar effect
The data set consist on 3 different gases at different con- on every class of data and, since it is so irregular, it may
centrations levels periodically measured during 10 months, be mostly attributed to environmental sources like changes
with an array of 17 conductive polymer sensors. The total in humidity and temperature, which are usual during an
number of samples in the dataset is 3415 samples assigned experiment of 10 months and to which conductive polymer
to 8 classes, which have been alternatively measured over the sensors are very sensitive.
whole duration of the experiment. The classes corresponded
to solutions of ammonia and propanoic acid, each one of
50
them at three concentration levels, and a third analyte (n- ammonia 0.01%
40 ammonia 0.02%
butanol) at two concentration levels (table I). ammonia 0.05%
For every sample, the full sensor response to the sampling 30 propanoic 0.01%
propanoic 0.02%
transient is recorded. The transient has a time duration of 20
propanoic 0.05%
8.38 %PC2
10
These three capabilities are evaluated on validation set num-
9 ber 1 and 9, both with 342 samples and 50% of the samples
with induced faults in one and two sensors failing at the
8
same time. The validation set number 9 is specially used to
7
ammonia 0.05% test the robustness of the methods to drift at detection and
propanoic acid 0.05%
n−butanol 1% identification points.
0 16 31 4959 80 97 118 142 178 303
time (days) III. M ETHODS
Fig. 3: Drift effect along the time over the signal of sensor A. Principal Component Analysis (PCA)
1. Samples belong to three classes. This approach was proposed by Dunia et al. [5]. The
detection of a fault is based on SP E value (Squared Pre-
diction Error), which represents the squared magnitude of
B. Simulation of sensor poisoning fault
the projection of every sample on the residual subspace.
The faulty sensors simulation are based on a change of Therefore, if x represents one sample in the sensors space, x̂
sensitivity of one or two sensors. To create a faulty sample, the projection of x on the PCA model and x̃ the projection
the response of the faulty sensor to a sample belonging to a on the residual space, with x = x̂ + x̃, then SP E = !x̃!2 .
certain class is changed by its response to any other of the A high value of SPE (over a certain threshold) indicates
remaining classes randomly selected (fig. 4). a faulty sample (fig. 5). To identify the faulty sensor, we
follow the identification by reconstruction approach in [5].
For this, the reconstructed vector with faulty sensor j, xj , is
considered:
xj = x − fj ξj
where ξj is the vector representing the faulty sensor, e.g. a
vector of length equal to the number of sensors in the sensor
array, with all elements 0 except the element j which is 1.
Fig. 4: Simulation of sensor poisoning fault. fj is an estimation of the magnitude of the fault on sensor j,
given by the minimization of SP Ej = !x̃j !2 which finally
gives:
C. Methodology f˜j = ξ˜joT x̃
To test the different algorithms along the time, the com- and, finally:
plete dataset (samples ordered in time) has been divided in
SP Ej = SP E − f˜j2
10 sections, with approx. 342 samples in each section (or set)
with approximatively the same amount of samples per class. where f˜j and ξ˜j are the corresponding elements projected
We refer the first section as calibration set and remaining on the residual space and ξ˜jo means ξ˜j of unit norm. When
ones as validation sets. From 342 samples of the calibration the corrected faulty sensor’s response is corrected, SPE is ex-
set, 311 samples are randomly selected, keeping the same pected to be reduced and its estimated f˜j is large. The index
amount of samples per class, to form the training set at every defined by ηj = SP Ej /SP E is used for the identification of
repetition of the experiment. the faulty sensor. Since for the faulty sensor f˜j is larger and
Comparison between both signal processing techniques is SP E smaller than for remaining sensors, the index ηj should
presented regarding three aspects: be close to zero for this sensor. Therefore, by comparing
• detection of faulty samples; by using receiver operating ηj for all sensors in a detected faulty sample, the one that
characteristic (ROC) curves, which represents the rate shows the minimum value of ηj is the malfunctioning one,
of true alarms versus rate of false alarms when varying or inversely, the index 1/ηj shows that the maximum value
the thresholds of a given detection index. The best corresponds to the faulty sensor.
performance would show the most abrupt corner at the Dunia et al. [24] also proposed a technique for an optimum
upper left of the figure, meaning high rate of true alarms choice of the number of Principal Components (PC’s) for
while low rate of false alarms. the PCA model. This method consist of the calculation
• identification of faulty sensors; by success rate on of the V RE (Variance of Reconstruction Error) index for
correct faulty sensor identification on 15 runs, where all sensors and all possible number of PC’s. A mean, or
one (or two) random sensors is set faulty at every run, weighted mean, value of all sensor’s indexes is minimum at
the best choice of the number of PC’s for the PCA model. instance, isolated singularities, like sharp signal transitions
The VRE in the direction of sensor j, ξj , is given by: which creates large amplitude wavelet coefficients, can be
detected by following across scales the local maxima of the
ξjT Rξj wavelet transform.
uj = var{ξjT (x − xj )} =
(ξjT ξj ) However, the resolution of the obtained information in
time and frequency domains is limited by the uncertainty
where R is the covariance matrix estimated from the training
principle. Resolution depends on the scale and it is propor-
data set.
tional to a in the time domain and to 1/a in the frequency
domain. Therefore, in a time-frequency plane the wavelet
Ψa,b is symbolically represented by a rectangle centered at
(b, fc ) with time and frequency spread proportional their
respective resolutions. When the scale a decreases, the time
support is reduced but the frequency spread increases and
covers an interval of higher frequencies.
Wavelets and its discrete dilations and translations generate
an orthonormal basis of the space of all square integrable
functions (L2 (R)). Therefore, any finite energy signal f
can be decomposed over this wavelet orthogonal basis
{Ψj,n , j, n ∈ Z} at certain scales or resolutions like:
∞
" ∞
"
f (t) = C(j, n)Ψj,n (t) (2)
j=−∞ n=−∞
C(a, b) is computed along the whole function time axis Therefore, the approximation coefficients define the pro-
(all possibles time locations b) at several scales a, it is jection of the signal onto the coarse space Vj preserving
obtained very useful time-frequency information content. For information about the lower frequency components, whereas
the details coefficients retain information about the high are set to 0 (fig. 7). This signal is therefore a representation
frequency components. of SPE as a function of the samples.
Later, the faulty sensor identification step is made by
C. Multiscale-PCA means of the PCA method explained above (III-A) applied
In the present work, multi-scale principal component anal- directly on the detail coefficients d1 matrix.
ysis (MS-PCA) is used for fault detection and identification.
This approach follows the scheme proposed by Barshki [1]
with some differences. MS-PCA extracts maximum infor-
mation from multivariate sensor data from signal trends and
correlation across sensors by combining two complementary
techniques; wavelets, which utilize of the individual sensors’
response to capture the time-frequency information within
the sensor, and PCA which capture correlation among sen-
sors. MS-PCA is built on two steps; the first one consists
on a wavelet decomposition of the individual sensors on a
selected family of wavelets into approximations and details
coefficients at different scales. In the second, every sen-
sor contributions from each scale are collected in separate
matrices, and a PCA model is then constructed to extract Fig. 7: Obtaining SPE vs. samples from SPE of PCA on
correlation at each scale. In fig. 6 a scheme of the method wavelets coefficients at the highest level of MS-PCA.
is shown.
Then, the progression of the measurements can be fol- Finally, the correction of the faulty sensor response is
lowed monitoring the scores and residuals of the wavelets performed by PCA in all matrices of detail coefficients. Since
coefficients of new samples on the PCA models at every the information about the lowest frequencies are contained
scale. If residuals violate certain limits at a given scale then within the approximation coefficients, they are set to 0 for
the sample data contains a fault corresponding to that scale, all sensors in order to improve the stability of their response
so the type of fault can be identified looking at the affected along the time. Then, IDWT is applied on the corrected
scales. For example, if a sudden and abrupt fault occurs it signals.
will be better reflected as a spike in the highest scales of the
MSPCA models since it is a high frequency change, or if IV. R ESULTS
the limits are violated at the approximation scale it implies To apply the MS-PCA approach to our data set, a wavelet
the presence of a slowly drifting fault. basis and scale were chosen: Daubechies 5 with 5 resolution
levels. Then, the multi scale development of the 17 sensors
signals belonging to the training set samples, was performed.
The following step consisted of building a PCA model of the
obtained coefficients at each of the 5 levels. The number of
PC’s of every PCA model varied depending on the validation
set, ranging from 3 to 5 PCs.
Then, the same multi-scale development was made on ev-
ery validation set of 342 samples. The calculated coefficients
were projected on the corresponding PCA model at every
level and then SPE was obtained.
A. Faulty samples detection
At each validation set, faulty samples were identified
Fig. 6: Scheme of the MS-PCA method. by means of the IDWT of the SPE previously obtained.
Fig. 8a shows ROC curves for both methods for validation
A fault is detected when the SPE calculated from the set 1. MS-PCA presents a more sensitive ability for faulty
PCA model of the finest level (highest frequency content) samples detection than PCA, also this method shows much
described by detail coefficients d1, exceed a given thresh- smaller degradation in time than PCA when validated with
old. Because of the nature of poisoning, this type of fault set number 9. This can be seen in fig. 8b.
can be detected at the highest levels of MS-PCA, e.g. d1
coefficients. Then, to identify the faulty samples, an inverse B. Faulty sensor identification
wavelet transform (IDWT) is applied on the SPE. For this, Highest ratios SP Ej /SP E identified the failing sensor
SPE is considered to be the coefficients d1 of a 1-dimensional in every validation set. In fig. 9 the ability of the methods
signal and length equal to the number of samples in the data to identify one or two faulty sensors by means of a percent
subset, whose remaining coefficients in the wavelet transform rate of accuracy is shown. In the case of one faulty sensor
(a) first validation set (a) One faulty sensor
Conclusions
AO systems based on an array of gas sensors and multivariate signal processing show
a lack of robustness. For instance, AO systems exhibit poor time stability and in an
array of sensors one ore more may fail due to a variety of reasons such as poisoning.
Also, chemical sensors are in general influenced by environmental variables such as
humidity and temperature, and the operation conditions can make the system to be
dependant on variables such as flow rate, heating time and heating cycle for MOX
sensors. Consequently, the metrologic characteristics of the instrument are degraded
and therefore need frequent re-calibrations, which is time consuming and expensive.
For example, paper 4.1 shows a degradation along the time in the detection of
moulds, which results in the need of a re-calibration of the system.
However, our results show that multivariate signal processing improve the robustness
of e-nose instruments and therefore extend the lifetime of the system between re-
calibrations. Along this thesis we have studied several techniques that are used with
this purpose. Our most important findings are summarised next.
Feature extraction is a key step to obtain maximum information from sensor re-
sponses. The particular data structure of sensor transient responses has lead us to
propose the use of multiway techniques. The ability of PARAFAC to provide parsi-
monious models has been shown in the analysis of the aroma concentration in potato
chips. Our analysis shows that it can be predicted from a single factor extracted by
PARAFAC algorithm. The analysis of the trilinear decomposition lead in our case
to direct interpretation of the role of the different decomposition modes, since it is
possible to identify the more relevant sensors and time window in transient signals
regarding predictive power.
MCR-ALS has been used for the first time for the analysis of temperature modulated
MOX sensors signals. We have observed that this technique provides an approximate
concentration and sensitivity profiles, even with data from a single sensor. Further,
results are improved by using the signals of the whole array of sensors and/or by
113
Conclusions Marta Padilla Ferran
We studied the ability of orthogonal projection (OP) methods to improve the time
stability of AO systems. In particular, our study has been based on a dataset
containing measurements from 17 conductive polymer sensors during 10 months.
The specific algorithms used to carry out the study were orthogonal signal correction
(OSC) and component correction (CC). We found that both OSC and CC improve
the time stability of the system. Additionally, OSC outperforms CC during the first
half of the experiment. We interpret that this is in part due to OSC having no
dependence on the choice of a reference class. The application of OSC resulted in
more compact clusters, which therefore implied an easier classification task. We have
also tested the robustness of the OP algorithms regarding the size of the calibration
set, and obtained that in our case results are robust up to a limit of about 12
samples per class. On the other hand, the study of the removed components by
both methods, revealed that the part of the sensors transient signal which most
contribute to sample dispersion are the rise and fall regions.
The detection of chemical sensor faults is challenging when the fault produces a shift
in the pattern of sensors sensitivity but the sensors keep responding to the target
sample. In this scenario, our findings demonstrate the ability of simple multivariate
techniques; PLS, PCA and structured residuals, for fault detection and identifica-
tion. In addition, these methods permit the reconstruction of the faulty sensor’s
signal, keeping the system working with minor performance degradation. The pres-
ence of drift may render the detection of faults more difficult. The use of wavelets
decomposition prior to the above mentioned techniques separates the slower changes
(drift) and therefore improve the ability of the method to detect the faults.
114
Capı́tulo 6
Resumen de la tesis
6.1 Introducción
Los sentidos humanos que nos permiten obtener información del exterior sin contacto
directo con la fuentes: vista, oído y olfato, han sido estudiados intensamente con
el objetivo de crear instrumentos que amplien nuestro conocimiento del entorno.
Especialmente, los más estudiados han sido la vista y el oído, mientras que el olfato
siempre se ha considerado un sentido primitivo y de poco interés. Sin embargo, en
2004 el Dr. Richard Axel y la Dra. Linda Buck compartieron el Premio Nobel de
Medicina por sus descubrimientos en este campo y, desde entonces, los estudios sobre
el olor y el sistema olfativo humano han recibido un nuevo interés.
Además, Persaud y Dodd, introdujeron por primera vez la idea de un dispositivo que
podía imitar la capacidad de discriminación del sentido del olfato de los mamíferos.
Las narices electrónicas evitaban a la vez los problemas que presentan los paneles
humanos, tradicionalmente encargados de la evaluación sensoriales de olores cuyas
respuestas están influenciadas por factores humanos como la fatiga, enfermedad,
estado de ánimo, etc. El término nariz electrónica se hizo popular algunos años
más tarde, cuando los investigadores esperaban desarrollar un dispositivo capaz de
115
Resumen de la tesis Marta Padilla Ferran
reconocer los olores y caracterizarlos con descriptores tales como afrutado, olor a
hierba, a tierra, etc. , al igual que el sentido humano del olor. Estos descriptores
serían de gran interés en las industrias de aromas, perfumes, vinos, alimentación,
etc.
Durante los años noventa, la nariz electrónica fue el más prometedor entre los nuevos
métodos analíticos para la evaluación objetiva del olor. Se esperaba incluso que, en
el futuro, el instrumento también fuera un sustituto completo para los métodos de
referencia que utilizan los paneles sensoriales humanos, ya que daría lugar a una
disminución en el costo y la arbitrariedad en los resultados de los paneles.
Sin embargo, pocos años después, la investigación sobre las capacidades de la nariz
electrónica hizo que algunos autores empezaron a cuestionar el nombre dado al
dispositivo, ya que: una nariz electrónica es, obviamente, electrónica, pero no nariz
[3]. Por eso, hoy en día se prefiere el término sistema o instrumento de olfacción
artificial (OA) al de nariz electrónica (e-nariz), aunque a lo largo de la tesis se
utilizan ambos términos indistintamente. Una vez aceptado que las expectativas
iniciales sobre imitar el sentido del olfato no eran realistas, se ha producido un cambio
en el punto de vista del problema. El enfoque actual es diseñar un instrumento
inspirado en la naturaleza, no una repetición de la misma. El término empleado
para esta filosofía es biomimética o bioinspiración.
Otra de las razones por las que en la actualidad, más de 25 años después del primer
dispositivo, las narices electrónicas no hayan tenido el éxito inialmente esperado,
reside en los problemas derivados de la tecnología de sensores. Los sensores quími-
cos de gases, comúnmente utilizados en las narices electrónicas, presentan proble-
mas como dependencias en factores ambientales (temperatura, humedad, etc. ) y en
exposiciones anteriores a otros analitos (efecto memoria), sensibilidades cruzadas,
degradación de su respuesta debido a contaminación de los sensores, cámara y fluí-
dica, etc. Por lo tanto, los instrumentos basados en estos sensores no son robustos
y no dan resultados suficientemente reproducibles.
116
Introducción
Los sensores MOX son probablemente los más ampliamente utilizados, debido a su
bajo precio y características bastante buenas en relación con el tiempo de vida y
sensibilidad; además, están disponibles comercialmente en muchos tipos y con distin-
tas especificidades. Normalmente, estos sensores están compuestos por un sustrato
y un elemento calefactor, cubiertos con el material sensor entre dos electrodos. Su
respuesta se basa en un cambio en la conductividad del material sensor cuando se
expone a los compuestos que inducen en él una reacción de oxidación o reducción.
Además, la sensibilidad que presentan depende de la temperatura [91, 92]. Entre
las desventajas que presentan se encuentran; el alto consumo de potencia del ca-
lenfactor, la posibilidad de envenenamiento con ácidos débiles o compuestos con
azufre, alta sensibilidad al etanol, a la humedad, lenta recuperación que depende de
la muestra, la alta temperatura de funcionamiento que hace que no sean adecuados
para aplicaciones que involucren gases inflamables, etc.
Aparte de mejoras tecnológicas que puedan disminuir las desventajas que presentan
los diferentes sensores, también se pueden usar en varios modos de operación que
117
Resumen de la tesis Marta Padilla Ferran
permitan extraer más información de las muestras. Éstos vienen dados por el diseño
del instrumento que a su vez depende de la aplicación y las cualidades técnicas
deseadas. Según los diseños, los instrumentos varían en coste y complejidad.
Se sabe que la forma del transitorio de un sensor es única para cada par sensor-
analito. También se sabe que la sensibilidad del sensor a un analito dado depende
de la temperatura. Por tanto, el análisis de la forma que tiene la respuesta del sensor,
transitoria o modulada, contiene más información que las medidas tradicionales por
las que se considera únicamente el máximo del transitorio o la respuesta a una
temperatura fija. Con esta información y los algoritmos adecuados de procesado de
señal, se pueden compensar en gran parte la inestabilidad de la respuesta de los
sensores a lo largo del tiempo y otros posibles problemas.
Otro problema importante que afecta a todo tipo de sensores químicos de gases,
aunque en diferentes grados, es la deriva. La deriva es un cambio en la sensibili-
dad de los sensores a lo largo del tiempo debido a los efectos del envejecimiento, el
118
Deriva y fallos en los sensores de una matriz
envenenamiento del material sensor y otros efectos a largo plazo. Algunos autores
incluyen factores ambientales tales como cambios en la humedad, temperaturas y
velocidad de flujo, como causas de la deriva, aunque se describen mejor como in-
terferencias cruzadas. Las derivas de los sensores alteran sus respuestas mediante
cambios de sensibilidad, selectividad, magnitud y línea de base. Estos cambios sue-
len ser graduales si la causa de la deriva es mayoritariamente el envejecimiento de
los sensores, pero pueden producirse bruscamente si los causantes mayoritarios son
cambios en variables medioambientales o un fuerte envenenamiento.
Cuando aplicamos un método de procesado de señal para corregir las derivas, éste
ha de evaluarse. Ésto se puede hacer mediante una comparación de las habilidades
predictivas del aparato antes y después de la corrección. Se pueden comparar las
tasas de clasificación, la relación de Fischer o el error cuadrático medio en predicción
(RMSEP) en aplicaciones de cuantificación.
En los sistemas de OA, se han realizado muchos experimentos donde se han ob-
servado derivas y fallos de los sensores. Para corregir las derivas se han propuesto
diferentes técnicas en la literatura: univaradas, multivariadas, adaptivas, lineales,
no lineales, mediante modelos de derivas, con medidas de gases de referencia a lo
largo del tiempo, con corrección de componentes, etc. Sin embargo, el problema no
está totalmente resuelto y, aunque se ha conseguido mejorar el rendimiento de los
sistemas, sigue siendo necesario la investigación de otros métodos de procesado de
señal.
Por otro lado, en una matriz de sensores, uno o más pueden fallar en un momento
dado debido a multitud de causas, tanto instrumentales como derivadas del propio
119
Resumen de la tesis Marta Padilla Ferran
sensor, y que pueden ser sutiles; es decir, fallos en los que parece que el sensor
funciona correctamente. Por lo tanto, es aconsejable que un sistema basado en una
matriz de gases sea capaz no sólo de detectar fallos en alguno de sus sensores, sino
que, además de identificar al sensor que falla, pueda corregir su respuesta y funcionar
hasta que el sensor pueda ser reemplazado. En la literatura referente a los sistemas
de OA no existen casi trabajos al respecto, aunque es un área de conocimiento
ampliamente desarrollada en otros campos.
Una manera interesante de trabajar con instrumentos que presentan fuertes derivas
es utilizar métodos que extraen información de las medidas, pero sin necesidad de
una calibración previa. En consecuencia, el rendimiento de estos métodos no depen-
den de la deriva, ya que no necesitan referencias de calibración. Un ejemplo es el uso
de métodos no supervisados, como las técnicas multiway, que permiten la creación
de modelos parsimoniosos y de interpretación más sencilla a partir de la información
que contienen las señales transitorias o moduladas de cada sensor. Este problema se
conoce en procesado de señal como separación ciega de fuentes (BSS), aunque en el
campo de la química analítica se le llama resolución de curvas multivariante (MCR).
En los métodos de MCR, una serie de algoritmos descomponen una matriz de datos
en componentes con signficado químico. Entre las técnicas multiway, utilizamos PA-
RAFAC para hacer un modelo con señales transitorias. También exploramos MCR
para obtener información sobre perfiles de concentraciones de analitos en una mez-
cla, y sensibilidades de los sensores a cada analito en función de la temperatura, sin
necesidad de una calibración previa.
Otro problema concreto estudiado en esta tesis, se relaciona con los fallos de sen-
sores en una matriz. Cuando esta situación ocurre, las medidas podrían conducir
120
Publicaciones y ponencias
Un primer objetivo en este trabajo, es realizar pruebas para determinar varias cues-
tiones en relación a la detección de moho por un conjunto de diez sensores MOX
comerciales. Entre las cuestiones estudiadas, se incluye la influencia de cada tipo
de material de construcción y la especie de moho, en el reconocimiento final del
substrato como infectados o no, y, también, si se puede detectar la presencia de
una especie desconocida de moho, así como la presencia de moho sobre substratos
121
Resumen de la tesis Marta Padilla Ferran
desconocidos.
122
Publicaciones y ponencias
mero de factores para el modelo PARAFAC. Para ello, se construyen varios modelos
con distinto número de factores a partir del conjunto de calibración. Luego, una vez
seleccionado el número de factores, se hace una prueba de robustez repitiendo el
procedimiento varias veces para confirmar esta selección.
Como se mencionó en la sección 6.3, los métodos de proyección ortogonal (OP) son
una familia de algoritmos interesantes para hacer frente a la deriva y la dispersión
local de muestras. Estos métodos se utilizan, también, en los problemas de transfe-
rencia de calibración, y como técnica de pre-procesado para simplificar un modelo de
calibración posterior y así aumentar su robustez. Sin embargo, estos métodos casi no
se utilizan en OA, sólo unas pocas referencias se pueden encontrar en la literatura,
por ejemplo [157, 168].
Como ya se mencionó en la sección 6.2, esta clase de métodos pueden extraer in-
123
Resumen de la tesis Marta Padilla Ferran
Por otra parte, MCR-ALS se aplica también sobre datos sintéticos. En primer lugar
se analizan las medidas de un sensor, lo que da lugar a una resolución de la mezcla
de gases: perfiles de concentración C y sensibilidad de temperatura S T . Posterior-
mente, las muestras de datos que contiene las respuestas de todos los sensores en
una matriz de tres vías (muestras × sensores × forma de onda de la modulación
en temperatura) se desdobla en una matriz de dos vías y se analiza. Los resultados
que se obtuvieron a partir de datos sintéticos fueron satisfactorios, sin embargo, se
observaron efectos no lineales sobre los datos reales.
Como en una matriz de sensores uno o más pueden fallar, es importante que se
detecte a tiempo el fallo, se identifique el sensor defectuoso y se corrija su respues-
ta hasta que el sensor pueda ser reemplazado por uno nuevo. En este artículo, se
simula un posible y sutil fallo que afecta típicamente a sensores químicos de gases.
El envenenamiento del sensor es un tipo de fallo que causa un cambio en el perfil de
sensibilidad del sensor, en este efecto se basa la simulación del fallo. El envenena-
miento es difícil de detectar en comparación con otro tipo de fallos, ya que el sensor
sigue trabajando en condiciones aparentemente normales.
En este trabajo, se propuso el uso de métodos estadísticos sencillos como PCA, PLS
y PCA con residuos estructurados, para realizar las tres tareas mencionadas ante-
124
Conclusiones
6.5 Conclusiones
Los sistemas AO basados en una matriz de sensores son poco robustos, presentan
poca estabilidad a lo largo del tiempo y uno o más de sus sensores pueden fallar
debido a muchas causas, como envenenamiento. Además, variables ambientales co-
mo la temperatura y la humedad afectan a los sensores químicos y las condiciones
125
Resumen de la tesis Marta Padilla Ferran
126
Conclusiones
de alrededor de 12 muestras por clase. Por otro lado, el estudio de los componen-
tes eliminados por ambos métodos, reveló que la parte de la señal de los sensores
transitorios que más contribuyen a la dispersión de la muestra son las regiones de
transición: subida y bajada de la señal.
127
Resumen de la tesis Marta Padilla Ferran
128
References
References
[1] K. Persaud and G. Dodd, “Analysis of discrimination mechanisms in the mammalian
olfactory system using a model nose,” Nature, vol. 299, pp. 352 – 355, Jan 1982.
[4] F. Rock, N. Barsan, and U. Weimar, “Electronic nose: Current status and future
trends,” Chemical Reviews, vol. 108, pp. 705–725, Jan 2008.
[6] R. Dutta, E. Hines, J. Gardner, and K. Kashwan, “Tea quality prediction using
a tin oxide-based electronic nose: an artificial intelligence approach,” Sensors and
Actuators B: Chemical, vol. 94, pp. 228–237, Jan 2003.
[7] M. Escuderos, S. Sánchez, and A. Jiménez, “Virgin olive oil sensory evaluation by
an artificial olfactory system, based on quartz crystal microbalance (qcm) sensors,”
Sensors and Actuators B: Chemical, vol. 147, pp. 159–164, Jan 2010.
[8] J. Lozano, T. Arroyo, J. Santos, J. Cabellos, and M. Horrillo, “Electronic nose for
wine ageing detection,” Sensors and Actuators B: Chemical, vol. 133, pp. 180–186,
Jul 2008.
[11] D. Kohl, L. Heinert, J. Bock, T. Hofmann, and P. Schieberle, “Gas sensors for food
aroma during baking and roasting processes based on selective odorant measurements
by an array (hrgc/sommsa),” Thin Solid Films, vol. 391, pp. 303–307, Jan 2001.
129
Resumen de la tesis Marta Padilla Ferran
[15] A. Gomez, J. Wang, G. Hu, and A. Pereira, “Monitoring storage shelf life of tomato
using electronic nose technique,” Journal of Food Engineering, vol. 85, pp. 625–631,
Jan 2008.
[18] L. Senesac and T. Thundat, “Nanosensors for trace explosive detection,” Materials
Today, vol. 11, pp. 28–36, Jan 2008.
[21] F. Bitter, B. Müller, and D. Müller, “Estimation of odour intensity of indoor air
pollutants from building materials with a multi-gas sensor system,” Building and
Environment, vol. 45, pp. 197–204, Jan 2009.
[22] S. Lim, L. Feng, J. Kemling, and C. Musto, “An optoelectronic nose for detection of
toxic gases,” Nature Chemistry, vol. 13, pp. 562–567, Jan 2009.
[23] A. Valleron, C. Pijolat, J. Viricelle, P. Breuil, and J.-C. Marchand, “Exhaust gas
sensor based on tin dioxide for automotive application,” Proceedings of 3th Interna-
tional Symposium on Olfaction and Electronic Nose ISOEN 2009, Brescia : Italy,
Jan 2009.
[25] S. Rappert and R. Müller, “Odor compounds in waste gas emissions from agricultural
operations and food industries,” Waste management, vol. 25, pp. 887–907, Jan 2005.
[26] A. Romain, D. Godefroid, M. Kuske, and J. Nicolas, “Monitoring the exhaust air
of a compost pile as a process variable with an e-nose,” Sensors and Actuators B:
Chemical, vol. 106, pp. 29–35, Jan 2005.
[27] O. Canhoto and N. Magan, “Electronic nose technology for the detection of microbial
and chemical contamination of potable water,” Sensors and Actuators B: Chemical,
vol. 106, pp. 3–6, Jan 2005.
130
References
[31] B. Kateb, M. Ryan, M. Homer, L. Lara, Y. Yin, K. Higa, and M. Chen, “Sniffing out
cancer using the jpl electronic nose: A pilot study of a novel approach to detection
and differentiation of brain cancer,” NeuroImage, vol. 47, no. Supplement 2, pp. T5–
T9, 2009.
[34] E. Thaler and C. Hanson, “Use of an electronic nose to diagnose bacterial sinusitis,”
American journal of rhinology, Jan 2006.
[35] J. Stetter and W. Penrose, “Understanding chemical sensors and chemical sensor
arrays(electronic noses): Past, present, and future,” Sensors update, vol. 10, no. 1,
p. 189, 2002.
[38] E. Schaller, J. Bosset, and F. Escher, “Electronic noses and their application to food,”
LWT-Food Science and Technology, vol. 31, no. 4, pp. 305–316, 1998.
[39] M. Peris and L. Escuder-Gilabert, “A 21st century technique for food control: Elec-
tronic noses,” Analytica chimica acta, vol. 638, no. 1, pp. 1–15, 2009.
131
Resumen de la tesis Marta Padilla Ferran
[41] S. Ampuero and J. Bosset, “The electronic nose applied to dairy products: a review,”
Sensors and Actuators B: Chemical, vol. 94, no. 1, pp. 1–12, 2003.
[42] E. Thaler and C. Hanson, “Medical applications of electronic nose technology,” Expert
Review of Medical Devices, Jan 2005.
[43] J. Gardner, H. Shin, and E. Hines, “An electronic nose system to diagnose illness,”
Sensors and Actuators B: Chemical, vol. 70, no. 1-3, pp. 19–24, 2000.
[44] A. Pavlou and A. Turner, “Sniffing out the truth: clinical diagnosis using the elec-
tronic nose,” Clinical Chemistry and Laboratory Medicine, vol. 38, no. 2, pp. 99–112,
2000.
[45] A. Turner and N. Magan, “Electronic noses and disease diagnostics,” Nature Reviews
Microbiology, vol. 2, no. 2, pp. 161–166, 2004.
[47] D. García-González and R. Aparicio, “Coupling mos sensors and gas chromatography
to interpret the sensor responses to complex food aroma: Application to virgin olive
oil,” Food Chemistry, vol. 120, pp. 572–579, May 2010.
[49] P. Wang, G. Xu, L. Qin, Y. Xu, Y. Li, and R. Li, “Cell-based biosensors and its
application in biomedicine,” Sensors and Actuators B: Chemical, vol. 108, no. 1-2,
pp. 576 – 584, 2005.
[50] Q. Liu, H. Cai, Y. Xu, Y. Li, R. Li, and P. Wang, “Olfactory cell-based biosensor: A
first step towards a neurochip of bioelectronic nose,” Biosensors and Bioelectronics,
vol. 22, no. 2, pp. 318 – 322, 2006.
132
References
[60] T. Nakamoto, “Odor handling and delivery systems,” In Handbook of Machine Ol-
faction - Electronic Nose Technology, pp. 55–78, (ch.3), T.C. Pearce, S.S. Schiffman,
H.T. Nagle, J.W. Gardner (Eds.),Weinheim, Germany: WILEY-VCH, 2003.
[62] G. Settles, “Sniffers: Fluid-dynamic sampling for olfactory trace detection in nature
and homeland security,” Journal of Fluids Engineering, vol. 127, p. 189, 2005.
[66] S. Scott, D. James, Z. Ali, and W. O’Hare, “Optimising of the sensing chamber of an
array of a volatile detection system,” Journal of Thermal Analysis and Calorimetry,
vol. 76, no. 2, pp. 693–708, 2004.
[67] Y. Kim, S. Ha, H. Yang, and Y. Kim, “Gas sensor measurement system capable of
sampling volatile organic compounds (vocs) in wide concentration range,” Sensors
and Actuators B: Chemical, vol. 122, no. 1, pp. 211–218, 2007.
133
Resumen de la tesis Marta Padilla Ferran
[73] D. Wilson, S. Hoyt, J. Janata, K. Booksh, and L. Obando, “Chemical sensors for
portable, handheld field instruments,” IEEE Sensors Journal, vol. 1, pp. 256–274,
Jan 2001.
[74] A. Ricco, R. Crooks, and J. Janata, “Chemical sensors: A perspective of the present
and future,” The Electrochemical Society Interface, vol. 7, pp. 18–24, Jan 1998.
[75] P. Moseley, “Solid state gas sensors,” Measurement Science and Technology, vol. 8,
no. 1, pp. 223–237, 1997.
[77] D. James, S. Scott, Z. Ali, and W. O’Hare, “Chemical sensors for electronic nose
systems,” Microchimica Acta, vol. 149, no. 1, pp. 1–17, 2005.
[78] C. Elosua, I. Matias, C. Bariain, and F. Arregui, “Volatile organic compound optical
fiber sensors: a review,” Sensors, vol. 6, pp. 1440–1465, Jan 2006.
134
References
[83] K. Arshak, E. Moore, G. Lyons, J. Harris, and S. Clifford, “A review of gas sensors
employed in electronic nose applications,” Sensor review, vol. 24, no. 2, pp. 181–198,
2004.
[85] S. Akbar, P. Dutta, and C. Lee, “High-temperature ceramic gas sensors: A review,”
International Journal of Applied Ceramic Technology, vol. 3, pp. 302–311, Jan 2006.
[86] H. Nagle, S. Schiffman, and R. Gutierrez-Osuna, “The how and why of electronic
noses,” IEEE Spectrum, vol. 35, pp. 22–34, Jan 1998.
[87] T. Seiyama, A. Kato, K. Fujiishi, and M. Nagatani, “A new detector for gaseous com-
ponents using semiconductive thin films,” Analytical Chemistry, vol. 34, pp. 1502–
1503, Jan 1962.
[89] N. Yamazoe and K. Shimanoe, “Theory of power laws for semiconductor gas sensors,”
Sensors and Actuators B: Chemical, vol. 128, no. 2, pp. 566–573, 2008.
[90] A. Lee and B. Reedy, “Temperature modulation in semiconductor gas sensing,” Sen-
sors and Actuators B: Chemical, vol. 60, no. 1, pp. 35–42, 1999.
[95] E. Schaller, J. Bosset, and F. Escher, “ ‘electronic noses’ and their application to
food,” LWT-Food Science and Technology, vol. 31, no. 4, pp. 305–316, 1998.
[97] R. Bogue, “Mems sensors: past, present and future,” Sensor review, vol. 27, no. 1,
pp. 7–13, 2007.
[98] G. Korotcenkov, “Metal oxides for solid-state gas sensors: What determines our
choice?,” Materials Science and Engineering: B, vol. 139, pp. 1–23, Jan 2007.
135
Resumen de la tesis Marta Padilla Ferran
[100] N. Barsan, D. Koziej, and U. Weimar, “Metal oxide-based gas sensor research: How
to?,” Sensors and Actuators B: Chemical, vol. 121, no. 1, pp. 18–35, 2007.
[103] R. Gutierrez-Osuna, “Pattern analysis for machine olfaction: a review,” IEEE Sen-
sors Journal, vol. 2, no. 3, pp. 189–202, 2002.
[104] M. Pardo and G. Sberveglieri, “Learning from data: A tutorial with emphasis on
modern pattern recognition methods,” IEEE Sensors Journal, vol. 2, pp. 203–217,
Jul 2002.
[107] E.Lobet, “Dynamic pattern recognition methods and system identification,” In Hand-
book of Machine Olfaction - Electronic Nose Technology, pp. 293–324, (ch.12), T.C.
Pearce, S.S. Schiffman, H.T. Nagle, J.W. Gardner (Eds.),Weinheim, Germany:
WILEY-VCH, 2003.
136
References
[113] D. Wilson and S. DeWeerth, “Signal processing for improving gas sensor response
time,” Sensors and Actuators B: Chemical, vol. 41, pp. 63–70, Jun 1997.
[114] D. Wilson and S. DeWeerth, “Odor discrimination using steady-state and transient
characteristics of tin-oxide sensors,” Sensors and Actuators B: Chemical, vol. 28,
no. 2, pp. 123–128, 1995.
[115] “Instrument manual for nst-3320,” Nordic Sensor Technologies (NST), 2000.
[116] P. G. Age Smilde, Rasmus Bro Multi-Way Analysis with Applications in the Chemical
Sciences, John Wiley & Sons, 2004.
[117] P. Clifford and D. Tuma, “Characteristics of semiconductor gas sensors ii. transient
response to temperature change,” Sensors and Actuators, vol. 3, pp. 255–281, Jan
1983.
[119] A. Ortega, S. Marco, A. Perera, T. Sundic, A. Pardo, and J. Samitier, “An intelli-
gent detector based on temperature modulation of a gas sensor with a digital signal
processor,” Sensors and Actuators B: Chemical, vol. 78, pp. 32–39, Aug 2001.
[120] S. Nakata, E. Ozaki, and N. Ojima, “Gas sensing based on the dynamic nonlinear
responses of a semiconductor gas sensor: dependence on the range and frequency
of a cyclic temperature change,” Analytica chimica acta, vol. 361, pp. 93–100, Mar
1998.
[124] X. Huang, F. Meng, Z. Pi, W. Xu, and J. Liu, “Gas sensing behavior of a single tin
dioxide sensor under dynamic temperature modulation,” Sensors and Actuators B:
Chemical, vol. 99, pp. 444–450, May 2004.
137
Resumen de la tesis Marta Padilla Ferran
[127] R. Gosangi and R. Gutierrez-Osuna, “Active chemical sensing with partially ob-
servable markov decision processes,” Proceedings 13th International Symposium on
Olfaction and Electronic Nose.
[130] C. Song, A. Aiyar, S. Kim, and M. Allen, “Exploitation of aeroelastic effects for drift
reduction, in an all-polymer air flow sensor,” Sensors and Actuators A: Physical,
no. Article in Press, Corrected Proof, 2010.
[131] A. Kersey, T. Berkoff, and W. Morey, “Fiber-optic bragg grating strain sensor with
drift-compensated high-resolution interferometric wavelength-shift detection,” Optics
letters, vol. 18, no. 1, pp. 72–74, 1993.
[133] M. Malátek, P. Ripka, and L. Kraus, “Temperature offset drift of gmi sensors,”
Sensors and Actuators A: Physical, vol. 147, no. 2, pp. 415–418, 2008.
[134] D.J.Hand Construction and Assessment of Classification Rules, John Wiley, Chich-
ester, 1997.
[135] K. Copsey and A. Webb, Classifier design for population and sensor drift. 2004.
[138] C. Stork and B. Kowalski, “Distinguishing between process upsets and sensor mal-
functions using sensor redundancy,” Chemometrics and intelligent laboratory systems,
vol. 46, no. 2, pp. 117–131, 1999.
[139] R. Dunia and S. J. Qin, “A unified geometric approach to process and sensor fault
identification and reconstruction: the unidimensional fault case,” Computers and
Chemical Engineering, vol. 22, pp. 927 – 943, Jan 1998.
[140] J. Haugen, O. Tomic, and K. Kvaal, “A calibration method for handling the temporal
drift of solid state gas-sensors,” Analytica chimica acta, vol. 407, no. 1-2, pp. 23–39,
2000.
138
References
[144] S. Labreche, S. Bazzo, S. Cade, and E. Chanie, “Shelf life determination by electronic
nose: application to milk,” Sensors and Actuators B: Chemical, vol. 106, pp. 199–206,
Apr 2005.
[147] M. Kuske, R. Rubio, A. Romain, J. Nicolas, and S. Marco, “Fuzzy k-nn applied to
moulds detection,” Sensors and Actuators B: Chemical, vol. 106, pp. 52–60, Apr
2005.
[148] W. Tsujita, A. Yoshino, H. Ishida, and T. Moriizumi, “Gas sensor network for air-
pollution monitoring,” Sensors and Actuators B: Chemical, vol. 110, pp. 304–311,
Jan 2005.
[150] A. Romain and J. Nicolas, “Long term stability of metal oxide-based gas sensors
for e-nose environmental applications: an overview,” AIP Conference Proceedings,
vol. 146, no. 2, pp. 502–506, 2009.
[151] A. Romain, P. Andre, and J. Nicolas, “Three years experiment with the same tin
oxide sensor arrays for the identification of malodorous sources in the environment,”
Sensors and Actuators B: Chemical, vol. 84, pp. 271–277, Jan 2002.
[152] J. Nicolas and A. Romain, “Establishing the limit of detection and the resolution
limits of odorous sources in the environment for an array of metal oxide gas sensors,”
Sensors and Actuators B: Chemical, vol. 99, pp. 384–392, Jan 2004.
[153] S. D. Vito, M. Piga, L. Martinotto, and G. D. Francia, “Co, no2 and nox urban
pollution monitoring with on-field calibrated electronic nose by automatic bayesian
regularization,” Sensors and Actuators B: Chemical, vol. 143, pp. 182–191, Jan 2009.
139
Resumen de la tesis Marta Padilla Ferran
[154] S. D. Vito, E. Massera, M. Piga, L. Martinotto, and G. D. Francia, “On field cali-
bration of an electronic nose for benzene estimation in an urban pollution monitoring
scenario,” Sensors and Actuators B: Chemical, vol. 129, pp. 750–757, Jan 2008.
[155] K. Pratt and D. Williams, “Self diagnostic gas sensitive resistors in sour gas appli-
cations,” Sensors and Actuators B: Chemical, vol. 45, no. 2, pp. 147–153, 1997.
[157] O. Tomic, H. Ulmer, and J. Haugen, “Standardization methods for handling instru-
ment related signal shift in gas-sensor array measurement data,” Analytica chimica
acta, vol. 472, no. 1-2, pp. 99–111, 2002.
[158] D. Bednarczyk and S. DeWeerth, “Smart chemical sensing arrays using tin oxide sen-
sors and analog winner-take-all signal processing,” Sensors and Actuators B: Chem-
ical, vol. 27, no. 1-3, pp. 271–274, 1995.
[159] M. Salit and G. Turk, “A drift correction procedure,” Analytical Chemistry, vol. 70,
pp. 3184–3190, Jan 1998.
[162] S. Marco, A. Ortega, A. Pardo, and J. Samitier, “Gas identification with tin oxide
sensor array and self-organizing maps: adaptive correction of sensor drifts,” Instru-
mentation and Measurement, IEEE Transactions on, vol. 47, no. 1, pp. 316 – 321,
1998.
[163] M. Zuppa, C. Distante, P. Siciliano, and K. Persaud, “Drift counteraction with multi-
ple self-organising maps for an electronic nose,” Sensors and Actuators B: Chemical,
vol. 98, pp. 305–317, Jan 2004.
[166] W. Ku, R. Storer, and C. Georgakis, “Disturbance detection and isolation by dy-
namic principal component analysis,” Chemometrics and intelligent laboratory sys-
tems, vol. 30, pp. 179–196, Jan 1995.
140
References
[167] E. Russell, L. Chiang, and R. Braatz, “Fault detection in industrial processes using
canonical variate analysis and dynamic principal component analysis,” Chemometrics
and intelligent laboratory systems, vol. 51, pp. 81–93, Jan 2000.
[169] M. Kermit and O. Tomic, “Independent component analysis applied on gas sensor
array measurement data,” Sensors Journal, IEEE, vol. 3, no. 2, pp. 218– 228, 2003.
[170] F. Hossein-Babaei and V. Ghafarinia, “Compensation for the drift-like terms caused
by environmental fluctuations in the responses of chemoresistive gas sensors,” Sensors
and Actuators B: Chemical, vol. 143, pp. 641–648, Jan 2009.
[171] D. hui, L. Jun-hua, and S. Zhong-ru, “Drift reduction of gas sensor by wavelet and
principal component analysis,” Sensors and Actuators B: Chemical, vol. 96, pp. 354–
363, Jan 2003.
[173] R. Gutierrez-Osuna, “Drift reduction for metal-oxide sensor arrays using canonical
correlation regression and partial least squares,” Proceedings of the 7th International
Symp. On Olfaction and Electronic Nose, pp. Brighton, UK, July 20–24„ 2000.
[175] Brahim-Belhouari, Bermak, Minghua, and Chan, “Fast and robust gas identification
system using an integrated gas sensor technology and gaussian mixture models,”
IEEE Sensors Journal, vol. 5, no. 6, pp. 1433–1444, 2005.
[176] B. Sisk and N. Lewis, “Comparison of analytical methods and calibration methods
for correction of detector response drift in arrays of carbon black-polymer composite
vapor detectors,” Sensors and Actuators B: Chemical, vol. 104, pp. 249–268, Jan
2005.
[177] C. Distante, P. Sicilian, and K. Persaud, “Dynamic cluster recognition with multiple
self-organising maps,” Pattern Analysis and Applications, vol. 5, pp. 306–315, Jan
2002.
[181] B. Dejaegher and Y. Heyden, “Ruggedness and robustness testing,” Journal of Chro-
matography A, vol. 1158, no. 1-2, pp. 138–157, 2007.
141
Resumen de la tesis Marta Padilla Ferran
[184] G. Box, “Non-normality and tests on variances,” Biometrika, vol. 40, pp. 318–335,
Dec 1953.
[186] S. Møller, J. von Frese, and R. Bro, “Robust methods for multivariate data analysis,”
Journal of Chemometrics, vol. 19, no. 10, pp. 549–563, 2006.
[188] H. Swierenga, A. D. Weijer, and R. V. Wijk, “Strategy for constructing robust multi-
variate calibration models,” Chemometrics and intelligent laboratory systems, vol. 49,
pp. 1–17, Jan 1999.
[190] S. Wold, H. Antti, F. Lindgren, and J. Öhman, “Orthogonal signal correction of near-
infrared spectra,” Chemometrics and intelligent laboratory systems, vol. 44, pp. 175–
185, Jan 1998.
[191] T. Fearn, “On orthogonal signal correction,” Chemometrics and intelligent laboratory
systems, vol. 50, pp. 47–52, Jan 2000.
[195] J. Trygg and S. Wold, “Orthogonal projections to latent structures (o-pls),” Journal
of Chemometrics, vol. 16, pp. 119–128, Jan 2002.
[196] J. Trygg, “O2-pls for qualitative and quantitative analysis in multivariate cali-
bration,” Journal of Chemometrics, vol. 16, no. 6, pp. 283–293, 2002.
142
References
[202] Y. Zhu, T. Fearn, D. Samuel, A. Dhar, O. Hameed, S. Bown, and L. Lovat, “Error
removal by orthogonal subtraction (eros): a customised pre-treatment for spectro-
scopic data,” Journal of Chemometrics, vol. 22, no. 2, pp. 130–134, 2008.
[203] B. Li, C. Wang, and D.-S. Huang, “Supervised feature extraction based on orthogonal
discriminant projection,” Neurocomputing, vol. 73, pp. 191–196, Oct 2009.
[205] A. Lorber, “Error propagation and figures of merit for quantification by solving ma-
trix equations,” Analytical Chemistry, vol. 58, pp. 1167–1172, Jan 1986.
[206] H. Goicoechea and A. Olivieri, “A comparison of orthogonal signal correction and net
analyte preprocessing methods. theoretical and experimental study,” Chemometrics
and intelligent laboratory systems, vol. 56, pp. 73–81, Jan 2001.
[207] A. Lorber, K. Faber, and B. Kowalski, “Net analyte signal calculation in multivariate
calibration,” Analytical Chemistry, vol. 68, pp. 1620–1626, Jan 1997.
[208] K. Park and C. Jun, “Hybrid linear analysis based on the net analyte signal in
spectral responses with orthogonal signal correction,” Near Infrared Analysis, vol. 1,
pp. 1–8, Jan 2000.
[209] P. Hansen, “Pre-processing method minimizing the need for reference analyses,” Jour-
nal of Chemometrics, vol. 15, no. 2, pp. 123–131, 2001.
143
Resumen de la tesis Marta Padilla Ferran
[216] Z. Liu, Z. Huang, F. Wang, and Y. Zhou, “Doa estimation with uniform linear arrays
in the presence of mutual coupling via blind calibration,” Signal Processing, vol. 89,
no. 7, pp. 1446–1456, 2009.
[218] A. de Juan and R. Tauler, “Multivariate curve resolution (mcr) from 2000: progress in
concepts and applications,” Critical Reviews in Analytical Chemistry, vol. 36, no. 3,
pp. 163–176, 2006.
[220] G. Wang, Q. Ding, and Z. Hou, “Independent component analysis and its applications
in signal processing for analytical chemistry,” TrAC Trends in Analytical Chemistry,
vol. 27, no. 4, pp. 368–376, 2008.
144
References
[228] K. Schostack and E. Malinowski, “Investigation of window factor analysis and ma-
trix regression analysis in chromatography,” Chemometrics and intelligent laboratory
systems, vol. 20, no. 2, pp. 173–182, 1993.
[229] R. Manne, H. Shen, and Y. Liang, “Subwindow factor analysis,” Chemometrics and
intelligent laboratory systems, vol. 45, pp. 171–176, Jan 1999.
[230] O. Kvalheim and Y. Liang, “Heuristic evolving latent projections: resolving two-way
multicomponent data. 1. selectivity, latent-projective graph, datascope, local rank,
and unique resolution,” Analytical Chemistry, vol. 64, pp. 936–946, Apr 1992.
[231] P. Paatero, U. Tapper, P. Aalto, and M. Kulmala, “Matrix factorization methods for
analysing diffusion battery data,” Journal of Aerosol Science, vol. 22, no. Supplement
1, pp. 273–276, 1991.
[232] M. Turk and A. Pentland, “Eigenfaces for recognition,” Journal of Cognitive Neuro-
science, vol. 3, pp. 71–86, Jan 1991.
[233] E. Acar and B. Yener, “Unsupervised multiway data analysis: A literature survey,”
IEEE Transactions on Knowledge and Data Engineering, vol. 21, no. 1, pp. 6–20,
2009.
[234] A. de Juan and R. Tauler, “Comparison of three-way resolution methods for non-
trilinear chemical data sets,” Journal of Chemometrics, vol. 15, no. 10, pp. 749–771,
2001.
[235] T. Skov and R. Bro, “A new approach for modelling sensor based data,” Sensors and
Actuators B: Chemical, vol. 106, pp. 719–729, May 2005.
[237] O. Shaham, L. Carmel, and D. Harel, “On mappings between electronic noses,”
Sensors and Actuators B: Chemical, vol. 106, pp. 76–82, Jan 2005.
145
Resumen de la tesis Marta Padilla Ferran
[240] P. Kadlec, B. Gabrys, and S. Strandt, “Data-driven soft sensors in the process in-
dustry,” Computers and Chemical Engineering, vol. 33, no. 4, pp. 795–814, 2009.
[241] S. Qin, H. Yue, and R. Dunia, “Self-validating inferential sensors with application
to air emission monitoring,” Industrial and Engineering Chemistry Research, vol. 36,
pp. 1675–1685, May 1997.
[242] G. Goodwin, “Predicting the performance of soft sensors as a route to low cost
automation,” Annual Reviews in Control, vol. 24, no. 1, pp. 55–66, 2000.
[243] M. Henry and D. Clarke, “The self-validating sensor: rationale, definitions and ex-
amples,” Control Engineering Practice, vol. 1, no. 4, pp. 585–610, 1993.
[244] S. X. Ding, P. Zhang, A. Naik, E. L. Ding, and B. Huang, “Subspace method aided
data-driven design of fault detection and isolation systems,” Journal of Process Con-
trol, vol. 19, pp. 1496–1510, Oct 2009.
[245] H. Wang, T.-Y. Chai, J.-L. Ding, and M. Brown, “Data driven fault diagnosis and
fault tolerant control: Some advances and possible new directions,” Acta Automatica
Sinica, vol. 35, pp. 739–747, Jun 2009.
146