marta tesis final

Universitat de BarcelonaFacultat de FísicaDepartament d’Electrònica

Memòria per optar al títol de Doctorper la Universitat de BarcelonaPrograma de doctorat: Enginyeria i tecnologia

electròniques (2003/2005)

Improving the Robustness of ArtificialOlfaction Systems by Multivariate Signal

Processing

Autora:Marta Padilla Ferran

Director:Dr. Santiago Marco Colás

July 27, 2010

Acknowledgements

Son muchas las personas a las que quiero agradecer por su influencia positiva en mivida durante estos largos años de tesis. No puedo nombrarlas a todas aquí y les pidodisculpas.

En primer lugar muchas gracias al Dr. Santiago Marco, director de mi tesis, que meha guiado, enseñado y ayudado para que, a pesar de las dificultades, este proyectosaliera (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 enel 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, ylos que se fueron: Benja, Ali, y Aina. Muchas gracias a todos por vuestra ayuda yánimos!!.

También quiero agradecer a la gente del departament d’electrònica, que me encuentroa menudo por los pasillos y en los comedores. Me alegro mucho de haber estadoestos años en el departamento con vosotros compartiendo impresora, picatrònics ycalçotadas, gracias por tan buen ambiente!. Especialmente a Pere Miribel, que meha ayudado con todos los papeles, sobretodo los necesarios para presentar esta tesis.

A lo largo de la tesis he tenido la oportunidad de hacer una estancia de dos meses enArlon, Bélgica, 15 días en Estocolmo y otros 15 días en Nápoles. En Arlon, estuveen el departamento de Surveillance de l’Environnement de la Universidad de Lieja,a cargo del Dr. Jacques Nicolas. Allí me acogieron muy bien, tuve unos compañerosestupendos y conocí gente muy interesante. Muchas gracias Jacques, Julien, Anne-Claude y Martyna!. En Estocolmo, estuve en el grupo del Brain Institute lideradopor el Dr. Anders Lansner, en el Royal Institute of Technology (KTH). Fui enfebrero, y tuve la suerte de coincidir con el invierno más suave de la historia de Sue-cia, menos mal!. También me trataron muy bien allí, muchas gracias Dr. Lansner,Simon y Malin!. La última estancia que hice fue en Nápoles, en ENEA, Agenzianazionale per le nuove tecnologie, l’energia e lo sviluppo economico sostenibile, bajola supervisión de Saverio De Vito. Muchas gracias Saverio!, a ti, a tu familia y a tuscompañeros.

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, amis principales amigos; Ceci, Bel, Lucía, Ani, Patricia, Paco ..... a todos muchísimasgracias por vuestros ánimos y por estar conmigo en los tiempos menos buenos!.

iii

Marta Padilla Ferran

Finalmente, hay un par de frases que creo que pasarán a la posteridad:

Todo es mejorable.Santi.

Real life is much more difficult.Menda.

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 ofhumans, this information is collected through the five senses ; sight, hearing, touch,smell and taste. Sight, hearing and smell senses are considered specially interestingbecause of their ability to get external information without direct interaction withthe sources. Sight is probably the most developed sense in humans, followed closelyby hearing. On the contrary, smell has always been considered the most primitiveand less important of those three senses, and was an almost total mystery until afew years ago. Recently, in 2004, Dr. Richard Axel and Dr. Linda Buck shared theNobel Prize in medicine for their discoveries of odorant receptors and the organizationof the olfactory system, since then, studies on the smell and human olfactory systemhave received a new interest.

With the aim of extending our senses abilities of obtaining information about oursurroundings, 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 veryextended use, like video cameras used for security issues in which image recognitioncan 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 formany qualitative and quantitative applications, since they were expected to providecharacteristics such as being of small size, low cost, fast and easy to use. Thesefeatures are specially interesting for on-field applications, compared to other well-established instruments for gas/volatiles analysis which are big, heavy, expensiveand 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 presenton the market. The main reason lies in the sensing area of the instrument, whichexhibits poor selectivity and bad stability. The chemical gas sensors used in elec-tronic noses present problems like cross sensitivities, time instability, dependenceon previous gas exposures, etc. Therefore, instruments based on these sensors arenot robust and do not give enough reproducible results. The nature of the prob-lems that influence chemical gas sensors is mainly technological, but affect sensorsof all state of the art technologies, though to different degrees. These deficienciescan be mostly overcome as more research is made on improving fabrication processor developing new technologies. However, while gas sensors technologies are beingimproved, statistical signal processing can help to mathematically compensate, orat least to reduce, the effect those mentioned issues have on the sensors responsesbefore 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 corrector 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 fivechapters. In the first one, the concepts of machine olfaction and electronic nosesare 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 bysignal processing techniques are presented in this chapter. Concepts and definitionsof 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 thepapers which shows many, but not all, of the results obtained during these years ofwork. Finally, conclusions are given in chapter five.

vi

Contents

1 Introduction 11.1 Electronic noses . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1

1.1.1 Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . 41.1.2 State-of-the-Art and future trends . . . . . . . . . . . . . . . 5

1.2 The electronic nose instrument . . . . . . . . . . . . . . . . . . . . . 61.2.1 Instrument general requirements . . . . . . . . . . . . . . . . 61.2.2 The sampling unit and delivery system . . . . . . . . . . . . 81.2.3 The sensors chamber . . . . . . . . . . . . . . . . . . . . . . 9

1.3 Gas sensors technologies . . . . . . . . . . . . . . . . . . . . . . . . 111.3.1 Metal Oxide Semiconductor sensors (MOX) . . . . . . . . . 131.3.2 Conductive Polymer Chemoresistors (CP) . . . . . . . . . . 16

1.4 Pattern recognition system . . . . . . . . . . . . . . . . . . . . . . . 171.5 Operation Modes . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18

1.5.1 Sampling Transient Analysis . . . . . . . . . . . . . . . . . . 191.5.2 Temperature-Modulation Analysis . . . . . . . . . . . . . . . 22

1.6 Drift and sensor faults in a gas sensors array . . . . . . . . . . . . . 241.6.1 Characteristics of drift and sensor poisoning . . . . . . . . . 241.6.2 Factors involved in measurements instability . . . . . . . . . 251.6.3 Effects of drift and poisoning on the sensors response . . . . 261.6.4 Evidence of drift and faulty sensors in electronic nose appli-

cations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 281.6.5 Drift correction and fault detection by signal processing tech-

niques . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31

2 Robustness 412.1 Robustness . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43

2.1.1 Definition of Robustness and Ruggedness . . . . . . . . . . . 432.1.2 Robustness and Ruggedness tests . . . . . . . . . . . . . . . 442.1.3 Robustness of calibration models . . . . . . . . . . . . . . . 45

2.2 Signal processing methods for improving robustness . . . . . . . . . 472.2.1 Orthogonal Projection (OP) methods . . . . . . . . . . . . . 472.2.2 Multivariate Curve Resolution (MCR) methods . . . . . . . 502.2.3 Multiway (N-way) methods . . . . . . . . . . . . . . . . . . 532.2.4 Calibration Transfer . . . . . . . . . . . . . . . . . . . . . . 552.2.5 Fault detection and isolation (FDI) methods . . . . . . . . . 58

vii

Contents Marta Padilla Ferran

3 Thesis objectives 61

4 Journal and main Conference Papers 634.1 Introduction to papers . . . . . . . . . . . . . . . . . . . . . . . . . 634.2 Full papers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67

I. Detection of diverse mould species growing on building materialsby gas sensor arrays and pattern recognition . . . . . . . . . 67

II. Feature extraction on three way enose signals . . . . . . . . . . . 76III. Drift compensation of gas sensor array data by Orthogonal Signal

Correction . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82IV. Multivariate curve resolution applied to temperature-modulated

metal oxide gas sensors . . . . . . . . . . . . . . . . . . . . . 90V. Poisoning fault diagnosis in chemical gas sensor arrays using mul-

tivariate statistical signal processing and structured residualsgeneration . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100

VI. Fault detection, identification, and reconstruction of faulty chem-ical gas sensors under drift conditions, using Principal Com-ponent Analysis and Multiscale-PCA . . . . . . . . . . . . . 106

5 Conclusions 113

6 Resumen de la tesis 1156.1 Introducción . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115

6.1.1 Tecnologías sensoras . . . . . . . . . . . . . . . . . . . . . . 1176.1.2 Modos de operación y procesado de señal . . . . . . . . . . . 117

6.2 Deriva y fallos en los sensores de una matriz . . . . . . . . . . . . . 1186.3 Propuesta de técnicas de procesado de señal para mejorar la robustez

del sistema . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1206.4 Publicaciones y ponencias . . . . . . . . . . . . . . . . . . . . . . . 1216.5 Conclusiones . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125

References 129

viii

Chapter 1

Introduction

In humans, the sense of smell gives place not only to instinctive responses butalso 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 beimitated by an artificial instrument with the actual technology. This reason, andproblems derived from sensors technologies have made that electronic noses havenot reached the initially expected success.

In this chapter, artificial olfaction systems, or electronic noses, are described. Itshistory 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 thisthesis are described, as well as strategies to extract maximum information from thesensors-analyte interactions. Finally, a deeper insight is made on the main problemsthat affect the chemical gas sensors; characteristics, evidence and state of the art incounteraction by signal processing methods.

1.1 Electronic noses

Any device that attempts to imitate, or is inspired by the biological mechanisms ofthe senses, must include elements equivalent to our sensory systems; specific sen-sors to convert a physical magnitude into signals (electrical), an electronic circuitry(nervous system) for sensor conditioning and signal processing, and a processor tocontrol the system and process the data (brain). Hence, an instrument that at-tempts to emulate the human sense of smell, must be designed keeping a certaindegree of parallelism with the human olfactory system. It must include three mainfunctional blocks; a sampling unit playing the role of the nose and the ventilatoryfunction, an array of sensors as the olfactory receptors and a data analysis/pattern

1

Introduction Marta Padilla Ferran

recognition unit to process sensors outputs and make a final decision, as the braindoes. 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 basedon an array of three gas sensors with partial selectivity together with a processingof the sensors response with a neural network, which discriminated among differentcompounds the sensor array was exposed to. In that work, the authors introducedfor the first time the idea of a device that could mimic the discrimination ability ofthe 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 themwith 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 interestin 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 wheresome gas compounds have to be differentiated. On the one hand, the objective ofgas 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 touse, and deliver an outcome after a long time, but, on the contrary, results providea lot of information about the sample composition. A gas chromatographer/massSpectrometer (GC/MS) is a gas analyser indispensable and irreplaceable in manyapplications. On the other hand, gas detectors are designed to fast detect a targetgas 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 specificapplications. Electronic noses intend to be small size devices, low cost, low powerconsumption, easy to use, portable and fast both for qualitative and quantitativeapplications, requiring minimal or even absent sample preparation like prior separa-tion in simpler compounds. Since these instruments were also expected to contributewith some particular abilities coming from the emulation of the human olfaction,they would offer an alternative technology for important applications that are notcovered by the mentioned (and others) well known gas instruments. The electronicnose would complement those instruments, even sometimes its capabilities wouldoverlap 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 substitutefor reference methods that use human sensory panels, since there would be a highdecrease in the cost and arbitrariness in the results of the panels.

During the nineties, given all these expected characteristics, the electronic nose wasthe 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

tial specificity and an appropriate pattern-recognition system, capable of recognizingsimple or complex odours.

However, a few years later, research on the capabilities of the electronic nose madethat 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 definitionhas been given: an attempt to mimic the principles of smelling that gives anotherview 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 cannotmimic olfaction. However, the term electronic nose is still in use, although the onlycommon 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 aredifferent. 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 applicationdue to technological limitations.

Today, once accepted that the initial expectations about mimicking the sense ofsmell 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 replicationof 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 thesmall 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 theseproblems. Nevertheless, also signal processing techniques can help to mathemati-cally compensate the degradation of the sensors responses and thus, to improve theperformance 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 frombiology and targeted to specific applications. In those applications these instrumentsshould fulfil the needs that other devices cannot cover, like low cost, portability, lowpower consumption, simplicity of use, etc. Effort is being made on the improvementof these characteristics and others related to the improvement of robustness andperformance.

3


1.1.1 Applications

E-nose instruments are used in many applications such as medical, pharmaceutical,cosmetics, food processing, safety, security, military, automotive, etc. In those ap-plications, AO systems do not attempt to give any human odour impression but todetect volatiles that give information about the samples.

For example, there are many references that study the quality of food products thatare 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 todifferentiate 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 duringstorage 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]. Environmentalmonitoring applications include the characterization of the pollutant level of ambientoutdoor [20] and indoor air [21]. Specifically the detection of toxic [22] and exhaustgases [23], leakages of certain gases [24] or the control of the ventilation of an indooratmosphere, determination of the odours emissions from animal production facilities,malodorous produced through industrial factories and from decomposition processof wastes in compost plants [25, 26]. Also the detection of microorganisms andresidues 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 allergieslike 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, alsodiabetes, halitosis [33], sinusitis [34], renal insufficiency, etc. However, given thevariability of human beings and the influence of lifestyle (diet, exercise, smoking,etc. ), more research have to be done in health applications.

In many of these examples, the target characteristics result from a combination ofmany volatile compounds on a background containing also many other volatiles.This is the case of coffee or tea, spoilage, malodorous, infections detection and soon. These applications in which complex samples are involved, are very suitable forelectronic noses, since other conventional analytical methods would try to give infor-mation about all the compounds when only a global evaluation is needed. However,the same applications turn difficult if the odour depends on minor compounds andmajor ones are highly variable.

Very good reviews on the applications of AO systems can be found in [35, 36], andmore recently [4, 37], namely, applications on food control [38–40], dairy products[41], medical applications [42] and disease diagnostics [43–45].

4

Electronic noses

1.1.2 State-of-the-Art and future trends

Many of the above mentioned applications are carried out in research laboratoriesand others require highly optimised systems for specific operations, but real-lifeapplications are still difficult to cover. In such applications, the high number ofinterferents may hide the target compound and thus it is needed an increase in thesensitivity and selectivity in the system. For this, several sensor technologies arecombined in the same instrument to provide complementary information, these arehybrid e-noses. However, these new devices obtain better selectivity range at theexpense 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 aresometimes considered as e-nose instruments, like ion mobility spectrometers (IMS)and fast gas chromatographers [4]. Nevertheless, the design of a commercial e-nosedevice depends on the target application or applications, since an ideal all-purposeinstrument is not possible nowadays due to technology limitations.

Actually, research is focused on new sensing materials, which provide new usefulinformation sometimes specific for a given type of target analytes, such as bio andcell-based biosensors which show high selectivity especially for certain organic com-pounds [49, 50], for example biological olfactory receptors [51]. Research is alsomade about other issues aimed to improve the device performance such as waysto increase the number of sensors in a system, micromachining to reduce powerconsumption [52–55], optimisation modelling, etc. (fig. 1.1).

Figure 1.1: Example of a monolithic multitransducer system. It contains four polymer-based sensors relying on two capacitive and two gravimetric transducers, two MOX-basedsensors on temperature-controlled (also modulated) microhotplates, the respective drivingand signal processing electronics, and a digital communication interface [52].

Rock et al. [4] give a complete overview of the current state of the art as well as

5


future trends in electronic noses. They also present a table containing commercialelectronic noses on the market in 2008.

1.2 The electronic nose instrument

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 amore detailed description about the sensors technologies used in this thesis is given,with emphasis in the transduction mechanisms and characteristics that lead to theirparticular advantages and disadvantages.

1.2.1 Instrument general requirements

As already mentioned, the most evident requirements for an electronic nose are alow price, high portability and quick response. The sensors are responsible for suchrequirements, since they are the heart of the device. Nevertheless, the remainingparts 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 becarefully 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), fastresponse and recovery times, robustness, time stability, reproducibility, small sizeand low power consumption, can be met in the same device. Therefore the idealsensors characteristics must be evaluated and prioritised, sometimes establishingcompromises among several of them, according to the application the instrument isgoing to be used for.

From the definition and philosophy of the electronic nose, it is obviously desiredthat sensors response and recovery time are short. Short recovery time would mean(almost) perfect reversibility of the physiochemical detection or sensing process, andthus long sensor life and the possibility of making many measurements in a shorttime. High sensitivity or low detection limits are also required in many applica-tions. Usually, gas sensors reach ppm to ppb detection limits. Linearity is anotherinteresting property, it is usually accomplished for low concentrations of analyte,but as concentrations increases non-linearity becomes more evident until a point ofsaturation. But it has to be taken into account that high analyte concentrationlevels also influences sensors response and recovery times, which may become muchlonger and may be not reversible. Large dynamic range, so that the sensors areable to respond from very low to very high concentrations with similar sensitivityand no saturation, is also highly desired. Other nice property include the possibilityto 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 foundthat the performance of sensor array can be improved by using a combination ofbroad selective and high selective sensors [56–58]. Finally, low cross-sensitivity tointerferents, and long-term stability of the sensor and sensing material are desiredas well [59].

Some of these properties, such us sensor sensitivity, selectivity, speed of response, andreversibility, 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 givenrequirement may imply difficulties to fulfil other. For instance, high sensitivity andselectivity on the one hand are typically associated with strong interactions, whereasperfect reversibility on the other hand requires weak interactions [59]. Consequently,a trade-off among these properties is necessary. However, selectivity and sensitivitycan be in turn improved by the signal processing of some of the sensors individualfeatures, as will be explained in sections 1.3 and 1.5.

Gas sensor devices often present lack of reproducibility in their responses. Thismakes difficult to establish a standard for measurements methodology and also de-creases the robustness of the instrument. For this reason, when a faulty sensor hasto be replaced by a new one, a recalibration has to be performed, although bothsensors are supposed to be identical. This loss of calibration is mostly due to thevariability 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 differenttransducing 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 sensorresponse 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 sensingmaterial and other long-term effects. Environmental agents such as changes in hu-midity, temperatures and flow rates, are also included by some authors into thecauses of drift, although they are better described as cross-interferences.

Besides, poisoning is a change in the response of a sensor to the same analytemeasured under the same conditions, due to irreversible interactions between thesensing material and the poisonous analyte. Further, poisoning can occur abruptlyor progressively depending on the poisonous substance and its concentration. Thepoisoned sensor responds with changed sensitivities and baseline which may lead toa degraded overall system performance.

Drift and related problems lead to periodical re-calibrations of the device that arecostly 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 andwill be explained in more detail in section 1.6.

7


An often underestimated issue concerns the mechanical design of the instrument.To perform dynamic measurements or apply modulation techniques, the instrumentmust be very carefully designed (see sections 1.5.1 and 1.5.2), since the sensor signalmust reflect the reaction characteristics between sensor and analyte, rather than thegas flow dynamics of the set-up and the measurement chamber [59].

1.2.2 The sampling unit and delivery system

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 areplaced in the gas flow, allowing the rapid exchange of gases and hence shorteningthe time of measure.

In the sample flow system, a flow control system is used to deliver the sample to thesensor chamber. In general, the flow system consists of a pneumatic pump or a gaspressure 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 MFCprecisely controls the flow rate independently of its pressure load, therefore this isa 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 betweenthem and the volatiles. Furthermore, the pipes and even the sensor chamber can beheated to avoid the condensation of the volatiles on the walls, which would changethe 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 theflow system by diffusion or be kept inside a tedlar bag for transportation from agiven place. Also for transportation of a gaseous sample, a technique based on thepre-concentration of the volatiles in a sorbent trap containing tenax, hidrocarbon,etc. which absorbs them, can be used. Later, these volatiles are desorbed by heatingthe pre-concentrator, which requires an additional hardware. The advantage of thismethod 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 whichvolatiles are evaporated. In these cases they are usually kept into a closed bottle orvial, 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, whichgives the name to this sampling method. Inside the vial, when the equilibriumbetween the sample phases is reached, the volatiles filling the headspace can beextracted. The time this equilibrium is reached depends on the sample and on thetemperature through the vapour pressure of the volatile compounds, which must bekept under the saturated point to avoid condensation into liquid drops. Therefore,

8

The electronic nose instrument

by keeping the vial under temperature control, temperature induced variations ofvapour pressure are avoided, and furthermore, it is possible either to accelerateemission of volatiles and thus the time of measurement by heating the sample vial,or to avoid the non desired condensation.

To drive the generated volatiles from the headspace to the sensors chamber, a systembased on two needles, which perforate the septum of every vial, is generally usedin 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 vialto carry the volatiles towards the sensors chamber at a constant (controlled) flowrate.

It has to be noted that the volatile concentration at the outlet of the vial graduallychanges its profile by dilution into the carrier gas, and this sometimes influences thewaveform of the sensor response. The volatile pulse supplied to the sensor must besufficiently short so that the concentration profile keeps constant, and sufficientlylarge so that the sensors are exposed to the analyte for a sufficient residence timeto absorb it. Therefore, to avoid the influence of the concentration profile on thesensor response, parameters such as temperature and flow rate must be carefullyselected [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 andreaction dynamics, and all the gas dead volumes in the flow system have to beminimized. Also, the time span between the entrance of the gas sample into theflow circuit and the moment at which the gas reaches the sensor, should be as shortas possible [59].

1.2.3 The sensors chamber

The design of the sensor chamber is an important issue that influences the sensorsresponse, although its importance have often been underestimated. Several studieson the of the airflow patterns through human nasal cavity have been carried outwithout 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 designof this part of the instrument would result in an improvement of its performance,by reducing important sensors drawbacks such as variability among responses ofsensors 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 sensorsand inlet and outlet position [63, 64]. These parameters play an important role inthe 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


its location inside the chamber. Indeed, an incorrect chamber design may lead tohigh sensitivity to slight changes in the flow rates that would directly influence sen-sors response characteristics such as stability, measurement reproducibility, responsetime and amplitude [65].

The design of the chamber must guarantee that the same volatile concentrationarrives to every sensor’s surface by a suitable geometry and sensors arrangementinside the chamber. At least this should not vary with input flow. Its geometry mustassure that there are no stagnant volumes or recirculating zones within the operatingflow rate range, and that the flow is homogeneous, so that the sensors have identicalexposure conditions. Also, a minimal chamber volume allows for rapid refresh rate,minimize any effect due to the sensor location and keeps the volatile concentrationconstant during the response time of the sensors [60]. Moreover, the condensationor 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 toaccurately reproduce the shape of the input concentration signal and obtain thesame concentration profile in repeated measurements. This would therefore resultin an improvement of measurements repeatability and a decrease in the variabilityin the response of sensors of the same type at different positions, since the systemwould be less dependant on flow rates.

Several studies about the sensors chamber characteristics have been made in order tooptimise the sensors response. These studies are based in dynamic flow simulations[63, 66] or comparison of different designs of sensors chambers [66, 67]. In theseworks, several sensors chambers are designed to fulfil the desired characteristics asmuch 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 chamberwith 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 optimumpackaging 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 lightershading represents higher concentration of the analyte. It is interesting to see howthe chamber design produces different analyte concentration arriving to the sensorssurface depending on their position.

Regarding the material, the sensors chamber is usually made of stainless steel butpolytetrafluoroethylene (PTFE) is also recommended. It must be a material chem-ically inert and non-adsorbent to avoid memory effects. Stainless steel has goodthermal conductivity, is tough, inert and non-absorbent but hard to machine. Be-sides, PTFE is reasonably tough, flexible, useful up to a maximum temperatureof 260oC, has poor thermal conductivity and is inert, but it is microporous, so itis absorbent and must be properly desorpted from any odorous materials betweensamples [66].

10

Gas sensors technologies

The temperature control is also recommended, since it allows the chamber to bekept at constant temperature, which is useful to avoid the influence of temperatureon the sensors, condensation of the volatile on the walls and helps sensor desorption.

Figure 1.2: Analyte concentration; a) rectangular chamber with sensors at 90o to flowdirection; b) semicircle chamber with sensors at 90o; c) semicircle chamber with sensors at0o; d) semicircle chamber with sensors at 0o with diffused flow [66].

1.3 Gas sensors technologies

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 thearray is exposed to a given compound or mixture, the response of all the sensorsas a whole forms a characteristic pattern, an electronic fingerprint, from which thegiven mixture or compound can be differentiated among other samples by the signalprocessing unit (fig. 1.3).

The arrays of gas sensors can be formed by sensors of the same technology withdifferent characteristics or by sensors of different technologies. It is known that thecombination of several technologies in a device gives complementary informationabout the sample, which therefore results in better characterization of the volatiles.However, the combination of several types of sensors makes the device be morecomplex and, consequently, more expensive. Furthermore, there is a limit to theuseful size of the array, since by increasing the dimensions of the array the noisemay be amplified, e. g. by sensitivity towards unimportant information, instead ofobtaining new information about the gaseous composition [4].

Chemical sensors mainly consist of a chemically sensitive material and a trans-ducer interface. The different types of sensing material and different transducerprinciples give place to the diversity of chemical sensor types, some of which arebeing described next. Recently, miniaturisation has allowed the fabrication of hand-held and portable micro-spectrometers. These new technologies and instruments,

11


Figure 1.3: Response from an non-dispersive infrared (NDIR) array of 16 sensors as afingerprint for diverse compounds [69].

such as mass spectrometers, ion mobility spectrometers, infrared analysers, fast gaschromatographers or optical sensors, complement the classical chemical sensors andsometimes they are also named e-noses.

The chemical gas sensors technologies used in this thesis, metal oxide (MOX) andconductive 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 suchas acoustic wave devices (quartz crystal microbalances (QCM) and surface acousticwave (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. ), electrochemicalsensors, biosensors, potentiometric sensors, optical based sensors (surface plasmonresonance (SPR), colorimetric, optical fiber (fig. 1.5(b)), etc. ), etc. can be found inliterature in AO systems reviews [4, 59] and books [70, 71], or specific reviews onsensor technologies [72–85].

(a) SAW, (b) QCM,

Figure 1.4: Acoustic wave devices schemes [86].

12


(a) MOSFET, (b) Optical fiber.

Figure 1.5: Other sensors [86].

1.3.1 Metal Oxide Semiconductor sensors (MOX)

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 differenttypes 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 1962by Seiyama [87], and hundreds of works to applications in gas sensing with thesesensors have appeared since then. The sensing mechanism involved in MOx sensorsis 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 byClifford and Tuma [88]. This formula reveals that MOXs sensors response is notlinear with the gas concentration:

R

Ro

= (1 + Kgas · Cgas)−β (1.1)

where R is the sensor resistance, Ro is the sensor resistance in air, Cgas is theconcentration of the target analyte, β is the power law characteristic of the particularsensor and the proportionality constant Kgas depends of the analyte [88, 89].

Normally, MOX sensors are composed of a substrate and a heating element, coatedwith a sensing material between two electrodes (fig. 1.8). Their response is basedon 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. Thesensing material consist of a semiconductor of n-type or p-type. The first group,

13


n-type semiconductors such as SnO2, ZnO, Fe2O3, TiO2 and WO3, respond mainlyto reducing compounds such as H2, CH4, CO, C2H5, or H2S by increasing theirconductivity. On the contrary, p-type semiconductors such as CuO, NiO or CoO,respond mainly to oxidising compounds such as O2, NO2, and Cl2. An advantageof MOX sensors is that the characteristics of these semiconductors can be tuned inseveral ways, such as changing the sensing film thickness, doping the semiconductorwith Pd or Pt metal catalysts and impurities, changing the operational temperatureor via the grain size of the material, in order to modify the sensitivity and selectivityof the final sensor.

The operation principle of the reaction between a n-type semiconductor, such asSnO2 which is the mostly used, and a volatile compound is based on the fact that an-type semiconductor provides an excess of negative electron charge carriers whichare trapped by molecules of O2 from air absorbed on the material surface, causingthe electrical conductivity to decrease. Later, when the sensor is exposed to an at-mosphere containing reducible gases, these react with the absorbed oxygen reducingthe electron trapping effect and thus increasing the conductivity. The mechanism issimilar for p-type semiconductors but of opposite sign [36, 77]. Consequently, thesetype of sensors need the presence of ambient oxygen to operate [88].

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 hightemperatures, greater than 300oC, to increase the reactivity of the semiconductorsurfaces. But it results in high power consumption, which is one of the disadvantagesof these sensors, since it reduces the instrument applicability to portable systems.The change in the sensor sensitivity with temperature can be used to extract moreinformation from a sensor by a thermal cycling technique. This extra informationhave been reported to improve either the sensors selectivity and sensitivity [90].Examples of the sensitivity dependence on temperature for several types of MOXsensors and analytes, can be seen in figures 1.6 and 1.7 [91, 92].

Regarding disadvantages, apart from a high power consumption, MOX sensors sufferfrom poisoning due to irreversible binding of compounds that contain sulphur orweak acids [93, 94] to the sensor oxide [80] and ethanol can also blind the sensorfrom other volatiles [83, 95] increasing their detection threshold. Slow base-linerecovery 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 byincreasing the sensor temperature of operation. Finally, other two drawbacks have tobe added; the high working temperature that makes MOX sensors inappropriate inenvironments containing flammable chemicals and the poor level of sensor to sensorreproducibility [96].

To overcome these problems, several technology approaches have appeared. An

14


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].

example is the miniaturisation of the sensors, which gives place to a a reductionin power consumption. This last feature enables rapid changes in the operatingtemperature, allowing more information to be obtained from a single sensor andfacilitating the use of smaller arrays. A further advantage of miniaturisation is thepossibility of combining an entire sensor array onto a single chip and the obtainingof more reproducible sensors [97].

More information can be obtained from previously mentioned reviews and also [98–100].

15


(a) Scheme [86], (b) Photo.

Figure 1.8: MOX sensor.

1.3.2 Conductive Polymer Chemoresistors (CP)

There are two basic conductive polymer chemorestistors: intrisically conductivepolymer and conductive composite polymer. Their response is based upon a changeon their conductivity as a consequence of exposure to a volatile and the differencebetween these two types is related to the way the conductivity characteristic of thepolymer is achieved.

Intrisically conductive polymer (ICP) consist of a linear backbone comprisedof repeating conjugated organic monomers which are the fundamental structuralunits of the polymer. In their neutral state these materials are insulating, however,they can became electrically conductive by a process of n-doping or p-doping, whichcreate charge carriers and transform its band structure. ICP are similar to semicon-ductors because their intrinsic electrical conductivity decreases as the temperature islowered. The ICP transduction principle is based upon the absorption of the vapoursample into the ICP, which induces a physical swelling of the material that affectsthe charier density on the polymeric chains, and thus a change in their conductivity.On the other hand the conductivity changes may, or may not, be linearly dependenton the concentration of analyte presented to the sensor, although some works havereported linear responses for different concentrations of vapours such as propanol,toluene, acetone, etc. [101]. ICP have been studied to provide changes also in thereactance, aside from resistance, by supplying the sensor with ac current at severalfrequencies. This can be used to increase both their sensitivity and selectivity [84].

On the other hand, Conducting polymer composites consist of conducting par-ticles such as polypyrrole and carbon black interspersed in an insulating polymermatrix [84]. The transduction mechanism for these sensors, especially for polypyr-role composites, is more complex in that the analyte can interact with both theinsulating matrix and the conducting particles, changing also their intrinsic conduc-tivity. On exposure to volatiles, the polymer swell to varying degrees depending onthe polymer-volatile interactions. This volatile-induced expansion of the polymer

16

Pattern recognition system

composite causes an increase in the electrical resistance because the the numberof conducting pathways for charge carriers is reduced. Some works have reportedlinear responses to concentration for various analytes and also good repeatabilityafter several exposures [83].

The use of CP sensors possess several advantages. First of all, a great range ofpolymers is available in the market, and can be easily tailored towards the detectionof 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 reversibleresponses. The reported response times vary considerably from seconds to minutes[83], generally their present faster response and recovery times for polar compoundsand larger responses to non-polar species. Their fabrication method is very flexiblebut has a considerable influence over the surface morphology and other critical pa-rameters that prevents the sensors to be batch-to-batch reproducible [102], althoughminiaturization 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, theyalso present some difficulties in resolving certain types of analytes. This last fact isdue to the relatively low diversity in affinity of the polymers toward a diverse setof analytes [84]. Other problem is the high sensitivity to humidity of the polymerresponse, 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].

1.4 Pattern recognition system

The signal processing unit in the electronic nose analyses the pattern formed by thesensors response in order to extract relevant information. There are many techniquesthat can be used in this stage, coming from a variety of fields that range fromchemometrics, machine learning to pattern recognition, only some used in this thesisare explained.

Pattern recognition follows typically a number of ordered steps shown in figure1.9; signal acquisition, signal preprocessing, dimensionality reduction, predictionand validation, previous to the final solution that can be descriptive, qualitativeor quantitative. Every step is optimised through proper validation and final test.However, in this thesis we propose to include additional blocks with the objectiveof reducing unwanted system variance due to uncontrolled perturbations and drift,which could reduce the prediction ability of the pattern recognition model.

Excellent reviews on signal processing tools used in the field of AO systems can befound in [103, 104].

17


Figure 1.9: Signal processing steps for pattern analysis in a smart sensor array [103].

1.5 Operation Modes

Diverse designs of individual parts of the AO instrument have been briefly describedabove. These different designs give place to systems that vary in cost and instrumentcomplexity. In turn, cost and complexity must be optimised in a compromise withthe desired performance for a given application. According to the variety of designsof possible e-nose instruments, we can distinguish different operation modes of thesesystems.

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 simpleand cheap since they do not need any sensors chamber, sampling unit and deliverysystem and also require simple signal processing (fig. 1.10). They are often usedfor detecting a given gaseous analyte, such as carbon monoxide or methane in anindoor environment.

Other AO systems include a delivery system to drive the target gas towards a sensorschamber at a constant flow. Some instruments even allow automatic dilution of thesample prior to exposure to the sensors, if high concentrations are detected [106]. InAO instruments with delivery system, the sensors respond faster and provide betterperformance 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 ofcleaning cycles provides certain advantages, such as slow down sensors degradation,avoid memory effects and increase information content for data analysis (see section1.5.1). Figure 1.10 shows two example of commercial portable e-noses. GDA2 (gasdetector array) model by Airsense Analytics [106] is a multiple technologies sensorarray 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 usedby 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 aredesigned to work in a network for applications in safety and security, environmentalmonitoring, industrial control and robotics.

(a) GDA 2 by AirSense Ana-lytics [106]

(b) SNet sensor module by JLM Innovation[105]

Figure 1.10: Portable AO instruments.

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 beconnected to vials and a system to automatically drive the volatiles from individualvials (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 latteris the Nordic Sensors Technologies NST-3320, where also a temperature control ofthe vials in the carousel is provided. These instruments are thought to be used inlaboratories because of their size and complexity.

Additionally, more information about the sensor-volatiles reactions can be extractedby modulation techniques, in which system variables such as flow rates, analyteconcentrations or internal parameter of the sensor, such as MOX temperature, arechanged in a controlled way. Among them, the most common techniques are two; acontrolled change (usually a step or a pulse) in the analyte concentration (samplingtransient or concentration modulation [107]) which gives place to the transient re-sponse (see section 1.5.1)), or a change in the temperature in MOX sensors (temper-ature modulation, see section 1.5.2). These operation modes also requires a controlunit for the valves and the MOX sensors heater. Both modulation techniques havebeen extensively used in the literature in order to improve the prediction abilitiesof the system and are briefly described next.

1.5.1 Sampling Transient Analysis

As already mentioned, it is known that the shape of the transient response of everysensor gives much information about the kinetics of the reactions between the target

19


(a) NST-3320 by Nordic Sen-sors Technologies

(b) Esense prototype

Figure 1.11: AO instruments with different sampling units.

volatile and the sensing layer at its surface. These dynamic processes are unique foreach sensor-analyte pair, therefore derived parameters are very useful and indeedlead to improved selectivity, reduced acquisition time, and even increased sensorlifetime [59, 108, 109]. However, traditionally, only one measure point per sensorand per sample is kept. This point corresponds either to the maximum point or tothe mean value among several points around the maximum point of the transientsignal 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 sensorsresponse 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 transientsignal is considerably higher than that from a steady-state signal, since, from theshape of these transient signals f(t) many parameters can be extracted to be usedas 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 sensitivelayer and surface or bulk reaction kinetics of the sensitive material, provided thediffusion processes in the measurement chamber are significantly faster [59].

Apart from the fact that by considering the transient signal f(t) the dimension ofthe feature space is increased and thus, in general, the performance of the followingprediction algorithms is much improved, a number of other advantages have beenreported. 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, theacquisition 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, irreversiblebindings 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, sincethe volatiles in the headspace may be depleted faster than they can be released fromthe 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 toexhibit better repeatability than static descriptors [109].

Computational methods to extract temporal characteristics can be mainly groupedinto (fig. 1.13) parameter-extraction and model-based. In parameter-extractionmethods a number of parameters are extracted from the shape of the transientsignal. These parameters include rise times, maximum/minimum responses, slopes,derivatives (first and second order), phase-space area or integral and curve integralscomputed at different time points during the exposure and recovery phases. Finally,Model-based techniques fit a theoretical model to the experimental transients and usethe model parameters as features [59, 111]. Several studies comparing transient andsteady-state parameters have shown that transient parameters highly outperformsteady-states ones [109, 112–114]. Indeed, discriminatory information is broadlydistributed in the exposure and desorption transients.

Besides, even the entire transient response may be processed with suitable classi-fication or regression models which include multiway techniques [116] (see section2.2.3).

Spite of the many advantages of processing the transient responses, this techniquealso presents some weakness, such as dependence in the sampling procedures andthe head-space generation for different products, in the fluid dynamics of the odourdelivery system and it is also known that the transient signals may also be influencedby previous measurements (memory effect) and by drift [107, 109].

21


Figure 1.13: Parameter-extraction approach [115].

1.5.2 Temperature-Modulation Analysis

It is known that the operating temperature of the MOX sensors influences stronglyits selectivity, since the kinetics of the reaction between the sensing material andthe different volatiles and the stability of the oxygen adsorbed in its surface, are afunction of the temperature. Consequently, much discriminatory and quantitativeinformation can be obtained by the modulation of the operating temperature, whichcan provide gas-specific temporal signatures. This dependence was first modelledby Clifford and Tuma [117] and Sears et al. [118] proposed for the first time the useof thermal modulation to improve the selectivity of a commercial MOX sensor.

There are two temperature-modulation categories for MOx sensors; thermaltransients and temperature modulation [59]. To extract information from thetemperature-modulation response, similar signal processing approaches to those usedin transient analysis can be used; simple parameters extraction of the signal’s wave-form, model based methods or the use of raw responses including multiway methods.Even combinations of parameters from transient and temperature-modulated sen-sors response can be used to build a model.

In thermal transient methods, a thermal transient is induced in the sensor by a fastchange in temperature, created by a step or pulse waveform in the heater voltage.The induced thermochemical transient contains all the discriminatory information.An advantage of the use of this technique is the significant reduction in powerconsumption that can be achieved by intermittently powering up and down thesensor [59].

In contrast, in temperature modulation methods, such temperature modulation isinduced on a sensor by continuous variation of the heater voltage. This variationcan be periodic and, given the dependence of selectivity on the temperature, during

22

Operation Modes

one cycle several peaks in sensitivity may appear. By this technique one sensor cangive place to several pseudo-sensors, since the sensor presents different sensitivitiesacting as several sensors. Therefore, a multivariate dynamic signature of the samplecan be obtained, where not only the magnitude of the conductance but also theshape of the dynamic response can be unique to each analyte.

Some issues regarding temperature modulation have been studied; the waveform ofthe modulation, its frequency, the harmonic content of the induced waveform or thepreferred range of temperatures. All studies conclude that temperature cycling is avery promising approach to improve the device performance [119].

It has been found that temperature-modulated patterns are markedly non-linear.Nakata et al. [120] analyzed these non-linear properties of chemical sensors, andconclude that this characteristic should be viewed as a property to be exploited fordiscrimination purposes. On the other hand, Ortega et al. [119] and Gutierrez-Osunaet al. [121] investigated the effect of the modulation frequency on the informationcontent and the stability of the sensor patterns. They showed that the classificationperformance decreased with increasing frequency, since if the heater waveform is slowenough it allows the sensor to settle at the respective temperatures. They could alsominimize drift effects at low modulation frequencies, where sufficient discriminatoryinformation is preserved in the shape of the response, but not at high frequencies.Finally, studies have also shown that the effects of relative humidity in the samplecan also be reduced and therefore, the repeatability of the responses are improvedover a long period of time [59, 122].

Regarding the best operating temperatures, studies have shown that, in the low-temperature ranges, the sensor response does not carry much information, sincemost reactions occur at the surface level. In general, at high temperatures, bulkreactions are increasingly involved, so the response patterns become more complexand more characteristic of the target gases [59]. However, it has been shown that theoptimum temperature range is dependent on the gas species to be detected [91, 92].

The usual employed waveform for temperature-modulation is sinusoidal, howeverother waveforms have been used and even comparisons among several of them havebeen performed. For instance, Perera et al. [123] used a single sensor modulatedwith a waveform close to sawtooth to find a correlation with the spoilage processof the fish samples. Huang et al. [124] included rectangular, triangular, sawtooth,sinusoidal, and trapezoidal shapes in his work. He showed that each waveform gaverise to a unique sensor-response pattern, which are due to characteristic changesin the actual surface temperatures of the sensor. The optimum selection of thetemperature-modulation profiles has also been studied, Kunt et al. [125], Vergara etal. [126] and Gosangi and Gutierrez-Osuna [127] have proposed some optimizationmethods. A review on this issue can be found in [90].

23


1.6 Drift and sensor faults in a gas sensors array

As a result of drift, models built with learned sensor response patterns may becomeobsolete 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 consideredto be one of the most serious limitations to the extensive use of the electronicnoses. Moreover, individual sensors in the array may fail and this introduce errorsin measurements that may give misclassification results.

The effects of drift on the sensor response, specifically from an array of sensors in ane-nose system, and the advances to compensate it by signal processing, are discussedin this section along with the subject of the presence of a failure in a sensor in thearray, 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 befollowed. For instance, the prediction capabilities of the classification/quantificationmodel 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 sincethe model complexity is increased. For classification applications, another measureof the effectiveness of the method is the relative change of the Mahalanobis distancebetween the different clusters, before and after drift reduction [128], or the evalu-ation of the Fisher ratio, again before and after drift correction. In quantificationapplications, 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 modelvalidation. If samples from the whole time period are taken for the model buildingand validation is based on techniques that do not consider the temporal ordering ofthe 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 evolutionof 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 toposterior samples for testing. This strategy must be followed to validate both, thecalibration model and the drift correction method.

1.6.1 Characteristics of drift and sensor poisoning

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

such as diverse techniques in spectroscopy or chromatography.

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 otherfields of data processing. This problem is generally referred to as population drift[134]. Some examples arise in medical applications, where there may be changesin the patient characteristics thus the patient population may drift, radar targetrecognition or object recognition in images, where the object changes its positionand ambient light conditions may change [135], email spam filtering systems whichneeds to be adapted because the types of spam changes over time, etc.

Drift is a common problem for all chemical sensors, although different causes fordrift dominate for different sensor types. In the field of chemical gas sensors, drift isconsidered strictly to be due to two main causes related to the sensor itself; ageingand 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 ofmetal particles, and changes in the number of reaction sites. These re-organizationmay happen spontaneously, but it may be speeded up if the sensor operates inreactive environments or at high temperatures [136].

Besides, poisoning consist of the irreversible chemical or physical interactions be-tween the molecules in the gas and the sensor surface or bulk material, by blockingor creating reaction sites on the sensor material (see section 1.2.1). As a result of thechange of the number of reaction sites, like in sensor ageing, the sensor sensitivityis also changed.

Much effort is being made to find sensor materials which interact reversibly withthe gas, such that the molecules that have reacted on the sensor will leave it assoon as the gas leaves the sensor surface. But this is a difficult task, instead andas mentioned previously, different operation modes of the sensors are used, suchas short time exposition to the gas analyte and study of the sensors transient ortemperature modulation of the sensors.

1.6.2 Factors involved in measurements instability

Some variables that influence the sensor measurements over time that are relatedto other sources than ageing and poisoning, such as environmental or measurementsystem, can be considered and treated as drift from the point of view of signalprocessing. Environmental sources include changes in temperature, pressure, orhumidity. Besides, a bad design of the measurement system may induce temperaturevariations in the measured head space or on the sensors, humidity variations in thesample or other physical and chemical processes, also fluctuations in flow rate oreven analyte condensation in the manifold, for example. Avoiding fluctuations ofthese important parameters is essential, this can be addressed by careful control of

25


sample and sensor parameters with a proper system design (see sections 1.2.2 and1.2.3) optimised for each application, or by measuring the variables that are knownto 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 effectswhich, 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 arepresent when the response of the sensor depends on the volatiles it has recently beenexposed to, giving an additional effect to the sensor response. This phenomenon isdue to the fact that remnants of previous gases may still be present either in thesampling system, or on the sensor surface itself. Memory effects can be consideredas an additional source of variance due to uncontrolled previous exposures. The bestway 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 suchas short pulses of clean air and/or high temperature annealing between the sampleswhen 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 orthermal trends, noise or even the degradation of the samples themselves. Also, theobserved sometimes called short-term drift or warm-up effect, by which the sensorsresponse 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 stableresponse after start-up of a measurement series, and thus its nature is different fromordinary drift [59, 136].

1.6.3 Effects of drift and poisoning on the sensors response

To reduce the effects of drift in the electronic nose by signal processing, differentmathematical models have been used to compensate for the changes in the sensorbehaviour, and therefore to maintain the gas identification capability. These meth-ods have been applied for different situations and thus they have made differentassumptions 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, thusthe 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 andmagnitude along the time, therefore the data series may present gradual (sometimesirregular) decreases or increases or even jumps, (usually of small size), also noisecan be superimposed on all the sensor signals and he speed of the response can bealso altered.

26


(a) Additive drift,

(b) Multiplicative drift,

(c) Drift corrected sequence.

Figure 1.14: Ilustration of additive and multiplicative drifts on a sensor transient responsein a sequence of measurements [137].

Furthermore, the characteristic pattern of a given target gas may not be preservedover time. The pattern considers relative relationship between different sensorsrather than their absolute output in a sample. If these relative relationships stayconstant over time, a simple normalization could correct the drift effect. However,drift may also modify the characteristic patterns of the analytes, since it influencesthe 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 allsimilar gas mixtures (fig. 1.15). This direction usually presents a big componentalong a PCA model space and a smaller one in the residual space, thus violatingHotelling’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 orseveral individual sensors, it has to be taken into account that drift and a processfault would be registered by a band of correlated sensors [138]. Therefore, a sensorfault would move the samples away from the PCA model space, surpassing mainlythe 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 tendto drift in a similar way [136]. However, different gas mixtures may drift in differentdirections. This has to be taken into account when choosing a reference gas forcompensating drift, thus this gas has to be very similar to the test gases in theapplication. In a situation where different sensor types are used, drift should bedescribed for each sensor type.

27


(a) drifting data, (b) drift corrected data,

Figure 1.15: PCA scores plot for data from 39 sensors, measuring on 9 different gasmixtures [128]. The direction of drift is similar for all mixtures.

1.6.4 Evidence of drift and faulty sensors in electronic nose applications

The effect of drift in array of gas sensors have been observed since the beginning ofthe 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 inthe 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 graphicallythe effect of drift on two experiments performed with a commercial device hybridgas-sensor array (NST 3220 by Nordic Sensor Technologies), which contained twoblocks of five MOS-FET sensors set at different temperatures and one block withfive Taguchi MOX sensors coupled in series. Indeed, they observed short-term driftwithin a measurement series lasting several hours, and distinguished it from longterm drift during several measurement series. The first measurements were obtainedfrom a fish storage experiment that lasted 5 days and the second one was performedon 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 pasteurisedmilk. Many other experiments related to food quality have been carried out wheredrift 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 measurementsof extra-virgin olive oils in 14 different geographical provenances, they observed alarge mixture of the different classes which they attributed to drift. They used sevencommercial sensors from Alpha M. O. S. and five from Figaro, during the time ofperforming 242 measurements. By testing several approaches they could reduce themisclassification 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 wereanalysed using a MOx sensor based electronic nose, to discriminate between theirflavours. A PCA scores plot was plotted to show the effect of drift on the samples.

28


Schaller and co-workers [142, 143] have reported large background drift of CP andMOS sensors, and also MOS sensor poisoning when attempting to analyse cheesesamples of Emmental type. Concita et al. [13] also observed drift on one of thesix thin film tin oxide sensors while detecting microbial contamination in processedtomatoes. The drifting sensor was finally discarded for that application. In otherexperiments, drift was taken into account and could be compensated by a referencegas between measurements; measurements of samples of milk during 20 days with acommercial 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 cultivarduring about 20 days [146], etc.

Kuske et al. [147] carried out work, the objective of which was to detect the presenceof moulds growing in building materials by using an e-nose with metal oxide sensorsand two different classifiers. In the experiment, two datasets were collected, thesecond one five months after the first one with a time duration of 3 months. Itwas found that drift had an important effect of both classifiers, which results werereduced from 4 up to 9 percent points in the final classification rate with respect tothe one obtained with data from first data subset. The authors also present a PCAscore plot where drift can be perceived visually, since both data sets show the samecluster 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 Tokyoover 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 necessaryone re-calibration every 10 days. Also, Nake et al. [149] measured volatile emissionsfrom wastewater plants with two commercial devices based on CP and MOx gassensors. They sampled air at different locations of the wastewater plant on differentdays. However, the conductive polymer signals formed groups correlated to the dayof measurement, thus showing a substantial influence of ambient humidity (whichcan be also considered as drift). Romain et al. [150, 151] have also reported strongdrift 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 carriedout by the same authors, included measurements from a compost plant [152] wheredrift was also present. Perera et al. [24] observed drift in the monitoring of the stateof 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 drifteffect. And also De Vito et al. [153, 154] had to deal with drift in an experimentduring 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 sensorarray device. They observed large variation of the sensors response due to seasonalmeteorological effects and drift, which is shown graphically by plotting for exampleabsolute and relative errors on benzene concentration versus time in weeks [154].

29


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 arraymay fail due to a variety of reasons, such as instrumental, or just the sensing materialdegradation (gradual or abrupt) produced by time or poisoning. Apart from driftduring their long experiment in monitoring odours on a landfill site, Romain etal. [150] also reported the need to replace some sensors. In experiments where sensorswere exposed to certain substances, such as sulphur compounds (e. g. H2S 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, failingsensors must be detected and identified in the array, in order to be intermediatelyreplaced. However, only few works in literature can be found which intend to detectand correct a fault in an array of gas sensors. In an experiment carried out byPardo 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 duringcontinuous operation, the outputs of the sensors were estimated from the signals andcompared to the real values. Finally, Tomic et al. [157] studied calibration transfer

30


techniques for their application in sensor replacement problems, (see section 2.2.4).

1.6.5 Drift correction and fault detection by signal processing techniques

In this section, a brief review of some techniques used in AO systems for driftcounteraction methods is made.

Robust features

A good starting point in the design of a technique for prediction which could workunder drift conditions, could be the identification of features that are (approxima-tively) stable over time. As it has already been commented, the modulation of thesensor operation temperature has been reported to generate features that are sig-nificantly more robust against drift than isothermal features. However, this is onlypossible in the case of MOx sensors (section 1.3), other extracted features such astransient responses have shown to be drift sensitive (see paper 4.1), although theyprovide more information for prediction than isothermal features.

Other approaches to find robust features used the pattern of responses rather thanindividual sensor responses. For example, robust features were found by Wilson etal. [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 sensorsat eight different temperatures. To binarise the sensors signals, the sensors responsewere arranged in ascending order and the output from each sensor was set to zeroif it was smaller than the median output, and set to one if it was larger than themedian. Therefore, the resulting output from the threshold function was a patternof zeros and ones and was more robust than the absolute sensor values since it isrelative and not as sensitive to noise. This technique removes much information,specially the one related to concentration levels, and therefore it is useful only forclassification 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, loserand its two nearest neighbours was calculated. Since Bednarczyk used ten tin oxidesensors, 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 stepswere involved; a first one using the winner and loser sensors which identify the familyof chemicals the sample belongs to, and a finer one with the calculated slopes whichfurther discriminates the chemical from others within the same family. However, the

31


authors mention that the system does not discriminate sufficiently among chemicalswithin a large range of concentrations, they must be further processed to eliminatethis 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 thecharacteristic patterns of the target analytes can also be altered (section 1.6.3).

Reference Gas Methods

Reference gas based methods are used to estimate drift. Usually, measurementson one or several reference gases are made at the beginning of the measurementseries and with some intervals as long as the sensors are used. The change in thesensor responses to the reference gas is taken as a measure of the change in thesensors response for all other measurements (fig. 1.17). Furthermore, the use ofa reference gas also allows to distinguish between sensor drift and changes in thesamples themselves.

In the use of these methods, it is assumed that there is a strong correlation amongthe 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 thosebeing measured, also close to the real samples in sensor response space, since thedrift in the reference gas should reflect the drift for all other samples. It must alsobe stable over time, without degradation or varying concentration, available andeasy to measure. Authors have followed different approaches to choose a suitablereference gas; some of them have used the same reference gas (often water) for allapplications, while others choose the reference gas depending on the application. Inother cases, replicates of the samples can be used as pseudo-reference gases, as longas 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 measurementbefore the target analyte measurement [137].

32


rected the additive influence of drift (xcorri (t) at every instant t and for every sensor i)by subtracting the response of the reference gas xiRefGas(t) from the sample responsesxi(t) sensor by sensor. Their results were very satisfactory along an experiment dur-ing 45 days. However, the authors highlight the fact that the experiment has beencarried out in laboratory conditions, without pollutants added to the target analytesand therefore results in less ideal situations may not be as good.

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, thetemporal 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 theadditive 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 referencegas xRefGasi (t)/xRefGasi (0) for each sensor i, and used it to compensate the responsesof the samples. Then, they fitted a linear trend to this ratio (at + b) to find thecorrection factor for the samples located between the reference gas measurements(eq. 1.2):

xcorri (t) = xi(t) · (at+ b)i (1.2)

This method is relatively simple and provided very good results. However, althoughthe technique was tested on complex mixtures, volatiles from food samples, the ex-periments did not last more than 10 days. Further, many calibration measurementshave to be made and this is not possible in many situations.

Finally, Arthursson et al. [128] devised a technique based on PCA and PLS, calledcomponent correction (CC). In this technique, a reference gas is used to estimatethe main direction of drift, done by computing the first principal component of itssamples. Then, this or these component/s, pref , are removed from the multivariatearray response x by:

xcorrected = x− (x · pref )pTref (1.3)

The vector pref contain the contribution of the original variables to the identifieddirection of drift, thus pref can be used also to see the most drifting parameters. Onthe other hand, the removal of this direction preserves all the other directions andkeeps the important variances that separate different clusters and concentrations,unless, of course, the information is found in the same direction as the drift. Resultswere satisfactory (see fig. 1.15) but this method depends strongly on the referencegas. If the analytes to be measure are similar, drift of all of them would show similartrends but this is not necessary true if the analytes are different (section 1.6.3).

33


(a) Before calibration,

(b) After short-term correction (within sequence),

(c) After long-term correction (between sequences).

Figure 1.18: Illustration of the multiplicative signal correction method of Haugen etal. [140]. Response of an individual sensor to the calibrant and target gases [59].

Non-linear and System-identification techniques

Examples of nonlinear models that follow an adaptive strategy are the use of a singleKohonen self-organizing map (SOM) for all the classes, [161, 162] or a separate SOMper 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 odoursusing 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 neighbourhoodconstraints 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 thesemaps 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 classesfor a classification task, that enable a self-adjusting process for each neuron in thelocal maps. Zuppa et al. [163] showed that mSOM provided small error rates inthe classification of six compounds in an experiment that lasted 4 weeks. However,for long time experiments, techniques that track the evolution of data with severalclasses need that all the classes are alternatively measured, otherwise it may losethe tracking of some of the classes.

System-identification techniques model sensor dynamics to correct drift effects, thus

34


Figure 1.19: Illustration of drift correction by an adaptive clustering method, which tracksthe evolution of the data due to the effect of drift [137].

assuming that sensors dynamics is more stable than the sensors response. Modelssuch as ARX, ARMA and Box-Jenkins were used by Holmberg et al. [164, 165]. Theyconcluded that Box-Jenkins was the more flexible to model the data because it doesnot contain an auto-regressive part (unlike ARX model), and also includes a noisemodel. They also found that the model was sensible to measurement pauses duringthe experiment, in which five analytes were classified along 45 days. In anotherwork, Holmberg et al. [165] used a recursive least squares algorithm to update themodel. This method proved to be suitable even for the periods of measurementpauses, thanks the ability of the updated model to adapt itself to the changes.Later, for classification, a separate dynamical model was built for each analyteclass, and unknown analytes were assigned to the class whose model displays thelowest prediction error. Finally, Perera et al. [24] have developed a fault detectionmethod based on recursive dynamic PCA [166, 167] that can operate under driftconditions. The proposed technique is adaptive, therefore it is very useful for thetarget application, in which a continuous monitoring is made to detect oil vapourleakages in air compressors. Results are very good, but in this work, the componentof drift is simulated and linear.

Other appraches [168, 169] have used independent component analysis (ICA) to findstatistically independent components of the sensor array response, which could beassociated to drift and discriminatory information. For instance, Kermit and Tomic[169] showed that ICA could separate three sources of information in the sensorresponse: analyte identity, analyte concentration, and drift effects. In contrast, PCAwas only able to separate concentration information (first principal component), butanalyte identity and drift kept mixed together in the second and third principalcomponents. However, in this case ICA was applied in an off-line way, withouttesting new incoming samples. In the experiment by Di Natale et al. [168], theyused the validation method of leave one out, which is not suitable for validation indrift 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


values of the sensor’s resistance, temperature and humidity as inputs and resulted ina very accurate estimation of the analyte concentration [170]. In long experimentsother 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 informationof 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 betested on real drift conditions.

Component and Orthogonal Signal Correction

Approaches based on component correction first identifies, from the sensor-arrayresponse, the components that captures as much of the variance as possible butuncorrelated with analyte information, concentrations in regression analysis or classlabels in discrimination problems. If the directions of drift and analyte informationare uncorrelated, the removal of the drift component would lead to very good results(fig. 1.20). Artursson et al. [128] follows this idea, as explained above. In the samework, he also proposed a method that selects the time as a dependant variable of apartial least square regression (PLS).

In a proposal by Gutierrez-Osuna [173], he first identified a set of variables y whosevariance can be attributed primarily to drift or interferents. These variables caninclude, e. g. the response to a purging or reference gas, the date and time when thesample was collected, or measurements from temperature, pressure, and humiditysensors. He then applied a Canonical Correlation Analysis (CCA) or Partial LeastSquares (PLS) to find directions in data x and information y, x = Ax and y = By,where x and y co-vary. And finally, Ordinary Least Squares (OLS) was used tofind a regression model xpred = Wx, that is subtracted from x to give the correctedmeasurement vector xcorrected = x − xpred. This method is also interesting, butpresent the drawback of the already commented techniques based on reference gases.

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 variancecommon 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 inimproving the time stability of the system. Results are presented in a figure showingthe evolution of the classification rates of the species of analytes along the time ofmeasurements. However, they use 1000 samples in the smallest training set, whichis not feasible in many situations.

36


Figure 1.20: Illustration of an ideal situation where the directions of drift and analyteinformation are orthogonal in a PCA model space [59].

Other methodologies to deal with drift

Brahim-Belhouari et al. [175] investigated on synthetic data, a way to obtain afast and robust system for combustible gas identification. The system consistedof a micromachined SnO2 integrated sensor array and a Gaussian Mixture Models(GMM) algorithm for gas classification. To deal with unavoidable drift, the authorssuggested the extension of the training set with simulated drift, which, in theiropinion, would result in more robust features and thus in an increase of the globalrobustness of the classifier. The simulated drift was linear; vd(t) = v(1 + at),where a was a random number varying from 0 to 30%. This strategy improvedthe classification performance. However, the authors also suggested that this driftcorrection approach should be assessed by its testing against real sensors data, inorder to check its real performance. In fact, the assumption of a linear behaviour ofthe drift can also be a matter of study.

Sisk et al. [176] studied the optimal calibration protocol. A condition-based cali-bration, in which calibration was only performed when the classification performancedecreased below a predetermined threshold value, was observed to be superior to atime-based calibration or to interval-based calibration protocols.

Comparison of different drift correction methods

Romain and Nicolas [150] compare three simple drift counteraction techniques fortwo arrays of six metal-oxide Figaro sensors continuously measuring during 7 years areal-life atmosphere, which was the odour generated by an urban waste compostingfacility. In that experiment, the temperature and humidity of the sensors cham-bers were controlled and a calibration gas was available. The tested solutions tocompensate the drift effect were: signal pre-processing, univariate sensor correction

37


[140] and multivariate array correction [128]. In the first method, they preprocessedevery sensor signal by subtracting the sensor response to pure reference air. Thissolution requires cycling operation between reference air and target gas. However, itis not convenient for on-site applications since heavy gas cylinders should be carriedin the field or, as an alternative, the reference air could be generated by filteringthe ambient atmosphere, but this produce a reference without much purity whichwould increase the data dispersion. Univariate sensor correction was performed byusing a multiplicative correction factor obtained from measurements of a referencegas at regular intervals within the target measurement, and finally, the Arthursson’sCC method was performed, which also required a calibration gas. In that work theyfound that the univariate approach was the most adequate for their problem, sincethey 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 anexperiment during four months. In that experiment, eight different analytes at lowconcentrations where exposed to the sensors. The most difficult separation tasks,e. g. the discrimination between hydrocarbons, were little improved by modellingthe dynamics of sensor drift, either through a linear regression or Fourier transformdecomposition of the individual responses of each sensor. Finally, a simple linearsensor-by-sensor calibration scheme proved effective for classification performanceof binary separation tasks.

Zuppa et al. [163] tested the performance of their proposed mSOM (with severalinternal parameters), mSOM with different filter types (Saviztky-Golay, DiscreteWavelet 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 foundthat their proposed mSOM with a filter was superior to the other tested techniquesand derived some conclusions about the internal parameters of the optimal mSOMtechnique.

Summary

In the field of gas sensor systems, drift correction is a very important task and shouldbe included into the general scheme of pattern recognition. Several and differentsignal 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 thespecific application; adaptive methods, where the prediction model is updated withrecently classified incoming samples, methods based on some measurements of areference gas distributed along the time, or methods which model a given trend inthe training subset that is extracted from new samples.

The adaptive methods strategy is useful in the case of an instrument working con-

38


tinuously and where target samples are known not to vary too much, as in processmonitoring 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 fromdrifting 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 giventrend, 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 thusto compensate for. All strategies are totally useless in cases where abrupt changesoccur, except when measures of a reference gas are made in that period. In thatcases only new re-calibrations can keep the instrument in use.

Anyway, no technique is valid for all situations and more research is needed to findmore methods that could be adjusted to specific situations.

39


40

Chapter 2

Robustness

In this chapter, an overview on some signal processing techniques that can be usedto 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 affectingarray of gas sensors; drift and failure of individual sensors in an array, is made insection 2.1.

As explained in previous chapter, electronic noses or artificial olfaction systems (AO)are instruments with very interesting capabilities for fast analysis of volatiles. How-ever, they present some shortcomings that have to be solved in order to become areliable instrument for industry. Different problems come from incorrect measure-ments procedure, sampling methods or system design. Some of the sources of theseproblems have been already revised in the previous chapter. Others come from thearea of sensor technology, especially sensors ageing and poisoning, which degradesthe sensors performance along the time and which, along with sensitivity to otherunwanted variables such as environmental, produce sensor drifting. Furthermore,since the core of an electronic nose is an array of gas sensors, an additional problemthat has to be considered is the possible failure of one or more sensors in an array.

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 tofulfil certain requirements, within which robustness is one of the most importantones. Every used instrumentation technology in industry is therefore rigorouslyvalidated through a proper study of the instrument and methodology robustness.A robustness test consists on a deep study of the behaviour of a given instrumentto variations in problematic variables, specific of every type of instrument, takinginto account not only the system but also the calibration method and algorithmsinvolved 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, theconcepts and methods of which are being continuously developing.

On the other hand, AO systems are relatively young, and general standards andconcepts for the study of their robustness have not been established yet. However,some of the concepts derived in other instrumentation areas, such as the varietyof methods in chromatography, spectroscopy, etc. could be adapted and transferredto 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 presentedin the following sections (sections 2.1.1 and 2.1.2). In particular, the robustness ofcalibration models is an important task for signal processing, thus this subject isdiscussed (section 2.1.3).

In the specific study of the counteraction of drift in AO systems by signal processingtechniques, different point of views can be adopted. First, it has been found thatdata affected by drift exhibit a high degree of dispersion in a feature space, whichis mainly oriented in few directions. Signal processing techniques that can reducethis dispersion by signal compensation or even removal of these directions would bevery useful. A group of techniques called orthogonal projection (OP) methods havebeen specially designed to carry out this task, and are already used in spectroscopy,chromatography and other instrumentation areas. These techniques consist on theremoval of data variability not correlated to a given information of interest. Theyare briefly explained in section 2.2.1.

Besides, an interesting way of dealing with instruments presenting strong drift, isto use methods that extract information from measurements but without the needof a previous calibration step, as non-supervised methods. Consequently, thesemethods are not influenced by drift, since they do not need calibration references.Such techniques are known in signal processing as blind source separation (BSS)techniques, though in the field of analytical chemistry they are called multivariatecurve resolution methods (MCR). In MCR methods, a series of algorithms attemptto decompose a data matrix into components with chemical meaning. They are alsoshortly presented in section 2.2.2.

Further, as drift causes measurements performed at the calibration phase and aftera certain period of time to differ considerably, this situation can be seen as havingtwo instruments of the same kind; the first one corresponds to the calibration phaseand the second one is considered to be out of calibration. Hence, in order to avoidto perform a full recalibration on the second device, techniques based on calibrationtransfer methods can be applied to make this second device work as expected. Theaim 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 expensivetask. 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 leadto 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 replacedwhen 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 sensorfault detection and isolation (FDI) and fault diagnosis, which involve faulty sensordetection, identification and correction of its response, have been largely developedfor statistical process control. Among the techniques used for such issue in systemcontrol, statistical methods are very interesting since they do not need to make amodel of the behaviour of the instrument, on the contrary, the model is made on itsmeasurement data. This topic is discussed in section 2.2.5.

2.1 Robustness

Industry is increasingly requiring for instruments able to give fast measurement withsuitable precision, accuracy and long-term stability under variable (industrial) condi-tions, which include variations of parameters, instruments, addition of uncontrolledfactors not studied during calibration, etc. At the same time, industry is subjectedto strict regulatory requirements (specially in pharmaceutical industry) and costs.This means that every used measurement method is required to be strictly validatedthrough a proper study of the robustness of the instrument and methodology. As aresult, the measurement of system robustness and model transferability have becometwo key points in the development of new measurement technologies.

2.1.1 Definition of Robustness and Ruggedness

Generally, an instrumentation technique consist on a measurement instrument com-bined with a (multivariate) calibration model. A validation of such technique mustbe made through a robustness study of all aspects of the measurement; sampling, in-strumental (mechanical and sensors), environmental characteristics and calibrationmodels.

However, there is not an official definition for the word robustness valid for all typeof measurements, although it has always been considered an important property ofa measurement method. Many similar definitions can be found in literature, mostof them related to specific problems in specific applications, and thus sometimes dif-ferent definitions can be found within one domain. Nevertheless, some internationalorganizations, such as International Organization for Standardization (ISO) [178] inthe 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 measurementmethod.

43


For instance, USP and ICH definition of robustness are equal: the robustness ofan analytical procedure is a measure of its capacity to remain unaffected by small,but deliberate variations in method parameters and provides an indication of itsreliability during normal usage. But there exist another term used for robustness orrelated properties, ruggedness. Also for this term, several definitions can be foundin literature, but also many times it is considered as a synonym of robustness. Anexample is ICH, which considers ruggedness and robustness as synonyms, while USPdefines it as the degree of reproducibility of test results obtained by the analysis of thesame 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 withtransferability problems, while robustness is computed from intra-laboratory trials(different temperature, concentration, etc. ) [182].

2.1.2 Robustness and Ruggedness tests

Robustness tests are widely applied to study the potential sources of variability inmethod results. The ICH guideline defines robustness testing as: an experimen-tal study in which one evaluates the influence of small changes in the operating orenvironmental conditions on measured or calculated responses. And state that oneconsequence 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 validityof the analytical procedure is maintained whenever used [181]. The methodology ofrobustness testing in chemometrics and statistics is in agreement with these ICHgidelines [182].

Indeed, another definition of a robustness test can be found in literature. It isgiven by the USP, and it is more associated with the term ruggedness. It describesthe ruggedness test as the evaluation of the degree of reproducibility of test resultsunder normal, expected operational conditions from laboratory to laboratory and fromanalyst to analyst [182].

In summary, robustness tests identify the parameters that can be varied withoutaffecting performance of an analytical method, providing reasonable operating tol-erances that may be expected during routine use. Moreover, robustness tests alsodetermine which parameters need to be tightly controlled, and include these criticalvariables in the procedural description of the analytical method. Hence, robustnesstests demonstrate that an analytical method is reliable during normal usage. Be-sides, ruggedness tests measure the reproducibility of results taking into accountnormal operating variations, and ensures that an analytical method will performsuitably each time it is used.

Therefore, as the USP/ISP definitions of ruggedness and robustness suggest, vali-dation tests involve a given methodology. On the one hand, according to the USP

44

Robustness

definition of ruggedness, the validation test of a method is made through severalrepetitions of the method being studied under different test conditions to examinethe effects of some non-procedure-related factors, such as laboratories, instrumentsor analysts, expected to run with normal operating conditions without changingthe procedure-related method parameters [181]. On the other hand, according tothe USP/ICH definition of robustness, to evaluate the influences of small deliberatechanges in procedure-related method parameters in a robustness test, two approachescan be followed; either an one-variable-at-a-time procedure (OVAT) or an experi-mental design procedure can be applied. In the OVAT procedure, to evaluate theeffect of a given factor on the method response, the levels of the former factor arevaried while keeping the other factors at nominal levels. In contrast, in an ex-perimental design procedure, several levels combinations of different parameters areconsidered. In fact, this is the recommended approach, since it represents more glob-ally what is happening around the nominal situation and requires a smaller numberof experiments that a study of every factor in a OVAT way. However, due to designrestrictions, the experimental design approach is mostly performed by applying alimited two-level screening designs, i. e. two laboratories, instruments, analysts, etc[181].

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 examplesof typical parameters that may need to be studied are; the sampling method (samplepreparation, sampling time), influence of temperature, flow rates and humidity asthese parameters change the sensors responses. The range used to evaluate theseparameters should reflect the potential operating conditions. An interesting robust-ness test performed on a electronic nose system is the study of the performanceand stability of the instrument when the operating conditions change. This is thecase of an array of metal oxide gas sensors when the sensors are heated by shortpulses instead of continuously. There is an considerable reduction in the power con-sumption of the instrument when using pulsed heating sensors in comparison withisothermal sensors, however the performance of the system may be degraded, as wellas the metrological characteristics of individual sensors, due to thermal effects thatdegrades the sensors when repetitively switching temperature [183].

On the other hand, it is important to note that a factor to be necessarily included ina robustness test is the time stability. In spectroscopy literature, statistical methodsapplied for improving the robustness of a calibration model (also used in calibrationtransfer problems) are also used to deal with problems of drift. These two subjectsare briefly discussed next.

2.1.3 Robustness of calibration models

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


community of artificial olfaction; the robustness of a multivariate calibration modelis the stability of its predictive capacity against perturbations centred on standardconditions.

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 theremight be violations in the assumptions upon which these procedures are based [184].The term robustness in robust statistics establish the resistance of estimators to theoccurrence of outliers [185]. Also, the model robustness refers to their insensitivity todeviations related to the non-conformity with statistical hypothesis (linearity, nor-mality, independence of the variables). In multivariate calibration, such a measureof robustness is equivalent to the maximal percentage of outliers that a calibrationmethod can handle [182, 186, 187].

In multivariate calibration, the used indices used for robustness assessment are basedon minimizing the objective function of prediction error [182]. Indeed, a calibrationmodel is mostly computed using the cross validation technique, so calibration modelswith less error of prediction are normally chosen and less attention is paid to theirrobustness. Later, models are commonly tested by the root mean square error ofprediction (RMSEP) index. Sometimes, there are also included other indices as ameasurement of robustness, such as the calibration model dependence on trainingsamples, on the training set size or results from permutation tests, all based onresulting prediction error.

However, in a study on the sensitivity of multivariate calibration models, the authors[188] consider that the robustness of models cannot be judged only in terms of theirprediction error; indeed, models may possess a small prediction error and at the sametime be very sensitive to small perturbations in experimental conditions. Therefore,the sensitivity of calibration models must be estimated through a robustness study,in which some robustness-indices would be calculated and in which it would befound the range, centred on the standard conditions, in which the model give similarresponses to those produced under standard conditions [182].

On the other hand, there exist some ways of improving the robustness of calibrationmodels. Several statistical techniques that are often used to enhance prediction-model performance, can also be used for this purpose. Zeaiter et al. [189] presentedan overview of these techniques in the field of near infra-red (NIR) spectroscopy. Inthat work, she considered multivariate calibration methods to be linear functionsrelating 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. Shethen stated that the robustness problem is due to variations in the measurementconditions caused by variations in influence factors that affect the measurementby adding a perturbation δx. This perturbation is represented in the prediction

46

Signal processing methods for improving robustness

responses of equation 2.1 as an error δy such that:

δy = δxT · b (2.2)

which gives:

|δy| = ‖δxT‖ · ‖b‖ · |cos(δx, b)| (2.3)

and therefore he conclude that, to minimize δy, one or more of the three terms in theright 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, Smoothingand Differentiation), which correct for baseline shift, curvi-linearity and noise-relatedterms, particularly reduce the ‖δx‖ term, smoothing reduces the second term ‖b‖giving an advantage over other geometric preprocessing methods but with a carefulselection of the width of the smoothing window. Also, ridge regression and regular-ization methods in general, decrease the term ‖b‖. Finally, Orthogonal projectionmethods (OSC, DOSC, OPLS, IIR, and EPO) minimized the third term, |cos(δx, b)|,representing the angle between the systematic variations and the model. Thesemethods seek to reduce |cos(δx, b)| to zero, to ensure the relative independence ofthe model from variations in the x data. Variable-selection methods (Wrappers, Fil-ters, and Embedded) affect the robustness via ‖δx‖ and |cos(δx, b)|. The term ‖δx‖could increase if the selected wavelengths are useless (or no information is givenby some gas sensors in the array). The term |cos(δx, b)| depends on the initiallyselected wavelengths (gas sensors in our case); if the number of useless wavelengths(sensors) increase, |cos(δx, b)| is increased. Hence, variable selection should be usedcarefully in seeking to enhance the robustness of the calibration model.

2.2 Signal processing methods for improving robustness

This section revises four areas of knowledge in signal processing, highly applied inother technologies, in which some developed algorithms can be very useful for theimprovement of the robustness of AO systems.

2.2.1 Orthogonal Projection (OP) methods

In the last decade, there has been a great effort to improve robustness in the areasof chromatography and spectroscopy. In these fields, several signal processing tech-niques have emerged to minimize these problems and improve the robustness of the

47


calibration methods (see section 2.1). Specifically, the family of orthogonal projec-tion (OP) techniques which include net analyte signal (NAS), are commonly usedsince 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], orthogonalprojection to latent structures (OPLS [195] and O2PLS [196]), OPLS discriminantanalysis (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 analytesignal (NAS) (also called net analyte projection (NAP)) [205] and derived algo-rithms [206–208] and independent interference reduction (IIR) [209]. In the field ofAO, the component correction method (CC) [128] can be considered as belongingto the OP family.

The main action of OP methods is a reduction of the data dimensionality, but ina different way that transformation methods or variable selection methods do. OPtechniques better define the dimensions of the subspace that describe the maximumvariance related or unrelated to an information of interest in the multivariate space,making the posterior prediction model independent of the influence of these nondesired variations in the data, and thus improving data and model interpretation.They also may show an improvement of the model predictive performance, evenoutside the calibration range, and in the presence of different factors that mayinfluence the signals.

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 Xdata structure to be related to Y , these approaches can be based on partial leastsquares (PLS).Several works comparing some of these techniques can be found inliterature [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 mainlycontaining effects due to structured systematic variations, and E containing effectsdue to variations of random noise (eq. 2.4). Thus, the data X, which constitute acloud 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 relatedto G, X− is the useless part of X related to N . Mathematically, this is representedby equation 2.4.

D = G+N + E with X = X+ +X− + E (2.4)

48


Orthogonal projection methods identify an orthonormal basis P− capturing themaximum of systematic variations of N and then project X onto the subspaceorthogonal to P−, and resulting in the corrected spectraX∗, which therefore containsX+ and a part of E:

X∗ = X ·[I − P−P−T ] (2.5)

Hence, according to Zeaiter [189], the different orthogonal projection methods arebased 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 methodsstarts by determining the structure of variance of X data by PCA. PCA scores arethen orthogonalized to Y giving T−. And finally, the matrix of latent structuresP− corresponding to T− is calculated using PLS or PCR. Besides, techniques thatidentify 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 byusing a set of dedicated experiments which incorporate systematic variations, theseare 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 ofX and Y , and an indirect approach, where X− is identified from the structure ofX. In the direct approach, it was expected to avoid the problem of overfitting thatpresent the methods following the indirect approach. It finds a subspace common toX and Y relative to X+. Then PCA is performed on this subspace to determine theorthonormal base P+. OSC by Fearn, OPLS, O2PLS, DO and an improved versionof POSC follow this strategy. In contrast, the indirect approach is the strategyused in projected orthogonal signal correction (POSC) [195] which is similar tothe 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−, whichcontains spectra acquired at different levels of variations of known influence fac-tors.Then, P− is identified using PCA on Z−. Different methods exist using thisdirect approach, such as independent interference reduction (IIR) or external pa-rameter orthogonalization (EPO). Most of these methods are able to remove onlythe effect of known influence factors (temperature, scale, instruments, etc. ), or acombination of them, this is the case of EPO, EROS or IIR. However, DOP is ableto remove variations related to unknown influence parameters and components inthe data that might affect the model. Thus, extrinsic methods have the advantageof not requiring the Y information to reduce the complexity of the model, doneby removing the effects of known or unknown influence factors [213]. Componentcorrection (CC) technique, used in e-nose applications, can be considered also tobelong to these extrinsic methods.

Despite the elimination of variability non related to the information of interest, many

49


authors have reported that most of the intrinsic algorithms are not able to reducethe error of prediction compared to ordinary PLS regression (only slight improve-ment in specific applications), even though the number of latent variables (LVs) isconsiderably decreased. Furthermore, these methods present the disadvantage of arisk of overfitting the number of OSC components, which need to be adjusted, andalso 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 inthe possible easier interpretation and understanding from the analysis of correcteddata, and the possibility of an analysis of the orthogonal information removed, whichmight 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 thinkthat parsimonious models obtained by applying intrinsic correction methods can beexpected to be more robust, even though prediction error is not reduced comparedto ordinary PLS regression on raw data.

In contrast, the use of most extrinsic methods have resulted in an improvement notonly in an easier interpretation of the calibration model, but also better predictionability 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 measurementsand 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 techniquemust be selected and combined, based on the nature of the data to be analysed andthe relevant features in the data.

However, it has to be noted that most of OP algorithms in spectroscopy have beenapplied on regression problems. Only few works can be found in literature in whichOP methods are used as preprocessing algorithms for classification problems, oftenrelating only two classes [214]. In these works the use of OP generally improvesresults. In contrast, most AO system problems involve classification of multipleclasses, and the removal of noisy parameters by OP techniques, specially intrinsicmethods, is interesting, since it would lead to more compact clusters and thereforeto easier cluster discrimination. And, surely, this improvement in the predictionability come along with an improvement in the interpretation of the model.

2.2.2 Multivariate Curve Resolution (MCR) methods

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 maybe prone to calibration errors, and these errors are one of the major obstacles totheir practical use. Calibrating every sensor by hand is infeasible if the networkscontain tens of devices and, consequently, automatic methods for jointly calibratingsensor networks in the field, without dependence on controlled stimuli data, is of

50


significant interest.In analytical instrumentation fields, such as techniques related tospectroscopy and chromatography, terms with similar meanings, such as calibrationfree [217], model free [218], soft or self-modelling [218], and self-modelling curveresolution [219] can be found in literature. The aim of all these methods is the blindextraction, thus requiring no prior information, of the spectral and concentrationprofiles of pure components from unknown samples mixture. In chemical processmonitoring, these techniques can also provide information about side reactions andintermediates over the course of the reactions involved in the process, and thereforea better understanding of such process is obtained.

The objective of both calibration free and blind calibration methods is similar to theso-called blind source separation (BSS) problem. BSS deals with situations that areclosely related to the cocktail-party problem, well known in signal processing, whichinvolve the recovery of the original speech signals of different speakers from the mixedsound. In signal processing, the set of algorithms named independent componentanalysis (ICA) are usual for solving such BSS problems. These methods have alsobeen used in other disciplines, including analytical chemistry [220] and electronicnoses [168, 169]. ICA algorithms uses high order statistics for de-mixing the originaldata, assuming that the data matrix contains linearly mixed signals coming frommutual statistically independent (non-Gaussian) and thus uncorrelated sources andwithout any prior information about the sources or reference materials. Like ICA,most calibration-free resolution techniques are based on the assumption that theinstrumental response in a mixture is an additive linear combination of the signalsfrom individual species, the pure components [217].

In contrast, in analytical chemistry, the set of multivariate curve resolution (MCR)methods are the mostly employed for BSS. The problem consists of blindly extractinginformation about the pure components from a mixture, which may contain manycomponents to be simultaneously analysed or may include a few interesting analytesin the presence of many other chemical interferences, as in environmental samples.MCR methods does not use high order statistics to extract the pure componentprofiles, but the simple least squares (LS) technique and some constraints extractedfrom prior problem knowledge. LS is performed in an iterative way in the mostcommon MCR algorithm named MCR alternating least squares (MCR-ALS) [221].

The correct performance of MCR methods depends strongly on the complexity ofthe multicomponent system. In particular, the degree of overlap among the pureprofiles of the different components and the specific way in which the regions orwindows of existence of these profiles are distributed along the row and columndirections of the data set, condition the ability to correctly recover pairs of pureprofiles and spectra for each of the components in the system [222].

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 componentsin the system by decomposing the initial mixture data matrix X into the product oftwo chemically meaningful data matrices C and ST . Each of these matrices contains

51


the pure response profiles of the n mixture components associated with the row andthe 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 modelof the form:

X = CST + E (2.6)

where X(r × c) is the original data matrix, C(r × n) and ST (n × c) are the ma-trices that contain pure-component profiles related to the data variation in the rowdirection 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 chemicalcontribution. The variables r and c represent the number of rows and the number ofcolumns of the original data matrix, respectively, and n is the number of chemicalcomponents in the mixture or process. Matrices C and ST often refer to concen-tration profiles and spectra, but they can refer to other factors in other problems[222].

However, the mathematical decomposition of the data matrix by equation 2.6, nomatter the method used, is ambiguous if no additional information is available,i. e. it is subject to ambiguities. This means that many pairs of matrices C and STcan be found that reproduce the original data set with the same fit quality, thusthere is a rotational and scale freedom in the solutions of equation 2.6, since theobtained component profiles may differ in shape (rotational ambiguity) or in mag-nitude (intensity ambiguity) from the true ones. Nevertheless, in order to restrictthe extent of these ambiguities, constraints derived from the physical nature of thesystem and from prior knowledge of the problem under study can be added. Mostcommonly, these constraints require non negativity (in particular for dealing withconcentration profiles in the elements of matrix C) or peak uni-modality (speciallyfor chromatographic problems, on matrix ST ), but other constraints can also beapplied [222].

On the other hand, MCR methods include two steps; a first step finds an initialestimates for matrix C or ST and the second one, often ALS, refines the previousestimation step and incorporate the required constraints. Initial estimates can beobtained 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 ofthe data set in order to know how the number and distribution of components evolvealong the data set. Evolving Factor Analysis (EFA) [223] and Fixed-Size WindowEvolving Analysis (FSWEFA) [224] are some examples. Furthermore, these tech-niques are particularly relevant in the study of processes, where the concentrationprofiles of the different components evolve smoothly and, often, following a sequen-tial pattern. On the other hand, methods of pure variable selection find the mostrepresentative row or column profiles for the different components in the data setto determine the number of components. Simple-To-Use Interactive Self-ModellingMixture Analysis (SIMPLISMA) [225] is the most commonly used technique of this

52


type [218].

Besides, other resolution methods, such as iterative target transformation factoranalysis (ITTFA) [226], Orthogonal Projection approach (OPA) [227], Window Fac-tor Analysis (WFA) [228], Subwindow Factor Analysis (SFA) [229], Heuristic Evolv-ing Latent Projections (HELP) [230], Positive Matrix Factorization (PMF) [231]and Non negative Matrix Factorization (NMF) [232], etc. or even ICA can be usedto provide the initial estimates for MCR-ALS. Triadaphillou et al. [217] investigatedand compared some of these techniques on synthetic spectral data, and obtainedthat they were all effective in terms of resolution, specially combinations of some ofthem with MCR.

In AO systems MCR has never been applied until now. In this thesis, we presentan application of MCR-ALS where, given a dataset of measurements by an array ofgas sensors, with pure and mixture samples, information about the concentrationof the pure analytes and temperature sensitivities are obtained in matrix C and Srespectively.

2.2.3 Multiway (N-way) methods

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 thetarget analyte, can be obtained (see section 1.5). The inclusion of any of these extrainformation 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 dataset can be organised in the form of a cube or a three way matrix (fig. 2.1), e. g. atensor where every edge or mode; row, column and tube (fig. 2.2)), refers to eitherthe 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 usingmultiway signal processing techniques. These methods are often used for preprocess-ing. They provide more parsimonious models than classical methods, and extractedfeatures are meaningful and easier to interpret. A deep study on multiway methods

Figure 2.1: Three-way matrix structure, modes and slices.

53


(a) Rows, (b) Columns, (c) Tubes.

Figure 2.2: Modes in a three-way data matrix [233].

can be found in [116]. Also, De Juan and Tauler make a comparison of three-wayresolution methods for non-trilinear chemical data sets in [234].

The most common multiway techniques are parallel factor analysis (PARAFAC) andTucker, and for both extensions and modifications have been proposed (PARAFAC2,Tucker3, etc. ).

For three way datasets PARAFAC is the alternative to PCA, however PARAFACrequires 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 onereferring to a mode and having all the same number of factors. PARAFAC2 is amodification 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 modelbuilding 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 andallowing different interactions among the modes. The decomposition of the three-way matrix X, according to some of the multiway methods, is given by:

PCA : xij =F∑

f=1

aifbjf + eij (2.7)

PARAFAC : xijk =F∑

f=1

aifbjfckf + eijk (2.8)

Tucker3 : xijk =F∑

d=1

F∑

f=1

H∑

h=1

aidbjfckhgdfh + eijk (2.9)

where xijk represents the ijkth element in the three-way data set, aif , bjf , ckf are theelements in the decomposition matrices A, B, and C (see fig. 2.3 where the differentmethods are represented graphically), eijk is the residual term and gdfh is the dfhthelement of a core array (in Tucker model) of size D × F × H. D, F and G are theselected number of factors in each decomposition matrix.

54


Skov and Bro [235] applied PARAFAC2 on an e-nose dataset with dimensionssamples × transient time × sensors. This method was preferred to classicalPARAFAC because of its ability to handle shifted sensors transient profiles. The au-thors compared this technique with PCA and conclude that, although PARAFAC2offers some advantages compared to PCA, the structure of the model required amore careful interpretation of the model parameters.

Figure 2.3: Graphical view of a three-way matrix decomposition by PARAFAC, PRAFAC2and Tucker3 methods respectivelly [234]. The term E is the residual error.

When classical methods are to be used on the three-way matrix, which is the usualapproach in AO, this is unfolded often in a row-wise way (fig. 2.4), and therefore therelationship among the three modes is destroyed and interesting information lost.

Figure 2.4: Column-wise unfolding of a three-way data matrix.

2.2.4 Calibration Transfer

Processes and applications that require continuous monitoring run over long periodsof time, during which drift is not the only problem the instruments may present, theyalso 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 thetasks involved in it; experimental design, new data acquisition, data evaluation andpreprocessing, model building and validation, and finally application of the selectedmodel to predict responses of unknown samples. However, good calibrations requiremany reference samples, depending on the instrument and the model complexity,

55


which must represent all possible situations that may arise during later operationof the instrument. Other situation may arise when, after considerable effort placedinto constructing a robust model, the samples to be predicted are measured on adifferent instrument or under differing environmental factors from those used to buildthe model. Further, many identical or almost identical instruments can operate inmany locations in an industry. In this case, one often has a master instrument anda number of field instruments, and thus, along their life time, regular re-calibrationshave to be run on each of the instruments.

Therefore, there exist many situations in which calibrations have to be repeated veryoften, 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 timeexpense of re-calibration by transferring the information contained in the model ofa first instrument to a model for a second instrument using as few measurements aspossible. This is a problem for many different instrument manufacturers, and thistopic is expected to expand as calibration spreads into more complicated industrialapplications of continuous monitoring.

The transfer of calibration is made from one instrument to a slightly different (almostidentical) instrument. Being that, for identical instruments calibration models areequal, and completely different instruments have no information in common, andthus all possible situations have to be measured also on the second instrument.When the instruments are similar, it is possible to assume that measurements insimilar 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 builtfor the second instrument [136].

The above mentioned situations, where calibration transfer is a useful solution tocheapen the calibration tasks, is directly applicable to gas sensor arrays. Further-more, in these devices, sensors deterioration is usual and may cause an initial cali-bration become totally invalid in a relative short time. After sensor replacement,the instrument performance may be substantially changed. This case is commonin gas sensor arrays, since sensor signals can differ significantly from those of thereplaced ones, although new and replaced sensors are supposed to be identical. An-other situation arise in the case on an array of gas sensors affected by drift. In sucha case, the problem of drift reduction and calibration transfer are similar, since thedrifting array can be seen as having one instrument at time t1, and another slightlydifferent instrument at a later time t2.

Different methodologies can be followed in order to counteract for the data shiftbetween two instants measurements in the same instrument or measurements madein two instruments. Feudale et al. [199] makes an interesting overview of thesemethodologies in the area of near-infrared (NIR) spectra.

In a mathematical form, if a calibration model f is made for mapping instrumentmeasurements X to predicted variables Y so that Y = f(X), these strategies can

56


be made; by transforming the model f , by pre-processing of the X-data, or bypost-processing of the Y-data. Preprocessing methods counteract for deficiencies insensor data before the proper calibration model is built, post-processing methodscounteracts the outcome of the calibration model (performed on deficient sensordata) and, when f is transformed, the counteraction in the process is obtained by aproper adjust of its coefficients or by the use of new incoming samples [136].

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 atcertain intervals or use procedures which automatically identify changes and ap-ply counteraction. Therefore, reference samples are needed in order to identify theundesired changes. These reference samples must be carefully chosen; they mustbe representative enough to describe the differences between the instruments, andshould ideally represent the entire experimental domain. Unfortunately, they tendto be hard to obtain because of different sample conditions, samples degradation, en-vironmental changes, etc. Methods that need such reference samples are often calledstandardization methods and many techniques coming from the field of chemometricscan be used, but, when transfer standards are not available, preprocessing methodscan be employed.

Research on calibration, recalibration and calibration transfer is very considerablein other instrumentation fields such as near-infrared (NIR), Raman and fluorescencespectroscopy, UV-visible spectrophotometry, etc. but not in artificial olfaction, andonly a few references can be found in literature. For example, Balaban et al. [236]applied three types of conversion functions and two neural networks on data fromtwo 12 conducting polymer gas-sensor array systems, to predict storage time of themeasured milk samples. The task of constructing a mapping between two electronicnoses that employ two very different sensor technologies, quartz microbalance andconducting polymers, is discussed by Shaham et al. [237]. Tomic et al. [238] studiedtwo approaches for recalibration of long-term measurement data from a solid-stategas-sensor array system; univariate multiplicative drift correction and multivariatecomponent correction. The efficiency of these methods was evaluated by classifyingrecalibrated measurement data using k-nearest neighbour (kNN) classifier and par-tial least-squares discriminant analysis (PLS-DA) trained with old measurements.Both methods proved to be useful for eliminating the signal shift caused by sen-sor replacement. The same author [157] proposed other two approaches to attemptto correct signal shift between measurements acquired with five commercial quartzmicro balance (QMB) gas-sensor array systems of identical construction. Both meth-ods worked by post-processing of the measurement data; straightforward univariatedirect standardization method (UDS), based on linear regression, and multivariatepartial least squares regression (PLS). Both methods proved to be very effective.

57


2.2.5 Fault detection and isolation (FDI) methods

As mentioned above, in situations where one or more sensors in an array fails, itis very important to detect and identify the failing sensor, and also to correct thissensor response, in order to avoid incorrect results that could lead to erroneousdecisions. Indeed, fault detection is a key point in industry, where the increasingdegree of automation and the growing demand for higher performance, efficiency,reliability and safety in their systems, have made the control research community toput a great effort in developing the areas of fault detection and identification (FDI)and fault detection and diagnosis (FDD). The purpose of FDI is mainly to useavailable sensor signals to detect and identify the faulty sensor, whereas FDD alsodetermine the cause of the fault. As in other subjects, other terms can be found inliterature, such as fault-tolerance [239], soft sensors [240], inferential sensors [241],observer-based sensors [242] and self-validating sensors (SEVA) [243]. All thesefields share similar objectives, but some of them also cover the reconstruction ofthe failing sensor’s response or the automation of the decision to be taken after thediagnosis of the fault, as SEVA and fault tolerant systems. A great progress havebeen done in these areas in comparison with classical fault detection methods, whichmethods are based on checking a threshold value of some important measurablevariables.

FDI methods can follow two different approaches; model-driven and data-driven.The model-driven family of FDI techniques is most commonly based on modelsthat describe the physical and chemical principles underlying the process. However,the development of these models requires a lot of process expert knowledge whichis not always available. Furthermore, since they are specifically developed for thedesign of the processing plant, they usually focus on the optimal steady-state ofthe 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 statespace or an input-output model, FDI can be classified into observer based, systembased and also adaptive observer based, this latter combines the best features of thetwo previous approaches. In general model based FDI methods can only be appliedto 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 caseswhere the system is time-varying and highly non-linear for example. At the sametime, data-driven FDI techniques can be divided in other three groups; signal based,multi-variable statistics based, and knowledge based. The common feature of thesemethods is that they all use raw system data and sometimes process knowledge tocarry 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. Knowledgebased methods include artificial intelligence algorithms, including expert systems

58


[245].

In multi-variable statistics based FDI methods, techniques like PCA can be usedto reduce the high dimensionality of the data in a few features containing most ofthe information of the original data. PCA divides the whole sensor space into aprincipal component (PC) subspace and a residual subspace, and then performs theFDI using Q-test and Hoteling T 2 test. This is done by checking whether or not thenew incoming data lies in the PC subspace through the mentioned statistics givenby the squared prediction error (SPE) or Q (eq. 2.10), and T 2 (eq. 2.11):

SPE = ‖x‖2 (2.10)

T 2 = xTPΛ−1P Tx (2.11)

where x is the new incoming data, x 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, thePCA residuals have to be analysed, even methods for structuring its residuals havebeen 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, whichlimits its application in complex industrial systems that exhibit a time-varying,non-Gaussian and non-linear characteristics. Alternatively, independent componentanalysis (ICA) does not assume the data to be Gaussian distributed and other meth-ods can deal with non-linear characteristics.

Multi-variable statistics based FDI methods are very interesting to implement inAO systems, since they are simple and robust, despite gas sensors response are nonlinear. Specially, it is worthy to test the performance of PCA since it has beenapplied in other fields with great success [244].

59


60

Chapter 3

Thesis objectives

The aim of this thesis is to investigate signal processing methods that may improvethe robustness of AO systems. In particular, the main focus is the time stabilityand sensor fault detection, identification and correction.

To improve time stability, we take a deeper look to techniques able to find restrictedsubspaces that reject drift while keeping the predictive power of the models. Wewill investigate the family of Orthogonal Projection methods, specially OrthogonalSignal Correction, in their capacity to reduce drift but also unwanted variance. Itsabilities will be compared to state of the art techniques in AO like ComponentCorrection.

Sensor poissoning detection has received poor attention by the AO community. Thisthesis proposes automatic methods to detect and correct faults in chemical sensorarrays based on the inner correlation of the sensor signals.

Additionally, we investigate the use of unsupervised feature extraction methods thatmake use of sensors time information either due to sampling transient or temperaturemodulation. A potential way to improve solution robustness is to introduce priorknowledge into the data matrix decomposition by means of constraints. Amongthese methods this thesis will investigate the ability of PARAFAC and MCR-ALSto model sensor array time responses.

61

Thesis objectives Marta Padilla Ferran

62

Chapter 4

Journal and main Conference Papers

Some of the results obtained during the course of this thesis are reported in fourjournal 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 ofthe journal papers.

4.1 Introduction to papers

Detection of diverse mould species growing on building materials by gassensor arrays and pattern recognitionM. Kuske, M. Padilla, A.C. Romain, J. Nicolas, R. Rubio, S. MarcoSensors and Actuators B: Chemical, vol. 119, no.1, pp. 33-40, 2006.

As already mentioned, chemical sensors are prone to drift, which causes statisticalmodels to degrade in time. This is shown in this paper, where an electronic nose isused to recognize building materials infected by moulds on two measurement seriesseparated by about four months.

A first goal in this work, is to perform several tests in order to ascertain severalquestions relative to the detection of moulds by an array of ten commercial MOXsensors. Studied issues includes the influence of every type of building material andthe 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 asthe presence of moulds could be detected when growing over an unknown buildingmaterial.

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, fromwhich a classification model was built. Results from this second data set revealsthe degradation of the classification model caused by drift. Although the overallclassification 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 ofthe substrate being considered. This also illustrates the fact that drift influencesindividual sensors in different ways, as was noted in section 1.6.3, since the evolutionof data along the time depends on the class they belong to.

It is interesting to observe the PCA scores plot of the distribution of infected orhealthy growing substrates. There, it is shown a high dispersion of the sampleswhich depends mostly on the type of building material, however scattering of thesamples is also present within every substrate class and this is mostly caused bydrift.

Feature extraction on three way enose signalsM. Padilla, I. Montoliu, A. Pardo, A. Perera, S. MarcoSensors and Actuators B: Chemical, vol. 116, no.1-2, pp. 145-150, 2006.

As seen in section 1.5.1, the use of the full transient signal of the sensors responseincreases the predictive capabilities of an electronic nose. When transient signalsfrom 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 parsimoniousand easier interpretation of the models. Extracted features can be used to buildposterior models for prediction.

In this paper, the multiway method parallel factor analysis (PARAFAC) is applied inan application for quality control of a processed food product, fried potato chips withdifferent concentration of a flavouring agent. The measurements were performedwith a commercial electronic nose containing 13 metal oxide sensors. Later, onlyone 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 canbe calculated. Remaining loadings from PARAFAC give information about theinfluence of every individual sensor on the model and identify the time windowwhen the the sensors are exposed to the volatiles.

It has to be noted that, before the application of PARAFAC, the trilinearity propertyrequired by PARAFAC must be tested. Furthermore, a number of factors for thePARAFAC model have to selected. For this, several models of the calibration set aremade 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 thischoice.

64

Introduction to papers

Drift compensation of gas sensor array data by Orthogonal Signal CorrectionM. Padilla, A. Perera, I. Montoliu, A. Chaudry, K. Persaud, S. MarcoChemometrics and Intelligent Laboratory Systems, vol. 100, no. 1, pp. 28-35, 2010.

As mentioned in section 2.2.1, orthogonal projection (OP) methods are a family ofinteresting algorithms to deal with drift and local sample dispersion. OP methodsare used in calibration transfer problems and as a preprocessing technique to simplifya posterior calibration model and thus to increase its robustness. However, thesemethods are almost not used in AO, only few references can be found in literature,like [157, 168].

With the aim of exploring the ability of OP techniques to diminish the influenceof drift, in this paper it is proposed the use of one of the OP methods, orthogonalsignal correction (OSC) algorithm, to compensate drift on an experiment lasting10 months. This experiment consists of measurements of three type of analytes atseveral concentration levels by using an array of 17 conducting polymer gas sensors.In addition to OSC, Arthursson’s component correction (CC) method, explainedin section 1.6.5, is tested on the same dataset in order to be compared with OSCand 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 valuesare used in the design of both CC and OSC techniques. Besides, it is also importantto note that the whole transient signals of the sensors have been used, since theyprovide additional information for class discrimination, as mentioned in section 1.6.

Additionally, it is also proposed a method to properly select the two parametersinvolved in the OSC algorithm. This method is based on the Fisher’s coefficientand attempts to show which parameters values could lead to an over-fitting, whichis a strong risk in the application of OSC. Finally, another objective in this workis to test the robustness of both methods, including preprocessing and classifieralgorithms, in relation with the sufficient size of the calibration set to provide stableresults.

Again, PCA score plots are very interesting, as they show firstly the effect on shortand long term drift on the data, and then the reduced class clusters obtained oncethe preprocessing methods are applied.

Multivariate curve resolution applied to temperature-modulated metal oxidegas sensorsI. Montoliu, R. Tauler, M. Padilla, A. Pardo, S. MarcoSensors and Actuators B: Chemical, vol. 145, no. 1, pp. 464-473, 2010.

In this paper, multivariate curve resolution alternating least squares (MCR-ALS) isapplied for the first time on data from a gas sensor array.

65

Journal and main Conference Papers Marta Padilla Ferran

As previously explained in section 2.2.2, self-modelling methods extract informationfrom data without the need of a calibration model and thus are not affected by drift.MCR-ALS is widely used in analytical instrumentation, where its capabilities forqualitative analysis have been proven. In this work, it is applied on data measuringthe evolution of two species of gases in a mixture with an array of four metal oxidesensors thermally modulated with a triangular waveform. Thermal modulation (seesection 1.5.2), provides higher selectivity and sensitivity than isothermal heating,since it gives more information by screening the thermal sensitivity of every sensorto the analytes in the mixture. Hence, in this experiment MCR-ALS determinesthe resolution of the gas mixture by means of the concentration C and also thetemperature sensitivity profiles ST to the both analytes in the mixture.

First, MCR-ALS is applied on synthetic data and then on real data. The analysisresults in a resolution of the gas mixture that gives C and ST profiles. Satisfyingresults are obtained from synthetic data, however in real data non-linear effects areobserved.

Further, the analysis is made on measures from a single sensor, and later on datacontaining responses from all sensors in the three way matrix (samples × sensors× temperature modulation waveform), which is unfolded into a two way matrixto be analysed.

Additionally, the concentration of the analytes present in the mixture is estimatedby building a linear quantitative model for each of them, although, in this paper,given the small size of the dataset, this model has to be validated with root meansquare of calibration (RMSEC).

Poisoning fault diagnosis in chemical gas sensor arrays using multivariatestatistical signal processing and structured residuals generationM.Padilla, A.Perera, I.Montoliu, A.Chaudry, K.Persaud, S.MarcoIEEE International Symposium on Intelligent Signal Processing, WISP 2007, pp. 1-6, 2007.

Since in an array of sensors one or more may fail, it is important to detect thefault on time, identify the failing sensor and correct its response until the sensor canbe replaced by anew one (section 2.2.5). In this work, a possible and subtle typeof fault typically affecting chemical gas sensors, is simulated. The poisoning faultcauses a change of the sensitivity profile of the affected sensor. Poisoning is difficultto detect compared with other type of faults, because the sensor keeps working underapparently normal conditions.

In this work, we proposed the use of simple statistical methods such as PCA, PLSand PCA with structured residuals, to perform the three above mentioned tasks ofdetection, identification and correction of one and two faulty sensors in an array of17 conductive polymer gas sensors. All three techniques are compared in every stepby 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 aftercorrection of the faulty sensor’s response.

Fault detection, identification, and reconstruction of faulty chemical gassensors under drift conditions, using Principal Component Analysis andMultiscale-PCAM.Padilla, A.Perera, I.Montoliu, A.Chaudry, K.Persaud, S.MarcoIEEE World Congress on Computational Intelligence, WCCI 2010, 2010.

Also in this work, we test the ability of a signal processing method to detect, identifyand correct one or two poisoned sensors in an array, but with an additional difficultyin which the system is subject to irregular drift. Therefore, since the method hasto deal with drift, the tasks of detection, identification and correction of the faultysensor become more challenging.

The dataset on which this situation is created consist of measurements of threeanalytes at different concentration levels along ten months of experiment. Again,the poisoning fault is simulated by changing the sensitivity profile of the failingsensor, which belongs to an array of 17 conductive polymer sensors.

In this work, we propose the use of multiscale-PCA for the detection, identificationand correction of a faulty sensor under drift conditions. The multiscale step consistson the prior decomposition of every sensor response along the time (sample mea-sures) by a discrete wavelet transform (DWT). DWT allows to isolate the slowervarying 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 allsensors are grouped in a matrix, to which the same PCA method that in section 4.1is applied to perform the detection, identification and correction tasks.

4.2 Full papers

67

Sensors and Actuators B 119 (2006) 33–40

Detection of diverse mould species growing on buildingmaterials by gas sensor arrays and pattern recognition�

M. Kuske a, M. Padilla b, A.C. Romain a, J. Nicolas a, R. Rubio b, S. Marco b,∗a Departement Sciences et Gestion de l’Environnement, Universite de Liege, Avenue de Longwy, 185, B6700 Arlon, Belgium

b Sistemes d’Instrumentacio i Communicacions, Departament d’Electronica, Universitat de Barcelona, Martı i Franques 1, 08028 Barcelona, Spain

Received 26 November 2004; accepted 16 February 2005Available online 6 January 2006

Abstract

This work explores the detection of moulds growing in different building materials by using a metal oxide sensor array. Four moulds specieshave been considered. Pattern classification provides detection rates on the order of 80–85% for different species. Drift degrades only slightly thesevalues subsequent test 4 months later.© 2005 Elsevier B.V. All rights reserved.

Keywords: Gas sensor arrays; Moulds; Sensor arrays; Pattern recognition

1. Introduction

One of the most frequent problems in buildings is the fungaldevelopment caused by excessive humidity. Fungal contamina-tion can produce infections, allergies, toxic effects and othersymptoms documented in many studies and characteristic forthe “sick building syndrome” [1–8].

Traditionally, fungal contamination in a building has beendescribed as quantity of viable spores determined from air, set-tled dust, surface and building material samples and results ofmeasures are obtained after several days. New methods involvesthe detection of fungal components, mycotoxins and micro-bial volatile organic compounds (MVOCs). This is interestingbecause compounds can penetrate barriers not penetrable byspores, facilitating the detection of hidden moulds. The detec-tion of these compounds was previously carried out by samplinginto carbon-based or TENAX adsorbents, and analysing bygas chromatography–mass spectrometry. However this method,very sensitive and specific, requires experience and special lab-oratory equipment, takes time and is relatively expensive. Theelectronic nose technology, in difference with gas chromatogra-phy, is simpler, cheaper, and the results can be obtained in situ.

� Submitted to Eurosensors ’04.∗ Corresponding author. Tel.: +34 93 402 9070; fax: +34 93 402 1148.

E-mail address: [email protected] (S. Marco).

Moulds produce a wide range of MVOCs: alcohols, ketones,terpenes, esters and sulphur compounds [9–11]. The produc-tion depends on environmental conditions, the species and thesubstrate on which the fungi grow [12–17]. However, no singleVOC seems to be a reliable indicator of biocontamination [18].For this reason fungal detection cannot be based on a singlesubstance but on the coexistence of several compounds.

The authors have already addressed the detection of a sin-gle mould species, namely Aspergillus versicolor, in a previouspaper [19]. In that paper a table with a detailed list of VOCsproduced by moulds can be found. Here this work will becompleted with more extensive measurements involving threeadditional mould species: Penicillium aurantiogriseum, Peni-cillium chrysogenum and Cladosporium sphaerospermum.

Fungal detection using electronic noses has been previouslyaddressed. Schiffman et al. [20] studied the ability of electronicnoses to detect moulds inside buildings. An array of 15 metaloxide sensors was capable of discriminating among the fungi andit was also able to recognise selected five volatile organic com-pounds emitted by fungi [21] studies showed that a conductingpolymer sensor array was able to distinguish the volatile patternsproduced by three species of moulds growing on three types ofpaper, permitting the early detection of fungal contamination inlibrary materials and archives [22] predicts the potential of elec-tronic noses for improving the possibility to detect individualfungal species as well as the degree of mycotoxin contamina-tion of food and animal feeds [23]. Investigated the possibility

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

of detecting fungal volatile metabolites as indicators of myco-toxins in cereal grain, using both an electronic nose and gaschromatography combined with mass spectrometry (GC–MS).Other studies [24–27] showed that sensor arrays and differenttechniques of statistical analysis allow detecting and classifyingselected fungal species growing on cereals. These techniquescan be used for quality classification of grain. An e-nose used byPersaud et al. [28] could discriminate between wood infected bySerpula lacrimans and uninfected samples. And [29] comparedtwo electronic nose systems employing conducting polymer sen-sor arrays for the early detection and discrimination betweenbacterial species, fungal spores and trace amounts of pesticides.

In this paper, the capability of a metal oxide gas sensor arrayto detect the presence of diverse mould species growing on avariety of building materials is investigated. Section 2 presentsthe materials and methods used in this work. Section 3 presentsthe methodology for pattern analysis. Section 4 presents theobtained results and finally Section 5 draws some conclusions.

2. Materials and methods

An electronic nose instrument was designed and constructedat the Universite de Liege. It consists of a sampling unit, a sen-sor array, and a signal processing system (Figs. 1 and 2). Thegas sensor array contains 12 metal oxide sensors in two sep-arate chambers placed in parallel. Each sensor chamber hasa volume of 210 cm3. The sensors were selected in order toobtain different responses for volatile compounds produced bymoulds. Six of them were manufactured by Capteur; CAP01,CAP03, CAP06, CAP07, CAP23, CAP25, and six from Figaro;TGS2620, TGS2180, TGS825, TGS822, TGS2600, TGS2602.

Four different materials typical of Belgian houses were usedas substrates for mould growing: plasterboard, particleboard (orchipboard), oriented strain board and wallpaper. Some combina-tions of them were also analysed; plasterboard with wallpaperand glue, particleboard with wallpaper and glue and orientedstrain board with wallpaper and glue. The four fungal specieswere separately incubated on culture medium, Malt Extract

Fig. 1. Schema of the instrument used in the study.

Fig. 2. Picture of the actual experimental setup showing the bubbler, the jarunder analysis, the mass flow controller and the sensor chambers.

Agar, for 1 week. After this period, moulds were inoculated intothe building materials. Contaminated samples were incubatedin airtight glass jars (500 ml) closed with a Teflon cover. Eachjar contained two pieces of the same material of dimensions5 cm × 6 cm contaminated with one fungal species and 40 ml ofdemineralised water in its bottom, for keeping the high degreeof humidity that moulds need to grow. The samples were placedon top of small glass vials to avoid direct contact with water.The jars were kept at room temperature in the darkness duringthe whole duration of the experiments.

For each fungal species, two sequential experiments werecarried out, resulting in two datasets comprising the growth ofthe moulds during 4 months. Clean and contaminated materials(corresponding to different species) were randomly presentedto the gas sensor array during this time. Contaminated sampleswere presented in order according to the species of moulds, inthis way firsts sets of A. versicolor followed by P. chrysogenumwere firstly measured, and then seconds sets of both speciesin the same order. After that first sets of C. sphaerospermumfollowed by P. aurantiogriseum were presented to the electronicnose and finally seconds sets of the two species were measuredin the same order. The contents of all datasets are summarizedin Table 1.

The electronic nose has the following operation; clean dryair is fed to a mass flow controller, then to a humidifier and thento a jar with samples to sweep the volatiles accumulated in theheadspace. Two three-way valves commute clean air or the jarheadspace to the gas sensors chambers (see Fig. 1). A samplecycle included a passage of clean humidified air (reference air),followed by a passage of air from the samples during 3 min (timeof signal stabilisation) at 200 ml/min of constant air flow.

3. Signal and data processing

The sensor signals were preprocessed by taking the ratio ofthe resistance at the end of the sampling period (3 min) and thebaseline resistance, just before sampling when exposed to the

M. Kuske et al. / Sensors and Actuators B 119 (2006) 33–40 35

Table 1Number of clean and contaminated samples for each substrate material and specie of mould in the two datasets

Materials Samples Type of mould

A. versicolor P. chrysogenum C. sphaerospermum P. aurantiogriseum

First set Second set First set Second set First set Second set First set Second set

Oriented strain board Clean 27 16 21 22 15 20 15 26Contaminated 27 16 11 18 16 16 12 12

Plasterboard Clean 27 16 21 22 18 22 15 27Contaminated 27 16 11 18 16 17 15 11

Particleboard Clean 27 15 21 21 16 19 14 27Contaminated 25 16 11 18 17 17 15 12

Wallpaper Clean 17 8 14 11 7 10 8 13Contaminated 17 8 7 8 9 7 7 7

Total Clean 98 55 77 76 56 71 52 93Contaminated 96 56 40 62 58 57 49 42

vapour saturated carrier gas. In this way every measurement wasconstructed with 12 sensors signals, assembling a 12 dimen-sional vector. However two sensors (CAP23 and TGS2180)were removed because of deficient behavior observed by visualinspection of the sensor patterns, therefore the formed patternvector has 10 dimensions.

As already mentioned, two datasets were collected for thisstudy for each species. First dataset was used for all the dataprocessing algorithm development, and the second one for aposterior evaluation (including the evaluation of drift), since itwas collected 4 months after the first one.

For validation of the classifier results on first dataset the fol-lowing strategy has been followed. This dataset (for each mouldspecies or all together) has been divided in two equally sized ran-dom partitions. One part has been used for classifier design andoptimization while the second half is used for classifier testing.This process is repeated ten times and the average classificationrate is retained. Also the standard deviation of the classificationrate is estimated.

Later, the robustness of the electronic nose for mould detec-tion was tested by including a wide variety of moulds andbuilding materials. Classification of each specie of mould is con-sidered, as well as the capability of a classifier to generalize tounknown mould species (not present in the training set). Finallyalso the classification of samples with an unknown substrate wasexplored.

Our primary objective was the binary classification of mouldsand no-moulds disregarding the particular fungal species and thesubstrate material. For this purpose a Mahalanobis classifier wasapplied on the raw data. The obtained results are on the sameorder than the ones seen in the previous paper [19]. Although thesame classifiers as in [19] were considered; namely k-NN, fuzzy-kNN with different initial membership functions on dimensionalreduced space and on raw data (a review of Fuzzy k-NN can befound in our previous paper [19]) we found that the applicationof complex classifiers on our small database did not provide anyadvantage; the obtained classification rates were not in generalbetter than the ones obtained using the Mahalanobis classifier.

For this reason only the Mahalanobis results are reported in thispaper.

4. Results

A PCA loadings plot on first two principal components(Fig. 3), for the ensemble of all moulds, shows one group of fiveFigaro sensors and one Capteur sensor on the left hand side andtwo groups of two Capteur sensors on the right. PCA loadingsplots for every specie of moulds are almost the same. Signalsprovided by the five Figaro sensors (series TGS) are very similarand anti-correlated to Capteur sensors signals on the first princi-pal component (except for CAP01). CAP01 signal is interestingbecause it varies inversely with the rest of Capteur sensors onPC1 and with all Figaro sensors on PC2. CAP03 and CAP25

Fig. 3. PCA projection, loadings plot for all moulds together. Samples fromfirst set of all moulds. PC1 and PC2 explains 82.27% and 10.88% of variancerespectively. The points corresponds to the five Figaro sensors and the plussymbol represent the Capteur sensors.


Fig. 4. PCA projection. Samples from first set of all moulds. PC1 and PC2explains 82.27% and 10.88% of variance respectively.

are not very significant for PC2, the rest of sensors gives infor-mation to both first two principal components. This loading plotalso shows that Figaro sensors are more correlated among themthan Capteur sensors, at least for this application.

A PCA projection, a scores plot in the first and secondprincipal components, seems to indicate that the classes (con-taminated/clean) are no linearly separable, non-gaussian andmulti-modal (Fig. 4). Multi-modality clearly arises from thepresence of different substrate materials. In fact, visual explo-ration of the results shows that the influence of the substrate ismuch larger than the influence of the presence of moulds. Fig. 5shows the distribution of the four substrates on the same PCAprojection than Fig. 4. The distribution of materials seems tobe more separable than the contaminated and clean classes in

Fig. 5. Same PCA projection as in Fig. 4. Samples from first set of all mould.Each color represents different substrate; plasterboard, particleboard, orientedstrain board, wallpaper.

Fig. 6. Same PCA projection as in Fig. 4 regarding the four different species ofmoulds on first dataset of every mould.

Fig. 4. Wallpaper and particleboard are mixed between thembut separated from oriented strain board and plasterboard. Dueto the important background variability the detection of mouldsbecomes a difficult problem.

In Fig. 6 the distribution of species of moulds on the PCAprojection is observed, the four types of moulds are completelymixed so we can think that the classifications rates of each speciewill be similar. This can be seen in Table 2, where the obtainedclassification rates for every type of mould are shown. All resultsare on the same order, with small differences. From these resultswe can also think that the detection of a specie of mould notpresent in the training set is possible.

To study the stability of the classifier a second experimentfor each specie of mould was performed in the same conditionsas the first one resulting in a second dataset. This experimentstarted 4 months after the first one and was carried out dur-ing more than 3 months. The Mahalanobis classifiers built withfirsts datasets were used to classify the samples from the sec-ond datasets. Table 2 shows the obtained classification rates forevery type of mould and all moulds together. These results show

Table 2Classification rates for every specie of mould and all moulds together, on firstand second datasets

Type of mould First set (%) Second set (%)

All moulds together 84 ± 1 81P. chrysogenum 80 ± 3 76A. versicolor 84 ± 2 77C. sphaerospermum 80 ± 7 55P. aurantiogriseum 83 ± 5 65

In ‘first set’ Mahalanobis model was built with half of the first dataset andvalidated with the other half of first dataset, this was repeated 10 times mak-ing random subsampling. In ‘second set’ the model was constructed with thecompleted first dataset and validated with the second dataset.


Table 3Classification of a specie of mould not considered in the training samples

Train samples (three species of moulds) Validation samples C.R. (%)

A. versicolor + C. sphaerospermum + P. aurantiogriseum P. chrysogenum 78C. sphaerospermum + P. aurantiogriseum + P. chrysogenum A. versicolor 86P. aurantiogriseum + A. versicolor + P. chrysogenum C. sphaerospermum 76A. versicolor + C. sphaerospermum + P. chrysogenum P. aurantiogriseum 73

in every case a decrease in the classification ability that can beattributed to the sensors drift. Sensors ageing or the change inthe average room temperature can be the main reasons of thesmall changes in the sensors responses. Fig. 7 shows a PCAprojection on the model made from first dataset on two firstprincipal components, for both datasets and for all moulds. Forevery type of mould the structure of the second dataset is similarto the one from first experiment, but a slight drift can be observed(Fig. 8).

One point to remark is that the most important decreaseappears in mould C. sphaerospermum, which is the specie thatpresents the highest standard deviation in the classification of itsfirst set. This mould presents a classification rate of only 55%,which means that this classifier is unable to differentiate mouldsand clean materials under drift conditions. C. sphaerospermumis especially sensible to sensors drift, therefore drift depends ondata.

The classifier ability to detect an unknown specie of mouldis shown in Table 3. This table shows the obtained results forclassification of firsts datasets of every mould. The detectionof a type of mould not present in the training set is possible,but it depends on the specie of mould being classified. Thealgorithm is able to detect A. versicolor a 13% better than P.aurantiogriseum, this is not a very big difference but it is desir-able, for having the best classifications results, to consider as

Fig. 7. PCA projection from first set of all moulds. Points are samples fromsecond set of all moulds, circles are samples from first set of all mould. PC1 andPC2 explains 82.27% and 10.88% of variance respectively.

Table 4Classification of moulds growing on substrates not considered in the trainingsamples

Train samples (three speciesof material)

Validation samples C.R. (%)

osb + cb + pb Wallpaper (wp) 67osb + cb + wp Plasterboard (pb) 69osb + wp + pb Chipboard (cb) 86wp + cb + pb Oriented strain board (osb) 78

many different species of moulds as possible in the design of theclassifier.

The larger influence of the substrate materials on the detectionof moulds can be seen in Table 4. This table shows the classi-fication of samples from first sets containing building materialsnot present in the training set. Classification rate of wallpapersamples is 67% whereas the one for chipboard is a 19% bigger.This difference between the best and the worst classified samplesis bigger in the case of detection of substrate materials than inthe case of detection of different species of moulds. Therefore,to design a good classifier is more important to have a trainingset that contains many growing substrates than many types ofmoulds.

Table 5 shows how the substrate materials influence the detec-tion of moulds. A big difference between the classification ratesof samples containing only wallpaper (74%) and only chipboard(89%) can be appreciated, while differences in classification ratefor the various types of fungi are very small (4%). Drift alsoaffects the classification results obtained 4 months after depend-ing on the building material.

Although we expect that the classification rates would beimproved in case of considering only one specie of mouldand one type of substrate, due to the lack of samples in oursmall database this cannot be demonstrated. Having only a fewsamples in each case causes problems with the ‘course of dimen-sionality’.

Table 5Classification of moulds according to growing substrates

Type of material First set (%) Second set (%)

Wallpaper 74 ± 6 72Plasterboard 83 ± 4 71Chipboard 89 ± 3 82Oriented strain board 85 ± 4 72

Samples from all species of moulds.


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%, PC211.51%.

5. Conclusions

In this work an electronic nose has been able to detect fourspecies of moulds growing on several building materials witha Mahalanobis classifier, obtaining an accuracy between 80and 85%. The use of other more complexes classifiers gave nomuch better results, possibly this is due to the small size of ourdatabase.

The system is able to detect unknown species of mould withaccuracy between 73% and 86%, depending on the specie ofmould. It can also detect contaminated samples of unknownbuilding materials, depending also on the type of material. Theobtained minimum accuracy was 67% for wallpaper samples andthe maximum was 86% for chipboard. This accuracy is worsethan the one for the detection of unknown moulds because, asseen in Fig. 5, the influence of the substrate materials is biggerthan the influence of the different species of moulds.

The stability of the designed classifier was tested with anexperiment made in the same conditions but 4 months later,in which a second set of data was collected. The observed drift

effect on the sensors depends on the specie of mould that is beingclassified; A. versicolor has a classification rate of 77% while C.sphaerospermum has 55%. However the obtained result for theset made from all moulds was 81%, so, in general, the sensorsdrift causes a small degradation without any further recalibrationof the system.

However, due to the modest values of classification rate weconsider that such a system could only be used for screeningpurposes followed for a more detailed analysis using alternativeanalytical techniques.

Acknowledgements

Spanish team has been funded by Cycit DPI2001-3213-C02-01 and Generalitat de Catalunya ACI2003-13. Belgian team hasbeen funded by Region Wallone. One of the authors R.R. alsoacknowledges a PhD grant from the Spanish Ministry of Scienceand Technology. The assistance of CERTECH is also acknowl-edged.


References

[1] B. Brunekreff, A.P. Verhoeff, R.T. van Strien, J.H. Wijnen, Home dampnessand childhood respiratory symptoms: the role of sensitization to dust mitesand moulds, in: Samson, et al. (Eds.), Health Implications of Fungi inIndoor Environments, 1994, pp. 189–199.

[2] E. Johanning, R. Biagini, D. Hull, P. Morey, B. Jarvis, P. Landsber-gis, Health and immunology study following exposure to toxigenic fungi(Stachybotrys chartarum) in a water-damaged office environment, Int.Arch. Occup. Environ. Health 68 (1996) 207–218.

[3] K. Klanova, The concentrations of mixed populations of fungi in indoor air:rooms with and without mould problems; rooms with and without healthcomplaints, Central Eur. J. Public Health 8 (2000) 59–61.

[4] O. Koskinen, T. Husman, A. Hyvarinen, T. Reponen, A. Nevalainen, Twomoldy day-care centers: a follow-up study of respiratory symptoms andinfections, Indoor Air 7 (1997) 67–76.

[5] D. Norback, G. Wieslander, K. Nordstrom, R. Walinder, Asthma symp-toms in relation to measured building dampness in upper concrete floorconstruction, and 2-ethyl-1-hexanol in indoor air, Int. J. Tuberculosis LungDis. 4 (2000) 1016–1025.

[6] R. Walinder, D. Norback, G. Johanson, Pulmonary reactions after exposureto 3-methylfuran vapour, a fungal metabolite, Int. J. Tuberculosis Lung Dis.2 (1998) 1037–1039.

[7] G.H. Wang, C.S. Li, Indoor endotoxin and glucan in association with airwayinflammation and systemic symptoms, Arch. Environ. Health 54 (1999)172–179.

[8] C.E. Wilson, Sudden infant death syndrome and Canadian aboriginals: bac-teria and infections, FEMS Immunol. Med. Microbiol. 25 (1999) 221–226.

[9] G. Fischer, T. Muller, R. Schwalbe, R. Ostrowski, W. Dott, Exposure toairborne fungi, MVOC and mycotoxins in biowaste-handling facilities, Int.J. Hygiene Environ. Health 203 (2000) 97–104.

[10] H. Kiviranta, A. Tuomainen, M. Reiman, S. Laitinen, J. Liesivuori, A.Nevalainen, Qualitative identification of volatile metabolites from twofungi and three bacteria species cultivated on two media, Central Eur. J.Public Health 6 (1998) 296–299.

[11] A.L. Sunesson, C.A. Nilsson, B. Andersson, G. Blomquist, Volatilemetabolites produced by two fungal species cultivated on building materi-als, Ann. Occup. Hygiene 40 (1996) 397–410.

[12] J. Bjurman, J. Kristensson, Volatile production by Aspergillus versicoloras a possible cause of odor in houses affected by fungi, Mycopathologia118 (1992) 173–178.

[13] T. Borjesson, U. Stollman, J. Schnurer, Volatile metabolites and other indi-cators of Penicillium aurantiogriseum growth on different substrates, Appl.Environ. Microbiol. 56 (1990) 3705–3710.

[14] K. Fiedler, E. Schutz, S. Geh, Detection of microbial volatile organic com-pounds (MVOCs) produced by moulds on various materials, Int. J. HygieneEnviron. Health 204 (2001) 2–3.

[15] W. Horner, A. Worthan, B. Epstein, M. Black, MVOC emission pat-terns of indoor fungi: strain and culture differences, 1999, Onlinehttp://www.aiha.org/abs99/9pos3.html.

[16] A. Korpi, A.-L. Pasanen, H. Viitanen, Volatile metabolites of Serpula lacry-mans, Coniophora puteana, Poria placenta, Stachybotrys chartarum andChaetomium globosum, Build. Environ. 34 (1999) 205–211.

[17] L.J. Rose, R.B. Simmons, S.A. Crow, D.G. Ahearn, Volatile organic com-pounds associated with microbial growth in automobile air conditioningsystems, Curr. Microbiol. 41 (2000) 206–209.

[18] A. Korpi, A.-L. Pasanen, P. Pasanen, Volatile compounds originating frommixed microbial cultures on building materials under various humidityconditions, Appl. Environ. Microbiol. 64 (1998) 2914–2919.

[19] M. Kuske, R. Rubio, A.C. Romain, J. Nicolas, S. Marco, Fuzzy k-NNapplied to moulds detection, Sens. Actuators B: Chem. 106 (1) (2005)52–60.

[20] S.S. Schiffman, D.W. Wyrick, G.A. Payne, G. O’Brian, H.T. Nagle, Effec-tiveness of an electronic nose for monitoring bacterial and fungal growth,in: Proceedings of the ISOEN, Brighton, July 20–24, 2000, pp. 173–180.

[21] O. Canhoto, F. Pinzari, C. Fanelli, N. Magan, Application of electronicnose technology for the detection of fungal contamination in library paper,Int. Biodeterior. Biodegrad. 54 (4) (2004) 303–309.

[22] J. Schnurer, J. Olsson, T. Borjesson, Fungal volatiles as indicators offood and feeds spoilage, Fungal Genet. Biol. 27 (2/3) (1999) 209–217.

[23] J. Olsson, T. Borjesson, T. Lundstedt, J. Schnurer, Detection and quantifi-cation of ochratoxin A and deoxynivalenol in barley grains by GC–MSand electronic nose, Int. J. Food Microbiol. 72 (3/5) (2002) 203–214.

[24] J. Olsson, T. Borjesson, T. Lundstedt, J. Schnurer, Volatiles for myco-logical quality grading of barley grains: determinations using gaschromatography–mass spectrometry and electronic nose, Int. J. FoodMicrobiol. 59 (2000) 167–178.

[25] A. Jonnson, F. Winquist, J. Schnurer, H. Sundgren, I. Lundstrom, Electronicnose for microbial quality classification of grains, Int. J. Food Microbiol.35 (1997) 187–193.

[26] G. Keshri, N. Magan, P. Voysey, Use of an electronic nose for the earlydetection and differentiation between spoilage fungi, Lett. Appl. Microbiol.27 (1998) 261–264.

[27] G. Keshri, N. Magan, Detection and differentiation between mycotoxi-genic and non-mycotoxigenic strains of two Fusarium spp. using volatileproduction profiles and hydrolytic enzymes, J. Appl. Microbiol. 89 (2000)825–833.

[28] K.C. Persaud, P.D. Wareham, P.A. Payne, H. Chueh, J.V. Hatfield, Aportable electronic nose for dry rot detection, in: Actes des conferences Lametrologie des odeurs et les nez electroniques, Eurodeur–Airodeur, Paris,2001.

[29] N. Canhoto, Magan, Electronic nose technology for the detection of micro-bial and chemical contamination of potable water, Sens. Actuators B:Chem. 106 (1) (2005) 3–6.

Biographies

Martyna Kuske is Doctor of medicine from the University of Warsaw (1990).She received the master in Environmental Sciences in 1999 and the DEA fromFUL in 2000. Since 2000 she works on mould detection in the frame of a Ph.D.thesis at FUL, which is now the department “Environmental Sciences and Man-agement” of the University of Liege. She is also involved in the diagnosis of“Indoor Air Pollution” in the frame of a contract with the official local serviceof intervention of the Belgian province of Luxembourg.

Marta Padilla was born in Sta. Cruz de Tenerife, Spain, on 1973. She receiveda degree in Physics from the University of La Laguna in 1999 and a degree inElectronic Engineering from University of Barcelona in 2003. In the same yearshe joined the ISPlab as a Ph.D. student.

Anne-Claude Romain graduated in chemical sciences from the Universityof Liege (Ulg, Belgium) in 1992. She received the master in Environmen-tal Sciences from the Fondation Universitaire Luxembourgeoise of Arlon(FUL, Belgium) in 1993. Since 1995, she is a searcher at FUL, whichbecame in 2004 the department “Environmental Sciences and Management”of the University of Liege. She has been working on the development ofa malodour detector. Currently, she finishes a Ph.D. thesis on the appli-cation of the electronic nose concept to the monitoring of odours in theenvironment.

Jacques Nicolas is Engineer in Physics. He received his Ph.D. degree in 1977 inSurface Physics in University of Louvain, Belgium. He joined Fondation Univer-sitaire Luxembourgeoise (FUL, Arlon, Belgium) in 1979, where he worked firston Solar Energy. He is currently the leader of the research group “EnvironmentalMonitoring” in the department “Environmental Sciences and Management” ofthe University of Liege. He gives lectures in the field of environmental param-eter measurement. His main research interest is the development of odour andindoor air pollution detectors usable in the field.

Rafael Rubio was born in Madrid, Spain, on 1978. He received his B.S. anddegree in physics from the University of Barcelona in 2001. In 2002 he joinedthe ISPlab as a Ph.D. student. His areas of interest are the signal processingapplied to gas sensors, MEMS and IR gas detection.


Santiago Marco Colas was born in Berga in 1965. Degree (1988) and Doc-tor in Physics (1993) at the University of Barcelona doing research on pres-sure sensors for biomedical uses which took place at the Centre Nacionalde Microelectronica. In 1994 he was post-doc at the University of Rome

“Tor Vergata” working on intelligent signal processing. Since 1995 he’san Associate Professor at the UB. His experience is centered in microsys-tem modelisation, and in signal processing-based intelligent instrumenta-tion.


Feature extraction on three way enose signals

M. Padilla ∗, I. Montoliu, A. Pardo, A. Perera, S. MarcoUniversitat de Barcelona, Departament d’Electronica, Sistemes d’Instrumentacio i Comunicacions,

Martı i Franques, 1, 08034 Barcelona, Spain

Received 10 July 2005

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 gettingmore reliable information, the proposal of incorporating the whole transitories of each gas sensor simultaneously recorded, to build a three-waydataset seems to be a good option. But, anyway, this strategy must be accompanied by suitable signal processing/feature extraction of the data inorder 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 3Ddata arrays are shown. To exemplify its performance, a quantitative case focused on food analysis has been selected. The results obtained point outthe suitability of the technique to achieve a good predictive ability by using a simple inverse least squares (ILS) calibration onto a set of syntheticsamples.© 2006 Elsevier B.V. All rights reserved.

Keywords: PARAFAC; ILS; Preprocessing; Feature extraction

1. Introduction

Thanks to electronic advances, modern instrumentationallows to measure complex magnitudes but, on the other hand,measurements and datasets have also grown in complexity.Under this paradigm, the extraction of embedded information inthe dataset demands advanced signal processing, and this is thecase of smart gas sensors arrays (also called electronic noses).It is true that the use of this kind of instrumentation for indus-trial process control is increasing but for an efficient instrumentsetting and calibration, suitable signal processing tasks must beimplemented. The extraction from a representative dataset of therelevant embedded information is one of these basic tasks.

Different methods, usually grouped under the general nameof feature extraction/selection techniques, can be used. Somepapers present simply methods based on linear algorithms asprincipal component analysis (PCA) or linear discriminant anal-ysis (LDA) [1–3]. Other simply approaches are related withthe use feature selection methods as stepwise forward search(SFS), stepwise backward search (SBS), advanced search based

∗ Corresponding author. Tel.: +34 934039148; fax: +34 034021148.E-mail address: [email protected] (M. Padilla).

on genetic algorithms [3–5] or, if the number of variables aresmall, an exhaustive search can be used.

Additionally, there have been efforts to extract useful infor-mation from the whole signal profile because dynamic signalsseem to contain useful information [6–9]. The signal profileobtained from gas sensors is strongly dependent of the generickind of volatiles to which they are exposed, the duration of thisexposition and the temperature of the system. In some cases, itcan be suitable to increase the amount of data in order to captureas much information as possible which properly describes thesystem, if more reliable models are desired.

One alternative to do this goes through the use of the three-way structure of the signal obtained from the recording of thedifferent sensor values during a period of time, for each sample.Following this strategy, the contribution of the inner relationshipbetween the three ways can be exploited to obtain more robustinformation about the system. However, when datasets are bigand complex, the computation time cost could be high and thereis a risk to obtain a performance decrease due to the curse ofdimensionality problem [3,10]. For this reason, more elaboratedalgorithms specific to multi-way data analysis must be used inorder to avoid or minimize these contingencies.

Parallel factor analysis (PARAFAC) is one of the most pop-ular multi-way data decomposition methods. Originated from


146 M. Padilla et al. / Sensors and Actuators B 116 (2006) 145–150

psychometrics [11,12] PARAFAC is gaining interest becauseit is a processing technique that simultaneously determinesthe pure contributions to the dataset, optimizing each factoras a time, in trilinear systems. Widely explored in differentfields, such spectrofluorimetry, spectrophotometry, chromatog-raphy and so on, it can deal with the increased complexity ofprocess control data in science and industry [13,14].

Towards other related decomposition methods, such Tucker3[15] and unfolded PCA [16,17], PARAFAC belongs to thesame family of multi- or bilinear methods of decompositionof multi-way data in a set of loading and score matrices. Thisprocess allows to describe the initial variability decreasing itsdimensionality, thus leading to a compressed form. In this pro-cess, PARAFAC uses less degree of freedom than Tucker3 orunfolded PCA methods. This intrinsic characteristic leads tosimpler models with less fit, but avoiding the incorporation ofnon-significative effects such noise and redundant informationto the model [18].

This paper presents an illustration of the use of PARAFACfor feature extraction on a multi-way dataset obtained from anelectronic nose. To check its performance, a quantization casefor quality control of processed food products is studied.

2. Case

2.1. Presentation

The strong competition in the snacks industry has leaded tothe companies to diversify their offer as well as possible. In thisway, the companies are offering variants of traditional productswith different tastes. Thus, at the present time, fried potato chipsbeing prepared with many tastes (‘ham’, ‘spicy’, ‘barbecue’,etc.)

The new products are often achieved through the additionof alimentary flavours with different formulations, which incor-porate salt in a certain number of cases, and are part of theknow-how of these companies and/or their suppliers. The addi-tion process of these flavours is of utmost importance in order toobtain a product of quality, with enough homogeneity betweenbatches. In spite of the importance of this phase, does not exista direct control of the addition of the alimentary flavour. At thepresent time, the process is only indirectly controlled in line byproperly weighting of the additive added, without any real con-trol of the taste of the final product. Although off line analysisare done, through salt control in the laboratory of selected pro-duction samples, this is only an indirect measure of the flavouraddition, and valid only if the flavouring agent contains salt.

Electronic olfaction appears as a possible alternative to quan-tify the addition of this flavouring agent, starting up from theunspecific measure of the detached volatiles. This kind of mea-surement presents the main advantage of its direct relationshipwith the magnitude to be determined (intensity of flavour/odour)and provide some possibilities to its continuous implementation,even tough it will come joined with appropriate modificationsof actual instrumentation. These systems are often composed ofa suitable sampling system, a measurement cell (containing thesensor array), a pumping device (responsible of introducing the

sample into the cell and further purging of the system), and allthe fluidics and valves necessary to properly deliver the differentgas flows. All this system is controlled by embedded electronics,which in addition is in charge of the first signal treatment.

2.2. Experimental

Thus, starting up from raw fried potato chips and a littleamount of flavouring agent submitted by a Spanish food industry,a set of 30 synthetic samples was prepared in different ses-sions. These samples were obtained by weighting of 1.00 g offried potato chips and further addition of different amounts offlavouring agent, in order to span the percent of additive desired:between 1% and 8% in weight. After proper weighting of theflavouring agent, the samples were tight closed in a glass vialand properly stored.

The measurement of the headspace of the vials wasdone using a Nordic Sensor Technologies System, NST3320(Linkoping, Sweden) with its accessories. This instrument con-tains a gas detector array of 23 sensors, divided in 10 MOSFETcatalytic field-effect sensors (not used in this work) and 13 MOXsensors. This unit is capable to deliver the headspace of thedifferent vials into the sensors chamber, after properly thermo-statization of the sample, and to perform appropriate cleaningoperations of the measurement chamber.

In this work, the measures were done keeping a con-stant temperature for all the samples of 40 ◦C. After half anhour of thermal stabilization, the data were acquired follow-ing a baseline–injection–recovery–flush–stabilization cycle of15–10–10–120–110 s each, employing a sampling rate of 1 Hz.The data thus obtained, corresponding to an increase in the con-ductance value of the sensors, was properly stored for each ofthe 13 MOX sensors and folded for each sample, in order togenerate a three-way matrix X of dimensions (30 × 13 × 265)(samples × sensors × time), containing the described sintheticsamples and its complete transitories. In a second step thisdatabase was split in two, to obtain two differentiated subsetsto be further processed. Both subsets were prepared spanninghomogeneously the same concentration intervals, and recordedin different sessions. The number of samples of each of the twosubsets (calibration and validation) was of 20 and 10 samples,respectively.

In Fig. 1 can be seen the plot of the raw signal for sample 1of the calibration subset. Data have been accordingly resized toavoid a confussing view. A transitory signal is obtained, present-ing different responses for each of the 13 MOX sensors, withdifferent behaviours along the measurement time.

3. Results

To test the applicabilty of the PARAFAC method as a featureextraction method from three-way datasets, the first step donewas to check the inner trilinearity of the data. With this aim,the calibration subset was properly row-wise (20 × (265 × 13)),column-wise ((20 × 265) × 13) and tube-wise ((20 × 13) × 265)unfolded to a two way array. If there exists trilinearity on the data,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-bration set.

be nearly the same [19]. After SVD calculation on the centereddata, they appeared to be the same (five) for each augmentedarray (see Table 1), thus confirming the trilinearity of the dataset.

Taking the calibration subset as a basis, and before PARAFACcompression were applied, several signal preprocessing tech-niques were tested. At last, among others (smoothing, deriva-tives), an offset correction based in the use of the average ofthe first 15 min of the transitories of each sensor was applied toremove constant effects from run to run. Following this first step,standard nomal variate (SNV) [20–22] was also applied to cor-rect the dispersion of the data. Both pretreatments were appliedto the third mode (time), and were the ones which offered morepromising results. Because processing of the datasets along oneor more modes can break its inner trilinearity condition, it mustbe ensured through a new check after the preprocessing step. Inthis case, the trilinearity of the dataset was re-evaluated render-ing results that can be assumed as sufficiently positive. Once thepretreatment was applied and to each variable slice, and veri-fied the trilinearity of the dataset, three-way data were centeredacross the first mode (samples) before the testing of PARAFACmodels.

Fig. 2. Determination of the number of factors. Core consistency test results ofeach of the three replicates calculated for each factor.

To do this test, starting up from centered data, an explorationof the different possible PARAFAC models was done, varyingthe number of factors and repeating the procedure by triplicate,in order to check the stability of the results. After using thecore consistency diagnostic [23] as an indicator of the successof the different models prepared (see Fig. 2), it was found thatthe number of factors which better describe the data should bethree.

Once this three factor model was selected, and after plottingthe different loadings (scores and second and third mode ones),interesting tendences are observed across the scores (loadingsof the first mode), as it can be seen in Fig. 3. There appearedthree factors which shown increasing behaviours as the sampleconcentration increased. These increases describe, but, differentprofiles, depending on the sample involved.

Second loading matrix plot show the contribution of eachsensor to the global signal. From this plot is possible to deter-mine which are the sensors which are contributing with moreinformation to each factor. As it can also be seen, these differ-ences become milder as the number of factors increases. In thecase of the third loading plot, it shows important differences

Table 1Trilinearity check

Eigenvalues of augmented data matrices

Before correction After correction

Column-wise Row-wise Tube-wise Column-wise Row-wise Tube-wise

1.0000 1.0000 1.0000 1.0000 1.0000 1.00000.5704 0.2329 0.3019 0.5931 0.6020 0.36700.4342 0.1621 0.1175 0.4437 0.3502 0.27600.3815 0.1339 0.0549 0.2827 0.2270 0.12790.3246 0.0490 0.0325 0.1820 0.1543 0.06920.2925 0.0270 0.0153 0.1541 0.1001 0.05990.2704 0.0244 0.0134 0.1292 0.0862 0.05420.2484 0.0137 0.0083 0.1066 0.0799 0.02120.2325 0.0104 0.0047 0.0911 0.0712 0.01690.2228 0.0085 0.0028 0.0496 0.0619 0.0095

Eigenvalues of the calibration dataset before and after preprocessing.


Fig. 3. Scores and loadings plot for the three factors PARAFAC model for eachof the three ways: samples (A), sensors (B) and recording time (C).

in the part of the time plot corresponding to the gas injection,comprised between time 15 and 25 min. There appear three par-tially overlapped factors, with different profiles, which seem tobe showing the averaged different contributions of the sensorsalong the time, and shortly after its contact with the blend ofvolatile analytes. Once this sample contact is finished, data arestill rendering information. There appear also differences in each

factor transitories looking to be important enough to play a roleinto the model.

After the calculation of the correlation coefficients betweeneach scores vector with the concentration of the calibrationdataset, and after its study, there were observed differencesbetween factor 3 (R = 0.8697) and factors 1 and 2 (R1 = −0.3564and R2 = −0.5184). Because of the differences in correlation forthe different factors towards the concentration, and the insuffi-cient correlation towards the third factor, a second order cal-ibration was initially discarded. For this reason, to obtain aquantitative model it was necessary to look for an alternative.Among the available ones, and trying to keep the model as sim-ple as possible, the use of the multiple linear regression toolwas taken as a basis, following a similar strategy to the one fol-lowed by principal components regression (PCR) [24], a wellknown regression method. Thus, an inverse least squares (ILS)[23] model was built, but taking the PARAFAC scores matrix asthe predictor variables matrix, and the concentration, expressedas grams of added flavour, as the response variable.

The results obtained provide a root mean square error of cal-ibration (RMSEC), as is expressed in Eq. (1), of 0.007. Thiscalibration error is almost the same as the experimental one,

Fig. 4. Quantitative results: (A) reference values vs. PARAFAC-ILS calculatedvalues and (B) residuals plot.

M. Padilla et al. / Sensors and Actuators B 116 (2006) 145–150 149

Table 2ILS quantitative results of PARAFAC processed data

Calibration Validation

Reference (g) Calculated (g) Reference (g) Calculated (g)

0.000 0.006 0.000 0.0090.041 0.044 0.042 0.0340.046 0.052 0.047 0.0490.053 0.060 0.050 0.0490.056 0.067 0.055 0.0580.061 0.060 0.060 0.0440.070 0.074 0.065 0.0590.072 0.063 0.071 0.0620.079 0.061 0.030 0.0280.006 −0.001 0.036 0.0470.012 0.0070.015 0.0200.019 0.0230.024 0.0300.029 0.0330.038 0.0350.039 0.0390.042 0.0440.047 0.0390.057 0.052

RMSEC 0.007 RMSEP 0.008

thus confirming the precision of the method. To check the accu-racy, a method comparison regression line (y = a + bx) betweenthe calculated and the reference values was calculated. The

RMSE =√∑I

i=1(Xi − Xi)2

I(1)

obtained results for the calibration set leaded to a linear equation:y = 0.004 + 0.902x. To define the suitability of this comparison,the confidence intervals for this line were checked to ensure thesolution did not show any kind of systematic deviation. Confi-dence intervals obtained in this case for each parameter were fora ±0.007 and b ±0.147. Both intervals confirm that comparedmethods (weighting of flavouring agent and ILS calibration onPARAFAC scores from enose data) leaded to similar results(Fig. 4).

To check the predictive ability of the model built, a validationstep using the second subset was performed. In this way, thefirst necessary point was to obtain the PARAFAC scores of thissecond subset. This was done, once it was properly centeredand corrected, through its projection onto the PARAFAC model.This scores estimation was done calculating the pseudoinverseof Khatri–Rhao product of both loadings matrices (C and B),multiplied by the vectorized form of the slice corresponding tothe first object to be processed [25] (Eq. (2)).

a = (C ⊗ B)+ vecXnew (2)

Once the scores of the validation set were obtained, its flavouringagent amount values were obtained using the former regressioncoefficients obtained from the ILS calibration. The results fromboth datasets, calibration and validation, are shown in Table 2.As it can be seen, properly validation results were also obtained

for this dataset, rendering a root mean square error of prediction(RMSEP) of 0.008.

After obtaining this results, which seem to be fairly good,a residuals plot was done. As it can be seen in Fig. 3, thedistribution presents an homogeneous pattern across the cali-bration range. The overall predictive results obtained (RMSECand RMSEP) agree with the absolute error (A.E. = ±0.01g) inthe determination of weights.

4. Conclusions

In this work it has been studied the use of PARAFAC as afeature extraction technique to perform a data reduction froma specific three-way dataset. Previously to this extraction, asuitable data pretreatment have been applied to obtain properlyquantitative results for our experimental case are desired. Eventhough SNV is known as a processing method in other fieldsprone to introduce nonlinearities in the datasets, it has revealeditself to be the most suitable one, jointly applied with offset cor-rection. Its application in this case has been effective, keepingto a certain extent the trilinearity of the processed dataset.

Taking the quantitative case results as indicators of the perfor-mance of this use of PARAFAC, the predictive errors providedacceptable RMSE values, taking in mind the sensitivity of tech-nique, and the amounts range. In this sense, PARAFAC performsa suitable data compression task, incorporating the most of theinformation of this complex data system. Although this work isfocused in a quantitative preliminary approach, further work inthis direction can be done on the application of PARAFAC asa feature extraction method for its use for pattern recognitionpurposes and on its application to more complex samples.

References

[1] E. Oja, Subspace Methods of Pattern Recognition, Research StudiesPress, Letchworth, Hertfordshire UK, 1983.

[2] E. Oja, J. Karhunen, Recursive construction of Karhunen–Loeve expan-sions for pattern recognition purposes, in: Proceedings of the FifthInternational Conference on Pattern Recognition, December, Springer-Verlag, NY, 1980, pp. 1215–1218.

[3] A. Pardo, S. Marco, C. Calaza, A. Ortega, A. Perera, T. Sundic, J.Samitier, Methods for sensors selection in pattern recognition, Proc.ISOEN (2000).

[4] T. Eklov, P. Martensson, I. Lundstrom, Selection of variables for inter-preting multivariate gas sensor data, Anal. Chim. Acta 381 (2–3) (1999)221–232.

[5] T. Sundic, S. Marco, A. Perera, A. Pardo, S. Hahn, N. Barsan, U.Weimar, Fuzzy inference system for sensor array calibration: predictionof CO and CH4 levels in variable humidity conditions, Chemomet-rics and Intelligent Laboratory Systems, October 2002. Corrected Proof,Available online 11 October 2002.

[6] E. Llobet, J. Brezmes, X. Vilanova, X. Correig, J.E. Sueiras, Qualitativeand quantitative analysis of volatile organic compounds using transientand steady-state responses of a thick- lm tin oxide gas sensor array,Sens. Actuators B, Chem. 41 (1997) 13–21.

[7] Y. Hiranaka, T. Abe, H. Murata, Gas dependant response in the temper-ature transient of WGXGYZ gas sensors, Sens. Actuators B, Chem. 9(3) (1992) 177.

[8] J. Samitier, J.M. Lopez-Villegas, S. Marco, L. Camara, A. Pardo, O.Ruiz, J.R. Morante, A new method to analyse signal transients in chem-ical sensors, Sens. Actuators B, Chem. (1994).


[9] R. Gutierrez-Osuna, H.T. Nagle, S.S. Schiffman,.Transient responseanalysis of an electronic nose using multi-exponential models, Sens.Actuators B, Chem. 61 (December) (1999) 170–182.

[10] A. Perera, T. Sundic, A. Pardo, S. Marco, Feature extraction by sen-sor waveform variance analysis, in: Proceedings of Ninth InternationalSymposium on Olfaction and Electronic Nose, Roma, 2002.

[11] R.A. Harshman, Foundations of the PARAFAC procedure: model andconditions for an ‘explanatory’ multi-mode factor analysis, UCLA Work-ing Papers in Phonetics 16 (1970) 1.

[12] J.D. Carroll, I. Chang, Analysis of individual differences in multidi-mensional scaling via an N-way generalization of and Eckart-Youngdecomposition, Psychometrika 35 (1970) 283.

[13] P. Geladi, Analysis of multi-way (multi-mode) data, Chemom. Intell.Lab. Syst. 7 (1989) 11.

[14] A.K. Smilde, Three-way analyses. Problems and prospects, Chemom.Intell. Lab. Syst. 5 (1992) 143.

[15] P.M. Kroonenburg, Three-mode principal component analysis, in: Theoryand Applications, DSWO Press, Leiden, 1983.

[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, Acomparison of principal components analysis, multi-way principal com-ponents analysis, trilinear descomposition and parallel factor analysis forfault detection in semiconductor etch process, J. Chemom. 13 (1999)379–396.

[18] Rasmus Bro. PARAFAC: tutorial and aplications. Chemom, Intell. Lab.Syst. 38 (1997) 149–171.

[19] A. de Juan, R. Tauler, Comparison of three-way resolution methods fornon-trilinear chemical data sets, J. Chemom. 15 (2001) 749–772.

[20] R.J. Barnes, M.S. Dhanoa, S.J. Lister, Standard normal variate transfor-mation and detrending of near-infrared diffuse reflectance spectra, Appl.Spectrosc. 43 (1989) 772–777.

[21] R.J. Barnes, M.S. Dhanoa, S.J. Lister, Correction to the description ofstandard normal variate (SNV) and de-trend (DT) transformations inpractical spectroscopy with applications in food and beverage analysis,second ed., J. Near Infrared Spectrosc. 1 (1993) 185–186.

[22] M.S. Dhanoa, S.J. Lister, R. Sanderson, R.J. Barnes, The link betweenmultiplicative scatter correction (MSC) and standard normal variate(SNV) transformations of NIR spectra, J. Near Infrared Spectrosc. 2(1994) 43–47.

[23] R. Bro, H.A.L. Kiers, A new efficient method for determining thenumber of components in PARAFAC models, J. Chemom. 17 (2003)274–286.

[24] B.G.M. Vandeginste, D.L. Massart, L.M.C. Buydens, S. De Jong, P.J.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 theChemical Sciences, Wiley, Chichester, 2004.

Biographies

M. Padilla was born in Sta. Cruz de Tenerife, Spain,on 1973. She received a degree in Physics from theUniversity of La Laguna in 1999 and a degree inElectronic Engineering from University of Barcelonain 2003. In the same year she joined the ISPlab as aPhD student.

I. Montoliu was born in Barcelona in 1969. Degree(1995) and PhD in Chemistry (2001) at the Uni-versitat Autonoma de Barcelona, doing research indevelopment and application of chemometrical meth-ods to agroalimentary analysis. In 2000 joins theofficial distributor of Dionex products in Spain, aschromatography applications and software special-ist. Since 2004 is staying as Post Doctoral ResearchAssociate Universitat de Barcelona. His areas of inter-est are Chemometrics and its application to signalsof different analytical instrumentation.

A. Pardo was born in 1967 in Barcelona. Receivedhis deegree in physics 1991 and his PhD in 2000 fromthe University of Barcelona. His research is focusedin system identification with applications in gas sen-sor systems, signal processing for gas sensors andpattern recognition.

A. Perera received his BS (1996) and his PhD (2003)in physics from University of Barcelona and his BSin Electrical Engineering (2001) from the same uni-versity. He was with Texas A&M University in theperiod from 2003 to 2005 as PostDoctoral ResearchAssociate. His research interests include applied pat-tern recognition focusing chemical sensors, machineolfaction and bioinformatics.

S. Marco was born in Berga in 1965. Degree (1988)and Doctor in Physics (1993) at the University ofBarcelona doing research on pressure sensors forbiomedical uses which took place at the CentreNacional de Microelectronica. In 1994 he was post-doc at the University of Rome “Tor Vergata” workingon intelligent signal processing. Since 1995 he’s anAssociate Professor at the UB. His experience iscentered in microsystem modelisation, and in signalprocessing-based intelligent instrumentation.

Drift compensation of gas sensor array data by Orthogonal Signal Correction

M. Padilla a,b,⁎, A. Perera a,b, I. Montoliu a,b, A. Chaudry c, K. Persaud d, S. Marco a,b

a Departament d'Electrònica, Universitat de Barcelona, Marti i Franquès 1, 08028 Barcelona, Spainb Institut de Bioenginyeria de Catalunya (IBEC), Baldiri i Reixac 13, 08028 Barcelona, Spainc Protea Ltd, 11 Mallard Court, Mallard Way, Crewe Business Park, Crewe, Cheshire, CW1 6ZQ, UKd School of Chemical Engineering and Analytical Science, The University of Manchester, PO Box 88, Sackville St, Manchester, M60 1QD, UK

a b s t r a c ta r t i c l e i n f o

Article history:Received 6 July 2009Received in revised form 7 October 2009Accepted 9 October 2009Available online 7 November 2009

Keywords:Gas sensor arrayDriftOrthogonal Signal CorrectionComponent CorrectionCross-validationElectronic noseData shift

Drift is an important issue that impairs the reliability of gas sensing systems. Sensor aging, memory effectsand environmental disturbances produce shifts in sensor responses that make initial statistical models forgas or odor recognition useless after a relatively short period (typically few weeks). Frequent recalibrationsare needed to preserve system accuracy. However, when recalibrations involve numerous samples theybecome expensive and laborious. An interesting and lower cost alternative is drift counteraction by signalprocessing techniques. Orthogonal Signal Correction (OSC) is proposed for drift compensation in chemicalsensor arrays. The performance of OSC is also compared with Component Correction (CC). A simpleclassification algorithm has been employed for assessing the performance of the algorithms on a datasetcomposed by measurements of three analytes using an array of seventeen conductive polymer gas sensorsover a ten month period.

© 2009 Elsevier B.V. All rights reserved.

1. Introduction

Gas sensor arrays are potentially economic and faster alternativesfor gas analysis and aroma evaluation compared to conventionalanalytical instruments such as gas chromatographs. Many differenttechnologies based on different principles for gas sensing areavailable. These include Metal Oxide (MOX), Quartz Microbalances(QMBs), Conductive Polymers (CPs), Surface Acoustic Waves (SAWs),and electrochemical sensors among others. While solid-state sensorsoffer very big advantages in terms of miniaturization, response timesand power consumption, those sensors show poor selectivity. Sincethe seminal paper of Persaud and Dodd [1], it has been known thatimproved selectivity may be achieved by combining different sensorswith partial and overlapping sensitivities and pattern recognitionsystems.

While a large number of successful applications for gas sensorarrays have been published, these are largely laboratory based andpractical applications of chemical sensor arrays in real commercialapplications are limited. This may be attributed to a number ofreasons, such as insufficient sensitivity (or limit of detection too high),lack of selectivity, and other problems. However, from the authors'point of view, the main reason is the lack of stability over time and thecost of recalibration. This paper addresses the application of different

methods to improve the stability over time of sensor arrays from asignal and data processing perspective.

Gas sensor drift consists of a random temporal variation of thesensor response when it is exposed to the same analytes underidentical conditions. This problem is generally considered to be due tosensors aging [5], but it has also been attributed to a variety of sources,like environmental factors such as humidity variations [2], systemsampling non-specific adsorption [3,4], variations on flow rate,thermo-mechanical degradation and poisoning among others [6,7].All of these factors can modify both the baseline and the sensitivity ofthe sensors in the array in different ways, depending on sensortechnology.

Therefore, for operation over long periods, the ability of theinstrument to recognize analytes is degraded, since statistical modelsbuilt in the calibration phase become useless after a short period oftime, in some cases weeks or few months. After that time, theinstrument must be completely re-calibrated, which is a time-consuming, laborious and expensive task, to ensure that thepredictions remain valid. The working hypothesis in this work isthat system stability may be improved using proper multivariate dataprocessing techniques.

Several methods have been reported in the literature in order toimprove stability over time bymodifications in sensor technology anddesign [8–10] or by the use of different sensor operation modes[11,12]. On the other hand, signal processing methods for driftcounteraction are based on different approaches: univariate [13,14] ormultivariate [15–17], linear [18,19] or non-linear [15–17], adaptive

Chemometrics and Intelligent Laboratory Systems 100 (2010) 28–35

⁎ Corresponding author. Departament d'Electrònica, Universitat de Barcelona, Marti iFranquès 1, 08028 Barcelona, Spain.

E-mail address: [email protected] (M. Padilla).

0169-7439/$ – see front matter © 2009 Elsevier B.V. All rights reserved.doi:10.1016/j.chemolab.2009.10.002

Contents lists available at ScienceDirect

Chemometrics and Intelligent Laboratory Systems

j ourna l homepage: www.e lsev ie r.com/ locate /chemolab

[15,16] or not, based in reference samples [14] or based in componentremoval [18,19]. Univariate corrections include basic baseline correc-tions [13], or more complex per-sensor correction by means of acalibration sample [14]. Thesemethods, however, do not take profit ofcorrelated drift effects. For this purpose, other multivariate methodshave been proposed. Non-linear multivariate adaptive algorithms likeSelf OrganizingMaps (SOMs) [15,16], multiple SOM (mSOM) [17] andsystem identification theory [15,18] have also been proposed in thepast. However, fully unsupervised adaptive SOM methods haveproblems in case of overlapping classes, since the reference patternupdating may follow the wrong class. Techniques based on systemidentification theory predict sensors outputs by means of dynamicalmodels for every sensor from the rest of the sensors' response,assuming the sensors behave very similarly. Other adaptiveapproaches include the use of sliding windowwavelet decompositionbased feature extraction for drift detection and compensation,sometimes post-processed by Principal Component Analysis orFuzzy Adaptive Resonance Theory networks (ARTMAP) based algo-rithms [19–21]. Data processing by wavelets decomposes data intomultiple time scales. Since drift is a slow process, it is better capturedby the coarsest scales, while noise and other events appear in thefinest scales. Provided that the separation of drift from real responsesis feasible, this procedure improves posterior classification orregression tasks. Linear methods like Component Correction (CC)based on Principal Component Analysis (PCA) [22,23], or CanonicalCorrelation Analysis (CCA) and Partial Least Squares (PLS) [24] havealso been reported to provide good results. In particular, ComponentCorrection has received considerable attention in the community.However, Component Correction methods assume that all chemicalclasses behave in the same way in the presence of drift and this is notalways the case (as it will be seen in the present study).

Our primary goal is to demonstrate thatOrthogonal Signal Correction(OSC) is suitable for drift compensation. OSC is a linear technique thatremoves components orthogonal to a variable Y indicative of the datastructure, which is correlated to the data. Thismethod is a very commontool used in spectroscopy to correct spectra by removingsystematic non-relevant information, such as baseline variation, but it is not socommonly used in the field of chemical sensors, although we findsome examples in literature [25,26] where it is used for removing localvariance. This technique will be compared to Artursson's ComponentCorrection (CC) method. The effectiveness of the proposed algorithmswill be evaluated on an experimental dataset composed by measure-ments of several analytes using an array of conductive polymer sensors.The time duration of experiments has been 10 months.

A secondary goal of this paper is to study the impact of the trainingset size on the ability of the methods for drift compensation. To be ofpractical interest, drift counteraction should be effective with only areduced set of training samples, spanning a limited time duration.

A key point in this paper is the validation methodology. Verycommon validation techniques, like k-fold cross-validation, randomsubsampling, bootstrap and leave-one-out, which abound on theliterature, totally neglect the influence of drift. In those techniquesvalidation and training samples are interleaved along the time axis.Then future evolution of the sensor responses is modeled by theclassifier and consequently provides overoptimistic results. Of course,this is not representative at all of real operation conditions. The authorswould like to emphasize that those validation techniques should neverbe used in conditions where drift is present. Instead, in this workvalidation samples are always in the future of training samples.

2. Theory

2.1. Component correction

Artursson proposed in 2001 the Component Correction method(CC) [22]. It is a signal processing technique based on a Principal

Component Analysis (PCA) decomposition of a reference class datasubset that is later used to correct undesired variance from the rest ofthe dataset. This method assumes that the reference class isrepresentative of the entire population. Therefore variation found inthe reference class will also be present in the rest of the dataset. Thismeans that, if the variance in the reference class is due to drift effects,drift will be also removed from the complete dataset.

2.2. Orthogonal Signal Correction (OSC)

Wold et al. firstly introduced Orthogonal Signal Correction (OSC)for its use on NIR spectra correction [27]. The main idea is to removevariance not correlated to variables in a vector (or matrix) Y, whichcontains some extra data information. This is done by constraining thedeflation of non-relevant information of X, so that only informationorthogonal to Y should be removed. The inclusion of the condition oforthogonality to Y ensures that the signal correction removes as littleinformation as possible.

After Wold's paper a number of OSC-like algorithms that tried toimprove the original OSC method were published [28–34]. Compar-isons among them can be also found in literature [35,36]. In this work,the version applied is based on the Wise implementation of thealgorithm [37]. This first searches for a direction of maximum varianceof the data X using PCA. The scores vector t corresponding to this firstprincipal component is then orthogonalized with respect to theinformation matrix Y, in order to obtain a new scores vector nt notcorrelated with Y that captures the highest possible amount ofvariance of X. A Partial Least Squares (PLS) step between initial data Xand nt, with a suitable number of latent variables (LV), gives scores T1and loadings P1 vectors that contain the information not related to Y.For the rest of the paper, we will refer to this step as the inner PLS. Thenumber of inner PLS latent variables, it is calculated in thisimplementation from an specification of the variance explained inthe X-block. We refer to this as the OSC tolerance and it will be givenin percentage values. In a final step T1 is again orthogonalized withrespect to Y and P1 is updated. The final T1 and P1 correspond to firstOSC factor and its products are then removed from original data. Toobtain a second OSC factor the complete treatment is then applied oncorrected data, therefore for n OSC factors corrected data XOSC isgiven by:

XOSC = X−∑n

i=1TiP

′i ð1Þ

3. Experimental

3.1. The dataset

The dataset was from Osmetech plc (Cheshire, UK). Three differentanalytes (ammonia, propanoic acid and n-butanol), at differentconcentrations levels, were periodically measured over 10 monthswith an array of 17 conductive polymer sensors, the total number ofsamples in the dataset being 3415. The concentration and number ofsamples for each analyte are shown in Table 1. We consider eachanalyte as a class. Additionally, we define a group as a particularanalyte at a given concentration. Hence 3 classes and 8 groups arepresent in the dataset.

For every sample, the full sensor response to a sampling transientis recorded. That is, the sensor array is initially exposed to clean air.Subsequently, the analyte at the desired concentration is introducedin the sensor chamber for 185 s. Finally, clean air is introduced again.Every transient signal lasts for 200 s at a sampling frequency of 1 Hz.All waveforms are baseline corrected so that starting baseline in allsamples is common at time 0 s. Fig. 1 shows an example of waveformfor one sensor and three classes. To apply the proposed drift

29M. Padilla et al. / Chemometrics and Intelligent Laboratory Systems 100 (2010) 28–35

counteraction techniques, the datamatrix is organized into a two-waymatrix with dimensions: number of samples x (sensors x transient time)or 3415 samples x 3400 variables.

At every sample interval of the transient signals, the response ofthe sensor array produces a characteristic pattern for every class. Fig. 2shows the patterns of three different classes at time 185 s. of thetransient waveform. Traditionally only one point of a transient signalis considered for every sensor, this point usually corresponds to themaximum value of the signal (in our case around instant 185 s inevery transient signal). When introducing the complete transient

waveform in the data matrix, dynamic information is also beingconsidered [38]. It is well known that transient information providesadditional discriminatory information.

3.2. Methodology

The proposed model validation procedure has been specificallyconceived to illustrate the capabilities of the different algorithmicsolutions towards drift rejection. The secondary goal is to ascertainthe behavior of the algorithmswhen a small sample training dataset isencountered. This is usually the case since, from a practical point ofview; calibration costs have to be reduced to a minimum.

As a figure of merit, the performance of a classifier in time-orderedvalidation subsets is evaluated. The classifier consists of a dimension-ality reduction step using Principal Component Analysis, followed by ak-NN classifier. While the input dimensionality is 3400, the outputdimensionality is limited to the number of Principal Components(PCs) that capture most of the X data variance. In this space a k-NNclassifier is used with k=3 nearest neighbors. The dataset has beenclassified in three classes, corresponding to the three chemicalspecies.

3.2.1. Validation methodologyIn order to carry out rigid validation, algorithm optimization has

been restricted to the use of calibration set information. Final testingwas performed with data subsets never used for algorithm buildingand optimization. To test the effectiveness of the drift counteractiontechniques, the complete dataset (3415 samples), was divided in 10sections (or subsets) of 342 samples each approximately. All samplesare ordered in time. A scheme of the validationmethodology is shownin Fig. 3. The first subset consists of samples measured during the first15 days of experiments. We will refer to this first section as thecalibration set.

In a first step, the algorithms parameters are optimized by using aninternal cross-validation within the calibration set. The last quarter ofthe calibration set is used for inner validation. The model is built witha random selection of 66% of the remaining calibration set samples(first 3/4). By repeating this process 10 times we evaluated therobustness of the models towards the particular selection of trainingdata. Graphics of results provide the corresponding error bars.

Once the algorithmswere optimized, stability over time evaluationwas done by assessing the performance with the remaining nine datasubsets ordered in time (Fig. 3).

To evaluate the performance of drift correction for smaller trainingsets, ten additional data subsets were built with random samples fromthe original calibration set. The sizes of these subsets range from 10%to 100% of its size. Therefore, the smallest calibration subset contains34 samples and the largest one 342. These ten subsets have been usedas training sets for building the algorithms models. Please note that inthis section, no further optimization of the algorithms innerparameters is done. Hence, these parameters stay fixed as the sizeof the training set varies.

3.2.2. Algorithm optimizationTo select the best internal parameters, we propose to use as a

figure of merit the Fisher ratio: ratio between inter-group and intra-group variance of the cross-validation dataset.

Table 1Measured compounds, concentrations and number of samples in the dataset.

Analyte Samples Concentration level

Ammonia 447 0.01%452 0.02%307 0.05%

Propanoic acid 458 0.01%457 0.02%423 0.05%

n-Butanol 446 0.01%425 1.00%

Fig. 1. Transient response of sensor 1 to 3 analytes.

Fig. 2. Examples of patterns corresponding to 17 sensors and samples of all classes atthe maximum time point of the transient signal (instant 185s). Fig. 3. Scheme of time stability evaluation.

30 M. Padilla et al. / Chemometrics and Intelligent Laboratory Systems 100 (2010) 28–35

In OSC, two internal parameters must be optimized; the number ofOSC components and the number of internal PLS latent variables.Here, instead of the number of inner PLS latent variables, we use theequivalent amount of captured variance in the X-block (or tolerance)of the inner PLS relation. This is the way original Wise's algorithm isimplemented [37].

On the other hand, in the PCA-CC method the number of principalcomponents and a reference group have to be selected as well. Thecriterion used to select the reference group was based on visualinspection of the score plots in the calibration set. This group shouldhave an inner covariance structure similar to most of the remaininggroups in the dataset. The reference group is modeled with a PrincipalComponents decomposition that captures a high amount of the groupinner variance in the calibration set. In fact, removingmanycomponentsin PCA-CC is risky, since some of those components may have infor-mation useful for posterior classifications. This is the case when driftdirection is parallel to discriminant direction. Because of these factorsonly three components were removed by PCA-CC.

4. Results and discussion

An exploratory analysis by PCA displays the initial distribution ofthe classes. Fig. 4 shows the PCA scores plot of the calibration set(solid symbols). In the figure, it can be seen that some classesapparently mix, there is scattering due to drift and also additionalintra-class variability. The direction of the main dispersion compo-nent in all classes is quasi parallel, except ammonia 0.05%. Somegroups, ammonia 0.01% and 0.02% and both concentration levels of n-butanol, present dispersion over mainly only one direction. Maindispersion directions are pointed out by an arrow for every type ofgas. A thick arrow shows the direction of the displacement of theeighth validation set (non-solid symbols) with respect to thecalibration set. Furthermore, in this data subset sensors are verycorrelated since only ten principal components capture more than99% of the total variance, and the first two PCs capture about 90%.

A posterior data subset of 342 samples, measured between day 97and day 118, has also been projected on the same PCA subspace.Scores of this projection are shown in the same figure in non-solidsymbols. The distribution of the groups in this latter subset is similarto the one in the first subset; only ammonia 0.05% shows an importantdeviation in its main dispersion directions. However, the whole subsetis completely shifted from the location of the first data subset.Therefore, additional to variability and scattering due to local noise/

drift, there is a long-term drift effect, which displaces and changes thevariance within the data set structure.

Fig. 5 shows the maximum of the transient for sensor one for thetotal duration of the experiment. In this figure, the effect of drift andintra-class variability in the measurements is clearly seen, and also aclear correlation between the different traces can be observed. Theirregular shape of these curves may be due to sensors aging and alsothe environmental changes that happened during the 10 months ofmeasurements, since conductive polymer sensors are stronglyperturbed by temperature and humidity.

4.1. Fitting PCA-CC and OSC parameters

In Fig. 6, the Fisher ratio is plotted for different numbers ofremoved components and tolerances for OSC and PCA-CC. Fisher ratiomeasures the ratio between mean values of the distances of groupsamong themselves and the groups' compactness:

FR = ∑N

i;j=1;i≠jjcij j = ∑

N

i=1mxi

where cij=|ci−cj| is the distance between the mean centers of groupsi and j, and mxi is the mean of the distances of the Nk elements ofgroup i to its center ci:

mxi =1Nk

� �∑Nk

k=1jxk−ci j ; with ci =

1Nk

� �∑Nk

k=1jxk j

Therefore, high values are desired since they would mean smallgroups and high separation among them. The proposed figure of merithelps to avoid over fitting, however it does not guarantee the bestchoice for drift rejection in the long term, since the cross-validationsamples are very close in time to the training samples.

Fig. 6 shows that the Fisher ratio is higher for OSC than for CC,suggesting that the former will outperform the later in the finalclassification task.

For the OSC, we observe a continuous increase in the Fisher ratiowith the number of extracted components at multiple levels of thetolerance value. Itmaybe surprising that this component removal doesnot saturate fast, but it has to be considered that this preprocessing iscarried out in an input space featuring high dimensionality, and thegraph only explores removal of up to 16 components. On the otherhand, the figure of merit shows more sensitivity to the trainingsamples (larger error bars) when the number of OSC componentsincreases. 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 (measurementsbetween day 97 and day 118, in non-solid symbols). Fig. 5. Responses of sensor 1 to 3 analytes along the time.


used. Additionally, the tolerance value chosen has a higher impact inthe results for higher number of OSC components, while it has littleinfluence for a reduced number of components. For further analysis,two cases are considered one and eight OSC components both with ahigh tolerance, 99.0% (other high tolerances around 99% could alsohave been chosen).

Regarding PCA-CC method, a reference group must be chosen. Thefirsts loadings of a PCA on this group are subtracted from all data, thusthis group must be representative of all groups. In Fig. 4, it has beenseen that all groups present main dispersion directions that are nearlyparallel, except for ammonia 0.05%. This direction is very well definedfor n-butanol and ammonia 0.01 and 0.02% but ammonia 0.05%, andpropanoic acid presents more dispersion components. Therefore,either n-butanol or ammonia 0.01 or 0.02%, can be a good reference.Finally we have selected n-butanol 1%. Fig. 6 shows Fisher ratio forPCA-CC computed with two groups; ammonia 0.05% and n-butanol1%. Best performance is obtained when using n-butanol 1% as areference. In fact, when selecting a reference group the dispersion ofwhich is not representative (ammonia 0.05%), Fig. 6 shows that CCdoes not improve Fisher ratio, but may even worsen it. This resultconfirms the choice of the reference group (n-buthanol 1%) from thevisual inspection of the PCA scores plot in Fig. 4. For both referencegroups, the Fisher ratio saturates sooner than in the OSC. Finally, Fig. 6also shows that the PCA-CC models are less sensitive to the particularchoice of training samples than OSC models (especially for highernumber of removed components). Three Principal Components havebeen removed, since they capture a high amount of inner variance(about 95%) in the reference group.

Fig. 7 shows one removed OSC component and only the firstremoved component in PCA-CC for the sake of a better visualization.An exploration on the shape of the loadings for the PCA of thereference group helps to understand where the drift appears (Fig. 7).This may have some impact on later instrument optimization.

It is interesting to observe, that the first principal component forPCA-CC drift correction almost equally weights all the sampleintervals in transient time. Regarding sensors, sensor 11 and 12contribute to drift slightly more than other sensors in this component.This component also shows high values (peaks) at points where thetransient signal changes abruptly, corresponding to the instant whenthe analyte comes into the sensors chamber and when dry air cleansthe chamber (at 2 s and 185 s approx.). It seems to reflect some jitterin this point and a better synchrony between signal acquisition andthe chemical sampling system is needed.

First OSC component is similar to first PCA-CC component but withemphasis on the last part of the transient. It also confirms that sensors11 and 12 are less stable than the others. As in PCA-CC, in general thesensors that contribute the most are 11, 12 and 17. Also the highestcomponents values (peaks) are located at the same transient instant,like in PCA-CC, showing that transition times contain most of thevariance, thus they are more affected by drift.

4.2. Data distribution of corrected data

A PCA scores plot for the calibration set and a posterior test set hasbeen shown on Fig. 4. Such test set consists on samples measuredbetween day 97 and day 118. PCA scores plot for corrected test setwith each method are shown on Figs. 8 and 9. In these figures, datafrom the test subset are projected on the PCA model (non-solidsymbols) built from corrected calibration set (solid symbols) like inFig. 4.

Distribution of data corrected by OSC (8 components and 99%tolerance, Fig. 8) shows little dispersion in its eight groups, which arewell separated, the better the ones corresponding to the highestconcentration of every analyte. Groups belonging to the same class arelocated along one direction in the figure, therefore a data setdistribution consist of clusters along three different directions, one

Fig. 6. Fisher ratio vs. number of OSC and PCA-CC removed components and internalOSC tolerances. PCA-CC is computed with two different groups of reference; ammonia0.05% (group 3) and n-butanol 1% (group 8). Error bars represents sensitivity to trainingsamples.

Fig. 7. First data removed component by PCA-CC and OSC.

Fig. 8. PCA scores of the calibration set (solid symbols) and eighth test set (non-solidsymbols) corrected by OSC-8 (8 removed components).


for each class, with decreasing concentrations towards the center ofthe data distribution. Corrected test set is displaced from the trainingset, presents higher dispersion than training data and shows a smallchange on the orientation of the branch corresponding to everychemical specie. All these three effects are due to drift that OSC hasnot modeled.

Drift corrected data by means of PCA-CC (3 components, Fig. 9)also shows a distribution along three main directions, with differentorientation according to the chemical species. It presents a very smallcluster for the reference group in the training set and other class of thesame corresponding chemical specie. Analytes whose initial trainingset distribution (Fig. 4) do not have parallel components to the ones ofthe reference group, like ammonia 0.05%, show larger variance. Due todrift, corrected validation data is also slightly displaced from trainingdata and it causes a change in the orientation of some class directions.The final eight groups are in general larger than in OSC plots andpoorer in separation.

It can also be noticed that the sparsest clusters in Figs. 8 and 9 arethe ones corresponding to ammonia. It has already been observed inFig. 4, that ammonia is the class whose covariance structure haschanged the most from first subset to seventh, due to the effect oflong-term drift that both signal processing methods are not able tocorrect.

4.3. Time stability

In Fig. 10 the performance of PCA-CC and OSC drift counteractiontechniques are plotted in the form of classification rates (CR) forcorrected and non-corrected data. Dataset dimensionality hasbeen reduced to 10 PCs, capturing about 99% of variance, previousto the k-NN classifier. CR is calculated over nine test sets ordered intime and models are built with 308 samples from calibration set.Results show that both methods systematically improve the classifi-cation rate. Obviously, the correction is not perfect, and the CR stilldecreases with time.

OSC-8 (8 components) method outperforms OSC-1 (1 component)and PCA-CC providing higher CR values along the first month ofmeasurements (first validation set). After first month, OSC-1clearly outperforms the other two methods until nearly day 100.OSC-8 outperforms PCA-CC until around day 100. Later, all methodsbecome comparable, although PCA-CC presents the more unstablebehavior. Regarding sensitivity to training set samples, in generalPCA-CC shows the smallest value and OSC-1 presents a lower error barthan OSC-8. This sensitivity was also observed in the Fisher ratiofigure (Fig. 6). We can conclude that OSC-1 is in general the most

suitable method for drift correction. However, if the data is very closein time to training data OSC-8 is themost effective one. Up to 100 daysboth OSC methods outperform PCA-CC.

The main reason of the different performance of both methods isthe way they calculate the components of variance to be removed. Themain component obtained by PCA-CC has an orientation given by thedirection of the largest variance of a reference group and, if morecomponents are to be calculated, these ones are orthogonal to it. Thismakes that the removed components may be parallel to importantdirections for posterior classification, since it is not possible to controltheir direction. On the other hand, PCA-CC only takes samples fromone group of reference to build a model, unlike OSC methods, whichtakes samples from all groups. PCA-CC reference group must berepresentative of all groups; therefore the more similar the groups arethe better this method performs. As a result PCA-CC depends stronglyin the selection of the reference.

Removed components in OSC are calculated considering informa-tion from all groups. The information regarding sample distribution isin matrix Y, which contains the sample memberships to each of the8 groups. Y is a matrix with dimensions number of samples x number ofgroups and binary elements; a ‘1’ at row i and column j means thatsample i belongs to group j. The condition of orthogonality assuresthat extracted components of variance are not parallel to importantdirections. Furthermore, OSC components are not forced to beorthogonal among them. These facts mean that OSC gives very goodresults for samples very close in time to the training set (little affectedby drift), outperforming PCA-CC until around day 100. Later itdegrades smoothly along the time resulting in a better and morestable behavior than non-corrected data, and even more stable thanPCA-CC.

For both methods, selection of parameters is a critical step. Morecomponents or more strict tolerance (in the case of OSC) would resultin a better-fitted model to training samples but in a loss of ability ofgeneralization to correct posterior data affected by drift. This fact canbe observed on comparing OSC-1 and OSC-8 results: OSC-8 outper-forms OSC-1 only on correcting samples from first validation set.

In fact, in both cases the main difficulty resides in estimating froma short period (calibration set duration) the directions of drift. For themethod to be effective, the system should exhibit at least stability inthe statistical properties of the variance structure. If the structure ofthe noise/drift variance changes in time beyond the calibration phase,obviously the methods have no capability to re-adapt.

On the other hand, OSC classification results show higheruncertainty than PCA-CC. We attribute this increased variance to thefact that it takes samples randomly from the eight groups in the

Fig. 9. PCA scores of the calibration set (solid symbols) and eighth test set (non-solidsymbols) corrected by PCA-CC (3 removed components, group of reference 8).

Fig. 10. Classification rates along the time for drift correction by PCA-CC and OSCmethods and non-corrected data (No Corr). Calibration set contains 308 samples.


training set, not only one as in PCA-CC. In consequence, the identifiedsources of variance may vary depending on the particular samplesthat are included in the training set.

The results indicate that this preprocessing is also useful to obtainbetter results also in times close to the training set. Methods areeffective not only correcting drift, but also other sources of variance.

4.4. Effect of the training set size

From a practical point of view, for these methods to be useful, theyhave to be able to correct data with a limited number of calibrationsamples. Fig. 11 shows the performance of both methods as a functionof the training set size for the first validation set. All curves presentbetter results with increasing number of training samples. However, itis remarkable that very good results can be obtained with very fewtraining samples. The number of training samples in Fig. 11corresponds to samples belonging to all classes, thus valid for OSCand the raw data case. However, PCA-CC only uses training samplesbelonging to one group; therefore, the effective number of trainingsamples is 1/8th of what is shown in horizontal axis. It is remarkablethat PCA-CC is able to improve the results only with 10 trainingsamples from the reference class.

PCA-CC classification rates stabilize with fewer training samplesthan OSC. The reason for this is the need of OSC to contain severalsamples from all classes to build a model. Again, PCA-CC only needsseveral samples from one class.

The apparent advantage for PCA-CC regarding the number ofsamples needed (1/8 of the total calibration set) is not absolutely true,because samples from the rest of the classes are also used to build theclassifier. However, in the current scenario OSC uses all theinformation available to build the model.

Nevertheless a different scenariomay be clearly envisioned, that ofusing additional future samples for recalibration. In PCA-CC recali-bration samples come only from a single reference analyte andconcentration, while OSC would require samples for all classes andconditions. Of course, the use of a single substance and not the wholeset of analyte species would be a clear advantage for choosing PCA-CCin this second scenario.

5. Conclusions

Lack of stability over time or drift, is a main drawback for the use ofchemical sensor array based instruments.

In this work, a drift compensation technique based on OrthogonalSignal Correction (OSC) has been proposed and compared to Compo-nent Correction (CC) method.

The results clearly show that the application of these preproces-sing techniques greatly improves the data distribution resulting insmaller clusters with better separation and consequently betterdiscrimination. Additionally, and from an interpretational point ofview, the analysis of the loadings of the removed components help toidentify the sources of unwanted variability. It may help to identifyparticularly unstable sensors, or areas of the transient signal that arenot stable. In this particular dataset, results show that times very closeto gas switching in the sampling system are rather unstable.

Results show that OSC outperforms PCA-CC for a limited period oftime (about 100 days in present example), while later on theadvantage is not clear. Complex OSC models do a better correctionof variance for shorter times, but they degrade faster than simpler OSCthat remain stable for a longer time.

It is also important to remark that both methods are relativelyrobust regarding small calibration set sizes and perform rather wellwith a reduced calibration set. OSC results show a higher variancethan PCA-CC, but this may be a result of the number of componentschosen in the models. On the other hand, PCA-CC needs a smallertraining set and a single chemical species. This advantage may turninto disadvantage if the reference class is not properly chosen.

Since gas sensor arrays systems are typically plagued with stabilityproblems, the authors would like to emphasize the importance ofusing validation methodologies that use test samples interspersed intime outside of the training set and acquired subsequent to thetraining set.

In summary, these two drift counteraction techniques providebetter performance over posterior classifiers or regression methodsby removing unwanted variance from the data. Although improvingthe performance of any posterior data processing, they do notcompletely solve the problem of drift. While recalibration of theinstrument is still necessary, the time between recalibrations can beextended.

Acknowledgment

This work was partially funded from the European Community'sSeventh Framework Programme (FP7/2007-2013) under grant agree-ment no. 216916: Biologically inspired computation for chemicalsensing (NEUROCHEM).

References

[1] Persaud, Dodd, Analysis of discrimination mechanisms in the mammalianolfactory system using a model nose, Nature 299 (1982) 352–355.

[2] R. Ionescu, A. Vancu, A. Tomescu, Time-dependent humidity calibration for driftcorrections in electronic noses equipped with SnO2 gas sensors, Sensors andActuators. B, Chemical 69 (2000) 283.

[3] C. Di Natale, E. Martinelli, A. D'Amico, Counteraction of environmental dis-turbances of electronic nose data by independent component analysis, Sensorsand Actuators. B, Chemical 82 (2002) 158.

[4] S. Roussel, G. Forsberg, P. Grenier, V. Bellon-Maurel, Optimisation of electronicnose measurements. Part II: influence of experimental parameters, Journal of FoodEngineering 39 (1999) 9.

[5] R.K. Sharma, P.C.H. Chan, Z.N. Tang, G.Z. Yan, I.M. Hsing, J.K.O. Sin, Investigation ofstability and reliability of tin oxide thin-film for integrated micro-machined gassensor devices, Sensors and Actuators. B, Chemical 81 (2001) 9.

[6] P. Mielle, Managing dynamic thermal exchanges in commercial semiconductinggas sensors, Sensors and Actuators. B, Chemical 34 (1996) 533.

[7] B. Ruhland, T. Becker, G. Muller, Gas-kinetic interactions of nitrous oxides withSnO2 surfaces, Sensors and Actuators. B, Chemical 50 (1998) 85.

[8] M. Burgmair, M. Zimmer, I. Eisele, Humidity and temperature compensation inwork function gas sensor FETs, Sensors and Actuators. B, Chemical 93 (2003) 271.

[9] R. Paris, S. Pawel, R. Herzer, T. Doll, P. Kornetzky, R.P. Gupta, G. Eranna, Proceedings ofIEEE Sensors, First IEEE International Conference on Sensors (Cat. No.02CH37394),2002.

[10] G. Muller, A. Friedberger, P. Kreisl, S. Ahlers, O. Schulz, T. Becker, A MEMS toolkitfor metal-oxide-based gas sensing systems, Thin Solid Films 436 (2003) 34.

[11] Gramm, A. Schutze, High performance solvent vapor identification with a twosensor array using temperature cycling and pattern classification, Sensors andActuators. B, Chemical 95 (2003) 58.

Fig. 11. Classification rates for corrected first test subset by PCA-CC and OSC methodsand non-corrected data (No Corr) vs. number of samples in the calibration set.


[12] M. Roth, R. Hartinger, R. Faul, H.E. Endres, Drift reduction of organic coated gas-sensors by temperature modulation, Sensors and Actuators. B, Chemical 36 (1996)358.

[13] E.L. Hines, E. Llobet, J.W. Gardner, Electronic noses: a review of signal processingtechniques, IEEE Proceedings—Circuits Devices and Systems 146 (1999) 297.

[14] J.E. Haugen, O. Tomic, K. Kvaal, A calibration method for handling the temporaldrift of solid state gas-sensors, Analytica Chimica Acta 407 (2000) 23.

[15] M. Holmberg, F. Winquist, I. Lundstrom, F. Davide, C. DiNatale, A. Damico, Driftcounteraction for an electronic nose, Sensors and Actuators. B, Chemical 36 (1996)528.

[16] S. Marco, A. Ortega, A. Pardo, J. Samitier, Gas identification with tin oxide sensorarray and self-organizing maps: adaptive correction of sensor drifts, IEEETransactions on Instrumentation and Measurement 47 (1998) 316.

[17] M. Zuppa, C. Distante, P. Siciliano, K.C. Persaud, Drift counteraction with multipleself-organising maps for an electronic nose, Sensors and Actuators. B, Chemical 98(2004) 305.

[18] M. Holmberg, F.A.M. Davide, C. DiNatale, A. Damico, F. Winquist, I. Lundstrom,Drift counteraction in odour recognition applications: lifelong calibration method,Sensors and Actuators. B, Chemical 42 (1997) 185.

[19] H. Ding, H.F. Ge, J.H. Liu, High performance of gas identification by wavelettransform-based fast feature extraction from temperature modulated semicon-ductor gas sensors, Sensors and Actuators. B, Chemical 107 (2005) 749.

[20] H. Ding, J.H. Liu, Z.R. Shen, Drift reduction of gas sensor by wavelet and principalcomponent analysis, Sensors and Actuators. B, Chemical 96 (2003) 354.

[21] E. Llobet, J. Brezmes, R. Ionescu, X. Vilanova, S. Al-Khalifa, J.W. Gardner, N. Barsan,X. Correig, Wavelet transform and fuzzy ARTMAP-based pattern recognition forfast gas identification using a micro-hotplate gas sensor, Sensors and Actuators. B,Chemical 83 (2002) 238.

[22] T. Artursson, T. Eklov, I. Lundstrom, P. Martensson, M. Sjostrom, M. Holmberg,Drift correction for gas sensors using multivariate methods, Journal of Chemo-metrics 14 (2000) 711.

[23] O. Tomic, T. Eklov, K. Kvaal, J.E. Haugen, Recalibration of a gas-sensor array systemrelated to sensor replacemet, Analytica Chimica Acta 512 (2004) 199.

[24] R. Gutierrez-Osuna, Drift reduction for metal-oxide sensor arrays using canonicalcorrelation regression and partial least squares, Electronic Noses and Olfaction:

Proceedings of the 7th International Symp. On Olfaction and Electronic Nose,Brighton, UK, July 20–24, 2000, Institute of Physics Publishing, 2000, p. 147.

[25] O. Tomic, H. Ulmer, J.E. Haugen, Standardization methods for handling instrumentrelated signal shift in gas-sensor array measurement data, Analytica Chimica Acta472 (2002) 99.

[26] J.E. Carlson, P.E. Martinsson, Exploring interaction effects in two-component gasmixtures using orthogonal signal correction of ultrasound pulses, Journal of theAcoustical Society of America 117 (2005) 2961.

[27] S. Wold, H. Antti, F. Lindgren, J. Ohman, Orthogonal signal correction of near-infrared spectra, Chemometrics and Intelligent Laboratory Systems 44 (1998) 175.

[28] T. Fearn, On orthogonal signal correction, Chemometrics and IntelligentLaboratory Systems 50 (2000) 47.

[29] J. Sjoblom, O. Svensson, M. Josefson, H. Kullberg, S. Wold, An evaluation oforthogonal signal correction applied to calibration transfer of near infraredspectra, Chemometrics and Intelligent Laboratory Systems 44 (1998) 229.

[30] J.A. Westerhuis, S. de Jong, A.K. Smilde, Direct orthogonal signal correction,Chemometrics and Intelligent Laboratory Systems 56 (2001) 13.

[31] Li, A.J. Morris, E.B. Martin, Orthogonal signal correction: algorithmic aspects andproperties, Journal of Chemometrics 16 (2002) 556.

[32] R.N. Feudale, H. Tan, S.D. Brown, Piecewise orthogonal signal correction,Chemometrics and Intelligent Laboratory Systems 63 (2002) 129–138.

[33] J. Trygg, O2-PLS for qualitative and quantitative analysis in multivariatecalibration, Journal of Chemometrics 16 (2002) 283.

[34] J. Trygg, S. Wold, Orthogonal projections to latent structures (O-PLS), Journal ofChemometrics 16 (2002) 119.

[35] O. Svensson, T. Kourti, J.F.MacGregor, An investigation of orthogonal signal correctionalgorithms and their characteristics, Journal of Chemometrics 16 (2002) 176.

[36] H.C. Goicoechea, A.C. Olivieri, A comparison of orthogonal signal correction andnet analyte preprocessing methods. Theoretical and experimental study, Chemo-metrics and Intelligent Laboratory Systems 56 (2001) 73.

[37] B.M. Wise, N.B. Gallagher, http://www.eigenvector.com/MATLAB/OSC.html.[38] J. Samitier, J.M. Lopez-Villegas, S. Marco, L. Camara, A. Pardo, O. Ruiz, J.R. Morante,

A new method to analyse signal transients in chemical sensors, Sensors andActuators. B, Chemical 18 (1994) 308.



Contents lists available at ScienceDirect

Sensors and Actuators B: Chemical

journa l homepage: www.e lsev ier .com/ locate /snb

Multivariate curve resolution applied to temperature-modulated metaloxide gas sensors

I. Montoliua, R. Taulerb, M. Padillaa,c, A. Pardoa, S. Marcoa,c,∗

a Departament d’Electrònica, Universitat de Barcelona, Martí i Franquès, 1, 08028 Barcelona, Spainb Institute for Research in Chemistry and Environment (IIQAB-CSIC), Jordi Girona 18-26, 08034 Barcelona, Spainc Institut for BioEngineering of Catalonia, Baldiri i Rexach 4, 08028 Barcelona, Spain

a r t i c l e i n f o

Article history:Received 25 July 2009Received in revised form10 December 2009Accepted 17 December 2009Available online 4 January 2010

Keywords:Temperature modulationMultivariate curve resolutionMCR-ALSMetal oxide sensors

a b s t r a c t

Metal oxide (MOX) gas sensors have been widely used for years. Temperature modulation of gas sensors isas an alternative to increase their sensitivity and selectivity to different gas species. In order to enhance theextraction of useful information from this kind of signals, data processing techniques are needed. In thiswork, the use of self-modelling curve resolution techniques, in particular multivariate curve resolution-alternating least squares (MCR-ALS), is presented for the analysis of these signals. First, the performanceof MCR in a synthetic dataset generated from temperature-modulated gas sensor response models hasbeen evaluated, showing good results both in the resolution of gas mixtures and in the determination ofconcentration/sensitivity profiles. Secondly, experimental confirmation of previously obtained conclu-sions is attempted using temperature-modulated MOX sensors together with MCR-ALS for the analysisof carbon monoxide (CO) and methane (CH4) gas mixtures in dry air. Results allow confirming the pos-sibility of using the proposed approach as a quantitative technique for gas mixtures analysis, and alsoreveal some limitations.

© 2009 Elsevier B.V. All rights reserved.

1. Introduction

Performance of semiconductor metal oxide gas sensors is stilla matter of current investigation and discussion [1]. These sensorsconsist of two electrical resistances: on the one hand, the sensingresistance made of the semiconductor metal oxide and the value ofwhich depends on the presence of the gas phase chemical speciesand the heater (typically platinum, although other choices are pos-sible). The aim of the heater is to increase the temperature of thesensor to the range where the chemical reactions at the sensingmaterial surface effectively occur. These kind of sensors featurepoor selectivity that constraints their use to low demanding (inanalytical terms)—low cost applications, as for instance, domesticgas alarms for combustible gases.

A first approach to improve this selectivity is to combine a diver-sity of sensors into a gas sensor array (the so-called electronicnoses), and then process the combined output using multivariatemethods. There are however, several additional ways to improvealso the selectivity of the individual gas sensors [2], such as theprevious separation of analytes [3], the investigation of differ-

∗ Corresponding author at: Departament d’Electrònica, Universitat de Barcelona,Martí i Franquès, 1, 08028 Barcelona, Spain. Tel.: +34 934 029 070;fax: +34 934 021 148.

E-mail addresses: [email protected], [email protected] (S. Marco).

ent semiconductor materials [4] (SnO2, WO3, In2O3, (Sn–In)O,etc.), the use of specific surface additives to dope the basis mate-rial employed in the sensor [5], and the proper selection of theoperating temperature [6]. To achieve the best operating condi-tions towards the detection of one specific analyte, temperaturecontrol/modulation of metal oxide (MOX) gas sensors has beenexplored during the last years and described as a good methodto provide differentiated response behaviors when MOX sensorsare exposed to different gas mixtures [7]. This discriminating effecthas been ascribed to the temperature dependence of the sensitivityto various chemical species as well as to the temperature depen-dence of the kinetics of the reaction on the sensor’s surface. To takeadvantage of this effect, temperature modulation and control of thesensor have followed basically two strategies: the pulsed temper-ature modulation, where a discrete voltage pulse is applied to thesensor’s heater and its response towards the analyte is monitoredduring the following time (transient recording); and the continuoustemperature modulation through the introduction of a controlledoscillation of the heater voltage, which allows the periodic heat-ing/cooling of the sensor.

The specific signal obtained from the sensors operated underthese conditions is complex and has been studied since yearsmainly using different approaches [8,9]. For instance, the appli-cation of gaussian deconvolution techniques to data from binarysystems has rendered interesting results in the interpretationof the oxidation mechanisms on the sensor’s surface and in


I. Montoliu et al. / Sensors and Actuators B 145 (2010) 464–473 465

simple calibrations by comparison to a reference standard. How-ever, these results have become of limited utility in quantitativeanalysis, mainly due to cross-sensitivity effects.

Other authors have proposed the use of relative intensities ofreal and imaginary components of the higher harmonics on theFast Fourier Transform (FFT) of gas sensor’s signals to distinguishbetween analytes (increasing selectivity) and this has been takenas a basis for quantitative purposes. The use of 3D plots [10] ofconcentrations vs. individual harmonics, with its further Artifi-cial Neural Networks (ANN) + Fuzzy Inference processing [11,12],have also provided partially successful qualitative and quantitativeresults.

In general, chemometric techniques [13] for mixture analysislook for the transformation of multivariate data into relevant infor-mation, allowing the determination of relationships among thedifferent individual system contributions to the global signal.

Multivariate curve resolution (MCR) is an algorithm for blindsource separation (BSS) [14] specifically adapted to chemical prob-lems. In this framework, the observed signals (sensor responses) areconsidered to be a mixture of source signals (concentrations of purecomponents). BSS separation techniques try to recover the timeevolution of the source signals (concentrations of pure analytes)without any knowledge of the mixing matrix. In other words, with-out any prior calibration step. BSS is typically attempted in signalprocessing by independent component analysis (ICA). ICA has beenapplied to chemical sensor signals by a number of authors [15,16].However, for chemical problems a number of techniques havebeen proposed that are able to perform this task (e.g. SIMPLISMA,MCR-ALS). Instead of imposing independence, other constraints aspositiveness and others are used. The aim of self-modelling mul-tivariate curve resolution (MCR) methods is to generate a bilinearmodel that best describes the individual contributions of a reducednumber of components to the global signal experimentally mea-sured. In other words, MCR methods seek for a bilinear modelwhich leads to the best fit of the original experimental data, ina least squares sense. The most obvious difference with princi-pal component analysis (PCA) is that the extracted componentsare not required to be orthogonal. MCR are powerful methodsbecause they are flexible and have low requirements. It is notnecessary to have an exhaustive previous knowledge of the behav-ior of the system, although additional information can be used ifit is available. The only initial assumption is the linear additiv-ity of the individual responses of the components of the analyzedmixture. Unfortunately, metal oxide sensors are generally consid-ered as non-linear sensing devices. However, in some particularcases where the concentration range is limited and the signal isexpressed in conductance (G), sensor array signals can be approxi-mately described by a linear additive model, individually weightedfor each of the individual contributions. Typically, the concentra-tion of these individual contributions or components can be used asa weighting element. In this work, we will explore to which extentthe non-linearity of the sensor responses allows the use of bilinearadditive models.

In this work, MCR methods are proposed to determine the purecomponent response (sensitivity) profiles as well as the concen-tration profiles from a set of temperature-modulated MOX sensorssignals. We have already mentioned that this approach has consid-erable interest since it can extract concentration profiles (signals)and sensitivities without prior calibration steps. The estimationof the sensitivity profiles has additional interest because it givesadditional insight regarding the sensor response and, moreover,it may help to set-up the optimum operational temperatures in aparticular scenario.

Two cases have been investigated: a synthetical dataset corre-sponding to a ternary gas mixture, and an experimental datasetcorresponding to the exposure of four commercial MOX gas sen-

sors to different binary gas mixtures. In a second part of the work,the exploration of the possibilities of the proposed approach forquantitative purposes is also investigated.

2. Experimental

2.1. Data description

Let us consider that we have N different temperature-modulated MOX sensors. Their signals are acquired during thetotal time of the experiment texp, with a sampling time of ts. Thesensor signals do produce then in general an array of dimen-sions (N, texp/ts). Temperature modulation has a period of tcycle(ts � tcycle � texp). To simplify the measurement conditions weassume that texp is an exact multiple of tcycle and tcycle is an exactmultiple of ts. We may consider that at the end of every modula-tion cycle, each sensor outputs a measurement vector of tcycle/ts

samples. Along the total duration of the experiment every sensorsproduces texp/tcycle vectors. In this way, and for the whole dura-tion of the experiment we may consider that each sensor producesa data matrix of dimensions (texp/tcycle, tcycle/ts). Since we haveN sensors, data can be organized as a cube (or three-way array)of dimensions (N, texp/tcycle, tcycle/ts) (see Fig. 1). In the follow-ing, we will refer to those dimensions as sensors, concentration(its evolution along the time) and sensitivities (due to temperaturemodulation).

Temperature modulation is accomplished by modulating thevoltage applied to the sensor heater. In this way, sensor oper-ation temperature is basically a low-pass version of the powersignal due to thermal inertia. MOX devices can exhibit diverse ther-mal time constants depending on their internal structure. Whilefast MOX devices fabricated using Microsystem technologies canachieve time constants �therm around 10 ms, older devices based inceramic substrates can have slower thermal dynamics with timeconstants on the order of 10 s. In any case, it is usually incorrect toassume that sensors temperature follow instantaneously the powersignal applied to the sensor. In order to have a significant variationof the temperature is necessary that tcycle � �therm.

Since the characteristic measurement time for every sensor isgiven by tcycle. This time should be such that the evolution of theconcentration within the gas chamber does not change dramati-cally. In other words, we will assume that chemical concentrationis basically constant during the cycle time.

2.2. Synthetic dataset

A synthetic dataset was generated, intending to simulate the sig-nals that can be obtained from the measure of a ternary mixture ofgases in air. While concentration of one of them was kept constantalong all the measurement period, the other gases were incorpo-rated and removed consecutively from the measurement chamberwith a certain time overlap. The conductance of gas sensors wasnormalized to the conductance in pure air. To do so, two datasetsfor each gas sample were generated. Thus, a first one was obtained,containing the theoretical signal when the sensor is exposed topure air, and a second one, giving its expected response when itis exposed to the ternary gas mixture.

In order to generate these datasets, the dynamic Clifford–Tumamodel [17,18] was taken as a basis, with suitable parameters todesign sensors based in n-type semiconductors. Thus, a model pro-viding responses with different sensitivities to air, CO, CH4 andethanol gases was obtained. It must be noted that all these gasesproduce an increase in the conductance of n-type sensors. The sim-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 bytemperature 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, renderingas 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 datapoints recorded.

While it is not the objective to describe here a detailed model ofthe sensors behavior, but only one capturing the complexity of thesensor signals, and for the sake of completeness, we present somehints on the implementation of the sensors models.

Gas cell was considered to be 100 ml in volume, and contin-uously fed with a gas flow of 300 ml/min. Sensor chamber wasconsidered to act as a low pass filter for the gas concentrationswith time constant 20 s.

The heating function, was numerically implemented using avoltage sawtooth function constrained between 0.2 and 0.9 V, dur-ing a period of tcycle = 10 s, producing a temperature excursion from200 to 450 ◦C in static conditions. Sensors were considered to have athermal time constant of �therm = 0.5 s. On the other hand, chemicalreactions at the sensor surface were considered to have a temper-ature dependent kinetics, being slow at low temperatures and fastat high temperatures. The overall contribution to the j sensor con-ductance (Gj) by the different analytes was combined according tothe expression 1.

Gj = G0,j(T)

(1 +∑

i

Si,j(T) × cˇii

)(1)

where G0,j is the static sensor conductance in air, Si,j is the sensitivityto gas i. ci is the concentration of gas i, T is temperature and ˇi is agas depending parameter with values around 0.5.

In order to simulate realistic noise sources, the followingassumptions were also made. Random gaussian noise with anamplitude of 2 ◦C was assumed for the sensors temperature, and

a 3% multiplicative random gaussian noise was assumed on thefinal conductance of each sensor. Signal conditioning was assumedto be obtained using a resistor half-bridge excited with 5 V anda loading resistance of 1 K�. Voltage was sampled with an 8-bitA/D converter using a 5 V reference introducing the correspondingquantization noise.

Sampling rate was also fixed to 10 Hz, thus resulting to an overallof 100 points/cycle. The synthetic experiment had a total durationof texp = 15 min, thus generating a total recording vector of 9000points for each sensor.

Following the general procedure presented in Section 2.1,synthetically obtained vectorized data were properly folded togenerate data matrices of 100 × 90 dimension, where the two direc-tions or ways of measurement of the data were sensitivity andconcentration for both datasets: air (A1, A2, . . . ,A11) and ternarygas mixtures (R1, R2, . . . ,R11). Each pair of matrices (air/gas mix-tures) were ratiored for each sensor (Xi = Ai/Ri), giving a new set ofdata matrices (X1, X2, . . . ,X11).

2.3. Experimental dataset

Experimental data was obtained using a set of four metal oxide(MOX) sensors (SB11A, SB15, SB31 and SB42A) provided by FIS Inc.(Hyogo Japan), built in a custom designed gas cell. In accordancewith their specifications, they are capable to detect hydrocarbons(FIS SB11A, FIS SB15), solvents (FIS SB31) and Hydrofluorocarbons(FIS SB42A). Data acquisition was performed using a DataTakerDT800 (dataTaker Pty Ltd., Australia) equipment, using a 2 Hz


Table 1Concentration of experimental gas mixtures.

CO (ppm) CH4 (ppm)

Standard 1 0 3600Standard 2 240 3000Standard 3 360 2400Standard 4 480 1800Standard 5 600 0

sampling rate. Sensors were operated under temperature mod-ulated conditions, applying a sawtooth waveform with a periodof tcycle = 20 s, by the application of different potentials within the0.1–0.8 V range, using a Hameg HM8142 (Hameg GmbH, Germany)adjustable power supply.

Under these sampling conditions, resistance signals wererecorded during texp = 80 min. Within this period of time, we intro-duced into the gas cell of several binary mixtures of two gases (COand CH4) in synthetic air. These mixtures were prepared at five con-centration levels as described in Table 1, delivering consecutivelysuitable pure gas amounts into the sensors chamber through a massflow controlled system.

Data preprocessing was done following the guidelines explainedin Section 2.1, now rendering in this case five datasets (R1,R2, . . . ,R4), with 272 × 40 dimension (concentration x sensitivity)for each sensor. The corresponding datasets were filtered with amedian filter of order 9 to remove spikes.

The response of the sensors to synthetic dry air using the sametemperature modulation was used to normalize the sensors con-ductance, rendering a set of 5 normalized conductance matrices foreach sensor (X1, X2, . . . ,X4).

3. Method

3.1. Multivariate curve resolution (MCR)

MCR-ALS (multivariate curve resolution-alternating leastsquares) [19,20], as stated above, looks for a bilinear data decom-position of the global response of each sensor (Xi) giving the purecontributions C (e.g. gas concentration profile) and S (e.g. sensorsensitivities) of each gas in the two modes of measurement.

X = CST + E (2)

In Eq. (2), X is the sensor response (conductance) data matrix, C isthe matrix containing for each component in the mixture how theconcentration changes over the course of the experiment and S thematrix of pure sensitivity profiles of the sensor for each gas compo-nent in the mixture during the temperature modulation cycle. Toachieve this bilinear decomposition of Eq. (2), an ALS algorithm isapplied, where the following two least-squares Eqs. (3) and (4) aresolved iteratively.

ST = C+X (3)

C = X(ST)+

(4)

The procedure is initialised using estimations about the pureprofiles, either the concentration profiles (C) or of the sensitivityprofiles (S). Several methods exist to provide initial estimationsof these profiles. Evolving factor analysis (EFA) [21], is commonlyemployed for time-evolving systems. The main idea of this methodis to examine the evolution and change in magnitude of the singularvalues associated with the components (factors) along the process.This provides an initial estimation of the analytes’ concentrationprofiles along the measure. Other possible alternatives, are basedin the detection of purest variables like in the simple-to-use inter-active self-modelling mixture analysis (SIMPLISMA) [22,23], oftenrecommended for the study of complex non-evolving systems [24].

Fig. 2. Data flow scheme of the followed MCR-ALS procedure.

For obtaining the initial estimations it is also necessary to knowthe number of factors present in the data. Assuming that most ofthe data variance is mainly due to the gas sensing process, principalcomponent analysis (PCA) [25] or the study of the singular valuescalculated by singular value decomposition (SVD) [26] are used todetermine this number.

Fig. 2 shows a scheme of the followed procedure. However,solutions obtained by MCR-ALS are not unique [27]. They presentrotation and intensity ambiguities. Rotation ambiguities refers tothe existence of different linear combinations of the true solutionfitting equally well the original data and fulfilling the constraintsapplied to the system. Intensity ambiguities are originated by theoccurrence of scale indeterminacy, which describe the original datawith the same fit.

In order to decrease the effect of rotation and intensity ambi-guities, Eqs. (3) and (4) are solved under a set of constraints,thus limiting the number of possible solutions. In this way, itis possible to tune the contribution of each pure profile, buttaking into account that the application of constraints must bedirectly linked to the physicochemical nature of every system.MCR-ALS can be also used as a quantitative tool for mixtures.This type of analysis is enabled by analyzing simultaneously theresponses of the standards and unknown samples profiles undersimilar conditions. Assuming a linear relationship between therelative areas/heights of the resolved profiles and the concen-trations of each component, MCR-ALS allows to obtain suitablecalibration curves and performs relative concentration estimations,which can be used for the quantification of unknown samples.Inclusion of some standards of known concentration into theresolution process is needed to perform accurate quantitative esti-mations.

ALS algorithms have important advantages. They are extremelyflexible towards the inclusion of constraints, such as nonnegativity,unimodality, equality and closure. In addition, they can be eas-ily adapted to the simultaneous analysis of several datasets. Bothadvantages are of utmost importance in the study of temperature-modulated gas sensors signals and for quantitative purposes. Inchemical sensing problems, the most common constraint is thatof the positiveness of the extracted concentration profiles. As such,


the use of this constraint is provided in most chemometric softwarepackages and it does not require specific expert knowledge.

On the other hand, uncertainty related to the obtained con-centrations and sensitivity profiles can be estimated for MCR-ALSmodels. Since solutions provided by MCR are often non-unique,uncertainty can be calculated in the form of feasible bands of solu-tions that fit the experimental data equally well. A method tocalculate these band boundaries has been proposed by Tauler [28].

3.2. Data preprocessing

As it was previously stated, in order to improve resolution, and toallow quantitative estimations, experimental data matrices can bearranged using different matrix augmentation schemes. Thus, thisprocess of matrix augmentation can be done in three sorts: columnwise (CW), row wise (RW) and column + row wise (CW + RW).

Column wise augmentation is performed as is stated in Eq. (5),where Xi are each one of the i (m × n) data matrices containingthe signal of each of the I sensors. As a result of this operation,the n number of columns is kept constant and the augmentationis achieved setting the Xi matrices one on top of each other. Thisprocedure gives a global column-wise augmented data matrix ofdimensions [m × I] × n.

XCW =

⎡⎣

X1...

XI

⎤⎦ =

⎡⎣

C1...

CI

⎤⎦ ST +

⎡⎣

E1...

EI

⎤⎦ (5)

Row wise augmentation is described in Eq. (6). In this case, the Jmatrices containing the data are arranged keeping constant the mnumber of rows of Xj. Thus, the augmentation is done setting eachXj data matrix one besides the other and the dimensions of the newrow-wise augmented matrix become m × [n × J].

XRW = [X1· · ·XJ] = C

⎡⎣

S1...

Sj

⎤⎦+ [E1· · ·EJ] (6)

Simultaneous augmentation in both row and column wisemodes, as is described in Eq. (7), is a combination of both typesof data arrangement. This procedure leads to a [m × I] × [n × J] aug-mented data matrix.

XA =

⎡⎣

X1,1· · ·X1,j...

XI,1· · ·XI,J

⎤⎦ =

⎡⎣

C1...

CI

⎤⎦⎡⎣

S1...

SJ

⎤⎦

T

+

⎡⎣

E1,1· · ·E1,j...

EI,1· · ·EI,J

⎤⎦ (7)

4. Results and discussion

4.1. Simulated data

With the purpose of resolving the pure sensitivity and concen-tration profiles of one single sensor during the measurement cycle,its conductance response was studied by MCR-ALS. Thus, the matrixcontaining data from sensor 5 was selected. The 3D plot of this sen-sor’s response is shown in Fig. 3a. Its singular value decomposition(SVD) indicated the presence of 2–3 major components.

As stated before, MCR-ALS algorithm initialization needs aninitial estimation either of the sensitivity, S, or concentration Cprofiles. To generate these initial profiles, EFA or SIMPLISMA canbe used. Due to the time evolving characteristics of this dataset inparticular, these pure profiles were generated by EFA.

MCR-ALS algorithm was then initialized applying nonnegativ-ity constraints to sensitivity and concentration profiles, assumingthat the concentration profiles cannot take negative values, andthat all the gases produce increases in the sensor conductance, so

Fig. 3. (A) Plot of sensor 5 response, corresponding to synthetic data. (B) Plot of FISSB31 MOX gas sensor response obtained in the analysis of a gas sample of 240 ppmof CO and 3000 ppm of CH4.

sensitivities are always positive. The results obtained for sensor 5are shown in Fig. 4. They confirmed the presence of three species inthe measurement interval, with variable concentrations dependingon time. However, pure concentration profiles initially obtained inthis way did not appear to be properly resolved. Although two ofthe compounds were sequentially added, they emerged at nearlythe same sampling time, and consequently with a high degree ofoverlap. On the other hand, the study of the sensitivity profilesshows maxima in different positions for two of the gases, beingthe component 2 the one showing the major sensitivity at a highertemperature. For component 3, the corresponding sensitivity pro-file shows a decreasing profile in the coincidence region with thesensitivities of the other two factors, thus suggesting a certain non-linear effect. However, the main trends observed in the true profilesagreed with the sequential addition of the gases and with the dif-ferent sensitivities of the simulated sensors to each of the species.

MCR-ALS procedure was extended to the simultaneous analy-sis of the response signals from 11 sensors, as it is described inthe method section. To do it so, a new row-wise (RW) augmented


Fig. 4. (A) Pure sensitivity and (B) concentration profiles obtained from MCR-ALSresolution of data from sensor 5.

data matrix was properly set following a row wise augmentation(Eq. (6)) procedure. This process leaded to an augmented matrix ofdimensions (90 × 1100). As it can be seen in Table 2, the determina-tion of the number of factors was not simple, leaving this selectionto 2–3 factors.

MCR-ALS was applied to this augmented data matrix, assumingagain nonnegativity constraints to both modes and normalizationof the data. Initial estimates were taken in this case from the anal-

Table 2Singular values for each row-wise augmented dataset.

Factor Synthetic data.augmented.(11 sensors)

Real data.augmented.(4 sensors)

Real data. augmented.(4 sensors × 5mixtures)

Singular values1 1625.18 1455.64 2748.562 152.25 231.44 693.433 61.28 128.12 264.854 36.55 79.29 148.085 32.38 53.01 124.536 28.23 13.56 58.257 23.89 9.86 46.088 21.58 7.20 37.779 20.18 4.83 29.67

10 17.82 2.69 21.74

Fig. 5. (A) Pure sensitivity and (B) concentration profiles obtained by MCR-ALS reso-lution of data form sensors 1–11. (C) Theoretical time/concentration profile of gasesinto sensors chamber. Both profiles have been displaced along the gas concentrationaxis to allow a better visualization.

ysis previously done, choosing the concentration profiles resolvedafter the analysis by MCR-ALS of the sensor 5 data. Additionallyin this case, equality constraints were applied to one of the con-centration profiles. As it can be seen in Fig. 5a, the study of thepure sensitivity profiles showed important differences among the3 different gases. For instance, it is observed a reversed behaviorof the sensitivity of the different sensors for components 1 and 2for each of the 11 sensors. On the other hand, the sensitivity profilecorresponding to component 3, which corresponds to the gas keptat a constant concentration during all the experiment, shows anincreasing response with the sensor number. Both observed order-ings are a direct consequence of the synthetic nature of the dataand must not be taken into consideration.

The incorporation and simultaneous analysis of the informationcontained in multiple sensors, following the data augmentationstrategy proposed in Eq. (6), allowed an improved resolution of theconcentration profiles of the different gas species. As it can be seenin Fig. 5b and c, now MCR-ALS pure and theoretical concentrationprofiles were agreeing. This improvement in resolution should berelated to the different relative sensitivities to gas species provided


Fig. 6. (A) Pure sensitivity and (B) concentration profiles obtained from MCR-ALSresolution of experimental data from sensor SB42A, for CO (solid line) and CH4

(dashed line).

by each of the MOX sensors operated under temperature modu-lation. These changes of sensitivity give to MCR-ALS the ability toresolve and discriminate among the different analyte contributionsto the global signal.

4.2. Experimental data

As it was described above, several gas standard mixtures ofCO and CH4 in air, at different relative concentrations were ana-lyzed. First, experimental data from the analysis of standard 2 (seeTable 1) using sensor SB42A, with dimension 272 × 38 (concentra-tion × sensitivity) were studied.

The number of factors was estimated to be 2–3 by SVD. EFA wasused to generate initial estimations of pure concentration profiles.Nonnegativity constraints were applied, rendering the pure pro-files given in Fig. 6. Sensitivity profiles showed a good resolution,but a deeper investigation of them showed some inconsistencieswith their foreseen behavior. Actually, sensitivity towards CO wasshifted towards higher heater voltage values [28], being the peakposition at higher temperatures than the maximum sensitivitytowards methane. Concentration pure profiles showed a generalagreement with the expected profiles from the experimental con-ditions, and allowed the identification of both components (CH4and CO). However, the concentration curves showed importantdistortions. For instance the initial introduction of CO followedthe expected transient due to chamber dynamics. However, when

Fig. 7. Multisensors case. (A) Sensitivity and (B) concentration profiles, obtainedfrom 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).

CH4 was introduced, the CO concentration apparently decreased animportant amount, which obviously is an artifact of the technique.Similarly, when the flow of CO was stopped, the CO concen-tration apparently increased and this was compensated with afalse decrease of the CH4 concentration. Finally when CH4 flow isstopped both concentrations displayed the final decrease to zero.In other words, pure concentration profiles of CO showed someinhibition of the signal in the overlapping area, corresponding tothe time interval in which both gases were simultaneously presentin the sensors chamber. The origin of this inhibition is unknown,and we cannot guarantee that it is originated by a non-additive(non-linear) behavior of the sensor responses that cannot be accu-rately modeled by MCR-ALS. However, it is also necessary to remarkthat this resolution was achieved using the different sensitivitiesof only one sensor, and this problem of lack of resolution for asingle sensor was also detected in the analysis of the syntheticdataset.

With the aim of enhancing the resolution of the concentrationprofiles, data matrices containing responses of the four sensorsexposed to standard 2 were simultaneously analyzed in accor-dance, to obtain a multisensor row-wise augmented data matrix,as it is described in Section 3 (Eq. (6)). MCR-ALS resolution wasperformed onto this augmented matrix, using in this case ini-tial estimations based on the detection of the purest variables(SIMPLISMA), due to the increasing complexity of the dataset. Nor-


Fig. 8. Multisensors + multiset case. (A) Pure sensitivity and (B) concentration pro-files obtained from MCR-ALS simultaneous resolution of experimental data fromsensors (1) SB11A, (2) SB42A, (3) SB15 and (4) SB31, exposed to five standard sampleswith different concentrations in CO (solid line) and CH4 (dashed line).

malization of the sensitivity profiles and nonnegativity constraintsto both modes were applied.

Pure profiles finally resolved by MCR-ALS are given in Fig. 7.Comparing these profiles with the previous ones using a single sen-sor, it can be observed now how the simultaneous analysis of thefour sensors provide a much better resolution of the gas concen-tration profiles and also that they were closer to the expected ones,specially for CH4. In particular, CH4 concentration profile agreedwell with what was expected from experimental conditions. In thecase of CO, the obtained concentration profile did not completelyagree with the expected one, since discordances appeared in theoverlapping region (time ≈ 19–34 min). Even though, as in the pre-vious cases, it is possible to identify both gases from its deviationsof the baseline (starting times).

On the other hand, sensors sensitivities presented the sameshapes as in the previous case, but now with different relative inten-sities for each of the four SB sensors. Results show that sensor 2is more sensitive to CO than to CH4 while, in sensor 3 sensitiv-ities are reversed. Resolution for the selectivity profiles towardseach gas due to the inclusion of sensors of different characteristicsimproved significantly. However, this improvement of resolutionwas not enough to provide sensitivity profiles in complete agree-ment with the foreseen ordering of sensitivities along the heatervoltage cycle. Again, they still were not presenting the expectedbehavior for CO, appearing in a reversed order as in the previous

Fig. 9. Data flow scheme for quantitative estimation of the analytes’ concentration.

case. Even so, the maxima of the different sensitivity profiles werenearly to be superimposed in the same heater voltage region, withdifferent intensities. Although these profiles were now improved,they still were not presenting the expected shape.

To improve the resolution and to investigate the possibility ofperforming quantitative estimations, two more standard mixtureswere simultaneously analyzed under the same experimental con-ditions, according to the concentrations specified in Table 1. Inaddition to these mixture standards, two pure samples of both gaseswere also recorded. Experimental data matrices obtained in thesemeasurements were ordered following a CW + RW procedure, as itis expressed in Eq. (7), now building a new multisensor + multisetaugmented data matrix to be resolved by ALS.

Multiset experimental dataset resulted to be well describedusing two components (see Table 2, 4 sensors × 5 mixtures aug-mented matrix). As in the previous case, initial estimations for bothcomponents were calculated by detection of purest variables. Non-negativity constraint to the concentration profiles was added. Timeintervals of resolved concentration profiles, where one of the gasspecies is known to be not present, were also constrained accord-ingly. MCR-ALS solution provided improved estimations of the pureprofiles for both concentration and sensitivity modes.

When the resolved profiles were compared, it was found thatthe different contributions from both individual gases, describedby the concentration and sensitivity profiles, were now in agree-ment with the expected ones. These shapes can be seen in Fig. 8. Inthe lower part of the figure in can be seen how the concentrationprofiles are showing distortions in the rising part of the peaks andon its overlapping regions (especially in the last mixture importantdistortions of both concentration profiles are observed). This meansthat the augmentation of the data matrix with additional record-


Fig. 10. Preliminary quantitative results obtained in the analysis of (A) CH4 and (B)CO mixtures.

ings and the consrtaints on the pure gases, still do not accomplishthe full resolution of the sensitivity and concentration profiles.

A little improvement of the results was obtained in the sensitiv-ity profiles, as is shown in the upper part of Fig. 8. In this case, the useof the proposed data augmentation strategies incorporating puregas responses, allowed the discrimination of the responses towardsboth gases in each sensor. Now, pure sensitivity profiles agreed bet-ter with the empirical description of these sensors provided by themanufacturer, pointing out the increased sensitivity to CO of MOXsensors at lower temperatures. The MCR-ALS resolved sensitivityprofiles showed also a differentiation on the relative intensities ofthe different peaks, thus reflecting the different sensitivities of thefour sensors towards both gas species. While the technique recov-ers the overall shape of the concentration profiles, some artefactsappear when both gases are present simultaneously. This is clearlydue to the non-linearity of the sensor response towards mixtures.

4.3. Quantitative model

To create a quantitative model of the concentrations of the stan-dards in the process, the area of the concentration profiles is useddue to its higher robustness (compared with the instantaenousvalues) towards the distortion observed in the peaks. Therefore,areas from resulting MCR-ALS concentration profiles were firstcalculated. Then, a linear calibration curve was obtained for eachcompounds’ (CH4 and CO) profiles, relating concentration profiles’areas and known concentration values (Fig. 9). Finally, these twolinear calibration models were used to estimate the concentrationsof both analytes in all samples and results are shown in Table 3.Since the number of available standards was quite low, the calibra-tion curves were calculated using all of them.

Table 3Quantitative results. Relative error and root mean standard errors of calibration forboth analytes.

Reference (ppm) Calculated (ppm) Relative error (%)

CH4 CO CH4 CO CH4 CO

0 0 197.30 15.78 – –1800 240 1330.44 254.08 −26.09 5.863000 360 3224.84 299.43 7.49 −16.823600 600 3647.41 630.71 1.32 5.12

RMSEC (CH4) 279.38RMSEC (CO) 35.36

Acceptable calibration errors are shown in Table 3, where itcan be seen that the lowest errors are obtained for the highestconcentrations of analytes. On the other hand, calibration models(Fig. 10) present high correlation coefficients (R > 0.95). Althoughthese results are good, we need more samples to build a betterregression model and to enable external validation.

5. Conclusions

As is described in previous works, the use of temperature mod-ulation techniques in MOX sensors can provide changes in itsspecificity that allow a differentiation in the response of thesedevices towards different gas species. To deal with these data,appropriate signal processing techniques must be employed.

MCR-ALS has been shown to be a good resolution technique toextract useful information from temperature-modulated gas sen-sor signals. On these systems, resolution can be approximatelyachieved with a single sensor by virtue of the sensitivity depen-dence on temperature. Resolution can be enhanced provided thata sufficient number of MOX sensors with different sensitivitiesis available. In this case, sensitivity profiles from different gasesand their concentration profiles in mixtures can be appropriatelyresolved. In the cases in which the number of sensors is limited,such in the real one, this resolution can be further enhanced withthe inclusion of mixture samples at different concentration lev-els into the resolution procedure. However, as seen in the studiedcase, these profiles can appear slightly distorted in the overlap-ping regions, due to effects not effectively modeled by MCR-ALS.This can be due to the similarity of the sensitivity profiles, ormaybe due to sensor non-linearities. A deeper study on theseeffects should need further efforts that are beyond the scope of thiswork.

Aside from resolution purposes of both sensitivity and concen-tration profiles in this kind of signals, the use of the proposedapproach allows a first attempt in the quantification of gasmixtures. The use of MOX pulsed gas sensors plus MCR-ALS res-olution becomes then a very promising line of application in thisdirection.

Acknowledgements

Authors acknowledges to GOSPEL NoE FP6-IST 507610 for theirsupport and CyCIT project TEC2004-07853-c02-01.

References

[1] I. Simon, N. Barsan, M. Bauer, U. Weimar, Micromachined metal oxide gas sen-sors: opportunities to improve sensor performance, Sens. Actuators B 73 (2001)1–26.

[2] U. Weimar, W. Göpel, A.C. measurements on tin oxide sensors to improveselectivities and sensitivities, Sens. Actuators B 26 (1995) 13–18.

[3] S. Zampolli, I. Elmi, J. Stürmann, S. Nicoletti, L. Dori, G.C. Cardinali, Selectivityenhancement of metal oxide gas sensors using a micromachined gas chromato-graphic column, Biosens. Bioelectron. 20 (2005) 1968–1976.

[4] G. Eranna, B.C. Joshi, D.P. Runthala, R.P. Gupta, Oxide materials for developmentof integrated gas sensors—a comprehensive review, Crit. Rev. Solid State Mater.Sci. 29 (3/4) (2004) 111–188.

[5] D.P. Mann, T. Paraskeva, K.F.E. Pratt, I.P. Parkin, D.E. Williams, Metal oxidesemiconductor gas sensors utilizing a Cr–zeolite catalytic layer for improvedselectivity, Meas. Sci. Technol. 16 (2005) 1193–1200.

[6] High performance solvent vapor identification with a two sensor array usingtemperature cycling and pattern classification, Sens. Actuators B: Chem. 95(1–3) (2003) 58–65.

[7] A.P. Lee, B.J. Reedy, Temperature modulation in semiconductor gas sensing,Sens. Actuators B 60 (1999) 35–42.

[8] H. Ding, H.F. Ge, J.H. Liu, High performance of gas identification by wavelettransform-based fast feature extraction from temperature modulated semi-conductor gas sensors, Sens. Actuators B: Chem. 107 (2005) 749–755.

[9] L. Ratton, T. Kunt, T. McAvoy, T. Fuja, R. Cavicchi, S. Semancik, A comparativestudy of signal processing techniques for clustering microsensor data (a firststep towards an artificial nose), Sens. Actuators B 41 (1–3) (1997) 105–120.


[10] S. Nakata, S. Akakabe, M. Nakasuji, K. Yoshikawa, Gas sensing based on a non-linear response: discrimination between hydrocarbons and quantification ofindividual components in a gas mixture, Anal. Chem. 68 (1996) 2067–2072.

[11] A. Heilig, N. Barsan, U. Weimar, M. Schweizer-Berberich, J.W. Gardner, W. Göpel,Gas identification by modulating temperatures of SnO2-based thick film sen-sors, Sens. Actuators B 43 (1997) 45–51.

[12] B. Yea, T. Osaki, K. Sugahara, R. Konishi, The concentration–estimation ofinflammable gases with a semiconductor gas sensor utilizing neural networksand fuzzy inference, Sens. Actuators B 41 (1997) 121–129.

[13] D.L. Massart, B.G.M. Vandeginste, L.M.C. Buydens, S. De Jong, P.J. Lewi, J.Smeyers-Verbeke, Handbook of Chemometrics and Qualimetrics, Part B, Else-vier, Amsterdam, 1997, p. 243.

[14] A. Cichoki, R. Zdunek, A.H. Phan, S.I. Amari, Non-negative Matrix Factoriza-tions: Applications to Exploratory Multi-way Data Analysis and Blind SourceSeparation, Wiley, Chichester, 2009.

[15] C. Di Natale, E. Martinelli, A. D’Amico, Counteraction of environmental dis-turbances of electronic nose data by independent component analysis, Sens.Actuators B 82 (2002) 158–165.

[16] M. Kermitt, O. Tomic, Independent component analysis applied on gas sensorarray measurement data, IEEE Sensor J. 3 (2003) 218–228.

[17] P.K. Clifford, D.T. Tuma, Characteristics of semiconductor gas sensors I. Steadystate gas response, Sens. Actuators 3 (1982–1983) 233–254.

[18] P.K. Clifford, D.T. Tuma, Characteristics of semiconductor gas sensors II. Tran-sient response to temperature change, Sens. Actuators 3 (1982–1983) 255–281.

[19] R. Tauler, D. Barceló, Multivariate curve resolution applied to liquidchromatography—diode array detection, Trends Anal. Chem. 12 (1993)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.

[23] W. Windig, Self-modeling mixture analysis of spectral data with continuousconcentration profiles, Chemom. Intell. Lab. Syst. 16 (1992) 1–16.

[24] E. Peré-Trepat, A. Hildebrandt, D. Barceló, S. Lacorte, R. Tauler, Fast chro-matography of complex biocide mixtures using diode array detection andmultivariate curve resolution, Chemom. Intell. Lab. Syst. 74 (2004) 293–303.

[25] I.T. Jolliffe, Principal Component Analysis, 2nd ed., Springer, New York, 2002.[26] G.H. Golub, C.F. Van Loan, Matrix Computations, 2nd ed., Johns Hopkins Uni-

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.

Biographies

Ivan Montoliu was born in Barcelona in 1969. Degree(1995) and Ph.D. in Chemistry (2001) at the UniversitatAutònoma de Barcelona, doing research in developmentand application of chemometrical methods to agroali-mentary analysis. In 2000 joins the official distributor ofDionex products in Spain, as chromatography applicationsand software specialist. From 2004 to 2007 he had a postdoc Position at the Department of Electronics at the Uni-versity of Barcelona. From 2008 he is with the NestleResearch Center in Laussanne, Switzerland. His areas ofinterest are Chemometrics and its application to signalsof different analytical instrumentation.

Roma Tauler obtained his Ph.D. in Chemistry at the Uni-versity of Barcelona in 1984. He is research professorat the Institute of Environmental Assessment and WaterResearch (IDÆA), CSIC, in Barcelona (Spain), since July2003. Previously he was associate professor of AnalyticalChemistry at the Analytical Chemistry Department of theUniversity of Barcelona during 1987–2003. He performedpostdoct stays at Institut fu¨r Anorg. u. Anal. Chemie,Univ. of Innsbruck, Innsbruck (Austria) in 1985 and 1989and a year’s sabbatical as a research scientist at the Cen-ter for Process Analytical Chemistry (CPAC), Departmentof Chemistry, University of Washington, Seattle, USA, in1992. At present, he is the Editor in Chief of the journal

Chemometrics and Intelligent Laboratory Systems He has published more than 200research papers, most of them in the field of chemometrics and its applications,and in particular in the development and applications of new multivariate reso-lution methods. In the recent years he has focused more on the investigation ofenvironmental problems.

Marta Padilla was born in Sta. Cruz de Tenerife, Spain, on1973. She received a degree in Physics from the Universityof La Laguna in 1999 and a degree in Electronic Engineer-ing from University of Barcelona in 2003. In the same yearshe became a Ph.D. student in Electronic Engineering inthe University of Barcelona.

Antonio Pardo received his diploma in Physics 1991 andhis Ph.D. in 2000 from the University of Barcelona. Duringhis Ph.D. studies he worked in system identification withapplications in gas sensor systems. Since 2005 he is asso-ciate professor at the University of Barcelona. His researchinterest focused on signal processing for gas sensors andpattern recognition as well as on hardware and softwaredevelopments for electronic nose instrumentation.

Santiago Marco is associate professor (Profesor Titu-lar) at the Departamento d’Electronica of Universitat deBarcelona since 1995. He received the degree in Physicsfrom the Universitat de Barcelona in 1988. From 1990 to1993, he was regular visitor of the Centro Nacional deMicroelectrònica, Bellaterra, Spain. In 1993, he receivedhis Ph.D. (honor award) degree from the Departamentode Física Aplicada i Electrònica, Universitat de Barcelona,for the development of a novel silicon sensor for in vivomeasurements of the blood pressure. In 1994, he wasa post doc researcher at the Department of ElectronicEngineering, Universita di Roma ‘Tor Vergata’, working indata processing for artificial olfaction. In 2004 he had a

Sabbatical Year at EADS Corporate Research in Munich working in Ion Mobility Spec-trometry. He has published about 70 papers in scientific journals and books, as wellas more than 130 conference papers. His current research interest is the applica-tion of signal and data processing to smart chemical instrumentation. Since 2008 heleads 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 alternative for gas analysis than conventional analytic instruments. .However they are prone to degradation because of sensor poisoning and drift. Statistical methods like Principal Component Analysis (PCA) and Partial Least Squares (PLS) have been proved to be very useful in the task of fault diagnosis of malfunctioning sensors. In this work we test the effectiveness of several techniques based on PCA and PLS on faults caused by sensor poisoning. These techniques will be evaluated on a dataset composed by the signals of 17 conductive polymers gas sensors measuring three analytes at several concentration levels. These techniques will be evaluated concerning their capabilities to detect the fault, identify the faulty sensor and correct their signal.

Keywords – fault diagnosis, gas sensor array, poisoning.

I. INTRODUCTION

The use of chemical gas sensors in an array format with a pattern recognition system improves the degree of selectivity of the individual gas sensors. In a sensor array each sensor has a partial specificity; the sensor array output produce a 'compound/ odour' fingerprint that can be identified by the pattern recognition system. A sensor array is the base of the instrument known as 'electronic nose' which is inspired by the mechanisms involved in human olfaction. Electronic noses have been used for the analysis of volatile organic compounds in applications that range from the food or medical industry to environmental monitoring and process control. Nevertheless problems like poisoning or drift degrade the ability of the instrument to recognize odours or gases. When using an array of sensors one or more of them can fail due to a great variety of faults like sensor failure, degradation, external interferences, bias, noise, occasional outliers (noise spikes). While these problems depend on the technology and fabrication process and may be overcome by technology

improvements, although signal processing can help to detect and compensate the output signal of the affected sensor.

Poisoning consists on a degradation of the sensing properties of a sensor when it is exposed to a gas compound that irreversibly interacts with its sensing material. It is not an easily detectable problem because the sensors keep their responses within normal levels, although with an altered profile of sensitivities to the different analytes. Poisoning is a surface phenomenon and is closely related to physical properties of the sensing material and the detecting gas, so for heterogeneous arrays it can affect only some of the sensors in an array, which have to be immediately detected and identified in order to correct their responses until they can be replaced by new ones.

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 time and the concentration of the poisonous substance [3], for example HMDS[3], SO2 [2], NO2 [1] or O3 [2] have been reported as poisoning gases for different technologies.

Event or fault detection on gas sensor arrays has been previously tackled by other authors. Perera et al. [4] presented an adaptive method based on a recursive dynamic principal component analysis (RDPCA) algorithm. This algorithm was applied on real signals arising by oil vapor leakages in an air compressor. Pardo et al. [5] used the correlation among the responses of five gas sensors detect a possible malfunctioning of one of the sensors during continuous operation. The outputs of the sensors were estimated as a function of the outputs of the remaining ones by Artificial Neural Networks.

On the other hand, the employed techniques of fault diagnosis can be based on statistical methods. Fault detection using Principal Component Analysis (PCA) or Partial Least Squares (PLS) has already been studied in applications like process monitoring, quality control, or sensor and process fault diagnosis [6]. Also modifications of PCA have been

1-4244-0830-X/07/$20.00 ©2007 IEEE.

developed to analyze non steady data; non-linear, multiblock, multiway [7] and dynamic PCA for example [8].

Dunia and Qin [9]studied the sensor and process fault detection, identification and reconstruction based on PCA. In a posterior work [10], they propose a method based in the use of structured residuals with maximized sensitivity and different indices for detecting four types of faults.

In this work we test three techniques: PCA, PLS and structured residuals of PCA with maximized sensitivity (PCA-MS), to detect faults, identify the faulty sensor and correct its response. The comparison among the algorithms will be made using three different criteria: The amount of false and true alarms by means of ROC representation (Receiver Operating Characteristic) to test their detection ability, rate of success on faulty sensor identification and percentage of success of classification on eight classes after correction of malfunctioning sensors.

In section II, the application of the three statistical methods on fault diagnosis is briefly presented. The dataset and the details of the procedure to simulate a 'sensor poisoning' fault is described in section III, Results are presented in section IV, followed by section V in which some conclusions are derived.

II. STATISTICAL METHODS FOR FAULT DETECTION, ISOLATION AND CORRECTION OF FAULTY SENSORS

PCA and PLS are very efficient in fault diagnosis when many correlated sensors are available. Sensor redundancy allows these methods to extract information about the data from the high correlation among the sensors responses, without the need of any model about the system or the process being measured.

A. Partial least squared (PLS) approach

The response of each sensor is predicted with the remaining sensors using a PLS model, then as many models as sensors are in the system are built using calibration data. To detect a faulty sensor the response of every sensor in the validation set is predicted with each one of the models. Then the real response of every sensor is compared to the predicted one in the validation set. If the sensors are unaltered, there should not be too much difference between the predicted and the real response. Therefore high residuals would denote changes on its normal response in a sample. Error plots (error2) for each one of the sensors can then be used as indicators of fault.

Once the error values are obtained, it is necessary to define a threshold value to define the conditions for the optimal detection of the faulty sensors. It has to be noted that the error values of all sensors increase when a fault occurs. The biggest increase obviously corresponds to the error of the faulty sensor. When two faulty sensors are present the increase of the error values of the remaining sensors will be bigger, then the chosen threshold for one faulty sensor will not be suitable. If the size of the fault is not big enough, as it is our case, it will be harder to identify two faulty sensors failing at the same time with this technique.

Therefore, if more than one faulty sensor has to be detected, the first faulty sensor detected has to be removed and the same procedure for the detection of the second faulty sensor has to be followed on the remaining set of sensors. Like this, a second malfunctioning sensor is identified and corrected and the resulting corrected signals are used to correct the signal of the first faulty sensor. This can be made for detecting several malfunctioning sensors but in every step the precision on the sensor correction degrades and the amount of calculus to be done increases .Another way to proceed in the case of the detection of two faulty sensors is to build models for every combination of two sensors, thus ( )

2N

models. This procedure is unpractical for three or more faulty sensors.

Fig.1. PLS predicts the response of one sensor from the response of the

remaining sensors. An error is obtained comparing the predicted with the actual response.

B. PCA approach

Dunia and Qin [9] give a detailed description of the use of Principal Component Analysis (PCA) for fault diagnosis. Here a brief summary is presented.

A raw measurement by N sensors is represented by a vector in a N dimensional space where the sensors are the axis with coordinates ξj=[0 0 ... 0 1 0...0], thus a vector of zeros with 1 at position j representing sensor j. In the original sensor space a faulty sample x is represented by:

jfxx ξ+= * (1)

where x* is the sample at normal conditions, which is unknown when a fault occurs. f is a scalar that represents the magnitude of the fault and ξj is a fault direction given by the sensors axis.

PCA decompose the original sensor space X into two orthogonal subspaces; model X and residual subspace X~ .

The detection of a fault in one sample x is based on its SPE value (Squared Prediction Error) which represents the magnitude of the projection of the sample on the residual subspace x~ :

2~xSPE = (2) A change in the correlation among the variables would

increase SPE value over a certain threshold. This would indicate the detection of a fault in a sample.

Sie →Error

),...,,...,(ˆ 111 niiii ssssfs +−=i=1, .., 17

± Ŝi

To identify the faulty sensor, the faulty sample has to be reconstructed along every direction ξj. The reconstruction of the faulty sample x in the direction of ξj is given by xj:

jjj fxx ξ−= (3)

where fj is an estimate of the fault magnitude in the direction ξj. The reconstructed vector xj is obtained minimizing the SPE of sensor j response for a sample xj (SPEj). The minimization of SPEj gives the value for fj:

2~jj xSPE = ,

2~/~0~/ jT

jjjj xffddSPE ξξ=⇒= (4)

where x~ , jx~ and jξ~ are the projection of the sample x, xj, and ξj respectively on the residual subspace.

Therefore the corrected component xj is the one that has the minimum size of its projection on the residual space. Then the relation SPEj/SPE (j=1,..,N) is calculated for the detected faulty sample. If sensor j is faulty its reconstruction will give a small SPEj, then SPEj/SPE will be close to zero, since SPE increases when a fault occurs. If sensor j is not faulty its reconstruction will not reduce SPEj, then SPEj/SPE will be close to 1. Therefore SPEj/SPE will present the minimum value for the direction that corresponds to the faulty sensor.

The technique for an optimum choice of the number of principal components for every PCA model is also explained by Dunia and Qin and can be found in [11].

To detect two faulty sensors the first faulty sensor detected has to be removed and a new model has to be rebuilt with the rest of the sensors. Then the same technique of detection and identification of a second faulty sensor has to be applied.

C. PCA structured residuals with maximum sensitivity approach (PCA-MS)

PCA with maximum sensitivity (PCA-MS) is a technique that directly uses the residuals of PCA. A method based on structured residual was introduced by Gertler et. al. in 1990 [12], and Qin and Li [10] proposed an extension of this method. The aim is to structure the residuals of a linear method like PCA in such a way that one residual present most sensitivity to one subset of faults while being insensitive to other faults. This is done by means of a transformation matrix, W which converts the primary residuals e of PCA into structured residuals r with:

Wer = (6)

The primary residuals e for a sample x with a fault e,

according to eq. (1), are given by:

iiii fbefBBxeBx +=+== ** ξ (7)

where bi is a column of B and the primary residuals e* are obtained applying the model matrix B to the training data x*.

The model matrix B represents the PCA matrix loadings of the model in the residual space.

The detection of a faulty sensor can be made using SPE index already explained for PCA or using d index, which is obtained from PCA residuals by:

eRed eT 1−= (8)

where Re is the covariance matrix of e*. This index d follows a Chi-square distribution when no fault is present.

This method allows the detection and identification of one or more faulty sensor in a single step; without having to remove a first faulty sensor detected and having to build a new model with the rest of the sensors. Then it is possible to know if there are one ore two malfunctioning sensors depending on the design of the structure of the transformed residuals.

To determine a suitable matrix W, a structure or incidence matrix has to be first designed. In this structure each fault has a characteristic response, a fault code. Then, in the incidence matrix, the rows are the residual structures while the columns are the fault codes. A 1 element means that the concerned residual does respond to the concerned fault while a 0 means it does not. It is possible to design a structure for a single fault or for multiple faults such as double or triple. In this case the incident matrix has to be carefully designed. More detailed explanations and conditions about structured residuals can be found in [13].

To calculate W relations between its components, vectors wi, and faults, bj, have to be established. The traditional structured residuals approach [12] chooses the vectors wi of matrix W to be insensitive to certain faults and sensitive to others, but this wi is not unique and doesn't maximize the response to those faults. Qin and Li propose [10] an algorithm (SRAM) that makes the residuals respond with maximum sensitivity to the considered faults being insensitive to the remaining ones. Mathematically it is equivalent to:

∑=≠

n

jjj

Tiij

bbw1

2)/(max with 0=iTi bw , 1=iw (9)

where wi is the ith row of matrix W and ri an ith element of r. Geometrically this means that wi and bi are orthogonal and the angle between wi and other faults directions bj (i≠j) is minimum. Finally, vector wi is given by the eigenvector that corresponds to the largest eigenvalue of:

iiT

ii wwBB λ=00 where 020 )/( BbbbIB jT

iii −= (10)

Later, some modifications have been introduced, for

example Li and Shah [14] use all the eigenvectors in eq. (10) 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 maximizing the sensitivity of the squares of fault responses as SRAM does. This technique is called Max-min method and it

is more powerful than SRAM because it maximizes the sensitivity of the response to the weakest of the faults of interest.

The obtained structured residuals are now used to build scalar fault identification indices which will indicate which sensor fails. In [10] Qin and Li investigate several types of indices but here we only consider one of them, given by ri

2. If the identification index is Ii= ri

2/ ηi , being ηi a confident limit determined by statistical techniques, it will be less than one in normal conditions and, if sensor i is faulty, it will be less than one but all the others Ij, with j≠i, will be larger than one.

Later, once the faulty sensor is identified, the fault can be estimated by minimizing the magnitude of the error of the corrected sample.

III. SENSOR DATA DESCRIPTION AND SIMULATION OF SENSOR POISONING

Measurements of three analytes at several concentrations levels (table I) with an array of 17 conducting polymer sensors were collected during one month. The total number of measurements in the dataset is 684 samples assigned to 8 classes, which have been alternatively measured over the whole duration of the experiment. The classes correspond to solutions in water of Ammonia and Propanoic acid, each one of them at 3 different concentration levels, and a third analyte (n-Butanol) at two concentration levels.

TABLE I.

SAMPLE DISTRIBUTION AMONG ANALYTES AND CONCENTRATION LEVELS

Analyte Ammonia Propanoic acid n-Buthanol

conc1 0,01 % 0,01 % 0,1 %

conc2 0,02 % 0,02 % 1,0 %

conc3 0,05 % 0,05 %

Figure 2 shows the sensor responses during a sampling

process. The sensor conductance increases and when the gas is removed it returns to the baseline. From this transient, only the maximum value is retained for further processing. At the considered point of the transient signals the response of every sensor has different values which together configure a characteristic pattern for every class (fig.3).

To make an exploratory analysis of the dataset a PCA was applied on the complete dataset after being autoscaled. The scores are shown in figure 4. In this dataset sensors are very correlated since only four PCs captures 99,30% of the total variance, and the first two PCs capture 96%. In the PCA scores plot, the scores in different colors and shapes according to the eight classes are shown. It can be seen a mixture of some classes, also several classes are scattered due to the effect of intra-class variability and drift.

A. Simulation of sensor poisoning fault

The usual types of faults considered in literature are those that affect sensors in general industrial process control

systems. These are commonly classified in four types; precision degradation, bias, drift and complete failure. The fault we consider in this work is characteristic of chemical sensors which interacts directly with the measured substance. The poisoning of the sensor sensing element would lead to a change in its sensitivity that can be not easily to detect, since the sensor behaviour is apparently normal.

0 20 40 60 80 100 120 140 160 180 200-2

0

2

4

6

8

10

12

time (s)

Co

nd

uct

ance

(1/O

hm

)

Transient response of sensor 1 per class

ammonia 0.01%

ammonia 0.02%ammonia 0.05%propanoic 0.01%

propanoic 0.02%propanoic 0.05%n-buthanol 0.1%

n-buthanol 1%

Fig. 2. Transient signals of sensor 1 responses to all analytes at different

concentration levels. The red dotted line shows the signals values at (aprox.) the maximum of the transient signals (time=186s).

1 2 3 4 5 6 7 80

2

4

6

8

10

12

14

clases

Co

ndu

ctan

ce (1

/Oh

m)

Pattern response at t=183s for all clases

Fig.3 Examples of patterns corresponding to samples of 8 different

classes.

-5 0 5 10

-1

0

1

2

3

4

5

91.2% PC1

5.1%

PC

2

PCA scores

ammonia 0.01%

ammonia 0.02%ammonia 0.05%

propanoic 0.01%propanoic 0.02%

propanoic 0.05%n-buthanol 0.1%n-buthanol 1%

Fig.4 PCA scores of the dataset.

In this work an abrupt but slight sensor poisoning fault is

simulated. After the fault the sensitivity pattern of the sensor to the different analytes is altered. To synthetically create a faulty sample, the response of the faulty sensor to a sample

belonging to certain class is changed by its response to any other of the remaining classes randomly selected (Fig. 5).

12 12

2

4

6

8

10

12

14

sensors

Condu

ctan

ce (1

/Ohm

)

Simulation of sensor 12 poisoning

ammonia 0.01%

n-buthanol 1%

Fig. 5. Simulation of sensor 12 poisoning fault

IV. RESULTS

In this section results from comparison between PCA, PLS and PCA-MS is presented regarding three aspects; detection of faulty samples by means of ROC curves, identification of faulty sensor/s by success rate on correct faulty sensor identification and correction of faulty sensors response by classification into eight classes after correction of faulty sensors.

These three capabilities are evaluated on one validation set with 342 samples and 50% of the samples with induced faults in one and two sensors failing at the same time. The training set contains also 342 samples.

A. Faulty samples detection:

The comparison on the detection ability of these three methods is made by ROC curves (Receiver Operating Characteristic) which represents true alarms vs. false alarms varying thresholds of the corresponding detection index (SPE for methods based in PCA or error2 for PLS). An ideal ROC curve would show maximum 100% of true alarms with minimum 0% of false alarms, therefore the curve would have the shape of a right angle with its corner in the upper left side.

In fig. 6 it can be compared the faulty samples detection ability by means of ROC curves for each of the three methods. In a region between 0 and 10 false alarms, PLS presents a higher number of true alarms than the other methods. This region corresponds to the higher values of the detection index. In a region in the middle, between 10 and 50 false alarms, PLS shows worse ability in detection than PCA and PCA-MS but in the last region, above 50% of false alarms, all methods present similar performance. PCA and PCA-MS show similar ability in detecting faulty samples in all regions.

B. Faulty sensor identification:

For the identification of one or two faulty sensors several structures of the incidence matrix have been tested. The first one is a 17x17 matrix of 1 with 0s in its diagonal and it is

suitable for one faulty sensor identification, since every row contains only one 0. Once all residuals have been calculated the minimum of them would directly identify the faulty sensor. Table II is used to identify two faulty sensors. Every row contains two 0 and there are as many rows as combinations of two of the 17 sensors, ( )

217 , therefore the

dimension of this matrix is 136x17. The last incidence matrix studied should be able to identify one or two faulty sensors. Its dimension is 17x17 and every row contains nine 0. It has been found that this structure cannot identify the faulty sensor or sensors because the sensor poisoning fault simulated in this work is not large enough. Further work has to be done in order to find an incident matrix capable to know how many sensors are failing and to identify them.

0 10 20 30 40 50 60 70 80 90 1000

10

20

30

40

50

60

70

80

90

100ROC

fp (%)

tp (

%)

pca

pcarespls

Fig.6. ROC representation for the detection index that presents the best ROC

curve for PCA, PCA-MS and PLS approaches.

TABLE II.

INCIDENT MATRIX FOR TWO FAULTY SENSORS IDENTIFICATION

residual\fault F1 F2 F3 ... F15 F16 F17 R1 0 0 1 ... 1 1 1 R2 0 1 0 ... 1 1 1 ... ... ... ... ... ... ... ...

R135 1 1 1 0 1 0 R136 1 1 1 ... 1 0 0

Also several algorithms for maximizing the sensitivity to

faults have been tested; SRAM proposed by Qin and Li [10], its extension by Li and Shah [14] and the Max-min algorithm by Xu and Kwan [15]. The best results were obtained with the extension of SRAM, but also Max-min algorithm outperforms the original SRAM.

Fig. 7 shows a comparison among PCA, PLS and PCA-MS using the corresponding incidence matrix for the detection of one or two faulty sensors. All three methods are very good in identifying one faulty sensor, in the case of two faulty sensors the best one is PCA-MS but also PLS results are close to 100% on success rate. It has to be noted that 88% of times PLS was able to correctly identify the second faulty sensor in the first step; without having to remove the first faulty sensor to identify the second one.

PCA PLS PCA-MS0

10

20

30

40

50

60

70

80

90

100

Su

cces

s ra

te (%

)

Success rate on correct faulty sensor identification

1 fault

2 faults

Fig.7. Percent of accuracy in one and two faulty sensors identification for

PCA, PLS and PCA-MS approaches.

C. Faulty sensor correction

After correcting the faulty sensors a reduction of dimensionality by PCA, from dimension 17 to 5, has been done. Later, a classifier (kNN with 3 nearest neighbors) has been applied in order to compare the correction ability of the algorithms. From figure 8 it can be seen that all methods present similar performance in correcting one faulty sensor signal, when two faulty sensors have to be corrected all the methods degrades and PCA-MS is the one that degrades the most. In any case the correction of the faulty signals clearly improves classification rates.

no fault 1 fault 2 faults0.91

0.92

0.93

0.94

0.95

0.96

0.97

0.98

0.99

1

CR

Normalized classification rates

non correctedPCA

PLS

PCA-MS

Fig.8. Normalized classification rates for the corrected responses by the three methods, having one and two faulty sensors and compared with non corrected

and no fault cases.

V. CONCLUSIONS

In this work the performance of PCA, PLS and PCA-MS for sensor poisoning fault diagnosis have been compared. These techniques have been applied on a dataset consisting on the measurements of three analytes at several concentration levels (8 classes) by 17 conducting polymer sensors. Sensor poisoning fault has been simulated by changing the faulty sensor response to one of the 8 classes by its response to any other class.

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

very good performance in identifying one faulty sensor and PCA-MS and PLS outperform PCA in the case of two faulty sensors failing at the same time. Finally the correction abilities of all methods are similar and necessary in order to improve results in classification.

PCA-MS is the most interesting technique because it is able to identify several faulty sensors in one step. It saves computation time since all the calculations are done in the previous calibration step.

ACKNOWLEDGMENTS

We want to acknowledge the partial funding of this

work by the Spanish project MCYT TEC2004-07853-C02-01 and GOSPEL IST 507610 Fp6.

REFERENCES

[1] B. Ruhland, Th.. Becker and G. Müller, "Gas-kinetic interactions of

nitrous oxides with SnO2 surfaces", Sensors and Actuators B, vol 50, pp. 85–94, 1998.

[2] L. Mazet, C. Varenne, A. Pauly, J. Brunet and J.P. Germain, "H2, CO and high vacuum regeneration of ozone poisoned pseudo-Schottky Pd–InP based gas sensor", Sensors and Actuators B, vol 103, pp. 190–199, 2004.

[3] M. Matsumiya, W. Shin, F. Qiu, N. Izu, I. Matsubara and N. Murayama, "Poisoning of platinum thin film catalyst by (HMDS) for thermoelectric hydrogen gas sensor", Sensors and Actuators B, vol. 96, pp. 516–522 , 2003.

[4] A. Perera, N. Papamichail, N. Barsan, U. Weimar, S. Marco, “On-Line Novelty Detection by Recursive Dynamic PCA and Gas Sensor Arrays Under Drift Conditions”, IEEE Sensors Journal, Vol. 6, pp.770-783, 2006.

[5] M. Pardo, G. Faglia, G. Sberveglieri, M. Corte, F. Masulli and M. Riani, "Monitoring reliability of sensors in an array by neural networks", Sensors and Actuators B, vol. 67, pp. 128-133, 2000

[6] J. F. MacGregor and T. Kourti “Statistical process control of multivariate processes“, Control Eng. Prac. Vol. 3, pp. 403-414, 1995.

[7] T. Kourti, P. Nomikos and J. MacGregor “Analysis, monitoring and fault diagnosis of batch processes using multiblock and multiway PLS” Journal of Process Control, Vol. 5, (1995), pp 277-284

[8] W. Ku, R. H. Storer and C. Georgakis, “Disturbance detection and isolation by dynamic principal component analysis” Chemometrics and Intelligent Laboratory Systems, vol. 30, pp. 179-196, 1995.

[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. vol. 20, pp. 713-S718, 1996.

[10] S. J. Qin and W. Li, “Detection, identification and reconstruction of 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. 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, pp. 381, 1990.

[13] J. Gertler,, W. Li, Y. Huang, and T. McAvoy, “Isolation-enhanced Principal Component Analysis”, AIChE Journal, Vol. 45, pp. 323-334, 1999.

[14] W. Li and S. Shah, “Structured residual vector-based approach to sensor fault detection and isolation”, Journal of Process Control, vol. 12, pp. 429-443, 2002.

[15] R. Xu and C. Kwan, "Robust isolation of sensor failures", Asian Journal of Control, vol. 5,pp. 12-23, 2003.

Fault detection, identification, and reconstruction of faulty chemicalgas 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 ComponentsAnalysis (PCA) or Partial Least Squares (PLS) and multiscaleapproaches, have been reported to be very useful in the taskof fault diagnosis of malfunctioning sensors for several typesof faults. In this work, we compare the performance of PCAand Multiscale-PCA on a fault based on a change of sensorsensitivity. This type of fault affects chemical gas sensors and itis one of the effects of the sensor poisoning. These two methodswill be applied on a dataset composed by the signals of 17 con-ductive polymer gas sensors, measuring three analytes at severalconcentration levels during 10 months. Therefore, additionallyto performance’s comparison, both method’s stability along thetime will be tested. The comparison between both techniqueswill be made regarding three aspects; detection, identificationof the faulty sensors and correction of faulty sensors response.

I. INTRODUCTION

Chemical gas sensors are a good alternative to conven-tional analytic instruments for gas analysis in terms ofpromptness of response and low cost. However, gas sensorssuffer from several drawbacks that prevent them from beingrobust, for example lack of selectivity, high sensibility tohumidity and temperature, drift and poisoning. The use ofchemical gas sensors in an array format with a patternrecognition system, improves the degree of selectivity ofthe individual gas sensors. In the sensor array, each sensorhas a partial specificity (non-specific) and its output producea compound/odor fingerprint that can be identified by apattern recognition system. A sensor array is the base ofthe instrument known as electronic nose, which is inspiredby the mechanisms involved in human olfaction.

In an array of gas sensors one or more sensors can fail dueto a great variety of faults like sensor failure, degradation,external interferences, bias, noise, or specific faults relatedto chemical sensors, like drift or poisoning among others.Although these problems can be overcome by technologyimprovements and careful system design, signal processing

S. Marco and M. Padilla, Departament d’Electronica, Universitat deBarcelona, Martı i Franques 1, 08028 Barcelona, Spain (phone: +34 93402 90 70 email: [email protected], [email protected] ).Institute for Bioengineering of Catalonia, Baldiri i Reixach, 4-6, 08028Barcelona, Spain

A. Perera, Centre de Recerca en Enginyeria Biomedica, UniversitatPolitecnica de Catalunya, Pau Gargallo 5, 08028 Barcelona, Spain

I. Montoliu, Nestle Research Center, Metabonomics Biomarkers, CH-1000 Lausanne 26, Switzerland

A. Chaudry, Crewe Business Park, Crewe, Cheshire, UK.K. Persaud, School of Chemical Engineering and Analytical Science, The

University of Manchester,UK.

can help to detect and compensate the output signal of theaffected sensor or sensors.

Poisoning consists of a degradation of the sensing prop-erties of a sensor when it is exposed to a gas compoundthat irreversibly interacts with its sensing material. Thesensor response gradually degrades over a period of timebut seem to be operating normally. Therefore it is not aneasily detectable problem, because the response of the faultysensor is kept inside its range with an apparently normalbehavior, although with an altered profile of sensitivities tothe different analytes. Poisoning is a surface phenomenonclosely related to physical properties of the sensing materialand the detecting gas, so it can affect only some of thesensors in an array, which have to be immediately detectedand identified in order to correct their responses until theycan be replaced by new ones. Effects of sensor poisoningare gradually and irreversible changes in its baseline resis-tance [26], [22], its sensitivity and recovery times [16]. Thedegree of poisoning depend on the exposure temperature,the time and the concentration of the poisonous substance[19][15], for example hexamethyldisiloxane (HMDS) [15],Sulfur dioxide (SO2) [19], [25], nitrogen dioxide (NO2) [16]or ozone (O3) [16], depending on the type of gas sensor. Toovercome poisoning new poison-resistant materials have tobe developed, however alternatives have been proposed likethe application of special treatments on the sensors [16], [19],[25], [3].

The detection and identification of a fault, identificationof a faulty sensor and correction of the fault (known asfault diagnosis) are issues very discussed in literature. Inliterature, a great variety of types of abnormal events havebeen considered and classified as faults, like external eventsor faults produced by the process or the equipment. However,less attention has been paid to faults related to the nature ofthe sensors.

On the other hand, the employed techniques of faultdiagnosis can be based on statistical methods which havethe advantage of extracting information from the processdata without taking into account the nature of the system.Techniques like Principal Components Analysis (PCA) ex-ploit the property of correlations among variables in order toextract this information. Fault detection using PCA or PartialLeast Squares (PLS) has already been studied in applicationslike process monitoring, quality control, or sensor and pro-cess fault diagnosis[23]. Also modifications of PCA havebeen developed to analyze non steady data; non-linear [28],

multiblock [8], [12], multiway [8], [27], dynamic or recursivePCA[9] and combinations [10]. More recently, the waveletsdecomposition of the sensor signals before the application ofstatistical techniques have reported very good performancein various tasks such as sensor validation [11] or processmonitoring [1]. Dunia and Qin [5], [4] study the sensorand process fault detection, identification and reconstructionbased on PCA, in another posterior work Qin comparedPCA with Multiscale-PCA (MS-PCA) [18]. In that work MS-PCA outperformed the conventional PCA based approach indetecting and identifying the faults.

Event or fault detection on gas sensor arrays has been pre-viously tackled by other authors. Perera et al.[21] presentedan adaptive method based on a recursive dynamic principalcomponent analysis (RDPCA) algorithm. This algorithm wasapplied on real signals arising by oil vapor leakages in an aircompressor. Pardo et al. [20] used the correlation among theresponses of five gas sensors detect a possible malfunctioningof one of the sensors during continuous operation. Theoutputs of the sensors were estimated as a function of theoutputs of the remaining ones by Artificial Neural Networks.

In this work, we compare two techniques PCA and MS-PCA to detect, identify and correct faulty sensors affectedby a simulated sensor poisoning fault along the time. Thesimulation of a sensor fault is based on a change on itssensitivity. The effectiveness of these methods are evaluatedon a dataset composed by measurements of three compoundsat different concentrations, using a 17 sensor array during 10months. The comparison between the algorithms at everystep of time is made regarding three aspects: detection,identification of the faulty sensors and correction of faultysensors response.

II. DATA SET DESCRIPTION AND METHODOLOGY

A. The dataset

The data set consist on 3 different gases at different con-centrations levels periodically measured during 10 months,with an array of 17 conductive polymer sensors. The totalnumber of samples in the dataset is 3415 samples assignedto 8 classes, which have been alternatively measured over thewhole duration of the experiment. The classes correspondedto solutions of ammonia and propanoic acid, each one ofthem at three concentration levels, and a third analyte (n-butanol) at two concentration levels (table I).

For every sample, the full sensor response to the samplingtransient is recorded. The transient has a time duration of200s at a sampling frequency of 1 Hz. At instant t=0s thecompound is introduced into the sensors chamber producinga change in every sensor signal, and at t=184s clean airremoves the analyte from the chamber (fig. 1). In this work,we follow the traditional approach of considering only onepoint value of every sensor response to one sample. Thispoint corresponds to the maximum value of the transientsignal.

In fig. 2 a PCA scores plot of the complete dataset can beseen. It shows the eight classes very mixed and scattered due

TABLE I: Sample distribution among analytes and concen-tration levels.

Analyte SamplesConcentration

levels447 0.01%

Ammonia 452 0.02%307 0.05%

Propanoic 458 0.01%acid 457 0.02%

423 0.05%n-Butanol 446 0.1%

425 1.0%

0 50 100 150 200

0

2

4

6

8

10

t (s)

arb

itra

ry u

nit

s

ammonia 0.05%

propanoic 0.05%

n!buthanol 1%

Fig. 1: Transient response of sensor 1 to three analytes. Thevertical dashed line shows the time instant where the threesignals are maximum.

to the effect of drift and intra-class variability. The irregularvariation, drift effect, along the time is shown in fig. 3 on thesignal of a single sensor (sensor 1) with the samples of onlythree classes for better visualization. Drift has similar effecton every class of data and, since it is so irregular, it maybe mostly attributed to environmental sources like changesin humidity and temperature, which are usual during anexperiment of 10 months and to which conductive polymersensors are very sensitive.

!100 !50 0 50 100 150

!40

!30

!20

!10

0

10

20

30

40

50

82.23 %PC1

8.3

8 %

PC

2

ammonia 0.01%

ammonia 0.02%

ammonia 0.05%

propanoic 0.01%

propanoic 0.02%

propanoic 0.05%

n!buthanol 0.1%

n!buthanol 1%

Fig. 2: PCA scores of the whole dataset.

0 16 31 49 59 80 97 118 142 178 303

7

8

9

10

11

12

13

time (days)

arb

itra

ry u

nit

s

ammonia 0.05%

propanoic acid 0.05%

n!butanol 1%

Fig. 3: Drift effect along the time over the signal of sensor1. Samples belong to three classes.

B. Simulation of sensor poisoning fault

The faulty sensors simulation are based on a change ofsensitivity of one or two sensors. To create a faulty sample,the response of the faulty sensor to a sample belonging to acertain class is changed by its response to any other of theremaining classes randomly selected (fig. 4).

Fig. 4: Simulation of sensor poisoning fault.

C. Methodology

To test the different algorithms along the time, the com-plete dataset (samples ordered in time) has been divided in10 sections, with approx. 342 samples in each section (or set)with approximatively the same amount of samples per class.We refer the first section as calibration set and remainingones as validation sets. From 342 samples of the calibrationset, 311 samples are randomly selected, keeping the sameamount of samples per class, to form the training set at everyrepetition of the experiment.

Comparison between both signal processing techniques ispresented regarding three aspects:

• detection of faulty samples; by using receiver operatingcharacteristic (ROC) curves, which represents the rateof true alarms versus rate of false alarms when varyingthe thresholds of a given detection index. The bestperformance would show the most abrupt corner at theupper left of the figure, meaning high rate of true alarmswhile low rate of false alarms.

• identification of faulty sensors; by success rate oncorrect faulty sensor identification on 15 runs, whereone (or two) random sensors is set faulty at every run,

• correction; by classification into three analytes after cor-rection of faulty sensors responses. The used classifierconsist on a PCA for dimensional reduction from 17to 5 and a kNN (k Nearest Neighbors) classifier withk = 3 nearest neighbors.

These three capabilities are evaluated on validation set num-ber 1 and 9, both with 342 samples and 50% of the sampleswith induced faults in one and two sensors failing at thesame time. The validation set number 9 is specially used totest the robustness of the methods to drift at detection andidentification points.

III. METHODS

A. Principal Component Analysis (PCA)

This approach was proposed by Dunia et al. [5]. Thedetection of a fault is based on SPE value (Squared Pre-diction Error), which represents the squared magnitude ofthe projection of every sample on the residual subspace.Therefore, if x represents one sample in the sensors space, xthe projection of x on the PCA model and x the projectionon the residual space, with x = x + x, then SPE = !x!2.

A high value of SPE (over a certain threshold) indicatesa faulty sample (fig. 5). To identify the faulty sensor, wefollow the identification by reconstruction approach in [5].For this, the reconstructed vector with faulty sensor j, xj , isconsidered:

xj = x " fj!j

where !j is the vector representing the faulty sensor, e.g. avector of length equal to the number of sensors in the sensorarray, with all elements 0 except the element j which is 1.fj is an estimation of the magnitude of the fault on sensor j,given by the minimization of SPEj = !xj!2 which finallygives:

fj = !oTj x

and, finally:SPEj = SPE " f2

j

where fj and !j are the corresponding elements projectedon the residual space and !o

j means !j of unit norm. Whenthe corrected faulty sensor’s response is corrected, SPE is ex-pected to be reduced and its estimated fj is large. The indexdefined by "j = SPEj/SPE is used for the identification ofthe faulty sensor. Since for the faulty sensor fj is larger andSPE smaller than for remaining sensors, the index "j shouldbe close to zero for this sensor. Therefore, by comparing"j for all sensors in a detected faulty sample, the one thatshows the minimum value of "j is the malfunctioning one,or inversely, the index 1/"j shows that the maximum valuecorresponds to the faulty sensor.

Dunia et al. [24] also proposed a technique for an optimumchoice of the number of Principal Components (PC’s) forthe PCA model. This method consist of the calculationof the V RE (Variance of Reconstruction Error) index forall sensors and all possible number of PC’s. A mean, orweighted mean, value of all sensor’s indexes is minimum at

the best choice of the number of PC’s for the PCA model.The VRE in the direction of sensor j, !j , is given by:

uj = var{!Tj (x " xj)} =

!Tj R!j

(!Tj !j)

where R is the covariance matrix estimated from the trainingdata set.

Fig. 5: Detection of faulty samples by SPE.

B. Discrete Wavelet Transform and Multi-Resolution Analy-sis

The theory of Wavelet Transform is a useful tool to analyzetime series [2]. A wavelet ! is a finite time function ofzero average and unit energy, which is dilated with a scaleparameter a and translated by b, !a,b.

!a,b(t) =1#a!(

t " b

a)

The function ! is named mother wavelet and the wavelets!a,b are the same functions centered in time at b, with spreadproportional to a, and centered in frequency at its namedcentral frequency, fc, with spread proportional to 1/a.

The continuous wavelet transform (CWT) work with timeseries defined over the entire real axis and the discretewavelet transform (DWT) deals with series defined over arange of integer values, dyades, where scale is given bya = 2j and b = na with n, j $ Z (Z is the set of integers).

Once defined a scale value a the discrete wavelet transformof a function f(t) at a time position b is given by:

C(a, b) =

! !

"!f(t)!#

a,bdt = %f(t),!#a,b& (1)

Then, the wavelet coefficient C(a, b) gives informationabout frequency content of value around fc of the functionf(t) in the neighborhood of the time location b. WhenC(a, b) is computed along the whole function time axis(all possibles time locations b) at several scales a, it isobtained very useful time-frequency information content. For

instance, isolated singularities, like sharp signal transitionswhich creates large amplitude wavelet coefficients, can bedetected by following across scales the local maxima of thewavelet transform.

However, the resolution of the obtained information intime and frequency domains is limited by the uncertaintyprinciple. Resolution depends on the scale and it is propor-tional to a in the time domain and to 1/a in the frequencydomain. Therefore, in a time-frequency plane the wavelet!a,b is symbolically represented by a rectangle centered at(b, fc) with time and frequency spread proportional theirrespective resolutions. When the scale a decreases, the timesupport is reduced but the frequency spread increases andcovers an interval of higher frequencies.

Wavelets and its discrete dilations and translations generatean orthonormal basis of the space of all square integrablefunctions (L2(R)). Therefore, any finite energy signal fcan be decomposed over this wavelet orthogonal basis{!j,n, j, n $ Z} at certain scales or resolutions like:

f(t) =!"

j="!

!"

n="!C(j, n)!j,n(t) (2)

The framework of multi-resolution analysis, developed byMallat [13], [14] and Meyer [17], is based upon the conceptthat a signal can be decomposed into components in nestedclosed subspaces at j = J resolution levels. A signal inspace Vj is broken down into two orthogonal components.One component lies in a subspace Vj"1, the other one lies ina subspace Wj"1 orthogonal to Vj"1, and the union of Vj"1

and Wj"1 gives the original space Vj . Then the subspaceVj"1 can be further decomposed into Vj"2 and Wj"2 andso on, until a satisfactory approximation is obtained. Finally,the transformation of the signals can be expressed in termsof the approximation coefficients aJ and of detail coefficientsdj (with j=1,..,J).

{aJ ; dJ , ... , d2, d1}where the details coefficients at level j are given by:

dj =!"

n="!C(j, n)!j,n(t) (3)

and the approximation coefficients at fixed J with j ' J :

aJ ="

j>J

dj (4)

being the relation between them:

aJ"1 = aJ + dJ

and the signal is the sum of all the details:

f(t) =

!"

j="!dj = aJ +

J"

j="!dj (5)

Therefore, the approximation coefficients define the pro-jection of the signal onto the coarse space Vj preservinginformation about the lower frequency components, whereas

the details coefficients retain information about the highfrequency components.

C. Multiscale-PCA

In the present work, multi-scale principal component anal-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 andcorrelation across sensors by combining two complementarytechniques; wavelets, which utilize of the individual sensors’response to capture the time-frequency information withinthe sensor, and PCA which capture correlation among sen-sors. MS-PCA is built on two steps; the first one consistson a wavelet decomposition of the individual sensors on aselected family of wavelets into approximations and detailscoefficients at different scales. In the second, every sen-sor contributions from each scale are collected in separatematrices, and a PCA model is then constructed to extractcorrelation at each scale. In fig. 6 a scheme of the methodis shown.

Then, the progression of the measurements can be fol-lowed monitoring the scores and residuals of the waveletscoefficients of new samples on the PCA models at everyscale. If residuals violate certain limits at a given scale thenthe sample data contains a fault corresponding to that scale,so the type of fault can be identified looking at the affectedscales. For example, if a sudden and abrupt fault occurs itwill be better reflected as a spike in the highest scales of theMSPCA models since it is a high frequency change, or ifthe limits are violated at the approximation scale it impliesthe presence of a slowly drifting fault.

Fig. 6: Scheme of the MS-PCA method.

A fault is detected when the SPE calculated from thePCA model of the finest level (highest frequency content)described by detail coefficients d1, exceed a given thresh-old. Because of the nature of poisoning, this type of faultcan be detected at the highest levels of MS-PCA, e.g. d1coefficients. Then, to identify the faulty samples, an inversewavelet transform (IDWT) is applied on the SPE. For this,SPE is considered to be the coefficients d1 of a 1-dimensionalsignal and length equal to the number of samples in the datasubset, whose remaining coefficients in the wavelet transform

are set to 0 (fig. 7). This signal is therefore a representationof SPE as a function of the samples.

Later, the faulty sensor identification step is made bymeans of the PCA method explained above (III-A) applieddirectly on the detail coefficients d1 matrix.

Fig. 7: Obtaining SPE vs. samples from SPE of PCA onwavelets coefficients at the highest level of MS-PCA.

Finally, the correction of the faulty sensor response isperformed by PCA in all matrices of detail coefficients. Sincethe information about the lowest frequencies are containedwithin the approximation coefficients, they are set to 0 forall sensors in order to improve the stability of their responsealong the time. Then, IDWT is applied on the correctedsignals.

IV. RESULTS

To apply the MS-PCA approach to our data set, a waveletbasis and scale were chosen: Daubechies 5 with 5 resolutionlevels. Then, the multi scale development of the 17 sensorssignals belonging to the training set samples, was performed.The following step consisted of building a PCA model of theobtained coefficients at each of the 5 levels. The number ofPC’s of every PCA model varied depending on the validationset, ranging from 3 to 5 PCs.

Then, the same multi-scale development was made on ev-ery validation set of 342 samples. The calculated coefficientswere projected on the corresponding PCA model at everylevel and then SPE was obtained.

A. Faulty samples detection

At each validation set, faulty samples were identifiedby means of the IDWT of the SPE previously obtained.Fig. 8a shows ROC curves for both methods for validationset 1. MS-PCA presents a more sensitive ability for faultysamples detection than PCA, also this method shows muchsmaller degradation in time than PCA when validated withset number 9. This can be seen in fig. 8b.

B. Faulty sensor identification

Highest ratios SPEj/SPE identified the failing sensorin every validation set. In fig. 9 the ability of the methodsto identify one or two faulty sensors by means of a percentrate of accuracy is shown. In the case of one faulty sensor

(a) first validation set

(b) nineth validation set

Fig. 8: ROC curves for two different validation sets.

(fig. 9a) MS-PCA has better performance that PCA andit shows no degradation along the time, unlike PCA whichpresents very good performance in firsts validation sets butdegrades later. When two sensors are failing at the sametime both methods get worse, although MS-PCA is stillbetter than PCA for all validation sets (fig. 9b), presentingless degradation along the time.

C. Faulty sensors correction

After correcting the faulty sensors’ responses on detailcoefficients at every resolution level, a classification step wasperformed in order to compare the correction performanceof both algorithms. Normalized results are presented in thefollowing figures. Fig. 10a shows similar performance forboth methods in first validation set; MS-PCA is slightlyworse than PCA and both degrades in the same way withthe number of faulty sensors, but in any case a correctionimprove results on classification. To better visualize thedegradation in time of both methods, the same calculationswere made on validation set 5. Fig. 10b shows that resultsare much worse for non corrected signals degrading fasterthat previous case with the number of faulty sensors. In thecase of correcting one faulty sensor the performance of bothmethods are again similar but PCA degradation is stronger

(a) One faulty sensor

(b) Two faulty sensors

Fig. 9: Success rate on correct faulty sensor identificationalong the time (9 validation sets) for one or two faultysensors.

in the case of two faulty sensors.

V. CONCLUSION

In this work, the performance of PCA and MS-PCA forfault diagnosis have been compared on a dataset consistingon the signals of 17 conducting polymer sensors along 10months. During these 10 months, the sensors have had avery irregular behavior given mostly by strong changes intemperature and humidity, to which polymer sensors arevery sensitive. Therefore, we have been able to compare theperformance of both methods along the time under severedrift conditions. The most important characteristic of MS-PCA is that, it does not only outperforms PCA in detectionability but it also shows a more robust performance alongthe time. This is a very useful feature for gas sensors, whichare very prone to drift.

ACKNOWLEDGEMENT

This work was partially funded from the European Com-munity Seventh Framework Program (FP7/2007-2013) undergrant agreement no. 216916: Biologically inspired computa-tion for chemical sensing (NEUROCHEM). M.P. and S.M.

(a) First validation set

(b) Fifth validation sets

Fig. 10: Normalized classification rates (CR) with 0, 1 or 2faulty sensors for non corrected and corrected signals in twodifferent validation sets.

are part of a consolidated research group of the Generalitatde Catalunya, SGR2009-753.

REFERENCES

[1] B. Bakshi. Multiscale pca with application to multivariate statisticalprocess monitoring. AIChE J., 44(7):1596–1610, 1998.

[2] B. Bakshi. Multiscale analysis and modeling using wavelets. J.Chemometrics, 13(3-4):415–434, 1999.

[3] E. Comini, G. Faglia, and G. Sberveglieri. Uv light activation oftin oxide thin films for no2 sensing at low temperatures. Sensors &Actuators: B. Chemical, 78(1-3):73–77, Jan 2001.

[4] R. Dunia and S. J. Qin. Joint diagnosis of process and sensor faultsusing principal component analysis. Control Engineering Practice,6(4):457–469, 1998.

[5] R. Dunia and S. J. Qin. A unified geometric approach to process andsensor fault identification and reconstruction the unidimensional faultcase. Computers and chemical engineering, 22(7-8):927–943, 1998.

[6] S. Fourie and P. de Vaal. Advanced process monitoring using an on-line non-linear multiscale principal component analysis methodology.Computers and chemical engineering, 24(2-7):755–760, 2000.

[7] J. Gardner. Intelligent gas sensing using an integrated sensor pair.Sensors & Actuators: B. Chemical, 27(1-3):261–266, Jan 1995.

[8] T. Kourti, P. Nomikos, and J. MacGregor. Analysis, monitoring andfault diagnosis of batch processes using multiblock and multiway pls.JOURNAL OF PROCESS CONTROL, 5(4):277–284, 1995.

[9] W. Li, H. Yue, S. Valle-Cervantes, and S. Qin. Recursive pca foradaptive process monitoring. JOURNAL OF PROCESS CONTROL,10(5):471–486, 2000.

[10] W. Lin, Y. Qian, and X. Li. Nonlinear dynamic principal componentanalysis for on-line process monitoring and diagnosis. Computers andchemical engineering, 24(2-7):423–429, Jan 2000.

[11] R. Luo, M. Misra, and D. Himmelblau. Sensor fault detectionvia multiscale analysis and dynamic pca. Ind. Eng. Chem. Res,38(4):1489–1495, Jan 1999.

[12] J. MacGregor, C. Jaeckle, and C. Kiparissides. Process monitoringand diagnosis by multiblock pls methods. AIChE J., 40(5):826–838,Jan 1994.

[13] S. Mallat. Multiresolution representations and wavelets. Dissertationsavailable from ProQuest, Jan 1988.

[14] S. Mallat and W. Hwang. Singularity detection and processing withwavelets. IEEE transactions on information theory, 38(2):617–643,Jan 1992.

[15] M. Matsumiya, W. Shin, F. Qiu, N. Izu, and I. Matsubara. Poisoning ofplatinum thin film catalyst by hexamethyldisiloxane (hmds) for ther-moelectric hydrogen gas sensor. Sensors & Actuators: B. Chemical,96(3):526–522, Jan 2003.

[16] L. Mazet, C. Varenne, A. Pauly, and J. Brunet. H2, co and highvacuum regeneration of ozone poisoned pseudo-schottky pd–inp basedgas sensor. Sensors & Actuators: B. Chemical, 103(1-2):190–199, Jan2004.

[17] Y. Meyer and J. Bartram. Wavelets and applications. Acoustical Societyof America Journal, 92(5):3023, Jan 1992.

[18] M. Misra, H. Yue, S. Qin, and C. Ling. Multivariate process monitoringand fault diagnosis by multi-scale pca. Computers and chemicalengineering, 26(9):1281–1293, Jan 2002.

[19] R. Moos, F. Rettig, A. Hurland, and C. Plog. Temperature-independentresistive oxygen exhaust gas sensor for lean-burn engines in thick-filmtechnology. Sensors & Actuators: B. Chemical, 93(1-3):43–50, Jan2003.

[20] M. Pardo, G. Faglia, G. Sberveglieri, M. Corte, F. Masulli, andM. Riani. Monitoring reliability of sensors in an array by neuralnetworks. Sensors & Actuators: B. Chemical, 67(1-2):128–133, 2000.

[21] A. Perera, N. Papamichail, and N. Barsan. On-line novelty detectionby recursive dynamic principal component analysis and gas sensorarrays under drift conditions. IEEE Sensors Journal, 6(3):770–783,Jan 2006.

[22] K. Pratt and D. Williams. Self diagnostic gas sensitive resistors insour gas applications. Sensors & Actuators: B. Chemical, Jan 1997.

[23] S. Qin. Statistical process monitoring: basics and beyond. J.Chemometrics, 17(8/9):480–502, 2003.

[24] S. Qin and R. Dunia. Determining the number of principal componentsfor best reconstruction. JOURNAL OF PROCESS CONTROL, 10(2-3):245–250, 2000.

[25] F. Rettig, R. Moos, and C. Plog. Sulfur adsorber for thick-film exhaustgas sensors. Sensors & Actuators: B. Chemical, Jan 2003.

[26] B. Ruhland, T. Becker, and G. Muller. Gas-kinetic interactions ofnitrous oxides with sno2 surfaces. Sensors & Actuators: B. Chemical,50(1):85–94, Jan 1998.

[27] B. Wise, N. Gallagher, S. Butler, D. White, and G. Barna. A compar-ison of principal component analysis, multiway principal componentanalysis, trilinear decomposition and parallel factor analysis for faultdetection in a semiconductor etch process. J. Chemometrics, 13(3-4):379–396, 1999.

[28] J. Zhang, E. Martin, and A. Morris. Process monitoring usingnon-linear statistical techniques. Chemical Engineering Journal,67(3):181–189, Jan 1997.

Chapter 5

Conclusions

AO systems based on an array of gas sensors and multivariate signal processing showa lack of robustness. For instance, AO systems exhibit poor time stability and in anarray 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 ashumidity and temperature, and the operation conditions can make the system to bedependant on variables such as flow rate, heating time and heating cycle for MOXsensors. Consequently, the metrologic characteristics of the instrument are degradedand 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 ofmoulds, which results in the need of a re-calibration of the system.

However, our results show that multivariate signal processing improve the robustnessof 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 withthis 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 topropose 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 potatochips. Our analysis shows that it can be predicted from a single factor extracted byPARAFAC algorithm. The analysis of the trilinear decomposition lead in our caseto direct interpretation of the role of the different decomposition modes, since it ispossible to identify the more relevant sensors and time window in transient signalsregarding predictive power.

MCR-ALS has been used for the first time for the analysis of temperature modulatedMOX sensors signals. We have observed that this technique provides an approximateconcentration 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

additional data from mixtures samples at different concentration levels. However,we have also observed that the concentration profiles are distorted in the overlappingregion. The reason for this is probably the non linearities of the sensors response. Itis important to remark that these profiles have been obtained without any previousrecalibration step.

We studied the ability of orthogonal projection (OP) methods to improve the timestability of AO systems. In particular, our study has been based on a datasetcontaining 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 improvethe time stability of the system. Additionally, OSC outperforms CC during the firsthalf of the experiment. We interpret that this is in part due to OSC having nodependence on the choice of a reference class. The application of OSC resulted inmore compact clusters, which therefore implied an easier classification task. We havealso tested the robustness of the OP algorithms regarding the size of the calibrationset, and obtained that in our case results are robust up to a limit of about 12samples per class. On the other hand, the study of the removed components byboth methods, revealed that the part of the sensors transient signal which mostcontribute to sample dispersion are the rise and fall regions.

The detection of chemical sensor faults is challenging when the fault produces a shiftin the pattern of sensors sensitivity but the sensors keep responding to the targetsample. In this scenario, our findings demonstrate the ability of simple multivariatetechniques; PLS, PCA and structured residuals, for fault detection and identifica-tion. In addition, these methods permit the reconstruction of the faulty sensor’ssignal, keeping the system working with minor performance degradation. The pres-ence of drift may render the detection of faults more difficult. The use of waveletsdecomposition prior to the above mentioned techniques separates the slower changes(drift) and therefore improve the ability of the method to detect the faults.

We performed a deep study on the ability of an e-nose instrument with 10 com-mercial MOX sensors to detect the presence of moulds growing in different buildingmaterials. We observed that the sample dispersion was mainly caused by the differ-ent growing substrates. The system was able to globally detect moulds with 84%accuracy. Also, individual species of moulds not included in the calibration model,as well as moulds growing on materials not considered in the calibration model,could be detected with good accuracies depending on the considered case. A degra-dation of the calibration model was observed when the experiment was performedfour months later, in particular, the decreased classification rates depended on thespecie of mould being considered. Additionally, we could ascertain that Capteur’ssensors were less correlated than Figaro’s.

114

Capıtulo 6

Resumen de la tesis

6.1 Introduccion

Los sentidos humanos que nos permiten obtener información del exterior sin contactodirecto con la fuentes: vista, oído y olfato, han sido estudiados intensamente conel objetivo de crear instrumentos que amplien nuestro conocimiento del entorno.Especialmente, los más estudiados han sido la vista y el oído, mientras que el olfatosiempre se ha considerado un sentido primitivo y de poco interés. Sin embargo, en2004 el Dr. Richard Axel y la Dra. Linda Buck compartieron el Premio Nobel deMedicina por sus descubrimientos en este campo y, desde entonces, los estudios sobreel olor y el sistema olfativo humano han recibido un nuevo interés.

En 1982, Persaud y Dodd [1] diseñaron un instrumento llamado nariz electrónicapara diferenciar olores. Este dispositivo se basa en una matriz de sensores de gasescon especificidad parcial junto con un sistema de reconocimiento de patrones. Lasnarices electrónicas fueron prometedores para muchas aplicaciones cualitativas ycuantitativas, ya que proporcionaban características tales como pequeño tamaño,bajo coste, rápido y fácil de usar. Estas cualidades son especialmente interesantespara aplicaciones de campo, en comparación con otros instrumentos bien establecidospara el gas y el análisis de volátiles, como el cromatógrafo de gases y espectrómetrode masas (GC / MS), que son grandes, pesados, caros, lentos y difíciles de usar,aunque proporcionan una mejor resolución química.

Además, Persaud y Dodd, introdujeron por primera vez la idea de un dispositivo quepodía imitar la capacidad de discriminación del sentido del olfato de los mamíferos.Las narices electrónicas evitaban a la vez los problemas que presentan los paneleshumanos, tradicionalmente encargados de la evaluación sensoriales de olores cuyasrespuestas están influenciadas por factores humanos como la fatiga, enfermedad,estado de ánimo, etc. El término nariz electrónica se hizo popular algunos añosmás 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 ahierba, a tierra, etc. , al igual que el sentido humano del olor. Estos descriptoresserían 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 nuevosmétodos analíticos para la evaluación objetiva del olor. Se esperaba incluso que, enel futuro, el instrumento también fuera un sustituto completo para los métodos dereferencia que utilizan los paneles sensoriales humanos, ya que daría lugar a unadisminución en el costo y la arbitrariedad en los resultados de los paneles.

Sin embargo, pocos años después, la investigación sobre las capacidades de la narizelectrónica hizo que algunos autores empezaron a cuestionar el nombre dado aldispositivo, 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 olfacciónartificial (OA) al de nariz electrónica (e-nariz), aunque a lo largo de la tesis seutilizan ambos términos indistintamente. Una vez aceptado que las expectativasiniciales sobre imitar el sentido del olfato no eran realistas, se ha producido un cambioen el punto de vista del problema. El enfoque actual es diseñar un instrumentoinspirado en la naturaleza, no una repetición de la misma. El término empleadopara esta filosofía es biomimética o bioinspiración.

Otra de las razones por las que en la actualidad, más de 25 años después del primerdispositivo, 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 enexposiciones 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 robustosy no dan resultados suficientemente reproducibles.

La naturaleza de los problemas de los sensores químicos de gases es principalmentetecnológica, aunque también afectan (en diferente grado) a todos los sensores delestado del arte actual. Estas deficiencias se pueden superar mayoritariamente conmás investigación en la mejora del proceso de fabricación o en el desarrollo de nuevastecnologías y materiales sensores. Sin embargo, mientras se mejoran las tecnologíasde sensores de gases, el procesado estadístico de señal puede ayudar a compensarmatemáticamente, o al menos reducir, el efecto de las deficiencias mencionadas enlas respuestas de los sensores antes de llevar a cabo el reconocimiento de patrones.

El objetivo de esta tesis es explorar la robustez de algunos modos de funcionamien-to del sensor, y proponer el uso de técnicas estadísticas de procesado de señalespara corregir o compensar las respuestas de los sensores afectados por problemasespecíficos, tales como las derivas y el fallo de uno o más sensores en la matriz.

116

Introducción

Sin embargo, en OA no se ha adoptado todavía una definición exacta del términorobustez ni una metodología para ponerla a prueba en un sistema o modelo decalibración. En esta tesis repasamos las nociones la misma y de los tests que se usanen otras áreas de instrumentación química, subrayando las partes que nos serían deutilidad en el campo de los sistemas de OA (ver sección 2.1).

6.1.1 Tecnologıas sensoras

Existen muchas tecnologías sensoras utilizadas en el campo de la OA. Entre ellas, enesta tesis utilizamos sólo dos clases: sensores de óxidos metálicos (MOX) y sensoresde polímero conductor (CP).

Los sensores MOX son probablemente los más ampliamente utilizados, debido a subajo precio y características bastante buenas en relación con el tiempo de vida ysensibilidad; además, están disponibles comercialmente en muchos tipos y con distin-tas especificidades. Normalmente, estos sensores están compuestos por un sustratoy un elemento calefactor, cubiertos con el material sensor entre dos electrodos. Surespuesta se basa en un cambio en la conductividad del material sensor cuando seexpone a los compuestos que inducen en él una reacción de oxidación o reducción.Además, la sensibilidad que presentan depende de la temperatura [91, 92]. Entrelas desventajas que presentan se encuentran; el alto consumo de potencia del ca-lenfactor, la posibilidad de envenenamiento con ácidos débiles o compuestos conazufre, alta sensibilidad al etanol, a la humedad, lenta recuperación que depende dela muestra, la alta temperatura de funcionamiento que hace que no sean adecuadospara aplicaciones que involucren gases inflamables, etc.

Existen dos tipos de sensores CP; polímero de conducción intrínseca y polímerobasado en compuestos conductores. Su respuesta se basa en un cambio en su con-ductividad al ser expuestos a un volátil, como consecuencia de la absorción de ésteque hace que el material se hinche. Las ventajas que presenta estos sensores son:la gran variedad de polímeros que pueden encontrarse en el mercado y que puedenser modificados para ajustarse a la aplicación, bajo consumo de energía ya que nonecesitan ser calentados, fabricación muy flexible que permiten miniaturización yproducción en masa, rápida respuesta y recuperación para algunos analitos, resis-tencia al envenenamiento, etc. Entre sus desventajas podemos encontrar diferenciasentre sensores supuestamente idénticos, derivas, gran sensibilidad a la temperaturay humedad, etc. [84, 93, 102].

6.1.2 Modos de operacion y procesado de senal

Aparte de mejoras tecnológicas que puedan disminuir las desventajas que presentanlos diferentes sensores, también se pueden usar en varios modos de operación que

117


permitan extraer más información de las muestras. Éstos vienen dados por el diseñodel instrumento que a su vez depende de la aplicación y las cualidades técnicasdeseadas. Según los diseños, los instrumentos varían en coste y complejidad.

Algunos sistemas OA carecen de fluídica y cámara de sensores, por lo que éstosestán expuestos directamente al ambiente. Estos son los sitemas más simples, unejemplo son los detectores de monóxido de carbono. Otros instrumentos poseen unsistema de control del flujo y por tanto fluídica y cámara de sensores. Algunos deestos aparatos disponen, también, de un mecanismo de conmutación y una fuente deaire limpio que permite la limpieza del sistema entre medidas, así se puede obtenerla señal transitoria de los sensores: limpieza, exposición al volátil, limpieza.

De forma adicional, cuando se tienen sensores MOX en el instrumento, éstos puedenser modulados en temperatura con diferentes duraciones de los ciclos y formas deonda del voltaje aplicado al calefactor. Este modo de operación requiere de unaunidad adicional de control de la tensión de los calefactores.

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 dependede 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 porlas que se considera únicamente el máximo del transitorio o la respuesta a unatemperatura fija. Con esta información y los algoritmos adecuados de procesado deseñal, se pueden compensar en gran parte la inestabilidad de la respuesta de lossensores a lo largo del tiempo y otros posibles problemas.

6.2 Deriva y fallos en los sensores de una matriz

Entre las propiedades ideales de los sensores se encuentran: tiempos de respuestay recuperación rápidos, alta sensibilidad o bajos límites de detección, linealidad,rango dinámico largo, posibilidad de alta selectividad o especificidad (para algunossensores en una matriz), bajas sensibilidades cruzadas a interferentes, alta estabili-dad en el tiempo, obtención de características iguales para sensores supuestamenteidénticos, etc. Como no todas estas propiedades ideales están presentes en un mismosensor, ya que el cumplimiento simultáneo de algunas puede ser física y químicamen-te contradictorio, hay que escoger y valorar los sensores en función de la aplicacióndeseada. Además, las partes restantes del instrumento, tales como cámara de senso-res y fluídica, no deben influir en la respuesta de los senores, por esta razón todaslas partes del instrumento deben ser cuidadosamente diseñadas [59].

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 autoresincluyen factores ambientales tales como cambios en la humedad, temperaturas yvelocidad de flujo, como causas de la deriva, aunque se describen mejor como in-terferencias cruzadas. Las derivas de los sensores alteran sus respuestas mediantecambios de sensibilidad, selectividad, magnitud y línea de base. Estos cambios sue-len ser graduales si la causa de la deriva es mayoritariamente el envejecimiento delos sensores, pero pueden producirse bruscamente si los causantes mayoritarios soncambios en variables medioambientales o un fuerte envenenamiento.

Por otro lado, el envenenamiento es un cambio en la respuesta de un sensor en lamedida de un mismo analito en las mismas condiciones. Este cambio es irreversibledebido a las interacciones entre el material sensor y el analito. Además, el enve-nenamiento puede ocurrir repentinamente o de manera progresiva en función de lasustancia y su concentración. El sensor envenenado responde con sensibilidades ylínea de base alteradas, lo que puede dar lugar a su pérdida.

La deriva y problemas relacionados dan lugar a la necesidad de re-calibracionesperiódicas del dispositivo, que son tareas costosas y requieren mucho tiempo. Estasre-calibraciones han de reducirse al mínimo. Uno de los principales objetivos de estetrabajo es el intento de compensar la respuesta degradada de los sensores, sobretodo debido a la deriva y envenenamiento, por medio de herramientas de procesadode señales.

Cuando aplicamos un método de procesado de señal para corregir las derivas, ésteha de evaluarse. Ésto se puede hacer mediante una comparación de las habilidadespredictivas del aparato antes y después de la corrección. Se pueden comparar lastasas de clasificación, la relación de Fischer o el error cuadrático medio en predicción(RMSEP) en aplicaciones de cuantificación.

En la evaluación del método corrector de derivas, se debe prestar atención al sistemade validación. Al construir el modelo de calibración hay que tomar muestras de unconjunto inicial de muestras ordenadas en el tiempo, ya que si se utilizan muestrasposteriores se puede llegar a resultados que no son reales. Por eso, en esta tesishemos descartado la validación por leave-one-out, k-fold, muestreo aleatorio, etc.

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 propuestodiferentes técnicas en la literatura: univaradas, multivariadas, adaptivas, lineales,no lineales, mediante modelos de derivas, con medidas de gases de referencia a lolargo del tiempo, con corrección de componentes, etc. Sin embargo, el problema noestá totalmente resuelto y, aunque se ha conseguido mejorar el rendimiento de lossistemas, sigue siendo necesario la investigación de otros métodos de procesado deseñal.

Por otro lado, en una matriz de sensores, uno o más pueden fallar en un momentodado debido a multitud de causas, tanto instrumentales como derivadas del propio

119


sensor, y que pueden ser sutiles; es decir, fallos en los que parece que el sensorfunciona correctamente. Por lo tanto, es aconsejable que un sistema basado en unamatriz de gases sea capaz no sólo de detectar fallos en alguno de sus sensores, sinoque, además de identificar al sensor que falla, pueda corregir su respuesta y funcionarhasta que el sensor pueda ser reemplazado. En la literatura referente a los sistemasde OA no existen casi trabajos al respecto, aunque es un área de conocimientoampliamente desarrollada en otros campos.

6.3 Propuesta de tecnicas de procesado de senal para mejorarla robustez del sistema

Se pueden adoptar diferentes puntos de vista en el estudio específico de medidasde corrección de derivas mediante procesado de señal. En primer lugar, se ha com-probado que las muestras que derivan presentan un alto grado de dispersión en unespacio de características. Esta dispersión está orientada, generalmente, en pocasdirecciones. Las técnicas de procesado de señal que puedan reducirla mediante com-pensación o incluso mediante la eliminación de estas direcciones pueden ser muyútiles. La familia de técnicas de proyección ortogonal (OP) fue especialmente dise-ñada para llevar a cabo esta tarea y ya se utiliza en campos como espectroscopía,cromatografía y otras áreas de instrumentación. Estos métodos consisten en la eli-minación de la variabilidad de los datos no correlacionada con una determinadainformación de interés y, por tanto, dan lugar a clases de datos más compactas ymás fáciles de predecir. En concreto, en esta tesis exploramos las habilidades delmétodo de corrección ortogonal de señal (OSC).

Una manera interesante de trabajar con instrumentos que presentan fuertes derivases utilizar métodos que extraen información de las medidas, pero sin necesidad deuna 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 usode métodos no supervisados, como las técnicas multiway, que permiten la creaciónde modelos parsimoniosos y de interpretación más sencilla a partir de la informaciónque contienen las señales transitorias o moduladas de cada sensor. Este problema seconoce en procesado de señal como separación ciega de fuentes (BSS), aunque en elcampo 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 datosen componentes con signficado químico. Entre las técnicas multiway, utilizamos PA-RAFAC para hacer un modelo con señales transitorias. También exploramos MCRpara 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, sinnecesidad 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

a resultados falsos y a grandes errores en la etapa de predicción. Por lo tanto, esmuy importante detectar el fallo de uno o más sensores a fin de ser reemplazadocuando sea posible y, mientras tanto, que el sistema sea capaz de funcionar con lasrespuestas de los sensores que fallan, reconstruídas a partir de los demás sensores dela matriz. Los problemas de detección de fallos y aislamiento (FDI) y de deteccióny diagnóstico de fallos (FDD), que implican la detección del fallo, la identificacióndel sensor defectuoso y la corrección de su respuesta, han sido desarrollado en granmedida para el control estadístico de procesos (SPC). Existen otras técnicas utili-zadas en el campo de sistemas de control, pero los métodos estadísticos son muyinteresantes, ya que no es necesario hacer un modelo del comportamiento del ins-trumento. Por el contrario, el modelo se construye con los datos de las medidas. Enesta tesis investigamos la capacidad de métodos sencillos como el análisis de com-ponentes principales (PCA) para detectar, identificar y corregir la respuesta de unsensor que falla debido a un envenenamiento. También estudiamos el mismo pro-blema cuando el sistema deriva, por lo que añadimos una descomposición de señalmediante wavelets antes de aplicar PCA.

6.4 Publicaciones y ponencias

Algunos de los resultados obtenidos a lo largo de esta tesis se presentan en cuatrorevistas y dos ponencias, donde el lector puede encontrar detalles concretos referentesa los experimentos, metodologías y técnicas sobre el tratamiento de datos empleados.A continuación, se presenta una breve introducción donde se detalla la contribuciónprincipal de cada trabajo a la mejora de la robustez de los sistemas de OA.

Detection of diverse mould species growing on building materials by gassensor arrays and pattern recognitionM. Kuske, M. Padilla, A.C. Romain, J. Nicolas, R. Rubio, S. MarcoSensors and Actuators B: Chemical, vol. 119, no.1, pp. 33-40, 2006.

Como ya se ha mencionado, los sensores químicos son propensos a la deriva, lo quehace que los modelos estadísticos sean inútiles después de un tiempo relativamentecorto. Esto se muestra en este trabajo, donde se usa una nariz electrónica para reco-nocer los materiales de construcción infectados por moho en dos series de medidasseparadas por unos cuatro meses.

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 MOXcomerciales. Entre las cuestiones estudiadas, se incluye la influencia de cada tipode material de construcción y la especie de moho, en el reconocimiento final delsubstrato como infectados o no, y, también, si se puede detectar la presencia deuna especie desconocida de moho, así como la presencia de moho sobre substratos

121


desconocidos.

El segundo objetivo es poner a prueba la estabilidad temporal del instrumento enesta aplicación. Para ello, un conjunto de datos similares se recogió cuatro mesesdespués del primero, a partir del cual se construyó el modelo de clasificación. Losresultados de este segundo conjunto de datos ponen de manifiesto la degradacióndel modelo de clasificación causados por la deriva. Aunque la tasa general de laclasificación no es demasiado baja en comparación con la obtenida en el primerconjunto de datos, algunas clases individuales se degradan fuertemente, en funciónde la especie de moho y del tipo de substrato. Esto también ilustra el hecho de quela deriva influye en sensores individuales de diferentes maneras, como se observó enla sección 6.2, ya que la evolución de los datos a lo largo del tiempo depende de laclase a la que pertenecen.

Es interesante observar la figura de scores PCA de la distribución de substratosde cultivo infectados o sanos. Allí, se muestra una alta dispersión de las muestrasque depende principalmente del tipo de material de construcción, sin embargo ladispersión de las muestras también está presente dentro de cada clase de substratoy esto se debe principalmente a la deriva.

Feature extraction on three way enose signalsM. Padilla, I. Montoliu, A. Pardo, A. Perera, S. MarcoSensors and Actuators B: Chemical, vol. 116, no.1-2, pp. 145-150, 2006.

Como se ha visto en la sección 6.2, utilizar la respuesta transitoria completa de lossensores mejora la capacidad predictiva de una nariz electrónica. Cuando las seña-les transitorias de todos los sensores de la matriz están dispuestos en una matrizde tres modos, los datos se pueden procesar mediante técnicas multiway que pro-porcionan modelos más parsimoniosos y de interpretación más sencilla. Además lascaracterísticas extraídas se pueden utilizar para construir modelos posteriores parala predicción.

En este trabajo, se aplica el método multiway de análisis de factores paralelo (PA-RAFAC) en el control de calidad de un producto alimenticio: las patatas fritas condiferentes concentraciones de un agente aromatizante. Las medidas se realizaroncon una nariz electrónica comercial que contiene 13 sensores de óxidos metálicos.Más tarde, sólo una de las características obtenidas del PARAFAC se utiliza paraconstruir un modelo de mínimos cuadrados inverso (ILS), a partir del cual se puedecalcular la concentración de aroma de nuevas muestras. Los restantes loadings dePARAFAC dan información sobre la influencia de cada sensor individual en el mo-delo e identifica la ventana de tiempo en la que los sensores están expuestos a losvolátiles.

Hay que señalar que, antes de la aplicación de este método, ha de confirmarse lapropiedad de trilinealidad requerida por PARAFAC. Además, se ha de elegir el nú-

122

Publicaciones y ponencias

mero de factores para el modelo PARAFAC. Para ello, se construyen varios modeloscon distinto número de factores a partir del conjunto de calibración. Luego, una vezseleccionado el número de factores, se hace una prueba de robustez repitiendo elprocedimiento varias veces para confirmar esta selección.

Drift compensation of gas sensor array data by Orthogonal Signal CorrectionM. Padilla, A. Perera, I. Montoliu, A. Chaudry, K. Persaud, S. MarcoChemometrics and Intelligent Laboratory Systems, vol. 100, no. 1, pp. 28-35, 2010.

Como se mencionó en la sección 6.3, los métodos de proyección ortogonal (OP) sonuna familia de algoritmos interesantes para hacer frente a la deriva y la dispersiónlocal 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 decalibración posterior y así aumentar su robustez. Sin embargo, estos métodos casi nose utilizan en OA, sólo unas pocas referencias se pueden encontrar en la literatura,por ejemplo [157, 168].

Con el objetivo de explorar la capacidad de las técnicas de OP para disminuir lainfluencia de la deriva, en este trabajo se propone el uso de uno de estos métodos,el algoritmo de corrección ortogonal de señal (OSC), para compensar la deriva enun experimento de una duración de 10 meses. Este experimento consiste en medidasde los tres tipo de analitos a varios niveles de concentración, utilizando una matrizde 17 sensores de gases de polímeros conductores. Además de OSC, el método deArthursson de corrección de componentes (CC), se prueba en el mismo conjunto dedatos con el fin de ser comparado con OSC y con la predicción cuando no se empleaetapa previa preprocesado para contrarrestar la deriva. Los resultados se presentanen forma de tasa de clasificación de los tres tipos de los analitos; sin embargo, hay queseñalar que la información relacionada con los valores de concentración también seutiliza en el diseño de los modelos, tanto de CC como de OSC. Además, es importantetener en cuenta que se han utilizado las señales transitorias completas del conjuntode los sensores, ya que proporcionan información adicional para la discriminaciónde clase, como se mencionó en la sección 6.2.

Multivariate curve resolution applied to temperature-modulated metal oxidegas sensorsI. Montoliu, R. Tauler, M. Padilla, A. Pardo, S. MarcoSensors and Actuators B: Chemical, vol. 145, no. 1, pp. 464-473, 2010.

En este artículo, se aplica por primera vez el método de resolución multivariantede curvas con mínimos cuadrados alternados (MCR-ALS) sobre los datos de unamatriz de sensores de gases.

Como ya se mencionó en la sección 6.2, esta clase de métodos pueden extraer in-

123


formación de los datos sin la necesidad de un modelo de calibración previo, y porlo tanto no se ven afectados por la deriva. MCR-ALS es ampliamente utilizado enel campo de instrumentación en química analítica, donde se han demostrado suscapacidades para el análisis cualitativo de los datos. En este trabajo, MCR-ALS seaplica sobre datos que miden la evolución de dos especies de gases en una mezclacon una matriz de cuatro sensores de óxidos metálicos modulados térmicamente conuna forma de onda triangular. La modulación térmica de los sensores, da lugar auna mayor selectividad y sensibilidad que los sensores isotermos (ver sección 6.1.1),ya que proporciona más información mediante un barrido de la sensibilidad térmicade cada sensor a los analitos en la mezcla. Por lo tanto, en este experimento MCR-ALS determina la resolución de la mezcla de gases por medio de la concentraciónC y también los perfiles de sensibilidad de temperatura ST para cada analito en lamezcla.

Por otra parte, MCR-ALS se aplica también sobre datos sintéticos. En primer lugarse analizan las medidas de un sensor, lo que da lugar a una resolución de la mezclade gases: perfiles de concentración C y sensibilidad de temperatura ST . Posterior-mente, las muestras de datos que contiene las respuestas de todos los sensores enuna matriz de tres vías (muestras × sensores × forma de onda de la modulaciónen temperatura) se desdobla en una matriz de dos vías y se analiza. Los resultadosque se obtuvieron a partir de datos sintéticos fueron satisfactorios, sin embargo, seobservaron efectos no lineales sobre los datos reales.

Además, la concentración de los analitos presentes en la mezcla se calcula medianteun modelo lineal cuantitativo para cada uno de ellos. Aunque en este trabajo, dadoel pequeño tamaño del conjunto de datos, este modelo tiene que ser validado con elerror cuadrático medio de calibración (RMSEC).

Poisoning fault diagnosis in chemical gas sensor arrays using multivariatestatistical signal processing and structured residuals generationM.Padilla, A.Perera, I.Montoliu, A.Chaudry, K.Persaud, S.MarcoIEEE International Symposium on Intelligent Signal Processing, WISP 2007, pp. 1-6, 2007.

Como en una matriz de sensores uno o más pueden fallar, es importante que sedetecte 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, sesimula 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 desensibilidad 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 sensorsigue trabajando en condiciones aparentemente normales.

En este trabajo, se propuso el uso de métodos estadísticos sencillos como PCA, PLSy PCA con residuos estructurados, para realizar las tres tareas mencionadas ante-

124

Conclusiones

riormente: detección, identificación y corrección de uno y dos sensores defectuosos enuna matriz de 17 sensores de gases de polímero conductor. Las tres técnicas se com-paran en cada paso por: la curva característica de receptor (ROC) para la evaluaciónde la curva de capacidad de detección, el porcentaje de acierto en la identificacióndel sensor defectuoso y las tasas de clasificación antes y después de la corrección dela respuesta del sensor defectuoso.

Fault detection, identification, and reconstruction of faulty chemical gassensors under drift conditions, using Principal Component Analysis andMultiscale-PCAM.Padilla, A.Perera, I.Montoliu, A.Chaudry, K.Persaud, S.MarcoIEEE World Congress on Computational Intelligence, WCCI 2010, 2010.

También en este artículo, se prueba la capacidad de un método de procesado deseñales para detectar, identificar y corregir uno o dos sensores envenenados en unamatriz, pero con una dificultad adicional en la que el sistema está sujeto a una derivairregular. Por lo tanto, dado que el método tiene que ver con la deriva, las tareas dedetección, detección y corrección de la falla del sensor serán más difíciles.

El conjunto de datos en el que se crea esta situación, consiste en medidas de tresanalitos a niveles de concentración diferentes a lo largo de diez meses de experi-mento. Una vez más, el fallo de envenenamiento se simula cambiando el perfil desensibilidad del sensor defectuoso, que pertenece a una matriz de 17 sensores depolímeros conductores.

En este artículo, proponemos el uso de PCA multi-escala para la detección, identi-ficación y corrección de un sensor defectuoso en condiciones de deriva. El paso demulti-escala consiste en la descomposición previa de la respuesta de cada sensor alo largo del tiempo (muestras) mediante una transformada wavelet discreta (DWT).DWT permite aislar la parte más lenta de las diferentes señales, las cuales estánrelacionadas con la deriva, de la parte relativa a la información. Luego, se agrupaen matrices los coeficientes correspondientes a cada escala de las señales de todoslos sensores. Finalmente, se aplica el mismo método de PCA que en la sección 6.4sobre la matriz de la escala más alta para llevar a cabo la detección, identificacióny corrección de tareas.

6.5 Conclusiones

Los sistemas AO basados en una matriz de sensores son poco robustos, presentanpoca estabilidad a lo largo del tiempo y uno o más de sus sensores pueden fallardebido a muchas causas, como envenenamiento. Además, variables ambientales co-mo la temperatura y la humedad afectan a los sensores químicos y las condiciones

125


de operación pueden hacer que el sistema dependa de otros parámetros como elflujo, tiempo y ciclo de calentamiento en los sensores MOX. En consecuencia, las ca-racterísticas metrológicas del instrumento se degradan y necesitan ser re-calibradosfrecuentemente, una tarea que es cara y laboriosa. Por ejemplo, en el artículo 4.1 semuestra cómo se degrada la habilidad el instrumento para detectar la presencia dehongos sobre ciertos materiales de construcción.

A pesar de la falta de robustez de los sensores químicos, nuestros resultados muestranque se puede extender la vida media del instrumento entre re-calibraciones utilizandoprocesado de señal multivariado. A lo largo de esta tesis hemos estudiado variastécnicas para este propósito y cuyas conclusiones resumimos a continuación.

La extracción de características es un paso clave para obtener la máxima informa-ción de los sensores. Dada la estructura particular de los datos que contienen lasrespuestas transitorias de los sensores, hemos propuesto el uso de técnicas multiway.En el análisis de la concentración de aroma en las patatas fritas hemos mostradola habilidad de PARAFAC de construir modelos parsimoniosos. Nuestros resultadosmuestran que es posible predecir la concentración de aroma a partir de un sólo factorcalculado mediante PARAFAC. El análisis de la descomposición trilinear da lugar,en nuestro caso, a la interpretación directa del papel de los diferentes modos dedescomposición, ya que se puede identificar los sensores más relevantes y la ventanade tiempo en las señales transitorias en relación a la capacidad predictiva.

Por primera vez en OA, se ha utilizado MCR-ALS para el análisis de señales desensores MOX modulados en temperatura. Hemos observado que esta técnica pro-porciona concentraciones y perfiles de sensibilidad aproximados, incluso con datosde un solo sensor. Además, los resultados se pueden mejorar mediante el uso de lasseñales de la matriz de sensores completa y/o mediante datos adicionales obtenidosa partir de mezclas de muestras a niveles de concentración diferentes. Sin embargo,también hemos observado que los perfiles de concentración están distorsionados enla región de superposición. Ésto probablemente se deba a la no linearidad de la res-puesta de los sensores. Es importante señalar que estos perfiles se han obtenido sinningún otro paso de recalibración anterior.

Hemos estudiado la capacidad de los métodos de proyección ortogonal (OP) paramejorar la estabilidad en el tiempo de sistemas de OA. En particular, nuestro estu-dio se ha basado en un conjunto de datos que contiene medidas de 17 sensores depolímeros conductores durante 10 meses. Los algoritmos específicos utilizados parallevar a cabo el estudio fueron corrección ortogonal de la señal (OSC) y la correcciónde componente (CC). Encontramos que tanto OSC como CC mejoran la estabili-dad temporal del sistema. Además, OSC supera a CC durante la primera mitaddel experimento. Interpretamos que esto se debe en parte a que OSC no dependede la elección de una clase de referencia. La aplicación de OSC da lugar a gruposmás compactos, que por lo tanto facilita la tarea posterior declasificación. Tambiénhemos comprobado la robustez de los algoritmos OP con respecto al tamaño delconjunto de calibración. En nuestro caso los resultados son robustos, con un límite

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 sensorestransitorios que más contribuyen a la dispersión de la muestra son las regiones detransición: subida y bajada de la señal.

La detección de fallos en sensores químicos es un reto cuando la falta produce uncambio en el patrón de sensibilidad de los sensores, pero los sensores siguen respon-diendo con aparente normalidad a la muestra gaseosa. En este escenario, nuestrosresultados demuestran la capacidad de las simples técnicas de análisis multivariante,PLS, PCA y residuos estructurados, para la detección e identificación. Además, estosmétodos permiten la reconstrucción de la señal del sensor defectuoso, manteniendoel sistema funcionando con mínima degradación en su rendimiento. La presencia dederivas pueden hacer que la detección de fallos sea más difícil. El uso de descompo-sición de la señal mediante wavelets, antes de las técnicas mencionadas, separa loscambios más lentos (deriva) y por lo tanto mejorar la capacidad del método paradetectar los fallos.

Realizamos un estudio profundo sobre la capacidad de una nariz electrónica con 10sensores de MOX comerciales para detectar la presencia de moho sobre diferentesmateriales de construcción. Observamos que la dispersión de la muestra se debióprincipalmente a la sustratos de cultivo diferentes. El sistema fue capaz de detectarhongos con una precisión global del 84%. Además, también pudimos detectar conbuena precisión, especies individuales de moho no incluidos en el modelo de calibra-ción, así como moho sobre materiales no considerados en el modelo de calibración.Se observó una degradación de éste cuando el experimento se llevó a cabo cuatromeses después. En concreto, la disminución de índices de clasificación depende dela especie de moho que se trate. Finalmente, pudimos comprobar que los sensoresCapteur estaban menos correlacionados que los de Figaro.

127


128

References

References

[1] K. Persaud and G. Dodd, “Analysis of discrimination mechanisms in the mammalianolfactory system using a model nose,” Nature, vol. 299, pp. 352 – 355, Jan 1982.

[2] J. Gardner and P. Bartlett, “Performance definition and standardization of electronicnoses,” Sensors and Actuators B: Chemical, vol. 33, pp. 60–67, Jan 1996.

[3] P. Mielle, F. Marquis, and C. Latrasse, “Electronic noses: specify or disappear,”Sensors and Actuators B: Chemical, vol. 69, pp. 287–294, Jan 2000.

[4] F. Rock, N. Barsan, and U. Weimar, “Electronic nose: Current status and futuretrends,” Chemical Reviews, vol. 108, pp. 705–725, Jan 2008.

[5] M. Pardo, G. Niederjaufner, G. Benussi, E. Comini, G. Faglia, G. Sberveglieri,M. Holmberg, and I. Lundstrom, “Data preprocessing enhances the classificationof different brands of espresso coffee with an electronic nose,” Sensors and ActuatorsB: Chemical, vol. 69, no. 3, pp. 397–403, 2000.

[6] R. Dutta, E. Hines, J. Gardner, and K. Kashwan, “Tea quality prediction usinga tin oxide-based electronic nose: an artificial intelligence approach,” Sensors andActuators B: Chemical, vol. 94, pp. 228–237, Jan 2003.

[7] M. Escuderos, S. Sánchez, and A. Jiménez, “Virgin olive oil sensory evaluation byan 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 forwine ageing detection,” Sensors and Actuators B: Chemical, vol. 133, pp. 180–186,Jul 2008.

[9] D. Ballabio, A. Mauri, R. Todeschini, and S. Buratti, “Geographical classification ofwine and olive oil by means of classification and influence matrix analysis (caiman),”Analytica chimica acta, vol. 570, pp. 249–258, Jan 2006.

[10] N. Bhattacharyya, S. Seth, B. Tudu, P. Tamuly, A. Jana, D. Ghosh, R. Bandyopad-hyay, and M. Bhuyan, “Monitoring of black tea fermentation process using electronicnose,” Journal of Food Engineering, vol. 80, no. 4, pp. 1146–1156, 2007.

[11] D. Kohl, L. Heinert, J. Bock, T. Hofmann, and P. Schieberle, “Gas sensors for foodaroma during baking and roasting processes based on selective odorant measurementsby an array (hrgc/sommsa),” Thin Solid Films, vol. 391, pp. 303–307, Jan 2001.

[12] E. Gobbi, M. Falasconi, I. Concina, G. Mantero, F. Bianchi, M. Mattarozzi,M. Musci, and G. Sberveglieri, “Electronic nose and alicyclobacillus spp. spoilageof fruit juices: an emerging diagnostic tool,” Food Control, vol. 21, pp. 1374–1382,Jan 2010.

[13] I. Concina, M. Falasconi, E. Gobbi, F. Bianchi, M. Musci, M. Mattarozzi, M. Pardo,A. Mangia, M. Careri, and G. Sberveglieri, “Early detection of microbial contamina-tion in processed tomatoes by electronic nose,” Food Control, vol. 20, pp. 873–880,Jan 2009.

129


[14] R. Paolesse, A. Alimelli, E. Martinelli, C. Natale, A. D’Amico, M. D’Egidio, G. Au-reli, A. Ricelli, and C. Fanelli, “Detection of fungal contamination of cereal grain sam-ples by an electronic nose,” Sensors and Actuators B: Chemical, vol. 119, pp. 425–430,Dec 2006.

[15] A. Gomez, J. Wang, G. Hu, and A. Pereira, “Monitoring storage shelf life of tomatousing electronic nose technique,” Journal of Food Engineering, vol. 85, pp. 625–631,Jan 2008.

[16] T. Rajamäki, H. Alakomi, T. Ritvanen, E. Skyttä, M. Smolander, and R. Ahve-nainen, “Application of an electronic nose for quality assessment of modified atmo-sphere packaged poultry meat,” Food Control, vol. 17, pp. 5–13, Jan 2006.

[17] D. Fernandes and M. Gomes, “Development of an electronic nose to identify andquantify volatile hazardous compounds,” Talanta, vol. 77, pp. 77–83, Jan 2008.

[18] L. Senesac and T. Thundat, “Nanosensors for trace explosive detection,” MaterialsToday, vol. 11, pp. 28–36, Jan 2008.

[19] A. Nimal, U. Mittal, M. Singh, and M. Khaneja, “Development of handheld sawvapor sensors for explosives and cw agents,” Sensors and Actuators B: Chemical,vol. 135, pp. 399–410, Jan 2009.

[20] S. Fuchs, P. Strobel, M. Siadat, and M. Lumbreras, “Evaluation of unpleasant odorwith a portable electronic nose,” Materials Science and Engineering: C, vol. 28,pp. 949–953, Jan 2008.

[21] F. Bitter, B. Müller, and D. Müller, “Estimation of odour intensity of indoor airpollutants from building materials with a multi-gas sensor system,” Building andEnvironment, vol. 45, pp. 197–204, Jan 2009.

[22] S. Lim, L. Feng, J. Kemling, and C. Musto, “An optoelectronic nose for detection oftoxic gases,” Nature Chemistry, vol. 13, pp. 562–567, Jan 2009.

[23] A. Valleron, C. Pijolat, J. Viricelle, P. Breuil, and J.-C. Marchand, “Exhaust gassensor based on tin dioxide for automotive application,” Proceedings of 3th Interna-tional Symposium on Olfaction and Electronic Nose ISOEN 2009, Brescia : Italy,Jan 2009.

[24] A. Perera, N. Papamichail, N. Barsan, U. Weimar, and S. Marco, “On-line noveltydetection by recursive dynamic principal component analysis and gas sensor arraysunder drift conditions,” Sensors Journal, IEEE, vol. 6, no. 3, pp. 770 – 783, 2006.

[25] S. Rappert and R. Müller, “Odor compounds in waste gas emissions from agriculturaloperations 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 airof 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 microbialand chemical contamination of potable water,” Sensors and Actuators B: Chemical,vol. 106, pp. 3–6, Jan 2005.

130

References

[28] R. Fend, A. H. J. Kolk, C. Bessant, P. Buijtels, P. R. Klatser, and A. C. Woodman,“Prospects for clinical application of electronic-nose technology to early detection ofmycobacterium tuberculosis in culture and sputum,” Journal of Clinical Microbiol-ogy, vol. 44, no. 6, pp. 2039–2045, 2006.

[29] S. Dragonieri, R. Schot, B. J. A. Mertens, S. Cessie, S. A. Gauw, A. Spanevello,O. Resta, N. P. Willard, T. J. Vink, K. F. Rabe, E. H. Bel, and P. J. Sterk, “Anelectronic nose in the discrimination of patients with asthma and controls,” J AllergyClin Immunol, vol. 120, no. 4, pp. 856–862, 2007.

[30] S. Dragonieri, J. Annema, R. Schot, M. van der Schee, A. Spanevello, P. Carratú,O. Resta, K. Rabe, and P. Sterk, “An electronic nose in the discrimination of patientswith non-small cell lung cancer and copd,” Lung Cancer, vol. 64, no. 2, pp. 166–170,2009.

[31] B. Kateb, M. Ryan, M. Homer, L. Lara, Y. Yin, K. Higa, and M. Chen, “Sniffing outcancer using the jpl electronic nose: A pilot study of a novel approach to detectionand differentiation of brain cancer,” NeuroImage, vol. 47, no. Supplement 2, pp. T5–T9, 2009.

[32] M. Shykhon, D. Morgan, R. Dutta, E. Hines, and J. Gardner, “Clinical evaluation ofthe electronic nose in the diagnosis of ear, nose and throat infection: a preliminarystudy,” The Journal of Laryngology & Otology, vol. 118, no. 09, pp. 706–709, 2004.

[33] G. Pennazza, E. Marchetti, M. Santonico, G. Mantini, S. Mummolo, G. Marzo,R. Paolesse, A. D’Amico, and C. D. Natale, “Application of a quartz microbalancebased gas sensor array for the study of halitosis,” Journal of Breath Research, vol. 2,no. 1, p. 017009, 2008.

[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 sensorarrays(electronic noses): Past, present, and future,” Sensors update, vol. 10, no. 1,p. 189, 2002.

[36] TC Pearce, SS Schiffman, HT Nagle, JW Gardner , “Applications and case stud-ies,” In Handbook of Machine Olfaction - Electronic Nose Technology, pp. 419–578,(part D), T.C. Pearce, S.S. Schiffman, H.T. Nagle, J.W. Gardner (Eds.),Weinheim,Germany: WILEY-VCH, 2003.

[37] A. D. Wilson and M. Baietto, “Applications and advances in electronic-nose tech-nologies,” Sensors, vol. 9, pp. 1–50, Jun 2009.

[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.

[40] M. Ghasemi-Varnamkhasti, S. Mohtasebi, and M. Siadat, “Odor and taste sensingsystems to food quality characterization: An overview on basic principles and recentachievements,” Journal of Food Engineering, Jan 2010.

131


[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,” ExpertReview 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 ReviewsMicrobiology, vol. 2, no. 2, pp. 161–166, 2004.

[46] C. Burian, J. Brezmes, M. Vinaixa, N. Cañellas, E. Llobet, X. Vilanova, and X. Cor-reig, “Ms-electronic nose performance improvement using the retention time dimen-sion and two-way and three-way data processing methods,” Sensors and ActuatorsB: Chemical, vol. 143, pp. 759–768, Jan 2010.

[47] D. García-González and R. Aparicio, “Coupling mos sensors and gas chromatographyto interpret the sensor responses to complex food aroma: Application to virgin oliveoil,” Food Chemistry, vol. 120, pp. 572–579, May 2010.

[48] E. Staples and S. Viswanathan, “Detection of contrabands in cargo containers usinga high-speed gas chromatograph with surface acoustic wave sensor,” Industrial &Engineering Chemistry Research, vol. 47, pp. 8361–8367, Nov 2008.

[49] P. Wang, G. Xu, L. Qin, Y. Xu, Y. Li, and R. Li, “Cell-based biosensors and itsapplication 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: Afirst step towards a neurochip of bioelectronic nose,” Biosensors and Bioelectronics,vol. 22, no. 2, pp. 318 – 322, 2006.

[51] V. Akimov, E. Alfinito, J. Bausells, I. Benilova, I. Paramo, A. Errachid, G. Ferrari,L. Fumagalli, G. Gomila, J. Grosclaude, Y. Hou, N. Jaffrezic-Renault, C. Martelet,E. Pajot-Augy, C. Pennetta, M.-A. Persuy, M. Pla-Roca, L. Reggiani, S. Rodriguez-Segui, O. Ruiz, R. Salesse, J. Samitier, M. Sampietro, A. Soldatkin, J. Vidic, andG. Villanueva, “Nanobiosensors based on individual olfactory receptors,” Analog In-tegrated Circuits and Signal Processing, vol. 57, pp. 197–203, Dec 2008.

[52] C. Hagleitner, A. Hierlemann, D. Lange, A. Kummer, N. Kerness, O. Brand, andH. Baltes, “Smart single-chip gas sensor microsystem,” Nature, vol. 414, pp. 293–296,Nov 2001.

[53] Y. Li, C. Vancura, D. Barrettino, M. Graf, C. Hagleitner, A. Kummer, M. Zimmer-mann, K.-U. Kirstein, and A. Hierlemann, “Monolithic cmos multi-transducer gassensor microsystem for organic and inorganic analytes,” Sensors and Actuators B:Chemical, vol. 126, no. 2, pp. 431 – 440, 2007.

132

References

[54] E. Zampetti, S. Pantalei, A. Macagnano, E. Proietti, C. D. Natale, and A. D’Amico,“Use of a multiplexed oscillator in a miniaturized electronic nose based on a multi-channel quartz crystal microbalance,” Sensors and Actuators B: Chemical, vol. 131,no. 1, pp. 159 – 166, 2008.

[55] W. Zhao, L. A. Pinnaduwage, J. W. Leis, A. C. Gehl, S. L. Allman, A. Shepp, andK. K. Mahmud, “Identification and quantification of components in ternary vapormixtures using a microelectromechanical-system-based electronic nose,” Journal ofApplied Physics, vol. 103, no. 10, 2008.

[56] T. K. Alkasab, J. White, and J. S. Kauer, “A computational system for simulatingand analyzing arrays of biological and artificial chemical sensors,” Chemical Senses,vol. 27, no. 3, pp. 261–275, 2002.

[57] M. Sánchez-Montañés and T. C. Pearce, “Why do olfactory neurons have unspecificreceptive fields?,” Biosystems, vol. 67, no. 1-3, pp. 229 – 238, 2002.

[58] S. D. Wilke and C. W. Eurich, “Representational accuracy of stochastic neural pop-ulations,” Neural Computation, vol. 14, no. 1, pp. 155–189, 2002.

[59] Hierlemann and R. Gutierrez-Osuna, “Higher-order chemical sensing,” Chemical Re-views, vol. 108, pp. 563–613, Jan 2008.

[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.

[61] J. W. Gardner, S. Nadarajan, and P. Kimber, “Modelling and measurement of odourtransportation within the human nasal cavity,” pp. 145–150, 2008.

[62] G. Settles, “Sniffers: Fluid-dynamic sampling for olfactory trace detection in natureand homeland security,” Journal of Fluids Engineering, vol. 127, p. 189, 2005.

[63] M. Falcitelli, A. Benassi, F. D. Francesco, C. Domenici, L. Marano, and G. Pioggia,“Fluid dynamic simulation of a measurement chamber for electronic noses,” Sensorsand Actuators B: Chemical, vol. 85, no. 1-2, pp. 166–174, 2002.

[64] A. Lezzi, G. Beretta, E. Comini, G. Faglia, G. Galli, and G. Sberveglieri, “Influence ofgaseous species transport on the response of solid state gas sensors within enclosures,”Sensors and Actuators B: Chemical, vol. 78, no. 1-3, pp. 144–150, 2001.

[65] F. D. Francesco, M. Falcitelli, L. Marano, and G. Pioggia, “A radially symmetricmeasurement chamber for electronic noses,” Sensors and Actuators: B. Chemical,vol. 105, no. 2, pp. 295–303, 2005.

[66] S. Scott, D. James, Z. Ali, and W. O’Hare, “Optimising of the sensing chamber of anarray 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 ofsampling volatile organic compounds (vocs) in wide concentration range,” Sensorsand Actuators B: Chemical, vol. 122, no. 1, pp. 211–218, 2007.

133


[68] R. Gmür, J. Goschnick, T. Hocker, H. Schwarzenbach, and M. Sommer, “Impact ofsensor packaging on analytical performance and power consumption of metal oxidebased gas sensor microarrays,” Sensors and Actuators B: Chemical, vol. 127, pp. 107–111, Oct 2007.

[69] R. Rubio, J. Santander, L. Fonseca, N. Sabaté, I. Gràcia, C. Cané, S. Udina, andS. Marco, “Non-selective ndir array for gas detection,” Sensors and Actuators B:Chemical, vol. 127, pp. 69–73, Oct 2007.

[70] J. R. S. H. Nanto, “Introduction to chemosensors,” In Handbook of Machine Olfaction- Electronic Nose Technology, pp. 79–104, (ch.4), T.C. Pearce, S.S. Schiffman, H.T.Nagle, J.W. Gardner (Eds.),Weinheim, Germany: WILEY-VCH, 2003.

[71] D. R. W. Todd A. Dickinson, “Optical electronic noses,” In Handbook of MachineOlfaction - Electronic Nose Technology, pp. 181–200, (ch.8), T.C. Pearce, S.S. Schiff-man, H.T. Nagle, J.W. Gardner (Eds.),Weinheim, Germany: WILEY-VCH, 2003.

[72] A. Hierlemann and H. Baltes, “Cmos-based chemical microsensors,” The Analyst,vol. 128, no. 1, pp. 15–28, 2003.

[73] D. Wilson, S. Hoyt, J. Janata, K. Booksh, and L. Obando, “Chemical sensors forportable, 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 presentand 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.

[76] J. Janata, M. Josowicz, and D. DeVaney, “Chemical sensors,” Analytical Chemistry,vol. 66, pp. 207R–228R, Jan 1994.

[77] D. James, S. Scott, Z. Ali, and W. O’Hare, “Chemical sensors for electronic nosesystems,” Microchimica Acta, vol. 149, no. 1, pp. 1–17, 2005.

[78] C. Elosua, I. Matias, C. Bariain, and F. Arregui, “Volatile organic compound opticalfiber sensors: a review,” Sensors, vol. 6, pp. 1440–1465, Jan 2006.

[79] B.R.Eggins, “Chemical sensors and biosensors,” In Practical Guide to Chemometrics,Wiley, 2002.

[80] T. Dickinson, J. White, J. Kauer, and D. Walt, “Current trends inartificial-nose’technology,” Trends in Biotechnology, vol. 16, pp. 250–258, Jan 1998.

[81] S. Capone, A. Forleo, L. Francioso, R. Rella, P. Siciliano, J. Spadavecchia, D. S.Presicce, and A. M. Taurino, “Solid state gas sensors: state of the art and futureactivities,” Journal of Optoelectronics and Advanced Materials, vol. 5, pp. 1335 –1348, Jan 2003.

[82] A. M. Azad, S. A. Akbar, S. G. Mhaisalkar, L. D. Birkefeld, and K. S. Goto,“Solid-state gas sensors: A review,” Journal of the Electrochemical Society, vol. 139,pp. 3690–3704, Jan 1992.

134

References

[83] K. Arshak, E. Moore, G. Lyons, J. Harris, and S. Clifford, “A review of gas sensorsemployed in electronic nose applications,” Sensor review, vol. 24, no. 2, pp. 181–198,2004.

[84] K. Albert, N. Lewis, C. Schauer, G. Sotzing, S. E. Stitzel, T. P. Vaid, and D. R. Walt,“Cross-reactive chemical sensor arrays,” Chemical Reviews, vol. 100, pp. 2595–2626,Jan 2000.

[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 electronicnoses,” 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.

[88] P. Clifford and D. Tuma, “Characteristics of semiconductor gas sensors. i- steadystate gas response,” Sensors and Actuators, vol. 3, no. 1, pp. 233–281, 1983.

[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.

[91] N. M. N. Yamazoe In Chemical Sensor Technology, pp. 19–42, (vol.4), S. Yamauchi(Ed.), Kodansha, 1992.

[92] S. Nakata, S. Akakabe, M. Nakasuji, and K. Yoshikawa, “Gas sensing based on anonlinear response: Discrimination between hydrocarbons and quantification of in-dividual components in a gas mixture,” Analytical Chemistry, vol. 68, pp. 2067–2072,Jan 1996.

[93] D. Strike, M. Meijerink, and M. Koudelka-Hep, “Electronic noses–a mini-review,”Fresenius’ Journal of Analytical Chemistry, vol. 364, no. 6, pp. 499–505, 1999.

[94] P. Mielle, “’electronic noses’: Towards the objective instrumental characterization offood aroma,” Trends in Food Science & Technology, vol. 7, pp. 432–438, Jan 1996.

[95] E. Schaller, J. Bosset, and F. Escher, “ ‘electronic noses’ and their application tofood,” LWT-Food Science and Technology, vol. 31, no. 4, pp. 305–316, 1998.

[96] D. García-González and R. Aparicio, “Sensors: From biosensors to the electronicnose,” Grasas y aceites- Sevilla, vol. 53, no. 1, pp. 96–114, 2002.

[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 ourchoice?,” Materials Science and Engineering: B, vol. 139, pp. 1–23, Jan 2007.

135


[99] G. Mandayo, “Gas detection by semiconductor ceramics: Tin oxide as improvedsensing material,” Sensor Letters, vol. 5, no. 2, pp. 341–360, 2007.

[100] N. Barsan, D. Koziej, and U. Weimar, “Metal oxide-based gas sensor research: Howto?,” Sensors and Actuators B: Chemical, vol. 121, no. 1, pp. 18–35, 2007.

[101] M. D. Wit, E. Vanneste, H. Geise, and L. Nagels, “Chemiresistive sensors of elec-trically conducting poly(2,5-thienylene vinylene) and copolymers: their responses tonine organic vapours,” Sensors and Actuators B: Chemical, vol. 50, pp. 164–172, Jan1998.

[102] J. Janata and M. Josowicz, “Conducting polymers in electronic chemical sensors,”Nature Materials, vol. 2, pp. 19–24, Jan 2003.

[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 onmodern pattern recognition methods,” IEEE Sensors Journal, vol. 2, pp. 203–217,Jul 2002.

[105] “Jlm innovation, germany,” http://www.jlm-innovation.de/.

[106] “Airsense analytics gmbh, schwerin, germany,” www.airsense.com.

[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.

[108] J. Samitier, J. López-Villegas, S. Marco, L. Cámara, A. Pardo, O. Ruiz, andJ. Morante, “A new method to analyse signal transients in chemical sensors,” Sensorsand Actuators B: Chemical, vol. 18, no. 1-3, pp. 308–312, 1994.

[109] E. Llobet, J. Brezmes, X. Vilanova, J. Sueiras, and X. Correig, “Qualitative andquantitative analysis of volatile organic compounds using transient and steady-stateresponses of a thick-film tin oxide gas sensor array,” Sensors and Actuators B: Chem-ical, vol. 41, pp. 13–21, Jun 1997.

[110] C. D. Natale, S. Marco, F. Davide, and A. D’Amico, “Sensor-array calibration timereduction by dynamic modelling,” Sensors and Actuators B: Chemical, vol. 25, no. 1-3, pp. 578–583, 1995.

[111] R.Gutierrez-Osuna, H.Troy Nagle, B.Kermani, S.S.Schiffman, “Signal conditioningand preprocessing,” In Handbook of Machine Olfaction - Electronic Nose Technol-ogy, pp. 105–132, (ch.5), T.C. Pearce, S.S. Schiffman, H.T. Nagle, J.W. Gardner(Eds.),Weinheim, Germany: WILEY-VCH, 2003.

[112] R. Cavicchi, J. Suehle, K. Kreider, M. Gaitan, and P. Chaparala, “Optimizedtemperature-pulse sequences for the enhancement of chemically specific response pat-terns from micro-hotplate gas sensors,” Sensors and Actuators B: Chemical, vol. 33,no. 1-3, pp. 142–146, 1996.

136

References

[113] D. Wilson and S. DeWeerth, “Signal processing for improving gas sensor responsetime,” Sensors and Actuators B: Chemical, vol. 41, pp. 63–70, Jun 1997.

[114] D. Wilson and S. DeWeerth, “Odor discrimination using steady-state and transientcharacteristics 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 BroMulti-Way Analysis with Applications in the ChemicalSciences, John Wiley & Sons, 2004.

[117] P. Clifford and D. Tuma, “Characteristics of semiconductor gas sensors ii. transientresponse to temperature change,” Sensors and Actuators, vol. 3, pp. 255–281, Jan1983.

[118] W. Sears, K. Colbow, and F. Consadori, “Algorithms to improve the selectivity ofthermally-cycled tin oxide gas sensors,” Sensors and Actuators, vol. 19, pp. 333–349,Sep 1989.

[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 signalprocessor,” 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 nonlinearresponses of a semiconductor gas sensor: dependence on the range and frequencyof a cyclic temperature change,” Analytica chimica acta, vol. 361, pp. 93–100, Mar1998.

[121] R. Gutierrez-Osuna, S. Korah, and A. Perera, “Multi-Frequency Temperature Mod-ulation for Metal-Oxide Gas Sensors,” Artificial chemical sensing: olfaction and theelectronic nose (ISOEN 2001): proceedings of the eighth international symposium,p. 212, The Electrochemical Society (Eds.), 2001.

[122] A. Schutze, A. Gramm, and T. Ruhl, “Identification of organic solvents by a virtualmultisensor system with hierarchical classification,” Sensors Journal, IEEE, vol. 4,no. 6, pp. 857 – 863, 2004.

[123] A. Perera, A. Pardo, D. Barrettino, A. Hierlermann, and S. Marco, “Evaluation offish spoilage by means of a single metal oxide sensor under temperature modulation,”Sensors and Actuators B: Chemical, vol. 146, pp. 477–482, Apr 2010.

[124] X. Huang, F. Meng, Z. Pi, W. Xu, and J. Liu, “Gas sensing behavior of a single tindioxide sensor under dynamic temperature modulation,” Sensors and Actuators B:Chemical, vol. 99, pp. 444–450, May 2004.

[125] T. Kunt, T. McAvoy, R. Cavicchi, and S. Semancik, “Optimization of temperatureprogrammed sensing for gas identification using micro-hotplate sensors,” Sensors andActuators B: Chemical, vol. 53, pp. 24–43, Nov 1998.

[126] A. Vergara, E. Llobet, J. Brezmes, P. Ivanov, X. Vilanova, I. Gracia, C. Cané, andX. Correig, “Optimised temperature modulation of metal oxide micro-hotplate gassensors through multilevel pseudo random sequences,” Sensors and Actuators B:Chemical, vol. 111-112, pp. 271–280, Nov 2005.

137


[127] R. Gosangi and R. Gutierrez-Osuna, “Active chemical sensing with partially ob-servable markov decision processes,” Proceedings 13th International Symposium onOlfaction and Electronic Nose.

[128] T. Artursson, T. Eklöv, I. Lundström, P. Mårtensson, M. Sjöström, and M. Holm-berg, “Drift correction for gas sensors using multivariate methods,” Journal ofChemometrics, vol. 14, no. 5-6, pp. 711–723, 2000.

[129] M. Matsuguchi, M. Yoshida, T. Kuroiwa, and T. Ogura, “Depression of a capacitive-type humidity sensor’s drift by introducing a cross-linked structure in the sensingpolymer,” Sensors and Actuators B: Chemical, vol. 102, no. 1, pp. 97–101, 2004.

[130] C. Song, A. Aiyar, S. Kim, and M. Allen, “Exploitation of aeroelastic effects for driftreduction, 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 withdrift-compensated high-resolution interferometric wavelength-shift detection,” Opticsletters, vol. 18, no. 1, pp. 72–74, 1993.

[132] H. Blanchard, C. de Raad Iseli, and R. Popovic, “Compensation of the temperature-dependent offset drift of a hall sensor,” Sensors and Actuators A: Physical, vol. 60,no. 1-3, pp. 10–13, 1997.

[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.

[136] T. A. M. Holmberg, “Drift compensation, standards, and calibration methods,” InHandbook of Machine Olfaction - Electronic Nose Technology, pp. 325–346, (ch.13),T.C. Pearce, S.S. Schiffman, H.T. Nagle, J.W. Gardner (Eds.),Weinheim, Germany:WILEY-VCH, 2003.

[137] R. Gutierrez-Osuna, “Course on Pattern Recognition,”http://research.cs.tamu.edu/prism/lectures.htm.

[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 faultidentification and reconstruction: the unidimensional fault case,” Computers andChemical Engineering, vol. 22, pp. 927 – 943, Jan 1998.

[140] J. Haugen, O. Tomic, and K. Kvaal, “A calibration method for handling the temporaldrift of solid state gas-sensors,” Analytica chimica acta, vol. 407, no. 1-2, pp. 23–39,2000.

138

References

[141] M. Pardo, G. Sberveglieri, A. Taroni, F. Masulli, and G. Valentini, “Decompositiveclassification models for electronic noses,” Analytica chimica acta, vol. 446, no. 1-2,pp. 221–230, 2001.

[142] E. Schaller, J. Bosset, and F. Escher, “Instability of conducting polymer sensors inan electronic nose system,” Analusis, vol. 28, pp. 217–227, Jan 2000.

[143] E. Schaller, J. Bosset, and F. Escher, “Practical experience with’electronicnose’systems for monitoring the quality of dairy products,” CHIMIA InternationalJournal for Chemistry, vol. 53, pp. 98–102, Jan 1999.

[144] S. Labreche, S. Bazzo, S. Cade, and E. Chanie, “Shelf life determination by electronicnose: application to milk,” Sensors and Actuators B: Chemical, vol. 106, pp. 199–206,Apr 2005.

[145] S. Saevels, J. Lammertyn, A. Berna, E. Veraverbeke, C. Natale, and B. Nicolaï, “Anelectronic nose and a mass spectrometry-based electronic nose for assessing applequality during shelf life,” Postharvest Biology and Technology, vol. 31, pp. 9–19, Jan2004.

[146] A. Berna, J. Lammertyn, S. Saevels, C. Natale, and B. M. Nicolaï, “Electronic nosesystems to study shelf life and cultivar effect on tomato aroma profile,” Sensors andActuators B: Chemical, vol. 97, pp. 324–333, Jan 2004.

[147] M. Kuske, R. Rubio, A. Romain, J. Nicolas, and S. Marco, “Fuzzy k-nn applied tomoulds detection,” Sensors and Actuators B: Chemical, vol. 106, pp. 52–60, Apr2005.

[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.

[149] A. Nake, B. Dubreuil, C. Raynaud, and T. Talou, “Outdoor in situ monitoring ofvolatile emissions from wastewater treatment plants with two portable technologiesof electronic noses,” Sensors and Actuators B: Chemical, vol. 106, pp. 36–39, Jan2005.

[150] A. Romain and J. Nicolas, “Long term stability of metal oxide-based gas sensorsfor 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 tinoxide 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 resolutionlimits 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 urbanpollution monitoring with on-field calibrated electronic nose by automatic bayesianregularization,” Sensors and Actuators B: Chemical, vol. 143, pp. 182–191, Jan 2009.

139


[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 monitoringscenario,” 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.

[156] M. Pardo, G. Faglia, G. Sberveglieri, and M. Corte, “Monitoring reliability of sensorsin an array by neural networks,” Sensors and Actuators B: Chemical, vol. 67, pp. 128–133, Jan 2000.

[157] O. Tomic, H. Ulmer, and J. Haugen, “Standardization methods for handling instru-ment related signal shift in gas-sensor array measurement data,” Analytica chimicaacta, 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.

[160] M. Fryder, M. Holmberg, F. Winquist, and I. Lundstrom, “Calibration technique foran electronic nose,” The 1995 8th International Conference on Solid-State Sensorsand Actuators and Eurosensors IX. Part 1(of 2); Stockholm, pp. 683—686, Jan 1995.

[161] F. Davide, C. D. Natale, and A. D’Amico, “Self-organizing multisensor systems forodour classification: internal categorization, adaptation and drift rejection,” Sensorsand Actuators B: Chemical, vol. 18, pp. 244–258, Jan 1994.

[162] S. Marco, A. Ortega, A. Pardo, and J. Samitier, “Gas identification with tin oxidesensor 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.

[164] M. Holmberg, F. Winquist, I. Lundström, F. Davide, C. DiNatale, and A. D’Amico,“Drift counteraction for an electronic nose,” Sensors and Actuators B: Chemical,vol. 36, pp. 528–535, Jan 1996.

[165] M. Holmberg, F. Davide, C. D. Natale, A. DAmico, F. Winquist, and I. Lundström,“Drift counteraction in odour recognition applications: lifelong calibration method,”Sensors and Actuators B: Chemical, vol. 42, pp. 185–194, Jan 1997.

[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 usingcanonical variate analysis and dynamic principal component analysis,” Chemometricsand intelligent laboratory systems, vol. 51, pp. 81–93, Jan 2000.

[168] C. D. Natale, E. Martinelli, and A. D’Amico, “Counteraction of environmental dis-turbances of electronic nose data by independent component analysis,” Sensors andActuators B: Chemical, vol. 82, no. 2-3, pp. 158–165, 2002.

[169] M. Kermit and O. Tomic, “Independent component analysis applied on gas sensorarray 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 causedby environmental fluctuations in the responses of chemoresistive gas sensors,” Sensorsand 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 andprincipal component analysis,” Sensors and Actuators B: Chemical, vol. 96, pp. 354–363, Jan 2003.

[172] M. Zuppa, C. Distante, K. Persaud, and P. Siciliano, “Recovery of drifting sensorresponses by means of dwt analysis,” Sensors and Actuators B: Chemical, vol. 120,no. 2, pp. 411–416, 2007.

[173] R. Gutierrez-Osuna, “Drift reduction for metal-oxide sensor arrays using canonicalcorrelation regression and partial least squares,” Proceedings of the 7th InternationalSymp. On Olfaction and Electronic Nose, pp. Brighton, UK, July 20–24„ 2000.

[174] A. Ziyatdinov, S. Marco, A. Chaudry, K. Persaud, P. Caminal, and A. Perera, “Driftcompensation of gas sensor array data by common principal component analysis,”Sensors and Actuators B: Chemical, vol. 146, pp. 460–465, Apr 2010.

[175] Brahim-Belhouari, Bermak, Minghua, and Chan, “Fast and robust gas identificationsystem 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 methodsfor correction of detector response drift in arrays of carbon black-polymer compositevapor detectors,” Sensors and Actuators B: Chemical, vol. 104, pp. 249–268, Jan2005.

[177] C. Distante, P. Sicilian, and K. Persaud, “Dynamic cluster recognition with multipleself-organising maps,” Pattern Analysis and Applications, vol. 5, pp. 306–315, Jan2002.

[178] “International organization for standardization (iso),” http://www.iso.org.

[179] “International conference on harmonisation of technical requirements for registrationof pharmaceuticals for human use (ich),” http://www.ich.org.

[180] “The united states pharmacopeial convention (usp),” http://www.usp.org.

[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


[182] M. Zeaiter, J. Roger, V. Bellon-Maurel, and D. Rutledge, “Robustness of modelsdeveloped by multivariate calibration. part i: The assessment of robustness,” TrACTrends in Analytical Chemistry, vol. 23, no. 2, pp. 157–170, 2004.

[183] I. Sayhan, A. Helwig, T. Becker, G. Mueller, I. Elmi, S. Zampolli, M. Padilla, andS. Marco, “Discontinuously operated metal oxide gas sensors for flexible tag microlabapplications,” Sensors Journal, IEEE, vol. 8, no. 2, pp. 176 – 181, 2008.

[184] G. Box, “Non-normality and tests on variances,” Biometrika, vol. 40, pp. 318–335,Dec 1953.

[185] M. Daszykowski, K. Kaczmarek, Y. V. Heyden, and B. Walczak, “Robust statisticsin data analysis-a review: Basic concepts,” Chemometrics and intelligent laboratorysystems, vol. 85, no. 2, pp. 203–219, 2007.

[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.

[187] P. Rousseeuw, M. Debruyne, S. Engelen, and M. Hubert, “Robustness and outlierdetection in chemometrics,” Critical Reviews in Analytical Chemistry, vol. 36, no. 3,pp. 221–242, 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.

[189] M. Zeaiter, J. Roger, and V. Bellon-Maurel, “Robustness of models developed bymultivariate calibration. part ii: The influence of pre-processing methods,” TrACTrends in Analytical Chemistry, vol. 24, no. 5, pp. 437–445, 2005.

[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 laboratorysystems, vol. 50, pp. 47–52, Jan 2000.

[192] J. Sjöblom, O. Svensson, M. Josefson, H. Kullberg, and S. Wold, “An evaluation oforthogonal signal correction applied to calibration transfer of near infrared spectra,”Chemometrics and intelligent laboratory systems, vol. 44, pp. 229–244, Jan 1998.

[193] J. Westerhuis, S. de Jong, and A. Smilde, “Direct orthogonal signal correction,”Chemometrics and intelligent laboratory systems, vol. 56, pp. 13–25, Jan 2001.

[194] C. Andersson, “Direct orthogonalization,” Chemometrics and intelligent laboratorysystems, vol. 47, no. 1, pp. 51–63, 1999.

[195] J. Trygg and S. Wold, “Orthogonal projections to latent structures (o-pls),” Journalof 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

[197] M. Bylesjö, M. Rantalainen, O. Cloarec, J. K. Nicholson, E. Holmes, and J. Trygg,“Opls discriminant analysis: combining the strengths of pls-da and simca classifica-tion,” Journal of Chemometrics, vol. 20, no. 8-10, pp. 341–351, 2006.

[198] M. Rantalainen, M. Bylesjö, O. Cloarec, J. K. Nicholson, E. Holmes, and J. Trygg,“Kernel-based orthogonal projections to latent structures (k-opls),” Journal ofChemometrics, vol. 21, pp. 376 – 385, Jan 2007.

[199] R. Feudale, N. Woody, H. Tan, A. Myles, S. Brown, and J. Ferré, “Transfer ofmultivariate calibration models: a review,” Chemometrics and intelligent laboratorysystems, vol. 64, no. 2, pp. 181–192, 2002.

[200] R. N. Feudale, H. Tan, and S. D. Brown, “Improved piecewise orthogonal signalcorrection algorithm,” Applied Spectroscopy, vol. 57, no. 10, pp. 1201–1206, 2003.

[201] J. Roger, F. Chauchard, and V. Bellon-Maurel, “Epo-pls external parameter or-thogonalisation of pls application to temperature-independent measurement of sugarcontent of intact fruits,” Chemometrics and intelligent laboratory systems, vol. 66,no. 2, pp. 191–204, 2003.

[202] Y. Zhu, T. Fearn, D. Samuel, A. Dhar, O. Hameed, S. Bown, and L. Lovat, “Errorremoval 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 orthogonaldiscriminant projection,” Neurocomputing, vol. 73, pp. 191–196, Oct 2009.

[204] M. Zeaiter, J. Roger, and V. Bellon-Maurel, “Dynamic orthogonal projection. a newmethod to maintain the on-line robustness of multivariate calibrations. application tonir-based monitoring of wine fermentations,” Chemometrics and intelligent laboratorysystems, vol. 80, no. 2, pp. 227–235, 2006.

[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 netanalyte preprocessing methods. theoretical and experimental study,” Chemometricsand intelligent laboratory systems, vol. 56, pp. 73–81, Jan 2001.

[207] A. Lorber, K. Faber, and B. Kowalski, “Net analyte signal calculation in multivariatecalibration,” Analytical Chemistry, vol. 68, pp. 1620–1626, Jan 1997.

[208] K. Park and C. Jun, “Hybrid linear analysis based on the net analyte signal inspectral 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.

[210] O. Svensson, T. Kourti, and J. F. MacGregor, “An investigation of orthogonal signalcorrection algorithms and their characteristics,” Journal of Chemometrics, vol. 16,pp. 176 – 188, Jan 2002.

143


[211] J. Gabrielsson and J. Trygg, “Recent developments in multivariate calibration,” Crit-ical Reviews in Analytical Chemistry, vol. 36, no. 3, pp. 243–255, 2006.

[212] B. M. Wise and N. Gallagher, “Eigenvector research, inc.,”http://www.eigenvector.com/MATLAB/OSC.html.

[213] D. R. M. Zeaiter, “Preprocessing methods,” In Comprehensive Chemometrics, vol. 3,pp. 123–228, (ch.15), S.D.Brown, R.Tauler, B.Walczak (Eds.), Elsevier, 2009.

[214] J. Trygg, E. Holmes, and T. Lundstedt, “Chemometrics in metabonomics,” Journalof proteome research, vol. 6, no. 2, pp. 469–479, 2007.

[215] R. L.Balzano, “Blind calibration of networks of sensors: Theory and algorithms,” InNetworked Sensing Information and Control, no. 1, pp. 9–37, (ch.1), V. Saligrama(Ed.), Springer US, 2008.

[216] Z. Liu, Z. Huang, F. Wang, and Y. Zhou, “Doa estimation with uniform linear arraysin the presence of mutual coupling via blind calibration,” Signal Processing, vol. 89,no. 7, pp. 1446–1456, 2009.

[217] S. Triadaphillou, I. Wells, A. Morris, and E. Martin, “Investigation of calibration-free resolution techniques and independent conmponent analysis,” Proceedings ofInternational Symposium on Advanced Control of Chemical Processes (ADCHEM),2003.

[218] A. de Juan and R. Tauler, “Multivariate curve resolution (mcr) from 2000: progress inconcepts and applications,” Critical Reviews in Analytical Chemistry, vol. 36, no. 3,pp. 163–176, 2006.

[219] B. Vandeginste, W. Derks, and G. Kateman, “Multicomponent self-modelling curveresolution in high-performance liquid chromatography by iterative target transfor-mation analysis,” Analytica chimica acta, vol. 173, no. 1, pp. 253–264, 1985.

[220] G. Wang, Q. Ding, and Z. Hou, “Independent component analysis and its applicationsin signal processing for analytical chemistry,” TrAC Trends in Analytical Chemistry,vol. 27, no. 4, pp. 368–376, 2008.

[221] R. S.C.Rutan, A.de Juan, “Introduction to multivariate curve resolution,” In Compre-hensive Chemometrics, vol. 2, pp. 249–258, (ch.15), S.D.Brown, R.Tauler, B.Walczak(Eds.), Elsevier, 2009.

[222] A. J. R.Tauler, “Multivariate curve resolution,” In Practical Guide to Chemometrics,pp. 417–467, (ch.11), P.Gemperline (Ed.), Taylor and Francis Group , 2006.

[223] H. Gampp, M. Maeder, C. Meyer, and A. Zuberbuehler, “Evolving factor analysis,”Comments on Inorganic Chemistry: A Journal of Critical Discussion of the CurrentLiterature, vol. 6, no. 1, pp. 41–60, 1987.

[224] D. H.R.Keller, “Peak purity control in liquid chromatography with photodiode-arraydetection by fixed size moving window evolving factor analysis,” Analytica chimicaacta, vol. 246, pp. 379–390, 1991.

144

References

[225] W. Windig and J. Guilment, “Interactive self-modeling mixture analysis,” AnalyticalChemistry, vol. 63, pp. 1425–1432, Jul 1991.

[226] B. Vandeginste, W. Derks, and G. Kateman, “Multicomponent self-modelling curveresolution in high-performance liquid chromatography by iterative target transfor-mation analysis,” Analytica chimica acta, vol. 173, no. 1, pp. 253–264, 1985.

[227] F. C. Sánchez, S. Rutan, M. G. García, and D. Massart, “Resolution of multicompo-nent overlapped peaks by the orthogonal projection approach, evolving factor anal-ysis and window factor analysis,” Chemometrics and intelligent laboratory systems,vol. 36, no. 2, pp. 153–164, 1997.

[228] K. Schostack and E. Malinowski, “Investigation of window factor analysis and ma-trix regression analysis in chromatography,” Chemometrics and intelligent laboratorysystems, vol. 20, no. 2, pp. 173–182, 1993.

[229] R. Manne, H. Shen, and Y. Liang, “Subwindow factor analysis,” Chemometrics andintelligent laboratory systems, vol. 45, pp. 171–176, Jan 1999.

[230] O. Kvalheim and Y. Liang, “Heuristic evolving latent projections: resolving two-waymulticomponent 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 foranalysing diffusion battery data,” Journal of Aerosol Science, vol. 22, no. Supplement1, 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 andActuators B: Chemical, vol. 106, pp. 719–729, May 2005.

[236] M. Balaban, F. Korel, A. Odabasi, and G. Folkes, “Transportability of data be-tween electronic noses: mathematical methods,” Sensors and Actuators B: Chemical,vol. 71, pp. 203–211, Jan 2000.

[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.

[238] O. Tomic, T. Eklöv, K. Kvaal, and J. Haugen, “Recalibration of a gas-sensor ar-ray system related to sensor replacement,” Analytica chimica acta, vol. 512, no. 2,pp. 199–206, 2004.

145


[239] M. Muenchhof, M. Beck, and R. Isermann, “Fault-tolerant actuators and drives-structures, fault detection principles and applications,” Annual Reviews in Control,vol. 33, no. 2, pp. 136–148, 2009.

[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 applicationto 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 costautomation,” 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 aideddata-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 andfault tolerant control: Some advances and possible new directions,” Acta AutomaticaSinica, vol. 35, pp. 739–747, Jun 2009.

146

marta tesis final

Documents