7/27/2019 On the Performance of Gas Sensor Arrays in Open Sampling Systems Using Inhibitory SVM

1/16

Sensors and ActuatorsB 185 (2013) 462477

Contents lists available at SciVerse ScienceDirect

Sensors and Actuators B: Chemical

journal homepage: www.elsevier .com/ locate /snb

On the performance ofgas sensor arrays in open sampling systems usingInhibitory Support Vector Machines

Alexander Vergara a,,1,3,Jordi Fonollosa a,,1,Jonas Mahiques a, Marco Trincavelli b,2,Nikolai Rulkov a, Ramn Huerta a

a BioCircuits Institute, University of California, SanDiego, La Jolla, CA 92093, USAb Mobile Robotics andOlfaction Lab, Centre for Applied Autonomous Sensor Systems, rebro, Sweden

a r t i c l e i n f o

Article history:Received 8 December 2012Received in revised form 2 April 2013Accepted 8 May 2013Available online 18 May 2013

Keywords:Metal-oxide sensorsSupport Vector MachinesSystem calibrationOpen sampling systemSensor arrayElectronic nose

a b s t r a c t

Chemo-resistive transduction presents practical advantages for capturing the spatio-temporal and struc-tural organization ofchemical compounds dispersed in different human habitats. In an open samplingsystem, however, where the chemo-sensory elements are directly exposed to the environment beingmonitored, the identification and monitoring ofchemical substances present a more difficult challengedue to the dispersion mechanismsofgaseous chemical analytes,namely diffusion, turbulence, and advec-tion. The success ofsuch actively changeable practice is influenced by the adequate implementation ofalgorithmically driven formalisms combined with the appropriate design ofexperimental protocols. Onthe basis ofthis functionaljoint-formulation, in this study we examine an innovative methodology basedon the inhibitory processing mechanisms encountered in the structural assembly ofthe insects brain,namely Inhibitory Support Vector Machine (ISVM) applied to training a sensor array platform and evalu-ate its capabilities relevant to odor detection and identification under complex environmentalconditions.We generated and made publicly available an extensive and unique dataset with a chemical detec-tion platform consisting of72 conductometric metal-oxide based chemical sensors in a custom-designedwind tunnel test-bed facility to test our methodology. Our findings suggest that the aforementionedmethodology can be a valuable tool to guide the decision ofchoosing the training conditions for a cost-

efficient system calibration as well as an important step toward the understanding ofthe degradationlevel ofthe sensory system when the environmental conditions change.

2013 Elsevier B.V. All rights reserved.

1. Introduction

The design of inexpensive, reliable, fast responding, highly sen-sitive, and low-power consuming chemosensory array systems also referred to as electronic nose systems (e-nose) haslongbeenrecognized as a primary goal for the chemo-sensing community.Since 1982 when the first scientific publication of artificial olfac-tion was presented, the e-nose [1,2] has evolved into a valuableinstrument envisioned for a wide variety of applications, includ-ing, but certainly not limited to, medical diagnostics [35], foodspoilage monitoring [68,52], quality control [911], and other

Corresponding author. Tel.: +1 8585346758. Corresponding author. Tel.: +1 3019756690.

E-mail addresses: vergara@ucsd.edu (A. Vergara), fonollosa@ucsd.edu(J. Fonollosa).

1 Joint first authors.2 MarcoTrincavellicarriedout hisresearchon thisarticlewhilevisitingUniversity

ofCalifornia San Diego (UCSD).3 Present address: Material and Measurement Laboratory, National Institute of

Standards and Technology, Gaithersburg, MD, USA.

types of event detectionin urban [12,13] or space [14,15] scenarios.However, as efficient and successful as these instruments may haveproved to be in most of the aforementioned chemo-sensing appli-cations, the e-nose systems have neither achieved the potential inapplication tasks that go beyond the identification and differenti-ation of chemical substances nor the market penetration expectedby the pioneers [16,17]. Part of the reasons responsible for thecrash of these applications are the natural mechanisms dominat-ing the dispersion of chemical gaseous analytes in environmentalconditions, namely diffusion, turbulence, and advection, as well asthe enormous vulnerability of chemical sensors to temperature,humidity, and air flow. Therefore, although numerous computa-tional algorithms have been proposed to enhance the performanceof e-nose systems, which are important for sensor drift reduction[18,19]aswellasminimizingtheimpactofsensorfailures [20,21], itstill remains unconclusive to whatextent the information obtainedfrom these chemosensory systems can be exploited to reliably dis-criminate gases in realistic sampling scenarios, e.g., wind tunnels.

Gas-phase chemical analysis by means of electrical transduc-tion involving the prediction of the identity and quantity of thechemical compound has been traditionally performed in highly

0925-4005/$ seefrontmatter 2013 Elsevier B.V. All rights reserved.

http://dx.doi.org/10.1016/j.snb.2013.05.027

http://localhost/var/www/apps/conversion/tmp/scratch_4/dx.doi.org/10.1016/j.snb.2013.05.027http://localhost/var/www/apps/conversion/tmp/scratch_4/dx.doi.org/10.1016/j.snb.2013.05.027http://www.sciencedirect.com/science/journal/09254005http://www.elsevier.com/locate/snbmailto:vergara@ucsd.edumailto:fonollosa@ucsd.eduhttp://localhost/var/www/apps/conversion/tmp/scratch_4/dx.doi.org/10.1016/j.snb.2013.05.027http://localhost/var/www/apps/conversion/tmp/scratch_4/dx.doi.org/10.1016/j.snb.2013.05.027mailto:fonollosa@ucsd.edumailto:vergara@ucsd.eduhttp://www.elsevier.com/locate/snbhttp://www.sciencedirect.com/science/journal/09254005http://localhost/var/www/apps/conversion/tmp/scratch_4/dx.doi.org/10.1016/j.snb.2013.05.027

7/27/2019 On the Performance of Gas Sensor Arrays in Open Sampling Systems Using Inhibitory SVM

2/16

A. Vergara et al. / Sensors and Actuators B 185 (2013) 462477 463

tight-controlled sensing test chambers that isolate the chemicalanalyte from its natural, predominantly complex environmentalcondition. Because it ensures a strict, tight control over somecritical sensing conditions, including environmental temperature,pressure, and ambient flow, such isolation enables the chemicalsensory system to exhibit chemical signatures that are, to a verylarge extent, specific to both, the kind of sensory elements usedand the chemical analyte being monitored [53]. Traditional meth-ods of analytical chemistry, including, but certainly not limitedto, mass spectrometry and gas chromatography, are other promi-nent examples of such isolating techniques that are very useful inthe identification and quantification of chemical analytes [2224].These types of analysis, however, not only require indirect, rathercomplicated sampling procedures, also including cases involvingthe destruction of the tested sample, but also, and most impor-tantly, they cannot reveal the spatial and temporal structure of thechemical stimulus in its natural ambient (e.g., in an open samplingscenario), which, as we will demonstrate here, is essential to thechemo-sensing task envisioned in this work.

A gaseous chemical plume emitted froma fixedlocationconveystwo critical pieces of information of the sensory world enclosedwithin its own volume: information about the analyte identityand, to some extent, information revealing the spatial coordi-

nates of, or distance between, the source point and the observer(e.g., a chemosensory mobile platform). The problems of analyteidentification and chemical source localization as a case of studyare individually not novel; they have been studied for over twodecades, particularly in the fields of artificial olfaction and mobilerobot olfaction [16,17]. Chemical analyte identification is a genuineclassification problem, in which various mainstream methods havebeen successful thanks to the effective quantitative metrics thathave been recently established in the chemosensory community[16,25,26,54]. Chemical source localization is another importantchemo-sensing inference problem that has been explored for overtwo decades, predominantly in the field of robotics [17]. A vastmajority of most of the reported successful attempts thus far hasbeen theuse oflocalconcentration gradients inone form or another

in the search. This generic attribute, naturally generated by therapidly decayingaverageconcentrationlevel of thechemicalplumewith distance away from its source or by local zigzagging move-ments of mobileagents at theedge of the plume,is easilyaccessiblein consecutive raw chemo-sensory samplings in a plume that canguide a mobile agent (e.g., a robot) upstream towards the centerpoint of the plume in a broad range of scenarios [2730,33,40,41].Although all these efforts have clearly been paid-off by the enor-mous expansion in the applications involving machine olfaction,chemo-sensory systems preferentially utilize a diverse array ofgeneral assumptions and additional instruments, such as sensingtest chambers, for their functioning in the search that collec-tively contribute to the successful completion of the chemical taskbeing pursued. Thebenefit of theseassumption-mediated practices

seemingly outweighs the cost to chemo-sensory systems of imple-menting these assumptions, but it remains controversial whetherthese apparently non-avoidable solutions are feasible for real-lifeapplication works, as of to date.

The particular point of view pursued in this paper is that theidentity of a chemical analyte in an open sampling scenario canitselfbeinferredfrompreviousobservationsmadeonthesamefieldunder similar ambient conditions. In other words, a gas chemicalplume can be characterized empirically in a data-driven fashion,based on pre-recorded, stationary representative measurementscollected at certain distances from the origin of coordinates andunder flow conditions resembling those of the test-field conditions.A similar idea satisfying the aforementioned mobility requirementin source seeking has been already embraced in the language

of analytical fluid dynamics by an accurate spatiotemporal

characterization of the plume, but the computational costs in theirdesign and simulation make it infeasible for the kind of problemsconsidered in this work and discouraging for todays technology[31]. Therefore, our suggested approach constitutes, to the best ofour knowledge, the first implementation of its kind that underl-ies a significant novelty in analyte identification for open samplingsystems, as it utilizes distant and stationary measurements, elimi-nating the need for movements within the test field or towards thesource. Drawing on this premise, our work in this paper focuseson three major aspects that have not been considered in previ-ous attempts. First, we introduce to the chemosensory communityan innovative formulation that takes benefit from the inhibitoryprocessing mechanisms encountered in the structural assembly ofthe insects brain, namely the Inhibitory Support Vector Machine(ISVM) [32], to select the optimal calibration process of e-nose sys-tems when operating in open sensing scenarios. Our first goal thusis to provide evidence that the suggested methodology can be avaluable tool to guide the decision of choosing the training condi-tions for a cost-efficient system calibration as well as an importantstep toward the understanding of the degradation level of the sen-sory system when environmental conditions change. Second, wecreate an analyte identification phenomenon affected by differentnon-predictable environmental parameters changing with time.

Thesuggestedscenariocontains cleanand representativemeasure-ments from 10 chemical agents that arehighly recognizable to posean immediate danger to life and health, namely, acetone, acetalde-hyde, ammonia, butanol, ethylene, methane, methanol, carbonmonoxide,benzene, and toluene, recorded utilizing chemoresistivemetal oxide based chemical sensors located at different positionsin the test field. Hence, the second goal of this study is to demon-strate that the e-nose system can indeed be utilized to effectivelyrepresent the information contained in thesensors signal responsesubject to such turbulent ambient to discriminate chemical ana-lytes reliably. And third, through the comprehensive experimentsconsideredinthispublication,wepresentandmakefreelyavailableto the chemo-sensory and artificial olfaction community a uniquecomprehensive dataset for comparison of their algorithms and

tools.It is ourexpectationthatthis posting will provide researchersin thisand other fieldsa repositoryfor alternative competitive solu-tions relevant to the chemo-sensing discrimination task in opensampling systems pursued here and/or navigation. Going forward,by utilizingthe introduced database along with the proposed algo-rithm, we believe us to be uniquely positioned in building morecost-efficient recalibration protocols for the e-nose, which willultimately result in more robust, more accurate, and more stablemodels for the specific application at hand. In the remainder ofthis manuscript, we first describe the experimental details utilizedin this work, including the e-nose system, the database and thegeometry of the problem pursued in this work (Section 2). We thenoutline some of the theoretical details of the ISVM methodologyin Section 3, followed by presenting the discrimination results of

the proposed methodology in Section 4. Finally, we will discuss ourresults and present some of the concluding remarks drawn fromthe results presented here in Sections 5 and 6, respectively.

2. Experimental protocol

2.1. Chemical detection platform

Conductometricsensingprinciples, in general,have beenwidelystudied in several types of gas sensing schemes because theyare stable in many environments and within a wide tempera-ture range, sensitive to many potential analytes at a wide varietyof concentrations, rapid and reversibly responding, and inexpen-

sive, while performing reasonably well in discriminating chemical

7/27/2019 On the Performance of Gas Sensor Arrays in Open Sampling Systems Using Inhibitory SVM

3/16

464 A.Vergara et al. / Sensors and Actuators B 185 (2013) 462477

Table 1

MOX sensors included in the 8-sensor array. The manufacturer [37] adapts thesensing layer to detect different target gases, however, all the sensors responddifferently to a large number of chemicals.

Sensor type Number of units Target gases

TGS2611 1 MethaneTGS2612 1 Methane, propane, butaneTGS2610 1 PropaneTGS2600 1 Hydrogen, carbon monoxide

TGS2602 2 Ammonia, H2S, volatile organiccompounds (VOC)TGS2620 2 Carbon monoxide, combustible

gases, VOC

analytes [36]. Although they, too, have been predominantly usedin isolated settings, their high sensitivity and rapid response toa wide variety of volatiles distinguishes them as suitable chemo-transducers for ambient conditions, especially for the predictionproblems addressed in this paper. Consistent with this approach,we designed a general purpose chemical sensing platform con-taining nineportable chemo-sensory modules, each endowed witheight commercialized metal oxide gas sensors, provided by FigaroInc. [37] (see Table 1 for more details of the manufacturers rec-

ommended specificity preference of each device), to detect a widevariety of various analytes and follow the changes of their con-centration in the environment depicted by a chemical analyte gasplume generated in our wind tunnel facility. The chemo-sensingprinciple of the sensors particularly considered here relies on theinteraction of reducing or oxidizing analyte gases at or with thesurface of the metal oxide film causing changes in electrical con-duction through the film that is measured in form of time-seriesacross the electrodes of the sensor, thereby exhibiting, as a whole,a multivariate response to the different exposed odorous/odorlessgas stimuli. The magnitude of the sensors response to the chemi-cal analyte is signaled by a change in the electrical conductivity ofthe sensor film, which is tightly correlated with the analyte con-centration present on its surface; hence, changes in the analyte

concentration (mostly due to patches and eddies in the chemicalplume) are reflected in the sensor response in real-time and arethe origin of the temporal resolution (i.e., fluctuations in the timeseries) that is desired for the chemo-sensing task pursued here, asattested by the time series illustrated in Fig. 4.

In addition to the natural characteristics (i.e., sensitivity, selec-tivity, relatively quick response, and spatio-temporal resolution),the custom design of the sensing technology used in our sensoryplatform makes use of two degrees of freedom to manipulate thesensing process. The first one is architectural, in which the activesurface chemistryis a decisive factorin both thesensitivityand theselectivity of a sensing element to certain analytes.In particular, thesensing layers used in each of our sensory modules represent sixdifferent sensitive surfaces listed in Table 1, widening thereby the

receptivefield of thesensor module substantially.The second oneisthe active surface temperature, which is consistently adjustable byapplying a voltage value within the admissible voltage range pro-vided by the manufacturer to the built-in, independently reachableheatingelement thatwe callheater voltageVHavailableineachsensor device.4 Thus, this active surface temperature a.k.a. the sen-sors operating temperature, is an effective primary conditioningparameter of the above indicated oxidation/de-oxidation process,which, in principle, affects all aspects of the sensor response,including those concerning the selectivity (i.e., the sensors abil-ity to encode the analyte information), sensitivity (i.e., the sensors

4 The manufacturersguide reports that the admissiblerange of voltages to attain

thesensorsfunctioning surface temperature is between 3.8V and 6.2V.

capability to sense the chemical substance at a low concentration),and, most importantly, response time of the sensor to certain ana-lytes. Accordingly, two identical sensors can respond substantiallydifferently to the same analyte when their active layers are heatedat different temperature values. The reader is referred to reviewRefs. [35,38,39] for an in-depth treatment of temperature depend-ence on the sensor response.

In our particular chemical sensing platform, each portablechemosensory module is integrated with a customized sensor con-troller implemented with a microprocessor MSP430F247 (TexasInstruments Inc.). The peculiarity of this controller is that itenables continuous data collection from the eight chemical sen-sors through a 12-bit resolution analog-to-digital converter (ADC)device at a sampling rate of 100 Hz, the control of the sensorheater temperature by means of 10ms period and 6 V amplitudepulse-width-modulated (PWM) driving signals, and the two-waycommunication with a computer to analyze the acquired data fromsensors and control the sensor heaters. The setup provides serialdata communication to the PC via either a USB and/or one of thewirelesscommunicationmodules(BlueToothor WiFi)connected tothe sensor array controller by the UART port. Fig. 1 shows a graph-ical illustration of the functional block diagram of the designedsensor controller (top panel) and an image of one of the nine fabri-

cated sensing module units integrating our 72-dimensional metaloxide based gas sensor platform (bottom panel).

2.2. Wind tunnel test-bed facility and data collection

The methodology proposed in this study requires an accuratecharacterization of chemical plumes in a data-driven fashion. Arepository of high-quality measurements representing differentplumes is the basis of all inference tasks, from the optimizationof all the arbitrary parameters of the measurement field to the cor-rect identification of the chemical event. A wind tunnel is such asetting that enables implementing environmental conditions accu-rately and with minimal disturbance from external flows. Drawingon this idea, the following subsections feature some of the design

details of the wind tunneltest-bed facility as well as thedescriptionof the dataset and measurement procedures followed to collect itin our particular case of study.

2.2.1. Description of the wind tunnel test-bed facilityTo conduct the required experiments of our study, we con-

structed a 2.5m 1.2m0.4 m wind tunnel, a research test-bedfacility endowedwith a computer-supervisedmass flowcontroller-based continuous flow gas delivery system specifically designedto evaluate our chemo-sensory platform capabilities of detectingand identifying potentially dangerous chemical gaseous sub-stances underrealistic, fairly complicated conditions of stimulationencountered in real environments. Specifically, our customizedwind tunnel test-bed facility has a flat, non-inclined floor, which

enables us to disregard the altitude dimension of the test fieldwhile preserving sufficient airflow to create a sheer, yet turbu-lent chemical airstream across the wind tunnel. Additionally, thetest-bed facility also contains a flexible nozzle located upstreamof the test section, constructed entirely of Polytetrafluoroethylenetubing and the appropriate stainless-steel fitting connectors (pro-vided by Swagelok [42]) coupled with a gas delivery system, i.e., afully computerized continuous flow gas delivery rig endowed withthree digital mass flow controllers (M100B Mass-Flo, MKS Instru-ments [43]) each with differentmaximum flowlevels (200 sccm,100 sccm, and 20sccm; accuracy 1% Full scale) for accurate con-centration resolution, which together with calibrated pressurizedgascylinders(providedbyAirgasInc. [44]) isutilizedtoconsistentlyprovide the annotated test section with the potentially dangerous

chemical substances of interest at relevant concentrations. Being

7/27/2019 On the Performance of Gas Sensor Arrays in Open Sampling Systems Using Inhibitory SVM

4/16

A. Vergara et al. / Sensors and Actuators B 185 (2013) 462477 465

Fig. 1. The custom-designed block diagram used as the chemical detection platform, containing nine portable chemosensory modules (top). Front and rear view of thecustom-designed portable metaloxide based chemical sensory module (bottom, left and right, respectively). Each chemosensorymodule is endowed with eight metaloxidebased gas sensors and the necessary signal conditioning electronics and can be placed in different locations of the wind tunnel. In the board, the operating temperature ofthe MOX sensors is controlled by means of a PWM whereas the sensor resistance response is indirectly measured from a standard measuring circuit for an optimal signalconditioning (10k load resistor)before thesignals are acquired with a 12-bit resolution ADC.

Fig.2. Wind tunneltestbed facilityused tocollect time seriesdatafrom sensor arraysfor theproblemof gasidentification.The redpoint in theschema indicatesthe locationofthe chemical source. The displacements of the six training lines (or positions) are labeled in the schema as P1P6. Each training line includes nine landmarks evenlydistributed along theline to complete a grid of 54 evaluationlandmarks. Noticealso that thesmoke distributionshownin thewind tunnelis a dramatization that illustratesthe dispersion mechanisms of thegas emissionswithin thewind tunnel. An actual distributionmap of thechemical analytesbeing released is graphically illustrated in Fig.5.Finally, to measure the ambient temperature and humidity during the entire experiment in the wind tunnel we utilized a sensor tagged by the manufacturer, Sensirion, asSHT15.

7/27/2019 On the Performance of Gas Sensor Arrays in Open Sampling Systems Using Inhibitory SVM

5/16

466 A.Vergara et al. / Sensors and Actuators B 185 (2013) 462477

operated by a fully computerized environment controlled by aplayer/stage robot server software [45] programmed on C++on aPC fitted with the appropriate serial cards and with minimumhuman intervention, the designed wind tunnel test-bed facilityprovides versatility in releasing the chemical substances of interestat the desired concentrations with high accuracy and in a highlyreproducible manner during the entire experiment and simulta-neously in preserving the appropriate environmental conditionsto generate chemical gas plumes exhibiting turbulent patterns. Agraphical illustration of the designed wind tunnel test-bed facilityconsidered in this study along with thecharacteristics of thegeom-etry of the problem as well as the exact locations of the chemicalanalyte source and chemo-sensory platform is presented in Fig. 2.Optical graphical images of thedesignedwind tunnel areillustratedin the Supplementary Material of this paper, Fig. S1.

Supplementary material related to this article found, in theonline version, at http://dx.doi.org/10.1016/j.snb.2013.05.027.

The resulting test-bed facility operates in a propulsion open-cycle mode, by continuously drawing external turbulent air intoand throughout the tunnel and exhausting it back to the outside,thereby creating a relatively less-turbulent airflow moving down-stream towards the end of the test field. This operational mode isparticularly crucial for applications that require injecting chemical

poisonous agents or explosivemixtures because it prevents satura-tion. To createvarious distinct artificialairflows in thewind tunnel,we utilize a multiple-step motor-driven exhaust fan located insidethe wind tunnel at the outlet of the test section rotating at differ-ent constant rotational speeds, ranging from 1500rpm (25 Hz) to5500rpm (91.66Hz).

To maximize data quality, the above described facility is con-trolled and monitored by a digital distributed control system,operated and pre-programmedon theC++player/stage robot serversoftware outlined above. Unique in testing the chemo-sensingcapabilities of our customized sensory platform in multiple-speedcontinuous air streams, the said facility includes, on one side, twostate-of-the-art 2D Ultrasonic Anemometers (wind speed: operat-ing range 060m/s, accuracy 2% of the reading; wind direction:

operating range 0359, accuracy3), provided by Gill Windsonic[46], each strategically placed at different equidistant positions atthe vertices of a measurement grid withinthe test sectionand withan equally distributed spacing of 0.25m and 0.20m along the x-andy-axis, respectively, used to visualize and quantitatively char-acterize the generated airflows in the wind tunnel. The averageinduced wind speeds are collected from the said anemometers inthe test field, processed, and visualized in a graphical format at anupdate rate of once-per-second during 100 s. Being able to oper-ate at, and reach, such rotational speeds rapidly and reliably, thetunnel exhaust fan can consistently induce air streams at differentspeeds of up to 0.34 m/s, thereby potentially generating different,yet reproducible turbulent flow patterns. Fig. 3 illustrates quanti-tative characterizations of the wind direction andvariance of three

generated wind flow speeds 0.10m/s, 0.21m/s, and 0.34m/s induced by three different rotational speeds of the exhaust fanof 1500rpm (25Hz), 3900rpm (65 Hz), and 5500rpm (91.66Hz),respectively.5 As observed, each of the three particular wind flowvelocities exhibits different patterns in the flow in terms of winddirection and variability, demonstrating thereby the sharp effectsof the magnitude of the velocity of the air stream in the test sec-tion to produce different turbulent motion patterns in the windtunnel. On the other side, the facility is also equipped with a tem-perature and relative humidity sensor module, SHT15 (providedby Sensirion [47]), which is available to control and monitor the

5 From now on, we will refer to low, medium, and high wind speed to the three

generated wind flow speeds.

environmental conditions of the wind tunnel during the creationof the experiments.6 This control data is transferred back to anddisplayed in the computerized control platform outside the windtunnel for further processing.

2.2.2. Dataset and measurement procedureWe compiled a very extensive dataset7 utilizing nine portable

sensor array modules, each endowed with eight metal oxide gassensors manufactured by Figaro Inc. [37] (cf. Table 1 for specifics

of the types of chemical sensors utilized), positioned at six differ-ent line locations normal to the wind direction, creating therebya total number of 54 measurement locations that we call uni-form measurement grid uniformly distributed throughout theentire wind tunnel test-bed facility, as illustrated in Fig. 2. In par-ticular, ourdatasetconsistsof a very extensiveselection of multiplemission/scenario-representative chemical analyte species, namely,acetone, acetaldehyde, ammonia, butanol (butyl-alcohol), ethyl-ene, methane, methanol, carbon monoxide, benzene, and toluene,which in addition to their industrial applications as precursorsin the manufacture of explosives, narcotics, and polymers, thesechemical agents are highly recognizable to pose an immediatedanger to life and health in public and military places [48]. Wecall the proposed dataset/problem scenario the detection and clas-

sification of high-priority chemical hazards dataset task, and itis motivated by two different, exemplary mission scenarios pro-posed by the Department of Defense (DoD) for the development ofchemo-sensing solutions and standards for early warning and pro-tection of military forces against potential chemical and biologicalattacks, namely, the Force Protection featured by the MovementtoContactandDeliberateDefensivescenariosandtheFixedFacil-ity Situation which is represented by the External Attack on aFixed Facility and Internal Attack on a Fixed Facility scenes.8 Theproposed dataset ultimately induces a 10-class gas discriminationproblem, in which the goal is to identify and discriminate the 10distinct, high-priority chemical analyte hazards at relevant con-centrations in real-world operatingenvironments regardless of thelocation of the sensory systemplatform within theannotated wind

tunnel test-bed facility. The entire list of chemical analyte hazardsas well as their nominal concentration values at the outlet of thegas source in parts-per-million by volume (ppmv) utilized in thedataset considered in this work is summarized in Table 2.

To construct our proposed dataset, we utilized the afore-mentioned wind tunnel test-bed facility, into which the gaseoussubstancesof interest are to be unrestrictedly released, resemblingthereby the realistic, fairly complicated conditions of stimula-tion encountered in real environments. Because different arbitraryparameters from the sensing test-bed setting are available to betailored primarily at the operational level, we began constructingourdatasetby predefining the context of theproblemin a restrictednumber of pre-established conditioning parameters. First, becausethe heater voltage (VH), i.e., the active sensitive surface tempera-

ture, is highly recognizable as a primary conditioning parameterof the sensing technology, we predefined five different individualoperating temperatures within the admissible temperature range

6 The values of the actual temperature and humidity of the environment werenot controlled in the wind tunnel; however, they were recorded along the entireexperiment and have been included in the database. The corresponding mean andstandard deviation values of the whole set of experiments are T=22.21 C andRH=57% 10%.

7 The dataset is freely available at the UCI repository at http://archive.ics.uci.edu/ml/.

8 Although we recognizethat theproposedscenarios do notinclude all plausiblechemo-sensing scenarios or spectrum of challenges pursued by the DoD, they doprovide anexemplary setting bywhichto describeparametervalues intermsof keysensormetricsto address present concernsof theDoD forvarious scenarios. A more

in-depth discussion of the selected scenarios can be found in [49].

http://dx.doi.org/10.1016/j.snb.2013.05.027http://archive.ics.uci.edu/ml/http://archive.ics.uci.edu/ml/http://archive.ics.uci.edu/ml/http://archive.ics.uci.edu/ml/http://dx.doi.org/10.1016/j.snb.2013.05.027

7/27/2019 On the Performance of Gas Sensor Arrays in Open Sampling Systems Using Inhibitory SVM

6/16

A. Vergara et al. / Sensors and Actuators B 185 (2013) 462477 467

Fig. 3. Average andvariance wind directionmeasuredat a lattice of 60 distinct position pointswithin thetest field of thewind tunnel. Theexhaust fanrotational speed wasset to three differentvalues1500rpm (25Hz), 3900 rpm (65Hz), and5500 rpm(91.66 Hz)to induce three differentwind flow speedsthat were measured and characterized

by theanemometers, obtaining the following mean and standard deviationacross thewind tunnel: 0.100.004m/s, 0.210.005m/s, and0.340.009m/s, respectively.

Table 2

Chemical analytes in gas-phase and corresponding concentrations at the outlet ofthe gas source. All thechemicals are released at a constant flow of 320sccm.

Chemical a nalyte n ame Molecular f ormula Concentration

Acetone C3H6O 2500ppmAcetaldehyde C2H4O 500 ppmAmmonia NH3 10,000ppmButanol C4H9OH 100 ppmEthylene C2H4 500ppmMethane CH4 1000ppmMethanol CH4O 200 ppmCarbon monoxide CO 1000 ppm/4000 ppmBenzene C6H6 200ppmToluene C7H8 200ppm

suggested by the manufacturer, each attained by setting five dis-tinct constant voltage values to the heating element of the sensorsVH{4.0V, 4.5V, 5.0V, 5.5V, 6V}.9 Second, we restricted our mea-surement setting to operate at three distinct airflow velocities 0.10m/s, 0.21m/s, and 0.34m/s each individually inducedby setting the abovementioned exhaust fan to operate at threedifferent rotational speeds: 1500rpm (25 Hz), 3900rpm (65 Hz),

9 We do nothave accessto theactualsensingsurface operatingtemperaturedueto the packaging of thesensordevice, but a one-to-one look-up tablemapping relat-ingthe said activesurface temperatureto theheatervoltage (VH) is obtainable upon

request from themanufacturer in Ref. [37].

and 5500rpm (91.66Hz), respectively. Once these conditioningparameters were established for our setting, we continue ourexperiments by adopting the following procedure. First, we pos-itioned our chemo-sensory platform in one of the fixed linepositions indicated in the wind tunnel (see Fig. 2), and set the plat-forms chemical sensors to one of the above predefined surfaceoperating temperatures. One of the predefined artificial airflowswas then individually induced into the wind tunnel by the afore-mentioned exhaust fan, generating thereby the turbulent airflowwithin the test section of the wind tunnel. This stage constitutes apreliminary phase, which was implemented,both, to reach a quasi-stationary situation between the airflow and remaining ambient

conditions within the wind tunnel defining the background of ourexperiments and to measure the baseline of the sensor time seriesdata before the actual exposure of the chemical analyte got startedfor exactly 20s. We then randomly selected one of the10 describedchemical hazards (cf. Table 2) and released it into the tunnel atthe source location landmark indicated in Fig. 2 for as close toexactly 3 min as possible for the computer platform controllingthe three MFC devices at maximum flow, thereby allowing thechemical analyte to circulate throughout the wind tunnel whilerecording the generated sensor time series data in response tothe chemical analytes. Note that the chemical gas exposures areheld at the source mark in the tunnel directly from a nozzle con-nected to the outlet of a set of MFCs through a flexible pipelinecontaining the targeted gas at the nominal concentration specified

on each pressurized cylinder. Therefore, the concentration dosage

7/27/2019 On the Performance of Gas Sensor Arrays in Open Sampling Systems Using Inhibitory SVM

7/16

468 A.Vergara et al. / Sensors and Actuators B 185 (2013) 462477

Fig.4. Multivariateresponse of theMOX gassensorarray whenmethaneis released

in thewind tunnel. The sensorsresponses are affected by theair turbulencepresentin the wind tunnel causing fluctuations in the acquired signals. The sensor array isplaced in a representative position line of the wind tunnel (0.98m from the gassource, in theaxis of thechemical plume).The experimentalprotocolcarried outtoacquire thesignals of thesensors under differentconditions consists of thefollow-ing steps. (i) Setting of the operating temperature, measurement locations of thesensors, and the wind speed of the fans. (ii) During 20s, measurement of the base-line of thesensors signalwhileno chemical compound is released.(iii)Releasing ofthe chemical compound during3 min. (iv) Circulation of clean air forone minutetoacquire the sensors recovery signals. (v) Purging of the wind tunnel at maximumspeed for two additional minutes to clean the test field. Each measurement in thedataset is considered to finish in point(iv) sincethe sensorssignalsare notrecordedduring theexecution of point (v).

reported in Table 2 for this particular work represents only theconcentration atthe outletof the gassource ratherthan allthe mag-nitudes, or average of concentration levels observed by the sensors

deployed in the wind tunnel, which in principle should disperse astime passes by. After that step, the chemical analyte was removedand the test section was ventilated utilizing outside clean air cir-culating through the sampling setting at the same wind speed foranother minute. Before and after each experiment, the wind tun-nel was operated at maximum speed (i.e., 0.34m/s, induced by theexhaust fast at5500rpm(91.66 Hz)) fortwomoreminutes(datanotrecorded) to fully reestablish the sensor response baseline. In theend, the acquisition of the entire measurement experiment cycletook approximately 260s to complete, i.e., the baseline recordings,gas exposure, and cleaning recovery phase, recorded at a samplingupdate rate of 100 Hz (see Fig. 4 for a visualization of a typical timeseries response after a complete measurement was performed).This measurement procedure was reproduced exactly for each gas

category exposure, landmarklocationin thewindtunnel,operatingtemperature, and airflow velocity in a random order and up untilall pairs were covered. The resulting dataset ultimately comprises18,000 72-dimensional10 time series recordings each dimensioncorresponding to each sensorutilized observed from the 20repli-cascollected foreach operatingparameter andat each of the pointscovered in the wind tunnel, and it took approximately 16 monthsto be compiled. Finally, note that although different induced windspeeds strongly influence the structure and spatial distribution ofthe generated gas plumes in the sense that slow fan speeds

10 The total number of measurements is distributed as follows: 3 different windspeeds 5 different sensors temperatures 10 gases 6 locations in the wind

tunnel

20 replicas.

Fig. 5. Normalized sensor response to ammonia when the chemical detection plat-form is placed in different locations of the wind tunnel. The chemical source islocated at the most-left s ide of the wind tunnel in front of the board # 5 in theposition line # 1 of thewind tunnel. Theartificial airflows move downstream to theend of the tunnel (left to right in the figure) at different velocities proportional tothe operating rotational speeds of the motor-driven exhaust fan, strongly influenc-ing thestructure andspatial distributionof thegenerated chemical gasplumes. The

panels illustrate thespatial distribution of the chemical analyte gas plume flowingacrossthe tunnelat slow (top panel)and fast (bottom panel)wind speeds. Thevari-ations in the air conditions change the gas plume and the system predictions maybe affected. Operating heater voltage, VH=6V .

induce less stable patterns of the air flow direction, resulting inwider gas plumes, whereas faster velocities in the wind generatenarrower gas plumes (see Fig. 5) there is no symmetry in thespatial distribution of the plume with respect to the main axis (i.e.,the line connecting the chemical analyte source to the exhaust).A plume demonstrating a perfect symmetry in real environmentalconditions is rare due to the existent non-symmetry of the volumeenclosing the field, the inhomogeneous temperature in the ambi-ent, and the variability of the flow direction. In the setup phase of

the data collection we found in all trials that the elected featuresevaluated at symmetry coordinates were strictly different. And yet,as we will demonstrate below, thisnon-symmetry plume structureis reproducible so that the predictions can be extended in time byreferring to earlier observations in the same environment.

3. Inhibitory SVM

Support Vector Machine (SVM) is a machine learning techniquethat builds a model based on a training dataset to classify the ele-ments of another dataset of interest, called validation dataset. Evenif SVM is essentially a binary classifier, traditionally, it has beenextended to multi-class problems by building a classifier for eachpair of classes (one versus one strategy)or a classifier for each class

with respect to the rest of the samples (one versus all strategy).

7/27/2019 On the Performance of Gas Sensor Arrays in Open Sampling Systems Using Inhibitory SVM

8/16

A. Vergara et al. / Sensors and Actuators B 185 (2013) 462477 469

Fig. 6. Accuracy of the e-nose system trained in one position and tested in the rest of the considered lines of the wind tunnel. The system is exposed to three wind speeds(0.1 m/s, 0.21 m/s, 0.34 m/s). VH=5V .

However, in this work we are using a variation of SVM calledInhibitory SVM (ISVM) [32] to provide a more robust multi-classclassifier inspired by the inhibition process present in animal neu-ral systems. The result is a classifier more robust to the selection ofthe metaparameters when the models are built with small numberof training examples.

ISVM trains one classifier, fj, for each of the classesj = 1, . . ., L.Then, ISVM essentially compares the classifier for class j to theaverage of the output of all the classifiers, whereas traditionalmulti-class SVM methodologies perform pairwise comparisons.However, in the ISVM methodology and likewise in the SVMapproach, slackvariables [50] are also included in thefunction to beoptimized to increase the flexibility of the model, and the standardkernel trick [51] is utilized to improve the accuracy of the classifier

for non-linear datasets using a Gaussian kernel. Therefore, thereare still two metaparameters to be selected for a better accuracyof the classifier: the cost parameter C that controls the contribu-tion of the misclassified samples during the training, and thatdefines the bandwidth of the Gaussian kernel function. To opti-mize the metaparameters selection, we used a software written inC/C++that provides the error in the cross-validation to estimate theaccuracy of the classifier. We searched the optimal values of themetaparameters in the 44 grid defined by{0.01, 0.05, 0.1, 1.0}and C{10, 100, 1000, 5000} by minimizing the error in the cross-validation. More details on the ISVM methodology along with theutilized C/C++software can be found in Ref. [32].

4. Results

4.1. Sensor in a turbulent plume

The complete chemo-sensory platform consists of nine sensorarrays, each of them composed of eight sensors, for a total of 72sensors contained in the complete platform. The chemo-sensoryplatform is positioned at six different lines located normal to thewind direction and covering the entire wind tunnel, making, thechemical detection platform work in a turbulent sampling setting,whereinthesensorsareexposedtodifferentairflowconditionsthatchange the distribution of the gas of interest in the sampled area.Fig. 4 shows a graphical illustration of the multivariate response ofthe metal oxide based chemical sensor module to methane.

Fig. 5 illustrates the maximum of the normalized response (to

the maximum signal to make the data lie between 0 and +1) to

Table 3Cross-validation for different selection of the metaparameters Cand . The ISVMmethodology is robustfor a broad selection of metaparameters.Trainingin row# 6,VH=6 V,low wind speed.

Cvalues Values

0.01 0.05 0.1 110 73.89 69.06 70.24 92.03

100 65.31 79.86 89.10 96.131000 81.83 93.18 98.21 98.155000 88.15 95.48 97.89 97.01

Ammonia of the individual sensors when located at different pos-itions of the wind tunnel and air circulates at low or high speeds.Notice that the measured Ammonia plume shows similar proper-

ties to plume measured by Justus et al. [34], where an increasein plume size results in a progressive reduction of the measuredgas concentration. The comparison between Fig. 5 (top vs. bot-tom panels) qualitatively shows that the different air circulation inthe wind tunnel affects the distribution of the measured chemical

Table 4

Accuracy of the e-nose system (in %) trained in one position line and tested in therest of the considered lines of the wind tunnel. The system is exposed to threewind speeds(0.1m/s, 0.21 m/s, 0.34 m/s) while thevoltage heater(VH) is s et to5V .(VH=5V).

Training position Wind speed (m/s) Mean (Min, Max)

1 0.10 36.48 (15.68, 99.84)0.21 36.07 (9.73, 99.90)0.34 43.30 (9.19, 99.92)

2 0.10 62.63 (34.05, 98.32)0.21 62.14 (35.61, 97.60)0.34 62.95(33.17,96.81)

3 0.10 66.74 (39.51, 99.99)0.21 69.11 (48.29, 100)0.34 68.05(50.24,99.94)

4 0.10 74.25 (51.89, 99.98)0.21 80.80 (64.88, 99.32)0.34 80.22 (61.46, 99.75)

5 0.10 69.58 (46.34, 98.82)0.21 66.11 (36.59, 99.24)0.34 66.88(40.49,98.75)

6 0.10 66.45 (45.37, 98.21)0.21 65.60 (35.12, 99.99)0.34 71.05 (46.34, 99.99)

7/27/2019 On the Performance of Gas Sensor Arrays in Open Sampling Systems Using Inhibitory SVM

9/16

470 A.Vergara et al. / Sensors and Actuators B 185 (2013) 462477

Fig. 7. Average accuracy of themodels trained in onepositionlandmarkand validated in therest of thepositions. Themodels aretrained andvalidatedat thesame sensorstemperature andwind speed.Models trained in positionlines# 1 and# 2 show poor performance when used to predictthe gasclassin other positions, while modelstrainedin line # 4 can be extended to other locations.

Fig. 8. Correlation between the prediction of models trained in one location and evaluated in the remaining position lines. The model trained in line #1 is significantlydifferentfrom therest of themodels, forboth lowwind speed (left)and high wind speed (right), andthereby therobustness of themodels trained in #1 is limited when thesystemis tested in other locations.

analytes in the following fashion: (i) the gas plume becomes widerfor low wind speed conditions and (ii) the chemical sensors tendto present weaker responses when exposed to higher wind speedsdue to the drop in the concentration of the gas of interest causedby the increase of the carrier gas flow. Therefore, we can concludethat the MOX gas sensors are influenced differently by the distinctwind conditions and that the prediction models may be affectedwhen tested in changing environmental conditions.

4.2. Single position training

The collected dataset composed of 10 different classes wasdivided into a training and a validation subsets to build an optimalcomputational model for the e-nose system and to test the accu-

racy in the predictions of the complete system, respectively. We

selected the optimal values ofCand by minimizing the cross-validation in the 44 grid defined by {0.01, 0.05, 0.1, 1.0} andC{10, 100, 1000, 5000}. Becuase, the ISVM methodology showshigh robustness in the selection of the metaparameters, as shownin Table 3, the estimated accuracy of the classifier remains high fora broad selection ofCand .

In order to find the best location to train the complete platformwhile having good accuracy in the system predictions across theentire wind tunnel, we trained the complete platform in each ofthelines andvalidated themodelin the rest of the considered loca-tions. Thus, we built a model searched the optimal Cand for the complete platform in one line, and evaluated the accuracyin the predictions of the built model using the data acquired inthe rest of the locations. We repeated the same methodology to

quantify the degradation of the system when the active layer of

7/27/2019 On the Performance of Gas Sensor Arrays in Open Sampling Systems Using Inhibitory SVM

10/16

A. Vergara et al. / Sensors and Actuators B 185 (2013) 462477 471

Fig. 9. Accuracy of the e -nos e s ystem when trained in all the lines , or in twolines (16, 25, or 34). Applied voltage in the sensor heater: VH= 6 V | Windspeed=0.21m/s.

the sensors was at different temperatures and when the systemwas exposed to different wind conditions. Fig. 6 and Table 4 showthe degradation in the predictions of the e-nose system for the

three considered wind speeds and when 5 V are constantly appliedto the sensor heaters when it is validated in a different loca-tion than it was previously trained. Figs. A1A4 and Tables A1A4show the same tendencies when the operating temperature of thesensor is different. Fig. 7 shows the mean accuracy in the modelspredictions when they are trained in one location and validated inother positions.The changesin thegasplume produced by differentwind conditions applied in the tunnel significantly alter the per-formance of the models when tested at different locations than thetraining position. However, the degradation of the systems trainedin the same line is similar when the sensors are operating at dif-ferent temperatures or are exposed to different wind conditions:the models trained in line # 4 (aboutthe middleof thewindtunnel)show betterrobustness when aretested in the rest of the wind tun-

nel, especially at higher sensor operating temperatures and windspeeds. On the other hand, the models trained in the line just infront of the chemical source have low prediction accuracy whenthey are extended to other locations. This is due to the particulargeometry of thewind tunneland the gascirculationinside of it thatmake the plume too narrowin line# 1 tobe recognized intheotherlocations of the tunnel.

We compared the similarity of the different models trained atdifferent locations by measuring the correlation of the respectiveaccuracies in the different locations. From Fig. 8 we can concludethat the model based on location line #1 is significantly differ-ent from the rest of the models built and thereby its robustnessis lower compared to the other models. Moreover, there is gradualdegradation of the performance when testing further away from

the training site.

4.3. Multiple position training

As we uncovered in the previous section, one single location fortraining the e-nose might be sufficient if we know that the oper-ating testing location of the sensor array is located in the vicinityof the training site. The question now is how we can improve therobustness of the sensor array. A natural manner to address thisissue is to train our e-nose system in several distinct locations ofthe wind tunnel, even if this strategy has a higher associated costdue to the effort to collect data in such different locations. Thus,in order to explore the robustness of the system when trained inseveral locations, we built new models trained with data acquired

in various locations.

Table 5

Accuracy of the models when are trained in several positions and validated in allthe lines of the wind tunnel. The robustness of the e-nose system is increased bytraining the system in more locations (no significant drop in the minimum of thesystem accuracy). When the system is trained in two different locations, the bestpair of training positions(indicated in bold) are at the beginning of thetunnel (P1)and at the very end (P6).

Parameters Training positions Mean (Min, Max)

VH=4 V; Wind= 0.10m/s All 87.14 (84.47, 92.23)

P1andP6 75.91 (62.14, 98.06)P3 and P4 73.33 (41.94, 99.03)P2 and P5 70.51 (56.31, 92.23)P3 68.40 (40.98, 99.84)

VH=4 V; Wind= 0.34m/s All 92.64 (85.44, 99.03)P1andP6 83.98 (59.22, 100)P3 and P4 78.83 (54.37, 100)P4 75.91 (52.20, 99.38)P2 and P5 67.75 (47.31, 91.26)

VH= 6 V; Wind = 0.21 m/s All 96.55 (95.1500, 97.09)P1andP6 89.15 (79.61, 100)P3 and P4 82.33 (58.06, 100)P4 80.81 (65.37, 99.85)P3 and P5 73.92 (48.39, 98.06)

Each newmodel obtained from the newtrainingconditionswas

tested in all the locations of the wind tunnel. In particular, webuilt one additional model trained in all the considered lines ofthe wind tunnel, and three additional models trained in two morelocations (16, 25, or 34). Fig. 9 shows the accuracy of the cor-responding models when the sensors are operated at 6V and thewind speed was set to 0.21m/s. As observed, the accuracy of themodel drops when itis testedin a positionnot included inthe train-ing set; yet, the models show qualitatively the same tendency fordifferent sensors operating temperatures and wind conditions (cf.Figs. A5andA6). Table 5 showsthe minimum, maximum, and aver-age accuracy of the models when trained and validated in differentlocations. For all the different models, the e-nose system showshigh accuracy in its predictions in the position lines where themodelwas previouslytrained.In particular, themodeltrained in all

the lines shows high performance in all the positions of the windtunnel, whereas the other models show a significant drop whentested in other locations than the ones used for training. Therefore,we can conclude that the robustness of the system is consistentlyimproved when increasing the number of locations for training ofthe model. Hence, the user can select the number of position lines

Fig. 10. Accuracy of the e-nose system trained in line # 4 at one windspeed (0.21 m/s, solid lines) or considering the three wind speeds(0.1m/s +0.21m/s +0.34m/s, dashed lines ). The models are validated at highwind speed (green), medium wind speed (blue) and low wind speed (red). Appliedvoltage in the sensing layer: VH= 5 V. (For interpretation of the references to color

in this figure legend, thereader is referred to theweb version of thearticle.)

7/27/2019 On the Performance of Gas Sensor Arrays in Open Sampling Systems Using Inhibitory SVM

11/16

472 A.Vergara et al. / Sensors and Actuators B 185 (2013) 462477

Table 6

Accuracy of the e-nose system trained in line #4 at one single wind speed or considering the three wind speeds. The models are validated at the three wind speeds. Themodel trained at different speeds show higher robustness. VH=5V .

Training Validation

0.10 m/s 0.21 m/s 0.34m/s

Mean Min Max Mean Min Max Mean Min Max

0.10m/s 74.25 51.89 99.98 72.75 54.05 99.51 54.66 42.44 74.630.21 m/s 69.29 42.44 97.07 80.80 64.88 99.32 70.20 52.43 92.68

0.31 m/s 66.12 43.41 85.85 78.48 61.95 94.15 80.22 61.46 99.75All 76.69 51.35 100.00 82.73 65.85 100.00 81.85 63.78 99.03

to include in the system calibration according to the resources, theavailability of the time to collect data, and the accuracy needed inthe area of interest.

In Table 5 we show that the best average performance can beachieved by training the sensor arrays at the most-left side of thewind tunnel (i.e., position line # 1) and at the most-right side ofthe wind tunnel (i.e., position line #6), which is in contrast withthe results shown in the previous section concerning training foronesingle linelocation. Tables A1A4 showthatline#1leadstotheworst possible performance. The combination of the worst locationwith a decent performer (note that it is not the best) leads to thebest overall performer in combination.

4.4. Multiple wind speed training

The operating temperature of commercial MOX gas sensors canbe easily controlled by applying the selected voltage to the built-inheater. However, in open sampling systems, the wind conditionsare hardly controlled and any variation of the air flow may affectthe predictions of the system.

We testedthe degradationof thesystem when the airflow is dif-ferent from the training conditions. In particular, we evaluated theperformanceofthemodeltrainedatmediumwindspeed(0.21m/s)and in the position line # 4 when exposed to low(0.1m/s), medium(0.21 m/s) or high (0.34 m/s) wind speeds. Fig. 10 shows the accu-racy of the system and its drop in the performance when the wind

conditions are changed. As observed, the accuracy of the e-nosesystem degrades as the testing position moves further from thetraining position line #4. However, in all the considered testingpositions, the performance of the system decreases systematicallywhen the tested wind conditions change with respect to the train-ing conditions. We trained a system at the three considered windspeeds to quantify the improvement in the robustness of the sys-tem. From Fig. 10 we can conclude that the accuracy of the systemremains high even when the sensors areexposed to anyof thewindspeeds used in the training; hence, the robustness of the system isimproved. This outcome is perhaps not such a surprising conclu-sion, but we had to verify that multiple speeds improve overallperformance of the e-nose or sensor array.

Table 6 summarizes a comparison between the accuracy of the

modelstrained atone specific wind speed and validatedat differentwind speeds with the performance of the model trained at threewind speeds. The accuracy of the model significantly decreaseswhen the wind speed in the test conditions is not included in thetraining conditions. However, when the model is built using thethree different wind speeds, the accuracy of the system remainshigh. Therefore, in order to increase the robustness of the systemagainst air flow variations, one may want to train the system in allthe expected system conditions.

5. Discussion

In this paper we presented a methodology based on ISVM toevaluate the accuracy in the gas discrimination predictions of an

e-nose system. We showedthat the most robust strategy is to train

in all the locations of interest and under all the possible environ-mental conditions. However, due to time and cost restrictions, thesystemcan often only be trained undera subsetof allthesepossibleconditions.

If we are constrained to a one single position calibration, weshowed that the accuracy in the e-nose predictions is optimal inthe middle transverse line of the wind tunnel. This position isthe most informative position. We also showed that the systemperformance is affected by the wind speed used during trainingconditions, and again, if the system can only be trained at onewind speed, the performance is maximal when the medium windspeed is used fortraining. We foundthatthe lowwindspeedgener-

ates a significantly different plume compared to medium and highwind speeds and the predictions of a system trained under lowwind speed conditions areimpairedwhenextendedto higherwindspeeds.

If the system can be trained in two differentpositions, the accu-racy of the system is best when it is trained in the two locationsfeaturing the highest variability (first and last lines). This out-come contrasts with the solution obtained for one single locationtransversely to the tunnel. Although the training position is moresignificant than the sensors operating temperature when evalu-ating the ability of the e-nose system to predict the nature of newsamples,we alsofound thatthe e-nose system shows higher perfor-mance when the sensors operate at higher temperatures. Althoughinformation-theoretic approaches have been presented to chose

the optimal operating temperature of MOX sensors [38,39], fur-ther work includes a more systematic study to select the sensorsbest operating temperature (or temperatures) and maximize thediscrimination accuracy of open sampling settings.

The optimal training conditions depend on the environmentalconditions of the area of interest and their variations, while theoptimal locations for training the system are dependent on the dif-ferent points of interest and the air circulation in the room. Thereis still a trade off between the effort/cost carried out to calibratethe system and the accuracy of the e-nose. However, we showedthat our methodology can quantify the degradation of the predic-tions when the system is working under new conditions and guidethe decision of choosing the number and position of the chemicalsensing devices.

Finally, accompanying the results presented in this manuscript,we made publicly available a database containing the multivariateresponses of a MOX-sensor-based chemical sensing platform oper-ating in a turbulent wind tunnel. The complete database is uniqueof this kind since it reproduces, on the one hand, the uncontrolledenvironmental conditions of open sampling systems by circulat-ing the gas sample at different flows and, on the other hand, thecontrolled operating conditions of MOX sensors by changing thetemperature of the sensing layer. The dataset includes the recor-dings of the sensors at several locations of the built wind tunnelwhen exposed to 10 different gases of special interest for a total of18000 different measurements. We believe the generateddatabasewill be of a high value for the Artificial Olfaction community forthe comparison of different algorithms when exploring new tech-

niques to overcome the variation of the environmental conditions,

7/27/2019 On the Performance of Gas Sensor Arrays in Open Sampling Systems Using Inhibitory SVM

12/16

A. Vergara et al. / Sensors and Actuators B 185 (2013) 462477 473

searchingfor the best position of the sensing platform, or monitor-ing the air conditions in open environments.

6. Conclusions

We presented a methodology that can be considered as a proofof concept to evaluate the performance of sensor arrays working inopen sampling settings. We validated our methodology discrim-

inating 10 different gases at various concentrations. Our resultsindicate that the sensor array, which is always calibrated forminga line nearly perpendicular to the wind direction, should be alwaystrained in the most representative location that, for this particularcase of study, happen to be in the middle of the wind tunnel at dif-ferent wind speeds. In contrast, if two arrays can be used, the bestcalibration locations are the ones that induce the highest variabil-ity of the sensor responses, which are at the entrance and outletextreme points of the wind tunnel (most-left and most-right sidesof the wind tunnel in Fig. 2, respectively). If several wind speedsare available, the recommendation is to use several locations for

training. If the experimental time is limited, though, lower speedsshould be avoided. In terms of sensors operating temperature,higher sensor temperatures are always the best.

Acknowledgments

This work has been supported by U.S. Office of Naval Research(ONR) under the contract number N00014-07-1-0741, by Jet

Propulsion Laboratory under the contract number 2012-1455933,and by the US Army Medical Research and Materiel Commandunder contract number W81XWH-10-C-0040 in collaboration withElintrix. The views, opinions and/or findings contained in thisreport are those of the author(s) and should not be construedas an official Department of the Army position, policy or deci-sion unless so designated by other documentation. The authorsalso thank Joanna Zytkowicz for proofreading and revising themanuscript.

Appendix A.

Fig. A1. Accuracy of the e-nose system trained in one line and tested in the rest of the considered lines of the wind tunnel. The system is exposed to three wind speeds(0.1 m/s, 0.21 m/s, 0.34 m/s). VH=4V .

Fig. A2. Accuracy of the e-nose system trained in one line and tested in the rest of the considered lines of the wind tunnel. The system is exposed to three wind speeds(0.1 m/s, 0.21 m/s, 0.34 m/s). V

H=4.5V.

7/27/2019 On the Performance of Gas Sensor Arrays in Open Sampling Systems Using Inhibitory SVM

13/16

474 A.Vergara et al. / Sensors and Actuators B 185 (2013) 462477

Fig. A3. Accuracy of the e-nose system trained in one line and tested in the rest of the considered lines of the wind tunnel. The system is exposed to three wind speeds(0.1 m/s, 0.21 m/s, 0.34 m/s). VH=5.5V.

Fig. A4. Accuracy of the e-nose system trained in one line and tested in the rest of the considered lines of the wind tunnel. The system is exposed to three wind speeds(0.1 m/s, 0.21 m/s, 0.34 m/s). VH=6V .

Fig. A5. Accuracy of the e-nose system when trained in all the lines, or in twolines (16, 25, or 34). Applied voltage in the sensor heater: VH= 4 V | Windspeed=0.10m/s.

Fig. A6. Accuracy of the e-nose system when trained in all the lines, or in twolines (16, 25, or 34). Applied voltage in the sensor heater: VH= 4 V | Windspeed=0.34m/s.

7/27/2019 On the Performance of Gas Sensor Arrays in Open Sampling Systems Using Inhibitory SVM

14/16

A. Vergara et al. / Sensors and Actuators B 185 (2013) 462477 475

Table A1

Accuracy of the e-nose s ys te m trained in one line and te sted in the re st of theconsidered lines of thewind tunnel. VH=4V .

Training position Wind speed (m/s) Mean (Min, Max)

1 0.10 33.71 (14.05, 98.98)0.21 33.73 (13.51, 99.66)0.34 43.07 (24.32, 98.51)

2 0.10 58.01 (34.59, 88.60)0.21 61.24 (34.05, 94.90)

0.34 60.30 (35.14, 94.92)

3 0.10 68.40 (40.98, 99.84)0.21 74.94 (42.93, 99.94)0.34 73.85 (50.73, 99.99)

4 0.10 66.96 (25.85, 98.12)0.21 77.79 (63.90, 99.95)0.34 75.91 (52.20, 99.39)

5 0.10 65.71 (43.90, 98.22)0.21 65.30 (31.22, 99.33)0.34 70.39 (46.83,98.29)

6 0.10 60.34 (25.37, 96.19)0.21 65.85 (36.10, 99.99)0.34 68.12 (44.88, 99.97)

Table A2

Accuracyof thee-nosesystemtrained inone line andtestedin therest oftheconsid-ered lines of thewind tunnel. Thesystem is exposed to three wind speeds(0.1m/s,0.21 m/s, 0.34 m/s). VH=4.5V.

Training position Wind speed (m/s) Mean (Min, Max)

1 0.10 40.80 (14.59, 99.97)0.21 35.79 (11.89, 99.98)0.34 42.72 (24.32, 99.36)

2 0.10 62.26 (36.22, 95.92)0.21 61.66 (36.76, 97.14)0.34 61.73 (38.92, 97.33)

3 0.10 63.50 (40.98, 99.44)0.21 73.89 (41.95, 99.94)0.34 68.53 (47.80, 99.89)

4 0.10 71.06 (40.54, 99.99)0.21 78.35 (61.62, 98.75)0.34 79.08 (51.22, 99.74)

5 0.10 65.26 (45.85, 98.39)0.21 72.32 (38.54, 98.72)0.34 66.54 (32.20, 99.20)

6 0.10 67.96 (41.46, 98.54)0.21 64.63 (33.17, 99.99)0.34 67.31 (41.95, 99.99)

Table A3

Accuracyof thee-nosesystemtrained inone line andtestedin therest oftheconsid-ered lines of thewind tunnel. Thesystem is exposed to three wind speeds(0.1m/s,0.21 m/s, 0.34 m/s). VH=5.5V.

Training position Wind speed (m/s) Mean (Min, Max)

1 0.10 38.74 (12.97, 99.97)0.21 36.14 (9.19, 99.90)0.34 46.77 (7.57, 99.90)

2 0.10 62.74 (35.68, 99.34)0.21 64.97 (43.78, 96.83)0.34 63.80 (40.54, 96.94)

3 0.10 68.47 (42.16, 99.93)0.21 70.09 (49.27, 100)0.34 70.09 (49.76, 99.94)

4 0.10 71.71 (46.49, 99.88)0.21 76.66 (50.24, 99.43)0.34 78.02 (53.66, 98.74)

Table A3(Continued)

Training position Wind speed (m/s) Mean (Min, Max)

5 0.10 66.02 (41.46, 99.13)0.21 67.61 (33.17, 99.80)0.34 68.80 (46.83, 99.31)

6 0.10 68.50 (40.49, 99.29)0.21 68.86 (46.34, 99.99)0.34 71.86(45.85,99.99)

Table A4

Accuracy ofthee-nose systemtrainedin onelineandtestedin therest oftheconsid-ered lines of thewind tunnel. Thesystem is exposed to three wind speeds(0.1m/s,0.21 m/s, 0.34m/s). VH=6V .

Training position Wind speed (m/s) Mean (Min, Max)

1 0.10 35.78 (11.35, 99.98)0.21 34.65 (8.11, 99.34)0.34 49.29 (8.11, 99.84)

2 0.10 65.87 (35.68, 97.14)0.21 63.32 (34.05, 99.07)0.34 63.24 (34.59, 98.02)

3 0.10 65.14 (42.16, 99.93)0.21 73.11 (53.66, 99.99)0.34 67.99 (51.22, 99.96)

4 0.10 72.16 (44.32, 99.89)0.21 80.81 (65.37, 99.85)0.34 77.40 (65.37, 99.83)

5 0.10 63.79 (41.46, 98.79)0.21 71.15 (52.20, 99.07)0.34 69.37 (42.93, 98.82)

6 0.10 68.08 (37.56, 99.25)0.21 69.02 (46.34, 99.99)0.34 72.11 (46.34, 99.99)

References

[1] K.C. Persaud, G. Dodd, Analysis of discrimination mechanisms in the mam-malian olfactorysystem using a model nose, Nature 299 (1982) 352355.

[2] J.W. Gardner, P.N. Bartlett, A brief history of electronic noses, Sensors andActuatorsB: Chemical 18 (1994) 210211.[3] E. Thaler, D. Kennedy, C. Hanson, Medical applications of electronic nose tech-

nology: review of current status, American Journal of Rhinology 15 (5) (2001)291295.

[4] V. Tran, H.P. Chan, M. Thurs ton, P. Jackson, C. Lewis , D. Yates , G . B ell, P.Thomas, Breath analysis of lung cancer patients usingan electronic nosedetec-tion system, Sensors Journal, IEEE 10 (9) (2010) 15141518, http://dx.doi.org/10.1109/JSEN.2009.2038356.

[5] A.D.Wilson,M. Baietto,Advancesin electronic-nose technologies developedforbiomedical applications, Sensors 11 (1) (2011) 11051176, http://dx.doi.org/10.3390/s110101105.

[6] J.A. Stinson, K.C. Persaud, G. Bryning, D. Stead, Detection of potato pathogensusing an SPME-enose, Chemical Senses31 (8) (2006).

[7] A. Vergara,E. Llobet,J.L. Ramrez,P. Ivanov, L.Fonseca,S. Zampolli,A. Scorzoni,T.Becker,S. Marco, J.Wllenstein,An RFIDreaderwithonboardsensingcapabilityfor monitoring fruit quality, Sensors and ActuatorsB 127 (2007) 143149.

[8] J . Fonollos a, B. Halford, L. Fons eca, J. Santander, S. Udina, M. Moreno, J.Hildenbrand, J. WAllenstein, S. Marco, Ethylene optical spectrometer for apple

ripening monitoring in controlled atmosphere store-houses, Sensors andActuators B: Chemical 136 (2) (2009) 546554, http://dx.doi.org/10.1016/

j.snb.2008.12.015 http://www.sciencedirect.com/science/article/pii/S0925400508008307

[9] A. Tomchenko, G. Harmer, B. Marquis, Detection of chemical warfare agentsusing nanostructured metal oxide sensors., Sensors And Actuators B: Chemi-cal 108 (12) (2005) 4155, 10th International Meeting on Chemical Sensors,Tsukuba, Japan, July 1114.

[10] P .H. Rogers, S . Semancik, Development of optimization procedure s forapplication-specific chemical sensing, Sensors and Actuators B: Chemical163 (1)(2012) 819, http://dx.doi.org/10.1016/j.snb.2011.11.015 http://www.sciencedirect.com/science/article/pii/S0925400511010124

[11] M.H. Hammond, K.J. Johnson, S.L. Rose-Pehrsson, J. Ziegler, H. Walker, K.Caudy, D. G ary, D. Til lett, A novel chemical dete ctor using cermet s en-sors and pattern recognition methods for toxic industrial chemicals, Sensorsand Actuators B: Chemical 116 (12) (2006) 135144, ISOEN 2005. SelectedPapers from the 11th International Symposium on Olfaction and ElectronicNoses Barcelona, Spain, 1315 April 2005. doi:10.1016/j.snb.2005.12.065.http://www.sciencedirect.com/science/article/pii/S0925400506001754

http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0005http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0005http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0005http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0010http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0010http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0010http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0015http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0015http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0015http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0015http://localhost/var/www/apps/conversion/tmp/scratch_4/dx.doi.org/10.1109/JSEN.2009.2038356http://localhost/var/www/apps/conversion/tmp/scratch_4/dx.doi.org/10.1109/JSEN.2009.2038356http://localhost/var/www/apps/conversion/tmp/scratch_4/dx.doi.org/10.3390/s110101105http://localhost/var/www/apps/conversion/tmp/scratch_4/dx.doi.org/10.3390/s110101105http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0030http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0030http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0035http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0035http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0035http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0035http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0035http://localhost/var/www/apps/conversion/tmp/scratch_4/dx.doi.org/10.1016/j.snb.2008.12.015http://localhost/var/www/apps/conversion/tmp/scratch_4/dx.doi.org/10.1016/j.snb.2008.12.015http://www.sciencedirect.com/science/article/pii/S0925400508008307http://www.sciencedirect.com/science/article/pii/S0925400508008307http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0045http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0045http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0045http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0045http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0045http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0045http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0045http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0045http://localhost/var/www/apps/conversion/tmp/scratch_4/dx.doi.org/10.1016/j.snb.2011.11.015http://www.sciencedirect.com/science/article/pii/S0925400511010124http://www.sciencedirect.com/science/article/pii/S0925400511010124http://www.sciencedirect.com/science/article/pii/S0925400506001754http://www.sciencedirect.com/science/article/pii/S0925400506001754http://www.sciencedirect.com/science/article/pii/S0925400511010124http://www.sciencedirect.com/science/article/pii/S0925400511010124http://localhost/var/www/apps/conversion/tmp/scratch_4/dx.doi.org/10.1016/j.snb.2011.11.015http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0045http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0045http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0045http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0045http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0045http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0045http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0045http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0045http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0045http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0045http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0045http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0045http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0045http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0045http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0045http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0045http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0045http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0045http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0045http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0045http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0045http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0045http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0045http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0045http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0045http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0045http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0045http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0045http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0045http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0045http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0045http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0045http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0045http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0045http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0045http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0045http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0045http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0045http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0045http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0045http://www.sciencedirect.com/science/article/pii/S0925400508008307http://www.sciencedirect.com/science/article/pii/S0925400508008307http://localhost/var/www/apps/conversion/tmp/scratch_4/dx.doi.org/10.1016/j.snb.2008.12.015http://localhost/var/www/apps/conversion/tmp/scratch_4/dx.doi.org/10.1016/j.snb.2008.12.015http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0035http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0035http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0035http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0035http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0035http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0035http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0035http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0035http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0035http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0035http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0035http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0035http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0035http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0035http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0035http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0035http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0035http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0035http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0035http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0035http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0035http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0035http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0035http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0035http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0035http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0035http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0035http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0035http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0035http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0035http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0035http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0035http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0035http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0035http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0035http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0035http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0035http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0035http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0035http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0035http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0030http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0030http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0030http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0030http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0030http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0030http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0030http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0030http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0030http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0030http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0030http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0030http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0030http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0030http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0030http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0030http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0030http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0030http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0030http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0030http://localhost/var/www/apps/conversion/tmp/scratch_4/dx.doi.org/10.3390/s110101105http://localhost/var/www/apps/conversion/tmp/scratch_4/dx.doi.org/10.3390/s110101105http://localhost/var/www/apps/conversion/tmp/scratch_4/dx.doi.org/10.1109/JSEN.2009.2038356http://localhost/var/www/apps/conversion/tmp/scratch_4/dx.doi.org/10.1109/JSEN.2009.2038356http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0015http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0015http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0015http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0015http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0015http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0015http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0015http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0015http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0015http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0015http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0015http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0015http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0015http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0015http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0015http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0015http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0015http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0015http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0015http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0015http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0015http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0015http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0015http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0015http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0015http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0015http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0015http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0010http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0010http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0010http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0010http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0010http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0010http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0010http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0010http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0010http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0010http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0010http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0010http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0010http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0010http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0010http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0010http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0010http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0010http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0010http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0010http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0005http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0005http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0005http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0005http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0005http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0005http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0005http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0005http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0005http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0005http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0005http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0005http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0005http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0005http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0005http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0005http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0005http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0005http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0005http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0005http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0005http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0005http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0005http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0005

7/27/2019 On the Performance of Gas Sensor Arrays in Open Sampling Systems Using Inhibitory SVM

15/16

http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0270http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0270http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0270http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0270http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0270http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0270http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0270http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0270http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0270http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0270http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0270http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0270http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0270http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0270http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0270http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0270http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0270http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0270http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0270http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0270http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0270http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0270http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0270http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0270http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0270http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0270http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0265http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0265http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0265http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0265http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0265http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0265http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0265http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0265http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0265http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0265http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0265http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0265http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0265http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0265http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0265http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0265http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0265http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0265http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0265http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0265http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0265http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0265http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0265http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0265http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0265http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0265http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0265http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0265http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0265http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0265http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0265http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0265http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0265http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0265http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0265http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0265http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0265http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0265http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0265http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0265http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0260http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0260http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0260http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0260http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0260http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0260http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0260http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0260http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0260http://refhub.elsevier.com/S0925-4005(13)00590-X/sbref0260http://refh