You are on page 1of 16

Sensors and Actuators B 185 (2013) 462–477

Contents lists available at SciVerse ScienceDirect

Sensors and Actuators B: Chemical
journal homepage:

On the performance of gas sensor arrays in open sampling systems using Inhibitory Support Vector Machines
Alexander Vergara a,∗,1,3 , Jordi Fonollosa a,∗∗,1 , Jonas Mahiques a , Marco Trincavelli b,2 , Nikolai Rulkov a , Ramón Huerta a
a b

BioCircuits Institute, University of California, San Diego, La Jolla, CA 92093, USA Mobile Robotics and Olfaction Lab, Centre for Applied Autonomous Sensor Systems, Örebro, Sweden

a r t i c l e

i n f o

a b s t r a c t
Chemo-resistive transduction presents practical advantages for capturing the spatio-temporal and structural organization of chemical compounds dispersed in different human habitats. In an open sampling system, however, where the chemo-sensory elements are directly exposed to the environment being monitored, the identification and monitoring of chemical substances present a more difficult challenge due to the dispersion mechanisms of gaseous chemical analytes, namely diffusion, turbulence, and advection. The success of such actively changeable practice is influenced by the adequate implementation of algorithmically driven formalisms combined with the appropriate design of experimental protocols. On the basis of this functional joint-formulation, in this study we examine an innovative methodology based on the inhibitory processing mechanisms encountered in the structural assembly of the insect’s brain, namely Inhibitory Support Vector Machine (ISVM) applied to training a sensor array platform and evaluate its capabilities relevant to odor detection and identification under complex environmental conditions. We generated — and made publicly available — an extensive and unique dataset with a chemical detection platform consisting of 72 conductometric metal-oxide based chemical sensors in a custom-designed wind tunnel test-bed facility to test our methodology. Our findings suggest that the aforementioned methodology can be a valuable tool to guide the decision of choosing the training conditions for a costefficient system calibration as well as an important step toward the understanding of the degradation level of the sensory system when the environmental conditions change. © 2013 Elsevier B.V. All rights reserved.

Article history: Received 8 December 2012 Received in revised form 2 April 2013 Accepted 8 May 2013 Available online 18 May 2013 Keywords: Metal-oxide sensors Support Vector Machines System calibration Open sampling system Sensor array Electronic nose

1. Introduction The design of inexpensive, reliable, fast responding, highly sensitive, and low-power consuming chemosensory array systems — also referred to as electronic nose systems (e-nose) — has long been recognized as a primary goal for the chemo-sensing community. Since 1982 when the first scientific publication of artificial olfaction was presented, the e-nose [1,2] has evolved into a valuable instrument envisioned for a wide variety of applications, including, but certainly not limited to, medical diagnostics [3–5], food spoilage monitoring [6–8,52], quality control [9–11], and other

∗ Corresponding author. Tel.: +1 8585346758. ∗∗ Corresponding author. Tel.: +1 3019756690. E-mail addresses: (A. Vergara), (J. Fonollosa). 1 Joint first authors. 2 Marco Trincavelli carried out his research on this article while visiting University of California San Diego (UCSD). 3 Present address: Material and Measurement Laboratory, National Institute of Standards and Technology, Gaithersburg, MD, USA. 0925-4005/$ – see front matter © 2013 Elsevier B.V. All rights reserved.

types of event detection in urban [12,13] or space [14,15] scenarios. However, as efficient and successful as these instruments may have proved to be in most of the aforementioned chemo-sensing applications, the e-nose systems have neither achieved the potential in application tasks that go beyond the identification and differentiation of chemical substances nor the market penetration expected by the pioneers [16,17]. Part of the reasons responsible for the crash of these applications are the natural mechanisms dominating the dispersion of chemical gaseous analytes in environmental conditions, namely diffusion, turbulence, and advection, as well as the enormous vulnerability of chemical sensors to temperature, humidity, and air flow. Therefore, although numerous computational algorithms have been proposed to enhance the performance of e-nose systems, which are important for sensor drift reduction [18,19] as well as minimizing the impact of sensor failures [20,21], it still remains unconclusive to what extent the information obtained from these chemosensory systems can be exploited to reliably discriminate gases in realistic sampling scenarios, e.g., wind tunnels. Gas-phase chemical analysis by means of electrical transduction involving the prediction of the identity and quantity of the chemical compound has been traditionally performed in highly

pressure. to some extent. as we will demonstrate here. The problems of analyte identification and chemical source localization as a case of study are individually not novel. we first describe the experimental details utilized in this work. in which various mainstream methods have been successful thanks to the effective quantitative metrics that have been recently established in the chemosensory community [16. the source point and the observer (e. namely. The suggested scenario contains clean and representative measurements from 10 chemical agents that are highly recognizable to pose an immediate danger to life and health. stationary representative measurements collected at certain distances from the origin of coordinates and under flow conditions resembling those of the test-field conditions. Hence. such isolation enables the chemical sensory system to exhibit chemical signatures that are. to a very large extent. ethylene. And third.26.. a robot) upstream towards the center point of the plume in a broad range of scenarios [27–30. which will ultimately result in more robust. benzene. and toluene. including. namely the Inhibitory Support Vector Machine (ISVM) [32]. 2. such as sensing test chambers. A similar idea satisfying the aforementioned mobility requirement in source seeking has been already embraced in the language of analytical fluid dynamics by an accurate spatiotemporal characterization of the plume. respectively. particularly in the fields of artificial olfaction and mobile robot olfaction [16.25. we present and make freely available to the chemo-sensory and artificial olfaction community a unique comprehensive dataset for comparison of their algorithms and tools. for their functioning in the search that collectively contribute to the successful completion of the chemical task being pursued. through the comprehensive experiments considered in this publication. more accurate. carbon monoxide. to select the optimal calibration process of e-nose systems when operating in open sensing scenarios. to the best of our knowledge. not only require indirect. including environmental temperature. First. A gaseous chemical plume emitted from a fixed location conveys two critical pieces of information of the sensory world enclosed within its own volume: information about the analyte identity and. have been widely studied in several types of gas sensing schemes because they are stable in many environments and within a wide temperature range.1.A. The particular point of view pursued in this paper is that the identity of a chemical analyte in an open sampling scenario can itself be inferred from previous observations made on the same field under similar ambient conditions. information revealing the spatial coordinates of. We then outline some of the theoretical details of the ISVM methodology in Section 3. Therefore. Chemical detection platform Conductometric sensing principles.17]. chemo-sensory systems preferentially utilize a diverse array of general assumptions and additional instruments. Going forward. but certainly not limited to. they have been studied for over two decades. the second goal of this study is to demonstrate that the e-nose system can indeed be utilized to effectively represent the information contained in the sensors’ signal response subject to such turbulent ambient to discriminate chemical analytes reliably. tight control over some critical sensing conditions. is essential to the chemo-sensing task envisioned in this work.. in general. eliminating the need for movements within the test field or towards the source.. the kind of sensory elements used and the chemical analyte being monitored [53]. our work in this paper focuses on three major aspects that have not been considered in previous attempts. predominantly complex environmental condition. Chemical analyte identification is a genuine classification problem. methane. Because it ensures a strict.g. This generic attribute. but it remains controversial whether these apparently non-avoidable solutions are feasible for real-life application works. These types of analysis.54]. including the e-nose system. by utilizing the introduced database along with the proposed algorithm. our suggested approach constitutes.g. they cannot reveal the spatial and temporal structure of the chemical stimulus in its natural ambient (e. recorded utilizing chemoresistive metal oxide based chemical sensors located at different positions in the test field. methanol. but the computational costs in their design and simulation make it infeasible for the kind of problems considered in this work and discouraging for today’s technology [31]. and inexpensive. butanol.41]. naturally generated by the rapidly decaying average concentration level of the chemical plume with distance away from its source or by local zigzagging movements of mobile agents at the edge of the plume. also including cases involving the destruction of the tested sample. Traditional methods of analytical chemistry. the database and the geometry of the problem pursued in this work (Section 2). acetaldehyde. acetone. but also. a chemosensory mobile platform). rapid and reversibly responding. predominantly in the field of robotics [17].33. Our first goal thus is to provide evidence that the suggested methodology can be a valuable tool to guide the decision of choosing the training conditions for a cost-efficient system calibration as well as an important step toward the understanding of the degradation level of the sensory system when environmental conditions change. ammonia. Vergara et al. we believe us to be uniquely positioned in building more cost-efficient recalibration protocols for the e-nose. Chemical source localization is another important chemo-sensing inference problem that has been explored for over two decades. as of to date. Although all these efforts have clearly been paid-off by the enormous expansion in the applications involving machine olfaction. Drawing on this premise. which. is easily accessible in consecutive raw chemo-sensory samplings in a plume that can guide a mobile agent (e. are other prominent examples of such isolating techniques that are very useful in the identification and quantification of chemical analytes [22–24]. and most importantly.40. Experimental protocol 2. Second. followed by presenting the discrimination results of the proposed methodology in Section 4. In other words. specific to both. The benefit of these assumption-mediated practices seemingly outweighs the cost to chemo-sensory systems of implementing these assumptions. we create an analyte identification phenomenon affected by different non-predictable environmental parameters changing with time. In the remainder of this manuscript. A vast majority of most of the reported successful attempts thus far has been the use of local concentration gradients in one form or another in the search. while performing reasonably well in discriminating chemical . in an open sampling scenario). a gas chemical plume can be characterized “empirically” in a data-driven fashion. as it utilizes distant and stationary measurements. we will discuss our results and present some of the concluding remarks drawn from the results presented here in Sections 5 and 6. / Sensors and Actuators B 185 (2013) 462–477 463 tight-controlled sensing test chambers that isolate the chemical analyte from its natural. we introduce to the chemosensory community an innovative formulation that takes benefit from the inhibitory processing mechanisms encountered in the structural assembly of the insect’s brain. Finally. rather complicated sampling procedures. sensitive to many potential analytes at a wide variety of concentrations. and more stable models for the specific application at hand. It is our expectation that this posting will provide researchers in this and other fields a repository for alternative competitive solutions relevant to the chemo-sensing discrimination task in open sampling systems pursued here and/or navigation. the first implementation of its kind that underlies a significant novelty in analyte identification for open sampling systems. mass spectrometry and gas chromatography. and ambient flow.g. or distance between. however. based on pre-recorded.

.2 m × 0.. a multivariate response to the different exposed odorous/odorless gas stimuli. fairly complicated conditions of stimulation encountered in real environments. as a whole.5 m × 1. The reader is referred to review Refs. most importantly.e. / Sensors and Actuators B 185 (2013) 462–477 Table 1 MOX sensors included in the 8-sensor array. relatively quick response. 2. the sensor’s capability to sense the chemical substance at a low concentration). and. 2. The manufacturer [37] adapts the sensing layer to detect different target gases. [44]) is utilized to consistently provide the annotated test section with the potentially dangerous chemical substances of interest at relevant concentrations. all the sensors respond differently to a large number of chemicals. the sensor’s ability to encode the analyte information). affects all aspects of the sensor response.2.a. The second one is the active surface temperature. and the two-way communication with a computer to analyze the acquired data from sensors and control the sensor heaters. propane. Drawing on this idea.4 m wind tunnel. 100 sccm. the sensing layers used in each of our sensory modules represent six different sensitive surfaces listed in Table 1. H2 S. our customized wind tunnel test-bed facility has a flat. the test-bed facility also contains a flexible nozzle located upstream of the test section. The first one is architectural. each endowed with eight commercialized metal oxide gas sensors. to detect a wide variety of various analytes and follow the changes of their concentration in the environment depicted by a chemical analyte gas plume generated in our wind tunnel facility. Accordingly. independently reachable heating element — that we call heater voltage VH — available in each sensor device. non-inclined floor. in which the active surface chemistry is a decisive factor in both the sensitivity and the selectivity of a sensing element to certain analytes. this active surface temperature a.4 Thus. sensitivity (i. . the sensor’s operating temperature. i. a research test-bed facility endowed with a computer-supervised mass flow controllerbased continuous flow gas delivery system specifically designed to evaluate our chemo-sensory platform capabilities of detecting and identifying potentially dangerous chemical gaseous substances under realistic. hence. In our particular chemical sensing platform. changes in the analyte concentration (mostly due to patches and eddies in the chemical plume) are reflected in the sensor response in real-time and are the origin of the temporal resolution (i. constructed entirely of Polytetrafluoroethylene tubing and the appropriate stainless-steel fitting connectors (provided by Swagelok [42]) coupled with a gas delivery system. Additionally. the custom design of the sensing technology used in our sensory platform makes use of two degrees of freedom to manipulate the sensing process. which together with calibrated pressurized gas cylinders (provided by Airgas Inc. Wind tunnel test-bed facility and data collection The methodology proposed in this study requires an accurate characterization of chemical plumes in a data-driven fashion. Consistent with this approach. yet turbulent chemical airstream across the wind tunnel. Although they.38. 4. butane Propane Hydrogen. response time of the sensor to certain analytes. too. The magnitude of the sensor’s response to the chemical analyte is signaled by a change in the electrical conductivity of the sensor film. Sensor type TGS2611 TGS2612 TGS2610 TGS2600 TGS2602 TGS2620 Number of units 1 1 1 1 2 2 Target gases Methane Methane. Vergara et al. widening thereby the receptive field of the sensor module substantially. Being 4 The manufacturer’s guide reports that the admissible range of voltages to attain the sensor’s functioning surface temperature is between 3.39] for an in-depth treatment of temperature dependence on the sensor response. Description of the wind tunnel test-bed facility To conduct the required experiments of our study. especially for the prediction problems addressed in this paper. In addition to the natural characteristics (i. Specifically. 1 shows a graphical illustration of the functional block diagram of the designed sensor controller (top panel) and an image of one of the nine fabricated sensing module units integrating our 72-dimensional metal oxide based gas sensor platform (bottom panel). The chemo-sensing principle of the sensors particularly considered here relies on the interaction of reducing or oxidizing analyte gases at or with the surface of the metal oxide film causing changes in electrical conduction through the film that is measured in form of time-series across the electrodes of the sensor. have been predominantly used in isolated settings. VOC analytes [36].. accuracy ±1% Full scale) for accurate concentration resolution.e.e. selectivity.2 V. is an effective primary conditioning parameter of the above indicated oxidation/de-oxidation process.2. as attested by the time series illustrated in Fig. two identical sensors can respond substantially differently to the same analyte when their active layers are heated at different temperature values.1. which enables us to disregard the altitude dimension of the test field while preserving sufficient airflow to create a sheer. we designed a general purpose chemical sensing platform containing nine portable chemo-sensory modules. we constructed a 2. Fig. in principle. and spatio-temporal resolution). The setup provides serial data communication to the PC via either a USB and/or one of the wireless communication modules (BlueTooth or WiFi) connected to the sensor array controller by the UART port. however. from the optimization of all the arbitrary parameters of the measurement field to the correct identification of the chemical event. and 20 sccm. fluctuations in the time series) that is desired for the chemo-sensing task pursued here. In particular. carbon monoxide Ammonia. A repository of high-quality measurements representing different plumes is the basis of all inference tasks. thereby exhibiting. including those concerning the selectivity (i. MKS Instruments [43]) — each with different maximum flow levels (200 sccm. volatile organic compounds (VOC) Carbon monoxide.8 V and 6.). which is tightly correlated with the analyte concentration present on its surface. provided by Figaro Inc.k. The peculiarity of this controller is that it enables continuous data collection from the eight chemical sensors through a 12-bit resolution analog-to-digital converter (ADC) device at a sampling rate of 100 Hz.e. sensitivity. [37] (see Table 1 for more details of the manufacturer’s recommended specificity preference of each device). which is consistently adjustable by applying a voltage value within the admissible voltage range provided by the manufacturer to the built-in.464 A. their high sensitivity and rapid response to a wide variety of volatiles distinguishes them as suitable chemotransducers for ambient conditions.. combustible gases. the following subsections feature some of the design details of the wind tunnel test-bed facility as well as the description of the dataset and measurement procedures followed to collect it in our particular case of study. each portable chemosensory module is integrated with a customized sensor controller implemented with a microprocessor MSP430F247 (Texas Instruments Inc.e. A wind tunnel is such a setting that enables implementing environmental conditions accurately and with minimal disturbance from external flows. a fully computerized continuous flow gas delivery rig endowed with three digital mass flow controllers (M100B Mass-Flo. which. the control of the sensor heater temperature by means of 10 ms period and 6 V amplitude pulse-width-modulated (PWM) driving signals. [35..

Sensirion. Each training line includes nine landmarks evenly distributed along the line to complete a grid of 54 evaluation landmarks. In the board. An actual distribution map of the chemical analytes being released is graphically illustrated in Fig. The custom-designed block diagram used as the chemical detection platform. Vergara et al. Front and rear view of the custom-designed portable metal oxide based chemical sensory module (bottom. respectively). Notice also that the smoke distribution shown in the wind tunnel is a dramatization that illustrates the dispersion mechanisms of the gas emissions within the wind tunnel. left and right. 2. The displacements of the six training lines (or positions) are labeled in the schema as P1–P6. Wind tunnel test bed facility used to collect time series data from sensor arrays for the problem of gas identification. .A. Each chemosensory module is endowed with eight metal oxide based gas sensors and the necessary signal conditioning electronics and can be placed in different locations of the wind tunnel. 1. as SHT15. to measure the ambient temperature and humidity during the entire experiment in the wind tunnel we utilized a sensor tagged by the manufacturer. The red point in the schema indicates the location of the chemical source. / Sensors and Actuators B 185 (2013) 462–477 465 Fig. 5. the operating temperature of the 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 signal conditioning (10 k load resistor) before the signals are acquired with a 12-bit resolution ADC. Finally. Fig. containing nine portable chemosensory modules (top).

Supplementary material related to this article found. On the other side.e.2. we will refer to low. Dataset and measurement procedure We compiled a very extensive dataset7 utilizing nine portable sensor array modules. the tunnel exhaust fan can consistently induce air streams at different speeds of up to 0. Being able to operate at. we predefined five different individual operating temperatures within the admissible temperature range 5 From now on. demonstrating thereby the sharp effects of the magnitude of the velocity of the air stream in the test section to produce different turbulent motion patterns in the wind tunnel.466 A. each endowed with eight metal oxide gas sensors manufactured by Figaro Inc.21 m/s. creating thereby a total number of 54 measurement locations — that we call uniform measurement grid — uniformly distributed throughout the entire wind tunnel test-bed facility. the designed wind tunnel test-bed facility provides versatility in releasing the chemical substances of interest at the desired concentrations with high accuracy and in a highly reproducible manner during the entire experiment and simultaneously in preserving the appropriate environmental conditions to generate chemical gas plumes exhibiting turbulent patterns. ranging from 1500 rpm (25 Hz) to 5500 rpm (91. positioned at six different line locations normal to the wind direction. thereby creating a relatively less-turbulent airflow moving downstream towards the end of the test field. ics. 0. in the online version. we began constructing our dataset by predefining the context of the problem in a restricted number of pre-established conditioning parameters. they were recorded along the entire experiment and have been included in the database.5 As observed. accuracy ±2% of the reading. [37] (cf. exemplary mission scenarios proposed by the Department of Defense (DoD) for the development of chemo-sensing solutions and standards for early warning and protection of military forces against potential chemical and biological attacks. S1. Vergara et al. 7 The dataset is freely available at the UCI repository at http://archive. provided by Gill Windsonic [46]. We call the proposed dataset/problem scenario the detection and classification of high-priority chemical hazards dataset task. medium. 2. acetaldehyde. and polymers. which in addition to their industrial applications as precursors in the manufacture of explosives. is highly recognizable as a primary conditioning parameter of the sensing technology. we utilized the aforementioned wind tunnel test-bed facility. and visualized in a graphical format at an update rate of once-per-second during 100 s. narcotics. respectively. fairly complicated conditions of stimulation encountered in real environments. respectively.2013.20 m along the xand y-axis. at http://dx. A graphical illustration of the designed wind tunnel test-bed facility considered in this study along with the characteristics of the geometry of the problem as well as the exact locations of the chemical analyte source and chemo-sensory platform is presented in Fig. Table 1 for specifics of the types of chemical sensors utilized). operated and pre-programmed on the C++player/stage robot server software outlined above. into which the gaseous substances of interest are to be unrestrictedly released. processed.34 m/s — induced by three different rotational speeds of the exhaust fan of 1500 rpm (25 Hz). Fig. To create various distinct artificial airflows in the wind tunnel. accuracy ±3◦ ).1016/j. The corresponding mean and standard deviation values of the whole set of experiments are T = 22.05.snb. The resulting test-bed facility operates in a propulsion opencycle mode. SHT15 (provided by Sensirion [47]). the active sensitive surface temperature. namely. resembling thereby the realistic. and 5500 rpm (91. This operational mode is particularly crucial for applications that require injecting chemical poisonous agents or explosive mixtures because it prevents saturation.34 m/s. by continuously drawing external turbulent air into and throughout the tunnel and exhausting it back to the outside. and toluene. Optical graphical images of the designed wind tunnel are illustrated in the Supplementary Material of this paper. such rotational speeds rapidly and reliably.6 This control data is transferred back to and displayed in the computerized control platform outside the wind tunnel for further m and 0. because the heater voltage (VH ). 2. A more in-depth discussion of the selected scenarios can be found in [49]. which is available to control and monitor the environmental conditions of the wind tunnel during the creation of the experiments. Unique in testing the chemo-sensing capabilities of our customized sensory platform in multiple-speed continuous air streams. we utilize a multiple-step motor-driven exhaust fan located inside the wind tunnel at the outlet of the test section rotating at different constant rotational speeds. To maximize data quality. thereby potentially generating different. benzene.10 m/s.027. as illustrated in Fig.66 Hz). used to visualize and quantitatively characterize the generated airflows in the wind tunnel. these chemical agents are highly recognizable to pose an immediate danger to life and health in public and military places [48].2 ± 1 ◦ C and RH = 57% ± 10 %.doi. 3900 rpm (65 Hz). 2. i. they do provide an exemplary setting by which to describe parameter values in terms of key sensor metrics to address present concerns of the DoD for various scenarios. carbon monoxide. In particular. yet reproducible turbulent flow patterns. however. 3 illustrates quantitative characterizations of the wind direction and variance of three generated wind flow speeds — 0. butanol (butyl-alcohol).uci. First. the said facility includes. each strategically placed at different equidistant positions at the vertices of a measurement grid within the test section and with an equally distributed spacing of 0. . methanol. and it is motivated by two different. The entire list of chemical analyte hazards as well as their nominal concentration values at the outlet of the gas source in parts-per-million by volume (ppmv) utilized in the dataset considered in this work is summarized in Table 2. namely. To construct our proposed dataset. the facility is also equipped with a temperature and relative humidity sensor module. Fig.66 Hz). the above described facility is controlled and monitored by a digital distributed control system. ethylene. and high wind speed to the three generated wind flow speeds. high-priority chemical analyte hazards at relevant concentrations in real-world operating environments regardless of the location of the sensory system platform within the annotated wind tunnel test-bed two state-of-the-art 2D Ultrasonic Anemometers (wind speed: operating range 0–60 m/s. each of the three particular wind flow velocities exhibits different patterns in the flow in terms of wind direction and variability. The average induced wind speeds are collected from the said anemometers in the test field. 8 Although we recognize that the proposed scenarios do not include all plausible chemo-sensing scenarios or spectrum of challenges pursued by the DoD. the Force Protection — featured by the Movement to Contact and Deliberate Defensive scenarios — and the Fixed Facility Situation — which is represented by the External Attack on a Fixed Facility and Internal Attack on a Fixed Facility scenes.2. wind direction: operating range 0–359◦ . acetone. / Sensors and Actuators B 185 (2013) 462–477 operated by a fully computerized environment — controlled by a player/stage robot server software [45] programmed on C++ on a PC fitted with the appropriate serial cards — and with minimum human intervention. and 0. in which the goal is to identify and discriminate the 10 distinct. on one side.. 6 The values of the actual temperature and humidity of the environment were not controlled in the wind tunnel. methane. our dataset consists of a very extensive selection of multiple mission/scenario-representative chemical analyte species.8 The proposed dataset ultimately induces a 10-class gas discrimination problem. ammonia. and reach. Because different arbitrary parameters from the sensing test-bed setting are available to be tailored primarily at the operational level.

respectively. Chemical analyte name Acetone Acetaldehyde Ammonia Butanol Ethylene Methane Methanol Carbon monoxide Benzene Toluene Molecular formula C3 H6 O C2 H4 O NH3 C4 H9 OH C2 H4 CH4 CH4 O CO C6 H6 C7 H8 Concentration 2500 ppm 500 ppm 10. and 5500 rpm (91.004 m/s. and set the platform’s chemical sensors to one of the above predefined surface operating temperatures. but a one-to-one look-up table mapping relating the said active surface temperature to the heater voltage (VH ) is obtainable upon request from the manufacturer in Ref. Table 2 Chemical analytes in gas-phase and corresponding concentrations at the outlet of the gas source. the concentration dosage . 3. and 5500 rpm (91. we positioned our chemo-sensory platform in one of the fixed line positions indicated in the wind tunnel (see Fig. Vergara et al. which was implemented. 5. generating thereby the turbulent airflow within the test section of the wind tunnel. 0.000 ppm 100 ppm 500 ppm 1000 ppm 200 ppm 1000 ppm/4000 ppm 200 ppm 200 ppm suggested by the manufacturer. 2 for as close to exactly 3 min as possible for the computer platform controlling the three MFC devices at maximum flow. both. 9 We do not have access to the actual sensing surface operating temperature due to the packaging of the sensor device.005 m/s. We then randomly selected one of the 10 described chemical hazards (cf. [37]. 4.66 Hz). First. and 0. we continue our experiments by adopting the following procedure.009 m/s. each attained by setting five distinct constant voltage values to the heating element of the sensors VH {4. Therefore.34 ± 0. we restricted our measurement setting to operate at three distinct airflow velocities — 0. 2). respectively.5 V.10 ± 0. This stage constitutes a preliminary phase. One of the predefined artificial airflows was then individually induced into the wind tunnel by the aforementioned exhaust fan. obtaining the following mean and standard deviation across the wind tunnel: 0.A.21 ± 0. The exhaust fan rotational speed was set to three different values 1500 rpm (25 Hz).66 Hz) to induce three different wind flow speeds that were measured and characterized by the anemometers. Once these conditioning parameters were established for our setting. thereby allowing the chemical analyte to circulate throughout the wind tunnel while recording the generated sensor time series data in response to the chemical analytes. Average and variance wind direction measured at a lattice of 60 distinct position points within the test field of the wind tunnel.0 V.34 m/s — each individually induced by setting the abovementioned exhaust fan to operate at three different rotational speeds: 1500 rpm (25 Hz). 5.0 V. 0.10 m/s. 6 V}.5 V. 3900 rpm (65 Hz). Table 2) and released it into the tunnel at the source location landmark indicated in Fig.21 m/s. Note that the chemical gas exposures are held at the source mark in the tunnel directly from a nozzle connected to the outlet of a set of MFCs through a flexible pipeline containing the targeted gas at the nominal concentration specified on each pressurized cylinder. and 0. All the chemicals are released at a constant flow of 320 sccm. / Sensors and Actuators B 185 (2013) 462–477 467 Fig. to reach a quasistationary situation between the airflow and remaining ambient conditions within the wind tunnel defining the background of our experiments and to measure the baseline of the sensor time series data before the actual exposure of the chemical analyte got started for exactly 20 s.9 Second. 3900 rpm (65 Hz).

34 m/s. (iii) Releasing of the chemical compound during 3 min. the chemical analyte was removed and the test section was ventilated utilizing outside clean air circulating through the sampling setting at the same wind speed for another minute. Operating heater voltage. After that step. gas exposure.. (ii) During 20 s. 5. And yet. In the end.000 72-dimensional10 time series recordings — each dimension corresponding to each sensor utilized — observed from the 20 replicas collected for each operating parameter and at each of the points covered in the wind tunnel. induce less stable patterns of the air flow direction. whereas faster velocities in the wind generate narrower gas plumes (see Fig. resulting in wider gas plumes. 4. 4 for a visualization of a typical time series response after a complete measurement was performed). measurement locations of the sensors.. landmark location in the wind tunnel. operating temperature. Even if SVM is essentially a binary classifier. The sensor array is placed in a representative position line of the wind tunnel (0. / Sensors and Actuators B 185 (2013) 462–477 Fig. Finally. Inhibitory SVM Support Vector Machine (SVM) is a machine learning technique that builds a model based on a training dataset to classify the elements of another dataset of interest. Multivariate response of the MOX gas sensor array when methane is released in the wind tunnel.e. The sensors’ responses are affected by the air turbulence present in the wind tunnel causing fluctuations in the acquired signals. Each measurement in the dataset is considered to finish in point (iv) since the sensors’ signals are not recorded during the execution of point (v). In the setup phase of the data collection we found in all trials that the elected features evaluated at symmetry coordinates were strictly different. The chemical source is located at the most-left side of the wind tunnel in front of the board # 5 in the position line # 1 of the wind tunnel.98 m from the gas source. the acquisition of the entire measurement experiment cycle took approximately 260 s to complete. 5) — there is no symmetry in the spatial distribution of the plume with respect to the main axis (i. 10 The total number of measurements is distributed as follows: 3 different wind speeds × 5 different sensors’ temperatures × 10 gases × 6 locations in the wind tunnel × 20 replicas.e. as we will demonstrate below. which in principle should disperse as time passes by. induced by the exhaust fast at 5500 rpm (91. traditionally. recorded at a sampling update rate of 100 Hz (see Fig. 3. this non-symmetry plume structure is reproducible so that the predictions can be extended in time by referring to earlier observations in the same environment. . The resulting dataset ultimately comprises 18. and the variability of the flow direction.66 Hz)) for two more minutes (data not recorded) to fully reestablish the sensor response baseline. it has been extended to multi-class problems by building a classifier for each pair of classes (one versus one strategy) or a classifier for each class with respect to the rest of the samples (one versus all strategy). VH = 6 V. This measurement procedure was reproduced exactly for each gas category exposure. and cleaning recovery phase. in the axis of the chemical plume). A plume demonstrating a perfect symmetry in real environmental conditions is rare due to the existent non-symmetry of the volume enclosing the field.468 A. Normalized sensor response to ammonia when the chemical detection platform is placed in different locations of the wind tunnel. 0. or average of concentration levels observed by the sensors deployed in the wind tunnel. strongly influencing the structure and spatial distribution of the generated chemical gas plumes. (v) Purging of the wind tunnel at maximum speed for two additional minutes to clean the test field. The artificial airflows move downstream to the end of the tunnel (left to right in the figure) at different velocities proportional to the operating rotational speeds of the motor-driven exhaust fan. called validation dataset. The panels illustrate the spatial distribution of the chemical analyte gas plume flowing across the tunnel at slow (top panel) and fast (bottom panel) wind speeds. the line connecting the chemical analyte source to the exhaust).e. measurement of the baseline of the sensors’ signal while no chemical compound is released. i. Vergara et al. the inhomogeneous temperature in the ambient. reported in Table 2 for this particular work represents only the concentration at the outlet of the gas source rather than all the magnitudes. and airflow velocity in a random order and up until all pairs were covered. the baseline recordings. and it took approximately 16 months to be compiled.. (iv) Circulation of clean air for one minute to acquire the sensors’ recovery signals. Before and after each experiment. the wind tunnel was operated at maximum speed (i. The experimental protocol carried out to acquire the signals of the sensors under different conditions consists of the following steps. The variations in the air conditions change the gas plume and the system predictions may be affected. note that although different induced wind speeds strongly influence the structure and spatial distribution of the generated gas plumes — in the sense that slow fan speeds Fig. and the wind speed of the fans. (i) Setting of the operating temperature.

. [34]. 98.81) 66. The comparison between Fig. Therefore.48 0. 99. for each of the classes j = 1.07 (9. (VH = 5 V). and that defines the bandwidth of the Gaussian kernel function.17. where an increase in plume size results in a progressive reduction of the measured gas concentration. fj . 0.10 0. 0.34 0. 98.0} and C {10. 0.34 m/s) while the voltage heater (VH ) is set to 5 V.01. 99.31 81.83 88. . Sensor in a turbulent plume The complete chemo-sensory platform consists of nine sensor arrays. 99. and the standard kernel trick [51] is utilized to improve the accuracy of the classifier for non-linear datasets using a Gaussian kernel. 97.34 Mean (Min. 99. Notice that the measured Ammonia plume shows similar properties to plume measured by Justus et al.88 (40.92) 62. 4. ISVM essentially compares the classifier for class j to the average of the output of all the classifiers.21 m/s. whereas traditional multi-class SVM methodologies perform pairwise comparisons.21) 65.21 0.10 0.21 0.58 (46. each of them composed of eight sensors.45 (45.34 m/s).21 0.22 (61.99) 71.1. 96.10 98.86 93. 99. L. The result is a classifier more robust to the selection of the metaparameters when the models are built with small number of training examples.1 m/s. However.32) 80. Training position 1 Wind speed (m/s) 0. 99.34 0.05 (50.82) 66.19. there are still two metaparameters to be selected for a better accuracy of the classifier: the cost parameter C that controls the contribution of the misclassified samples during the training.88.63 (34.68.98) 80.24) 66.10 0. we used a software written in C/C++that provides the error in the cross-validation to estimate the accuracy of the classifier.89. the chemical detection platform work in a turbulent sampling setting.29. 5000} by minimizing the error in the crossvalidation.05. low wind speed.11 (48. wherein the sensors are exposed to different airflow conditions that change the distribution of the gas of interest in the sampled area. 99. for a total of 72 sensors contained in the complete platform. . slack variables [50] are also included in the function to be optimized to increase the flexibility of the model.80 (64.21 0.10 0.90) 43.21 m/s.99) 2 4. 5 illustrates the maximum of the normalized response (to the maximum signal to make the data lie between 0 and +1) to 3 4 5 6 .95 (33. Vergara et al.89 65.21 97.73..46.1. Max) 36. 99. 1000.01 73.89 1 92.84) 36.24.48 (15.1 70. 0. in the ISVM methodology and likewise in the SVM approach.03 96. 0.34.1 m/s. 100) 68. Results Ammonia of the individual sensors when located at different positions of the wind tunnel and air circulates at low or high speeds.60 (35. Training in row # 6.60) 62. Accuracy of the e-nose system trained in one position and tested in the rest of the considered lines of the wind tunnel. / Sensors and Actuators B 185 (2013) 462–477 469 Fig.A.05 (46.21 0. [32]. 5 (top vs.15 0. ISVM trains one classifier. VH = 6 V. C values 10 100 1000 5000 Values 0. VH = 5 V.05.32) 62.15 97. 100.13 98. bottom panels) qualitatively shows that the different air circulation in the 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 the rest of the considered lines of the wind tunnel.74 (39. 1. The chemo-sensory platform is positioned at six different lines located normal to the wind direction and covering the entire wind tunnel.06 79. 98.11 (36.24 89. The ISVM methodology is robust for a broad selection of metaparameters.34 0. 98.99) 69. 4 shows a graphical illustration of the multivariate response of the metal oxide based chemical sensor module to methane. Table 3 Cross-validation for different selection of the metaparameters C and .05 69.75) 66. The system is exposed to three wind speeds (0. The system is exposed to three wind speeds (0.59.30 (9. 0. To optimize the metaparameters selection.21 0.94) 74. 99. We searched the optimal values of the metaparameters in the 4 × 4 grid defined by {0.34. Then.37. 99.49.18 95.10 0. 6.25 (51. Fig.34 0. making.12. 99.14 (35. Fig.61.34 0.75) 69.10 0. More details on the ISVM methodology along with the utilized C/C++software can be found in Ref.51. in this work we are using a variation of SVM — called Inhibitory SVM (ISVM) [32] — to provide a more robust multi-class classifier inspired by the inhibition process present in animal neural systems.01 However.

as shown in Table 3. we built a model — searched the optimal C and — for the complete platform in one line.0} and C {10. The models are trained and validated at the same sensors’ temperature and wind speed.05. 4. 0. We selected the optimal values of C and by minimizing the crossvalidation in the 4 × 4 grid defined by {0. 1. In order to find the best location to train the complete platform while having good accuracy in the system predictions across the entire wind tunnel. and evaluated the accuracy in the predictions of the built model using the data acquired in the rest of the locations. Fig. Models trained in position lines # 1 and # 2 show poor performance when used to predict the gas class in other positions. the estimated accuracy of the classifier remains high for a broad selection of C and . 1000. 5000}. we trained the complete platform in each of the lines and validated the model in the rest of the considered locations. 8.01. Correlation between the prediction of models trained in one location and evaluated in the remaining position lines. Becuase. we can conclude that the MOX gas sensors are influenced differently by the distinct wind conditions and that the prediction models may be affected when tested in changing environmental conditions. the ISVM methodology shows high robustness in the selection of the metaparameters. 7.1. and thereby the robustness of the models trained in # 1 is limited when the system is tested in other locations. 0. while models trained in line # 4 can be extended to other locations. Therefore. We repeated the same methodology to quantify the degradation of the system when the active layer of . analytes in the following fashion: (i) the gas plume becomes wider for low wind speed conditions and (ii) the chemical sensors tend to present weaker responses when exposed to higher wind speeds due to the drop in the concentration of the gas of interest caused by the increase of the carrier gas flow. for both low wind speed (left) and high wind speed (right). Average accuracy of the models trained in one position landmark and validated in the rest of the positions. The model trained in line # 1 is significantly different from the rest of the models. respectively. / Sensors and Actuators B 185 (2013) 462–477 Fig. Single position training The collected dataset — composed of 10 different classes — was divided into a training and a validation subsets to build an optimal computational model for the e-nose system and to test the accuracy in the predictions of the complete system.2.470 A. Thus. 100. Vergara et al.

one single location for training the e-nose might be sufficient if we know that the operating testing location of the sensor array is located in the vicinity of the training site. Vergara et al. In particular. A1–A4 and Tables A1–A4 show the same tendencies when the operating temperature of the sensor is different.21 m/s.83 (54. Max) 87. For all the different models. 7 shows the mean accuracy in the models’ predictions when they are trained in one location and validated in other positions. 100) 82.34 m/s. Accuracy of the e-nose system when trained in all the lines. 99.1500. 9 shows the accuracy of the corresponding models when the sensors are operated at 6 V and the wind speed was set to 0. Multiple position training As we uncovered in the previous section.33 (41. 99. the reader is referred to the web version of the article. Table 5 shows the minimum. or 3–4). The robustness of the e-nose system is increased by training the system in more locations (no significant drop in the minimum of the system accuracy). the accuracy of the model drops when it is tested in a position not included in the training set.10 m/s Training positions All P1 and P6 P3 and P4 P2 and P5 P3 All P1 and P6 P3 and P4 P4 P2 and P5 All P1 and P6 P3 and P4 P4 P3 and P5 Mean (Min. (For interpretation of the references to color in this figure legend.91 (62.51 (56.94. maximum. the models trained in the line just in front of the chemical source have low prediction accuracy when they are extended to other locations. we built one additional model trained in all the considered lines of the wind tunnel. the user can select the number of position lines Fig. In particular.) .31. Parameters VH = 4 V. 6 and Table 4 show the degradation in the predictions of the e-nose system — for the three considered wind speeds and when 5 V are constantly applied to the sensor heaters — when it is validated in a different location than it was previously trained.1 m/s + 0. yet.75 (47.31.06) VH = 4 V. Each new model obtained from the new training conditions was tested in all the locations of the wind tunnel. 99. 100) 75. 98.34 m/s Fig.21 m/s. or in two lines (1–6. From Fig.38) 67. On the other hand.21 m/s the sensors was at different temperatures and when the system was exposed to different wind conditions.21 m/s. This is due to the particular geometry of the wind tunnel and the gas circulation inside of it that make the plume too narrow in line # 1 to be recognized in the other locations of the tunnel.26) 96.64 (85. Fig.92 (48.15 (79. dashed lines). However. medium wind speed (blue) and low wind speed (red).03) 70. we can conclude that the robustness of the system is consistently improved when increasing the number of locations for training of the model.A. in order to explore the robustness of the system when trained in several locations. Moreover. 98. 4. or 3—4). Figs. Wind = 0. the e-nose system shows high accuracy in its predictions in the position lines where the model was previously trained. A natural manner to address this issue is to train our e-nose system in several distinct locations of the wind tunnel.14 (84. When the system is trained in two different locations.44. 100) 80.84) 92. 99.85) 73.22. Fig. Wind = 0. and three additional models trained in two more locations (1—6.23) 68. the models show qualitatively the same tendency for different sensors’ operating temperatures and wind conditions (cf. whereas the other models show a significant drop when tested in other locations than the ones used for training. As observed. the degradation of the systems trained in the same line is similar when the sensors are operating at different temperatures or are exposed to different wind conditions: the models trained in line # 4 (about the middle of the wind tunnel) show better robustness when are tested in the rest of the wind tunnel. The changes in the gas plume produced by different wind conditions applied in the tunnel significantly alter the performance of the models when tested at different locations than the training position. 2–5. we built new models trained with data acquired in various locations.03) 83. The models are validated at high wind speed (green).06) 73.3. A5 and A6).21 m/s + 0. (65. Figs. even if this strategy has a higher associated cost due to the effort to collect data in such different locations. there is gradual degradation of the performance when testing further away from the training site. the best pair of training positions (indicated in bold) are at the beginning of the tunnel (P1) and at the very end (P6).98. the model trained in all the lines shows high performance in all the positions of the wind tunnel. 100) 78. Hence. Applied voltage in the sensor heater: VH = 6 V | Wind speed = 0.55 (95. 97. Fig. especially at higher sensor operating temperatures and wind speeds.37.91 (52.98 (59. Wind = 0.33 (58. Applied voltage in the sensing layer: VH = 5 V. 91.20.40 (40. and average accuracy of the models when trained and validated in different locations. Thus. 8 we can conclude that the model based on location line # 1 is significantly different from the rest of the models built and thereby its robustness is lower compared to the other models. 2—5. 10. Therefore. 92. We compared the similarity of the different models trained at different locations by measuring the correlation of the respective accuracies in the different locations.23) 75. Accuracy of the e-nose system trained in line # 4 at one wind speed (0.14. VH = 6 V.09) 89. / Sensors and Actuators B 185 (2013) 462–477 471 Table 5 Accuracy of the models when are trained in several positions and validated in all the lines of the wind tunnel. 92. solid lines) or considering the three wind speeds (0. The question now is how we can improve the robustness of the sensor array.

/ Sensors and Actuators B 185 (2013) 462–477 Table 6 Accuracy of the e-nose system trained in line # 4 at one single wind speed or considering the three wind speeds. and the accuracy needed in the area of interest. In particular.89 42. 4. The dataset includes the recordings of the sensors at several locations of the built wind tunnel when exposed to 10 different gases of special interest for a total of 18000 different measurements. Fig.63 92. This outcome contrasts with the solution obtained for one single location transversely to the tunnel. the uncontrolled environmental conditions of open sampling systems by circulating the gas sample at different flows and. This outcome is perhaps not such a surprising conclusion. while the optimal locations for training the system are dependent on the different points of interest and the air circulation in the room. the performance of the system decreases systematically when the tested wind conditions change with respect to the training conditions.69 Min 51. There is still a trade off between the effort/cost carried out to calibrate the system and the accuracy of the e-nose.85 100. Finally. The optimal training conditions depend on the environmental conditions of the area of interest and their variations. the performance is maximal when the medium wind speed is used for training. 10 we can conclude that the accuracy of the system remains high even when the sensors are exposed to any of the wind speeds used in the training. further work includes a more systematic study to select the sensors’ best operating temperature (or temperatures) and maximize the discrimination accuracy of open sampling settings. position line # 1) and at the most-right side of the wind tunnel (i.07 85. we made publicly available a database containing the multivariate responses of a MOX-sensor-based chemical sensing platform operating in a turbulent wind tunnel. If we are constrained to a one single position calibration.34 m/s Mean 54.1 m/s).85 Max 99. We also showed that the system performance is affected by the wind speed used during training conditions. This position is the most informative position..21 m/s Mean 72. Therefore. hence. 10 shows the accuracy of the system and its drop in the performance when the wind conditions are changed. Training Validation 0. if the system can only be trained at one wind speed. due to time and cost restrictions. the controlled operating conditions of MOX sensors by changing the temperature of the sensing layer.44 43.44 52. the system can often only be trained under a subset of all these possible conditions.e. The model trained at different speeds show higher robustness.03 to include in the system calibration according to the resources. Vergara et al. We believe the generated database will be of a high value for the Artificial Olfaction community for the comparison of different algorithms when exploring new techniques to overcome the variation of the environmental conditions. but we had to verify that multiple speeds improve overall performance of the e-nose or sensor array.4.472 A.35 Max 99.20 80. We found that the low wind speed generates a significantly different plume compared to medium and high wind speeds and the predictions of a system trained under low wind speed conditions are impaired when extended to higher wind speeds. on the other hand. Discussion In this paper we presented a methodology based on ISVM to evaluate the accuracy in the gas discrimination predictions of an e-nose system.34 m/s) wind speeds.e. Multiple wind speed training The operating temperature of commercial MOX gas sensors can be easily controlled by applying the selected voltage to the built-in heater.00 0.46 63. 5. However. on the one hand. when the model is built using the three different wind speeds. one may want to train the system in all the expected system conditions. the accuracy of the e-nose system degrades as the testing position moves further from the training position line # 4. we evaluated the performance of the model trained at medium wind speed (0.43 61.29 66.88 61. the accuracy of the system remains high.95 65. we showed that the accuracy in the e-nose predictions is optimal in the middle transverse line of the wind tunnel.51 99. Table 6 summarizes a comparison between the accuracy of the models trained at one specific wind speed and validated at different wind speeds with the performance of the model trained at three wind speeds. . the accuracy of the system is best when it is trained in the two locations featuring the highest variability (first and last lines). in order to increase the robustness of the system against air flow variations.25 69.12 76.75 80.31 m/s All 74.22 81.39].41 51. accompanying the results presented in this manuscript. Although the training position is more significant than the sensors’ operating temperature when evaluating the ability of the e-nose system to predict the nature of new samples. The combination of the worst location with a decent performer (note that it is not the best) leads to the best overall performer in combination.15 100.10 m/s 0.68 99. Although information-theoretic approaches have been presented to chose the optimal operating temperature of MOX sensors [38.98 97. We trained a system at the three considered wind speeds to quantify the improvement in the robustness of the system.80 78. The accuracy of the model significantly decreases when the wind speed in the test conditions is not included in the training conditions.21 m/s) or high (0. However. However. we also found that the e-nose system shows higher performance when the sensors operate at higher temperatures. position line #6). We tested the degradation of the system when the air flow is different from the training conditions.32 94. From Fig. However.73 Min 54.21 m/s) and in the position line # 4 when exposed to low (0. However. we showed that our methodology can quantify the degradation of the predictions when the system is working under new conditions and guide the decision of choosing the number and position of the chemical sensing devices.10 m/s Mean 0. The complete database is unique of this kind since it reproduces.00 0. We showed that the most robust strategy is to train in all the locations of interest and under all the possible environmental conditions.66 70. Tables A1–A4 show that line #1 leads to the worst possible performance. As observed. VH = 5 V.05 64. in open sampling systems.75 99. If the system can be trained in two different positions. The models are validated at the three wind speeds. medium (0.85 Min 42.21 m/s 0. which is in contrast with the results shown in the previous section concerning training for one single line location. the availability of the time to collect data. in all the considered testing positions. In Table 5 we show that the best average performance can be achieved by training the sensor arrays at the most-left side of the wind tunnel (i.78 Max 74. and again.. the robustness of the system is improved.48 82. the wind conditions are hardly controlled and any variation of the air flow may affect the predictions of the system.

0.34 m/s). Fig.34 m/s). the best calibration locations are the ones that induce the highest variability of the sensor responses. should be always trained in the most representative location that. Fig. VH = 4 V. The views. If the experimental time is limited. The authors also thank Joanna Zytkowicz for proofreading and revising the manuscript.1 m/s. 0. . Vergara et al. VH = 4. policy or decision unless so designated by other documentation.21 m/s. the recommendation is to use several locations for This work has been supported by U. and by the US Army Medical Research and Materiel Command under contract number W81XWH-10-C-0040 in collaboration with Elintrix. Our results indicate that the sensor array.1 m/s. which is always calibrated forming a line nearly perpendicular to the wind direction.5 V. Appendix A. In terms of sensor’s operating temperature. 0. higher sensor temperatures are always the best.S. lower speeds should be avoided. opinions and/or findings contained in this report are those of the author(s) and should not be construed as an official Department of the Army position. Acknowledgments We presented a methodology that can be considered as a proof of concept to evaluate the performance of sensor arrays working in open sampling settings. Conclusions training. by Jet Propulsion Laboratory under the contract number 2012-1455933. happen to be in the middle of the wind tunnel at different wind speeds. though. In contrast. Office of Naval Research (ONR) under the contract number N00014-07-1-0741. if two arrays can be used. A1. Accuracy of the e-nose system trained in one line and tested in the rest of the considered lines of the wind tunnel. 6. for this particular case of study. Accuracy of the e-nose system trained in one line and tested in the rest of the considered lines of the wind tunnel. A2. If several wind speeds are available.21 m/s.A. 0. 2. / Sensors and Actuators B 185 (2013) 462–477 473 searching for the best position of the sensing platform. respectively). We validated our methodology discriminating 10 different gases at various concentrations. The system is exposed to three wind speeds (0. or monitoring the air conditions in open environments. The system is exposed to three wind speeds (0. which are at the entrance and outlet extreme points of the wind tunnel (most-left and most-right sides of the wind tunnel in Fig.

Accuracy of the e-nose system trained in one line and tested in the rest of the considered lines of the wind tunnel. Fig. A3. 2–5.21 m/s. VH = 6 V. 0.34 m/s). or in two lines (1–6. A6. / Sensors and Actuators B 185 (2013) 462–477 Fig. Accuracy of the e-nose system when trained in all the lines. 0. The system is exposed to three wind speeds (0. or 3–4).474 A. Applied voltage in the sensor heater: VH = 4 V | Wind speed = 0.34 m/s. Fig. Fig. VH = 5. . Accuracy of the e-nose system trained in one line and tested in the rest of the considered lines of the wind tunnel.21 m/s. 0. A5.34 m/s). Accuracy of the e-nose system when trained in all the lines.1 m/s.10 m/s.1 m/s. A4. Applied voltage in the sensor heater: VH = 4 V | Wind speed = 0. 0. The system is exposed to three wind speeds (0. or in two lines (1–6.5 V. or 3–4). 2–5. Vergara et al.

Hammond.3390/s110101105.doi.22.16. 0.21 0.66 (36. 98.34 0. Udina. 99. IEEE 10 (9) (2010) 1514–1518.14 (42.79) 71.24 (34.27.20) 67.21 0. Baietto.73 (38. Marco. 99.66.74 (12. 99.98) 34.10 0. 99. [8] J.34 0. Sensors and Actuators B: Chemical 163 (1) (2012) 8–19. Marco. doi:10.12. B. G. Caudy.86 (45.21 m/s. Stinson.54 (32.34 m/s).84) 74. Stead. [6] J.43) 78. G. D.85 (50. A novel chemical detector using cermet sensors and pattern recognition methods for toxic industrial chemicals.77 (7.96) 72.75) 79.02 (41.14 (9. 99. S. http://dx.20.21 0. 0. 97. 99.A. The system is exposed to three wind speeds (0.16.C.98) 33.21 0.21 0.35. 98. 99. Breath analysis of lung cancer patients using an electronic nose detection system.99 (51.34 0.80) 68. 99.56. Kennedy. J. L.21 0. 99.88) 76.14) 63. 99. D.34 Mean (Min.92) 61.16 (44. B. 98. 99.95) 75.34 0. 94.99) 67. Medical applications of electronic nose technology: review of current status.37.07) 63.97) 36. 99.10 0.snb.54.31 (41.1 m/s. Max) 33.90) 62. 99. 0.89 (41.83) 63. WAllenstein.21 0.73.66 (50.10 0.40 (65. 99. Sensors and Actuators B: Chemical 116 (1–2) (2006) 135–144. Advances in electronic-nose technologies developed for biomedical applications. Development of optimization procedures for application-specific chemical sensing.66.88.66) 43.34 0.68. M. VH = 4 V.10 0. 99.1016/j. J. Hanson.02 (46.73 (13. The system is exposed to three wind speeds (0.85) 77. 10.85. S.20. Gary.46. M. 98. Thurston. [5] A.J.95.39) 72.98. 99. Wilson. L. J. 99. P.35 (61. Semancik.21 0. 98.50 (40.25) (51. 99. 98.H.015 http://www. 95. 99.94 (42.34 0. Llobet. S.34 Mean (Min.53 (47.12) 77. http://www. Rogers.5 V. Sensors and Actuators B: Chemical 18 (1994) 210–211.60) 61. 99.97 (43.14) 61. Sensors And Actuators B: Chemical 108 (1–2) (2005) 41–55.L. S. 99.47 (42.21 0. [2] J.34) 49.24.79 (11.17.33) 63. sciencedirect.89) 71. Zampolli.34 0.32. 88.22.34 (25.21 0.065.40 (40. Marquis. 99. July 11–14. Ramírez.99) 475 6 2 3 Table A4 Accuracy of the e-nose system trained in one line and tested in the rest of the considered lines of the wind tunnel.85.37.11 (46.10 0. P.46. K. Max) 40.24 (34.1016/ http://www. Max) 66. M.10 0. Tsukuba.39 (46. 99.92.51. 99. ISOEN 2005. 98.34 Mean ( [11] M.33) 70. Japan.10 0. 99.85.34 0. 99.34.71 (43. 96.62. / Sensors and Actuators B 185 (2013) 462–477 Table A1 Accuracy of the e-nose system trained in one line and tested in the rest of the considered lines of the wind tunnel.79 ( 10.21 0.99) 66.22. Walker. 98.sciencedirect.2009.34 0. An RFID reader with onboard sensing capability for monitoring fruit quality. Sensors 11 (1) (2011) 1105–1176.1 m/s.34 0.99) 4 5 6 2 References [1] K.10 0. Nature 299 (1982) 352–355. D. 99.81 (65.99) 71.96 (25. http://dx. A brief history of electronic noses.90. Fonseca.36) 62.95. [3] E.A.93) 70.50 (40. E. American Journal of Rhinology 15 (5) (2001) 291–295. 98. C.1 m/s.H.31) 68.85 (36.10 0.21 m/s. 98. 99.06 (40.34.21 0. P.86 (46. 99. http://dx.29) 60. Chemical Senses 31 (8) (2006). K. Bell.90) 60.96 (41.54) 64.5 V.21 0.29 (8..92) 68. Max) 38.93) 73. 99.49.90) 46. Sensors and Actuators B: Chemical 136 (2) (2009) 546–554.02 (53.11. Max) 35.97.30 (35.78.99) 72.78 (11. 99. The system is exposed to three wind speeds (0. 99.34 0.07) 69. C. Analysis of discrimination mechanisms in the mammalian olfactory system using a model nose. 99.34 0. Lewis.80 (14. S. Thomas.98) 42.97) 35.11 (53.32 (34.10 0. 97.10 0. 0.snb. VH = 5. Bartlett.99) 68.97) Table A3 (Continued) Training position 5 Wind speed (m/s) 0.12 (44.10 0. Persaud.22.94) 73. Santander.26 (45.72 (24.doi.49.10 0. Training position 1 Wind speed (m/s) 0. 99. 99.015 S0925400508008307 [9] A.76. Vergara et al. http://dx.13) 67.93.21 0.87 (35.91 ( Sensors Journal.09 (49. Rose-Pehrsson.10 0.D. J.11. Dodd.44) 73. 99.26 (36.71 (46. Moreno. 0.20. Bryning.39) 65. 99.90. Training position 1 Wind speed (m/s) 3 4 5 6 Table A3 Accuracy of the e-nose system trained in one line and tested in the rest of the considered lines of the wind tunnel. Detection of potato pathogens using an SPME-enose.10 0. Fonollosa. 98.05. 99.21 0. VH = 6 V. G. Spain.07 (24.10 0. Vergara.34 m/s). 100) 70.89. P.2005. Becker.59.N.doi. Thaler. [10] P. 99.99) 78.09 (49. D. [7] A. 96. Ethylene optical spectrometer for apple ripening monitoring in controlled atmosphere store-houses.61 (33.10 0.29) 68.34 0. 94.32 (38.54. 96. Jackson.2008.10 0.C. G.05. [4] V.34 0.12. D.74 (35. Gardner.74) 65.10 0. T. Chan.21 0.34 0.72) 66. 99.59. Tillett.22) 65.83) 63. Scorzoni. A. 98. VH = 4.80 (40.82) 68.21 0. Training position 1 Wind speed (m/s) 0. 10th International Meeting on Chemical Sensors. 99.17.99) 67.10 0. J.10.37 (42.19) 65.98.94) 68.65 (8. K. Ivanov.W.02) 65. 13–15 April 2005.80.01 (34. 98.21 0. Selected Papers from the 11th International Symposium on Olfaction and Electronic Noses – Barcelona.10 0. Detection of chemical warfare agents using nanostructured metal oxide sensors. Training position 1 Wind speed (m/s) 0.30 (31.2011. 97. Persaud.76. 99.83.10 0.51) 58.71 (14.21 0. Sensors and Actuators B 127 (2007) 143–149.37.21 0.21 0.10 0. Tomchenko.1016/j.74) 2 3 4 .84) 65. Tran.93. Wöllenstein. 99.08 (37.2038356. 99.34 Mean (Min.94) 71.79 (41. Johnson.34 0.34 Mean (Min. 99. Hildenbrand. Ziegler.21 0. S. Fonseca.1109/JSEN.34 0.L.94) 68.34 0. Halford.15 (52. Harmer.32. J.34) 64. j.doi.21 m/s.19.80 (46.32.34 0. 0. 99.89) 80.34 m/s).57. Yates. 99.99) 4 5 2 6 3 Table A2 Accuracy of the e-nose system trained in one line and tested in the rest of the considered lines of the wind tunnel.21 0.sciencedirect.snb.05.P.63 (33.59. H.

M. 100. [46] Gill Windsonic.C. Nature 445 (7126) (2007) 406–409.. Kientz. [42] Swagelok Company. E. [51] A. Trincavelli. Blanco.. Farrell. in: Proceedings of the 11th International Conference on Advanced Robotics. Proceedings of the Ninth International Meeting on Chemical Sensors. Chaudry.sciencedirect. Inc. Jonas Mahiques was a visiting student at the BioCircuits Institute. http://www. In autumn 2010 he spent another period as a guest researcher at the BioCircuits Institute at the University of California. H. [23] G.mksinst. Francia. J. His areas of interest also include signal processing.K. http://www. Sensors Journal. M. Sweden. Chromatography and mass spectrometry of chemical warfare agents.cdc. Cardé.org/10. 827275..T. 2009.2012. Information-theoretic optimization of chemical sensors. Signal and data processing for machine olfaction and chemical sensing: a review. Huerta. ensors and Actuators B: Chemical 166 (2012) 320–329. IEEE 4 (3) (2004) 337–347. pp.1109/IROS. Chaudry. Tian. Kadri. M. 2010 — Örebro University) received both his BSc degree (2003) and his MSc degree (2006) in computer engineering from the Politecnico di Milano. Hugues. Nakamoto. S. M. K. Amigó. Shraiman. Sensors and Actuators B: Chemical 137 (2) (2009) 507–512. A.00019/abstract [27] D. Gonzalez. Rochel. Ishida. Moriizumi. Massera.026 http://www. E. T. Sweden. Murlis. M.1109/JSEN. Vaughan. Abdelghani. Odor-source localization in the clean room by an autonomous mobile sensing system. [21] J. Russia) is an associate research scientist at the BioCircuits Institute. Vito. using principal component analysis and multiscale – PCA. Milan. Nakamoto. http://s. F. Vaˇ ızquez.5509617.1021/cr068117e. Villermaux. Linnell.2012.1021/cr068117e. Nie. Vergara. 2009 — University of Barcelona) is a Postdoctoral Researcher at the BioCircuits Institute.doi. Environmental Fluid Mechanics 2 (2002) 115–142. Martinez. A. Rabinovich. [22] D. J.. Autonomous Robots 20 (2006) 185–195. Hines.airgas. Collins.doi. Zhang. [33] J. E. Biographies Alexander Vergara (Ph. Selverston.2/ADA458370 [50] L. Buttner. Marco. Environmental Fluid Mechanics 2 (2002) 143–169. [31] A.sciencedirect. [44] Airgas. E. On time series features and kernels for machine olfaction. org/doi/abs/10.. R. University of California San Diego. Marco Trincavelli (Ph. ‘infotaxis’ as a strategy for searching without gradients. A.1016/j. Lund. Sensors and Actuators B: Chemical 93 (1–3) (2003) 7–16.1016/S0925-4005(03) article/pii/0925400596019077 [29] H. Manatt. Persaud.T. Inhibition in multiclass classification.1021/ac060638e. Methods in Molecular Biology 609 (2010) 223–239. [19] A. S. His work mainly focuses on the use of dynamic methods for the optimization of micro gas-sensory systems and on the building of autonomous vehicles that can localize odor sources through a process resembling the biological olfactory processing. science/article/pii/S0925400512007010 [14] M. Nehorai.1109/JSEN. chemical sensor arrays. T.09. IROS 2009.. R. R. Bhushan.. 2006 — Universitat Rovira i Virgili) is a NRC research associate jointly working at the National Institute of Standards and Technology (NIST) and the National Institutes of Health (NIH) and a visiting research scholar at the BioCircuits Institute. Marco.A. UC San Diego and a research scientist at Information Systems Labs Gardea-Torresdey.html [49] Chemical and biological sensors standards [20] M.060 http://www. A statistical approach to gas distribution modelling with mobile robots — the kernel dm + v algorithm. Murlis.doi. P. R.365304. A. [26] [45] B. S. O. http://dx. http://dx. J Erik Jonsson CTR. B. R. R. G. 1–7. C.acs. V. Huerta.. http://dx.doi. G.G. Frontiers in Neuroengineering 4 (19) (2011). Muezzinoglu. Vembu. 2010. Mowrey. A. J. Neural Computation 24 (9) (2012) 2473– [48] Niosh pocket guide to chemical hazards. FernA-ndez.R. Italy. Vergara. and reconstruction of faulty chemical gas sensors under drift conditions. He additionally received a MSc degree in electrical engineering and computer science from the Lund Tekniska Högskola. In 2007–2010 he has been a graduate student at the Center for Applied Autonomous Sensor Systems. [34] K.D. N. [28] H.doi. N. Electronic nose for space program applications. Measurement of odorplume structure in a wind tunnel using a photoionization detector and a tracer gas. Sensors and Actuators B: Chemical 129 (2) (2008) 750–757. Sensors and Actuators B: Chemical 146 (2010) 460–465. D. Huerta. M. Knowledge discovery with support vector machines. doi:10. I.12. Huerta. Analytical chemistry 84 (17) (2012) 7502–7510.doi. Jones. Homer. Martinotto. Padilla. Caminal. Filament-based atmospheric dispersion model to achieve short time-scale structure of odor plumes.021 http://www. His areas of expertise include nonlinear . in 2006. / Sensors and Actuators B 185 (2013) 462–477 [36] M. K. [39] J.D. M. Other strong interests include biologically inspired algorithms. where he was a postdoctoral researcher until summer 2011.1023/A:1016283702837.doi.2004. http://pubs. acs.1016/0925-4005(96)01907-7. Li. Ishida. [12] S.5354304.snb. and machine olfaction.2012. Vembu. Dorman. R. Young. [53] A. Janata. Ramírez. Sensor selection and chemo-sensory optimization: toward an adaptable chemo-sensory system.1109/ROBOT. Paldi. Ayhan. Weston. Sensors and Actuators B: Chemical 148 (1) (2010) 298– windsonic. Metal oxide-based gas sensor research: how to? Sensors and Actuators B: Chemical 121 (2007) 18–35.sciencedirect. Perera. pMID: 18220426. science/article/pii/S0925400503003381 [16] A.. Multifrequency interrogation of nanostructured gas sensor arrays: a tool for analyzing response kinetics. A.D. science/article/pii/S0925400512013354 [40] A. Chemical sensing in spatial/temporal domains.sciencedirect. Montoliu. T. Chemical Reviews 108 (2) (2008) 680–704. Marco. 1991 — Nizhny Novgorod University. Reducing power consumption via a discontinuous operation of temperature-modulated micro-hotplate gas sensors: Application to the logistics chain of fruit.figarosensor.D. arXiv:http://pubs.2192920. R. Signal Processing. Workshop on Invertebrate Sensory Information Processing: Implications for Biologically Inspired Autonomous science/article/pii/S0925400507007691 [13] L.campbellsci. Dharmasena. E. toxins and related compounds: state of the art and future prospects. Howard. T. Nikolay Rulkov (Ph. http://www. [35] N. Sensors and Actuators B: Chemical 33 (1–3) (1996) 115–121.D. L. International Conference on IEEE/RSJ. Inc. [37] Figaro USA. 2009.H. Vergara. Hamel. [30] Massachusetts. http://dx. Algorithmic mitigation of sensor failure: is sensor replacement really necessary? Sensors and Actuators B Chemical 183 (2013) 211–221. Koziej. B. Marco. Ziyatdinov. Journal of Chromatography A 814 (1–2) (1998) 1–23.2007. J. IEEE Transactions on 43 (1) (1995) 243–253. Huerta. Vergara. 2009. in: IJCNN. 317–323. J.doi. Örebro. and infrared sensing Huerta. U. Weimar. Y. W. M. Kisor. Eiceman. Abarbanel. Li. Vembu. Sensors and Actuators B: Chemical 174 (0) (2012) 114–125. Ryan. A. Sensors [38] A.1016/j. J. in: 2010 IEEE International Conference on Robotics and Automation (ICRA). A. Vergara. http://dx. Sensors Actuators B 174 (2012) 535–546. L. Buehler. Analytical Chemistry 78 (12) (2006) 3985–3996. Gerkey. Huerta. April 15–17. Chemical gas sensor drift compensation using classifier ensembles. [32] R. doi:10. Gutiérrez-Gà A.I. http://dx. UC San Diego. A biomimetic robot for tracking specific odors in turbulent plumes.D. The player/stage project: tools for multirobot and distributed sensor systems. Lee. S. J. Zhou. Sensors Actuators B 129 (2007) 311–318. Fault detection.doi. R. Piga. Inc. H. [17] T. http://dx.2010. Fonollosa. Kagawa. K. Trincavelli. Sensors and Actuators B: Chemical (0) (2012). Gutierrez-Osuna.sensirion. [54] S.doi. http://dx. Nakamoto.snb.1016/j. H. Q.doi. Persaud. and information theory applied to chemical sensing. E. frontiersin. Ryan. http://dx. Monitoring space shuttle air quality using the jet propulsion laboratory electronic nose. Wood Hole. Ye. S. During autumn 2009 he visited the Tokyo University of Agriculture and Technology as a guest researcher. T. Ramesham.07.1021/cr068117e [18] Temperature optimization of metal oxide sensor arrays using mutual information.2009. Drift compensation of gas sensor array data by common principal component analysis. Ben-Hur. identification. Ramón Huerta. [43] MKS Instruments.00019 http://www. Plume-tracking robots: a new application of chemical sensors. doi:10. Classification of multiple indoor air contaminants by an electronic nose and a hybrid support vector machine. Porat. http://www. http://dx. http://www. http://www. Mozalev. San Diego. Eurosensors IX. Calavia. Natl Acad Sci. A. pp. C. Hierlemann. Biological Bulletin 200 (2) (2001) 222– 10. Mehmet K Muezzinoglu. [15] R. J. IEEE. pp. [52] A. C. Inc. R.-L. His research interests include machine learning and artificial olfaction with particular focus on mobile robotics applications. Feature selection for gas identification with a mobile robot. Cardé. Rulkov. Rulkov. Barsan.476 A. pattern recognition. http://www.3389/fneng. support vector machines. M. signal recovery systems.D.doi.1023/A:1016227601019. Jordi Fonollosa (Ph. Acceleration of chemo-sensory information processing using transient features. F. Justus. M.sciencedirect.sciencedirect. 2852–2857. H. T. http://dx. UC San Diego. [24] C.snb.1016/S0021-9673(98)00338-0 article/pii/S0021967398003380 [25] S.doi. Loutfi. Moriizumi. Nowotny.dtic. Vergara et al. Örebro University. A. T.pdf [47] Sensirion AG. E. R. R.. Lilienthal. Vergara. J. feature extraction. in: Intelligent Robots and Systems. A. W. Developments in ion mobility spectrometry–mass spectrometry.1109/78.M. Elkan. A.doi. A.2307/1543320. E. On field calibration of an electronic nose for benzene estimation in an urban pollution monitoring default. H.3389/fneng.K. Perera. X. http://www.2011. Vergara. A user’s guide to support vector machines. Analytical and Bioanalytical Chemistry 372 (1) (2002) 66–73. M. Fonollosa. Vergassola. Ishida. http://handle.2011. Shevade. Llobet. Muezzinoglu. Overton. Kikas. E. A.1007/s00216-001-1195-5. pp.P. ACS Chemical Reviews 108 (2008) 563–613. Vergara. R. Gutierrez-Galvez. IEEE PP (99) (2012) 1. His research is focused on gas sensor array robustness and optimization. 2010. Higher-order chemical sensing. Wiley Online Library. http://dx. doi:10. 570–576. http://dx.A. Detection and localization of vapor-emitting sources. Llobet.A. [41] M. San Diego.L. Gas chromatography.

D. Dr. Dr. UC San Diego. model design and analysis of dynamical systems. Prior his current appointment. His areas of expertise include dynamic systems. . Rulkov has published in over 60 articles in peer-reviewed journals and holds 17 US patents. electronics design. and biology.A.. Ramón Huerta (Ph. biomimetic robotics. data fusion. artificial intelligence. and chemical sensing and machine olfaction applications based on bio-inspired technology. 1994 — Universidad Autónoma de Madrid) is a research scientist at the BioCircuits Institute. model building to understand the information processing in the brain. physics. Vergara et al. His work deals with the development of algorithms for signal processing. and neuroscience. development of custom control and signal processing algorithms. Huerta’s research work gathers in a publication record of over 90 articles in peer-reviewed journals at the intersection of computer science. His work deals with the development algorithms for the discrimination and quantification of complex multidimensional time series. he was 477 an associate professor at the Universidad Autónoma de Madrid (Spain). artificial intelligence. and computational neuroscience. / Sensors and Actuators B 185 (2013) 462–477 dynamics and its applications.