Available online at www.sciencedirect.com Sensors and Actuators B 131 (2008) 100–109 Exploratory data analysis for industrial safety application M. Vezzoli ∗ , A. Ponzoni, M. Pardo, M. Falasconi, G. Faglia, G. Sberveglieri CNR-INFM Sensor Laboratory, Department of Chemistry and Physics, University of Brescia, Via Valotti 9, I-25123 Brescia, Italy Available online 28 December 2007 Abstract We tested the detection properties of four MOX sensors toward different ozone mixtures to identify sets of sensing layers and interfering compounds concentrations most suitable for a reliable detection of ozone. The measurement campaign lasted 1 year divided in four sessions. We collected a substantial amount of measurements (more than 500) with diverse interfering gases: ammonia, ethanol, ethylene, carbon monoxide and humidity. Due to the dimension of the data set it could not be analyzed using the conventional methods generally applied for characterizing gas sensors: evaluating the sensor performance by visual inspection of the sensors responses is unfeasible. For this reason we systematically applied the exploratory data analysis methodology. We used some simple but effective statistical techniques to insight the data. This approach allows us to draw sound conclusions about the causes of variation in the data, e.g. time (sensors’ long-term stability) or interfering effects of different chemical compounds. All the analysis techniques employed in this work are implemented in a software package developed at our laboratory. We concluded that the two best stable and sensitive sensors are based on WO3 and SnO2 (Au catalyzed). We ranked the contributions of different gases on sensor responses, deducing that out sensors are suitable to detect steps of 50 ppb of ozone when ethylene is less than 10 ppm. Carbon monoxide does not affect the measurements still, the strongest interfering compound is humidity that needs to be controlled or parallely measured also in a preliminary stage. © 2008 Elsevier B.V. All rights reserved. Keywords: Exploratory data analysis; Industrial application; Ozone detection; Sensors array 1. Introduction Ozone detection is not only an environmental priority but also an industrial requirement to keep workplaces healthy. Catalytic ozonation is widely used in industrial processes and the trend is proceeding upwards. Typical applications include water and wastewater disinfection (wastewater plants, hospitals) and food sterilizing (approved in 2001 in the US) where ozonation systems are used, e.g. in modified air packaging lines and fruit storages. Ambient ozone level in workplaces where ozonation is used has to be monitored. The US Occupational Safety and Health Association (OSHA) has set strict exposure limits to assure workplace safety [1]. OSHA has defined the ozone exposure limits as follows: threshold limit value (TLV) is 100 ppb, shortterm exposure limit is 300 ppb The normal background level of ozone in workplaces is 30 ppb [2]; thus for indoor applications it is important to investigate and characterize sensitive layers ∗ Corresponding author. Tel.: +39 030 3715789; fax: +39 030 2091271. E-mail address: marco.vezzoli@ing.unibs.it (M. Vezzoli). 0925-4005/$ – see front matter © 2008 Elsevier B.V. All rights reserved. doi:10.1016/j.snb.2007.12.047 for ozone concentration higher than 30 ppb. For the industrial application envisaged, the effect of interfering gases can have a relevant importance due to its presence in the environment as will be shown in this paper. Up to now, the only solution for ozone monitoring is given by ozone analyzers that are highly selective and sensitive, but also quite expensive (about D 10,000o). Gas sensors arrays (or electronic nose) offer a cheaper alternative approach. The need to get portable, user friendly, cheap and low power consumption devices for gas detecting drives the market trend. The technological improvements occurred in the last decade in system miniaturization is leading towards small and smart devices containing a reduced number of sensitive gas sensors coupled to pattern recognition software implemented in a microprocessor. Our paper shows that this trend might be followed also for ozone monitoring. Portable devices can be proposed to the final user for a dedicated application with reduced price and more specific sensing capability with respect to electronic noses. During the last years different metal oxide sensors such as tungsten oxide, indium oxide, mixed indium and iron oxide, revealed suitable for ozone sensing ([3–8]). 101 M. Vezzoli et al. / Sensors and Actuators B 131 (2008) 100–109 This work is motivated by the lack of exhaustive studies in monitoring ozone in realistic conditions: long measurements time, diverse interfering gases and different humidity levels. These operative conditions are not generally used in the preliminary sensor testing stage but they are the most indicative about the actual sensor working capabilities. We tested an array composed of four metal oxide gas sensors toward mixtures of ozone and four interfering compounds, carbon monoxide, ethylene, ammonia and ethanol at different concentrations. Also two different humidity levels are monitored to evaluate the influence of that parameter on sensing properties. We carried out four measurement sessions, lasting from 2 to 4 days each, over a 1 year period in order to evaluate sensor stability. Testing different interfering species at different humidity concentrations over a long period of time permits to determine critical parameters for a possible industrial application of such sensors. The measurements’ variability depends on a high number of variables: sensor type, ozone concentration, interfering species and their concentrations, humidity level and time progression. In order to visually understand how the variables affect the sensor response, we applied exploratory data analysis techniques. Exploratory data analysis (EDA) is a fundamental step in the data analysis cycle (the cycle consists of: data acquisition, data preprocessing, exploratory data analysis and classification) [12]. The aims of explorative analysis are manifold: maximize insight into a data set, uncover underlying structure, extract important features and detect outliers. A most valuable outcome of EDA is to check for prior assumptions and determine optimal experimental settings. With EDA we identified the sensing layers and the interfering compounds’ concentrations most suitable for our specific industrial target. We end up with a two-sensors array dedicated to ozone detection. Finally, a quantitative evaluation of the ozone concentration in different mixtures with interfering gases is performed with the multi-linear regression (MLR) method. 2. Experimental and methods Materials, deposition methods and working temperature are reported in Table 1 together with codes that will be used in the following to identify each sensor. The working temperature has been chosen as the best compromise between enhanced sensitivity (improved by decreasing the working temperature) and fast response/recovery times (improved by increasing the working temperature) [6]. Measurements were carried out with the flow-through technique in a temperature-stabilized sealed chamber (volume of 1 L) at 20 ◦ C under controlled humidity, working with a constant flux of 0.2 standard litres per minute (s.l.m.). Gas mixtures were generated by certified dry air bottles with diluted target gases concentrations and a humidity control system. A multiple automatic mass flow controller system pilots the correct mixture composition before injection. Ozone was generated through UV lamp discharge and the concentration was measured at the chamber outlet by a detector based on the wet chemical Brewer–Milford principle. A commercial readout electronic has been used to measure sensor resistance values. We initially tested four interfering gases: ammonia, ethylene, ethanol and carbon monoxide. For three compounds, ammonia, ethanol and carbon monoxide, we observed a very low interfering behaviour with respect to ozone so, after some preliminary measurements, we decided to discard ammonia and ethanol. Carbon monoxide was chosen as representatives of this class of low-interfering gases. Ethylene showed a stronger interfering effect. Thus, we prepared samples mixing dry air and ozone or dry air, ozone and an interfering compound (ethylene or carbon monoxide). No ternary mixtures have been examined. The concentrations of the analytes employed are: ozone (0, 70, 140, 280, 560 ppb), carbon monoxide (0, 5, 10 ppm), ethylene (0, 5, 10, 30, 60 ppm). The measurements are performed at different humidity concentrations: 3 and 20% at 20 ◦ C. Different humidity levels are considered to increase the system complexity and match more realistic industrial environments. The detailed measurement table is reported in Table 2. We tested 581 samples divided in 30 different binary mixtures. We performed at least two repetitions for each gas mixture. Four blocks of measurements were carried out during an eleven months campaign with the same sensor array at two different humidity levels. During the last session we observed the poisoning of two sensors, CoO and InFe. As a consequence the measurements collected with such sensors were eliminated and were not considered for data analysis. We designed the measurement protocol with up and down concentration ramps. The ramps are formed with increasing and decreasing steps of ozone concentrations. At each step (fixed ozone concentration) the concentration of a second component is changed (Fig. 1). The concentration of mixture constituents is kept constant for 30 min in order to stabilize the sensors response. The first measurement of each session measures the baseline (i.e. just air) which is than used to calculate the (Rss − R0 ) feature as we described afterwards. Table 1 Description of the sensors composing the array Code Material Working T (◦ C) Deposition method Reference WHT SnAu InFe CoO WO3 Au catalyzed SnO2 Mixed In and Fe oxides CoO 450 450 350 400 Thermal evaporation from metallic W source RGTO SnO2 layer + sputtered Au RGTO from sputtered In target with Fe insets Reactive sputtering from Co target [9] [10] [6] [11] The materials, the working temperatures, the deposition methods and the references for further synthesis details are reported in the table. 102 M. Vezzoli et al. / Sensors and Actuators B 131 (2008) 100–109 Table 2 The table reports the number of measurements divided for different sessions Session (interfering compound) Ozone (ppb) Interfering compound concentration (ppm) 0 5 Total number of measurements 10 30 60 2 2 3 3 3 7 3 3 3 7 0 35 70 140 280 560 60 16 29 15 46 14 April 2006 (ethylene) 0 70 140 280 12 24 24 24 12 12 12 12 12 12 July 2006 (CO) 0 70 140 280 9 18 19 27 16 17 35 16 17 November 2006 (CO) 0 70 140 280 2 3 4 8 3 3 4 3 3 4 December 2005 (ethylene) Humidity level (%) 3 20 214 73 141 156 – 156 174 99 75 37 – 37 In each session we summarize how many measurements we carried out with respect to different ozone concentrations and the concentration of the interfering compound. No ternary mixtures have been examined. The interfering compound used in each session is declared in the first column. The table also reports the total number of measurements per session and the number at each humidity level (3 and 20%). Fig. 1. (a) Figure shows the responses of two sensors (WHT and SnAu) toward mixtures of ozone and ethylene over time. Mixtures are obtained using concentration ramps as shown in (b) and (c). We reported an ozone ramp at three concentrations: 70, 140, 280 ppb (b). At each ozone step, the ethylene concentration is changed between 0 and 10 ppm (c). Two different curve features were extracted from each sensor response: Rss (steady state resistance) and R = Rss − R0 . The first feature is the mean value of the sensor response between 10 and 20 min after injection. Averaging is done to minimize the fast variation of sensor response, caused by noisy and spurious signals that can invalidate a punctual sensor response sampling. R0 (baseline) value was evaluated for all the sensors only once at the beginning of each measurement session and then applied to the calculation of R for all the measurements in that session. The exploratory data analysis methods used are: feature plot, box plot and principal component analysis (PCA) that are implemented in the EDA software package. The EDA software is a Matlab toolbox developed over the years at our lab. The most common and widely applied descriptive statistics functions (e.g. box plot) are already included in the original version of the Matlab software [13]. The contribution of our lab is the definition of a data structure (both the measurement matrix and the data covariates), the development of utilities for easy data manipulation (e.g. data sub-sampling, data set fusion) and plots customization. For example we introduce some useful plotting feature as the possibility to describe data points with a double labeling. This means that the legend is characterized by two labels: the first label refers to a data category, e.g. ozone concentration, so that data points are coloured with different colours relating to different ozone concentrations. The second label refers to a diverse category, e.g. humidity level, and points are marked with diverse markers relating to diverse humidity levels. This simple customization is useful to retrieve more information out from the same figure reducing the number of needed plots. In particular, the advantage of this graphical mode is that the user is able to observe how data points are spatially located referring M. Vezzoli et al. / Sensors and Actuators B 131 (2008) 100–109 simultaneously to two categories and to recognize particular data structures more easily. 3. Results and discussion We divided the data analysis in three sections: in the first one we assess which sensors in the array better discriminate ozone. In the second section we explain the effect of factors such as long measurement time and interfering gases that are responsible for the data variance, whilst the last section is devoted to the quantitative assessment of ozone concentration using the best sensors. 3.1. Sensor selection Initially, with feature plots and box plots, we visually selected the best sensors. Feature plots depict the value of a feature extracted from a sensor versus time or another data category. For example in Fig. 2 the feature is plotted with respect to the category ozone concentration and category classes are 35, 70, 140, 280 and 560 ppb). The subset of measurements depicted are (1) the measurement session of December 2005 (the other sessions give similar results), (2) all the sensors composing the array, (3) at fixed humidity (RH = 20%), and (4) without inter- 103 fering gases. Each subplot refers to a sensor while points are grouped with respect to different ozone concentrations on the x axis. WHT and SnAu are the two best sensors. For the WHT sensor, measurements at different ozone concentrations do not overlap (except for a few measurements at 35 and 70 ppb). For the other three sensors an overlapping is present. Yet, for SnAu the within class spread is noticeably minor that for the other two. A further indicator of sensor performance is stability over different sessions. In Fig. 3 we expand the values of Fig. 2 with the second, third and fourth sessions, depicting feature distributions with their mean values and standard deviations. We also change the label which now reflects the session name. This plot confirms that sensor WHT is performing best and the sensor SnAu is the second best. Such deduction is also supported by the box plot of the sensor WHT and CoO (Fig. 4). The box plot summarizes different properties of a data distribution: (1) the box has lines at the lower or first quartile (bottom blue line), median or second quartile (red line in the middle) and upper or third quartile (top blue line) values; (2) whiskers are lines extending from each end of the box showing the extent of the tails of the sample distribution. Whiskers extend from the box out to the most extreme data value within 1.5*IQR, where IQR is the inter-quartile range (i.e. difference between 3rd and 1st quartile values) of the sample; (3) Fig. 2. Feature plot of the R feature. Ozone concentration is on the x axis (from 35 to 560 ppb). The sensors that better discriminate different ozone concentration are the WHT and SnAu. All the measurements reported were collected only in December 2005. The measurements collected at 35 and 560 ppb of ozone were collected only in the first session. 104 M. Vezzoli et al. / Sensors and Actuators B 131 (2008) 100–109 Fig. 3. (Top) Feature plots of the R feature extracted from WHT and SnAu sensors over four measurement sessions: November 2006, July 2006, April 2006, and December 2005. (Bottom) The same feature extracted from CoO and InFe sensors presents a visible drift. Measurements with 35 and 560 ppb ozone have been performed only in the first measurement session. Feature value distributions are summarized with mean values and standard deviations. outliers are data with values beyond the ends of the whiskers and they are marked with a red cross. If there is no data outside the whisker, a dot is placed at the bottom whisker. Box plots convey more synthetically the different performance of WHT and CoO. You should compare these figures with the relative plots in Fig. 3. For the WHT, the boxes are well separated and a neat increasing relation with respect to ozone concentration is observed. The contrary is true for CoO. As you note in Fig. 4 (bottom), there is only a slight negative correlation between the ozone concentration and the sensor response. Remember that the CoO sensor behaviour is opposite to WHT sensor behaviour: WHT is an n-type sensor whilst the CoO is a p-type sensor. Therefore, the negative correlation is not the problem. The problem is clear from Fig. 3: the dependence on ozone is less than the dependence on the session number (in this paper we do not attempt to counterbalance drift). M. Vezzoli et al. / Sensors and Actuators B 131 (2008) 100–109 Fig. 4. Box plots of RSS feature extracted from the sensors WHT (top) and CoO (bottom), one for each ozone concentration (from 0 to 560 ppb). The points marked with a red cross are outliers. 3.2. Explaining the variance The three interfering compounds play an important role in the determination of the ozone concentration. In fact, as we will see in this section, in a mixture of ozone, ethylene, carbon monoxide and water steam at the concentrations of interest in this study, the data variability depends, sorting with descending importance, on: 1. 2. 3. 4. 5. humidity level; sensor drift; ozone concentration (with a resolution of circa 50 ppb); ethylene concentration (if less than 10 ppm); CO concentration (any). The interfering effect of humidity is well known for all metal oxide sensors and represents the major drawback of this type of thin film sensors. The a priori humidity level knowledge is almost necessary when sensors are placed in a workplace in which humidity variations are not negligible. Humidity can be 105 measured with a dedicated sensor or it can be estimated using a model built on the experimental points, also in a preliminary stage. We will not show the effect of humidity in the following. In the following we perform successive principal component analysis plots on the two best sensors: WHT and SnAu. In two dimensions principal component analysis just performs a rotation so that the first axis is the direction of the greatest variance. All the data plotted are scaled with respect to the sensor baseline R0 to minimize the influence of sensor drift, i.e. we consider the feature R. In Fig. 5 we plot all the measurement sessions carried out with humidity level at 20%. Here distinct colours refer to distinct ozone concentrations whilst distinct markers are relating to distinct sessions. From this big dataset we visually recognize three main aspects: (1) clusters are ordered in increasing order with respect to ozone concentration as indicate by the curved arrow; (2) clusters are also spread with respect to different measurement session as indicate by the thick arrow; (3) the right half of the figure is quite confused: there is overlapping between points at different ozone concentration (different colour markers). In order to make sense of the confusion in the bottom right corner, we plot a subset of the data and change one of the two labels. This is easily carried out with the EDA software by changing two lines in the parameter file. In this way we highlight in Fig. 6 the interfering effect of ethylene in the first measurement session. We see that measurements with 280 ppb of ozone (green marker) are divided in three sub-clusters due to different concentrations of ethylene. In particular, the sub-cluster at 280 ppb of ozone and 30 ppm of ethylene overlaps those relating to 140 ppb of ozone (blue circle marker). Following the green arrow (increasing ethylene concentration) we end up with the sub-cluster (280 ppb of ozone, 60 ppm of ethylene) being almost superposed the zero ozone cluster (red markers). The same tendency is true for measurements with 35 and 70 ppb of ozone. The ethylene–ozone mixture is known to be strongly unstable. Measurements carried out with an ozone analyzer revealed that the ozone concentration decreases of about 2–5% once ethylene is added. However sensor response is strongly decreased more than the above 2–5% when ethylene is added to ozone. It is reasonable to ascribe this effect to the catalytic effect of the metal oxide surface, which enhances the ozone–ethylene interaction. It is not thus a critical issue to recognize the ozone concentration inside the test chamber because of the catalytic reaction occurring at the sensor surface. Only this surface phenomenon is important for our purpose. It is apparent that the sensor system is not able to detect ozone in presence of up to 30 ppm of interfering of ethylene. For this reason, we performed measurements with mixtures containing high concentration of ethylene (30 and 60 ppm) just in the first session and we decided to test only mixtures with low concentrations of ethylene (≤10 ppm) in the last three sessions. Extremal concentrations of ozone (35 and 560 ppb) were also not measured in the last sessions: we preferred to test fewer ozone concentrations but with all the possible combinations of concentrations of the interfering gases. Fig. 7 shows the second measurement session at (1) 0, 5 and 10 ppm of ethylene and 0, 70, 140; (2) 280 ppb of ozone. 106 M. Vezzoli et al. / Sensors and Actuators B 131 (2008) 100–109 Fig. 5. Different ozone concentrations (from 0 to 560 ppb) are reported with different colours. The clusters are ordered and quite separated with respect to ozone concentration and session apart those located close to 0 ppb of ozone (dash circle). The anomalous measurements are due to the presence of high ethylene concentrations as shown in the next figure. If the ethylene is present at low concentrations (up to 5 ppm) ozone concentration differences are still bigger than ethylene differences. Ten ppm of ethylene, at least for higher ozone concentrations, still represents a big interference. Until now we did not show the effect the three different levels of CO present in the mixtures. In Fig. 8 you can see that carbon monoxide does not affect the sensors’ response towards ozone. Different CO concentrations are represented with points that fall within the same ozone-dependent cluster (having the same colour). Thus no evidence of interference behaviour is shown by the oxide. Finally if we consider all sessions without the interfering effect of high ethylene concentrations (10, 30 and 60 ppm) and without the 35 and 560 ppb of ozone, we obtained Fig. 9. The measurements sessions are ordered with respect to (1) the ozone concentration and (2) measurement session as indicated with the Fig. 6. This figure is a zoom of the portion of relating to the first measurement session. The first column of each entry in the legend refers to the ozone concentration (ppb) whilst the second one to the ethylene concentration (ppm). The arrows show the effect of increasing ethylene concentrations (0 to 30 to 60 ppm). Fig. 7. This figure is a zoom of the portion of relating to the second measurement session. The first column of each entry in the legend refers to the ozone concentration (ppb) whilst the second one to the ethylene concentration (ppm). The spreading of clusters at different ozone concentrations is still bigger than those of ethylene (up to 5 ppm). M. Vezzoli et al. / Sensors and Actuators B 131 (2008) 100–109 107 arrows. We note that the right half of the plot is not confused anymore as it was in Fig. 5. 3.3. Quantitative ozone determination We initially verified that the response of the selected sensors had linear relation when seen in a bi-logarithmic plot. To this end we limited the observation to the best operative conditions: (1) without ethylene; (2) all the carbon monoxide concentrations tested; (3) fixed humidity level (RH = 20%); and (4) one measurement session. In Fig. 10 we see that the expected relation Fig. 8. Different colours refer to different ozone concentrations, whilst different markers refer to 0, 5 and 10 ppm of CO. Fig. 9. Considering four sessions with ozone (70, 140, 280 ppb) and ethylene (0, 5 ppm) at fixed humidity (20%), measurements are ordered with respect to session number and ozone concentration. Fig. 10. On the x axis is reported the logarithmic value of three ozone concentrations: 70, 140 and 280 ppb. On the y axis is reported the logarithmic value of the WHT sensor response with the error bars. The black line is the linear fitting line. Fig. 11. Multi-linear regression obtained with the sensor WHT and SnAu. On the x axis is reported the real ozone concentration, on the y axis the estimated concentration obtained applying the regression coefficients. (Top) The regression model is built on measurements without ethylene (first session). The four error bars are relating to 35 ppb (σ = 2 ppb), 70 ppb (σ = 11 ppb), 140 ppb (σ = 8 ppb) and 280 ppb (σ = 13 ppb) of ozone. (Bottom) It shows the estimated ozone on measurements with 5 ppm of ethylene (error bars: 8, 23, 55 ppb). 108 M. Vezzoli et al. / Sensors and Actuators B 131 (2008) 100–109 is satisfied for the WHT sensor. The linear fitting line (black line) is a good approximation of the true relation between the ozone concentration and the sensor response. The same result was obtained for the remaining sessions considered separately and also for the second best sensor, SnAu. We performed the multiple linear regression (MLR) and calculated the relative standard deviations for the estimation of three different ozone concentrations. Such result was obtained considering the two best sensors (WHT and SnAu) in the two operative conditions: without ethylene and with low concentration of ethylene. In the first case (see Fig. 11(top)) there is a very good linear relation between the true ozone concentration (reported on x axis) and the estimated ozone concentration (reported on y axis, pink crosses). The standard deviation, that is a measure of the precision we reach in the determination of ozone concentration, is equal to 2, 11, 8 and 13 ppb, respectively, for samples containing 35, 70, 140 and 280 ppb of ozone in the first session. If we consider only the best sensor WHT the performance we reach are slightly worse than those relative to WHT and SnAu sensors together: 5, 17, 10 and 28 ppb for the same group of ozone concentrations. The SnAu sensor increases the capability of a single sensor to estimate the concentration of the target gas. In presence of ethylene (second session), the linear fitting of the data distributions at 70, 140 and 280 ppb of ozone is good but the presence of the interfering ethylene affects the estimation. In particular, it splits the cluster relating to 280 ppb in two subclusters and the standard deviation for the estimation of ozone is worse: 8, 23 and 55 ppb, respectively. Such splitting is observed in Fig. 11(bottom) at 280 ppb of ozone due to the presence of 5 ppm of ethylene. 4. Conclusions We tested four sensors for 1 year collecting more than 500 measurements in atmospheres composed of ozone and different interfering gases in order to get realist information about sensor ozone discrimination capabilities. Due to the amount of measurements, they could not be analyzed using the conventional methods generally applied for characterizing gas sensors. Thus we systematically applied the exploratory data analysis methodology. Through the visual inspection of feature plots and box plots we selected the two more stable and sensitive sensors: WHT (based on WO3 ) and SnAu (Au catalyzed SnO2 ). Principal component analysis allows us to investigate and draw sound conclusions that explain the data variability. Data measured over 1 year period reveal sensing performances suited for industrial safety application. The selected sensors are able to track ozone at concentrations close to both the STEL and TWA limits, with an accuracy lower than 30 ppb, also in presence of interfering gases. Acknowledgments The measurements sessions have been supported by the FP6 European Project “Nano-structured solid-state gas sensors with superior performance” (NANOS4) No. 001528. The EDA software has been supported by the FP6 European Project “Mobile system for non-invasive wound state monitoring” (WoundMonitor) IST-2004-27859. References [1] OSHA – Occupational Safety and Health Standards - Hazardous Materials, List of Highly Hazardous Chemicals, Toxics and Reactives (2000) 1910.119 App. A. [2] A.G. Clarke, in: R.M. Harrison (Ed.), Understand Our Environment, The Royal Society of Chemistry, 1992. [3] K. Seung-Ryeol, Hong Hyung-Ki, Kwon Chul Han, Yun Dong Hyun, Lee Kyuchung, Sung Yung Kwon, Ozone sensing properties of In2O3-based semiconductor thick films, Sens. Actuators B 66 (2000) 59–62. [4] M. Suchea, N. Katsarakis, S. Christoulakis, M. Katsarakis, T. Kitsopoulos, G. Kiriakidis, Metal oxide thin films as sensing layers for ozone detection, Anal. Chim. Acta 573–574 (2006) 9–13. [5] T. Becker, L. Tomasi, Chr. Braunmuhl Bosch-v, G. Muller, G. Sberveglieri, G. Faglia, E. Comini, Ozone detection using low-power-consumption metal-oxide gas sensors, Sens. Actuators 74 (1999) 229–232. [6] C. Baratto, M. Ferroni, G. Faglia, G. Sberveglieri, Iron-doped indium oxide by modified RGTO deposition for ozone sensing, Sens. Actuators B 118 (1–2) (2006) 221–225. [7] Hosoya Yuuki, Itagaki Yoshiteru, Aono Hiromichi, Sadaoka Yoshihiko, Ozone detection in air using SmFeO3 gas sensor, Sens. Actuators B 108 (2005) 198–201. [8] A. Gurlo, N. Barsan, M. Ivanovskaya, U. Weimar, W. Gopel, In2 O3 and MoO3 –In2 O3 thin film semiconductor sensors: interaction with NO2 and O3, Sens. Actuators B 47 (1998) 92–99. [9] A. Ponzoni, E. Comini, M. Ferroni, G. Sberveglieri, Nanostructured WO3 deposited by modified thermal evaporation for gas-sensing applications, Thin Solid Films 490 (1) (2005) 81–85. [10] G. Faglia, E. Comini, M. Pardo, A. Taroni, G. Cardinali, S. Nicoletti, G. Sberveglieri, Micromachined gas sensors for environmental pollutants, Microsyst. Technol. 6 (2) (1999) 54–59. [11] M. Ferroni, E. Comini, G. Faglia, M. Sacerdoti, G. Sberveglieri, Structural and electrical characterization of cobalt oxide p-type gas sensor, in: Proceedings of the 4th Ieee Conference on Sensors, 2005, pp. 1323– 1325. [12] M. Pardo, G. Sberveglieri, Learning from data: a tutorial with emphasis on modern pattern recognition methods, IEEE Sens. J. 2 (3) (2002) 203–217. [13] Statistics Toolbox, User’s Guide, Vers. 5., The Mathworks. Biographies Marco Vezzoli received his degree in material science in 2002 from the University of Milan. In 2007, he obtained a PhD degree in materials engineering from the University of Brescia with a dissertation on exploratory data analysis for gas sensors arrays. His research interests include statistical data analysis and pattern recognition for chemical sensors. Andrea Ponzoni was born in 1976. He received the degree in physics from the University of Parma, Italy, in 2000 and the PhD degree in material engineering from the University of Brescia in 2006 with a thesis on nanostructured metal oxides for gas sensing applications. His major research activity concerns synthesis and electrical characterization of metal oxides for gas sensing applications. Matteo Pardo got a degree in physics (summa cum laude) in 1996 with a thesis in theoretical surface physics at the University of Milano. In March 2000, he obtained the PhD in computer engineering with a dissertation on multivariate data analysis for gas sensor arrays. Since 2002, he is a researcher of the National Institute for Matter Physics (INFM), now part of the Italian National Research Council (CNR). His research interest is data analysis and in particular the applications of machine learning and pattern recognition techniques for the analysis of chemical sensor arrays and, recently, DNA microchips data. He was an invited lecturer at three international conferences and co-director of the M. Vezzoli et al. / Sensors and Actuators B 131 (2008) 100–109 Short Course on Fundamentals of signal and data processing for the 2nd EU Network of Excellence on Artificial Olfactory Sensing. He is the winner of the 2003 Gopel award. Matteo Falasconi received his degree in physics (summa cum laude) in 2000 from the University of Pavia. In 2005, he obtained a PhD degree in materials engineering from the University of Brescia with a dissertation on the development of an electronic nose for the food industry. At present he is member of the research staff of the SENSOR Lab at the University of Brescia. His research interests include chemical sensor devices and statistical data analysis for electronic noses. Guido Faglia has received an MS degree from the Polytechnic of Milan in 1991 with a thesis on gas sensors. In 1992, he has been appointed as a researcher by the Thin Film Lab at the University of Brescia. He is involved in the study of the interactions between gases and semiconductor surfaces and in gas sensors electrical characterization. In 1996, he has received the PhD degree by discussing a thesis on semiconductor gas sensors. In 2000, he has been appointed associate professor in experimental physics at University of Brescia. During his career 109 Guido Faglia has published more than 50 articles on International Journals with referee. Giorgio Sberveglieri received his degree in physics cum laude from the University of Parma (Italy), where in 1971 he started his research activities on the preparation of semiconductor thin film solar cells. In 1994, he was appointed full professor in physics. At present he is director of the CNR– INFM SENSOR Lab established in 1988 at the University of Brescia. SENSOR Lab is devoted to the preparation and characterization of thin film chemical sensors based on nanostrucured metal oxide semiconductors and to development of electronic noses. He has been the General Chairman of the 11th International Meeting on Chemical Sensors in 2006, now he is the Chair of the Steering Committee of the IMCS series Conference. He is evaluator of European Union, in the area of nanoscience and nanomaterials, and the Coordinator of the EU Project NANOS4 (Nano-structured solid-state gas sensors with superior performance) and several Italian projects on gas sensors. During 30 years of scientific activity, he published more than 250 papers in international journals and presented more than 250 oral communications to international congresses (12 plenary talks and 45 invited talks).