T

ERa

b

a

ARR2AA

KTEEN

1

uigoswntt[dbpida[pp

e

0h

Sensors and Actuators A 192 (2013) 1– 8

Contents lists available at SciVerse ScienceDirect

Sensors and Actuators A: Physical

jo u rn al hom epage: www.elsev ier .com/ locate /sna

NT detection using a voltammetric electronic tongue based on neural networks

duardo Garcia Breijoa,∗, Cristian Olguin Pinatti a, Rafael Masot Perisa, Miguel Alcaniz Fillol a,amón Martínez-Máneza,b, Juan Soto Caminoa,b

Institute of Molecular Recognition and Technological Development, Unidad Mixta UPV-UV, Universitat Politècnica de Valéncia, Camino de Vera s/n, 46022, Valencia, SpainCIBER de Bioingeniería, Biomateriales y Nanomedicina (CIBER-BBN), Spain

r t i c l e i n f o

rticle history:eceived 4 July 2012eceived in revised form7 November 2012ccepted 27 November 2012vailable online 21 December 2012

a b s t r a c t

We report here the use of a voltammetric electronic tongue based on simple metallic electrodes forthe detection and discrimination of different concentrations of 2,4,6-trinitrotoluene (TNT) in acetoni-trile:water 1:1 (v/v) mixtures. The tongue consisted of noble working electrodes made of iridium,rhodium, platinum and gold and non-noble electrodes including silver, copper, cobalt and nickel. Both theself-organizing map (SOM) and multi-layer feed-forward network (MLFN) neural networks were applied

eywords:NTxplosiveslectronic tongueeural networks

to the data obtained from the electronic tongue and TNT solutions. From SOM analysis it was establishedthat a suitable response in terms of a correct classification of the TNT concentration was observed whenusing only noble metal electrodes and only 5 selected pulses. Similar good classifications were foundwhen using MLFN. Moreover, the algorithm of neural network MLFN was embedded in a microcontrollerin order to obtain a smart portable system for discrimination of TNT. In this case an R2 of 0.993 wasobtained for predicted vs observed graphs of concentrations of TNT concentrations.

. Introduction

Current international public awareness on terrorist attackssing explosives has resulted in the particular interest of develop-

ng quick and reliable methods to detect these chemicals. Moreover,iven the widespread use of explosive formulations, the analysisf explosives is also of interest in landmine detection, foren-ic research, and to study environmental problems associatedith explosive residues. Among the different chemical explosives,itroaromatics are perhaps the most commonly used and fromhem probably 2,4,6-trinitrotoluene TNT is the most representa-ive. TNT is well known not only for its use for military purposes1] but it also has a great industrial interest [2]. With the pro-uction, storage and use of TNT, an environmental problem haseen created because this mutagenic, toxic, and persistent com-ound has been reported to leach from soils and to accumulate

n the food chain [3–6]. Current methodologies employed for theetection of nitroaromatic explosives such as TNT are enzymaticssays [7], gas and liquid chromatography [8], mass spectrometry9], ion-mobility spectroscopy [10], optodes [11], electrochemicalrocedures [12] and fluorescent and colorimetric probes [13]. Some

ortable systems have also been reported [14].

In recent years, electronics tongues have appeared as anxcellent alternative to traditional analysis methods in different

∗ Corresponding author. Tel.: +34 963877608; fax: +34 963877609.E-mail address: egarciab@eln.upv.es (E.G. Breijo).

924-4247/$ – see front matter © 2012 Elsevier B.V. All rights reserved.ttp://dx.doi.org/10.1016/j.sna.2012.11.038

© 2012 Elsevier B.V. All rights reserved.

areas as food, pharmaceutical industries, environment and others[15–17]. These systems combine electrochemical techniques suchas potentiometry, voltammetry or impedance spectroscopy withmultivariate analysis tools (neural networks, principal componentanalysis (PCA), partial least squares (PLS), fuzzy logic, etc.) in orderto classify samples or quantify their physicochemical properties[18]. Their main advantage compared to traditional methods is thatthey allow the implementation of fast and low-cost measurementssystems avoiding pre-processing of the samples and the need ofqualified personnel to carry out the analyses.

Traditional electronic tongue systems have been mainlydesigned using potentiometric and voltammetric techniques[19–21]. In particular voltammetric procedures have recentlybecome popular for the design of electronic tongues using arraysof electrodes suitable for voltammetric experiments with very fineresults [21–25]. Voltammetry covers a group of electro-analyticalmethods in which the information of the analyte is derived fromthe measurement of the current vs the applied potential underconditions that helps the polarization of a working electrode.

In this particular work we report the design of a voltammet-ric electronic tongue based on simple metallic electrodes and thestudy of its use as a suitable system for the detection of TNT inaqueous samples. Moreover we were also interested in testingthe use of neural networks to study the electrochemical response

of the electrodes in the presence of this chemical. In particu-lar, neural networks allows a facile implementation of algorithmsin microprogrammable system such as field-programmable gatearray (FPGA), digital signal processor (DSP) or microcontrollers,

2 s and Actuators A 192 (2013) 1– 8

as

2

2

dvaenvsatt2cwoacrtoaaaatdtf

2

eovTotdbrs

2

oessuaiaiasn

E.G. Breijo et al. / Sensor

nd this is the basis for the further design of easy-to-use portableystems for in situ or at site detection applications [15,25–33].

. Materials and methods

.1. Electronic system

The electronic system used in the electronic tongue waseveloped in the IDM Research Institute at the Polytechnic Uni-ersity of Valencia (Spain). The system consisted of a softwarepplication, which runs on a PC, and electronic equipment. Thelectronic system allows to work with three electrochemical tech-iques: impedance spectroscopy, cyclic voltammetry and pulseoltammetry. For impedance spectroscopy the system generatesinusoidal signals with frequencies in the range 1 Hz to 200 kHznd an amplitude up to 500 mV. For cyclic voltammetry the ampli-ude of the triangular potential waveforms can be configured inhe range of −2 V to +2 V and the scan rate can be established form

mV/s to 10 V/s. Finally for pulse voltammetry up to 50 pulsesan be programmed with amplitude from −2 V to +2 V and a pulseidth from 1 ms to 800 ms. The system carries out measurements

n up to ten multiplexed electrodes for voltammetric techniquesnd on one electrode for impedance spectroscopy. A potentiostatontrols the voltage applied to the electrodes and measures theesulting current. The potentiostat can be configured to work inwo-electrodes mode (counter electrode and working electrode)r in three-electrodes mode (counter electrode, reference electrodend working electrode). The user configures the test in the softwarepplication where the data required to carry out the measurementre prepared and this is sent to the electronic equipment through

USB bus. The electronic equipment applies the voltage signalso the electrodes and measures the generated current signals. Theigitalized values of the signals from the samples are then sento the PC where the software application stores them in a file forurther processing.

.2. Electrodes

The electronic tongue system is designed to work with differ-nt electrodes configuration. Voltammetric studies can be carriedut in up to ten working electrodes. Saturated calomel or sil-er chloride electrodes are normally used as reference electrodes.he configuration of the counter and working electrodes dependsn the applied technique. For voltammetric techniques the elec-rodes used are based on the voltammetric electronic tongue (VET)escribed by Winquist et al. [21]. Two types of electrodes haveeen used in this work; noble working electrodes made of iridium,hodium, platinum and gold and non-noble electrodes includingilver, copper, cobalt and nickel (see Fig. 1).

.3. Implementation of MLFN in a microcontroller

Most of the systems of electronic tongues remain in the lab-ratory version, which requires the presence of a computer and,specially above all, two separate processes, one for taking mea-urements and another for data processing. If it is desired for theseystems to have industrial application however, it is necessary tonify these two phases into a single system. The best method forchieving this goal is the use of microcontrollers in systems which,n addition to the measurement process, are able to perform thenalysis of relevant data using a software program implemented

n the microcontroller memory. Thus portable electronic tonguesre becoming popular as they offer simplicity, reliability and use initu. Pattern recognition algorithms have become a critical compo-ent in the implementation of electronic tongues and noses, and

Fig. 1. Electrodes.

have been used successfully in these applications. For implemen-tation in portable equipment the algorithm must be transferable toa microcontroller which has a limited amount of memory. Thus theperfect pattern recognition algorithms will require high accuracy,to be fast, work in real-time and have low memory requirementsin order to be implemented in a microcontroller.

The embedded system was built around a Microchip PIC24FJ256microcontroller. The PIC24FJ256 is a PIC24/16-bit family microcon-troller and has 16KB of RAM and 256KB of reprogrammable flashmemory. The software was coded in C language for the microcon-troller and consists of two main routines: (i) overall system controland (ii) implementation of the pattern recognition algorithm. Inthe implementation of the pattern recognition algorithm, weights(Wji) and biases (Bj) of trained MLFN were used in order to programthe MLFN into the microcontroller memory. Using the normalizedinput values, weights and biases, the microcontroller calculated theoutput for each hidden node by using a sigmoid transfer function.With these outputs the microcontroller calculated the output byusing a linear transfer function in fitting case and a sigmoid transferfunctions in classification case. This routine was coded in C lan-guage and was converted to HEX code using the cross compiler.The HEX file is downloaded into the flash memory of the micro-controller.

2.4. Experimental

The designed electronic tongue system was applied to thedetection and quantification of 2,4,6-trinitrotoluene (TNT). Thereactives used for samples preparation were a commercial TNTsolution 0.13 M in acetonitrile and potassium nitrate (KNO3).A background electrolyte was prepared with 50% of potassiumnitrate KNO3 0.01 M in distilled water and 50% of acetonitrile.This background electrolyte was used for the preparation of theTNT samples. Different solutions having different concentrationsof TNT were prepared in acetonitrile (3 ml). The final TNT sampleswere prepared by mixing these 3 ml with 22 ml of the backgroundelectrolyte. The final concentrations for the tested TNT solu-tions were 3.9 × 10−7 M, 6.0 × 10−7 M, 1.2 × 10−6 M, 6.0 × 10−6 M,1.2 × 10−5 M, 6.0 × 10−5 M, 1.2 × 10−4 M and 5.9 × 10−4 M. Fifteensamples were prepared for each concentration and three cycles of

pulse voltammetry measurements were carried out on each sam-ple.

Two sets of electrodes were used: one with noble metals (Ir,Rh, Pt and Au) and one with non-noble metals (Ag, Co, Cu and

E.G. Breijo et al. / Sensors and Actuators A 192 (2013) 1– 8 3

0 10 0 200 300 40 0 500 600 70 0 800

e (ms)

-200

-400

-600

-800

-100 0

-1200

Am

plit

ude

(mV)

P1

P2

P3

P4

P5 P7 P11 P1 3 P1 5 P1 7

P6

P8

P9

P10

P12

P14

P16

P18

P20

P19

or Ir, R

NdAaarAo

B5o

2

nTtwascmttw

Tim

Fig. 2. Pulses array f

i). Different pulses arrays were applied to the working electrodesepending on their nature. The applied pulses array for Ir, Rh, Pt,u, Ag and Ni is shown in Fig. 2. The pulse sequence used for Cund Co is similar but the pulses P2, P8 and P12 have −300 mV ofmplitude. The pulse arrays were designed according to a methodecently described by us for voltammetric electronic tongues [29].s example, Fig. 3 shows the response of the Cu electrode for eachf the concentrations.

A saturated calomel electrode was used as a reference electrode.efore each measurement the sample was bubbled with argon for

min. The measurement process was carried out at a temperaturef 25 ◦C.

.5. Data preprocessing

In pulse voltammetry the measured current is related with theature and concentration of the species present in the solution [34].he electronic tongue system collects the samples correspondingo the temporal evolution of the current circulating through theorking electrodes. Neural networks were then used to establish

correlation between the collected data and properties the corre-ponding sample. For each electrode the electronic tongue systemollected 1000 data points of the current signal. Considering that in

ost of the experiments 8 electrodes were used, neural networks

ools have to deal with a large amount of data. In order to reducehe number of data to be processed, a compression algorithmas developed. This algorithm applied a 4th order polynomial

Fig. 3. The response of the Cu electrod

h, Pt, Au, Ag and Ni.

approximation to the current samples for each pulse resulting inthe final measurement of the corresponding area. The effect of theapplied algorithm was not only the compression of the data butalso the filtering of the current signals. Finally for each electrode20 data (areas) were obtained for each experiment (Fig. 4). Theinput data were 20 values of pulse areas (considering each area tobe the only input). Thus, with 8 electrodes and 20 pulses appliedto each electrode, 160 input data were obtained. With 15 measuresfor each of 3 cycles (45 measures) for each of the 8 concentrations,a total of 360 measures (with 160 input data each measure) wereobtained.

In order to validate this data compression algorithm a principalcomponent analysis (PCA) was carried out. A total of 24 sam-ples (3 measures for each concentration) were selected for thisstudy. PCA results are shown in Figs. 5 and 6. For the first PCAplot (Fig. 3) raw data (i.e. the values corresponding to the tem-poral evolution of the current signal) were used while the secondPCA plot (Fig. 6) was generated using the coefficients of the 4thorder polynomial approximation to the current samples for eachpulse.

Both PCA plots present a different distribution of the samples,but their relative position in both graphs is very similar. In bothcases the samples corresponding to the background electrolyte are

clearly separated from the TNT samples. Besides in both plots TNTsamples are organized in the same way: lower TNT concentrationsolutions are located in the upper area while higher concentra-tion solutions are in lower area. We can then conclude that the

e for each of the concentrations.

4 E.G. Breijo et al. / Sensors and Actuators A 192 (2013) 1– 8

e each

du

2

itwabcitnnnbb

Fig. 4. Set of data for Cu electrode (D1–D20 corresponding to the area of th

iscrimination capabilities of this technique are maintain whensing the data compression algorithm.

.6. Neural networks

Neural networks are constituted of simple components operat-ng in parallel simulating a biological nervous system. As in nature,he connections between components largely determine the net-ork function and the neural network can be trained to perform

particular function by adjusting the values of the connectionsetween the elements. Neural networks have been used to performomplex functions in various fields, including pattern recognition,dentification or classification. Data matrixes generated by elec-ronic tongue systems have generally a considerable size and neuraletworks tools are suitable to process them. Two types of neural

etwork, very commonly used, are supervised and unsupervisedeural networks. There are several types of both neural networksut among them SOM (unsupervised) and MLFN (supervised) haveeen used in this work.

Fig. 5. PCA plot fro

one of the pulses obtained from the 4th order polynomial approximation).

MLFN has proved to be a very useful tool for pattern recognitionand classification problems and has been used to investigate prob-lems that cannot be easily solved by traditional methods in severalfields. This network is particularly powerful for pattern classifica-tion and function approximation. An advantage of MLFN is thatonce a network is well trained, it can retain excellent performanceeven if degraded, noisy, or missing data are used [35]. An elemen-tary MLFN with R inputs is shown in Fig. 7. Each input is weightedwith an appropriate W. The sum of the weighted inputs and thebias forms the input to the transfer function f. Neurons can use anydifferentiable transfer function f to generate their output.

The SOM, also known as the Kohonen Map, is an unsupervisedneural model of widespread use in areas such as pattern recog-nition. The SOM is a neural network model that projects a highdimensional input space usually onto a one or two dimensionaloutput space. Because of its typical two-dimensional shape, it is

also easy to visualize. This architecture (Fig. 8) is similar to that ofa competitive network, except no bias is used here. Each neuronin output layer is a cell containing a template against which inputpatterns are matched. All cells are presented with the same input

m raw data.

E.G. Breijo et al. / Sensors and Actuators A 192 (2013) 1– 8 5

Fig. 6. PCA plot from the coefficients of the 4th order polynom

f

p1p2p3

pR

W1,1

W1,R

1

b

n a

Input General Neuron

WhereR = numberofelementsin inputvectorn=inputs to thetransfer func�on f.

pptp

3

3

w

a = f(Wp+b)

Fig. 7. Multilayer neural network architecture.

attern in parallel and compute the distance between their tem-late and the input in parallel. Also all cells compete so that onlyhe node with the closest match between the input and its templateroduces an active output.

. Results and discussion

.1. Classification

The electronic tongue, using both non-noble and noble metals,as applied to a total of 8 concentrations of TNT employing the

Input

Dat

a In

put

Compe�ve/Output Layer

Weights

Fig. 8. Self-organizing map architecture.

ial approximation to the current samples for each pulse.

pulse sequences shown in Fig. 2. Moreover Table 1 shows the rela-tion between the concentrations and the name of the assignedclass, see also Section 2.4 for details. As a result of the applicationof the set of pulses the corresponding pulse-signal diagrams wereobtained. This corresponded to a total of 1000 data points that werereduced using a compression algorithm that allowed calculatingthe corresponding area for each current–time curve related withthe pulses. Using this approach, a total of 20 areas were calculatedfor each electrode and each sample (see Section 2.4 for details).The solutions of TNT were prepared in acetonitrile:water 1:1 (v/v)mixtures containing potassium nitrate 0.01 M as supporting elec-trolyte. Both the SOM and MLFN neural networks were applied tothe data obtained from the electronic tongue.

3.1.1. Study with SOMThe program SOMmine5 from Viscovery Software GmbH was

employed to carry out the study using SOM. The number of nodesused in all the studied cases was 1000. The program calculates aneighbourhood of at least 50 nodes with a linear neighbourhoodweight function. As output the program divides a map into regions,called clusters.

The SOM neural network allows carrying out a relatively sim-ple study of the contribution that the different pulses have on theresponse of a selected electrode and the weight that a certain elec-trode has in the overall response obtained in the electronic tongue.In relation to the first issue it was found that for most electrodes thepulses corresponding to a potential of 0 V (i.e. P1, P3, P5, P7, P11,P15, P17 and P19) gave practically no information. This poor con-tribution of the 0 V pulses means that the response observed forthese pulses does not depend significantly on the previous pulseapplied to the electrode. In fact for most of the electrodes it wasobserved that a worse classification of the TNT concentration wasobtained considering all the pulses that when some selected pulsedwere applied. These studies also allow concluding that the set ofpulses could be reduced. For a more detailed study it was foundthat pulses P1, P2, P4, P6 and P20 have a significant contributionfor all the electrodes. Although some of the remaining pulses alsocontributed to the classification, it was found that similar accurate

classifications were reached when only using pulses P1, P2, P4, P6and P20 or when those are combined with others. As a conclusion itwas obvious from the study that a significant reduction in the trainof pulses can be carried out.

6 E.G. Breijo et al. / Sensors and Actuators A 192 (2013) 1– 8

Table 1Concentration named as classes.

Class 1 2 3 4 5 6 7 8Concentration [M] 3.9 × 10−7 6 × 10−7 1.2 × 10−6 6 × 10−6 1.2 × 10−5 6 × 10−5 1.2 × 10−4 5.9 × 10−4

SOM-Ward-Clusters

1,2E-0066E-007

6E-006

3,9E-007 1,2E-005

0,00012

0,00059 6E-005

1

23

4

5

67 8

SOM-Ward-Clusters

3,9E-007

6E-007

6E-006

1,2E-005

1,2E-006

0,000126E-005

0,00059

12

3

4

5

67 8

SOM

twbprocc

ioP

3

w

Nobles

Fig. 9. Comparison between

Once the set of pulses was selected, a study of the contribu-ion of different electrodes to the response of the electronic tongueas also carried out. In a first study nobles and non-nobles have

een studied separately and Fig. 9 shows the SOM obtained (onlyulses P1, P2, P4, P6 and P20 were used). It can be observed a cor-ect classification in both cases was found and with a distributionf concentration form above to below and from left to right. Theoncentrations 3 (1.2 × 10−6 M) and 7 (1.2 × 10−4 M) are the worstlassified.

From these studies it was established that a suitable responsen terms of a correct classification of the TNT concentration wasbserved when using only noble metal electrodes and the pulses1, P2, P4, P6 and P20.

.1.2. Study with MLFNA study of the classification of the TNT’s samples using an MLFN

as also carried out. The network had 160 nodes in the input layer,

Fig. 10. PlotConfusion and PlotRoc of the

Non nobles

for nobles and non-nobles.

and the exit layer was formed by 8 neurons corresponding to the 8concentrations used. Studies with different numbers of hidden neu-rons were done, obtaining an optimum result with 8 secret neurons.The study was performed using the toolbox MATLAB’s NPRTOOL,the standard network that was used for pattern recognition was atwo-layer feedforward network, with sigmoid transfer functions inboth the hidden layer and the output layer.

A network has been built up using only noble metals and withonly the most significant pulses (i.e. P1, P2, P4, P6 and P20). Usingthese parameters, Table 2 shows the value of mean squared error(MSE) and the percentage of error (%E) of the study. Whereas Fig. 10shows the confusion matrix and the receiver operating character-istic (ROC), where it is observed that all the samples are classified

between 82.2% and 100%. In general the recognition rate was 94.2%.

From the studies with neuronal networks SOM and MFLN it hasbeen deduced the conclusion that it is not necessary to carry out themeasure with all the types of electrodes and pulses. By using only

concentrations with noble metal.

E.G. Breijo et al. / Sensors and Actuators A 192 (2013) 1– 8 7

Table 2MSE and %E del training, validation and testing of all metal nobles samples.

Samples MSE %E

Training 252 5.2E−3 1.6

fimc

3

tbcTfrtfifo

oo9hrd(13

otmcq

pdDl

Ft

Validation 54 4.03E−3 14.8Testing 54 2.87E−3 16.7

ve of the pulses and together with the noble metal a satisfactoryeasure is obtained, and therefore the circuit and the electrode

an/could be simplified in future applications.

.2. Fitting

An MLFN has been used for determining the relation of the elec-rodes signal with concentration of TNT. The used program haseen the NeuralTools from Palisade Corporation, carrying out theomparative for different number of neurons in the hidden layer.he network has 160 nodes in the input layer and the exit layer isormed by 1 neuron. Studies with different number of secret neu-ons have been done and the most ideal case was for 8 neurons inhe hidden layer. The standard network that was used for functiontting was a two-layer feedforward network, with a sigmoid trans-

er function in the hidden layer and a linear transfer function in theutput layer.

The study was done with only noble metals and consideringnly the relevant P1, P2, P4, P6 and P20 pulses. In this case theptimal number of hidden neurons calculated by the program was

and the number of input nodes was 24. Training and validationave been carried out using 70% (252) and 30% (108) of the samplesespectively. Samples for training and validation were chosen ran-omly. For the 252 training samples an R2 of 0.996 was obtainedsee Fig. 11) with an MSE of 1.1362 × 10−5. In the case of the08 validation samples an R2 of 0.974 was found with a MSE of.0069 × 10−5.

With that study, the weight and bias matrix from MLFN wasbtained and a similar MFLN was implemented in a microcon-roller. In order to test the implementation of this MLFN in a

icrocontroller, a new array of samples were measured. In thisase an R2 of 0.993 was obtained (see Fig. 12) which was a valueuite similar to that obtained using NeuralTools in a PC.

Both Figs. 11 and 12 show a quite large spread in the actual vsredicted data especially at low concentration, whereas the pre-

iction is much more accurate when the concentration increases.ata dispersion at low concentration is most likely related with the

imit at which the electronic tongue was able to detect TNT

R² = 0,996

1,0E-07

1,0E-06

1,0E-05

1,0E-04

1,0E-03

1,0E-07 1,0E-06 1,0E-05 1,0E-04 1,0E-03

Pred

icte

d

Observed

ig. 11. Predicted vs observed graph of concentrations using noble metals elec-rodes and pulses P1, P2, P4, P6 and P20 (training set).

Fig. 12. Predicted vs observed graph of concentrations with data gathered by micro-controller.

4. Conclusion

A voltammetric electronic tongue that used simple metallicelectrodes has been designed for the detection and discriminationof different concentrations of TNT in acetonitrile:water 1:1 (v/v)mixtures. The tongue was built with noble (iridium, rhodium, plat-inum and gold) and non-noble electrodes (silver, copper, cobaltand nickel) and its response in the presence of TNT was measured.Both, the SOM and MLFN neural networks were applied to the dataobtained from the electronic tongue in order to select pulses andelectrodes which gave a suitable response and to obtain models forclassification applications. From SOM analysis it was found a correctclassification of the TNT concentration when using only noble metalelectrodes and only 5 selected pulses. Similar good classificationswere found when using MLFN. The algorithm of neural networkMLFN was embedded in a microcontroller in order to obtain a smartportable system for discrimination of TNT. Although we are awarethat some additional studies should be carried out in order to detectTNT at very low concentrations, we believe that our results suggestthat it might be possible to develop simple and easy-to-use portableequipment for the detection of TNT in real samples. Moreover weare currently designing further studies directed to the detectionsand discrimination of other nitrated explosives.

Acknowledgement

We gratefully acknowledge financial support from the SpanishGovernment project MAT2009-14564-C04-02.

References

[1] T.F. Jenkins, J.C. Pennington, T.A. Ranney, T.E. Berry Jr., P.H. Miyares, M.E.Walsh,A.D. Hewitt, N.M. Perron, L.V. Parker, C.A. Hayes, E.G. Wahlgren, Characteri-zation of explosives contamination at military firing ranges, Technical ReportERDC-TR-01-5, 2001.

[2] S.I. Sax, R.J. Lewis, Hawley’s Condensed Chemical Dictionary, 11th ed., VanNostrand, New York, 1987.

[3] A. Hilmi, H.T. Luong, A.L. Nguyen, Determination of explosives in soil andground water by liquid chromatography – amperometric detection, Journalof Chromatography A 844 (1999) 97–110.

[4] P.G. Rieger, H.J. Knackmus, Biodegradation of Nitroaromatic Compounds,Plenum Press, New York, 1995, pp. 1–18.

[5] J.E. Walker, D.L. Kaplan, Biological degradation of explosives and chemicalagents, Biodegradation 3 (1992) 369–385.

[6] W.D. Won, L.H. DiSalvo, J. Ng, Toxicity and mutagenicity of 2,4,-6-trinitrotoluene and its microbial metabolites, Applied and Environmental

Microbiology 31 (1976) 576–580.

[7] R.G. Smith, N. D’Souza, S. Nicklin, A review of biosensors and biologically-inspired systems for explosives detection, Analyst 133 (2008) 571–584.

[8] D.S. Moore, Instrumentation for trace detection of high explosives, Review ofScientific Instruments 75 (2004) 2499–2513.

8 s and A

[

[

[

[

[

[

[[

[

[

[

[

[

[

[

[

[

[

[

[

[

[

[

[

[

[

in 1986 at the same University. He is currently a full pro-fessor in the Department of Chemistry at the UniversidadPolitécnica de Valencia (UPV). His main area of interest isthe development of chemical chemosensors and probes,especially those based on electrochemical processes.

E.G. Breijo et al. / Sensor

[9] K. Hakansson, R.V. Coorey, R.A. Zubarev, V.L. Talrose, P.J. Hakansson, Low-massions observed in plasma desorption mass spectrometry of high explosives, MassSpectrometry 35 (2000) 337–346.

10] E. Wallis, T.M. Griffin, N. Popkie Jr., M.A. Eagan, R.F. McAtee, D. Vrazel, J. McKinly,Instrument response measurements of ion mobility spectrometers in situ:maintaining optimal system performance of fielded systems, Proceedings ofthe SPIE – The International Society for Optical Engineering 5795 (2005) 54–64.

11] M.E. Germain, M.J. Knapp, Optical explosives detection: from color changes tofluorescence turn-on, Chemical Society Reviews 38 (2009) 2543–2555.

12] E.S. Forzani, D.L. Lu, M. Leright, A.D. Aguilar, F. Tsow, R. Iglesias, Q. Zhang, J.Lu, J.H. Li, N.J. Tao, A hybrid electrochemical-colorimetric sensing platform fordetection of explosives, Journal of the American Chemical Society 131 (2009)1390–1391.

13] Y. Salinas, R. Martínez-Mánez, M.D. Marcos, F. Sancenón, A.M. Costero, M. Parra,S. Gil, Optical chemosensors and reagents to detect explosives, Chemical Soci-ety Reviews 41 (2012) 1261–1296.

14] J. de Sanoita, E. Vanhovea, P. Mailleyb, P. Bergonzoa, Electrochemical diamondsensors for TNT detection in water, Electrochimica Acta 54 (2009) 5688–5693.

15] M. Alcaniz, J.-L. Vivancos, R. Masot, J. Ibanez, M. Raga, J. Soto, R. Martínez-Mánez,Design of an electronic system and its application to electronic tongues usingvariable amplitude pulse voltammetry and impedance spectroscopy, Journalof Food Engineering 111 (2012) 122–128.

16] K. Toko, Taste sensor, Sensors and Actuators B 64 (2000) 205–215.17] M.S. Cosio, D. Ballabio, S. Benedetti, C. Gigliotti, Evaluation of different storage

conditions of extra virgin olive oils with an innovative recognition tool builtby means of electronic nose and electronic tongue, Food Chemistry 101 (2007)485–491.

18] Yu.G. Vlasov, A.V. Legin, A.M. Rudnitskaya, A. D’Amico, C. Di Natale, «Electronictongue» – new analytical tool for liquid analysis on the basis of non-specificsensors and methods of pattern recognition, Sensors and Actuators B 65 (2000)235–236.

19] R. Martínez-Mánez, J. Soto, E. Garcia-Breijo, L. Gil, J. Ibánez, E. Llobet, An elec-tronic tongue design for the qualitative analysis of natural waters, Sensors andActuators B 104 (2005) 302–307.

20] L. Lvova, E. Martinelli, E. Mazzoneb, A. Pede, R. Paolesse, C. Di Natale, A. D’Amico.,Electronic tongue based on an array of metallic potentiometric sensors, Talanta70 (2006) 833–839.

21] F. Winquist, R. Bjorklund, C. Krantz-Rülcker, I. Lundströma, K. Östergren, T.Skoglund, An electronic tongue in the dairy industry, Sensors and Actuators B111–112 (2005) 299–304.

22] P. Ivarsson, S. Holmin, N.-E. Höjer, C. Krantz-Rülcker, F. Winquist, Discrimina-tion of tea by means of a voltammetric electronic tongue and different appliedwaveforms, Sensors and Actuators B: Chemical 76 (2001) 449–454.

23] V. Martina, K. Ionescu, L. Pigani, F. Terzi, A. Ulrici, C. Zanardi, R. Seeber, Develop-ment of an electronic tongue based on a PEDOT-modified voltammetric sensor,Analytical and Bioanalytical Chemistry 387 (2007) 2101–2110.

24] L. Gil-Sánchez, J. Soto, R. Martínez-Mánez, E. Garcia-Breijo, J. Ibánez, E. Llobet, Anovel humid electronic nose combined with an electronic tongue for assessingdeterioration of wine, Sensors and Actuators A: Physical 171 (2011) 152–158.

25] L. Gil, J.M. Barat, E. Garcia-Breijo, J. Ibanez, R. Martínez-Mánez, J. Soto, E. Llobet,J. Brezmes, M.-C. Aristoy, F. Toldrá, Fish freshness analysis using metallic poten-tiometric electrodes, Sensors and Actuators B: Chemical 131 (2008) 362–370.

26] R.H. Labrador, R. Masot, M. Alcaniz, D. Baigts, J. Soto, R. Martínez-Manez, E.García-Breijo, L. Gil, J.M. Barat, Prediction of NaCl, nitrate and nitrite contentsin minced meat by using a voltammetric electronic tongue and an impedimetricsensor, Food Chemistry 122 (2010) 864–870.

27] I. Campos, R. Masot, M. Alcaniz, L. Gil, J. Soto, J.L. Vivancos, E. García-Breijo, R.H.Labrador, J.M. Barat, R. Martínez-Manez, Accurate concentration determinationof anions nitrate, nitrite and chloride in minced meat using a voltammetricelectronic tongue, Sensors and Actuators B: Chemical 149 (2010) 71–78.

28] I. Campos, M. Alcaniz, D. Aguado, R. Barat, J. Ferrer, L. Gil, M. Marrakchi, R.Martínez-Manez, J. Soto, J.L. Vivancos, A voltammetric electronic tongue as toolfor water quality monitoring in wastewater treatment plants, Water Research46 (2012) 2605–2614.

29] I. Campos, M. Alcaniz, R. Masot, J. Soto, R. Martínez-Mánez, J.L. Vivancos, L. Gil,A method of pulse array design for voltammetric electronic tongues, Sensorsand Actuators B: Chemical 161 (2012) 556–563.

30] L. Gil-Sanchez, E. Garcia-Breijo, J. Garrigues, M. Alcaniz, I. Escriche, M. Kadar,Classification of honeys of different floral origins by artificial neural networks,IEEE Sensors (2011) 1780–1783.

31] E. Garcia-Breijo, J.K. Atkinson, J. Garrigues, L. Gil, J. Ibanez, M. Glanc, C. Olguin, Anelectronic tongue for monitoring drinking waters using a fuzzy ARTMAP neuralnetwork implemented on a microcontroller, IEEE International Symposium onIndustrial Electronics (ISIE) (2011) 1270–1275.

32] J. Ibánez, E. Garcia-Breijo, N. Laguarda Miró, L. Gil, J. Garrigues, I. Romero Gil,R. Masot, M. Alcaniz, Artificial neural network onto eight bit microcontrollerfor Secchi depth calculation, Sensors and Actuators B: Chemical 156 (2011)132–139.

33] E. Garcia-Breijo, J. Atkinson, L. Gil-Sanchez, R. Masot, J. Ibanez, J. Garrigues, M.Glanc, N. Laguarda-Miro, C. Olguin, A comparison study of pattern recognitionalgorithms implemented on a microcontroller for use in an electronic tongue

for monitoring drinking waters, Sensors and Actuators A: Physical 172 (2011)570–582.

34] A.J. Bard, L.R. Faulkner, Electrochemical Methods: Fundamentals and Applica-tions, John Wiley & Sons, Inc, 2001.

ctuators A 192 (2013) 1– 8

35] H. Yang, P.R. Griffiths, Application of multilayer feed-forward neural networksto automated compound identification in low-resolution open-Path FT-IR spec-trometry, Analytical Chemistry 71 (1999) 751–761.

Biographies

Eduardo Garcia Breijo has obtained his MS degree inelectronic engineering from the University Polytechnicof Valencia, Spain (UPV) in 1997, and received his PhDin 2004 from the University Polytechnic of Valencia. Heis an assistant professor of the Electronics EngineeringDepartment of the UPV. He is a member of the Instituteof Molecular Recognition and Technological Development(IDM) of UPV. His main areas of interest are the develop-ment of multisensors in thick-film technology, design ofelectronic systems and neural networks.

Cristian Olguin Pinatti has received his MS degree inautomation and industrial electronics engineering fromthe University Polytechnic of Valencia, Spain, (UPV) in2010. Nowadays he is developing his Thesis at the Instituteof Molecular Recognition and Technological Development(IDM) of the UPV. His research focuses on the electro-chemical sensors, electronic noses and design of electronicsystems for electrochemical measurements techniques.

Rafael Masot Peris has obtained his MS degree in physicsat the University of Valencia, Spain (UV) in 1991, his MEdegree in Electronic Engineering at the UV, 1996 andreceived his PhD in 2010 from the University Polytech-nic of Valencia (UPV). He is an assistant professor ofthe Electronics Engineering Department of the UPV. Heis a member of Institute of Molecular Recognition andTechnological Development (IDM) of UPV. His researchinterest includes EIS, multivariable analysis and designof electronic systems for electrochemical measurementstechniques.

Miguel Alcaniz Fillol has received his MS degree inphysics from the University of Valencia, Spain (UV) in1991, and received his PhD in 2011 from the UniversityPolytechnic of Valencia (UPV). He is an assistant professorof the Electronics Engineering Department of UPV. He isa member of the Institute of Molecular Recognition andTechnological Development (IDM) of UPV. His researchinterest includes design of electronic systems for electro-chemical measurements techniques, digital systems andmultivariable analysis.

Ramón Martínez-Mánez graduated in chemistry fromthe Universidad de Valencia (UV) in 1986, and receivedhis PhD in 1990 at the same university. After a postdoc-toral period at Cambridge (UK), he joined the Departmentof Chemistry at the UPV. He became a full professor in2002. His main areas of interest lie in the field of chromo-fluorogenic and electrochemical sensors and molecularprobes for anions, cations and neutral chemical species.

Juan Soto Camino graduated in chemistry from the Uni-versidad de Valencia (UV) in 1981 and received his PhD