Sensors and Actuators B 44 (1997) 423 428

Multicomponent analysis on polluted waters by means of an

electronic tongue
C. Di Natale a,*, A. Macagnano a, F. Davide a, A. DAmico a, A. Legin b, Y. Vlasov b,
A. Rudnitskaya b, B. Selezenev b
a

Department of Electronic Engineering, Uni6ersity of Rome Tor Vergata, 6ia della Ricerca Scientifica, 00133 Rome, Italy
b
Department of Chemistry, Uni6ersity of St. Petersburg, St. Petersburg, Russia
Accepted 14 May 1997

Abstract
In this paper the simultaneous measurements of the concentrations of a number of chemical species in solutions performed by
a sensor array of ion sensitive electrodes are presented and discussed. By analogy with the well known electronic nose this sensor
array operating in solutions, will be here called electronic tongue. In order to extract optimized information from the electronic
tongue output data, many different techniques have been applied; they were based on chemometrics, non-linear least squares and
neural networks. The best results have been achieved by the introduction of modular models which make use, at the same time,
of both qualitative and quantitative information. 1997 Elsevier Science S.A.
Keywords: Electronic tongue; Multicomponent analysis; Chemometrics

1. Introduction
Environmental monitoring requires on-site simultaneous measurements of a number of different chemical
species. A classic approach to this problem is to consider
a number of selective sensors one for each chemical
species. Along with this in presence of complex solutions,
sensors tend to lose their own specificity and their
response is no more directly related to the concentration
of the species for which they have been designed, and they
become rather influenced by the presence of the other
species. This process can be represented as a sort of
convolution between the sensor selectivity (namely the
sensitivity to the species present in the environment) and
the chemical pattern occurring in the environment under
measurement.
Multicomponent analysis is an analytical procedure
allowing the extraction of qualitative and quantitative
information from an array of non-selective sensors. It is
based on the utilization of an array of sensors matched
with a suitable data analysis procedure.

* Corresponding author. Tel.: + 39 6 7254408; fax: + 39 6

2020519; e-mail: [email protected]
0925-4005/97/$17.00 1997 Elsevier Science S.A. All rights reserved.
PII S 0 9 2 5 - 4 0 0 5 ( 9 7 ) 0 0 1 6 9 - X

Multicomponent analysis provides a sensor array

model from a calibration data-set, which should be large
enough to cover the concentration range of each species,
and to cope the non-linearity in the sensor responses.
Many data analysis techniques can be utilized to
disentangle the information from sensor array outputs;
they can be grouped in four classes: chemometrics,
artificial neural networks, non-linear least squares and
some other exotic methods (such as genetic algorithms
or multi-dimensional splines).
In past years the possibility to perform measurements
of concentrations of a number of chemical species in
complex solution has been attempted [1 3]. By analogy
with the natural olfaction, where sensor arrays operating
in air which are labelled as electronic noses, sensor arrays
operating in liquid media will be here called electronic
tongues.
In this paper the measurement of quantities relevant
for pollution monitoring of internal waters by means of
an array of chemically sensitive electrodes will be presented and discussed. To this scope samples of waters
from Neva river (a river flowing through the town St.
Petersburg) has been artificially polluted with ionic
metals in order to simulate generic industrial waste
pollution.

C. Di Natale et al. / Sensors and Actuators B 44 (1997) 423428

424

Table 1
Ranges of the total concentrations of the eight chemical species in
solution
Species

Concentration range (log[C])

Cu
Cd
Zn
Cr
Fe
Cl
SO4
H

68
6 7
3 5
4 7
47
3 5.3
3 3.9
2.6 8

Different data analysis methods have been utilized in

order to select the most fitting one. Substantial improvement in accuracy has been obtained with modular
models that make use, at the same time, of qualitative
and quantitative information.
It is worth noting that regression methods generally
do not make any use of qualitative information on the
data. This means that an investigation of the data
distribution may reveal, sometimes, the existence of
different qualitative classes. In these cases it is straightforward to suppose that the development of a many
regression model, one for each class, should perform
better relative to a unique regression model holding
over the whole concentration ranges. Data classification
can be achieved in several ways, among the others in
this paper principal component analysis and self organizing maps have been considered.

addition of Cl, SO4 and H was done and that therefore

their concentrations actually occurred, in the Neva
river, at the sampling locations. The range of concentrations, for each species, is shown in Table 1.
Samples have been measured by an electronic tongue
formed by 22 electrodes mainly based on chalcogenide
glasses variously doped and conventional electrodes.
The following chalcogenide glass systems have been
used for the sensor preparation: AgI Sb2S3, Ag2S
As2S3, GeS GeS2 Ag2S, Ge Sb Se Ag with different
components content. Also commercial copper-, leadand cadmium-selective chalcogenide glass sensors were
incorporated. Glasses were prepared in evacuated
quartz ampoules at 1000 K from high purity components. 5 7 mm diameter and 3 5 mm thickness disks
were cut and sealed into plastic tubes. Both liquid and
solid-state inner contacts were used. The details of glass
synthesis and sensor preparation are described in Ref.
[4].
Commercial chalcogenide glass sensors were provided by Analytical Systems (St. Petersburg, Russia).
Their compositions are summarized in Ref. [4]. Conventional crystalline sensor was used basically for chloride ion determination. Two sensors of every type have
been used in the array.
In order to determine a robust regression model for
the sensor array about 150 chemical solutions of different concentrations were prepared and measured.
All sensors, including those with a nominal ion selectivity, have shown a strong cross-selectivity that did not
allow any direct measurement of concentrations.

2. Experimental

3. Data analysis

To simulate real conditions samples of Neva river

waters have been artificially polluted with ionic metals
in order to reproduce generic industrial wastes. Neva
river waters were taken at three different locations. In
each sample the following elements were added in ionic
form: Cu, Cd, Fe, Cr and Zn. The solutions were left at
room temperature for one week to approach, as much
as possible, an equilibrium state. The scope of the work
was to measure the total concentrations of: Cu, Cd, Fe,
Cr, Zn, Cl, SO4 and H. It should be noted that no

In order to utilize the sensor array to retrieve an

estimate of the concentrations of the relevant species in
unknown samples a calibration of the sensor array is
required. For this scope it is necessary to collect a
calibration data-set of measurements performed at
known environmental conditions and then carry out a
data analysis procedure.
As mentioned in the previous paragraph about 150
calibration measurements were performed. The whole
data-set has been split in two parts, one for calibration

Table 2
Mean relative absolute error (RAE) obtained analysing the data with various methods

MLR
PLS
NLLS
BP NN

Cu (%)

Cd (%)

Zn (%)

Cr (%)

Fe (%)

Cl (%)

SO4 (%)

H (%)

40
7
9
7

25
5
4
5

47
7
7
7

208
15
13
15

73
8
6
7

6
2
3
1

11
4
3
1

11
4
6
4

It is worth to note that, apart MLR which is only a reference to evaluate the non-linearities, the other methods show about the same
perfromances.

C. Di Natale et al. / Sensors and Actuators B 44 (1997) 423428

425

f(c1,...,c8) = k0 + % k1, j ci +k2, j c 2j + k3, j c 3j

(1)

j=1

The calibration of the array consists then in the estimation of 25 parameters for each electrode for a total of
550 parameters. The whole array is then represented by
a vectorial function describing the relation between
each sensor output and the concentrations of the eight
species. Since each electrode is represented by Eq. (1)
the array function is written as:

Fig. 1. PCA plot of concentration values in which the occurrence of

two classes of data is observed.

(to determine the model) and one to test the capability

of the model to predict correct concentrations from
unknown samples.
Many different approaches to the data analysis have
been utilized. An extensive comparison of the performances of many techniques have been accomplished in
order to obtain the best possible estimation of the
concentrations. The following methods were utilized:
multiple linear regression (MLR), partial least squares
(PLS), non-linear least squares (NLLS) and back-propagationneural network (BP NN).
MLR approach has been utilized to evaluate how far
from linearity the sensor response was. Electrodes operating in single-component solutions behave according
to the Nernst law, and cross-selectivities are taken into
account for mixed solutions by the Nikolskij extension.
In solutions characterized by the presence of many
different compounds a deviation form the linearity is
expected also if models for the sensor response are not
currently available. So that the errors found by linear
modelling give a figure for the entity of the deviation
from linear behaviour.
PLS is a tool extensively utilized, for quantitative analysis, in chemometrics and in multisensors
applications. Although it is based on a linear approach
it achieves results which are substantially better
than those obtained by MLR. Nevertheless the nonlinearities involved can, sometimes, be so large that
different approaches, non-linear in character, are
required.
From an analytical point of view the non-linear
approach is represented by the non-linear least squares.
The sensor array is modelled as a system of non-linear
equations where the number of equations is equal to
the number of sensors and the variables are the concentrations.
Since no theoretical models are available each electrode has been fitted by a polynomial function of the
third order.

k 1, 1
Fb (c1,...,c8) = k0 +
k 22, 1

1

c
1

k 1 82
c1
*

k 22, 25 2
c
83
c
1

3
c8
(2)

The inversion of this equation allows to estimate, from

the output of the sensors, the concentrations of interest.
These operations (parameters estimation and array
function inversion) can be performed using the LevenbergMarquardt algorithm which provides an iterative
solution for the solution of redundant systems of nonlinear equations. [5]
Non-linear least squares approach (NLLS) requires
much effort in terms of calculus and furthermore the
Levenberg Marquardt algorithm can also have convergence problems which gives rise to a limitation of the
accuracy of the estimations.
In recent years artificial neural networks have gained
popularity to solve non-linear modelling problems.
Their successes are mainly due to the fact that, from the
user point of view, they can be utilized as a regression
machine able to establish correlations between blocks
of data.
Among the neural networks able to perform regression, back-propagation based networks are currently
widely utilized. Basically these networks provide nonlinear models whose parameters are optimized by an
algorithm whose convergence is extremely favoured by
the particular arrangement of the network in feed forward layers [6].
The network here utilized was a 22:15:8 feed forward
network where the neuron transfer function was the
hyperbolic tangent.
MLR, PLS and NLLS were implemented on Mathematica For the ANN the Professional II Software
tool by NeuralWare was utilized. All software ran on
Apple PowerMacintosh 7500. PLS was implemented
following the kernel approach described in Ref [7]. The

426

C. Di Natale et al. / Sensors and Actuators B 44 (1997) 423428

Fig. 2. Radar plot of the concentration ranges in the two classes. The main difference among the classes is represented by the value of pH.

performances of the three methods in predicting unknown concentrations have been evaluated through the
percentage relative absolute error (RAE) in retrieving
the concentration in the test data-set. Results are listed
in Table 2.
Disregarding MLR which, as said before, gives only
information on the entity of non-linearities involved in
the array, the other three methods exhibit similar performances towards all the concentrations. It is remarkable the fact that although the results are comparable
the effort, in terms of calculation and theoretical complications, is rather different. Indeed PLS can be considered as the less expensive technique, it can be
implemented in any mathematical oriented programming language (such as Mathematica or Matlab) and
the calculation time is rather fast also on medium sized
personal computers. NLLS and BP NN, on the other
hand, require longer time of calculus and, due to their
iterative nature, present convergence problems.
A substantial improvement of the data analysis performances can be obtained taking into consideration
the qualitative character of the data. This can be easily
performed by using PCA, a method widely utilized in
chemometrics and then on electronic nose data analysis
to display multidimensional data in a sub-space formed

by the principal components, namely those directions

along which the variance of the data is maximum.
Fig. 1 shows the PCA plot of the concentrations; the
distribution of points reveals that the concentrations
were not homogeneously distributed in the eight-fold
space of concentrations but that they are rather clustered in two classes. The concentration range, inside
each class, is shown as a radar plot in Fig. 2. The main
difference among the two classes of concentrations lies
in the pH value that, in one case, spans from 2.6 to 3.5
and, in the other class, from 6.5 to 8.
Fig. 3 shows the PCA plot of the sensors outputs. It
is clear that, from a qualitative point of view the
sensors are able to discriminate between the two classes.
This circumstance suggests that it should be possible to
obtain a better performing sensor array model including these qualitative information and taking into account one sub-model for each qualitative class.

Fig. 3. PCA plot of the sensor outputs. The electronic tongue

correctly classify the data in two classes.

Fig. 4. SOM lattice plot of the sensor outputs. A net distinction of the
two classes is observed.

C. Di Natale et al. / Sensors and Actuators B 44 (1997) 423428

427

Fig. 5. Architecture of the two modular models. On the left side the chemometric based one is shown. It is formed by a PCA layer as classifier
and on two PLS block one for each classes. Depending on the side of the PCA plot plane on which a data falls one of the two block is activated.
On the right side the neural network based model is represented. In this case the classification is accomplished by a SOM while two BP NN
produce the estimation of the concentrations.

Two modular models have been considered and

compared. The first was based on chemometrics
blocks, composed by one PCA and two PLS blocks,
and the second was based on neural networks, using a
Self Organizing Map (SOM) and two BP NN [8].
In the neural networks model a SOM is utilized as
a classifier. Beside its many features [9,10] SOM can
be considered as a data modelling tool which gives a
bidimensional representation. It has been demonstrated that SOM is a neural implementation of a sort
of principal curve analysis. Fig. 4 shows the SOM
lattice plot of the sensor outputs; as in PCA, also in
this case the two classes are clearly separated.
Fig. 5 shows the architecture of the two modular
models, while their results, expressed by the mean
RAE, are listed in Table 3.
Modular models strongly reduces the error in the
concentration estimates in respect to the conforming
models operating without the qualitative information.
Furthermore the modular model formed by two kinds
of neural networks behaves better than that based on
chemometrics blocks.

4. Conclusions
The electronic tongue, developed in the St. Petersburg Rome collaboration, has been shown to be an
useful tool for simultaneous measurement of several
species in environmental applications, such as the waters of a river. The accuracy of the measurement can
be improved choosing the proper data analysis technique. A comparison among several methods, based
on conceptually different approach, shows that the
performances obtained by PLS, NLLS and BP NN
are basically similar.
A decisive improvement of the predictions is obtained enlarging the amount of considered information taking into account also the qualitative
aspects of the data. To this regard two different
modular models based on chemometrics (PCA+PLS)
and neural networks (SOM+ BP NN) have been
introduced. It has been proved that these models significantly improve the accuracy of the estimations
in respect to models operating without qualitative
information.

Table 3
Mean RAE obtained by modular models

PCA+PLS
SOM+BP NN

Cu (%)

Cd (%)

Zn (%)

Cr (%)

Fe (%)

Cl (%)

SO4 (%)

H (%)

4
2

2
B1

3
1

6
3

4
1

1
B1

2
B1

2
1

A factor two of gain in respect to the value listed in table 2 is obtained for (PCA+PLS) model, while more than a factor four is reached by the
(SOM+GP NN) model.

C. Di Natale et al. / Sensors and Actuators B 44 (1997) 423428

428

References
[1] C. Di Natale, F. Davide, A. DAmico, A. Legin, Y. Vlasov,
Multicomponent analysis of heavy metal cations and inorganic
anions in liquids by a non-selective chalcogenide glass sensor
array, Sensors and Actuators B 34 (1996) 539 542.
[2] M. Otto, J.D.R. Thomas, Model studies on multiple channel
analysis of free magnesium, calcium, sodium and potassium at
physiological concentration levels with ion-selective electrode,
Anal. Chem. 57 (1985) 2647.
[3] M. Bos, A. Bos, W.E. Van der Linden, Processing of signals
from ion selective electrode array by a neural network, Anal.
Chem. 233 (1990) 31.
[4] Yu.G. Vlasov, E.A. Bychkov, Chalcogenide glass ion-selective
electrodes, Ion Selective Electrode Rev. 9 (1987) 5.
[5] C. Di Natale, F. Davide, in: A. DAmico and G. Sberveglieri

[6]

[7]
[8]

[9]
[10]

(Eds.), Proceedings of the 1st European School on Sensors,

World Scientific, Singapore, 1995.
D.E. Rumelhart, J.L. McClelland (Eds.), Parallel Distributed
Processing: Explorations in the Microstructures of Cognition,
vol. 1, Foundations, MIT Press, Cambridge MA, 1986.
F. Lindgren, P. Geladi, S. Wold, A kernel algorithm for PLS, J.
Chemometrics 7 (1993) 45.
C. Di Natale, F. Davide, A. DAmico, W. Gopel, U. Weimar,
Sensor arrays calibration with enhanced neural networks, Sensors and Actuators B 19 (1994) 654657.
T. Kohonen, Self Organizing Map, Springer-Verlag, Berlin,
1995.
F. Davide, C. Di Natale, A. DAmico, Self organizing multisensor systems for odour classification: internal categorization,
adaptation and drift rejection, Sensors and Actuators B 18
(1994) 244 258.