Measurement 28 (2000) 219–224www.elsevier.com/ locate /measurement

Water quality monitoring using a smart sensing systema , a b a*Azedine Charef , Antoine Ghauch , Patrick Baussand , Michel Martin-Bouyer

a ´ ´Laboratoire de Chimie et d’Ingenierie de l’Environnement (L.C.I.E), ESIGEC, Universite de Savoie, 73376 Le Bourget du Lac Cedex,France

b ´ ´Groupement de Recherche sur l’Environnement et la Chimie Atmospherique (GRECA), Universite Joseph Fourier, 38000 Grenoble,France

Received 22 October 1999; received in revised form 2 March 2000; accepted 2 March 2000

Abstract

This paper presents an innovative smart sensor used for water quality monitoring. This smart sensing system utilizesspectroscopic techniques combined with the measurement of physico-chemical variables, to estimate global pollutionparameters in water samples, particularly the Chemical Oxygen Demand (COD). This estimation is computed using amultisensor fusion approach, by means of an artificial neural network algorithm. The smart sensor has been tested

´successfully over a set of 71 wastewater samples of the city Pierre Benite (France) and it was shown that the estimatedvalues of COD were in good agreement with the observed COD values, which were measured with a conventional method. 2000 Elsevier Science Ltd. All rights reserved.

Keywords: Smart sensor; Water quality; Monitoring; Sensor fusion; Spectroscopic techniques; Neural networks

1. Introduction important to make a distinction between smartsensors and integrated sensors, terms which are often

There are many real-time (rapid response) and confused. An integrated sensor is a sensing system inremote sensing applications that require an inexpen- which signal conditioning circuits are an integral partsive, compact, and automated system for identifying of the sensing element itself, the sensor being madean object. The development of computer technology in one piece either by silicon or by thick-filmopens a wide area in the field of modern and deposition on a ceramic structure. Such sensors areintelligent sensing systems [1]. As a matter of fact, not to be considered as ‘smart’, unless it incorporatesthere is no universal definition of what constitutes an a programmable microprocessor [3]. Thus, an inte-intelligent or a smart sensor. Nevertheless, a smart grated sensor may or may not be a smart sensingsensor can be defined as a device containing the system. In the context of this study, the term smart isprimary sensing element, a signal amplification and used to signify that the sensing element is intimatelyfiltering system combined with dedicated software associated with a microprocessor, and specificallyfor data processing and compensation [2]. It is programmed with algorithms derived from Artificial

Intelligence, such as the Backpropagation ArtificialNeural Networks algorithm [4].*Corresponding author. Tel.: 1 33-4-7975-8844; fax: 1 33-4-

Artificial Neural Networks (ANN) are sometimes7975-8843.E-mail address: charef@univ-savoie.fr (A. Charef) used in optical sensing and spectroscopic applica-

0263-2241/00/$ – see front matter 2000 Elsevier Science Ltd. All rights reserved.PI I : S0263-2241( 00 )00015-4

220 A. Charef et al. / Measurement 28 (2000) 219 –224

tions [5]. Indeed, there are many examples of smart sensing system that uses the multisensor fusionsuccessful use of ANN for qualitative and/or quan- approach [12] to evaluate, from changes in bothtitative analysis of chemical species in solution [6,7]. UV–Vis spectra and physico-chemical parameters,In environmental engineering, one of the research the concentration of global organic pollution parame-area deals with the development of optical smart ters such as Chemical Oxygen Demand (COD) insensors that use ultraviolet–visible (UV–Vis) and water and wastewater samples. This smart sensingnear-infrared spectroscopy, for the monitoring of system was developed in the CIE laboratory andchemical pollutants in water [8]. In this way, Ben- consists of an integrated smart multisensor whichjathapanun et al. [9] investigated the use of artificial measures both the spectral and the physico-chemicalneural network algorithms to determine the nature of characteristics of the water samples (Fig. 1). Thespecies in water mixtures, from UV–Vis absorption goal is to fuse the spectral and physicochemicalspectra. In 1994, Keller et al. [10] developed information by means of a backpropagation neuralprototypes of neural networks based sensors for network algorithm, in order to estimate the CODenvironmental monitoring. They proposed a smart concentration of the wastewater samples.system that employs an array of optical sensors toidentify chemical dyes in liquids, by means of neuralnetworks algorithms. They also developed an effec- 2. Experimentaltive neural network based technique that uses acompact and portable gamma-ray spectrometer for To carry out the study, 71 real samples of urban

´monitoring radioactive isotopes in the nuclear indus- wastewater (Pierre Benite city, in France) were used.try. This system was trained with the spectra of eight Data for the study were obtained from the integratedradioactive isotopes and the operation consisted of sensing system described in Fig. 1. For each waste-presenting an unknown sample to the system, water sample, three physico-chemical parametersgenerating a gamma-ray spectrum and passing the (temperature, pH and conductivity) were measuredspectrum to the ANN which produces a classification together with the UV–Vis absorption spectrum of theof the unknown sample. A review of optical based sample. The physico-chemical parameters were mea-

sensors for environmental monitoring was reported sured with two Ingold electrodes whereas theby Grattan [11], who described the technological absorption spectra were obtained by means of aadvances and the challenges of water real-time Jobin-Yvon CP140 imaging system, combined with aanalysis. photodiode array and equipped with a 10-mm path

This present work deals with the development of a quartz cell for optical measurement. This system

Fig. 1. Schematic view of the smart sensing system for water quality monitoring.

A. Charef et al. / Measurement 28 (2000) 219 –224 221

optimized during the training phase, to guaranteegood learning and prediction of the COD pollutionparameter. The choice of the input variables of theneural model was determined using the PrincipalComponent Analysis (PCA) method [13]. This PCAstatistical treatment has led to the building of aninput vector composed of 15 variables: 13 absorb-encies and two physico-chemical parameters, thetemperature and the conductivity (Fig. 3). The goalof this PCA statistical treatment was to choose the

Fig. 2. Measurement of an absorption spectrum of a Pierre Benitemost pertinent input variables among those available.wastewater sample.

allowed us to obtain the whole UV–Vis absorption 3. COD estimation using a sensor fusionspectrum of a sample, from 200 to 800 nm. The approachdimension of the photodiode array is 512 pixels. Inthis work, only the UV part of the absorption The spectral and physico-chemical variables ob-spectrum is analyzed because the organic pollution tained at the previous steps were used by a back-has the property to absorb light in the ultraviolet propagation ANN implemented model, trained with azone of a spectrum, representing the spectral region set of 30 observations (Fig. 4), to evaluate the CODfrom 200 to 350 nm (Fig. 2). concentration in the wastewater samples. The train-

The overall system is coupled with a micropro- ing phase led to the development of a three-layeredcessor in which a backpropagation neural network neural network with three nodes in the hidden layer.algorithm has been implemented for data analysis Table 1 gives the statistical results of this trainingand treatment. The data used to train and test the stage. It can be seen that the mean value of thebackpropagation neural network model are sub- estimated COD is very close to the conventional onedivided into two sets: a first set of 30 observations, (observed one). The good correlation between theto train the neural model and a second set of 41 learned and the conventionally measured COD isobservations to test the model (validation set). The expressed by the high value of the adjustment

2structure of the network (number of layers, number coefficient (R 5 0.95). The estimated and observedof hidden nodes, learning rate, momentum values) is standard deviation (S.D.) are very close to each

Fig. 3. Measurement of physico-chemical parameters for 71 wastewater samples.

222 A. Charef et al. / Measurement 28 (2000) 219 –224

Fig. 4. Learning results for 30 observations. The continuous line deals with the laboratory-measured COD values whereas the dashed linedeals with the neural model learned values.

Table 1Statistical results of the learning stage for 30 observations (learning set)

Variable Mean value Standard deviation Minimum Maximum21 21 21 21 21COD (mg l ) (mg l ) (mg l ) (mg l ) (mg l )

Measured 286.90 81.02 112.00 422.00Estimated 285.38 75.43 134.25 394.61Residual 2 1.52 2 17.632326 48.50534 2 47.124863

2R 0.952

other, which indicates a fine global modeling (Table 4. Discussion and conclusion1).

Once the structure of the neural model was In this study, spectroscopic and physico-chemicaldetermined, it was later tested over a set of 41 data were used to estimate the COD pollutionobservations (Fig. 5). The comparison between the parameter of real wastewater samples. To make thisestimated and the conventionally measured COD estimation possible, the spectral and physico-chemi-shows that the estimated values are in good agree- cal data were fused by means of a backpropagationment with the observed ones. Indeed, the adjustment neural network in order to predict the COD values.

2correlation coefficient (R ) between the estimated The use of this multisensor fusion approach for theand the measured COD values is quite satisfactory. estimation of the COD water pollution parameterThe accuracy of the estimated values is confirmed by was motivated by the need to take into accountthe absence of large exceptional errors. The statisti- nonlinear effects such as chemical interactions, non-cal results of this sensor fusion estimation approach linear absorbency effects in water samples, andare shown in Table 2. It can be seen that both mean possible instrumental artifacts. The results obtainedand standard deviation values of the estimated COD with this method are interesting because they showare close to the observed ones, and the minimum and the ability of this smart sensing system to estimatemaximum are well matched by this method. correctly the COD values. It also shows the accuracy

A. Charef et al. / Measurement 28 (2000) 219 –224 223

Fig. 5. Validation results for the 41 observations. The continuous line deals with the laboratory-measured COD values whereas the dashedline deals with the estimated values.

Table 2Statistical results of the prediction for the 41 observations (validation set)

Variable Mean value Standard deviation Minimum Maximum21 21 21 21 21COD (mg l ) (mg l ) (mg l ) (mg l ) (mg l )

Measured 259.27 84.10 115.00 427.00Estimated 263.73 81.81 137.37 418.89Residual 4.46 2 28.30 70.89 2 54.62

2R 0.88

´ ´and the simplicity of the method which is innovative and the ‘Centre Europeen en Genie de l’Environne-and attractive for a wide area of on-line environmen- ment’ (CERGE, France) for its financial support.tal measurements and control applications. Indeed,the use of such a system, to control a wastewatertreatment plant is now conceivable [14]. This smart Referencessensing has been patented and research concerningthe reduced cost in the basic device itself, the [1] J.E. Brignell, Measurement and control feature on intelligentimprovement of the technological and software instruments, Measurement and Control 29 (1996) 164.

[2] A.H. Tanner, N.M. White, Virtual instrumentation: a solutionperformances, and the miniaturization, are in pro-to the problem of design complexity in intelligent instru-gress, and will be exposed soon. In a future article,ments, Measurement and Control 29 (1996) 165–171.the ability of this smart sensor to estimate other

[3] J.C. Carson, D.R. Shostak, D.E. Ludwig r, Smart sensorspollution parameters such as suspended solids and integrate detectors and processors, Laser Focus World 32 (7)nitrates will also be discussed. (1996) 113–115.

[4] D.E. Rumelhart, J.L. McClelland, Foundation: Forminginternal representation by error propagation, Parallel Distrib-uted Processing, Vol. 1, MIT Press, Boston, 1992, Chapter 8.Acknowledgements

[5] Y. Liu, B.R. Upadhyaya, M. Naghedolfeizi, Chemometricdata analysis using artificial neural networks, Applied Spec-

The authors wish to thank C. Dedieu, C. Agnel, C. troscopy 47 (1) (1993) 12–13.Fachinger and J. Suptil for their technical support, [6] F.M. Ham, G.M. Cohen, B. Cho, Neural network based real

224 A. Charef et al. / Measurement 28 (2000) 219 –224

time detection of glucose using a non-chemical optical based sensors for environmental monitoring, in: Proc. IEEEsensor approach, in: Proc. 12th Annual Conf. IEEE En- Electro’94 Conf., Boston, Miller Freeman, 1994, pp. 378–gineering in Medicine and Biology Society, Vol. 2, LRW 382.Associates, Arnold, MD, PA, 1990, pp. 480–482. [11] K.T.V. Grattan, Water quality monitoring—Technological

[7] M. Beltz, W.J.O. Boyle, K.F. Klein, K.T.V. Grattan, Smart advantages and potential developments in the use of fibresensor approach for a fibre optic-based residual chlorine optic technology, Water Sci. Tech. 37 (12) (1998) 247–253.monitor, Sensors and Actuators B39 (1–3) (1997) 380–385. [12] J. Ghosh, R. Lyon, Multisensor fusion using neural net-

[8] K.T.V. Grattan, Principles of optical fibre sensing for water works, in: Proc. Symp. on Parallel and Distributed Process-industry applications, Measurement 20 (2) (1997) 109–119. ing, Dallas, 1990, pp. 812–815.

[9] N. Benjathapanun, W.J.O. Boyle, K.T.V. Grattan, Binary [13] I.T. Jollife, Principal component analysis, Series in Statistics,encoded 2nd-differential spectrometry using UV–Vis spectral Springer Verlag, 1986.data and neural networks in the estimation of species type [14] A. Charef, A. Ghauch, M. Martin-Bouyer, An adaptive andand concentration, IEE Proc. Sci. Meas. Technol. 144 (2) predictive control algorithm for an activated sludge process,(1997). Bioprocess Engineering (1999), in press.

[10] P.E. Keller, R.T. Kouzes, L.J. Kangas, Three neural network