322 ieee transactions on instrumentation and measurement...

322 IEEE TRANSACTIONS ON INSTRUMENTATION AND MEASUREMENT, VOL. 54, NO. 1, FEBRUARY 2005

Self-Organizing Maps Application in a Remote WaterQuality Monitoring System

Octavian Adrian Postolache, Member, IEEE, Pedro M. B. Silva Girão, Senior Member, IEEE,J. Miguel Dias Pereira, Member, IEEE, and Helena Maria Geirinhas Ramos, Member, IEEE

Abstract—This work was developed in the context of a systemfor remote water quality monitoring based on a wireless local areanetwork (WLAN) and includes a Kohonen self-organizing map(K-SOM) implementation in order to perform sensor data vali-dation and reconstruction and sensor failure and pollution eventdetections. Simulation and experimental results are presented.

Index Terms—Self-organizing maps, telemetry, water qualitymonitor.

I. INTRODUCTION

SEA AND RIVER water quality monitoring is one of theimportant activities in the environment-monitoring domain.

The number of research and development activities in that areais extremely large. The assessment of water river basin condi-tions for drinking water is reported by the Guilikeng investi-gation group [1]. Two networks based on supervision controland data acquisition (SCADA) systems were used to collectdata from automatic monitoring stations located on riverbanks.Meanwhile, different commercially available monitoring sys-tems such as a remote underwater sampling station (RUSS) [2]or YSI [3] are used in different water quality monitoring appli-cations in order to collect multiple water quality parameters (pH,temperature, conductivity, turbidity, heavy metal concentration,etc.) from rivers, lakes, or sea water. The acquired data is usu-ally sent to a central location using mobile phone [global systemfor mobile communication (GSM), global packet radio service(GPRS) [4], or personal handy-phone system (PHS) [5])], satel-lite, or VHF [6] technologies.

Water quality monitoring hardware (sensors, conditioningcircuits, acquisition, and communication) must usually becomplemented with processing blocks (WQ-PB) to performdifferent tasks associated to one-dimensional or multidimen-sional data that flows on the system measuring channels.Important processing tasks of the WQ-PB are data valida-tion, data linearization [7] and data compensation [8], shortand long term prediction of pollution events (duration andconcentration), data fusion and data compression, fault andpollution detection, and data recovery. One of the most suc-cessful solutions for advanced WQ data processing is neuralnetworks. Thus, different applications of neural network on

Manuscript received June 15, 2003; revised May 29, 2004. This work wassupported by Portuguese Science and Technology Foundation PRAXIS XXIProgram FCT/BPD/2203/99 and by Project FCT PNAT/1999/EEI/ 15052.

The authors are with the Instituto de Telecomunicações, Centro de Elec-trotecnia Teórica e Medidas Eléctricas, DEEC—Instituto Superior Técnico,1049-001 Lisboa, Portugal (e-mail: [email protected]).

Digital Object Identifier 10.1109/TIM.2004.834583

water treatment [9] and on river water for drinking purpose [1],are reported in the literature.

This paper deals with the design and implementation of adistributed measuring system for water quality monitoring andemphasizes multilayer perceptron and Kohonen maps neuralnetwork processing structures implemented for advanced dataprocessing. The system is characterized by high accuracy ofwater parameter measurements, data validation, data recon-struction, sensor fault detection, and pollution events signaling.Data flow obtained from the primary acquisition and processingunits is distributed in a wireless local area network (WLAN).The WLAN includes a personal computer (PC) as a centralprocessing unit that manages the distributed measuring systemand analyzes the received data.

II. WQ MONITORING SYSTEM DESCRIPTION

The remote distributed measuring system includes a set ofFieldPoint primary acquisition and processing units (PAPs).Each PAP contains a processing unit with an ethernet controllerinterface (FP-2000) and a four–channel analog input module(FP-TB-10) that performs the water quality parameter acquisi-tion. Communication between the PAPs and the central controland processing unit (CCP) is based on WLAN and GPRS[10] technologies. Thus, the FieldPoint units that include theRS232 and ethernet interfaces can be directly connected to theCCP in order to transmit the numerical values of water qualityparameters from the zones to monitor. In the following twosubsections, two network and communication architectures arepresented: a GPRS polling architecture (GPRS-PA), which wasinitially selected, and a WLAN-GPRS hybrid architecture thatis the upgrade of GPRS-PA finally implemented.

A. GPRS Smart Polling Architecture

The GPRS polling architecture includes a set of Siemens Cel-lular Engines M35 [11] that are class B GPRS mobile stations.The M35’s only port (RS232) is used to connect it to the FP2000of the PAPs (Fig. 1).

Data acquired by the PAPs is sent to the CCP using GPRSM35 terminals and is then processed by the CCP’s personalcomputer.

The CCP interrogates the PAPs from time to time accordingto the derivative values of the measured quantities in order to ob-tain more accurate data for fault and pollution events detection(smart polling). The instantaneous sampling rateof a water quality parameter has a nominal value of fora stationary signal . However, the

0018-9456/$20.00 © 2005 IEEE

POSTOLACHE et al.: SELF-ORGANIZING MAPS APPLICATION IN A REMOTE WATER QUALITY MONITORING SYSTEM 323

Fig. 1. Water quality monitoring architecture based on FieldPoint embeddedsystems and GPRS modems.

Fig. 2. Water quality parameter linear variation.

sampling rate is proportionally incremented by the relation be-tween the absolute value of the water quality parameter varia-tion and its nominal value . The factor , where

is the maximum sampling interval , worksas a speed-up coefficient for the sampling rate when large vari-ations of water quality parameters occur. In fact, this speed-upeffect is limited by the value of that depends on systemhardware and system configuration, namely, the number of PAPunits. In our practical implementation, ms.

Considering a linear approximation for the water quality pa-rameter as the one represented in Fig. 2, the absolute value ofthe straight line slope is equal to

(1)

The normalized slope used to increment the sampling rate indynamic conditions is given by

(2)

Finally, the sampling frequency can be obtained using theequation

with (3)

Equation (3) ensures that the computational and communicationload of the system is very low under normal working conditionsbut increases with the time rate of the measured quantity. How-ever, a maximum sampling interval must be established inorder to detect global communication failures and to assure thata minimum amount of data is gathered for WQ parameters timemapping definition.

B. WLAN-GPRS Hybrid Architecture

In order to reduce costs, to be able to increase the distancebetween the CCP and the monitored area and, at the same time,to increase the communication rate up to 11 Mbps betweenfield units should communications between the PAPs and theCCP fail, a hybrid architecture based on a WLAN and on GPRSmodems was designed and implemented. Data acquisition ratesare slow, but increasing the communication rate capabilityamong PAP units can be very important to detect system faults,to identify geographical and temporal trends associated topollution events, and also to increase system protection againstfaults (robustness) if GPRS communication failures betweenthe access point (AP) and the CCP unit occur. In this case, oneof the PAP units assures the management and the control ofthe measuring system and temporarily performs the functionsassociated with the CCP unit.

The communication hardware is WLAN-based and includesa set of 2.4-GHz ethernet-to-wireless bridges (D-LinkAir DWL-810) that convert the FP2000 ethernet port into an IEEE 802.11bwireless network device. Because wireless outdoor transmissionrange is lower than several hundred meters (e.g., 300 m), addi-tional wireless bridges can be used to extend the range, avail-ability, and functionality of the implemented WQ wireless net-work. For the same reason, and in order to increase the dis-tance between CCP and PAPs (e.g., range m), two GPRSmodems (Siemens M35) are used: one connected to the CCP andthe other to the PPP Gateway. The WLAN-GPRS compatibilityis assured by the D-Link DWL-900AP wireless access point, bythe ipEther232.PPP PPP-Gateway, and by a M35 GPRS modem(Fig. 3).

The PAP information is sent wirelessly to the access point(AP) that communicates via GPRS with the CCP. The APprotocol is compatible with IEEE 802.11b, and its operating fre-quency and modulation are 2.4 GHz and Direct Sequence SpreadSpectrum (DSSS), respectively. The PAP works as a bridgebetween the FieldPoint WLAN and the wired PPP Gateway(ipEther232.PPP). The gateway enables the transmission ofnetwork packets via the serial interface of the GPRS modem.The serial port of the PPP Gateway accommodates baud ratesfrom 2400 to 115 200 baud and is connected to an M35 GPRSmodem that assures the communication with the CCP. TheGPRS modem enables a transparent connection between theCCP and the PAPs that are identified by their IP addresses.


Fig. 3. Upgrade of the architecture of Fig. 1. Water quality monitoring system based on a WLAN-GPRS hybrid architecture.

C. Primary Acquisition and Processing Units

Each PAP incorporates several hardware blocks:

i) sensing block;ii) acquisition block;iii) primary processing and communication control unit;iv) communication block.

The sensing block includes a set of sensors and conditioningcircuits associated to water quality (WQ) parameters measure-ment such as turbidity (TU), pH, temperature (T), and conduc-tivity (C). The WQ values are transmitted as a current ampli-tude (4–20 mA) to PAPs analog inputs. The sensors used areGlobal Water WQ770 for turbidity measurement, ISI OLS50 forconductivity measurement, ISI-11 for pH measurement, and aPt1000 for temperature measurement.

The acquisition block incorporates two FieldPoint dualchannel analog input modules FP-AI-V10B (12 bits, up to2 kS/s) mounted on a FP-TB-10 terminal base and a set of4–20 mA receivers RCV420 that convert the current signalsinto voltages , , and .

The acquired data is delivered via the FieldPoint bus to a pri-mary processing and communication control unit (FP2000) thatperforms embedded measurement, data logging, and communi-cation tasks.

The acquired data is stored in nonvolatile memory for reli-able data logging or is transmitted to the CCP, as described be-fore, using the serial RS232 interface (first architecture) or theethernet communication interface (second architecture) that are

part of the communication block. Other elements of this com-munication block are a GPRS modem, for the first consideredmonitoring system architecture, and a wireless bridge for thesecond one.

D. Central Control and Processing Unit

The central control and processing unit includes a PC (Pen-tium III, 128 Mb RAM) and several communication interfacecomponents. Thus, for the first proposed architecture a GPRSmodem (Siemens M35) assures the direct communication be-tween the PC and the GPRS compatible field units (M35 GPRSmodem included). In the second architecture, the CCP commu-nicates via GPRS with the networked PAPs using a PPP Gateway(ipEther232.PPP) and a M35 GPRS modem that perform theWLAN-GPRS conversion. The data flow, GPRS received by theCCP, is distributed over the Internet, the CCP working like a WQInternetServer. Based on the GPRS connection, theCCP controlsthe WLAN elements (PAPs) in order to acquire new WQ valuesthat permit the actualization of a WQ Net Page. Anomalous func-tioning of the PAP’s channels is detected by the CCP, whichthen issues commands to switch off the anomalous channels.

III. SOFTWARE COMPONENTS OF WQ MONITORING SYSTEM

PAPs and CCP software components were mainly developedusing LabVIEW 6.1 and LabVIEW Real Time [12]. MATLAB6.0 modules for offline MLP-NN design and optimization wereimplemented using the Neural Network toolbox [13] and the


TABLE IMLP-NNS PARAMETERS: n —NUMBER OF INPUT NEURONS;

n -NUMBER OF HIDDEN LAYER NEURONS (ONE HIDDEN LAYER) AND

n —NUMBER OF OUTPUT NEURONS

SOM toolbox [14] used for the K-SOM’s design and onlineoperation. The connection between MATLAB and LabVIEWmodules was realized using MATLAB Script LabVIEW struc-tures [15] that permit the use of the MATLAB functions withinLabVIEW programs.

The tasks performed by the software are data acquisition andprocessing at PAPs’ level (including PAPs’ measuring channelsmodeling [16] and compensation of temperature effects [17]),advanced analysis of WQ data at the CCP level, data communi-cation, and data publishing.

A. PAPs’ Software

The software associated with PAP units includes the acquisi-tion subroutines, based on the FP Read LabVIEW—Real Timefunction, and neural processing blocks [multilayer perceptronneural networks, (MLP-NN)], designed using MATLAB, thatconvert the , , , and acquired voltages intovalues of TU, pH, C, and T, respectively. Thus, two singleinput-single output neural networks (one for andanother for conversion) and two dual input-singleoutput neural networks [one for temperaturecompensated pH and another temperature com-pensated C conversion] were designed and implemented.The networks have one single hidden layer, the training al-gorithm in all four cases is Levenberg Marquardt [18] andsum-square errors (SSE) of 1E-4 are the stop condition. TheMLP-NN’s training and test are performed using a data setof voltage values delivered by the transducers during PAP’scalibration phase. The calibration solutions used were formazin

NTU for the TU trans-ducer, buffer solutions for the pH transducer,and uS/cm for conductivity. Calibra-tions were performed in a laboratory for different temperatures,

C.Taking into account the FP2000 storage memory limitation,

an optimization of the MLP-NNs was carried out in order to ob-tain high conversion accuracy using a small number of neurons.Table I presents not only the results obtained in terms of thenumbers of neurons for each of the three layer MLP-NNs butalso of neurons transfer functions, tan-sigmoid (tansig ( )), forthe hidden neurons, and linear, for the output neurons.

After MLP-NNs design in MATLAB, the calculated neuronweights and biases are then used to implement the online neuralconversion and compensation blocks based on the followingrelation:

(4)

TABLE IIPERCENTAGE ERROR OF WQ PARAMETERS VALUES FOR THE DATA RECEIVED

FROM THREE PAPS AFTER NEURAL NETWORK PROCESSING. IN BOLD, AND

FOR COMPARISON PURPOSES, ARE THE PERCENTAGE ERROR OF THE RAW WQPARAMETER VALUES (WORST CASE ONLY)

where and represent the weights matrices, and andrepresent the biases matrices of the neural network designedin order to obtain the WQ parameter value using the ac-quired normalized voltages , , pH, and included invector . The conversion and compensation blocks are parts ofLabVIEW Real Time program uploaded to the PAP’s FP2000memory.

The global errors associated to WQ measuring channels arelower than 1% (Table II) in Tagus River WQ operational mea-surement ranges TU NTU conductivity

mS/cm temperature C and confirm that MLP-NNsdo increase the accuracy of WQ evaluation, particularly whenthe measuring channels are highly nonlinear (e.g., conductivitymeasurements).

The PAPs also include signaling blocks associated to pollu-tion events and faulty operation detection. For that purpose, acomparison is made between the current TU, pH, C, and T mea-sured values and the historical values (last 2 h) of the measuredvalues and the WQ parameter limits for the monitored area (e.g.,estuary of Tagus river) according to the European EnvironmentAgency (EEA).

Several LabVIEW real time blocks were designed and imple-mented at the FP2000 level for communication purposes. Forthe first architecture, the main software blocks are associatedwith RS232 port configuration and with sending and receivingdata. The Serial Init LabVIEW function was used to set the serialcommunication port (e.g., 57.6 kbaud, one stop bit, non parity,no flow control), while the Serial Write function was used toprogram the GPRS modem using the AT commands accordingto the communication requirements. The Serial Read LabVIEWfunction was used to receive the commands issued by the CCPto control data flow and on/off switching of the measuringchannels. For the second communication architecture, known asDataSocket, a programming technology based on industry-stan-dard TCP/IP is used to perform data read and data write actions.

Taking into account possible network failures, PAP unit soft-ware also includes a watchdog time function that autoreconfig-ures the network if faulty communication between the PAPs andthe CCP are detected. As mentioned before, in this case, the con-nection with the CCPis disabled, and one of the PAPs assumes themanagement and control of the measuring system. After fault re-covery, theCCPstopsallPAPsprograms, savesmanagementdatafrom “master” PAP, and reinitializes activities (Fig. 4).

B. CCP Processing Component

The CCP receives the WQ parameter values based on thePAP’s interrogation, processes the data, and publishes the WQcurrent values on the PAP’s web pages. Data processing is mainly


Fig. 4. WQ network reconfiguration when CCP-PAPs connection fails.

Fig. 5. PAP1 web page including the TU, C, pH, and T information.

related with multichannel data analysis and is based on Kohonenmaps. The main objective of such processing is to reduce all WQmeasured parameters to a format better suited to the extraction ofinformation concerning water quality in the monitored area andthe performance of the overall telemetric system.

Kohonen maps are neural networks widely used in both dataanalysis and vector quantization because they compress the in-formation while preserving the most important topological andmetric relationship of the primary data and also for their ab-straction capabilities. These characteristics of Kohonen mapsare used in the present application for representation and anal-ysis of WQ data received via WLAN—GPRS from the PAPs.Each K-SOM defines a mapping from the input space (TU,C, pH, T, and WQ parameters) on to a regular two-dimensionalarray of nodes. The number of designed K-SOMs is equal to thenumber of PAP units; three, in the present case.

The K-SOMs were trained using the WQ parameters (TU, C,pH, and T) acquired by the PAPs. The used values correspondto the following conditions: “Normal Functioning (NF)” (theWQ parameters values are inside the EEA limits [19]; “FaultEvent (FE)” (the WQ values corresponds to faulty functioningof one or several PAPs measuring channels); and “PollutionEvent (PE)” (the WQ values correspond to either low values ofpH or high water conductivity and turbidity values). Differentnumbers of cells were considered accordingto the size of the training set.

While MLP-NNs used for data conversion and temperaturecompensation were trained using a supervised training strategybased on a gradient descent type algorithm, K-SOMs are ob-tained using an unsupervised learning process, which is an in-cremental-learning SOM algorithm in our application. Thus, theprototype vector of cell of a K-SOM network randomlyinitialized is updated according to the following learning rule:

(5)

where represents an input vector randomly drawn from theinput data set at time , TU pH , andis called the neighborhood kernel around the winner cell anddefined by

(6)

is the learning rate at time , is thedistance between cells and within the output space (map),and corresponds to the width of the neighborhood function[20]. A batch-training algorithm was also applied, and a com-parison in terms of resolution and topology preservation of thedesigned K-SOMs was carried out. The map training was im-plemented using the MATLAB SOM toolbox.

As mentioned before, K-SOMs are used in the present appli-cation mainly for classification purposes: NF, FE, and PE. Thedegree of confidence of the classification depends on the com-paration between the WQ vector received from a PAP (in-cluding TU, pH, C, and T current values) and a reference vector

using a Gaussian kernel defined by

(7)


Fig. 6. PAP’s K-SOM graphical representation.

The reference vector is the vector resulting from the trainingphase that describes the cell whose distance to the current inputvector is a minimum. The corresponding cell is identified bythe so-called best match unit (BMU). The closer to unit is,the higher the probability that the input vector is assigned to thecorrect cluster. For , a weak association between theinput vector and the correspondent BMU must be assumed.

Based on the contribution of to the distance, the faulty channel of a PAP and pollu-

tion events can be detected. In order to distinguish betweenfaulty events and pollution events, a comparison between thevalues acquired at the same instant in different PAPs must beperformed.

C. WQ Data Publishing

The water-quality data publishing is based on the LabVIEWweb publishing tool. Values transmitted by the PAPs to the CPPare published on pap1.htm, pap2.htm, and pap3.htm pages. InFig. 5, the PAP1 web page associated to TU, C, pH, and T datapublishing is shown.

Each PAP can directly generate its own web pages that canbe visualized using an ordinary web browser (e.g., Internet Ex-plorer). PAP’s web page generation offers the possibility to ac-cess online the evolution of water quality parameters using thePAP IP address (e.g., IP 193.136.143.199 for PAP1). However,the limited processing capability of the field units means that thePAP’s pages contain only information related with the currentvalues of WQ parameters and, thus, do not offer additional in-formation about global evolution of water quality based on Ko-honen maps. Also, pollution events or anomalous functioningof PAPs channels are not analyzed at the PAPs level and, thus,are not presented on their web pages. Only the CCP with its

higher computation capability can perform advanced data anal-ysis including data mapping, data recovering [21], and fault andpollution alarm generation.

IV. RESULTS AND DISCUSSION

The WQ information is transmitted to the CCP where the self-organizing maps are offline designed and implemented in orderto perform the mapping of the WQ values received from thePAPs’ measuring channels in the operational phase.

Based on training data associated to the PAP units (PAP1,PAP2, PAP3), a set of three K-SOMs was designed and is pre-sented in Fig. 6.

In that figure, the K-SOM cells correspond to the followingsituations: normal functioning (NF) of PAPs and normal waterconditions (WQ inside EEA limits); fault events (FEs), relatedto anomalous operation of one or several PAP measuring chan-nels (e.g., pH, C, T, or TU channel); and pollution events (PEs),related to WQ parameter values out of EEA specifications.

In order to express the map quality dependence on the sizeof the training data, PAPs’ maps were designed using differentnumbers of WQ parameters measurements (4 624 for PAP1,4 360 for PAP2, and 4 174 for PAP3). The number of mapcells was experimentally optimized in order to obtain maps withthe same number of cells but well-defined clusters, which leadto the 6 12 cells using MATLAB SOM Toolbox som_make( )function. In Fig. 6, it is possible to observe the division of mapsin clusters of cells: clusters that mainly corresponds to NF, FE,and PE situations and whose number of useful cells depends onthe size of the training set (47 for PAP1, 34 for PAP2, and 32for PAP3). Fig. 7 shows that the cluster borders are characterizedby larger distances between neighboring cells, while the cluster


Fig. 7. PAPs D-MATRIX graphical representation: black = 1 and white=0.

areas are characterized by low distance between neighboringcells.

Distance D between a cell and its neighboring cells is cal-culated taking into consideration all the elements of the unifieddistance matrices , , and of a map. In order to obtainthe D values, the MATLAB som_umat( ) function was used.

Once the map designed, visualization of each WQ inputvector implies the determination of its localization in the map.For that purpose, the so-called best-matching unit (BMU) mustbe evaluated. BMU is the value of that minimizes the distance

where takes all values between 1 and 72, in our case. If thereis more than one with the same distance, then the winning cellis randomly chosen among them.

TABLE IIIK-SOM1: EXAMPLE OF WATER PARAMETERS VALUES, CORRESPONDING

BMUS, AND CONDITIONS

TABLE IVPAP’S K-SOM QUALITY EVALUATION

In order to express the correlation between PAPs operatingcondition and BMU values, a set of three WQ vectors associatedwith NF, FE, and PE conditions was used. For the particular caseof K-SOM1, the correspondence between WQ parameter valuesand BMU is presented in Table III.

The BMU value in time gives global information about thetime evolution of WQ parameters acquired by the PAP units.The global information provided by the time evolution of theBMU value can be used to predict undesired events such as pol-lution events. Comparing the evolution of the BMU position inthe maps, false pollution alarms can be reduced and, at the sametime, the WQ evolution for a larger area can be evaluated. Tran-sient alarm conditions are always very difficult to detect, espe-cially when the working condition corresponds to a BMU that isin the border of different clusters (e.g., “PE” and “FE” zones).However, some success can be obtained in false alarm detectionby analyzing the trajectory of the BMU in a given map or cor-relating the measurement data information from different maps.Normal system behaviors correspond to exchanges in a BMUposition between the adjacent cells. Thus, when large jumps be-tween BMU locations are detected, this is generally associatedto a fault event condition.

PAP’s K-SOMs quality evaluation criteria used in the presentwork were resolution and topology preservation. Thus, calcu-lation of the average quantization error (average distancebetween each data vector and its BMU) and topographic error

(percentage of data vectors for which the first- and second-BMUs are not adjacent cells) was carried out. The results arepresented in Table IV and show that the iterative training algo-rithm is better because a) it leads to smaller values of the to-pographic error, which means a better separation between NF,PE and, FE clusters, and b) the training time and computationalload required are lower.


V. CONCLUSION

Even if some interesting solutions in the domain of waterquality monitoring are addressed—use of FieldPoint and Wire-less LAN technology, WEB publishing capabilities using datasocket transfer protocol implemented in LabVIEW, increase ofthe measuring channel accuracy using Multilayer Perceptronneural networks—the main contribution of this work is relatedwith the utilization of Kohonen self-organizing maps (K-SOMs)for multidimensional data representation, data validation, andreconstruction.

The results obtained are encouraging and point to the possi-bility of implementing accurate, standalone smart (self tested)water-quality monitoring systems with space and time inte-grating capabilities and, thus, of water-quality forecasting.

REFERENCES

[1] A. M. Guilikeng, “Real-time monitoring and assessment of river basinconditions for drinking water intake protection,” in Proc. Water QualityResource Management, REC-Telematics, Brussels, Belgium, 1999.

[2] Remote Underwater Sampling Stations. Apprise Inc, Duluth, MN. [On-line]. Available: www.apprisetech.com

[3] YSI Environment—YSI-6600 EDS—YSI Catalog, Yellow Springs, OH.[Online]. Available: www.ysi.com

[4] J. Hoffman, GPRS Demystified. New York: McGraw-Hill, 2002.[5] Public Personal Handy-Phone System, PHS MoU Group. Gen-

eral Description of Interworking/Internetworking of Public PHSNetwork, Tokyo, Japan. [Online]. Available: www.phsmou.or.jp/docu-ments/pdf/B-IW0.00-03-TS.pdf

[6] “Spread-Spectrum radio modem for fieldpoint,” in Meas. AutomationCat. Austin, TX: Nat. Instruments Inc., 2002, pp. 508–509.

[7] J. Patra, A. Kot, and G. Panda, “An intelligent pressure sensor usingneural networks,” IEEE Trans. Instrum. Meas., vol. 49, pp. 829–834,Aug. 2000.

[8] O. Postolache, P. Girão, and M. Pereira, “Neural network in automaticmeasurement system: State of art and new research trends,” in Proc.IEEE—IJCNN, vol. 3–4, 2001, pp. 2310–2315.

[9] N. Valentin, T. Denoeux, and F. Fotoohi. A Hybrid Neural NetworkBased System for Optimization of Coagulant Dosing in a Water Treat-ment Plant. Univ. Technol. Compiegne, Compiegne, France. [Online].Available: www.hds.utc.fr/~tdenoeux/congres/icnn99.pdf

[10] O. Postolache, P. Girão, M. Pereira, and H. Ramos. An internet andmicrocontroller-based remote operation multi-sensor system for waterquality monitoring. presented at Proc. IEEE Sensors. [CD-ROM] 2002.

[11] M35 GPRS Modem- Data Sheet. Siemens, Munich, Germany. [Online].Available: www.siemens.com

[12] J. Travis, LabVIEW for Everyone. Upper Saddle River, NJ: Prentice-Hall, 2002.

[13] (2003) Network Network Toolbox. Mathworks Inc., Natick, MA. [On-line]. Available: www.mathworks.com

[14] J. Vesanto, J. Himberg, E. Alhoniemi, and J. Parhankangas. (2000)SOM Toolbox for Matlab, Helsinki, Finland. [Online]. Available:www.cis.hut.fi/projects/somtoolbox

[15] R. Bitter, T. Mohiuddin, and M. Nawrocki, LabVIEW: Advanced Pro-gramming Techniques. Boca Raton, FL: CRC, 2001.

[16] J. C. Patra, A. Bos, and A. C. Kot, “An ANN-based smart capacitivepressure sensor in dynamic environment,” Sens. Actuators, vol. 86, pp.26–38, 2000.

[17] O. Postolache, M. Pereira, P. Girão, M. Cretu, and C. Fosalau, “Ap-plication of neural structures in water quality measurements,” in Proc.IMEKO World Congr., vol. IX, 2000, pp. 353–358.

[18] S. Haykin, Neural Networks—A Comprehensive Foundation. UpperSaddle River, NJ: Prentice-Hall, 1999.

[19] Water Temperature Level of Pollution in Surface Waters for QualityClass I. European Environment Agency, Copenhagen, Denmark. [On-line]. Available: http://www.eea.eu.int/main_html

[20] T. Kohonen, Self Organizing Maps, 3rd ed. New York: Springer-Verlag, 1997.

[21] T. Bohme, “Sensor failure and signal reconstruction using auto- associa-tive neural networks,” in Proc. ICSC/FAC Symp. Neural Computation,1998, pp. 220–225.

Octavian Adrian Postolache (M’99) was born in Pi-atra Neamt, Romania, on July 29, 1967. He receivedthe electrical engineering degree from the Faculty ofElectrical Engineering, Technical University of Iasi(TUI), Iasi, Romania, in 1992.

In 1992, he joined the Faculty of ElectricalEngineering Iasi, Department of Electrical Mea-surements,TUI, as an Assistant Professor, wherehe is currently Auxilliary Professor. In the last fouryears, he has developed research activity at InstitutoSuperior Técnico of Lisbon, Lisbon, Portugal and the

Instituto de Telecomurncações, where is currently Senior Researcher. His mainresearch interests concern intelligent sensors, laser systems, and intelligentprocessing in distributed measurement systems.

Pedro M. B. Silva Girão (M’00–SM’01) was born inLisbon, Portugal, on February 27, 1952. He receivedthe Ph.D. degree in electrical engineering from theInstituto Superior Técnico, Technical University ofLisbon (IST/UTL), in 1988.

In 1975, he joined the Department of ElectricalEngineering at IST/UTL, first as an Assistant Pro-fessor and, since 1988, as an Associate Professor.Presently, his main research interests are instru-mentation, transducers, measurement techniques,and digital data processing. Metrology, quality, and

electromagnetic compatibility are also areas of regular activity, mainly asauditor for the Portuguese Institute for Quality (IPQ), Lisbon.

J. Miguel Dias Pereira (M’02) received the de-gree in electrical engineering from the InstitutoSuperior Técnico (IST), Technical University ofLisbon (UTL), Lisbon, Portugal, in 1982. In 1995,he received the M.Sc. degree, and in 1999, thePh.D. degree in electrical engineering and computerscience from IST.

He worked for eight years for Portugal Telecom,Lisbon, in digital switching and transmission sys-tems. In 1992, he returned to teaching as AssistantProfessor in Escola Superior de Tecnologia of

Instituto Politécnico de Setübal, Setübal, Portugal, where he is, at present, aCoordinator Professor. His main research interests are in the instrumentationand measurements areas.

Helena Maria Geirinhas Ramos (M’03) was bornin Lisbon, Portugal, in October, 1957. She receivedthe M.Sc. and Ph.D. degrees in electrical engineeringfrom the Instituto Superior Técnico of the TechnicalUniversity of Lisbon (IST/UTL) in 1987 and 1995,respectively.

In 1981, she joined the Department of ElectricalEngineering at IST/UTL, first as Assistant and, since1995, as a Professor. Her main research interests arein the area of instrumentation, transducers, and mea-surement techniques.

322 ieee transactions on instrumentation and measurement...

Documents