for oil-well monitoring

8/8/2019 For Oil-Well Monitoring

1/8

484 IEEE TRANSACTIONS ON INDUSTRY APPLICATIONS, VOL. 36, NO.2, MARCH/APRIL 2000

A Novel Competitive Learning Neural Network BasedAcoustic Transmission System

for Oil-Well MonitoringMarcelo Godoy Simoes, Senior Membet; IEEE, Celso Massatoshi Furukawa, Alexander T. Mafra, andJulio Cezar Adamowski, Membet; IEEE

Abstract--=rhe optimal operation of an oil well requires the pe-riodic measurement of temperature and pressure at the downhole.In this paper, acoustic waves are used to transmit data to the sur-face through the pipeline column ofthe well, making up a wirelesstransmission system. Binary data istransmitted in two frequencies,using frequency-shift keying modulation. Such transmission facesproblems with noise, attenuation, and, at pipeline joints, multiplereflections and nonlinear distortion. Hence, conventional demod-ulation techniques do not work well in this case. The neural net-work presented here classifies signals received by the receiver to es-timate transmitted data, using a linear-vector-quantization-basednetwork, with the help of a preprocessing procedure that trans-forms time-domain incoming signals in three-dimensional images.The results have been successfully verified. The neural network es-timation principles presented in this paper can be easily applied toother patterns and time-domain recognition applications.

Index Terms-Acoustic data transmission, neural network, oilpipeline.

I. INTRODUCTIONA SIMPLIFIED diagram of an oil extraction column usedin deep-sea exploration is shown in Fig. 1. Oil flows upthrough the oil pipeline, i.e., the inner tube, which is made upof tube segments connected by thread joints. The figure alsoshows an external tube, called a jacket, which surrounds thepipeline. The space between the pipeline and the jacket is filledwith water. A column length used in deep-sea exploration mightordinarily be longer than 3000 m. Two different types of equip-ment are used to monitor the temperature and pressure at theoil-well downhole.

I) One is an autonomous data logger, which is used fromthe beginning of the well operation, to store data fromthe initial conditions up to a stabilized situation. Sincethe logger cannot be used online in the measurement, it isimpossible to know the time that the well enters the steadystate, leading to overestimation of the acquisition time.

Paper MSDAD- I 99--01, presented at the 1998 Industry Applications SocietyAnnual Meet ing, St . Louis, MO, October 12-16, and approved for publ icationin the IEEE TRANSACTIONSONINDUSTRYAPPLICATIONSby the Industrial Au-tomat ion and Control Committee of the IEEE Industry Applications Society.Manuscript submitted for review October 15, 1998 and released for publicationOctober 13, 1999. This work was supported by CNPq under Research Grant300.227/96-5, by the FINEP-RECOPE Project, and by NeuralWare, a sub-sidiary of Aspen Technology, Inc.The authors are with the University of Sao Paulo, 05508-900 Sao Paulo,Brazil (e-mail: [email protected]).Publisher Item Identifier S 0093-9994(00)01003-3.

Receiver(Top of pipeline)-----,.,..,r--"T'"rT----- Sea surface

Jacket

Fig. 1. Permanent downhole sensors for oil -well monitoring.

Therefore, the data logger remains in operation longerthan necessary, delaying the beginning of full operationof the well.

2) Another conception of monitoring system replaces thedata logger for the whole operating life of the well; it iscalled a permanent downhole sensor (PDS). The operatormeasures the temperature and pressure every day and in-jects water into the system in order to keep the well op-erating under optimum conditions. However, temperatureand pressure sensors are usually connected to the surfacewith cables, which makes the system prone to damage andexpensive to install and maintain.

This work describes a battery-powered wireless transmissionsystem for this application. With the elimination of cabling, sig-nificant cost savings and increased reliability can be attained. Inaddition, a wireless downhole sensor could be used to monitorthe well at the beginning of operation, replacing the data logger.To implement wireless transmission along an oil-wellcolumn, underground sensors must be placed at a depth of

several meters. Radio waves are not appropriate in this case,since they would have to propagate in the water betweenthe pipeline and the external jacket with consequent largeattenuation. On the other hand, acoustic waves constitute analternative for transmitting data in this situation. Acousticmodems that transmit data in water have been successfullydeveloped for submarine applications and ocean oil platforms[1]. Acoustic transmission through the walls of a drilling tubehas been reported where data were received from tilt sensors

0093-9994/00$10.00 c 2000 IEEE


2/8

SIMOES et ai.: A NOVEL NEURAL NETWORK BASED ACOUSTIC TRANSMISSION SYSTEM 485

Transducer (wellhead) Threaded tool joint Pipe Coupling. . . . . . . . . . . . . . .~ - + - - - - - - - - - O - - - - - - - - - - O l - - :- - - - - ~ - : : :: = : : : : : : : :E : ~ = - - = r = T f l

Fig. 2. Oil production pipeline.

that measured the tilt angle of the drill tip [2], [3]. In the caseof the downhole sensor, data would be traveling along the oilpipeline during full operation of the well.A simple modulation technique used to transmit binary data

is frequency-shift keying (FSK), that employs two different fre-quencies, it and [z- A burst signal at frequency it is emittedwhen a bit" 1" has to be transmitted, while frequency [z is usedfor transmitting a bit "0." Although the encoding principle issimple, the correct detection of ones and zeros at the receiverside of a downhole measuring system is quite difficult. Theacoustic signal is strongly distorted and attenuated after trav-eling along the oil pipeline. Multiple reflections take place atthe pipeline junctions, and propagation through thread jointspresents nonlinear characteristics, resulting in a very complexacoustic signal both in time and frequency domains. The oil flowalso produces vibration, which, in its turn, constitutes a sourceof acoustic noise that influences the arriving signal. Such char-acteristics strongly deteriorate the signal-noise ratio.The usual correlation techniques for demodulation are not ap-

propriate for this case, since their utilization would require in-creased transmission power to reduce the high error rate of thecommunication channel, consequently reducing the autonomyof the downhole sensor. Such nature motivated the developmentof a neural network, to automatically cope with the signal trans-mission and overcome its nonlinear nature.Two fundamental aspects must be considered critical in thedesign of this system: the piezoelectric acoustic transducer,responsible for generating the acoustic waves, and the under-standing of the wave propagation through the steel pipelineelements. Inaddition, operating temperature and pressure (100C/300 atm), size (within a cylinder of 30-mm diameter andmaximum length of 8 m), and noninterrupted operation of upto five years are challenging requirements. Much care must betaken for power consumption optimization, by designing theacoustic transducers and managing the sleeping state (idling)of electronic circuits. A wakeup signal sent from the top of thepipeline commands the system to get out of idling.

II. PIPELINE ACOUSTIC TRANSMISSIONFig. 2 illustrates the oil pipeline, actually the inner tube inside

the external jacket where the oil flows. The pipeline is assem-bled from steel tubes, with length d1 and cross-sectional area aI,connected by threaded tool joints, with length d2 and cross-sec-tional area a2. The acoustic wave propagating on such mechan-ical structure has phase and group velocities depending on thefrequency; some frequencies are blocked for propagation withincertain periodic bands. Therefore, the acoustic wave propagatesat the expense of high distortion due to the variation of phaseand group velocities in terms of frequency. The relationship be-

Transducer (downhole)

200 400 600 800 1000 1200 1400F requency (H z)

Fig. 3. Phase velocity behavior.

tween the angular frequency wand the wave number k is givenby [3]

where Cl = C2 is the extensional wave velocity in steel, cf =w / k is the phase velocity, and cg = ~~is the grpoup ve-locity. Fig. 3 shows the behavior of the phase velocity for an oilpipeline, composed of pipes of9-m length. The acoustic wavescan propagate within the frequency range indicated in Fig. 3where the phase velocity changes from imaginary (representedby the value zero) to infinity, being blocked after the phasecomes back to imaginary again. The figure indicates severalbandpass ranges, the fourih range has been analyzed, showing a180-Hz width. The pipe length might vary up to 0.5 m, leading toshifting bandpass ranges around 50 Hz. For longer lengths, thecentral frequency range is smaller and the bandwidth has suchnature along all the frequency range. In order to choose the FSKoperating frequencies, a tradeoff study of transducer power andattenuation has been made, and the frequencies 4375 and 4425Hz have been selected.

III. COMPETITNE LEARNING FUNDAMENTALSNeural networks are normally used in applications where it

is required to learn hidden patterns and store them in an asso-ciative memory. If the neural network has self-organizing capa-bilities, it can modify the connection strengths based only oninput characteristics; in this case, the storage is autoassociative.The Kohonen network is a self-organizing system, with a singlelayer, having highly interconnected neurons. There is an inverse


3/8


Ouput vector y' for recalling

Outstar Layer

Instar Layer

Input Layer

Input vector x Output y for training

Fig. 4. CPN.

distance function that provides a feedback signal so as to renderexcitatory connections to the neurons in the vicinity, as wellas inhibitory connections to the neurons further away [4], [5].The combination of a Kohonen network and a Grossberg out-star makes up the counterpropagation network (CPN) shown inFig. 4. Such ablending creates a powerful network that can func-tion as an adaptive lookup table in pattern recognition, patterncompletion, and signal enhancement. It contains a supervisedlearning process by virtue of the association of input vectorswith the corresponding output vectors [6]. Even though it is asrobust as a regular backpropagation neural network, it has rapidtraining and saved computational time, via the construction of astatistical model of the input vector environment.Given Xi and Yi as correspondent ith input and output vec-

tors, a training data base is constituted by the set of patterns( X l , yd, ( X 2 ' Y 2 ) , , ( X L , yd. The CPN learns to associatethe input layer vector X with the ouput layer vector y. If the rela-tionship between them could be described by continuous func-tion < I > , as Y = < I > ( X ), the CPN could learn the mapping of any Xwithin the range of the training input and also the inverse map-ping. When an input vector is applied to layer #1, each neuronon layer #2 calculates the net input value, and there is a compe-tition among the neurons to determine which one will send anexcitation to the output neurons.Four modules are working in this neural network: The input

layer-which processes input data, the instar layer, the innercompetitive network, and the outstar layer-the output struc-ture. The input layer might contain buffers responsible for de-livering data to the neurons in the following layer; it can alsopreprocess incoming data, as well as performing the necessaryinput data scaling. Fig. 5 shows the usual way of representing aprocessing element; on the other hand, if such a neuron is seenas an instar element as in Fig. 5(b), the output is no longer dif-ferent from the inputs.Therefore, (2) determines the input pattern of the total inten-

sity for each neuron. To every input pattern is assigned a vector(e 1 , e2 , , e n )T in (3), called the normalized reflected pattern

nI=i* w =LZjWj

j=l

y

(a) (b)Fig. 5. Input processing element. (a) Usual representation. (b) Instarrepresentation.

Fig. 6. Instar learning by approximation of tbe weight vector to tbe inputpattern.

(3)

(2)

where W is the weight vector and iis the input vector.The hidden layer in the CPN is constructed out of instarprocessing elements. The net value is calculated by the innerproduct i* w, assuming that input data has been prescaledto unity length. Therefore, both iand W could be representedin a multidimensional unit of space vector, as in Fig. 6. Thequantity (i - w) is a vector that points the weight vector towardthe input, and the instar can learn such a pattern by rotating theweight vector until it becomes aligned to the input pattern, asindicated by

w(k + 1) =w(k ) + a( i(k ) - w (k) ) . (4 )Although this simple learning algorithm assumes the instarlayer as a "winner-takes-all" competition, there are othertraining methods termed LVQ2, LVQ2.1, and LVQ3, wherethe network structure remains intact [5], but the training isno longer restricted to the winning neuron alone. In the finalimplementation of this work, only the neuron closest to thepattern sends a single +1 output to the outstar layer, shown inFig. 4. The outstar neurons are then trained by either delta orHebbian rule [6].This CPN is done in two steps. First, the Kohonen middle

layer is trained until it recognizes the input patterns and catego-rizes them into an input feature map. It selects the input vectorpattern randomly and then calculates a(i-w) to update theweight vector in accordance with (4). After presentation of a


4/8

SIMOES et ai.: A NOVEL NEURAL NETWORK BASED ACOUSTIC TRANSMISSION SYSTEM

pattern in the input layer, the units in the hidden layer sum theirinputs and then compete among them to respond to the inputpattern. The unity with the highest net input wins and its acti-vation is set to 1 while all others are set to 0; the process is re-peated to cover all input space vectors and after competition theoutput layer sums weighted outputs of the hidden layer. Oncethe middle layer is adequately trained, the weights between theinput layer and the middle layer are frozen and the outstar istrained with supervised Hebbian training [6], [7], i.e., by in-creasing the learning constants until it can reproduce the out-puts. When the network is fully trained, the presentation of apattern in x gives trained y' . A commercial neural network sim-ulator-NeuralWorks Professional-was used to initially de-velop the topology and strategies before writing the final code.Issues like initialization of the neural network, pruning, gener-alization, and tests are hard to do without a graphical simulator.The assumption that all input vectors are of constant norm 1is achieved by inserting a projection layer, which projects theinput vectors to a hypersphere of radius R in a space one di-mension higher than the input space; then, the Euclidean dis-tance between a training vector and a cluster center is given by

liC k - XI12 =2(R2 - Ck . X) (5 )where X is the input vector, Ck is the cluster center, and R is thehypersphere radius.In the context of NeuralWorks Professional, a unidirectional

CPN can be implemented like a radial basis function (RBF)network through the setting of the hidden layer for competi-tive learning with P nearest neighbors heuristic, i.e., a givencluster center Ck, let kl' ... ,kp indexes the P nearest neigh-boring cluster centers. The width of a Gaussian transfer func-tion (O"k) is then set to the root-me an-square distance of a givencluster center to the P nearest neighboring cluster centers

I I'P Llick - ckp 1 1 2 .p=l

O" k =After developing the neural network strategies, a final cus-tomized algorithm was implemented in C++ language, so as torun in the microcomputer embedded acoustic communicationreception system.

IV. INPUTSIGNALPREPROCESSINGFORCPN ESTIMATIONIn order to improve the training convergence of the CPN net-

work and the correspondent classification performance, the rawacoustic signal was preprocessed before being fed to the net-work inputs. A time window was programmed to find significantultrasound activity, i .e., the transmitted burst was sampled andacquired. The discrete vector of n, samples, f, was transformedinto a three-dimensional image with m x m pixels, where m isthe number of discrete levels used to sample the signal. Two newvectors, u and v, were created from f, as indicated by

u [ k ] =f [ k ]

v[k] = f [k + lJ, for k =0,1"", ns.

487

The ordered pairs (u[kJ, v[k]) defined ti; coordinates of theimage domain (u, v). A histogram matrix H(u, v) was con-structed using these coordinates, by counting how many timeseach pixel was addressed by u and v. Such histograms werethen used as the images to be classified initially. However, it wasfound that they showed pronounced peaks around the origin,masking out other features of the image. Therefore, a geometricseries generator was used to compress the histogram peaks andreinforce other points of the image [8]. The geometric series hastwo main advantages over other compression functions (like thelog function 1n): 1) it is less computationally expensive to cal-culate and 2) it offers control on the asymptotic behavior. Thecompressor Z applied to the histogram H( u, v) is given by

1- 0.5H(u,v)Z(u,v) = 1-0.5 . (9)Equation (9) makes up a geometric series with initial value 1and ratio 0.5. The series convergence is 2, i.e., for large valuesofH(u,v).

(6)

A. Application to Distorted Wavefonn EstimationBefore applying this technique to the present work, some

tests were performed in the study of ac line harmonics estima-tion. The standards adopted by IEC-555 and IEEE-519 strictlylimit the total harmonic distortion and maximum individualharmonics injected by the power consumers. Harmonic cur-rents may cause distortion of the voltage waveform via thepower system series impedance, exciting resonances in somedistance far from their source, while odd triple harmonics leadto large neutral currents in three-phase systems. The techniquepresented here of: 1) using the compression function Z(u, v)and 2) employment of a CPN was used to detect the thresholdsstandardized by IEEE-5l9. Fig. 7(a)-(c) shows the mapping ofa 60-Hz sinusoidal waveform and the correspondent third andfifth harmonics pollution. It is easily seen that the surface gen-erated by Z(u , v) encoded the signal dynamics. A descriptionof such investigation is not within the scope of this paper, but itwas very important for debugging the proposed methodology;it successfully confirmed the proposed ideas of this paper.

(7)

V . EXPERIMENTALSETUPA transducer system with all of the required electronics was

installed in the laboratory. This system was used to study thetransmission and also to settle down the hardware and softwarebefore installation in the oil extraction system. An oil pipeline ofapproximately 100 m in length, made of9.5-m-long steel tubesegments which were connected in a chain by means of threadjoints, was used in the laboratory. The diameter of the tubes was4 in and the wall thickness was 0.5 in. The whole structure restedhorizontally on the ground.Acoustic signals were generated by a piezoelectric trans-

ducer with output power of 150 W , connected to one end of thepipeline through a coupler, as shown in Fig. 8. On the otherend, an accelerometer was responsible for signal reception anddelivery to an analog-digital converter interfaced to a personalcomputer. The transducer emitted one bit as a burst of tencycles using 4375 kHz for a "zero" and 4425 kHz for a "one."Bits were separated by 500 ms of gaps to avoid intersymbolic(8)


5/8


Samp le # (k) Sam p le # (k )

(a) (b)

~~~ .----+---t.-------~ :.. . ~. . _- ~. - .. - ---- ' = . : ~ . ' ' ' ll _.__. '

... I " __ , ' . _ . ,_, _Qj . - - !.: -1--"1-----

....lSam p le # (k )

Fig. 7. Distorted waveform estimation. (a) SinusoidaL (b) Third harmonic distortion. (c) Third and fif th harmonic distortion.(c)

Transducer Coupler~1" """[::3F Tube TubeFig. 8. Coupling of the ultrasonic transducer to the oil pipeline.

1 2

. ................ ,. . . . . "0" ~ " 0' ';; 10~;>5- 8.~~ 6"0""0~ 4c..8.. .:2

. ................ . . ..:..... :.... :.... :............................ . . .

v , ..:.... !.... :..... :....

4 250 4300 4350 44 00 4450 4500 4550 46 00 4650Frequency (Hz )

Fig. 9. Transmission channel frequency response.

interference. The transducer was installed 200 m down thepipeline in a prototype installation.Fig. 9 shows the frequency response of the transmission

channel in such conditions. Fig. 10 shows a signal received bythe accelerometer. The amplitude samples had a resolution of8 bits, i.e., in a binary range of from -128 to + 128. Multiplereflections due to the pipeline joints are clearly observed inFig. 10. The signal envelop presents several local maxima andthe time span of the signal is quite long when compared to theshort excitation (ten cycles) imposed by the emitter. Ithas beenobserved that those interactions at the joints added nonlineardistortion, leading to difficulties in using standard correlator

detectors for FSK estimation. The three-dimensional figureshown in Fig. 11 is an example of the images obtained withthe described algorithm for the acoustic signal presented inFig. 10. The accelerometer signal was sampled with 5 bits (i.e.,m = 32). The surface in Fig. 11 was padded with a contour ofzeros to permit a better view of the inner features.A. Algorithm ImplementationThe network was implemented in C++. Such decision was

due the flexibility of operating system implementation and tothe function overloading possibilities, i.e., allowing functionsthat perform similar tasks operating with different data types ofobjects; in addition, C++ provides encapsulation, inheritance,and dynamic run-time binding, allowing reusable code [9] .A flow chart for the network program is shown in Fig. 12,where two learning phases, namely, the Kohonen layer andthe Grossber outstar, are described. The network weights wereinitialized with the average pattern of the whole data set. Itwasobserved that this procedure helped the training convergence.The training was controlled by the learning parameter (a)which was decreased as the training went on. Convergence wasattained when the system classified the stimulus pattern cor-rectly, or eventually stopped by a prescribed maximum numberof training steps. In tum, the program saved the weights usedfor recalling. Therefore, the network received a preprocessedpattern, classifying it into a stored pattern that resembled theinput.The neural network data structure is given in Fig. 13. CPN

is the main variable, which is made up of three lists linked dy-namically: CPN[inputj, CPN[hiddenj, and CPN[outputj, eachbeing a record with the following elements: A (neuron activa-tion) and Wij (connection weights from ielements and j ele-ments on the preceding layer). The software modules determinememory allocation, image width and height, scaling of images,and the class routines for classification and training control. Thenetwork was trained using a set of 40-bit transmission experi-ments, each with 8192 sampled values. Frequencies indicating


6/8

SIMOES et al.: A NOVEL NEURAL NETWORK BASED ACOUSTIC TRANSMISSION SYSTEM

80:; : 40.~~.~ 0"og 40i:


7/8


CPNInputHiddenOutput

Fig. 13. Data structure organizat ion.

Training Parameters:Standard Width: 70Standard Height: 70Scale: 1:1Kohonen learning gain: 0.010000Learning gain scheduling:1.000000Maximum training steps: 5STARTING TRAINING: 11h 41min 23sEND TRAINING: 11h 43min 12sRESULTS:F4375 was recognized 99.17 %F4425 was recognized 97.50 %

Fig. 14. Screen output for the network implementation.

It is possible for some promising methods for generalizationoptimization to be further addressed, like noise injection andweight decaying [10], or like statistics tests to estimate whethera certain weight will become zero during a future training step[11]. Those suggestions are being taken into consideration forfuture improvements in this estimation algorithm.

VI. CONCLUSIONA CPN neural network algorithm was designed for the es-

timation of acoustic signal transmission in an oil-pipeline ex-traction system, where the well pressure and temperature weresensed and sent to the top of the line by the ultrasonic signalsthat performed a FSK-type of digital modulation, with two fre-quencies within the ultrasonic transducer range. The oil pipeline

has junctures around every 10m, which cause multireflectionsand nonlinear distortion, resulting in a complex acoustic signalin the time and frequency domains. The usual correlation tech-niques are not efficient enough for demodulation, and one needsseveral field measurements for setting up the channel commu-nication model. A CPN neural network was used to classify theacoustic signals into two classes of binary transmission, over-coming such nonlinearities and model settings. A preprocessingmethod of raw input acoustic signal was proposed with a chaoticdiscrete mapping, so as to build a three-dimensional image ofthe incoming signal, relating the surface to the time signal.The system was tested in a laboratory with a 100-m-Iong

oil-pipeline. Acoustic signals were generated by a 150-W piezo-electric transducer and an accelerometer was responsible forsignal reception and delivery to an analog-digital converter in-terfaced to a PC. The neural network algorithm was written inC++ and ran in the PC for training and estimation. The inputdata signals were deliberately corrupted with 40% of noise andthe network was still able to correctly estimate original data with2.5% of average correctness, demonstrating its robustness andstability. This present system opens new horizons for oil-wellPDS monitoring systems. Periodically, the PDS can send knownpatterns for the equipment on the top of the pipeline, which willtrain the CPN-based estimation algorithm, making the systemparameter insensitive and more reliable in regard to the naturalenvironmental degradation. Itis important to observe that, in anacoustic transmission system with a neural network capabilitylike this, the channel communication model is not necessary.The CPN-based system can be easily translated to any oil wellas well, and even retrofitted to the existing ones. The system isbeing fitted into a Brazilian sea oil extraction system and can bealso used in other pattern recognition applications.

REFERENCES

[1] R. F. W. Coates, "The des ign of transducers and arrays for underwaterdata transmission," IEEE 1Oceanic Eng., vol. 16, pp. 123-135, Jan.1991.

[2] D. S. Drumheller, "Acoustical properties of drill strings," 1 Acoust. Soc.Amer., vol. 85, no. 3, pp. 1048-1064, 1989.

[3] D. S. Drumheller and S. D. Knudsen, "The propagation of sound wavesin drill strings," 1Acoust. Soc. Amer, vol. 97, no. 4, pp. 2116-2125,Apr. 1995.

[4] T. Kohonen, Self-Organization and Associative Memory. Berlin, Ger-many: Springer-Verlag, 1984.[5] --, "Improved versions oflearning vector quantizations," in Proc. Int.Joint Coni Neural Networks, vol. 1, 1990, pp. 545-550.

[6] L.H. Tsoukalas and R. E. Uhrigh, Fuzzy and Neural Approaches in En-gineering. New York: Wiley, 1997.[7] A. J. Maren, C. T. Harson, and R. M. Pap, Handbook of Neural Com-puting Applications. New York: Academic, 1994.

[8] G. Strang, Introduction to Applied Mathematics. Cambridge, MA:Wellesley-Cambridge, 1986.[9] J. A. Freeman and D. M. Skapura, Neural Networks-Algorithms,Applications and Programming Techniques. Reading, MA: Ad-dison-Wesley, 1992.

[10] A. S. Pandya and R. B. Macy, Pattern Recognition with Neural Networksin C++. Boca Raton, FL: CRC, 1996.

[11] G. Thimm and E. Fies ler, "Neural network pruning and pruning param-eters," 1st Online Workshop on Soft Comput ing 1996.


8/8

SIMOES et al.: A NOVEL NEURAL NETWORK BASED ACOUSTIC TRANSMISSION SYSTEM

Marcelo Godoy Simoes (S'89-M'95-SM'98)received the B.S. and M.S. degrees from the Uni-versity of Sao Paulo, Sao Paulo, Brazil, the Ph.D.degree from the University of Tennessee, Knoxville,and the "Livre-Docencia" degree (D.Sc.) from theUniversity of Sao Paulo, in 1985, 1990, 1995, and1998, respectively.For the last ten years, he has been a Professor of

Engineering at the Polytechnic School, University ofSao Paulo (EPUSP), involved in organizing under-graduate courses in the automation and systems area,

teaching graduate courses in power electronics and drives, and also conductingresearch on applications of intelligent systems. He was a Visiting AssociateProfessor at Georgia Institute of Technology, Atlanta, from January to March1999, teaching graduate courses and conducting research. He has been workingon research concerning fuzzy logic and neural networks appl ications to powerelectronics, dr ives, and machines control. He has authored numerous paperspubl ished in international journals and conference proceedings. His interestsinclude research and development of intelligent applications (neural network,fuzzy logic, and genet ic algorithms) for renewable energy systems, diagnos-tics, sensor fus ion techniques, robotics, rapid prototyping, manufactur ingsystems, and industrial applications. In addition to his academic career, hehas initiated several intelligent-based control system industrial applicationsin Brazil , including applications in the oil extraction industry, photovoltaicapplications, and electric vehicles and drive systems applications. He organizedthe SBAI-Brazil ian Intell igent Automation Seminar, sponsored by IFAC andIFSA, held in September 1999.Dr. Simoes organized a session on "Industrial Applications ofIntelligent Sys-tems" at the IEEE 2nd International Conference on Knowledge-Based Intelli -gent Electronic Systems, organized by IEEE South Australia in 1998 in Ade-laide, Australia. He is also involved with the IEEE Industry Applications Society(lAS), leading technical activit ies in the lAS Sao Paulo Chapter. He organizedthe 1998 Brazilian IEEE-lAS INDUS CON Meeting in Sao Paulo, Brazil. He isa Registered Professional Engineer in the State of Sao Paulo, Brazil.

Celso Massatoshi Furukawa received the Engi-neering degree and the Masters degree in electronicsfrom the University of Sao Paulo, Sao Paulo, Brazil,in 1987 and 1992, respectively, and the Doctoraldegree from The Universi ty of Tokyo, Tokyo, Japan,in 1996.Since 1989, he has been with the PolytechnicSchool, University of Sao Paulo, where he iscurrently an Assistant Professor, engaged in teachingthe undergraduate course on mechatronics inthe Department of Mechanical Engineering. His

research interests include industrial and medical appl ications of ult rasound,industrial automation, electronic sensors, and solar vehicles.Prof. Furukawa is a member of the American Society of Mechanical Engi-

neers.

491

Alexandre T. Mafra received the B.E. degree fromthe University of Sao Paulo, Sao Paulo, Brazi l, wherehe is currently working toward the Master degree.During his undergraduate studies, he was engaged

in research on neural networks applications to patternrecognit ion, such as fingerprint recognition and dig-i tal fi ltering for FSK signals with high noise levels.For the last two years, he has been an IT Consul-tant, developing factory floor monitoring and controlreal-time integrated systems and warehouse manage-ment systems at Uniconsult Systems and Services,

Sao Paulo, Brazil. In addition, for eight years, he has managed his own highschool (Exatus Colegio e Vestibulares), where he teaches physics. His interestsinclude research on neural network applications for pattern recogni tion in dy-namic signals.

Julio Cezar Adamowski (M'95) was born inParana, Brazil, in 1954. He received the Aeronau-t ical Mechanical Engineer degree from the InstitutoTecnologico de Aeronautica, Sao Jose dos Campos,Brazil, the Masters degree in precision machineryengineering from the Faculty of Engineering, TheUniversity of Tokyo, Tokyo, Japan, and the Doctordegree from University of Sao Paulo, Sao Paulo,Brazil, in 1980, 1985, and 1993, respectively.He joined the Mechanical Engineering Depar t-ment, Polytechnic School , University of Sao Paulo,

in 1988 and became a Full Professor in 1998. He has worked on severalresearch projects concerning applicat ions of sound and ultrasound since 1988.His research interests include the modeling of elastic-wave propagation insolids, the ultrasonic characterization of liquids, and ultrasonic NDE.

for oil-well monitoring

Documents