Proceedlngs of the z= EAA InternationalSymposium on Hydroacoustics24-27 May 1999, Gdańsk-JurataPOLAND

The Method of Underwater Object Identification Using Multi-IayerPerceptron

P. Soszyński, A Stateczny

Naval Academy, ul. Śmidowicza 17, Gdynia, POLAND

e-mail: soszynski@amw.gdyniapI

Maritime University of Szczecin, ul. Zielona 7, Gdynia, POLAND

e-mail: astat@amw.gdyniapl

An al?orit~m for. detecti0n/fdentijication of underwater objects is proposed. The algorithm is based uponthe classification ability of a simple multi-layer neural network. also called a Perceptron. A signal recordedby a hydrophone and preprocessed by a computer is supplied to the ANN, which classijies it according to thepossessed information encoded within its structure and an array of weighs. Anałysis of ejJectiveness isconducted depending on the vanables pertainmg lo the neural network.

1. IntroductioD

A lot of attention bas been given in recent yearsto the problem of detection and classification ofunderwater targets from the backscattered acousticsignals. Schemes were proposed to facilitate thedifferentiation between mines, partially buriedobjects and solid rocks. In those cases it wasnecessary to transmit acoustic signals and analyzethe reflection. Things are quite different ona battlefield when one needs to determine whethera moving submerged vehicle (for example asubmarine) belongs to friend1y or hostile forces.Now, in order to remain undiscovered, it isdetrimental that detection and classification be donesecretly. To maintain acoustic silence the passiveanalysis of the target' s signature has to be employed,the assmnption being that each vessel bas its own,unique acoustic characteristic. Unfortunately, thereare factors that make such classification difficult:

• variation of target' s signature depending onaspect angIe and range of the target from thehydrophones;

• ambient noise caused by various biological andhuman-made sources [3],[9];

• scattering;

• improvements in silencing of vessel' s machinery(especially the propellers).

Traditionally, two approaches have been used toachieve the purpose of classification: StatisticalDecision Theory [3},[4],[6],[13] and DiscriminantAnalysis [8]. These methods, however, are faidycomplex, time and memory consuming.Alternatively, the artificial neural networks seem tobe an adequate technique for classification of sound-producing vehicles. To determine whether passiveacoustic identification or subrnerged objects byartificial neural networks (ANN) producessatisfactory effects and further investigation hasmerit an experiment was conducted. A simple ANNwas trained to classify nine classes of objects(vehicles) and an analysis of effectivenessconducted.

2. Data Obtaining And Pre-processiug

Ali the information 'necessary to conduct theexperiment were collected at the Polish NavyResearch Center located at the Naval Base, Gdynia.The center maintains a database of acoustic

157

signatures of polish naval vessels recorded at thenearby measuring range and stored in the analogform on tapes. Data regarding three different vessels(A, B, C) at three different speeds (a, b, c) wereaccessed thus creating 9 distinct categories (Aa, Ah,Ac, Ba, ... ). Prograrnmable sigoal analyzer GC-89was used to sample the signals obtaining time-domain records in form of computer files. Sincetime-domain signa1s are meaningless from theclassification perspective the transformation to thedomain of frequency was necessary. FFT wasemployed to go from 1024 element vector in time-domain to 256 element vector in the frequencydomain using MATLAB environment A number ofsuch transformations is shown on Fig. l.

Fig. 1Samples ofFFT

The resulting vectors consisting of 200 elementsin the frequency domain will be now caI1ed records.Each record belongs to one of 9 categories based onthe type and velocity of a vessel. Averaging ofsignaIs was used to m.inimize the ambient noise.NormaIization of records was conducted in order toprevent the ANN learning the relative amplitudes,but rather to "remember" the overall shape of thesignaI.

3. The Algorithm

The expeńment was based upon an algorithmconsisting of four steps:

• registering of the acoustic signature of a vessel,• transformation of this sigoal to the frequency

domain,• decision by the ANN,• interpretation of the result

Of course, for this algorithm to function it wasnecessary to have a well trained neural network,which is the most important aspect of the wholemethod, because the overall quality of theclassification depends on it. As have been mentionedbefore a simpłe multi-layer network was used, but a

158

decision still bad to be made regarding the variablespertaining to this ANN.

Determining tbe stnlcture ol the neurał network(OptimizatioD). Decision to use the Perceptron wasnot all. To achieve the best possible results thefolIowing determinations still bad to be made:

a) how to define the input vector,b) how many neurons should be in the hidden layer,c) how to code the output vector?

The number of elements in the input vectordepended on the number of elements in vectorsobtained after using the FFT, namely 256 values,each corresponding to a different frequency. It wasnoticed that in all the records the amplitudes of thespectral lines with indexes between 200 and 256were low and remained almost constant; thereforethose 56 vaIues were disregarded, thus creating a200-element input vector. The number of neurons inthe hidden layer was changed between 20 and 50and the effectiveness of classification observed. Asfor the last determination, the following codingscheme for the 0Ińput vector was used: the outputvector consisted of five elements, three of whichdetermined the type of avehicie (Ą B ar C) and theremaining two its velocity. The resulting neuralnetwork is shown on Fig. 2.

lnput vector200 clcm<UIs

Hidden Jay..-var.numbec

OuIput JayerS elemems

Fig. 2. The Perceptron

Training ol tbe ANN. The neural network wastaugbt in the following process:

a) A raudom but ba1anced set of records representingeach class called the training set was assembled (thenumber of records for each class was altered.).

b) After each record from the training set was shownto the ANN, its weight matrix was updated (gainingknowledge) and an error was calculated. The error isa measure of how much the response of the ANNdiffered from the expected response. When allrecords in the training set were presented to theneural network (an epoch) an average error for thatepoch was calculated.

c) The process described in the second step wasrepeated until the average epoch error droppedbelow an arbitrarily chosen value. The trainingprocedure was finished, the neural network hadconverged.

Due to the generalization ability of neural networksthe Perceptron was able to leam to discem thenine distinct classes of objects in the above process.The detailed description regarding the techniques ofupdating the values in the weight and the proof forthc convergence of the Perceptron is offered in thefoUowing positions [l], [11], [13],[16].

4. Results

The effectiveness of the method was inspectedaccording to three different factors:

• learning rate,• momentum,• number of neurons in the hidden layer.

Ali these factors affect the speed of reaching theconvergence which is applicable during the trainingphase (acquiring of the knowledge by the ANN). Inthe first analysis the leaming rate and momentumwere changed and the speed of reaching theconvergence in tenns of a number of epochsobserved. Fig. 3 shows the obtained results.

r5 .~~....,.~~.~=iił

Fig. 3. Speed ofleaming by ANN

The changing parameters are the learning rateand the momentum. The constant parameters are thenumber of neurons in the hidden layer - 20 and thenumber of records for each class in the training set -5. It can be inferred that the optimal conditions inrespect to the speed of leaming of the multi-layerperceptron network occurs for the leaming ratebetween 0.2 - 0.3 and the momentum between 0.4-0.6.

Figure 4 shows how the speed of leaming isaffected by the number of hidden neurons. It tums

out that if we choose this number too srnall it willtake a long time for the ANN to converge. lf thenumber is too big however, then the complexity ofthe structure will also hinder the convergence rate.Please note, that the thick lines in the next twofigures represent the trends for the given data.

340'".<:o 2908.Gl

'O~Gl.oE:JC

---~-----_V_------------'~----liIi

number of neurons in the hidden layer

Fig. 4. Speed ofleaming by ANN.

The second part of the analysis of the robustnessof the method concems the effectiveness ofclassification itself which is the primary criteriondetennining the quality of the knowledge imbeddedwithin the neural network. The variable that affectsthis effectiveness within the ANN is the number ofneurons in the hidden layer. The foUowing figuredepicts the effectiveness of classification if we setall the variabies constant expect for the number ofhidden neurons.

~~~~'],~'O~~~

number ot neurons in the hidden layer

Fig. 5. Effectiveness of classification

Agam, it can be inferred that there is a certainargument on the number of hidden neurons axis forwhich the effectiveness of classification reaches itsmaximum value (for the conditions in theexperiment it is about 36 neurons).

5.Conclusions

The primary conclusion is that the method ofclassification based on a neural network classifiergives promising results and should be subject tofurther investigation. Achieved effectiveness, of upto 80%, is representative for the conditions of theexperimental set-up and the data available. It may be

159

different for different conditions and data. It wasa certain restriction that tbe testing was done onlyon the records recorded at similar conditions as tberecords used. to train the ANN. This was due toextema1 limitations regarding the access to thepertaining data.

Reference

[l] 1. Balicki, M przyborski, A Stateczny. Seaobjea identification wuh neural networksapplication. Zeszyty Naukowe AMW Gdynia,pp. 5-18, (1997).

[2] M. Barski, W. Jędruch, 1. żurada, Sztucznesieci neuronowe. Wydawnictwo NaukowePWN, Warszawa 1996.

[3] M.Bouvet, S.C. Schwartz, Underwater Noises:Statistical Modeling Detection andNormalization; JASA, Vol. 83, (pp. 1023-1033) 1988.

[4] P. Devijver, 1. Kittler, Pattern Recognition:A Stalistical Approach, Prentice-Hall, 1982.

[5] H.Demuth, M Beale, Neural Network Toolbox- for use with MATLAB. The Math Works Inc.,1994.

[6] RO. Duda, P.E. Hart, Pattem classificattonand scene analysis. John Wiley & Sons, NewYork 1986.

[7] G. Grelowska, P. Bittner, I. Gloza,Experimenta/ investigation of the underwateracoustical disturbances by the moving ship.Proceedings of the IX symposium onhydroacoustics, Gdynia-Jurata 2-5.06.1992, pp.125-132, (1992).

[8] D.J. Hand, Discrimination and classification.Wileyand Sons, New York 1981.

[9] W. Kiciński, Method of shippłng noiserecognition, Proceedings of the internationalsymposium on hydroacoustics and u1trasonics,Gdańsk-Jurata 12-16.05.1997, pp. 157-162,(1997).

160

[10] E. Kozaczka, G. Grelowska, 1. Gloza,Determmation ofthe ship detection area in thecoastal region, ProceOOings of the internationalsymposium on hydroacoustics and ultrasonics,Gdańsk-Jurata 12-16.05.1997, pp. 169-172,(1997).

[ll] J.L. Kulikowski, Cybernetyczne systemyrozpoznawania, PWN, Warszawa 1972.

[12) S. Osowski, Sieci neuronowe w ujęciualgorytmicznym, Wydawnictwo Naukowo-Techniczne, Warszawa 1996.

[13] z. Paszotta, Zastosowanie metody Bayes'aw identyfikacji wielospektralnych danychteledetekcyjnych. Rozprawa doktorska, OlsżtynART 1983.

[14] M Przyborski., Neural method foridentification ofmoving objects for the needs of(he Navy. Doctorał dissertation, Polish NavalAcademy, Gdynia 1998.

[15] 1. Schurmann, Paitern Classificanon,A unifled view of statistical and neuralapproaches, A Wiley-Interscience Publication,1996. \

[16] J. Szabatin, Podstawy teorii sygnałów,Wydawnictwa Komunikacji i Łączności,Warszawa 1990.

[17] Z. Świątnicki, R Wantoch-Rekowski, NeuralNetworks: Introduction, Bellona, warszawa1999.

[18] J.T.Tou, RC. Gonzalez. Pattem RecognitionPrinciples, Addison Wesley Pub!. Co.,Reading, Ma. 1983.

[19] A Zahallca, J. Principe, Transieni detectionusing neural networks: the search for thedesired signal, in Advanced NeuralInformation Proc. Systems 5, Ed. Hanson,Cowan, Giles, pp. 688-695, (1993).