Bionic Sonar for Target Detection

Bionic Sonar for Target Detection Dr. Gee-In Goo IEEE member Morgan State University School of Engineering, EE Department Baltimore, Maryland Address for correspondence: Dr. Gee-In Goo Morgan State University, School of Engineering, EE Department Cold Spring Lane and Hillen Road, Baltimore, Maryland 2239 USA Running Head: AIR; Minelike Target Identification; Resonance Scattering Theory; Target Recognition through Resonance; Mine Detection. Acknowledgement The author would like to recognize and thank to Dr. Joel Davis of ONR for his confidence and the support in form of research grant (NO 004-94 - -0560) for "Bionic Sonar Target Identification Systems'. The author is also indebted to Dr. Whitlow W. L. Au of the University of Hawaii for his continue support in providing recorded target echoes for analysis and investigations in the paper. Additionally, the author extends his appreciation to Morgan State University for providing this research opportunity and facility. O-894-838-2/9S/$6.OO SPIE Vol. 2485 / 8S

Bionic Sonar for Targets Detection A variety of experimental results indicate that Dolphins possess a unique and sophisticated sonar system. In addition, this sonar system is highly adaptive in detecting, discriminating and recognizing objects in highly reverberating and noisy environments. In this paper a new approach using Resonance Scattering Theory in target detection and recognition is presented. The results seems to imply that this approach may be useful in minelike target detection and identification. Introduction Over the past three decades, many sonar experiments have been conducted in the US, Russia and other countries indicating that dolphins possess a superior sonar system which can detect, discriminate and recognize object structures, shapes, sizes and material compositions. Other animals such as whales, seals, and bats also possess unique biological sonar systems which are highly adapted to their environments. For example, echolocating bats possess a highly sophisticated air acoustic sonar system which enable them to hunt for insects and to avoid obstacles and walls in caves. Similarly, dolphins, porpoise, blue whales and sea lions possess unique sonar systems which are highly adapted to their aquatic environments. Using their sonar systems like the dolphins, these sea mammals can identify characteristics of a submerged object by transmitting a wideb and signal and processing the returning echoes from these objects. These experimental observations have been reported by investigators Altes, Au, Floyd, Moore, Nachtigall, Pawloski, etc. Most of these reports have been reviewed by Dr. Nachtigall in 980, reference 4. In general, these experiments have been very successful in determining if a dolphin can discriminate between two or more targets of differing material compositions, shapes, sizes, or structures. These targets may be presented one at a time or simultaneously. In addition, in these experiments, the dolphins are "trained" to recognize the targets individually or simultaneously. Frequently, these results are compared with researchers' effort to emulate the mammals' capabilities in target identification. These emulation frequently require some signal processing 86 / SPIE Vol. 2485

techniques and the use of neural networks or A!. These signal processing techniques have been investigated by Dr. Nachtigall and su m marized as: and a. Filter Bank Model b. Spectrograin Correlation Model c. Time-Domain Highlight Structure Model d. Resonance Scattering Theory. Dolphin Signal Processing Models To better understand the signal processing model used in the dolphin experiments, first let's consider the signals in their respective domains in figures and 2. In figure, the top plot is a dolphin transmitted signal in the time domain, 5(t). The bottom plot is the same transmitted signal in the frequency domain, 5(w), (fourier transform of 5(t)). Notice that the dolphin transmits a sinusoidal like signal with a very short duration. However, the bottom plot shows that this short transmission is a wideband acoustic signal containing frequencies from 80 to 60 KHz, 5(w). Figure 2 are plots of four typical echoes from four cylindrical objects made of four different materials. The left column contains the four time domain signals of the echoes, e(t); the middle column shows the frequency domain signals of the respective echoes, E(w); and the right column are the match filter outputs of these echoes. The ' match filter" output is a signal processing operation, and it is also presented in time domain. However, it is a "time delayed or time shift" (t ) domain resulting from the auto-correlation of the time signal, e(t). Filter Bank Model An interesting and important experiment was performed by Johnson in 968 reference 2. As the name implies, the signal processing is done in the frequency domain, S(w), or the fourier transform of the time echo, e(t). It seems to indicate that the auditory system of a dolphin can be models by a bank of continguous filters. It is frequently known as "comb filters where the filters are connected from lowest frequency (fl) of interest to the highest frequency (fh). Additionally, the filters are frequently designed with a constant "Q" from the lowest to the highest frequency of interest. In this case the filter is known as "constant Q filter". In either case, echoes from targets are analyzed using these filters. The energy content in each of the filters from an echo from a known object is recorded. In an experiment, the unknown object echo is processed and SPIE Vol. 2485 /87

compared with the records for identification as reported by Hammer and Au, and Chesternut et al, in references 4, 6 and 7. A similar experiment was repeated in 993 by Au, where a neural networks was trained to perform the identification part of the experiment. These experiments have reported excellent recognition rate, results of 75 to 95 % correct. Spectrogram Correlation Model Similarly, Altes in 980, references 2 and 3, demonstrated that a mammalian auditory systems consist of a continguous bank of overlapping filters. By determining the energy in each of the filters, and displaying them with overlaps in a time-frequency representation, a high resolution S pectro-gram is generated. This 3 -D s pectrogra m contains characteristic high-lights in the echo relating to the acoustic object. Thus, though decorrelation of the echo spectrogram with known spectrograms in system me mory, the acoustic object can be identified. Excellent results were observed in identification of target in noise, reference 2. Time-Domain Highlight Structure Model Au and Martin's studies, reference 5, indicated that the time domain cues in the "match filter" outputs (right column of figure 2) can be used for discriminating targets. These includes, the time separation between pitches, the rise time, and the duration of these predominate cues. In addition, the amplitude ratio between the predominate pitches are also important characteristics for the target discrimination process. In Au's paper, reference 5, very successful results were reported. Resonant Scatering Theory The resonant scatering theory (RST) is a physical phenomena in nature. It can be simply stated that all object have resonant frequencies or all object will resonate to some natural frequencies. In Newtonian Physics, a beam has a natural frequency depending on its length, width and thickness. Similarly, all objects have natural frequencies depending on the object's size, shape, and structure. Furthermore, it is common knowledge that material composition effects natural resonance. The resonance theory derived by Flax et al in reference 8 and later by Flax, Gaunaurd and Uberall in reference 9, presented the mathematical proof of RST. Since these results are mathematical, the proof can only provide examples of symmetrical objects which can be expressed conveniently, such as plates, spheres and cylinders. These spheres and 88 / SPIE Vol. 2485

cylinders results were also experimentally verified by Tsui and Reid, references 5. These results showed that the frequency resonance appears as the modulations of the frequency spectrum. Since the dolphin transmits a wideband signal as shown in figure, then the modulations on the frequency spectrum of the echoes in the center column of figure 2, are the resonant characteristic of differing material cylinders in this experiment. Therefore, RST can be used to discriminate acoustic targets. Application of Resonance Scattering Theory The Resonance Scattering Theory (RST) was developed theoretically by investigators Doolittle, Flax, Gaunaurd, & Uberall. RST was experimentally verified independently by Brill, Goo, Tsui, Reid, and others. These theoretical and experimental results demonstrated that the size, shape and material composition can be observed by analyzing the spectral content of acoustic echoes. Particularly, the modulations on the spectral signatures are due to resonance. On the basis of these results, it is conclusive that size, shape and material compositions of objects immersed in water can be identified through the processing of the spectral content of wideband acoustic echoes from the test targets. However, it was not obvious that a solution is available, because the modulation on the spectrogram varies from echo to echo as shown in figure 3. Figure 3 shows three echoes from a foam cylinder. The second row in each column is the frequency content of the echoes above. Readers can observe the variation in both the time echoes and the spectrum plots. Readers can also observe the variation in the third row which is the "matched filter' outputs where the Time-Domain Highlight Structure Model is based. Row four is a transform of the frequency domain echo signatures in row two. With a careful inspection of row four in figure 3, one will find that the signatures did not change. If there is changes, it is much less than row one, time echo; row two, frequency echo; and row three, the "match filter" output. If the experimental results from processing signatures in rows two and three are superior, than the results from processing signatures in row four should be even greater. Figure 4 are the transformed signatures of eight echoes from eight cyliners with different material compositions. The reader will notice that each of the signatures are different. It is obvious that a neural networks can be "trained" easily to identify these targets. In addition, figure 5 is a plot of sphereical targets versus cylindrical targets of the same material. The reader can see that the signatures of spheres on the left column have two peaks (not including the DC component at zero); while there is only one peak in the signatures of cylinders in the right column. SPIE Vol. 2485 /89

Based on these signatures in figures 4 and 5, one can conclude that a neural networks can be sitidli for target identification. Conclusion The results seem to imply that the modulations on the frequency spectrum due to target resonance can be obtained from a "Spectral Transform". The signatures from these transforms can uniquely represent the resonance characteristic of the acoustic targets due to its structure, shape, size, and/or material compositions. These signature properties seem to indicate I that neural networks can be trained for identification of these objects; 2/ the spectral transform can also minimize noise in the data; thus improving the performance against noise. 3/ possibly fewer inputs are required for the neural networks because the spectral transform signatures seem to be more compact versus the envolops of the matched filter outputs. These properties requires further investigation during the coming year. Follow-on Investigations When possible, the author would like the opportunity:. to "train" a neural network with 20-25 inputs with 8 hidden nodes, reference 0 & for target identification and evaluate the transformed signatures with noise for its effectiveness against noisy data. 2. to develop an understanding of the relationship between the variance of the spectrum signatures with the transformed signatures. 3. to collaborate with NSWC and/or other researchers in the field to collect minecase data for evaluation and to develop effective techniques for mine detection and identification. 90/SPE Vol. 2485

References. Altes, R. A.,( 980). "Detection, estimation and classification with spectrograms," Journal of the Acoustical Society of America, 67, pp.232-246. 2. Altes, R. A. & Faust, W. J. (978). "Further Development and New Concepts for Bionic Sonar". Naval Ocean System Center Technical Report 404, October, 978. 3. Altes, R.A. ( 980) Models for Echolocation," in Animal Sonar System, ed. by R.G. Busnel and J.F. Fish, (Plenum, New York), pp. 625-27. 4. Au, WWL. ( 988). Detection and recognition models of dolphin sonar system. in P. E. Nachtigall & P. W. Moore (Eds.), Animal Sonar Processes and Performance (pp. 753-768). NY: Plenum Press. 5. Au, W.W.L. & Martin, D. W. ( 986). "Sonar Discrimination of Matallic Plates by Dolphins and Humans," Presented at Animal Sonar System Symposiu m III, Helsingor, Denmark, Sept. 0-9. 6. Au, W. W. L. & Hammer, C.E. Jr.. (980). "Target Recognition Via Echolocation by Tursiops truncatus" in Animal Sonar System, ed. by R.G. Busnel and J.F. Fish, (Plenum, New York), pp. 855-858. 7. Chestnut, P, et al, ( 979 ) "A Sonar Target Recognition Experiment" Journal of the Acoustical Society of America, 66, pp.40-47. 8. Flax, L. et al,. ( 978 ). "Theory of Resonance Excited by Sound Scattering" J. of the Acoustical Society of Am., 63, pp.723-73. 9. Gaunaurd, G. & Uberall. ( 978 ). "Theory of Resonance Scattering from Spherical Cavities in Elastic and Viscoelastic Media," Journal of the Acoustical Society of America, 63, pp. 699-7 2 0. Goo, G. ( 99 ) "Back Propagation Neural Networks Trained to classify Sonar Objects', The International Conference on DSP Appi. and Tech. Oct. 28-3, Hotel Palace in Europa Center, Berlin, Germany. I. Goo, Gee-In, Chang, C-I, and Goo, Heather ( 992) "A Novel Approach to Sonar Target Identification Using Back Propagation Neural Networks", SPIE OR'9 2 Symposiu m on A utomatic Obj ect Identification Applications of Artificial Neural Networks Session, Marriott's World Center, Orlando, FL. April 992. 2. Johnson, "Animal Sonar Systems, Biology and Bionics", R. G. Busnel, Ed., Laboratory de Physiologie Acoustique, Jouy-en-Josas, France. 3. Maze, G. et al., (985). "Resonance of plates and cylinders: Guided waves," Journal of the Acoustical Society of America, 77, 352-357. 4. Nachtigall, P. E. (980) Odontocete echolocation performance on object size, shape and material. In R. G. Busnel & J. F. Fish (Eds.), Animal Sonar Systems, pp. 7-95, New York: Plenum Press. 5. Tsui, C. Y., Reid, G. N. & Gaunaurd, G. C. (986). "Resonance scattering by elastic cylinders and their experimental verification". Journal of Acoustical Society of America, 80(2), pp. 382-390. SPIE Vol. 2485/9

Co U, I CD H -t I =.0 U TRINSMED SIGNAL 0 FREQUENCY SPECTRUM 00 FREIIJENCY (KHZ) 250 isec 200

Ui (JJ ECHO FREQUENCY SPECTRUM ENVELOPE OF MATCHED FILTER RESPONSE a. A/Alumènum ill. ij. '32 b. SlIP, I flionse 52 I - 0 c. IiiP /GIsII 20.0 I,sI.I d. SliP3 %I Ipuvi III' TiME 500JJSEC00 00 FREQUENCY (khz) 200 25OJJSEC 250,USEC Figure 2. Received Echoes from Cylinder of Different Material

Echo # Echo #2 Echo #3 _ - IhA. o I 200 400 200 400 4 _ I 200 400 4[ 2 o ' 200 400 I I 50 00 200 400 50 00 200 400 I II 50 00 'I 0 50 00 0 50 00 0 50 00 Figure 3 Examples of Foam Target Echoes

J J sm bronze, 0 0 50 00 50 0 50 00 50 smalum 0 50 00 50 0 50 00 50 0 0 smsteel. 50 00 50 0 foam 0 50 00 50 wwsscyll 0 50 00 50 0 0 50 00 50 Figure 4 Signatures of Cylinders with Different Material

C' salumspl o 0 0 50 00 50 wwsssph2 iaiumsp2 0 50 00 50 0 salumcyl 0 50 00 50 iiumcy2 0 50 00 50 wwsscyt2 0 50 00 50 0 50 00 50 Figure 5 Signatures of Spherial Vs. Cylinders