# Full text of "DTIC ADA275112: VLF Source Localization with a Freely Drifting Acoustic Sensor Array"

## See other formats

I — AD-A275 112 0 Pubi'C mamt sugg( 22202 4302 ana to iMe >j.i TATION PAGE Form Approved 0MB No 0704 owe nou» pet response ir»ciijd»ng the lime tor reviewing instructions searching o*isiing naia sources gathering ar'O «»oh Sena comments teqaramg this Duroon estimate o» any other aspect oi this cotieciion o' intorrnation mciuoing nreclorate lor information Operations ana Reports t2t5 Jefferson Davis Highway Suite '204 Arlington VA Project (0704 0^661 Washington DC 20503 AGENCY USE ONLY (Leave Oiankt 2 REPORT DATE December 1993 3 REPORT TYPE AND OATES COVERED Professional Paper 4 title AND Subtitle VLF SOURCE LOCALIZATION WITH A FREELY DRIFTING ACOUSTIC SENSOR ARRAY 6 AUTHORISI G. Chen and W. Ilodgkiss 5 FUNDING NUMBERS PR; SUBS PE:0602314N WN. DN308105 7 PERFORMING ORGANIZATION NAME(S| AND AOORESSIES) Naval Command, Control and Ocean Surveillance Center (NCCOSC) RDT&E Division San Diego. CA 92152-5001 8 PERFORMING ORGANIZATION REPORT NUMBER 9 SPONSORING/MONITORING AGENCY NAME(S| AND ADDHESS(ESI Naval Command, Control and Ocean Surveillance Center (NCCOSC) RDT&E Division San Diego, CA 92152-5001 iTsPONSORIF^ AGENCY REll>RT,IIUM{ER I i f f: L- -MO- i 1T' ^ t1 SUPPLEMENTARY NOTES c t2a DtSTRtBin’tON/AVAlLABlUTY STATEMENT Approved for public release; distribution is unlimited. 94-01572 5 3 abstract (Maximum 200 worefsj This paper describes the source localization and tracking capability of a freely drifting volumetric array with matched- field processing (MFP) using experimental data. Published in IEEE Journal of Oceanic Engineering, July 1993, pp. 209-223. Accosion F for ACCOS NTIS CRA&I 't DllC TAB Q I Jor.tifiC;itio* By__ Ui..i' lOL ti'j;'. ( Dist A,/ I)-; (v FY.fJc'S ' Av 'il ) 0 / or i‘\o S;; .■cul 14 SUBJECT TERMS environmental adaptation freely drifting sensor array matched-field processing (MFP) simulated annealing source localization t5 numBeF^ 16 PRJCECODE 17 SECURITY CLASSIFICATION OF report UNCLASSIFIED 18 SECURPTY CLASSIFICATION OF THIS PAGE UNCLASSIFIED 19 SECURITY CLASSIFICATION OF ABSTRACT UNCLASSIFIED 20 limitation of ABSTRACT SAME AS REPORT NSN 7540-01-280-5500 Siandurd lorm 296 (FRONT) NSN 7540-01 -280-5500 Standard lorm 298 (BACK) il.JI !()IK\\l < H (K{ AMC KN(;iNh:KKINC.. V( )| IS, \m • If I > :•**>; VLF Soul xe Localization with a Freely Drifting Acoustic Sensor Array George C. Chen, Member. IEEE, and William S. Hodgkiss, Member. IEEE (Invited Paper) Abstract —Source localization and tracking capability of the freely drifting Swallow float volumetric array is demonstrated with matched-field processing (MFP) technique using the 14- Hz cw data collected during the 1989 Swallow float experiment conducted in the northeast Pacific by Marine Physical Labora¬ tory. Initial MFP of the experimental data revealed difficulties in estimating the source depth and range while the source az¬ imuth estimate was quite successful. The main cause of the MFP performance degradation was incomplete knowledge of the enrironnient. An environment adaptation technique using a global optimization algorithm was developed to alleviate the environmental mismatch problem. With limited knowledge of the environment and a known location of the 14-Hz source during a selected time interval according to the source log, the ocean- acoustic environment can be adapted to the acoustic data in a matched-field sense. Using the adapted environment, the 14-Hz source was successfully localized and tracked in azimuth and range within a region of interest using the MFP technique at a later time interval. Two types of environmental parameters were considered, i.e., sound speed and modal wave number. While both approaches yield similar results, the modal wave number adaptation implementation is more computationally efficient. Index Terms — Matched-field processing, source localization, freely drifting sensor array, environment adaptation, simulated annealing. I. Introduction RADITIONALLY, source localization has relied on the processing of assumed plane-wave fronts received by spatially distributed sensors to estimate the source bearing or vertical angle of arrivals. In reality, the ocean acoustic channel is extremely complex due to refractive and multipath effects. Assumption of plane-wave arrivals in the processing scheme in some cases can lead to severe degradation of the estimate. Matched-field processing (MFP) has been proposed [I] to actually u.se the complex ocean acoustic properties to improve source detection and localization. MFP involves the correlation of the actual acoustic pressure field mea¬ sured at the array with a predicted field due to a source at an assumed location deriving from an acoustic propagation model. A high degree of correlation between the measured field and the predicted field indicates a likely source loca- Manuscript received January 20. 19*12. revised May 7. 199.2 This work was supported by the Oflice of Naval Techimlugv under NRI, Contract N000I4- 88-K-2040 G C Chen iv wnh the Nav.d i. i.i\'u.,iiKt Control and Ocean Surveillance Center. RDTVI. DiviMon, San Dice" ('\ '*2152 .2001 W. S Hodgkiss IS with the Marine I'hv-ical l aboratory. Scripps Institution of Oceanograpliv liniversiiv of ( ilr m. i San Diego. CA 92152-6400. lliBL l.og Niiinbct 92IOI.''K tion. MFP of the acoustic wavciield has shown that when sufficient environmental characterizations (e.g., sound-speed profile, bathymetry, sediment properties) are available, rather remarkable detection and localization results can be obtained. Most available matched-field work has been for rather simple propagation situations (e.g., range-independent environment) and much of the work has been re.stricted to vertical-line arrays (!)-[91. Although matched-field processing offers an appealing approach to the underwater source detection and estimation problem, a common difficulty with this technique occurs when the environment information is inaccurate. A “mismatch" occurs between the measured data and the modeled pressure field, and the performance of the MFP is degraded and leads to errors in the e.stimation of the source location [10H141. Several previous studies such as self-cohering [15], envi¬ ronmentally tolerant beamfoiming (161, acoustic tomography (17], Realization 118). and MV beamformer with sound-speed perturbation constrains (19) have been proposed to combat the environmental mismatch problem so as to improve the localization performance. The focus of this paper is twofold: 1) to demonstrate the match-field source localization and tracking capability of the Swallow float freely drifting volumetric array using experi¬ mental data, and 2) to propose and demonstrate an environment adaptation technique that may minimize the effect caused by imprecise knowledge of the environment and thereby lead to MFP localization performance enhancement. This paper is organized as follows. Section II gives a brief description of the Swallow float system and a summary of the July 1989 Swallow float experiment. Section III presents the results of initial matched-field processing on the 14-Hz continuous wave (cw) tone collected by the Swallow floats during the 1989 experiment. Controlled simulations also are presented to aid in interpreting the experimental data priKessing results. Section IV proposes an environment adaptation technique to enhance the MFP localization performance The technique is illustrated using both simulation and experimental data. La.stly, a summary of Ihe paper is given 11 1989 SWAl.I.OW l lO.XI H.XI’HRIMKNT A. Swallow Ebuir Svsiem l)i u "".'imi Over the last several yeais \li'l li.is designed and devel¬ oped a set of 12 acoustic scum ili.ii are neutrally buoyant and freely drifting when dejilov il m the ocean. The sensors are called Swallow floats m iimi.ii of J C. Swallow who 0264 90.59/9.2$03 00 © 1992 IKBI ii i-i J(M K\ \\!( 1 \(.IM I KIN(;. VOI. IS, NO V JOl V nil N A Fig. I. Schematic drawing of a typical Swallow float. used freely drifting floats to measure deep ocean currents. The MPL Swallow floats are used to measure acoustic energy in the very low frequency (VLF) band from 1 to 25 Hz. The Swallow floats are designed to operate without tether so that their measurements are not contaminated by tether strumming noise and flow noise. As illustrated in Fig. 1 (20), (21), each Swallow float is a 17-in. diameter glass sphere containing three orthogonally oriented geophones used as the acoustic particle velocity sensor, a compass, the electronics, and power supply. External to the sphere are a hydrophone for measuring VLF acoustic pressure, an 8-kHz acoustic transducer for transmitting and receiving acoustic ranging signals, an optical flash, a radio beacon, and the relea.se ballast. In operation, the floats are deployed and ballasted to neutral buoyancy at the desired depths. While deployed, each float records signals from the three geophones and the hydrophone sampled at 50 Hz, the compass, and the 8-kHz range puKse arrival times. The acoustic transducer with source strength 192 dB re I /iPa at 1 m generates and receives 8-kHz 10-ms pulses in a programmed sequence. A different float transmits every 45 seconds. When 12 floats are deployed, each float transmits every 9 minutes. The floats are listening whenever they are not transmitting. They receive pulses transmitted by other floats as well as surface/bottom reflections of their own pul.ses. The interfloat and float-to-surface acoustic travel times can be used to determine the float positions as a function of time with a least-squares-based postprocessing procedure 120| B. 1989 Experiment Summary The 12 Swallow floats were deployed for a 24-h period on 8-9 July 1989 near 34°50'N, 122'’20'W, about 150 km wc.si northwest of Pt. Arguello, CA (22). Of the 12 floats. 9 v.crc Irecly drifting in the water column, and } wore tethered to the ocean bottom by 3.05-m lines with lO- to 1.5-lb anchors The ^ Ifue Nonh e * T Y 9 -8 -7 6 5 -4 5 : 1 0 1 2 3 Dtsunce (km) Fig. 2 Roa( honzontal displacement estimates using least-squares filter dunng the July 1989 experiment The circles mark the starting positions. I three bottom-tethered floats were positioned at the comers of a triangle with sides about 6.3 km long in order to provide an absolute reference for the float localization The nine midwater floats were deployed in a quasi-vertical line array geometry | with a vertical float separation of about 400 m, starting at : about 600-m depth to about 3800 rn. The midwater floats were put into the water at about the geometric center of the bottom-float triangle. Fig. 2 shows the horizontal position estimates of the midwater floats from the least-squares method between 00:00-13:58 PST, 9 July 1989 (records 1003 and 2120). The position estimates indicates that the freely drifting floats dispersed away from the center of the float triangle with floats 0 and 1 (the shallower floats) moving to the northwest, floats 2 and 3 to the west, float 4 to the southwest, and floats > 5, 6, 7. and 8 (the deeper floats) to the southeast. The drifting pattern was probably due to the complex water movement near the experiment site. The float depth estimates from the i least-squares Alter are also plotted in Fig. 3. The estimate of rms float position error is less than 4.6 m [23). The float j localization procedure appeared to be capable of estimating float positions to within the desired accuracy of one-tenth of a wavelength at the highest frequency of interest 25 Hz (6 m) in order to effectively beamform the VLF acoustic data. ; As part of a companion experiment, a VLF source was being j towed approximately 2500 km west of the Swallow float array at an average speed of 8 knots and depth of 90 m. Fig. 4 shows the VLF source relative to the Swallow float deployment site. The VLF source was to transmit 14 Hz for a half-hour, then j 8 Hz for a half-hour, then 14 Hz, then 11 Hz and then repeat the pattern. The power spectral estimates from data collected i by float I's hydrophone during records 1040 to 1680 (00:28- } 08:28 PST, 9 July 1989) are presented in Fig. 5 in spectrogram format. Evidently, the Swallow floats can see quite clearly the j 14-flz line projected by the VLF source and, at times, the 11 Hz line A large volume of environmental data such as AXBT, XB3 . and CTD measurements was collected by the companion c\ IH'iimeni during 8-10 July 1989 beiween ,34"r)0'N, 122°2{l'W Fig 3. The hor number z U 1 ■a J Fig. 4. of mar and 3'. profilt 111 . Thi Hz-to’ float ( The fi the sc vector the CO surfac array (7), I that L' conve \vhere -ure f H k , ( I ( H \l 1 / M K >N ‘illcr ons rs of Je an •vater netry ng at Hoals ■:r of >ition Mhod and ifting with west, floats ifting ment n the ite of float ating ith of 6 m) Teing array hows site then cpeai ccted 0:28- ■gram V the .. the \BT. II ex -’O'W 9 . 10.11 JSOO.. i.. 1 .: 900 ICO'' !iX) 1600 1800 300 C 2200 K-v Surrifx’f I ig ^ I'loal depth estimates uMfig least ^quarcs filter. July 1989 expcnmeni The ht'ri/onlal axis <'l'the figure gixoN the time m units of Swallow fh^al record numher l ights records represent one tmui Fig 4 VLF source—Swallow float arras gesmietry. 9 July 1989 The line of triangles marks the AXBT measurements taken between H-IO July 1989 and 32°00'N. 150°0()'SV. Fig. 6 is the summary of sound speed profiles derived from the measurements. III. Source Locai.i/.ation SVith Swai.i.ow Ft.OAT Array This section presents the results from initial MF'P of the 14- Flz-tone propagation eolleeied during the July 1^)8^ Swallow float experiment. There are three steps involved in the MFP. The first step is the estimation of array covarianee matrix at the source frequency, the second is the prediction of replica vectors for all assumed source locations; and the Iasi step is the computation of amhiguiis surfaces a peak m the ambiguity surfaces indicates a likely source location In this studs, three array processing structures commonly reported m the literature |7|, |24| are used to compute the amhicmts surfaces ,o that their results can he compared I he It u': -;; method is a conventional technique that is rohusi h :' . lesoliiiion’ l \ unun {'. .. «) /';"///■; (I) where r is the beamformer output power. I-' ■ -phea pres¬ sure field vector due to a narrow band smiic ■ H). II - X k 7 k V. ■ i i.HunwiUSM 1! ! Y l-mjueocy ill;' l-ig 5 VLF acouMic spcclrograni ol lloal I. (XI 2S--0S F’S I. u .lul> IVSn The vertical aris of ihc figure goes ihe limc m iinils ol S^.iHow lloat record number Eighty records icpreseni one Ik'ui E\XX^] is the cross-spectral densit; matrix or array covan- ance matrix of the sensor outputs,' nd X is vector ot the array Fourier coefficients computed at me frequency of interest. The minimum variance method is a data adaptive technique that yields higher resolution ff| = -^7 - k" n ( 2 ) and the eigenvector methtid exploits ihe orihogou.iliis princi¬ pal that achieves even higher resolution /’vii sictr. ; 'I 1.^) where R\ is the noise cov:in.mce m.iiiiv A. Estimating the Arnn CinniKinn Mann Although each .Swallow iloat colicvo ' geophone data and one channel oi In.i the omnidirectional hydrophone dai.i !!.•■ floats are studied and analyzed m this p.ipe contains its own clock for timing, the In the acoustic data is to align the time Ic! ■ path travel times arc combined to estim.i: uinels of i.ii.i. only ■ ilrifting c.ieh float '■I'Cessing leeipriKal dive clock >•» <M } Wll- I N(ilM-,KKIN(J. VOL IS. \() L Jl LV I'sr, Fig 7 Acoustic pressure 3 too \ Fto»l7 • 40 ■ S A 1 ■ H \ A A Z » ^ X, 0 S 10 IS 20 25 Frequency (Hi) specira for mjdwaier floais acceleration, clock rate, and the offset between two floats (25|. The lime bases are then aligned by choosing one float as a reference whose time base is not shifted and shifting the time base of the rest of the floats’ time series. To ensure the quality of the array covariance matrix estimate, the data records need to be qualified and selected with care. Two criteria used in selecting the data record are S/N and spatial coherence at the frequency of interest. Fig. 7 is the power spectra obtained by processing 3 min of data (records 1143 to 1146). The power spectra are derived from incoherently averaging 28, 50% overlapped, 512-point FKls (97-mHz bin width). A Kaiser-Bessel window with a parameter of 2.5, yielding a sidelobe level of -57 dB (26), weights the data prior to each FFT. Power values are calibrated in decibels re 1 fiPa. The 90% confidence level in these spectra is about ±1 dB. The 14 Hz was being transmitted during the time when data records 1120 through 1160 (01:27-01:57 PST, 9 July 1989) were collected. A line at 14 Hz, about 10 to 15 dB above the background noise, can be clearly scon in all freely drifting floats’ pressure spectra. The high S/N at 14 Hz observed in the power spectra illustrates the good quality of the 1989 Swallow float data sets. The second criterion in selecting the data is to estimate the spatial coherence between floats. The spatial coherence or the magnitude-squared coherence (MSC) is defined as |7xv(/)|" \Sry{f)\^ 5xx(/)5„(/) (4) where Szy{f) is the cross-spectral density at frequency / between x{t) and y{t) with power spectra S^xif) and Syy{f). The MSC function evaluated at / is a real value, conveniently normalized to lie between zero and unity. High coherence of a line frequency harmonic among all float pairs not only indicates the signals originate from the same source but assures high array gain when the individual sensor outputs are combined to form a beamformer. Fig. 8 shows the magnitude- squared coherence and the phase difference between floats 0 and 2 during 00:27-02:57 PST, 9 July 1989 (records 1040-1240). The MSC functions are calculated by averaging over 40.96 s of data, 128-point FFT’s with 50% overlap, providing 31 averages. The high coherence during records 1120 to 1160 is a confirmation that the signal originates from the same source, and the smooth measure of the phase differential suggests that beamforming of this data would be successful. 1 ,L I •2 - I04( Fip 8 Estin three f estimat Hz wic FFT’s . windov I j a 40.9< {XX” In prai reliable the nui will su proces; well ki to ensu 31) use invertil numbe the lai is requ loadin) by add is the ■ systen- sensor B. Act The replicc by a 5 vi*.J \ i HI \ Wl) HOlHiKlSs \\ \ S(tlk(| l(N \ 1 I/Alt(*\ 1040 1060 1080 1100 1120 1140 1160 1180 1200 1220 1240 Record Number acduslit.- iMnni'aii .:i! ''riu cl-ii the .iiiil the vciimm I le <■> IS .1 sciics oi -.pi'L'il piolik's ileriNi’il ti'im ilie t, I D s, the XHI's. .Hill liiL’ A.XlJT's taken alone the a|ipro\iniaie path between the souree ;inj the Swiilluu lloat ;ina> uithin 4X h ol' the .Sv^allov^ lloat experiment All of the proliles along the propagation path exhibit a depth excess that is a necessary condition for long range-propagation. The itteso scale change in the ocean temperature structure across range on the order of several thousand kilometers results in range- dependent variations in the sound-speed profile as shown in Fig. 6. The slow varying sound speed structure variation, especially in the upper water column between 120°W and l.T'5‘'W. is believed to be influenced by the cold waters of the California Current coming dowm from the north. The sound speed structure between 140°W and 15()°W remains relatively stable; this pan of the ocean is known to be environmentally benign. We use the adiabatic mode theory to model the acoustic pressure field. The adiabatic mode theory solution in a 2-D environment is: 1040 1060 10*0 1100 1120 1140 1160 1180 1200 1220 1240 Record Number Fig. 8 (a) Spaiial coherence al 14 Hz and (b) Phase difference beiween floats 0 and 2 at 14 Hz during records 1040-1240 Estimating the array covariance matrix is required by all three processors. The array covariance ntatrix R can be estimated from the measurement data. Frequency bins of 0.4- Hz width centered at 14 Hz were extracted from 128-poini FFT's (2.56 s of data) using a Kaiser-Bessel (with a = 2.5) window and 50% overlap for each of the nine floats over a 40.96-s data record. This results in a 31 dyad produefe (XX^) being averaged for the covariance matrix estimate. In practice, the number of averages required to produce a reliable covariance matrix estimate is about two to three times the number of sensors in the array. In our case, 31 averages will suffice. Since the matrix must be inverted for the MV processor, it should be well conditioned and of full rank. It is well known that as many dyads as sensors must be averaged to ensure full rank. Although a large number of snapshots (i.e.. 31) used to compute the array covariance matrix R ensures the invertibility of the covariance matrix in theory, the condition number of the matrix given by the ratio of the smallest to the largest eigenvalue is below 10“^’. and additional effort is required. We use the white-noise slabili/alion or 'diagonal loading" method (27J. The covariance matrix R is stabilized by adding to the main diagonal the quantity 10 ' (ir/f/\/)(tr is (he trace operator) which corresponds to intrxKlucing m the system an uncorrelated .sensor noise 10 dB below ihc average sensor power level (9). li. Acoustic Propagation Modeling Tlic second step toward MFP is the calciilaiion >1 replica vectors. To predict the acoustic pressure field reccnc.l by a .sensor due to an assumed source, one mu.si model ihe /)(r. = ,-1 ^ !/,,,(0)u, j(c. r j . ■— The modal sum is over the number of propagating modes. A/, that exist between the source and the receiver The modal function involves the depth eigenfunctions at the source u„,(Zo; 0). the receiver u,„(z: r). and the horizontal wave numbers function (,n(-^) reflecting the range dependence of the mediurr along the path between the source and the receiver. The implementation strategy for (,„(■•<) ds is to divide the full range into a number of segments; the horizontal wa\e numbers are calculated at a di.screie sets of ranges where sound speed measurements are available. To determine the number of propagating modes. M, for the oceanic waveguide formed by the source and the Swallow float array, we need to examine the modal depth eigenfunctions. The modal depth eigenfunctions at a frequency of 14 Hz for the environment near the VLF sourcc and the Swallow float deployment site are computed with the ATLAS normal mode model (28), [29], The first 30 modal depth eigenfunctions are gray-leveled and displayed in Fig. 9. Examining the upper portion of the figure that corresponds to the environment where the .source is ioeaied. we see that there are about 22 modes trapped in the water column and are non-bottom interacting; also, for the voutee depth at 90 m, the first 6 modes arc very weaklx exviied For the low-order modes to be strongly excited, the Mmive would have to be placed between its turning points vDH'ie the mode peaks up. The lower portion of the figure eonv'poiuis to the environment at the array site; given the water d. pih .n this liKation, non-bottom interacting modes are limited fo iIk- first 16 modes. Note that the depth functions entei le ■ > pressure field calculation in (5) as product u ,Oo where the one depth function is evaluated at source dc the other depth function is evaluated at the receivei , ; ' As result of the product j ), and .i-.- the bottom interacting modes arc unable to propagate long range due to bottom attenuation, only the first lo u : - ■ ■ i CIU-.N u i-l lOUkNAl. OF CKTANKM.NGtNH RtNCi. VOL IH NO ^ l sSf ,?t S.Hjfce USO i!^ 1520 1 540 Sound Speed ( ir /« c ) 0 5 10 15 20 25 '0 Mode Nunbcf Sound Speed (m/sec) S 10 IS 20 25 Mode Nmnber 00 02 04 06 Of 10 Unevtcik Fig 9 Modal eigenfunctions at frequency of 14 Hz. modes were used in the calculation of the acoustic pressure field. Also, in the mode calculation, earth curvature correction was incorporated due to the propagation distance of 2500 km. In this paper, the 3-D replica pressure field was approximated by evaluating the adiabatic model along a series of N bearings in the range-dependent environment assuming the environment is a^imuth-invariant, to give an /V x 2-D description of the field [30]. C. Experimental Results After aligning the float time bases and performing data quality checks, the array covariance matrix for record 1145 (01-.47 PST, 9 July 1989) was estimated. The problem of grating lobes in beamforming using a sparse array was not addressed; we therefore limited our focus to a region of interest. The replica vectors were computed using (5) for a hypothetical source in a spatial window extending in range from 2300 km to 2700 km, in depth from 0 to 300 m, and in azimuth from 166° to 176° (refer to Fig. 2, we use the mathematical convention with -X pointing 0° as reference and rotating counterclockwise). The sampling intervals in range, depth, and azimuth were 1000 m, 10 m, and 0.1°, respectively. MFP was performed with all three processors; Bartlett, MV, and MUSIC. The peak value in the region of interest was recorded and normalized to yield power in dB re /rPa. Table 1 summarizes the results of MFP on the experimental data along with the true source location and the expected bcamformer output power. Note that power levels were reported for the Bartlett and MV processors onlv since (he MUSIC method does not yield the true power l ip. 10 present the Bartlett TABl.F. I MFP Rf.si i IS Banleli MV MUSIC 1 ',. ■ ' Pr(Kesst>r l^roccssor ProcossDf IV pill t't Max 10 m 1.70 m 10 ni Range Max 26.S9 km 2659 km 2659 km 2-i^t ' km A/imuih Max 172.r 172 1° 172.r 1717 l\»ucf Max 82.7 dB 73 2 dB - .S7 dB matched-field ambiguity surfaces. The upper panel was the range-azimuth ambiguity surface evaluated at the depth where the highest peak occurred while the lower panel was the range-depth ambiguity surface evaluated at the azimuth where the highest peak occurred. The surface was normalized to its highest peak and was marked with this symbol. *. For comparison, the true source location was marked with a A. While mismatch existed, all three processors were in good agreement except for their depth estimates. In fact, all three lacked the depth resolving power [31]. The highest peak in the ambiguity surface differs from the true location by 0.6° in azimuth and 166 km in range. The large number of sidelobes observed in the ambiguity surfaces was thought to be due to imperfect modeling and the sparseness of the array. The MV and MUSIC processors suffered large losses due to mismatch since the replica were imperfect [31]. Ambiguities (or sidelobes) in range were the result of the repetitive structure of the acoustic field in a convergence zone environment. ' D. Controlled Simulation ' While source localization in azimuth was somewhat sue- , cessful, localization in range and depth seemed to be a problem. Simulations are presented here in an effort to under¬ stand the experimental ambiguity surfaces. Three simulation cases were studied; the ideal simulation, uncertainty in float | positions, and uncertainty in sound speed structure. j 1) No Mismatch Simulaton: Assuming there was no mis¬ match and input S/N was 10 dB, the simulated "acoustic data” and replica vectors were generated using the same environment model. A 14-Hz source was simulated at a range of 2493 km, an azimuth of 171.5°, and a depth of 90 m, which is the true source location at record 1145 according to the source ship log. The Bartlett ambiguity functions were evaluated at a depth of 90 m and an azimuth of 171.5° degrees and plotted in Fig. 11. As expected, the source was correctly localized. The high sidelobes found predominately for the ambiguity surfaces produced by the Bartlett processor were believed to be due to the nature of the processor and the array geometry. Also,' the pressure field calculated using the adiabatic normal mode model was composed of only the first 16 low-order waterborne modes. Thus, the acoustic pressure field was less complex or less unique at the simulated source liKation. The poor depth resolution observed in the ambiguity surfaces was thought to be due to the combination of few propagating modes and the low source depth (90 m) to wavelength (105 m) ratio. Fig. 1 2 ) posit no p the 5 simu array sensi over had i The the s posit case in Fi are s Th simu proa [31]. the h SUCCl The rang! simu less CH1-:N AM) HOlKiKISS Ml S(>|!R(| i « k M;.'\U()\ Kange-A/iniuih anthigmr. sum.*. led U km IB IS the where IS the where ced to For I a A. good I three . peak lon by iber of ight to • array, due to (guities ructure nl. Kange-.A/iniutii a;nhik:t< 2700 26S0 2^ 2SS0 2500 2450 2400 2350 2300 fUsseOaa) flSO & 250 300 RA&ge-Oepth ambiguirv surface at izunutfa 171.5 degree (Banlen) . M i 4 . ^ - ^ . 2700 2650 2600 2550 2500 2450 2400 2350 1300 Rjuge 0cm) •3 -2 Rc1m«c Pawn (dB) Fig. 10. Matched-field processing results using the Bartlett method for the Fig. 11. Matched-field processing simulation result.s (no mismatch) using 14-Hz source during record 1145. the Bartlett method. at suc- ) be a under- lulation in float lo mis- ic data” onment f93 km, ^ is the ; source ited at a lotted in '.ed. The surfaces ■ be due y. Also, lal mode iterbome nplcx or lor depth ■ought to > and the o. 2) Uncertainty in Float Positions: The sensitivity to sensor position mismatch was next investigated. Assuming there were no position errors, the replica field were generated using the same conditions as in the real data case. However, the simulated “acoustic data” was computed using a perturbed array geometry where the amount of perturbation to each sensor position had been drawn from a uniform distribution over (4 m, 10 m), similarly the perturbation related to direction had also been drawn from a uniform distribution over (0, 2n). The rms position error for the particular realization used in the simulation was 7 m. The impact of mismatch due to float position error was investigated again in a 10-dB input S/N case. The Bartlett range-azimuth ambiguity surface is plotted in Fig. 12(a) and the all three matched field processing results are summarized in Table II. The sidelobe structures were very similar to those of ideal simulation, but the peak value for the MV and for the MUSIC priKessor was much reduced from that of the ideal simulation (31). Although the mismatch reduces the dynamic range of the MV and MUSIC processors tremendously, the source was succes,!ully located in range and azimuth with reduced power. The ONiiin.iied source range was identical to the “true” source range with a minor discrepancy in the azimuth estimate. These simui.i.. ; results suggest that slight float position error, i.e., less than one tenth of the wavelength (10.7 m) might be the cause of the mismatch in azimuth but cannot be responsible for the mismatch in range. 3) Uncertainty in Sound-Speed Structure: We then Studied the simulation of sound speed mismatch. For simplicity, the sound-speed profiles collected during the experiment were modified by adding a linear function of the form (2}: m/s (6) to the sound-speed profiles collected between 140°W and 1 j 0°W. In this simulation, we used 6 = -3 and D = 2000 m so that at the surface, the sound speed was decreased by 3 m/s. while below 2000 m. there was no change. The specific form (6) is justified for two reasons; 1) The sound-speed profiles collected in this track during the experiment were south of ihc signal propagation path (refer lo Fig. 4) and 2) The general climatic change as one goes from south to north is such ihai the sound speed in the upper part of the water column l.creases. The replica vectors were generated using the ongiii.il profiles while the simulated “acoustic data" were generated using the modified sound speed profiles reflecting (fi) lii. iiarilett range-azimuth ambiguity surfaces for the sound speed mismatch simulation are plotted in Fig. 12(b) and n i l- JOUkNM Oh (K h.ANIC I N(.IM I Kl\< (. lU N AN U nccna inty in Sensor Po&iuons |.j i I- /Ml 2700 2650 2600 2550 2500 2450 2400 25^0 2300 Range (kis) (a) 166 168 |,70 -e J- 174 176 2700 2650 2600 2550 2500 2450 2400 2350 2300 Range (kn) (b) Uocenainty in Sound Speed Stnicture / 1 I f / f » I I Fig. 12. Maiched-field sitnulaiion using the Bartlett Method (a) uncertainty in sensor positions and (b) uncertainty in sound-speed structure. TABLE II Matched-Field Si.mulation of Sensor Position Errors Bartlett Processor MV Processor MUSIC Processor Depth of Max. no m 110 m 110 m Range of Max. 2493 km 2493 km 2493 km Azimuth of Max. 171.7° 171.7° 171.7° Power of Max. -0.16 dB -7.25 dB - the MFP results are summarized in Table III. The results show the source peak shifted in range with much reduced power for the MV and MUSIC processors [31]. The estimated source range is off by 117 km from the "true” source range, and the source azimuth is slightly shifted by 0.1°. These simulated results confirmed that uncertainty in sound-speed structures can be responsible for the large mismatch observed in the real-data ambiguity surfaces, particularly the range error. The environmental mismatch problem will be further investigated in Section IV. IV. Environment Adaptation Matched-Reld Processing MFP has been proposed and developed for localizing un¬ derwater acoustic sotirces by comparing acoustic data with predicted replica pressure fields [Ij. The inputs to MFP are the acoustic parameters of the ocean for predicting the replica U T -1 ';•! I TAIil I. Ill M-m'i iii.ti iTiiii .SiMiT Arms m s,,. Banleii iV'vL - Sl( IX‘pth ol Max no 1 il ;n Range ol Max 2610 ■f.in A/.imuth of Max 171 4° i:i 1 ■ 1 1 Power ('f Max -0 5 dB - in 7 .iH pressure fields and the acoustic data rcceiscd h> an arras of hydrophones. In order to compute the rcplic.t sectors for the MFP, the knowledge of sound-speed structure and bottom characteristics as a function of depth and range are required. As shown in the previous section, if the available environmental information is not sufficiently accurate, MFP can be degraded even if the signal-to-noise ratio is high. Thus, calibration of the environmental parameters so as to improve the matched-processing performance is of special interest. In this section, we propose an environment adaptation technique which has the potential to enhance the matched-field localization performance. This technique is illustrated using both simulation and experimental data. A. Environment Adaptation Technique We envision that the way that the environmental mismatch can be reduced is a two-phase process [15). 1) Adaptation phase: During this phase, a narrowband signal with frequency of interest at a known location is transmitted to probe the oceanic waveguide. The signal could be a surface ship of opportunity or a broadband source with good S/N at the frequency of interest such as air-deployed shots [17], The matcned-field processor is configured in a feedback loop fashion as in Fig. 13 to adjust the environmental parameters with the goal of causing the predicted pressure field to match the measured pressure field, i.e., maximize P(r„, ^o) (7) r Monte minimi many t of exp problet of new allow t by a t tempet less ar Our in uses tl (becau cost ft algorit procec 1 ) 2 ) 3) 4) 5) where F is the environmental parameter set, and Zg, 6„) is the matched-field processor output power due to source at a known location (tq, Zo, Oo). 2) Localization phase: When the environment adaptation phase is completed, the optimized environmental parameter set Fopt is then used to compute the replica pressure fields and normal matched-field processing can resume to search for an unknown target of interest in the vicinity of the reference source. In this study, two types of environmental parameters, sound-speed structure, and modal horizontal wave number, are considered. We use matched-field processor output as a performance function or cost function, since the cost function P is nonlinear and may have many focal maxima (or local minima if we use the negative of P as the cost function); a global optimization technique is required for searching the optimum environmental parameters. /) Global Optimization Method: In this Study, we use the fa.st annealing mcthiHl (IS], [32] for searching the envi¬ ronmental pariimcier spaces. The method is based on the conventional simulated annealing technique [33], a heuristic 6 ) 7) 2 ) rial c cient lead i [18], inter way. ; princ \ find I'Wl array ;ctors ; and ■e are ilable MFP high, as to pecial lation 1-field using match )tation jUency Te the nip of at the |. The loop meters match (7) urce at When imized ute the ;essing in the pes of modal ;d-field nction, ■ many )f P as ^uired use the ; envi- on the euristic CHI N AMI IIOIKIKISS Ml SOl'KCI I ( K' Al 1 / \ I H Sensor Dat* f ' 1 \ .. 1 -^ i-f-T J- •“[l-. 1 \N I Fig. 13. Environment adaptation technique bltKk diagram Monte Carlo search method for the determination of global minimum of combinatorial optimization problems involving many degrees of freedom. Its basic feature is the possibility of exploring the parameter search space of the optimization problem allowing “hill climbing” moves, i.e., the generation of new states (or parameter sets) which increase the cost to I allow escape from local minima. These moves are controlled I by a dynamic variable called temperature, in analogy with I temperature in annealing process. The hill climbing moves are ! less and less probable as the temperature decrea.ses or cools. Our implementation of the fast simulated annealing algorithm uses the negative of the MUSIC matched-field output power (because of its high resolution capability) as the energy or cost function to be minimized. The fast simulated annealing algorithm for this study is encapsulated in the following procedure. 1) Set the initial temperature T„ to a large value. 2) Determine the number of parameters needed to be per- ' turbed using a random number generator with Poisson distribution and select the parameter to be varied using a random number generator with uniform distribution. 3) Generate the perturbations for the selected parameter(s) using a random number generator with Cauchy distribu¬ tion [32]. I 4) Perturb the current parameter set and compute the new : cost, f is the iteration number. 5) If E, < £,-i, then accept the new parameter set and i proceed to (7). I 6) If £, > £._i, then accept the new parameters with a probability given by the Boltzmann distribution, e\p{-AE/kT), the quantity /c is a constant and its dimension depends on the dimensions of A£ and T. 7) Decrea.se the temperature according to the fast cooling schedule [32] which is inversely linear in time, i.e., = 7'„/?, and repeat the procedure, terminate the procedure when no new parameters are accepted for a large number of interactions. 2) Reducing the Parameter Search Spaces: For combinato¬ rial optimization problems of very many parameters, an effi¬ cient characterization to reduce the parameter search space will lead to fast convergence and more uniqueness in the solution [18]. Thus we wouki like to characterize the environment of intere.st in as few parameters as possible yet in a meaningful way. A common method from statistics for analyzing data is principal component analysis. The essence of this method is to find a set of K onhogonal vectors in data space that account loi HA much as possible of the dai.i's '..ui.hkc l’iii|c>,liiig the tlat.i trom their original A'-dimension.il sp.i, c mito ihc A dimensional subspace spanned by these vciHU' ilien peiiornis a dimensionality reduction that often leitmis most of the inirm sic infomiation in the data. Typically. A' <t. .\ . thus making the reduced data much easier to handle. Similar to the principal component analysis method, oceanographers have developed a method for deriving efficient basis functions, known as empirical orthogonal functions (EOF’s), for measured physical quantities such as temperature, salinity, or sound speed as a function of depth [17], [18], [34], [3.“)]. In an effort to reduce the ocean-acoustic parameter spaces, one can describe the parameter set as a sum of EOF’s. The EOF's are defined as the eigenvectors V, of the parameter covariance matrix R: RV, = X,V, (8) where A, is the ith eigenvalue or the variance associate with V,. The covariance matrix for the environmental data R is defined as R = iJlf F] (9) where F I' - £[r], and F is the measured environmental parameter values. E\ ] is the expectation operator. The pa¬ rameter search spaces can then be represented or spanned as a linear combination of appropriate EOF's. A’ Ein-t-^QnV'n (10) n = l where N is the total number of eigenvalues, the V^„'s are indexed so that A^ > A„ + i, and rt„’s are the EOF coefficients. In practice, a high degree of accuracy can be achieved with only two or three EOF’s for representation of ocean-acoustic parameters [17], [35]. Using the EOF approach to parameterize the environment, (7) can now be expressed as maximize P{r„, Zo, Oo) (11) {oi, fc = l, K ) where K < .V and is related to F by K r = E(F) + ^a,V^,. (12) t—i B. Simulation Results The environment adaptation technique described above is first applied to simulation data. To make the simulation as realistic as possible. Ave modeled the source-array geometry corresponding to the Inly 1989 experiment with a 14-Hz source deploved ;ii the locations according to the source ship log. The freely drifting sensor array geometries corresponding to the navigation lesiilis during the two time intervals, i.e., 01:47 and 04:38 I’.S f. July 1989 were used in the simulation CllLN Fig. 14. Matched-rteld simulation of environmental mismatch for the two time intervals corrsponding to the experimental data collected during 1989 experiment. study. To simulate the environmental mismatch, similar to that of last section, we modified the profiles between 34°N, 140° W and 32°N, 150°W by adding a function of the form; AC(2) = - Cmean(^) ftVS (13) where C 340 N, hoovv is the sound-speed profile taken at 34°N, 140°W, and Cmean is the mean profile between 34°N, 140°W and 32°N, 150°W (refer to the solid line in Fig. 18(b) for the shape of AC(a)). Although arbitrary, this perturbation was justified by the fact that the sound-speed profiles collected in this track were off the signal propagation path. The simulated “acoustic data” were generated with an adiabatic normal mode model using the sound speed profiles reflecting (13), while the replica vectors were generated using the original measured sound-speed profiles. The range-azimuth ambiguity surfaces for the mismatched case during the two time intervals are plotted in Fig. 14. The A’s are the “true” source locations according to the source log, and the *’s arc the MFP esti¬ mated source locations, i.e., the highest peak in the ambiguity surfaces. As expected, the estimated source locations using the mismatched environment were off in range by roughly 100 km to 150 km (2 to 3 CZ's) for both time intervals. I) Sound-Speed Adaptation: We now proceed to the envi¬ ronment adaptation technique and assume that the source-array geometry is known exactly during the first time interval (01:47). Under the assumption that the adiabatic approximation is adequate to predict acoustic pressure field in a slowly varying range dependent environment for MFP, one can show that there will be some equivalent range independent sound- speed profile [31). The first simulation case is to invert for the range-independent sound-speed profile representing the environment along the signal propagation path between 140°W and 150°W in a matched-field sense. Fig. 15(a) shows the excess (demeaned) sound speed profiles for the upper ocean derived from 30 AXBT measurements made between 34°N, 140°W and .32°N, 150°W in July, 1989. Table IV lists the five largest eigenvalues obtained through eigen-decomposition of the sound speed covariance matrix. As can be .seen, only the first few eigenvalues are significant. Wc thus use the eigenvectors or KOl-'s eone-.p.svlnu' lo ihe two largest eigen¬ values in spanning the ve.iul. -p.i. ■ - \unn.ili/ed versions of the F.OF's corresponding 10 the iwu largest eigenvalues are shown as solid and doiieil tuiww lespeeiively. in Fig. 15(b). The eigenvalues or v anaiiees .s'l i. spondiiig to these FOF's were Ai - np = 54.(lti and \ . respectively To ensure all possible parameiei v.ilues are reachable in the search spaces, we allow the values of the liOF coefficients to vary within ±6rr„. Using the mean sound-speed profile, the first two FlOF's. and the initial solution of the EOF coefficients of (0, 0). the optimi/ation is performed with the fast annealing procedure. Fig 16 shows the convergence properties of a typical annealing nin. In this example, the temperature was initialized at 2(X) and was reduced to 0.2 | over 1000 iterations. The first and second panels are the trajectories of the EOF coefficients while the third panel is the cost function learning curve as the annealing proceeds. As expected, the cost function learning curve declines as the iteration goes on, but occasionally the curve increases to a higher energy state thus indicating the escaping out of the local minima. After 350 iterations, the optimization process converges to the minimum energy state, and the solution of the EOF coefficients found at the 1000th iteration is (9.00, 21.91). To validate the optimal solution, the energy (cost function) surface as a function of the EOF coefficient values is calculated exhaustively on a regular grid in ni and 03 (with a grid size of 1) bounded by (-50, 50). which corresponds to 10201 evaluations of the cost function as displayed in Fig, 17. The surface shows a global minimum at f>i = 9, 02 = 22 with numerous local minima scattering around. The joint trajectories of the EOF coefficients as the annealing proceeds are overlaid on the energy surface plot. A good agreement between the solutions obtained by grid search and optimization is observed. It is of interest to know whether the optimal solution of the EOF coefficients varies from run to run. The environment adaptation procedure is then repeated for nine times, each time with 1000 iterations. The solutions are given in Table V. The convergency properties and the consistent agreement among all runs are again observed, which confirms the fitness of the global optimization algorithm. The adapted sound-speed profile derived from the mean profile and the two EOF coefficients found at the 1000th iteration is plotted by dotted line in Fig. 18. For comparison, the modified sound-speed profiles (i.e., the “true” profiles) used to simulate the “acoustic data” are averaged and plotted by solid line, and the measured profiles that were used to model the replica vectors are averaged and plotted by dashed line. As can be seen, the adapted profile tracks the “true” profile closely. We now proceed to the localization phase of the technique. The measured sound-speed profiles between 140°W and 150°W are replaced by the single adapted sound speed profile, and normal matched-field processing resumes. Fig. 19 shows the optimized range-azimuth ambiguity surfaces for both time intervals. The environmental adapted source locations match those of an ideal simulation (no mismatch). 2} VV'rtit’ Number Adaptation: The second implementation of the environment adaptation technique is to invert for the modal wave numbers at the source frequency (14 Hz) for 2 ( I' i> 1 2 ' a 1 I Fig. 15 AXBT Normal E the ac The a speed wave To CO the sa execs; numb AXB' of the showi the h now I surf at joint .'.nnca I '‘'i ’» > ■ ( Ml N AND HOrXiKISS M l SOUKC'I: I.CXTALIZATION ’IM ;Ji-n IS Ol ■ .III' S(h) JPs •ely. I the tents )file, EOF with ence . the > 0.2 the el is jeds. s the to a " the •cess in of 900. (cost diues I with I ds to joint j eeds j ment j ition I imal I The I nine 'jven stent hrms nean OOth ison, hies) jtted d to shed true" X of vecn ound I mes. faces •urcc ;h). ation r the ) for .Oi 06 0.4 -03 0.0 0.2 0.4 06 06 Nonnalized Souod Speed EOFs (o/s) (b) Fig. 15. (a) Excess (demeaned) sound-speed profiles computed from 30 AXBT measurements made between 140°W and 150°W. July 1989. (b) Normalized versions of the first and second EOF's derived from (a). TABLE IV The Five Largest Eigenvalues Obtained From ElGEN-DECOMPOSmON OF THE SOUND-SPEED COVARIANCE MATRIX Number Eigenvalue 1 54.06 2 35.96 3 11.93 4 2.81 5 1.81 the acoustic waveguide of interest in a matched-field sense. The assumptions and conditions are identical to the sound speed adaptation case with one exception, that is, the modal wave number EOF coefficients are now the search parameters. To compare these two implementations, we followed exactly the same path of investigation. Fig. 20(a) shows the first 16 excess (demeaned) wave number estimates of the modal wave number at the source frequency (14 Hz) obtained from 30 AXBT measurements made in July 1989. Normalized versions of the EOF’s corresponding to the two largest eigenvalues are shown as solid and dashed curves in Fig. 20(b). Note that the lag in the wave number covariance matrix estimate is now the mode instead of the depth. Fig. 21 shows the energy surface computed by exhaustive .search and an example of the joint trajectories of the wave number EOF coefficients as the annealing proceeds. Using as a reference the averaged model 0 lOD»DKfi400SOOfiOOX>OK»«» 1000 Iwva* 0 lOO»)M0400SOO«00 100tDO«nKn) 0 lOOXlOWDaOOSOOOOOIOOiOOQOO tooo Imwom Fig. 16. Trajectories of the sound-speed EOF coefficients and cost function learning curve for a typical annealing run. 12 10 I 6 4 2 0 Relative (dRt Fig. 17. Energy surface as a function of sound speed EOF coefficients computed by exhaustive search (simulation). wave numbers derived from the measured profiles, the “true” minus the measured and the adapted minus the measured are plotted by solid and dotted lines, respectively, in Fig. 22. TTie adapted wave numbers track the "irue ' vsave numbers nicely except for the first five modes. A possible explanation of the I Ihih IIMKNAI <)( (XI:aNIC KNOlNI'.hRING. VOL IS, NO LILLY 1995 IABI.I-; V O' liMi.'-MKi.N Rim I IS iso.M Nink Runs Using i' Si'i I D LOf- Cm H U ii NTS AS Skarch Parameters l;()I Cud # 1 EOFCoef. #2 M7S ISM I4«S 1490 I49S ISOO ISOS ISIO ISIS IS20 ISIS 5»auod Spcied (a^) Sound Speed Difiereacc (oA) Fig. 18. (a) Sound-speed profile derived from Che environmenl adapcation technique (dotted line), “true” profile (solid line), and measured sound-speed profile (dashed line); (b) the difference version of (a), i.e., using the measured profile as reference (dashed line), the “true" minus the measured and the adapted minus the measured are in solid and dotted lines, respectively (simulation). Fig. 19. Self-cohered source locations by envu'onmeni adaptation technique using sound-speed EOF coefficients as search parameters. ■ 0.6 - 0.4 -oa ao 0.2 04 0.6 Ncnalbed Wtve Nnster BOfh (nd/B) deviation from the “true" wave numbers is that these modes are weakly excited, thus less weight is given during the adaptation process. Again, the environment adapted locations match those of the ideal simulation [31]. It is of interest to know whether the true source location is unique in the neighborhood of the source location. The adaptation procedure is then repeated for each assumed source location within a spatial window extended 16 km in range and 1.5** in azimuth (approximately 60 km in cross range) enclosing the true source location. Fig. 23 confirms that the range-azimuth maximum energy surface indeed has a peak that corresponds to the true source location. This suggests that a Fig. 20. (a) Excess (demeaned) borizonlal wave numbers at 14 Hz derived from 30 AXBT measurements made between 140“W and I50°W. July 1989. (b) Normalized versions of the first and second EOF's derived from (a). weakly range-dependent environment can be approximated by a range-independent environment in a matched-field sense. Although both implementations of the environment adapta¬ tion technique produce similar results under the example given above, the computation burden for the two approaches differs significantly. For the sound speed case, evaluation of the cost function requires the invocation of the acoustic propagation model to compute the modal eigenfunctions and wave numbers > I ' I, I f I I derived y l9Xy .11 led by l^e dapia- eivcn li Iters e cost e.ition tubers I Ml N \M I IK ll U ;KIS^ \ 1 I Si H Kl I I I H \1 1 / \ l II IN , - j r --,-^ -, •jO ^ M 10 10 0 10 K) 10 40 W bOFC«tr ti IS 10 S 0 RelJtive Power (OBl Pig. 21. Energy surface a.s a function of wave number [X)P coefficients computed by exhaustive search (simulaiionf •10* -4-ia’ •2Ma' 0 2M0' 4-10' 6*10' Wtve Number Differaice (r«dAn) Fig. 22. Deviations in adapted wave numbers and ’ime” wave numbers from the measured model wave numbers (simulation). (this is computation intensive). In the wave number adaptation case, the wave numbers are readily available, and updating the wave numbers according to the perturbations involves only a few trivial algebraic steps. The difference in computation performance for the two implementations is at least 3 to 4 orders of magnitude. C. Experimental Results Data collected by the Swallow float array during the July 1989 experiment in the northeast Pacific were processed in the same fashion as the simulated data The acoustic modeling was the same as in the simulation. The data considered here were recorded at 01:47 and 04:38 PST. 9 July P1S9. Fig. 24 plots the matched-field rangc-a/.imuth ambiguitv surfaces for the two intervals with the highest peak marked with ‘ s and the true source locations marked with A s. The shilt m the source locations was due to the environmental mismatch as diagnosed through simulation in the last section. We empirically selected MW nm IM) M91 um MT MS IM) Sows tait 10 •% -6 -4 -2 0 Power (<JB) f'Dj 2^ Range a/imuih energy surface derived from rcp‘-*aling ihe tion priKcdure for each rangc^azirnuth cell (simulationi the reference source location that corresponded to 01:47 PST. 9 July 1989, by repeating the adaptation procedure for all assumed range-azimuth cells in the neighborhood of the source location obtained from the source log. The highest peak in the range-azimuth maximum energy surface was found at a range of 2489 km and an azimuth 172.1° 131). We then entered into the adaptation phase of the technique. The sound-speed EOF coefficients were used as search parameters to determine the optimal sound speed structure. Fig. 25 shows the optimized range-azimuth ambiguity surface for both time intervals using the single adapted sound speed structure. As can be seen, the environmental adapted source track mimics the track derived from the source log. The modal wave number adaptation implementation also produced the same results [31], Table VI lists the adapted source locations versus the source locations according to the VLF source log. The minor discrepancy, a 0.2° to 0.6° shift between the true and adapted source locations, is thought to be due to the slight uncertainty in selecting the coordinate system with respect to true north for the float localization (refer to Fig. 2). since the onentation of the X axis of the coordinate system was taken to bo the ship's position when bottomed floats 9 and 10 were pul into the water. A relative motion of approximating 60 ni in the Y direction between floats 9 and lO while the floats descended to the bottom would cause the O.ti' rotain'ii in the MFP results. This order of error between ship position and true float position on the bottom has been noted in other Swallow float experiments. In addition, uncertainty in niidw aicr sensor positions might also contribute to the azimiiili error .i- seen in the simulation study. The minor discrepance m r.inec error, a 4-km shift, is thought to be due to errors in the ,l^^l;nlpllons made in modeling the replica vectors. Noneihclcs, 'iie relative source movement, a 17-km separation between ih, a vourcc locations, is found to be consistent with that ol die mucc log. f ig 24 Matchcd held processing of experimental data during the two time intervals. Fig. 25 Self-cohered source locations by environment adaptation tech¬ nique using sound speed EOF coefficients as the search parameters. TABLE VI Environment Adaptation MFP Results Time True source True source Adapted source range Adapted source azimuth range azimuth 00;47 2493 km 171.5° 2489 km 172.1° 04:38 2476 km 170.7° 2472 km 170.9° V. Conclusions In this paper, we have demonstrated the source localization and tracking capability of a freely drifting volumetric array with MFP using experimental data. We have also proposed and demonstrated an environment adaptation technique to deal with the environmental mismatch problem. Data collected during the 1989 Swallow float experiment were used to perform the study. The geometries of the Swallow float array as a function of time during the 1989 experiment have been estimated using the 8-kHz range measurement with a least- squares based float localization method. The rms position errors estimated by the float localization method is less than 4.5 m, which is within the desired accuracy of one-tenth of a IHl nUKN.M 01 IHT-ANIC INI'ilNlt RlNt'. Vlll IS. NO T )l I I A.i'l'Icnjflh at the highcM rrequency of intcresl 25 11/ ((i im la .icr III cticcTncly process the VI.I- acoustic data cotieieiul\ I urihemiore. analysis of the experiment acoustic data shoued !.:eii Munal-to noise raiio and high coherence ai 14 11/ aiiione ...I nine lreel> driliing floats during some time inlervals I he 14 (1/ was a coniinuoiis wave tonal projected by a VLI- source iiiNolved in a companion experiment. The high coherence among the floais provided an opportunity for matched field beanilorming of the VLF acoustic data. The replica vectors were modeled with an adiabatic normal mode numerical technique using the environmental data collected during the experiment. Inttial MFP of the experimental data expertenced diflicultics in estimating the source depth and range while the source azimuth estimate was somewhat successful. Controlled simulation using the same conditions as in the real data has revealed that 1) depth resolution indeed is a difficult problem for a shallow VLF source in a long-range environment due to fewer modes being available and due to low source depth-to-wavelength ratio, 2) the range estimate is sensitive to environmental mismatch, and 3) the azimuth estimate is robust. The main cause of the performance degradation has thus been identified to be uncertainty in the environment (i.e , sound-speed mismatch). An environment adaptation technique using a global optimization algorithm was proposed and de¬ veloped to counteract the MFP performance degradation due to uncertainty in the ocean acoustic environment. We have demonstrated through simulation that with limited a priori knowledge of the environment and with a reference source at a known location, the environment can be adapted in a matched-field sense. Using the adapted environment, other unknown source(s) of interest in the vicinity of the reference source can be correctly localized. Applying the environment adaptation technique to experimental data has shown that ttje 14-Hz source was successfully localized and tracked in azimuth and range within a region of interest using the MFP technique at a later time interval. While both types of environmental parameters, i.e., sound speed and wave number, provided similar results, the modal wave number adaptation implementation has proven to be computationally efficient. 112) R. or of IS I'31 Acknowledgment The authors are indebted to G. D’Spain and J-M. Tran of the Marine Physical Laboratory for several useful discussions concerning the issues of Swallow float system and random array matched-field proce.ssing. They also thank H. Bucker of the Naval Command Control and Ocean Surveillance Center. RDT&E Division, for many useful discussions on the self- cohering technique. References ni H. P, Bucker. “Use of calculated sound fields and malchcd-ficld detec¬ tion to locate sound sources in shallow water," J. Acousl. Soc Am., vol 59. pp. 368-373, 1976. (2| M. B. Porter. R. L Dicus. and R. G. Fizcll, "Simulations of matched- field processing in a deep-water Pacific environment," It-'HF. J Ocean Eng., vol. OF.-12, pp. 173-181. Jan. 1987. 13) R. G. Fizell, "Application of high-resolution processing to range and depth estimation using ambiguity function methods," J Acoust .Soc Am , vol. 82, pp 606-613. 1987. XT S I ? “ 'Z'5 o Il.\ 1993 m) in ;rently. -howed among Is. The source lerence . id-field vectors Tierical ing the rienced tile the itrolled al data lifficult onment source insitive nate is on has nt (i.e., hnique ind de- on due ' e have ' priori source ! id in a j . other I Terence i tnment n that ked in | ng the pes of umber, ptation ient. Fran of jssions andom cker of Tenter, e self- id detec- tm., vol Hatched r Ocean CHI-.N AND MOIKiKISS \l| SODKCI- 1 (K AI l/.ATION (4) A B Ba'iiicroor. \V A Kupciin.tn. and H Schmidt. ■‘M-Hchod-lield pro cessing. Source loc.ili/.iiion in correlated noise as an opiiinuiii parameter estimation problem.' / Ac iiusi Soc. Am . vol. X.t, pp S7I S87, 1988 |.S| J. M. 0^ard, Maiched held proces.sing in shallow water for range, depth, and bearing determination: Results of experiment and Simula- tion," J. Acau.u Soc Am., vol 86, pp. 744-753, Aug 1989 (6| C. Feuillade, W A Kinney, and D. R. Del Balzo "Shallowr water matched-field localization off Panama City, Rorida," J Acousi. Soc. Am., vol. 88, pp. 423-433, July 1990. (7) C. Z. Zala and J. M. 07.ard, “Matched-field processing in a range- dependent environment," J. Acousi. Soc. Am., vol. 88, pp. 1011-1019, Aug. 1990. [8| B. J. Sotirin, J-M. Q. D. Tran, and W. S. Hodgkiss, "Matched-field processing of deep-water ambient noise," /£££ J Oceanic Eng., vol OE-15, pp. 316-323, Oct. 1990. (9) J-M. Q. D. Tran, and W. S. Hodgkiss, “Matched-field processing of 200 Hz continuous wave (cw) signals," J. Acousi. Soc. Am., vol. 89, pp. 745-755, Feb. 1991. (101 A. Tolstoy, “Sensitivity of matched-field processing to sound speed pro¬ file mismatch for vertical arrays in a deep water Pacific environment," J. Acousi. Soc. Am., vol. 85, June 1989. (11] C. Feuillade, D. R. Del Balzo, and M. M. Rowe, "Environmental mis¬ match in shallow water matched-field processing: Geoacoustic parameter variability," J. Acousi. Soc. Am., vol. 85, June 1989. (12] R. M. Hamson and R. M. Heitmeyer, “Environmental and system effects on source localization in a shallow water by the matched-field processing of a vertical array," / Acousi. Soc. Am., vol. 86, pp. 1950-1959, Nov. 1989. (13] D. F. Gingras, “Methods for predicting the sensitivity of matched-field processors to mismatch," J. Acousi. Soc. Am., vol. 86, pp. 1940-1949, Nov. 1989. (14] J. R. Daugherty and J. F. Lynch, “Surface wave, internal wave, and source motion effects on matched-field processing in a shallow water waveguide," J. Acousi. Soc. Am., vol. 87, pp. 2503-2526. June 1990. (15] H. P. Bucker, “Self-coherent matched-field processing," Founh Malched-FieUl Processing Workshop. Defence Research Establishment Pacific. Sept. 1989. (16] H. Schmidt. A. B. Baggeroer, W. A. Kuperman, and E. K. Sheer, "Environmentally tolerant beamfomting for high-resolution matched- field processing: Deterministic mismatch," J. Acousi. Soc. Am., vol, 88, pp. 1802-1810, 1990. (17] A. Tolstoy, O. Oiachok, and L. N. Frazer, “Acoustic tomography via matched-field processing.” / Acousi. Soc. Am., vol. 89. pp. 1119-1127, Mar. 1991. (181 M. D. Collins and W, A. Kuperrttan, "Focalizalion. environmental focusing and source localization," J. Acousi. Soc. Am., vol. 90, pp. 1410-1422, Sept, 1991. (19) J. K. Krolik, "Match-field minimum variance beamforming in a random ocean channel.” J. Acousi. Soc. Am., vol. 92. pp. 1408-1419, Sept. 1992. (20) R. L. Culver and W. S. Hodgkiss. Comparison of Kalman and least squares filters for locating autonomous very low frequency acoustic sensors,” IEEE J. Oceanic Eng., vol. OE-13, pp. 282-290, 1988. (21) G. L. D'Spain, W. S. Hodgkiss, and G. L. Edmonds, "Ttie simultaneous measurement of infrasonic acoustic particle velcxity and acoustic pres¬ sure in the ocean by freely drifting Swallow floats,” IEEE J. Oceanic Eng., vol. 16. pp. 195-207, Apr. 1991. (22) G. C. Chen, G. L. D'Spain, W. S. Hcxlgkiss, and G. L. Edmonds, "Freely drifting Swallow float array: July 1989 trip report," MPL TM- 420, Marine Physical Laboratory, Tripps Institution of Oceanography. San Diego, CA, 1990. (23) G. C. Chen and W. S. Hodgkiss, “Localizing Swallow floats during the July 1989 experiment,” MPL TM-421, Marine Physical Laboratory. Scripps Institution of Oceanography, San Diego, CA. 1990. (24) D. H. Johnson. “The application of spectral estimation methrxls to bearing estimation problems.” Proc. IEEE, vol. 70, pp. 1018-1028. Sept. 1982. (25) R. L. Culver. G. L. D'Spain, W. S. Hodgkiss, and G. L. Edmonds. “Estimating 8 kHz pulse travel times and travel times error from Swallow float localization system measurements," Unpublished Notes. Marine Physical l.aboiah'r^ S^iipps Insiiliilion nt Ove.inngraphy. San Diego, CA, 1989 (26| F J Harris, “On the use ol Windows lor h.uiiionic an.ilssis wilh the discrete Fouricr-transforni." I'mc IEEE, vol 66, pp 51 83. 1978 (27| B. D. CarLson. “Covartance matnx estimation errors and diagonal loading in adaptive arrays." IEEE Trans Aerosp. Electron. Svsi. vol. 24. pp. 397-^1, 1988. (28| D F. Gordon and H. P. Bucker. “Arctic acoustic propagation model with ice scattering," NOSC Tech Rep 985. Naval Ocean Systems Center, San Diego. CA. 30 Sept. 1984. (29| F. Ryan. “ATLAS Normal Mode Model," Version 2.19 , Naval Ocean Systems Center, San Diego, CA. (301 J S. Perkins and R. N. Baer, "An approximation to the three-dimensional parabolic-equation method for acoustic propagation." J. Acoustic Soc Am., vol. 72, pp. 515-522, Aug. 1982. (31) G. C. Chen, “VLF source localization with a freely-drifting sensor array," MPL-U-45/92. Marine Physical Laboratory, Scripps Institution of Oceanography, San Diego. CA. 1992. (32) H. Szu and R. Hartley, “Fast simulated annealing," Phys. Lett. A, vol. 122. pp. 157-162, June 1987. (33) S. Kirkpatrick, C. D. Gcllal, and M. P. Vecchi, “Optimization by siroulat^ annealing," Science, vol. 220, pp. 671-680, May 1983. (34) E. R. Davis, “Predictability of sea surface temperature and sea level pressure anomalies over the North Pacific Ocean." J Phvs Ocean, pp. 249-266. May 1976. (35) L, R. LeBlanc and F. H. Middleton. “An underwater acoustic sound velocity data model," J. Acousi. Soc. Am., vol. 67, pp. 2055-2062, June 1980. George C. Chen (M'89) received the M.S. de¬ gree in computer science from the University of California at Los Angeles in 1981 and the Ph D degree in electrical engineering from the University of California at San Diego in 1992. Prior to 1985, he was a software engineer with the Computer Sciences Corporation, El Segundo, CA. and worked on computer networking and database systems. Since 1985 he has been a research engi¬ neer with die Naval Command, Control and Ocean Surveillance Center’s RDTicE Division (formally Naval Ocean Systems Center). San Diego, CA. His research interests include digiul signal processing, communication systems, and neural networks. William S. Hodgkiss (S'68-M'75) received the B.S.E.E. degree from Bucknell University. Lewis- burg, PA, in 1972, and the M S. and Ph D. degrees from Duke University, Durham, NC, in 1973 and 1975, respectively. From 1975 to 1977 he worked with the Naval Ocean Systems Center, San Diego, CA. From 1977 to 1978 he was a faculty member in the Electn cal Engineering Department. Bucknell University. Lewisburg, PA. Since 1978 he has been a member of the faculty of the Scripps Institution of Oceanog¬ raphy and on the staff of the Marine Physical Laboratory, University of California, San Diego. He is the Applied Ocean Science curricular group coordinator. Graduate Department of the Scripps Institution of Oceanography His present research interests are in the areas of adaptive digital signal processing, adaptive array processing, applications of these to underwater acoustics, the propagation of acoustic energy in the ocean and its interaction with the sediments beneath the cxean and the sea surface, and the staiisiic.)l properties of ambient ocean noise. nge and usl. Soc.