Acoustic Source Localization and Cueing from an Aerostat during the NATO SET- 093 Field Experiment

Transcription

1 Acoustic Source Localization and Cueing from an Aerostat during the NATO SET- 93 Field Experiment Christian G. Reiff Army Research Laboratory, 28 Powder Mill Rd., Adelphi, MD, USA ABSTRACT The US Army Research Laboratory has conducted experiments using acoustic sensor arrays suspended below tethered aerostats to detect and localize transient signals from mortars, artillery, and small arms fire. The airborne acoustic sensor array calculates an azimuth and elevation to the originating transient, and immediately cues a collocated imager to capture the remaining activity at the site of the acoustic transient. This single array s vector solution defines a groundintersect region or grid coordinate for threat reporting. Unattended ground sensor (UGS) systems can augment aerostat arrays by providing additional solution vectors from several ground-based acoustic arrays to perform a 3D triangulation on a source location. The aerostat array s advantage over ground systems is that it is not as affected by diffraction and reflection from man-made structures, trees, or terrain, and has direct line-of-sight to most events. Keywords: Acoustic, elevated array, aerostat, localization 1. INTRODUCTION The US Army Research Laboratory (ARL) acoustics and signal processing group has been involved in the development of site protection and persistent threat detection systems using acoustic arrays. Acoustic array systems have proven their capability for omni-directional detection and localization of impulsive acoustic events [1] such as weapons fire or explosions and continuous acoustic events such as engines [2]. The initial development of acoustic detection and localization took place using ground systems and the placement of the acoustic arrays was generally on high ground to decrease terrain effects and scattering of the acoustic signal. The next step was to try elevating the arrays using a small balloon and then an aerostat. The improvement in the signal frequency content and amplitude as well as the ability to localize and cue a camera to an acoustic impulsive source was shown in the results of three field experiments [3][4][5]. The NATO SET-93 experiment provided a chance to experiment with design changes in the aerostat acoustic array and payload as well as providing different terrain and climate conditions from the earlier aerostat experiments. The acoustic array and payload directly were mounted directly to the aerostat envelope to decrease the dramatic payload motion seen in the earlier experiments. The acoustic array was changed to an eight microphone array to allow the comparison of wind noise at three distances from the aerostat and microphone configuration effectiveness. 2. SENSORS AND SENSOR SYSTEMS The sensor systems that provided data for this report were the aerostat array and payload sensors, the ground acoustic array near the winch for the aerostat, four distributed ground acoustic arrays, and the 5 and 15 meter meteorological sensors supplied by ETBS. The aerostat acoustic array (Figures 1 and 2) consisted of 8 Knowles model VEK-323- microphones. Microphones 1-4 were arranged in a tetrahedral identical to the ground array. The tetrahedral portion of the array was inverted with the center microphone (1) below the three microphones forming the base of the tetrahedral (2-4). The tetrahedral was mounted.76 meters below the plane of microphones (5-8) which were arranged in a circle with a diameter of 1.44 meters. The tetrahedral array was offset forward.2 meters from the circle array. The microphones were protected from wind noise with 6 inch foam windscreens. The acoustic data was recorded at 1Hz with 24 bit resolution and stored on a flash drive for later retrieval via LAN connection. The aerostat payload motion was measured with a True North Technologies Revolution GS compass and stored as heading, pitch, and roll at a rate of 1 Hz. A small Garmin GPS 35-HVS measured the payload latitude, Unattended Ground, Sea, and Air Sensor Technologies and Applications XI, edited by Edward M. Carapezza, Proc. of SPIE Vol. 7333, 7333M 29 SPIE CCC code: X/9/$18 doi: / Proc. of SPIE Vol M-1

2 Camera 3D wind sensor Figure 1. ARL Aerostat payload inverted on the ground before mounting to the aerostat with the tetrahedral microphone arms not attached for handling. Figure 2. ARL Aerostat with payload attached. The microphones in the acoustic array can be identified by the wind screens. longitude, and height at a rate of 1 Hz. The GPS Pulse per Second (PPS) signal reference was used in correcting the system timing. This system time was used for all sensor data file time marks. An ultrasonic anemometer was mounted between the tetrahedral array and the ring array about a foot above the system box. This anemometer was a Young model 81V which provided wind azimuth and elevation, velocity, and temperature. The anemometer data was recorded at a rate of 1 Hz to allow the wind turbulence values to be calculated. A Sony SNC-RZ3N ballcam was mounted on the payload to demonstrate cueing capability. The localization solutions were sent to the camera through the network connection. The camera would pan and tilt to the solution azimuth and elevation in about a second and take programmed sequence of images. The camera was normally set to zoom for the first image and back out for a second. These images were stored on 2GB flash drives for later retrieval. The ground acoustic array consisted of four microphones arranged in a tetrahedral array located approximately 1 meter above the ground (Fig. 3). The microphones, wind screens, GPS timing, and data recording parameters are identical to those used on the aerostat payload for simplifying comparisons between data and solutions. The temperature is also measured at the array to calculate the sound speed at the array. Figure 4 shows one of four other similar ground systems Figure 3. ARL Tetrahedral acoustic microphone array and recorder on ground at aerostat site. Figure 4. ARL PXI Tetrahedral acoustic microphone array and seismic template on ground. Proc. of SPIE Vol M-2

3 that were located around the test site. These acoustic arrays are identical to the ground array at the aerostat site and the data is recorded at 2 khz and 24 bit precision. 3. ANALYSIS Most of the analysis will consist of comparing system performance of the aerostat and ground arrays. A comparison of the acoustic signals measured at the elevated and ground arrays from several events will demonstrate the effects of the environment on the propagating acoustic signal. Analysis of how the differences in wind noise and array motion can affect the solution accuracy and detection sensitivity will be done. 3.1 Elevated and Ground Signal Comparison The path a sound takes from its source to the microphone recording it for analysis determines the amount of information that can be inferred about the source. Acoustic localization for our purposes most often occur at over a kilometer from the microphone array which allows the original signal to be attenuated, scattered, refracted, etc. The signal passing above the ground to the aerostat array passes through a less turbulent environment and so retains more of the original signal. The lower frequencies are attenuated by scattering from the turbulent air near the ground as over long distances as seen in Figure 4 and shorter distances with a small arm source as seen if Figure 5. Aerostat Ground.4 81mm tube 81mm tube.2 Pascals Seconds Figure 4. An 81mm launch at 488m recorded by the aerostat array at an elevation of 384m and the ground array below. The retention of the high frequency content of the signal measured at the aerostat array allows the event to stand out in the low frequency wind noise environment common at higher elevations. The noise levels in figure 4 seem to be higher at the ground arrays during this event even though the wind was 7.8 m/s at the aerostat and 3.8 m/s on the ground. The retention of the high frequency content of the original signal allows the detection of small arms, which have their peak frequency content above 15 hertz, at greater ranges. An example of the air and ground detection of a 12.7mm event is shown in Figure 5. The 12.7mm is one of the larger small arms and contains enough low frequency energy to be detected by both systems. The data in Figure 5 was measured at a distance of 1333m from the source and the high frequency loss is already evident in the ground signal. The large amplitude difference indicates the original signal has a lot of frequency content in the range that is attenuated propagating near the ground. There is also an indication of a reflection delayed by Proc. of SPIE Vol M-3

4 around 6ms in the aerostat signal which could be caused by the muzzle being around 1 meter above the ground. This type of information could be used to estimate the height of the shooter but also cause problems with classifiers Pascals Seconds Pascals Seconds Figure 5. The acoustic signal of a 12.7mm shot measured at the aerostat array elevation of 5m and at a distance of 1333m. Another example of reflection is shown in Figure 6. The first impulse in both the aerostat and ground signals is the shock 1 Shock wave 5 Shock wave reflection Muzzle Pascals Shock wave 5 Muzzle Pascals Seconds Figure 6. Acoustic signal of a 155mm shot at a distance of 7756m recorded on the aerostat array at an elevation of 344m and the ground array below. Proc. of SPIE Vol M-4

5 wave from a supersonic artillery shell traveling past the arrays producing a shock cone. The artillery shell was supersonic at charge 5 for this weapon and was not supersonic at charge 3. The solution elevation for the first impulse is positive indicating the path of the artillery shell creating the shockwave passes over the aerostat array. The second impulse in the aerostat signal is the shockwave reflection from the ground and has a negative solution elevation. The ground signal does not show the reflection impulse as expected. The muzzle blast from the 155 shot is the third impulse in the aerostat and second impulse in the ground array signals with a bearing and elevation solution that agrees with the ground truth. The timing and incidence angle of the shock wave across the aerostat array along with the additional information provided by the ground reflection can improve the localization of the source as well as classification. The amplitude of the muzzle blast was larger at the ground array than at the aerostat array. The upward refractive temperature profile seen in the meteorological data for this event normally enhances the elevated signal so the wind must be affecting the propagation. Both the aerostat and ground arrays detected the shell detonation from at a range of 9385m around 5 seconds later with azimuth and elevation solutions that agree with ground truth. The shell detonation amplitude is larger at the aerostat than the ground indicating the wind and the temperature profile aided the signal propagation in that direction. 3.2 Camera Cueing A ball camera mounted on the array captured images when the system detection algorithm produced solutions. Figure 7 shows three representative images captured after impulse solutions with event ground truth locations identified. The first image in the upper left corner is a detonation in an unknown series of detonations that created solutions which cued the Unknown Event at 9:2:13.93 on 6/26 Zoomed In Det Soln Error Camera? ?? Shot time 13:24:37.53 on 6/25 Zoomed Out Type Loc Soln Error Camera 12.7mm S o o x Tube shot time 13:45:57.6 on 6/25 Zoomed Out Type Loc Soln Error Camera 81mm M Detonation time 14:3:52.8 on 6/25 Zoomed In Type Loc Soln Error Camera 12mm M o o x Figure 7. These four images are results of the aerostat camera cued to real time impulse solutions from the localization algorithm. Each solution has an azimuth and elevation error compared to ground truth. The camera delay contributes an additional azimuth and elevation offset. The blue o indicates the center of the image and a red x or arrow indicates the source. Proc. of SPIE Vol M-5

6 camera. The heading in the images of ground truth events shows the type and test location of the source. The image heading also shows the corrected solution azimuth and elevation, the degrees error of the solution related to the ground truth, and the degrees of payload drift in the time the camera takes to pan to the uncorrected solution vector. Because the camera used the uncorrected solution azimuth and elevation, the payload drift while the camera was panning and the actual solution error and not the compass corrections affected the accuracy. The blue o in the images is the center of the image that would be the camera was pointing after it panned to the uncorrected solution vector. The red x or arrow indicates the source of the impulse. The detection algorithm should reject continuous acoustic sources; however, the camera cued to several that occurred during the test. The nearby airport had many planes raising the noise floor with flying, take off, landing, and general engine noise. Images of a plane and a tractor captured from detection of an abrupt change in the continuous signal that would create the broadband impulse characteristics that would cause detection. The plot of the time history of the frequency content of the recorded acoustic signal that includes the detection which cued the camera is below the captured image. The arrows show the time of capture in the plot corresponds to a broadband impulse like signal. The color bar for the plots relate to the FFT magnitude of the signal in Pascals. A tracking algorithm could find continuous sources such as generators and determine changes in atmospheric conditions by monitoring changes in the signal. Tracking also identifies aircraft that are flying too close and warns them of the aerostat and tether. Detection time 11:9:12 on 6/25 Az El 11.1 Detection time 8:31:24 on 6/26 Az 8.81 El I Figure 8. Images captured after detection and cueing on an abrupt change in a continuous sound source from a tractor and a prop plane. Spectrograms below the images have arrows indicating the time of detection. (for color image see electronic version) Proc. of SPIE Vol M-6

7 3.3 Wind Noise Wind speed was measured at 5m, 15m, and aerostat height during the test and was consistently 2 or 3 m/s higher at the aerostat. This wind at the aerostat height is faster but less turbulent than the wind near the ground. The increased turbulence near the ground causes more noise below 25 Hz on the recordings even though the wind speeds are normally lower than at the aerostat. Figure 9 compares the aerostat and ground array sound levels at four different wind speeds at the aerostat while there were no other obvious sound sources. The aerostat and ground environments are compared in the aerostat and ground array plots because the data was recorded at the same time. The sound levels are consistently higher at the aerostat above 25 Hz. The sound levels above 25 Hz stay very close to the same for the three higher wind speeds and drop around 2 db for both aerostat and ground at the lowest wind speed. This decrease in sound levels in the absence of wind seems to be the decrease of micro turbulence that would cause the broadband change. Sound Level in db Wind 1 Wind 1 gnd Sound Level in db Wind 8 Wind 8 gnd Sound Level in db Wind 5 Wind 5 gnd Sound Level in db Wind.2 Wind.2 gnd Frequency Figure 9. The power spectral density (PSD) for Mic 1 on the aerostat array in four different wind speeds is plotted against the PSD for mic 1 on the ground array at the same time. The wind conditions on the ground would normally be 2 to 3 m/s slower. The wind noise at the microphones mounted at 1.6m and 2.6m from the aerostat envelope were very similar at 5m/s wind speed. The microphones mounted just above the ring at.3m have a slightly higher broadband noise level above 75 Hz possibly caused by wind moving faster near the aerostat envelope. The aerostat oriented itself into the wind because of the sail which because of the way the array was mounted positions mics 6 and 7 in the front of the ring and in undisturbed air. This is the case for the tetrahedral mics as well. Mics 5 and 8 on the aft part of the ring usually have mics 6 and 7 and the array frame disturbing the air as the wind blows past which causes an increase in the sound level below 75 Hz. The aerostat tilts away from the wind as the wind speed increases. This increased the pitch to around 25 degrees at when the wind was at 1 m/s. This tilting of the array actually exposed mics 5 and 8 to undisturbed air which removed the increased noise level below 75 Hz. 3.4 Detection and Localization Azimuth and elevation solutions were calculated on the aerostat and aerostat ground system as well as the four PXI ground systems in real time using an executable C algorithm and the results were stored to disk. This data from the four PXI systems was available for use by the SPIDER network. The aerostat communication link with the SPIDER network Proc. of SPIE Vol M-7

8 was never successfully established. The logged results from these algorithms for the June 25 th test day as well as result from Matlab algorithms that processed the stored acoustic data are in Table 1. The bearing error and elevation error are calculated by subtracting the solution value from the ground truth value and preserving the sign. Because the detonation ground truth locations may not be as accurate as the weapon locations, the detonation errors are not included in the table. The mean and standard deviation of the bearing and elevation errors used in the table are helpful in showing an error offset and the consistency of the offset. June 25 8:46:53 to 15:5:3 Aerostat Array Ground Array NI PXI Ground Arrays / Range meters C-4 1Kg C-Detect M-Det 8 M-Det 4 C-Detect M-Detect 2 / / / / 6221 Detected Total 3 Brng Ave Charge Error StdDev n/a Range 2497 Elev Ave n/a n/a n/a n/a n/a n/a Loc c Error StdDev n/a n/a n/a n/a n/a n/a n/a 81mm Mortar C-Detect M-Det 8 M-Det 4 C-Detect M-Detect 2 / / / / 8357 Detected Total 11 Brng Ave Charge 3,6 Error StdDev Range 4773 Elev Ave n/a n/a n/a n/a n/a n/a Loc M1 Error StdDev n/a n/a n/a n/a n/a n/a 12mm Mortar C-Detect M-Det 8 M-Det 4 C-Detect M-Detect 2 / / / / 8358 Detected Total 9 Brng Ave Charge 2,7 Error StdDev Range 4753 Elev Ave n/a n/a n/a n/a n/a n/a Loc M1 Error StdDev n/a n/a n/a n/a n/a n/a 12.7mm small arm C-Detect M-Det 8 M-Det 4 C-Detect M-Detect 2 / / 76 5 / / 2949 Detected Total 1 Brng Ave Charge Error StdDev Range 1251 Elev Ave n/a n/a n/a n/a n/a n/a Loc S1,S2 Error StdDev n/a n/a n/a n/a n/a n/a 7.62 small arm C-Detect M-Det 8 M-Det 4 C-Detect M-Detect 2 / / 76 5 / / 2949 Detected Total 11 Brng Ave n/a n/a Charge Error StdDev n/a n/a n/a Range 114 Elev Ave n/a n/a n/a n/a n/a n/a Loc S1,S2 Error StdDev n/a n/a n/a n/a n/a n/a n/a Total False Detections Table 1. This table is a comparison of detections, solution errors, and false detections for the aerostat and ground arrays on June 23, 28. The solutions for the aerostat and ground arrays are the results from the resident C and Matlab algorithms. Both 4 and 8 microphones are used in calculating the Matlab solutions for the aerostat array. For each type of event, the table lists the total number of that type of event, the number of events detected by each algorithm, and the total number of false detections by each algorithm. The false detections include all detections with no ground truth during the day between the times shown on the table header. The lunch time from 1: to 12: GMT is not included. Other impulses from the test area and multiple detections on a source and included in the false detections. The detection algorithm will reject a trigger the does not meet a required Least Squares Error (LSE) level or a calculated sound speed range. The rejection limits were the same for the Matlab algorithms with the LSE limit being.12. The C algorithms had different LSE levels with the aerostat system being the most sensitive (.5), the aerostat ground system next (.2), and the PXI systems being the least sensitive (.1). The initial analysis of the solution results indicated that the solutions had consistently positive errors that are evident in the results from June 19, 2, 25, and 26 as a positive (clockwise from ground truth) offset from 16 to 2 degrees. Each Proc. of SPIE Vol M-8

9 day mentioned has a consistent error throughout the day of 1, 16, 2, and 7 degrees respectively. The meteorological effects on the propagation normally deflect the bearing only a few degrees which does not account for the total offset. The bearing error for June 23 rd is inconsistent with the other test days by having a negative (counter clockwise) offset for most of the day. The wind was changing direction and speed throughout June 23 rd which may have contributed to the inconsistency during the day. June 25 th had steady wind from around 29 degrees at around 6m/s which held the aerostat stable possibly resulting in the bearing error being comparable to the ground systems. Another possible contribution to a consistent error offset would be that the array shape might change from day to day when stressed by mounting on the aerostat. The aerostat array consisted of two relatively rigid sub arrays. One of which is the 4 mic tetrahedral array which is the same array used for the aerostat ground array. The other is the four mics mounted on the struts holding the system box to the mounting ring. Most of the flexing in the aerostat array would be an offset between the two relatively rigid arrays. A Matlab algorithm similar to the resident C algorithm was written to calculate solutions using the recorded acoustic data from the four mics of the tetrahedral portion of the aerostat array. The same Matlab algorithm processed the acoustic data from all 8 mics in the aerostat array to calculate solutions to compare directly with the 4 mic Matlab algorithm. A comparison of the aerostat and ground array with identical algorithms and arrays is made possible by processing the aerostat ground array data with the Matlab algorithm. Another possibility is that the electronic compass contributed errors for some reason. With the combined contributions to the solution error, the localization analysis is limited. Meteorological corrections could make the situation clearer when applied in the future. The detection rate for the aerostat array was not as good as the aerostat ground array for the larger impulses. The aerostat ground array is not near its range limit for the large low frequency impulses and will detect them. The aerostat array algorithm rejected most of the missed detections because solution LSE and sound speed limits were exceeded and not S/N issues. The small arms events were detected more often by the aerostat array than the aerostat ground array because of the enhanced S/N at the elevated array. The aerostat ground array had many more false detections than the aerostat array during the test because of activity on the ground. The high amount of aerostat ground system false detections on June 26 th were caused by mowing operations in the area. See the tractor in figure 8. The impulses for most events the aerostat did not detect had reasonable S/N ratios and were readily identifiable on the plots and spectrograms of the data. The rejection of detection is usually caused by the LSE or sound speed values exceeding the algorithm limits. A change in mic position caused by the flexing of the array can cause the rejection of a solution if the LSE exceeds the limit. A comparison of he Matlab aerostat 8 mic and 4 mic solutions should indicate increased rejections and decreased accuracy for the 8 mic array if flexing is an issue. The detections for each array during the test show no trend that could identify flexing as an issue. An elevation accuracy comparison shows the 4 mic array to be slightly better overall which is counter to the expectations for the 8 mic array. The azimuth accuracy is roughly the same for both arrays. A modification to the aerostat algorithm and acoustic signal sample rate may be needed to improve the LSE values. The detection algorithm was developed for ground systems that normally receive signals that are low pass filtered by the atmospheric turbulence in the ground layer during propagation. The elevated array receives the higher frequency portion of the signals as well as complicated reflections from the ground. A high sample rate would help the cross correlation portion of the algorithm calculate more accurate time differences between microphones. A sample rate should be used in future testing to allow the determination of the best sample rate. 4 SUMMARY The aerostat acoustic array was able to provide calibrated acoustic data from an 8 microphone array along with cued images, meteorological data, orientation, and position information. The images were captured after panning to a solution vector calculated from the acoustic data. No corrections for payload drift while panning were made for this test but will be implemented in the future. Solutions were corrected for the payload orientation and made available for transmission to the ground via a wireless link where they were to be sent to the SPIDER network. Communications with the payload were established reliably late in the test however the SPIDER node was too far away to connect. The enhanced signal that is expected at an elevated array compared to a ground array was seen in most of the aerostat acoustic data. Some data indicated the presence of an inverted sound speed profile when the elevated signal amplitudes were smaller than the signal amplitudes on the ground. Proc. of SPIE Vol M-9

10 When compared to the ground truth, the aerostat solution was not as accurate as expected. Application of meteorological corrections may improve the understanding of the error sources. A better magnetic compass is a probable solution. The aerostat had no problem detecting events but the detections were sometimes rejected because of high LSE values. The cause of this will be investigated to determine if the algorithm or sample rate needs modification. The elevated array did detect small arms at a greater range than the aerostat ground array because of the propagation path above the ground effects. Wind noise analysis demonstrated the ground arrays have higher noise levels below 25Hz because of increased turbulence in the ground layer. The elevated array had higher noise levels above 25Hz because of higher wind levels. More work can be done on this dataset in the future with the NI array in the farm area, the propane cannon experiment, and the meteorological data. This work will be documented in future papers. REFERENCES [1] [2] [3] [4] [5] [6] Scanlon, M., Tran-Luu, D., Acoustic Mortar Detection System, MSS 23. Pham, T., Sadler, B., Wideband Array Processing Algorithms for Acoustic Tracking of Ground Vehicles, 1 st ARL Sensors and Electron Devices Symposium: College Park, MD, January 1997, pp Reiff, C., Pham, T., Scanlon, M., Noble, J., Acoustic Detection from Aerial Balloon Platform, 23 rd Army Science Conference 24. Reiff, C., Scanlon, M., Noble, J., Acoustic Detection and Localization from a Tethered Aerostat during the NATO TG-53 Test, SPIE Defense & Security Symposium 26 Reiff, C., Scanlon, M., Noble, J., Acoustic Transient Source Localization from an Aerostat, 25 th Army Science Conference 26. Tran-Luu, D., Fractional Time Delay for Direction-of-Arrival Estimation, MSS BAMS 27. Proc. of SPIE Vol M-1