HSTAMIDS with Acoustic Vibration Sensing 14 February 2006

Transcription

1 HSTAMIDS with Acoustic Vibration Sensing 14 February 2006 US Army CE-LCMC Acquisition Center - Washington 2461 Eisenhower Avenue Alexandria, VA

2 1.0 Program Overview CyTerra Corporation has developed and integrated a wideband radar vibrometer mode into the Handheld Standoff Mine Detection System (HSTAMIDS). Modeled after a radar vibrometer, this device uses the mine detector s ground penetrating radar (GPR) to measure and classify vibration signatures of objects buried within a few inches of the ground surface that are induced by an acoustic transducer placed in contact with the ground. Responding to vibration signatures unique to mines, the vibrometer mode adds a confirmation capability to the standard HSTAMIDS detection modes, providing discrimination between mines and clutter and thereby improving probability of detection and false alarm performance. Analysis of data collected using a trained classification algorithm shows that mines can be distinguished from clutter better than 85% of the time, while maintaining the HSTAMIDS initial detection at 100%, and hence validates the feasibility of implementing a handheld radar vibrometer within the HSTAMIDS platform. Additional data collection on different mine, clutter, and soil types is required to further improve the robustness of the classification algorithm and to assess its capabilities with high confidence. A photograph of the HSTAMIDS vibrometer is shown in Figure 1. It is comprised of a modified HSTAMIDS unit, laptop PC for data collection and acoustic waveform generation, audio power amplifier, and acoustic transducer. The acoustic waveform is comprised of a series of 1 or 4-second pulses, linearly swept in frequency from 50 Hz to 1000 Hz over the duration of each pulse. It is created in MATLAB, stored as a WAV-file, and played in Windows Media Player. The power amplifier and transducer, are standard commercial equipment used for vibration testing. Figure 1 HSTAMIDS Vibrometer. 2.0 Mode of Operation 2

3 Vibrometer mode processing, including data acquisition, signal processing, and classification, can take upwards of 20 seconds to complete per target, depending upon the waveform selected. As such, it is envisioned that the vibrometer will be employed in a confirmation role after the primary detection has been made using standard HSTAMIDS mine detection modes. Under this scenario, the system operates as follows: 1. The transducer is placed on the ground several feet from the target of interest. 2. The PC is set to start generating the acoustic waveform. 3. The acoustic waveform is amplified in the power amplifier and provides the drive signal to the transducer. 4. The HSTAMIDS radar head is placed over the target of interest and the PC is set to collect data. Data from several acoustic pulses is collected, which for 4-second pulses can take upwards of 20 seconds. 5. Signal and classification processing take place off-line using algorithms developed in MATLAB. The sound pulses launched by the transducer travel along the surface of the ground and under the radar antenna head. As the pulses pass over objects buried at or near the surface, they excite vibration modes within these objects that propagate up to the ground surface and perturb the radar signal. Mines, made up of complex structures including hollow cavities, firing mechanisms, and pressure plates, produce unique resonant signatures when exposed to acoustic energy. These signatures are sufficiently different than those observed from common buried clutter such as metal fragments, wood, and rocks, and thus provide a useful feature set for target/clutter discrimination algorithms. 3.0 Summary of Hardware Changes Peak mine and ground surface displacements resulting from vibration are very small, on the order of 10-8 to 10-6 meters, and create narrowband phase modulation on the radar signal with a power level more than times smaller than the static ground return. In order to measure these low-level signals, modifications to the HSTAMIDS radar transceiver were required. The HSTAMIDS receiver function was replaced with two assemblies: an externally mounted direct down-conversion receiver and an internally mounted 4-channel analog-to-digital converter circuit card assembly (A/D CCA). The direct down-conversion receiver improves sensitivity by lowering system noise 40 db. The 4-channel A/D CCA offers 20 db more dynamic range than its HSTAMIDS counterpart by providing separate processing paths for the static ground return signal and the low-level phase modulation. Together, these modifications provide for the concurrent linear processing of ground and phase modulation signals with increased signal-to-noise ratio (SNR). The direct down-conversion receiver along with it improved noise floor performance are shown in Figure 2. 3

4 Figure 2 New Direct Down-Conversion Receiver Improves Noise Floor. The two modulation sidebands 100 Hz from the carrier are the result of observing a large base speaker being driven by a 100 Hz sine wave. Observe that the sideband SNR improves by more than 40 db in the new receiver. The lower noise floor also reveals 60 Hz modulation sidebands and a continuum of harmonics. These are caused by the fluorescent lighting used in the laboratory. When a fluorescent light is on, the gas within is ionized and becomes a strong reflector of radar energy. The lights in the laboratory flash on and off in sync with the 60 Hz AC power grid presenting an effect to the radar much like the light reflecting off a signaling mirror. Fortunately, there are not many fluorescent lights in the typical minefield! The 4-channel A/D CCA employs filters to separate the phase modulation from the static ground return signal and amplification to boost the signal levels into the operating range of the A/D converters. In-phase and quadrature (IQ) data from the direct down-conversion receiver are first processed in a low-pass filter that performs the channel select function: both static ground return and phase modulation are passed through while out-of-band interference is rejected. The data is filtered again, this time by a high-pass filter to strip out the large static ground return allowing for 50 db additional gain to boost the level of the phase modulation prior to A/D conversion. This process is depicted in Figure 3. 4

5 Figure 3 Four-Channel A/D CCA Separates and Amplifies Ground Return and Phase Modulation Sidebands. Two minor changes were also made to the HSTAMIDS programmable transmit waveform. The transmitter is capable of generating frequencies in the range 1120 MHz to 2510 MHz in 10 MHz steps, but for vibrometer operation, a fixed frequency of 2000 MHz was selected. While the specific frequency is not particularly important, the fact that it does not change over the data acquisition time is. A constant frequency provides the phase coherence necessary to detect and process the phase modulation signal. Additionally, the transmit power was increased to +15 dbm (32 mw) to further improve the target SNR. A summary of significant hardware parameters is provided in Table 1. 5

6 Table 1 Hardware Parameters. Parameter Frequency Source Frequency Synthesis Transmit Power Transmit Frequency Receiver Type Receiver Gain Low Pass Channel Band Pass Channel Receiver Bandwidth RF Low Pass Video (I&Q) Band Pass Video Receiver Dynamic Range A/D Size A/D Sample Rate Acoustic Waveform Swept Frequency Range Pulse Width Signal Processing Gain Value 70 MHz TCXO Indirect, Phase Locked Loop (PLL) +15 dbm 2 GHz ( GHz is available) Direct Down-Conversion (Homodyne) 47 db 97 db 1 GHz ( GHz) 2 khz (0 2 khz) 1.96 khz (40 Hz 2 khz) 108 db 16 Bits Hz Linear Frequency Modulation (LFM) 50 Hz 1000 Hz 1 sec & 4 sec 30 db (1 sec pulse) 36 db (4 sec pulse) 4.0 Software Algorithms The software is divided into two distinct algorithms: signal processing and classification. Both are currently implemented in MATLAB and run offline (i.e. non-real time) on data collected with the PC. The signal processing algorithm is modeled after the algorithm developed at GA Tech. The primary functions include calibration, extraction of phase modulation and pulse compression. A block diagram of the signal processing algorithm is shown in Figure 4. Figure 4 Vibrometer Signal Processing Algorithm. 6

7 Calibration corrects for the non-ideal characteristics of the transceiver hardware including DC offset at the mixer output, IQ gain and phase imbalance, and frequency drift between the radar oscillator and the acoustic source oscillator. Offline calibration procedures determine correction coefficients that are applied to the IQ data as it enters the signal processing. Of the three corrections, the DC offset and frequency drift corrections are most significant. Left uncorrected, DC offset adds error to the computed ground displacement values, and frequency drift degrades pulse compression performance and affects an overall reduction in sensitivity. The LPF & Decimate block reduces the data sample rate, consistent with the Nyquist requirement, reducing both data throughput and processing load. The 20 Hz digital low pass filter (LPF) provides additional smoothing of the static ground return signal to support calculation of the carrier magnitude and phase angle. The 20 Hz cutoff allows the algorithm to follow carrier phase angle changes caused by operator motion without tracking out the vibration-induced phase modulation. Extraction of the phase modulation from the band pass data is achieved using the calculated ground return signal magnitude and phase angle. Since the phase modulation is narrowband, its signal vector is oriented 90 degrees away from the ground return vector. Thus, by applying a vector rotation to the band pass data equal in magnitude to the angle of the ground return signal, the phase modulation is moved entirely to the quadrature channel. The phase modulation data extracted from the quadrature channel and subsequently normalized by the magnitude of the ground return signal yields a real signal with amplitude units of radians. The phase modulation is then processed in a matched filter to implement pulse compression and maximize SNR. The one and four-second acoustic pulses received by the radar are compressed to a width of approximately 1 millisecond, affecting a processing gain of 30 db and 36 db respectively. Eighty milliseconds of compressed output data surrounding the peak of each compressed pulse are output to the classification algorithm. A sample of the signal processing output is shown in Figure 5. The response from clear ground is represented by the solid line and the response from a VS-2.2 anti-tank mine buried 3.25 inches deep is represented by the dashed line. Notable are the increased peak displacement over the mine and the resonance displacement seen only from the mine. Displacement (nanometers) Mine resonance displacement. Time (seconds) Figure 5 Signal Processing Output for VS-2.2 Mine. 7

8 The classification algorithm uses a quadratic classifier trained on n features from two classes of targets: mines and clutter. (The specific n features will be described shortly.) The equation for the quadratic classifier is given by where r r T 1 r r q = ( x μ ) K ( x μ ) + ln( K i i q i is the statistical distance from the target under test to the i th apriori trained class, x r r is the ( 1 n) vector of feature values for the target under test, μ is the ( 1 n) vector of mean feature values for the i th apriori trained class, and i i K is the ( n n) covariance matrix for the i th apriori trained class. i A graphical representation of this process with two trained classes and two feature values is shown in Figure 6. i i ) Figure 6 Graphical Representation of Quadratic Classifier The two ellipses identify the one-standard deviation (1σ) boundaries of the feature values for each class. The target under test, X, has feature values that place it within the 1σ boundary of Class #1. Evaluation of the quadratic classifier equation for the two classes yields statistical distances q 1 and q 2. With q 1 < q2 as shown, the algorithm classifies the target as belonging to Class #1. The short time response of the mine resonance shown in Figure 5 suggests the use of a discrete wavelet transform to provide the n features that characterize each target class. Based upon analysis of data collected, the first 8 coefficients of a 64-point discrete Haar wavelet transform were selected as the feature set for this classifier. A block diagram of the classification algorithm is shown in Figure 7. 8

9 Figure 7 Target Classification Algorithm. Classification begins with extraction of 64 data samples starting with the peak output from each compressed pulse. This corresponds to 18 milliseconds of data and encompasses the majority of mine resonance signatures seen to date. The 64-point Haar wavelet transform is calculated and the first 8 wavelet coefficients are extracted to form the feature vector. The feature vector is applied to the quadratic classifier equations for mine and clutter and the one giving the shortest statistical distance, Q min e or Q clutter, is selected as the target class. 5.0 Data Collection and Analysis A data collection was performed on May 10 through May 12, The goal was to collect acoustic vibration data from a variety of mine and clutter types to support the evaluation of hardware and algorithm performance, development and training of the target classifier, and ultimately, to assess the feasibility of implementing a handheld vibrometer mode within the HSTAMIDS. Figure 8 shows several pictures of the data collection effort in progress. The small wagon carried the laptop PC used to collect radar data and generate the acoustic waveform, the power amplifier that drives the acoustic transducer, and a car battery and power inverter to run the amplifier. 9

10 Figure 8 Data Collection at Ft. AP Hill in May Acoustic vibration data was collected from a variety of mine and clutter types: PMD-6, VS-2.2, M19, M14, TS-50, VS-50, VAL-69, splayed round, clay brick, miscellaneous metal, and irregular plastic. At each target location (mine or clutter), multiple data files were recorded to explore variations in the process including aspect angle, acoustic transducer type (large and small), and pulse length (1 second and 4 seconds). Data was collected both with the radar head over the target and again with the radar head over blank ground for comparison. Analysis of the processed data revealed that the vibration signatures are very consistent from pulseto-pulse and are not greatly affected by the physical orientation of the operator and acoustic source relative to the target location. Further, little difference was observed in the target signatures between data collected with the 1-second acoustic pulse and the 4-second acoustic pulse. However, while all of the data collected using the large transducer (the one shown in Figure 8) exhibited strong 10

11 vibration responses, none of the data collected using the small transducer were viable. It is believed that the small shaker may have been over-driven and breaking contact with the ground surface as a result of its much smaller mass. Under these conditions, sound transfer to the ground is very inefficient with little energy reaching the target. Several samples of the processed vibration data are shown in Figure 9 through Figure 12. Observe the following characteristics: Consistency in vibration response from pulse-to-pulse. Larger displacement when over the mine than when over blank ground. Large secondary peaks when over the mine that are not present when over blank ground or clutter. Little difference between 1-second and 4-second acoustic pulse data. These characteristics are fairly consistent across all the data collected and provide the basis for the feature-based classification approach. 11

12 Figure 9 VS-2.2 Mine Vibration Signature. 12

13 Figure 10 TS-50 Mine Vibration Signature. 13

14 Figure 11 Metal Clutter Mine Vibration Signature. 14

15 Figure 12 TS-50 Mine Vibration Signatures Using 1 and 4 Second Acoustic Pulses. 15

16 To derive features for the classifier, the target data was separated into the two chosen classes: mine and clutter. For each target, the pulse compressed output from the signal processing was normalized by the peak amplitude and 64 samples starting at the peak (see Figure 13) were extracted and processed by the Haar wavelet transform. The first 8 wavelet coefficients constitute the feature vector for that target. 64 samples comprise about 20 milliseconds of data. Figure 13 Data Selected for Haar Wavelet Transform. Classifier training consisted of calculating the ( 1 8) mean feature vector r μ and the ( 8 8) feature covariance matrix K for the two target classes. Finally, the classifier performance was evaluated by testing each target feature vector against the trained classifier. Testing was done two ways to determine the upper and lower performance bounds. First, each target feature vector under test was also included in the training set. This is known as resubstitution and establishes the upper bound on performance. And second, each target feature vector under test was not included in the training set. This is known as leave-one-out and establishes the lower bound on performance. The test results are reported in the form of a confusion matrix. With two classes, the confusion matrix is of size ( 2 2) and defined as follows: Selected Class Mine Clutter Actual Target Mine Clutter Probability of calling mine a mine. Probability of calling clutter a mine. Probability of calling mine a clutter. Probability of calling clutter a clutter. Figure 14 Confusion Matrix Definition. 16

17 The upper left and lower right matrix elements give the probabilities of correct classification while the upper right and lower left elements give the probabilities of incorrect classification. The overall probability of correct classification, P cc, is defined as the average of the individual probabilities of correct classification. Classifier testing was performed using 35 feature vectors from mines and 18 feature vectors from clutter objects. The confusion matrices and probabilities of correct classification for the resubstitution and leave-one-out test methods are as follows: Resubstitution Selected Class Leave-One-Out Selected Class Mine Clutter Mine Clutter Actual Target Mine Clutter 0 1 Actual Target Mine Clutter P = 0.96 = cc P cc Lower and upper bounds on the probability of correct classification are thus 0.86 and 0.96 respectively. In addition, note from the leave-one-out data that the declaration of mine as mine is maintained at 100% with a reduction in false alarms (calling clutter a mine) by nearly a factor of four. Although these results are based upon a limited set of 53 feature vectors, the demonstrated ability to correctly distinguish mines from clutter better than 85% of the time is quite promising and gives validity to the overall approach. Additional data collection on different mine, clutter, and soil types is required to further improve the robustness of the classification algorithm and to assess its capabilities with high confidence. 5.0 Conclusion and Suggestions for Future Efforts This project has demonstrated the feasibility of implementing a vibrometer mode within the HSTAMIDS platform. The key enablers to this capability are the modifications to the receiver that increase sensitivity by reducing system noise and increasing the instantaneous dynamic range. Signal processing and a quadratic classifier were exercised on mine and clutter data collected, demonstrating better than 85% probability of correct classification. As modifications to the hardware and software were implemented in stages over the course of the project, they are not fully integrated within the HSTAMIDS architecture. The direct downconversion receiver is mounted externally to the electronics unit and the 4-channel A/D CCA replaces the standard HSTAMIDS receiver CCA thereby rendering inoperable the standard mine detection modes. Software is written in MATLAB and runs as post-processing on data collected 17

18 with a PC. Naturally, the next step would be to fully integrate the vibrometer mode allowing the operator to seamlessly switch between acoustic vibration detection and standard mine detection modes operating in real time. The band of frequencies around 2 GHz is well suited for integration. With the widespread development of products to support the cellular communications industry, low-cost, highlyintegrated RF components are in abundance. Options for integration include: (1) modification of the existing HSTAMIDS receiver CCA to support both receiver architectures, and (2) complete redesign of the receiver to incorporate only the direct down-conversion architecture. The increased sensitivity of the later option may in fact improve the standard mine detection modes. Fully integrated hardware will also improve the efficiency of data collection and analysis. Extensive data collection is still required to further develop the capabilities of the classifier. This includes data from different mine and clutter objects in varied soil conditions. Currently, the classifier is trained on only two classes of targets: mines and clutter. It may be advantageous to create sub-classes, e.g. AT mines, AP mines, metal clutter, plastic clutter, etc. With this approach, a signature unique to AT mines, for example, will more strongly affect the training of that class and improve overall classification performance. In summary, acoustic vibration detection using a modified HSTAMIDS has been successfully demonstrated. Future efforts should include complete integration of the mode within the HSTAMIDS architecture and extensive data collection to further develop the robustness of the classifier. 18

19 Appendix A Radar Vibrometry Overview A short overview of radar vibrometry is given here to assist in understanding the hardware modifications made to the HSTAMIDS receiver and signal processing algorithm. A radar is able to measure small displacements in target range, such as those induced by vibration, by taking advantage of the doppler effect: a change in target range causes a corresponding change in the round-trip phase of the radar signal. This round trip phase, as measured by the phase detector in the radar receiver, is given by: 4π f RF R( t) ϕ( t) = (radians) c where ϕ(t) is the measured phase angle as a function of time, f RF is the RF transmit frequency in hertz, R(t) is the target range as a function of time, and c is the velocity of light. Sinusoidal vibration with peak amplitude r P will produce sinusoidal phase modulation (PM) with peak phase deviation 4π f RF rp β = (radians). c As the ground vibration peak displacement is quite small, the resulting narrowband PM will have peak amplitude with respect to the carrier. ( β ) μ = 20 log (dbc) This PM signal is the part of the received radar signal that contains the ground vibration displacement information and must be extracted from the carrier by the signal processing function. For this to happen properly, it is required that the radar receiver has sufficient dynamic range to linearly process both the carrier and the much smaller PM signals simultaneously. The low end of peak ground displacement values reported by Dr. Scott is on the order of tens of nanometers. At 30 nm peak displacement and 2 GHz radar transmit frequency (HSTAMIDS vibrometer frequency), the peak modulation amplitude is 4π μ 30 nm) = 20log = 112 (dbc). 9 9 ( 2 10 )( ) ( Assuming 30 db minimum signal processing gain and 20 db minimum signal-to-noise ratio (SNR) required for reliable detection and classification, the instantaneous dynamic range required for the HSTAMIDS radar receiver is Required Instantaneous Dynamic Range = = 102 (db). 19

20 This exceeds the available dynamic range of the current HSTAMIDS receiver and required adding a second receive channel to the hardware: one to process the carrier and one to process the PM. It is possible to reduce the required dynamic range by increasing the transmit frequency. In the above equations, it can be seen that the required dynamic range is inversely proportional to the peak phase deviation β, which is in turn directly proportional to the radar transmit frequency f RF. Each doubling of the transmit frequency reduces the required dynamic range by 6 db. In fact, the GA Tech radar transmit frequency is 8 GHz or a factor of four higher than HSTAMIDS. Thus, the dynamic range required by the GA Tech radar is 12 db less than HSTAMIDS. While a smaller dynamic range is certainly attractive, an 8 GHz transmit frequency can not be readily generated by the current HSTAMIDS transmitter. On the other hand, the lower HSTAMIDS frequency has advantages of its own. For one, the lower frequency achieves better penetration through surface vegetation and soil, making it possible to detect vibration features not only off the ground surface, but also directly off the mine structure. Also, operating within this frequency band allows the hardware designer to leverage the wide variety of low-cost, highly-integrated RF components made available by the cellular phone industry. Extraction of the PM from the radar signal is a simple matter given the narrowband nature of the modulation. Referring to Figure 15 below, the PM vector with amplitude a0ϕ ( t) is in quadrature to the carrier vector with amplitude a 0. And, in accordance with the small angle approximation, the amplitude of the PM signal with respect to the carrier is a0ϕ ( t) = ϕ( t), a 0 i.e. the length of the PM vector relative to the carrier is equal to the magnitude of the phase modulation angle ϕ(t) in radians. This is the same as β, the peak phase deviation for sinusoidal PM discussed earlier. Im a 0 ϕ(t) (PM signal) ϕ(t) a0 (carrier) θ (carrier angle) Re Figure 15 PM Phasor Diagram By applying a vector rotation to the radar signal equal in magnitude to the angle of the carrier signal θ, the carrier becomes aligned with the real axis and the phase modulation becomes aligned with the quadrature axis. The phase modulation data can then be extracted from the quadrature data, and when normalized by the magnitude of the carrier, yields a real signal with amplitude units of radians. 20