HIGH PERFORMANCE RADAR SIGNAL PROCESSING

Transcription

1 HIGH PERFORMANCE RADAR SIGNAL PROCESSING Justin Haze Advisor: V. Chandrasekar Mentor: Cuong M. Nguyen Colorado State University ECE 401 Senior Design 1

2 Objective Real-time implementation of Radar Data Product Generation and Clutter Filtering Processing for Weather Radar. 3

3 Outline Background Software Design Testing Results Conclusions and Future Work 2

4 Outline Background 2

5 Introduction Abstract: The radar research group at Colorado State University under the direction of Dr. V. Chandrasekar participated in the development of the NASA D3R Radar, a ground validation of the GPM Dual-frequency Precipitation Radar system. Dr.V. Chandrasekar and Cuong Nguyen developed an innovative ground clutter filtering algorithm, Gaussian Model Adaptive Processing in the Time Domain (GMAPTD), that can be applied to the staggered pulse repetition time technique that the D3R radar uses. The goal of this project is to implement this algorithm in real-time and make some other improvements to the D3R signal processing software. Milestones: 1. Implementation of Clutter Filtering for Weather Radar 2. Testing/Verification 3

6 Budget Since this project involves mostly software development, no material cost has been incurred, just cost in time and effort. 3

7 Introduction to Weather Radar Pulse Doppler weather radar radiates a high-power microwave pulse through a directional antenna Duration of each pulse is τ Pulses are separated by an interval T s Received signal is a convolution between transmitted pulse and the volume distribution of the target Radar Transmitter Radar Receiver Distributed Target May 24, 2011 CSU-CHILL Radar Tour, REU '11 7

8 Introduction to Weather Radar A Doppler radar compares the received signal with the frequency of the transmitted signal and measures the frequency shift, giving the speed of the target... 1 Dual-Polarization radar can transmit and receive separate horizontally and vertical polarizations to determine the orientation, shape, and size of weather targets 1. Rinehart, 11 May 24, 2011 CSU-CHILL Radar Tour, REU '11 8

9 D3R Project Developed as part of NASA s Global Precipitation Measurement Project Dual-frequency Dual-polarized Doppler Radar (D3R) Collaborative effort between NASA Goddard Space Flight Center, Remote Sensing Solutions, Inc., Orbital System, Ltd., and Colorado State University Radar Research Group led by Dr. V. Chandrasekar Ground validation of DPR Radar on the GPM Core Observatory Satellite scheduled for launch in Image Courtsey NASA science/ground-validation/ground-instruments 4

10 D3R Specs Ku & Ka frequency bands (13.9 GHz, 35.6 GHz) 150m Resolution 30 Km Maximum Range Solid State Transmitter 90 db Dynamic Range Staggered PRT Transmission 4

11 Staggered PRT Most Radars use uniform transmission, ie 1 period However, radars operating at high frequencies have limited maximum unambiguous velocity that they are able to detect Because of this, the D3R uses a staggered PRT transmission technique sampling time sampled sequence n T u T 1 = 1 T2 = n2tu T = 1 n1t u T2 = n2tu τ +T1 τ +T 1 + T2 τ V stg V ( τ ) ( τ ) V ( + + ) V +T 1 τ T 1 T 2 time This technique uses two alternating periods The maximum unambiguous velocity is measured using the difference of the two periods * λ arg( R( T1 ) R ( T v = 4 π ( T2 T1 ) Velocity Calculation 2 )) v a = λ 4( T 2 T1 ) Maximum Unambiguous Velocity r a = ct 1 2 Maximum Range 4

12 Ground Clutter Weather radars often scan at varying elevations due to the curvature of the earth Low elevation scan can hit trees, building, mountains, etc Most of these objects have close to zero velocity 3. Image Courtesy of Australian Bureau of Meteorology australia/radar/about/what_is_radar.shtml 12 4

13 Clutter Example Velocity PPI Scan from CSU CHILL showing 0 velocity for ground clutter from Rocky Mountains Reflectivity PPI Scan from CSU CHILL showing ground clutter from Rocky Mountains

14 GMAPTD Traditionally, standard IIR and FIR High Pass Filters have been used to remove ground clutter. However, if the radar signal and ground clutter overlap, these filters also remove part of the weather data. Can not be used with staggered PRT. Gaussian Model Adaptive Processing 4 (frequency domain) was developed to remove the ground clutter and recover the lost signal. However, this method has limitations due to spectral leakage and finite data length. It works well with moderate clutter to signal ratio (CSR) but can not be used with staggered PRT. Gaussian Model Adaptive Processing in Time Domain 5 to solve these problems. It runs on time domain data which eliminates problems due to frequency domain processing. It works with staggered PRT. It also allows data to be split into blocks for fast processing. 4. GMAP, Siggia and Passarelli, GMAPTD Nguyen and Chandrasekar,

15 Signal model GMAP-TD filter matrix Clutter time variant filter A R y = = E y = Ax Auto-covariance matrix for the filtered signal { } { H } H yy = E ( Ax)( Ax) H ( 2 ) H H A = A R + σ I A + AR A AR x c N m p Filter solution 2 σ I N m transform clutter to noise A ( 2 R / + ) c σ I = N m 1/ 2 15

16 Example of transforming clutter to noise Original power spectrum Power (db) clutter echo precipitation echo -20 After applying GMAP-TD filter Power (db) Velocity (m/s) (a) precipitation echo Velocity (m/s) (b) 16

17 Signal interpolation procedure in time domain x / ˆ R x Filter A Rˆ y + + Σ Estimate spectral moments from cov. matrix Pˆ P, vˆ P, wˆ P Gaussian Model R P Filtering cov. matrix + - Σ H AR P A Iterate until ΔP p is less than the threshold. Cuong Nguyen 17

18 GMAP-TD algorithm for Staggered PRT 2/3 will be implemented on D3R radar Input Digital radar signals x H (H) and x V (V) Compute sample auto-cov. matrix (H) R ˆ = x ( h ) x H H Estimate clutter power and noise floor ( 1 ) F(x H ) P xx h f 1 (P xxh ) σ N(h) 2 f 1 (P xxh ) P c(h) H x Compute clutter Gaussian covariance matrix model R c from P c(h) and pre-set clutter width w c Note: ( 1 ) f 1 : dynamic noise floor estimation ( 2 ) R c (0.4) and R c (0.8) are shifted versions of R c at 0.4v a and 0.8v a.(v a is the unambiguous velocity) Cuong Nguyen Compute filter matrix A (0.4) ( 0.4) R c + Rc + Rc + 2 A = / σ + (0.8) ( 0.8) N I Rc + Rc (for R (0.4) c..., see note ( 2 )) 1 1/ 2

19 1 Compute filtered auto-cov. matrix Rˆ = ( 0) ˆ y AR ( h) x A H diag Compute NCP from (0) ( ) ( ˆ (0) Rˆ,0 and diag R,1) y y NO NCP > threshold? Return parameters NO Overlap echo exist? YES YES set n=0 Recover the overlap echo 3 Cuong Nguyen

20 3 Estimate signal spectral moments from the filtered auto-cov. matrix for n ˆ ( n) ˆ ( n) P diag R,0 wˆ p vˆ ( y ) ( ˆ ( ) diag R y,1) ( ˆ ( n),1) and ( ˆ ( ) R diag R,2) ( n) n p ( n) n p diag y y Compute signal Gaussian model matrix from estimated spectral moments R p (n) Update filtered auto-cov. matrix n n n h Rˆ ( + 1) ( ) ( ) ( ) = Rˆ + R ARˆ A y y p x H Estimate signal spectral moments from the filtered auto-cov. matrix for (n+1) ˆ ( n+ 1) ˆ ( n+ 1) P diag R,0 wˆ p vˆ ( y ) ( ˆ ( n 1) diag R y,1) ( ˆ ( n+ 1),1) and ( ˆ ( n 1) R diag R,2) ( n+ 1) + p ( n+ 1) + p diag y y n:=n+1 NO v P ( n+ 1) p ( n+ 1) p v P ( n) p ( n) p and/or < threshold YES Cuong Nguyen Return parameters ˆ ( n+ 1) ( n+ 1) = P ; vˆ = vˆ ; wˆ = wˆ ˆ ( n+ 1) p p p p p p P

21 Outline Background Software Design Testing Results Conclusions and Future Work 2

22 Software Design Overview Challenge: Run GMAPTD for each ray in 64ms (real time). A ray is a collection of 400 range gates. A range gate is one 150m scan volume. Each range gate contains 128 samples. This means that each gate needs to be processing in 0.16ms or 160µs. Each gates goes through a GMAPTD clutter filtering loop. Written in C Fast Many libraries and tools available in Linux Maintains some compatibility with existing real-time moment processing code Reads time series data from the network Contains infinite loop for real-time processing Uses GNU Scientific Library for Linear Algebra computations, easy to manage memory for any size matrix or vector, methods for fast matrix and vector computations 6

23 Outline Background Software Design Testing Results Conclusions and Future Work 2

24 Time Series Streaming Code We can not test this software on the D3R, because the D3R is expensive! We don t want to break it. Set up network environment to simulate the D3R Signal Processing. The D3R has already collected thousands of file of data. The time series streaming code reads one of these file and streams it over the network. The GMAPTD moment processing code receives this data over the network as if it was running in real time 7

25 Misc Testing Track memory usage with System Monitor and Valgrind Track change history using Git 8

26 Outline Background Software Design Testing Results Conclusions and Future Work 2

27 Results I have finished writing the program in C. However, I am still trying to debug the output of the program. I have a Matlab program I am using to check the output. Initial testing indicates that GMAPTD program is not running in real-time. It is running in approximately 67ms, really close to the needed 64ms! Still need to add multithreading. If I split the processing into 4 threads. Each thread will process approximately 100 gates. This will allow 4 threads to be processed simultaneously on a 4 core SMT CPU. This should make the program approximately 4 times faster 17ms per ray. Therefore we should be able to beat the goal of 64ms. 9

28 Outline Background Software Design Testing Results Conclusions and Future Work 2

29 Summary and Future Work Summary: GMAPTD is an algorithm developed for ground clutter filtering of weather data The iterative GMAPTD algorithm is being implemented in C to run in real time on the D3R Radar Future Work: Add dual-polarization GMAPTD processing More debugging Multithreading 35

30 References 4. A. D. Siggia, A. D., Passarelli, R. E. Jr., Gaussian model adaptive processing (GMAP) for improved ground clutter cancellation and moment calculation, Cuong M. Nguyen and V. Chandrasekar, Time domain GMAP clutter filter for weather radars, Proc. 34th Conf. on Radar Meteorology, Amer. Meteor. Soc., Oct , Williamsburg, VA. 2. D3R, NASA, Accessed Nov. 30, How Radar Works, Australian Bureau of Meteorology, Accessed December 5, Rinehart, Ronald E., Radar For Meteorologists 4 ed. Rinehart Publications. 2004, p 11 Simultaneous multithreading, Intel, Accessed Dec. 5,

31 Libraries GNU Scientific Library (GSL), GNU General Public Licence (GPL), ISO/IEC 9899:TC3 C99 Standard, International Organization for Standardization, International Electrotechnical Commission, Language Standards Supported by GCC, GNU Operating System, GNU General Public Licence, The Single Unix Specification, Version 2, The Open Group, 1997, 35

32 Acknowledgements I would like to express my gratitude to Dr. Chandra for creating this senior design project and giving me the opportunity to expand my skills in signal processing, high performance computation, and radar systems. Thanks to Cuong for being patient with me when I am learning new things and helping me on all aspects of the project. I could not have accomplished much with this help. Thanks to Olivera Notaros for managing the senior design program and helping me to prepare for my future career. 35

33 Thank you! Any Questions?