0 ratings0% found this document useful (0 votes) 226 views175 pagesDOA Calculation
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content,
claim it here.
Available Formats
Download as PDF or read online on Scribd
DIRECTION OF ARRIVAL ESTIMATION USING
ANTENNA ARRAYS
by
Rias Muhamed
Thesis submitted to the Faculty of the
Virginia Polytechnic Institute and State University
in partial fulfillment of the requirements for the degree of
MASTER OF SCIENCE
in
Electrical Engineering
oN :
\ ?
OW
VS) :
Dr. Theodore S. Rappaport, Chairman
. Jeffrey H. Reed Dr. Brian D. Woermer
January 1996
Blacksburg, Virginia
Keywords: Direction of Arrival. DOA, Adaptive ArraysAbstract
RRIVAL ESTIMATION USING
ANTENNA ARRAYS
by
Rias Muhamed
Theodore S. Rappaport, Chairman
Bradley Department of Electrical Engineering
ABSTRACT
The objective of this thesis is to design, build and evaluate a direction-of-arrival
measurement system using an antenna array operating at 2050 MHz. The first phase of the
project is an extensive study of various high resolution DOA estimation algorithms. The
algorithms studied include subspace-based techniques such as the MUSIC (MUItiple
Signal Classification) and ESPRIT (Estimation of Signal Parameters via Rotational
Invariance) algorithm, and the integrated approaches which combine property restoral-
based techniques such as the Iterative Least Squares Projection-based Constant Modulus
Algorithm (ILSP-CMA) with subspace-based techniques. All these algorithms were
simulated in MATLAB and their performance under different conditions were tested and
compared. In the second phase of the project a six element uniformly spaced linear array
receiver was built. Three Ariel DSP-96 boards based on the Motorola DSP96002 are used
to simultaneously sample, collect, and store data from each of the array elements. The data
collected by the DSP boards are processed off-line to estimate the directions of arrival.
Several experiments are conducted to test the functioning of the system under various
conditions, and demonstrate the performance of different algorithms. Results of these
experiments are also presented in this thesis.Acknowledgements
Acknowledgments
I would like to thank my academic advisor, Dr. Theodore S. Rappaport, for the invaluable
support, guidance and encouragement he has provided over the past two years. I thank my
committee members, Dr. Jeffrey H. Reed and Dr. Brain D. Woerer, for carefully
reviewing my thesis report and providing useful suggestions.
1am grateful to Joe Liberti for his help and guidance during the course of this research. I
sincerely appreciate the time and effort he had spent in getting me up to speed on this
project. I also thank Tom Beidka for his valuable advice and suggestions.
This project could not have been completed without the help of several students at MPRG.
Mansoor Ahmed modeled the measurement site in the site specific propagation prediction
tool. Keith Blankenship and Bill Newhall provided the power delay profiles taken at the
measurement site. Zhigang Rong, Mansoor Ahmed, Francis Dominique, Paul Petrus, Nitin
Mangalvade, Kevin Saldanha, and Anil Doradla assisted me during the measurement
campaign. I wish to thank each one of them for the help they provided.
{wish to thank Francis Dominique and Paul Petrus for providing several ideas and
suggestions throughout the course of this work.
1 would like to thank Prab Koushik for providing the necessary support for my
computational demands.
I would like to thank the Advanced Research Project Agency’s (ARPA) Global Mobile
(GLOMO), the Office of the National Drug Control Policy (ONDCP), and the MPRG
industrial affiliates for providing financial support for this work.
Most of all, | would like to thank my parents and family without whose continuous and
immeasurable support I could not have completed my degree
itiTable of Contents
Contents
1. Introduction 1
1d Motivation ..... i
1.2. Objective and Outline of Thesis, : o2
2. Adaptive Arrays - Background Theory 5
2.1 Fundamentals of Antenna Arrays......
2.1.1 Uniformly Spaced Linear Array
2.1.2 Spatial Signature. te
2.2. Beamforming and Spatial filtering
2.3. Wideband Arrays
2.4 Adaptive Arrays ..
2.5 The LMS Algorithm mn
26 Property Restoral Algorithms
2.6.1 The Constant Modulus Algorithm .
6.2 Spectral Coherence Restoral Algorithm
2.7
3. Overview of Direction of Arrival Estimation Algorithms 29
3.1 Introduction. ...... 06.606 c ee cece e eee ee ene eee +29
3.2. Conventional Methods for DOA Estimation.
3.2.1 Delay-and-Sum Method .
3.2.2 Capon’s Minimum Variance Method -
33 Subspace Methods for DOA Estimation.
7 1 The MUSIC Algorithm. .
2 Improvements to the MUSIC Algorithm -
3 3.3. The ESPRIT Algorithm .
3.4 Maximum Likelihood Techniques 49
3.5 DOA Estimation under Coherent Signal Condi ions
3.5.1 Spatial Smoothing Techniques . .
3.5.2. Multidimensional MUSIC
3.6 The Iterative Least-Squares Projection Bas
3.7 The Integrated Approach to DOA Estimation ...
38 Detection of Number of Sources .
3.8.1 | The SH, MDL and AIC Criteria .
3.8.2. Order Estimation Using Transforme
3.9 Summary 2.0... cece eee eeesTable of Contents
4. Simulation Based Performance Comparison of DOA Estimation
Algorithms 2
4A Introduction... 000... 0 0c eee cece eeeeeeeeeeeeeeeee 72
4.2 Resolution and Sensitivity to Calibration of MUSIC type algorithms. . . 73
4.3 Performance Comparison of DOA Estimation Algorithms . n
4.4 Performance of Source Order Estimation Algorithms . .
4.5 Conclusion
5. Description of the DOA Measurement System 100
5.1 Introduction. . - 100
52 Antenna Array and RF front-end . - 102
5.3. Sampling and Data Collection . 103
5.3.1 Configuring Multiple DSP-96 boards for 5 Symehronows Seeing
5.4. RF Chain Phase Calibration . = 108
5.5 Receiver Sensitivity and Dynamic Range. 110
5.6 Processing for DOA Estimation . 112
6. DOA Measurement Experiments and Results.
6.1 Imroduction. ......
6.2 Description of the DOA Measurements .
6.2.1 Measurements with a Single Transmitter Antenna. .
6.2.2 Measurements With Multiple Transmitter Antennas Fed From
Uncorrelated Sources . 7
6.2.3 Measurements With Multipie Transmitter “Antennas Fed Froma
Single Source ............ 126
6.3 Sources of Error in DOA Estimation . . - 130
6.4 Results of the Simulation Based Study. - 137
6.5 Conclusion APRS - 146
7, Conclusion and Future Work ........0.0.eeeeeeeeee se eeenees 146
71 Conclusion . 146
7.2. Future Work .. - 148
Appendix 1 List of Abbreviations ..........6.0eeeeeeee ween 148
References ....... . veeee 149
Programs seeeeee 155List of Figure
Figure 2.1: Illustration of an array with antenna elements located at arbitrary locations in
space... 26
Figure 2.2: Illustration of a plane wave incident on a linear equispaced array. The dotted
lines represent the phase fronts of the incident wave.............-22++ . eee
Figure 2.3: Direct and multipath signals from multiple sources impinging on a receiver an-
tenna array. aa oe a Tiotfatd2
Figure 2.4: A narrowband beamformer whose output is a complex weighted sum of the
spatially sampled input data......0000000.cceeeeeeeeeeeeeeeeeeeeeenes vee 5
Figure 2.5: Illustration of a wideband array with a transversal filter attached to each anten-
na element... 17
Figure 2.6: Illustration ofa simple adaptive array.......002220000000eeeeeee 219
Figure 2.7: An example of a two dimensional quadratic performance surface. ........ 20
Figure 3.1: Illustration of the classical beamforming structure. .
Figure 3.2: Comparison of resolution performance of delay-and-sum method and Capon’s
minimum variance method. Two signals of equal power at an SNR of 20 dB arrive at a 6-
element uniformly spaced array with an interelement spacing equal to half a wavelength at
angles 90 and 100 degrees respectively........
Figure 3.3: Comparison of MUSIC and Capon’s minimum variance method. Two signals
of equal power at an SNR of 20 dB arrive at a 6-element uniformly spaced array with an
interelement spacing equal to ha\fa wavelength at angles 90 and 95 degrees respectively.40
Figure 3.4: Illustration of ESPRIT array geometry [Roy90]. 145
Figure 3.5: Comparison of MUSIC with and without forward/backward averaging in co-
herent multipath. Three coherent signals of equal power at an SNR of 20 dB arrive at a 6-
vielement uniformly spaced array with an interelement spacing equal to half a wavelength at
angles 60, 90 and 120 degrees respectively...........222000000005 ; 22.57
Figure 3.6: Example of spatial spectrum estimated using ILSP-CMA for spatial signature
estimation, followed by MUSIC with forward backward averaging. The six element array
is able to resolve 24 direct and multipath components and associate each component to the
appropriate signal (user).......2.22.2-. Tiatclotedatat ct tateltatal balsa claliatalalt 64
Figure 4.1: Resolution performance of MUSIC and the integrated ILSP-CMA with MU-
SIC approaches. The resolution is plotted here as the angle Dq such that two equal powered
signals impinging on the six element uniformly spaced linear array at 90 and 90 + Dg re-
spectively were consistently resolved. Resolution is plotted as function of both SNR and
number of samples of data used for DOA estimation, for the case of both coherent and un
correlated signals. For the coherent cases, forward/conjugate backward averaging was
used BreToer 274
Figure 4.2: Sensitivity of the MUSIC spectrum to spacing errors. Four noncoherent signals
arrive at angles 10,30,50,70 degrees, respectively, on a six element uniform linear array
with half wavelength interelement spacing. This figure illustrates how the MUSIC spec-
trum becomes corrupted due to imperfections in constructing the array. The first figure is
without any spacing errors, and the subsequent ones are with +/- 1, 2, and 5 percent spacing
errors respectively. ............ Pde dete tat dat totaled wine 76
Figure 4.3: Performance of various DOA estimation algorithms. RMS error in DOA esti-
mation is plotted as a function of number of samples of data used to estimate the DOA. In
all simulations, two uncorrelated (non-coherent) signals with equal power are incident on
a six element uniform linear array at 90 and 95 degrees, respectively, at a 204B SNR. The
RMS error is averaged over 200 repeated DOA estimation trials... 79
Figure 4.4: RMS Error in DOA Estimation as a function of number of samples. Simulation
parameters same as Figure 4.3, except that the SNR here is only SB. ............ 79
Figure 4.5: Performance of various DOA estimation algorithms. RMS error in DOA esti-
mation is plotted as a function of signal-to-noise ratio. In all simulations, 50 samples of two
uncorrelated (non-coherent) signals with equal power incident on a six element uniform
linear array at 90 and 95 degrees respectively are used for DOA estimation. The RMS error
is averaged over 200 repeated DOA estimation trials. ........ ale Reen een 80
Figure 4.6: Simulation parameters same as in Figure 4.5 with the two sources being coher-
ent. .. 80
Figure 4.7: Performance of various DOA estimation algorithms. RMS error in DOA esti-
mation is plotted as a function of number of samples of data used to estimate DOA. In all
simulations, two coherent signals with equal power are incident on a six element uniform
linear array at 90 and 95 degrees respectively, at a 20dB SNR. The RMS error is averaged
over 200 repeated DOA estimation trials, All subspace based algorithms are used along
vii81
with forward/backward averaging.
Figure 4.8: RMS Error in DOA Estimation as a function of number of samples. Simulation
parameters same as figure 4.7, except that the SNR is SB. 0.2.0... 0... cece eee 81
Figure 4.9: RMS Error in DOA estimation as a function of the angle of arrival. A single
sinusoidal tone signal is incident on a six element uniformly spaced linear array at 20 dB
SNR. 500 samples of data is used for each DOA estimation and the rms error is obtained
by averaging over 200 repeated trials. 00... 00.5 cceeev cece eeee nese ee eee 83
Figure 4.10: RMS Error in DOA estimation as a function of angle of arrival. Simulation
parameters same as in Figure 4.9, except that forward/backward averaging (spatial smooth-
ing) is used. ............. Set et se taletslet it tet teats atdtttettedat 84
Figure 4.11: RMS error in DOA estimation as a function of angle of separation (Dg) and
number of sources for TLS-ESPRIT and the integrated ILSP-CMA with TLS-ESPRIT al-
gorithms. All sources are assumed to be uncorrelated (non-coherent) and at the same power
level with an SNR = 20 dB. The estimations are made using 500 samples of data and the
rms error is computed by averaging over 200 repeated trials. The top figure shows the case
where the incident signals arrive at 10, 10+A8, 10 +2A0, and 10+ 340 degrees, and the bot-
tom figure shows the case where the incident signals arrive at 30, 30+A0, 30+2A0 and
30+3A0 degrees. 85
Figure 4.12: RMS error in DOA estimation as a function of the angle of separation be-
tween adjacent signals. The top figure shows the case where the incident signals arrive at
60, 60+A0, 60 +2A0, and 60+ 3A0 degrees, and the bottom figure shows the case where the
incident signals arrive at 90, 90+A, 90+2A0 and 90+3A0 degree: ; 86
Figure 4.13: RMS error in DOA estimation as a function of the angle of separation be-
tween adjacent signals and the number of signals. Simulation parameters are the same as
used in Figure 4.11, except that the signals are perfectly correlated as would be the case in
a coherent multipath environment. Forward/backward averaging is used to decorrelate the
coherent signals. .......2..2.4. ae 2.87
Figure 4.14: RMS error in DOA estimation as a function of angle of separation between
adjacent signals and the number of signals. Simulation parameters are the same as used in
Figure 4.12, except that the signals are perfectly correlated as would be the case in a coher-
ent multipath environment. Forward/backward averaging is used to decorrelate the coher-
ent signals. . bectecee eee eee es 88
Figure 4.15: RMS error in DOA estimation as a function of the difference in power be-
tween the uncorrelated (non-coherent) signals. All the sources except the first one is as-
sumed to be at the same power level. The estimations are made using 500 samples of data
and the rms error is computed by averaging over 200 repeated computations, .......90
Figure 4.16: RMS error in DOA estimation as a function of the difference in power be-
tween the incident coherent signals. All the sources except the first one is assumed to be at
villthe same power level. .........ccccceeeeeeeeeeeeeeeeeseeeeeeeseeeees cece OL
Figure 4.17: RMS error in DOA estimation as a function of angle of arrival for different
spacing errors. Spacing errors were simulated by perturbing the location of the sensors. All
the sources incident on the array are uncorrelated with one another. The estimations were
made using 500 samples of data and the rms error was computed by averaging over 200
repeated trials. . 7 92
Figure 4.18: RMS error in DOA estimation as a function of the angle of arrival for differ-
ent spacing errors for the case of coherent (perfectly correlated) signals incident on the ar-
ray. Spatial smoothing through forward/backward averaging is used to decorrelate the
signals. The estimations are made using 500 samples of data and the rms error is computed
by averaging over 200 repeated computations. 93
jure 4.19: Comparison of MDL.AIC, GMDL, GAIC detection for a six element uniform
linear array with equal powered uncorrelated sources incident 10 degrees apart from each
other with the first source at 90 degrees, second source at 100 degrees, and so on. The prob-
ability of detection error computed as a relative frequency over 200 trials is plotted as a
function of the number of samples used in the detection scheme. The signal-to-noise ratio
2.95
‘omparison of MDL,AIC, GMDL, GAIC detection for a six element uniform
linear array with uncorrelated sources incident 10 degrees apart from each other with the
first source at 90 degrees, second source at 100 degrees, and so on. The probability of de-
tection error computed as a relative frequency over 200 trials is plotted as a function of the
signal-to-noise ratio. 100 samples of the input data was used to estimate the DOA... 96
Figure 4.21: Comparison of the modified MDL,AIC detection for a six element uniform
linear atray with coherent sources of equal power incident 10 degrees apart from each other
with the first source at 90 degrees, second source at 100 degrees, and so on. The probability
‘of detection error computed as a relative frequency over 200 trials is plotted as a function
of number of samples of data used. The input signal-to-noise ratio was 10 dB. ...... 97
Figure 4.22: Comparison of MDL,AIC detection for a six element uniform linear array
with coherent sources incident 10 degrees apart from each other with the first source at 90
degrees, second source at 100 degrees, and so on. The probability of detection error com-
puted as a relative frequency over 200 trials is plotted as a function of the signal-to-noise
MBO toatl altel at lata latdaleletalatetalaledaddalaladala teh deelctdly vecee eee
Figure 5.1: A block diagram of the DOA measurement system. .............-. 101
Figure 5.2: Photograph of the six element uniformly spaced linear array receiver used for
DOA measurements... 102
Figure 5.3: Flowchart illustrating the frame synchronization process. ........... 106
Figure 5.4: Hardware setup for the RF chain phase calibration ...............-5 109Figure 5.5: Flowchart illustrating the off-line processing steps... 2S
Figure 5.6: Flowchart illustrating the TLS-ESPRIT algorithm for DOA estimation 114
Figure 5.7: Flow chart illustrating the integrated [LSP-CMA with TLS-ESPRIT algorithm
for DOA estimation...............0 0.0000 e eee tees tenet eens sees ANS
Figure 6.1: DOA Measurement Results - RMS Error in DOA Estimation as a function of
the angle of arrival. RMS error is computed as the absolute difference between the estimat-
ed and actual angle of arrival 120
Figure 6.2: Illustration of a the DOA measurement setup with three uncorrelated signals
being transmitted at different angles with respect to the receiver array location. .... 121
Figure 6.3: DOA Measurement Resullts - RMS Error in DOA estimation as a function of
angle of separation between the two uncorrelated sources. The first signal arrives at 10 de-
grees and the second signal arrives at 10+A@ degrees. ...2..2....00:eceseeeeereeeee 122
Figure 6.4: DOA Measurement Results - RMS Error in DOA estimation as a function of
angle of separation between the two uncorrelated signals. The first signal arrives at 30 de-
grees and the second signal arrives at 30+A0 degrees. +123
Figure 6.5: DOA Measurement Results - RMS Error in DOA estimation as a function of
angle of separation between adjacent uncorrelated signals. The first signal arrives at 30 de-
grees, the second signal arrives at 30+A0 degrees and the third signal arrives at 30+2A0 de-
BOS ec e eee ee eee ee eec tenes eeeteeeeetieeteeeeenetteeteees 123
Figure 6.6: DOA Measurement Results - RMS Error in DOA estimation as a function of
angle of separation between the two uncorrelated signals. The first signal arrives at 60 de-
grees and the second signal arrives at 60+A0 degrees. ...........:0c000eeeereeeee esl 24
Figure 6.7: DOA Measurement Results - RMS Error in DOA estimation as a function of
angle of separation between adjacent uncorrelated signals. The first signal arrives at 60 de-
grees, the second signal arrives at 60+A0 degrees, and the third signal arrives at 60+2A9
degrees. ........0 Telatetalatadahcldadtahatatdfotedalataledadaledadaded elated 124
Figure 6.8: DOA Measurement Results - RMS Error in DOA estimation as a function of
angle of separation between the two uncorrelated signals. The first signal arrives at 90 de-
grees and the second signal arrives at 90+A0 degrees.
Figure 6.9: DOA Measurement Results - RMS Error in DOA estimation as a function of
angle of separation between adjacent uncorrelated signals. The first signal arrives at 90 de-
grees, the second signal arrives at 90+A@ degrees, and the third signal arrives at 90+2A0
egret tre oa oft tattle ba dled atte edalattate leat 125
Figure 6.10: Illustration of a the DOA measurement setup with three correlated (fully co-
herent) signals being transmitted at different angles with respect to the receiver array loca-
xFigure 6.11: DOA Measurement Results - RMS Error in DOA estimation as a function of
angle of separation between the two fully correlated (coherent) signals. The first signal ar-
rives at 30 degrees and the second signal arrives at 30+A0 degrees. ... 128
Figure 6.12: DOA Measurement Results - RMS Error in DOA estimation as a function of
angle of separation between the two fully correlated (coherent) signals. The first signal ar-
rives at 60 degrees and the second signal arrives at 60+A0 degrees. 128
Figure 6.13: DOA Measurement Results - RMS Error in DOA estimation as a function of
angle of separation between the two fully correlated (coherent) signals. The first signal ar-
rives at 60 degrees, the second signal arrives at 60+A8 degrees, and the third signal arrives
at 60+2A0 degrees... eee veeeeeeeceeeeeeeeeseeesetereeeeeteee een es 129
Figure 6.14: DOA Measurement Results- RMS Error in DOA estimation as a function of
angle of separation between the two fully correlated (coherent) signals. The first signal ar-
rives at 90 degrees and the second signal arrives at 90+A0 degrees. 2.129
Figure 6.15: DOA Measurement Results - RMS Error in DOA estimation as a function of
angle of separation between the two fully correlated (coherent) signals. The first signal ar-
rives at 90 degrees, the second signal arrives at 90+A0 degrees, and the third signal arrives
at {0F2A0 degrees. 02... eee eee c eee eceeeeeeeeeeeeeeeee este eter erties 130
Figure 6.16: Figure showing sample power delay profiles corresponding to two transmitter
locations. The first profile was obtained when the transmitter was located at 20 degrees
with respect to the receiver, and the subsequent ones were measured when it was located
at 120, 90, and 150 degrees respectively. 0... 00. .0.ccceeceeeeeeeeeesetenees 134
Figure 6.17:Figure showing the single bounce multipath rays incident on the array when a
single transmitter antenna is located at 20 degrees with respect to the receiver array axis.
Using ray tracing it was found that for all transmitter locations, there were at least three
multipath components that were about 15 to 20 dB below the line-of-sight component. 136
Figure 6.18: Simulation Results - Figure illustrates the RMS error in estimating the DOA
of a single source under RF chain phase errors, sensor location errors, and | multipath in
channel... cece ee eee eeeeceeeeecseeeseeeeeeeeeeeeeeeaes Tdeteledelele 137
Figure 6.19: Simulation Results -Figure illustrates the RMS error in estimating the DOA
of two uncorrelated signals arriving at 10 and 10+A8 degrees respectively, under RF chain
phase errors, sensor location errors, and multipath in channel. ...........0000000085 138
Figure 6.20: Simulation Results - Figure illustrates the RMS error in estimating the DOA
of two uncorrelated signals arriving at 30 and 30+A0 degrees respectively, under RF chain
phase errors, sensor location errors, and multipath in channel. 140
Figure 6.21: Simulation Results - Figure illustrates the RMS error in estimating the DOA
of three uncorrelated signals arriving at 30, 30+A0, and 30+2A0 degrees respectively, under
RF chain phase errors, sensor location errors, and multipath in channel...............140
xiFigure 6.22: Simulation Results - Figure illustrates the RMS error in estimating the DOA
of two uncorrelated signals arriving at 60 and 60+A0 degrees respectively. under RF chain
phase errors, sensor location errors, and multipath in channel. .........0....000266 0141
Figure 6.23: Simulation Results - Figure illustrates the RMS error in estimating the DOA
of three uncorrelated signals arriving at 60, 60+A0, and 60+2A8 degrees respectively, under
RF chain phase errors, sensor location errors, and multipath in channel. ..............141
Figure 6.24: Simulation Results - Figure illustrates the RMS error in estimating the DOA
of two uncorrelated signals arriving at 90 and 90+A0 degrees respectively, under RF chain
phase errors, sensor location errors, and multipath in channel....... 142
Figure 6.25: Simulation Results -Figure illustrates the RMS error in estimating the DOA
of three uncorrelated signals arriving at 90, 90+0, and 90+2A8 degrees respectively, under
RF chain phase errors, sensor location errors, and multipath in channel.............. 142
Figure 6.26: Simulation Results -Figure illustrates the RMS error in estimating the DOA.
of two coherent signals arriving at 30 and 30+A0 degrees respectively, under RF chain
phase errors, sensor location errors, and multipath in channel. .........2000+e¢00+++ 0143
Figure 6.27: Simulation Results -Figure illustrates the RMS error in estimating the DOA
of two coherent signals arriving at 60 and 60+A@ degrees respectively, under RF chain
phase errors, sensor location errors, and multipath in channel. ......222002000000000. 144
Figure 6.28: Simulation Results -Figure illustrates the RMS error in estimating the DOA
of three coherent signals arriving at 60, 60+A0, and 60+2A0 degrees respectively, under RF
chain phase errors, sensor location errors, and multipath in channel. .............-.-144
Figure 6.29: Simulation Results -Figure illustrates the RMS error in estimating the DOA
of two coherent signals arriving at 90 and 90+A0 degrees respectively, under RF chain
phase errors, sensor location errors, and multipath in channel. 2145
Figure 6.30: Simulation Results -Figure illustrates the RMS error in estimating the DOA
of three coherent signals arriving at 90, 90+A0, and 90+2A0 degrees respectively, under RF
chain phase errors, sensor location errors, and multipath in channel. .........- 145
xiiChapter 1: Introduction
Chapter 1
Introduction
1.1 Motivation
Direction of arrival (DOA) estimation or direction finding has been an active area of
research for a long time. Historically, direction finding techniques have found application
in the field of radar, sonar, electronic surveillance and seismic exploration. In radar appli-
cations, they are useful for air traffic control and target acquisition{Hay85]. Intelligence
agencies use them for covert location of transmitters and signal interception. Direction
finding also finds application in position location and tracking systems.
More recently, direction of arrival estimation has become important in mobile radio com-
munications [Mat89]. For example, it is useful in determining the multipath structure of
radio channels. Channel models incorporate the angle of arrival statistics along with the
time of arrival statistics to more accurately characterize multipath radio channels [Lib96].
Such characterization is especially important in analyzing communication systems which
employ some form of spatial filtering using sectorized directional antennas, switched
beam antennas, or adaptive antennas[Lib95].Chapter 1: Introduction
Estimation of angle of arrival of multiple signals and their multipath components is impor
tant in systems employing adaptive antenna arrays for signal extraction in interference
environments. Adaptive antenna systems are capable of automatically forming beams in
the directions of the desired signal and steering nulls in the direction of the interfering sig-
nals. Direction finding techniques can be used to estimate the directions of the desired and
interfering signals, so that they can be separated using appropriate spatial filtering. DOA
based beamforming techniques are particularly suited for wireless communication sys-
tems which utilize different frequencies for uplink and downlink transmissions. Due to the
difference in frequencies, the beamforming weights for the smart uplinks cannot be
directly mapped to the downlink. However, since the directions of arrival of a signal does
not change as rapidly as the spatial signature of the signal does, DOA based beamforming
offers some advantages over spatial signature based beamforming [Big95][Xu94C}.
With the FCC mandating that the Enhanced 911 (E-911) services, which include location
identification, be made available to mobile users, there is a great deal of research interest
in developing reliable position location techniques [B1294]. One technique which offers
great promise is that based on direction finding algorithms [Ken95][TR45].
1.2 Objective and Outline of Thesis
The aim of this project is to build a direction-of-arrival measurement system using an
antenna array operating in the 2050 MHz band and verify the performance of some high
resolution DOA estimation algorithms.
The first phase of the project is an extensive study of various high resolution DOA estima-
tion algorithms. The algorithms studied include subspace-based techniques such as the
MUSIC (MUltiple Signal Classification) and ESPRIT (Estimation of Signal Parameters
via Rotational Invariance) algorithms, and the integrated approaches which combine prop-
erty restoral-based techniques such as the Iterative Least Squares Projection-based Con-
2Chapter 1: Introduction
stant Modulus Algorithm (ILSP-CMA) with the subspace-based techniques. All these
algorithms were simulated in MATLAB and their performance under different conditions
were tested and compared. These algorithms were evaluated for their accuracy in DOA
estimation, ability to resolve closely spaced signals, performance under various angle
spread and power spread, and performance under coherent signal conditions.
In the second phase of the project a six element uniformly spaced linear array receiver was
built. Three Ariel DSP-96 boards based on the Motorola DSP96002 are used to simulta-
neously sample, collect, and store data from each of the array elements. The data collected
by the DSP boards are processed off-line to estimate the directions of arrival, Several
experiments are conducted to test the functioning of the system under various conditions,
and demonstrate the performance of different algorithms. The results of these experiments
are presented in this thesis.
This thesis is organized as follows. Chapter 2 introduces the fundamentals of antenna
arrays, and some basic terminology and notations. It also provides an overview of some
adaptive algorithms. Chapter 3 provides a detailed survey of various direction-of-arrival
estimation algorithms. The algorithms detailed include the conventional beamforming
type of algorithms, the high resolution subspace based algorithms, maximum likelihood
algorithms, and the integrated approaches which combine property restoral techniques
with the subspace based techniques. Spatial smoothing techniques that are necessary for
direction finding in coherent multipath are also described. Chapter 3 also provides a brief
survey of source order estimation algorithms.
Chapter 4 presents the results of various simulations that were conducted to evaluate the
performance of various direction-of-arrival and source order estimation algorithms. A
detailed comparison of the performance of these algorithms under different conditions is
made in this chapter.Chapter 1: Introduction
Chapter 5 provides a detailed description of the hardware and software developed for
direction-of-arrival measurements at 2050 MHz. Chapter 6 describes the experiments con-
ducted to perform direction-of-arrival measurements, and presents the results of the mea-
surements. Chapter 6 also presents a simulation based study of the effect of various errors
in DOA estimation.
‘A brief summary and conclusion is provided in Chapter 7 along with some suggestions for
future work.Chapter 2: Antenna Arrays - Background Theory
Chapter 2
Antenna Arrays - Background Theory
2.1 Fundamentals of Antenna Arrays
An array antenna or an array is a configuration of multiple antenna elements arranged
and interconnected in space to obtain a directional radiation pattern. Arrays built using
small antenna elements achieve the same level of performance as that of a single large
antenna, by trading the electrical problems of feeding for mechanical simplicity. It is
possible, using arrays, to electronically scan the main beam and/or place nulls in any
direction by changing the phase of the exciting currents in each of the antenna elements.
An array which does this is called a phased array [Stu8 1]
Arrays can be constructed in various types of geometric configurations. A linear array is
the most elementary form of arrangement in which the centers of the elements of the
array are aligned along a straight line. A planar array is one in which the centers of the
array elements lie on a single plane. Planar arrays could be circular, rectangular or
arbitrarily shaped. Arrays whose element locations conform to a given nonplanar surface
are called conformal arrays.
The radiation pattern of an array is determined by the radiation pattern of the individual
5Chapter 2: Antenna Arrays - Background Theory
elements, their orientation and relative positions in space, and the amplitude and phase of
the feeding currents. If each element of the array is an isotropic point source, then the
radiation pattern of the array will depend solely on the geometry and feeding current of the
array, and the radiation pattern so obtained is called the array factor. If each of the
elements of the array are similar but non-isotropic, by the principle of pattern
multiplication, the radiation pattern can be computed as a product of the array factor and
the individual element pattern [Stu81]
v 1
(py pzy)
plane wavefront
\\.\. incident ray
4
(xa¥az4)
Cive7)
Figure 2.1: Illustration of an array with antenna elements located at arbitrary locations in space.
Figure 2.1 shows an array of M elements distributed arbitrarily in space. Consider a
narrowband signal 3(1) having a frequency @o, and arriving at an angle @ and 6 with
respect to the z-axis and x-axis, respectively. The narrowband signal may be expressed asChapter 2: Antenna Arrays - Background Theory
S(t) = u(t)cos(@pt + v(2)) (2.1)
where u(i) and v(t) are slowly varying functions of time that define the amplitude and
phase, respectively, of 3(r). In the array context, slowly varying implies that the
amplitude and phase variations as functions of spatial position for fixed ¢ are negligible
over the extent of the array. That is, the approximation
5(t—1) = u(t) C08 { @p(t—t) + v(t) } (2.2)
is valid for the all time t required by the signal to travel over the extent of the array. It is
convenient to express the narrowband signal in complex envelope representation, in which
5(0) = Re{s(t)}, where
s(t) = u(dexp{—jl@ot + ()]} (2.3)
Assuming that the propagation medium does not significantly effect the signal as it
propagates from one end to the array to the other, the signal received at one sensor differs
from that of the other only by a delay. As seen from Figure 2.1, the delay depends on the
relative position of the sensors and on the angle of arrival. If we take the origin of the
rectangular coordinate system as the reference location, and the ith antenna element to be
located at (x},yj,z)), the delay 1, of the signal at the ith sensor relative to the signal at the
reference location can be expressed as
__Ly,sin0cos¢ + y,sinO sing + z,cos0]
c
(2.4)
ty
where c is the velocity of light. However, since the signal is a narrowband complex signal
(analytic signal), the effect of the propagation delay 1, is simply a phase shift ¢; = —agt,,
ie,
s(t—t)) = s(exp(7G;) = s(t)exp(—/@ot,) (2.5)
where the phase shift ¢, is given by [Col69]Chapter 2: Antenna Arrays - Background Theory
2(~sindcos@ + y;sinOsing + z,cos0) (2.6)
22(, sinOcosd + y,sinOsing + z,cos®)
and 2 is the wavelength of the incident signal. Now, if the signal received at sensor
(antenna element) locations /,2,...,M are designated x ,x>, . . ., X\ Tespectively, the analytic
signal received at the array can be expressed in vector form as
xO] fe
xy = [2] = Jel scn (2.7)
md) | Jou
The vector x(t) is often referred to as the array input data vector or the illumination vector.
In equation (2.7) the phase shifts are assumed to be solely due to the spatial separation
between the array elements. In a more general sense, the array elements will themselves
have a directional and frequency dependent response. This can be modeled by applying
differing gains and phases to the elements of the vector in (2.7). If the direction and
frequency dependent gain and phase of the ith antenna element is denoted by g,(@, 8,6) .
the analytic signal at the array output can be expressed as
16 \(@, 8, 6)
xo] | £1008.
50 JEq(@. 8, 4)
x(r) = [72 | = | (0,8, )e = a(@,0,6)8(t) (2.8)
bal) Tylor 86)
glo. 0, )e
where the vectorChapter 2: Antenna Arrays - Background Theory
15,(0.9,9)]
81(0.0. be
iEy(0.0, 6)
a(, 6,6) = | g3(0.0,d)e (2.9)
JEyplo- 0,0)
Suh. 8. )e
i ©
j2(x; sinOcos6 +»; sindsing + =,c0s9)
g1(0.8. Oe
jap sindcosd+ ysinOsing +25c0s0)
870.8, )e © ~
J 2(x sin cos + yyy sinBsind +z 440058)
1g. 0, )e ©
is called the steering vector. Equation (2.9) represents the general form of the steering
vector for an array. As seen from (2.9), the steering vector (also called direction vector,
array vector ot aperture vector) is a function of the individual element response, the array
geometry, signal frequency, and the angle-of-arrival. The collection of array vectors for all
angles (6, 6) and frequencies « is referred to as the array manifold. (Some authors in the
literature refer to the individual steering vectors themselves as the array manifold).
Though for many simple arrays the array manifold can be computed analytically, in
practice, the array manifold is measured as point source respons
s of the array at various
angles and frequencies. This process of obtaining the array manifold is called array
calibration.
If the frequency band of interest is sufficiently narrow, the steering vector @/,0,6) is
approximately constant with respect to @ over the band of interest for all angles (0, 6).
This condition is satisfied, if the sensor characteristics do not vary significantly across this
bandwidth. Such an array is called a narrowband array. In most of the discussion that
follows, the array is assumed to be narrowband unless specified otherwise, and forChapter 2: Antenna Arrays - Background Theory
simplicity the steering vectors of narrowband arrays are denoted by a/0,$) with the
frequency dependence dropped.
The input data model of equation (2.8) can be extended to the more general case of
multiple signals and interference impinging on the array by decomposing the input signal
in the frequency domain and using linear supersposition.
>
x(0) = Y alo, 8; ))s() + mo) (2.10)
Using matrix notation, this can be represented as
5\(@)
X(@) = |a(c, 0,01) » a(0, 8p. op)}| .. | #A(@) (2.11)
Sp())
or
x(@) = A(o, 0, ®)s() + n(@) (2.12)
where D signals s,().....Sp(@) arrive from angles (8;.4,).-...(8p. dp)
respectively and m(«) represents the interference and noise components. If the frequency
band of interest is sufficiently narrow to categorize the array as a narrowband array, the
dependence on can be dropped and the array data can be modeled in the time domain as
the analytic signal
D
x(t) = Ya, 6;)s,(1) + (0) (2.13)
ist
which can be written using matrix notation as,
x(1) = A(O)s(t) + a(t) (2.14)
Equation (2.14) represents the most commonly used narrowband input data model for
analyzing data obtained from an antenna array.Chapter 2: Antenna Arrays - Background Theory
2.1.1 Uniformly Spaced Linear Array
incident plane wave
Figure 2.2: Illustration of a plane wave incident on a linear equispaced array. The dotted lines
represent the phase fronts of the incident wave.
For the case of the uniformly spaced linear array lying along the z-axis as shown in Figure
2.2, the steering vector can be computed by simplifying equation (2.9) as follows. Let the
first antenna element be located at the reference location, i.e (x), ¥;,2;) = (0, 0,0). The
incoming plane wave at element 2 travels a distance longer by an amount equal to deos0
with respect to the wave arriving at the first element, and hence arrives later. Using (2.6),
the corresponding phase lag of the plane wave arriving at element 2 with respect to that at
element 1 is 6=Bdeos0, where B is the phase propagation factor given by 2n/A. By the
same argument, C;=B2dcos®, C4=B3dcos®, and so on. Hence, using equation (2.9) the
steering vector for the uniformly spaced linear array comprising of isotropic antenna
elements can be written as
I
Fa
aoy=| or (2.15)
eiBlM~ 1)de0s0
iChapter 2: Antenna Arrays - Background Theory
Note that for the case of the linear array, the steering vector is independent of the elevation
angle 4. Further, the steering vector of the uniformly spaced linear array of isotropic
:
elements is a Vandermonde vector (vector having the form {y 4 4293 .. !M]_), and
[Lad ‘|
this fact is exploited by many array-based direction of arrival estimation algorithms
2.1.2 Spatial Signature
\-—mn.
Source | \ Source 2
si) multipath
sx)
multipath direct path
x 1 ae
1 2 --—- M2
Receiver Antenna Array
Figure 2
array.
Direct and multipath signals from multiple sources impinging on a receiver antenna
Consider an M-element antenna array receiving signals from users located at different
positions. The signals impinging on the array may contain both direct path and multipath
components of multiple sources, and each of these components may be from different
directions of arrival. Following the narrowband data model discussed in section 2.1, the
array input data vector corresponding to one source, say the kth source, may be written asChapter 2: Antenna Arrays - Background Theory
N mp
X= AO), + LY ca(O/)5y = ae pS (2.16)
1
where Npy-/ is the total number of multipath signals (excluding direct path). c; is the
amplitude and phase difference between the /th multipath and the direct path, a/®) is the
array steering vector corresponding to the angle of arrival 8, and sy is the first (line-of-
sight) component of the signal from the kth source. The vector a... defined by
N
5,4 = (01) + Y) ea(O)) (2.17)
i=
is called the spatial signature of the signal s. Note that the above model assumes that all
the multipath components arrive within a time window which is much less than the inverse
bandwidth of the signal so that the multipath components differ only in phase shift, and
not in delay. This of course follows directly from the narrowband assumption.
Now, if there are D sources present, then the array input data vector can be expressed as
D
x(k) = Yo ag, (k)s/(k) + n(k) (2.18)
or equivalently,
X(k) = A(O)s(k) + a(k) (2.19)
where A(®) is the spatial signature matrix and s(k) is the Dx 1 signal vector
comprising of signals from D different sources.
The correlation between the signals impinging on the array has a significant effect on the
performance of algorithms used to process the received signal. For example, many
direction-of-arvival estimation algorithms will fail completely when there are two or more
signals that are fully correlated. The correlation coefficient may be used to quantify the
degree of correlation between signals. Two zero mean signals s; and s; are said to be
uncorrelated or noncoherent if their correlation coefficient defined byChapter 2: Antenna Arrays - Background Theory
(2.20)
with values of |p| in between
is equal to zero, and fully correlated ot coherent if |p| =
zero and unity indicating partial correlation (Assuming ergodicity, the expectation
operation may be replaced by a time averaging operation in the above definition). For
example, multipath signals which are attenuated and phase shifted versions of one another
are coherent signals, and signals from two different users are in general uncorrelated.
2.2 Beamforming and Spatial filtering
An antenna array essentially provides a means of sampling the received signal in space.
By processing the sampled signals, it is possible to estimate the parameters of the signal
such as the direction-of-arrival, and also achieve spatial discrimination through filtering.
The techniques employed for parameter estimation and spatial filtering have a lot in
common.
Analogous to performing filtering on a set of data sampled in time to obtain a specific
frequency response, it is possible to process the spatially sampled data to obtain a desired
spatial response through a process called spatial filtering. Through spatial filtering it is
possible to separate a desired signal from interference even if they occupy the same
frequency band at the same time, as long as they are spatially separated. A processor used
in conjunction with an array of sensors to obtain some form of spatial filtering is called a
beamformer. Typically a beamformer linearly combines the spatially sampled data from
each sensor to obtain a scalar output in the same manner that an FIR filter linearly
combines temporally sampled data [Van88].
4Chapter 2: Antenna Arrays - Background Theory
Figure 2.4: A narrowband beamformer whose output is a complex weighted sum of the spatially
‘sampled input data.
Figure 2.4 shows a linear combiner beamformer which is typically used to process
narrowband signals. As shown in Figure 2.4, the output of each element is multiplied by a
complex factor called the element weights. The weights associated with all the elements
are collectively represented as the weight vector
W = [Wp anes) (2.21)
Note that the weights are represented as complex conjugates of the elements of the weight
vector, and for the sake of consistency this notation shall be adhered to throughout this
document. Also note that throughout this document lowercase bold letters represent
vectors, and uppercase bold letters represent matrices. Superscripts *, ”, and ” represent
conjugate, transpose, and hermitian transpose (conjugate transpose) respectively.
In practice, the weights are built using attenuators and phase shifters at the RF front-end or
IF, or as complex multiplying factors in baseband using digital processing. The output of
the beamformer can be expressed as
M
vk) = Yo wix (6) (2.22)Chapter 2: Antenna Arrays - Background Theory
or in matrix notation as
yk) = wl x(k) (2.23)
By appropriately choosing the weight vector, the output of the array can be adjusted to
discriminate between spatially separate signals. The resulting radiation pattern of the array
is directly related to the weight vector and is given by
F(0, 6,0) = |wa(®, 6, 0] (2.24)
If the array used in a narrowband M-element uniformly spaced linear array with an
interelment spacing d, using (2.24) and (2.15) and assuming isotropic elements, the
radiation pattern of the beamformer can be expressed as
M
F(Q) = |S wel Dace (2.25)
As clearly seen from equations (2.24) and (2.25), the radiation pattern of the array can be
adjusted by appropriately selecting the weight vector. Hence it is possible to form beams
or steer nulls in any required direction by appropriately adjusting the complex weight
vector.
2.3 Wideband Arrays
When only a narrowband of frequencies are of interest an array can be steered by simple
phase shifting operations. Therefore, in narrowband arrays, the weights are made up of
simple attenuators and phase shifters. However, when signals of significant frequency
extent (wideband signals) are of interest, simple phase shifting will not suffice. Instead
physical delay is required to accommodate the wide band of frequencies. Hence wideband
systems need to sample the incoming wave in both space and time. Thus the wideband
system requires spatio-temporal sampling and the processing of a large number of samples
from each sensor, and as a result is more complex. Temporal sampling can be achieved by
means of a tapped delay line in each antenna element [Gri83].
16Chapter 2: Antenna Arrays - Background Theory
xo(h)
w,
Figure 2.5: _ Illustration of a wideband array with a transversal filter attached to each antenna
element,
Figure 2.5 shows an example of a wideband antenna array with a tapped delay line
attached to each antenna element, The output of the array can be expressed as [Van88]
Mk
vk) = YY wi php) (2.26)
i=lp=0
where K-/ is the number of taps in the transversal filter. Consistent with the notation used
in the narrowband case, the output of the wideband array can equivalently be expressed asChapter 2: Antenna Arrays - Background Theory
yh) = wl x(a) (2.27)
where w and x(k) are of dimension I x KM and given by
Tr
We DW oe Mi tee Ma Kasse eee Wao ae, Ma, K-11 (2.28)
and
XC) = [ey (Ds ty 1a Eas tang UO] (2.29)
2.4 Adaptive Arrays
An adaptive array is a system in which the radiation pattern of the array is adjusted
continually to meet a certain performance criterion. Some of the most frequently used
performance criteria are: minimum mean squared error (MSE), maximum signal-to-noise
and interference ratio (SNIR), maximum likelihood (ML), minimum noise variance,
minimum output power, maximum gain, etc. [Nic88]. Adaptive arrays are particularly
suited for applications where the statistics of the signals and interference change
continuously and are not known in advance. An adaptive array consists of an array of
sensor elements whose weights are controlled by an adaptive processor which performs
the optimization based on computations performed on the incoming signal.
‘An adaptive array as shown in Figure 2.6 is an example of an adaptive signal processing
system [Wid89]. An M element array can be considered as a multiple input system, where
at the Ath sampling instant
mee ie Xm Xu)” 2.30)
is the array input vector or the array illumination vector, and
T
= DM, Woe me Wa) 2.31)
is the weight vector, Here the first subscript refers to the sensor number and the second
subscript refers to the time index. Using this notation, the array output can now be written
18(Chapter 2: Antenna Arrays - Background Theory
(2.32)
Wik’
XK
X2k
To Detector
a
Desired Signal
5s
ME BA
4 ‘Adaptive __|Exor
Algorithm *¥
Figure 2.6: Illustration of a simple adaptive array
As illustrated in Figure 2.6, the weight vectors are computed by an adaptive algorithm. In
general, the adaptive algorithm works to minimize a cost function which is set up to
achieve a certain performance criterion. For the particular array shown in Figure 2.6 the
weight vector computation depend on the output signal as well as a known “desired
response” at the receiver. During the adaptation process a feedback system adjusts the
weight vector such that the output, y,, agrees as closely as possible with the desired
response, d;. This is done by computing the error signal, which is the difference between
the output and the desired response, and adjusting the weight vector to minimize the mean
squared error signal.
The mean square error in the adaptive system shown in Figure 2.6 is given by
19Chapter 2: Antenna Arrays - Background Theory
MSE = Els;] = Eld;]+wR,.w—2r,j4w (2.33)
From equation (2.33) it is clearly seen that the mean square error is a quadratic function of
the weight vector w. Figure 2.7 shows a typical two-dimensional mean square error
function. The bowl shaped surface, called the performance surface, formed here is a
paraboloid (a hyperparaboloid if there are more than two weights). With a quadratic
performance surface there is only a single global minimum; no local minima exists. The
goal of adaptive algorithms is to find the optimal weight vector w,,, which corresponds
to this minima.
The weight vector wp, which minimizes the mean squared error is given by the Wiener-
Hopf solution [Wid89}.
-1
Wop = Rated (2.34)
MSE
Figure 2.7; An example of a two dimensional quadratic performance surface
where R,, is the input correlation or covariance matrix, and r,q is the cross correlation
20Chapter 2: Antenna Arrays - Background Theory
matrix between the input and the desired response. That is,
ik MaDe ial XM
Rao = EGA = Fleet, dy atl 2.35)
XX ik *MiX2k si *Mk
ya = Eldyxy) = Eldyxyp dgXry, 1yray] (2.36)
Ry, and ryq are in practice computed as time averages using samples collected over an
observation period. If the input data is stationary, the adaptive array only needs to acquire
enough data to make an adequate estimate of Ry and rq. Once Ry, and ryq are known, it is
conceptually a simple matter to invert R,, and compute the optimum weight vector.
However, since matrix inversion is a computationally expensive operation, the optimum.
weight vector is almost never computed this way. Iterative techniques such as the steepest
descent and stochastic gradient methods are used instead [Wid89]. The steepest descent
method computes the optimum weight vector using the following iterative equation.
Wey = tHE) (237)
where is a constant that regulates the step size, and V, is the gradient of the mean
square error with respect to the weight vector (the slope of the performance surface).
Stochastic gradient methods such as the Least Mean Squares (LMS) algorithm uses the
same iterative equation with the true gradient replaced by the instantaneous gradient
[wia67}.
2.5 The LMS Algorithm
As seen from section 2.4, the design of an adaptive array involves computing the optimum
weight vector w,,,. In order to compute the optimum weight vector it is required to find
the minimum of the error performance surface. In most applications of interest, the
parameters of the performance surface are not available in advance, and hence need to be
estimated by averaging the squared error over a period of time. Various algorithms have
2Chapter 2: Antenna Arrays - Background Theory
been developed to search the performance surface to locate the optimum weight vector.
Iterative methods of searching the performance surface entail the use of gradient estimates
to indicate the direction in which the minimum of the surface lies. The steepest descent
method of searching the performance surface has been widely used due to its ease of
implementation. In the method of steepest descent, all components of the weight vector
are changed in the direction of the negative gradient of the performance surface at each
iteration. Moving in the direction of the negative gradient leads toward the minimum as
ong as the origin lies on one of the principal axes of the surface.
The method of steepest descent can be expressed in the form of the following iterative
equation,
Wea = Wet WV) (2.38)
where 1 is a constant that regulates the step size and has dimensions of reciprocal signal
power,
‘As seen from equation (2.38), the steepest descent method requires the estimation of the
gradient of the performance surface. There are various procedures to obtain the gradient
estimate, The most general method, called the derivative method, estimates the gradient of
the mean square error (MSE) by taking the differences between short term averages of the
squared error, £47.
The Least Mean Square (LMS) algorithm is based on the method of steepest descent
which uses the instantaneous value of ¢,? itself as an estimate of the mean square
error{Wid67]. Therefore, at each iteration in the adaptation process, the gradient estimate
is given by
22Chapter 2: Antenna Arrays - Background Theory
= 26x, (2.39)
The derivatives of c, with respect to the weights follow directly from
& = dy—y, = dy —xflw = d,—w"' x, (2.40)
Using this estimate of the gradient, the LMS algorithm can be obtained from equation
(2.38) as
Wea = Wye— UV, = wy + 2MEpy (2.41)
Since it is based on a simple estimate of gradient, the LMS algorithm can be implemented
without squaring, averaging, or differentiation, as would be required by the derivative
methods. The greatest appeal of the LMS algorithms is its computational simplicity.
However, since the gradient estimates are obtained without averaging. there would be a
significant amount of noise in the estimate. The noise is attenuated with time by the
adaptation process which acts as a low pass filter.
It can be shown that as the number of iterations increases without limit, the expected value
of the weight vector computed by the LMS algorithm converges to the optimum Wiener-
Hopf solution. Convergence to the optimum solution is ensured only if step size parameter
wis defined such that [Wid67]
[l+2ua, (2.42)
mas| <1
or
<< (2.43)
where 2ay is the maximum eigen value of the input covariance matrix Ry,.
23(Chapter 2: Antenna Arrays - Background Theory
2.6 Property Restoral Algorithms
The computation of the weight vector using the techniques described so far require that
the desired signal be known at the receiver. This entails an overhead in the form of training
sequences or pilot signals. A relatively new class of algorithms referred to as property
restoral algorithms or blind algorithms which do not require training, has been proposed
and studied extensively. These algorithms are designed to exploit the fact that most signals
used in communication systems, have known modulation properties that can be used to
adapt the receiver processor. The properties commonly exploited include the constant
envelope property of FM, PM. and FSK type of signals, finite alphabet property of digital
signals, cyclostationarity of digital modulation signals, etc. The basic idea behind property
restoral-based approaches is to find a weight vector w such that the array output
y(k) = w"lx(k) has the desired property. Property restoral based methods can be used
for signal extraction as well as parameter estimation,
2.6.1 The Constant Modulus Algorithm
The constant modulus algorithm was introduced as an adaptive filtering technique for
correcting multipath and interference-induced degradations in constant envelope
waveforms such as FM and QPSK signals {Tre83]. Since its inception, the CMA
algorithm has been readily applied for spatial filtering as well [Goo86]. This algorithm is
based on the premise that multipath and additive interference causes unwanted amplitude
modulation on the transmitted signal, and hence destroys its constant envelope property.
Therefore, by sensing the received signal envelope variations, it is possible to build an
adaptive filter which will remove the channel-induced envelope variations and restore the
constant modulus property of the signal, thereby removing the interference components
from the desired signal. This is achieved by developing an algorithm to minimize a cost
function which penalizes envelope variations,
24Chapter 2: Antenna Arrays - Background Theory
The CMA cost function J has the general form
J = Elly? 8") (2.44)
where p and q are positive integers, and 6 > 0 is the desired nominal level of the received
signal envelope. Usually, 8 is set to unity, and p and q take on values of one or two. The
*yand “2-
algorithm is generally referred to be of a “p-q” form such as “1-1”, “1-2”, “2-1
2”.
A simple gradient search algorithm can be used to minimize the CMA cost function J. The
method of steepest descent as given in equation (2.37) may be applied to obtain the
following algorithm.
w(k+1) = w(K) pV, (2.45)
For the case of p = 2, q = 1, d =I, it can be shown that VJ with respect to the weight
vector w is given by [Tre83]
Vd = RE tIpGo—11VLW" xx") (2.46)
= E{[v)?=1xGox"(yw"}
= E{[lv(@))?—1y")x0)}
Removing the expectation operation from (2.46) gives the instantaneous gradient estimate.
Substituting the instantaneous gradient estimate in (2.45) the steepest descent CMA
algorithm is obtained as
w(k+1) = w(K) —néLlyGoP ~My" @x()} (2.47)
By defining an error term
e(k) = {Iv - Ly) (2.48)
equation (2.47) can be compactly expressed as
w(k+ 1) = w(k)—pe (K)x(K) (2.49)
The 2-1 form of CMA as described above closely resembles the complex version of the
2sChapter 2: Antenna Arrays - Background Theory
LMS algorithm, The only actual difference between the LMS algorithm and this version
of CMA is that instead of using an externally applied reference signal to compute the error
signal, the apriori knowledge that y(k) would be unit modulus in the absence of multipath
and interference is used to compute the error signal.
The other forms of the CMA algorithm corresponding to different values of p and q can be
similarly derived. The resulting weight update equations are summarized below [Lar83}.
“Lr wk +1) = w(ky— xo 2 sgn(lsol-8) (2.50)
“Q" w(k-+ 1) = w(k)—2ux(h)y" (A) sgn (lip? — 8°) (2.51)
“1.2” w(k+1) = w(K) — 2x CLC 8) (2.52)
2.2" w(k +1) = w(k)~4pxtk)y" Gop? —8) (2.53)
Ever since its inception, the CMA algorithm has been studied in great detail
(Lar83][Lar85][Tre85A][Tre85B][Kik94]. Due to the non-quadratic nature of the CMA
cost function its convergence properties are not as intuitive as that of the LMS type
algorithms. Existence of troughs in the cost function can cause the convergence to be
extremely slow giving an appearance of a false solution. The convergence behavior of
CMA under various conditions has been studied by (Lar83), [Age88], and [Tre85A]
The convergence of CMA can be significantly improved by pre-multiplying the input
vector x(k) by the inverse of the input correlation matrix R,, = E[xx""]. This has the
effect of orthogonalizing the input correlation function. The resulting algorithm is called
the orthogonalized CMA (O-CMA), and can be summarized as follows [Goo86]
w(k+1) = wk) + uRA(k+ De (x(k) (2.54)
where
26Chapter 2: Antenna Arrays - Background Theory
Rutk+ 1) = (2.55)
RAG) 1 fa Riix()x R100
Tra So heereecrranen
Various other modifications to the CMA algorithm have been proposed to speed up the
convergence rate, and to enable capture of multiple signals. Real arithmetic
implementations of CMA and extensions of CMA to non-constant known modulus signals
have also been studied by [Tre84}, [Tre8SB], and[Lun$8]. Some of the improved CMA
algorithms include techniques such as the Least-Squares CMA, Multitarget CMA
(Age89a], Multistage CMA[Shy94][Sub89], and the Iterative Least Squares Projection-
Based CMA [Tal94][Par95]. The Iterative Least Squares Projection Based CMA is
discussed Chapter 3
2.6.2 Spectral Coherence Restoral Algorithm
Most communication signals exhibit a property called cvelostationarity which can be
exploited to achieve blind adaptive beamforming and improve the performance of signal
parameter estimation algorithms. Cyclostationarity is a term used to describe the repetitive
or cyclic nature of the statistics associated with communication signals. Periodicities in
the second order statistics of a signal leads to the existence of a correlation between the
random fluctuations of different frequency components. This property is called spectral
coherence. A signal s(t) is said to exhibit spectral correlation if it is correlated with a
frequency shifted version of itself. That is, if the cyclic autocorrelation function, defined
by [Gar91][Gar94]
RX(t) = (s(e+ Dae ) (2.56)
is not identically zero for some cycle frequency a and some delay parameter t. Most
communication signals exhibit non-zero spectral correlation at one or more cycle
frequencies (the doubled carrier frequency or the baud rate or chip rate, etc.), and are said
to be cyclostationary [Gar87A}{Gar87B]. The Fourier transform of the cyclic
27Chapter 2: Antenna Arrays - Background Theory
autocorrelation function called the cyclic spectral density, and defined by
SLD =f RDeP Mae (257)
provides a clear measure of both the location and degree of spectral correlation in a
cyclostationary signal
Various cost functions based on the cyclic autocorrelation function and the cyclic spectral
density were developed by Agee [Age89b], and the class of algorithms which are based on
these cost functions are called the Spectral COherence REstoral (SCORE) algorithms.
2.7 Summary
In this chapter we introduced the basic concepts and terminology related to antenna arrays
and direction of arrival estimation, A brief overview of adaptive arrays and algorithms was
also presented. In the next chapter, we will present algorithms specifically for direction-of-
arrival estimation.
28Chapter 3: Overview of DOA Estimation Algorithms
Chapter 3
Overview of Direction of Arrival Estimation
Algorithms
3.1 Introduction
This chapter provides a detailed survey of the various methods available for estimation of
angle of arrival of a radio signal using an antenna array. The array-based direction-of-
arrival (DOA) estimation techniques considered here can be broadly divided into four
different types: conventional techniques, subspace based techniques, maximum likelihood
techniques and the integrated techniques which combine property restoral techniques with
subspace based techniques. Conventional methods are based on classical beamforming
techniques and require a large number of elements to achieve high resolution. Subspace
based methods are high resolution sub-optimal techniques which exploit the eigen
structure of the input data matrix. Maximum likelihood techniques are optimal techniques
which can perform well even under low signal-to-noise ratio conditions, but are in general
computationally very intensive. The integrated approach use property restoral based
techniques to separate multiple signals and estimate their spatial signatures from which
their directions of arrival can be determined using subspace techniques.
29Chapter 3: Overview of DOA Estimation Algorithms
3.2 Conventional Methods for DOA Estimation
Conventional methods for direction-of-arrival estimation are based on the concepts of
beamforming and null-steering, and do not exploit the nature of the model of the received
signal vector x(k) or the statistical model of the signals and noise. Given the knowledge of
the array manifold, an array can be steered electronically just as a fixed antenna can be
steered mechanically. Conventional techniques used for DOA estimation consists of
electronically steering beams in all possible directions, and looking for peaks in the output
power [Sch93]. The conventional methods discussed here are the delay-and-sum method
(classical beamformer) and the Capon’s minimum variance method.
3.2.1 Delay-and-Sum Method
x(k)
26K
xk)
Figure 3.1: Illustration of the classical beamforming structure.
The delay and sum method, also referred to as the classical beamformer method or Fourier
method, is one of the simplest techniques for DOA estimation. Figure 3.1 shows the
classical narrowband beamformer structure, where the output signal y(k) is given by a
30Chapter 3: Overview of DOA Estimation Algorithms
linearly weighted sum of the sensor element outputs. That is,
yk) = we(ay GB.1)
The total output power of the conventional beamformer can be expressed as
Por = Et] = Ete" x(a)] = wl EL x! Uw = "Rw G2)
where Ry, is the autocorrelation matrix of the array input data as defined in (1.34).
Equation (3.2) plays a central role in all the conventional DOA estimation algorithms. The
autocorrelation matrix Rj, contains useful information about both the array response
vectors and the signals themselves, and it is possible to estimate signal parameters by
careful interpretation of Ryy
Consider a signal s(k) impinging on the array at an angle 8p. Following the narrowband
input data model discussed in section 2.1, the power at the beamformer output can be
expressed as
Ethw xy = El}w™(a(O))s(k) + aC) T 63)
= (w%a0,) (2
Pep Oo
+0,))
where a(Q,) is the steering vector associated with the DOA angle Op, n(k) is the noise
vector at the array input, and o, = E[s(k)°] and o, = E[n(k)°] are the signal power
and noise power respectively. 1t is cleasly seen from (3.3) that the output power is
maximized when w = a(8p) . Therefore, of all the possible weight vectors, the receiver
antenna has the highest gain in the direction 8, when w = a(,). This is because
w = a(@q) aligns the phases of the signal components arriving from 8p at the sensors,
causing them to add constructively.
In the classical beamforming approach to DOA estimation, the beam is scanned over the
angular region of interest in discrete steps by forming weights w = a(@) for different 0,
and the output power is measured. Using equation (3.2), the output power at the classical
beamformer as a function of the angle of arrival is given byChapter 3: Overview of DOA Estimation Algorithms
P40) = wR, = a!"(0)R,,a(8) (3.4)
Therefore, if we have an estimate of the input autocorrelation matrix, and know the
steering vectors a/Q) for all 6's of interest (either through calibration or analytical
computation), it is possible to estimate the output power as a function of the angle of
arrival ®. The output power as a function of angle of arrival is often termed as the spatial
spectrum. Clearly, the directions of arrival can be estimated by locating peaks in the
spatial spectrum defined in (3.4).
The delay and sum method has many disadvantages. The width of the beam and the height
of the sidelobes limit the effectiveness when signals arriving from multiple directions and/
‘or sources are present because the signals over a wide angular region contribute to the
measured average power at each look direction. Hence, this technique has poor resolution.
Although it is possible to increase the resolution by adding more sensor elements,
increasing the number of sensors increases the number of receivers and the amount of
storage required for the calibration data, i.c., a0).
3.2.2 Capon’s Minimum Variance Method
The delay-and-sum method works on the premise that pointing the strongest beam in a
particular direction yields the best estimate of power arriving in that direction. In other
words, all the degrees of freedom available to the array were used in forming a beam in the
required look direction. This works fine when there is only one signal present. But, when
there are more than one signal present the array output power contains contribution from
the desired signal as well as the undesired ones from other directions.
Capon’s minimum variance technique [Cap69] attempts to overcome the poor resolution
problems associated with the delay-and-sum method. The idea is to use some (not all) of
the degrees of freedom to form a beam in the desired look direction while simultaneouslyChapter 3: Overview of DOA Estimation Algorithms
using the remaining degrees of freedom to form nulls in the direction of interfering
signals. This technique minimizes the contribution of the undesired interferences by
minimizing the output power while maintaining the gain along the look direction to be
constant, usually unity. That is,
minE[\y(QP1 = min w' Rw — subject to wa(Oo) = 1 G5)
” a
The weight vector obtained by solving (3.5) is often called the minimum variance
distortionless response (MVDR) beamformer weights since, for a particular look
direction, it minimizes the variance {average power) of the output signal while passing the
signal arriving in the look direction without distortion (unity gain and zero phase shift),
Equation (3.5) represents a constraint optimization problem which can be solved using the
method of Lagrange multipliers. This approach converts the constraint optimization
problem into an unconstrained one, thereby allowing the use of least squares techniques to
determine the solution. Using a Lagrange multiplier, the weight vector that solves (3.5)
can be shown to be [Hay91]
Ra(0)
a(8)R,.a@)
(3.6)
Now the output power of the array as a function of the angle of arrival, using the Capon’s
beamforming method, is given by the Capon’s spatial spectrum,
1
a (8)R.
P capon(®) 3.7)
2 4(0)
By computing and plotting the Capon's spectrum over the whole range of ®, the DOA’s
can be estimated by locating the peaks in the spectrum.
Although it is not a maximum likelihood (ML) estimator, Capon’s method is sometimes
referred to as an ML estimator since for any choice of ®, PCapon(®) is the maximum
likelihood estimate of the power of a signal arriving from the direction 0 in the presence of
white Gaussian noise having arbitrary spatial characteristics (Cap79].
33Chapter 3: Overview of DOA Estimation Algorithms
20. $$$ + —}
| delay and sum method |
Ld
yan |
I
1
|
|
|
g ee |
2 hI
é hi
3 | "| Capon’s method
2 |
|
WI
iy
ta
Io
I
tt
a)
oy
a ee eT)
Angle of Arrival in Degrees
Figure 3.2: Comparison ot resolution performance of delay-and-sum method and Capon's
minimum variance method. Two signals of equal power at an SNR of 20 dB arrive at a 6-element
uniformly spaced array with an interelement spacing equal to half a wavelength at angles 90 and
100 degrees respectively.
Figure 3.2 illustrates the performance improvement obtained by Capon’s method over the
delay-and-sum method, Computer simulations show that using a six element uniformly
spaced linear array with half wavelength interelement spacing, Capon’s method is able to
distinguish between the two signals arriving at 90 and 100 degrees respectively, while the
delay-and-sum method fails to differentiate between the two signals.
Though it provides a better resolution when compared to the delay-and-sum method,
Capon’s method suffers from many disadvantages. Capon’s method fails if other signals
that are correlated with the signal of interest are present because it inadvertently uses that
correlation to reduce the processor output power without spatially nulling it [Sch93]. In
other words, the correlated components may be combined destructively in the process of
34Chapter 3: Overview of DOA Estimation Algorithms
izing the output power. Also, Capon’s method requires the computation of the
matrix inverse which can be expensive for large arrays
3.3 Subspace Methods for DOA Estimation
Though many of the classical beamforming based methods such as the Capon’s minimum
variance method are often successful and widely used, these methods have some
fundamental limitations in resolution. Most of these limitations arise due to the fact they
do not exploit the structure of the narrowband input data model of the measurements.
Schmidt [Sch79] and Bienvenu and Kopp [Bie79] were the first to exploit the structure of
a more accurate data model for the case of sensor arrays of arbitrary form. Schmidt
derived a complete geometric solution to the DOA estimation problem in the absence of
noise, and extended the geometric concepts to obtain a reasonable approximation to the
solution in the presence of noise. The technique proposed by Schmidt is called the
MuUltiple Signal Classification (MUSIC) algorithm, and has been thoroughly investigated
since its inception [Bar84][Sto89][Pil89B]. The geometric concepts upon which MUSIC
is founded form the basis for a much broader class of subspace-based algorithms
[Pau93][Joh86}. Apart from MUSIC, the primary contributions to the subspace-based
algorithms include the Estimation of Signal Parameters via Rotational Invariance
Technique (ESPRIT) proposed by Roy et. al.,{Pau86B][Roy89][Roy90], and the
minimum-norm method proposed by Kumaresan and Tufts [Kum83].
3.3.1 The MUSIC Algorithm
The MUSIC algorithm proposed by Schmidt in 1979 {Sch79\[Sch86A] is a high
resolution multiple signal classification technique based on exploiting the eigen structure
of the input covariance matrix. MUSIC is a signal parameter estimation algorithm which
provides information about the number of incident signals, direction of arrival (DOA) of
each signal, strengths and cross correlations between incident signals, noise power, etc.
35Chapter 3: Overview of DOA Estimation Algorithms
While the MUSIC algorithm provides very high resolution, it requires very precise and
accurate array calibration. The MUSIC algorithm has also been implemented and its
performance experimentally verified [Sch86B]
The development of the MUSIC algorithm was based on a geometric view of the signal
parameter estimation problem. Following the narrowband data model discussed in section
1.1, If there are D signals incident on the array, the received input data vector at an M-
element array can be expressed as a linear combination of the D incident waveforms and
noise. That is,
x} 0 si] [ay
= |a(0,) .. a(Bpy) b+). G8)
XM] Sp} [Ma]
x= Astn (3.9)
where s = [s; s2 ... Sp] isthe vector of incident signals, m = [n, 7) ... Mp] is the
noise vector, and a(Q)) is the array steering vector corresponding to the direction of arrival
of the jth signal. In geometric terms, the received vector x and the steering vectors a(8))
can be visualized as vectors in M dimensional space. From (3.8), it is seen that the
received vector x is a particular linear combination of the array steering vectors, with s;,
53...u8p being the coefficients of the combination. In terms of the above data model, the
input covariance matrix R,, can be expressed as
R,, = Elxx"] = AE[ss]4" + Elan} (3.10)
Reo = AR A + No icel GB.)
where R,, is the signal correlation matrix E[ss"].
Since A is comprised of steering vectors which are linearly independent, it has full column
rank, and the signal correlation matrix R,, is non-singular as long as the incident signals
are only at most partially correlated
36Chapter 3: Overview of DOA Estimation Algorithms
A full column rank and nonsingular R,, guarantees that, when the number of incident
signals D is less than the number of array elements M, the MxM matrix AR,.4” is
positive semidefinite with rank D, From elementary linear algebra, this implies that M-D
of its eigenvalues are zero. If 41, do
Ay and v;, ¥2..... ¥y¢ denote the eigenvalues and
eigenvectors, respectively, of Ryx it is clear from (3.11) that the 4-D smallest eigenvalues
of Ry are all equal to oygise”. That is,
7
=o, G.12)
min = ONoise
thos = Rpv2 = M
In practice, however, when the autocorrelation matrix R,, is estimated from a finite data
sample, all the eigenvalues corresponding to the noise power will not be identical. Instead
they will appear as a closely spaced cluster, with the variance of their spread decreasing as
the number of samples used to obtain an estimate of Ryy is increased. Once the
multiplicity, K, of the smallest eigenvalue is determined, an estimate of the number of
signals, D, can be obtained from the relation M = D+K. Therefore, the estimated
number of signals is given by
D=M-K G.13)
From the definition of eigenvalues and eigenvectors (Gol89}, the eigenvectors
corresponding to the smallest eigenvalues (non-principle eigenvalues) must satisfy
2 7
Ryg¥i = ONipice”i + i= D+1,..,M G.14)
From (3.11), this means that
AR, Ay, = 0, i= D+,
M (3.15)
Now since A is full column rank, and R,, is non-singular, it follows that
i=D+l,.
M 3.16)
From (3.16), since the dot product of A" and the eigenvectors vp4),....¥4y are Zero, it
P ig Deb ™M
implies that the column vectors of A are perpendicular to those eigenvectors.
37Chapter 3: Qverview of DOA Estimation Algorithms
£4(8)), 5 @Op)} L Lp epson My 3.17)
In short, the above analysis has shown that the eigenvectors of the covariance matrix Ryy
belong to either of the two orthogonal subspaces, called the principal eigen subspace
(signal subspace) and the non-principal eigen subspace (noise subspace). The steering
vectors corresponding to the direction of arrivals of signals lie in the signal subspace and
are hence orthogonal to the noise subspace. By searching through all possible array
steering vectors to find those which are perpendicular to the space spanned by the non-
principal eigenvectors, the direction of arrivals 0,’s can be determined.
The direction of arrival of the multiple incident signals can be estimated by locating the
peaks of a MUSIC spatial spectrum given by
Puusic® = For Pao) G.18)
or
Pyusic®) = oe G9)
a ov yV aC)
where
wo yg} (3.20)
Equations (3.18) and (3.19) are two possible measures of the closeness of an element of
the array manifold to the signal subspace. The denominator of (3.18) and (3.19) is
essentially a scalar measure of the distance between the steering vectors, @/@), in the array
manifold, and the estimated noise subspace spanned by the eigenvectors, Vy. The product
VyVy" represents the projection matrix on the noise subspace. Orthogonality between
(8) and Vy will minimize the denominator and hence will give rise 10 peaks in the
MUSIC spectrum defined in (3.18) and (3.19). The D largest peaks in the MUSIC
spectrum correspond to the directions of arrival of the signals impinging on the array.Chapter 3: Overview of DOA Estimation Algorithms
Once the directions of arrival @,’s are determined from the MUSIC spectrum the signal
covariance matrix R,, can be determined from the following relation [Sch86A].
Ry = (AN Ay AM (Re pin AANA) (3.21)
From (3.21), the powers and cross correlations between the various input signals can be
readily obtained.
The MUSIC algorithm may be summarized as follows:
Step 1. Collect input samples {x/k), k = 1... ., N} and estimate the input covariance
matrix
ee
R= yD (3.22)
kel
Step 2. Perform eigen decomposition on Ry
RuV = VA (3.23)
where A = diag{hy,Aoy..uhybs Ay 2%22-.. Zh are the eigenvalues and
V = (v1, vp... Vag} ate the corresponding eigenvectors of R,. .
Step 3. Estimate the number of signals D, from the multiplicity K, of the smallest
eigenvalue Xypin aS
D=M-K (3.24)
Step 4. Compute the MUSIC spectrum
a"(@)a(0)
(3.25)
a (OW yVia(0)
Puusrc(9) =
= ra
where Vy = (5, jy)
Step 5. Find the D largest peaks of Pyyspc(®) to obtain estimates of the direction-of-
arrival.Chapter 3: Overview of DOA Estimation Algorithms
50+
al |
\y
Vy
|
MUSIC algorithm |
|
!
al
——
0 20 40 60 80, 100 120 140 160, 180,
Angle of Arrival in Degrees
Figure 3.3: Comparison of MUSIC and Capon’s minimum variance method. Two signals of
equal power at an SNR of 20 dB arrive at a 6-element uniformly spaced array with an interelement
spacing equal to haif a wavelength at angles 90 and 95 degrees respectively.
Figure 3.3 shows a comparison between the resolution performance of MUSIC and the
Capon’s minimum variance method. As seen clearly from the plot, MUSIC can resolve
closely spaced signals which cannot be detected by Capon’s method. Simulation results
show that two signals arriving at angles 90 and 95 degrees, respectively, at the input ofa 6-
element uniformly spaced linear array can be detected by MUSIC, while Capon’s
minimum variance method fails to differentiate between the two signals.
It should be noted that unlike the conventional methods, the MUSIC spatial spectrum does
not estimate the signal power associated with each arrival angle. Instead, when the
ensemble average of the artay input covariance matrix is known exactly, under
uncorrelated and identical noise conditions, the peaks of Pyyyjsjc(8) are guaranteed to
40Chapter 3: Overview of DOA Estimation Algorithms
correspond to the true angles of arrival. Since these peaks are distinct irrespective of the
actual separation between arrival angles, in principle, with perfect array calibration, these
estimators can distinguish and resolve arbitrarily closely spaced signals.
3.3.2. Improvements to the MUSIC Algorithm
Various modifications to the MUSIC algorithm have been proposed to increase its
resolution performance and decrease the computational complexity. One such
improvement is the Root-MUSIC algorithm developed by Barabell [Bar83]. which is
based on polynomial rooting, and provides higher resolution but is applicable only to a
uniform spaced linear array. Another improvement proposed by Barabell uses the
properties of the signal space eigenvectors (principal eigenvectors) to define a rational
spectrum function with improved resolution capability [Bar83]
CYCLIC MUSIC which exploits the spectral coherence properties of the signal to
improve the performance of the conventional MUSIC algorithms has been proposed in
[Sch89]. Fast Subspace Decomposition techniques have also been studied to decrease the
computational complexity of MUSIC [Xw94B].
Root-MUSIC Algorithm
For the case of a uniformly spaced linear array with interelement spacing d, the mth
element of the steering vector a(®) may be expressed as (see (2.15)):
a,,(8) = exp{ j2nm(Z)c0s8); .M (3.26)
The MUSIC spectrum given by (3.18) is an all-pole function of the form
1
a (3.27)
a"(8)VyVia(8)
1
Pyusic® =
a”"(0)Ca(0)
41Chapter 3: Overview of DOA Estimation Algorithms
where € = VyV . Using equation (3.26) the denominator of (3.27) may be written as
2) 030) Ca exp( ~/280( £Je0s0) (3.28)
MoM
i Cet
Pausic(®) = Dy YL exp|—j2nm
m=in=l
where Cy, is the entry in the mth row and nth column of C. Combining the two
summations into one, (3.28) can be simplified as
Me
1
Pye) = > Cyexp(-72n{)coso) (3.29)
tame
where C, = S) Cyyy is the sum of the entries of C along the ith diagonal.
mont
By defining a polynomial D(z) as follows,
Mat
Dey= YS Cet (3.30)
[=-M41
evaluating the MUSIC spectrum Pyysc(8) becomes equivalent to evaluating the
polynomial Dz) on the unit circle, and the peaks in the MUSIC spectrum are due to the
roots of D(z) lying close to the unit circle. (Ideally (no noise) the poles will lie exactly on
the unit circle at locations determined by the angle of arrival). In other words, a pole of
|exp(jarg(z,)) will result in a peak in the MUSIC spectrum at
cos0 = (545 are(=)) G31)
Barabell [Bar83] showed through simulations that the ROOT-MUSIC algorithm has better
resolution than the spectral MUSIC algorithm, especially at low SNR conditions.
Cyclic MUSIC Algorithm
Cyclic MUSIC isa signal selective direction finding algorithm which exploits the spectral
coherence of the received signal as well as the spatial coherence. By exploiting spectral
correlation along with MUSIC, it is possible to resolve signals spaced more closely than
the resolution threshold of the array when only one of them is an signal-of-interest (SOI)
a2Chapter 3: Overview of DOA Estimation Algorithms
[Sch89][Sch94]. Cyclic MUSIC also circumvents the requirement that the total number of
signals impinging on the array (including both SOI and interference) be less than the
number of sensor elements [Gar88].
Consider an array of M sensors which receives Dz signals which exhibit spectral
correlation at a cycle frequency a, and an arbitrary number of interferers that do not
» Dy be the
exhibit spectral correlation at that particular frequency. Let si(1), i
desired signals, and n/t) the noise and interference vector incident on the array. The
received signal vector x(t) can then be expressed as
Pa
x(t) = DY) a(8,)s,(1) + (0) (3.32)
i=)
= As(t) +n(0)
Since only the desired signals exhibit spectral correlation at a, the cyclic autocorrelation
matrix R&,(z) of the received signal x(t) defined as
bee oloe
can be expressed as
4
Jese(2xan} ) (3.33)
RE(t) = ARR (DAM (3.34)
where R(t) is the cyclic autocorrelation matrix of the desired signals, and defined as
= (ee Dhol
”
$)erwtanu} ) 3.35)
where
r
On = fim fae (3.36)
2
Clearly, the matrix R&,() has rank Dj. For Dy s4(exp[joyrcos(!)]a(0,) +m, (0) (3.41)
f
where 0, is the direction of arrival of the kth source relative to the direction of the
45Chapter 3: Overview of DOA Estimation Algorithms
translational vector A, and D is the number of signals incident on the array. Now, using
matrix and vector notation, the received signal vector at the two subarrays can be written
as follows:
Xp(t) = As(t) + ng(0) (3.42)
XD = A®s(1) +0, (1) (3.43)
where © is a D x D diagonal unitary matrix whose diagonal elements represent the phase
delays between the doublet sensors for the D signals. The matrix @ relates the
measurements from subartay X, to those from subarray X), and is given by
A
® = diagiexp(jy;) exP(/¥)s--..expUYp)]. Where yy, = —Zcos(0,) (3.44)
Though in the complex field, the matrix & is a simple scaling operator, it is similar to the
real two-dimensional rotation operator. The total array output vector x/t) can be written as
x(t) = [OO] = Aste) + min) (3.45)
x (1)
where
a- 4) and n(t) = ee (3.46)
AQ} ay
The basic idea behind ESPRIT is to exploit the rotational invariance of the underlying
signal subspace induced by the translational invariance of the sensor array[Roy89]. The
relevant signal subspace is the one that contains the outputs from the two subarrays Xj and
X;. Simultaneous sampling of the output of the arrays leads to two sets of vectors Vp and
V,, that span the same signal subspace.
The signal subspace can be obtained from the knowledge of the input covariance matrix
R.
AR,,A+ ovpisel. If D... > Ay)
48Chapter 3: Overview of DOA Estimation Algorithms
[, |
oH Volf. . H
Vivo = i = VAV' 3.56
oi Mor li [i #1] GB.56)
1
and partition V into D x D submatrices,
y
v= |Mn Ya (3.57)
Vay Vy
1
Vy,
Step 6. Calculate the eigenvalues of ¥ =
ty = eigenvalues of(-V V3), Vk oD (3.58)
Step 7. Estimate the angle-of-arrival as
4 IP (arg (b,))
8 = cos |e J} (3.59)
As seen from the above discussion, ESPRIT eliminates the search procedure inherent in
most DOA estimation methods. ESPRIT produces the DOA estimates directly in terms of
the eigenvalues.
3.4 Maximum Likelihood Techniques
Maximum Likelihood (ML) techniques were one of the first techniques to be investigated
for DOA estimation. Since ML techniques were computationally intensive, they were less
popular than suboptimal subspace techniques. However, in terms of performance, the ML
techniques are superior to the subspace based techniques, especially in low signal-to-noise
ratio conditions or when the number of samples is small [Zis88]. Moreover, unlike
subspace based techniques, ML based techniques can perform well in coherent signal
conditions as well,
To derive the ML estimator, data collected over a block of N snapshots is formulated as
49Chapter 3: Overview of DOA Estimation Algorithms
X = A(@)S+N (3.60)
where X = [x(1)mu.X(N)] is the array data input vector matrix of dimension
MxN, A(@) = [a(0,
4(0p)] is the spatial signature matrix of dimension
MxD, S = [s(D ous 8(N)] is the signal waveform matrix of dimension D x N, and
N= [n(1).
the angles of arrival 8... .. 8 of the D sources, the maximum likelihood estimator makes
n(N)] _ is the noise matrix of dimension M x N. In order to estimate
some assumptions about the signals and noise. First, it is assumed that the number of
signals is known or estimated, and is smaller than the number of sensors. Second, every set
of D steering vectors are assumed to be linearly independent. The noise component is
assumed to be stationary and ergodic complex valued Gaussian process of zero mean and
covariance oY, where o” is an unknown scalar and J is the identity matrix. Finally it is
assumed that the noise samples are statistically independent. it should be noted that the
ML estimator is meaningful even when the assumptions made about noise do not hold, in
which case it coincides with the Least-Squares estimator [Zis88].
The derivation of the ML estimator described here regards the signals to be sample
functions of unknown deterministic sequences, rather than random processes. Based on
the assumptions made about the nature of noise, the joint probability density function of
the sampled data as given by equation (3.60) can be expressed as [Zis88]
»
sy = FT —
0x0 ( xcs) --4(@)s(49") 6.61)
jerrdetfo TN] \o*
where det{ ] denotes the determinant. Ignoring the constant terms, the log likelihood
function is given by
N
D eG) -4)sHI? (3.62)
Cnet
1
J = —NDlogo*
To compute the maximum likelihood estimator, the log likelihood function of (3.62) has to
be maximized with respect to the unknown parameters. This yields the following
maximization problem:
50Chapter 3: Overview of DOA Estimation Algorithms
N
Dv x5 xy ix) 00"} (3.63)
=
The logarithm being a monotonic function, maximizing (3.63) is equivalent to the
following minimization problem:
¥
min| x x(a) —ac@ystanrt (3.64)
(8) LO J
Fixing © and minimizing with respect to S, yields the well known least squares solution
5k) = (4%(0)4(0)) (Ox) (6.65)
Substituting (3.65) into (3.64), we obtain
y
min Y, |x(k) Pox’) (3.66)
kel
where Pyg) is the projection operator which projects vectors onto the space spanned by
the columns of 4(@), and is given by
a
Po = A(0)(A"(0)4(0)) 4) 6.67)
Therefore, the ML estimate of the directions of arrival © = {0,,..., 8p} is obtained by
maximizing the log-likelihood function
x
IO) = ¥ [Paeyx/? (3.68)
fel
Equation (3.68) can be interpreted in a geometric way such that the ML technique appears
as a variant of the subspace based method. Viberg and Oterson [Vib91 ]presented a
generalized framework to highlight the similarities between the various subspace based
DOA estimation techniques and the maximum likelihood technique. In geometric terms, it
follows from (3.68) that the ML estimator is obtained by searching over the array
manifold, of those D steering vectors that form a D-dimensional signal subspace which is
closest to the vectors (x(k), k=/,.... N}, where closeness is measured by the modulus of
slChapter 3: Overview of DOA Estimation Algorithms
the projection of the vectors onto this subspace.
It can be shown that equation (3.68) can be equivalently written as
trace[P 4.@)Rex) (3.69)
where R,, is the sample covariance matrix
y
Ry = xz x(ox"(k) (3.70)
‘The maximization of the log-likelihood function in (3.69) is # nonlinear, multidimensional
maximization problem which is computationally very intensive, Many computationally
efficient algorithms have been developed to simplify the solution to the maximization
problem [Fed88][Zis88][Li93].
The Alternating Projection algorithm developed by Ziskind and Wax [Zis88] is an iterative
technique which reduces the maximization problem from a multi-dimensional problem to
a one-dimensional problem. The idea is to perform the maximization with respect to a
single parameter while holding the remaining parameters fixed. That is, the value of 8, at
the (n+/)th iteration is obtained by solving the following one-dimensional maximization
problem:
67 = arg max trace[P,
(4.09). 08) 4) 2)
a(n) ct . ;
where 6{") denotes the vector comprising of the direction of arrival angle estimates of all
signals other than the one being computed.
f= (0)... 1, OF... .637) (G.72)
Since the log-likelihood function J(@) may have multiple local maximas, proper
initialization is critical for global convergence. The initialization procedure suggested by
Ziskind and Wax begins by solving the maximization problem for a single source:Chapter 3: Overview of DOA Estimation Algorithms
A. = arg max tracel Py, Rex) (3.73)
o,
Using the estimated 0;, 0 is calculated as,
400
@S = arg max trace[P
5
rac 6%", aay) Ree] a7)
- ara
Continuing in this fashion, @{° = [0\6S
65)'] is computed. After proper
initialization, the altemating projection algorithm can be used to maximize the log
likelihood function.
Further reduction in computational complexity can be achieved by taking advantage of the
properties of the projection matrix [Zis88]. It can be shown that, by using the properties of
the projection matrix, (3.71)can be written as
Of"? = arg max trace| Pa a] (3.75)
8; ie")
where
0, = UP. a0 3.76
200) gg) = EPyg0 OD (3.76)
By defining a unit vector,
ow
»(0,, 8 6.77)
equation (3.75) can be rewritten as
max 60, 6\")R6(0,,0;") G78)
8
The alternating projection based maximum likelihood estimator algorithm may be
summarized as follows:
Step 1: InitializationChapter 3: Overview of DOA Estimation Algorithms
(0) (0). a”
6, = max 60, 6;")R0(8,, 6 sfori=1.....D (3.79)
Step2: Main Loop
umit {a afr] Peele
"> (" ) fori=1,...,D. (3.80)
6 = grax 0 "(0,, 0)" R6(0, 0;
3.5 DOA Estimation under Coherent Signal Conditions
As mentioned in the Section 3.3.1, the MUSIC algorithm works on the premise that the
signals impinging on the array are not fully correlated, or coherent. Only under
noncoherent conditions does the source covariance matrix R,, satisfy the full rank
condition which is the basis of the MUSIC eigen decomposition. The performance of
MUSIC degrades severely in a coherent or highly correlated signal environment as
encountered in multipath propagation. Many modifications to the MUSIC algorithm have
been proposed to make it work in the presence of coherent signals. Many of these
techniques involve modification of the covariance matrix through a preprocessing scheme
called spatial smoothing. One method of spatial smoothing proposed by Evans et. al
[Eva82] and further elaborated by Shan et.al [Sha85] is based on averaging the covariance
matrix of identical overlapping arrays. This method requires an array of identical elements
built with some form of periodic structure, such as the uniformly spaced linear array. An
adaptive spatial smoothing technique was proposed by Takao and Kikuma [Tak87], which
is useful for interference cancellation in multipath environments. Another form of spatial
smoothing proposed by Haber and Zoltowski [Hab86] involves moving the entire array
structure during the time interval in which the covariances are estimated. A similar
technique based on moving the array was proposed by Li and Compton [Li94]. Spatial
smoothing techniques always impose restrictions of some form or the other on the type
and structure of the array. For the general case, coherent signal detection involves
employing a multidimensional search through all possible linear combination of steering
34Chapter 3: Overview of DOA Estimation Algorithms
vectors to find those orthogonal to the noise subspace [Zol86].
3.5.1 Spatial Smoothing Techniques
The idea behind the spatial smoothing scheme proposed by Evans et. al., [Eva82] is as
follows. Let a linear uniform array with M identical sensors be divided into overlapping
forward subarrays of size p, such that the sensor elements {1,... , p} forms the first
forward subarray, sensors {2,..., p+} forms the second forward subarray, etc. Let x;(t),
denote the vector of received signals at the Ath forward subarray. Based on the notation of
equation, (3.9) we can write
x[(0 = AFO sr) + ne) 81)
where F” denotes the kth power of the diagonal matrix
» \}
F= diag xo(-so(doxe0,)) 7 wexp(-Jo0fe0885)) (3.82)
The covariance matrix of the Ath forward subarray is therefore given by
RE = APO R FD AT oT (3.83)
where R,, is the covariance matrix of the sources.
Based on the above, the forward averaged spatially smoothed covariance matrix Rf is
defined as the sample mean of the subarray covariance matrices:
R
L
1 f
TUR (3.84)
kel
where L=M-p+1 is the number of subarrays. Now, substituting (3.83) in (3.84), we obtain
L \
1 1 =
AR RGR Pal orl 8s)
where Rf, is the modified covariance matrix of the signals, given by
55Chapter 3: Overview of DOA Estimation Algorithms
L
ia! kt) H(k-1)
Ri. = ize RF (3.86)
For L 2 D, the covariance matrix Rf, will be nonsingular regardless of the coherence of
the signals [Pil89A].
The price paid for detection of coherent signals using forward averaging spatial smoothing
is the reduction in the array aperture. An M element array can only detect M/2 coherent
signals using MUSIC with forward averaging spatial smoothing as opposed to M-/
noncoherent signals that can be detected by conventional MUSIC.
Pillar and Kwon [Pil89A] proved that by making use of a set of forward and conjugate
backward subarrays simultaneously, it is possible to detect up to 2/3 coherent signals. In
this scheme, in addition to splitting the array into overlapping forward subarrays, it is also
split into overlapping backward arrays such that the first backward subarray is formed
using elements {M, M-/..... M-p+/}, the second subarray is formed using elements {M-1,
M-2,..., M-p}, and so on.
Similar to (3.81), the complex conjugate of the received signal vector at the kth backward
subarray can be expressed as
7 : : r
XP [Ex ike, Mok oe x poke] (3.87)
= AF '(FM"'s) +n, (1ski
os f
i os, 3
| i ae
. os © @ 6 wo www wm Se 0 @ © mm wm wm
Angle of Arrival (deg.) ‘Angle of Arrival (deg.)
8
1 } | oNospacing. errors] 7 ‘0 No spacing. errors
* 1% spacing errors
* 1% spacing errors | /
+ 5% spacing errors
+ 5% spacing errors | /
TLS-ESPRIT | | | TLS-ESPRIT
+ ILSP-TLSESPRIT| | 1+ 2% ... ILSP-TLSESPRIT|
3 signals | * nals
RMS Error in DOA Estimation (deg.)
RMS Error in DOA Estimation (deg.)
2 «oo @ @ wm © ¢ # @ @ w wm www
Angle of Arrival (deg.) Angle of Arrival(deg.)
Figure 4.17: RMS error in DOA estimation as a function of angle of arrival for diferent spacing
errors. Spacing errors were simulated by perturbing the location of the sensors. All the sources
incident on the array are uncorrelated with one another. The estimations were made using 500
samples of data and the rms error was computed by averaging over 200 repeated trials. (These
plots were generated using Programs [46][48][50)
92Chapter 4: Simulation Based Performance Comparison of DOA Estimation Algorithms
1 ‘0 No spacing. errors T ‘0 No spacing. errors | | 7
* 1% spacing errors , & * 1% spacing errors | |
25 +.5% spacing errors ; 4 + 5% spacing errors | |)
5 TLS-ESPRIT t| TLS-ESPRIT | |
24 ILSP-TLSESPRIT| /, 38 bail ILSP-TLSESPRIT| }
é 1 signal 3 2signals | 3
oo a
< |
Z a
e. 5
cm \es
g S os,
7 et
oo 2 eo @ » Se 4 8 © @ 2 we
Angle of Arrival Angle of Arrival
w 7 7
7 TT |, Jo No spacing errors
| 1 \\|*" 1% spacing. error j
‘No spacing. errors], + 5% spacing errors |
* 1% spacing errors | 3 pF
\ + 5% spacing errors | 2 /|
\ |_TLS-ESPRIT | 5 WP
é | -lLSP-TLSESPRIT| SSS rseisPiar
< \ 3 signals ae ILSP-TLSESPRIT
= T 5 4 signals
= i a
a, \ | lé
7 \ 2 let
2
z 2.
to. (ey
a ‘veo ene? J
° em a meee)
Angle of Arrival
Figure 4.18:
=o © 6
eee
ie
Angle of Arrival
RMS error in DOA estimation as a function of the angle of arrival for different
spacing errors for the case of coherent (perfectly correlated) signals incident on the array. Spatial
smoothing through forward/backward averaging is used to decorrelate the signals. The estimations.
are made using 500 samples of data and the rms error is computed by averaging over 200
repeated computations. (These plots were generated using Programs (47][49][50))
93Chapter 4: Simulation Based Performance Comparison of DOA Estimation Algorithms
As seen clearly from Figures 4.17 and 4.18, the DOA estimation error increases rapidly
with increasing imperfections. Perturbations in sensor element location cause greater
errors on end-fire estimates than on the broadside estimates. Also the errors increase
rapidly with increasing number of incident signals. While for estimating the DOA of a
single signal, perturbations in the sensor element location do not produce much error, the
error grows very rapidly as the number of signals increase. For the case of uncorrelated
signals, the integrated approach is more robust to array imperfections than the
conventional TLS-ESPRIT algorithm. For the coherent signal case, both the algorithms
are equally sensitive. Also, array imperfections cause larger errors in DOA estimation of
coherent signals than for non-coherent signals. Though the integrated approach is more
robust to array imperfections than the conventional TLS-ESPRIT, to mininimize the error
in DOA estimation, for both the algorithms, it is required that the array imperfections be
kept at a minimum.
4.4 Performance of Source Order Estimation Algorithms
The source order estimation (detection of number of sources) techniques compared here
are all based on analyzing the spread of the eigen values of the autocorrelation matrix. The
techniques compared are the Maximum Descriptive Length Criterion (MDL), Akaike
Information Theoretic Criterion (AIC) [Wax85], and both MDL and AIC modified by
incorporating the Gershgorin radii information, i.e, the GMDL and GAIC criteria
respectively (see Section 2.7) [Wu95]. Each of these techniques were tested with different
number of signals and their detection error probabilities are plotted as function of number
of samples and signal-to-noise ratio respectively in Figures 4.19 and 4.20. The detection
error probability was computed as the relative frequency of error over 200 Monte Carlo
trials, The accuracy of the order estimation algorithms improves with increasing signal-to-
noise ratio and number of data samples used.
94Chapter 4: Simulation Based Performance Comparison of DOA Estimation Algorithms
ee
MDL
7 .! ~. GMDL
E B fy on GAIC |
5 5B )\ wee ey
g g
3 # of sources = 1 305 # of sources
5 3
2 3
2 z
= 0) ea SO BB
1 2 3 1+ 1 2 3 4
10 10 10 10 10 10 10 10
Number of samples of data used Number of samples of data used
AIC
- MDL
» 4 “ot tha ge ex
é ann GAIC
BR
o % eo
2 \ ke
205 \ altel
s x # of sources = 3
0 +# of sources = 4
10' w 1 10
‘Number of samples of data used
Figure 4,19: Comparison of MDL,AIC, GMDL, GAIC detection for a six element uniform linear
array with equal powered uncorrelated sources incident 10 degrees apart from each other with the
first source at 90 degrees, second source at 100 degrees, and so on. The probability of detection
error computed as a relative frequency over 200 trials is plotted as a function of the number of
samples used in the detection scheme. The signal-to-noise ratio was 10 dB. (These plots were
generated by Programs [51][52][53][54[63))
4
4
av 10 10
Number of samples of data usedChapter 4: Simulation Based Performance Comparison of DOA Estimation Algorithms
7 8
5 oat
= oy
2 06; “®-*---"-__y
Sal # of sources =2
2 02;
|)
0 10 20 3 40 10 20 3% 40
Signal-to-noise ratio (4B) Signal-to-noise ratio (4B)
o
AIC
MDL.
5 eo , Wy eK
: Pte Cee ee teeta
a” | 2° MDL
z £06 snoxee: GMDI|
= d : --* = x 0... GAIC
1 # of sources = 3 04 _x-2T¥
# of sources = 4
“0 10 20 3% 40 0 10 20 2% 4
Signal-to-noise ratio (4B) Signal-to-noise ratio(dB)
Figure 4.20: Comparison of MDL,AIC, GMDL, GAIC detection for a six element uniform linear
array with uncorrelated sources incident 10 degrees apart from each other with the first source at
90 degrees, second source at 100 degrees, and so on. The probability of detection error computed
as a relative frequency over 200 trials is plotted as a function of the signal-to-noise ratio. 100
samples of the input data was used to estimate the DOA. (These plots were generated using
Programs [55][56][571[581(64))
96Chapter 4: Simulation Based Performance Comparison of DOA Estimation Algorithms
{ | AIC
5 1
: : Mt
2 05 Hofsources = 205 of sources =?
ce rae
3 kK K HEX MORK OK y
é NAA
0 0
0 0 10 ow ww 10
Number of samples ‘Number of samples
wd ho eee
E E HR ORK ORK
2 os ¥ og totsouces=4
7 3
=o = 9
10° 10 10 i 0 10
‘Number of samples ‘Number of samples
Figure 4.21: Comparison of the modified MDL,AIC detection for a six element uniform linear
array with coherent sources of equal power incident 10 degrees apart from each other with the first
source at 90 degrees, second source at 100 degrees, and so on. The probability of detection error
computed as a relative frequency over 200 trials is plotted as a function of number of samples of
data used. The input signal-to-noise ratio was 10 dB. (These plots were generated using Programs
{9]160]163))
07Chapter 4: Simulation Based Performance Comparison of DOA Estimation Algorithms
03
E
202
#ofsourees= 1
0 +
0 10 a 30 10 20 30
Signal-to-noise ratio Signal-to-noise ratio
1 1
. 0g? ** | y 7
EO :
2 E
g 06 y\tmen |
3 04 3
3 3
2 02 ne 2 02| —+_alc ‘\
\
0 0
0 10 20 30 0 10 20 30
Signal-to-noise ratio Signal-to-noise ratio
Figure 4.22: Comparison of MDL,AIC detection for a six element uniform linear array with
coherent sources incident 10 degrees apart from each other with the first source at 90 degrees,
second source at 100 degrees, and so on. The probability of detection error computed as a relative
frequency over 200 trials is plotted as a function of the signal-to-noise ratio. (These plots were
generated using Programs [61][62][64))
98Chapter 4: Simulation Based Performance Comparison of DOA Estimation Algorithms
Figures 4.19 to 4,22 show the performance of the order estimation algorithms under the
following sample test conditions: All signals incident on the array were at an equal power
with the first signal arriving at 90 degrees, the second signal at 100 degrees, the third
signal at 110 degrees and the fourth signal at 120 degrees. For the case of the six element
linear array for which the performance was evaluated, it is found that the GMDL and
GAIC criterion provided more consistent and accurate detection when the number of
sources were less than or equal to 3. As can be seen from Figure 4.19 and 4.20, when the
number of sources is greate: than three, the AIC criterion performs better. Figure 4.19 and
4.20 also shows that the MDL criterion provides highly inconsistent and inaccurate
estimates.
For the case of coherent sources the performance of the modified AIC and MDL criterion
(see Table 2.1) was compared. Figures 4.21 and 4,22 show that the AIC technique
performs better than the MDL technique even under coherent source conditions.
4.5 Conclusion
This chapter presented the results of simulations that compared the performance of
various direction-of-artival and source order estimation algorithms. The simulation results
demonstrate that the integrated approach which combines ILSP-CMA with TLS-ESPRIT
outperforms the conventional subspace based techniques in most situations. The
performance improvement is more significant when multiple signals are incident at closely
spaced angles, low signal-to-noise ratio, and when a small number of data samples are
used (i.e small collect interval) to estimate the direction of arrival. It is also found that the
integrated approach is more robust to imperfections in building the array.
99.Chapter 5: Description of the DOA Measurement System
Chapter 5
Description of the DOA Measurement System
5.1 Introduction
In the previous chapters, we presented the background theory and simulation results
comparing the performance of different algorithms. In this chapter, we provide a detailed
description of the DOA measurment system devoloped to test the performance of different
algorithms. Both the hardware and software developed for performing direction of arrival,
measurements are detailed in this chapter. The hardware consists of a six element antenna
array, three Ariel DSP-96 boards, and an IBM compatible PC which is used to interface
with the DSP-96 and store the sampled data collected from the six channels. The system is
a narrowband array operating at 2050 MHz.
A block diagram and a photograph of the DOA measurement system is shown in Figure
5.1 and 5.2 respectively. The DOA measurement system may be divided into three
sections. The first section described in Section 5.2 consists of the antenna array and the RF
front end. The second section does the sampling and data collection, and is described in
Section 5.3. The third section does the off-line processing of the collected data. All the
DOA estimation algorithms are coded in MATLAB and runs on the host PC attached to
the system. All MATLAB codes are also available on the Sun Workstations.Chapter 5: Description of the DOA Measurement System
60 MHz PC
Inteface witl
Processing
Estimation
element 2
LNA Mixer LPF ate
ZEL-I724LN[—™]zBspD |—®) sup-s |~P) channel A
To fip= 12H Ariel
DSP-96
Board #1
LNA Mixer LPF
> >| |} apc
ZEL-1724LN ZB8PD SLPS | Ne
Pte J
[Host 486
LNA Mixer LPF ane
) zeL-1724LN/ |} zespp sip-s (P| channel A
DSP-96,
Pee ei Ariel boards
DSP-96
Board #0
LNA Mixer LPF Off-line
~ ~ | ADC
ZEL-1724LN ZB8PD SLP-5 chamnetB | {for DOA
i 2 kHz
LNA Mixer LPF Ape
fr | ZEL-1724LN zpspD [—™) sip-s |] channel
i ie = 12 iz Ariel
DSP-96,
Board #2
LNA Mixer LPF
| | st L—et ips ey ADC |
ZEL-1724LN ZB8PD SLP-S channel B
ie = 12 kz L___|
L.
8 way Power Splitter
ZB8PD-2
Local Oscillator
HP 8673G
fio = 2049.988MHz
Figure 5.1: A block diagram of the DOA measurement system.
101Chapter 5: Description of the DOA Measurement System
Figure 5.2: Photograph of the six element uniformly spaced linear array receiver used tor DOA
measurements.
5.2 Antenna Array and RF front-end
The antenna array is a six element uniformly spaced linear array. It consists of six
quaterwave monopole elements spaced one half wavelength apart. Since the array is
designed to operate at 2050 MHz, this corresponds to a spacing of 7.3 cms. Each
quaterwave monopole element is built using a 3.7 cm long hollow brass tubing. Each brass
tubing is soldered to the center conductor of a SMA jack chassis mount connector. The
SMA connectors are bolted to a brass ground plane such that the connections to the
antenna elements can be made from the bottom of the ground plane.
The RF section consists of a single downconversion stage built with no front-end filtering.
‘The RF hardware is extremely simple and built using off-the-shelf components purchased
102Chapter 5: Description of the DOA Measurement System
from Mini-Circuits, to reduce cost and complexity. Each monopole antenna element is
connected to a Low Noise Amplifier (LNA) using a flexible SMA cable. The LNA’s used
are the Mini-Circuits ZEL-1724LN, which have a gain of 20 dB in the 1.7 to 2.4 GHz
band, and a noise figure of 1.5 dB. The LNA’s have a maximum power output rating of 10
dBm and a maximum input power rating of 13 dBm. The output of each amplifier is
connected to the RF port of a Mini-Circuits ZEM-4300 mixer. The mixer has a conversion
loss of 6.65 dB, making the overall gain in the RF chain equal to 13.35 dB. The local
oscillator port of each mixer is connected to a 2049.98 MHz carrier which is generated
by the Hewlett Packard Synthesized Signal Generator HP 8673G. The same local
oscillator signal is connected to all the six mixers through an 8 way power splitter which
has two of its output ports terminated with a 50 ohm load. The LO signal power is set at 13
dBm and the power splitter contributes a loss of ]0 dB, which makes the LO signal power
at the LO port of each mixer equal to 3 dBm. The mixer downconverts the input RF signal
to a very low IF at 12 kHz. The IF output of each mixer is connected to a Mini Circuit
SLP-5 Low Pass Filter (0 -5 MHz). The output of the IF filters are connected using BNC
cables to the input of the A/D convertors on the DSP-96 boards. All the RF hardware is
mounted on a heavy gauge steel plate. The monopole elements with the ground plane is
mounted at a height of 31 cms above the plate using aluminum angle stock.
5.3 Sampling and Data Collection
Three Motorola DSP96002 based Ariel DSP-96 boards are used to sample the data from
all the six channels synchronously, store them in a buffer and save them back to a file on
the host PC. The software developed to set up the DSP-96 boards along with the host PC
to collect data from the six channels of the antenna array receiver consists of many parts.
One is the C code which runs on the host PC and the others run on the DSP-96 boards.
The host program [65] is a C-program which is developed and compiled using Borland C.
The program downloads the DSP codes on to the DSP-96 boards, and uses the m96bcs.
103Chapter 5: Description of the DOA Measurement System
library supplied by Ariel Corp. to interact with the DSP-96 boards. The code which runs
on the DSP-96 boards are written in two parts. The higher level functions such as host
interface, and other less speed sensitive functions are written in C and compiled using the
Intermetrics C compiler for DSP-96. Through function calls, the C code [68][69] accesses
macros written in a separate file [66][67] in the MOTOROLA DSP 96002 assembly
language. The assembled object modules and C compiled object module are linked along
with the required libraries and locator files to obtain a single object module [72][73]. The
locator file [70][71] specifies exactly where different segments of code need to be located
‘on the DSP-96 board memory. The locator file gives us the flexibility to locate different
sections of the code in internal, external, SRAM, DRAM, X,Y,L,P memory according to
the speed and ease of access requirements. For more information on these topics, refer to
the Ariel DSP-96 users manual [Are93].
One of the three DSP-96 boards, designated board #0, is set up as the master board, and
the other two DSP-96 boards, designated board #1 and #2, are set up as slave boards. The
master board handles all the communication with the host PC, controls the DSPNet, and
co-ordinates the sampling on all the three boards. As shown in Figure 5.1, the IF signals
obtained at the output of the RF chain from elements 1 and 6 are connected to the A/D
convertors on the master board #0, and the IF signals from elements 2,3 and 4,5 are con-
nected to the slave board #1 and #2 respectively. This rather strange assignment of DSP
boards to the different antenna elements is made to maintain compatibility with the 4 ele-
ment adaptive array program that runs on the same hardware.
The analog daughter card on each of the DSP-96 boards is programmed to generate a
hardware interrupt (IRQA) every time a dual channel sample is available at the analog-to-
digital convertor output. Samples are latched simultaneously on the A and B channels.
Whenever this hardware interrupt occurs, the processor jumps to program memory loca-
tion $8 (the IRQA ISR vector). An instruction at $8 instructs the processor to jump to the
location that handles the interrupt (ISR). The ISR fetches the two 32 bit samples from the
104Chapter 5: Description of the DOA Measurement System
two A/D convertor outputs on each board and stores them in a buffer.
Once the buffer is full, the data is transferred to the host PC on request. While data in the
master board buffer is send directly to the host, data from the slave boards are transferred
to the master board via DMA which in turn sends it to the host.
It is required that all the A/D convertors on the three DSP-96 boards sample at the same
time. Failure to do so, will produce a frequency dependent phase shift between each ele-
ment due to the delay. Therefore it is required that the sampling on all the three boards be
synchronized, This requires that the sampling clocks be tied together, and that all the three
boards start sampling (collecting data) at the same time instant. Section 5.3.1 details how
the sampling on the three boards (six channels) are synchronized.
§.3.1 Configuring Multiple DSP-96 boards for Synchronous
Sampling
When multiple DSP-96 boards are installed in the same PC, it is required to adhere to the
following guidelines:
1. Each DSP-96 board must be set at a different I/O base address.
2. There must be a record in the configuration file M96.CFG corresponding to each board
and its particular /O address.
3. Multiple DSP-96 boards may not be set to share the same DMA channel.
4. A DSP-96 board may not be set to share the same AT-bus interrupt with another DSP-96
board.
5. When using DSPnet, each board must have a unique DSPnet setting (0 to 9).
To achieve synchronous sampling of data from all the three boards, it is required to ensure
the following:
105Chapter 5: Description of the DOA Measurement System
1. Tie up the sample clocks of all the three boards.
2. Ensure that the corresponding samples from each board were picked at the same sam-
pling instants. That is, the sampling and data collection process on all three boards should
begin at the same sampling instant (Frame synchronization).
Slave Board (#0)
Master Board (#1)
Write a value to
Slave Board (#2)
slave boards
i
I
7
Y v y
Reset Buffers
Read a value Set sampling rate to 6 kHz | | Reet a aie
; |
| I
y | y ' Y
| it til inte it -
Reset Buffers 1 eee cP ako | | | Reset Butters
| | sampling clock)
; i
y Y i
targdt ,
Wait until read LQ"8S' [Read value back from | “Set [~ Wait until read
targeted slave boards | __ targeted
Write a value Reset sampling rate to |g! | Write a value
tomaster board [read desired value (64 kHz)“ [| _to master board
r
Y
y
Enable Sampling,
Enable Sampling
Enable Sampling
Figure 5.3: Flowchart illustrating the frame synchronization process
Two headers on each board provide access to the master clock and sample clock. The mas-
ter clock header, P3, has three pins. The first pin of P3 is the programmable master clock
106Chapter 5: Description of the DOA Measurement System
signal generated by each board by setting the sampling rate in software. The second pin
accepts an external master clock. Under normal operation, the programmable master clock
from each board is fed to its own A/D convertor by placing a jumper between the two
points. This clock is 256 times higher than the desired sample rate. A second header, P2,
allows similar access to the sample clock which is 1/256 of the master clock. Under nor-
mal operation, the programmable sample clock on pin 1 of P2 is fed to the sample clock
input on pin 2 of P2. To tie up the sampling clocks of the slave boards to that of the master
board, the following hardware connections need to be made. Short pins | and 2 on header
P3 on master board (see Ariel manual [Are93] for P3 location). Connect a wire from P3
pins | and 2 on the master board to P3 pin 2 on the slave board. Short pins 1 and 2 on
header P2 on master board (see Ariel manual [Are93] for P2 location). Connect a wire
from P2 pins | and 2 on the master board to P2 pin 2 on the slave board. For both connec-
tions a shielded cable type with the shield connected to pin 3 of the particular header need
to be used.
‘A proper handshaking protocol is used to ensure that the sampling and data collection pro-
cess on all three boards begin at the same sampling instant, and hence have the data frame
from all the six A/D convertors synchronized. As mentioned earlier, one of the boards is
configured as the master board and the other two as slave boards. The master board writes
a value to each of the slave boards, and that value is read by the slave boards. The slave
boards then write back a retum value to the master board which is read only at a known
precise time. After writing a value to the slave boards the master board reduces its sam-
pling rate to the lowest possible value, i.e. 6 kHz. Since the sampling clocks of all the
boards are tied together, this implies that all the boards are now operating on a 6 kHz sam-
pling clock. The sampling clock is reduced to this low rate to ensure that there is sufficient
time to execute many instructions between sampling instants. Having done this, the sam-
pling is momentarily enabled on the master board. The master board now waits for the first
sampling interrupt (IRQA) to arrive. Once the first interrupt arrives we know that we are
now at the beginning of a sample clock and have time to execute plenty of instructions
107Chapter 5: Description of the DOA Measurement System
before the arrival of the next interrupt. We actually have 167 j1s (2755 instruction cycles)
before the next interrupt arrives. At this time, sampling is disabled on the master board,
and the returned values from the slave boards are quickly read. Once the returned values
from the slave boards are read, sampling is enabled on all the three boards at the desired
rate, This ensures that the first ADC interrupt on all the three boards occur at the same
time and hence the data collected from all three boards correspond to the same time
instants. This procedure is shown to work for sampling frequencies upto 64 kHz.
The boards were checked for proper synchronization by simultaneously feeding a square
wave from a single source to all six ADC inputs (two on each board). By looking at the
zero-crossings, one can decide if all the channels are synchronized ot not. Presence of any
offset between channels will clearly indicate mis-synchronization. Figure 5.3 shows a
flowchart detailing the entire frame synchronization process.
5.4 RF Chain Phase Calibration
Since all the direction of arrival estimation techniques rely on the phase difference
between the signals on the individual antenna elements to estimate the angle of arrival, itis
necessary that the RF chain connected to each of the element does not introduce different
phase shifts. In other words, the phase difference between the signals at each element
should be solely due to the spatial separation of the antenna elements and a function of the
angle of arrival. Therefore, any phase difference between the signals on each element
caused by the RF chain has to be corrected using a calibration procedure.
The basic idea behind phase calibration is to determine the phase correction factors 0; to
compensate for phase differences in the RF chain of the six channels. The phase correction
factor 0;, is the phase shift which must be applied to channel 7 so that when a signal at
2050 MHz is introduced at the monopole element, the signals measured at the output of
the RF chain | and RF chain i have equal phase. The calibration procedure uses the setup
108Chapter 5: Description of the DOA Measurement System
shown in Figute 5.4. A 2050 MHz input signal is fed to all the six RF chains through an 8
way power splitter which has two of its unused output ports terminated with a 50 ohm
load. It was found that the power splitter introduces identical phase shifts from the input
port to each of its output port, and hence it is ensured that the signal fed to each of the RF
chain has identical phase. Once this is ensured, the phase difference between the IF signals
at the end of each RF chain is solely due to the phase differences caused in the RF chain.
IF chan. 1
—=<—| LPF
X}——|
LNA Mixer
IF chan, 2
~<«e—| LPF
x)—
Power LNA Mixer Power
Splitter : Splitter (9)
I
HP 8647¢ |ZB8PD-2 | ZBS8PD-2| HP 8673G
Synthesizer ' Synthesizer
2050 MHz 2049.988 MHz
IF chan. 6
<«— LPF
x)—— =,
LNA Mixer
LNA : Mini Circuits ZEL-1724LN_
‘Mixer: Mini Circuits ZEM-4300
LPF : Mini Circuits SLP-5
Figure 5.4: Hardware setup for the RF chain phase calibration
Using sampled data obtained from the calibration set up, the phase difference between
109Chapter 5: Description of the DOA Measurement System
each element is determined by finding the arctangent between the complex samples from
each channel. The real samples from each channel can be converted to complex samples
by performing I-Q IF to baseband downconversion on the 12 kHz IF signal. The MAT-
LAB program calib.m [74] determines the phase difference of each channel with respect to
the channel corresponding to the first antenna element, and stores the phase correction fac-
tors, 6), to a calibration file. This calibration file is read by the DOA estimation programs
and the necessary phase corrections are applied to all data collected before performing
DOA estimation using them. Phase corrections are applied by multiplying the ith channel
complex baseband signal by e 7”
5.5 Receiver Sensitivity and Dynamic Range
The sensitivity and dynamic range of the receiver is determined by the dynamic range of
the analog-to-digital convertors, the operating range of the front-end amplifiers and
mixers, the IF amplification obtained at the amplifier in front of the ADC on the DSP-96
analog daughter card.
The ADC’s on the DSP-96 board are 16 bit ADC’s and hence the signal-to-quantization
ratio is 96 dB when operated at full scale. Since we do not have an automatic gain control
circuit (AGC) to ensure that the ADC’s always operate at full scale. we are not be able to
guarantee the full dynamic range of the ADC. Depending on the input signal level relative
to the full scale level, the actual number of bits used for representing the sampled signal
will vary, and hence the signal-to-quantization noise ratio also varies.
The useful dynamic range of the ADC is determined by the signal-to-quantization noise
ratio that can be tolerated. In other words, it depends on how many bits of resolution we
need to be able to successfully process the sampled data. A simple experiment was con-
ducted to determine how many bits of resolution is required. A transmitter antenna was
10Chapter 5: Description of the DOA Measurement System
placed at a known angie and the DOA estimation algorithms were ran to determine the
direction-of-arrival. The experiment was repeated with lower and lower transmitter power
until the DOA estimation algorithm began to produce large errors (i.e the DOA estimate
vaired by greater than 2 degrees from the estimate made when the received signal was
large enough to use the full dynamic range of the ADC). It was found that, if the received
signal was large enough to swing over at least the 7 least significant bits of the ADC the
DOA estimation errors were within +/- 2 degrees.
The ADC’s on the DSP-96 board have a full scale input equal to 4V peak-to-peak. An
adjustable gain amplifier in from of the ADC can vary the sensitivity of the analog-to-dig-
ital convertors. Adjusted to maximum sensitivity, an input of 500 mv peak-to-peak will
give a full scale reading on the ADC. This corresponds to obtaining a voltage gain of 8. At
this level of sensitivity, the step size of the A/D convertor is equal to
= 500mV
16
4 7.62uV, and for a signal to swing over a 7 bit range it should have a
peak-to-peak voltage equal to A x (2”—1) = 0.96774mV. Therefore, the minimum sig-
nal level at the input of the A/D convertor board (i.e., minimum IF signal out of the LPF in
the RF chain) to do usefal processing is equal to 0.9677mV peak-to-peak.
Therefore, from the point of view of the A/D convertor, the maximum IF signal out of the
LPF is equal to 500 mV peak-to-peak (i.e equal to -8 dBm) and the minimum signal level
is equal to 0.9677 mV peak-to-peak (i.e equal to -62 dBm). Now, in order to find the
dynamic range and sensitivity of the overall receiver, we need to compute the input RF
signal power levels that will fall within the above computed maximum and minimum IF
signal levels at the ADC input.
The front-end LNA has a 20 dB gain, the mixer has a 6.65 dB conversion loss and the LPF
has a loss which is less than 1 dB, making the overall front-end gain equal to 20-6.65-1 =
12.35 dB. Therefore an RF signal power of -8 - 12.35 = -20.35 dBm will produce a full
miChapter 5: Description of the DOA Measurement System
scale reading on the ADC, and the minimum required RF input signal power is equal to -
62 -12.35 = -74.35 dBm. This corresponds to an overall dynamic range of -20.35 - (-
74.35) = 54 dB.
5.6 Processing for DOA Estimation
The sampled data from the six element linear array receiver is processed off-line for DOA
estimation. The DOA estimation algorithms were coded in MATLAB and runs on the PC
attached to the measurement system, as well as on the Sun workstations. Figure 5.5 shows
a flow chart illustrating the off-line processing steps. The processing program [75][76]
reads the user specified file which contains sampled data from the six element linear array.
The file also contains any comments which may be entered by the user during data
collection. Data samples corresponding to the six different channels are separated and
each channel data is passed through an 8-16 kHz band pass filter to remove any harmonics
or spurious signals. The BPF used is anFIR filter of order 20.
The filtered signal is then downconverted to base band using I-Q demodulation. Downcon-
version is achieved by multiplying the IF signal with a complex exponential at 12 kHz.
Downconversion is followed by low pass filtering to remove the harmonics. The low pass
filter is an FIR filter of order 20. The output of the low pass filter is the complex baseband
equivalent of the sampled IF signal.
The phase calibration correction factors are read from the file containing it, and the com-
plex baseband signals are phase shifted by the amount specified in the file. After applying
these phase corrections, the complex baseband data matrix, X, is formed. The direction-of-
arrival estimation algorithms can now be applied on the complex baseband data matrix.
For the DOA algorithms to work under coherent signal conditions, forward and conjugate
backward averaging is used as a spatial smoothing technique to decorrelate the coherent
signals. ESPRIT and the integrated ILSP-CMA with ESPRIT algorithms are used for esti-
2Chapter 5: Description of the DOA Measurement System
mating the directions of arrival from the data matrix. Figure 5.6 and 5.7 illustrates the
‘Total Least Squares ESPRIT (TLS-ESPRIT) and the integrated ILSP-CMA with TLS-
ESPRIT algorithms For more details on these algorithms see Chapter 3.
Read file /
/ containing
oH data
Pass each channel data through
8-16 kHz Band Pass Filter to
| temove any harmonics or spurious |
signals present.
1
Down convert to complex baseband
by multiplying the signal by a 12 kHz
complex exponential and passing it
through a low pass filter.
t
Get file containing calibration correction
factors and apply them to the complex
baseband signal trom each element,
Form complex baseband data matrix |
and apply the DOA estimation algorithm,
ESPRIT or ISLP-CMA with ESPRIT
Figure 5.5: Flowchart illustrating the off-line processing steps
113Chapter 5: Description of the DOA Measurement System
M=# of antenna elements (6)
L= # of element in the subarray (5)
i N= # of data samples pes element
Get X= MXN data matrix and form
the forward and conjugate backward
subarray data matrices
Xp = XCGLAN), Xp¢ = X(2:M,1:N)
Xp, = X(Ms1:2,E:N). X3, = X(Leel:1l:N)
‘Compute the spatially smoothed covariance matrix
1 HW
Rn = aH XE +X pXhb + Xap Xp)
and its eigen values {%) je.) and corresponding
eigenvectors {Vj}... Vi}.whete 2) > 2... > Ap
SS
Compute the number of sources D and obtain the signal
subspace estimate V, = [Vj. .. Vp]
Compute |
| pe MabtrL) lv
(LLL) V2L1:L)]
LIL)
vi
Compute the eigen values; of E and estimate
the angles of arrival as
{atta
Inf od
iency = 2050 MHz
h= 7.317 cms
Figure 5.6: Flowchart illustrating the TLS-ESPRIT algorithm for DOA estimation. This is
implemented in Program [76].
14Chapter 5: Description of the DOA Measurement System
[M = # of antenna elements (6)
of data samples per clement
‘Number of sources |
n= 1; iteration count
Chose an M.X’D random spatial
signature wattix A = Ay
Get the MAN data matrix X
S, = (ALA, 1) 4M
S, =
BI
XSUS, Sh)
nent!
‘Compute the spatially smoothed spatial signature covariance matrix
2
RAK) = YA EM 241, AEM —2 + 1,49")
k=kt+l
isl ‘
Rpfk) = RAK) + IRA)
Jisa pefenutation ‘matrix with all
iagonaf elements
"ros except ones in the
Estimate DOA using TLS-ESPRIT on Rygih)
CEnd_)
Figure 5. Flow chart illustrating the integrated ILSP-CMA with TLS-ESPRIT aigorithm for
DOA estimation. This is implemented in Program (75].
usChapter 6: DOA Measurement Experiments and Results
Chapter 6
DOA Measurement Experiments and Results
6.1 Introduction
This chapter describes the direction-of-arrival measurement experiments conducted using
the six element linear array operating at 2050 MHz. The objective of the experiments was
to test the functioning of the measurement system and verify the performance of DOA
estimation algorithms. The two algorithms whose performance is studied are the TLS-
ESPRIT and the integrated ILSP-CMA with TLS-ESPRIT algorithms. The experiments
were designed such that the DOA estimation accuracy of these algorithms under different
signal conditions are studied. Tests were conducted with multiple uncorrelated signals as
well as coherent multipath signals. This chapter also presents the results of the
measurements along with an analysis of the errors in measurement, and simulations which
corroborate the measurements and verify the most significant error sources.
6.2 Description of the DOA Measurements
Since the objective is to verify the functioning of the system, DOA measurements were
made under known signal conditions. That is, multiple transmitter antennas were placed at
n6Chapter 6: DOA Measurement Experiments and Results
known locations such that the direction of arrival of the plane waves at the receiver
antenna is known beforehand. According to the number and type of signals transmitted,
the experiments can be grouped into three different categories.
1. A single antenna positioned at locations with known angles with respect to the receiver
array, and transmitting a CW tone signal at 2050 MHz. Under this set up, in the absence of
multipath, at the receiver array we have a single signal (ray) incident at a known direction-
of-arrival.
2. Multiple (up to three) antennas located at different locations at known angles with
respect to the receiver array, and connected to different signal sources transmitting CW
tone signals at 2050.000, 2049.99 and 2050.001 MHz respectively. Under this set up, at
the receiver array, we have multiple uncorrelated signals (rays) incident at known
directions-of -arrival.
3. Multiple (up to three) antennas located at different positions at known angles with
respect to the receiver array, and connected to the same source transmitting a CW tone
signal at 2050 MHz. Under this set up, at the receiver array, we have multiple signals
(rays) that are perfectly correlated incident on the array at known directions-of-arrival.
This
is identical to a coherent multipath situation where the multiple rays incident on the
receiver array are attenuated and phase shifted versions of the same signal.
The DOA experiments were conducted in the Western Sizzling parking lot near the MPRG
building. The antenna array receiver was placed approximately half way along one edge of
the parking lot. Ata distance of 60 feet from the receiver array location, using a protractor
and a rope, points were marked out every five degrees along a semicircle on the field. The
angles were measured with respect to the receiver array axis. These marking were made so
that we can position the transmitter antennas at known angles with respect to the receiver
array axis.
117Chapter 6: DOA Measurement Experiments and Results
A transmitter receiver (T-R) separation of 60 feet was chosen as a compromise between
various factors. First, to keep the transmit power to a minimum it is required to reduce the
TR separation distance. Also, since spatially separated antennas need to be fed from the
same source via coaxial cables (for the case of coherent signals), keeping a large T-R
separation distance implies that we will need longer cables to obtain the required angular
separation between multiple transmitter antennas. At the same time, we cannot keep the T-
R separation to be very small, as this will invalidate the plane wave assumptions.
‘The DOA estimation algorithms described in chapter 3 are based on modeling the incident
signals as plane waves, and assumes that the propagation conditions do not change as the
ray traverses the extent of the array. If the incident wave is a spherical wave instead of a
plane wave, the direction of arrival of the wave with respect to each individual receiver
antenna element will be different from one another, and the difference is directly
proportional to the physical dimension of the array and inversely proportional to the radius
of the spherical wave (which is equal to the T-R separation distance). A good rule of
thumb to ensure the validity of the plane wave assumption is to have the transmitting
source at a distance of at least 50 times the maximum dimension of the receiver array
[Don93]. Under such conditions, the difference between the DOA's relative to the closest
and most distant antenna elements is less than about I degree.
Since the array operates at 2050 MHz (2. = 0.14634 m) and has six elements, each spaced
one half wavelength apart, the total extent of the array is equal to 2 = 0,36585m.
Therefore, the T-R separation should be at least 50 x 0.36585 = 18.2925m, which is
approximately equal to 60 feet.
As discussed in section 5.5, the minimum required received power for the antenna array
receiver is approximately -75 dBm. Assuming free space propagation, the minimum
required transmitter power can be calculated using the Friis free space formula [Rap95]
18,Chapter 6: DOA Measurement Experiments and Results
P,G,Gie
(4nya
(6.1)
For a T-R separation distance of 60 feet, using equation (6.1), at 2050 MHz the minimum
required transmitter power level is equal to -10 dBm. For all the experiments, the
transmitter signal generator power was kept at 10 or 15 dBm. With the cable connecting
the source to the antenna having a loss of 10 dB, the effective transmitted power was equal
to 0 or 5 dBm.
Quarterwave monopoles were used as transmitting antennas. The quarter wave monopoles
were constructed using 3.7 cm long hollow brass tubing soldered to an SMA jack which
was bolted to a brass ground plane. The monopole antennas were mounted on PVC pipes
at a height of 1.37 m (4.5 fi.). The receiver array was kept on a trolley such that the array
was 1.37 m above the ground.
6.2.1 Measurements with a Single Transmitter Antenna
A CW tone signal at 2050 MHz was generated using the HP83620 signal generator. The
signal generator output power was set to 15 dBm, and the RF signal was fed to a monopole
antenna using a flexible SMA cable which has a 10 dB loss. The transmitter antenna was
placed at different positions such that the receiver array has a line-of-sight wave impinging
on it at known angles. Measurements were made at every five degrees, beginning with 0
degrees up to 180 degrees. For each DOA measurement, 2048 (2 K) samples of the signal
was collected from each antenna element at a sampling rate of 64 kHz[65].
Figure 6.1 shows the results of the DOA measurements made with a single transmitter
antenna. The RMS error is plotted as a function of the angle of arrival. The RMS error for
the single plane wave case is simply the difference between the estimated and actual DOA.
The actual DOA is taken as what is measured using a protractor and rope and marked on
the field. It should be noted that inaccuracies in the measurement of actual DOA will also
119Chapter 6: DOA Measurement Experiments and Results
manifest as DOA estimation error in all the plots presented. Since angle marking were
made at a distance of 60 feet from the receiver array, one degree corresponds to a distance
of approximately one foot. Therefore, if the marked positions are in error by one foot, it
will correspond to an error in actual DOA measurement of approximately one degree.
As seen from Figure 6.1, for the case of a single transmitter, the DOA estimated by the two
algorithms (TLS-ESPRIT and ILSP-CMA with TLS-ESPRIT), are within +5 degrees of
the actual DOA, except for the case when the angle of arrival is 0 or 180 degrees (endfire).
where the error is within 10 degrees.
7 se. TLS-ESPRIT
B° ILSP-CMA with TLS-ESPRIT
&
Ze
B7
5
6
|
V
Angi of Arfival in d>ees
RMS Error in DOA E:
».
20 40 120. «140~=«160~—~SC«8D
DOA Measurement Results - RMS Error in DOA Estimation as a function of the
angle of arrival. RMS error is computed as the absolute difference between the estimated and
actual angle of arrival.
6.2.2 Measurements With Multiple Transmitter Antennas Fed
From Uncorrelated Sources
These measurements were set up to test the performance of the DOA estimation
120Chapter 6: DOA Measurement Experiments and Results
algorithms under uncorrelated signal conditions. ‘This is similar to having signals from
multiple users impinging on the array. DOA measurements were made using two and three
transmitters, with each of them spatially separated by different angles. As shown in
Chapter 4, the accuracy of the DOA estimation algorithms is a function of the angular
separation between the multiple signals and also the angle of arrival itself (broadside
giving better performance than endfire). The DOA measurement experiments were
designed such that the performance of the DOA estimation algorithms could be studied as
a function of the angular separation between multiple (two or three) signals and also the
angle of arrival itself.
Source #1
HP 83620
flexible cable
Source #2
HP 8647C
Receiver Array
Source #3
flexible caBE Flake 6062
Figure 6.2: Ilustration of a the DOA measurement setup with three uncorrelated signals being
transmitted at different angles with respect to the receiver array location,
The three uncorrelated signals were generated using separate signal sources at different
frequencies. The first signal is a CW tone at 2050 MHz generated by the HP83620 signal.
generator. The second signal is a CW tone at 2049.999 MHz generated by the HP8648C
signal generator, and the third signal is a CW tone at 2050.001 MHz generated by the
121Chapter 6: DOA Measurement Experiments and Results
Fluke 6062 signal generator. All the signal generators were set up for a power output of 10
dBm (the maximum power output that can obtained from the Fluke 6062 is 10 dBm).
Since all the sources were connected to their antennas using a cable of 10 dB loss, the
effective transmitter power was equal to 0 dBm (1mW). Figure 6.2 illustrates the
experimental setup for the case of three uncorrelated signals.
B 99 TLS-ESPRIT :
3 ___ILSP-CMA fo
= 20] - ;
é with TLS-ESPRIT :
5 70 a \ 4
g First signal arrives at 10 deg. 7
2 0) Second signal aves at 10-40 deg .
GZ s0- \
a
= a0] .
E 20-
a .
g ‘
————
x 25 30 35
5
Angle of | Separation (a0) ‘between the two signals in degrees
Figure 6.3: DOA Measurement Results - RMS Error in DOA estimation as a function of angle of
separation between the two uncorrelated sources. The first signal arrives at 10 degrees and the
second signal arrives at 10+A0 degrees.
Figures 6.3 to 6.9 show the results of the DOA measurements performed with multiple
uncorrelated signals. In ali the figures, the RMS error in DOA estimation is plotted as a
function of the angular separation between adjacent signals, 1 Ojo.» Op are the actual
. Oy are the angles estimated by the
measured angles of arrival of the D signals, and 6) ,.
algorithms, the RMS error (RMSE) in DOA estimation is defined as
(6.2)
D
1
RMSE = Db» (84
122Chapter 6: DOA Measurement Experiments and Results
2 a TLS-ESPRIT
= ost ILSP-CMA with TLS-ESPRIT
= : First signal arrives at 30 deg.
2 a . Second signal arrives at 30+A0 deg.
£ |
z :
15 :
< 1
3 :
a \
= 10 \
5 :
Eos woos e ee SSS ese
g
gf. —e —__—!
eos 7 75 3 25 30
Angle of Separation (Ad) between the adjacent signals in degrees
Figure 6.4: DOA Measurement Results - RMS Error in DOA estimation as a function of angle of
‘separation between the two uncorrelated signals. The first signal arrives at 30 degrees and the
second signal arrives at 30+ degrees.
~ 30 :
2 : __ TLS-ESPRIT
ae \ ____1LSP-CMA with TLS-ESPRIT
= : First signal arrives at 30 deg.
& \ Second signal arrives at 30+A0 deg,
z” \ Third signal arrives at 30+240 deg.
2 i. :
< \
Q -
a mse
Z 10 te
E To
Es wk
a
=
4
9,
5 25
70 15 20
Angle of Separation (AQ) between the adjacent signals in degrees
Figure 6.5: DOA Measurement Results - RMS Error in DOA estimation as a function of angle of
separation between adjacent uncorrelated signals. The first signal arrives at 30 degrees, the
second signal arrives at 30+A0 degrees and the third signal arrives at 30+2A0 degrees.
123Chapter 6: DOA Measurement Experiments and Results
: ’ a TLS-ESPRIT
7h ILSP-CMA with TLS-ESPRIT
a First signal arrives at 60 deg.
sf Second signal arrives at 60+A0 deg,
RMS Error in DOA Estimation (in deg.)
v0 75 2 25 30 35 30
Angle of Separation (AQ) between the adjacent signals in degrees
Figure 6.6: DOA Measurement Results - RMS Error in DOA estimation as a function of angle of
separation between the two uncorrelated signals. The first signal arrives at 60 degrees and the
second signal arrives at 60+A0 degrees.
—..._ TLS-ESPRIT
me ILSP-CMA with TLS-ESPRIT
Te First signal arrives at 60 deg.
a Second signal arrives at 60+0 deg.
E és Third signal arrives at 60+240 deg.
2 20} as Saeed
3 ~s.
Sis 7
a .
= A
= 10 me
5 ~
& Te
a, oe
g ~
= oo =
a :
10 75 20 25 30
Angle of Separation (A®) between the adjacent signals in degrees
Figure 6.7; DOA Measurement Results - RMS Error in DOA estimation as a function of angle of
separation between adjacent uncorrelated signals. The first signal arrives at 60 degrees, the
second signal arrives at 60+A0 degrees, and the third signal arrives at 60+2A0 degrees.
124Chapter 6: DOA Measurement Experiments and Results
2 7 ._ TLS-ESPRIT
S ost ___ILSP-CMA with TLS-ESPRIT
2 First signal arrives at 90 deg.
S oS Second signal arrives at 90+A0 deg.
2| ™
£ .
1 BL
RMS Error in DOA E:
5 70 cr ee 30
Angle of Separation (A®) between the adjacent signals in degrees
Figure 6.8: DOA Measurement Results - RMS Error in DOA estimation as a function of angle of
separation between the two uncorrelated signals. The first signal arrives at 90 degrees and the
second signal arrives at 90+40 degrees.
7 ._ TLS-ESPRIT
ILSP-CMA with TLS-ESPRIT
First signal arrives at 90 deg
Second signal arrives at 90+40 deg.
Third signal arrives at 904240 deg.
RMS Error in DOA Estimation (in deg.)
5 76 25 30
Angle of Separation (A@) between the adjacent signals in degrees
Figure 6.9: DOA Measurement Results - RMS Error in DOA estimation as a function of angle of
separation between adjacent uncorrelated signals. The first signal arrives at 90 degrees, the
second signal arrives at 90+A0 degrees, and the third signal arrives at 90+240 degrees.
125Chapter 6: DOA Measurement Experiments and Results
For all the test situations, the angular separation between all the adjacent transmitter
antennas were kept equal. Figure 6.3 shows the RMS error in DOA estimation as a
function of angular separation A@, with the first signal arriving at 10 degrees and the
second signal arriving at 10 + AO degrees. Figures 6.4, 6.6, and 6.8 show the same with the
first signal arriving at 30, 60, and 90 degrees respectively. Figures 6.5, 6.7, and 6,9 show
the RMS error in DOA estimation error for the case of three uncorrelated signals.
As seen from Figures 6.3 to 6.9, the integrated ILSP-CMA with TLS-ESPRIT algorithm is
able to estimate the directions-of-arrival for the case of two signals within less than +5
degrees for most of the angles and angular separations at which the measurements were
made. The TLS-ESPRIT algorithm produced large errors when there are two uncorrelated
signals, one arriving at 10 degrees and the other arriving at angle separated from the first
signal by less than 30 degrees (see Figure 6.3). The TLS-ESPRIT algorithm also produced
large errors when there are three uncorrelated signals (see Figures 6.5., 6.7, and 6.9).
Section 6.3 provides an analysis of these errors.
6.2.3. Measurements With Multiple Transmitter Antennas Fed
From a Single Source
These measurements were set up to test the performance of the DOA estimation
algorithms under correlated signal conditions. This is similar to having multipath signals
from a single user incident on the receiver array. DOA measurements were made with two,
three, and four signals, with each of them separated from one another by different angles.
‘As mentioned in Section 6.2.2, the experiments are designed to verify the performance of
the DOA algorithms as a function of angle of arrival and the angular separation between
adjacent sources.
All the required signals were generated by a single source. The HP 83620 signal generator
‘was set up to generate a CW tone at 2050 MHz at an output power level of 15 dBm. This
126Chapter 6: DOA Measurement Experiments and Results
signal was fed to multiple transmitter antennas via power splitters and cables. Equal loss
cables were used such that the power transmitted by all the antennas are approximately
equal. With the splitter and the cables providing an attenuation of approximately 13 dB,
the power transmitted by the antennas was approximately equal to 2 dBm,
Source #1
HP 83620
Splitter
flexible cable
flexible cable,
Receiver Array
flexible cable
Figure 6.10: Illustration of a the DOA measurement setup with three correlated (fully coherent)
signals being transmitted at different angles with respect to the receiver array location.
Figures 6.11 to 6.15 show plots of the RMS error in DOA estimation as a function of angle
of separation for various angles of arrival. It is seen that the performance of the TLS-
ESPRIT and the integrated ILSP-CMA with TLS-ESPRIT is more or less the same. For
the case of two coherent signals being transmitted, both the algorithms estimate the angle
of arrivals within an RMS error of 5 degrees for most cases (see Figures 6.11, 6.12, and
6.14), When the angle of separation between the coherent signals is only five degrees, the
rms error is greater than 5 degrees. For the case of three coherent signals being
transmitted, both the algorithms produce large errors except when the signals are
separated by more than 30 degrees (see Figures 6.13 and 6.15). Section 6.3 provides a
detailed analysis of the errors in DOA measurement.
127Chapter 6: DOA Measurement Experiments and Results
TLS-ESPRIT
: ILSP-CMA
Tee with TLS-ESPRIT|
First signal arrives at 30 deg.
Second signal arrives at 30+A0 deg|
RMS Error in DOA Estimation (in deg.)
a
Angle of Separation (40) between the adjacent signals in degrees
Figure 6.11: DOA Measurement Results - RMS Error in DOA estimation as a function of angle
of separation between the two fully correlated (coherent) signals. The first signal arrives at 30
degrees and the second signal arrives at 30+A0 degrees.
* ___ TLS-ESPRIT
ILSP-CMA
with TLS-ESPRIT|
First signal arrives at 60 deg.
Second signal arrives at 60+A6 deg|
RMS Error in DOA Estimation (in deg.)
so SSCS BO
Angle of Separation (AQ) between the adjacent signals in degrees
Figure 6.12: DOA Measurement Results - RMS Error in DOA estimation as a function of angle
of separation between the two fully correlated (coherent) signals. The first signal arrives at 60
degrees and the second signal arrives at 60+A6 degrees.
128Chapter 6: DOA Measurement Experiments and Results
__ TLS-ESPRIT
ILSP-CMA with TLS-ESPRIT
First signal arrives at 60 deg.
Second signal arrives at 60+40 deg.
Third signal arrives at 60+240 deg.
RMS Error in DOA Estimation (in deg.)
'S 70 15 20 3 30
Angie of Separation (AQ) between the adjacent signals in degrees
Figure 6.13: DOA Measurement Results - RMS Error in DOA estimation as a function of angle
of Separation between the two fully correlated (coherent) signals. The first signal arrives at 60
degrees, the second signal arrives at 60+A0 degrees, and the third signal arrives at 60+2A0
degrees.
eee
‘LS-ESPRIT
ILSP-CMA with TLS-ESPRIT
First signal arrives at 90 deg.
Second signal arrives at 90+A0 deg.
RMS Error in DOA Estimation (in deg.)
ss 08 aa B0
Angle of Separation (A@) between the adjacent signals in degrees
Figure 6.14: DOA Measurement Results - RMS Error in DOA estimation as a function of angle
of separation between the two fully correlated (coherent) signals. The first signal arrives at 90
degrees and the second signal arrives at 90+A0 degrees.
129Chapter 6: DOA Measurement Experiments and Results,
—_—__ TLS-ESPRIT
______ ILSP-CMA with TLS-ESPRIT
First signal arrives at 90 deg.
Second signal arrives at 90+A0 deg.
Third signal arrives at 904240 deg.
RMS Error in DOA Estimation (in deg.)
40
Angle of Separation (A®) between the adjacent signals in degrees
Figure 6.15: DOA Measurement Results - RMS Error in DOA estimation as a function of angle
of Separation between the two fully correlated (coherent) signals. The first signal arrives at 90
degrees, the second signal arrives at 90+A0 degrees, and the third signal arrives at 90+20
degrees.
6.3 Sources of Error in DOA Estimation
The DOA measurement results presented in section 6.2 show that both the TLS-ESPRIT
and the integrated ILSP-CMA with TLS-ESPRIT algorithm were able to estimate the
directions-of -arrival in most cases within an accuracy of +5 degrees. For the case of
coherent signals, the algorithm produced large errors when more than two transmitter
antennas were used. It was also found that the TLS-ESPRIT algorithm performed worse
than the integrated approach in the case of multiple uncorrelated signals. This was true,
especially when the signals arrived close to the endfire angles. For coherent signals, the
performance of both the algorithms was more or less equal
As seen from the simulation results presented in Chapter 4, under ideal conditions, the
TLS-ESPRIT algorithm and the integrated ILSP-CMA with TLS-ESPRIT algorithm
130Chapter 6: DOA Measurement Experiments and Results
should perform significantly better than it did in the DOA experiments. The poorer
performance of these DOA estimation algorithms may be attributed to several causes.
First, the ESPRIT approach requires that the array be comprised of two identical
overlapping subarrays. Therefore, if the two subarrays are not identical due to physical
perturbations of the array geometry, the performance of the algorithm can be significantly
degraded [Sch93]. Second, any phase difference induced in the RF chains of the six
channels can cause errors in the DOA estimate. Though the phase differences between the
individual RF chains are compensated for through the RF chain calibration procedure (see
Section 5.4), phase variations with time due to component drift can degrade the
performance of the system. Third, multipath that already exists in the measurement area
may overload the array, and lead to the failure of the DOA estimation algorithms. That is,
for every transmitter antenna positioned at a known angle, multipath reflections in the
channel would produce multiple rays that will arrive at the array at unknown angles, and
this would cause errors in the estimate of the direction-of-arrival of the known signal.
Further, mutual coupling between the array elements may also cause errors.
Simulations were conducted to investigate the effect of these error sources on the DOA
estimation algorithms for all the signal scenarios under which the measurements were
conducted. The goal was to understand the reasons for the failure of the algorithms under
those situations where large errors were produced. Separate simulations were run to study
the effect of (1) perturbation in antenna element locations (2) RF phase calibration errors
and (3) multipath existing in the channel.
Simulation of Perturbation in Antenna Element Locations: Ideally, the antenna
elements are each supposed to be spaced a half wavelength apart. Therefore, if the first
clement location is taken as the origin ofthe coordinate axis, and the array is aligned along
a straight line, the individual element locations may be expressed in vector form as
or2343] "In order to study the effect of perturbation of element locations,
simulations were ran using a randomly perturbed location vector,
131Chapter 6: DOA Measurement Experiments and Results
0
1 + 0.002(rand—0.5)
yprenarendas
4+ 0.008(rand—0.5)
5 +0.065(rand —0.5)
instead of the ideal location vector (“rand” is a uniform random number between 0 and 1,
and hence rand-0.5 is a uniform random number between -0.5 and 0.5). Notice that the
random perturbations on individual elements are weighted by different coefficients. These
weighting coefficients are proportional to the actual perturbation on the array element
locations, so that the simulations would duplicate as closely as possible, the actual
perturbations present in the array. For example, in the actual array, it was found that the
fourth and sixth element were displaced the most from their ideal location, whereas the
second and fifth were the least perturbed. Hence the fourth and sixth element in (6.3) have
larger coefficients than the second and fifth element. After incorporating these
perturbations into the simulation, the DOA estimation algorithms were tested to determine
the error in DOA estimation for each of the different scenarios at which the actual
measurements were made, Section 6.4 presents the results obtained through 100 Monte
Carlo runs of these simulations.
Simulation of RF Chain Phase Calibration Errors: Though the phase difference
between the six RF channels in the DOA measurement system may be measured and
compensated for, it was found that the phase difference varies with time. The time varying
phase difference between the RF channels, and the errors in the measurement of the phase
difference may lead to performance degradation of the DOA measurement system. Phase
differences between the RF channels were measured several times over a period of four
hours to observe the amount of phase drift in each channel. Phase difference
measurements were made using the RF chain phase calibration procedure detailed in
Section 5.4. To simulate the random phase variations in the RF chain, the complex
baseband array input data vector X was multiplied by a random phase shift vector
132Chapter 6: DOA Measurement Experiments and Results
€XP(JP pang)» Where
0 7]
0.0497 (rand —0.5)|
0.0279(rand —0.5) (64)
0,0458(rand—0.5)
0.0835(rand—0,5)
0.0590(rand—0.5)
Prana =
The coefficients in (6.4) were chosen as the difference between the minimum and
maximum phase value obtained during the four hour observation period. Since the phase
drifts observed in the six RF channels were different from one another, the coefficients in
(6.4) are not all equal. Notice that some RF channels have larger phase drifts than the
others. Therefore, using the random phase drift given by (6.4), the simulations can
duplicate the actual phase drift observed in the measurement system. After incorporating
these perturbations into the simulations, the DOA estimation algorithms were tested to
determine the error in DOA estimation for each of the different scenarios at which the
actual measurements were made. Section 6.4 presents the results obtained through 100
Monte Carlo runs of these simulations.
Simulation of Multipath due to Reflections in Channel: When a transmitter antenna is
placed at a known angle with respect to the receiver array, apart from the line of sight ray,
reflected rays from the surrounding buildings will also be incident on the array. The
presence of these reflected rays can cause errors in the estimation of the direction of the
desired line-of-sight component. The errors caused by the existence of these multiple rays
will depend on the number of multipath rays, their angles-of-arrival, and relative power
levels.
In order to investigate if the multipath that exists at the DOA measurement site was a
cause for the large errors seen in some of the DOA measurements, we made a study of the
multipath structure of the measurement site. The multipath structure of the measurement
133Chapter 6: DOA Measurement Experiments and Results
site was characterized through power delay profile measurements and ray tracing.
-70}-
-75|
80)
Power (dBm)
-95)
100
Figure 6.16:
Tx located at 20 deg.
with respect to Rx.
‘Tx located at 120 deg.
with respect t0 Rx
ower (dBm)
&
R
500 10001500
Delay (ns)
0 500 1000-1500
Delay (ns)
85]
~90}°
Tx located at 90 deg.
with respect to Rx
500 1000*1500
Delay (ns)
Tx located at 150 deg.
with respect to Rx
Power (dBm)
i
eb \
See bob
SR SaaS
0 500 10001500
Delay (ns)
Figure showing sample power delay profiles corresponding to four transmitter
locations. The first profile was measured when the transmitter was located at 20 degrees with
respect to the receiver, and the subsequent ones were measured when it was located at 120, 90,
and 150 degrees respectively.
134Chapter 6: DOA Measurement Experiments and Results
Power delay profile measurements using the MPRG channel sounder [New96] were
conducted at the measurement site to check for the existence of multipath components.
The channel sounder receiver was positioned at the same point at which the antenna array
receiver was placed for DOA measurements. The channel sounder transmitter was kept at
different angles with respect to the receiver, and power delay profiles were measured at
every 10 degrees. It was found that for each of the transmitter locations, there were at least
one multipath component, arriving with a delay of approximately 200 ns, which was about
10 to 20 dB below the line-of-sight component. For many locations, there were additional
multipaths too. Figure 6.16 shows the measured power delay profiles for some of the
transmitter antenna locations.
As described above, the power delay profile measurements show that multipath does exist
at the measurement site. In order to study the effect of these multipath rays on DOA
estimation algorithms, we also need to know the angle of arrival of each of these multipath
rays. We used ray tracing to determine the angle of arrival of the multipath rays. To
facilitate ray tracing, the measurement site was modeled using MPRG’s site specific
propagation prediction software [Kau95]. The dimensions of the four buildings around the
measurement area were measured, and the buildings were modeled as four rectangular
concrete boxes. The terrain was assumed to be flat.
For each transmitter antenna location, the number of multipath reflected rays incident on
the array, their angles of arrival and the relative power levels were determined using ray
tracing techniques. It was found that for each of the transmitter antenna locations at which
measurements were made, apart from the line of sight (LOS) ray and a ground reflected
ray, thete were at least three other significant (about 15 to 20 dB below the line-of-sight
ray) multipath rays, arriving at the receiver. For example, Figure 6.17 shows that when a
single transmitter antenna is placed at an angle of 20 degrees with respect to the receiver
array axis, there is an LOS ray, a ground reflected ray, and three single bounce multipath
rays at 76.9, 177and -77 degrees arriving at the receiver array. It was also predicted thatChapter 6: DOA Measurement Experiments and Results
there are many more multipath components due to multiple bounces, but all of them were
at least 30 dB below the LOS component.
iyeare Cente
Figure 6.17: Figure showing the single bounce multipath rays incident on the array when a
single transmitter antenna is located at 20 degrees with respect to the receiver array axis. Using
ray tracing it was found that for all transmitter locations, there were at least three multipath
components that were about 15 to 20 dB below the line-of-sight component.
Through ray tracing and power delay profile measurements, it is seen that multipath does
exist at the DOA measurement site, Simulations were conducted incorporating, three
additional multipath components for each LOS ray that is incident on the array (the
number of LOS rays is equal to the number of transmitter antennas used). Each of these
136Chapter 6: DOA Measurement Experiments and Results,
components were set up to have a power level 15 dB below the LOS component and
angles of arrival as predicted by the ray tracing software. After incorporating these
multipath rays into the simulation, the DOA estimation algorithms were tested to
determine the error in DOA estimation for each of the different scenarios at which the
actual measurements were made. Section 6.4 presents the results of the simulations.
6.4 Results of the Simulation Based Study
| TLS-ESPRIT
ot -.-. ILSP-CMA with TLS-ESPRIT
8 1 0 With RF chain Phase Errors
at * With Sensor Location Errors
1 + With Multipath in Channel
RMS Error in DOA Estimation (in deg.)
80 100 +120 +140 160 180
|
0 20 40 60
Angle of Arrival in degrees
Figure 6.18: Simulation Results - Figure illustrates the RMS error in estimating the DOA of a
single source under RF chain phase errors, sensor location errors, and multipath in channel. (The
plots were generated by Programs (79](81][83](85][871[69][90))
In Section 6.2, we presented the results of the DOA measurements that were conducted.
This section presents the results of simulations that recreated the measurement scenarios
incorporating the possible sources of error. Separate simulations were run for each of the
three possible sources of error, and DOA estimation error due to each of the three possible
137Chapter 6: DOA Measusement Experiments and Results
sources of error are presented separately. Figures 6.18 to 6.30 present simulation results
corresponding to each of the measurement results presented in Section 6.2 (Figures 6.1,
6.3 through 6.9, 6.11 through 6.15).
Figure 6.18 shows the RMS error in DOA estimation for the case of a single transmitting
antenna. It is clearly seen that perturbation in antenna element location, RF chain phase
dri
angles (9 and 180 deg.). It is also seen that RF chain phase drift and perturbation in
as well as the existence of multiple reflected rays produce large errors at the end-fire
antenna element location does not cause any significant errors in the estimated DOA at
angles other than the endfire angles. The existing multipath in the channel is the most
prominent source of error.
o With RF chain Phase Errors TLS-ESPRIT
With Sensor Location Errors 7 ILSP-CMA
With Multipath in Channel with TLS-ESPRIT
7) First signal arrives at 10 deg.
BEES SERIE SE oe Second signal arrives at 10+A@ deg
eB
RMS Error in DOA Estimation (in deg.)
io 8 25 35 30
Angie of Separation (AQ) between the adjacent signals in degrees
Figure 6.19: Simulation Results - Figure illustrates the RMS error in estimating the DOA of two
uncorrelated signals arriving at 10 and 10+A0 degrees respectively, under RF chain phase errors,
sensor location errors, and multipath in channel. (The plots were generated by Programs
[78][80][82](84)[86][88][90})
Figures 6.19 through 6.25 shows the results for the case of two or three transmitting
antennas fed from uncorrelated sources, and located at different angles and angular
138