Thesis 2020

You might also like

Download as pdf or txt
Download as pdf or txt
You are on page 1of 88

Master of Science Thesis in Electrical Engineering

Department of Electrical Engineering, Linköping University, 2020

Time Synchronization of
TDOA Sensors Using a
Local Reference Signal

Alfred Hult
Master of Science Thesis in Electrical Engineering
Time Synchronization of TDOA Sensors Using a Local Reference Signal:
Alfred Hult
LiTH-ISY-EX--20/5318--SE

Supervisor: Johannes Lindblom


Swedish Defense Research Agency (FOI)
Unnikrishnan Kunnath Ganesan
isy, Linköping University

Examiner: Mikael Olofsson


isy, Linköping University

Division of Communication Systems


Department of Electrical Engineering
Linköping University
SE-581 83 Linköping, Sweden

Copyright © 2020 Alfred Hult


Sammanfattning
Synkronisering av ett distribuerat nätverk av sensorer för tidsdifferensmätning
(TDOA-sensorer) kan göras med referenssignaler från GPS-satelliter. Den meto-
den ger hög noggrannhet, men är känslig för störning, och är inte tillräckligt på-
litlig för att användas i militära tillämpningar. En lösning som inte beror på någ-
ra signaler utsända från externa aktörer är att föredra. Ett sätt att åstadkomma
detta är att använda referenssignaler som sänds ut från en UAV. En UAV är lämp-
lig eftersom endast lokal synkronisering för ett geografiskt begränsat område är
nödvändig. Den lokala synkroniseringen görs genom estimering av tidsfördröj-
ningen mellan sändningen och mottagandet av en referenssignal. Den estimerade
tidsfördröjningen kan användas för att upptäcka drivningar i TDOA-sensorernas
klockor. Detta arbete analyserar vanliga referenssignaler, för att utvärdera vilka
som ger tillräckligt hög noggrannhet vid estimering av tidsfördröjningar, och vil-
ka egenskaper hos signalerna som påverkar noggrannheten hos estimeringarna
mest. Simuleringarna visar att prestandan kan uppnå samma noggrannhet som
synkronisering gjord mot GPS för flera typer av signaler. Det är viktigare att öka
bandbredden än att öka signallängden eller signal-till-brus-förhållandet för att
höja noggrannheten hos estimeringen.

iii
Abstract
Synchronization of distributed time difference of arrival (TDOA) sensor networks
can be performed using reference signals from GPS satellites. This method pro-
vides high accuracy, but is vulnerable to jamming, and is not reliable enough to
be used in military applications. A solution that does not depend on any sig-
nals transmitted from external actors is preferred. One way to achieve this is to
use reference signals transmitted from a UAV. A UAV is suitable since only lo-
cal synchronization for a geographically restricted area is necessary. The local
synchronization is achieved by estimating the time-delay between the transmis-
sion and reception of a reference signal. The estimated time-delay can be used
to detect drifts in the clocks of the TDOA sensors. This thesis analyzes com-
mon reference signals, to evaluate which provide high accuracy for time-delay
estimation, and what properties of the signals influence the estimation accuracy
the most. The simulations show that the time-delay estimation performance can
reach the same accuracy as synchronization against GPS for different types of sig-
nals. An increased bandwidth is more important than an increased signal length
or signal-to-noise ratio to improve the estimation accuracy.

v
Acknowledgments
Thanks
to my supervisor at FOI, Johannes, who showed patience while explaining things
I should already know.

Thanks
to Per at FOI, for helping with the laboratory tests, and Tommy, for the code used
to generate the Gold sequences.

Thanks
to the colleagues at FOI, employees and other thesis students, for making it such
a joy to be at the office.

Thanks
to my supervisor Unnikrishnan and examiner Mikael at the university, for feed-
back on the report, and help along the way.

Linköping, 2020
Alfred Hult

vii
Contents

List of Figures xii

List of Tables xv

Notation xvii

1 Introduction 1
1.1 Purpose . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.2 Problem Statements . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.3 Limitations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4

2 Theoretical Background 5
2.1 Positioning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
2.1.1 TOA . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
2.1.2 TDOA . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.2 Methods for Synchronization . . . . . . . . . . . . . . . . . . . . . . 8
2.2.1 GPS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
2.2.2 DCF77 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
2.2.3 Wired Synchronization . . . . . . . . . . . . . . . . . . . . . 10
2.3 Signal Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
2.3.1 Time-delay in Frequency Domain . . . . . . . . . . . . . . . 11
2.4 Free-space Propagation . . . . . . . . . . . . . . . . . . . . . . . . . 11
2.5 Time-delay Estimation . . . . . . . . . . . . . . . . . . . . . . . . . 12
2.5.1 Correlation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
2.5.2 Cramér-Rao Bound for Time-delay Estimation . . . . . . . 13
2.5.3 Maximum Likelihood Estimator . . . . . . . . . . . . . . . . 14
2.5.4 Sub-sample Accuracy . . . . . . . . . . . . . . . . . . . . . . 15
2.6 Reference Signals . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
2.6.1 Frequency Band . . . . . . . . . . . . . . . . . . . . . . . . . 16
2.6.2 Signal-to-noise Ratio . . . . . . . . . . . . . . . . . . . . . . 16
2.6.3 Ambiguity Function . . . . . . . . . . . . . . . . . . . . . . 16
2.6.4 General Expression for the Cramér-Rao Bound for Time-
delay Estimation . . . . . . . . . . . . . . . . . . . . . . . . 17

ix
x Contents

2.6.5 Pulse . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
2.6.6 Chirp . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
2.6.7 White Gaussian Noise . . . . . . . . . . . . . . . . . . . . . 21
2.6.8 Gold Sequences . . . . . . . . . . . . . . . . . . . . . . . . . 22

3 Method 25
3.1 Setup . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
3.1.1 Parameter Restrictions . . . . . . . . . . . . . . . . . . . . . 26
3.2 Analytical Evaluation . . . . . . . . . . . . . . . . . . . . . . . . . . 26
3.3 Simulations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
3.3.1 Time-delay Estimation in the Simulations . . . . . . . . . . 27
3.3.2 Generation of Gold Sequences . . . . . . . . . . . . . . . . . 27
3.4 Laboratory Tests . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28

4 Results 31
4.1 Cramér-Rao Bounds for Time-delay Estimation . . . . . . . . . . . 31
4.2 Simulations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
4.2.1 Varied Bandwidth . . . . . . . . . . . . . . . . . . . . . . . . 33
4.2.2 Varied Signal Length . . . . . . . . . . . . . . . . . . . . . . 35
4.2.3 Varied Signal-to-noise Ratio . . . . . . . . . . . . . . . . . . 37
4.3 Tests in Laboratory . . . . . . . . . . . . . . . . . . . . . . . . . . . 39

5 Discussion 43
5.1 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
5.1.1 Estimation Accuracy of the Waveforms . . . . . . . . . . . . 43
5.1.2 Effect of the Signal Properties . . . . . . . . . . . . . . . . . 44
5.2 Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
5.2.1 Chosen Waveforms . . . . . . . . . . . . . . . . . . . . . . . 46
5.2.2 Theoretical Tools . . . . . . . . . . . . . . . . . . . . . . . . 47
5.2.3 Simulations . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
5.2.4 Laboratory Tests . . . . . . . . . . . . . . . . . . . . . . . . . 48
5.2.5 Sources . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
5.3 The Thesis in a Larger Perspective . . . . . . . . . . . . . . . . . . . 49

6 Conclusions 51
6.1 Problem Statements . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
6.1.1 Type of Waveform . . . . . . . . . . . . . . . . . . . . . . . . 51
6.1.2 Effect of Signal Properties . . . . . . . . . . . . . . . . . . . 52
6.1.3 Performance Compared to GNSS Synchronization . . . . . 52
6.2 Future Improvements . . . . . . . . . . . . . . . . . . . . . . . . . . 52

A Calculations for Cramér-Rao Bounds 53

B Cramér-Rao Bounds for Time-delay Estimation Compared to Simula-


tion Errors 57

C Additional Simulation Results 61


Contents xi

D Additional Laboratory Results 65

Bibliography 69
List of Figures

2.1 The two possible locations for the target, TA and TB, using two
TOA sensors, R1 and R2. Since no unambiguous estimate is given,
the target can be at either TA or TB. . . . . . . . . . . . . . . . . . . 6
2.2 The estimated location for the target using three TDOA sensors R1,
R2, and R3. Three pairs of receivers generate three hyperbolas, for
which the intersection is the estimated position of the target, TA. . 7
2.3 Example showing auto-correlation of 20 000 samples of white Gaus-
sian noise, w ∼ N (0, 1). Since there is a distinct peak at lag 0, it can
be concluded that none of the signals are time-delayed relative to
the other. If one of the signals was time-delayed, the peak would
have been at a lag corresponding to the delay measured in samples. 13
2.4 Ambiguity function of a pulse with A = 1, T = 100 µs and fs = 2
MHz. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
2.5 Spectrum of a baseband pulse signal with A = 1, T = 100 µs and fs
= 2 MHz. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
2.6 Ambiguity function of a chirp with B = 1 MHz, T = 100 µs and fs
= 2B. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
2.7 Spectrum of a baseband chirp signal with bandwidth B = 1 MHz,
T = 100 µs and fs = 2B. . . . . . . . . . . . . . . . . . . . . . . . . . 20
2.8 Ambiguity function of complex WGN with B = 1 MHz, T = 100 µs,
fs = 2B and variance σ 2 = 1. . . . . . . . . . . . . . . . . . . . . . . 21
2.9 Spectrum of one realization of complex WGN with bandwidth B =
1 MHz, T = 100 µs, fs = 2B and variance σ 2 = 1. . . . . . . . . . . . 22
2.10 Ambiguity function of a Gold sequence with B = 1 MHz, T = 100
µs and fs = 2B. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
2.11 Spectrum of a Gold sequence with bandwidth B = 1 MHz, T = 100
µs and fs = 2B. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23

3.1 Schematic drawing of the setup of the TDOA receivers and the
UAV used in this thesis. . . . . . . . . . . . . . . . . . . . . . . . . . 25

4.1 The CRBs for the time-delay estimation for different bandwidths,
signal lengths, and SNRs. If the parameter is not varied, it is set to
B = 10 MHz, T = 1 ms, and SNR = 0. The sampling frequency is
set to fs = 2B for all three plots. . . . . . . . . . . . . . . . . . . . . 32

xii
LIST OF FIGURES xiii

4.2 The effect of different bandwidths on the synchronization error,


with T = 1 ms, SNR = 0 dB, and fs = 2B. . . . . . . . . . . . . . . . 33
4.3 The effect of different bandwidths on the synchronization error,
with T = 1 ms, SNR = 0 dB, and fs = 4B. . . . . . . . . . . . . . . . 34
4.4 The effect of different bandwidths on the synchronization error,
with T = 1 ms, SNR = 0 dB, and fs = 2B. The distance between the
UAV and the receiver for channel 1 is 5 km and 15 km for channel
2. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
4.5 The effect of different signal lengths on the synchronization error,
with B = 10 MHz, SNR = 0 dB and, fs = 2B. . . . . . . . . . . . . . 35
4.6 The effect of different signal lengths on the synchronization error,
with B = 10 MHz, SNR = 0 dB and, fs = 4B. . . . . . . . . . . . . . 36
4.7 The effect of different signal lengths on the synchronization error,
with B = 1 MHz, SNR = 0 dB and, fs = 2B. The distance between the
UAV and the receiver for channel 1 is 5 km and 15 km for channel
2. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
4.8 The effect of different SNRs on the synchronization error, with B =
10 MHz, T = 1 ms, and fs = 2B. . . . . . . . . . . . . . . . . . . . . 38
4.9 The effect of different signal-to-noise ratios on the synchronization
error, with B = 10 MHz, T = 1 ms, and fs = 4B. . . . . . . . . . . . . 38
4.10 The effect of different signal-to-noise ratios on the synchronization
error, with B = 10 MHz, T = 1 ms, and fs = 2B. The distance be-
tween the UAV and the receiver for channel 1 is 5 km and 15 km
for channel 2. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
4.11 Resulting differences in error from the laboratory test for band-
widths 500 kHz and 1 MHz, with T = 1 ms and Pt = 10 dBm. . . . 40
4.12 Resulting differences in error from the laboratory test for different
signal lengths, with B = 1 MHz and Pt = 10 dBm. . . . . . . . . . . 40
4.13 Resulting differences in error from the laboratory test for different
output powers, with B = 1 MHz and T = 1 ms. . . . . . . . . . . . . 41

B.1 Comparison between the square root of the Cramér-Rao bound for
the time-delay estimation and the synchronization error for three
reference signals, with T = 1 ms, SNR = 0, fs = 2B, and varying
bandwidth. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
B.2 Comparison between the square root of the Cramér-Rao bound for
the time-delay estimation and the synchronization error for three
reference signals, with B = 10 MHz, SNR = 0, fs = 2B, and varying
signal length. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
B.3 Comparison between the square root of the Cramér-Rao bound for
the time-delay estimation and the synchronization error for three
reference signals, with B = 1 MHz, T = 1 ms, fs = 2B, and varying
SNR. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60

C.1 The effect of different bandwidths on the synchronization error,


with T = 100 µs, SNR = 0 dB, and fs = 2B. . . . . . . . . . . . . . . 61
xiv LIST OF FIGURES

C.2 The effect of different signal lengths on the synchronization error,


with B = 1 MHz, SNR = 0 dB, and fs = 2B. . . . . . . . . . . . . . . 62
C.3 The effect of different SNRs on the synchronization error, with B =
1 MHz, T = 1 ms, and fs = 2B. . . . . . . . . . . . . . . . . . . . . . 62
C.4 The effect of different SNRs on the synchronization error, with B =
10 MHz, T = 100 µs, and fs = 2B. . . . . . . . . . . . . . . . . . . . 63

D.1 Resulting differences in error from the laboratory test for different
output powers, with B = 1 MHz and T = 10 µs. . . . . . . . . . . . 65
D.2 Resulting differences in error from the laboratory test for different
output powers, with B = 1 MHz and T = 100 µs. . . . . . . . . . . . 66
D.3 Resulting differences in error from the laboratory test for different
output powers, with B = 500 kHz and T = 10 µs. . . . . . . . . . . 66
D.4 Resulting differences in error from the laboratory test for different
output powers, with B = 500 kHz and T = 100 µs. . . . . . . . . . . 67
D.5 Resulting differences in error from the laboratory test for different
output powers, with B = 500 kHz and T = 1 ms. . . . . . . . . . . . 67
List of Tables

3.1 Table showing the maximal length linear feedback shift register
polynomials used to generate the Gold codes. . . . . . . . . . . . . 28

4.1 Table showing the ratios between the errors for doubling of the
bandwidth. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
4.2 Table showing the ratios between errors for doubling of the signal
length. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36

xv
Notation

Defined parameters

Parameter Description
A Amplitude
Cx Covariance matrix
c Speed of light
d Distance between transmitter and receiver
f0 Instantaneous frequency
fc Carrier frequency
fs Sampling frequency
I(θ) Fisher information
IN Identity matrix of size N × N
LB Path loss
Np Noise power
Pr Received signal power
Pt Transmitted signal power
s Reference signal vector
w Additive white Gaussian noise vector
x Received signal vector
xτ Time-delayed signal vector
 Rate of frequency increase for chirp signal
θ Estimation parameter
θ̂ Estimated value
µ Mean value
σ2 Variance
τ Estimated delay

xvii
xviii Notation

Mathematical Operations

Operator Description
( · )∗ Conjugate operator
diag( · ) Diagonal matrix
F{·} Discrete Fourier transform
E{ · } Expectation
( · )H Hermitian operator
Re[ · ] Real part operator
tr( · ) Trace operator

Abbreviations

Abbreviation Description
awgn Additive white Gaussian noise
comint Communications intelligence
crb Cramér-Rao bound
dft Discrete Fourier transform
fft Fast Fourier transform
fi Fisher information
foi Totalförsvarets forskningsinstitut (Swedish Defense
Research Agency)
gnss Global navigation satellite system
gps Global positioning system
los Line-of-sight
mle Maximum likelihood estimator
rf Radio frequency
snr Signal-to-noise ratio
tde Time-delay estimation
tdoa Time difference of arrival
toa Time of arrival
uav Unmanned aerial vehicle
utc Coordinated universal time
wgn White Gaussian noise
Introduction
1
In military applications, it is important to have compromising information about
the enemy. On the battlefield, for example, that can mean knowing the coordi-
nates of enemy forces or vehicles. The process of determining the coordinates of
a target is called positioning. One common positioning method is time of arrival
(TOA). A known signal is emitted at a known time. The positioning is performed
by reckoning the time difference from emission to reception of the signal, so that
the distance to the target can be calculated. If two sensors are performing TOA,
the two resulting time differences can be compared. This process is called time
difference of arrival (TDOA), and gives information about how much closer the
target is to one of the sensors than the other. A difference from TOA is that TDOA
does not have to have prior knowledge about the signal, or when it was transmit-
ted. TDOA can achieve high positioning accuracy, and is, therefore, suitable to
be used in military applications, where incorrect information can lead to catas-
trophic consequences.
For TDOA to achieve accurate positioning of an object at range, a widely dis-
tributed network of positioning sensors is needed. However, as is the case for
most sensor networks, synchronization is an issue. The sensors need to work ac-
cording to a common time, meaning that the sensor clocks have to have the same
perception of when measurements should be made. If the sensors in a TDOA sen-
sor network are not synchronized, the measurements will be inaccurate, because
the measurements will not give coherent information about where the target was
located at a certain time. There are several methods used to achieve synchroniza-
tion for a network of TDOA sensors. For example, the timing signals transmitted
from global navigation satellite systems (GNSS) like global positioning system
(GPS) can be used as a reference, so that all the sensors in the network work ac-
cording to a common time. The reference signals transmitted from GPS satellites
contain universal coordinated time (UTC) timestamps [19]. UTC is seen as the

1
2 1 Introduction

official time scale of the world [2], and via the GPS satellites, it can be distributed
to the positioning sensors on the ground.
But communication systems intended for civilian or military use impose dif-
ferent challenges related to security. The need for stability and security increases
from preferable, for systems intended for civilian use, to absolutely crucial. This
means that the systems used regularly by civilians might be unusable by the mil-
itary. For the military forces of a nation it is also important with self-sufficiency,
and it is preferable not to rely on technology that is sustained or operated by
other nations or organizations. Moreover, it is undesirable to use technical solu-
tions for which the specifications are common knowledge. Such systems can be
easily jammed by external forces.
GPS reference signals suffer from the above-mentioned vulnerabilities [9].
The reference signals are vulnerable to jamming, or might disclose the position
of the user. GPS might even be shut off completely. Ideally, it would be possi-
ble to achieve the same synchronization accuracy with a solution that does not
depend on any external system. Additionally, it would be of great benefit if the
signals only reached sensors in a geographically restricted area. One solution
to this problem is to mount a clock on an unmanned aerial vehicle (UAV). The
clock on the UAV would act as the reference clock for all the sensors in the net-
work. The benefit of having the reference signal being transmitted from a UAV
is that the height can make line-of-sight (LOS) transmission possible, making the
connection more precise. The positioning sensors could use the reference signals
from the UAV to achieve local synchronization. Local synchronization means
three things:
• Only the relative time between the receivers is of importance. The TDOA
receivers do not need to know the correct time according to UTC to work,
only synchronization with the other sensors in the network is required.
• The reference signal from the UAV only need to reach receivers in the geo-
graphically restricted area in which the TDOA sensors are located.
• No external systems are needed. The sensor should not need any more in-
put besides the reference signals from the UAV to achieve synchronization.
To achieve synchronization between the TDOA sensors, the UAV will with pre-
determined intervals transmit reference signals to the positioning sensors. The
UAV and the TDOA receivers will be in fixed positions, so that the position of the
UAV is known at the receivers. To perform positioning, it is a requirement that
every receiver knows the position of all the other sensors in the network, includ-
ing their own. Naturally, there will be a delay between the transmission of the
reference signal and the reception at the receiver. However, given the positions
of the UAV and the receivers, it is possible to estimate the time-delay for a signal
transmitted from the UAV, since the distance between the UAV and receiver can
be calculated.
The estimation is performed using correlation of the original reference sig-
nal and the received signal. If a signal is correlated with itself (so-called auto-
correlation), the correlation reaches its maximum when the two signals are com-
1.1 Purpose 3

pletely overlapping. This peak is said to occur when the lag is zero, since the
signals are not shifted in time relative to each other [11, Ch. 4]. But if the orig-
inal reference signal is correlated with a time-delayed version of itself, the peak
occurs at a lag equal to the corresponding delay, measured in number of samples.
If the delay is larger or smaller than what is expected according to the distance
between the UAV and the receiver, the receiver clock is no longer synchronized
with the reference clock on the UAV. Given the estimated time-delay, and the
knowledge of the expected delay, the receiver clock can be adjusted. If the TDOA
receivers (independently of each other) are synchronized with the UAV, conse-
quently, they will be synchronized with each other.
For this method of estimation to be accurate and unambiguous, the peak
that occurs at maximum correlation has to be narrow, meaning that it has a large
amplitude relative to the rest of the correlation function. The width of the peak
also has to be small, in terms of samples. Some signals and waveforms yield a
more narrow correlation peak than others, and are therefore more well-suited
for time-delay estimation.
This thesis looks into how well different signals work when used for time-
delay estimation. The goal is to find what type of signals give the best estima-
tion accuracy, and how different parameters affect the estimation accuracy. The
accuracy of the estimation is measured as how close the estimation is to the ex-
pected delay. The signals are analyzed theoretically, to enable explanations for
the behavior. The estimation accuracy is evaluated by simulations performed in
Matlab, where the effect of different bandwidths, sampling frequencies, signal
lengths, and signal-to-noise ratios are studied. To verify whether the simulation
results would hold for a real communication channel, tests were carried out in a
laboratory.

1.1 Purpose
The overall purpose of this thesis is to compare signals and waveforms that have
good auto-correlation properties. The influence of different signal parameters on
the behavior of the time-delay estimation is studied. The goal is to find out which
signal gives the lowest error when used for time-delay estimation for a certain set
of time-bandwidth restrictions. The thesis also aims to highlight the practical
aspects of using the analyzed waveforms as synchronization signals.

1.2 Problem Statements


The problems that this thesis investigates can be summarized in three problem
statements. The thesis aims at answering these questions.
• What kind of signal should be used to synchronize two TDOA sensors as
accurately as possible, using a local reference signal from a UAV?
• What signal properties are of the highest importance when performing
time-delay estimation using correlation (bandwidth, signal length, etc.)?
4 1 Introduction

• Can two TDOA receivers be locally synchronized as accurately as if GNSS


would have been used?

1.3 Limitations
Signal theory and signal processing are large subjects, and covering every aspect
of the synchronization process in this thesis would be impossible. Therefore, to
keep the scope of the project at a reasonable size, this thesis focuses on the design
and correlation properties of the reference signal which is transmitted from the
UAV. This means that subjects like timing recovery algorithms are left out. To
make the comparisons of the signals meaningful, only properties that are com-
mon for all the signals are varied, such as bandwidth and signal length.
One reason to not use external reference signals is to achieve robustness
against jamming. However, this thesis will not investigate what can be done to
achieve higher resistance against such disturbances.
Theoretical Background
2
This chapter will provide the underlying theory of the topics in this thesis. In
Section 2.1, the positioning methods TDOA and TOA will be explained. Since
the synchronization signals should replace reference signals from external sys-
tems, Section 2.2 describes a few common systems used today. Section 2.3 gives
a mathematical description of the signal model, and Section 2.4 explains the free-
space propagation channel model. In Section 2.5, the theory of correlation and
time-delay estimation (TDE) is explained. Lastly, in Section 2.6, important sig-
nal properties when performing correlation are described, as well as the chosen
reference signals examined in this thesis.

2.1 Positioning
Positioning is the process of estimating the position of a target, in relation to
one’s own position. In military applications this could for example be detection
of approaching enemy surveillance drones or determining the coordinates of an
artillery target. Positioning can be made in either two dimensions (coordinates
on a plane) or three dimensions (exact position). In the following sections, only
positioning in terms of determining coordinates in two dimensions is considered.
There are several approaches to the problem of positioning, based on what
property of the received signal the estimator utilizes. Depending on what sig-
nal property is used, the approaches are typically divided into three observation
models: waveform observations, timing observations, and power observations [7,
Ch. 4]. An observation model can be viewed as a function that takes the measure-
ments (the observations) as input and yield the relevant positioning information
as output. The information the chosen observation yields, is usually categorized
as either range or bearing measurements. This thesis studies two types of range
measurements using timing observations: TOA and TDOA. Both TOA and TDOA

5
6 2 Theoretical Background

measures the timing of the arrival of the signal at the receiver (timing observa-
tions) and use that information to give information about the distance to the tar-
get (range measurements).

2.1.1 TOA
One method for estimating range is time of arrival (TOA). This is the basic prin-
ciple behind radar [17, Ch. 1]. TOA measurements utilize the fact that radio
frequency (RF) signals travel through space with a known velocity, the speed of
light. If a transmitter emits a pulse, that pulse will be reflected off of the target
object. The reflected version of the pulse returns to a receiver placed in the same
place as the transmitter. Then, the distance from the transmitter and the receiver
to the target can be calculated. Mathematically, this distance to the target is cal-
culated as:
ct
d= , (2.1)
2
where c is the speed of light and t is the propagation time from the transmitter to
the object. Note that the transmitted signal has to be known, so that its echo can
be recognized. To get accurate estimates of the TOA, a characteristic reference
signal is usually embedded into the signal [7, Ch. 4]. This known reference is
correlated with the received echo of the pulse at the receiver, to enable a more
accurate estimation of d. More information about correlation can be found in
Section 2.5.1.

d1 TA
d2

R1
R2
TB

Figure 2.1: The two possible locations for the target, TA and TB, using two
TOA sensors, R1 and R2. Since no unambiguous estimate is given, the target
can be at either TA or TB.

A setup with two TOA sensors is shown in Figure 2.1. The estimated distances
d1 and d2 do not contain any information about the direction of the target is,
which is different from radar, where direction is also measured. Since the pulse
is emitted in every direction at the same time, the constraint of possible locations
for the target looks like a circle around the receiver. Therefore, two sensors are
not enough to get an unambiguous estimate of the position of the target, since
the circles which mark the possible locations for the target intersect at two points.
2.1 Positioning 7

Therefore, at least three TOA measurements by three separate sensors are needed.
Then, the three circles would only intersect at one point.

2.1.2 TDOA
Instead of calculating the absolute distance to the target, TDOA utilizes the dif-
ference between two TOA measurements from a pair of sensors [7, Ch. 4]. This
means that it is not the actual TOAs that are of interest, the measure that matters
is the difference between them. Hence, one TDOA measurement means the dif-
ference between two TOA measurements from two different sensors. For regular
TOA to work, the transmitted signal has to be known, as well as the the time of
transmission. However, TDOA need only to calculate the difference between at
what time a signal was received at two receivers, no further knowledge of the
signal is needed. Hence, TDOA can be performed with signals not transmitted
from the TDOA sensors.
The constraint of possible locations for the target for one TDOA measure-
ment is a hyperbola, marking all the positions where the difference in distance
to the two sensors is the same. Thus, three TDOA measurements (from at least
three sensors) are required to give an unambiguous estimate of the position of
the target, because, in that scenario, three hyperbolas of possible locations are
calculated. The estimated target position is the intersection between the three
hyperbolas, as shown in Figure 2.2.

R3
R1-R3 TA

R1 R2

R2-R3 R1-R2

Figure 2.2: The estimated location for the target using three TDOA sensors
R1, R2, and R3. Three pairs of receivers generate three hyperbolas, for which
the intersection is the estimated position of the target, TA.

TDOA can achieve very high positioning accuracy, but with strict requirements
on the synchronicity of the sensors. Moreover, the shorter the distance between
the sensors, the more accurate the synchronization needs to be [9]. For example,
8 2 Theoretical Background

with a distance of 10 m between the sensors, a synchronization error of only 10 ns


can lead to a deviation from the correct target angle of up to ±27◦ . If the sensors
are placed 10 km apart, the angle error decreases to only ±0.0027◦ for the same
synchronization error.

2.2 Methods for Synchronization


Synchronization of widely distributed sensor networks is a well-known problem,
and several methods have been used to solve it. Usually, an oscillator is used to
generate a given frequency which in turn is used to provide a timing signal [18].
In the following section, a few common solutions are described. Solutions
which incorporate individual oscillators in each sensor are omitted, since those
are either not accurate enough to be used for TDOA, or very expensive [9]. The
methods described are divided into two categories: common-view time transfer
models and master-slave models. For common-view time transfer, each sensor
individually receives the same synchronization signal from a common transmit-
ter [4]. Master-slave models have one sensor (the master) that controls the other
sensors in the network, i.e., either forwards a received synchronization signal or
transmits a signal generated by the clock in the master device.

2.2.1 GPS
Today, there are several global navigation satellite systems (GNSS) in use. GNSS
is the term used for any navigation system which uses satellites and has global
or regional coverage. Examples of GNSS systems are GLONASS, which is owned
and operated by Russia, and Galileo, operated by the European Union [13]. All
GNSS systems could be used for time synchronization, but this thesis will focus
on GPS, which is operated by the United States. The method for synchronization
against GPS signals is a common-view time transfer method.
GPS satellites keep time using atomic clocks, which are made out of cesium
and rubidium oscillators [3]. Though very accurate, the clocks can drift, and are
therefore regularly controlled and adjusted by monitor stations on Earth. The
clocks follow universal coordinated time (UTC). According to the current speci-
fication for GPS, the time GPS transmits to receivers on Earth has a maximum
error of 40 ns 95% of the time [19], relative to UTC.
GPS satellites transmit carrier signals to GPS receivers on two frequency
bands: 1575.42 MHz (called L1), and at 1227.6 MHz (L2) [19]. Both of these
carrier signals are encoded with the current time and the position of the transmit-
ting satellite. In total GPS satellites transmit three codes:
• P code on both the L1 and L2 channel,
• C/A code on the L1 channel,
• the navigation message on the L1 channel.
The P code, or precision code, is restricted to be used only by the United States
military and NATO. The C/A (coarse/acquisition) code is available to the general
2.2 Methods for Synchronization 9

public. The navigation message is also publicly available, and contains general
information about the satellite [3].
The publicly available C/A code is a pseudo-random noise code, specifically,
a Gold sequence, which is unique for every GPS satellite. Gold sequences are
described in Section 2.6.8. The signal includes a timestamp for when the signal
is generated in the satellite [3]. To synchronize a receiver against a GPS satel-
lite, the received Gold sequence should be correlated with a copy of the same
Gold sequence, unspoiled by noise and delay. Since the signals for each satel-
lite are unique, the receiver can determine what signal it should correlate with,
knowing from which satellite the received signal came from. When correlation is
performed between the received signal and the copy, the delay from generation in
the satellite to reception on Earth is estimated. This means that the receiver cal-
culates the TOA for the received signal using t = dc . This expression is similar to
Equation (2.1), but without the factor 12 since the signal is only transmitted from
the satellite to the sensor. The estimated delay multiplied with the speed of light
is sometimes called the pseudo-range, implying that it is not the true range. The
pseudo-range is longer than the true range, because it is derived from the total
time from generation of the signal to finished estimation in the receiver. However,
since the position of the satellite is known, the true range can be calculated, and
subtracted from the pseudo-range. Knowing what delay is caused by the pseudo-
range, any undesirable clock deviations can be calculated and accounted for. For
this kind of common-view time transfers, the accuracy is roughly 1 - 10 ns [4].
However, the common-view method is most effective when the distance from the
receivers to the transmitter is very large [15].
As mentioned in the introduction, it is preferable not to depend on a system
operated by foreign states. This is the main drawback of using GPS for synchro-
nization in military applications. Apart from that, the accuracy is reasonable for
the intended purpose of synchronizing a network of TDOA sensors. The cost
can be kept low, as only GPS receivers are needed, rather than expensive atomic
clocks for each reciever [9].

2.2.2 DCF77
DCF77 is a long-wave (77.5 kHz) transmitter, which can be used as a time ref-
erence in most of Europe [5]. It transmits a signal that has a reach of roughly
1500 - 2000 km from Frankfurt am Main, where it is located. That means that
the most northern parts of Sweden are not covered. To keep time according to
UTC, the DCF77 uses cesium and rubidium atomic clocks. It is maintained by
the Physikalisch-Technische Bundesanstalt (PTB), the national physics laboratory
of Germany. It is commonly used to synchronize regular consumer clocks, like
alarm clocks and wristwatches.
A major advantage of using the time signal and standard frequency of the
DCF77 is that the long wavelength (≈3868.3 m) of the emissions makes them
invulnerable to obstacles. This is of great benefit, and is a reason why it is used
by clocks indoors. In the case of GPS, the receiver needs LOS communication
with the satellite [12].
10 2 Theoretical Background

The transmitted signal deviates from UTC by less than 2 · 10−12 s in one day
on average [5]. However, since the distance to the receiver is very long, reflections
in the ionosphere make the DCF77 an unreliable distributor of time reference, es-
pecially when the sensors are widely separated geographically [9]. The accuracy
is not high enough to motivate the use of DCF77 as an external reference. It is
also geographically restricted to central Europe. Furthermore, it has one of the
same drawbacks as GPS, namely that it is controlled by another state.

2.2.3 Wired Synchronization


One way to achieve robust time distribution between the sensors is to physically
connect them, using either coaxial or fiber optic cables. Using this solution, very
high accuracy can be achieved, e.g., 10 - 50 ps for optical fiber [2]. However, the
system becomes very inflexible, and almost unmovable if the distance between
the sensors is large.
Either a master/slave or a common-view model can be applied. For the
common-view model, a separate control station is needed, which contains the
oscillator. For a master/slave model, one of the TDOA receivers hold the oscil-
lator. One benefit, which applies both models, is that since sensors in a TDOA
network only need to be synchronized relative to each other. This means that
cheaper oscillators can be used, since the oscillator does not need to keep UTC
for long periods.

2.3 Signal Model


This section describes the notation in the thesis. It also provides the assumptions
made about the signals. Section 2.3.1 describes how time-shifts are applied, as
well as an additional method used to express the time-delay of a signal. The
complex baseband representation of signals is used.
The signal transmitted from the UAV is denoted s, and the signal received
by the TDOA sensor is denoted x. The received signal x is both time-shifted and
affected by noise. The noise is modeled as complex additive white Gaussian noise
(AWGN), denoted w. The noise samples in w are independent and identically
distributed according to w ∼ CN (0, σ 2 ), where σ 2 is the noise variance.
In this thesis, the generated signals are time-discrete, and are also sampled
at the reciever. However, the delay due to the propagation is continuous. If the
transmitted signal is denoted s, the received version might be delayed by τ, which
is continuous. The notation used to describe a continuous delay in a time-discrete
signal is x( fn − τ) = xτ [n], where n = − N2 , − N2 + 1, . . . , N2 − 1 with N = fs T . Here,
s
fs is the sampling frequency, and T is the duration of the signal in seconds. Thus,
the received time-shifted and noisy signal is denoted xτ = sτ + w, where sτ is a
time-delayed version of the original reference signal.
In the simulations, the number of samples to generate was calculated ac-
cording to N = fs T . However, for the WGN signal and the Gold sequences, BT
samples are generated, where B is the bandwidth of the signal. These were then
2.4 Free-space Propagation 11

upsampled to fs . This way of generating the WGN and Gold sequences is made
because the symbol rate is equal to the bandwidth for these signals.

2.3.1 Time-delay in Frequency Domain


Since many fast Fourier transform (FFT) algorithms are very computationally ef-
ficient, it can be beneficial to do certain operations in the frequency domain [8,
Ch. 2]. For a time-discrete signal x[n], − N2 , − N2 + 1, . . . , N2 − 1, a delay τ in the time
domain equals to
k
F {x[n − τ fs ]} = X[k]e−j2πτfs N , (2.2)
where fs is the sampling frequency and F { · } denotes the discrete Fourier trans-
form (DFT). The time-shift performed in 2.2 is circular. That means that if the
signal is shifted n samples to the right, the n samples farthest to the right will end
up outside of the signal length. However, those are appended to the signal as the
n samples farthest to the left. Therefore, to avoid unwanted samples added to the
left of the signal (for a shift to the right), zeros have to be added at the beginning
and end of the signal.
If the signal cannot be expressed as a function of time, derivation with re-
spect to the time-delay might be difficult. To make the derivation easier, a method
described in [20] can be used to multiply the signal with a delay factor. This
method also performs the time-shift in the frequency domain, but is only ap-
proximately equal to the time-shift performed in Equation 2.2. The benefit is
that the time-delay τ is introduced as a factor multiplied with the signal x, mak-
ing derivation with respect to the time-delay easy. The formula is, with n =
h iT
− N2 , − N2 + 1, . . . , N2 − 1, ,
xτ ≈ F H Dτ F x (2.3)
where
1 2π T
F = √ e−j N nn , (2.4)
N
and  n 
Dτ = diag e−j2πτfs N , (2.5)
where diag( · ) creates a diagonal matrix, with the argument vector as the diag-
onal. In the expressions above, F is the unitary N × N DFT matrix, while Dτ
can be compared to the delay factor in (2.2). This method of expressing the time-
delay makes derivation of the signal with respect to the time-delay much easier,
especially for signals not easily expressed by mathematical formulas.

2.4 Free-space Propagation


There are many methods used to model the effects of a wireless channel on a
signal. The energy the signal loses during propagation is called the path loss [1,
Ch. 2]. The path loss used for the simulations in this thesis is free-space prop-
agation. Free-space propagation path loss disregards any possible obstacles in
12 2 Theoretical Background

the propagation path, and also neglects effects of multi-path propagation. It is


therefore suitable to model LOS communications and estimate received signal
strengths.
Mathematically it is formulated as
!2
4πfc d
Lb = , (2.6)
c

where fc is the carrier frequency, and d is the distance of the propagation path.
The received signal power is
!2
E{|x|2 } c
Sp = = E{|x|2 } , (2.7)
Lb 4πfc d

where E{|x|2 } denotes the expectation of |x|2 .

2.5 Time-delay Estimation


This thesis focuses on estimation of one parameter, namely time-delay. When
electromagnetic waves propagate through air, they propagate at the speed of
light. Knowing the propagation speed means that if the position of both the trans-
mitter and the receiver (or a least the distance between them) is known, the TOA
can be calculated. Additionally, as always for TOA, the signal must be known.
If the clocks of the transmitter and receiver are synchronized, the signal will be
time-shifted according to the propagation time. However, if the clocks are out of
sync, the time-shift will be either larger or smaller than what is expected. Usually
what is meant by time-delay estimation, is the estimation of the total time-shift
from generation of the signal to the arrival at the receiver. Knowing the total
time-shift, any deviation from the expected can be detected. To reach very high
accuracy, other sources of delay, such as the time from generation of the signal
to the actual time of transmission, must be accounted for. The following sections
describe the fundamental concepts of TDE, such as correlation, the Cramér-Rao
bound for TDE, and maximum likelihood estimators.

2.5.1 Correlation
In signal processing applications, correlation means finding the similarity be-
tween two signals. The higher the similarity, the higher the correlation. Math-
ematically, this is described as the expectation of the product of two stochastic
processes X(t) and Y (t) [14, Ch. 4]. For complex signals, the expression is

E{X(t1 )Y ∗ (t2 )}, (2.8)

where ( · )∗ denotes the complex conjugate. This is called cross-correlation. Cor-


relation between a signal with itself is called auto-correlation. More accurately,
2.5 Time-delay Estimation 13

auto-correlation means comparing two realizations of a single stochastic process


at different time instances, and it is described by

E{X(t1 )X ∗ (t2 )}, (2.9)

In this thesis, cross-correlation between the transmitted and received version of


the signal will be performed.
Figure 2.3 visualizes why TDE can be performed using correlation. The fig-
ure shows auto-correlation of 20 000 samples of white Gaussian noise with vari-
ance σ 2 = 1, but in the same time instance t1 = t2 . Maximum correlation occurs
when the two signals completely overlap, meaning that the first sample of one
signal is correlated with the first sample of the other signal, etc. In Figure 2.3,
maximum correlation happens at lag 0. Lag means how many samples one of the
signals is time-shifted compared to the other. If instead, one of the WGN signals
had been time-delayed, the peak would have been located at a lag corresponding
to the number of samples equal to the delay.
Correlation

-2 -1.5 -1 -0.5 0 0.5 1 1.5 2


Lag [samples] 104

Figure 2.3: Example showing auto-correlation of 20 000 samples of white


Gaussian noise, w ∼ N (0, 1). Since there is a distinct peak at lag 0, it can
be concluded that none of the signals are time-delayed relative to the other.
If one of the signals was time-delayed, the peak would have been at a lag
corresponding to the delay measured in samples.

2.5.2 Cramér-Rao Bound for Time-delay Estimation


Because of the additive white Gaussin noise w, the received signal can be seen
as a realization of a random process. To map this realization to an estimated
value for the time-delay, an estimator is used. An estimator gives a value for
14 2 Theoretical Background

the delay for each realization of the received signal [10, Ch. 1]. If, on average,
an estimator yields the true value for the estimated parameter, it is said to be
unbiased [10, Ch. 3]. If the true value of the estimated parameter θ̂ is θ, then the
chosen estimator is unbiased if
E{θ̂} = θ. (2.10)
In estimation theory, a theoretical lower bound on the variance of an unbiased
estimator is called the Cramér-Rao bound, or CRB for short. It is one of the most
interesting and useful metrics when performing estimation. If, for a certain set
of parameter values, the CRB for the TDE is var(
qθ̂), the root mean squared error
(RMSE) on the estimate will not be lower than var(θ̂). The CRB is very useful
when different estimators are compared, but can also be used as a benchmark to
evaluate how close a chosen estimator is to the theoretical limit of its variance.
The goal is to find the minimum variance unbiased estimator (MVU) and see if it
satisfies the CRB. If it does, the MVU estimator is said to be efficient, meaning the
estimator is unbiased and attains the CRB for all possible values of the estimation
parameter. For an estimator to be efficient, it is required that the signal length
tends to an infinite number of samples, N → ∞. It is also required that the noise
variance σ 2 → 0.
One way to describe the CRB is as the inverse of the Fisher information (FI)
I(θ). If θ is the real-valued parameter that should be estimated, then
1
var(θ̂) ≥ . (2.11)
I(θ)
The CRB depends on the probability density function (PDF) of the samples of the
signal. For samples distributed according to a complex Gaussian PDF, as in this
thesis because of the complex AWGN, the expression for the FI is
" # " H #
−1 ∂Cx −1 ∂Cx ∂µ (θ) −1 ∂µ(θ)
I(θ) = tr Cx (θ) C (θ) + 2Re Cx (θ) (2.12)
∂θ x ∂θ ∂θ ∂θ
where Cx is the covariance matrix of the data and µ(θ) is the mean of the received
signal xτ . The operator tr[ · ] is the trace operator, ( · )H is the Hermitian operator,
and Re[ · ] outputs the real part of the argument. In this thesis, the time-delays of
received signals are estimated. Therefore, θ = τ.

2.5.3 Maximum Likelihood Estimator


Estimators that attain the theoretically lowest possible variance of the estimation
parameter according to the CRB are called minimum variance unbiased (MVU)
estimators [10, Ch. 2]. However, MVU estimators can be very hard to derive, or
in some cases, do not exist. Preferably, there would then exist an estimator which
is almost as accurate, but more practical to find. Often, that estimator is the
maximum likelihood estimator (MLE). The principle behind the MLE is to find
for which value of the parameter θ the PDF for xτ is maximized. The MLE is
asymptotically efficient for large data sets, and is in some cases easier find for
complicated estimation problems [10, Ch. 7].
2.6 Reference Signals 15

Let the original reference signal be denoted s[n], n = 0, 1, . . . , N − 1, and


let the received time-shifted version be denoted x[n − n0 ]. The delay n0 is also
time-discrete. Then, the MLE for time-delay estimation is
+N −1
n0X
n̂0 = argmax s[n]x[n − n0 ], (2.13)
n0 n=n0

where n̂0 is the estimated values for the true delay n0 [10, Ch. 7]. The sum in
(2.13) does describe cross-correlation between s and x. This means that the MLE
is equal to finding the delay for which the correlation between the reference sig-
nal with the received signal is maximized.

2.5.4 Sub-sample Accuracy


There is a limit to how accurately it is possible to estimate the delay for a signal
sampled with a certain sampling frequency fs , if the MLE is used. The sampled
signal is time-discrete, while the delay is continuous. No matter how large fs is,
the delay might not be a multiple of the sampling period T . It is necessary to be
able to find values between samples to find an accurate estimate of the delay.
Interpolation is a well-known method used for the problem of finding new
data points in the interval between already known data points. Following is a
description of an interpolation method that can be used to achieve sub-sample
accuracy for estimation. The idea is to choose the sample for which the correla-
tion is maximized, ni , as well as the two adjacent samples, ni−1 and ni+1 . The
three samples are then used to calculate the coefficients of the quadratic polyno-
mial that fits the samples best in a least-squares sense [16, Ch. 16].
Assume that the polynomial that is to be fitted to the samples is f (t) = at 2 +
bt + c. Then the maximum of that polynomial (assuming that the middle sample
b
ni is larger than the other two, and thus a < 0) will be at t = − 2a . Since f (t) is
known for three values of t = t1 , t2 and t3 , it is possible to calculate a, b and c
using the matrix equation
 2 
 t1 t1 1   a   f (t1 ) 
  
 2
 t2 t2 1   b  =  f (t2 )  . (2.14)
   
 2  
c f (t3 )
  
t3 t3 1

This is a system of linear equations with three unknown variables a, b and c, so


only one solution exists. The resulting values for the coefficients a and b are then
b
used to calculate t = − 2a .

2.6 Reference Signals


In this thesis, different reference signals used for synchronization are compared.
This section will describe those signals, as well as signal properties which affect
the estimation accuracy. The choice of what waveforms to compare is made based
on the knowledge presented here.
16 2 Theoretical Background

The signals in this section are baseband signals. A signal which is not modu-
lated up to a carrier frequency fc , and therefore has its energy contents centered
around f = 0, is called a baseband signal [11, Ch. 2]. A signal centered around
a carrier frequency is instead called narrow-band, or radio frequency (RF). The
effect of different modulation techniques are not accounted for in this thesis.

2.6.1 Frequency Band


The frequency band which is suitable for a signal to be transmitted over is deter-
mined by the application. Certain frequency bands are suitable for long-range
communications, while other frequency bands are suitable to achieve high data
rates. The path loss (2.6) used to model free-space propagation depends on the
carrier frequency fc . In the simulations, fc = 300 MHz was used. This frequency
belongs to the very high frequency band (30 - 300 MHz) or the ultra high fre-
quency band (300 - 3000 MHz). These frequency bands are usually used for
transmission with directional antennas and fixed point-to-point communications
over large distances [1, Ch. 1].

2.6.2 Signal-to-noise Ratio


As the name suggests, the signal-to-noise ratio (SNR) is the ratio between the
total received signal power and the noise power [1, Ch. 3]. In this thesis, it is
formulated as
P E{|s|2 }
SNR = r = (2.15)
Np σ2
where, Pr is the total signal power, and Np is the noise power at the receiver. It is
often expressed in dB, for which an SNR > 0 dB means that the signal energy is
greater than the noise energy for the signal. It is desirable to maximize the SNR.
If the SNR is small, noise will introduce a lot of disturbance to the signal. The
maximum achievable SNR depends only on the signal energy, so the SNR has no
theoretical upper limit.

2.6.3 Ambiguity Function


The width of the peak at lag 0 in Figure 2.3 is very small, only a few samples
wide, and the amplitude of the peak is large compared to the rest of the values.
If the auto-correlation of a signal has those properties, the signal is suitable to
be used for time-delay estimation. But not all signals have an auto-correlation
curve with a narrow peak and a large amplitude for the peak sample. Rather, it is
only for certain types of signals that the peak becomes distinct when performing
correlation. Since this property is desirable when estimating time-delay, the auto-
correlation curve is used as a performance metric, and is called the ambiguity
function. In literature, the ambiguity function sometimes is formulated so that
it also measures the Doppler resolution [17], but that version of the ambiguity
function is omitted from this thesis, as only time-delay resolution is of interest.
2.6 Reference Signals 17

It is a well-known fact that a waveform that has an ambiguity function with


a single, narrow peak is well-suited for time-delay estimation [11, Ch. 4]. In the
time-continuous case, the ambiguity function is defined as

Z∞
rx (τ) = x(t − τ)x∗ (t)dt. (2.16)
−∞

This expression is analogous to the auto-correlation of a function, which for some


signals produces a peak with a small width and a large amplitude for signals
suitable for TDE. The narrower the peak, and the higher the amplitude, the better
the signal is suited for TDE. Therefore, the ambiguity function is an intuitive way
of evaluating if a certain signal is appropriate to use for TDE.

2.6.4 General Expression for the Cramér-Rao Bound for


Time-delay Estimation
For some signals, deriving analytical closed-form expressions for the CRB for
TDE can be difficult. For random signals such as white Gaussian noise (WGN),
it’s impossible, since the signal cannot explicitly be expressed analytically. A
general expression for the CRB for TDE for the signals examined in this thesis is
however possible, given that the expression depends on the transmitted signal s.
Using Equation (2.12) and the approximation in (2.3), the result is

N 2σ 2
var(τ) ≥ , (2.17)
8π2 fs2 Re s H F H G H GF s


h iT
where G = diag(n) with n = − N2 , − N2 + 1, . . . , N2 − 1, , and var(τ) is the CRB of
the estimated delay τ. The approximation in (2.3) is used so that τ is canceled
out from the expression. Full calculations can be found in Appendix A.
It is hard to see how (2.17) depends on the bandwidth B and signal length T .
Therefore, one more approximation is made. The full calculations for this step
can also be found in Appendix A. The approximation is

h i B3 N 3 P
s H F H G H GF s ≈ (2.18)
12fs3

where P = |Sk |2 for |k| ≤ 2fB N , S being the discrete Fourier transform (DFT) of
s
the signal. Combining (2.17) and (2.18), the result is

3σ 2
var(τ) ≥ . (2.19)
2B3 T π2 P

This expression conveniently describes the relation between bandwidth, signal


length and the CRB.
18 2 Theoretical Background

2.6.5 Pulse
Mathematically, a pulse is expressed as

A,

 0≤t<T
x(t) =  (2.20)
0,
 otherwise

where A is the amplitude of the pulse. As described in Section 2.6.3, it is prefer-


able to have an ambiguity function which gives a peak with a small width and
large amplitude when performing time-delay estimation. As showed in Figure
2.4, the ambiguity function of a pulse has the shape of a triangle. The shorter the
pulse, the shorter the base of the triangle will be.
Correlation

-250 -200 -150 -100 -50 0 50 100 150 200 250


Lag [samples]

Figure 2.4: Ambiguity function of a pulse with A = 1, T = 100 µs and fs = 2


MHz.

The CRB for TDE for an ideal pulse exists, but is zero, which is a result that
give no meaningful information [17, Ch. 7]. The CRB for TDE is zero because an
ideal rectangular pulse has an infinite effective bandwidth, which makes (2.19)
tend to zero. This means that to be able to calculate an expression for the CRB
for TDE with a pulse, the pulse has to be filtered with a low-pass filter. Low-
pass filtering will spread the pulse in the time-domain, which will decrease the
bandwidth. Figure 2.5 shows the spectrum for an ideal pulse of length T = 100
µs and sampling frequency fs = 2 MHz.
2.6 Reference Signals 19

10 1

10 0

10 -1

10 -2
Power [W]

10 -3

10 -4

10 -5

10 -6

10 -7
-1 -0.8 -0.6 -0.4 -0.2 0 0.2 0.4 0.6 0.8 1
Frequency [MHz]

Figure 2.5: Spectrum of a baseband pulse signal with A = 1, T = 100 µs and


fs = 2 MHz.

2.6.6 Chirp
A chirp waveform is a pulse whose frequency increase linearly with time. It is
given by
 2
x(t) = ej2π(f0 t+ 2 t ) . (2.21)
In this thesis, the chirp sweeps over a frequency interval as large as the signal
bandwidth. With t = fn , n = − N2 , − N2 + 1, . . . , N2 − 1, assume that the instantaneous
s
frequency of the sweep f0 + 2 t is set to 0 at t = 0. The reason why  is divided by
2 in (2.21) is because then the term 2 denotes the rate of frequency increase per
unit time. At t = 2fN
− f1 the instantaneous frequency is set to B. Consequently,
s s
the rate of frequency increase per unit time is set to 2 = 2T
B
.
In Figure 2.6, the ambiguity function for a chirp signal is presented. The
width of the ambiguity function for the chirp signal is very small. This can be
understood intuitively, by realizing that since the frequency constantly differs,
the correlated signals will not have matching frequencies, except for lag 0.
20 2 Theoretical Background

Correlation

-250 -200 -150 -100 -50 0 50 100 150 200 250


Lag [samples]

Figure 2.6: Ambiguity function of a chirp with B = 1 MHz, T = 100 µs and


fs = 2B.

Figure 2.7 shows the spectrum of a baseband chirp signal with bandwidth B = 1
MHz. It is clear that the power is approximately the same for the whole frequency
range that the chirp signal sweeps over.

10 1

0
10

10 -1

-2
10
Power [W]

10 -3

10 -4

10 -5

10 -6

10 -7
-1 -0.8 -0.6 -0.4 -0.2 0 0.2 0.4 0.6 0.8 1
Frequency [MHz]

Figure 2.7: Spectrum of a baseband chirp signal with bandwidth B = 1 MHz,


T = 100 µs and fs = 2B.
2.6 Reference Signals 21

2.6.7 White Gaussian Noise


Signals that are random, or appear random, are usually called noise, or noise-like
waveforms. Noise-like waveforms has good correlation properties [11, Ch. 4].
The effect of the randomness on the ambiguity function is that the signal does
not contain repeating intervals, which means that the correlation will only be
large at lag 0. The result is an ambiguity function with a peak which is very
narrow, which can be seen in Figure 2.8, which shows the ambiguity function
for 200 samples of WGN. The complex WGN signal is generated from a complex
Gaussian distribution
1 − 12 |x|2
p(x) = e σ . (2.22)
πσ 2
Correlation

-250 -200 -150 -100 -50 0 50 100 150 200 250


Lag [samples]

Figure 2.8: Ambiguity function of complex WGN with B = 1 MHz, T = 100


µs, fs = 2B and variance σ 2 = 1.

True WGN can not be generated in reality, because the large and rapid changes in
amplitude would require unreasonably precise and powerful amplifiers. It has
only a practical meaning if it has been filtered [11, Ch. 3]. The randomness of
WGN can be seen in Figure 2.9, where the spectrum for 200 samples of complex
WGN is plotted. The spectrum do not follow any pattern, and the frequency
contents are random.
22 2 Theoretical Background

10 1

10 0

10 -1

10 -2
Power [W]

10 -3

10 -4

10 -5

10 -6

10 -7
-1 -0.8 -0.6 -0.4 -0.2 0 0.2 0.4 0.6 0.8 1
Frequency [MHz]

Figure 2.9: Spectrum of one realization of complex WGN with bandwidth B


= 1 MHz, T = 100 µs, fs = 2B and variance σ 2 = 1.

2.6.8 Gold Sequences


The correlation properties of random noise signals are favorable, but WGN sig-
nals are impossible to generate in practice. Therefore, pseudo-random sequences
are easier to use. One set of pseudo-random sequences is Gold sequences [6].
Gold sequences are designed to resemble noise, but contains only binary values.
However, since Gold sequences are only pseudo-random, they can be generated
in practice. Every GPS satellite has its own Gold sequence for its C/A code, which
makes it possible for the GPS receivers on Earth to uniquely identify each satellite
[3]. Gold sequences are generated by using maximal linear-feedback state regis-
ters. The maximal linear-feedback state register generates two binary pseudo-
random number sequences. Those sequences are then used as input to an XOR
gate, and the output is a Gold code. An in-depth description of generation of
Gold codes can be found in [6].
The Gold sequences are tested for different bandwidths and signal lengths
in this thesis. But Gold sequences have lengths described by 2n − 1, where n is
a positive integer (2n − 1 = 255, 511, etc.). This introduces a dilemma: should a
shorter Gold sequence be generated, and then be extended with a part of the same
sequence to achieve the desired length, or should a part of a longer sequence be
used instead? The solution used in this thesis is the latter; for a time-bandwidth
product BT , n was calculated as n = blog2 (BT )c + 1, so that a larger sequence than
needed always was generated.
2.6 Reference Signals 23

Figure 2.10 shows the ambiguity function for a Gold sequence of length 200
samples. The behavior is similar to that of the WGN signal in Figure 2.8.

Correlation

-250 -200 -150 -100 -50 0 50 100 150 200 250


Lag [samples]

Figure 2.10: Ambiguity function of a Gold sequence with B = 1 MHz, T =


100 µs and fs = 2B.

10 1

0
10

10 -1

-2
10
Power [W]

10 -3

10 -4

10 -5

10 -6

10 -7
-1 -0.8 -0.6 -0.4 -0.2 0 0.2 0.4 0.6 0.8 1
Frequency [MHz]

Figure 2.11: Spectrum of a Gold sequence with bandwidth B = 1 MHz, T =


100 µs and fs = 2B.
24 2 Theoretical Background

Figure 2.11 shows the spectrum for a Gold sequence of length 200 samples. The
spectrum is symmetrical around f = 0, since the signal is real-valued (binary).
Except for the symmetry, the spectrum is random.
3
Method

In this chapter the method used to answer the problem statements in Section 1.2
is described. Section 3.1 provides the assumed setup for the UAV and the TDOA
receivers, as well as the specifications set by the Swedish defense research agency
(FOI). The analytical evaluation of the signals is described in Section 3.2. Section
3.3, explains the Matlab simulations, and Section 3.4 contains a description of
the laboratory tests.

3.1 Setup
Figure 3.1 presents the assumed setup of the TDOA receivers and the UAV. In
this thesis, only two TDOA receivers are considered. However, for TDOA to work,
three receivers are required.

Reference signal Reference signal


UAV

TDOA receiver  max 20 km TDOA receiver

Figure 3.1: Schematic drawing of the setup of the TDOA receivers and the
UAV used in this thesis.

25
26 3 Method

The main focus is the fundamental theory of TDE, so implementation and scala-
bility were not accounted for. The maximal distance between the TDOA receivers
were set to 20 km, with the UAV in the intermediate area. However, the effect of
different distances between the sensors and the UAV was also tested.

3.1.1 Parameter Restrictions


The restrictions set on the parameter values are listed below. The restrictions
were set based on how current communications intelligence (COMINT) systems
work. For the tested parameters, the limits were:

• maximum signal bandwidth: 10 MHz,

• maximum signal length: 1 ms,

• maximum output power: 5 W (≈ 37 dBm).

3.2 Analytical Evaluation


The four chosen signals were analyzed theoretically by deriving the Cramér-Rao
bound (CRB) for the time-delay estimation. Based on the CRBs, predictions about
the TDE accuracy could be made. The signal which would yield the lowest CRB
is in theory the most suitable for estimating time-delays. Matlab was used to
compute the expression in (2.17). The TDE performance was evaluated from the
CRB for the TDE, and the ambiguity function.
From (2.19), it can be found that the CRB decreases with a factor B13 and T1 .
Since the CRB is a measure of variance, the actual error (the standard deviation)
would therefore decrease with a factor √1 3 and √1 . This fact was used to verify
B T
the simulations. If the bandwidth is doubled, for example, from 1 MHz to 2 MHz,
the error is multiplied with a factor √1 3 ≈ 0.35. Using the same method, the error
2
is multiplied with a factor √1 ≈ 0.71 for each doubling of the signal length. As
2
mentioned in Section 2.6.5, the CRB for TDE for a rectangular pulse does not give
any meaningful information, so the CRB for TDE for the pulse was not derived.

3.3 Simulations
To verify the TDE performance of the reference signals, simulations were per-
formed in Matlab. Four signal parameters were varied: the bandwidth, the sam-
pling frequency, the signal length, and the SNR. The parameters were varied ac-
cording to the restrictions in Section 3.1.1. Simulations were also performed for
different UAV positions.
In the simulations, the sampling frequency is a multiple of the bandwidth.
Two levels of oversampling were tested, most simulations were performed with
fs = 2B, and a few were performed with fs = 4B. The speed of light was set to
c = 2.99792458 · 108 m/s.
3.3 Simulations 27

First, the reference signals were generated for every set of parameter values.
For WGN, 1000 realizations were generated for each parameter configuration.
fs T
After the signals were generated, 10 zeros were added to the beginning and end
of every signal, as the signals could not be assumed to be strictly time-limited.
The next step was pulse-shaping, where the signals were filtered with a raised-
cosine filter. All the simulations were made over two channels, so the next step
was to simulate transmission over the channels. First, the signals were delayed
in the frequency domain, as explained in Section 2.3.1. The size of the delay
corresponded to the distance between the UAV and the receivers. Then the output
signal power Pt was normalized
√ to Pt = 1 W. The next step was to divide the
signal with the path loss Lb , as described in Section 2.4. The carrier signal
was set to fc = 300 MHz, as it is a reasonable carrier frequency for signals with
bandwidths in the simulated interval. The path loss also depends on the distance
between transmitter and receiver, d. To model different positions of the UAV, the
simulated distances between the sensors and the UAV were varied. Finally, WGN
was added, with a variance set to achieve the specified SNR. If the UAV was not
placed with equal distance to the receivers, the noise variance was calculated
based on the received signal power for the corresponding distance, with the SNR
fixed for both channels.

3.3.1 Time-delay Estimation in the Simulations


The estimation of the time-delay in the simulations was performed as described
in Section 2.5. The received signal was correlated with the original signal. To
get sub-sample accuracy, interpolation with a quadratic polynomial was made
(see Section 2.5.4). The error was calculated as the deviation from the expected
delay with regards to the distance between UAV and receiver. For example, if the
distance between the receiver and the UAV was 10 km, the expected delay was
τ = 10000
c ≈ 33 µs. If then the estimated delay was, e.g., 33.4 µs, the error was 0.4
µs. If the estimated delay was, e.g., 32.6 µs, the error was still considered to be
0.4 µs, i.e., no consideration was taken to if the delay was positive or negative.
For WGN, the deviation from the expected delay was calculated for each
realization, and the final error was set to the mean error for the 1000 delays.

3.3.2 Generation of Gold Sequences


In Table 3.1, the maximum length polynomials used to generate the Gold se-
quences are listed. The generation was performed using code provided by FOI.
Due to restrictions in the generation code, sequences of lengths where n is divisi-
ble with 4 were not generated. Instead, a longer code of length 2n was generated
and truncated. Example 3.1 explains how to read Table 3.1.
28 3 Method

Example 3.1
If a polynomial in Table 3.1 is written as [5 4 2 1], then the polynomial used is
x5 + x4 + x2 + x + 1. The last term (+1) is always included.

Table 3.1: Table showing the maximal length linear feedback shift register
polynomials used to generate the Gold codes.

Maximum length polynomials


Bits in feedback term (n) 2n − 1 Polynomial
2 3 [2 1]
3 7 [3 1]
4 15 –
5 31 [5 4 2 1]
6 63 [6 5 2 1]
7 127 [7 6 5 3 2 1]
8 255 –
9 511 [9 8 6 5 3 1]
10 1023 [10 9 8 6 5 1]
11 2047 [11 10 6 5 4 1]
12 4095 –
13 8191 [13 10 4 3 2 1]
14 16383 [14 9 7 4 2 1]

3.4 Laboratory Tests


The laboratory test were performed using antennas on the roof of the FOI build-
ing. The setup was similar to the simulations; the signals were transmitted from
one source antenna, and were received by two antennas, roughly 0.5 m apart. The
transmitted signals were 2 s long, containing 998 repetitions of the same wave-
form. The first and last millisecond of the transmitted signals were set to zeros
to account for packet losses at the receiver. The parameter values tested in the
laboratory tests were
• Bandwidth: B = 1 MHz and B = 500 kHz.
• Signal length: T = 10, 100 and 1000 µs.
• Output power: Pt = −10, 0 and 10 dBm.
In total, 54 combinations were tested. The signals were sampled with fs = 16B.
For transmission, the signals were modulated up to fc = 432 MHz, which is not
the same carrier frequency as in the simulations, where fc = 300 MHz was used.
The reason for this difference is that since the laboratory signals were transmitted
outdoors from real antennas, where there are limitations regarding what frequen-
cies can be used. Which carrier frequency could be used for the laboratory tests
was not known when the simulations were performed.
3.4 Laboratory Tests 29

The raw recorded files were converted to Matlab files and downsampled to
fs = 2B, to make the data sets reasonably small. Then the synchronization errors
were calculated. Since the distances between the transmitter and the receiver
were unknown for the laboratory tests, the absolute time-delay was not calcu-
lated. The deviation from the correct time-delay can only be calculated if the ex-
pected delay is known. Therefore, the synchronization performance was consid-
ered to be the difference between the estimated time-delays for the two receivers.
A small difference in estimated time-delay between the receivers was considered
as good synchronization performance. Since there were 998 recordings of each
signal, the error was calculated as the mean error for the 998 resulting differences
for each waveform and parameter configuration.
4
Results

In this section, the results are presented. In Section 4.1, the CRBs for TDE is pre-
sented. The results from the Matlab simulations are found in Section 4.2. Finally,
the results from the laboratory tests are presented in Section 4.3. Additional re-
sults from the simulations and the laboratory tests can be found in Appendix C
and D, respectively.

4.1 Cramér-Rao Bounds for Time-delay Estimation


Since Equation (2.19) uses such a rough approximation for P , the results from
that expression are not included as results. The expression was only used to
predict the behavior for different bandwidths and signal lengths. The formula
used for the CRBs for the TDE is Equation (2.17),

N 2σ 2
var(τ) ≥ .
8π2 fs2 Re s H F H G H GF s


Figure 4.1 shows the CRBs for the TDE for different bandwidths, signal lengths,
and SNRs. The pulse is omitted from the plots for reasons explained in Section
2.6.5. The CRBs for TDE are very similar. Though not visible, there is a negligible
difference between the CRBs for the TDE.

31
32 4 Results

10 -16
Chirp
WGN
Gold
CRB [s2 ]

10 -18

10 -20

10 -22
0 1
10 10
Bandwidth [MHz]
10 -16
Chirp
WGN
Gold
CRB [s2 ]

10 -18

10 -20

10 -22
10 -1 10 0
Signal length [ms]
-16
10
Chirp
WGN
Gold
CRB [s2 ]

10 -18

10 -20

10 -22
-30 -20 -10 0 10 20 30
SNR [dB]

Figure 4.1: The CRBs for the time-delay estimation for different bandwidths,
signal lengths, and SNRs. If the parameter is not varied, it is set to B = 10
MHz, T = 1 ms, and SNR = 0. The sampling frequency is set to fs = 2B for
all three plots.

Plots comparing the root of the CRBs for the TDE with the estimation errors can
be found in Appendix B. The estimation accuracy does not reach the theoretical
limit set by the CRB for the TDE, indifferent of what type of signal is used.
4.2 Simulations 33

4.2 Simulations
This section shows the most essential results for the simulations. The section is di-
vided into three parts, depending on which parameter was varied: the bandwidth
(Section 4.2.1), the signal length (Section 4.2.2) and the SNR (Section 4.2.3).
This is a selection of all the plots generated in the simulations, and additional
plots can be found in Appendix C. The error on the vertical axis is the mean
deviation from the expected delay, as explained in Section 3.3.1.

4.2.1 Varied Bandwidth


Figure 4.2 shows the effect of different bandwidths on the synchronization accu-
racy. The sampling frequency is set to fs = 2B, and the UAV is placed with a
distance of 10 km to both receivers.

Channel 1 Channel 2
10 3 10 3
Pulse Pulse
Chirp Chirp
WGN WGN
Gold Gold

10 2 10 2
Error [ns]

Error [ns]

10 1 10 1

10 0 10 0

10 -1 10 -1
0 1 0 1
10 10 10 10
Bandwidth [MHz] Bandwidth [MHz]

Figure 4.2: The effect of different bandwidths on the synchronization error,


with T = 1 ms, SNR = 0 dB, and fs = 2B.

All signals except the pulse give similar errors, as shown by the CRBs for the
TDE. For larger bandwidths, the errors oscillate slightly. That is because a lack
of sampling resolution at high frequencies, and can be avoided with a higher
sampling frequency relative to the bandwidth.
In Table 4.1, the relation between errors for different bandwidths are listed.
In theory, the ratio should be roughly 0.35 ≈ √1 for a doubled bandwidth, as
8
described in Section 3.2.
34 4 Results

Table 4.1: Table showing the ratios between the errors for doubling of the
bandwidth.

2 MHz/1 MHz 4 MHz/2 MHz 8 MHz/4 MHz


Pulse 0.573 0.470 0.522
Chirp 0.366 0.336 0.439
WGN 0.380 0.347 0.415
Gold 0.363 0.357 0.372

Channel 1 Channel 2
10 3 10 3
Pulse Pulse
Chirp Chirp
WGN WGN
Gold Gold

10 2 10 2
Error [ns]

Error [ns]

1 1
10 10

0 0
10 10

-1 -1
10 10
10 0 10 1 10 0 10 1
Bandwidth [MHz] Bandwidth [MHz]

Figure 4.3: The effect of different bandwidths on the synchronization error,


with T = 1 ms, SNR = 0 dB, and fs = 4B.

In Figure 4.3, the sampling frequency is fs = 4B. All the other values are set as
before. The higher sampling frequency stabilizes the oscillation for higher band-
widths. This is because when performing the TDE, the cross-correlation can with
more accuracy produce a peak at the correct number of samples, corresponding
to the delay. However, the increased sampling frequency does not result in a
lower estimation error, except for the pulse.
In both Figure 4.2 and 4.3, the distance between UAV and receiver is the
same for both channels. In Figure 4.4, the UAV is moved 5 km closer to one of
the receivers, so that the distance between the UAV and the receiver is 5 km for
channel 1, and 15 km for channel 2. Again, the sampling frequency is set to
twice the bandwidth. The error is notably smaller for channel 1, for which the
propagation distance is smaller.
4.2 Simulations 35

Channel 1 Channel 2
10 3 10 3
Pulse Pulse
Chirp Chirp
WGN WGN
Gold Gold

10 2 10 2

Error [ns]

Error [ns]
10 1 10 1

10 0 10 0

10 -1 10 -1
10 0 10 1 10 0 10 1
Bandwidth [MHz] Bandwidth [MHz]

Figure 4.4: The effect of different bandwidths on the synchronization error,


with T = 1 ms, SNR = 0 dB, and fs = 2B. The distance between the UAV and
the receiver for channel 1 is 5 km and 15 km for channel 2.

4.2.2 Varied Signal Length


Figure 4.5 shows the effect of different signal lengths on the synchronization per-
formance. The sampling frequency is set to fs = 2B, and the UAV is placed with
a distance of 10 km to both receivers.

Channel 1 Channel 2
10 4 10 4
Pulse Pulse
Chirp Chirp
WGN WGN
Gold Gold

10 3 10 3

10 2 10 2
Error [ns]

Error [ns]

10 1 10 1

0 0
10 10

-1 -1
10 10
10 -1 10 0 10 -1 10 0
Signal length [ms] Signal length [ms]

Figure 4.5: The effect of different signal lengths on the synchronization error,
with B = 10 MHz, SNR = 0 dB and, fs = 2B.
36 4 Results

The chirp, WGN and Gold sequences achieve almost the same error. Worth noting
is that the performance for the pulse increases with increasing signal length. That
is because for a shorter pulse, the width of the ambiguity function peak is smaller,
making TDE more precise.
As in Section 4.2.1, the simulations are verified by controlling the relation
between errors for different signal lengths. The results are listed in Table 4.2.
In theory, the ratio should be roughly 0.71 ≈ √1 for a doubled signal length, as
2
described in Section 3.2.

Table 4.2: Table showing the ratios between errors for doubling of the signal
length.

0.2 ms/0.1 ms 0.4 ms/0.2 ms 0.8 ms/0.4 ms


Pulse 1.174 1.078 1.117
Chirp 0.764 0.788 0.748
WGN 0.784 0.772 0.772
Gold 0.767 0.745 0.765

The results above are achieved with a sampling frequency which is double the
bandwidth. In Figure 4.6, the sampling frequency is fs = 4B. All the other values
are set as before. The increased sampling frequency introduces a slight decrease
for the estimation error for the pulse. The other signals are unaffected.

Channel 1 Channel 2
10 4 10 4
Pulse Pulse
Chirp Chirp
WGN WGN
Gold Gold
3 3
10 10

10 2 10 2
Error [ns]

Error [ns]

1 1
10 10

10 0 10 0

10 -1 10 -1
10 -1 10 0 10 -1 10 0
Signal length [ms] Signal length [ms]

Figure 4.6: The effect of different signal lengths on the synchronization error,
with B = 10 MHz, SNR = 0 dB and, fs = 4B.
4.2 Simulations 37

In Figure 4.7, the UAV has been moved 5 km closer to one of the receivers. Hence,
channel 1 in Figure 4.7 has a distance between UAV and TDOA receiver of 5 km,
while the distance is 15 km for channel 2. The shorter distance in channel 1
improves the estimation accuracy notably. Additionally, the error for the pulse
does not increase as much for the shorter distance.

Channel 1 Channel 2
10 4 10 4
Pulse Pulse
Chirp Chirp
WGN WGN
Gold Gold

10 3 10 3

10 2 10 2
Error [ns]

Error [ns]
1 1
10 10

0 0
10 10

-1 -1
10 10
10 -1 10 0 10 -1 10 0
Signal length [ms] Signal length [ms]

Figure 4.7: The effect of different signal lengths on the synchronization error,
with B = 1 MHz, SNR = 0 dB and, fs = 2B. The distance between the UAV
and the receiver for channel 1 is 5 km and 15 km for channel 2.

4.2.3 Varied Signal-to-noise Ratio


Figure 4.8 shows the effect of different signal-to-noise ratios on the synchroniza-
tion performance. The sampling frequency is set to fs = 2B, and the UAV is
placed with a distance of 10 km to both receivers.
Up to SNR ≈ 5 dB, the error decreases log-linearly with increasing SNR for
the chirp signal, WGN and the Gold sequences. For larger SNRs, the error does
not decrease, instead it flattens at 0.2 ns. This is because for very large SNR,
the lack of sampling resolution influence the TDE accuracy more than the SNR.
The error for the pulse still decrease at SNR = 30 dB, but the error has flattened
slightly.
38 4 Results

Channel 1 Channel 2
10 5 10 5
Pulse Pulse
Chirp Chirp
WGN WGN
10 4 Gold 10 4 Gold

10 3 10 3

10 2 10 2
Error [ns]

Error [ns]
1 1
10 10

10 0 10 0

10 -1 10 -1

10 -2 10 -2
-30 -20 -10 0 10 20 30 -30 -20 -10 0 10 20 30
SNR [dB] SNR [dB]

Figure 4.8: The effect of different SNRs on the synchronization error, with B
= 10 MHz, T = 1 ms, and fs = 2B.

Channel 1 Channel 2
10 5 10 5
Pulse Pulse
Chirp Chirp
WGN WGN
10 4 Gold 10 4 Gold

10 3 10 3

10 2 10 2
Error [ns]

Error [ns]

1 1
10 10

10 0 10 0

10 -1 10 -1

10 -2 10 -2
-30 -20 -10 0 10 20 30 -30 -20 -10 0 10 20 30
SNR [dB] SNR [dB]

Figure 4.9: The effect of different signal-to-noise ratios on the synchroniza-


tion error, with B = 10 MHz, T = 1 ms, and fs = 4B.
4.3 Tests in Laboratory 39

In Figure 4.9, the bandwidth is increased to fs = 4B. Here, the error does not
flatten out, instead it continues to decrease for larger SNR. For SNRs below zero,
the error is unchanged except for the pulse, which has a lower error than was the
case for fs = 2B.
The UAV is moved 5 km closer to one of the receivers in Figure 4.10. Hence,
channel 1 has a distance between the UAV and the TDOA receiver of 5 km, while
the distance for channel 2 is 15 km. The curves are similar to those in Figure 4.8,
but for channel 1, the error flattens out at a lower error, just below 0.1 ns for all
signals except the pulse. The difference in error can be explained by the fact that
in channel 1, the received signal power is larger, which increases the SNR.

Channel 1 Channel 2
10 5 10 5
Pulse Pulse
Chirp Chirp
WGN WGN
10 4 Gold 10 4 Gold

10 3 10 3

10 2 10 2
Error [ns]

Error [ns]

1 1
10 10

0 0
10 10

-1 -1
10 10

-2 -2
10 10
-30 -20 -10 0 10 20 30 -30 -20 -10 0 10 20 30
SNR [dB] SNR [dB]

Figure 4.10: The effect of different signal-to-noise ratios on the synchroniza-


tion error, with B = 10 MHz, T = 1 ms, and fs = 2B. The distance between the
UAV and the receiver for channel 1 is 5 km and 15 km for channel 2.

4.3 Tests in Laboratory


This section will show the results from the tests performed in the laboratory. In
Figure 4.11 the resulting differences in error for the two bandwidths are plotted,
with T = 1 ms and Pt = 10 dBm. As expected, the pulse gives a larger synchro-
nization error than the other signals. However, the difference is not particularly
large. A doubled bandwidth only improves the difference slightly.
40 4 Results

10 3

Difference in estimated delay [ns]


Pulse
Chirp
Gold

10 2

10 1
0.5 0.6 0.7 0.8 0.9 1
Bandwidth [MHz]

Figure 4.11: Resulting differences in error from the laboratory test for band-
widths 500 kHz and 1 MHz, with T = 1 ms and Pt = 10 dBm.

3
10
Difference in estimated delay [ns]

Pulse
Chirp
Gold

10 2

10 1
10 -2 10 -1 10 0
Signal length [ms]

Figure 4.12: Resulting differences in error from the laboratory test for differ-
ent signal lengths, with B = 1 MHz and Pt = 10 dBm.

For different signal lengths, the result can be seen in Figure 4.12. An increased
signal length has only a mild effect on the TDE in these tests, except for the
pulse. As was the case in the simulations, a larger signal length results in worse
synchronization performance for the pulse.
4.3 Tests in Laboratory 41

10 4

Difference in estimated delay [ns]


Pulse
Chirp
Gold

10 3

10 2

10 1
-10 -5 0 5 10
Output power [dBm]

Figure 4.13: Resulting differences in error from the laboratory test for differ-
ent output powers, with B = 1 MHz and T = 1 ms.

In Figure 4.13, the difference in error for different output powers is shown. The
difference in synchronization error for the chirp and the Gold sequence is almost
the same for the three output powers, while the error for the pulse decreases
for larger output power. As can be seen in the additional figures in Appendix
D, the variations in output power did only affect the results for the chirp and
Gold sequences for B = 500 kHz, while the behavior of the pulse corresponds to
variations in both bandwidth and signal length.
Discussion
5
This chapter discusses the results presented in Section 4. The simulation results
are discussed in Section 5.1. In Section 5.2, the method used to achieve and verify
the results in this thesis is commented on. Section 5.3 contains a discussion about
the thesis from an ethical and societal perspective.

5.1 Results
The general results are discussed in this section. The section is divided into two
parts: Section 5.1.1, where the waveforms and their performance is discussed,
and Section 5.1.2 where the effect of the simulated sets of parameter values are
commented on. This separation is made to make a clear distinction between the
effect on the estimation error of the waveforms and the effect that is caused the
signal properties.

5.1.1 Estimation Accuracy of the Waveforms


The synchronization errors for the tested waveforms are similar, except for the
pulse. This result matches the CRBs for the time-delay estimation, which are
very similar for the same three signals. Additionally, the peaks of the ambiguity
functions for these three waveforms are all very narrow. The time-delay estima-
tion not attaining the CRB for the TDE was expected, as the expression used for
the CRB was not based on the MVU estimator, and it included an additional ap-
proximation (see Section 2.6.4).
It is surprising that the TDE performance of the chirp signal so closely re-
sembles that of the noise-like waveforms WGN and the Gold sequences. The
chirp and the noise-like waveforms are very different types of waveforms, but
for all of the simulations, the estimation accuracy is very similar. The three well-

43
44 5 Discussion

performing waveforms are affected by changes in sampling frequency and the


distance to the receiver in the same way.
The pulse performs significantly worse than the three other signals. Except
for very low SNR, and lower effective bandwidth (Figure C.3), or shorter signal
length (Figure C.4), the pulse yields a larger estimation error. However, it reaches
the goal of a synchronization error of 20 ns at SNR ≈ 1 dB for the highest possi-
ble time-bandwidth product BT = 104 . As predicted, a shorter signal length
results in a lower error, although negligible. It is worth noting that an increased
bandwidth only means a higher sampling frequency, as the shape of the pulse is
unchanged for variations in bandwidth.
Gold sequences are designed to have beneficial TDE properties. However,
the Gold sequences can not be fully utilized in this thesis, as many signal lengths
are compared. An alternative method would have been to perform the tests with
the lengths most favorable to the Gold sequences. It is unclear what the effect of
such tests would have been, since the correlation properties for shortened Gold
sequences still are excellent.
In the laboratory tests, the chirp signal and the Gold sequences give roughly
the same difference in estimated time-delay for the tested bandwidths and signal
lengths at output power Pt = 10 dBm. As in the simulations, the pulse does not
give the same accuracy, put in the laboratory tests, the difference is much smaller.
However, since simulations uses another metric to evaluated TDE performance,
it is difficult to compare the results. But the laboratory results do correspond
with the simulation results.
In summary, all signals except the pulse give nearly the same estimation
accuracy for the same set of parameter values. However, for high SNR and high
sampling frequency, the pulse also gives very low TDE error.

5.1.2 Effect of the Signal Properties


An increase in bandwidth has a positive effect on the estimation error for all
of the waveforms. As can be seen in Table 4.1, a doubling of the bandwidth
result in an error 35% as large. It is the most important parameter for the time-
delay estimation performance. It can be explained by the fact that an increased
bandwidth result in a signal containing "more detail". For the chirp signal, the
bandwidth is equal to the frequency range of the frequency sweep. Hence, the
chirp will contain more frequencies in the same amount of time for an increased
bandwidth, which results in a more distinct peak in the ambiguity function. For
WGN and Gold sequences, the bandwidth is equal to the symbol rate, meaning
that longer Gold sequences and more samples of WGN can be transmitted during
the signal length, also resulting in a more distinct peak in the ambiguity function.
For higher bandwidths the error starts to oscillate slightly, as can be seen in
Figure 4.2. The reason is limitations in the sub-sample resolution, and can, there-
fore, be solved by increasing the sampling frequency, relative to the bandwidth.
When the sampling frequency is doubled relative to the bandwidth, from fs = 2B
to fs = 4B, the oscillation for the error is removed. In conclusion, fs = 2B do not
seem to be enough for the larger bandwidths.
5.1 Results 45

The TDE accuracy for the pulse also increases for increased bandwidth, al-
though not as much as the other three waveforms. With T = 1 ms and SNR =
0 dB, a bandwidth of 10 MHz is not enough to achieve the target error of 20 ns.
However, as mentioned in Section 5.1.1, the bandwidth has not the same meaning
for the pulse. In the simulations, an increased bandwidth only means a higher
sampling rate, but will not change the waveform itself. However, the other three
waveforms performs well enough (error < 20 ns), already at B = 1 MHz for T = 1
ms and SNR = 0 dB.
The performance benefit of increasing the signal length was, as shown by the
CRBs for the TDE and the ambiguity function. A doubled signal length results
in an error ≈ 77% as large for all signals except the pulse. A longer pulse gives a
larger error than a shorter one. This was expected, since the peak of the ambigu-
ity function for the pulse is more narrow for shorter pulses. The opposite is true
for the other three waveforms. An increased signal length has the same effect on
the synchronization error for those signals.
The fact that higher SNR gives increased estimation accuracy is intuitive.
There is a risk that the disturbances caused by the AWGN are mistaken for the
peak caused by maximum correlation when comparing the signals. For lower
SNR, the AWGN disturbs the received signal more, which means that the maxi-
mum amplitude for the correlation peak will not be as high as for auto-correlation.
Since correlation is measuring similarity between signals, correlation between a
signal and a noisy version itself will give lower correlation for lower SNR.
If the sampling frequency is not large enough relative to the bandwidth, in-
creased SNR does not give better estimation accuracy. Therefore, the conclusion
can be drawn that a certain degree of oversampling is necessary to increase the es-
timation performance for high SNR. However, for very high SNR, the estimation
errors are already so small that an increased sampling frequency not necessarily
is needed.
The distance between the UAV and receiver only have a large impact on the
synchronization error, especially for smaller SNR. When the UAV is moved 5 km
closer to the receiver, the difference in error is notable. The difference is smaller
for high SNR, where a shorter distance only gives a slightly increased estimation
accuracy. A longer distance means higher attenuation due to path loss, hence
lower SNR.
In the laboratory tests, nethier bandwidth, signal length nor output power
had any significant impact on the results, except for the pulse. The effect on the
pulse did correspond to the behavior in the simulations. The reason why the
tested parameters did not affect the performance much can be manifold. Only
two bandwidths are tested, and the largest bandwidth is only 1 MHz. More vari-
ations would be required to get more trust-worthy results. This also applies to
the signal lengths and the output powers. The channel model in the simulations
is simple, and in the laboratory tests, multi-path propagation might have had a
larger effect than what was modeled in the simulations. Additionally, there is
a risk that the clocks of the receivers drifted to a degree where the effect of the
drift were noticeable for the longer signals, thus explaining the negligible gain in
estimation accuracy for increased signal lengths.
46 5 Discussion

In summary, the most important parameter for the time-delay estimation


accuracy is the bandwidth of the reference signal. Both longer signal length and
higher SNR is preferable to increase the accuracy, but have a smaller effect. The
distance between the UAV and the receiver is not of great importance.

5.2 Method
The method used to answer the problem statements in Section 1.2 will be dis-
cussed in this section. First, the central decision regarding what waveforms to
analyze will be discussed in Section 5.2.1. A motivation for the use of the CRB
and the ambiguity function can be found in Section 5.2.2. The simulations and
laboratory tests will be commented on in Section 5.2.3 and 5.2.4. The sources
used will be discussed in 5.2.5.

5.2.1 Chosen Waveforms


The decision to test four different waveforms, was made so that enough care could
be taken to each of them. Additionally, when simulating for different sets of
parameters, the number of simulations soon grow very large, and had to be kept
at a reasonable level.
When it came to the choice of what signals that would be examined and
evaluated, four possibly conflicting approaches were considered:

• Choose a set of similar signals, such that a fixed set of parameters easily
could be varied and compared.

• Choose the signals that have the highest possibility of achieving the greatest
performance.

• Choose the most commonly used signals.

• Choose signals that are largely different.

If the chosen waveforms are dissimilar, it is difficult to draw conclusions about


the effects other parameters have, except the TDE performance under different
SNRs. However, choosing variations of signals which are largely similar intro-
duces the risk of missing many well-working waveforms. As for the second ap-
proach, the goal of the thesis is not only to achieve the highest possible estima-
tion performance, but also to study the behavior of synchronization signals un-
der varying parameter configurations. Due to the problem of determining which
signals are the most popular, the third approach could not be fully used. The
selection in this thesis mostly resembles the fourth approach. The idea was that
largely different signals would more clearly show what properties of the wave-
forms made them suitable for synchronization.
The pulse was chosen because of the ease with which it could be imple-
mented and tested. However, it turned out to be hard to analyze theoretically,
because it is not continuous. Even though the peak of the ambiguity function for
5.2 Method 47

the pulse is a lot broader than that of the other signals, it was considered to be of
interest to evaluate if it could reach a synchronization accuracy ≤ 20 ns.
The chirp waveform was chosen because of its peaky ambiguity function,
that in theory makes it suitable for use as a reference signal. The same statement
holds for both WGN and the Gold sequences. The WGN and the Gold sequences
are in a way similar. Both have elements of randomness, and both have very simi-
lar ambiguity functions (see Section 2.6). However, WGN is not possible to use in
practice. WGN is included as a benchmark against which the TDE performance
of the Gold sequence was compared. The goal was to evaluate if the pseudo-
randomness of the Gold sequences would decrease its estimation accuracy. It can
be concluded that the difference in synchronization accuracy between the Gold
sequences and the WGN is negligible.
It could be considered counter-intuitive to include the Gold sequences, which
are used by GPS satellites, when the purpose of the thesis is to find signals suit-
able to replace synchronization against GPS. However, the problem with GNSS-
based solutions in this thesis is not that they do not work well when used for
synchronization. In fact, such solutions are preferable when widely distributed
sensor networks are to be synchronized [9]. The problem is that GPS signals are
vulnerable to jamming. GPS has functionality which makes it more secure, but it
is still the U.S. that has full control over who is allowed to use that functionality.
Therefore, it is suitable to study how synchronization using GNSS systems work,
to gain inspiration on how well-working reference signals are designed. There-
fore, Gold sequences were included.
That two out of the four waveforms probably were not realistic alternatives
in practice was known early in the project. However, the purpose of the thesis is
to study be behavior for synchronization signals for different parameter setups,
rather than developing the highest performing signal.

5.2.2 Theoretical Tools


For parameter estimation, the Cramér-Rao bound is a very common tool to eval-
uate estimation performance. It gives information both about the estimation per-
formance of the chosen estimator (MLE in this thesis), and how well the chosen
estimator works for different signals. The two expressions for the CRB for the
TDE, (2.17) and (2.19), were included for two different reasons. The former was
included as a benchmark the performance could be compared to. The idea was
also that the resulting CRBs from (2.17) would be compared to each other. How-
ever, as can be seen in Figure 4.1, the CRBs for the TDE were so similar that
the comparison said that the signals, in theory, should approximately give the
same estimation accuracy. This also proved to be true in the simulations, where
the chirp, the WGN, and the Gold sequences resulted in approximately the same
synchronization error when used for TDE.
The reason for using such a rough approximation as in (2.19), was because
it was easier to interpret, and the dependencies on bandwidth and signal length
became more clear. Despite the approximations, it accurately described the influ-
ence of increasing bandwidth and signal length on the estimation accuracy.
48 5 Discussion

The ambiguity function does not give any specific information about the
TDE performance, but can be useful as a tool used to determine if a signal is
suitable for TDE. For example, it is clear from the ambiguity function that the
pulse will perform worse. However, it could not be concluded from the ambigu-
ity functions whether the chirp signal, the WGN or the Gold sequences would
work better.
In conclusion, the CRB for the TDE and the ambiguity function proved to be
useful tools in order to understand the behavior of the reference signals.

5.2.3 Simulations
Simulations in Matlab is a useful tool for testing hypotheses. The problem is that
every step is idealized, and there is a risk that the results are not achievable in
practice. One assumption made in the simulations is that both the UAV and the
receivers are perfectly still, so that the position is known with high accuracy. This
is crucial since the distance between them determines the expected delay. As an
example, if the position of the UAV differs 30 cm from what is expected, the offset
will be 1 ns. It is an unreasonable assumption that the UAV will be perfectly still
in the air. However, to get an offset of 20 ns (the maximum allowable error), the
error for position can be up to 6 m. The example also assumes that the UAV will
be closer or further away from the receiver than expected. But the error might
as well be angular, or a difference in height. In summary, it is impossible to
have perfect knowledge of the distance between UAV and receiver, but it can be
reasonably assumed.
The path loss is modeled with the free-space propagation model, which is a
highly idealized model of a wireless communication channel. It does not take the
effects of multi-path propagation into account. However, LOS communication
can be reasonably assumed, since the UAV is assumed to transmit the reference
signals from a height above the TDOA sensors.
As mentioned in Section 3.1.1, the maximum output power allowed was 5 W,
but for all the simulations, the power is set to 1 W. This is because an increase in
output power would naturally only result in higher SNR. Different levels of SNR
is already tested by adjusting the variance of the AWGN, which makes variation
of the signal power redundant.

5.2.4 Laboratory Tests


The step from Matlab simulations to tests in a practical environment is a vast
one, and it could be assumed that the performance in the simulations would
not be achieved in the laboratory. Since the distances between the transmitter
and receivers were unknown, the estimation error could not be calculated in the
same way as in the simulations. This makes comparisons between the simulations
and the laboratory tests difficult. Despite the difference in method, the results
generally correspond to the simulation results.
The laboratory tests involve several elements that could be potential sources
of error. Most importantly, any tests which involve hardware might skew the
5.3 The Thesis in a Larger Perspective 49

results, and it might be difficult to identify what element of the process that is
the problem. Regardless, tests in realistic conditions are important to validate
if the simulation results are achievable in practice. In the tests carried out for
this thesis, the absolute estimation error could not be verified, but the effect of
different signal properties is, to a degree, corroborated.

5.2.5 Sources
The sources used in this thesis can be considered highly trust-worthy, as mostly
books, research articles and official websites have been used. When websites were
used, care was taken to make sure that those were recently updated, and that the
websites were run by official agencies and organizations, as opposed to regular
individuals. The majority of sources are books written for educational purposes.
Since this thesis compares already well-known signals, rather than examining
a new method, most of the information could be found in books, rather than
research articles.

5.3 The Thesis in a Larger Perspective


The technology evaluated in this thesis is meant to be used in combat. The TDOA
sensors are used to find the position of targets, so that coordinates can be given
to the artillery. These kinds of positioning systems are in an indirect sense used
as weapons, just like the sight of a rifle. This fact introduces an ethical dilemma;
how should one relate to the fact that ones seemingly harmless comparison of
synchronization signals, in the end, can be used in combat? It is up to each and
everyone to make that decision for themselves. The research carried out at FOI
is mainly used by the Swedish military, which is currently not engaged in war.
The fact that FOI is a public agency whose research and development is done
for the benefit of Sweden as a nation, strengthens the belief that the results are
produced with no intention of profiting from conflicts. The purpose of improving
the synchronization accuracy is to help the TDOA sensors make better estimates.
One positive side of that is that better accuracy might decrease collateral damage.
Even if war might be inevitable, civilian casualties should not be.
Additionally, a general improvement of synchronization accuracy in sensor
networks is beneficial for countless civilian applications as well. Technology
originally developed for military purposes can unexpectedly be found useful in
widely different areas. For example, the police or fire brigade should find the
technology discussed in this thesis useful.
Conclusions
6
In this section, the results are summarized in regard to the problem statements.
Section 6.1 concludes whether the initial problem statements have been answered,
and Section 6.2 discusses what could be further improved or expanded upon.

6.1 Problem Statements


This section will give answers to the three problem statements. For convenience,
they are repeated here, with a reference to the corresponding section.
• Section 6.1.1: What kind of waveform should be used to synchronize two
TDOA sensors as accurately as possible, using a local reference signal from
a UAV?
• Section 6.1.2: What signal properties are of highest importance when per-
forming time-delay estimation using correlation (bandwidth, signal length,
etc.)?
• Section 6.1.3: Can two TDOA receivers be locally synchronized as accu-
rately as if GNSS would have been used?

6.1.1 Type of Waveform


In the simulations, the chirp signal, WGN, and the Gold sequences give similar
estimation accuracy. However, since WGN can not be realized in practice, the
chirp signal or the Gold sequences are the best choices. Generally, it can be con-
cluded that the signal type is not of great importance when choosing a reference
signal, as long as the peak of the ambiguity function has a small width and a
large amplitude relative to the rest of the curve. The laboratory tests confirm this
conclusion, as the three tested waveforms gave similar results.

51
52 6 Conclusions

6.1.2 Effect of Signal Properties


The parameter that has the largest effect on the synchronization error is the band-
width. Generally, a maximized time-bandwidth product BT is desirable. Addi-
tionally, the sampling frequency becomes very important for high SNR. However,
for high SNR, the estimation errors are already insignificantly small. In the lab-
oratory tests, the bandwidth did not affect the difference in estimation notably.
However, only two bandwidths were tested.

6.1.3 Performance Compared to GNSS Synchronization


This thesis examines if the time-delay estimation could be performed with an er-
ror smaller than 20 ns. In the simulations, with BT = 104 , this is possible for
the chirp signal, WGN, and Gold sequences, even under very low SNR. The labo-
ratory tests do not give information about the absolute estimation error for each
receiver, but the smallest achieved difference was roughly 30 ns, which suggests
that estimation performance of that magnitude is possible.

6.2 Future Improvements


This thesis focuses on the theoretical properties of the synchronization signals,
rather than the implementation. The natural next step would be to further inves-
tigate if the simulated performance could be achieved in practice. If there was
more time, more tests would have been made in the laboratory, as a step in this
direction. When working with the implementation, it would also be highly fit-
ting to add more than two TDOA sensors. This might be trivial, but it would be
reasonable to test it thoroughly. Also, it would be appropriate if the performance
could be evaluated if some information was to be added between the reference
signals, to see if, e.g., decoding of information could be performed. The added
information could be, for example, timestamps or coordinates for the position for
the UAV.
Initially, the tests were supposed to be made with the transmitter mounted
on a UAV. Such tests would be of great interest. Ideally, tests should be carried out
over longer periods of time, to see if a network can hold synchronization using the
method of transmitting reference signals from a UAV. Furthermore, it would be
interesting to research the subject of making the synchronization against a local
UAV as discrete as possible, and/or more robust against jamming. Potentially,
beamforming could be used to direct the signals towards the TDOA receivers, so
that emission are not leaked to the surrounding environment.
Calculations for Cramér-Rao Bounds
A
The expression for the Fisher information,

∂µH (θ) −1
" # " #
∂C ∂C ∂µ(θ)
I(θ) = tr Cx−1 (θ) x Cx−1 (θ) x + 2Re Cx (θ)
∂θ ∂θ ∂θ ∂θ

is used for any signal with complex Gaussian PDF. However, for all signals con-
sidered in this thesis, the first term of the FI is zero. This is because Cx (θ) is an
N × N diagonal matrix with the noise variance σ 2 as element on the diagonal.
It is assumed that the SNR is known. Therefore, the first term of (2.12) is zero,
because ∂C∂θ
x
= 0. Also, Cx−1 (θ) = σ12 IN , where IN is an N × N identity matrix. The
mean µ is computed as

µ[n] = E{sτ + w[n]} = E{sτ } + E{w[n]} = sτ .

Using the approximation from Equation 2.3, µ is written as sτ ≈ F H Dτ F s. As


there is only one real parameter that should be estimated, θ = τ. The expression
for the FI thus simplifies to

∂µH (τ) −1 ∂µ(τ)


" #
I(τ) = 2Re Cx (τ) (A.1)
∂τ ∂τ

The derivative of the mean µ with respect to the estimation parameter is

∂µ(τ) 2πfs H
= −j F GDτ F s (A.2)
∂τ N

53
54 A Calculations for Cramér-Rao Bounds

where G = diag(n). Thus, the FI is


" H #
∂µ (τ) −1 ∂µ(τ)
I(τ) = 2Re Cx (τ)
∂τ ∂τ
" H #
2 ∂µ (τ) ∂µ(τ)
= 2 Re
σ ∂τ ∂τ
 !2 
2  2πfs  H H
H 
= 2 Re − j
 F GDτ F s F GDτ F s
σ N
2
8π f 2 h i
= 2 s2 Re s H F H DτH G H F F H GDτ F s
N σ

Since both F and Dτ are unitary [20], its conjugate transpose is equal to its inverse,
F F H = IN , and therefore the FI is
8π2 fs2 h H H H i
I(τ) = Re s F G GF s , (A.3)
N 2σ 2
resulting in the CRB for the TDE
N 2σ 2
var(τ) ≥ . (A.4)
8π2 fs2 Re s H F H G H GF s

h i
Furthermore, s H F H G H GF s can be rewritten so that it depends on the band-
width B and signal length T . Since F is the DFT matrix,
h i
, , N /2
X
s H F H G H GF s = S = F s = S H G H GS = |Sk |2 k 2 . (A.5)
k=−N /2

Then, assume that √


BN
 P, |k| ≤

2fs

|Sk | =  (A.6)
0,
 otherwise.
This can be assumed if s is limited in both time and bandwidth, which is only
approximately true. The assumption in (A.6) yields
BN
X /2fs BN
X /2fs
2
H H
S G GS = P k = 2P k2
k=−BN /2fs k=1

n(n + 1)(2n + 1)
= 2P
6 n= BN
2fs

2n3 P
, ,
= Assume n  1 ≈
3 n= BN
2fs

B3 N 3 P
=
12fs3
55

Since N = T fs , the resulting simplification is


h i B3 T 3 P
s H F H G H GF s ≈ . (A.7)
12
Combining (A.7) and (A.4), the final expression for the CRB is

3σ 2
var(τ) ≥ . (A.8)
2B3 T π2 P
Cramér-Rao Bounds for Time-delay
B
Estimation Compared to Simulation
Errors
In this section, the square root of the CRBs for the TDE are plotted against the
signal error for different bandwidths, signal lengths and SNR. The CRB for the
TDE is calculated using Equation (2.17). For all the following plots, the sampling
frequency is set to fs = 2B and the UAV is placed with equal distance to both
receivers. Because the two channels are equal when the distance to the UAV is the
same for both receivers, the plotted simulation error from channel 1 is presented.

57
58 B Cramér-Rao Bounds for Time-delay Estimation Compared to Simulation Errors

Chirp
101
Error [ns]

100

10-1
100 101
Bandwidth [MHz]
WGN
101
Error [ns]

100

10-1
100 101
Bandwidth [MHz]
Gold
101
Error [ns]

100

10-1
100 101
Bandwidth [MHz]

Figure B.1: Comparison between the square root of the Cramér-Rao bound
for the time-delay estimation and the synchronization error for three refer-
ence signals, with T = 1 ms, SNR = 0, fs = 2B, and varying bandwidth.
59

Chirp

100
Error [ns]

10-1

10-1 100
Signal length [ms]
WGN

100
Error [ns]

10-1

10-1 100
Signal length [ms]
Gold

100
Error [ns]

10-1

10-1 100
Signal length [ms]

Figure B.2: Comparison between the square root of the Cramér-Rao bound
for the time-delay estimation and the synchronization error for three refer-
ence signals, with B = 10 MHz, SNR = 0, fs = 2B, and varying signal length.
60 B Cramér-Rao Bounds for Time-delay Estimation Compared to Simulation Errors

Chirp
Error [ns]

100

10-2
-30 -20 -10 0 10 20 30
SNR [dB]
WGN
Error [ns]

100

10-2
-30 -20 -10 0 10 20 30
SNR [dB]
Gold
Error [ns]

100

10-2
-30 -20 -10 0 10 20 30
SNR [dB]

Figure B.3: Comparison between the square root of the Cramér-Rao bound
for the time-delay estimation and the synchronization error for three refer-
ence signals, with B = 1 MHz, T = 1 ms, fs = 2B, and varying SNR.
Additional Simulation Results
C
Channel 1 Channel 2
10 3 10 3
Pulse Pulse
Chirp Chirp
WGN WGN
Gold Gold

10 2 10 2
Error [ns]

Error [ns]

10 1 10 1

10 0 10 0

10 -1 10 -1
0 1 0 1
10 10 10 10
Bandwidth [MHz] Bandwidth [MHz]

Figure C.1: The effect of different bandwidths on the synchronization error,


with T = 100 µs, SNR = 0 dB, and fs = 2B.

61
62 C Additional Simulation Results

Channel 1 Channel 2
10 6 10 6
Pulse Pulse
Chirp Chirp
WGN WGN
10 5 Gold 10 5 Gold

10 4 10 4

10 3 10 3
Error [ns]

Error [ns]
2 2
10 10

10 1 10 1

10 0 10 0

10 -1 10 -1
-1 0 -1 0
10 10 10 10
Signal length [ms] Signal length [ms]

Figure C.2: The effect of different signal lengths on the synchronization er-
ror, with B = 1 MHz, SNR = 0 dB, and fs = 2B.

Channel 1 Channel 2
10 6 10 6
Pulse Pulse
Chirp Chirp
WGN WGN
Gold Gold
10 5 10 5

10 4 10 4
Error [ns]

Error [ns]

10 3 10 3

10 2 10 2

10 1 10 1

10 0 10 0
-30 -20 -10 0 10 20 30 -30 -20 -10 0 10 20 30
SNR [dB] SNR [dB]

Figure C.3: The effect of different SNRs on the synchronization error, with B
= 1 MHz, T = 1 ms, and fs = 2B.
63

Channel 1 Channel 2
10 5 10 5
Pulse Pulse
Chirp Chirp
WGN WGN
10 4 Gold 10 4 Gold

10 3 10 3

10 2 10 2
Error [ns]

Error [ns]

1 1
10 10

0 0
10 10

-1 -1
10 10

-2 -2
10 10
-30 -20 -10 0 10 20 30 -30 -20 -10 0 10 20 30
SNR [dB] SNR [dB]

Figure C.4: The effect of different SNRs on the synchronization error, with B
= 10 MHz, T = 100 µs, and fs = 2B.
Additional Laboratory Results
D
4
10
Difference in estimated delay [ns]

Pulse
Chirp
Gold

10 3

10 2

1
10
-10 -5 0 5 10
Output power [dBm]
Figure D.1: Resulting differences in error from the laboratory test for differ-
ent output powers, with B = 1 MHz and T = 10 µs.

65
66 D Additional Laboratory Results

Difference in estimated delay [ns] 10 4


Pulse
Chirp
Gold

10 3

2
10

1
10
-10 -5 0 5 10
Output power [dBm]

Figure D.2: Resulting differences in error from the laboratory test for differ-
ent output powers, with B = 1 MHz and T = 100 µs.

10 4
Difference in estimated delay [ns]

Pulse
Chirp
Gold

10 3

10 2

10 1
-10 -5 0 5 10
Output power [dBm]
Figure D.3: Resulting differences in error from the laboratory test for differ-
ent output powers, with B = 500 kHz and T = 10 µs.
67

10 4
Difference in estimated delay [ns] Pulse
Chirp
Gold

10 3

10 2

1
10
-10 -5 0 5 10
Output power [dBm]
Figure D.4: Resulting differences in error from the laboratory test for differ-
ent output powers, with B = 500 kHz and T = 100 µs.

10 4
Difference in estimated delay [ns]

Pulse
Chirp
Gold

10 3

10 2

10 1
-10 -5 0 5 10
Output power [dBm]

Figure D.5: Resulting differences in error from the laboratory test for differ-
ent output powers, with B = 500 kHz and T = 1 ms.
Bibliography

[1] Lars Ahlin, Jens Zander, and Ben Slimane. Principles of Wireless Commu-
nication. Studentlitteratur, first edition, 2015.
[2] David W. Allan, Neil Ashby, and Clifford C. Hodge. The Science of Time-
keeping Application. Technical report, Hewlett-Packard, 1997.
[3] Geoffrey Blewitt. Basics of the GPS Technique: Observation Equations. Tech-
nical report, Department of Geomatics, University of Newcastle, 1997.
[4] Common View GPS Time Transfer. https://www.nist.gov/pml/time-and-
frequency-division/atomic-standards/common-view-gps-time-transfer,
2019. Accessed: 2020-02-07.
[5] DCF77. https://www.ptb.de/cms/en/ptb/fachabteilungen/abt4/fb-44/ag-
442/dissemination-of-legal-time/dcf77.html. Accessed: 2020-04-20.
[6] Robert Gold. Optimal binary sequences for spread spectrum multiplexing.
IEEE Transactions on Information Theory, pages 619–621, February 1967.
[7] Fredrik Gustafsson. Statistical Sensor Fusion. Studentlitteratur, third edi-
tion, 2018.
[8] Fredrik Gustafsson, Lennart Ljung, and Mille Millnert. Signal Processing.
Studentlitteratur, first edition, 2010.
[9] Daniel Johansson. Tidsynkronisering av ett TDOA-baserat positioner-
ingssystem. Technical report, Totalförsvarets Forskningsinstitut, 2002.
[10] Steven M. Kay. Fundamentals of Statistical Signal Processing Vol. 1: Estima-
tion Theory. Prentice Hall, 1993.
[11] Erik G. Larsson. Signals, Information and Communications. Department of
Electrical Engineering, Linköping university, 2018.
[12] Michael A. Lombardi, Lisa M. Nelson, Andrew N. Novick, and Victor S.
Zhang. Time and Frequency Measurements Using the Global Positioning
System. Technical report, National Institute of Standards and Technology,
Time and Frequency Division, 2001.

69
70 Bibliography

[13] Navipedia. https://gssc.esa.int/navipedia/index.php/Main_Page, 2018. Ac-


cessed: 2020-02-06.

[14] Mikael Olofsson. Signal Theory. Studentlitteratur, first edition, 2011.


[15] One Way GPS Time Transfer. https://www.nist.gov/pml/time-and-
frequency-division/atomic-standards/one-way-gps-time-transfer, 2019. Ac-
cessed: 2020-02-05.

[16] Lennart Råde and Bertil Westergren. Mathematics Handbook for Science
and Engineering. Studentlitteratur, fifth edition, 2004.
[17] Mark A. Richards. Fundamentals of Radar Signal Processing. McGraw-Hill
Education, second edition, 2014.
[18] Y. Saigusa. Quartz-Based Piezoelectric Materials. In Advanced Piezoelectric
Materials, pages 197–233. Elsevier, January 2017.
[19] Global Positioning System Standard Positioning Service Performance Stan-
dard. United States Department of Defense, fourth edition, 2008.
[20] Arie Yeredor and Eyal Angel. Joint TDOA and FDOA estimation: A condi-
tional bound and its use for optimally weighted localization. IEEE Transac-
tions on Signal Processing, pages 1612–1623, April 2011.

You might also like