YA DOA Dual Antenna

You might also like

Download as pdf or txt
Download as pdf or txt
You are on page 1of 8

EUROPEAN COOPERATION CA20120

IN SCIENCE TD(22)01TD(22)01049
AND TECHNOLOGY Bologna, IT
———————————————— February, 28th, 2022
EURO-COST
————————————————
SOURCE: Université libre de Bruxelles (ULB)

Direction-of-arrival estimation with virtual multi-antenna arrays using dual-antenna


receivers : algorithms and controlled experiments.
Youssef Agram, Jianqiao Cheng, François Quitin

50 Av. Roosevelt, 1050 Brussels, Belgium


Phone: +3226502803
Fax: —
Email: Youssef.Agram@ulb.be
Direction-of-arrival estimation with virtual
multi-antenna arrays using dual-antenna
receivers : algorithms and controlled
experiments.
Youssef Agram* , Jianqiao Cheng* and François Quitin*
*
Brussels School of Engineering, Université libre de Bruxelles (ULB), 50 Av. Roosevelt, 1050 Brussels, Belgium

Abstract—Localization of radio-frequency (RF) trans- Previous work advised the use of virtual antenna
mitters can be achieved using virtual multi-antenna arrays. arrays for DoA estimation [7]–[10]. Virtual antenna
The method relies on a mobile, single-antenna receiver that arrays rely on moving, single-antenna receivers that cap-
captures successive messages from a transmitter, thereby
emulating a multi-antenna array. However, two main chal- ture multiple messages from the transmitter along their
lenges emerge from this technique: 1) Successive positions trajectory, as shown in Figure 1. The successive received
and orientations of antennas have to be determined, meet- messages will be processed similarly to the messages
ing spatial Nyquist criterion; 2) the local oscillator (LO) received by different antennas in a multiple-antenna
frequency offset between transmitter and receiver adds a array, allowing to recover the DoA of the transmitter.
drifting phase component to the received signal on each
antenna of the array. In this paper, we extend the algorithm Since virtual antenna arrays do not require any addi-
to dual-antenna receivers, which allows to recover the tional hardware in RF chipsets, they offer an attractive
azimuth (without ambiguity) and the elevation. A linear alternative for DoA estimation to large, expensive multi-
model of the LO frequency drift is also used, allowing for antenna arrays.
low-quality, drifting LOs. Controlled experimental results
from a software-defined radio testbench are presented.
Increased performance is noted when moving from a single
antenna to a multi-antenna system.
Index Terms—Direction of arrival, MUSIC, Virtual
Multi-antenna array, RF transmitter localization.

I. I NTRODUCTION

Localization of radio frequency (RF) transmitters is an


important service of current and future wireless systems,
and an indispensable feature in many applications, such
as sensor networks, vehicular networks and autonomous Fig. 1. Virtual array concept for DoA estimation. A virtual array
is created by the receiver’s motion, where each antenna element
robotics navigation systems [1], [2]. corresponds to a packet received from the transmitter.
Several metrics can be used for determining infor-
mation about a transmitter’s location, such as time- One problem when using virtual antenna arrays for
of-flight, time-difference-of-arrival or direction-of-arrival DoA estimation is related to the observability of the DoA
(DoA) [3]–[5]. The latter is often used because it does estimation problem for certain receiver trajectories [10].
not require any synchronization between anchor nodes, If the receiver performs a uniform linear motion, it
and the exact waveform of the transmitter need not becomes impossible to recover the DoA unambiguously.
be known [1]. If multiple anchor nodes perform DoA This can be solved by using multi-antenna receivers
estimation, the location of the transmitter can be easily whose antennas span a 2D plane, as will be proposed
deduced through triangulation [6]. The major drawback in this paper.
of DoA estimation is the need of physically large multi- This paper builds on the work on virtual multi-antenna
antenna arrays, which makes the technique unsuitable arrays proposed in [9] and [10], adding the following
for portable consumer electronics. main contributions:
• we consider the use of dual-antenna receivers n-th received packet (where n = 1, · · · , N ), which can
(whose physical size constraints are compatible be written as:
with consumer electronics), and extend the virtual
DoA estimation algorithms to account for two an- rA [n, m] = hA [n, m] ∗ s[m]
tenna elements; tn +mT
R s
j(ϕA
0 +2π f0 (τ )dτ )
• we consider a LO model that accounts for a linear ·e 0 + ω A [n, m] (1)
frequency drift of the LO, as can often be observed
in low-quality oscillators. where hA [n, m] is the wireless channel impulse response
Note that our previous work [10] still considered a stop- seen by antenna A, ϕA 0 is the phase of the initial
and-start approach to estimate the LO frequency offset packet received by antenna A, f0A is the frequency offset
and the DoA, placing severe constraints on the move- between front-end of the transmitter and front-end of
ment of the receiver. This approach has been abandoned the receiver, tn is the elapsed time between the first
in favour of a joint estimation approach that does not packet and the n-th packet, Ts stands for the receiver
place any constraints on the receiver movement. sample time and ω A [n, m] represents an independent and
The paper is organized as follows. Section II lays the identically distributed Gaussian noise with distribution
theoretical foundations of the DoA estimation algorithm ω A [n, m] ∼ CN (0, σ 2 ). Equation (1) contains the time
for virtual antenna arrays and explains main differences integral of the LO frequency offset which corresponds to
between single and multi-antenna systems. Section III its cumulative effect on the phase of the received signal
describes the hardware and software setup used to emu- by antenna A. As long as the transmitter up-conversion
late wireless scenarios. It also presents the data process- chain and the receiver down-conversion chain are not
ing, and shows examples of data processing for a single turned off between packets, the term ϕA 0 will remain
measurement. All results are presented in Section IV, constant between multiple received packets.
in which a performance comparison between single and Let us only consider the case of the line-of-sight path
dual-antenna receivers is given. The differences between for a narrowband channel such that hA [n, m] can be
the different emulated scenarios is presented, as well as simplified as follows:
the effect of receiver trajectory size. ⃗ A
hA [n, m] = α · ej β·⃗r [n]
(2)
II. V IRTUAL AOA ESTIMATION ALGORITHM where α denotes the complex channel amplitude, β ⃗ the
A. Single-antenna system wave vector and ⃗rA [n] are the relative coordinates of
antenna A when receiving the n-th packet with respect
We start by introducing the basics of DoA estimation to the initial one.
with virtual antenna arrays, based on the the work in
In a general 3D configuration, phase-shift due to
[10]. Let us consider a dual-antenna receiver (as shown
position ⃗rA [n] can be rewritten as follows:

2π A
β⃗ · ⃗rA [n] = r [n] cos(θ) cos(φ)
λ E
A A

+rN [n] sin(θ) cos(φ) + rU [n] sin(φ) (3)

where θ and φ respectively represents the azimuth and


elevation angles-of-arrival of the transmitter, λ is the
wavelength at the carrier frequency and (rE , rN , rU )A
denotes coordinates of antenna A in global E-N-U frame
Fig. 2. Example of a scenario with a dual-antenna receiver moving in
with respect to its initial position. The final expression
the North-East plane. of received signal at antenna A can be derived using
equations (1), (2) and (3):
in Figure 2) moving in a 3D East-North-Up (E-N-U)
tn +mT
R s
space. In the case of a single-antenna system A, let j(ϕA
0 +2π f0 (τ )dτ )
A
us consider the transmitter sends successive identical r [n, m] = α · s[m] · e 0

packets, whose baseband representation is given by s[m] ⃗ A


· ej β·⃗r [n]
+ ω A [n, m] (4)
(where m = 1, · · · , M ) represents the m-th sample of
the packet. These packets could be distress beacons, The purpose is to perform a DoA estimation based
or simply the fixed header of a given communication on the model in (4), using existing techniques for
standard (e.g. the PSS sequence in a LTE base station). conventional multi-antenna arrays. However, two major
Let us denote rA [n, m] the m-th baseband sample of the differences arise in the virtual array formalism:
1) Positions of antenna A (rE [n], rN [n], rU [n])A where A(β0 , β1 , θ, φ) is the steering vector of the vir-
should be estimated for each received packet i.e. tual, dual-antenna array:
with a fraction-of-a-wavelength accuracy.
2) The frequency offset f0A [n] (and it’s accumulated A(β0 , β1 , θ, φ) =
tn +mT
R s    t2 [1]
  
j 2π β0 t[1]+β1 ⃗ r A [1]
+β(θ,φ)·⃗
phase offset f0 (τ )dτ ) should be estimated 2

0
 e 
as well.
 .. 

   .  


2
The first problem can be solved with the use of inertial 
 j 2π β0 t[N ]+β1 t 2[N ] +β(θ,φ)·⃗
⃗ r A [N ] 

measurements units (IMUs). Acceleration and gyroscope e 
   
 j 2π β t[1]+β t2 [1] +β(θ,φ)·⃗
  (9)
measurements can be integrated through an Extended 0 1 2
⃗ B
r [1]

 e
 
Kalman filter (EKF) or Unscented Kalman filter (UKF), 
 .. 
providing a 3D navigation solution [11]. 
   . 


The second limitation requires to estimate the LO fre-

t2 [N ]

j 2π β0 t[N ]+β1 B ⃗
+β(θ,φ)·⃗
r [N ]
2
quency offset (and it’s accumulated phase) based on the e
received signal. We will see in Section III that a time
The terms wA [m] and wB [m] are the N × 1 white
linear model is appropriate to model the drift of the LO
Gaussian noise vectors for antenna A and B respectively.
frequency offset:
Finally, the term x[m] is defined as:
f0 [n] = β0 + β1 · t[n] (5) x[m] = α · s[m] · ejϕ0
For this final equation, it is important to note that
where β0 and β1 are the unknown parameters. If we the system should be calibrated appropriately, such that
make the assumption that f0 [n] is constant within packet ϕA B
0 = ϕ0 = ϕ0 .
n, the final expression for (4) can be derived: Conventional DoA estimation algorithms, such as
   beamforming or MUSIC, can be used with the signal
t2 [n]
A
j ϕA
0 +2π. β0 t[n]+β1 2
system model (8). Let S = E{YY∗ } the 2N × 2N
r [n, m] = α · s[m] · e covariance matrix (where E. and (.)∗ are the expectation
⃗ A
· ej β·⃗r [n]
+ ω A [n, m] (6) and the Hermitian operator, respectively). In the case of
the MUSIC algorithm, we search for the K components
B. Multi-antenna system and joint estimation algorithm orthogonal to the noise subspace [12] (K = 1 if there is
only one signal source). Let Ew be the 2N × (2N − K)
Let us consider the scenario where we have a dual- matrix of eigenvectors corresponding to the 2N − K
antenna receiver moving in the 3D E-N-U space, as smallest eigenvalues of S. The MUSIC spectrum is
shown in Figure 2. computed as follows (where the dependence of A on
The signal received by the second antenna (denoted (β0 , β1 , θ, φ) has been omitted for clarity):
antenna B is given by): 1
PM U (β0 , β1 , θ, φ) = (10)
  
A∗ Ew E∗w A
t2 [n]
j ϕB
0 +2π β0 t[n]+β1 Therefore, the problem of finding the DoA (along-
rB [n, m] = α · s[m] · e
2

side the parameters β0 and β1 ) is reduced to a four-


⃗ B
· ej β·⃗r [n]
+ ω B [n, m] (7) dimensional search:
 
Since the front-ends associated to antennas A and B (β̂0 , β̂1 , θ̂, φ̂) = arg max PM U (β0 , β1 , θ, φ)
β0 ,β1 ,θ,φ
are driven by the same local oscillator, we can assume (11)
that their frequency offset is identical. If the receiver has In the next section, we will present how the two scenarios
been calibrated appropriately, the initial phase difference of Fig.3 have been emulated in a controlled laboratory
between the two front-ends can be estimated and com- experiment.
pensated, such that ϕA B
0 = ϕ0 = ϕ0 . The received packets
on both antennas can be stacked into a column vector III. E XPERIMENTAL SETUP
and expressed as follows: A. Hardware and software setup
The wireless scenarios from Figure 3 have been emu-
yA [m]
   A   A 
A [m] w [m] lated through a controlled cabled experiment. An USRP-
= · x[m] + (8)
yB [m] AB [m] wB [m] B205 mini software-defined radio (SDR) has been used
| {z } | {z }
y[m] A(β0 ,β1 ,θ,φ) as a transmitter and USRP-E310 SDR has been used as
received packets (at a rate of 1 ms) are dumped into a
data file for offline processing. The other parameters are
set as follows: the carrier frequency is set to 2 GHz, and
the sampling frequency is set to fs = 614.4 kHz. Packets
are sent over a period of 10s. Regarding the receiver
trajectory, the rail-mounted receiver travels a distance of
1.22m along the East axis for about 2s. The USRP-E310
is configured to wait for 10 s before it starts to record
packets.
B. Data processing
In this section, a single data set for the 2nd scenario
(with the 90°-splitter) has been isolated to give an
example of the data processing. Figure 4 illustrates the
measured accelerations after correction of the initial
orientation. The receiver waits for 10 s before it starts
Fig. 3. Illustration of the two considered scenarios with dual-antenna recording data packets, and it waits for another 5 s before
receiver (A and B) moving in the North-East plane.
it starts moving. It can be observed that only the local y-
axis includes acceleration and deceleration, since it was
a receiver. Both were attached to a mechanical guided oriented in the east direction.
rail. Two types of RF power splitters were used split the
signal from the transmitters to the two front-ends of the
USRP-E310, thereby emulating a wireless Tx-to-dual-
antenna-Rx situation. More specifically:
1) 0°-splitter ZFRSC-4-842-S+, adding no phase-
shift between front-ends A and B of the USRP-
E310 (top scenario of Figure 3).
2) 90°-splitter ZX10Q-2-25-S+, adding a phase-shift
of +90° to front-end A with respect to front-end
B (bottom scenario of Figure 3).
The introduced RF power splitters emulate the spatial Fig. 4. Example of IMU measurements - accelerations of receiver
(90°-splitter).
phase-shifts that would have been introduced by the
positions of antenna A and B as shown in Figure 3. In
The LO frequency offset f0 [n] is computed by looking
fact, if d denotes the relative distance between antennas
at the derivative of the phase of the received packets.
A and B, and is set to d = λ2 , then their relative phase-
Figure 5 confirm the time-linear nature of the drift of
shift can be computed:
the LO frequency offset. The estimated linear regression
∆ϕAB = β⃗ · ∆⃗rAB during motion is shown in black dashed line.
−2π
= . d sin(γ)
λ
A phase shift of 0° between front-ends A and B cor-
responds to an inclination of the dual-antenna array of
0°, whereas a phase shift of +90° between front-ends A
and B corresponds to an inclination of the dual-antenna
array of -30°. The transmitter sends periodic LTE PSS
sequences with a period of T0 = 1 ms. The receiver
USRP-E310 is equipped with an MPU-9150 Inertial Fig. 5. Example of computed and estimated LO frequency offset based
Measurement Unit (IMU) for measuring accelerations on linear regression (90°-splitter).
and angular speeds along all 3 axes. It also contains
a customized FPGA that implements the first stages of The final MUSIC spectrum depends on 4 parameters
an LTE receiver, i.e. a correlator and a peak detector which makes it difficult
to illustrate. We decided to plot
to detect the boundaries of the received packets. The

it as PM U (θ, β1 )
i.e. as function of θ and β1
customized FPGA forwards the received packets to the ˆ
(φ̂,β0 )
host software. The IMU data (at a rate of 5 ms) and the considering the estimated φ̂ and βˆ0 (see Figure 6 ). Note
that elevation angle φ̂ has been set to 0° to simplify the
search. Three observations can be made:
• A main lobe appears at an azimuth of 93°, which
is closed to the predicted value.
• No lobe appears around −90° (which corresponds
to the ambiguity we observe with a single-antenna
receiver).
• A side lobe is present at −135°, for a slightly
different value of β1 .

Fig. 6. Example of MUSIC spectrum in (θ, β1 , PM U ) space (90°- Fig. 7. Comparison of PDFs - single vs. dual antenna receiver (35
splitter). Main lobe at (93◦ , 3.41 s12 , −23.70 dB) and side lobe at processed data).
(−135◦ , 3.61 s12 , −26.88 dB)

The Root Mean Square Error (RMSE) of the DoA


IV. M EASUREMENT RESULTS estimates (after removing outliers and ambiguities) is
Multiple measurements were conducted for each of estimated, and shown in Figure 8. The following RMSE
the emulated scenarios: are computed for both systems:
• 28 measurements using the 0°-splitter.
0°-splitter 90°-splitter
• 35 measurements using the 90°-splitter.
1 antenna 8.96° 8.96°
The following sections aim to compare single-antenna 2 antennas 6.01° 4.20°
and dual-antennas receivers, both scenarios from Fig-
ure 3 and to study the effect of travelled distances. The results show a significant decrease of the RMSE
with an additonnal antenna at the receiver.
A. Single-antenna and dual-antennas receivers
We start by comparing a single-antenna receiver (by
considering only the measurements from the first an-
tenna) with a dual-antenna receiver. The probability
density function of the estimated DoA for the 90°-splitter
is given in Figure 7. For the single-antenna receiver, we
can see that 28.57% of the DOA estimates are in the
interval 90◦ ±20◦ , 51.43% are in the interval −90◦ ±20◦
and 20% are considered to be outliers. As expected, a
single-antenna receiver moving along a linear trajectory
is not able to resolve the ambiguity between +90◦ and
−90◦ .
For the dual-antenna receiver, we can see that 71.43% Fig. 8. Comparison of RMSE - single vs. dual-antenna receiver.
of the DOA estimates are in the interval 90◦ ± 20◦ ,
without any results in −90◦ ± 20◦ , which confirms the
ambiguity removal. However, we obtained 28.57% in B. 0° and 90°- splitters
[−160◦ , −140◦ ]. It seems that side lobes identified in We now compare the estimated DoA results for a dual-
previous Section might be higher than the main lobes in antenna array, when using the 0°- and 90°-splitters. The
some cases, leading to a false result. results are shown in Figure 9. It can be seen that for
the 0°-splitter, there is an ambiguity between +90◦ and
−90◦ , which can be easily explained when analyzing
the emulated scenario. Using a 0°-splitter emulates a
scenario where the dual-antenna receiver is oriented in
parallel to the direction of the movement. The virtual
array only spans a one-dimensional space (as opposed
to a two-dimensional space), which does not allow to
raise the ambiguity in the estimated DoA. This is not
the case with the 90°-splitter (which emulates a scenario
where the dual-antenna receiver has an angle of -30°
with respect to the direction of the movement). Note
that, as mentioned previously, a sidelobe of the steering Fig. 10. Comparison of CDFs - effect of travelled distances for dual-
vector does produce some erroneous results as well for antenna system with 90°-splitter.
the 90°-splitter.

V. C ONCLUSION
This paper investigates the use of dual-antenna re-
ceivers for performing virtual DoA estimation, as well
as the use of low-quality oscillators with drifting LO
frequency offset. We propose a DoA estimation model
that can account for dual-antenna receivers, and LO fre-
quency offset that has a linear drift over time. We present
a proof-of-concept experiment to show the feasibility of
the proposed algorithm. The experiment results show
that dual-antenna receiver outperforms single-antenna
receivers, and are able to resolve the remaining ambi-
Fig. 9. Comparison of CDFs - 0°- vs. 90°-splitters for dual antenna guity in the case of linear virtual arrays. However, the
system. appearance of high side lobes in the multi-dimensional
steering vector (that spans the angles-of-arrival, as well
as the LO frequency offset parameters) cause a certain
amount of misdetections. Our future work will investi-
C. Effect of travelled distance gate which trajectories, from a theoretical point-of-view,
allow to obtain the best DoA estimation performances.
Finally, we investigate the effect of travelled distance We will also test our proposed system in fully-wireless
on the DoA estimation using measurements from exper- scenarios, as opposed to the controlled emulations that
iments with 90°-splitter. For this, only 400 packets have were performed in this paper.
been selected to generate a virtual array of size 29.7 cm.
Subsampling has then been applied to generate two other R EFERENCES
virtual sizes i.e. 57.8 cm and 112.5 cm. To have the same
[1] Klaus Witrisal and Carles Antón-Haro, “Whitepaper on New
number of elements in all three virtual arrays, it implies Localization Methods for 5G Wireless Systems and the Internet-
that: of-Things,” Tech. Rep., COST Action CA15104 - IRACON,
2018.
• Selecting one packet out of 4 allows to cover 112.5 [2] Claude Oestges and Francois Quitin, Inclusive Radio Communi-
cm with the virtual array. cations for 5G and Beyond, Elsevier, 2021.
[3] N. Patwari, J.N. Ash, S. Kyperountas, A.O. Hero, R.L. Moses,
• Selecting one packet out of 2 allows to cover 57.8 and N.S. Correal, “Locating the nodes: cooperative localization
cm. in wireless sensor networks,” IEEE Signal Processing Magazine,
• No subsampling is needed for 29.7 cm. vol. 22, no. 4, pp. 54–69, 2005.
[4] Hui Liu, Houshang Darabi, Pat Banerjee, and Jing Liu, “Survey
The results are shown in Figure 10. While one would of wireless indoor positioning techniques and systems,” IEEE
expect that larger arrays would have lower errors, we Transactions on Systems, Man, and Cybernetics, Part C (Appli-
cations and Reviews), vol. 37, no. 6, pp. 1067–1080, 2007.
can see that the amount of misdetection increases with [5] Henk Wymeersch, Jaime Lien, and Moe Z. Win, “Cooperative
travelled distance. This is due to the fact that, for larger localization in wireless networks,” Proceedings of the IEEE, vol.
array, the sidelobes in the array steering vector are larger 97, no. 2, pp. 427–450, 2009.
[6] Kutluyil Doğançay, “Bearings-only target localization using total
than for physically smaller arrays, increasing the risks of least squares,” Signal Processing, vol. 85, no. 9, pp. 1695–1710,
misdetection. 2005.
[7] M. A. Yaqoob, A. Mannesson, N. R. Butt, and F. Tufvesson,
“Source localization using virtual antenna arrays,” in 2015
International Conference on Localization and GNSS (ICL-GNSS),
2015, pp. 1–6.
[8] A. Mannesson, M. A. Yaqoob, B. Bernhardsson, and F. Tufves-
son, “Tightly coupled positioning and multipath radio channel
tracking,” IEEE Transactions on Aerospace and Electronic
Systems, vol. 52, no. 4, pp. 1522–1535, 2016.
[9] François Quitin, Philippe De Doncker, François Horlin, and
Wee Peng Tay, “Virtual multiantenna array for estimating the
direction of a transmitter: System, bounds, and experimental
results,” IEEE Transactions on Vehicular Technology, vol. 67,
no. 2, pp. 1510–1520, 2018.
[10] Jianqiao Cheng, Ke Guan, and François Quitin, “Direction-
of-arrival estimation with virtual antenna array: Observability
analysis, local oscillator frequency offset compensation, and
experimental results,” IEEE Transactions on Instrumentation and
Measurement, vol. 70, pp. 1–13, 2021.
[11] Rudolph van der Merwe, Eric Wan, and Simon Julier, Sigma-
Point Kalman Filters for Nonlinear Estimation and Sensor-
Fusion: Applications to Integrated Navigation, 2004.
[12] R. Schmidt, “Multiple emitter location and signal parameter
estimation,” IEEE Transactions on Antennas and Propagation,
vol. 34, no. 3, pp. 276–280, 1986.

You might also like