GPS Solutions (2020) 24:109

https://doi.org/10.1007/s10291-020-01020-y

ORIGINAL ARTICLE

A subspace‑based code tracking loop design for GPS multi‑antenna

receiver in multipath environment
Xi Hong1 · Wenjie Wang1 · Ning Chang1 · Qinye Yin1

Received: 14 December 2019 / Accepted: 3 August 2020

© Springer-Verlag GmbH Germany, part of Springer Nature 2020

Abstract
Global Navigation Satellites Systems have been greatly developed in recent years, but receivers usually suffer the undesired
measurement bias errors caused by multipath incidence in an urban area. We propose an effective subspace-based estimator
cooperating with the conventional delay-locked loop (DLL) for Global Positioning System L1 C/A Signal. With the help of
the multi-antenna, the incident rays could be distinguished in space. A forward and backward space spatial smooth technique
is then taken to solve the coherent problem in signal subspace estimation. By using the structure of the receiving array, a
subspace rotation invariance can be built, and the angles of arrival and relative delays of incident rays could be estimated
jointly. After combining the estimates from DLL and subspace-based estimators, we can enhance the strength of the line-
of-sight ray and achieve a modified delay tracking output without multipath bias effectively. Simulation results prove that
compared with the existing methods, such as the narrow early-minus-late and the multipath estimating delay lock loop, the
proposed method has the best multipath mitigation in the delay estimation.

Keywords Multi-antenna receiver · Multipath mitigation DLL · Subspace-based estimator · FBSS · Joint delay · AOA
estimation

Introduction and generate a bias error in pseudorange measurement. This

error is environment-depended, so it is hard for the receiver
Within recent decades, the development of the Global Navi- to reduce the effects by using aid information from reference
gation Satellite Systems (GNSSs) has made much progress. stations and local area enhancement systems. It means that
The USA first developed its Global Positioning System the multipath effects shall be suppressed by the receiver in
(GPS) from the 1970s, and the new Block-III satellites were real-time applications.
launched last year with additional L1C and L5 broadcasting. There is much research on multipath mitigation. Van
China, Russia, and E.U. also developed Beidou, GLONASS, Dierendonck et al. (1992) mitigates the multipath effects
and Galileo systems, respectively. For the purposes of posi- by using a narrow early–late discriminator (nEML). This
tioning improvement, Japan and India also have their own method limits the discriminator error within a small area,
space-based local area GNSS enhancement systems (Fernan- and the tracking bias is also reduced effectively. However,
dez-Prades et al. 2011). Although these systems are essen- nEML does not cancel any multipath component so that the
tial to location navigation applications, they usually suffer bias still exists in tracking. The multipath estimating delay
performance degradations from the ionosphere, troposphere, lock loop (MEDLL) (Van Nee 1992; Townsend et al. 1995)
jammer, multipath, and other bad effects. Since more GNSS- is another well-known multipath mitigation method. It sets
based applications have been used in cities, multipath miti- up many code and phase tracking loops for different rays and
gation is now an important research point. Multipath rays distinguishes them by using a maximum-like (ML) iteration
are mainly caused by reflections surrounding the receiver search, where the effects from other rays are successively
subtracted by using their estimates. The MEDLL could sup-
press the medium and large relative delay multipath rays, but
* Xi Hong it is sensitive to the rays with short relative delays. Other
[email protected] multi-correlator techniques, such as double delta, strobe cor-
1 relator, and vision correlator, are also proposed by Garin and
Xi’an Jiaotong University, Xi’an, People’s Republic of China

13
Vol.:(0123456789)
109 Page 2 of 17 GPS Solutions (2020) 24:109

Rousseau (1997), Jia et al. (2016), Li et al. (2015), Lohan section, we first examine the multipath separation capability
and Lakhzouri (2006), and Fenton and Jones (2005). They of the proposed method when four rays are impinging the
mainly make use of the slopes or peaks of GNSS autocor- receiver. The multipath mitigation performances of different
relation functions (ACF) to mitigate multipath further in methods are then compared by the maximum error envelop
tracking results but still cannot achieve the bias-free per- (MEE) versus relative delays. The delay estimation accuracy
formance. Hsu et al. (2014) and Xu and Hsu (2019) use the of the proposed method is also compared with other exist-
vector delay lock loop to reduce the part of positioning error ing methods by the root-mean-squared errors (RMSEs) in a
in urban areas, and the 3D building model is also taken into multipath environment.
account by Kumar and Petovello (2017). These methods are
too complicated and hardly used in real time. Antenna tech-
nique is another effective way to reduce the multipath effects Received signal model
in space. With the help of antenna design (Khosravi et al.
2015), the multipath rays from low elevation angles can be The following notations are used throughout this paper. Bold
suppressed. For the multi-antenna GNSS receiver, the signal and lowercase letters denote vectors, whereas bold and capi-
in the line-of-sight (LOS) direction can also be strengthened tal letters denote matrices. The notations (⋅)†, (⋅)H , and (⋅)T
by using adaptive beamformers (Amin and Sun 2005; Lu denote the pseudoinverse, Hermitian, and transpose of a vec-
et al. 2013). An antenna array is also used to detect weak tor or matrix, respectively, and (⋅)−1 denotes the inverse of a
GPS signals in radar (Pui and Trinkle 2013) and suppress matrix. The notations ⊗ and ‖⋅‖ denote the Kronecker prod-
interference during GPS tracking (Xu et al. 2015). In order uct and the vector norm, respectively. The notation diag{⋅}
to take advantage of signal spatial and temporal characters denotes the diagonalization operation.
jointly, some subspace-based methods are also extended The multi-antenna GNSS receiver can make use of the
to estimate different ray for multi-antenna GNSS receiver space properties of signals. Apart from the jammer sup-
parameters (Chang et al. 2018a, b; Hong et al. 2018). These pression by beamforming techniques (Amin and Sun 2005;
subspace-based methods have a higher delay estimation Cuntz et al. 2016; Lu et al. 2013; Zoltowski and Gecan
accuracy than other iterated ML searching methods in the 1995), the antenna array can be also used to enhance the
static signal model due to the spatial separation capacity spatial separation ability for the different signal rays (Chang
of the array. However, they do not take the ACF code shift et al. 2018a, 2018b; Hong et al. 2018). ULA is a most typi-
caused by the Doppler effect into consideration, so it is hard cal and easiest array, and it has been wildly used in wireless
to use them directly in practice. communication (Chenu-Tournier et al. 2000), radar (Abou-
We analyze the temporal character of multipath bias for tanios et al. 2017), and GNSS signal processing (Amin and
multi-antenna GPS receiver and propose a new joint spa- Sun 2005; Chang et al. 2018a, b; Cuntz et al. 2016; Hong
tial and temporal estimator to modify the tracking result. et al. 2018; Lu et al. 2013; Zoltowski and Gecan 1995).
Unlike the existing methods, we do not focus on improving For ease of discussion, we consider that a GPS receiver is
the delay lock loop (DLL) temporal resolution, but admit the equipped with a ULA as shown in Fig. 1.
fact that the DLL follows the code phase shift (CPS) with The carrier frequency offset (CFO) of the GPS signal
bias. This bias holds the code phase difference between the is caused by Doppler effects and oscillator instability. The
local code and the received signal constant within a short
time, and it is suitable to make use of subspace techniques
to distinguish all received rays in space and time. In order
to deal with the coherent problem caused by the fact that
the multipath rays share the same navigation bits, we give a Satellite k
forward and backward spatial smooth processing in signal
subspace estimation. Afterward, a subspace rotation invari-
ance is established by the property of uniform linear array
(ULA) and steering vectors, and it also provides the joint dsin(­ k )
estimates of angles of arrival (AOA) and relative delays for
all detected rays from the eigenvalues and eigenvectors of
the cross-correlation matrix. After picking up the param-
­k ­k
eters of the LOS ray by using its temporal character, we d d
..
can finally modify the DLL tracking result and eliminate its 1 2 N-1 N
.
bias. Besides, the application of multi-antenna strengthens
the received signal power, which also improves the estima-
tion accuracy. In the simulation and performance analysis Fig. 1  GPS receiver with uniform linear array (ULA) structure

13
GPS Solutions (2020) 24:109 Page 3 of 17 109

CFO usually changes very slowly, and it could be assumed and Benson 2017; Zhu and Fan 2015), so all GPS signals in
to be constant within tens of seconds (Mahmud et al. (3) are thought to be visible and have been captured.
2016). Compared with the satellite velocity, the relative The Gold codes of GPS signal have a large spreading
movement between receiver and reflection is so little that gain (IS-GPS-200J 2018), so the signals from different sat-
different rays from one satellite are thought to share nearly ellites can be assumed to be uncorrelated with the near-far
the same CFO with different carrier phase. Therefore, the effect cancellation. Here, the kth satellite signal is set to be
received signals from ULA could be expressed as of interest in tracking. With the help of the carrier tracking,
a cross-correlation (CC) is( always) made between 𝐱(n) and
∑ ∑
K Lk
( ) ( ) the local spreading code ck nTs , and the pseudorange meas-
𝐱(t) = 𝛼k,l 𝐚k,l ck t − 𝜏k,l dk t − 𝜏k,l ej2𝜋fk (t−𝜏k,l )+j𝜑k,l + 𝐰(t)
k=1 l urement of the kth satellite is often based on its correlation
(1) peak. Then, the CC results from (3) could be approximately
where K GPS satellite signals are received by the ULA with shown by
N antennas, and all receiving channels have been calibrated � �
perfectly. 𝐰(t) is the received noise of ULA. For the kth ⎛ gk �nTs − 𝜏k,1 � ⎞
⎜ g nTs − 𝜏k,2 ⎟ � �
received satellite signal, ck (t), dk (t), and fk are its spread- 𝐫k (n) ≈ 𝐀k 𝛬𝛼,k ⎜ k ⎟dk nTs − 𝜏k,1 + 𝐰(n)
̃ (4)
⋮
ing code, navigation bit, and CFO, respectively, and it is ⎜ � �⎟
⎝ gk nTs − 𝜏k,Lk ⎠
assumed that there are Lk multipath rays impinging the ULA
from different directions. The normalized spatial steering where g(t) is the autocorrelation function (ACF) of ck (t) and
vector for each ray could be given by ̃ is the noise item. Multipath rays are often generated by
𝐰(n)
� �T the reflections around the receiver so that all rays are thought
1 2𝜋d 2𝜋d
𝐚k,l = √ 1, ej 𝜆 sin (𝜃k,l ) , … , ej 𝜆 (N−1) sin (𝜃k,l ) (2) to share the same delay effect 𝜏k,1 on their navigation bits in
N
(4) approximately.
where d is the distance between the adjacent antennas, 𝜆 is
the wavelength of the GPS signals, and 𝜃k,l is the AOA of the
lth ray of the kth signal. Its corresponding amplitude, carrier
phase, and delay are also represented by 𝛼k,l , 𝜑k,l , and 𝜏k,l , Conventional multi‑antenna receiver
respectively. The delays of different rays are set to follow problems
the relationship: 𝜏k,1 ≤ 𝜏k,2 ≤ ⋯ ≤ 𝜏k,Lk . Because the LOS
ray is always the first one arriving at the ULA, 𝜏k,1 is thought As shown in Fig. 2, the majority of the existing works often
to be its delay, and 𝜏k,l (l ≠ 1) denotes the delay of multipath process the received signal from antennas by two cascaded
ray. There is always some power loss by the reflections, then parts: spatial part and temporal tracking. The spatial part
|𝛼 | > |𝛼 |(l ≠ 1) could also be met. could be realized by some RF antenna techniques (Caiz-
| k,1 | | k,l |
The receiver structure in Fig. 1 follows the zero-IF RF zone et al. 2019) or many adaptive beamformers in base-
strategy with I/Q problem calibration, and the outputs of band signal processing (Amin and Sun 2005; Lu et al. 2013;
the ULA are down-converted into baseband orthogonally Zoltowski and Gecan 1995). With the help of the receiving
and sampled by parallel ADCs simultaneously. Then, the array, the signals can be distinguished in space and their
discretized signal could be written as characters could also be taken to improve the signal-to-
� � � � interference-and-noise ratio (SINR).
⎛ ck �nTs − 𝜏k,1 �dk �nTs − 𝜏k,1 � ⎞ For the spatial part, no matter which kinds of spatial
� K
⎜ c nTs − 𝜏k,2 dk nTs − 𝜏k,2 ⎟ j2𝜋fk nTs
𝐀k 𝛬𝛼,k ⎜ k beamforming techniques used, they are all equivalent to a
𝐱(n) = ⎟e + 𝐰(n)
k=1 ⎜ � �⋮ � �⎟
⎝ ck nTs − 𝜏k,Lk dk nTs − 𝜏k,Lk ⎠
(3)
( )
where 𝐀k = 𝐚k,1 , … , 𝐚k,Lk is the steer ing matr ix
of all rays of the kth satellite signal. The corre-
sponding complex amplitudes are represented by
𝛬𝛼,k = diag{𝛼k,1 ej𝛽k,1 , … , 𝛼k,Lk ej𝛽k,Lk } , in which
( 𝛬𝛼,k is a)
diagonal matrix, and 𝛽k,l = −2𝜋fk 𝜏k,l + 𝜑k,l l = 1, … , Lk
are the phases of its diagonal elements. There is much
research discussing GPS signal acquisitions on L1, L2, and
L5 bands (Leclere et al. 2017; Mao and Chen 2009; Qaisar
Fig. 2  Conventional spatial and tracking processing structure for
multi-antenna receiver

13
109 Page 4 of 17 GPS Solutions (2020) 24:109

linear spatial combination with a weight vector 𝐯k . Then, the Fig. 2 structure takes the signal temporal and spatial free-
signal after beamforming can be shown as doms cascaded, which is not optimal in the view of signal
� � processing. Therefore, the multipath mitigation might be
⎛ g�nTs − 𝜏k,1 � ⎞ improved if the signal properties in time and space are taken
⎜ g nTs − 𝜏k,2 ⎟ � �
𝐯H H H into account jointly.
k k (n) ≈ 𝐯k k 𝛼,k ⎜ ⎟dk nTs − 𝜏k,1 + 𝐯k 𝐰(n)
𝐫 𝐀𝛬 ̃
⋮
⎜ � �⎟
⎝ g nTs − 𝜏k,Lk ⎠
(5) Tracking bias analysis and novel code loop
The item 𝐯H 𝐀 𝛬 here is a row vector, and its elements
k k 𝛼,k design
are corresponding to the complex amplitudes of all rays
after beamforming. In open environments, there is usually Before discussing the details of the joint spatial and temporal
only LOS ray impinging the ULA. By using a spatial weight tracking loop design, we would like to analyze more signal
steering to the LOS direction, the signal strength could be temporal characters in a multipath environment. Due to the
improved effectively; then, the tracking results could be weak power of the GNSS signals, the multipath rays are
upgraded obviously. However, when the receiver is in the mainly caused by the reflections around the receiver, and
urban areas, although the beamformer in (5) can enhance the relative delays between LOS and multipath rays (RDLM)
ray strength of the LOS and suppress components from are not too large. From the ACF of the GPS L1 C/A code,
other directions, the beamformed signal may still suffer the it could be found that the ACF waveform will have an obvi-
effects from multipath rays. For example, there are an LOS ous distortion if the RDLM is less than 2Tc , which means
ray and a reflection ray impinging the receiver ULA, whose that the receiver cannot distinguish LOS ray easily. It was
adjacent antennas distance is set (to 𝜆∕ 2. )The temporal
( and
) given by Van Nee (1992), Townsend et al. (1995), and Van
spatial parameters of the ray are 0Tc , 0◦ and 0.5Tc , 40◦ , Dierendonck et al. (1992) that the conventional tracking loop
respectively. If the 𝐯k is the steering vector of the LOS ray, has to deal with the delay estimation bias when the RDLM
although the multipath ray is suppressed, there is still 23.1% is less than Tc, so we mainly focus on the multipath problem
unwanted component left after beamforming, which causes with the RDLM less than Tc.
an ACF distortion as shown in Fig. 3. The RDLM often causes an unwanted positioning bias
Generally, the multipath distortion after beamforming in error, but it may change slowly due to the small Dop-
Fig. 3 can be further relieved by some temporal mitigation pler frequency difference between different rays (Irsigler
methods, such as nEML, MEDLL, and WRELAX (Jia et al. 2010). According to the measurements in an urban can-
2016). These existing methods are proposed for temporal yon by Peng and Petovello (2015), the differential Doppler
processing with a single antenna, so they do not consider frequency between the LOS ray and the multipath ray is
the signal spatial characters. Therefore, the similar disadvan- always observed within 10 Hz when the multipath signal
tages of these methods, such as poor mitigation performance power is higher than 25 dB Hz . It means that the RDLM
with short relative delay, shall also be met, regretfully, when change between rays shall be smaller than 6.6 × 10−4 Tc
they are used in Fig. 2 structure. It is mainly caused by that during 100 ms tracking. The observations by Peng and
Petovello (2015) were taken from a moving GPS receiver,
so it can be implied that the RDLM may remain unchanged
1400 if the receiver is in low dynamic. Therefore, the RDLMs
1200 are assumed here to hold constant within a short time
approximately.
1000 Based on this assumption, although the ACF has been
Amplitude

800 affected in a multipath environment, it may hold the dis-

torted waveform unchanged and make the bias not change
600 in a short time. It means that the DLL may suffer the delay
400 estimate bias, but it still makes the local code follow the
ACF before Beamforming CPS of the satellite signal. Besides, the temporal relation-
200 ACF after Beamforming ships between received rays at the DLL discriminator are
also held. Therefore, a novel joint spatial and temporal loop
0
-3 -2 -1 0 1 2 3 structure is proposed in Fig. 4.
Delay(Tc) There are two main functional parts in this structure:
a DLL with a beamformer and a subspace-based spatial
Fig. 3  ACF distortion before and after beamforming when the LOS and temporal estimator. The former is used to enhance the
ray and multipath ray are from 0° and 40°, respectively LOS ray’s power during tracking, and the latter is used to

13
GPS Solutions (2020) 24:109 Page 5 of 17 109

Fig. 4  Subspace-based joint spatial and temporal loop structure

distinguish the received multipath rays. In Fig. 4 structure, Cross-Correlation at 0.0ms

the received signal was firstly beamformed according to the 1200
Cross-Correlation at 100.0ms
last time AOA estimate for LOS ray, which could increase
1000
the SINR of the specific ray and reduce the effect from
multipath. The beamformed signal is then tracked by a con- 800
Amplitude

ventional DLL, and a coarse delay estimate 𝜏̂k,DLL could be

generated by the loop. Even though there might still be an 600
unwanted bias in this coarse tracking result, the CPS has
been followed by the DLL. 400
This code phase tracking is vital to the subspace-based
200
method because the CPS usually causes the ACF peak dis-
persion, but the existing methods (Chang et al. 2018a, b; 0
Hong et al. 2018) are given in the static signal model with- -2 -1 0 1 2
out CPS. In these methods, the local code does not follow Delay(Tc)
the received signal code phase, so the direct use of them to
the signal in (4) is catastrophic. For example, according to Fig. 5  CPS effect on tracking GPS L1 C/A Signal
the relationship between CPS and Doppler frequency offset
(DFO) which is given by Mao and Chen (2009), if a received
GPS L1 C/A signal has a 5 kHz DFO, its code phase shall
shift about 0.325Tc during 100 ms as shown in Fig. 5. The A subspace-based method can be finally extended here
shifted ACF waveform makes them fail to build a stable to distinguish all received rays and estimate their signal
signal subspace and hard to distinguish different rays clearly. parameters in space and time. According to the discussion
With the help of the DLL with the beamformer in above, a joint spatial and temporal subspace-based estima-
Fig. 4, although there is a residual bias in the coarse delay tion is provided as follows.
estimate of the DLL, the local code is generated according
to 𝜏̂k,DLL , and its code phase will surely follow the beam-
formed signal. It means that if we make use of this local Spatial and temporal snapshot construction
code to get the CC signal in (4), all relative code phases
between different rays could be seen to be constant within We define 𝛿k,l as the relative delay with respect to the 𝜏k,DLL
a short period of time. Therefore, the original dynamic for each ray. Because the local code is generated based on
GNSS signal model has been transformed into an approxi- the 𝜏̂k,DLL , 𝜏k,l = 𝜏k,DLL + 𝛿k,l is hence satisfied. The joint
mate static model with the DLL, and the ACF peak disper- spatial and temporal CC snapshot corresponding to the
sion problem caused by the CPS is thought to be solved. mth navigation bit could be written by

13
 109 Page 6 of 17 GPS Solutions (2020) 24:109

⌣ ( ( ) ( ))T and it is helpful to improve the temporal resolution in delay

𝐫 k (m) = 𝐫 T mNc Ts , … , 𝐫 T (m + 1)Nc Ts − Ts
(6) estimation at the cost of computation burden. Fewer rows of
𝛤f are useful to relieve the computation complexity, but
⌣
= 𝐇k 𝛬𝛼,k dk (m) + 𝐰(m)
( ) some temporal resolution has to be paid. It means that there
where 𝐇k = 𝐡k,1 , … , 𝐡k,Lk is a (spatial and temporal
) channel is a trade-off between the complexity and delay estimation
matrix, and its each column 𝐡k,l l(= 1,)… , Lk is correspond- accuracy. Let 𝐅L = 𝛤f 𝐅𝛤t be the frequency and time (selec-)
ing to the signature as 𝐡k,l = 𝐠 𝛿k,l ⊗ 𝐚k,l , in which the tion matrix, and the spectrum response becomes 𝐅L 𝐠 𝛿k,l .
( ) ( ( ) (( ) ))T
𝐠 𝛿k,l = g −𝛿k,l , … , g Nc − 1 Ts − 𝛿k,l is a circle For the simplification of( the)expression, ( )we still define the
convolution ACF and Nc is the sample number of the spread- spectrum response as 𝐠̃ 𝛿k,l = 𝐅L 𝐠 𝛿k,l and normalize it
ing code ‖ ( )‖
( with
( the )Ts sampling ( interval. The noise
))T item is to satisfy ‖ ‖
‖𝐠̃ 𝛿k,l ‖ = 1. The 𝐰(m)
̃ is noise components in
𝐰(m)= 𝐰̃ T mNc Ts , … , 𝐰̃ T (m + 1)Nc Ts − Ts .
⌣
‖ ‖
We are mainly concerned about the case that the mul- 𝐫̃ k (m). Compared with the length of the GPS spreading code,
tipath rays arrive at ULA later than the LOS signal by within the duration of the CC main lobe is quite short. It implies
Tc , so that there is only one main lobe in time above the that the sample selection in (9) could not only mitigate the
noise floor after the cross-correlation. According to the noise impact but be a reduced-rank projection operation as
property of spreading code, the samples away from the main well. It makes the proposed algorithm have a low complexity
lobe are just multi-access noise between the local code and in real-time implementations.
the received signals. These samples have few useful compo- In the view of the subspace signal processing, the
nents, so they could be discarded to suppress the impact of received(signal always) lies in the signal and noise subspace.
noise. Therefore, we here only pick up P samples to cover The ̃𝐡k,l l = 1, … , Lk can be seen as the spatial and spec-
the main lobe for each antenna. The smaller P can reduce the trum steering vectors, and the signal subspace is also built
noise but may not cover the whole CC main lobe; the larger by them. These subspaces can be estimated by a matrix
P can take more multipath rays into consideration but also decomposition
aggravate the estimation accuracy. By experience, P shall be { }
𝐑k = E 𝐫̃ k (m)̃𝐫kH (m)
set large enough to cover 4–6 chips. { }
̃ k 𝛬𝛼,k 𝟏L ×L 𝛬H 𝐇̃ H E dk (m)d∗ (m) + 𝜎 2 𝐈N N (10)
Let 𝛤t be the diagonal selection matrix for samples of =𝐇 k k 𝛼,k k k w sub F

each antenna. If 𝛤t is even, it can be expressed as ̃ k 𝛬𝛼,k 𝟏L ×L 𝛬H 𝐇

=𝐇 ̃H 2
k k 𝛼,k k + 𝜎w 𝐈Nsub NF
( )
𝐈P∕ 2 𝟎P∕ 2×(Nc −P) 𝟎P∕ 2 which is the auto-covariance matrix of 𝐫̃ k (m). The navigation
𝛤t = (7)
𝟎P∕ 2 𝟎P∕ 2×(Nc −P) 𝐈P∕ 2 bits and the noise{ are assumed} to be uncorrelated, and it is
assumed that E dk (m)dk∗ (m) has been normalized to one;
when 𝛤t is odd, it can be shown as 𝟏Lk ×Lk is an all-one matrix. In practical usage, the estimate
( ) of 𝐑k is usually gotten by
𝐈(P+1)∕ 2 𝟎(P+1)∕ 2×(Nc −P) 𝟎(P+1)∕ 2
𝛤t = (8) ∑
M
𝟎(P−1)∕ 2 𝟎(P−1)∕ 2×(Nc −P) 𝐈(P−1)∕ 2 ̂ k= 1
𝐑 𝐫̃ (m)̃𝐫kH (m) (11)
M m=0 k
After sample selection and a Fourier transform, the spatial
and spectrum response of (6) can be given by in which M is the number of CC snapshots. For expression
( )⌣ simplification, 𝐑
̂ k is still represented by 𝐑k.
𝐫̃ k (m) = 𝐅𝛤t ⊗ 𝐈N 𝐫 k (m) If the received signals are not correlated, the signal sub-
(9)
=𝐇̃ k 𝛬𝛼,k dk (m) + 𝐰(m)
̃ space could be spanned by the eigenvectors of 𝐑k , which
( ) are corresponding to its largest eigenvalues. However, the
𝐇̃ k = ̃𝐡k,1 , … , ̃𝐡k,L is the spatial and spectrum channel LOS and multipath rays share the same navigation bit in (9),
k ( )
matrix, and its column is ̃𝐡k,l = 𝐠̃ 𝛿k,l ⊗ 𝐚k,l , where so they are naturally coherent sources when impinging the
( ) ( ( ))
𝐠̃ 𝛿k,l = 𝐅𝛤t 𝐠 𝛿k,l and 𝐅 is a P × P Fourier transform receiver antennas. It means that a rank deficiency problem
will be introduced if the signal subspace is gotten by the
matrix.
matrix decomposition on 𝐑k directly.
The received signal is oversampled to meet the Nyquist
sampling law so that there are only a few dots of main lobe
in frequency having large values. It means that these large Coherent problem and spatial smooth processing
dots may play an important role in rays distinguishing, and
we could only choose them by a row selection matrix 𝛤f . Inspired by the existing spatial smooth techniques (Cad-
More rows of 𝛤f are able to cover a wider frequency span, zow 1988; Chang et al. 2018a, b; Hong et al. 2018; Pillai

13
GPS Solutions (2020) 24:109 Page 7 of 17 109

and Kwon 1989), we would like to extend the forward and where 𝐑s,k denotes the equivalent source covariance matrix
backward spatial smooth (FBSS) here to solve this coherent as
problem. As shown in Fig. 6, we divide the ULA into Q sub-
1
arrays, and each subarray has Nsub =(N − Q + 1) antennas. 𝐑s,k (Q) = 𝐄(Q)𝐄(Q)H (17)
Q
The spatial response of lth ray on the ith subarray can be
denoted by 𝐚(i)
k,l
as and 𝐄 is the equivalent source matrix given by
⎛ 1 ⎞ 2𝜋d 2𝜋d
⎛ 1 ej 𝜆 sin (𝜃k,1 ) ⋯ ej 𝜆 (Q−1) sin (𝜃k,1 ) ⎞
⎜ j 2𝜋d sin (𝜃k,l ) ⎟ 2𝜋d ⎜ ⎟
1 ⎜ e 𝜆
⎟ej 𝜆 (i−1) sin (𝜃k,l )
2𝜋d 2𝜋d
1 ej 𝜆 sin (𝜃k,2 ) ⋯ ej 𝜆 (Q−1) sin (𝜃k,2 ) ⎟
𝐚(i) = √ (12) 𝐄(Q) = 𝚲𝛼,k ⎜ (18)
k,l
N ⎜ ⋮ ⎟ ⎜⋮ ⋮� � ⋮ � �⎟
⎜ j 2𝜋d (Nsub −1) sin (𝜃k,l ) ⎟ ⎜
⎝ e 𝜆
⎠ j 2𝜋d 2𝜋d
sin 𝜃k,Lk ⎟
⎝1 e 𝜆 sin 𝜃 k,L k
j
⋯e 𝜆
(Q−1)
⎠
where i = 1, … , Q. The spatial and spectrum matrix corre- Usually, the LOS signal is coming from the sky, but the mul-
sponding
( to the ith) subarray could (be )expressed by tipath rays are mainly caused by the reflections around the
𝐇k = 𝐡k,1 , … , 𝐡k,L , in which ̃𝐡(i)
̃ (i) ̃ (i) ̃ (i)
k,l
= 𝐠̃ 𝜏k,l ⊗ 𝐚(i)
k,l
. Then, receiver. Hence, we assume that all the received rays have
k

the subarray’s CC snapshot is given by different AOAs here. Then, it could be seen from (18) that, if
the number of subarrays
( )Q is not smaller than Lk , 𝐑s,k will be
𝐫̃ k(i) (m) = 𝐇
̃ (i) 𝚲𝛼,k dk (m) + 𝐰̃ (i) (m)
k (13) full rank, and rank 𝐑s,k = Lk is satisfied. Because the row
of 𝐅L needs to be large enough to cover the spectrum main
Based on (12) and ̃𝐡(i) ,𝐇
̃ (1) and 𝐇
̃ (i) have the following
k,l k k lobe of the CC result, with the help of the multi-antenna, the
relationship:
row number of 𝐇 ̃ (1) is generally much larger than Lk . The
k
̃ (i) = 𝐇
𝐇 ̃ (1) 𝚽i−1 (14) different AOAs of the received rays and their delays both
k k k
lead to that the columns of 𝐇 ̃ (1) cannot be coherent, so 𝐇
k
̃ (1)
k
where 𝚽i−1 is the full column rank by Lk . All the above guarantees that
k is the (i − 1)th power of the k
L × Lk diagonal
the rank of 𝐑k shall be not smaller than Lk if the Q ≥ Lk is
f
matrix
{ ( )} held. It means that, when the power of the received rays is
greater than the noise, the largest eigenvectors of 𝐑k , which
2𝜋d 2𝜋d j 2𝜋d sin 𝜃k,Lk f
𝚽k = diag ej 𝜆 sin (𝜃k,1 ) , ej 𝜆 sin (𝜃k,2 ) , … , e 𝜆
are corresponding to the largest Lk eigenvalues, can be seen
(15) as the estimate of the signal subspace that is very useful in
Similar to the methods by Chang et al. (2018a, 2018b), by detecting and distinguishing the LOS and multipath rays.
using (13), (14), and (15), the forward smooth covariance An additional backward smooth covariance is provided in
matrix could be gotten by (19) to increase the number of detectable rays.

( )H ( )T
1 ∑ ̃ (i)
Q f
f
𝐑k = 𝐇k 𝚲𝛼,k 𝟏Lk ×Lk 𝚲H ̃ (i) + 𝜎 2 𝐈N N
𝐇 𝐑bk = 𝐉 𝐑k 𝐉T
Q i=1 𝛼,k k w sub F
( )∗ ( )T (19)
=𝐉 𝐇 ̃ (1) 𝐑T (Q) 𝐇̃ (1) 𝐉T + 𝜎 2 𝐈N N
( )H
1 ∑ ̃ (1) i−1
Q k s,k k w sub F
= 𝐇k 𝚽k 𝚲𝛼,k 𝟏Lk ×Lk 𝚲H 𝐇̃ (1) 𝚽i−1 + 𝜎 2 𝐈N N
Q i=1 𝛼,k k k w sub F
where 𝐉 denotes an anti-diagonal matrix, and the elements
( )H
̃ (1) 𝐑s,k (Q) 𝐇̃ (1) + 𝜎 2 𝐈N N of the anti-diagonal are all ones. As shown in
=𝐇 k k w sub F
( )∗
(16)
2𝜋d
𝐉 𝐚(1)
k,l
= e−j 𝜆 (Nsub −1) sin (𝜃k,l ) 𝐚(1)
k,l (20)

The Receiving ULA the ULA spatial response 𝐚(1) has the symmetry character.
...
k,l
We define NF as the row (number ) of 𝐅L (and set it to be odd
1 2 3 4 N-2 N-1 N )
here. Let the element of 𝐠̃ 𝛿k,l be gk,l (j) j = 0, ⋯ , NF − 1 .
Subarray 1 In general, the response of RF front
( end) is thought to be flat
Subarray 2 in the band, and the elements of 𝐠̃ 𝜏k,l satisfy the following
equation as
( )
Subarray Q 1 Nsub |gk,l (j)| = ||gk,l NF − j || = |g(j)|, 1 ≤ j < NF (21)
| | | |

Fig. 6  ULA is divided into many subarrays

13
 109 Page 8 of 17 GPS Solutions (2020) 24:109

where g(j) is the jth element of 𝐠̃ (0)

( and
) |g(j)| is its amplitude. Joint spatial and temporal subspace‑based
The phases of the elements of 𝐠̃ 𝛿k,l are mainly decided by estimator
the delay specific ray and follow
Based on the FBSS discussion above, when N (and/Q )are
⎧ � �
2𝜋 𝛿
j Tk,l ,
N +1 set large enough to meet the requirement Q ≥ Lk 2 , a
⎪ ∠g(j) + 0 ≤ j < F2
P � � (22) subspace rotation invariant property could be established
2𝜋 � � 𝛿
∠gk,l (j) = ⎨ s
N +1
⎪ ∠g(j) + j − NF Tk,l , F2 ≤ j < NF by the phase relationship between the first (Q − 1) and
⎩ P s
the last (Q − 1) subarrays. Based on (13), the forward CC
so that they also have symmetry. matrix of them can be denoted by
( ) According to (21) and (22),
the spectrum response 𝐠̃ 𝜏k,l also has a similar symmetry Q−1 { }
property as 1 ∑
E 𝐫̃ k(i+1) (m)̃𝐫k(i) (m)H
f
𝐑YX,k =
( ) ( ) Q − 1 i=1
(29)
𝐉̃𝐠∗ 𝛿k,l = 𝐠̃ 𝛿k,l (23) ( )H
̃ (1) 𝚽 𝐑s,k (Q − 1) 𝐇
≈𝐇 ̃ (1)
k k k

Then, the symmetry of ̃𝐡(1) can further be derived by

k,l
where the noise impact has been neglected for the simplifica-
( )∗ ( )
𝐉 ̃𝐡(1)
2𝜋d
= e−j 𝜆 (Nsub −1) sin (𝜃k,l ) 𝐠̃ k 𝜏k,l ⊗ 𝐚(1) tion of expression; it will also be ignored in the following
k,l k,l
(24) equations.( Define
) the backward CC matr ix as
2𝜋d T
= e−j 𝜆
(Nsub −1) sin (𝜃k,l )̃𝐡(1) 𝐑bYX,k = 𝐉 𝐑XY,k
f
𝐉T . With the use of (24), 𝐑bYX,k can be
k,l
expressed by
Taking (24) into (19), 𝐑bk can be rewritten as
( )T
( )H f
𝐑bYX,k = 𝐉 𝐑YX,k 𝐉T
𝐑bk = ̃ (1) 𝚽1−Nsub 𝐑T (Q)𝚽Nsub −1
𝐇 𝐇̃ (1) + 𝜎 2 𝐈N N (25)
k k s,k k k w sub F ( )∗ ( )∗ (30)
≈𝐉 𝐇 ̃ (1) 𝐑T (Q − 1)𝚽 𝐇 ̃ (1) 𝐉T
Let define the FBSS matrix 𝐑FBSS as k s,k k k
k
f Taking (24) into (30), 𝐑bYX could be rewritten by
𝐑k + 𝐑bk
𝐑FBSS
k
= ( )H
2
( )H (26) ̃ (1) 𝚽1−Nsub 𝐑T (Q − 1)𝚽Nsub 𝐇
𝐑bYX,k = 𝐇 ̃ (1) (31)
k k s,k k k
̃ (1) 𝐑FBSS (Q) 𝐇
=𝐇 ̃ (1) + 𝜎 2 𝐈N N
k s,k k w sub F
Similar to (26)–(28), the FBSS CC matrix between the first
and its equivalent source covariance matrix is updated by and the last (Q − 1) subarrays could then be given by
1 f
𝐑FBSS 𝐒(Q)𝐒(Q)H (27) 𝐑YX,k + 𝐑bYX,k
s,k (Q) = 2Q 𝐑FBSS =
YX,k 2
( )H (32)
where the FBSS equivalent source matrix is = ̃ (1) 𝚽 𝐑FBSS (Q
𝐇 − 1) 𝐇̃ (1)
k k s,k k
( )
𝐒(Q) = 𝐄(Q) 𝚽1−N sub
𝐄(Q)∗
k
(28) Correspondingly, the forward and backward auto-covariance
( )
= 𝐄(Q) 𝚽Q−N 𝐄(Q)∗ matrices are, respectively, defined by
k
Q−1 { }
From a comparison between (18) and (28), it could be found 1 ∑
E 𝐫̃ k(i) (m)̃𝐫k(i) (m)H
f
𝐑XX,k =
that the Vandermonde structure of equivalent source matrix Q − 1 i=1
(33)
is expanded by the phases of different rays after FBSS, ( )H
̃ (1) 𝐑s,k (Q − 1) 𝐇 ̃ (1)
and it( also
/ )relieves the requirement of Q from Q ≥ Lk to
≈𝐇 k k
Q ≥ Lk 2 , which is beneficial to the usage in the receiver
with small antenna array. Similar to the case of 𝐑k , the (Q−1 )
f
∑ { (i+1) } ∗
decomposition of 𝐑FBSS
k
provides the largest eigenvalues and 𝐑bXX,k =
1
𝐉 (i+1)
E 𝐫̃ k (m)̃𝐫k (m) H
𝐉T
corresponding eigenvectors, and they could be used for the Q−1 i=1 (34)
signal subspace estimation and multipath detection. ( )H
(1) −N
̃ 𝚽 sub 𝐑∗ (Q − 1)𝚽Nsub 𝐇 ̃ (1)
≈𝐇 k k s,k k k

The FBSS auto-covariance matrix is then presented by

13
GPS Solutions (2020) 24:109 Page 9 of 17 109

f ( ( ))
𝐑XX,k + 𝐑bXX,k 𝜆
𝜃̂k,l = asin angle 𝛷̂ k (l, l) (41)
𝐑FBSS
XX,k = 2𝜋d
2
( )H (35)
̃ (1) FBSS ̃ (1)
= 𝐇k 𝐑s,k (Q − 1) 𝐇 (1)
k ̃̂𝐡 = 𝐔̂ s 𝐓̂ s (∶, l) (42)
k,l
It could be found that the signal components in (32) and (35)
where asin(⋅) is an arcsine function, angle(⋅) is an operation
have a DOAM-like structure (Yin et al. 1989; Chang et al.
to get the phase of a complex number, 𝛷̂ k (l, l) denotes the lth
2018a, b; Hong et al. 2018), and they also build a subspace
diagonal element of 𝛷̂ k , and 𝐓̂ s (∶, l) is the lth column of 𝐓̂ s .
rotation invariance about the AOA. This property could be
Meanwhile, the estimate of 𝐚̂ (1) is also given. Based on the
used to acquire the estimates of AOA and delay jointly, but k,l
structure of 𝐠̃ (𝛿) and ̃𝐡(1) which have been given in (9), (21),
the signal subspace shall be estimated firstly. Based on the k,l
and (22), the vector in (42) can be seen as a joint spatial and
discussion above, the largest eigenvectors of 𝐑FBSS can be
XX,k temporal response as
seen as the estimate of the signal subspace. There is much
research (Cozzens and Sousa 1994; Wax and Ziskind 1989; (1) ( )
̃̂𝐡 = ej𝜂k,l 𝐠̃ 𝛿̂ ⊗ 𝐚̂ (1) (43)
Xin et al. 2007; Xu et al. 1994) focusing on the source num- k,l k,l k,l

ber estimation, and their good estimation performances have

been widely admitted. Therefore, we can assume that the where 𝜂k,l is an unknown phase shift introduced by the matrix
number of multipath rays has been gotten precisely from the decomposition. 𝐠̃ (𝛿) is the ACF waveform corresponding to
analysis of 𝐑FBSS , and just pick up the eigenvectors, which the delay 𝛿 , so we could estimate the delay for the lth ray
XX,k
are corresponding to the largest Lk eigenvalues, to make up without extra phase searching by
the estimate of the signal subspace. ‖( )H (1) ‖
Let 𝐔s and 𝐔o be the signal and noise subspace, respec- 𝛿̂k,l = arg max ‖
‖ 𝐮(𝛿) ⊗ 𝐚̂ (1) ̃̂𝐡 ‖
k,l ‖ (44)
𝛿 ‖ k,l
‖
tively. Then, 𝐑FBSS
XX,k
can be decomposed by
( )( H ) where 𝐮(𝛿) is a phase vector, and its elements follow
( ) 𝚲s 𝐔s
𝐑FBSS = 𝐔 𝐔 (36) �� ��
XX,k s o
𝚲o 𝐔Ho ⎧ 2𝜋n𝛿 NF +1
⎪ exp j ∠g(n) + , 0≤n<
un (𝛿) = ⎨ �� PTs �� 2
where 𝚲s and 𝚲o are diagonal eigenvalue matrices corre- 2𝜋 (n−NF )𝛿 NF +1
⎪ exp j ∠g(n) + , ≤ n < NF
sponding to 𝐔s and 𝐔o , respectively. When the LOS and ⎩ PTs 2

multipath rays here are thought to be visible and not be bur- (45)
ied by the noise, because the steering matrix 𝐇 ̃ (1) is in the
k
signal subspace, the relationship between 𝐇k and 𝐔s could
̃ (1) (The delay) of any ray here is around the DLL result within
−Tc , Tc , and it could be gotten efficiently by using some
be given by classic low-complexity numerical searching methods. For
̃ (1) = 𝐔s 𝐓s the sake of saving space, we ignore the searching details,
𝐇 (37)
k but suppose the 𝛿̂k,l can be achieved effectively.
where 𝐓s is a Lk × Lk full-rank matrix. Taking (37) into (35),
𝐑FBSS (Q − 1) can be expressed by Modified delay estimation generation
s,k

𝐑FBSS −1
s,k (Q − 1) = 𝐓s
𝚲s 𝐓−H
s (38) It should be mentioned that the estimate of the delay from
(45) is just a relative result with respect to the tracking output
Then, taking (37) and (38) into (32), we have 𝜏̂k,DLL. All spatial and temporal parameters of the distinguish-
able ray have been estimated; then, the LOS ray’s delay can be
𝐑FBSS ≈ 𝐔s 𝐓s 𝚽k 𝐓−1 𝚲s 𝐔H (39)
YX,k s s selected by its property that it is always the first one arriving
the receiver array as
Since the estimates of 𝐔s and 𝚲s have been gotten from (36),
a projected formula could be given by { }( )
𝛿̂k,1 = min 𝛿̂k,l l = 1, ⋯ , Lk (46)
𝐔̂ H 𝐑FBSS 𝐔̂ ̂ −1 ≈ 𝐓̂ s 𝛷̂ k 𝐓̂ −1
s YX,k s 𝛬s s (40)
It also represents the tracking bias compared with the actual
It can be found from (40) that its eigenvalues and eigenvec- LOS arrival time, so the modified delay could be gotten by
tors are made up with 𝛷̂ k and 𝐓̂ s , respectively. And they the sum of the 𝜏̂k,DLL and 𝛿̂k,1 as
could provide the estimate of AOA and the correspond-
ing spatial and spectrum steering vector jointly for the lth 𝜏̂k,1 =𝜏̂k,DLL + 𝛿̂k,1 (47)
received ray of the kth satellite GPS signal by

13
109 Page 10 of 17 GPS Solutions (2020) 24:109

It is an absolute delay estimate that has been protected from Table 1  Simulation profile description
the multipath bias by using the proposed subspace-based Parameter Value
method. Meanwhile, the corresponding AOA estimate 𝜃̂k,1
has also been gotten, and it can be used to update the beam- Bandwidth 4 MHz
forming weights to track the LOS ray in space. Sampling rate 5 MSPS
Antenna number, N 4
Computation complexity analysis DLL bandwidth 10 Hz
PLL bandwidth 20 Hz
Except for the conventional tracking part in Fig. 4, the DLL and PLL order 2nd order
computation burden of the subspace-based processing can Coherent integration 20 ms
be divided into three parts: the spatial smoothed matrices Non-coherent integration No
construction, the signal subspace estimation, and the delay E–L spacing for DLL 1 Chip
searching. Since the searching for 𝛿̂k,l in (44) can be carried The CC snapshots number, M 5
out by many numeral calculation methods efficiently, such The temporal length of the CC snapshots, P 32
as binary searching, the primary complexity is mainly taken The row number of FL, NF 9, 31
by the first two parts. The direct eigenvalue The FBSS factor, Q 2
( decomposition) to
estimate 𝐔s in (36) has a (complexity by o )(N − Q) NF , and
3 3

another complexity by o (N − Q)2 NF2 MQ has to be paid to

construct the smoothed matrices 𝐑FBSS YX,k
and 𝐑FBSS
XX,k
. It seems
to be a heavy computation burden, but the complexities can 80
( )
be greatly reduced to o (N − Q)NF MQ by an efficient mul- 60
tistage Wiener decomposition (Goldstein et al. 1998), which
helps our proposed method be applied in practice easily. 40
AOA(Degree)

20

Simulation and performance results 0 Multipath Rays

-20
In this section, we provide three simulation experiments to
evaluate the proposed method performance without and with -40
LOS Ray
multipath rays impinging. The method is firstly tested by a -60
numeral simulation to show its distinguishable rays num- 0 0.2 0.4 0.6 0.8
ber capability. Afterward, an MEE experiment is arranged Delay(Tc)
to show its maximum delay bias caused by the different
RDLM. An additional experiment is set at last to evaluate Fig. 7  Proposed estimator for an LOS ray and three multipath rays
the RMSE performance of the proposed method in different
kinds of situations.
A four-antenna ULA with a half-wavelength antenna number of 𝐅L is set to NF = 9. In order to examine the sepa-
spacing is used here for GPS L1 C/A signal receiving. The ration ability of the proposed method, it is assumed that
receiver takes the zero-IF RF structure for convenience. In there are one LOS ray and three multipath rays imping-
all cases, the RF front end of each antenna is set to have ing the receiving( array, ) (and their temporal
)( and spatial
)
a 4 MHz bandwidth, and the complex signals at 5 MSPS parameters
( are
) 0T c , 0◦
, 0.25T c , 40 ◦
, 0.5T c , −50◦ , and
with 10 s duration are generated by MATLAB software. The 0.75Tc , 60◦ . The corresponding multipath-to-direction
bandwidth of the DLL and phase lock loop (PLL) is set to amplitude ratio (MDR) of amplitude is set to be 0.707,
be 10 Hz and 20 Hz, respectively. The simulation profile is 0.5, and 0.5. When the chip signal-to-noise power ratio
summarized in Table 1. (SNR) of the LOS ray is − 15 dB, the results from the
proposed estimator for all rays are shown in Fig. 7. It can
Number of incident multipath rays be seen from the figure that the proposed estimator is able
to distinguish different rays by their spatial and temporal
The parameters N = 4 and Q = 2 both have been set, so the parameters effectively. The little shift in time is caused by
proposed joint spatial and temporal subspace-based esti- the bias from DLL, and it will eventually be compensated
mator could be used to detect and distinguish four rays at by (47). Besides, the proposed estimator only needs Q ≥ 2
maximum according to the discussion in the section of the to solve the coherent problem with the four rays inci-
coherent problem and spatial smooth processing. The row dences, but the subspace methods by Chang et al. (2018a,

13
GPS Solutions (2020) 24:109 Page 11 of 17 109

b) require Q ≥ 4 . It means that the estimator can work with curves of the conventional DLL in Figs. 8 and 9 show that,
a small-size antenna array, which saves much device cost. after beamforming, the residual components of multipath
still cause quite large MEE. The nEML and MEDLL also
Multipath error envelopes experiments mitigate more multipath than the beamformed DLL, but they
cannot suppress them completely. Their performances even
MEE is the most typical way to evaluate the performance of get worse when the RDLM is small. In contrast, the pro-
the multipath mitigation ability. Typically, an LOS ray and posed method achieves the best MEE performances and it
a multipath ray are arranged to be present, and the tracking seems nearly no multipath effects in these figures. The rea-
errors are acquired for RDLMs up to 1.5 chip at maximum. son for it is that the proposed method is able to distinguish
The phase between rays is set to be in-phase and out-of- different incident rays by their AOAs. The delay of LOS ray
phase to obtain the upper and lower MEE (Bhuiyan and can also be gotten from the corresponding delay estimates;
Lohan 2010). MDR is set to be 0.707, and the ideal noise- then, the tracking bias is then canceled by (47). Besides,
free assumption is made here only to evaluate the multipath as shown in Fig. 9, even though the received two rays are
mitigation abilities of different methods. In order to simplify closed in space, the proposed method can still distinguish
the experiments, we fix the AOA of the LOS ray coming at them and modify the tracking result effectively.
0° but set the AOA of the multipath ray at 40°, 10°, and 30°
in Figs. 8, 9, and 10, respectively. The NF = 9 and Q = 2 are
still set here for the proposed method. In order to make a fair Delay RMSE evaluation
comparison, we make beamforming to the LOS direction
before the use of conventional DLL, nEML, and MEDLL. The proposed method is then evaluated by RMSE in the
The early–late separation of nEML is 0.1 chip. The MEE environments without and with multipath. The simulation
results with different multipath AOAs are shown in Figs. 8, conditions are still set according to Table 1. The Q = 2 is
9, and 10. held, and different ­NF are set to compare the frequency com-
As shown in Figs. 8, 9, and 10, with the aid of ULA, even ponents effects in the proposed method. The main lobe is
the conventional DLL could achieve a lower bias than the covered when NF = 9 , and the whole spectrum is almost
DLL with a single antenna. This improvement benefits from taken into account when NF = 31. The chip SNR of the
that the beamformer before tracking reduces the multipath LOS ray is changed from − 30 to 0 dB. Similar to the MEE
effects from other directions. Especially when the spatial experiments, the proposed method is also compared with
responses of two rays are orthogonal, as shown in Fig. 10, the conventional DLL, nEML, and MEDLL, which are still
the multipath is filtered completely in space. However, the working with a beamformer to the direction of the LOS ray.

Fig. 8  MEE comparison 0.4 0.02

when LOS and multipath rays
are from 0° and 40°, respec-
tively, with different RDLMs:
RDLM = 0–1.5Tc (left panel), 0.2 0.01
RDLM = 0–0.3Tc (right panel)
RMSE(Tc)

RMSE(Tc)

0 0

-0.2 -0.01

-0.4 -0.02
0 0.5 1 1.5 0 0.1 0.2 0.3
RDLM(Chip) RDLM(Chip)

N=1, Conventional DLL N=4, Proposed Method

N=4, Proposed Method N=4, Beamformed nEML
N=4, Beamformed DLL N=4, Beamformed MEDLL
N=4, Beamformed nEML
N=4, Beamformed MEDLL

13
109 Page 12 of 17 GPS Solutions (2020) 24:109

Fig. 9  MEE comparison 0.4 0.06

when LOS and multipath rays
are from 0° and 10°, respec-
tively, with different RDLMs:
RDLM = 0–1.5Tc (left panel), 0.2 0.04
RDLM = 0–0.3Tc (right panel)

RMSE(Tc)

RMSE(Tc)
0 0.02

-0.2 0

-0.4 -0.02
0 0.5 1 1.5 0 0.1 0.2 0.3
RDLM(Chip) RDLM(Chip)

N=1, Conventional DLL N=4, Proposed Method

N=4, Proposed Method N=4, Beamformed nEML
N=4, Beamformed DLL N=4, Beamformed MEDLL
N=4, Beamformed nEML
N=4, Beamformed MEDLL

Fig. 10  MEE comparison 0.4 0.001

when LOS and multipath rays
are from 0° and 30°, respec-
tively, with different RDLMs:
RDLM = 0–1.5Tc (left panel), 0.2 0.0005
RDLM = 0–0.3Tc (right panel)
RMSE(Tc)

RMSE(Tc)

0 0

-0.2 -0.0005

-0.4 -0.001
0 0.5 1 1.5 0 0.1 0.2 0.3
RDLM(Chip) RDLM(Chip)

N=1, Conventional DLL N=4, Proposed Method

N=4, Proposed Method N=4, Beamformed DLL
N=4, Beamformed DLL N=4, Beamformed nEML
N=4, Beamformed nEML N=4, Beamformed MEDLL
N=4, Beamformed MEDLL

In LOS environment, the delay estimation RMSEs of easily. The CRBs of the nEML and the DLL with a single
different methods are evaluated in Fig. 11. The proposed antenna and four antennas are both plotted as references
method makes use of the AOA estimate of the LOS ray to (Van Dierendonck et al. 1992).
make beamforming before the DLL part in Fig. 4. The spa- As shown in Fig. 11, compared with the single antenna
tial beamforming only strengthens the received signal power case, with the help of the beamformer, nEML, MEDLL, and
in the environment without multipath, so the Cramér–Rao the proposed method, all could achieve obvious performance
bound (CRB) of the four-antenna DLL could be derived improvements. Meanwhile, MEDLL and the proposed

13
GPS Solutions (2020) 24:109 Page 13 of 17 109

10-1 10-1

10-2 10-2
RMSE(Tc)

RMSE(Tc)
10-3 10-3

10-4 10-4
-30 -25 -20 -15 -10 -5 0 -30 -25 -20 -15 -10 -5 0
Chip SNR(dB) Chip SNR(dB)
Conventional DLL Theory (N=1) Beamformed DLL Theory (N=4,LOS)
Beamformed DLL Theory (N=4) Beamformed DLL(N=4)
Beamformed DLL(N=4) Beamformed nEML Theory (N=4,LOS)
Conventional nEML Theory (N=1) Beamformed nEML (N=4)
Beamformed nEML Theory (N=4) Beamformed MEDLL(N=4)
Beamformed nEML (N=4) Proposed Method (N=4,N F =9)
Beamformed MEDLL (N=4)
Proposed Method (N=4,N F =31)
Proposed Method(N=4,NF =9)
Proposed Method(N=4,NF =31)
Fig. 12  Delay RMSE comparison when LOS and multipath rays are
from 0° and 30°, respectively, with RDLM = 0.5Tc
Fig. 11  Delay estimation RMSE in LOS environment
techniques could mitigate some multipath effects by spa-
method with NF = 9 both have nearly the same RMSE per- tial beamforming. In Figs. 13 and 14, nEML, MEDLL, and
formances with the beamformed conventional DLL. The the proposed method all have better multipath mitigation
proposed method with NF = 31 has about 4 dB better per- performances than beamformed conventional DLL, and the
formance than the method with NF = 9. The reason for this proposed method has the best tracking accuracies. Although
improvement is that the NF = 31 case covers more compo- nEML can reduce the tracking error by its small early–late
nents in the frequency domain so that the higher temporal separation, it still suffers a bad estimation result because
resolution can be achieved. The nEML has the best accurate tracking bias is not eliminated completely. MEDLL has a
estimates due to its much narrow discriminator space. It can better performance than nEML in Figs. 13 and 14. It can-
be implied that, if the DLL in Fig. 4 is replaced by nEML or cels multipath rays by the iterative successive interference
other peak tracking techniques (Bhuiyan and Lohan 2010; cancelation technique, but it always has to keep with an
Chen et al. 2013), the proposed method RMSE performance undesired iteration residual.
can be further improved. In Fig. 13, benefiting from the separation ability in space
We next evaluate different methods in the environment and time, the proposed method is able to achieve nearly the
with multipath incidence. The LOS ray’s AOA is fixed at same RMSE performance with the case when no multipath
0° for convenience. It is arranged that there is one multipath is incident. Compared with the LOS case in Fig. 11, there
ray impinging the receiver ULA, and its MDR is set to be is only about 0.2 dB performance loss. The larger NF still
0.707 as a strong reflection. The AOA of the multipath ray is provides a higher resolution in time and offers better RMSE
changed at 30°, 40°, and 10° in Figs. 12, 13, and 14, respec- performance.
tively, and the delay estimation RMSEs of different mul- The space difference between LOS and multipath rays
tipath mitigation methods are also provided. becomes smaller in Fig. 14. The proposed method with
In Fig. 12, because the spatial steering vectors of LOS NF = 9 has a better performance than MEDLL when the
and multipath rays are orthogonal to each other, all methods RDLM is 0.1Tc, but its RMSE performance is worse than
have nearly the same RMSE performances with the only MEDLL at some SNR points when the RDLM is 0.5Tc.
LOS case. It proves that the application of the multi-antenna The reason for it is that the MEDLL is sensitive to the

13
109 Page 14 of 17 GPS Solutions (2020) 24:109

Fig. 13  Delay RMSE com-

10-1 10-1
parison when LOS and
multipath rays are from 0°
and 40°, respectively, with
RDLM = 0.1Tc (left panel) and
RDLM = 0.5Tc (right panel)
10-2 10-2

RMSE(Tc)

RMSE(Tc)
10-3 10-3

10-4 10-4
-30 -20 -10 0 -30 -20 -10 0
Chip SNR(dB) Chip SNR(dB)
Beamformed DLL Theory(N=4,LOS) Proposed Method(N=4,NF =9)
Beamformed DLL(N=4) Proposed Method(N=4,NF =31)
Proposed Method with nulling
Beamformed nEML(N=4) (N=4,N F =31)
Beamformed MEDLL(N=4) Beamformed MEDLL(N=1)

Fig. 14  Delay RMSE com-

100 100
parison when LOS and
multipath rays are from 0°
and 10°, respectively, with
RDLM = 0.1Tc (left panel) and 10-1 10-1
RDLM = 0.5Tc (right panel)
RMSE(Tc)

RMSE(Tc)

10-2 10-2

10-3 10-3

10-4 10-4
-30 -20 -10 0 -30 -20 -10 0
Chip SNR(dB) Chip SNR(dB)
Beamformed DLL Theory(N=4,LOS) Proposed Method(N=4,NF =9)
Beamformed DLL(N=4) Proposed Method(N=4,NF =31)
Proposed Method with nulling
Beamformed nEML(N=4) (N=4,N F =31)
Beamformed MEDLL(N=4) Beamformed MEDLL(N=1)

short relative delay multipath effect. Owing to the higher encounters some performance degradations by different
temporal resolution, the proposed method with NF = 31 levels. The reason for it is that, when the distance between
can still achieve the best performances among these AOAs and delays are getting small, the joint spatial and
methods in Fig. 14. Compared with the results in Fig. 13, temporal steering vectors of incident rays are heavily cor-
no matter which NF is, the proposed method in Fig. 14 related. It usually makes the FBSS matrices (32) and (35)

13
GPS Solutions (2020) 24:109 Page 15 of 17 109

ill-conditioned, and there are some small eigenvalues, References

which are corresponding to the signal subspace, of 𝐑FBSSXX,k
after decomposition. These small eigenvalues are sensi- Aboutanios E, Hassanien A, Amin MG, Zoubir AM (2017) Fast itera-
tive to the noise so that the estimated signal subspace gets tive interpolated beamforming for accurate single-snapshot DOA
estimation. IEEE Geosci Remote Sens Lett 14(4):574–578
more perturbed. Therefore, the proposed method must Amin MG, Sun W (2005) A novel interference suppression scheme for
suffer some degradation if the AOAs and delays are both global navigation satellite systems using antenna array. IEEE J Sel
too closed. This degradation gets rapid when the signal Areas Commun 23(5):999–1012
strength is weak as in Fig. 14 (left panel). It can also be Bhuiyan MZH, Lohan ES (2010) Advanced multipath mitigation tech-
niques for satellite-based positioning applications. Int J Navig
found in Fig. 14 that the proposed method with NF = 31 Observ 2010:412393
is less sensitive to the closed parameter distance than the Cadzow JA (1988) A high-resolution direction-of-arrival algorithm for
case with NF = 9 , which benefits from its higher resolu- narrow-band coherent and incoherent sources. IEEE Trans Acoust
tion in time. Figures 13 and 14 also evaluate the proposed Speech Signal Process 36(7):965–979
Caizzone S, Buchner G, Circiu MS, Cuntz M, Elmarissi W, Perez Mar-
method with the steering beam nulling to multipath cases. cos E (2019) A miniaturized multiband antenna array for robust
The figures show that compared with the steering-only navigation in aerial applications. Sensors (Basel) 19(10):2258
beam, the beam with nulling to multipath cases could Chang N, Hong X, Wang W, Wang Z (2018a) Joint delay and angle
achieve a better estimation accuracy when the spatial sep- estimation for GNSS multipath signals based on spatial and fre-
quential smoothing. In: 2018 14th IEEE international conference
aration between two incident rays is large. The beam has on signal processing (ICSP), Aug. 12–16, pp 203–207
an obvious distortion when the LOS and multipath rays Chang N, Hong X, Wang W, Wang Z (2018b) Subspace based joint
are too closed in space, so the RMSE performances suffer delay and direction of arrival estimation for GNSS multipath
degradation by 1–3 dB in Fig. 14. signals. In: China satellite navigation conference (CSNC) 2018
proceedings. Harbin, May 23–25, pp 189–199
As shown in Figs. 13 and 14, the nEML and MEDLL Chen X, Dovis F, Peng SL, Morton Y (2013) Comparative studies
curves both get flat when SNR is high. It means that the of GPS multipath mitigation methods performance. IEEE Trans
biases caused by multipath still exist after processing. Aerosp Electron Syst 49(3):1555–1568
On the contrary, the RMSEs of the proposed method Chenu-Tournier M, Chevalier P, Barbot J (2000) A parametric spatio-
temporal channel estimation technique for FDD UMTS uplink.
are always decreasing with the SNR increasing. It can In: Proceedings of the 2000 IEEE sensor array and multichan-
be implied that the proposed method is able to eliminate nel signal processing workshop. SAM 2000 (Cat. No.00EX410),
delay bias effectively in any cases. March 17, pp 12–16
Cozzens JH, Sousa MJ (1994) Source enumeration in a correlated sig-
nal environment. IEEE Trans Signal Process 42(2):304–317
Cuntz M, Konovaltsev A, Meurer M (2016) Concepts, development,
and validation of multiantenna GNSS receivers for resilient navi-
Conclusion gation. Proc IEEE 104(6):1288–1301
Fenton PC, Jones J (2005) The theory and performance of NovAtel
Inc.’s Vision Correlator. In: Proceedings of ION GNSS 2005,
We propose a new subspace-based code loop design for a Institute of Navigation, Long Beach, CA, USA, September 13–16,
multi-antenna GPS receiver to reduce the multipath effect. pp 2178–2186
With the help of the DLL tracking part, the receiver could Fernandez-Prades C, Lo Presti L, Falletti E (2011) Satellite radiolo-
follow the code phase with a bias in a short time so that calization from GPS to GNSS and beyond: novel technologies and
applications for civil mass market. Proc IEEE 99(11):1882–1904
the signal model can be seen to be static at discriminator Garin L, Rousseau J M (1997) Enhanced strobe correlator multipath
and the subspace-based estimator can be used to distin- rejection for code and carrier. In: Proceedings of ION GPS 1997,
guish received rays. By using an FBSS processing and the Institute of Navigation, Kansas City, MO, USA, September 16–19,
spatial property of the ULA, the signal subspace is esti- pp 559–568
Goldstein JS, Reed IS, Scharf LL (1998) A multistage representation
mated without rank deficiency and a rotation invariance of the Wiener filter based on orthogonal projections. IEEE Trans
property between adjacent subarrays is also established. Inf Theory 44(7):2943–2959
After decomposing the CC matrix, we estimate the AOAs Hong X, Chang N, Wang W, Yin Q (2018) Subspace-based joint DOA,
from its eigenvalues and get the relative delays from the delay and DFO estimation for GNSS multipath signals. In: Pro-
ceedings of ION GNSS 2018, Institute of Navigation, Miami,
corresponding eigenvectors at the same time. The param- Florida, USA, September 24–28, pp 3788–3801
eters of the LOS ray are easily picked up by its first arrival Hsu L-T, Jan S-S, Groves PD, Kubo N (2014) Multipath mitigation
property, and a modified tracking result is finally gotten by and NLOS detection using vector tracking in urban environments.
combining the subspace and the DLL estimates together. GPS Solut 19(2):249–262
Irsigler M (2010) Characterization of multipath phase rates in different
Simulations results prove that the proposed method is able multipath environments. GPS Solut 14(4):305–317
to not only achieve better tracking accuracy than the con- IS-GPS-200J (2018) NAVSTAR GPS space segment/navigation user
ventional DLL in LOS environment but also have better segment interfaces. Global positioning system directorate systems
MEE and RMSE performance than nEML and MEDLL in engineering and integration. 25–April. https:​ //www.gps.gov/techn​
ical/icwg/IS-GPS-200J.pdf
the environment with multipath.

13
109 Page 16 of 17 GPS Solutions (2020) 24:109

Jia Q, Wu R, Wang W, Lu D, Wang L, Li J (2016) Multipath Yin QY, Newcomb RW, Zou LH (1989) Estimating 2-D angles of
interference mitigation in GNSS via WRELAX. GPS Solut arrival via two parallel linear arrays. In: International confer-
21(2):487–498 ence on acoustics, speech, and signal processing, May 23–26,
Khosravi F, Moghadas H, Mousavi P (2015) A GNSS antenna pp 2803–2806
with a polarization selective surface for the mitigation of low- Zhu C, Fan X (2015) A novel method to extend coherent integra-
angle multipath interference. IEEE Trans Antennas Propag tion for weak GPS signal acquisition. IEEE Commun Lett
63(12):5287–5295 19(8):1343–1346
Kumar R, Petovello MG (2017) 3D building model-assisted snapshot Zoltowski MD, Gecan AS (1995) Advanced adaptive null steering
positioning algorithm. GPS Solut 21(4):1923–1935 concepts for GPS. In: Proceedings of MILCOM ‘95, Nov. 5–8,
Leclere J, Botteron C, Farine PA (2017) High sensitivity acquisition pp 1214–1218
of GNSS signals with secondary code on FPGAs. IEEE Aerosp
Electron Syst Mag 32(8):46–63 Publisher’s Note Springer Nature remains neutral with regard to
Li S, Lv J, Chang J, Tian X, Tong K (2015) Optimization selection jurisdictional claims in published maps and institutional affiliations.
of the HRC correlator spacing for different navigation signals.
In: China satellite navigation conference (CSNC) 2015 Proceed-
ings. Xi’an, May 13–15, pp 3–13
Lohan ES, Lakhzouri A, Renfors A (2006) Feedforward delay estima- Xi Hong received a B.S. degree
tors in adverse multipath propagation for Galileo and modern- from the School of Electronics
ized GPS signals. EURASIP J Appl Sig Process 2006(1):1–19 and Information Engineering,
Lu D, Liu H, Wu R (2013) Global positioning system anti-jamming Xi’an Jiaotong University, Xi’an,
algorithm based on period repetitive CLEAN. IET Radar Sonar China, in 2012, where he is cur-
Navig 7(2):164–169 rently pursuing the Ph.D. degree.
Mahmud MS, Qaisar SU, Benson C (2016) Weak GPS signal detec- His main research interests
tion in the presence of strong signals with varying relative Dop- include array signal processing,
pler and long integration gain. In: Proceedings of IEEE/ION signal processing in communica-
PLANS 2016, April 11–14, pp 1015–1020 tion systems, multipath mitiga-
Mao WL, Chen AB (2009) New code delay compensation algorithm tion in navigation, and GNSS
for weak GPS signal acquisition. AEU Int J Electron Commun physical layer interference
63(8):665–677 detection.
Peng X, Petovello MG (2015) Measuring GNSS multipath distribu-
tions in urban canyon environments. IEEE Trans Instrum Meas
64(2):366–377
Pillai SU, Kwon BH (1989) Forward backward spatial smoothing Wenjie Wang received the B.S.,
techniques for coherent signal identification. IEEE Trans Acoust M.S., and Ph.D. degrees in infor-
Speech Signal Process 37(1):8–15 mation and communication engi-
Pui CY, Trinkle M (2013) GPS bistatic radar using phased-array neering from Xi’an Jiaotong
technique for aircraft detection. In: 2013 international confer- University, Xi’an, China, in
ence on radar, Adelaide, SA, pp 274–279 1993, 1998, and 2001, respec-
Qaisar SU, Benson CR (2017) Processing cost of Doppler search in tively, where he is currently a
GNSS signal acquisition: measuring Doppler shift in navigation Professor. His main research
satellite signals. IEEE Signal Process Mag 34(5):53–58 interests include information
Townsend BR, Fenton PC, Van Dierendonck KJ, Van Nee DJR theory, broadband wireless com-
(1995) Performance evaluation of the multipath estimating munications, signal processing
delay lock loop. Navigation 42(3):502–514 in communication systems, and
Van Dierendonck AJ, Fenton PAT, Ford TOM (1992) Theory and array signal processing.
performance of narrow correlator spacing in a GPS receiver.
Navigation 39(3):265–283
Van Nee DJR (1992) The multipath estimating delay lock loop. In:
IEEE second international symposium on spread spectrum tech-
Ning Chang received B.S. and
niques and applications, Nov. 29–Dec. 2, pp 39–42
M.S. in communication engi-
Wax M, Ziskind I (1989) Detection of the number of coherent sig-
neering from Information Engi-
nals by the MDL principle. IEEE Trans Acoust Speech Signal
neering University, China, 2011,
Process 37(8):1190–1196
and Xidian University, China,
Xin JM, Zheng NN, Sano A (2007) Simple and efficient nonparamet-
2014, respectively, and is cur-
ric method for estimating the number of signals without eigen-
rently pursuing a Ph.D. degree in
decomposition. IEEE Trans Signal Process 55(4):1405–1420
information and communication
Xu B, Hsu L-T (2019) Open-source MATLAB code for GPS vector
engineering at Xi’an Jiaotong
tracking on a software-defined receiver. GPS Solut 23(2):46
University. Her research interests
Xu GH, Roy RH, Kailath T (1994) Detection of number of sources
include compressed sensing,
via exploitation of centro-symmetry property. IEEE Trans Sig-
array signal processing, param-
nal Process 42(1):102–112
eter estimation, and multipath
Xu Z, Trinkle M and Gray DA (2015) Weak interference direction
mitigation for GNSS signals.
of arrival estimation in the GPS L1 frequency band. In: 2015
IEEE international conference on acoustics, speech and signal
processing (ICASSP), Brisbane, QLD, pp 2649–2653

13
GPS Solutions (2020) 24:109 Page 17 of 17 109

Qinye Yin received the B.S., M.S.,

and Ph.D. degrees in communi-
cation and electronic systems
from Xi’an Jiaotong University,
Xi’an, China, in 1982, 1985, and
1989, respectively. Since 1989,
he has been a Faculty Member
with Xi’an Jiaotong University,
where he is currently a Professor
with the Department of Informa-
tion and Communications
Engineering.

13