A Subspace-based Spatial and Temporal Multipath

Mitigation for Multi-antenna GPS Receiver

Xi Hong, Tian Gan, Menghan Lin, Ning Chang, Wenjie Wang, Qinye Yin
School of Information and Communications Engineering
Xian Jiaotong University
Xi’an, China
[email protected]

Abstract—In this paper, a novel subspace-based spatial and errors in position, and it now is one of the major positioning
temporal mitigation method is proposed for GPS signal tracking error sources that has been highly concerned.
by a multi-antenna receiver in multipath environment. Since the
tracking result from the delay-lock-loop has been polluted by the In order to reduce the effects reflection rays in GNSS signal
multipath incidence, we only use it to follow the code phase shift tracking, there is much research attempting to mitigate the
but design a subspace-based estimator to distinguish different multipath effects from different aspects. The [2] mitigates the
incident rays by their spatial and temporal signatures. The multipath effects by using a narrow early-late discriminator
carrier phase is firstly extended into the signatures because of the (nEML). This method aims to reduce the tracking error by
signal BPSK modulation. A subspace rotation invariance in space
limiting the discriminator error within a small area, but it does
and DOAM-like correlation matrices are both established by the
property of the receiving array, and a forward and backward
not cancel any multipath component, which makes the bias
space smooth is also taken here to solve the coherent problem by error still existed. The multipath estimating delay lock loop
multipath. The angle-of-arrivals and signatures are estimated (MEDLL) sets up several code and phase tracking loops for
from the eigenvalues and eigenvectors from the cross-correlation different rays and distinguishes them by using a maximum-like
matrix, the signatures also provide the relative delay and carrier (ML) iteration search, where the effects from other rays are
phase estimates with respect to the tracking loop. The line-of- successively subtracted by using their estimates [3], [4]. The
sight ray’s estimates are picked up by its first arrival character, MEDLL could suppress the medium and large relative delay
and a modified delay tracking output without multipath bias is multipath rays, but it is sensitive to the rays with short relative
finally achieved by combining tracking loop and the subspace
delays. Other multi-correlator techniques [5]-[8], such as
estimator results together. Simulation results prove that
compared with the existing methods, such as the narrow early-
“Double Delta” and “Strobe Correlator”, make use of the
minus-late and the multipath estimating delay lock loop, the slopes or peaks of GNSS auto-correlation functions (ACF) to
proposed method has the best multipath mitigation in the delay mitigate multipath further but still cannot achieve the bias-free
estimation. performance. [9] and [10] use the vector delay lock loop to
reduce the part of positioning error in urban areas, and the 3D
Keywords—Multi-antenna receiver; Multipath mitigation; DLL; building model is also taken into account in [11]. These pose-
Subspace-based estimator; forwar and backward space smooth; processing methods are too complicated to be used in real-time.
Joint delay and AOA estimation
Antenna technique is another way to decrease the multipath
I. INTRODUCTION effects in space. With the aid of antenna design, the rays from
low elevation directions can be suppressed [12]. For the multi-
Global Navigation Satellite Systems (GNSS) have been antenna GNSS receiver, adaptive beamformers can be used to
wildly used in many civil and military applications, such as strengthen the signal in the specific direction [13], [14]. In
auto-driving and precision agriculture services. Within decades order to take the signal spatial and temporal characters jointly,
developments, USA, Russia, European Union and China all set some subspace-based methods [15]-[17] are also extended to
up their own GNSS systems as GPS, GLONASS, Galileo and estimate different rays’ parameters. These subspace-based
Beidou respectively [1]. Japan and India also launch their own methods have a higher delay estimation accuracy than other
local area positioning enhancement systems [1]. Since the iterated ML searching methods in the static signal model due to
GNSS are increasingly involved in urban areas, the high-rise their spatial separation capacities, but they are hardly used
buildings bring more challenges in positioning and navigation directly in practices for the reason that the code phase shift
services. The buildings arounds the GNSS receivers usually not (CPS) is not taken into consideration.
only block the satellite signal propagation from the sky, but
also generate multipath reflections, which is harmful to signal In this paper, we propose a novel subspace-based spatial
tracking and pseudo-range measurement. According to the and temporal multipath mitigation method, and modify the
discussion in [2]-[4], the multipath may cause tens of meter code tracking result to compensate the multipath bias. This

978-1-7281-0244-3/20/$31.00 ©2020 IEEE 1256

Authorized licensed use limited to: University of Exeter. Downloaded on June 20,2020 at 08:08:40 UTC from IEEE Xplore. Restrictions apply.
method is cooperating with the delay lock loop (DLL) so that K LK
j 2 f k  t  k ,l   j k ,l
the CPS is followed with the unwanted bias but the subspace- x  t     k ,l a k ,l ck  t   k ,l  d k  t   k ,l  e
k 1 l =1 (1)
based method can be applied without the code phase dispersion.
Because of the BPSK modulation of the GPS signal, we are  w t 
able to firstly construct a phase extended spatial and temporal
where w (t ) is the noise from the array; there are K satellite
snapshots based on the tracking results. By taking the receiving
signals impinging the receiving ULA, and each of them has Lk
uniform linear array structure into account, a subspace rotation
rays; the ck (t ) , d k (t ) and f k are spreading code, navigation
property can be built, and the DOAM-like correlation matrices
bit and CFO of the kth satellite signal, respectively; the  k ,l ,
are also generated. Additional forward and backward spatial
 k ,l and k ,l represent the delay, amplitude and phase of the
smooth (FBSS) is then given on the correlation matrices to
lth ray of the kth satellite signal respectively, and its
solve the coherent problem by the same navigation bits of the
normalized spatial steering vector is given by
multipath rays. After the eigenvalue decomposition on the
cross-correlation matrix, the estimates of the angle-of-arrivals T
1  j 2 d sin k ,l  j
2 d
 N 1 sin  k ,l  
(AOA) and the phased extended signatures of the incident rays a k ,l  1, e , , e   (2)
can be acquired by the eigenvalues and corresponding N 
eigenvectors, which are utilized to provide the relative delays
and carrier phases estimation with respect to the DLL and PLL where N is the antenna number of the ULA and d is the
tracking results. The line-of-sight (LOS) ray’s parameters are distance between the adjacent antennas;  is the wavelength
picked up by its first arrival character, and we can finally of the signal, and  k ,l is the AOA of the lth ray of the kth
modify the DLL tracking result and eliminate its bias. Besides, satellite signal. For ease of the discussion, l  1 here indicates
the application of multi-antenna strengthens the received signal the LOS ray’s vectors and parameters. Because the LOS ray is
power, which also improves the estimation accuracy. The always the first one arriving the receiver, the relationship as
multipath mitigation performances of the proposed method are
 k ,1   k ,2     k , Lk is surely met, and  k ,1 is also the largest
evaluated in simulated multipath environments by maximum one among the rays for its direct incidence.
error envelop (MEE) and the root mean squared errors (RMSE). After complex sampling, the baseband signal in (1) is then
via acquisition and tracking digitally. The spreading codes of
II. RECEIVING SIGNAL MODEL GPS signal have a large spreading gain so that different
satellite signals can be seem as unrelated sources [19]. Here the
Multi-antenna GNSS receiver has abilities to enhance the
kth satellite signal is the one of interest. With the help of the
signal strength, cancel the jammer impacts, and distinguish
phase-lock-loop (PLL) and DLL, the cross-correlation (CC)
different incident rays in space. For ease of discussion, a GNSS
results can be approximately shown by
receiver with a uniform linear array (ULA) is considered in this
paper as shown in Fig. 1. The carrier frequency offset (CFO) is  g  nTs   k ,1  
mainly caused by the doppler effect oscillator instability. Both  
of them usually change slowly so the CFO can be thought to be  g  nTs   k ,2  
constant within tens of seconds receiving [18]. Besides, the rk  n   A k Λ , k Dk   d k  nTs   k ,1   w
  n  (3)
  
satellites are far away from the receiver and reflections on the
 
ground so that the CFO differences between incident rays that  
g nTs   k , Lk 
 
from the specific satellite are assumed to be same during the
receiving. Therefore, the received signals in baseband can be where Ts is the sampling interleave; A k  (a k ,1 , a k ,2 , , a k , Lk ) is
expressed by the spatial steering matrix; g (t ) is the ACF of the GPS signal;
Λ ,k and Dk are the diagonal matrices to represent the
amplitude  k ,l and the carrier phase e j 2 fk k ,l  jk ,l respectively;
 (n) is the noise item after CC. Because the multipath rays
w
Satellite k are usually generated by the reflections around the receiver, the
relative delays between different rays are quite small in
contrast to the duration of the d k (t ) so that all rays in (3) are
dsin( θ k ) thought to share the same navigation bit with the LOS ray
approximately.

III. CPS ANALYSIS AND CODE LOOP DESIGN

θk θk The doppler effect not only makes the carrier frequency
d d shifted, but generates the CPS effect on code. The relationship
1 2 ... N-1 N between the doppler frequency offset (DFO) and CPS has been
given in [20]. The PLL is able to track the received signal
Fig. 1 GPS receiver with ULA structure carrier effectively, but the DFO difference between two rays
may still make their code phases shift not at the same rate. This

1257

Authorized licensed use limited to: University of Exeter. Downloaded on June 20,2020 at 08:08:40 UTC from IEEE Xplore. Restrictions apply.
 ˆk ,DLL
Spatial v kH x  t  GNSS
Beamforming
Temporal ˆk ,1
ck ˆk ,DLL 
Tracking
Part
Part
vk
Local Code
Generation

x t  Cross- The CC ˆk ,1

Baseband Subspace-
rk  m 
Correlation Matrices FBSS
Signals from (CC) for Construction
R XX,k based Joint
ULA Spatial And
Each &
Antenna Spatial R FBSS
YX,k
Temporal
Output Smooth
Estimator
ˆk ,1
Fig. 2 Subspace-based joint spatial and temporal loop design

effect can be found in long-time observations or in urban areas Therefore, we can give a subspace-based code loop design as
[21], and it makes the signal model not static. Therefore, the in Fig. 2.
CPS effects on incident rays should be discussed further before
There are two main functional parts in this loop design: a
we design the subspace-based method and the corresponding
DLL with a beamformer and a subspace-based spatial and
code tracking loop.
temporal estimator. The former part is used to enhance the
signal from the LOS ray’s direction before the DLL tracks the
A. CPS on Incident Rays
code phase of the received signal; the latter one is designed to
Because of the weak strength of GPS signal on the ground, distinguish the incident rays in space and time. In Fig. 2, the
the multipath rays are often generated by the reflections around beamformer can improve the specific satellite signal’s signal to
the receiver, and the relative delays between the LOS and interference and noise ratio (SINR). The residual components
multipath rays (RDLM) are always not very large. For GPS L1 from multipath could still cause bias error in tracking, but the
C/A signal, its ACF waveform determines that the obvious CPS of the received signal has been followed.
distortion occurs when the RDLM is less than 2Tc, in which Tc
is the duration of the chip. It has been given by [2]-[4] that the It is vital to subspace-based method due to that the CPS
conventional DLL has to put up with the significant bias error usually cause the ACF peak dispersion. The existing method
when the RDLM is within on chip, so we also mainly focus the [15]-[17] are proposed without the consideration of CPS, and
multipath problem with the RDLM less than Tc. use the fixed local code to calculate the CC snapshots. If they
are taken in practices, the shifting ACF peak will greatly
Although the RDLM often cause unwanted bias in tracking, degrade their temporal resolution and make them fail to
it may change slowly due to the little DFO difference between distinguish the rays in time. As shown in Fig. 2, with the help
incident rays [22]. The measurements in [21] also show that the of the DLL, the local code is now generated on the DLL results
DFO difference between the LOS and multipath rays are ˆk ,DLL , and the relative delays of incident rays with respect to
always within 10Hz when the multipath rays strength are above the DLL shall be held in a short time. It transfers the original
25 dB·Hz. It means that the RDLM changes shall be smaller received dynamic signal model into a static one, and a
than 6.6  104 Tc during 100 ms tracking. The observation of subspace-based method can be finally extended here to
[21] is acquired by a moving platform in the urban areas. distinguish the incident rays in space and time.
Therefore, it can be implied that the RDLM may even remain
unchanged if the receiver is in low dynamic, and it is assumed
IV. PHASE EXTENDED SUBSPACE-BASED ESIMATOR
in this paper that the RDLM hold constant within a short time.
This assumption indicates that the receiving signal model can The code loop in Fig. 2 maintains the temporal relationship
be approximately treated as a static model if the duration of the of the incident rays with respect to the DLL. It implies that the
processed samples is not very long, and it also guarantees that bias in tracking can be compensated if the LOS relative delay
the subspace-based method could be extended with DLL in the can be estimated. The nearly static signal model keeps different
following subsection. rays’ signatures constant during a short-time processing, then
the subspace-based method can be introduced to estimate each
B. Subspace-based Code Loop Design rays’ spatial and temporal parameters as follows.
Based on the discussion above, although the waveform of
A. Spatial and Temporal Snapshot Construction
ACF has been affected by the multipath incidence, it can be
thought to be constant within a short time. It implies that the We define the  k ,l as the relative delay of the lth with
multipath bias in tracking not changed, and meanwhile the respect to the coarse code tracking result  k ,DLL . Because the
temporal relationship between different rays are also held. local code is yielded on the DLL,  k ,l = k ,DLL   k ,l is satisfied.

1258

Authorized licensed use limited to: University of Exeter. Downloaded on June 20,2020 at 08:08:40 UTC from IEEE Xplore. Restrictions apply.
The samples from (3), which are corresponding to the mth Because the typical modulation of the GNSS signal is

navigation bit, are collected to construct a snapshot as always BPSK, the navigation bits satisfy d k (m)=d k (m) and
this property can be used to further construct a phase extended

   m  1 N T  Ts  
T
rk  m   rkT  mN cTs  , , rkT c s snapshot. We arrange the rk (m) and its conjugation as the
 (4) following expression
 H k Λ , k Dk 1Lk 1 d k  m   w  m 
 r  m    H  D 
where Nc is the sample number of the spreading code with the zk  m   k*     * *  Λ , k 1Lk 1 d k  m 
k k
(7)
sampling interleave Ts; H k  (h1 , ,h Lk ) is the joint spatial  Jrk  m    JH k Dk 
and temporal channel matrix, and its each column
where J denotes an NL  NL anti-diagonal eye matrix. As
hl (l  1, , Lk ) is the signature as hl =g k ,l  a k ,l , in which
 shown by (2), the spatial steering vector has the following
g k ,l  (g ( k ,l ), , g (( N c  1)Ts   k ,l ))T ; w (m) is the noise
relationship with its conjugation as
item; 1Lk 1 is the all ones vector.
2  N 1 d
j sin  k ,l 
We mainly concern the case than the RDLM is less than J N ak ,l =e  a k ,l (8)
one chip so that there is only one main temporal lobe above the
noise floor after the CC in (4). According to the property of the where J N is an N  N anti-diagonal eye matrix. The vector
spreading code, the samples away from the CC main lobe can g k ,l is based on the standard spectrum of ACF at the delay  k ,l .
be seem as the multi-access noise between the received signal Because the standard ACF of GPS C/A signal both has the
and the local code. These useless samples can be discarded for symmetry in time and frequency, the element of g k ,l maintains
their few desired signal components, and only P samples that the relationship as
covers the main lobe of g k ,l in time are selected here for each 2 i
 k ,l
antenna output in (4). The selection for P is arbitrary by g k ,l  i   g 0  i  e
j
L
 g k*,l  i  (9)
experience: the larger P can take more multipath rays into
account but too much noise impact is also introduced; the in which g 0 (i ) and g k ,l (i ) denote the ith elements of the
smaller P shall relief sort of the computation burden but may standard ACF after the Γ L FP projection and g k ,l respectively.
not cover the response of the main lobe which also aggravates A conjugated symmetry of g k ,l can be found by (9). For sake
the estimation accuracy. Here, as for the GPS L1 C/A signal, P of ease discussion, we just set L to be odd, and arrange the row
shall be set to cover 4~6 chips. sequence of Γ L FP to make the g k ,l as the structure as
Let FP denote the operation that the main lobe samples of
the g k ,l are chosen then a P  P discrete Fourier Transform is   L  1   j 2L  L21  k ,l 
 L  1    g 0  
 
also applied to acquire the spectrum for each antenna. Because  e

g
 k ,l      2  
the received signal is often oversampled to meet the Nyquist   2   

Sampling Theorem, there is a frequency main lobe for GPS L1     
2
C/A signal for its square window shaping, and it means that the     j  k ,l 
 g k ,l  1   g 0 1 e L
spectrum of FP g k ,l can also be selected for saving the 
g k ,l   g k ,l  0     g 0  0   (10)
complexity. We define the Γ L as the selection matrix to    
choose L samples of the frequency lobe of the FP g k ,l . More  g k ,l 1   2
j  k ,l 
rows of Γ L can cover a wider frequency span but the higher    g 0 1 e L

 
computation burden has to be paid; Fewer rows of Γ L could    
save the complexity at the cost of temporal resolution. It is   L 1    
 g k ,l      L  1  j 2L  L21  k ,l 
clearly a tradeoff between the complexity and the delay   2    g 0  
e
estimation accuracy. The operation Γ L FP transforms hl into a   2  
joint spatial and spectrum form as The above equation makes J L g *k ,l  g k ,l satisfied where J L is
h l  g k ,l  a k ,l (5) an L  L anti-diagonal eye matrix. The (8)-(10) confirms that
JH * could be rewritten by
k

where  denotes the Kronecker product; g k ,l  Γ L FP g k ,l is the  * =H

 Φ1 N
JH (11)
spectrum response, and it is normalized to satisfy g k ,l  1 for k k k

ease of the discussion. Make the above operations on rk (m) where Φ1k N is the (1-N) th power of the diagonal AOA matrix
and the processed snapshot can be expressed by
  j 2 d sin k ,1  j
2 d
sin  k , Lk  
rk  m    Γ L FP  I N  rk  m  Φ k  diag e  , , e   (12)
 Λ D 1 d  m
(6)  
H k  , k k Lk 1 k
Taking (11) into (7), the extended snapshot can be written by
in which H  k  (h 1 , ,h Lk ) is the channel matrix in space and
frequency, and the noise item in (6) has been ignored for H D 
z k  m    k *k 1 N  Λ , k 1Lk 1 d k  m  (13)
simplification of the expression. 
 H k Dk Φ k 

1259

Authorized licensed use limited to: University of Exeter. Downloaded on June 20,2020 at 08:08:40 UTC from IEEE Xplore. Restrictions apply.
It can be seen from (13) that the original channel matrix has
been expanded by Dk and Φ1k N , and the phase extended  1 
signature has the formula as  2 d 
sin  k ,l 

1  e
j
  j 2 d i 1 sin k ,l 
 h l e j 2 fk k ,l  jk ,l  ak ,l
i
 e 
vl   2 d
 (14) N   (16)
   j 2 fk k ,l  jk ,l  j  1 N  sin k ,l    2 d 
 hl e  e 
j  Nsub 1 sin  k ,l 

 
Vk  ( v1 , , v Lk ) is set to denote the extended channel matrix. 2 d
j  i 1 sin  k ,l 
In the view of subspace, the received signals always lie in  ak1,l e 

the signal and noise subspaces, and the signature v l is always Then the channel matrix H  (ki ) satisfies H
 (ki )  H
 (1) i 1
k Φ k . This
in the signal subspace. These subspaces are able to provide the relationship can be applied to the extended snapshot z k (m) by
estimate of v l and the incident rays’ parameters can also be
gotten from them. In practice, the signal subspace is estimated  Γ i  
z k   m    s  zk  m
i
by the eigenvectors, which are corresponding to some largest 
eigenvalues of the auto-correlation matrix of the received  Γ s 
i 
(17)
signals. However, the impinging rays in (13) have the same  Vk1 Φik1 Λ , k 1Lk 1 d k  m 
navigation bits, and they are all equivalent to the Lk coherent
sources. This phenomenon always yields a significant rank where Vk(1) is the extended channel matrix denoted as
insufficiency problem when the direct matrix decomposition is
taken on the E{z k (m)z kH (m)} , and its largest eigenvectors now  1 D
H 
1
cannot span the whole signal subspace, which is not able to Vk =  k1 k  (18)
H
   * 1 N 
distinguish all rays clearly. Therefore, an additional smooth  k Dk Φ k 
process is needed to solve this coherent problem.
Similar to [15]-[17], the forward spatial smooth auto-
covariance matrix can be constructed as
B. Forward and Backward Spatial Smooth
Inspired by [15]-[17], the Vandermonde structure of ULA 1 q i 
R FZZ  q    z k  m  z ki   m 
H
can be used to enlarge the uncorrelated level between incident q i 1
rays from different directions. This process is also able to be
1 q 1 i 1
 
H
applied on the extended snapshot in (13) because of the same
array type that we use in this paper.
  Vk Φk Λ ,k 1Lk Lk Λ ,k Vk1Φik1
q i 1
(19)

 
H
 Vk  R s , k  q  Vk 
1 1
As shown in Fig. 3, we divided the array into Q subarrays,
and each of them has N sub  ( N  Q  1) antennas. Denote the
Γ (si ) as the row selection matrix for the ith subarray, and the The R s , k  q  above is the equivalent source auto-covariance
CC snapshot of the subarray can be given by matrix as

Γs  rk  m   Γs  H

 D Λ 1 d  m 1
i i
R s,k  q   S qS q
H
k k  , k Lk 1 k
(15) (20)
 i  D Λ 1 d  m  q
H k k  , k Lk 1 k

 (ki )  (h 1(i ) , ,h (Lik) ) is the channel matrix of the ith and E  q  is the equivalent source given by
where H
subarray, and h l(i )  g k ,l  a(ki,)l , in which a (ki,)l is the spatial  j
2 d
sin  k ,1  j
2 d
 q 1 sin  k ,1  
steering vector of the subarray. The ULA can provide the 1 e   e  
following relationship between different subarrays as  j
2 d
sin  k ,2  j
2 d
 q 1 sin  k ,2 

 
 e  
S  q   Λ , k  1 e  (21)
    
The Receiving ULA  2 d
sin  k , Lk 
2 d
 q 1 sin  k , Lk 

... 1 e j
 e
j
 
1 2 3 4 N-2 N-1 N  
In (19)-(20), the smooth factor q is not greater than Q. It can be
Subarray 1 seen from the (20) and (21) that, if the incident rays have
Subarray 2 different AOAs, the R s , k (q) shall be full of rank when q is not
smaller than the number of the rays Lk. Because the row of g k ,l
should be set large enough to cover the main lobe of cross-
Subarray Q 1 Nsub correlation in frequency domain, with the help of the multi-
antenna, the row of the Vk(1) shall be significantly greater than
Fig. 3 ULA is divided into many sub-arrays Lk then q  Lk is able to guarantee the rank of R FZZ (q) to Lk.
This derivation proves that the coherent problem can be solved

1260

Authorized licensed use limited to: University of Exeter. Downloaded on June 20,2020 at 08:08:40 UTC from IEEE Xplore. Restrictions apply.
through the spatial smooth if the number of subarrays is large The similar averaging is also taken for R FYZ (q) and R BYZ (q) as
enough.
R FYZ  q   R BYZ  q 
The above smooth can also be used on the cross-correlation YZ  q  
R FBSS
matrix between adjacent subarrays. We define this matrix as 2 (29)
1 1
 
H
Vk Φ k E ZZ  q  E ZZ  q  Vk1
H
q =
1
 zki 1  m  z ki   m  2q
H
R FYZ (q)  (22)
q i 1
The E ZZ (Q  1) guarantees that the matrices R FBSS
ZZ (Q  1) and
Taking (17)-(20) into (22), the matrix can be written by YZ (Q  1) share the same rank to Lk when Q is not smaller
R FBSS
than ( Lk  1) , which implies that the FBSS in (25) and (29) are
 
H
R FYZ (q)  Vk1 Φ k R s , k  q  Vk1 (23) able to estimate the complete signal subspace to distinguish the
LOS and multipath rays.
The requirement to make R FYZ (q) full of rank is same with the
case of R FZZ (q) . C. Joint Spatial and Temporal Subspace-based Estimator
An addition backward spatial smooth can be taken to make From (25) and (29), the DOAM-like [15]-[17] structure has
ZZ (Q  1) and R YZ (Q  1) , and a subspace
been built by R FBSS FBSS
the correlation matrices more stable to noise and snapshots
number. Denote J E as the 2 N sub L  2 N sub L anti-diagonal eye rotation invariance property on AOA is also constructed by
matrix. Unlike the above forward process, we choose the q their relationship. Denote the Us and Uo as the signal and noise
subarrays from the right side in Fig. 3, and their auto- ZZ (Q  1) respectively, then the matrix can be
subspaces of R FBSS
covariance matrix can be given by decomposed by
T Σ   U sH 
ZZ (Q  1)   U s Uo   s
1 q  R FBSS   (30)
R  q   J E   z kQ  i 1  m  z kQ  i 1  m 
B H
ZZ  JE  Σ o   U oH 
q
 i 1 
where Σ s and Σ o are the diagonal eigenvalue matrices
 V  
q
1 H
  Vk1 Φ qk i Λ , k 1Lk  Lk Λ , k Φik q k
1
(24) corresponding to Us and Uo respectively. When the LOS and
q i 1
multipath rays both are visual to the receiver and not buried by
 
H
 Vk  Φ kq 1R s , k  q  Φ1k q Vk  the noise, the estimate of Us is able to be acquired by the
1 1

eigenvectors, which are corresponding to the largest Lk

After averaging the (19) and (24) by (25), the equivalent source ZZ (Q  1) . The received signal signatures
eigenvalues of R FBSS
matrix is also expanded as (26). always belong to the signal subspace so that the relationship
between Us and Vk(1) can be given by
R FZZ  q   R BZZ  q 
ZZ  q  
R FBSS Vk  =U s Ts
1
2 (31)
(25)
1 1
 
H
= Vk E ZZ  q  E ZZ  q  Vk1
H where Ts is the full rank linear matrix. Taking (31) into (25),
2q s , k (Q  1) is provided by
another expression of R FBSS

  s , k  Q  1  Ts Σ s Ts
1 H
E ZZ  q   S  q  Φ qk 1S  q 

(26) R FBSS (32)

Take (31) and (32) into (29), then we have

From (26), the column rank of E(q ) is decided by the subarray
YZ  Q  1  U s Ts Φ k Ts Σ s U s
1
number q. It means that the requirement to make the matrix R FBSS H
(33)
s , k ( q )  E ZZ ( q )E ZZ ( q )
R FBSS H
2q is full of rank by Lk is still
q  Lk . When q is equal to (Q  1) , the equivalent source is We estimate the signal subspace Uˆ s and the eigenvalue matrix
expressed by ˆΣ s from R FBSS
ZZ (Q  1) , then a projected formula of (33) is
derived by

E ZZ  Q  1  S  Q  1 ΦQk  2 S  Q  1

 (27)
YZ  Q  1 U s Σ s  Ts Φ k Ts
ˆ sH R FBSS
U ˆ ˆ 1 ˆ ˆ ˆ 1 (34)
F
The backward spatial smooth can also be taken for R (q) . YZ This equation reveals that the estimates of Φˆ k and T
ˆ s could be
Since the adjacent subarrays have been taken into account in gotten by the eigenvalue decomposition of (34), and the AOA
(23), the backward expression of the cross-correlation of them and the corresponding phase extended signature for the lth
can be derived by incident ray therefore are estimated respectively by
R BYZ (q)  J E  R YZ (q)  J E
T
 
 
F
ˆk ,l  asin  ˆ l, l  
angle Φ (35)
(28) 
 2 d
k

V 
H
 Vk Φ1 q 1
k R 
s,k qΦ q2
k
1
k
ˆ Tˆ :, l 
vˆ l1 =U (36)
s s

1261

Authorized licensed use limited to: University of Exeter. Downloaded on June 20,2020 at 08:08:40 UTC from IEEE Xplore. Restrictions apply.
where asin() is an arcsine function; angle() is the function to The ˆk ,1 and ˆk ,1 are also picked up at the same time for their
get the phase of a complex number; Φ ˆ k (l , l ) is the lth diagonal correspondence. Since aˆ k ,1 has been built by ˆk ,1 , and it can be
element of Φ ˆ k , and T
ˆ s (:, l ) is the lth column of the T ˆ s . The as the beamformer before the DLL in Fig. 2.
spatial steering vector estimates aˆ k ,l and aˆ k ,l both are built by
(1)
The  k ,1 also represents the negative response of the bias
ˆk ,l . According to the structure in (18), the carrier phase of the
by multipath, then the DLL tracking results can be modified by
lth ray can be gotten by the cross-correlation between the upper
the summation as
part vˆ l(1) and lower part vˆ l(1)  of (36), which are corresponding
to the first and the last N sub L elements of vˆ l(1) . For sake of ˆk ,1  ˆk ,DLL  ˆk ,1 (42)
simplification expression, we define  k ,l  2 f k k ,l   k ,l , and
it is estimated by This is an absolute bias-free result, and therefore the multipath
bias has been finally compensated.
1   1  j 2 d  N 1 sin ˆk ,l   H 1  
ˆ  angle   vˆ l  e  
 vˆ l  (37)
k ,l
2    V. SIMULATION AND PERFORMANCE ANALYSIS
 
We arrange two kinds of simulations in this section: the
This phase result can be used for the precise positioning or the MEE simulation is carried out for test the multipath mitigation
receiver RF channel calibration in multipath-free environment. ability of the proposed subspace-based method, and the RMSE
Because of the ambiguous weakness of the phase,  k ,l seems simulation is also set to evaluate the accuracy of the proposed
too little to directly cancel the multipath bias in pseudo-range method in the environments without and with multipath input.
measurement, but it can involve in other existing carrier phase
smooth based multipath mitigation methods. As for the relative A ULA with four elements is equipped for the receiver, and
delay of the specific ray, since the estimates of aˆ (1) ˆ
k , l ,  k , l and
the adjacent antenna distance is set to half-wavelength of the
ˆ
carrier phase have been gotten, the  k ,l is then decided by a GPS L1 C/A signal. The receiver takes zero-IF structure for
searching as each antenna, and the channel bandwidth of every RF front-end
is set to 4 MHz. All RF channels here are assumed with the
 hˆ l e k ,l  perfectly calibration, and there is no non-ideal mismatch in the
j ˆ

   
H
1
ˆ  arg max vˆ l (38) following simulations. The received signals are generated at
k ,l
  ˆ  j ˆk ,l  j 2 d 1 N  sin ˆk ,l   5 MSPS orthogonal sampling rate for 10 s by MATLAB

 hl e

 software. It is assumed that the acquisition for the specific
where the estimate of the signature hl is defined by satellite signal has been done. The subspace-based method is
sensitive to the high order perturbation of noise, so the
hˆ l   g 0  u     aˆ (1)
k ,l (39) snapshots that belong to the same navigation bit are summed
for the stability in calculation. There are several source number
in which  denote the Hadamard product; g 0 is the standard estimation methods having been given in [23]-[25], and they
ACF after the Γ L FP projection; u( ) is the complex cosine could be directly used for R FBSS ZZ (Q  1) in its eigenvalue
vector relying on  as decomposition. Therefore, the number of the incident rays is
T
also assumed known in the simulations. The phase extended
  j 2  L 1  j
2
j
2 j   
2  L 1 
snapshots are collected for every 200 ms, which means the
u     e L  2  , , e L , 1, e L , , e L  2   (40) number of the snapshots is M  10 . The simulation profile is
 
  summarized in TABLE. 1.
The mainly concerned relative delay is within (Tc, Tc) in this
paper so that the searching in (38) can be realized by some A. Multipath Error Envelopes Experiments
classic low complexity numerical methods. For sake of saving MEE is a common way to evaluate multipath mitigation
space, the detail of the searching is ignored here but the ability. Typically, an LOS ray and a multipath ray are arranged
estimate of  k ,l is thought to have been gotten. Because of the to be present with in-phase and out-of-phase to record the
correspondence between the eigenvalues and eigenvectors, the upper and lower MEE versus different RDLM. The multipath-
estimates ˆk ,l , ˆk ,l and ˆk ,l above are paired automatically, to-direction amplitude ratio (MDR) is set to 0.707 to simulate
and this is very useful to distinguish different rays in parameter the case with high-reflection incidence. The noise-free setting
domain. is made here to evaluate the mitigation abilities of different
methods fairly without the noise impacts. For simplification of
D. Modified Delay Estimation Generation the experiments, the LOS direction is fixed at 0°, but the AOA
It should be mentioned that the results from the proposed of multipath ray is changed at 40° and 10°. The L  9 is set for
subspace-based estimator are just the relative delay estimates the proposed method, and other simulation parameters are
with respect to ˆk ,DLL . That the LOS ray is always the first one configured according to the TABLE. 1. In order to make a fair
impinging the ULA, so it can be picked up by comparison, we make beamforming to the LOS direction
before the use of conventional DLL, nEML and MEDLL. The
 
ˆk ,1  min ˆk ,l  l  1, , Lk  (41) discriminator distance of nEML is 0.1 chip. The MEE results
with different multipath AOAs are shown in Fig. 4 and Fig. 5.

1262

Authorized licensed use limited to: University of Exeter. Downloaded on June 20,2020 at 08:08:40 UTC from IEEE Xplore. Restrictions apply.
 The nEML and MEDLL both mitigate more multipath impacts
TABLE 1 SIMULATION PROFILE DESCRIPTION than the conventional DLL, but they cannot eliminate the bias
completely. Besides, their mitigation performances suffer an
Parameter Value obvious degradation when the RDLM is smaller than 0.1 chip.
Bandwidth 4 MHz In contrast, the proposed method is able to obtain the best MEE
mitigation performance than others. By taking the spatial and
Sampling Rate 5 MSPS temporal characters jointly, this method is able to distinguish
Antenna Number, N 4 different rays and cancel the tracking bias errors at all test
RDLM settings. The curves of the proposed method in Fig. 4
DLL Bandwidth 10 Hz and Fig. 5 look like no multipath errors even when RDLM is
small and when spatial steering vectors are too correlated.
PLL Bandwidth 20 Hz

DLL and PLL Order 2nd Order B. Delay Estimation RMSE Evaluation
Coherent Integration 20 ms We next evaluate the proposed method accuracy in the
cases with and without multipath incidence. All the simulation
Non-coherent Integration 1 Block settings are same with the previous subsection, but L  31 is
E-L Spacing for DLL 1 Chip added here to find out how much delay estimation can be
influenced by the width of used spectrum of g k ,l . The chip
The Snapshots Number, M 10 signal-to-noise ratio (SNR) is increased from -30dB to 0dB at
The Row Number of FP, P 32
3dB step, and ten times Monte Carlo experiments are executed
for every SNR pinot. Similar to the MEE experiments, the
The Row Number of ΓL, L 9, 31 beamformer to the LOS direction is still kept before the
conventional DLL, nEML and MEDLL.
The Subarray Number, Q 3
In LOS environment, without the interference from the
multipath, the beamforming here only strengthens the received
signal, so the CRB of the four-antenna DLL could be derived
With the help of the ULA, as shown in Fig. 4 and Fig. 5, easily from the single antenna case in [2]. Both the CRBs of
even the conventional DLL achieves smaller bias errors than the nEML and the DLL with single antenna and four antennas
the single antenna GPS receiver, which benefits from that some are plotted as references. For different methods, the RMSEs of
multipath component has been suppressed by the beamformer. their delay estimation are compared in Fig. 6.
RMSE(Tc)

RMSE(Tc)

Fig. 4 MEE comparison when LOS and multipath rays are from 0° and 40° respectively with
different RDLM: RDLM=0~1.5 Tc (left panel), RDLM=0~0.3 Tc (right panel)

1263

Authorized licensed use limited to: University of Exeter. Downloaded on June 20,2020 at 08:08:40 UTC from IEEE Xplore. Restrictions apply.
 RMSE(Tc)

RMSE(Tc)
Fig. 5 MEE comparison when LOS and multipath rays are from 0° and 10° respectively with
different RDLMs: RDLM=0~1.5 Tc (left panel), RDLM=0~0.3 Tc (right panel)

As shown in Fig. 6, compared with the single antenna case, better improvement than the method with L  9 . The reason
the beamformed conventional DLL, nEML and the proposed for it is that the case with L  31 covers more components in
method all achieve obvious performance improvements. The the frequency domain so that the higher temporal resolution is
MEDLL and the proposed method with L  9 have nearly the achieved. The nEML has the best accuracy due to its much
same RMSE performances with the beamformed conventional narrow discriminator space. It implies that, if the DLL in Fig. 2
DLL, but the proposed method with L  31 has about a 2.5dB is replaced by nEML or other peak tracking techniques [26],
the proposed method accuracy might be further improved.
10-1 The RMSE performances of different methods are then
evaluated in the environment with multipath incidence. The
LOS direction is fixed at 0° for simplification. There is only
one multipath ray incident with MDR  0.707 to be a strong
10-2 reflection. Its AOA is changed at 40° and 10°, and the delay
estimation RMSEs are correspondingly plotted in Fig. 7 and
Fig. 8.
In Fig. 7 and Fig. 8, nEML, MEDLL and the proposed
10 -3 method all have better multipath mitigation performances than
beamformed conventional DLL. The proposed method has the
best tracking accuracies among them. Similar to the discussion
in the MEE subsection, although the nEML reduce the tracking
error by its narrow discriminator distance, the uncancelled bias
10-4 still significantly deteriorates the accuracy. The MEDLL has a
-30 -25 -20 -15 -10 -5 0
Chip SNR(dB) better RMSE performance than nEML in Fig. 7 and Fig. 8. It
cancels multipath rays by the iterative successive interference
Conventional DLL Theory (N=1) cancelation, but the undesired iteration residual always is also
Beamformed DLL Theory (N=4) brought in the delay estimation.
Beamformed DLL(N=4)
Conventional nEML Theory (N=1) In Fig. 7, due to the good distinguishing ability in space and
Beamformed nEML Theory (N=4) time, the curve of the proposed method with L  9 is nearly
Beamformed nEML (N=4) the same with its RMSE when no multipath is incident as in
Beamformed MEDLL (N=4) Fig. 6. The performance loss between two curves is only 0.2dB.
The Proposed Method(N=4,L=9)
The Proposed Method(N=4,L=31)
Besides, the Fig. 7 also proves that the proposed method with
larger L can achieve better delay accuracy and higher temporal
Fig. 6 Delay RMSE comparison when LOS and multipath rays are from 0° resolution.
and 30°respectively with RDLM=0.5 Tc

1264

Authorized licensed use limited to: University of Exeter. Downloaded on June 20,2020 at 08:08:40 UTC from IEEE Xplore. Restrictions apply.
 10-1 10-1

10-2 10-2

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)
Beamformed DLL Theory (N=4,LOS)
Beamformed DLL (N=4)
Beamformed nEML (N=4)
Beamformed MEDLL (N=4)
The Proposed Method (N=4,L=9)
The Proposed Method (N=4,L=31)

Fig. 7 Delay RMSE comparison when LOS and multipath rays are from 0° and 40°
respectively with RDLM=0.1 Tc (left panel) and RDLM=0.5 Tc (right panel)

100 100

10-1 10-1

10-2 10-2

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)
Beamformed DLL Theory (N=4,LOS)
Beamformed DLL (N=4)
Beamformed nEML (N=4)
Beamformed MEDLL (N=4)
The Proposed Method (N=4,L=9)
The Proposed Method (N=4,L=31)

Fig. 8 Delay RMSE comparison when LOS and multipath rays are from 0° and 10°
respectively with RDLM=0.1 Tc (left panel) and RDLM=0.5 Tc (right panel)

The space distance between incident rays becomes smaller MEDLL cannot work well with small RDLM as the discussion
in Fig. 8. The performance of the proposed method with L  9 of the MEE. Besides, the proposed method becomes sensitive
is better than MEDLL when RDLM is 0.1Tc, but it is getting when the distance between different rays in parameter domain
worse at some SNR points with the RDLM increasing. The is too closed. Little AOA difference often makes R FBSSZZ (Q  1)
method with L  31 is still the best one among these curves YZ (Q  1) ill-conditioned during the FBSS process, and
and R FBSS
but some degradation has been paid compared with the results the situation gets worse as that the short relative delay between
in Fig. 6 and Fig. 7. The reason for this phenomenon is that incident rays causes their joint spatial and temporal signatures

1265

Authorized licensed use limited to: University of Exeter. Downloaded on June 20,2020 at 08:08:40 UTC from IEEE Xplore. Restrictions apply.
still too correlated to be separated. Therefore, the estimate of [8] E. S. Lohan, A. Lakhzouri, and A. Renfors, "Feedforward delay
estimators in adverse multipath propagation for Galileo and modernized
signal subspace is severely perturbated by noise, and the delay GPS signals," Eurasip Journal on Applied Signal Processing, vol. 2006,
estimation accuracy degradation has to be paid. The higher no. 1, p. 050971, December 01 2006.
temporal resolution, introduced by the wider components in [9] L.-T. Hsu, S.-S. Jan, P. D. Groves, and N. Kubo, "Multipath mitigation
frequency, is able to relief some part of the above problem as and NLOS detection using vector tracking in urban environments," GPS
the curves of the proposed method with L  31 . Solutions, vol. 19, no. 2, pp. 249-262, 2014.
[10] B. Xu and L. T. Hsu, "Open-source MATLAB code for GPS vector
As shown in Fig. 7 and Fig. 8, the nEML and MEDLL tracking on a software-defined receiver," GPS Solutions, vol. 23, no. 2,
curves both get flat when chip SNR is high. It means that there Apr 2019.
are some biases not cancelled. On the contrary, the RMSEs of [11] R. Kumar and M. G. Petovello, "3D building model-assisted snapshot
positioning algorithm," GPS Solutions, vol. 21, no. 4, pp. 1923-1935,
the proposed method are always decreasing with the SNR Oct 2017.
increasing. It can be implied that the delay biases have been [12] F. Khosravi, H. Moghadas, and P. Mousavi, "A GNSS Antenna With a
eliminated by the proposed method in any circumstances. Polarization Selective Surface for the Mitigation of Low-Angle
Multipath Interference,", IEEE Transactions on Antennas and
Propagation, vol. 63, no. 12, pp. 5287-5295, Dec 2015.
VI. CONCLUSION
[13] M. G. Amin and W. Sun, "A novel interference suppression scheme for
We propose a novel phase extended subspace-based spatial global navigation satellite systems using antenna array," IEEE Journal
and temporal mitigation method for GPS receiver with ULA in on Selected Areas in Communications, vol. 23, no. 5, pp. 999-1012, May
2005.
multipath environment. By using the signal BPSK modulation,
[14] D. Lu, H. Liu, and R. Wu, "Global positioning system anti-jamming
the carrier phase is firstly extended to construct the phase algorithm based on period repetitive CLEAN," IET Radar, Sonar &
extended signatures. with the use of the ULA structure, a Navigation, vol. 7, no. 2, pp. 164-169, 2013.
subspace rotation invariance and DOAM-like correlation [15] N. Chang, X. Hong, W. Wang, and Z. Wang, "Subspace Based Joint
matrices are both established. The FBSS is then introduced to Delay and Direction of Arrival Estimation for GNSS Multipath Signals,"
China Satellite Navigation Conference (CSNC) 2018 Proceedings,
solve the coherent problem in eigenvalues decomposition by Singapore: Springer Singapore, 2018, pp. 189-199.
multipath. The extended signatures and AOAs of the incident [16] N. Chang, X. Hong, W. Wang, and Z. Wang, "Joint Delay and Angle
rays are estimated from the eigenvectors and eigenvalues of Estimation for GNSS Multipath Signals Based on Spatial and
the DOAM-like cross-correlation matrix, and the signatures Frequential Smoothing," in 2018 14th IEEE International Conference on
also provide the relative delay and carrier phase estimates with Signal Processing (ICSP), 2018, pp. 203-207.
respect to the tracking loop. The LOS ray is easily picked up [17] X. Hong, N. Chang, W. Wang, and Q. Yin, "Subspace-based Joint DOA,
Delay and DFO Estimation for GNSS Multipath Signals," presented at
by its first arrival character, and finally the biased DLL results the Proceedings of the 31st International Technical Meeting of The
are modified by the subspace estimator results. Simulation Satellite Division of the Institute of Navigation (ION GNSS+ 2018),
results prove that the proposed method has the much better Miami, Florida, September 2018, 2018.
MEE and RMSE performances than nEML and MEDLL in [18] M. S. Mahmud, S. U. Qaisar, and C. Benson, "Weak GPS signal
multipath environment. detection in the presence of strong signals with varying relative Doppler
and long integration gain," in 2016 IEEE/ION Position, Location and
Navigation Symposium (PLANS), 2016, pp. 1015-1020.
REFERENCES [19] IS-GPS-200J, NAVSTAR GPS space segment/navigation user segment
interfaces. Global positioning system directorate systems engineering &
[1] C. Fernandez-Prades, L. Lo Presti, and E. Falletti, "Satellite
integration: https://www.gps.gov/technical/icwg/IS-GPS-200J.pdf
Radiolocalization From GPS to GNSS and Beyond: Novel Technologies
and Applications for Civil Mass Market," Proceedings of the IEEE, vol. [20] W. L. Mao and A. B. Chen, "New code delay compensation algorithm
99, no. 11, pp. 1882-1904, Nov 2011. for weak GPS signal acquisition," Aeu-International Journal of
Electronics and Communications, vol. 63, no. 8, pp. 665-677, 2009.
[2] A. J. Van Dierendonck, P. A. T. Fenton, and T. O. M. Ford, "Theory and
Performance of Narrow Correlator Spacing in a GPS Receiver," [21] X. Peng and M. G. Petovello, "Measuring GNSS Multipath Distributions
Navigation, vol. 39, no. 3, pp. 265-283, 1992. in Urban Canyon Environments," IEEE Transactions on Instrumentation
and Measurement, vol. 64, no. 2, pp. 366-377, 2015.
[3] R. D. J. v. Nee, "The Multipath Estimating Delay Lock Loop," in IEEE
Second International Symposium on Spread Spectrum Techniques and [22] M. Irsigler, "Characterization of multipath phase rates in different
Applications, 1992, pp. 39-42. multipath environments," GPS Solutions, vol. 14, no. 4, pp. 305-317,
Sep 2010.
[4] B. R. Townsend, P. C. Fenton, K. J. Van Dierendonck, and D. J. R. Van
Nee, "Performance Evaluation of the Multipath Estimating Delay Lock [23] J. H. Cozzens and M. J. Sousa, "Source Enumeration in a Correlated
Loop," Navigation, vol. 42, no. 3, pp. 502-514, 1995. Signal Environment," IEEE Transactions on Signal Processing, vol. 42,
no. 2, pp. 304-317, Feb 1994.
[5] L. Garin and R. J. M, "Enhanced Strobe Correlator Multipath Rejection
for Code & Carrier," presented at the Proceedings of the 10th [24] G. H. Xu, R. H. Roy, and T. Kailath, "Detection of Number of Sources
International Technical Meeting of the Satellite Division of The Institute Via Exploitation of Centro-Symmetry Property," IEEE Transactions on
of Navigation (ION GPS 1997), Kansas City, MO, September 1997, Signal Processing, vol. 42, no. 1, pp. 102-112, Jan 1994.
1997. [25] J. M. Xin, N. N. Zheng, and A. Sano, "Simple and efficient
[6] Q. Jia, R. Wu, W. Wang, D. Lu, L. Wang, and J. Li, "Multipath nonparametric method for estimating the number of signals without
interference mitigation in GNSS via WRELAX," GPS Solutions, vol. 21, eigendecomposition," IEEE Transactions on Signal Processing, vol. 55,
no. 2, pp. 487-498, 2016. no. 4, pp. 1405-1420, Apr 2007.
[7] S. Li, J. Lv, J. Chang, X. Tian, and K. Tong, "Optimization Selection of [26] X. Chen, F. Dovis, S. L. Peng, and Y. Morton, "Comparative Studies of
the HRC Correlator Spacing for Different Navigation Signals," China GPS Multipath Mitigation Methods Performance," IEEE Transactions on
Satellite Navigation Conference (CSNC) 2015 Proceedings: Volume II. Aerospace and Electronic Systems, vol. 49, no. 3, pp. 1555-1568, Jul
Berlin, Heidelberg, 2015, pp. 3-13: Springer Berlin Heidelberg. 2013.

1266

Authorized licensed use limited to: University of Exeter. Downloaded on June 20,2020 at 08:08:40 UTC from IEEE Xplore. Restrictions apply.