754 IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS, VOL. 25, NO.

1, JANUARY 2024

Channel Parameter Estimation of mmWave MIMO

System in Urban Traffic Scene: A Training
Channel-Based Method
Han Wang , Senior Member, IEEE, Pingping Xiao, and Xingwang Li , Senior Member, IEEE

Abstract— In frequency selective channel environment, channel 6GHz, which seriously restricts the development of commu-
estimation in hybrid precoding millimeter-wave (mmWave) mas- nication industry. The implementation of wireless commu-
sive multiple input multiple output (MIMO) system is a challenge nication in millimeter wave (mmWave) band has attracted
issue. To solve this problem, we propose an effective channel
estimation scheme for frequency selective channel, which is based extensive research interest [4]–[7]. However, the development
on the training channel model in urban traffic environment. and application of Massive MIMO also face many problems.
Considering that the practical mmWave MIMO channel is For example, the large increase in the number of antennas
sparsity and the subcarrier multi-channels have the same sparse makes the system need a large amount of computation, which
structure, we regard the channel estimation problem as the puts forward higher requirements for the hardware platform.
sparse channel recovery, and propose a multipath simultaneous
matching tracking estimation method. It is assumed that the noise The pilot overhead of traditional channel estimation (CE) will
between the practical channels has a certain correlation, and the also be enlarged with the increase of the size of the trans-
noise correlation has an impact on the selection of the optimal ceiver antenna, resulting in the reduction of the system band
atomic support set in the process of channel recovery. Therefore, utilization. Millimeter wave technology improves the network
noise weighting is introduced in our proposed method. The transmission rate by increasing the spectrum bandwidth. It has
simulation results prove the validity of this proposed method in
frequency selective mmWave MIMO channel. Without increasing the advantages of high gain, strong anti-interference, narrow
the complexity of the algorithm, the proposed method can achieve beam and strong directivity. However, it will be affected by
better local performance than the traditional classical methods. factors such as propagation path loss, building penetration loss
Index Terms— Urban traffic scene, mmWave MIMO, training and rain and snow fading. It is facing great challenges in
channel model, channel estimation. practical application [8].
The mmWave communication can be organically integrated
with massive MIMO technology. The mmWave has shorter
I. I NTRODUCTION working wavelength, which can effectively reduce the size of
A. Background devices and systems. The array gain brought by large-scale
MIMO beamforming can make up for the disadvantage of
M ASSIVE multiple input multiple output (MIMO) has
become one of the core technologies of the fifth gen-
eration communication system (5g) [1]–[3], because it has
poor millimeter wave penetration. The mmWave Massive
MIMO wireless transmission [9], [10] can expand the use of
the advantages of improving system capacity, user experience new spectrum resources, further tap spatial wireless resources
rate, enhancing full-dimensional coverage and saving energy and greatly improve the wireless transmission rate. It is one
consumption. On the other hand, it is difficult to obtain a of the most potential research directions of future wireless
wide continuous spectrum in the gold frequency band below communication systems. However, the performance of the
wireless communication system is affected by wireless chan-
Manuscript received 5 October 2021; revised 4 December 2021; accepted nel. Accurate channel parameter estimation is a key technology
19 January 2022. Date of publication 1 February 2022; date of current to realize wireless system. Therefore, the research on CE is
version 17 January 2024. This work was supported in part by the National
Natural Science Foundation of China under Grant 61901409, in part by the of great significance. In practical mmWave Massive MIMO
Open Project of State Key Laboratory of Marine Resources Utilization in system, hybrid precoding technology is usually used to reduce
South China Sea under Grant MRUKF2021034, and in part by the Operation its hardware cost and energy loss. The design of hybrid
Mechanism and Active Control of Urban Traffic System under Intelligent
Connected Conditions (FDCT-MOST) Joint under Grant 0091/2019/AMJ. precoder requires accurate channel state information (CSI).
The Associate Editor for this article was L. Wan. (Corresponding author: Due to the limited number of RF links, the baseband cannot be
Han Wang.) connected to all antennas at the same time. Under the condition
Han Wang is with the College of Physical Science and Engineering, Yichun
University, Yichun 336000, China, also with the Faculty of Data Science, City of limited hardware, the accurate CE of mmWave Massive
University of Macau, Macao, China, and also with the State Key Laboratory of MIMO system is a difficult problem. On the other hand, due
Marine Resource Utilization in South China Sea, Hainan University, Haikou to the influence of large-scale antenna arrays, the traditional
570228, China (e-mail: [email protected]).
Pingping Xiao is with the College of Physical Science and Engineering, CE methods are not suitable for mmWave Massive MIMO
Yichun University, Yichun 336000, China (e-mail: [email protected]). systems. For example, the number of pilots required by the
Xingwang Li is with the School of Physics and Electronic Information traditional least square method is directly proportional to the
Engineering, Henan Polytechnic University, Jiaozuo 454000, China (e-mail:
[email protected]). number of antennas. In Massive MIMO systems, the difficulty
Digital Object Identifier 10.1109/TITS.2022.3145363 of CE increases sharply with the increase of the number of
1558-0016 © 2022 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See https://www.ieee.org/publications/rights/index.html for more information.

Authorized licensed use limited to: ULAKBIM UASL - ANKARA UNIVERSITESI. Downloaded on December 16,2024 at 18:46:06 UTC from IEEE Xplore. Restrictions apply.
 WANG et al.: CHANNEL PARAMETER ESTIMATION OF mmWave MIMO SYSTEM IN URBAN TRAFFIC SCENE 755

antennas and users. In order to realize mmWave Massive

MIMO communication system, the problem of effective CE
in its communication process must be solved.

B. Related Work
Massive MIMO technology utilizes ultra-high antenna
dimension to fully tap space resources, mmWave communica-
tion uses ultra-large bandwidth to improve network through-
put, which will produce a large amount of wireless data. It is
of great practical significance to explore the corresponding low
complexity robust CE methods combined with machine learn-
ing theory [11], [12]. Compressed sensing (CS) essentially
transforms the input features to form a new feature expression,
so it can also be regarded as a method of feature learning. But
unlike deep learning (DL) method, CS itself is only a single-
layer algorithm. After a set of basis vectors (usually called
as dictionary) in a new feature expression space, for the new
Fig. 1. Urban Traffic Scene.
input features, CS needs to select a part of the vectors from
these basis vectors to complete the reconstruction of the input
features. The main feature of CS is that it requires as few
base vectors as possible to be selected in the reconstruction Bayesian CS algorithm can achieve higher precision estimation
process, that is, the reconstruction coefficients are as sparsity when the prior knowledge of sparse channel is incomplete.
as possible. In this way, the signal can be compressed better, The above CE methods for mmWave massive MIMO sys-
and the expression ability of the feature can be improved to tem can obtain better CE performance than the traditional
some extent, so that the feature can achieve better results in pilot based algorithm by developing the sparse statistical
some tasks, such as channel estimation. At present, scholars’ characteristics of the channel. However, most traditional CS
research on CE of mmWave massive MIMO systems mainly algorithms seek the optimal solution by greedy iteration and
focuses on estimation approaches using its sparsity character- rely on sparse channel prior information, and the channel
istics. In Refs. [13]–[18], the authors studied CS based CE model used in most studies is flat fading multipath channel.
approach for mmWave massive MIMO systems. In Ref. [13], But the practical frequency selective channel is characterized
the CE problem is solved by the orthogonal matching pur- by random uncertainty, and the channel sparsity cannot be
suit (OMP) algorithm, which employed a redundant dictionary known as a priori. Therefore, it is necessary to further study
consisting of array response vectors with finely quantized new methods to solve the parameter estimation problem of
angle grids. Simulation results verified the superiority of the practical mmWave MIMO channels.
CS based CE method. Most existing research on CE of In Refs. [19]–[23], the authors studied the CS based CE
mmWave systems utilizes the sparse scattering characteristics for hybrid mmWave MIMO systems, in which the wireless
of the channel. In order to effectively leverage the joint channel is frequency selective. In Refs. [19], [20], utilizing
sparse and low-rank structure, the authors in [14] developed the sparse angle domain structure of mmWave frequency
a two-stage CS method for mmWave systems, in which the selective fading channel, the authors proposed a CE method
sparse and low-rank properties were respectively used in two based on distributed CS and adopted grid matching track-
consecutive stages. Simulations verified the superiority of the ing strategy to solve the problem of power leakage caused
proposed method. Ref. [15] proposed a sparse CE method by continuous angles of arrival or departure (AOA/AOD).
based on regularization. Simulation experiments show that In Ref. [19], the authors proposed a multiuser uplink CE
this method can obtain better estimation performance than method for mmWave MIMO channel. The authors in [21], [22]
traditional methods. Traditional spatial meshing leads to the proposed a frequency-domain strategy to estimate mmWave
trade-off between CE accuracy and complexity. Ref. [16] channels. The channel sparsity in the two above methods is
proposed a new sparse CE algorithm, which has good per- assumed to be known as a priori. Since the classical OMP
formance in off grid scenarios. In Ref. [17], the authors algorithm needs to run the algorithm as many times as the
proposed a low complexity OMP algorithm by using the number of subcarriers. The authors in [23] considered an
internal correlation and statistical scattering information in additional assumption to solve the problem, which avoids the
the signal angle domain. The complexity of the proposed need to run K sparsity times OMP algorithms in parallel
algorithm is reduced by more than 90% while compared with as proposed in [21]. The authors in [23] considered a fre-
the traditional OMP method, and the estimation performance quency selective channel model, and proposed a simultaneous
is better than the existing OMP series algorithms. Ref. [18] weighted orthogonal matching pursuit (SWOMP) CE method.
is about the research on Bayesian CS based CE methods for The SWOMP generalizes the algorithm to account for the
mmWave massive MIMO systems. The authors concluded that correlated noise after combining, and the channel sparsity level

Authorized licensed use limited to: ULAKBIM UASL - ANKARA UNIVERSITESI. Downloaded on December 16,2024 at 18:46:06 UTC from IEEE Xplore. Restrictions apply.
 756 IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS, VOL. 25, NO. 1, JANUARY 2024

Fig. 2. Block diagram of how the datasets are used for the channel estimation challenge [24].

does not need know as a priori. Most of prior research on 1) We propose a multipath simultaneous weighted match-
CE for hybrid mmWave architectures supposes a frequency ing pursuit (MSWOMP) algorithm, the simulation
selective channel model. There are differences between the results verify its superiority.
traditional classical channel model and the practical channel 2) Most of the traditional methods adopt the classical chan-
model. In Ref. [24], the authors presented a method to generate nel model for simulation. In this paper, we utilize the
channel realizations that represents 5G scenarios with the practical millimeter wave frequency selective channel to
mobility of both transceivers and objects. Then, the authors study the channel estimation based on the CS method.
provided a specific dataset for investigating beam selection The experimental simulation verifies the effectiveness
techniques on vehicle to infrastructure utilizing millimeter of the proposed CS method for the practical channel
waves. The above mentioned channel training dataset can be recovery.
used for CS based frequency selective CE. The method used 3) As a special kind of machine learning method, the results
to generate the frequency selective channels can be found of this paper can provide a reference for further explo-
in [24], and it is summarized in the following two figures. ration of machine learning CE approach in mmWave
Fig. 1 shows the Urban Traffic Scene. The rays were obtained frequency selective channel.
in a traffic jam situation. Fig. 2 gives the block diagram of how The paper is organized as follows. We first introduce the
the datasets are used for the CE challenge. The Raymobtime system model in Section II. In Section III, CS based CE
datasets are generated by ray tracing, which can be obtained in frequency selective mmWave communication is analyzed,
from Fig. 1. and the MSWOMP method is proposed for mmWave CE.
In traditional CS greedy algorithm, the performance of Section IV provides the CE performance comparisons of
the algorithm is usually improved by adjusting the atomic the proposed method and other classical methods. Finally,
selection strategy. For example, OMP algorithm selects a Section V gives the conclusions.
single atom in each iteration, compressive sampling matching
pursuit (CoSaMP) algorithm selects multiple atoms in each
II. S YSTEM M ODEL
iteration, and the other algorithms select atoms through thresh-
old. Ref. [25] proposed a multipath matching pursuit (MMP) A. The mmWave MIMO System
algorithm, which provided a new idea for atomic selection. We consider an orthogonal frequency division multiplex-
The MMP modeled the atomic selection problem of traditional ing (OFDM) based mmWave massive hybrid MIMO system,
greedy algorithm as the search problem of combinatorial tree. with Nc subcarriers to send Ns data streams. The mmWave
Simulation results show that this method can achieve higher hybrid system is shown in Fig. 3, with Nt the transmitter
precision signal reconstruction. Ref. [26] proposed a regular- antennas and Nr the receiver antennas, L t and L r are the
ized MMP CE method for mmWave Massive MIMO system. radio frequency (RF) chains at the transmitter and receiver
The proposed regularized MMP algorithm outperforms the sides. A frequency selective hybrid precoder is adopted in
existing OMP and orthogonal least squares algorithms. the hybrid structure, with F[k] = FRF FBB [k] ∈ C Nt ×Ns ,
Based on the above analysis, combined with the atomic k = 0, . . . , Nc − 1, where FRF and FBB [k] represent analog
selection strategy of MMP method and considering the influ- precoder and digital precoder. It should be noted that the
ence of channel noise correlation on atomic selection, we pro- digital precoder is different for each subcarrier, and the analog
pose a new method to estimate the practical frequency selec- precoder is frequency flat. The RF precoder and combiner
tive channel. The main contributions of this paper are given are realized by using a fully connected network of phase
below. shifters. We use L t parallel K point inverse fast fourier

Authorized licensed use limited to: ULAKBIM UASL - ANKARA UNIVERSITESI. Downloaded on December 16,2024 at 18:46:06 UTC from IEEE Xplore. Restrictions apply.
 WANG et al.: CHANNEL PARAMETER ESTIMATION OF mmWave MIMO SYSTEM IN URBAN TRAFFIC SCENE 757

Fig. 3. The hybrid mmWave MIMO architecture.

transform (IFFT) to transform the symbol blocks into the Then, the channel at subcarrier k with different delay taps
time domain. The complex baseband signal at discrete-time can be expressed as
subcarrier k can be given as c −1
N
Hd e − j = A R [k]A∗T
2π k
H[k] = K d
x[k] = FRF FBB [k]s[k] (1)
d=0
 N −1 
where s[k] is the transmitted symbol sequence at k-th subcar- c
2π k d

rier of size Ns × 1. ≈ Ã R vd e− j K Ã∗T ≈ Ã R v [k]Ã∗T (5)

d=0
The MIMO channel between the transmitter and the receiver
is assumed to be frequency selective, and has a delay tap The hybrid combiner is W[k] = W R F W B B [k], with W R F
length of Nc . Hd is the Nr × Nt matrix, which represents the the RF combining matrix and W B B [k] the baseband combining
d-th delay tap of the channel, with d = 0, 1, . . . , Nc − 1.The matrix. The received signal at subcarrier k can be expressed
channel model can be written as [23] as
 y[k] = W∗B B [k]W∗R F H[k]F R F F B B [k]s[k]
Nt Nr 
L
Hd = αl prc (d Ts − τl )a R (φl )a∗T (θl ) (2) +W∗B B [k]W∗R F n[k] (6)
Lρ L
l=1
where n[k] is the Gaussian additive noise.
where L denotes the number of multi-paths, Ts is the sampling
period, and ρ L is the path loss between the transmitter and
B. CS-Based CE for mmWave MIMO System
receiver. prc is a filter, and αl is the complex gain of the
l-th path, τl is the delay of the l-th path. φl ∈ [0, 2π) In this section, the CS approach is introduced to estimate the
and θl ∈ [0, 2π) are the AoA/AoD of the l-th path, a R (φl ) sparse mmWave MIMO channel vector in frequency domain.
and aT (θl ) are the array steering vectors for the receive and At the transmitter and receiver, L t and L r RF chains are
transmit antennas. The superscript (·)∗ denotes conjugation. utilized. The m-th training frame of the received sample at
The channel Hd can be written in a more compact manner as the k-th subcarrier in frequency domain can be written as [22]
(m)∗ (m)
y(m)[k] = Wtr H[k]Ftr q(m) t(m) [k] + nc(m) [k] (7)
Hd = A R d A∗T (3)
(m)
where Wtr ∈ C Nr ×L r is the training combiner, and F(m)
tr ∈
where A R ∈ C Nr ×L and AT ∈ C Nt ×L embody the receive and C N t ×L t is the training precoder. H[k] is the frequency domain
transmit array steering vectors a R (φl ) and aT (θl ), respectively. mmWave MIMO channel. It is assumed that the channel
d is diagonal with non-zero complex entries. The channel Hd coherence time is larger than the frame duration and that
can be approximated by using the extended virtual channel the same channel can be considered for several consecutive
model in [9] as frames. The transmitted symbol s(m) [k] is decomposed as
s(m) [k] = q(m) t(m) [k], with t(m) [k] a pilot symbol know at
Hd ≈ Ã R vd Ã∗T (4) (m)
the receiver, and q(m) a frequency flat vector. nc [k] =
(m)∗ (m)
where vd is a sparse matrix containing the path gains of Wtr n [k]. The vectorized received signal can be given
quantized spatial frequency in non-zero elements, with vd ∈ as
CG r ×G t , a grid size G r for the AoA and a grid size G t for the vec{y(m) [k]} = (q(m)T F(m)T
tr
(m)∗
⊗ Wtr )vec{H[k]} + nc(m) [k]
AoD. Ã R and Ã∗T are the array response matrix of the receiver (8)
and transmitter. The sparse assumption for vd ∈ CG r ×G t is
generally accepted since there are few scattering clusters in Considering the expression in formula (5), the vector-
mmWave channel. ized channel matrix can be given as vec{H[k]} = (ÃT ⊗
Authorized licensed use limited to: ULAKBIM UASL - ANKARA UNIVERSITESI. Downloaded on December 16,2024 at 18:46:06 UTC from IEEE Xplore. Restrictions apply.
 758 IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS, VOL. 25, NO. 1, JANUARY 2024

à R )vec{[k]}. Hence, we define the matrix as measurement III. P ROPOSED M ULTIPATH S IMULTANEOUS W EIGHTED
(m)T (m)∗
matrix, with (m) ∈ C L r ×Nt Nr = q(m)T Ftr ⊗ Wtr . The M ATCHING P URSUIT M ETHOD
dictionary matrix is ∈C Nt Nr ×G t G r = ÃT ⊗ Ã R . Formula
(8) can be given as In Ref. [22], the authors leveraged the sparse nature of
mmWave channel and the common support property for all
vec{y(m)[k]} = (m)
hv [k] + nc(m) [k] (9) sub-channel, and considered the correlation noise after com-
where hv [k]
= vec{[k]}, hv [k]
∈ CG r G t ×1
is the sparse vec- bining in mmWave system, the proposed SWOMP method
tor containing the channel gains. In order to obtain sufficient can reach the Cramer-Rao Lower Bound (CRLB). The authors
measurements and accurately rebuilt the sparse vector hv [k], in [24] provided a multipath matching pursuit strategy, which
multiple training frames need to be used. We extend formula can achieve higher precision signal reconstruction. Inspired
(9) to by the above two strategies, we exploit the above mentioned
⎡ (1) ⎤ ⎡ (m) ⎤T ⎡ (m) ⎤ characteristics of mmWave channels, and propose a multipath
y [k] nc [k]
⎢ . ⎥ ⎢ . ⎥ ⎢ . ⎥ simultaneous weighted matching pursuit (MSWOMP) algo-
⎢ . ⎥=⎢ . ⎥ h [k] + ⎢ .
v ⎥ (10) rithm for mmWave sparse channel estimation. We first give the
⎣ . ⎦ ⎣ . ⎦ ⎣ . ⎦
support calculation of correlation noise. Define c[k] ∈ C M L r
y(M) [k] (M)
nc(M) [k] as the correction vector, with
     
y[k] nc [k]
where M is the training steps. c[k] = ϒ ∗ y[k] (14)
Finally, hv [k] can be obtained by solving the problem of
compressive sensing sparse signal reconstruction where ϒ ∈ C M L r ×G t G r is the equivalent measurement matrix,
   2 and ϒ = . y[k] is the received signal. If there is a
min hv [k]1 subject to y[k] − hv [k]2 < ε (11) correlation between the noise components, the atom estimated
where ε is the adjustable parameter that defines the maxi- from (14) should not be an actual atom. We need to adopt
mum error between the measurement and the received signal. the specific form of the noise covariance matrix to introduce
(i)
Because the sparsity of the channel is usually unknown, appropriate corrections in the correlation. Denote nc [k] the
choosing a suitable ε is the key to solving formula (11). combined noise, with the given subcarrier k and training step
In fact, the matric [k] shows an interesting property, which i.
can be used to solve the problem of compressive channel (i)∗ (i)
estimation, which is defined in (11). Define the vectorized nc(i) [k] = Wtr n [k] (15)
channel matrix for a given delay tap as
where nc(i) [k] ∼ N(0, σ 2 I Lr ), Wtr (i)
is the combiner.
hdv  vec{vd } (12) (i) ( j )∗ (i)∗ 2 ( j)
E{nc [k]nc [k]} = Wtr σ δ[i − j ]Wtr represents the
since formula (9) has been given hv [k] = vec{[k]}, with noise cross-covariance matrix. The noise covariance matrix
c −1
N
[k] = vd e− j N d , k = 0, . . . , K − 1, obviously,
2π k of y[k] can be given as a block diagonal matrix C, with
(1)∗ (1) (2)∗ (2) (M)∗ (M)
d=0 Cw = blkdiag{Wtr Wtr ,Wtr Wtr , . . . , Wtr Wtr }. Use
in general the Cholesky factorization to express Cw as Cw = D∗w Dw ,
 c −1
 N   with Dw ∈ C M L r ×M L r the upper triangular matrix. Therefore,
supp hv [k] = supp hdv (13) the correlation noised vector is given below
d=0
among them, the support union of time domain virtual channel c[k] = ϒw∗ yw [k] (16)
matrix comes from the additive property of Fourier transform.
Therefore, the channel model in (5) shows the same sparse where ϒw = D−∗ w ϒ is the whitened measurement matrix,
structure for all K subcarrier channels, while AoA and AoD do yw [k] = D−∗w y[k] is the whitened received signal, and. Here,
not change with the frequency in the transmission bandwidth. is regarded as a baseband combiner used in the m-th training
Note that AoA/AoD in formula (2) does not depend on delay step. In this way, by considering the correlation noise, and
tap d, hence
 formula (13) can be written as supp {hv [k]} = simultaneously estimating the most likely support atom, the
v
supp hd . Since in the mmWave channel, L  G r G t , for sparse channel hv [k] can be reconstructed.
a given delay tap hdv , we usually assume that the vectorized Based on the above analysis and combined with multi-path
channel matrix is sparsity. Therefore, the mmWave channel matching tracking strategy, we proposed the SMWOMP algo-
can be sparsely recovered by the CS sparse reconstruction rithm, the proposed algorithm flow is given below.
approach. It should be noted that the above proposed algorithm
There are many sparse reconstruction algorithms to solve requires the hyper-parameter L. In conventional SWOMP
formula (11), OMP algorithm is one of the classical CS meth- method, it should either know the noise variance or the sparsity
ods. However, for the massive MIMO channel, the computa- level of the signal for the stopping criteria. We used the chan-
tional complexity of OMP will be greatly improved. Therefore, nel training data to learn the sparsity level of the channel which
we can consider using the characteristics of mmWave channel is an input to our proposed MSWOMP algorithm. By using the
structure to study the low complexity sparse reconstruction training data, the optimized sparsity level is learned for three
algorithm. SNR ranges using the training datasets generated at −15dB,

Authorized licensed use limited to: ULAKBIM UASL - ANKARA UNIVERSITESI. Downloaded on December 16,2024 at 18:46:06 UTC from IEEE Xplore. Restrictions apply.
 WANG et al.: CHANNEL PARAMETER ESTIMATION OF mmWave MIMO SYSTEM IN URBAN TRAFFIC SCENE 759

Algorithm 1 TABLE I
S IMULATION PARAMETERS
Require: y, , , Dw , K (sparsity), L (number of path)
Ensure: estimated Ĥ
Set ϒw = D−∗ w
Initialization: i = 0 (iteration index), yw = D−∗
w y, r = y
0

(initial residual), S = { }.
0

while k ≤ K do
k = k + 1, u = 0, Sk = φ
c = ϒw r (Distributed
  correlation)
for i = 1 to Sk-1  do
TABLE II
 2 T HE C OMPARISON OF RUNNING T IME (S ECOND )
p∗ = arg max c p 2 (Select L best indices)
p=L
for j = 1 to L do
st emp = sik−1 { p ∗j } (Temporary path)
if st emp ∈ / Sk then
u =u+1 squared error (NMSE) is utilized to evaluate the simulation
suk = st emp
 performance. It is defined as
Sk = Sk {suk } (Update the set of the path)
xuk = ([ϒw ]:,suk )† yw (Project input signal using  K −1 

2

k=0  Ĥ [k] − H [k]
weighted least squares)
NMSE =  K −1
F
(18)
ruk = yw − [ϒw ]:,suk x (Update residual) k=0 H [k] 2
F
end if
end for where Ĥ [k] is the estimated channel.
end for In the mmWave MIMO system, it is assumed that both
end while   the transmitter and receiver use uniform linear array (ULA)
u ∗ = arg min r K 
2
u 2 with half wavelength separation. We take Nt = 16, Nr =
s ∗ = suk∗ 64,and G t = 64, G r = 256.The simulation parameters are
return ĥ = [ ]:,s ∗ xuk ∗ and Ĥ = vec−1 (ĥ) summarized as follows.
In order to evaluate the accuracy of the proposed method,
we present the NMSE performance of the proposed method
−10dB, −5dB, respectively. The optimized L is given below and other four methods on the validation dataset. The val-
⎧ idation dataset of size is Nval = 100, which is extracted
⎨0, for γtrain = −15d B(−20d B  γt est < −11d B) from the provide dataset. The sparsity level L is estimated
L = 1, for γtrain = −10d B(−11d B  γt est < −6d B) from the training dataset. Fig. 4 gives the NMSE versus
⎩
3, for γtrain = −5d B(−6d B  γt est < −0d B) SNR curves from different algorithms, with training frames
(17) M = 20, 40, 80, respectively.
It can be seen from Fig. 4, with the increase of the number
where γt est denotes the SNR of the signal in the test dataset
of training frames ( M ), the performance of the five algorithms
which is unknown.
has developed and changed. The NMSE performance of the
five algorithms has improved, especially the M-SWOMP and
IV. S IMULATION R ESULTS our proposed algorithm. In the three sub-graphs of Fig. 4,
In this section, the mmWave MIMO CE performance of the M-SWStOMP algorithm has floor effect. It may due to
our proposed algorithm is evaluated. We utilize a prac- the reason that the number of iterations is too small, and the
tical frequency selective channel model, and the channel atoms selected in each iteration are not the best atoms, which
training dataset that we utilized consist of a collection of causes the performance deterioration.
10000 channels in HDF5 format, which can be found from In Fig. 4 (a), the M-SWCoSaMP method has the best NMSE
Ref. [24]. The methodology used to generate the channels performance under the moderate SNR of SNR= −10dB.
is summarized in Fig. 2. Some traditional methods, such as The proposed method can provide better NMSE values than
SWOMP, modified SWOMP(M-SWOMP), modified simul- other methods in the SNR range of −10dB to −5dB. When
taneous weighted stagewise OMP (M-SWStOMP), modified SNR= −5dB, the NMSE performance of the proposed method
simultaneous weighted CoSaMP (M-SWCoSaMP), are used is about 0.8dB better than the M-SWOMP method. However,
to compare with our proposed method. Note that in SWOMP in the low SNR range (−10dB to −5dB), MSWOMP algo-
algorithm, the stopping condition is determined by the noise rithm obtains the best NMSE performance.
variance. The M-SWOMP is modified SWOMP, and the With the number of M increases to 40, the NMSE per-
proposed MSWOMP algorithm is a multipath simultaneous formance of the five methods is improved. Compared with
weighted matching pursuit algorithm, they are two different the other three methods, the performance superiority of
algorithms. In modified algorithms, the stopping condition is M-SWOMP and the proposed algorithms are more obvious.
determined by the number of path L. The normalized mean The M-SWCoSaMP does not maintain the best performance

Authorized licensed use limited to: ULAKBIM UASL - ANKARA UNIVERSITESI. Downloaded on December 16,2024 at 18:46:06 UTC from IEEE Xplore. Restrictions apply.
 760 IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS, VOL. 25, NO. 1, JANUARY 2024

Fig. 5. Comparison of the NMSE performance of the proposed algorithm

and other four algorithms, with M = 40, Nt = 8, Nr = 32.

When the number of M increases to 80, it can be found

from Fig. 4(c) that in the case of low SNR range (−15dB
to −10dB), the proposed method shows the best NMSE
performance. However, the performance of M-SWOMP algo-
rithm is better than the proposed method in the SNR range of
−10dB to −5dB. Compared with Fig. 4(a) and (c), increasing
the number of M, the NMSE performance of our proposed
algorithm, M-SWOMP algorithm and M-SWCoSaMP algo-
rithm is significantly improved. When SNR= −10dB, the
performance of the proposed method under M = 80 is about
4.5dB higher than that under M = 20.
Increasing the number of M can promote the NMSE per-
formance of the proposed algorithm, but it will raise compu-
tational complexity and cause the higher overhead. Then, we
observe the performance change of the proposed method by
reducing the number of transmitting and receiving antennas.
Here, we choose. Fig. 5 shows the comparison of NMSE
performance curves for the five algorithms. Compared with
Fig. 4(b), careful observation exhibits that NMSE values of
the proposed method is slightly improved by 0.7dB when the
number of transmitting and receiving antennas is reduced.
Nevertheless, the performance of other traditional methods
has decreased. In other words, the proposed method can
achieve better performance improvement under the condition
of reducing the number of transmitting and receiving antennas.
In Fig. 6, we analyze the NMSE versus M under different
SNR for the three SWOMP, M-SWOMP and proposed algo-
rithms. Compared with the other two methods, the SWOMP
methods has the worst NMSE performance. In Fig. 6(a), when
SNR= −15dB, the NMSE performance of SWOMP algorithm
Fig. 4. Comparison of the NMSE performance of the proposed algorithm
and other four algorithms, with (a) training frames M=20, (b) training frames tends to be stable with the increase of the number of M. In the
M=40 and (c) training frames M=80. range of 20 to 70, the performance of M-SWOMP algorithm is
better than the proposed method. With the continuous increase
as it in M = 20. Under high SNR values, the proposed method of the number of M, the superior of the proposed method
can offer only a little better performance than M-SWOMP. becomes more obvious.

Authorized licensed use limited to: ULAKBIM UASL - ANKARA UNIVERSITESI. Downloaded on December 16,2024 at 18:46:06 UTC from IEEE Xplore. Restrictions apply.
 WANG et al.: CHANNEL PARAMETER ESTIMATION OF mmWave MIMO SYSTEM IN URBAN TRAFFIC SCENE 761

By increasing the SNR value to −10dB, the NMSE per-

formance of the three methods has been improved. As it
can be seen from Fig. 6(b), the curves of proposed method
and M-SWOMP method are consistent in the range of M =
20 to M = 80. With the number of M increasing, our
proposed method offers significantly better performance than
M-SWOMP method.
In Fig. 6 (c), when SNR= −5dB, compared with
Fig.6 (a) and (b), and starting from M = 40, the proposed
method has obvious better performance than other two meth-
ods with the increase of M.
Combined with the analysis of Fig. 4 and Fig. 6, we can
conclude that with the increase number of M, the proposed
method can provide better NMSE performance than the other
four methods under high SNR values. In the case of low
SNR, the proposed method can reflect the optimal NMSE
performance only at high M values, but this will raise the com-
putational complexity. Finally, we also compare and analyze
the simulation times of the five algorithms. The simulation
software is Matlab version R2016a on an Intel(R) Xeon(R)
W-2150B CPU with 64GB of RAM. As shown in Table II,
with M = 40, and SNR= −10dB, the SWOMP algorithm
has the longest simulation time, and M-SWCoSaMP has the
shortest simulation time. The proposed method takes less time
than M-SWOMP method. Combined with the previous analy-
sis, it could be concluded that compared with the traditional
methods, the proposed method can obtain the best NMSE
performance in the range of low and moderate SNR without
increasing the complexity of the algorithm.

V. C ONCLUSION
In this paper, we studied the sparse CE method for hybrid
mmWave massive MIMO systems. Considering the frequency
selective characteristics of practical mmWave channel, the
sparse structure of the channel was analyzed, a CS based
CE approach was proposed. By considering the correlation
noise after combining in mmWave system and combining
with multi-path matching tracking strategy, we proposed the
SMWOMP algorithm. The site-specific training dataset with
channel realizations was used to learn the number of dominant
paths in the channels. The simulation results showed that the
proposed method can achieve effective CE in the practical fre-
quency selective channel. Compared with traditional methods,
our proposed method had better CE performance under low
and moderate SNR. The proposed approach is more robust
without increasing the complexity of the algorithm.

R EFERENCES
[1] J. G. Andrews et al., “What will 5G be?” IEEE J. Sel. Areas Commun.,
vol. 32, no. 6, pp. 1065–1082, Jun. 2014.
[2] S. A. Busari, K. M. S. Huq, S. Mumtaz, L. Dai, and J. Rodriguez,
“Millimeter-wave massive MIMO communication for future wireless
systems: A survey,” IEEE Commun. Surveys Tuts., vol. 20, no. 2,
pp. 836–869, 2nd Quart. 2018.
[3] L. Wan, K. Liu, Y.-C. Liang, and T. Zhu, “DOA and polarization
Fig. 6. The NMSE versus M under different SNR for SWOMP, M-SWOMP estimation for non-circular signals in 3-D millimeter wave polarized
and proposed algorithms, with (a) SNR=-15dB, (b) SNR=-10dB and massive MIMO systems,” IEEE Trans. Wireless Commun., vol. 20, no. 5,
(c) SNR=-5dB. pp. 3152–3167, May 2021.

Authorized licensed use limited to: ULAKBIM UASL - ANKARA UNIVERSITESI. Downloaded on December 16,2024 at 18:46:06 UTC from IEEE Xplore. Restrictions apply.
 762 IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS, VOL. 25, NO. 1, JANUARY 2024

[4] S. Rangan, T. S. Rappaport, and E. Erkip, “Millimeter-wave cellular [24] A. Klautau, P. Batista, N. Gonzalez-Prelcic, Y. Wang, and R. W. Heath,
wireless networks: Potentials and challenges,” Proc. IEEE, vol. 102, “5G MIMO data for machine learning: Application to beam-selection
no. 3, pp. 366–385, Feb. 2014. using deep learning,” in Proc. Inf. Theory Appl. Workshop (ITA),
[5] Z. Gao, L. Dai, D. Mi, Z. Wang, M. A. Imran, and M. Z. Shakir, San Diego, CA, USA, Feb. 2018, pp. 1–9.
“MmWave massive-MIMO-based wireless backhaul for the 5G ultra- [25] S. Kwon, J. Wang, and B. Shim, “Multipath matching pursuit,” IEEE
dense network,” IEEE Wireless Commun., vol. 22, no. 5, pp. 13–21, Trans. Inf. Theory, vol. 60, no. 5, pp. 2986–3001, May 2014.
Oct. 2015. [26] J. Tao, C. Qi, and Y. Huang, “Regularized multipath matching pursuit for
[6] T. Rappaport, Y. Xing, G. R. MacCartney, A. F. Molisch, E. Mellios, sparse channel estimation in millimeter wave massive MIMO system,”
and J. Zhang, “Overview of millimeter wave communications for fifth- IEEE Wireless Commun. Lett., vol. 8, no. 1, pp. 169–172, Feb. 2018.
generation (5G) wireless networks-with a focus on propagation mod-
els,” IEEE Trans. Antennas Propag., vol. 65, no. 12, pp. 6213–6230,
Dec. 2017.
[7] J. G. Andrews, T. Bai, M. N. Kulkarni, A. Alkhateeb, A. K. Gupta,
and R. W. Heath, Jr., “Modeling and analyzing millimeter wave cel-
lular systems,” IEEE Trans. Commun., vol. 65, no. 1, pp. 403–430,
Jan. 2016. Han Wang (Senior Member, IEEE) received the
[8] M. Shafi et al., “Microwave vs. millimeter-wave propagation channels: B.S. degree in electrical engineering from the Hubei
Key differences and impact on 5G cellular systems,” IEEE Commun. University of Nationalities, China, in 2009, and the
Mag., vol. 56, no. 12, pp. 14–20, Dec. 2018. M.S. and Ph.D. degrees in information and com-
[9] A. L. Swindlehurst, E. Ayanoglu, P. Heydari, and F. Capolino, munication system from Hainan University, Haikou,
“Millimeter-wave massive MIMO: The next wireless revolution?” IEEE China, in 2013 and 2017, respectively. He has
Commun. Mag., vol. 52, no. 9, pp. 56–62, Sep. 2014. worked as a Network Engineer at China Mobile
[10] R. W. Heath, N. González-Prelcic, S. Rangan, W. Roh, and Jiangxi Branch for one year. He is currently a
A. M. Sayeed, “An overview of signal processing techniques for mil- Post-Doctoral Research Fellow with the City Uni-
limeter wave MIMO systems,” IEEE J. Sel. Topics Signal Process., versity of Macau. He is also an Associate Professor
vol. 10, no. 3, pp. 436–453, Apr. 2016. with Yichun University. He is the author of over
[11] W. H. Chin, Z. Fan, and R. Haines, “Emerging technologies and research 30 papers published in related international conference proceedings and
challenges for 5G wireless networks,” IEEE Wireless Commun., vol. 21, journals. His current research interests include intelligent communication,
no. 2, pp. 106–112, Apr. 2014. filter-bank multicarrier communications, and information theory.
[12] L. Wan, Y. Sun, L. Sun, Z. Ning, and J. J. P. C. Rodrigues, “Deep
learning based autonomous vehicle super resolution DOA estimation
for safety driving,” IEEE Trans. Intell. Transp. Syst., vol. 22, no. 7,
pp. 4301–4315, Jul. 2021.
[13] J. Lee, G.-T. Gil, and Y. H. Lee, “Channel estimation via orthogonal
matching pursuit for hybrid MIMO systems in millimeter wave com-
Pingping Xiao received the B.S. and M.S. degrees in
munications,” IEEE Trans. Commun., vol. 64, no. 6, pp. 2370–2386,
optical engineering from Jiangxi Normal University,
Jun. 2016.
China, in 2003 and 2006, respectively, and the
[14] X. Li, J. Fang, H. Li, and P. Wang, “Millimeter wave channel estima- Ph.D. degree from Shanghai Jiaotong University,
tion via exploiting joint sparse and low-rank structures,” IEEE Trans. China, in 2012. He is currently a Professor with
Wireless Commun., vol. 17, no. 2, pp. 1123–1133, Feb. 2018. Yinchun University. His current research interests
[15] Z. Y. Zhang et al., “1/2 -regularization-based super-resolution sparse include wireless communications and optical sensor
channel estimation for mmWave massive MIMO systems,” IEEE Access, applications.
vol. 7, pp. 75837–75844, 2019.
[16] B. Qi, W. Wang, and B. Wang, “Off-grid compressive channel estimation
for mm-wave massive MIMO with hybrid precoding,” IEEE Commun.
Lett., vol. 23, no. 1, pp. 108–111, Jan. 2019.
[17] X. D. Wu, G. H. Yang, F. Hou, and S. D. Ma, “Low-complexity
downlink channel estimation for millimeter-wave FDD massive MIMO
systems,” IEEE Wireless Commun. Lett., vol. 8, no. 4, pp. 1103–1107,
Mar. 2019.
[18] Y. You, L. Zhang, F. Hou, and S. D. Ma, “Bayesian matching Xingwang Li (Senior Member, IEEE) received the
pursuit-based channel estimation for millimeter wave communica- M.Sc. degree from the University of Electronic Sci-
tion,” IEEE Wireless Commun. Lett., vol. 24, no. 2, pp. 344–348, ence and Technology of China in 2010 and the Ph.D.
Nov. 2019. degree from the Beijing University of Posts and
[19] Z. Gao, C. Hu, L. Dai, and Z. Wang, “Channel estimation for millimeter- Telecommunications in 2015. From 2010 to 2012,
wave massive MIMO with hybrid precoding over frequency-selective he worked at Comba Telecom Ltd., Guangzhou,
fading channels,” IEEE Commun. Lett., vol. 20, no. 6, pp. 1259–1262, China, as an Engineer. From 2017 to 2018, he was
Jun. 2016. as a Visiting Scholar at Queen’s University Belfast,
[20] R. Zhang, J. Zhang, Y. Gao, and H. Zhao, “Distributed compressed Belfast, U.K. He is currently an Associated Professor
sensing-based multiuser channel estimation for frequency-selective with the School of Physics and Electronic Infor-
hybrid mmWave MIMO system,” Int. J. Commun. Syst., vol. 31, no. 18, mation Engineering, Henan Polytechnic University,
p. e3816, Dec. 2018. Jiaozuo, China. His research interests include MIMO communication, cooper-
[21] K. Venugopal, A. Alkhateeb, N. G. Prelcic, and R. W. Heath, “Channel ative communication, hardware constrained communication, NOMA, physical
estimation for hybrid architecture based wideband millimeter wave layer security, UAV communication, and the Internet of Things. He has
systems,” IEEE J. Sel. Areas Commun., vol. 35, no. 9, pp. 1996–2009, served as many TPC members, such as the IEEE Globecom, IEEE WCNC,
Sep. 2017. IEEE VTC, IEEE ICCC, and so on. He has also served as the Co-Chair for
[22] J. P. González-Coma, J. Rodríguez-Fernández, N. González-Prelcic, the IEEE/IET CSNDSP 2020 of the Green Communications and Networks
L. Castedo, and R. W. Heath, Jr., “Channel estimation and hybrid Track. He also serves as an Editor on the Editorial Board for IEEE A CCESS ,
precoding for frequency selective multiuser mmWave MIMO systems,” Computer Communications, Physical Communication, KSII Transaction on
IEEE J. Sel. Topics Signal Process., vol. 12, no. 2, pp. 353–367, Internet and Systems, and IET Quantum Communication. He is also the
May 2018. Lead Guest Editor of the Special Issue on UAV-enabled B5G/6G networks:
[23] J. Rodríguez-Fernández, N. González-Prelcic, K. Venugopal, and Emerging Trends and Challenges of Physical Communication, Special Issue
R. W. Heath, Jr., “Frequency-domain compressive channel estima- on Recent Advances in Physical Layer Technologies for the 5G-Enabled
tion for frequency-selective hybrid millimeter wave MIMO systems,” Internet of Things of Wireless Communications and Mobile Computing, and
IEEE Trans. Wireless Commun., vol. 17, no. 5, pp. 2946–2960, Special Issue on Recent Advances in Multiple Access for 5G-enabled IoT of
May 2018. Security and Communication Networks.

Authorized licensed use limited to: ULAKBIM UASL - ANKARA UNIVERSITESI. Downloaded on December 16,2024 at 18:46:06 UTC from IEEE Xplore. Restrictions apply.