sensors

Article
Deep Learning-Based Channel Estimation for mmWave
Massive MIMO Systems in Mixed-ADC Architecture
Rui Zhang 1 , Weiqiang Tan 1,∗ , Wenliang Nie 2 , Xianda Wu 3 and Ting Liu 4,5

1 School of Computer Science and Cyber Engineering, Guangzhou University, Guangzhou 510006, China;
[email protected]
2 School of Electronic and Information Engineering, Chongqing Three Gorges University,
Chongqing 404000, China; [email protected]
3 School of Electronics and Information Engineering, South China Normal University,
Foshan 528000, China; [email protected]
4 School of Artificial Intelligence, Nanjing University of Information Science and Technology,
Nanjing 210044, China; [email protected]
5 National Mobile Communications Research Laboratory, Southeast University, Nanjing 210096, China
* Correspondence: [email protected]

Abstract: Millimeter-wave (mmWave) massive multiple-input multiple-output (MIMO) systems

can significantly reduce the number of radio frequency (RF) chains by using lens antenna arrays,
because it is usually the case that the number of RF chains is often much smaller than the number of
antennas, so channel estimation becomes very challenging in practical wireless communication. In
this paper, we investigated channel estimation for mmWave massive MIMO system with lens antenna
array, in which we use a mixed (low/high) resolution analog-to-digital converter (ADC) architecture
to trade-off the power consumption and performance of the system. Specifically, most antennas
are equipped with low-resolution ADC and the rest of the antennas use high-resolution ADC. By
utilizing the sparsity of the mmWave channel, the beamspace channel estimation can be expressed
as a sparse signal recovery problem, and the channel can be recovered by the algorithm based on
Citation: Zhang, R.; Tan, W.; Nie, W.;
compressed sensing. We compare the traditional channel estimation scheme with the deep learning
Wu, X.; Liu, T. Deep Learning-Based
Channel Estimation for mmWave
channel-estimation scheme, which has a better advantage, such as that the estimation scheme based
Massive MIMO Systems in on deep neural network is significantly better than the traditional channel-estimation algorithm.
Mixed-ADC Architecture. Sensors
2022, 22, 3938. https://doi.org/ Keywords: millimeter-wave; massive MIMO; channel estimation; deep learning; mixed resolution
10.3390/s22103938 ADC; approximate message passing; compressed sensing

Academic Editor: Davy P. Gaillot

Received: 20 April 2022

Accepted: 19 May 2022 1. Introduction
Published: 23 May 2022
Millimeter-wave (mmWave) massive multiple-input multiple-output (MIMO) technol-
Publisher’s Note: MDPI stays neutral ogy can significantly improve data transmission rate with wider bandwidth and higher
with regard to jurisdictional claims in spectral efficiency, becoming one of the key technologies for the sixth generation (6G)
published maps and institutional affil- wireless communication [1]. However, digital beamforming and full-resolution ADCs are
iations. not suitable in mmWave massive MIMO system, due to the power consumption of ADCs
scales being exponential with the number of quantization bits, leading to high hardware
cost, power consumption and system complexity. Therefore, the use of current high-speed
and high-resolution ADCs (8–12 bits) for each antenna array would become a great burden
Copyright: © 2022 by the authors.
to the base station (BS) [2–4]. Consequently, the use of low-cost and low-resolution ADCs
Licensee MDPI, Basel, Switzerland.
(1–4 bits) is promoted as a potential solution to this problem.
This article is an open access article
distributed under the terms and 1.1. Related Works
conditions of the Creative Commons
In order to reap the maximum benefits of mmWave massive MIMO, the researchers
Attribution (CC BY) license (https://
creativecommons.org/licenses/by/
have proposed hybrid precoding, which involves a massive number of antennas attached
4.0/).
with a few RF chains [5,6], which effectively improved the energy efficiency of systems.

Sensors 2022, 22, 3938. https://doi.org/10.3390/s22103938 https://www.mdpi.com/journal/sensors

Sensors 2022, 22, 3938 2 of 16

In addition, the authors of [7] utilized the lens antenna array at the BS in the mmWave
massive MIMO system, in which signals from different directions can be concentrated on
different antennas, and the spatial channel can be converted into a beamspace channel.
Thus, the mmWave beamspace channel is sparse and there are only a few main propagation
path gains. By selecting a small number of main beams, the number of RF chains required
by the massive MIMO system can be significantly reduced [8,9]. Unfortunately, the beam
selection requires the BS to obtain the channel state information (CSI), and this is difficult to
achieve, especially when the number of RF links is limited. Furthermore, the performance
of hybrid precoding systems relies heavily on the precise control of the analog components,
and also the selection of the optimal beam will be more difficult if the beam width is small.
In parallel, there is a common solution to replace high-resolution ADCs with low-
resolution ADC (1–4 bits), which means the deployment of pure low-resolution ADCs at
the BS. However, low-resolution ADCs are prone to severe non-linear distortion, which
inevitably causes several problems, including high pilot overhead for channel estima-
tion [10], the system performance loss and signal detection [11]. In order to balance the
cost and performance of the system, a mmWave massive MIMO system with a mixed ADC
architecture was proposed in [12,13], which replaces the low-resolution ADCs with the
partial high-resolution ADCs on the original basis. As reported in [14], channel estimation
in a mixed-resolution ADC architecture is easier to process than in a pure low-resolution
ADC system. The authors of [15] derived a closed-form approximation of the reachable rate
of massive MIMO uplink under Rician fading channels with mixed-ADC. The work in [16]
analyzed the sum-rate performance of the multi-user massive MIMO relaying system
equipped with mixed-ADC architecture in the BS. Authors of [17] proposed the channel
estimation algorithm for uplinking massive MIMO systems with mixed-ADC architecture.
For ease of understanding, the advantages and limitations of the two solutions mentioned
above are summarized in Table 1.

Table 1. Multiple solutions to the problem of overwhelming hardware costs and power consumption
in mmWave massive MIMO system.

Solutions Advantages Limitations

Beam selection requires the BS
Significant reduction in the number
to obtain the CSI of beamspace,
Lens antenna arrays of RF chains by adopting a hybrid
and system performance is
precoding scheme.
vulnerable to bandwidth.
Reduce power consumption by
Low-resolution Low-resolution ADCs are prone
reducing the high resolution of the
ADCs to severe non-linear distortion.
ADCs to low resolution.

Different from using mixed-ADC architecture to reduce cost, another research direction
focuses on using the sparsity of the mmWave channel to estimate the channel by using some
classical schemes based on compressed sensing (CS) [18–21]. Specifically, the work in [18]
used the orthogonal matching pursuit (OMP) algorithm to detect the dominant entries of
multiple channel paths. The authors of [19] proposed a simultaneous weighted orthogonal
matching pursuit (SWOMP) channel estimation scheme. Furthermore, the authors of [20]
proposed stage wise orthogonal matching pursuit. Using the sparse characteristics of the
mmWave channel, the work in [21] designed a compressive sampling matching pursuit,
which aims to reduce system complexity. Unfortunately, the estimation accuracy of the
greedy algorithm is not ideal in the low signal-to-noise ratio (SNR) range. As a powerful
sparse signal recovery algorithm, the approximate message passing (AMP) algorithm was
proposed in [22], which can be used to estimate the beamspace channel, but it is difficult to
find the optimal solution for the shrinkage parameters of the algorithm.
Deep learning (DL) has achieved great success in speech recognition and image
processing. The advantages of DL are expected to bring changes to the communication
system. Compared with traditional methods, DL can reveal the internal characteristics of
end-to-end collected data or signals, so as to better solve various complicated problems
Sensors 2022, 22, 3938 3 of 16

encountered in wireless communication [23–27]. Inspired by the powerful learning ability

of deep neural networks (DNNs), some methods based on DL have been applied to
channel estimation and achieved good results. For example, the work in [28] applied to
channel estimation in the wireless downlink transmission system, which is superior to
the traditional scheme in estimation accuracy and energy acquisition. The use of neural
networks can help improve channel estimation performance was also demonstrated in [29],
where the method was designed by the structure of the minimum mean squared error
(MMSE) channel estimator. In addition, the authors of [30] proposed a framework based on
DL for direction-of-arrival estimation and the massive MIMO system. In [31], the authors
proposed an iterative channel estimation scheme, in which a denoising neural network
is used to update the estimated channel in each iteration. The authors in [32] proposed a
two-stage massive MIMO channel estimation process based on DL, including pilot-aided
and data-aided estimation stages. In [33], a deep neural network based on long short-
term memory (LSTM) was introduced to develop a more effective CSI feedback channel
compression and recovery method. To improve the estimation accuracy, the authors of [34]
developed a learned AMP (LAMP) network for channel estimation. Compared with the
original AMP algorithm, the estimation accuracy of both algorithms is improved.

1.2. Contributions
In this paper, we investigate channel estimation for mmWave massive MIMO system
with lens antenna array, in which we use a mixed (low/high) resolution ADC architec-
ture. To the best of our knowledge, there is no previous work exploring a mixed-ADC
architecture of mmWave massive MIMO system and how to estimate CSI based on the DL
method. We first study that the mixed-ADC architecture is practically useful because it
can provide performance comparable to the ideal high-resolution ADC architecture, while
reducing the complexity and power consumption of signal processing. Specifically, the
main contributions of this work can be summarized as follows:
• We study the DL-based channel estimation for a mmWave massive MIMO system with
mixed-ADC architecture, where most antennas are equipped with low-resolution ADC
and the rest of the antennas use high-resolution ADC. In terms of spectral efficiency,
the results showed that the mixed-ADC architecture is practically useful and it can
provide performance comparable to the ideal high-resolution ADC architecture.
• We compare the performance of traditional beamspace channel estimation algorithms
and DL-based schemes in mixed-ADC architecture. By applying the sparsity of the
mmWave channel, the DL-based channel estimation scheme performs significantly
better than the conventional channel estimation algorithm and its complexity outper-
forms the conventional estimation algorithm.
• We evaluate the effect of different quantization bits of low-resolution ADC in the
mixed-ADC architecture on the channel estimation and the sum rate. The results show
that when the number of quantization is about 4 bit, the estimation error with the
mixed-ADC architecture is small compared to that with the high-resolution ADC.
Notation: We use boldface letters to denote vectors and capitals to denote matrices. For
matrix A, Aij is the (i, j) th entry of A. A T , A H represent A’s transpose, complex conjugate.
Moreover, a circularly symmetric complex Gaussian random vector with zero mean and
covariance matrix R is denoted v ∼ CN (0, R), In is the identity matrix of size n, ⊗ is the
Kronecker product and k · k2 is the Euclidean norm.

2. System Model
We consider a time division duplex (TDD) mmWave massive MIMO system, in which
the BS is equipped with N antennas and NRF RF chains, serving K single antenna users
simultaneously In this section, we first introduce the traditional mmWave massive MIMO
channel and the space channel after adding the lens antenna array. In the second section,
we introduce the mixed resolution ADC architecture. Finally, we express the beamspace
channel estimation problem as the sparse signal recovery problem.
Sensors 2022, 22, 3938 4 of 16

2.1. Beamspace Channel Model

We start with the traditional massive MIMO mmWave channel, and in this work, we
adopt the Saleh–Valenzuela channel model, which is widely used in the frequency domain.
The channel vector with the size of N × 1 between the kth (k = 1, 2, . . . , K ) user and the BS
can be expressed as
s s
N Lk N Lk
Lk l∑ Lk l∑
hk = β k,l a ( φ , θ
k,l k,l ) = ck,l , (1)
=1 =1
where N is the total number of antennas, Lk is the number of resolvable paths and β k,l ,
φk,l and θk,l are the complex gain, azimuth and elevation on the lth path, respectively.
a(φk,l , θk,l ) is the N × 1 array steering vector, which depends on the array geometry, and
ck,l = β k,l a(φk,l , θk,l ) denotes the lth path component. For a typical uniform planar arrays
(UPAs) with N1 × N2 ( N = N1 × N2 ) antennas, the array steering vector can be expressed as

1 h i h i
a(φ, θ )= √ e− j2πd sin(φ) sin(θ )n1 /λ ⊗ e− j2πd cos(θ )n2 /λ , (2)
N
where n1 = [0, 1, · · · , N1 − 1] T and n2 = [0, 1, · · · , N2 − 1] T , λ is the wavelength of the
carrier, and d is the antenna spacing usually satisfying d = λ/2 in mmWave communica-
tions. Then, we can define ψa = d sin(φ) sin(θ )/λ and ψe = d cos(θ )/λ, respectively, as the
spatial angles for UPAs configuration [35].
Traditional channels in the spatial domain can be converted to beam spatial channels
by using a lens antenna array. In fact, the lens antenna array plays the role of the spatial
discrete Fourier transform (DFT) matrix U of size N × N. For UPA configuration, the matrix
U can be expressed as




 H
U = a(ψa,1 , ψe,1 ), · · · , a ψa,1 , ψe,N2 , · · · , a ψa,N1 , ψe,1 , · · · , a ψa,N1 , ψe,N2 , (3)
   
where ψa,n = N1 n − N12+1 for n = 1, 2, · · · , N1 and ψe,n = N12 n − N22+1 for
1
n = 1, 2, · · · , N2 are the spatial direction corresponding to the azimuth and elevation
predefined by the lens antenna array, respectively. Therefore, the beamspace channel vector
hk with a size of N × 1 between the kth user and BS with N antennas can be written as
r
N L
L l∑
hk = Uhk = ck,l (4)
=1

where ck,l = Uck,l represents the lth path component in the beamspace channel.

2.2. Mixed Resolution ADC Architecture

In this section, we consider a mixed resolution ADC architecture on the part of the
lens antenna array. We divide the antenna into two parts, where the N0 = σN antennas
are connected to the high-resolution ADC, and the N1 = (1 − σ ) N antenna is connected
to lower ADC. In order to facilitate calculation, the antenna number N0 and N1 in the
simulation are integers, and the coefficient σ is also limited to a certain rational number.
In this experiment, σ is set to 0.25, so we can redivide the channel matrix and rewrite the
above Formula (4) as
   
hk,0 hk,0
h̃k = =U , (5)
hk,1 hk,1

where hk,0 represents the channel matrix associated with N0 antennas connected to a
high resolution ADC, and hk,1 represents the channel matrix associated with N1 antennas
connected to a low resolution ADC.
Sensors 2022, 22, 3938 5 of 16

Due to the reciprocity of the TDD channel, after the pilot sequence is transmitted to BS,
the user can obtain the downlink channel through the estimated uplink channel. This paper
adopts the widely used orthogonal pilot transmission strategy, and the channel estimation
of each user is independent, so the subscript k in Formula (4) can be omitted.
Assuming that all users transmit known pilot sequences through instance Q for
channel estimation, the measured signal yq with a size of N × 1 passing through the RF
chains at the qth instant can be expressed as

yq = Aq h̃sq + nq , q = 1, 2, · · · , Q, (6)

where Aq is the NRF × N beam-selection network, sq is the pilot transmitted symbol at the
qth instant, nq = Aq nq is the effective noise vector, where nq ∼ CN 0, σn2 I N is the N × 1

noise vector with σn2 representing the noise power.

After the pilot transmission of Q instances, we can obtain the QNRF × 1 overall
measurement vector y by assuming sq = 1 for q = 1, 2, · · · , Q as
 T
y = y1 , y2 , . . . , yQ = Ah̃ + n, (7)
h iT
where A = A1T , A2T , · · · , AQ T is the QNRF × N overall combining matrix, and
h iT
n = n1T , n2T , · · · , nQ
T is the effective noise vector for Q instants.

2.3. Problem Formulation

According to Formula (7), we can now recover channel h̃ from y and A. Due to the
limited scattering at mmWave frequency, there are only a few propagation paths, and the
beam space channel h̃ is approximately sparse. This problem can be solved by the sparse
signal recovery algorithm in compressed sensing (CS), in which matrix A in Formula (7)
can be regarded as the sensing matrix in CS.

min kh̃k0 , s.t. ky − Ah̃k2 ≤ ε, (8)

where kh̃k0 is the number of non-zero elements of h̃, ε is the error tolerance parameter.
Because the l0 -norm minimization problem is a NP-hard problem in practice, so in
many cases, the l0 -norm optimization problem will be converted to a higher-dimension
norm problem, such as replacing l0 -norm with l1 -norm for the convex optimization problem.
At present, some traditional greedy algorithms are commonly used to solve this problem;
for example, the OMP [18] and compressive sampling matching pursuit (CoSaMP) [21].
However, using these greedy algorithms it is difficult to find the global optimal solution
and the estimation accuracy is not ideal.

3. Traditional Channel Estmation Algorithems

3.1. Compressed Sensing
Compressed sensing is a new technology for finding sparse solutions of underde-
termined linear systems. By using the characteristics of signal sparsity, compared with
Nyquist theory, the technology can restore the original signal to be recognized from fewer
measured values. The theory mainly includes three parts: signal-sparse representation,
reconstruction conditions and signal recovery algorithm. Suppose a signal x ∈ R N ×1 can
be sparsely represented by an orthonormal basis Ψ ∈ R N × N , i.e,

x = Ψs, (9)
Sensors 2022, 22, 3938 6 of 16

where s is a sparse column vector, the sparse coefficient of x is K and K is far less than N.
Consider a linear measurement process; x is represented by an orthonormal basis Ψ and a
coefficient vector s, and the measured value y ∈ R M×1 ( M << N ) can be written as

y = Φx = ΦΨs = As, (10)

where Φ represents the measurement matrix, and A = ΦΨ is the sensing matrix.

In general, taking the sparse of the reconstructed signal in a certain transform domain
as a priori information, the original signal is observed with the measurement matrix, and
the complete measured signal is reconstructed from the observed value combined with the
reconstruction algorithm.

3.2. OMP (Orthogonal Matching Pursuit)

Orthogonal Matching Pursuit is one of the classic algorithms in the field of compressed
sensing. It is the basis of many commonly used efficient algorithms. This algorithm has the
characteristics of simplicity and efficiency. The essence of the OMP algorithm is to select
the columns in the sensor matrix by greedy iteration so that the selected column is most
related to the current redundant vector in the process of each iteration, subtract the relevant
part from the original signal vector, and iterate repeatedly until the number of iterations
reaches the sparsity K, and then stop the iteration.
We define A ∈ C M× N as a sensing matrix and ai is the ith column of A, observation
vector y ∈ C M×1 and sparsity K, the initialization residual r0 = y. The initial iterated
index collection Λ0 = ∅, and initial value of iteration k = 1. To solve the first problem, we
attempt to look for an iterative index as follows:

λk = arg max |hrk−1 , ai i|. (11)

i =1,··· ,N

Adds the index of the most relevant dictionary element found to the index set, while
the set of reconstructed atoms in the sensing matrix is updated:

Λ k = Λ k −1 ∪ { λ k } , (12)

A Λ k = A Λ k −1 ∪ a Λ k , (13)
Update the residual:
  −1
T T
rk = y − AΛk AΛ k
A Λk AΛ k
y = y − AΛk sk . (14)

After executing K cycles, the reconstructed sparse coefficient sk can be obtained. For
more specific OMP algorithm flow, please refer to Algorithm 1.

3.3. AMP (Approximate Message Passing)

The approximate message passing (AMP) algorithm is an iterative compressed sensing
approach based on a probability graph to predict the next iteration through state evolution
and to de-noise through soft threshold iteration. In this section, we will describe how the
AMP algorithm estimates the beamspace channel, as shown in Algorithm 2.
In Algorithm 2, α in Step 6 is a tuning parameter and usually takes a fixed value during
iteration, the term bt vt−1 in Step 3 is called Onsager Correction, which is introduced into
the AMP algorithm to accelerate the convergence. The key step of the AMP algorithm is
step 6, in which the estimated channel ĥt+1 is obtained by soft threshold shrinking function
ηst during the tth iteration. The shrinkage function ηst is non-linear element-wise operation,
due to the sparsity of the channel, the channel vector ĥt+1 updated in each iteration will
Sensors 2022, 22, 3938 7 of 16

be more sparse. For the i th element rt,i = |rt,i |e jωt,i (i = 1, 2, · · · , N ) of input vector rt , we
have  
[ηst (rt ; α, λt )]i = ηst |rt,i |e jωt,i ; α, λt
(15)
= max(|rt,i | − αλt , 0)e jωt,i ,
where ωt,i is the phase of complex-valued element rt,i , α is the fixed parameter in the T
iteration, and λt is updated with iteration process in Step 5. In addition, b is obtained by
calculating the element-wise derivatives of the shrinkage function at the input vector r in
Step 7.

Algorithm 1 OMP Algorithm in mmWave Channels

Require:
sensing matrix A, measurement vector y, and the threshold δ
1: Λ0 = ∅, r0 = y, r−1 = 0, and k = 1
2
2: while k rk−1 − rk−2 k2 > δ and k ≤ k max do
3: λk = argmax A H (i ), rk−1
i =1,··· ,N
4: Λ k = Λ k −1 ∪ { λ k }
5: A Λ k = A Λ k −1 ∪ a Λ k
  −1
6: sk = AΛ T A T y
AΛ
k
Λ k k
7: rk = y − AΛk sk
8: k = k+1
9: end while
10: return sk

Algorithm 2 AMP Algorithm in mmWave Channels

Require:
sensing matrix A, measurement vector y, and the number of iterations T
1: Initialize ĥ0 = 0, v−1 = 0, and b0 = 0
2: for t = 0, 1, · · · , T − 1 do
3: vt = y − A ĥt + bt vt−1
4: rt = ĥt + A T vt
5: λt = √1 k vt k2
M
6: ĥt+1 = ηst (rt ; α, λt )
∂ηst (rt ;α,λt )
7: 1
bt + 1 = M ∑iN=1 ∂rt,i
8: end for
9: return ĥ = ĥ T .

Although the AMP algorithm can effectively deal with massive sparse signal problems,
and performs well in many traditional channel-estimation algorithm schemes, for sparse
beamspace channel estimation, many problems still exist, such as that the AMP algorithm
has a high requirement for i.i.d. sub-Gaussian matrix A. Otherwise, the algorithm itself is
prone to divergence. There are two key problems that restrict the performance of the AMP
algorithm: (1) The shrinkage parameters in the AMP algorithm usually take the same value
in the whole iterative process; (2) the general AMP algorithm cannot make full use of the
prior distribution of beamspace channels.

4. AI Channel Estmation Algorithems

4.1. Deep Learning
Deep learning is a kind of representation learning method based on data in machine
n o D
learning. Its basis is the neural network. In DL, training data yd , xd consisting of
d =1
feature and label pairs are used to train the parameters of the deep neural network, which
Sensors 2022, 22, 3938 8 of 16

aims to accurately predict the unknown label x associated with the newly acquired feature
y. Depth networks accept y and process it in many layers, each of which usually consists of
a linear transformation followed by a simple non-linear transformation. Unlike traditional
shallow learning, in-depth learning emphasizes the depth of the model structure, usually
with five, six or even ten layers of hidden nodes, and transforms the feature representation
of samples in the original space into a new feature space through layer-by-layer feature
transformation, which makes classification or prediction easier.
In general, the label space is discrete; for example, y is an audio or a picture; x is a class
of cat, dog, or some other type. However, for the sparse linear problem in beam space, label
x is continuous, but before that, some authors have demonstrated that a well-constructed
deep neural network can predict such tags accurately.

4.2. LAMP Network

Recently, the authors of [34] proposed a LAMP network scheme based on the classical
AMP algorithm. The results showed that the contraction parameters of the AMP algorithm
can only take the same value during the iterative process. Therefore, it essence is to map
each iteration of the AMP algorithm to each layer of the LAMP network and optimize the
non-linear parameters α in each iteration.

+LGGHQ 2XWSXW
,QSXW +LGGHQ +LGGHQ OD\HU OD\HU
OD\HU OD\HU OD\HU
T T

  
͘͘͘͘͘͘

͘͘͘͘͘͘
,QSXW 2XWSXW
͘͘͘

͘͘͘
͘͘͘
͘͘͘

  

T T

Figure 1. LAMP network structure with a total of T layers.

Figure 1 shows the network structure of the LAMP scheme with a total of T layers.
Specifically, the LAMP network processes signals in the same way as the AMP algorithm,
where the input of tth layer is y, ĥt and vt . It is worth mentioning that y is the measurement
signal, and both ĥt and vt are (t − 1)th layer outputs, so the processing of signals in tth
layer can be summarized as follows:

ĥt+1 = ηst (rt ; αt , λt ), (16)

vt+1 = y − Aĥt+1 + bt+1 vt , (17)
where
1
λ t = √ k v t k2 , (18)
M
rt = ĥt + B T vt , (19)
N
1 ∂ηst (rt ; α, λt )
bt + 1 =
M ∑ ∂rt,i
. (20)
i =1

From the above Formulas (16) and (19), we see the LAMP network involved in the
learnable parameters compared with Algorithm 2. In the tth iteration, operations involving
(α, A T ) are replaced by (αt , B T ), and the shrinkage function ηst of the AMP algorithm
plays a role in the non-linear activation function in the conventional DNN. It is worth
Sensors 2022, 22, 3938 9 of 16

noting that selecting A T in the AMP algorithm is only for the convenience of formula
derivation. If enough training data are given, the LAMP network can use DNN’s powerful
learning ability to find better shrinking parameters. Thus, the performance of the original
AMP algorithm can be further improved by optimizing the linear transformation coefficient
B T and non-linear shrinking parameter αt .

4.3. GM-LAMP Network

The LAMP network solves the problem of taking fixed shrinkage parameters in the
AMP algorithm, but it does not make good use of prior information on the beamspace
channel. The GM-LAMP network proposed by [36] solves this problem. The GM-LAMP
network derives a new shrinking function by considering the Gaussian mixing distribution
of the elements of the beamspace channel. Before that, we first introduce the expression of
the probability density function of the element h̃ in the beamspace channel:

Nc −1  
p(h̃; θ) = ∑ pk CN h̃; µk , σk2 (21)
k =0
n o
where θ = p0 , · · · , p Nc −1 , µ0 , · · · , µ Nc −1 , σ02 , · · · , σN
2
c −1
is a set of all the parameters,
pk is the probability of the kth Gaussian component and Nc is the number of Gaussian
components in the Gaussian mixture distribution. µk and σk2 represent the mean and
(h̃−µk )∗ (h̃−µk )
−
1 σ2
h̃; µk , σk2


variance of the kth Gaussian component, respectively. CN = πσk2
e k

denotes the probability density function of the k th Gaussian component.

It is interesting that when the mean and variance of Gaussian components are zero,
the probability density function of Gaussian distribution can be rewritten as

CN (h̃; 0, 0) = δ(h̃),

where the δ(h̃) is the Dirac delta function, which means that the variable h̃ will be exactly
zero. Therefore, the sparsity of beam space channel can be described as a special case
by using Gaussian mixture distribution. Finally, the Gaussian Mixed Shrinkage Function
considering the prior distribution of the beamspace channel can be written as
N −1
∑k=c 0 pk µ̃k (r )CN r; µk , σ2 + σk2
 

2
ηgm r; θ, σ = N −1
, (22)
∑k=c 0 pk CN r; µk , σ2 + σk2

where a set of all distribution parameters θ can also be called as the shrinkage parameters.
GM-LAMP network is still based on AMP algorithm, and it has T uniform layers, in
which the input and output of each layer are the same as that of LAMP network, and the
difference from LAMP network is that its soft threshold shrinkage function is replaced by
the Gaussian mixture shrinkage function. The channel estimation for the tth layer can be
written as
rt = ĥt + B T vt , (23)
 
ĥt+1 = ηgm rt ; θt , σ2 , (24)

where the linear transformation coefficient is B T and the non-linear shrinkage parameter θt
is the variable that can be optimized in the training stage.
The GM-LAMP network is mainly divided into two stages: offline training and online
estimation. During the offline training phase, a large amount of known training data
are provided to optimize the overall trainable variable by minimizing the loss function.
In the online estimation phase, the new measurement data can be input into the trained
GM-LAMP network and the corresponding channel estimates can be obtained directly.
Sensors 2022, 22, 3938 10 of 16

In the offline training stage, supervised learning is adopted to train the GM-LAMP
n oD
network, and the training dataset can be expressed as yd , h̃d , where yd is the input of
d =1
the GM-LAMP network, h̃d is the corresponding label and D represents the number of the
training data.
In order to avoid over-fitting, a layer-by-layer training method is adopted. Specif-
ically, the whole training process can be divided into T training subroutines according
to 0, 1, · · · , T − 1 sequence. For the tth training subprocess, its objective is to optimize
trainable variable Ωt = {Bi , θi }it=0 of the i = 0, · · · , i = tth layer. In this simulation, we
refer to the GM-LAMP algorithm that was first proposed in the literature [36]. In the model
training stage, the huber loss is used to define two loss functions of linear transformation
coefficient Bt and non-linear shrinkage parameter θt :
  
 1 (rd yd , Ωt − h̃d )2 V1 ≤ δ
2 t
Llinear
t (Ωt ) =   , (25)
δ|r y , Ωt − h̃ | − δ otherwise
d d d 1 2
t 2
  
 1 (ĥd y d, Ω − h̃d )2 V2 ≤ δ
2 t + 1 t
Lnon-linear
t (Ωt )=   , (26)
t+1 y , Ωt − h̃ | − 2 δ
δ|ĥd d d 1 2
otherwise
   
where V1 = |rdt yd , Ωt − h̃d |, V2 = |ĥdt+1 yd , Ωt − h̃d |, δ is a hyperparameter, rdt is
the output of linear transformation operation in Formula (23), and ĥdt+1 is the output of
non-linear contraction operation in Formula (24). Based on these two loss functions, the
training sub-process at the tth layer can be divided into linear training to minimize Llinear
t
and non-linear training to minimize Lnon-linear
t .
Algorithm 3 represents the specific layer-by-layer training method of the GM-LAMP
network that was provided in [36]. To avoid the trapped-in local optimization caused by
over-fitting, we firstly adopt the separate optimization method in step 2, and then jointly
optimize the method in step 3 and 4, which are set to B0 and θ0 , respectively. Then the
training is carried out in sequence from the 1th layer to the ( T − 1)th layer. For the training
sub-process of the tth layer, the training variable is set to the value of the training variable
of the (t − 1)th layer before training. Steps 7–8 indicate that the linear transformation
coefficient Bt is optimized separately, and then Bt and Ωt are optimized jointly. Similarly,
steps 9–10 indicate that the non-linear shrinkage parameter θt is optimized separately first,
and then Bt , Ωt and θt are optimized jointly.

Algorithm 3 Layer-by-Layer Training Method

1: Initialize B0 = A T and θ0
2: Learn B0 by minimizing Llinear t
3: Learn θ0 with fixed B0 by minimizing Lnon-linear
t
4: Relearn Ω0 = {B0 , θ0 } to minimize Lnon-linear
0
5: for t = 1, 2, · · · , T − 1 do
6: Bt = Bt−1 , θt = θt−1
7: Learn Bt with fixed Ωt−1 by minimizing Llinear
t
8: Relearn {Bt , Ωt−1 } by minimizing Llinear
t
9: Learn θt with fixed {Bt , Ωt−1 } by minimizing Lnon-linear
t
10: Relearn {Bt , Ωt−1 , θt } by minimizing Lnon-linear
t
11: end for
12: return Ω T −1

A trained GM-LAMP network can be obtained after optimizing the overall trainable
variables Ω T −1 of T layers. In the online estimation phase, the corresponding estimates
can be directly generated by inputting new measurement signals into the trained GM-
LAMP network.
Sensors 2022, 22, 3938 11 of 16

4.4. Complexity Analysis

In this subsection, we perform the complexity analysis of the channel estimation
algorithms used in the simulation process. Considering the OMP algorithm, when the
sparsity of the channel vector is set to K, the computational complexity in the atom selection
step is about O(KMN ), and the computational complexity of the k-th iteration is at least
O(K3 ). Therefore, we ignore trivial
operations; the computational complexity can be
calculated as O(KMN ) + O K3 M . In addition, since the LAMP network and GM-LAMP
network are constructed that both are based on the AMP algorithm, the computational
complexity of the AMP algorithm, LAMP network and GM-LAMP network is roughly the
same values that equals to O( TMN ).

5. Numerical Results
In this section, we compare the performance of traditional beamspace channel estima-
tion algorithms and DL-based schemes in mixed-ADC architecture. Specifically, DL-based
schemes include the LAMP network and the GM-LAMP network. This experiment provides
a widely used Saleh–Valenzuela channel model.

5.1. Parameter Setting

In our simulations, we consider that the BS is deployed a lens antenna array, where
the numbers of antennas and RF chains are set to N = 256 and NRF = 16, individually. The
number of single-antenna users is set to K = 16, and the number of measurements is set to
M = 128. For the the Saleh–Valenzuela channel model, we set the same channel parameters
for each user k, where the number of channel paths Lk = 3, the complex gain on the lth
path satisfies β k,l ∼ CN (0, 1) for l = 1, 2, 3, and the range of azimuth φk,l and elevation θk,l
on each path is between −π/2 and π/2. In order to train and test the LAMP network and
the GM-LAMP network, we generate 80,000, 2000 and 2000 samples as the training, the
validation and the testing set based on the above setup, respectively. Then the number of
training layers is set as T = 8 for both network schemes, where the number of nodes in each
layer depends on the number of measurements M and the number of dimensions N of the
beamspace channel. Finally, we use the normalized mean square error (NMSE) to quantify
the accuracy of channel estimation for each user, which is mathematically defined as
" #!
kĥ − h̃k2
NMSE = 10 log10 E , (27)
kh̃k2

where ĥ is the recovered channel matrix using the channel estimation algorithms.

5.2. Simulation Results on the Saleh–Valenzuela Channel Model

Figure 2 shows the NMSE performance comparison of various algorithms under the
considered UPA and mixed-resolution ADC quantization architecture. It can be seen from
the figure that the performance of the traditional two algorithms is poor, and in general,
the performance of the AMP algorithm is better than the OMP algorithm. For the DL-based
LAMP network and GM-LAMP network, the performance of the scheme has been greatly
improved. In addition, considering the prior distribution of beamspace channels, the
GM-LAMP network has better channel estimation accuracy than the LAMP network.
Figure 3 shows a comparison of the complexity of the above four kinds of channel
estimation schemes. We can find that the traditional OMP algorithm requires more complex
multipliers than other schemes, because the OMP algorithm is a kind of greed-tracking
algorithm, the number of complex multiplications increases exponentially with the number
of antennas, so the OMP algorithm has an excessive demand on the number of antennas.
In addition, thanks to the powerful learning of DNNs, the LAMP network and GM-LAMP
network can converge faster than the AMP algorithm. Therefore, the complex multiplier
required by the LAMP network and GM-LAMP network is smaller than that required by
the AMP algorithm.
Sensors 2022, 22, 3938 12 of 16

Next, we discuss the effects of resolution of ADC and channel sparsity on NMSE
performance for two schemes based on DL. Figure 4 shows the performance changes of
LAMP network schemes with different resolution ADC quantization. It can be seen that
lower estimation error can be obtained by using higher-resolution ADC quantization, and
the estimation performance after using mixed resolution ADC is better than four-bit ADC.
With the increase in SNR, the performance difference between mixed resolution ADC and
other low-resolution ADCs increases.
5
OMP
AMP
LAMP
0 GM-LAMP

-5
NMSE (dB)

-10

-15

-20
0 5 10 15 20
SNR (dB)
Figure 2. NMSE performance comparison of each channel estimation scheme for UPAs based on the
Saleh-Valenzuela channel model in the mixed resolution ADC architecture.

10 6
OMP
8 AMP
The number of complex multiplications

LAMP
7 GM-LAMP

0
32 80 128 176 224 272 320 368 416 464 512
The number of antennas N at the BS
Figure 3. The number of complex multiplications against the number of antennas N.

Figure 5 plots the variation of estimation error in low-resolution ADCs from one to
four bits and mixed resolution after adopting the GAMP network scheme. The sparsity in
this figure is equivalent to the number of channel paths (Lk ). With the increase in sparsity,
Sensors 2022, 22, 3938 13 of 16

the channel estimation performance of all quantization resolutions will decline. It can be
seen from the four curves in the comparison that the higher the resolution, the smaller the
estimation error.

-2
One-bit ADC
-4 Two-bit ADC
Perfect ADC
-6 Mixed-ADC

-8
NMSE (dB)

-10

-12

-14

-16

-18

-20

-22
0 2 4 6 8 10 12 14 16 18 20
SNR (dB)
Figure 4. Channel estimation error is a function of SNR. Variations of estimation errors in different
resolution ADCs quantification in LAMP network schemes are plotted.
-4

-6

-8

-10
NMSE (dB)

-12

-14

-16 One-bit ADC

Two-bit ADC
Three-bit ADC
-18
Four-bit ADC
Mixed-ADC
-20
2 3 4 5 6 7 8
The number of paths
Figure 5. Channel estimation error is a function of channel sparsity. Variations of estimation errors in
different resolution ADCs quantification in GAMP network schemes are plotted.

In addition, we evaluate the effects of NMSE and SNR on the total rate of beam
selection in beamspace channel estimation. In this simulation, we reference the parameters
of [37] to model the estimated beamspace channel Ĥ in case of imperfect CSI as
√
Ĥ = αH̃ + 1 − αE (28)

where α ∈ (0, 1) is the error parameter, E denotes the error matrix, whose elements satisfies
the independent and identical distribution of the zero means and variance error, i.e., CN (0,
Sensors 2022, 22, 3938 14 of 16

 
NMSE ), and H̃ = h̃1 , h̃2 , · · · , h̃K denotes the channel matrix with perfect CSI for K users.
In order to demonstrate the effectiveness of this work, we use the results provided in [37];
that is, AI beam selection with perfect CSI as the benchmark.
In Figure 6, we consider UPA based on the Saleh-Valenzuela channel model, and
compare the sum-rate between an imperfect CSI and a perfect CSI with a downlink SNR
of 10 dB. As shown in Figure 6, for the imperfect CSI under the mixed-resolution ADC
architecture, when the NMSE is about −23 dB, the total rate-loss due to beam selection
is less than 5% compared to the perfect CSI, and the beam-selection rate quantized by a
one-bit ADC is always quite different from that of the mixed-resolution ADC. Figure 7
shows the effect of the change of SNR on the total rate of beam selection for an imperfect
CSI and a perfect CSI when the NMSE is set to −10 dB. We can clearly see that with the
increase in SNR, the beam selection rate of all three increases, but the gap between the rate
of imperfect CSI and perfect CSI gradually increases.

80

70
Achievable sum-rate (bits/s/Hz)

60

50

40

30 UPA with perfect CSI, high ADC

UPA with imperfect CSI, mixed ADC
UPA with imperfect CSI, 1bit ADC
20
0 -5 -10 -15 -20 -25 -30
NMSE (dB)

Figure 6. Total rate for beam selection against different NMSE for the beamspace channel estimation.

100
UPA with perfect CSI, high ADC
UPA with imperfect CSI, mixed ADC
90 UPA with imperfect CSI, 1bit ADC
Achievable sum-rate (bits/s/Hz)

80

70

60

50

40
5 10 15 20 25
SNR(dB)

Figure 7. Total rate for beam selection against different SNR for the beamspace channel estimation.
Sensors 2022, 22, 3938 15 of 16

6. Conclusions
In this paper, the channel estimation for a mmWave massive MIMO system with
a mixed-resolution ADC architecture was investigated, where the BS deployed the lens
antenna arrays. By using the sparsity of the mmWave channel, the beamspace channel
estimation can be expressed as a sparse signal-recovery problem, and the channel can be
recovered by the algorithm based on compressed sensing. We compare the traditional
channel-estimation scheme with the channel-estimation scheme using DL. Simulation
results showed that the performance of the DNN-based estimation scheme is significantly
better than that of the traditional estimation scheme in the same configuration. Furthermore,
the performance gap between using mixed-resolution ADC and using high-resolution ADC
is slowly closing as the SNR increases. In future work, we will investigate the performance
of IRS-assisted mmWave massive MIMO systems with mixed-ADC architectures, which
aims to study the channel estimation problems in cascaded channels.

Author Contributions: W.T. conceived and designed the idea. R.Z. and X.W. performed the experi-
ments and analyzed the data. W.N. and T.L. gave valuable suggestions on the structuring of the paper
and assisted in the revising and proofreading. All authors have read and agreed to the published
version of the manuscript.
Funding: This work was supported by the Natural Science Foundation of Guangdong Province of
China under Grant 2020A1515010484, the Guangzhou Science and Technology Planning Project under
Grant 202102020879, the 2023 Guangzhou Basic Research Program Municipal School (College) Joint
Funding Project, the 2022 Joint Research Fund for Guangzhou University and Hong Kong University
of science and technology, the open research fund of National Mobile Communications Research
Laboratory, Southeast University under Grant 2021D03, the National Nature Science Foundation
of China under Grant 62101274, the Nature Science Foundation of Jiangsu Province under Grant
BK20210640, and Scientific Research Cultivation Project for Young Scholars of South China Normal
University under Grant 21KJ07.
Institutional Review Board Statement: Not applicable.
Informed Consent Statement: Not applicable.
Data Availability Statement: Not applicable.
Acknowledgments: The authors would like to extend their gratitude to the anonymous reviewers
and the editors for their valuable and constructive comments, which have greatly improved the
quality of this paper.
Conflicts of Interest: The authors declare no conflicts of interest.

References
1. Giordani, M.; Polese, M.; Mezzavilla, M.; Rangan, S.; Zorzi, M. Toward 6G networks: Use cases and technologies. IEEE Commun.
Mag. 2020, 58, 55–61. [CrossRef]
2. Fozi, M.; Sharafat, A.R.; Bennis, M. Fast MIMO Beamforming via Deep Reinforcement Learning for High Mobility mmWave
Connectivity. IEEE J. Sel. Areas Commun. 2022, 40, 127–142. [CrossRef]
3. Zhang, J.; Liu, J.; Ma, S.; Wen, C.K.; Jin, S. Large System Achievable Rate Analysis of RIS-Assisted MIMO Wireless Communication
With Statistical CSIT. IEEE Trans. Wirel. Commun. 2021, 20, 5572–5585. [CrossRef]
4. Chen, S.; Zhang, J.; Jin, Y.; Ai, B. Wireless powered IoE for 6G: Massive access meets scalable cell-free massive MIMO. China
Commun. 2020, 17, 92–109. [CrossRef]
5. Zeng, Y.; Zhang, R. Millimeter wave MIMO with lens antenna array: A new path division multiplexing paradigm. IEEE Trans.
Commun. 2016, 64, 1557–1571. [CrossRef]
6. Zeng, Y.; Zhang, R.; Chen, Z.N. Electromagnetic lens-focusing antenna enabled massive MIMO: Performance improvement and
cost reduction. IEEE J. Sel. Areas Commun. 2014, 32, 1194–1206. [CrossRef]
7. Brady, J.; Behdad, N.; Sayeed, A.M. Beamspace MIMO for millimeter-wave communications: System architecture, modeling,
analysis, and measurements. IEEE Trans. Antennas Propag. 2013, 61, 3814–3827. [CrossRef]
8. Gao, X.; Dai, L.; Chen, Z.; Wang, Z.; Zhang, Z. Near-optimal beam selection for beamspace mmWave massive MIMO systems.
IEEE Commun. Lett. 2016, 20, 1054–1057. [CrossRef]
9. Amadori, P.V.; Masouros, C. Low RF-complexity millimeter-wave beamspace-MIMO systems by beam selection. IEEE Trans.
Commun. 2015, 63, 2212–2223. [CrossRef]
Sensors 2022, 22, 3938 16 of 16

10. Mo, J.; Schniter, P.; Heath, R.W. Channel estimation in broadband millimeter wave MIMO systems with few-bit ADCs. IEEE
Trans. Signal Process. 2017, 66, 1141–1154. [CrossRef]
11. Zhang, J.; Dai, L.; Sun, S.; Wang, Z. On the spectral efficiency of massive MIMO systems with low-resolution ADCs. IEEE
Commun. Lett. 2016, 20, 842–845. [CrossRef]
12. Tan, W.; Li, S.; Zhou, M. Spectral and energy efficiency for uplink massive MIMO systems with mixed-ADC architecture. Phys.
Commun. 2022, 50, 101516. [CrossRef]
13. Zhou, M.; Zhang, Y.; Qiao, X.; Tan, W.; Yang, L. Analysis and optimization for downlink cell-free massive MIMO system with
mixed DACs. Sensors 2021, 21, 2624. [CrossRef] [PubMed]
14. Liang, N.; Zhang, W. Mixed-ADC massive MIMO uplink in frequency-selective channels. IEEE Trans. Commun. 2016,
64, 4652–4666. [CrossRef]
15. Zhang, J.; Dai, L.; He, Z.; Jin, S.; Li, X. Performance analysis of mixed-ADC massive MIMO systems over Rician fading channels.
IEEE J. Sel. Areas Commun. 2017, 35, 1327–1338. [CrossRef]
16. Liu, J.; Xu, J.; Xu, W.; Jin, S.; Dong, X. Multiuser massive MIMO relaying with mixed-ADC receiver. IEEE Signal Process. Lett.
2016, 24, 76–80. [CrossRef]
17. Yuan, J.; He, Q.; Matthaiou, M.; Quek, T.Q.; Jin, S. Toward massive connectivity for IoT in mixed-ADC distributed massive MIMO.
IEEE Internet Things J. 2019, 7, 1841–1856. [CrossRef]
18. Venugopal, K.; Alkhateeb, A.; Heath, R.W.; Prelcic, N.G. Time-domain channel estimation for wideband millimeter wave systems
with hybrid architecture. In Proceedings of the 2017 IEEE International Conference on Acoustics, Speech and Signal Processing
(ICASSP), New Orleans, LA, USA, 5–9 March 2017; pp. 6493–6497.
19. Rodríguez-Fernández, J.; González-Prelcic, N.; Venugopal, K.; Heath, R.W. Frequency-Domain Compressive Channel Estimation
for Frequency-Selective Hybrid Millimeter Wave MIMO Systems. IEEE Trans. Wirel. Commun. 2018, 17, 2946–2960. [CrossRef]
20. Lee, D. MIMO OFDM channel estimation via block stagewise orthogonal matching pursuit. IEEE Commun. Lett. 2016,
20, 2115–2118. [CrossRef]
21. Jayanthi, P.; Ravishankar, S. Sparse channel estimation for MIMO-OFDM systems using compressed sensing. In Proceedings of
the 2016 IEEE International Conference on Recent Trends in Electronics, Information & Communication Technology (RTEICT),
Bangalore, India, 20–21 May 2016; pp. 1060–1064.
22. Donoho, D.L.; Maleki, A.; Montanari, A. Message passing algorithms for compressed sensing: I. motivation and construction. In
Proceedings of the 2010 IEEE information theory workshop on information theory (ITW 2010, Cairo), Cairo, Egypt, 6–8 January
2010; pp. 1–5.
23. Qin, Z.; Ye, H.; Li, G.Y.; Juang, B.H.F. Deep learning in physical layer communications. IEEE Wirel. Commun. 2019, 26, 93–99.
[CrossRef]
24. O’shea, T.; Hoydis, J. An introduction to deep learning for the physical layer. IEEE Trans. Cogn. Commun. Netw. 2017, 3, 563–575.
[CrossRef]
25. Liang, F.; Shen, C.; Wu, F. An iterative BP-CNN architecture for channel decoding. IEEE J. Sel. Top. Signal Process. 2018,
12, 144–159. [CrossRef]
26. Kim, M.; Kim, N.I.; Lee, W.; Cho, D.H. Deep learning-aided SCMA. IEEE Commun. Lett. 2018, 22, 720–723. [CrossRef]
27. He, H.; Wen, C.K.; Jin, S.; Li, G.Y. Model-driven deep learning for MIMO detection. IEEE Trans. Signal Process. 2020, 68, 1702–1715.
[CrossRef]
28. Kang, J.M.; Chun, C.J.; Kim, I.M. Deep-learning-based channel estimation for wireless energy transfer. IEEE Commun. Lett. 2018,
22, 2310–2313. [CrossRef]
29. Neumann, D.; Wiese, T.; Utschick, W. Learning the MMSE channel estimator. IEEE Trans. Signal Process. 2018, 66, 2905–2917.
[CrossRef]
30. Huang, H.; Yang, J.; Huang, H.; Song, Y.; Gui, G. Deep learning for super-resolution channel estimation and DOA estimation
based massive MIMO system. IEEE Trans. Veh. Technol. 2018, 67, 8549–8560. [CrossRef]
31. He, H.; Wen, C.K.; Jin, S.; Li, G.Y. Deep learning-based channel estimation for beamspace mmWave massive MIMO systems.
IEEE Wirel. Commun. Lett. 2018, 7, 852–855. [CrossRef]
32. Chun, C.J.; Kang, J.M.; Kim, I.M. Deep learning-based channel estimation for massive MIMO systems. IEEE Wirel. Commun. Lett.
2019, 8, 1228–1231. [CrossRef]
33. Wang, T.; Wen, C.K.; Jin, S.; Li, G.Y. Deep learning-based CSI feedback approach for time-varying massive MIMO channels. IEEE
Wirel. Commun. Lett. 2018, 8, 416–419. [CrossRef]
34. Borgerding, M.; Schniter, P.; Rangan, S. AMP-inspired deep networks for sparse linear inverse problems. IEEE Trans. Signal
Process. 2017, 65, 4293–4308. [CrossRef]
35. Tan, W.; Yang, X.; Ma, S. Spectral Efficiency of Massive MIMO With LMMSE Receivers Under Finite Dimensional Channels. IEEE
Wirel. Commun. Lett. 2021, 11, 18–22. [CrossRef]
36. Wei, X.; Hu, C.; Dai, L. Deep learning for beamspace channel estimation in millimeter-wave massive MIMO systems. IEEE Trans.
Commun. 2020, 69, 182–193. [CrossRef]
37. Gao, X.; Dai, L.; Han, S.; Chih-Lin, I.; Heath, R.W. Energy-efficient hybrid analog and digital precoding for mmWave MIMO
systems with large antenna arrays. IEEE J. Sel. Areas Commun. 2016, 34, 998–1009. [CrossRef]