IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 69, NO.

8, AUGUST 2020 9217

Deep Learning and Compressive Sensing-Based CSI some approaches by using channel correlation and statistics have been
Feedback in FDD Massive MIMO Systems proposed. In [4], channel extrapolation utilizes correlations between
the downlink CSI and the uplink CSI in the FDD systems to reduce
Peizhe Liang, Jiancun Fan , Wenhan Shen, Zhijin Qin , the feedback overhead. In [5], the user statistical CSI is exploited to
and Geoffrey Ye Li design a statistical 3D downlink beamforming transmission scheme.
Compared with traditional multiuser MIMO transmission algorithms,
which require instantaneous CSI at the BS, this scheme can reduce
Abstract—To fully utilize multiplexing and array gains of massive
multiple-input multiple-output (MIMO), the downlink channel state in-
the channel estimation and feedback overhead greatly. However, the
formation (CSI) must be acquired at the base station (BS). In frequency extrapolated channel suffers from the estimation errors and the use
division duplexing (FDD) massive MIMO systems, the downlink CSI is of statistical CSI cannot achieve the same performance as the use of
generally estimated at the user equipment (UE) and then fed back to the instantaneous CSI.
BS. The huge number of antennas at the BS leads to overwhelming feedback To further reduce CSI feedback overhead and guarantee the accu-
overhead. To address this issue, we propose a framework, named CS-ReNet.
In this framework, the CSI is first compressed at the UE based on the racy of acquired CSI, CS has been widely applied for CSI feedback
compressive sensing (CS) technology and then reconstructed at the BS compression and recovery to reduce the overhead. For massive MIMO,
using a deep learning (DL)-based signal recovery solver, named ReNet. We the channel tends to be sparse due to the limited local scatters at the
analyze the CSI quality at the BS in terms of the normalized mean-squared BS compared with the huge number of transmit antennas [6], [7]. CS
error (NMSE) and cosine similarity. Simulation results demonstrate that the
proposed method outperforms the existing CS-based and some DL-based
can exploit the sparsity of the massive MIMO channel to acquire the
methods. CSI efficiently. In particular, CS can be used to reduce the overhead of
CSI feedback by compressing the CSI at the UE and recovering it at
Index Terms—Massive MIMO, FDD, CSI feedback, compressive sensing, the BS [8]. Inspired by this method, Bayesian CS has been used in [9]
deep learning.
to recover sparse channels in time domain, a joint orthogonal matching
pursuit has been proposed in [2] to enhance the CSI estimation quality
I. INTRODUCTION in the angular domain. As for an unknown statistical information of
the channel in OFDM system, [10] parameterizes the wideband MIMO
To maximize spectrum efficiency and minimize interference among channel by a limited number of distinct paths, each characterized by
different users in a massive multiple-input multiple-output (MIMO) path delay, path angle and path gain and a CS-based algorithm is
system, the base station (BS) needs accurate channel state information proposed to obtain the estimation of the path number and the latency.
(CSI) for precoding, adpative coding, multi-user scheduling, or other Many popular CS reconstruction algorithms, including approximate
operations [1]. In a time-division duplexing (TDD) massive MIMO message passing (AMP) [11], LASSO 1 -solver [12], TVAL3 [13],
system, the BS usually gets the downlink CSI by utilizing the channel BM3D-AMP [14] and OMP-US [15], need multiple iterations, thus
reciprocity while in a frequency division duplexing (FDD) massive their convergence speed is relatively slow.
MIMO system, users have to estimate the downlink CSI first and then As a state-of-the-art technology, deep learning (DL) plays an im-
feed it back to the BS [2]. When the number of antennas at the BS is portant role in solving the problems aforementioned. In recent years,
large, the overhead for CSI feedback in FDD massive MIMO systems intelligent communications have attracted intensive attention since ma-
could be huge. Since the deployed systems are mainly in the FDD chine learning (ML) techniques have solved many extremely complex
mode rather than the TDD mode, how to obtain the CSI at the BS scenarios [16]. As the most important part of ML, DL has been widely
becomes a challenging problem. To address this issue, vector quanti- investigated to tackle communication problems [17]–[22]. As for the
zation (VQ) and codebook-based approaches [3] have been proposed. channel estimation and CSI feedback, a DL-based scheme for real-time
In particular, the codebook size becomes huge if the high accuracy channel recovery leverages the power of the deep neural network (DNN)
CSI is desired at the BS. In order to reduce the feedback overhead, in training and prediction to reduce feedback overhead [23]. A DL
architecture, named CsiNet, uses an encoder to transform channel
Manuscript received January 29, 2020; revised May 18, 2020; accepted June matrices into codewords at the UE and a decoder to reconstruct the CSI
10, 2020. Date of publication June 25, 2020; date of current version August from these codewords at the BS [22]. Because the codewords are too
13, 2020. This work was supported in part by the National Natural Science short for some existing approaches, they cannot represent the complete
Foundation of China under Grant 61671367, in part by the Key Research features of channel matrices and lead to severe overfitting during the
and Development Plan of Shaanxi Province under Grant 2018GY-003, in part
by the Research Foundation of Science and Technology on Communication training process and low recovery quality of the CSI. A convolutional
Networks Laboratory, and in part by the Fundamental Research Funds for the long short-term memory network (ConvLSTM-net)-based DL method
Central Universities. The review of this article was coordinated by Dr. B. Mao. has been proposed to predict the downlink CSI from the uplink CSI [24].
(Corresponding author: Jiancun Fan.) The novel DL-based lightweight automatic modulation classification
Peizhe Liang and Jiancun Fan are with the School of Information and
Communication Engineering, Xi’an Jiaotong University, Xi’an 710049, China
(AMC) method in [18] is with smaller model sizes and faster compu-
(e-mail: [email protected]; [email protected]). tational speed.
Wenhan Shen and Zhijin Qin are with the School of Electronic Engineering To apply DL, the CSI matrices can be treated as two-dimensional
and Computer Science, Queen Mary University of London, London E1 4NS, images with real and imaginary parts, thus some DL-based image
U.K. (e-mail: [email protected]; [email protected]). processing methods, such as the convolutional neural networks (CNNs),
Geoffrey Ye Li is with the School of Electrical and Computer Engi-
neering, Georgia Institute of Technology, Atlanta, GA 30332 USA (e-mail: enable us to reconstruct the image from compressive CSI measurements
[email protected]). in a more efficient and accurate manner. However, these methods can
Digital Object Identifier 10.1109/TVT.2020.3004842 only be applied at the BS, since the training overhead is always huge

0018-9545 © 2020 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: MAULANA AZAD NATIONAL INSTITUTE OF TECHNOLOGY. Downloaded on February 05,2024 at 10:22:53 UTC from IEEE Xplore. Restrictions apply.
 9218 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 69, NO. 8, AUGUST 2020

due to a large amount of parameters for DL model. Moreover, it is very we will reduce the channel feedback overhead with the aid of CS and
challenging to deploy the DL-based methods at the UEs, which are DL technologies.
equipped with the limited device memories and computation power. We assume a uniform linear array (ULA) model in the massive
In order to take advantage of deep learning and tackle the problems MIMO system. Using the representation in the angular-delay domain,
like the limited device memories at the UE, in this paper, we propose the channel matrices can be described as
a two-step approach based on DL, named CS-ReNet, to compress
and reconstruct the CSI. In this framework, channel matrices are first Ha = Ad HAH
a , (2)
compressed by using random projection at the UE, and then the com-
pressed measurements are fed back to the BS where channel matrices where Ad and AH a are Nc × Nc and NBS × NBS unitary matrices,

are reconstructed using a trained model, called ReNet. We introduce respectively, and Ha denotes the channel matrix in size of Nc × NBS
the spatial pyramid pooling (SPP) layer in ReNet, which can generate over the angular-delay domain. The channel matrix in the angular-delay
fixed-dimensional vectors regardless of the input size [25]. In other domain is sparse, that is, with only a few non-zero coefficients. In the
words, the SPP-net enables our model to accept arbitrary input sizes delay domain, only the first Ñc rows of Ha may be non-zero and the
and use multi-size training that can reduce the overfitting and improve rest rows are all zeros because the multipath delay lies within a limited
the reconstruction accuracy of the CSI. The main contribution of this period. Therefore, we just retain the first Ñc rows of Ha and omit the
paper is summarized as follows: others. In this case, we use H̃a to denote the Ñc × NBS truncated
1) We develop a framework for CSI feedback, named CS-ReNet. In matrix.
this framework, we first compress the CSI at the UE and then use
a trained model, which is a DL-based CS-reconstruction model, III. CHANNEL COMPRESSION AND RECONSTRUCTION
named ReNet, to reconstruct the CSI at the BS. In this section, we will propose the framework of CS-ReNet, which
2) At the UE side, we adopt the CS technology to compress the CSI. includes two steps. In the first step, we adopt the compressive sensing
In this stage, we only need to generate a Gaussian random matrix to compress the CSI at the UE. In the second step, the compressed CSI
to compress the initial CSI to eliminate the special requirements is first fed back to the BS and then is used as the input to the ReNet, a
for the storage space and communication scenario of UEs com- well-trained model, to reconstruct the CSI at the BS.
pared with some existing approaches, such as autoencoder-based
one.
3) At the BS side, we utilize ReNet to recovery the CSI. In this stage, A. CSI Compression
we first use a seven-layer convolutional network to extract the The CSI compression problem based on CS at the UE can be
features of the compressed measurements and then use the SPP- summarized as
net for multi-size training to reduce the overfitting and perform
fine-tuning for further boosting performance of the network. By Y = ΦX, (3)
using the ReNet, the CSI can be accurately recovered and the
great performance is achieved even at excessively low compres- where Y and Φ are the CS measurements and the measurement
sion ratio (CR) where the other methods fail or perform badly. matrix, respectively, and X is the original signal. For the proposed
4) We show the performance gain of the proposed method over CSI compression scheme shown in Fig. 1(a), the original signal, X,
the existing ones by experiments with the indoor picocellular can be obtained from the estimated channel matrix, H̃a , which is
scenario of COST 2100 channel model under different CRs. calculated from the perfect CSI. We assume that perfect CSI is avail-
The rest of this paper is organized as follows. Section II provides the able through pilot-based training [26]. Once we obtain the estimated
system model. Section III presents the method on compressing channel channel matrix H̃a , we first add 0.5 to the real and imaginary parts
information and the structure of the proposed DL-based model for of H̃a and then reshape its real and imaginary parts into a 2 Ñ × 1
channel reconstruction. In Section IV, simulation results are presented real-valued vector X, where N = Ñc NBS . The detailed process is
and finally Section V concludes this paper. as shown in Fig. 1(a). After getting the vector X, a Gaussian ran-
dom matrix Φ is multiplied to X to get the compressed feedback
II. CHANNEL MODEL data Y = ΦX. In this processing, X ∈C 2 Ñ ×1 , Φ∈C M ×2 Ñ , the CR
We consider a massive MIMO orthogonal division multiplexing M/(2 Ñ ) ∈ {1/4, 1/8, 1/16, 1/32, 1/64}, and Y ∈C M ×1 is the com-
(OFDM) system operating in the FDD mode. There are NBS antennas pressed CSI that needs to be fed back to the BS once obtained at the
equipped at the BS and there are K users, each with the single antenna UE side, where M is the dimension of the feedback CSI and is decided
and working at non-overlapped frequency band. In the frequency do- by the CR. Obviously, M is is much less than 2 Ñ , hence the feedback
main, the received signal at the nth subcarrier for the kth user can be overhead could be reduced.
expressed as
ykn = hH kn wkn skn + vkn , (1) B. Channel Reconstruction With the Proposed
where hkn ∈ C NBS ×1 , wkn ∈ C NBS ×1 , skn ∈ C, and vkn ∈ C de-
DL-Based Model
note the channel vector, precoding vector, data symbol in the fre- In this section, we exploit the popular convolutional neural networks
quency domain, and additive noise over the nth subcarrier, respectively. (CNNs) to reconstruct the CSI. However, the prevalent CNNs require a
For simplicity, we omit subscript k in the subsequent discussion, fixed input size and limit the scale of the input size, we therefore intro-
thus the channel matrix in the frequency domain can be expressed duce a SPP-net between the last convolution layer and fully-connected
as H = [h1 , h2 , ..., hNc ]H ∈ C Nc ×NBS , where Nc is the number of layer to generate a fixed-dimensional outputs but with flexible input
subcarriers. In this case, there are NBS × Nc feedback elements for sizes. To further improve its performance, we add the fine-tuning,
each user if no compression is performed. For the massive MIMO in which parameters of a model are adjusted in order to fit certain
OFDM system with a large number of antennas and subcarriers, the observations. The overview of our developed model, named ReNet, is
feedback overhead consumes huge spectrum resource. In this paper, illustrated in Fig. 1(b). The input of the first convolutional layer is the

Authorized licensed use limited to: MAULANA AZAD NATIONAL INSTITUTE OF TECHNOLOGY. Downloaded on February 05,2024 at 10:22:53 UTC from IEEE Xplore. Restrictions apply.
 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 69, NO. 8, AUGUST 2020 9219

Fig. 1. The CS-ReNet framework for CSI feedback. (a) The process of compressing the CSI at the UE. (b) ReNet Model for CS reconstruction.

TABLE I add it to the output of the fourth convolutional layer. Once the channel
SIZES OF 2D MATRICES AT DIFFERENT CRS matrix has been refined by all Resblock, the channel matrix is inputted
into the final convolutional layer, and the sigmoid function is used to
scale values to the [0, 1] range. The output of the last convolutional
layer is the final reconstruction of X. Finally, we transform X to H̃a
linearly. Experiments show that two Resblocks perform better since
further increasing Resblock does not improve the reconstruction quality
but increase the computational complexity.
Thus, the training procedure could be summarized into two steps as
following:
real and imaginary parts of CS measurements obtained at the BS and 1) At the begining, we perform the training with inputs obtained at
the output is the reconstructed X. In order to expand the receptive field various CRs to train the ReNet without fine-tune to generate the
without excessively increasing the training time, our model contains initial estimation and save the model as a pre-trained model as
seven convolutional layers, each uses kernels with the dimension of shown in Fig. 1(b).
3 × 3. The Leaky rectified linear unit (LeakeyReLU) is used as the 2) Then we use single-size training to train the whole ReNet with
activation function and we add batch normalizaion to each layer. the given pre-trained model and generate the final results. In this
Inspired by [25], we add the SPP-net between the last convolutional training process, we only use size1 shown in Table I to train
layer and the fully-connected layer to train the model with different ReNet.
input sizes. It means that this is a multi-size training. By using this multi- In this training, the adaptive moment estimation (ADAM) algorithm
size training, we can reduce overfitting so that the CSI recovery accuracy and the loss function in terms of the mean-squared error (MSE) are
can be improved. For the multi-size training, we consider a set of pre- exploited. The MSE is calculated as follows:
defined sizes. As shown in Fig. 1(b), we can obtain several different
1
T
compressed measurements at each CR at the BS and crop a small region
L(Θ) = f (Y; Θ) − X22 , (4)
of each compressed measurement at the same CR. Then we reshape T i=1
them into 2D-matrices as multi-size inputs and their size is shown in
Table I. Following the last convolutional layer, SPP-net generates fixed- where .2 is the Euclidean norm, Θ is the set of parameters, T is
length representation of feature maps. The representation is then fed the total number of samples in the training set, and f (Y; Θ) is the
into a fully-connected layer to produce a vector and reshaped into size output of the trained model. At the BS, the general CS techniques
Ñc × NBS × 2. This reshaped matrix is used as the initial estimation exploit the reconstruction algorithm like orthogonal matching pursuit
of X. (OMP) to obtain H̃a from Y and Φ. In this paper, we exploit ReNet to
For multi-size training, we use the combination of seven convo- obtain H̃a from Y and then use the inverse discrete Fourier transform
lutional layers, SPP-net, and a fully-connected layer as one single (IDFT) to acquire the original channel matrix. Compared with other
network, which accepts inputs with different sizes. In each epoch, we conventional CS-based algorithms, the proposed reconstruction method
first train the network with a given input size, i.e., a small number of based on ReNet does not require iteration so it can reduce computational
compressed measurements (size2), and then switch it to another input complexity.
size (size1). Once the stable output results are obtained, the multi-size
training will be stopped.
To produce more precise results, we introduce residual-learning
IV. SIMULATION RESULTS
based networks [27], named Resblock, to learn the difference between We use the original data set in [22], which are generated by the
the initial estimation and the original X and then generate the final COST 2100 channel model [29] in the indoor picocellular scenario at
estimation. This method is inspired by fine-tuning (details in [28]). Each 5.3 GHz band, to compare the proposed method with the other ones
Resblock consists of four convolutional layers, all with 3 × 3 kernels. fairly. In this scenario, the BS at the center of a square has a length of
The second, third and fourth layer generate 8, 16 and 2 feature maps, 20 m for an indoor environment. The UEs randomly locate in the square
respectively. We consider the input of the Resblock as a shortcut and area. The BS is equipped with Nt = 32 antennas and there are Nc =

Authorized licensed use limited to: MAULANA AZAD NATIONAL INSTITUTE OF TECHNOLOGY. Downloaded on February 05,2024 at 10:22:53 UTC from IEEE Xplore. Restrictions apply.
 9220 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 69, NO. 8, AUGUST 2020

1024 subcarriers in this system. In addtion, the initial channel matrices TABLE II
are transformed in the angular-delay domain and the first 32 rows are PERFORMANCE COMPARISON BETWEEN CS RECONSTRUCTION ALGORITHMS
retained. Then the magnitude of channel matrix H̃a is transformed into
a 32 × 32 × 2 matrix with real and imaginary parts.
Keras and Tensorflow with a GPU backend are used to implement our
proposed scheme. In the offline stage, the loss function of the network is
measured by MSE and the exploited optimization algorithm is ADAM.
For the multi-size training, the batch size is 200 with the maximum
epochs of 800. And for the single-size training, the batch size is 200
with the maximum epochs of 150. The training, validation and testing
sets contains 130,000, 10,000 and 10,000 samples, respectively.
The recovery performance is quantified by normalized MSE (NMSE)
and cosine similarity, which is shown as following
 2 
  2
NMSE = E H − H  / H2 , (5)
2
⎧ ⎫
⎨ 1 Nc hn h H ⎬
n
ρ=E  
⎩ Nc   ⎭ , (6)
n=1 hn 2 hn 
2

where H and H  denote the original and the recovered channel matrices,
respectively, and  hn is the reconstructed channel vector at the nth
subcarrier. We use the NMSE to measure the difference between H 
and H. Moreover, we consider the cosine similarity to measure the
precoding performance. As the CSI feedback serves as a precoding
vector. If wn = h  n /h
 n  is used as a precoding vector, then we can
2
obtain the equivalent channel hH  
n hn /hn 2 at the UE side. The output
of CS-ReNet is the normalized channel matrix, therefore, the closer the
cosine similarity approaches 1, the better precoding performance is.
We compare the CSI reconstruction accuracy for the proposed re-
TABLE III
construction method with three state-of-the-art CS methods i.e. LASSO PERFORMANCE COMPARISON OF CHANNEL FEEDBACK SYSTEM BASED ON
1 -solver [12], TVAL3 [13], and BM3D-AMP [14]. We also compare DEEP LEARNING
the proposed ReNet with CS-CsiNet [22], which only uses several
convolutional layers learning to recover the CSI from the compressed
measurements. In this procedure, we use the same algorithm to com-
press CSI. In Table II, we compare the NMSE and cosine similarity of
the reconstructed CSI under four different CRs. Table II shows that the
proposed ReNet achieves a better performance than other methods with
the lowest NMSE and gets the maximum value in cosine similarity. This
means that the proposed CS-ReNet can recuce the overhead compared
to the conventional methodes when targeting the same recovery accu-
racy. As the CR recovery accuracy increases, the performance of ReNet
is better and the difference between ReNet and the benchmarks becomes
more significant. That is because with the increase of the CR, more
original information can be retained by the compressed measurements.
In addition, the speed of ReNet reconstruction is much faster than other
methods because ReNet does not need iteration. the code rate to 1/3. In Figs. 2–4, the BER with respect to Eb /N0
We also compare the proposed framework for CSI feedback with is illustrated when the CR is 1/4, 1/16, and 1/32, respectively. Three
CsiNet [22]. CsiNet is closely to autoencoder and can generate code- traditional state-of-the-art methodes in [12], [13] and [14] and two
words as representation of channel matrix and recover the CSI from DL-based [22] are presented for comparison. Note that the performance
these codewords. The two frameworks are both DL-based. All corre- of precoding is strongly associated with the cosine similarity. As
sponding NMSE and cosine similarity are summarized in Table III. shown in Table II, when the CR increases, the cosine similarity of
From Table III, the proposed framework outperforms the CsiNet at all methods increases and BER of all methods becomes on the order of
high CR like 1/4, 1/8, 1/16 and 1/32. When the enviroment changes, magnitudes better. Moreover, CS-ReNet-based precoding outperforms
the proposed CS-ReNet only needs to train deep neural network at the other schemes, in all cases, because the value of cosine similarity is
the BS while the UEs with limited storage space only need to save largest.
limited parameters to generate five Gaussian random matrice, i.e. five
measurement matrice and do not require any update.
V. CONCLUSION
Furthermore, we present the precoding performance of the estimated
CSI. The data symbols are drawn independently from a binary quadra- In this paper, we have proposed a new CSI feedback framework,
ture amplitude modulation (4QAM) constellation. We employ turbo named CS-ReNet, which can significantly reduce the feedback over-
code for channel coding. We set the packet length to 678 bits and head and lower the complexity of implementing CS at the UE. In this

Authorized licensed use limited to: MAULANA AZAD NATIONAL INSTITUTE OF TECHNOLOGY. Downloaded on February 05,2024 at 10:22:53 UTC from IEEE Xplore. Restrictions apply.
 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 69, NO. 8, AUGUST 2020 9221

framework, we exploited the DNN and SPP-net to bulid a novel CSI

recovery model, named ReNet, which can accurately recover the CSI
and achieve great performance with low complexity even at excessively
low CRs. The stimulation results have verified the performance gains
achieved by the proposed CS-ReNet in terms of recovery accuracy.

REFERENCES
[1] Q. Shi, M. Razaviyayn, Z. Luo, and C. He, “An iteratively weighted
MMSE approach to distributed sum-utility maximization for a MIMO
interfering broadcast channel,” IEEE Trans. Signal Process., vol. 59, no. 9,
pp. 4331–4340, Sep. 2011.
[2] X. Rao and V. K. N. Lau, “Distributed compressive CSIT estimation and
feedback for FDD multi-user massive MIMO systems,” IEEE Trans. Signal
Process., vol. 62, no. 12, pp. 3261–3271, Jun. 2014.
[3] Q. Sun, D. C. Cox, H. C. Huang, and A. Lozano, “Estimation of continuous
flat fading MIMO channels,” IEEE Trans. Wireless Commun., vol. 1, no. 4,
pp. 549–553, Oct. 2002.
[4] F. Rottenberg, R. Wang, J. Zhang, and A. F. Molisch, “Channel ex-
Fig. 2. BER performance comparison between the proposed scheme and the trapolation in FDD massive MIMO: Theoretical analysis and numerical
other schemes when CR is 1/4. validation,” 2019, arXiv:1902.06844. [Online]. Available: http://arxiv.org/
abs/1902.06844
[5] X. Li, S. Jin, H. A. Suraweera, J. Hou, and X. Gao, “Statistical 3-D
beamforming for large-scale MIMO downlink systems over rician fad-
ing channels,” IEEE Trans. Commun., vol. 64, no. 4, pp. 1529–1543,
Apr. 2016.
[6] F. Kaltenberger, D. Gesbert, R. Knopp, and M. Kountouris, “Correlation
and capacity of measured multi-user MIMO channels,” in Proc. IEEE 19th
Int. Symp. Pers., Ind. Mob. Rad. Commun., Sep. 2008, pp. 1–5.
[7] J. Hoydis, C. Hoek, T. Wild, and S. ten Brink, “Channel measurements
for large antenna arrays,” in Proc. Int. Symp. Wireless Commun. Syst.,
Aug. 2012, pp. 811–815.
[8] Z. Qin, J. Fan, Y. Liu, Y. Gao, and G. Y. Li, “Sparse representation for
wireless communications: A compressive sensing approach,” IEEE Signal
Process. Mag., vol. 35, no. 3, pp. 40–58, May 2018.
[9] M. Masood, L. H. Afify, and T. Y. Al-Naffouri, “Efficient coordinated
recovery of sparse channels in massive MIMO,” IEEE Trans. Signal
Process., vol. 63, no. 1, pp. 104–118, Jan. 2015.
[10] D. Hu and L. He, “Channel estimation for FDD massive MIMO OFDM
systems,” in Proc. IEEE 86th Veh. Technol. Conf., 2017, pp. 1–5.
[11] B. C. Jung and W. Chang, “A message passing algorithm for compressed
sensing in wireless random access networks,” in Proc. 19th Asia-Pacific
Conf. Commun., Aug. 2013, pp. 463–464.
[12] I. Daubechies, M. Defrise, and C. De Mol, “An iterative thresholding
algorithm for linear inverse problems with a sparsity constraint,” Commun.
Fig. 3. BER performance comparison between the proposed scheme and the Pure Appl. Math.: A J. Issued Courant Inst. Math. Sci., vol. 57, no. 11,
other schemes when CR is 1/16. pp. 1413–1457, 2004.
[13] C. Li, W. Yin, and Y. Zhang, “User’s guide for TVAL3: TV minimization
by augmented lagrangian and alternating direction algorithms,” [Online].
Available: http://www.caam.rice.edu/∼optimization/L1/TVAL3/
[14] C. A. Metzler, A. Maleki, and R. G. Baraniuk, “From denoising to com-
pressed sensing,” IEEE Trans. Inf. Theory, vol. 62, no. 9, pp. 5117–5144,
Sep. 2016.
[15] M. J. Azizipour and K. Mohamed-Pour, “Compressed channel estimation
for FDD massive MIMO systems without prior knowledge of sparse
channel model,” IET Commun., vol. 13, no. 6, pp. 657–663, 2019.
[16] N. Kato, B. Mao, F. Tang, Y. Kawamoto, and J. Liu, “Ten challenges
in advancing machine learning technologies toward 6G,” IEEE Wireless
Commun., vol. 27, no. 3, pp. 96–103, Jun. 2020.
[17] B. Mao, F. Tang, Z. M. Fadlullah, and N. Kato, “An intelligent route com-
putation approach based on real-time deep learning strategy for software
defined communication systems,” IEEE Trans. Emerg. Topi. Comput., to
be published, doi: 10.1109/TETC.2019.2899407.
[18] Y. Wang, J. Yang, M. Liu, and G. Gui, “LightAMC: Lightweight automatic
modulation classification via deep learning and compressive sensing,”
IEEE Trans. Veh. Technol., vol. 69, no. 3, pp. 3491–3495, Mar. 2020.
[19] Z. Qin, H. Ye, G. Y. Li, and B. F. Juang, “Deep learning in physical layer
communications,” IEEE Wireless Commun., vol. 26, no. 2, pp. 93–99,
Apr. 2019.
[20] H. Ye, G. Y. Li, and B. Juang, “Power of deep learning for channel esti-
Fig. 4. BER performance comparison between the proposed scheme and the
mation and signal detection in OFDM systems,” IEEE Wireless Commun.
other schemes when CR is 1/32.
Lett., vol. 7, no. 1, pp. 114–117, Feb. 2018.

Authorized licensed use limited to: MAULANA AZAD NATIONAL INSTITUTE OF TECHNOLOGY. Downloaded on February 05,2024 at 10:22:53 UTC from IEEE Xplore. Restrictions apply.
 9222 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 69, NO. 8, AUGUST 2020

[21] P. N. Alevizos, X. Fu, N. D. Sidiropoulos, Y. Yang, and A. Bletsas, “Lim- [25] K. He, X. Zhang, S. Ren, and J. Sun, “Spatial pyramid pooling in deep
ited feedback channel estimation in massive MIMO with non-uniform convolutional networks for visual recognition,” IEEE Trans. Pattern Anal.
directional dictionaries,” IEEE Trans. Signal Process., vol. 66, no. 19, Mach. Intell., vol. 37, no. 9, pp. 1904–1916, Sep. 2015.
pp. 5127–5141, Oct. 2018. [26] J. Choi, D. J. Love, and P. Bidigare, “Downlink training techniques for
[22] C. Wen, W. Shih, and S. Jin, “Deep learning for massive MIMO CSI FDD massive MIMO systems: Open-loop and closed-loop training with
feedback,” IEEE Wireless Commun. Lett., vol. 7, no. 5, pp. 748–751, memory,” IEEE J. Sel. Topics Signal Process., vol. 8, no. 5, pp. 802–814,
Oct. 2018. Oct. 2014.
[23] H. Sun, Z. Zhao, X. Fu, and M. Hong, “Limited feedback double direc- [27] K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning
tional massive MIMO channel estimation: From low-rank modeling to for image recognition,” in Proc. IEEE Conf. Comput. Vision Patt.
deep learning,” in Proc. IEEE Workshop Signal Process. Adv. Wireless Recognit., Jun. 2016, pp. 770–778. [Online]. Available: https://doi.
Commun., 2018, pp. 1–5. ieeecomputersociety.org/10.1109/CVPR.2016.90
[24] J. Wang, Y. Ding, S. Bian, Y. Peng, M. Liu, and G. Gui, “UL-CSI data [28] Y. Guo, H. Shi, A. Kumar, K. Grauman, T. Rosing, and R. Feris, “Spottune:
driven deep learning for predicting DL-CSI in cellular FDD systems,” Transfer learning through adaptive fine-tuning,” in Proc. IEEE Conf.
IEEE Accel., vol. 7, pp. 96 105–96 112, 2019. Comput. Vision Pattern Recognit., 2019, pp. 4805–4814.
[29] L. Liu et al., “The COST 2100 MIMO channel model,” IEEE Wireless
Commun., vol. 19, no. 6, pp. 92–99, Dec. 2012.

Authorized licensed use limited to: MAULANA AZAD NATIONAL INSTITUTE OF TECHNOLOGY. Downloaded on February 05,2024 at 10:22:53 UTC from IEEE Xplore. Restrictions apply.