International Journal of Reconfigurable and Embedded Systems (IJRES)

Vol. 13, No. 2, July 2024, pp. 271~277

ISSN: 2089-4864, DOI: 10.11591/ijres.v13.i2.pp271-277  271

Deep learning-based channel estimation with application to 5G

and beyond networks

Parinitha Jayashankar, B. N. Shobha

Department of Electronics and Communication Engineering, Institut Teknologi Sri Jagadguru Chandrashekaranatha, Chikkaballapur, India

Article Info ABSTRACT

Article history: Channel state information (CSI) feedback estimation for a downlink medium
in a massive multiple input multiple output (MIMO) system is an essential
Received Jan 25, 2023 and critical task to improve channel capacity and performance yield,
Revised Nov 12, 2023 especially in a frequency division duplex (FDD) multiplexing system.
Accepted Dec 22, 2023 However, spectral efficiency degradation is a massive issue due to high
channel feedback overhead. This work proposes a deep learning-based
channel estimation (DLCE) model to improve channel reconstruction
Keywords: efficiency and channel overhead reduction accuracy. The proposed deep
learning (DL) mechanism consists of encoder and decoder network where
Channel state information encoder network is utilized to compress CSI matrices whereas decoder
DLCE network is used to decompress obtained CSI matrices. Here, inverse discrete
IDFT Fourier transform (IDFT) method is utilized to convert CSI matrices of
MIMO frequency domain into CSI matrices of delay domain. Simulation results are
Normalized mean square error evaluated between uplink and downlink medium in the massive MIMO
system considering a co-operation in science and technology (COST) 2,100
model. Here, a significant improvement in correlation and normalized mean
square error (NMSE) results is observed. The proposed DLCE model shows
superior performance against varied channel estimation techniques in terms
of NMSE and correlation efficiency.
This is an open access article under the CC BY-SA license.

Corresponding Author:
Parinitha Jayashankar
Department of Electronics and Communication Engineering
Institut Teknologi Sri Jagadguru Chandrashekaranatha
Chikkaballapur, Karnataka, India
Email: [email protected]

1. INTRODUCTION
The telecommunication industry is likely to cultivate endlessly at an emerging rate. This works as a fuel
to the modernization and yield of varied industries like agriculture, wireless networks, transportation, finance,
consumer electronics, and health care [1]. More specifically, the recent pandemic-coronavirus disease 2019
(COVID-19) outbreak have caused massive disruption and shows how important the telecommunication and
information industry is to our society [2]. Thus, a top rated technology, which has gained immense praise from
all over the world since last three decades and provided greater yield in many industries is cloud computing
technology. This technique has provided immense strength to the information and communication societies. The
key features of this technique are high elasticity, on-demand resource access, and large resource storage. Thus,
this technique is capable of handling high market demands and expectations without any hassle and management
problems. This technique mainly works on principle of pay per use model [3], [4]. However, telecommunication
and wireless communication system must handle the ever-rising demands of evolving techniques, devices, and
networks to keep up with these trends, which cannot be met by third or fourth generation cellular networks.
Therefore, current research mainly focuses on 5 generation (5G) cellular networks or beyond 5G (B5G) networks.

Journal homepage: http://ijres.iaescore.com

272  ISSN: 2089-4864

Further, the future 5G or B5G networks can provide data rates up to gigabit-per-second and latency
up to few milliseconds with billions of wireless devices remain linked with each other [5]. However,
implementation of 5G cellular network requires high bandwidth spectrum as well as lightning speed. Thus,
massive multiple input multiple output (MIMO) technique is a top-rated and efficient mechanism to provide
immense strength towards future 5G or B5G networks [6], [7]. The main advantages of massive MIMO systems
are enhancement of channel capacity and spectral efficiency of a wireless network. The massive MIMO system
consists of a base station (BS) and user equipment (UE) in which BS is heavily equipped with hundreds of
antennas whereas UE transmit compressed information towards BS for the reconstruction of received
information. However, the potential advantages of massive MIMO systems are exploited by the acquisition of
channel state information (CSI). However, acquisition of CSI is a critical and channeling task. Further, the
performance of massive MIMO system is heavily affected by the CSI acquisition efficiency [8]. Thus,
acquisition of CSI in uplink and downlink medium is essential as well as challenging. The CSI acquisition can
take place in frequency division duplex (FDD) as well as time division duplex (TDD) domain. Further, in FDD
domain, operating frequencies for both uplink and downlink medium are different whereas both uplink and
downlink medium remain separated from each other considering varied time-slots. However, CSI acquisition
in FDD domain is quite challenging and complex task due to frequency separation of uplink and downlink
medium. Thus, downlink medium needs to be predicted at UE and then fed back to the BS via a feedback
channel. However, this process can cause high channel overhead. Thus, several researchers have provided their
efforts to mitigate channel overhead and improve spectral efficiency of channel estimation.
A CSI compression method is introduced to estimate CSI and compress channel matrices in a massive
MIMO systems based on randomized low-rank approximation [9]. This method enhances accuracy precision,
reduces computational complexity, and exploits characteristics of channel matrices. According to research by
Choi et al. [10], a zero-feedback FDD system is adopted to estimate channel matrices in a massive MIMO
system. Here, vector spatial signature (VSS) model is also utilized to obtain data of channel measurements.
Simulation results are computed in terms of beamforming efficiency and normalized mean square error
(NMSE). A deep learning (DL) based mechanism is introduced to estimate channel matrices in a millimetre-
wave massive hybrid MIMO system [11]. Here, sparsity of channel matrices is exploited to mitigate channel
overhead. Further, a priori and learning model is introduced to study approximated vectors and recover multiple
subcarriers. In [12]–[14], a channel reconstruction model is presented for a downlink medium in a massive
MIMO system to acquire CSI. Here, reference signals of CSI feedback medium are utilized to back spatial
multiplexing to counter power consumption and computational complexity problems. Here, a downlink CSI is
reconstructed using sounding reference signals of uplink medium. In [15]–[18], a wideband channel estimation
technique is proposed for spatial multiplexing in a massive MIMO system with low-resolution analog-to-digital
converters (ADCs). This technique reduces propagation delay and improves inter-symbol interference. Further,
a maximum posteriori (MAP) criterion is adopted for channel estimation. In [19], [20], a convolutional neural
network architecture is presented to compress channel matrices in a massive MIMO system. Further, a
compression sensing based method is utilized to quantize CSI and improves reconstruction accuracy. Here, a
CsiNet+ method is adopted to parameter efficiency by 38.0% and 46.7%.
However, implementation of these techniques in practical wireless system is quite challenging due to
high interference, massive channel overhead, and lower reconstruction accuracy. Thus, a deep learning-based
channel estimation (DLCE) model is proposed in this article in a massive MIMO system. The proposed DLCE
technique is utilized to estimate downlink CSI matrices by exploiting correlation between uplink and downlink
medium in the massive MIMO system. These channel matrices are obtained from UE feedback and downlink
channel matrices are acquired from CSI estimation of uplink medium at source and UE. The proposed DL
mechanism consists of encoder and decoder network where encoder network is utilized to compress CSI
matrices whereas decoder network is used to decompress obtained CSI matrices. A high quality training is
conducted on a large database like co-operation in science and technology (COST) 2,100 dataset using a DL
model to improve channel estimation efficiency. Here, efficient CSI feedback code-words are decoded and
channel correlation is utilized to improve performance of massive MIMO system. Experimental results are
computed using a well-trained DL model in terms of NMSE and channel correlation and compared against
several traditional CSI estimation methods.
This article is arranged in following manner. Section 2, discusses about the mathematical
representation of channel estimation in a massive MIMO system. In section 3, demonstrates simulation results
and their comparative analysis with classical CSI estimation methods and section 4 concludes the paper.

2. MODELLING OF PROPOSED DLCE MODEL

This section discusses about the mathematical representation of DLCE model to estimate channel
matrices efficiently and reduce channel overhead in a massive MIMO system. Here, DL based model is adopted

Int J Reconfigurable & Embedded Syst, Vol. 13, No. 2, July 2024: 271-277
Int J Reconfigurable & Embedded Syst ISSN: 2089-4864  273

to train large dataset and improve channel reconstruction efficiency. Here, multiple antennas are utilized at the
BS and channel estimates are acquired at UE and fed back to the BS to reconstruct CSI feedback. Furthermore,
downlink CSI medium is estimated using CSI estimation of uplink medium at UE as well as uplink CSI
estimation at BS. Figure 1 demonstrates block diagram of downlink CSI feedback mechanism. A detailed
mathematical representation of channel overhead reduction and spectral efficiency enhancement in a massive
MIMO system is presented in the following section.

Figure 1. Block diagram of DL based downlink CSI feedback

Here, a study is conducted on a massive MIMO system with a singular cell. Here, the massive MIMO
system consists of a BS and UE. Further, a BS consists multiple number of antennas Cd ≫ 1 and UE is
equipped with a singular antenna. Here, orthogonal frequency division multiplexing (OFDM) is adopted in a
massive MIMO system considering Cl subcarriers. Then, the received code-words for downlink medium
considering c − th subcarrier is given by (1).

(c) (c)M (c) (c) (c)

k g = mg rS yg + cg (1)

(c)
Where, vector coefficients of a channel considering c − th subcarrier are denoted as mg and
(c) Cd ×1 (c)
proportional to mg ∈ 𝔼 . Further, broadcasted code-word is given by yg ∈ 𝔼 and broadcaster beam-
(c) (c) Cd ×1 (c)
former is given by rS and proportional to rS ∈ 𝔼 . Moreover, the additional noise is given by cg ∈
𝔼 and transpose of conjugates are given by (∙)M. Then, the received code-words for uplink medium considering
c − th subcarrier is given by (2).

(c) (c)M (c) (c) (c)M (c)

k n = rF mn yn + rF cn (2)

(c)
Where, vector coefficients of a channel considering c − th subcarrier are denoted as mn for an uplink
(c) (c)
medium and proportional to mn ∈ 𝔼Cd×1 . Further, received code-word is given by yn ∈ 𝔼 and receiver
(c) (c)
beam-former is given by rF and proportional to rF ∈ 𝔼Cd×1 . Moreover, the additional noise is given by
(c)
cn ∈ 𝔼 and transpose of conjugates are given by (∙)M in an uplink medium. Then, the CSI matrices of uplink
channel medium in a frequency domain is given by (3).

Deep learning-based channel estimation with application to 5G and beyond … (Parinitha Jayashankar)
274  ISSN: 2089-4864

(Cl ) M
̃ n = [m(1)
M n , … … . , mn ] (3)

̃ n ∈ 𝔼Cl×Cd . Further, the CSI matrices of downlink

Where, uplink CSI matrices are proportional to M
channel medium in a frequency domain is given by (4):

(Cl ) M
̃ g = [m(1)
M g , … … . , mg ] (4)

where, downlink CSI matrices are proportional to M ̃ g ∈ 𝔼Cl×Cd . Further, it is considered that the uplink CSI
matrices M̃ n and downlink CSI matrices M̃ g are perfectly obtained by the BS and UE respectively. However,
the focus of this study remains on the acquisition of downlink CSI matrices. As mentioned, downlink CSI
̃ g ∈ 𝔼Cl×Cd and due to large size of Cd , the feedback overhead becomes very large
matrix is proportional to M
in a massive MIMO system. Thus, CSI feedback reduction is handled using a property of channel matrices in
which common sparsity is utilized which is present in uplink and downlink medium at the time of CSI
estimation in a time-domain. Then, firstly, channel matrix in frequency domain Ml is converted into channel
matrix of time domain Ms using an inverse discrete Fourier transform (IDFT) method:

Ml LM = Ms (5)

where, L is expressed as a singular DFT matrix which is proportional to L ∈ Cd × Cd . Whereas, matrix obtained
after IDFT process Ms is remain proportional to Cl × Cd . Further, majority elements of matrix Ms are zero or near
to zero except the first Ĉl row elements. Thus, zero value elements are discarded so that uplink CSI matrix M ̃n
̃ ̂
and downlink CSI matrix Mg are expressed as Mn and Mg after IDFT process, respectively. Still, Cl × Cd is
considered as large quantity. Thus, CSI matrix Mg is further compressed at UE.
Thus, the proposed massive MIMO system contains encoder and decoder mechanism in which
encoder is used to compress CSI matrices and decoder is utilized to reconstruct CSI matrices, respectively.
However, the proposed DLCE model estimates downlink CSI feedback matrices with the help of uplink CSI
matrices. Assume that the recovered CSI matrix in a downlink medium is denoted as M ̂ g . Then, encoder and
decoder networks are given by (6) and (7), respectively.

h = lb (Mg ) (6)

̂ g = la (h, Mn )
M (7)

Thus, the proposed DLCE model shows encoder decoder configuration in which encoder performs
compression of CSI matrices and decoder performs reconstruction of obtained CSI matrices and fed them back
to the BS from UE. In this way, efficient CSI feedback reconstruction and channel overhead reduction
mechanism is presented. Further, CSI matrices of uplink medium acquired at BS and UE are utilized to
reconstruct CSI matrices of downlink medium.
Here, the CSI matrices are studied in time domain with the help of correlation coefficients obtained
considering uplink and downlink CSI feedback matrices. The obtained CSI matrices are complex in nature in
which correlation coefficients are separated into real and imaginary portions. The acquired real and imaginary
portions from correlation coefficients considering uplink and downlink medium are uneven and inconsistent. It is
examined that channel estimation considering two varied carrier frequency for an FDD system, can contribute to
random phase differences between those two carrier frequencies. Thus, it is analysed that from the FDD channel
estimation, the magnitude obtained in time domain has higher correlation than the phase obtained in time domain.
As a result, correlation of magnitude and phase is evaluated separately by conversion of CSI matrices into polar
coordinates. Furthermore, correlation acquired is higher for both uplink channel model and downlink channel
model considering magnitude whereas both uplink channel model and downlink channel model, correlation
obtained is lower considering phase. Further, real values show better correlation results in contrast to their
respective signs considering channel coefficients of both uplink and downlink medium.
Further, the proposed DLCE model evaluates real and imaginary portions of channel correlation
coefficients as well as their respective signs which are separately fed back to the BS, unlike classical channel
estimation methods. Therefore, this concept is an essential step to mitigate CSI feedback overhead. The main
architectural characteristics are utilization of residual blocks, convolutional layer, fully linked layer, and
generation of feature maps from obtained high quality features. Here, correlation coefficients of uplink and

Int J Reconfigurable & Embedded Syst, Vol. 13, No. 2, July 2024: 271-277
Int J Reconfigurable & Embedded Syst ISSN: 2089-4864  275

downlink medium are transformed into real and imaginary portions. As a result, the size of fully linked layers
are enhanced and the number of generated feature maps are increase in number. Thus, efficient CSI feedback
mechanism is designed by compression of phase and their respective signs considering downlink CSI matrices
at the UE. This is due to the low correlation generated by phase. Furthermore, the compressed phase and signs
of downlink CSI matrices are fed back to BS from UE and uplink CSI matrices present at the BS, both are
decompressed together to achieve downlink CSI channel matrices. In this way, efficient channel reconstruction
is achieved as well as channel overhead is mitigated using proposed DLCE model.

3. RESULT AND DISCUSSION

This section discusses about the performance efficiency enhancement using proposed DLCE model
in a massive MIMO system. Here, channel estimation is achieved at a rapid rate with the help of DL model,
which performs high quality training of channel parameters and coefficients. The proposed DL model reduces
utilization of channel parameters to mitigate computation complexity issues. The proposed DLCE model
exploits absolute and imaginary portions of correlation coefficients obtained considering uplink and downlink
CSI feedback matrices. Besides, correlation coefficients are computed using CSI matrices of uplink and
downlink medium. Here, multiple antennas are equipped at BS in a massive MIMO station. Here, efficient CSI
feedback code words are compressed at the encoder side and decompressed at the decoder side and channel
correlation is utilized to improve performance of massive MIMO system as shown in Figure 1. Here, IDFT
method is utilized to convert CSI matrices of frequency domain into CSI matrices of delay domain.
Here, correlation is evaluated between uplink and downlink medium in the massive MIMO system
considering a COST 2,100 model [21]. Further, the data present in the COST 2,100 mode is divided into two
type of environments such as indoor environment and outdoor environment. Nevertheless, indoor environment
is considered in an article in which frequency of uplink medium is set as 5.1 GHz and frequency of downlink
medium is fixed at 5.3 GHz. Further, uniform linear array antennas are adopted to improve performance
efficiency. In future work, an extension of proposed DLCE model and downlink CSI feedback architecture is
presented and experimental results are obtained considering outdoor environment and multi-cell scenarios. The
total number of uniform linear array antennas are utilized at BS are Cd = 32 and number of subcarriers utilized
are Cl = 1,024. The number of epochs considered for large training are 700 with batch size as 200. The data
present in the COST 2,100 dataset is massive and requires efficient training, which is conducted using DL
model. The number of parameters utilized in the proposed DLCE model to improve CSI reconstruction
accuracy are huge. However, parameter utilization are reduced in comparison with traditional channel
estimation methods.
In this section, the obtained experimental results using proposed DLCE model and their comparison
with classical channel estimation methods are examined considering COST 2,100 channel model. The data
utilized to measure CSI reconstruction accuracy is obtained by efficiency training of COST 2,100 model. The
performance accuracy is measured using proposed DLCE model in terms of NMSE and correlation obtained
between downlink and uplink CSI matrices against varied compression ratios and the performance efficiency
is compared against varied traditional CSI estimation techniques. The obtained simulation results shows
efficient results in terms of performance accuracy, NMSE and correlation coefficients and obtained simulation
results outperforms traditional CSI feedback estimation techniques.
The simulation results are demonstrated in Table 1 for NMSE against CsiNet [22], [23], CRNet [24],
and CsiNetPlus [20] whereas Table 2 shows correlation efficiency against CsiNet and normalized-CsiNet [22].
Further, overall average NMSE is reduced is 9.60% against previous best CSI estimation technique in terms of
NMSE. Here, Figure 2 demonstrates graphical representation of performance accuracy using proposed DLCE
model against traditional channel estimation methods such as CRNet-Const, CsiNet, DS-NLCSiNet [25] in terms
of NMSE results against varied compression ratios. Here, a significant reduction in NMSE results is observed
which results in a higher performance accuracy. In contrast to other traditional channel estimation techniques,
proposed DLCE model shows lower NMSE results. Here, a significant improvement in NMSE and correlation
results is observed. This conclude that the proposed DLCE model efficiently reconstruct downlink CSI feedback
matrices as well as reduces channel overhead in contrast to several classical channel estimation techniques.

Table 1. NMSE (dB) using proposed DLCE model for varied compression ratios
CR CsiNet CRNet CsiNetPlus DLCE
4 -17.36 -26.99 -27.37 -28.983
8 -12.70 -16.01 -18.29 -19.47
16 -8.65 -11.35 -14.14 -16.47

Deep learning-based channel estimation with application to 5G and beyond … (Parinitha Jayashankar)
276  ISSN: 2089-4864

Table 2. Correlation efficiency using proposed DLCE model for varied compression ratios
CR CsiNet Norm-CsiNet DLCE
4 0.99 0.99 0.99
16 0.93 0.94 0.97
32 0.89 0.91 0.96
64 0.87 0.84 0.90

NMSE performance against varied compression ratios

for outdoor scenarios
0
CRNet-Const CsiNet DS-NLCsiNet DLCE
-2
-4
NMSE (DB)

-6
-8
-10
-12
-14
32 64

Figure 2. Graphical representation of performance accuracy using proposed DLCE model

4. CONCLUSION
Massive MIMO is quite essential for the practical implementation of future generation high speed
networks. Therefore, in this article, a DLCE model is adopted in a massive MIMO system to estimate CSI
reconstruction accuracy and mitigate channel overhead reduction. A comprehensive mathematical modelling
is presented for the evaluation of correlation coefficients considering uplink and downlink channel medium.
The proposed DLCE model exploits absolute and imaginary portions of correlation coefficients obtained
considering uplink and downlink CSI feedback matrices. The main architectural characteristics are utilization
of residual blocks, convolutional layer, fully linked layer, and generation of feature maps from obtained high
quality features. Here, proposed DLCE model utilizes COST 2,100 channel dataset to train CSI matrices. The
performance accuracy is measured using proposed DLCE model in terms of NMSE and correlation obtained
between downlink and uplink CSI matrices against varied compression ratios. Further, overall average
performance accuracy is reduced is 9.60% against previous best CSI estimation technique in terms of NMSE.
Thus, the proposed DLCE model outperforms classical channel estimation techniques in terms of NMSE and
performance efficiency.

REFERENCES
[1] M. Liyanage, I. Ahmad, A. B. Abro, A. Gurtov, and M. Ylianttila, Eds., A comprehensive guide to 5G security, 1st edition. Hoboken:
Wiley, 2018.
[2] Oxford Analytica, “COVID-19 will probably accelerate remote working trend,” Emerald Expert Briefings, Jan. 2020, doi:
10.1108/OXAN-ES251161.
[3] Y. Zhang, X. Lan, J. Ren, and L. Cai, “Efficient computing resource sharing for mobile edge-cloud computing networks,”
IEEE/ACM Transactions on Networking, vol. 28, no. 3, pp. 1227–1240, Jun. 2020, doi: 10.1109/TNET.2020.2979807.
[4] M. Shojafar, C. Canali, R. Lancellotti, and J. Abawajy, “Adaptive computing-plus-communication optimization framework for
multimedia processing in cloud systems,” IEEE Transactions on Cloud Computing, vol. 8, no. 4, pp. 1162–1175, Oct. 2020, doi:
10.1109/TCC.2016.2617367.
[5] C. Zhang, Y.-H. Huang, F. Sheikh, and Z. Wang, “Advanced baseband processing algorithms, circuits, and implementations for 5G
communication,” IEEE Journal on Emerging and Selected Topics in Circuits and Systems, Aug. 2017, doi:
10.1109/JETCAS.2017.2743107.
[6] T. L. Marzetta, “Massive MIMO: an introduction,” Bell Labs Technical Journal, vol. 20, pp. 11–22, 2015, doi:
10.15325/BLTJ.2015.2407793.
[7] V. W. S. Wong, R. Schober, D. W. K. Ng, and L.-C. Wang, Eds., Key technologies for 5G wireless systems. Cambridge: Cambridge
University Press, 2017, doi: 10.1017/9781316771655.
[8] E. G. Larsson, O. Edfors, F. Tufvesson, and T. L. Marzetta, “Massive MIMO for next generation wireless systems,” IEEE
Communications Magazine, vol. 52, no. 2, pp. 186–195, Feb. 2014, doi: 10.1109/MCOM.2014.6736761.
[9] Z. Wei, H. Li, H. Liu, B. Li, and C. Zhao, “Randomized low-rank approximation based massive MIMO CSI compression,” IEEE
Communications Letters, vol. 25, no. 6, pp. 2004–2008, Jun. 2021, doi: 10.1109/LCOMM.2021.3065751.

Int J Reconfigurable & Embedded Syst, Vol. 13, No. 2, July 2024: 271-277
Int J Reconfigurable & Embedded Syst ISSN: 2089-4864  277

[10] T. Choi et al., “Experimental investigation of frequency domain channel extrapolation in massive MIMO systems for zero-feedback
FDD,” IEEE Transactions on Wireless Communications, vol. 20, no. 1, pp. 710–725, Jan. 2021, doi: 10.1109/TWC.2020.3028161.
[11] X. Ma, Z. Gao, F. Gao, and M. Di Renzo, “Model-driven deep learning based channel estimation and feedback for millimeter-wave
massive hybrid MIMO systems,” IEEE Journal on Selected Areas in Communications, vol. 39, no. 8, pp. 2388–2406, Aug. 2021,
doi: 10.1109/JSAC.2021.3087269.
[12] H. Lee et al., “Downlink channel reconstruction for spatial multiplexing in massive MIMO systems,” IEEE Transactions on
Wireless Communications, vol. 20, no. 9, pp. 6154–6166, Sep. 2021, doi: 10.1109/TWC.2021.3072158.
[13] I.-S. Kim and J. Choi, “Spatial wideband channel estimation for mmwave massive MIMO systems with hybrid architectures and
low-resolution ADCs,” IEEE Transactions on Wireless Communications, vol. 20, no. 6, pp. 4016–4029, Jun. 2021, doi:
10.1109/TWC.2021.3054998.
[14] H. Huang, J. Yang, H. Huang, Y. Song, and G. Gui, “Deep learning for super-resolution channel estimation and DOA estimation
based massive MIMO system,” IEEE Transactions on Vehicular Technology, vol. 67, no. 9, pp. 8549–8560, Sep. 2018, doi:
10.1109/TVT.2018.2851783.
[15] Y. Yang, F. Gao, X. Ma, and S. Zhang, “Deep learning-based channel estimation for doubly selective fading channels,” IEEE
Access, vol. 7, pp. 36579–36589, 2019, doi: 10.1109/ACCESS.2019.2901066.
[16] Q. Bai, J. Wang, Y. Zhang, and J. Song, “Deep learning-based channel estimation algorithm over time selective fading channels,”
IEEE Transactions on Cognitive Communications and Networking, vol. 6, no. 1, pp. 125–134, Mar. 2020, doi:
10.1109/TCCN.2019.2943455.
[17] X. Zhu, A. Koc, R. Morawski, and T. Le-Ngoc, “A deep learning and geospatial data-based channel estimation technique for hybrid
massive MIMO systems,” IEEE Access, vol. 9, pp. 145115–145132, 2021, doi: 10.1109/ACCESS.2021.3121750.
[18] Y. Zhang, J. Sun, J. Xue, G. Y. Li, and Z. Xu, “Deep expectation-maximization for joint MIMO channel estimation and signal
detection,” IEEE Transactions on Signal Processing, vol. 70, pp. 4483–4497, 2022, doi: 10.1109/TSP.2022.3205478.
[19] X. Cheng, D. Liu, C. Wang, S. Yan, and Z. Zhu, “Deep learning-based channel estimation and equalization scheme for
FBMC/OQAM systems,” IEEE Wireless Communications Letters, vol. 8, no. 3, pp. 881–884, Jun. 2019, doi:
10.1109/LWC.2019.2898437.
[20] J. Guo, C.-K. Wen, S. Jin, and G. Y. Li, “Convolutional neural network-based multiple-rate compressive sensing for massive MIMO
CSI feedback: design, simulation, and analysis,” IEEE Transactions on Wireless Communications, vol. 19, no. 4, pp. 2827–2840,
Apr. 2020, doi: 10.1109/TWC.2020.2968430.
[21] L. Liu et al., “The COST 2100 MIMO channel model,” IEEE Wireless Communications, vol. 19, no. 6, pp. 92–99, Dec. 2012, doi:
10.1109/MWC.2012.6393523.
[22] H. Ye, F. Gao, J. Qian, H. Wang, and G. Y. Li, “Deep learning-based denoise network for CSI feedback in FDD massive MIMO
systems,” IEEE Communications Letters, vol. 24, no. 8, pp. 1742–1746, Aug. 2020, doi: 10.1109/LCOMM.2020.2989499.
[23] C.-K. Wen, W.-T. Shih, and S. Jin, “Deep learning for massive MIMO CSI feedback,” IEEE Wireless Communications Letters, vol.
7, no. 5, pp. 748–751, Oct. 2018, doi: 10.1109/LWC.2018.2818160.
[24] Z. Lu, J. Wang, and J. Song, “Multi-resolution CSI feedback with deep learning in massive MIMO system,” arXiv, May 27, 2021,
doi: 10.48550/arXiv.1910.14322.
[25] Z. Lu, X. Zhang, H. He, J. Wang, and J. Song, “Binarized aggregated network with quantization: flexible deep learning deployment
for CSI feedback in massive MIMO system,” arXiv, May 01, 2021, doi: 10.48550/arXiv.2105.00354.

BIOGRAPHIES OF AUTHORS

Parinitha Jayashankar earned her bachelors of engineering in telecommunication

engineering and master’s degree in digital communication and networking from Visvesvaraya
Technological University (VTU), Belgavi. Currently she is a pursuing her Ph.D. in electronics
and communication engineering from Institut Teknologi Sri Jagadguru Chandrashekaranatha,
VTU, Belgavi. She is working as an assistant professor with a total experience of 11 years. She
has guided over 11 under graduate projects and published 10 papers in various journals. She can
be contacted at email: [email protected].

Dr. B. N. Shobha is professor and head of the Department of Electronics and

Communication Engineering at Institut Teknologi Sri Jagadguru Chandrashekaranatha,
Chickbaballpur, Karnataka, Visvesvaraya Technological University, Belgavi. She is having a
total experience of 29 years. She is qualified in bachelor and master degrees in electronics from
Bangalore University, and Ph.D. from Karpagam University in the field of electronics. She has
guided over 60 undergraduate projects, 45 post graduate projects, and guiding 4 Ph.D. students.
She has published over 28 papers in various journals and have 6 patents in her name. Her area
of interest are bio sensor, image processing, computer vision, embedded system, and IoT. She
can be contacted at email: [email protected].

Deep learning-based channel estimation with application to 5G and beyond … (Parinitha Jayashankar)