CC BY
17
CT) I
CT) £±
CO
Deep learning channel estimation for 5G wireless communications
Mohammed Zouaoui M. Laidounia, Taki-eddine Ahmed A. Benyahiab, Boban Z. Pavlovicc, Salem-Bilal B. Amokraned, Touati B. Adlie °
University of Defence in Belgrade, Military Academy, Department of Telecommunications and Informatics, Belgrade, Republic of Serbia, £
e-mail: [email protected], corresponding author, |
ORCID iD: ©https://orcid.org/0009-0008-6042-0513 b e-mail: [email protected], ORCID iD: ©https://orcid.org/0009-0006-6025-6915
e-mail: [email protected], ¿)
ORCID iD: ©https://orcid.org/0000-0002-5476-7894 d e-mail: [email protected], ORCID iD: ©https://orcid.org/0009-0009-7588-5708
a
e-mail: [email protected], E
ORCID iD: ©https://orcid.org/0009-0000-2673-6954 to
DOI: 10.5937/vojtehg71-46057;https://doi.org/10.5937/vojtehg71-46057
FIELD: computer sciences, telecommunications ARTICLE TYPE: original scientific paper ^
'c
Abstract: <ô
e
5
c
Introduction/purpose: In recent years, deep learning techniques, particularly Convolutional Neural Networks (CNNs), have demonstrated re- ^ markable performance in 5G communication systems by significantly improving the accuracy of channel estimation compared to conventional methods. This article aims to provide a comprehensive review of the existing literature on CNN-based channel estimation techniques, as well as to enhance the state-of-the-art CNN-based channel estimation methods o
"O
by proposing a novel method called VDSR (Very Deep Super Resolution), inspired by Image Super-Resolution techniques. Methods: To evaluate the effectiveness of various approaches, we conduct a comprehensive comparison considering different scenarios, including low Signal-to-Noise Ratio (SNR) and high SNR, as well as Line-of-Sight (LOS) and Non-Line-of-Sight (NLOS) scenarios. Through this comparative analysis, we assess the performance of the existing methods and highlight the advantages offered by the proposed VDSR-based technique.
Results: Our findings reveal a significant potential of CNN-based channel estimation in 5G communication systems, with the VDSR method demonstrating a consistent performance across all scenarios. This re-
search contributes to the advancement of channel estimation techniques in 5G networks, paving the way for enhanced wireless communication systems with improved reliability.
Conclusion: The VDSR architecture demonstrates remarkable adapt> ability to different types of channels, which results in achieving requested co performances for all analyzed SNR values.
o
™ Key words: deep learning, CNN, 5G communication systems, very deep
w super resolution.
oE o
^ Introduction
o
With the advent of 5G communication technology, the demand for highspeed, low-latency, and reliable wireless communication is increasing exponentially (Albreem, 2015). The key enabler for 5G communication is £ accurate channel estimation, which refers to the process of estimating the wireless channel parameters between the transmitter and the receiver (Morocho-Cayamcela et al., 2019). Accurate channel estimation is critical for improving the performance of 5G communication systems, including data rates, spectral efficiency, and reliability (Ma et al., 2015). In recent iH years, convolutional neural networks (CNNs) have emerged as a promising technique for channel estimation in 5G communication systems (James y et al., 2011). CNNs are powerful deep learning algorithms that can learn and extract complex features from large amounts of data. By leveraging o the power of CNNs, channel estimation in 5G communication systems can achieve high accuracy, robustness, and efficiency (Ye et al., 2017; Kaur et al., 2021).
This work aims to investigate the effectiveness of CNN-aided channel estimation in 5G communication systems. Specifically, through exploring the existing literature on CNN-aided channel estimation, a novel architecture for CNN-based channel estimation is being proposed, and the performance of the suggested approach will be assessed through simulations.
The rest of this paper is structured as follows. Section 2 describes the 5G new radio SISO-OFDM system. Section 3 provides a literature review on CNN-aided channel estimation and describes the architecture used for our method. Section 4 presents the results and analysis of the simulation. Section 5 discusses the implications of our findings and provides recommendations for future research. Finally, Section 6 presents the conclusions of this work.
X
o
LU
<
CO
o
5G new radio SISO-OFDM system
The focus of this paper is on analyzing a SISO-OFDM system that employs a single antenna at both the transmitter and receiver. This system is depicted in the diagrams shown in Figure 1 and Figure 2 , and the channel model is constructed accordingly.
о
CT> I
CT) £± Œ
СЛ С
о
Transmitter
Figure 1 shows the architecture of the transmitter, which involves converting serial binary input bits (a sequences of zeros and ones) into a parallel form. Based on the chosen modulation scheme, the binary bits are then mapped onto symbols, with each symbol being K - dimensional and the binary bits selecting one of M constellation points. Typically, K is 2, and M is determined by the modulation scheme chosen at the higher layer. Additionally, intermittent pilot symbols are inserted among the modulated symbols, which serve as a reference for channel estimation and are also recognizable to the receiver.
E E о о
(Я (Я jD Ф
CD
Ю
го E
(Я
ф
Serial to Parallel Symbole Add pilot symbole IFFT Parallel to Serial Add cyclic prefix
mapping
Figure 1 - Block diagram for the OFDM transmitter model
Рис. 1 - Блок-схема модели передатчика OFDM Слика 1 - Блок-д^аграм за модел OFDM предаjника
о CT
го
Œ Ф Ф
а
"то ф
Let XS e {Xm,Xp} where Xm e {s0,si,s2, ,sM-1} is the modulated symbol selected by log2 M binary input bits and Xp e {p0,p1,p2, ...,pK-1} are pilot symbols respectively. Equation 1 in the digital domain is Inverse Discrete Fourier Transform operation which can be efficiently realized by the Inverse Fast Fourier transform (IFFT) before adding the cyclic prefix (Banerjee etal., 2022).
о
"О
1 ^ k xs(n) = N ¿2 Xs(k)exP(j2^—n)-s k=i s
(1)
where NS is the IFFT length. A parallel to serial converter is present after the IFFT operation to serialize the output.
913
ф
Receiver
о >
PO (M о (M
ОС Ш 0£ ZD О О
-J
<
о
X
о ш
I—
>-
Q1 <
Figure 2 shows the architecture of the receiver, which includes a process for estimating the timing of the received signal.
I
Cyclic prefix removal Serial to Parallel Equalise Parallel to Serial
FFT Channel and symbol De-
estimation
mapping
Interpolation linear
Deeplearning
Figure 2 - Block diagram for the OFDM receiver model
Рис. 2 - Блок-схема модели приемника OFDM Слика 2 - Блок-д^аграм за модел OFDM преемника
гл
CD S2 'У z х ш I—
о
о >
This process involves cross-correlating the input waveform with a reference waveform and compensating for any timing offset. Once the timing offset has been accounted for, the cyclic prefix is removed from the received waveform. If YS is the received OFDM symbol and yS is the output of the FFT operation, then yS can be expressed in the following manner (Banerjee etal., 2022):
1 Ns - I
Vs{n) = ntY^ 2nNn)
s 1=1 s
(2)
The pilot samples, which are located at predetermined positions, are extracted from the signal and utilized to estimate the channel characteristics. This channel estimation information is then used to equalize the output yS(n). After equalization, the signal is demodulated based on the modulation scheme that was employed at the transmitter.
Signal model
In an OFDM system (Soltani et al., 2019), for the kth time slot and the ith subcarrier, the input-output relationship is represented as:
Yi.k = XLk + Z
Ji,k
(3)
914
Considering an OFDM subframe of size NSND, the time slot index k is ^
5G data architecture
The physical layer of the 5G NR is based on resource blocks allowing the NR physical layer to adapt to various spectrum allocations. A resource block spans 12 subcarriers with a given sub-carrier spacing. A radio frame has a duration of 10 ms and consists of 10 sub-frames with a sub-frame duration of 1ms as shown in figure 3 . A sub-frame is formed by 1 or multiple slots each having 14 adjacent symbols (a variable number of OFDM symbols per subframes, different from LTE) (3GPP. 2018).
Subcarrier spacing
slot: 0,25 ms
Figure 3 - Sub-frame architecture in 5G Рис. 3 - Подкадровая архитектура в 5G Слика 3 - Архитектура подоквира у 5G
between [0, ND - 1], and the range of the subcarrier index i is [0, NS - 1]. Yikk: The received signal Xikk: Transmitted OFDM symbol
Zikk: white Gaussian noise §
Hi,k: the (i,k) element of H e CNsNd . H represents time-frequency response of the channel for all subcarriers and time slots.
en £±
СЛ
E E о о
(Л (Я jD <D
CD
Ю
го E
(Я
<u
о CT
го
JD
Π<U <U
Q
о
"О
Subframe
slot:1ms
15 KHz
In 5G NR, the pilot symbols are referred to as demodulation reference symbols (DMRS) and this is used by the receiver for radio channel estimation. The DMRS symbols are uniformly placed within sub-carriers as shown in figure 4. We assume the DMRS symbols used in the 3GPP specification (3GPP. 2020a).
Figure 4 shows the DM-RS pattern and frequency for type 1 and type 2. Type 1 on the left corresponds to every other resource element in the frequency being occupied by a DM-RS symbol. Type 2 on the right shows two consecutive resource elements occupied by the DM-RS symbols out of
о >
00 C\l о (M
ОС Ш 0£ ZD
о
О <
О
х о ш
I—
>-
Q1 <
ОТ
CD
S2 ■О
х ш н
о
о
OFDM Symbol duration
rV
Subcarrier-j^
DM-RS Symbol
un s l
О. С
DMRS Configuration type 1
14 symbols slot period
DMRS Configuration type 2
Figure 4 - Representational figure of the distribution of DM-RS Рис. 4 - Репрезентативное изображение распределения DM-RS Слика 4 - Репрезентативни приказ дистрибуц^е DM-RS
each group of six resource elements. Therefore, type 1 has a denser occupancy at 50% of the resource estimates, versus one-third of the resource elements for type 2. On the other hand, you can only have two such columns of type 1 DM-RS, whereas there can be three different sets of type 2 DMRS as there are two more possible positions for a set of two DM-RS in each group of six resource elements. This means that type 2 supports a larger number of orthogonal signals, which is more suitable for multi-user MIMO. These two types correspond to a trade-off between density and frequency and the number of orthogonal DM-RS sequences supported.
Channel model
In wireless communication channels, the signal transmitted from a base station to user equipment not only includes a direct line-of-sight (LOS) component, but also other components that are reflected off scatterers, leading to a multipath propagation environment. Each path of the signal experiences different amounts of attenuation and delay (Wang et al., 2018). The channel's impulse response can be expressed as:
o ^f en I
L-1
h(t) = £ at5(r - ti), (4) |
co c o
i=0
where ai is the attenuation and ti is the delay in the ith path.
CNN-aided channel estimation
o o tn (0 JD <D
Tap Delay Line models |
In TDL models, the channel impulse response (CIR) is represented by a linear finite impulse response (FIR) filter. Each tap of the TDL model is composed of several multipath component (MPCs) with non-resolvable delays. Tap weights are modeled by a random process with amplitudes S following Rayleigh, Rician, or Weibull distributions (Wang etal., 2018).
A TDL (Tap Delay Line) profile in 5G communication represents a specific channel model that simulates the characteristics of radio wave propa- | gation in a wireless communication system. Three TDL models, namely TDL-A, TDL-B and TDL-C, are constructed to represent three different channel profiles for NLOS while TDL-D and TDL-E are constructed for LOS (3GPP. 2020b). »
ro
o
ro
a.
<u <u
"O
In recent times, there has been a significant surge in interest in channel Q estimation techniques based on deep learning. This is due to their ability to adapt and learn from data, as opposed to conventional estimation techniques that rely on a model-based approach.
A convolutional neural network (CNN) approach is chosen because the o channel estimation problem can be modelled as an image-processing problem (Banerjee et al., 2022; Soltani et al., 2019; Gizzini et al., 2021). The CNN-based deep learning approach has proven to be efficient for handling image processing problems as it keeps the number of parameters in weight matrix less in comparison to a fully connected neural network model by making use of parameter sharing and sparsity of connections.
Recently, the channel estimation in OFDM systems has been approached using a deep learning-based framework, where the time-frequency grid of the channel response is represented as a 2D-image that is only available at the pilot positions. (Soltani et al., 2019) presented a deep learning-based framework for channel estimation in OFDM systems,
917
O >
00 C\l o (M
yy
0£ ZD O o
-J
<
X
o
LU
s
CD
O >
which proposed an image super-resolution (SRCNN) and image denoising (DnCNN) algorithms to estimate the channel. In (Banerjee et al., 2022) a CNN model for Over-the-Air channel estimation has been applied, and the model is proposed by Matlab.
In this paper, we present a novel method for channel estimation that utilizes a very deep convolutional network inspired by VGG-net used for ImageNet classification; the method was proposed by (Kim etal., 2016) and presents a highly accurate single-image super-resolution (SR) technique. The next sections will provide detailed explanations of the three methods.
y Method 1: channel estimation using super-resolution (SRCNN) and denoising techniques
The method treats the channel grid with several pilots as a low-resolution (LR) image and aims to estimate the high-resolution (HR) channel. To achieve this, the framework models the channel response as a super-resolution image problem (Soltani et al., 2019).
The channel grid estimation is performed using two phases. In the first phase, the image super-resolution (SR) CNN-based (Convolutional Neural Network) algorithms (Dong et al., 2015), SRCNN, are implemented to ins? crease the resolution of the low-resolution (LR) input. The schema for the >o
CNN-based (Convolutional Neural Network) algorithms is shown in Figure 5.
o In the second phase, an image restoration (IR) method based on CNN
(Figure 6 ) is utilized to eliminate the noise effects and improve the quality of the estimated channel grid (Zhang et al., 2017).
Network architecture for SRCNN and DnCNN
The SRCNN technique involves utilizing an interpolation technique to estimate the high-resolution image (channel) values initially, and then refining the resolution by employing a three-layer convolutional network as shown in Figure 7:
• The first convolutional layer uses 64 filters of size 9 x 9 followed by ReLu activation,
• The second layer uses 32 filters of size 1 x 1 followed by ReLu activation.
? ? 7 ? 7
? P ? ? P
7 ? ? ? ?
f ? P 7 f
f ? ? ? ?
P ? 7 P ?
Super Resolution Network
(a)
(b)
Figure 5 - Super-resolution based CNN, (a) 2D-image which is known only at the
pilot positions, (b) estimated channel as a high-resolution Рис. 5 - CNN на основе сверхразрешения, (а) 2D-изображение, известное только на позициях пилота, (b) оцениваемый канал высокого разрешения Слика 5 - CNN заснован на супер резолуци]и, (а) 2D-^u^ ко}а jе позната само на пилот позици]ама, (b) проце^ени канал високе резолуцще
• The final layer uses only one filter of size 5 x 5 to reconstruct the grid channel.
The DnCNN technique in Figure 8 is a residual-learning based network composed of 20 convolutional layers:
• The first layer uses 64 filters of size 3 x 3 x 1 followed by a ReLU,
• Each of the succeeding 18 convolutional layers uses 64 filters of size 3 x 3 x 64 followed by batch-normalization and ReLU, and
• The last layer uses one 3 x 3 x 64 filter to reconstruct the output.
CO I
СЛ С
о
го о
Е Е о о
(Л
(Я ф
CD ю
го Е
(Я ф
о CT
го
Œ Ф Ф
Q
"то ф
Method 2: channel estimation using a regression method
The approach used for channel estimation is the same as the first method; the channel estimation problem was considered as an image processing problem by viewing the resource grid as a 2D image. A regression method based on deep learning is used in (Banerjee et al., 2022) to estimate a perfect channel. The input to the deep learning model is the LS channel estimated data and the CNN model can be trained against a perfect channel estimate as a reference, based on the statistical information available. CNN operates by applying convolution operations between images and kernels of different sizes to extract feature information. This process occurs in a multilayered system where the output of the convolution opera-
o
"O
'ro
о >
00 CM о CM
ОС Ш
0£ ZD
о
О <
О
X
о ш
I>-
Q1 <
ОТ
CD
S2 ■О
х ш н
о
о
(a)
(b)
Figure 6 - Denoising based CNN, (a) estimated channel which is considered as a noised image, (b) estimated channel Рис. 6 - CNN на основе шумоподавления, (a) оценочный канал, который рассматривается как зашумленное изображение, (b) оценочный канал Слика 6 - CNN заснован на смаъеъу шума, (a) процеъени канал ко\и се сматра сликом са шумом, (b) процеъени канал
Figure 7- SRCNN architecture Рис. 7 - Архитектура SRCNN Слика 7 - Архитектура SRCNN
tion is passed through an activation function, which is a non-linear function that transforms data. In regression problems, the final output layer is a regression layer that calculates the half-mean-squared-error loss. Finally, an optimization function is used to optimize the multilayered system, and the choice of optimization function is determined by the user.
Network architecture for the regression technique
The CNN model consists of 5 hidden layers as shown in Figure 9, where the first four hidden layers are associated with a ReLU activation function.
The fifth layer is associated only with the regression layer, as in regression problems the CNN output does not require an activation function.
DepthD() 1
Estimated channel
Figure 8 - DnCNN architecture Рис. 8 - Архитектура DnCNN Слика 8 - Архитектура DnCNN
Figure 9 - Regression technique architecture Рис. 9 - Архитектура метода регрессии Слика 9 - Архитектура технике регреси]е
The layers are ordered as follows:
• The first convolutional layer uses 64 filters of size 9 x 9 followed by ReLu activation,
• Each of the succeeding 2 convolutional layers uses 64 filters of size 5 x 5 followed by ReLu activation,
• The fourth layer uses 32 filters of size 5x5 followed by ReLu activation, and
• The final layer uses only one filter of size 5 x 5 followed by the regression layer to reconstruct the grid channel.
CO I
CP CP
СЛ с о
го о
Е Е о о
(Л (Я JD Ф
О ю
го Е
(Я ф
о
го
CP ф
ф а
о
"О
'го
Method 3: channel estimation using Very Deep Convolutional Networks
The channel estimation problem in this method was also modelled as an image-processing problem, the main difference being that this technique
о >
00 C\l о (M
ОС
УУ 0£ ZD
о
О <
О
х о ш
I—
>-
Q1 <
ОТ
S
CD
S2 >о
is using a very deep convolutional network to improve the performance. The SRCNN technique failed to create deeper models for super resolution with superior performance. However, (Kim et al., 2016) presented a method (VDSR: Very Deep Super-Resolution) that utilizes a very deep convolutional network inspired by VGG-net used for ImageNet classification, and it is found that increasing the depth significantly boosts the estimation performances. Given that VDSR shows a highly accurate single-image super-resolution, we want to apply this technique in the channel estimation problem.
Network architecture for VDSR
The VDSR (Kim et al., 2016) technique uses a very deep convolutional network inspired by Simonyan and Zisserman (Simonyan & Zisserman, 2014). The network structure cascades a pair of layers (convolutional and nonlinear) repeatedly. An interpolated low-resolution (CLR) channel goes through the layers and transforms into a high-resolution (HR) channel. The network predicts a residual image and the addition of CLR and the residual gives the desired output.
Network Depth
__I_.
X Ш I—
о
о >
Estimated channel which is considered as a noised image
,3 4 Pi 2 4 Pi + у Pi + 7 s ci + y с о 'и ai S, +
& m t m t m t
и В о о 1 о 1 О
Estimated channel
Figure 10- VDSR architecture Рис. 10 - Архитектура VDSR Слика 10- Архитектура VDSR
The VDSR architecture, depicted in Figure 10, consist of 20 layers where layers except the first and the last, are of the same type:
• The first layer operates on the input grid channel,
• Each of the 18 convolutional layers uses 64 filters of size 3 x 3 x 64 followed by ReLU, and
The last layer, used for grid channel reconstruction, consists of a sin- ^
o
co I
CO £± cp
<0 c o
Ln
gle filter of size 3 x 3 x 64 followed by the regression layer.
Results and discussion
In this section, all the networks introduced in Section 3 were trained. Following that, the Mean Squared Error (MSE) was evaluated across a range of Signal-to-Noise Ratios (SNRs). The setup involves a single antenna as ! both the transmitter and the receiver. The 5G Toolbox in Matlab was used ° for the channel modeling and pilot transmission. The training, testing, and validation sets comprised 40000, 5000, and 5000 channels respectively.
For the purpose of creating test scenarios, a slot period of resource grid cd consisting of 51 resource blocks was selected to form PDSCH data, forming a matrix of resource elements with dimensions 612 by 14. In order to map the pilots, a slot-wise type A mapping solution was adopted with | the DM-RS symbol position set to 2. Furthermore, a single DR-MS symbol was introduced, featuring an additional position of 1.It is worth noting that these parameters and decisions were made in accordance with the rigorous guidelines set forth by the 3GPP standard (3GPP. 2020b). ^
The parameters used for data generation are presented in Table 1. During this process, a sub-carrier spacing of 30 kHz was maintained, and the g actual data symbols were set to zero. Instead, only the DM-RS symbols were embedded in the data as displayed in Figure 11. For the data trans- a mission, a repeated transmission approach was employed. This involved looping through the data of a single slot period, which lasts 0.5 ms. By repeating the transmission within this time frame, the integrity and continuity of the data were effectively maintained. Finally, the collected data was par- i titioned into the training, validation, and test sets in order to train the CNN models.
"O
Training CNN based channel models
The performance of the neural network-based channel estimation methods relies on the SNR value. Ideally, the weights of the neural network should be optimized for each SNR value to achieve the best performance. However, in practice, this approach is not feasible since the SNR value is continuous, and retraining the network for every possible SNR value is computationally intensive.
о >
00 CM о CM
ОС
УУ 0£ ZD
о
О -j
<
О
х о ш
н
>-
СС <
ОТ
CD
S2 ■О z х ш н о
о
(a) DM-RS in transmitter (b) DR-MS in receiver
(c) interpolation
2 4 6 8 10 12 14 2 4 6 8 10 12 14 2 4 6 8
(d) estimated channel with (e) estimated channel with (f) estimated channel with SRCNN + DnCNN regression CNN VDSR
(g) perfect channel
Figure 11 - Resource grid images Рис. 11 - Изображения сети ресурсов Слика 11 - Слике мреже ресурса
Fortunately, training the neural network for a few representative SNR values can still yield satisfactory performance. In such cases, the neural network can estimate the channel for SNR values that are close to the ones it was trained on, and can interpolate to SNR values that are not covered in the training. Therefore, in our work, we have selected two ranges of representative SNR values for training the neural network, a range of discrete values [0,5] for low SNR and [20,25] for high SNR.
It is worth noting that for each of three methods, two models have been trained, one for low SNR and the other for high SNR values Also, the models were trained using the parameters specified in Tables 2,3,4,5, for each range of the Signal-to-Noise Ratio (SNR)
Table 1 - Parameters for PDSCH DM-RS data generation Таблица 1 - Параметры генерации данных PDSCH DM-RS Табела 1 - Параметри за генерисаше PDSCH DMRS података
Parameters value
PDSCH Mapping Type Type A
DR-MS TypeA Position 2
DM-RS Additional Position 1
DM-RS Configuration Type 1
Subcarrier Spacing 30 kHz
Cyclic Prefix Normal
Bandwidth in number of resource blocks 51
Model Channel TDL
Power Delay Profile All profiles
Table 2 - Training parameters for the SRCNN method Таблица 2 - Параметры обучения по методу SRCNN Табела 2 - Параметри обуке за SRCNN метод
Training Parameters Value
Solver for training network Adam (Adaptive Moment Estimation)
Batch Size 128
Initial Learn Rate 0.001
Max Epochs 5
CT) I
CT) £± Œ
(Л С
о
Е Е о о
(Л
(Л ф
CD ю
го Е
(Л ф
о CT
Training progress for low SNR values
From the Loss graph in figures below (12,13 and 14 ), we can see that both the training and validation losses decrease steadily over iterations, indicating that the model is learning effectively without over-fitting. The validation loss is consistently similar to the training loss, which suggests that the model is generalizing well to new data.
ro
JD
Œ Ф Ф
Q "rö ф
о
"О
о >
PO C\l о (M
ОС Ш 0£ ZD
о
О -j
<
О X
о ш
I—
>-
Q1 <
ОТ
CD
S2 ■О
х ш I—
о
о >
Table 3 - Training parameters for the DnCNN method Таблица 3 - Параметры обучения по методу DnCNN Табела 3 - Параметри обуке за DnCNN метод
Parameters Value
Solver for training network Sgdm (Stochastic Gradient Descent with Momentum)
Momentum 0.9
Initial Learn Rate 0.001
Learn Rate Schedule piecewise
Gradient Threshold Method absolute-value
Gradient Threshold 0.005
L2Regularization 0.0001
Batch Size 128
Max Epochs 30
Table 4 - Training parameters for the regression CNN method Таблица 4 - Параметры обучения по регрессионному методу CNN Табела 4 - Параметри обуке за регресиону CNN методу
Training Parameters Value
Solver for training network Adam (Adaptive Moment Estimation)
Batch Size 32
Initial Learn Rate 0.0003
Max Epochs 5
<и 2000
to >
§ 1000
Loss Training Loss Validation
60
> '
ш
от s ai
20
— RMSE Training -RMSE Validation
5000 iteration (a) Loss
10000
5000 10000 iteration (b) RMSE
15000
Figure 12- Training progress for regression model Рис. 12 - Прогресс обучения по регрессионной модели Слика 12 - Напредак у фази обучаваъа за регресиони модел
0
0
0
0
Similarly, from the RMSE graph, we can see that both the training and validation RMSEs displayed a consistent downward trend, indicating good learning and that the models were gradually fitting the training data.
Table 5 - Training parameters for the VDSR method Таблица 5 - Параметры обучения по методу VDSR Табела 5 - Параметри обуке за VDSR метод
Parameters value
Solver for training network Sgdm (Stochastic Gradient Descent with Momentum)
Momentum 0.9
Initial Learn Rate 0.1
Learn Rate Schedule piecewise
Learn Rate Drop Period 10
Learn Rate Drop Factor 0.1
L2Regularization 0.0001
Batch Size 32
Max Epochs 100
Gradient Threshold Method l2norm
Gradient Threshold 0.01
СП I
со £± Œ
(Л с о
Е Е о о
(Л
(Л ф
CD ю
Ф 1000
J3
(Л
>
(Я (Я
о
500
6000
e
al > 4000
s
s
o L 2000
1000 2000 iteration
(a) Loss for SR
3000
Loss Training Loss Validation
5000 10000 iteration
(c) Loss for Dn
15000
—Loss Training •Loss Validation 60 <D -3 50 TO — RMSE Training -RMSE Validation
1...... LU 40 w I30 1
20 IL.....
1000 2000 iteration
(b) RMSE for SR
3000
100
LU
w
cc
50
— RMSE Training -RMSE Validation
0.5
1
1.5
iteration (d) RMSE for Dn
x104
Figure 13 - Training progress for the SRDn model Рис. 13 - Прогресс в обучении по модели SRDn Слика 13 - Напредак у фази обучаваъа за SRDn модел
го Е
(Я ф
о CT
го
Œ Ф Ф
а
"то
Ф
О "О
0
0
0
0
0
0
2
0
In the initial epochs, the loss and RMSE for Regression, SRDn and VDSR show a rapid drop, suggesting that the models quickly learned from the training samples. However, after that, the rate of improvement slowed
down, and the training was stopped after the loss and RMSE curve flattened, indicating that the model had reached the limit of learning from data.
3000 2000 1000 0
0
5 10
iteration
x 10
15 4
80 J60
то
ш 40
со
|20 0
— RMSE Training -»-RMSE Validation
¡вам
0
(a) Loss
5 10
iteration
(b) RMSE
x 10
15 4
Figure 14 - Training progress for the VDSR model Рис. 14 - Прогресс в обучении по VDSR Слика 14 - Напредак у фази обучаваъа за VDSR модел
It is worth noting that for VDSR, the RMSE curve experienced some fluctuations, which could be attributed to the complexity of the dataset. However, the Regression and SRDn models could not capture this complexity.
Training progress for high SNR values
The figures below (15,16 and 17) present the Loss and RMSE progress
5000 iteration
(a) Loss
10000
80
e з 60
то
>
III 40
со
^
20
0
5000 10000
iteration
(b) RMSE
Figure 15- Training progress for the regression model Рис. 15 - Прогресс в обучении по регрессионной модели Слика 15 - Напредак у фази обучаваъа за регресиони модел
0
0
Comparing the Loss and RMSE graphs in the preceding figures (15,16 and 17 ), it is clear that the trends follow a similar pattern. The models show promising results, with no evidence of over-fitting or under-fitting.
As in the case of the low SNR, the training process presents a downward trend of the loss and RMSE functions, showing that the models were gradually fitting the training data in the same way as in the low SNR.
CO I
CO s± Œ
СЛ с о
<d 1000
to >
w 500 о
Loss Training Loss Validation
o, 0.0- 0 0 о ® t°* f *°t ° •Of • r°' 9 '01 °
0 1000 2000 3000
iteration
(a) Loss for SR
6000
<D
£ 4000
<f> U)
Я 2000
Loss Training Loss Validation
5000 10000 iteration
(c) Loss for Dn
15000
60
Ï540
>
ш
Ю
; 20
IT
100
<u
ZD TO
£ 50
5
or
0
RMSE Training RMSE Validation
- 0 1000 2000 iteration ° ' " 3000
(b) RMSE for SR
RMSE Training RMSE Validation
0
0.5
1
1.5
iteration (d) RMSE for Dn
x104
Figure 16- Training progress for the SRDn model Рис. 16 - Прогресс в обучении по модели SRDn Слика 16- Напредак у фази обучаваъа за SRDn модел
Е Е о о
(Л
(Я ф
CD ю
го Е
(Я ф
о CT
го
Œ Ф Ф
Q
"то ф
о
"О
0
0
0
2
0
The RMSE curve presents some fluctuations in the cases of Regression, SRDn and VDSR model training, which indicates the ability of models to capture the complexity of the channel in the high SNR.
ф
о >
00 C\l о (M
ОС Ш 0£ ZD
о
О <
О
х о ш
I—
>-
Q1 <
8000
6000
to
м 4000 ф
°2000 0
0
Loss Training Loss Validation
150
<D
100
>
ш
ю 5 QT
50
— RMSE Training -RMSE Validation
.—--
10 iteration
15
0.5
1
1.5
x104
(a) Loss
iteration (b) RMSE
x105
Figure 17- Training progress for the VDSR model Рис. 17 - Прогресс в обучении по модели VDSR Слика 17 - Напредак у фази обучаваъа за VDSR модел
In summary, it could be seen in both low and high SNR values that the VDSR presented fluctuation in RMSE during the training, which indicates a high adaptability to the complexity of the channel.
CO
s
CD
S2 >o z
X Ш I—
о
о >
Performance evaluation of CNN models using test data
The three methods (SRCNN + DnCNN, Regression CNN and VDSR) are evaluated on 5000 random channels in both low and high SNR conditions. Based on the provided RMSE (Root Mean Squared Error) values, their performance can be compared with the traditional method of LS (Least Squares). The results are presented in Table 6.
Table 6 - Performance evaluation of the CNN models Таблица 6 - Оценка производительности моделей CNN Табела 6 - Процена перформанси CNN модела
0
5
0
2
Model RMSE (Low SNR) RMSE (Hight SNR)
Least Square 2.0850 0.2425
Methodl: SRCNN + DnCNN 0.4776 0.1299
Method2: Regression CNN 0.4942 0.1006
Method3: VDSR 0.4797 0.0968
For the low SNR, the SRCNN + DnCNN method and the VDSR method have similar performances, with the RMSE values of 0.4776 and 0.4797, re-
spectively. The Regression CNN method has a slightly higher RMSE. How- ^
ever, all three methods significantly outperform the Least Square method.
For the high SNR, the VDSR method has the best performance followed by the Regression CNN method and the SRCNN + DnCNN method. Again, all three methods significantly outperform the Least Square method. o
In summary, the deep learning-based methods (SRCNN + DnCNN, Regression CNN, and VDSR) are more effective than the traditional Least Square method for channel estimation in both low and high SNR conditions. | Among the deep learning-based methods, VDSR appears to be the most effective for high SNR conditions, while SRCNN + DnCNN and VDSR have similar performance for low SNR conditions. The Regression CNN method has slightly lower performance than the other two deep learning-based $ methods, but is still significantly better than the Least Square method. These results demonstrate the effectiveness of deep learning-based methods for channel estimation in wireless communication systems.
CO £± cp
<0
ro E
(0 <u
Channel Estimation MSE in terms of SNR for different channel profiles
The accuracy of channel estimation can be evaluated using the mean square error (MSE) metric. The MSE is a measure of the average differ- <jj ence between the estimated channel and the actual channel, and it is com- g-monly used to compare different channel estimation methods. The MSE of channel estimation is affected by several factors, including the channel profile and the signal-to-noise ratio (SNR)
To illustrate the impact of channel profile and SNR on channel estimation for each of the three methods mentioned before, we have calculated the MSE for each of the scenarios, Non-Line-of-Sight NLOS (TDL-A, TDL-B and TDL-C) and Line-of-Sight LOS (TDL-D and TDL-E), in both low and high SNR conditions.
Channel Estimation MSE for NLOS communication
In the context of NLOS communication, where there is no direct line-of-sight between transmitting and receiving antennas, the signal travels along multiple paths to reach the receiver, causing severe signal attenuation, delay spread, and inter-symbol interference. The performance of the three
о >
00 C\l о (M
ОС Ш 0£ ZD
о
О <
О
х о ш
I—
>-
Q1 <
aforementioned channel estimation methods is impacted by the SNR values.
0.080
0.060
0.040
0.020
•DL Regression Я DL SRDn —DLVDSR
0 5 10
SNR
(a) SNR from 0 to 15
15
1 2 3 SNR
(b) SNR from 0to 5
Figure 18- Channel Estimation MSE in terms of a low SNR for NLOS Рис. 18 - Оценка канала MSE с точки зрения низкого SNR для NLOS Слика 18 - Процена канала MSE у код ниског SNR за NLOS
гл
CD S2 'У z х ш I—
о
о >
In very low SNR conditions (Figure 18), with a high number of multi-paths, the SRCNN + DnCNN and VDSR methods outperform the CNN regression method, with VDSR exhibiting slightly better performance. The superior performance of these deep architectures can be attributed to their ability to better capture the complexity of the channel model.
8.000 х10-3
\ -•-DL Regression
6.000 —DL SRDn
^ \ -•-DL VDSR
« 4.000 х » "
2.000 8 " ' > , , 8 8 - . ' 8 » 8 . , ,
0.000 ............
15 20 25
SNR
(a) SNR from 15 to 30
30
4.000 x10"3
• DL Regression DL SRDn
3.000 • DLVDSR
w 2.000
1.000 —^—^
20 21 22 23 24 25
SNR
(b) SNR from 20 to 25
Figure 19- Channel Estimation MSE in terms of a high SNR for NLOS Рис. 19 - Оценка канала MSE с точки зрения высокого SNR для NLOS Слика 19 - Процена канала MSE у код високог односа SNR за NLOS
о
4
5
However, as SNR values increase (Figure 19), the performance of the SRCNN + DnCNN method decreases drastically in comparison to the remaining methods. In contrast, the VDSR method continues to outperform all other methods.
Channel Estimation MSE for LOS communication
Line-of-Sight (LOS) scenarios are often preferred due to a clear, unobstructed path between transmitting and receiving antennas. In such scenarios, the signal travels directly between the antennas without being scattered or reflected by obstacles, resulting in minimal attenuation and distortion. As a result, channel estimation in the LOS scenarios is less challenging than in the NLOS scenarios.
CO I
CO £± cp
(Л с о
E E о о
(Я (Я JD <D
CD
Ю
0.060
ш 0.040
w
0.020
0 5 10
SNR
(a) SNR from 0 to 15
0.030
0.020
-•-DL Regression
- DL SRDn
—DLVDSR
в ^
• _
" s
8
15 0 12 3
SNR
(b) SNR from 0to 5
Figure 20 - Channel Estimation MSE in terms of a low SNR for LOS Рис. 20 - Оценка канала MSE с точки зрения низкого SNR для LOS Слика 20 - Процена канала MSE у смислу ниског SNR за LOS
However, even in the LOS scenarios (Figure 20), the accuracy of channel estimation is still impacted by SNR values. In a very low SNR values (SNR < 2), the deep CNN architectures (SRCNN + DnCNN and VDSR) outperform the simplistic architecture of CNN regression, due to their ability to capture the complexity of the channel model. The SRCNN + DnCNN and VDSR methods are better suited for achieving accurate channel estimation in such scenarios.
ro E
(Я
<u
о
ro
JD
cp
<U <U
Q "ro
<u
о
"О
4
5
о >
00 CM о CM
ОС
УУ 0£ ZD
о
О -j
<
О
х о ш
н
>-
СС <
ОТ
CD
S2 ■О z х ш н о
о
20 25
SNR
(a) SNR from 15 to 30
(b) SNR from 20 to 25
Figure 21 - Channel Estimation MSE in terms of a high SNR for LOS
Рис. 21 - Оценка канала MSE с точки зрения высокого SNR для LOS Слика 21 - Процена канала MSE код високог односа SNR за NLOS
As SNR values increase (Figure 21), the performance of the CNN regression method becomes more favorable, due to its simplistic architecture being well adapted to the low complexity of the channel. On the other hand, the performance of SRCNN + DnCNN decreases significantly due to the negative impact of its deep denoising architecture (DnCNN). The VDSR architecture, however, demonstrates remarkable adaptability to the channel complexity, resulting in stable performance across a range of SNR values.
Conclusion
By leveraging the power of deep learning algorithms such as CNNs, channel estimation in 5G communication systems can be improved significantly. This work has showcased the potential that CNN offers compared to the traditional method of the Least square for an accurate channel estimation.
First, by conducting a comprehensive review of the existing literature on CNN-based channel estimation, two of widely used methods were chosen, namely the super-resolution and denoising method (SRCNN+DnCNN) and the CNN regression method. Besides that, a novel method (VDSR, Very Deep Super Resolution) was proposed in order to improve the accuracy of the state-of-the-art CNN based channel estimation methods. The three CNN models were trained on a large dataset in both low and high SNR conditions.
The trained models were evaluated and the results were compared to ^
CO £± cp
<0
the traditional method of Least Square. The compared results have demonstrated the superiority of deep learning-based methods under varying SNR conditions. Moreover, the novel method exhibits the best overall performance in comparison to the two other deep learning-based methods.
Further, the impact of channel complexity on estimation accuracy was investigated in the case of the CNN based methods. The results highlighted the importance of selecting an appropriate channel estimation model based | on the specific communication scenario's complexity and SNR values.
In NLOS scenarios with very low SNR values and a high number of multi-paths, deep architectures such as SRCNN + DnCNN and VDSR outperform the CNN regression method due to their ability to capture the complexity of $ the channel model.
In contrast, in LOS scenarios, signal attenuation and distortion are minimal, making channel estimation less challenging. Nonetheless, the accuracy of channel estimation is still heavily impacted by SNR values, and deep CNN architectures such as SRCNN + DnCNN and VDSR remain better suited for achieving accurate channel estimation in very low SNR values.
As the SNR values increase, the CNN regression method exhibits improved performance due to its simplistic architecture that is well-suited to the low complexity of the channel. Conversely, the performance of SRCNN + DnCNN deteriorates significantly due to the adverse impact of its deep Q denoising architecture (DnCNN).
Notably, the VDSR architecture demonstrates remarkable adaptability to the channel complexity, resulting in consistent performance across all range of SNR values. This makes it a promising method for channel esti- ° mation in diverse 5G communication scenarios (NLOS and LOS).
In future work, we propose to extend the evaluation of the proposed method, VDSR (Very Deep Super Resolution), to Single-Input Multiple-Output (SIMO) and Multiple-Input Multiple-Output (MIMO) channel models for 5G wireless communication. The performance of VDSR has shown promising results in our current research, particularly in terms of its adaptability to varying channel complexities and SNR values. The extended evaluation will provide valuable insights into the performance and robustness of VDSR across different wireless communication setups, further enhancing its applicability and potential for real-world 5G deployments. Additionally,
O >
00 CM o CM
investigating the impact of various system parameters, such as the number of Additional DM-RS and DM-RS configuration types, on the performance of VDSR in SIMO and MIMO models will enable to optimize and tailor the method for specific wireless communication scenarios, paving the way for improved channel estimation techniques in future 5G networks.
q:
yy
id 3GPP. 2018. 5G, NR, Physical layer, General description, Technical spec-
References
3GPP. 2018. 5G
§ ification (3GPP TS 38.201 version 15.0.0 Release 15) [online]. Available at: ^ https://portal.3gpp.Org/desktopmodules/Specifications/SpecificationDetails.a
0 spx?specificationId=3211 [Accessed: 10 August 2023].
z
1 3GPP. 2020a. 5G, NR, Physical channels and modulation, Technical Spec-w ification (3GPP TS 38.211 version 16.2.0 Release 16) [online]. Available at: >_ https://portal.3g pp.org/desktopmodu les/Specifications/Specification Details. a !< spx?specificationId=3213 [Accessed: 10 August 2023].
3GPP. 2020b. 5G, Study on channel model for frequencies from 0.5to 100 GHz, Technical Report (3GPP TR 38.901 version 16.1.0 Release 16) [online]. Available at: https://portal.3gpp.org/desktopmodules/Specifications/SpecificationDetails.a spx?specificationId=3173 [Accessed: 10 August 2023].
Albreem, M.A.M. 2015. 5G wireless communication systems: Vision and challenges. In: 2015 International Conference on Computer, Communications, and o Control Technology (I4CT). Kuching, Malaysia, pp.493-497, April 21-23. Available at: https://doi.org/10.1109/I4CT.2015.7219627.
Banerjee, B., Khan, Z., Lehtomaki, J.J. & Juntti, M. 2022. Deep Learning Based Over-the-Air Channel Estimation Using a ZYNQ SDR Platform. IEEE Access, 10, o pp. 60610-60621. Available at: https://doi.org/10.1109/ACCESS.2022.3180352.
Dong, C., Loy, C.C., He, K. & Tang, X. 2015. Image Super-Resolution Using Deep Convolutional Networks. IEEE Transactions on Pattern Analysis and Machine Intelligence, 38(2), pp. 295-307. Available at: https://doi.org/10.1109/TPAMI.2015.2439281.
Gizzini, A.K., Chafii, M., Nimr, A., Shubair, R.M. & Fettweis, G. 2021. CNN Aided Weighted Interpolation for Channel Estimation in Vehicular Communications. IEEE Transactions on Vehicular Technology, 70(12), pp. 12796-12811. Available at: https://doi.org/10.1109/TVT.2021.3120267.
James, A.R., Benjamin, R.S., John, S., Joseph, T.M., Mathai, V. & Pillai, S.S. 2011. Channel estimation for OFDM systems. In: 2011 International Conference on Signal Processing, Communication, Computing and Networking Technologies. Thuckalay, India, pp.587-591, July 21-22. Available at: https://doi.org/10.1109/ICSCCN.2011.6024619.
Kaur, J., Khan, M.A., Iftikhar, M., Imran, M. & Haq, Q.E.U. 2021. Machine
en
Learning Techniques for 5G and Beyond. IEEE Access, 9, pp. 23472-23488. 1 Available at: https://doi.org/10.1109/ACCESS.2021.3051557.
Kim, J., Lee, J.K. & Lee, K.M. 2016. Accurate Image Super-Resolution Using £ Very Deep Convolutional Networks. In: 2016 IEEE Conference on Computer Vi- £ sion and Pattern Recognition (CVPR). Las Vegas, NV, USA, pp.1646-1654, June ° 27-30. Available at: https://doi.org/10.1109/CVPR.2016.182.
Ma, Z., Zhang, Z., Ding, Z., Fan, P. & Li, H. 2015. Key techniques for 5G | wireless communications: network architecture, physical layer, and MAC layer perspectives. Science China Information Sciences, 58(4), pp. 1-20. Available at: https://doi.org/10.1007/s11432-015-5293-y.
Morocho-Cayamcela, M.E., Lee, H. & Lim, W. 2019. Machine Learning for 5G/B5G Mobile and Wireless Communications: Potential, Limitations, and Future Directions. IEEE Access, 7, pp. 137184-137206. Available at: https://doi.org/10.1109/ACCESS.2019.2942390.
Simonyan, K. & Zisserman, A. 2014. Very Deep Convolutional Networks for Large-Scale Image Recognition. arXiv:1409.1556. Available at: e
https://doi.org/10.48550/arXiv.1409.1556.
Soltani, M., Pourahmadi, V., Mirzaei, A. & Sheikhzadeh, H. 2019. Deep Learning-Based Channel Estimation. IEEE Communications Letters, 23(4), pp. 652-655. Available at: https://doi.org/10.1109/LC0MM.2019.2898944. Ц
Wang, C.X., Bian, J., Sun, J., Zhang, W. & Zhang, M. 2018. A Survey of 5G Channel Measurements and Models. IEEE Communications Surveys & Tutorials, 20(4), pp. 3142-3168. Available at: https://doi.org/10.1109/C0MST.2018.2862141.
Ye, H., Li, G.Y. & Juang, B.H. 2017. Power of Deep Learning for Channel Esti- Q mation and Signal Detection in OFDM Systems. IEEE Wireless Communications Letters, 7(1), pp. 114-117. Available at: https://doi.org/10.1109/LWC.2017.2757490.
Zhang, K., Zuo, W., Chen, Y., Meng, D. & Zhang, L. 2017. Beyond a Gaussian Denoiser: Residual Learning of Deep CNN for Image Denoising. IEEE transactions on image processing, 26(7), pp. 3142-3155. Available at: https://doi.org/10.1109/TIP.2017.2662206.
Оценка канала глубокого обучения в 5G беспроводной связи
Мохамед Зуауи М. Лаидуни, корреспондент, Таки-эддине Ахмед А. Беняхия, Бобан З. Павлович, Салем-Билал Б. Амокране, Туати Б. Адли
Университет обороны в г. Белград, Военная академия, Департамент телекоммуникации и информатики, г Белград, Республика Сербия
ю
го
о >
00 гм о (М
CD
О >
РУБРИКА ГРНТИ: 49.33.29 Сети связи,
20.23.25 Информационные системы с базами знаний
ВИД СТАТЬИ: оригинальная научная статья
Резюме:
Введение/цель: За последние годы методы глубокого сс обуче- ния, в частности сверточные нейронные сети
^ (CNN), показали высокую производительность в системах
g связи 5G, значительно повысив точность оценки канала по
о сравнению с обычными методами. Целью данной статьи
< является всесторонний обзор существующей литерату-
° ры по методам оценки канала на основе CNN. Помимио то-
g го, статья нацелена на усовершенствование современных
ш методов оценки канала на основе CNN путем предложе-
>_ ния нового метода под названием VDSR (Very Deep Super
<с Resolution), вдохновленного методами изображения Super-
Resolution.
Методы: Для того чтобы оценить эффективность различных подходов было проведено всестороннее сравнение Тп с учетом различных сценариев, в том числе с низким со-
отношением сигнал-шум (SNR) и высоким SNR, а также в условиях прямой видимости (LOS) и вне прямой видимо-
>о сти. Учитывались также сценарии вне пределов видимо-
сти (NLOS). С помощью сравнительного анализа была произведена оценка эффективности существующих методов о и выявлены преимущества предлагаемого метода, осно-
ванного на VDSR.
Результаты: Результаты данного исследования показывают значительный потенциал оценки канала, основанного на CNN в системах связи 5G, при этом метод VDSR демонстрирует стабильную производительность во всех сценариях. Данное исследование способствует совершенствованию методов оценки каналов в сетях 5G, прокладывая путь усовершенствованным системам беспроводной связи с повышенной надежностью.
Выводы: ААрхитектура VDSR прекрасно приспособлена к сложности канала, что обеспечивает стабильную производительность во всем диапазоне значений SNR.
Ключевые слова: глубокое обучение, CNN, системы связи 5G, сверхглубокое сверхвысокое разрешение.
Процена канала дубоког уче^а за 5G бежичне 5
ОБЛАСТ: телекомуникаци|е, рачунарске науке
СП £± CP
<л"
комуникац^е
Мохамед Зуауи М. Лаидуни, аутор за преписку, Таки-еддине Ахмед А. Бен]ахи]а, Бобан З. Павлович, Салем-Билал Б. Амокране, Туати Б. Адли о
Универзитет одбране у Београду, Во^а академи]а, Катедра телекомуникаци]а и информатике, Београд, Република Срби]а
Е Е о о
(Я (Я
КАТЕГОРША (ТИП) ЧЛАНКА: оригинални научни рад ф
го
о
Сажетак: ю
Увод/цил>: Технике дубоког уче^а, посебно конволуционе неуронске мреже (CNN), последних година показале су из-узетне перформансе у 5G комуникационим системима та- £ ко што су знача]но побоъшале тачност процене канала у поре^ек>у са конвенционалним методама. У овом раду представлен jе свеобухватан преглед посто]еПе литературе о тех- никама процене канала заснованих на CNN-у. Поред тога, основни циъ рада }есте унапре^ива^е на}са-времени]их метода за процену канала заснованих на CNN-у, што jе резултирало предлагаем нове методе под нази-вом VDSR (Very Deep Super Resolution), инспирисане тех- щ никама Super Resolution слике. а
Методе: Да би се извршила процена ефикасности различи-тих приступа, спроведено jе свеобухватно поре^е^е ра-зличитих сценари}а, укл>учу]уГ1и низак однос сигнал-шум (SNR) и висок SNR, као и лини]у оптичке видъивости (LOS) и сценарио без видъивости (NLOS). Кроз ову компаратив-ну анализу проце^ене су перформансе посто}еЬих метода и истакнуте предности ко}е нуди предложена техника заснована на VDSR.
Резултати: На основу доби}ених резултата откривен jе знача]ан потенциал процене канала заснованог на CNN-у у 5G комуникационим системима, при чему VDSR метод по-казу}е константну предност у свим сценари}има. Основни циъ истражива^а }есте унапре^е^е техника процене канала у 5G мрежама, чиме се да}у основе побоъшаним бе-жичним комуникационим системима са веЬом поузданошПу.
о
"О
о >
00 C\l о (M
ОС ш
Закъучак: VDSR архитектура показу]е изузетну прила-годъивост различитим врстама канала, што резултира обезбе^е^ем захтеваних перформанси за све анализира-не вредности SNR.
Къучне речи: дубоко уче^е, CNN, 5G комуникациони систе-ми, веома дубока супер резолуци}а.
ос Paper received on / Дата получения работы / Датум приема чланка: 18.08.2023. О Manuscript corrections submitted on / Дата получения исправленной версии работы / ° Датум достав^а^а исправки рукописа: 27.11.2023.
< Paper accepted for publishing on / Дата окончательного согласования работы / Датум коначног прихвата^а чланка за об]ав^ива^е: 29.11.2023.
q © 2023 The Authors. Published by Vojnotehnicki glasnik/Military Technical Courier ш (http://vtg.mod.gov.rs, http://втг.мо.упр.срб). This article is an open access article distributed under
the terms and conditions of the Creative Commons Attribution license QC (http://creativecommons.org/licenses/by/3.0/rs/).
© 2023 Авторы. Опубликовано в "Военно-технический вестник / Vojnotehnicki glasnik / Military Technical Courier" (http://vtg.mod.gov.rs, httpV/втг.мо.упр.срб). Данная статья в открытом доступе и распространяется в соответствии с лицензией "Creative Commons" (http://creativecommons.org/licenses/by/3.0/rs/).
© 2023 Аутори. Обjавио Воjнотехнички гласник/Vojnotehnicki glasnik / Military Technical Courier (http://vtg.mod.gov.rs, httpV/втг.мо.упр.срб). Ово jе чланак отвореног приступа и дистрибуира се у складу са Creative Commons лиценцом (http://creativecommons.org/licenses/by/3.0/rs/).
<
ОТ
S
CD
-О icü ©
I ш I—
о
о >
