An o몭cial website of the United States government Here's how you know

Log in

 

As a library, NLM provides access to scientific literature. Inclusion in an NLM database

does not imply endorsement of, or agreement with, the contents by NLM or the National
Institutes of Health.
Learn more: PMC Disclaimer | PMC Copyright Notice

Heliyon. 2022 Apr; 8(4): e09317. Published online 2022 Apr 22.
doi: 10.1016/j.heliyon.2022.e09317

PMCID: PMC9061263 | PMID: 35520616

A review on machine learning and deep learning for various

antenna design applications
Mohammad Monirujjaman Khan,a,∗ Sazzad Hossain,a Puezia Mozumdar,a Shamima Akter,a
and Ratil H. Ashiqueb

▸ Author information ▸ Article notes ▸ Copyright and License information PMC Disclaimer

Associated Data

▸ Data Availability Statement

Abstract Go to: ▸

The next generation of wireless communication networks will rely heavily on

machine learning and deep learning. In comparison to traditional ground‑
based systems, the development of various communication‑based
based systems, the development of various communication‑based
applications is projected to increase coverage and spectrum ef몭iciency.
Machine learning and deep learning can be used to optimize solutions in a
variety of applications, including antennas. The latter have grown popular for
obtaining effective solutions due to high computational processing, clean
data, and large data storage capability. In this research, machine learning and
deep learning for various antenna design applications have been discussed in
detail. The general concept of machine learning and deep learning is
introduced. However, the main focus is on various antenna applications, such
as millimeter wave, body‑centric, terahertz, satellite, unmanned aerial
vehicle, global positioning system, and textiles. The feasibility of antenna
applications with respect to conventional methods, acceleration of the
antenna design process, reduced number of simulations, and better
computational feasibility features are highlighted. Overall, machine learning
and deep learning provide satisfactory results for antenna design.

Keywords: Deep MIMO, Beam‑forming, Machine learning, LOS, NLOS,

Antenna, DNN, CDF, GSCM, PDP, CNN, Millimeter wave, THz communications,
Body‑centric, Radio frequency, THz DL CT, Frequency, RFC, Meta‑material
identi몭ication

Deep MIMO, beam‑forming, machine learning, LOS, NLOS, antenna, DNN, CDF,
GSCM, PDP, CNN, millimeter wave, THz communications, body‑centric, radio
frequency, THz DL CT, frequency, RFC, meta‑material identi몭ication, THz‑TDS,
deep learning, 6G, Satellite, arti몭icial intelligence, Beam hopping, re몭lectarray,
Direct Broadcast Satellite (DBS), space communications, 5G, UAV, Cellular
networks, radio access network, reinforcement learning, unmanned aerial
vehicle, Radio Frequency Identi몭ication (RFID), GPS receiver, GPS spoo몭ing,
GPS meaconing, textile, RT, Nyusim, deep neural network, RSU, channel
estimation, large intelligent surfaces, channel extrapolation and FDD.

1. Introduction Go to: ▸

Machines are developing the abilities of humans, such as problem solving,

decision‑making, and learning. ML automates analytical model building
through data analysis. On the other hand, DL is an ML skill that helps
machines mimic human behavior by processing data. Through using ML and
DL, many applications can gain an advantage, and the antenna is one of them.
DL, many applications can gain an advantage, and the antenna is one of them.
As the complexity of antennas increases, ML and DL are used to optimize the
performance of the antennas. ML and DL have been used to create multiple
trained models for antenna design applications, allowing antenna design
applications to become more ef몭icient and rapid. With the help of high
computing power and software engineering capabilities, ML and DL for
different antenna designs have become the most important 몭ields of recent
research. Millimeter wave, body‑centric, terahertz, satellite, UAV, GPS, and
textile are just a few of the antenna design topics. Body‑centric is human to
human communication with the assistance of wearable antennas. Terahertz
frequencies are used for spectroscopy in different areas. Satellites are objects
that orbit around the earth and send communication signals. UAVs are
aircraft with ground‑based control, and textile technology focuses on textile
fabric made from textile 몭ibers.

Without the utilization of machine learning and deep learning algorithms,

antenna design is hard to design and maintain. There is a problem with
acceleration in antenna design without ML and DL. Maintaining low errors
and high productivity is dif몭icult without ML and DL. Without the support of
DL and ML reduction in simulation, preserving work feasibility and
calculation of antenna behavior is a very hard job to do.

In this section, the authors introduce all the applications for various antenna
design 몭ields by using ML and DL. Firstly, in [1], To increase the mean data
rate of a multi‑antenna wireless system and implement hybrid beam forming
in mmWave frequency bands, a Reinforcement Learning (RL) algorithm was
used to speed up the selection process of spatial beams. RL is an area of ML
used to maximize the notion of growing reward. In [7], taking maximum
advantage of ML with previous beam training information using locations,
nearest vehicles, and sizes of the receiver were used to learn the optimal
beam pair index. For research into mmWave or massive MIMO antennas, a
dataset is needed, so in [18], the dataset for mmWave or massive MIMO
antennas has been described. In [19], authors describe a hybrid beam
forming (BF) design for the downlink of multi‑user mmWave systems, in
which the number of AEs used at the base station to achieve BF bene몭its per
user is proportional to the user's distance. A machine learning framework for
learning environment‑aware beam‑forming codebooks for large‑scale MIMO
systems was developed. In [25], authors provide an overview of millimeter
wave channel concepts as well as an explanation of how map‑based channels
are classi몭ied. In [34], a system for future body‑centric communication was
developed using off‑the‑shelf non‑wearable devices such as Wi‑Fi routers,
network interfaces, and an omnidirectional antenna. In [25], authors provide
an overview of millimeter wave channel models as well as an indication of
how map‑based channels are classi몭ied. A THz DL computing Tomography
(CT) system is presented in section [37], capable of visualizing hidden objects
using a variety of material systems. A ML model Support Vector Machine
(SVM) was used to design and optimize re몭lect arrayantennas. To simplify
feasible beam hopping (BH) in multibeam satellite systems, a Dl‑based path
was developed in [47]. A full description of the ML design, the design of
collectors and relays, and a brief description of the choice of UAV types have
been given in section [73]. A machine learning‑based hybrid framework for
propagating both aleatory and epistemic uncertainties in antenna design is
proposed in [78]. In [93], a neural network (NN) dependent delay locked loop
(DLL) is established in the GPS receiver for multipath reduction. An overview
of the applications of machine learning and deep learning to the development
of various antenna designs has been presented in this paper. The used
methods and their outcomes have been presented. A comprehensive review
of different antenna designs, the general concept of machine learning and
deep learning, and ways of electromagnetic computation are also studied in
this paper.

The paper is structured in the following way‑Section 1 introduces the topic,

Section 2 discusses machine learning and deep learning for various antenna
design applications, Section 3 discusses analysis, and Section 4 concludes the
presented study.

2. Machine learning and deep learning for various antenna Go to: ▸

design application

Machine learning and deep learning are showing wonderful results in various
applications such as UAV, THz, textile, GPS, and Satellite. Its excellent
capabilities for learning representations in real environments make it more
suitable for applications. By using machine learning, the UAV is used for many
civilian purposes and many other purposes. On the other hand, body‑centric
communication systems also use machine learning and deep learning to
increase their capabilities. A review of recent reported uses and applications
of machine learning and deep learning for various antennas has been
performed in this paper.

2.1. Machine learning and deep learning for millimeter wave for antenna
design applications

The 30–300 GHz frequency band or the 1 cm to 1 mm wavelength range is

the millimeter wave (mmWave) region of the electromagnetic spectrum. For
the design of data transmission and sensing systems, the use of this
millimeter wave frequency has many advantages. Machine learning and deep
learning for millimeter wave applications have been described for different
applications. mmWave technology is used in various types of 몭ields. It has
huge unlicensed bandwidth and has great 몭lexibility and capability. Many
types of wireless applications use mmWave antennas, and by using machine
learning algorithms with the system, it becomes more 몭lexible. Machine
learning tools are 몭inding useful applications in both mm‑wave and massive
MIMO antenna design. The large antenna array designed for mm‑wave
systems is required for hybrid beam‑forming.

2.1.1. Wireless systems hybrid beamforming algorithm using reinforcement

learning To maximize the mean data rate of a multi‑antenna wireless system,
hybrid beamforming was implemented in millimeter wave (mmWave)
frequency bands. A Reinforcement Learning (RL) algorithm is also
demonstrated to speed up the selection of spatial beams.

There are two parts to a hybrid beamforming architecture. One is the analog
beamformer, and the other is the digital pre‑coder. The pre‑coding weights
for the various frequency portions of the baseband signal and 몭lexibility
happen when the digital pre‑coder connects parallel streams of input
symbols to RF transmission chains. The analog beamformer connects the
transmit antennas with the output of the RF blocks. For wideband RF signals,
because of its analog characteristics, the beamformer applies the same phase
shift to each antenna [2]. The hybrid beamforming scheme can be
implemented in many ways and in [3, 4], and [5], those ways are reported. A
hybrid beamforming algorithm, which is the focus of this section, can
maximize the earnable sum data rate of a mmWave Massive MIMO system.
The weights of the analog beamformer can only belong to a set of uniformly
quantized phase shift values for this purpose, and it is done jointly by the
digital pre‑coder and the analog beamformer to be used in transmission [6].
It is assumed that for a particular analog beam former, a lower dimensional
It is assumed that for a particular analog beam former, a lower dimensional
wireless channel can be gained. Using Singular Value Decomposition (SVD), a
lower‑dimensional wireless channel's capacity to transmit digital pre‑coder
can be derived. By using a brute force search for the given channel state, the
analog beamformer can be gained. It is possible to achieve this by moderating
the number of transmitted antennas and phase shift values of each antenna.
Reinforcement Learning (RL)was used to speed up the selection of the analog
beamformer in this section. RLis mentioned based on a Machine Learning
(ML) algorithm. The ML algorithm has earned experience from previous
work and, using that experience, this RL algorithm assesses the execution of
the candidate solution in every case of the process. Brute force search and the
mentioned RL algorithm show similar sum data rates. But for the RL
algorithm, fewer iterations are required.

2.1.2. Millimeter wave vehicular beam training with situational awareness In

this study, a feasible machine learning framework with situational realization
has been proposed. This study also proposed vehicle locations and sizes in
the domain to estimate the most favorable beam pair indexes. Three different
ways have been applied to with appropriate feature ordering in polar
coordinates, in Cartesian coordinates, and in grids of occupancy. Based on the
situational features to forecast the best beam pair index, various
classi몭ication methods have been compared. The remarkable development of
prediction accuracy has occurred because of model use of GPS inaccuracies,
the frequency of vehicle location reporting location errors in realistic
implementations, and connected vehicles' various penetration rates.

A wide evaluation has been introduced for the newly introduced beam
selection path. In this study, importance was given to the alignment
probability and achieved throughput matrices. The beam selection
performance has been compared with various machine learning models. The
prediction performance at various levels is then evaluated by varying the
number of vehicles in the feature. Noisy features of some realistic issues are
also discussed in the study. With no GPS error, a straightforward feature set is
encoded in Cartesian coordinates. Table 1 shows the classi몭iers utilized in this
investigation. From the table, it can be seen that the random forest got an
85.14% alignment probability, which is a better result than other classi몭iers.
With the alignment probability, the earned throughput does not scale, and
though Naı̈ve Bayes or gradient boosting have less alignment probability than
random forest, there is no main difference between their throughput. This
happens because several beams of power are almost the same. The model is
good enough to 몭ind good beams, though the accuracy is not 100%.

Table 1
Beam selection alignment probability and achieved throughput ratio with different
classi몭iers [7].

PA (%) RT (%)

Naı̈
ve‑Bayes 59.31 91.14
AdaBoost 45.80 75.05
RBF‑SVM 55.89 89.32
Gradient Boosting 69.05 96.49
Random forest 85.14 98.32

Open in a separate window

It is advantageous to sacri몭ice some optimality in order to achieve low

overheads, and this is especially important for millimeter‑wave vehicular
systems. In the study, system performance is compared and evaluated at
various levels of situational awareness.

2.1.3. Beam alignment in millimeter wave massive multiple input multiple

output (MIMO) To earn data transmission and directional beam alignment
(BA), a massive multiple input multiple output (MIMO) antenna array can be
utilized to reduce the high path loss of the mmWave signal. Being packed into
small form factors can be done by large antenna arrays because the
millimeter wavelength is short enough. Multiple users with multiple beams
are all served by the base station (BS) in an mmWave multi‑user multi‑
stream system. Beam training based on a theoretical code‑book is generally
utilized to align beams for various users. For instance, in [9], authors
suggested a method named "adaptive compress sensing (ACS). In [10], there
is a method that works faster than the ACS method, and that method is the
hierarchical search (HS) method. In [11], to combine the bene몭its of the ACS
and the HS, a multi‑path decomposition and recovery (MDR) method has
been proposed. Using a hierarchical codebook to generate beam training is
not unimportant for aligning beams for a large number of users. Using a
hierarchical codebook to generate beam training is not unimportant for
aligning beams for a large number of users. For all users, the BS must
position beams sequentially. Time is wasted when, for every layer of the
hierarchical codebook, the optimal code word index is fed back. Sensors or
radars might provide the BS with information about the user's location. In
mmWave systems, those sensors and radars are used for beam alignment,
and they will have greater hardware overhead. An alignment process with
partial beams leveraging AMPBML has been proposed in this paper for the
multi‑user mmWave massive MIMO system. The NN for the AMPBML is
trained of몭line using simulated environments according to the mmWave
channel model. The NN is then launched live to estimate the beam
distribution vector using incomplete beams. The beams for all users are then
aligned at the same time using the obtained indices of the dominating entries
of the beam distribution vector. In this study, beam alignment for all users at
the same time has been done by a hierarchical codebook and remarkably
saves the entire training time. In this study, there is no need for previous
knowledge of user location details to train NN, and remarkably, the system
overhead will be reduced. The AMPBML shows better results than existing
methods in terms of spectral ef몭iciency and total training time slots. The
better results of AMPBML also include hierarchical search, multi‑path
recovery, multi‑path decomposition, and adaptive compressed sensing.

2.1.4. Millimeter wave massive multiple input multiple output for hybrid
precoding For future communications, millimeter wave massive multiple
input multiple output (MIMO) is a remarkable solution. In mmWave MIMO, to
minimize the complexity of the hardware and energy spending related to
components of mixed signal, digital pre‑coding and hybrid analog are
important methods. Present hybrid pre‑coding schemes have high
computational complexity and cannot use spatial information. In the
proposed solution, every speci몭ication of the precoders to get the most
effective use of a deep neural network, the decoder is treated as a mapping
relation (DNN). In this study, hybrid precoding and a deep learning
architecture that supports mmWave massive MIMO were applied. The
importance of this study is that it can enhance the spectrum feasibility of
mmWave massive MIMO and also minimize the bit error ratio (BER). For
those above results, the required computational complexity becomes less,
and hybrid pre‑coding gives better performance than conventional schemes.

In [12], a framework for the mmWave MIMO system is being created by

In [12], a framework for the mmWave MIMO system is being created by
combining deep learning (DL) with hybrid pre‑coding. Here, this model is
viewed as a deep network and a black box (DNN) as an auto coder. In the
black box, activation functions create corresponding mapping relations by
optimizing multiple layers of the network. Through the training stage, DNN is
able to lower the computational time. This happens when, through the
training stage, DNN captures structural details of the hybrid pre‑coding
scheme.

The performance of the DNN‑based mmWave massive MIMO method has

been explored in [12] using numerical analysis. The DNN framework is a
process and is built by Keras. Observing the models in [13], the channel
model was created without loss of generality. The BER performance is
compared with many standard methods. The BER was also analyzed using
different batch sizes of the training dataset and learning rates. In the
simulation, the network has been trained for about 45,000 iterations. To
verify better results, the fully digital SVD‑based pre‑coding method,
completely GMD‑based pre‑coding method, SVD‑based hybrid pre‑coding
scheme, and the traditional schemes have failed to maintain performance
against the deep learning‑based method. The BER result is compared with
several batch sizes so that the mmWave massive MIMO scheme's
performance can be analyzed. Then, with various learning rates, the DNN
based mmWave massive MIMO scheme analyzes the BER versus SVR. The
spectrum feasibility result has been given in [12] for the completely digital
GMD based pre‑coding scheme, the SNR of the DNN‑based hybrid pre‑coding
scheme, and the spatially sparse pre‑coding method. As the SNR is enhanced
in all the schemes, the spectrum feasibility also gets better. Exploration of the
relationship between the iterations of the deep learning‑based strategy
analyzed with the analog pre‑coding scheme and MSE. The above
investigation was carried out to determine the stability and performance of
the suggested hybrid pre‑coding approach.

2.1.5. Hybrid precoding for wideband millimeter wave massive MIMO systems
Millimeter wave (mmWave) massive multiple input multiple output (MIMO)
has been proposed as a viable solution for future Internet of Things (IoT)
data rates. Hybrid precoding is a viable result for mmWave large MIMO
systems without a notable sum rate loss to reduce the number of radio
frequency (RF) chains. The current study is evaluated using an unrealistic
narrowband mmWave channel basedon using hybrid precoding or, on the
other hand, the high resolution (HR) phase shifters (PSs) with huge power
other hand, the high resolution (HR) phase shifters (PSs) with huge power
waste on hybrid precoding. For practical frequency selective wideband
mmWave large multiple input multiple output systems, an energy ef몭icient
hybrid pre‑coding approach based on one bit PSs has been investigated. A
cross‑entropy optimization (CEO) based hybrid pre‑coding strategy to
optimize the earnable sum rate of the reviewed system has also been
published as the CEO algorithm for machine learning breakthroughs. In the
case of HR‑PSs in general, the suggested CEO‑dependent hybrid precoding
plot from the event with one‑bit PSs has been enlarged to show that the
solution may be used in various plots. In terms of energy feasibility and near‑
optimal sum rate, the presented systems outperform some convocational
methods.

The millimeter wave has unexplored and good spectrum resources. To ful몭ill
the high data rate, the necessity for IoTmmWave is considered. Extreme
propagation loss because of short wavelengths happens when high frequency
mmWave signals are at 30–300 GHz, and this problem can be solved by high
antenna array gain. The short wavelengths of mmWave communications are
assigned high antenna array gain [15]. A new problem could arise from the
use of a massive antenna array. In a sub‑6 GHz MIMO system, for example, a
dedicated radio frequency (RF) chain is typically required for each antenna to
implement fully digital pre‑coding [16]. For RF chains in mmWave, totally
digital pre‑coding is not sustainable [17]. A solution can be obtained from
analog pre‑coding with a RF chain, but it does not support multiplexing.
Hybrid pre‑coding has been suggested for multiplexing and sum rate of total
digital precoding. A portable‑sized digital pre‑coder is enough to work out
spatial multiplexing because of the low rank of mmWave channels. This
hybrid pre‑coding can get an almost optimal sum rate.

In this study, proposed solutions were compared with various solutions. In

[14], energy feasibility and in terms of sum rate, CEO‑based hybrid pre‑
coding has been explained. ACEO‑based hybrid precoding with one bit PSs
has been proposed for routinely selectable wideband mmWave large multiple
input, multiple output systems. According to the results of the investigation,
one bit PSs based hybrid pre‑coding uses less power. Utilizing one‑bit‑PSs to
experience array gain loss is done, and the result remains limited and
constant. This answer proposes a CEO‑based low complexity method to
address the sum rate maximization problem. For the sum rate maximization
problem, one bit PSs dependent hybrid precoding was built under practical
control.

2.1.6. A generic deep learning dataset for massive MIMO antenna The
researchers presented the deep MIMO dataset, which is essential for any
research. They provide a dataset of mm waves or a massive MIMO antenna
design dataset. This is actually the generic dataset for mm wave‑antennas.
They also give detailed information about the structure of the generic dataset
of the massive MIMO antenna. They provide some information about the
channel's dataset design.

2.1.7. Multi‑user hybrid beam‑forming relyingon learning‑aided link‑

adaptation This study is based on a hybrid beam‑forming architecture for
multi‑user mmWave systems' downlink, in which the number of antenna
elements used at the base station to achieve beam‑forming gains per user is
proportional to the user's distance. The design is based on simulations that
show that the proposed learning assisted in adapting the target bit error rate,
resulting in a much higher data rate than traditional link‑adaptation based on
signal to noise ratio threshold values.

2.1.8. Learning beam codebooks with neural networks: towards environment‑

aware In [20], a machine learning framework designed to learn
environment‑aware beam‑forming codebooks for large‑scale MIMO systems
is presented. It is based on a neural network model that employs hardware
limitations as well as learning beam codebooks from the environment and
the user's location. This learning platform aids in the reduction of codebook
size and can result in signi몭icant improvements over traditional codebook
design.

In [20], the authors provide a hardware constraint on large‑scale MIMO

systems as well as an arti몭icial neural network‑based framework for learning
environment‑aware beam‑forming codebooks. For the surroundings and user
location, machine learning patterns have been used. Designing beam‑forming
codebooks has become an important research topic in academia and industry
[21]. However, in large scale MIMO systems, the hardware limitations of
mmWave/THz and the use of analog‑only or hybrid transceiver architectures
imposed new constraints on codebook design problems. This motivated the
development of new beam‑forming codebooks [22, 23], and [24]. In the
system model, a millimeter wave BS (base station) equipped with M antennas
can communicate with a single antenna user. In the machine learning
approach, supervised learning is used. For the solution, there have been two
approach, supervised learning is used. For the solution, there have been two
communication scenarios. The 몭irst scenario is the Line of Sight (LOS)
scenario, which is an outdoor scenario when users meeta LOS connection
with BS. The other one is the None Line of Sight (NLOS) scenario, which is an
indoor scenario and will happen when a user does not get a LOS connection.

In the end, the learned codebook in the LOS scenario with 64 beams hits
about 90% of its upper bound, and with 128 beams, it reaches about 95% of
its upper bound. It is very important in areas where analog phase shifters'
resolution is limited. From the NLOS results, we can see that with 64 beams,
learned codebooks reach about 90% of their upper bound. These results
assure the similarity of the framework to learning beam codebooks that can
optimize the size and beam patterns.

2.1.9. Map‑based millimeter‑wave channel models In [25], an overview of

millimeter wave channel models is offered, as well as the classi몭ication of
map‑based channels. Map‑based models should be used in different modeling
applications in the millimeter wave‑range and they can be used as a
supplement to SW test beds as they can support HW measurement.

Mobile broadband, mMTC (massive machine‑type communication), and

ultra‑reliable low‑latency communication are the technologies preferred for
mobile broadband. Researchers have been debating potential frequency
bands to serve such applications [26]. Map‑based mm‑wave design channel
models that can utilize RT (Ray‑tracing) have gathered momentum. It can
also model irregular cell layouts and support new application link types such
as D2D (device to device), V2X (vehicle to everything), and A2X (application
to everything) (air to everything). RT is used to produce multipath channel
parameters in map‑based channel models (also known as site‑speci몭ic
propagation models) [27, 28]. Figure 1, shows the map‑based model with its
characteristics. It represents the area and scenario of selected applications
and adopted technologies.

 Figure 1

Map‑based model with its characteristics [25].

The database used contains a large number of snapshots for training sets
used in machine learning methodologies [7, 28]. In the dataset, a DNN (deep
neural network) based beam selection algorithm is used, which is a machine
learning algorithm. Figure 2 shows that there is an overall performance
evolution of map‑based proposed models. In part 2, they used DNN‑based
beam selection algorithms and simulations of parameters. In the proposed
algorithm, CDFs (Cumulative Distribution Functions) are used, and GSCM
(Geometry‑based Stochastic Channel Model) is used in the training sets and
database.

Figure 2

Performance evolution of map‑based channels [25].

Different models show different results. The beam selection algorithm has
low accuracy, which is 12.8% using CDF, and the DNN‑based beam algorithm
using PDP (Power Day Pro몭ile) accuracy is 45.2%, which is more 몭lexible. So,
it will be very ef몭icient.

2.1.10. Network analysis using millimeter‑wave narrow‑band energy traces In

[29], a model for evaluating a machine learning framework for performing
protocol layer analysis and diagnosing physical layer faults in 60 GHz
networks was provided. The major goal is to provide a machine learning
framework that can appropriately classify transmitted networks and aid in
the detection of network faults. The main focus of this type was on millimeter
wave antennas and large bandwidth.

In essence, this model is a machine learning framework that uses template

matching and EDHMM to infer protocol layer information automatically. The
major aim was to determine the structural elements of the unpredictable
behavior by analyzing the variability of channel traces. The challenge was
solved using a directional antenna and a machine learning system.

2.1.11. Long‑range gesture recognition using millimeter wave radar A long‑

range gesture recognition model based on mm‑wave radar is provided in this
research. It is 몭lexible in human‑computer interaction (HCI).

Contactless gesture detection is a common way to achieve natural human‑

computer interaction (HCI) for a better experience, hence “in air” gestures
will increasingly replace external physical gesture devices [30, 31]. Wireless
communication has become the main focus of HCI. From all the other waves
(sonic wave, WIFI signal, ultrasound wave), mm wave is the most suitable
choice for this model. This design is based on a machine learning algorithm
called CNN (Convolutional Neural Network). The 몭irst step is to create a long‑
range gesture detection model by extracting spatial‑temporal aspects of the
hand's re몭lection spots. After that, CNN is utilized to learn the attributes of
the points for recognition. It can then recognize gestures automatically by
using a millimeter wave radar sensor to implement the model. Figure 3
shows three modules of the proposed model. In Figure 3, they are signal
transformation, information extraction, and the neural network. In hardware,
Section MM Wave radar is used. In radar, the 3TX and 4RX antennas are used
in sensing.
 
Figure 3

Three modules (signal transformation, information extraction, and the neural

network)of the proposed model [102].

To verify the model, three real scenarios are used. For the test, some furniture
was used in it. In the 몭irst scene, two participants stood 2.4 m away from the
radar and were required to repeat each of four gestures for 30 min in order
to collect 60 gesture data points. The 몭irst scenario is shown in Figure 4 (a).
Then, in the 2nd and 3rd scenes, the two participants were at the same
distance and just the placement of the chair and table was changed, as shown
in Figure 4 (b‑c).

Figure 4

Three real test sceneries of the model [2]. (a) First Scenario, (b) Second Scenario, (c)
Third Scenario.
From the output, the accuracy result was as in Table 2. The results of rotation
scenarios are almost the same and have a high degree of accuracy.

Table 2
Accuracy result of three scenarios [102].

Types/accuracy Knock Left swipe Right swipe Rotate

Scene 1 48.33% 38.33% 51.67% 98.00%

Scene 2 35.00% 5.00% 23.33% 96.67%
Scene 3 38.33% 3.33% 30.00% 98.33%

Open in a separate window

The 몭irst three gestures' accuracy decreased greatly, and it is concluded that
the external environment affects the model the most.

2.1.12. Large intelligent surfaces aided mm‑wave massive MIMO systems by

deep channel learning In this paper, a deep learning strategy for channel
estimation in large intelligent surfaces (LIS) with massive MIMO is proposed
(multiple input multiple output). A twin convolutional neural network (CNN)
architecture is created and maintained with the received pilot signals to
estimate both direct and cascaded channels, and each user gets access to the
CNN to estimate their own channel if there is a multi‑user situation. With
state‑of‑the‑art deep learning based techniques, the performance of the
proposed deep learning system is compared, and it shows better results. To
obtain a vigorous estimation execution, several channel realizations are used
to train the deep network. A separate set of test data is generated. Training
data in the prediction stage, is used to con몭irm the performance. The existing
DL based techniques [104, 105] are outperformed by the proposed deep
learning framework and achieve reasonable channel evaluation accuracy. As a
result, the suggested DL approach has demonstrated robust channel estimate
performance, which is tolerant of changes in user positions of up to four
degrees.

2.1.13. Deep learning based antenna selection for channel extrapolation in FDD
2.1.13. Deep learning based antenna selection for channel extrapolation in FDD
massive MIMO Massive multiple‑input multiple‑output (MIMO) systems
include a large number of antennas, which makes obtaining accurate channel
state information dif몭icult, especially in the frequency division duplex mode.
As demand for data transfer rates rises, massive multiple‑input multiple‑
output (MIMO) systems have emerged as a crucial technology for the next
generation of wireless communication [107]. Massive MIMO presents a
signi몭icant issue for the base station (BS) in obtaining correct channel state
information (CSI), particularly in frequency division duplex (FDD) mode
[108]. For the FDD massive MIMO system, Yu et al. [109] designed an ef몭icient
downlink channel reconstruction approach. The distance between antenna
elements in a massive MIMO system can be narrow enough that the channels
have a signi몭icant correlation. The fundamental advantage of DL‑based
channel extrapolation is that it does not require an accurate model and may
easily combine existing antenna selection approaches [110, 111, 112, 113].
As a result, proper DL‑based antenna selection should be designed, as should
effective downlink channel extrapolation from partial uplink channels. In this
study, we used probabilistic sampling theory to characterize discrete antenna
selection as a continuous and differentiable function. For comparison, the
performance of the DL and uniform antenna selection‑based channel
extrapolation is studied. Our proposed method outperformed the DL and
uniform antenna selection‑based schemes in simulations, and it was able to
properly handle signi몭icant frequency gaps and uplink channel estimate
inaccuracies.

2.1.14. 5G MIMO data for machine learning: application to beam‑selection

using deep learning In [114], a speci몭ic dataset for investigating beam‑
selection techniques on vehicle‑to‑infrastructure using millimeter waves has
been described. It presented a traf몭ic simulator that combined a vehicle
traf몭ic simulator with a ray tracing simulator for generating channel
realizations that are represented in the 5G scenario. It was designed to
update the features of the traf몭ic simulator. A speci몭ic dataset was used for
investigating beam selection techniques. In this, many other modeling
techniques using nyusim and quadriga [115, 116] were compared with RT. In
this research, RT (Ray‑tracing simulation) was used. RT can generate data for
two key requirements. RT was convenient in this scenario, and different types
of deep learning algorithms were used in the data processing. Among all of
them, random forest and deep neural networks have given almost 60%
accuracy. As the main focus was on mmWave MIMO, RSU antenna arrays were
used for transmitting and receiving data. Future work on this paper is to
make it more convenient and cost‑effective.

2.2. Machine learning for body‑centric communications

In the last few years, wearable body‑centric communication systems have

increased their applications and their areas. These systems are used in
various applications like healthcare, sports, military, identi몭ication systems,
smart phones etc. The applications are given below.

2.2.1. Body‑centric for THz networks THz communications are being

celebrated as the key enablers for wireless communication systems. Recently,
THz has been used so much in in‑body and on‑body communications. In this
design, they actually talked about the THz band for body‑centric
communications and its technologies, channel, noise modeling, modulation
schemes, and network topologies. From this paper you get some description
of the THz sensing and imaging applications in the healthcare sector. It is very
necessary to think about the RHz band to 몭ight this pandemic. COVID‑19 is
impacting humans and the overall worldwide economy. Finally, the body‑
centric THz application design gives knowledge about using the THz band for
in‑body and on‑body communications. In this paper they examine the THz
band noise, modeling, modulation etc.

2.2.2. Human muscle mass measurement through passive 몭lexible UWB‑

myogram antenna sensor to diagnose Sarcopenia Human muscle mass
measurements are a hot issue in the antenna area these days. As a result, the
researchers discussed Sarcopenia in this design. The researchers
demonstrate a non‑invasive, passive 몭lexible Ultra‑Wide Band (UWB)
myogram antenna sensor for predicting sarcopenia using human muscle
mass assessment in this design. The result is a change in muscle 몭iber shape
[60].

In this case, a non‑invasive method for predicting sarcopenia by measuring

skeletal muscle mass using passive 몭lexible UWB‑Myogram antenna sensor
signals from various muscle locations was employed. Sarcopenia is diagnosed
using a variety of diagnostic criteria, including quantitative and qualitative
muscle mass measures. Furthermore, earlier detection of skeletal muscle
mass measurement prevents metabolic side effects such as diabetes,
depression, abnormal cholesterol levels, and weight gain. They use three
ways to assess proteins, all of which are associated with using machine
learning, such as linear regression for prediction data.

Sarcopenia is caused by a lack of skeletal muscle mass in humans. Dual

energy X‑ray absorption has recently been used to quantify skeletal muscle
mass with some restrictions. The use of a passive 몭lexible UWB‑Myogram
antenna to quantify skeletal muscle mass has been demonstrated. Sarcopenia
can be predicted by measuring lean mass after fat signals have been
separated from skeletal muscle mass using an NMF 몭ilter. For qualitative
assessments, the proposed method eliminates the use of empirical
calculations of lean mass. According to the prediction equation, a protein
value of fewer than six indicates the presence of Sarcopenia, while a value of
less than 몭ive indicates a severely affected person.

2.2.3. Privacy‑preserving non‑wearable occupancy monitoring system

exploiting Wi‑Fi imaging for next‑generation body centric communication The
main focus of this application was on new, non‑wearable, device‑free,
privacy‑preserving wi‑몭i imaging‑based occupancy detection systems for
future smart buildings. Wireless and wearable gadgets are being developed
to develop the next generation of communication networks. They discussed
the detection of a person's existence during their daily activities without
deploying on the person's body in this study.

2.2.4. Deep learning framework for subject‑independent emotion detection

using wireless signals In this method, unique noise 몭iltering techniques are
used to gather most of the individuals' heartbeat and breathing signals from
radio frequency re몭lections off the body. Deep learning approaches are also
employed for comparing their 몭indings. Their proposed wireless emotion
detection gadget could also be used with ECG data in this paper.

2.2.5. Antennas and propagation for body‑centric communications Body‑

centric communication systems play an important role in the 4G generation
of mobile communication systems. In this paper, authors provide details
about the current position of body‑centric communication systems.

Antennas for body‑centric communications have been summarized well in

recent publications [73, 74], including antennas for 10 MHz body surface
communications [75] and button antennas [76, 77, 78]. Bandwidth is
determined by the system or by spectrum allocations. It is very dif몭icult to
specify the radiation pattern requirements. The use of various sectors, like
medical sensing and support, with either skin‑mounted sensors or implants,
is also getting attention. In this design, they summarize the antennas and
propagation for body‑centric communication systems.

2.3. Machine learning for THZ communication system

The frequencies of THz are used for spectroscopy in different types of areas.
The THz antenna can transmit and receive THz electromagnetic waves in the
THz system as it has some features like wide frequency, small size bandwidth,
and high rate. The THz frequency has signi몭icance in met material
identi몭ication, the 6G network, visualizing hidden objects and beam selection.
As we know, network systems are improving day by day, so 6G is the future
network technology and is 몭lexible. Met material and other hidden objects are
important as we are not able to visualize some of the objects that are hidden.
Beam selection is important as hybrid beamforming is very important for
overcoming the attenuation that is created by the extremely high frequency
in the THz band.

2.3.1. Terahertz deep learning computed tomography A THz DL CT (Terahertz

Deep Learning Computing Tomography) system that is capable of visualizing
hidden objects with multiple‑material systems is presented. Figure 5 shows
the schematic diagram of the experimental setup, and Figure 6 is the
schematic diagram of the THz DL CT model. 

Figure 5

Schematic diagram of THz TDS system [37].

 
Figure 6

Schematic diagram of the THz DL CT model [37].

The 몭inal results of the comparison of THz CT and THz DL CT models. The
THz DL CT model may use kernel 몭ilters to recreate superior images in the
high spatial frequency area, which is useful for visualizing the interior
structure of 3D objects. As previously stated, the MSE (mean square error) of
the THz DL CT model was 1.86 percent, which is lower than the standard THz
CT model. The 몭inal result, shown in Figure 7 (a‑c), demonstrate that THz DL
CT is a model capable of visualizing concealed objects using material systems.

 Figure 7

(a) Comparison between THZ CT and THz DL‑ CT. (b) Numerical metrics on two
algorithms, (c) Visible image and 3D THz images by THz DL‑CT on a testing object [37].

2.3.2. Low complexity beam selection scheme for terahertz systems When
compared to some existing beam selection schemes, a proposed beam
selection model that uses a machine learning algorithm that is an RFC
(Random Forest Algorithm) based beam selection scheme is capable of
providing a better arrangement between sum‑rate and complexity by
choosing the proper parameter settings. In this paper, it is considered that
there is a THz multi‑user uplink system featuring hybrid beamforming
architecture on both the base station and user sides. The channel
characteristics of L propagation pathways are assumed to be approximated
and known on the base station side in this model, and the channel‑related
additional content is not taken into account. There is an exhaustive search
approach for the maximum sum rate that can compute the sum rate under all
beam combinations and locate the best transmitter and receiver pair. They
also utilized the SVM model to see the results in this study, and after applying
it, it was discovered that the SVM model causes data bias, which undermines
the balance of two data sets. As a result, the RFC model's training set is
similar to the SVM model's. This paper covered the communication problems
and the machine learning approach that will help to improve the 5G
communication system as well.

2.3.3. Secure deep learning for intelligent terahertz metamaterial identi몭ication

It's a system that uses the THz technology and the crypto‑oriented CNN
model to detect the presence of metamaterial in mixtures. Figure 8 shows the
work몭low of private‑preservation THz metamaterial identi몭ication. In the THz
work몭low of private‑preservation THz metamaterial identi몭ication. In the THz
TDS (Time Domain Spectroscopy), a commercial photoconductive antenna is
used to get electromagnetic response signals. The THz wave passes through
the two lenses. To obtain huge amounts of data, random augmentation was
employed in accordance with probable noise, and these augmented signals
were translated to the frequency domain using a quick Fourier
transformation. In the training step, CNN can learn to discriminate between
features. After the network has been trained, it will be able to detect the
presence of metamaterial. They encrypted the original data before passing it
through the network and receiving all the results in ciphertext. This
ciphertext can also be decrypted independently.


Figure 8

Work몭low of private preserving THz metamaterial identi몭ication [39].

In the end, to compare the results, they used the SVM algorithm, the human
baseline, and CNN. In the human baseline, the mean accuracy was 56.97%,
the SVM method's accuracy was 87.9%, and CNN had 100% accuracy on
every fold. These experiments were already done by SVM, but in this paper
they proved that deep learning with CNN gives better accuracy in identifying
the existence of metamaterials in mixtures.

2.3.4. 6G wireless communications: vision and potential techniques It is a

design of the potential requirements and an overview of 6G mobile networks.
As we can see from 2G to 5G, the progress of mobile communication
networks is centered on serving people. 5G technology allows for a latency
time of 1 ms, and for that, 6G aims to make it less than 1 ms or nonexistent or
undetectable latency. 6G is designed so that it can be more 몭lexible than 3G to
5G. To increase data throughput, 6G will use a higher frequency spectrum
5G. To increase data throughput, 6G will use a higher frequency spectrum
than previous generations. It will outperform all other technologies in terms
of latency and architectural adjustments. For THz communication, UM‑MIMO
and PM‑MIMO techniques are used. In Figure 9, it is the design of 6G based on
time‑frequency‑space resource utilization.

Figure 9

6G based on the time‑frequency‑space resource utilization [40].

It will be possible to obtain it using machine learning approaches such as

classi몭ication or neural networks, rather than any other calculation. As a big,
data‑driven network, 6G will be able to manage large amounts of data.
Figure 10 has some promising features.

 Figure 10

Some promising techniques of 6G network [40].

As given techniques, there are some power supply issues, network security
issues, and hardware design issues. Millimeter waves and THz bands need to
be recreated again for joint use. If the issues can be overcome, then it will be
몭lexible.

2.3.5. Next generation terahertz communications: a rendezvous of sensing,

imaging, and localization It's a comprehensive and forward‑thinking vision of
THz communications that incorporates machine learning and THz antennas.
As may be seen, the majority of current THz transceiver designs are electrical
and photonic. While photonic technologies offer a data rate advantage,
electronic platforms have a higher power generation capability (100 mW–1
mW compared to typical tens of microwatts in photonics) [42]. THz‑band
communication is trying to play a vital role in future 6G technology. THz
communications, unlike mmWave communications, may take advantage of
the enormous accessible bandwidths in the THz band to attain a terabit per
second data rate without the use of any extra spectral ef몭iciency
augmentation techniques [43]. In this paper, we present some advanced
technology areas that can be developed using THz technology. For example, in
the areas of imaging, localization, and sensing. Previously, THz signals were
used in imaging and sensing applications like food security, gas detection, and
water dynamics. THz TDS is the most popular method for THz sensing [45].
THz can be easily used in imaging applications. The high gain of directional
antennas allows for substantially higher spatial resolution (sub‑millimeter
antennas allows for substantially higher spatial resolution (sub‑millimeter
spatial differentiation [44]) and targeted directed sensing and imaging. As
future generations are expected to make wireless services location based, 6G
networks that are based on mm‑wave and THz frequencies are needed. For
high‑speed drone communications, THz bands like antennas are needed to
detect location. Machine learning is needed for these services as the next
generation is almost entirely based on arti몭icial intelligence. Machine learning
can handle large data sets. For instance, the THz‑based localization system
can locate any place. We need to use an antenna for collecting data, and
machine learning will help with map interpolation. So, we can say that THz
applications will be many in future communication systems.

2.4. Machine learning and deep learning for satellite antenna design
application

A satellite is a spacecraft that travels around the planet in orbit. These

satellites deliver signals to a central station, which develops programming for
smaller stations that broadcast the signals locally through cables or the
airwaves. For various applications, the use of machine learning and deep
learning for satellites is explained below.

2.4.1. Acceleration of design and cross polar optimization for shaped beam
re몭lect array antennas for space applications A machine learning approach
called Support Vector Machine (SVM) was used to develop and optimize
re몭lectarray antennas in [46]. As a result, computing time is reduced without
a decrease in accuracy. The main signi몭icance of [46] is that by using SVMs,
crosspolar isolation and crosspolar discrimination can be improved, and the
computing time remains the same.

As indicated in [46], arti몭icial neural network (ANN), kriging, and support

vector machines (SVM) algorithms can be used to develop surrogate models
for the characterization of unit cells, so that the FW‑LP (full‑wave
electromagnetic tool based on local periodicity) can be replaced. Then, based
on a number of input parameters such as frequency, substrate, and
geometrical characteristics, the FW‑LP tool is used to create the unit cell
pattern of electromagnetic behavior. A function was obtained from ANN or
SVM and to train ANN or SVM those patterns were used. The function has a
high degree of resemblance to training patterns and can predict unit cell
patterns for fresh values. To characterize the whole matrix of re몭lection
coef몭icients, SVMs are applied. ANN was utilized to determine the phase
response of the re몭lectarray unit cell. In [46], it is stated that ANNs have
recently been employed for the substrate, cross polarization for dual
polarized unit cells, and phase shift to project the losses of the unit cell.
Kriging has also been used to project the phase replay and associated losses.
SVMs, Kriging, and ANN accelerated computation were used to analyze
re몭lectarray antennas. An application used machine learning methods to
create the design of re몭lectarray antennas. The analysis algorithm is called
hundreds or even thousands of times to obtain the unit cell geometry that
gives each re몭lectarray element the required phase shift. SVM is used in this
case to make re몭lectarray design and to preserve ef몭iciency of cresspolar and
copolar manufacturing, as well as to signi몭icantly speed up operations. Many
methodologies for training SVMs have been devised for the analysis of a large
re몭lectarray for Direct Broadcast Satellite (DBS) applications, as detailed in
[46]. The design process is detailed and carried out using a Method of
Moments based on Local Periodicity to gain access to accuracy and
computation time acceleration (MoM‑LP). ML algorithms are a hopeful
system to speed up re몭lectarray study.

2.4.2. Beam hopping in multibeam satellite system A Deep Learning (DL)

based path has been developed to simplify feasible beam hopping (BH) in
multibeam satellite systems. In the satellite coverage area, to manage time
variants and irregular traf몭ic requests, BH can maintain 몭lexibility. A learning
and optimization approach was used to create a rapid, near‑optimal, and
viable solution for BH scheduling. In [47], it is mentioned that getting optimal
performance within a reasonable time is not possible for traditional frequent
optimization paths and data‑driven techniques. Many studies suggest that the
path taken in [47] can improve the solution's performance and capability by
using the optimization component, while the learning component can speed
up the process of BH pattern selection and allocation.

The fundamental purpose of this work is to ensure that satellite resource

distribution in multibeam system situations with non‑uniform traf몭ic needs is
optimized by utilizing data‑driven pathways. In [48], it is mentioned that for
future broadband multibeam satellites, it is important to have the ability to
control the system resources over the service coverage area. The demand for
some spot beams in large multibeam satellites surpasses the limit and they
are known as hot spots. On the other hand, the demand for some spot beams
is less than the service they can provide, known as cold spots. So this causes a
is less than the service they can provide, known as cold spots. So this causes a
problem because where demand is high, capacity is low, and where capacity
is high, demand is low as well. Because the resources per beam are set and
consistently distributed throughout beams, standard payloads provide the
same capacity for each beam. But in this solution with 몭lexibility, the system
capacity can be maximized for allocation in different areas where needed. In
many ways, the satellite's payload 몭lexibility can be ensured by the
distribution of power, time, and bandwidth. In [47], it is mentioned that in the
satellite coverage area, a high level of 몭lexibility to maintain time variants and
irregular traf몭ic requests can be provided by the Beam Hopping (BH) system.
For some period of time, all the satellite resources can be focused on
providing service to a selected subset of beams with BH until the demand is
met. Depending on a space time‑dispatch pattern, in every time slot, the
group of illuminated beams swaps and it is replayed from time to time. In
[49], an iterative algorithm for BH illumination design has been proposed. On
the other hand, in [50], BH for power minimization has been proposed, so it
is clear that different works show the beam illumination pattern design in
different ways. Gaining feasible BH patterns while search space expands
more rapidly is hard for pattern design inBH. A satellite system requires a
long computation time to render a complex optimization process because it is
composed of hundreds or thousands of beams. To get a better solution, the
complexity of BH design is expected. In wireless networks, complex resource
management can be done by creating ef몭icient algorithms with DL. In [47],
researchers looked into the role of DL in BH optimization [47]. A better
methodology to combine learning and optimization methods for BH has been
presented to overcome the shortcomings of conventional algorithms and
classical learning models. The number of elements in the beam patterns for
BH gives the top forecast precision in deep learning. The combination of
optimization components and DL can be done simply by the proposed
algorithm. In [47], using computational time and optimality, an analogy has
been made between the proposed DL‑based optimization algorithm (DBO),
optimal and suboptimal paths. This solution shows better BH pattern
selection, feasibility, and enhancement in performance.

2.4.3. Tuning optical beamforming networks A deep neural network

representation of adjusting Optical Beamforming Networks (OBFNs) was
developed in this part. To get signals from particular projections, phased
array antennas (PAAs) are level and small, requiring OBFNs to be tuned so
that communication between satellites and plains can be done. This topic is
important because tuning large scale OBFNs for any delays can be done with
the help of deep neural networks.

The planes should focus their transmission beams in the direction of the
satellite so that they can receive or transmit Radio Frequency (RF) signals to
or from the satellite. Antennas with omnidirectional are not good for low
gain. In [53, 54], it is mentioned that steering dish antennas mechanically is
the usual solution. This solution has drawbacks like increased drag force,
high maintenance costs, and a large dimension. In [55, 56], it is mentioned
that because of low maintenance costs, agility, and reduced drag forces, the
potential solution is the Phased Array Antenna (PAA) system. A PAA system is
developed using a beamforming network and an array of antenna elements
(AEs), as shown in Figure 11. From a speci몭ic angle, a time‑delayed version of
the aimed signal is received by every AE. With predetermined delay values,
the received signal will be delayed while it goes through RF paths. In [53], for
simpli몭ication of the tuning process to the bottommost path, the signal
arrives 몭irst. As it is discussed [57], the delay points are tuned, so that the
aimed delays can be matched while the signal passes with the help of a
beamforming network. In [57] mentioned, the delay points are tuned, so that
the aimed delays can be matched while the signal passes with the help of a
beamforming network. According to [58], the aimed signal adds up in phase
using a delay and a combination characteristic of the beamforming network.
The intended signal and its time‑delayed form are well de몭ined in [52]. An
optical beamforming network was employed in [59, 60], and [61] with optical
ring resonators (ORRs) as con몭igurable delays. The ORRs genetic method [64]
and nonlinear programming optimization [62, 63] were employed to 몭ind the
optimum parameter.


Figure 11

Antenna elements (AEs), a phased array antenna (PAA) system, and an optical
beamforming network (OBFN) are all examples of optical beamforming networks [52].
 beamforming network (OBFN) are all examples of optical beamforming networks [52].

The OBFN structure is designed to cover a wide bandwidth while maintaining

low cost, scalability, and a low number of ORRs. For the same bandwidth and
antenna speci몭ications, Nugroho et al. [53] discovered that the asymmetrical
binary tree‑structured OBFNs and their neural network representations as
shown in Figure 12 are scalable and have the least amount of ORR.


Figure 12

The right diagram is its neural network con몭iguration and in the left there is OBFN
system (4 × 1) [52].

Figure 12 shows the diagram of the neural network con몭iguration, and on the
left there is the OBFN system (4 × 1). From training examples of a neural
network, we can obtain input vectors and their corresponding aimed output.
The signal received by each antenna element is the neural network's input in
Figure 12. The aimed output is acquired via the reference path, and this is the
signal.

To take advantage of OBFN, a deep learning algorithm is suggested so that the

OBFN system can be tuned in a big way. The unique structure of OBFNs can
be viewed using a deep neural network. A deep learning algorithm works for
a speci몭ic OBFN structure to 몭ind ideal ORRs parameters for 4 × 1, 8 × 1, or 16
× 1 OBFN. Utilization of quanti몭iable signals by deep learning methods can be
done as a preparation paradigm since it has data‑driven skills. For future
development, it is important for online tuning to use real data as a
measurable signal, as used in [52].

2.4.4. Mobile tracking and antenna pointing in satellite terrestrial network

Growing mobile services have been unable to provide facilitation for
conventional satellite terrestrial networks in recent years. Reduction of the
communication load by processing collected data through accurate mobile
terminal location is the main problem mentioned in [65]. An arti몭icial
intelligence (AI) based pointing and tracking method for mobile terminals
and stations in satellite terrestrial networks has been developed to ensure
that mobile stations and terminals experience minimal communication
interference and can access ideal antenna signals from other stations or
terminals. An AI‑based self‑learning (ASL) network framework has been
designed for data sampling and 몭iltering in this solution. The framework also
supports unsupervised satellite selection, antenna adjustment schemes, and
mobile terminal and station tracking by mobiles.

With the rise of mobile services supported by satellite terrestrial networks,

both data analysis and data transmission need greater time and resources
[66, 67]. In [68, 69], it is mentioned that the distribution of high‑quality
service on satellite terrestrial networks has become more dif몭icult.
Conventional satellite terrestrial network service quality has been declining
because of the increased number of devices and mobile stations. It is quite
dif몭icult to communicate with several mobile devices quickly over a standard
satellite terrestrial network. Mobile device movement pathways, on the other
hand, are quite complicated. For terminals and stations, mobile pointing,
tracking, and data analysis based on AI are very important.

While talking among satellites, we must follow and precisely point the mobile
target for ground stations and terminals in satellite terrestrial networks. To
get ideal signal reception requirements, stations or terminals require
changing the pointing of satellite antennas on time while relative movement
happens between satellites. Because of high turbulence in the mobile carrier
for mobile services in the satellite terrestrial network, the elevation of the
ground mobile station and terminal, as well as the antenna azimuth, will
change rapidly. By aiming antenna beams at compatible satellites, we can
increase communication. Arti몭icial intelligence is used to evaluate the quality
of mobile communication on satellite terrestrial networks, engage with
mobile targets, and explore a new road to integrative collaboration. Multi‑
mode perception information must be considered when using different types
of sensors to detect information about moving things. Then pointing and
tracking were introduced to mobile phones. In [70], it is mentioned that to
get data from various sources and then learn to schedule tasks, optimize
allocation of resources, and train needs unsupervised learning with satellite
terrestrial networks.

2.4.5. Satellite communication In the satellite communication (SatCom)

system, the use of arti몭icial intelligence (AI) mechanisms was explored in this
study. Conventional SatCom is dependent and controlled totally by human
intervention. The AI equipment can do SatCom related work and can deal
with those challenges. In the SatCom 몭ield, long‑term AI ‑related development
has been discussed.

A reduced client standard and high operational expenditure (OPEX) have

been encountered because satellite operators' teleports need human
involvement. Satellites are built to recon몭igure their communication budget
on a millisecond basis with the imminent placement of a 몭lexible payload. To
maintain the client's service level covenant, short design time, response time,
frequency, power, and beamforming are needed for radio resource
management algorithms. By adding the requirement to coordinate space
networks of hundreds or thousands of satellites, the emergence of satellite
mega constellations only creates problems. Soon, everyday satellite
operations are expected to include automation algorithms. Deep learning has
the ability to simulate any nonlinear function, and it will play a prime role. AI
may have an impact on 몭lexible payload optimization, beam congestion
prediction, interface detection and classi몭ication, and anomaly detection in
telemetry data.

2.5. Machine learning for unmanned aerial vehicle

Wireless communication networks will rely heavily on unmanned aerial

vehicles (UAVs). When compared to older ground‑based technologies, their
acceptance in various communication‑based applications is predicted to
improve coverage and spectral ef몭iciency. This new strategy, on the other
hand, will introduce fresh changes to the network's communication
mechanisms. In this case, the machine‑learning structure should be able to
solve the different issues that have already been recognized when UAVs are
solve the different issues that have already been recognized when UAVs are
used for communication. A comprehensive review is presented, including all
relevant research papers in which machine learning approaches have been
applied to UAV‑based communications to improve different design and
functional elements such as channel modeling, resource management, and
security.

2.5.1. UAV in the machine learning environment For academic and industrial
research, unmanned aerial vehicles and machine learning are key
applications. This paper's main focus is on applying machine learning and its
techniques to many 몭ields. This is a very useful application for UAVs, which
are used in the environment. The unmanned aerial vehicle (UAV) and
machine learning are two of the most important aspects of the fourth
industrial revolution. This was actually created to research the importance of
machine learning and the scope of use for the UAV. The UAV was actually used
because of its low altitude, high resolution, 몭lying capability, and probability.
In this application, they focus on the implementation of the UAV
environment. Finally, they proved that the UAV and machine learning both
have a huge scope for scienti몭ic research.

2.5.2. UAV based 5G radio access networks The authors explain why, how, and
which sorts of machine learning approaches are effective for constructing
UAV‑based radio access networks in this application. They concentrated on
supervised and reinforced learning systems in particular. They also discussed
radio access networks and compared them to radio access networks based on
unmanned aerial vehicles.

2.5.3. Construction resource localization based on UAV‑RFID platform They

discuss data collection via a UAV‑RFID integrated platform and data analysis
utilizing the k‑nearest neighbors machine learning technique in this UAV
application. This paper characterizes localization as a classi몭ication problem
by discretizing the location regions and applying a machine learning
technique to solve the problem. They discuss data collection via a UAV‑RFID
integrated platform and data analysis utilizing the k‑nearest neighbors
machine learning technique in this UAV application. This paper characterizes
localization as a classi몭ication problem by discretizing the location regions
and applying a machine learning technique to solve the problem. Preliminary
tests show that using the UAV‑RFID platform to locate construction materials
is a viable option.
Understanding the context of building sites necessitates having location data
for construction resources. To obtain location data, most sites still rely on
human observations. Technology and research, on the other hand, have
limitations when it comes to locating construction resources, as most of them
have a restricted recognition range and accuracy for outside building sites. By
combining UAV with RFID platforms, the limiting identi몭ication range of RFID
might have been overcome due to the higher agility of UAV. Preliminary tests
show that using the UAV‑RFID platform to locate construction materials is a
viable option.

Understanding the context of building sites necessitates having location data

for construction resources. The majority of sites still rely on human‑centered
observations to obtain location data. Technology and research, on the other
hand, have limitations when it comes to locating construction resources, as
most of them have a restricted recognition range and accuracy for outside
building sites. By combining UAV with RFID platforms, the limiting
identi몭ication range of RFID might have been overcome due to the higher
agility of UAV.

2.5.4. Arti몭icial intelligent for UAV enabled wireless networks The authors
provide a detailed summary of current research in the area of arti몭icial
intelligence‑enabled UAV networks in this article. They also go through some
of the current research's limitations and present some prospective concepts
that could be pursued in the near future. They also report on some of the
work done in Florida for UAV‑based networks in order to examine the
deployment of intelligence at the boundary of UAV networks. Furthermore,
they provide a thorough introduction to each arti몭icial intelligence topic
covered in this work, allowing readers from a variety of backgrounds to
comprehend it.

Smart cities and aerial base station deployment are two examples of UAV
applications that provide motivation. The researchers looked at how machine
learning techniques are utilized to improve the performance of UAV networks
in these applications. They also provide some insight into how FL techniques
are applied to UAV networks.

2.5.5. Predicting within‑몭ield variability in grain yield and protein content of

winter wheat using UAV‑based multispectral imagery and machine learning
approaches Crop yield and quality forecasting are essential for pro몭itable
agriculture. Commercialization has resulted in low‑cost multispectral
cameras being attached to UAVs, and the development of machine learning
algorithms has made the prediction process more valuable. Machine learning
is used to forecast wheat grain production and protein content by using
spectral re몭lectance and plant height. In this research, they compared the
performance of machine learning based on re몭lectance and classic linear
regression models for forecasting wheat grain yield and protein content.

2.5.6. Cattle detection and counting in UAV images based on convolutional

neural networks The authors proposed using UAV photos to detect and count
cattle in this investigation. The targets all look to be almost the same size,
which is a peculiarity of UAV photos. Furthermore, introducing the concept of
domain adaptation could help improve the performance of a slightly different
dataset. Other slow‑moving animals can also be detected and counted using
cattle detection and counting devices.

2.6. Machine learning for textile communication systems

These antennas are designed with textile materials. These antennas are
important for developing wireless electronic textiles. It helps communicate
between garments and sensors with external devices. Wearable antennas are
몭lexible, and washable, and the following papers are focused on these.
Sensors and techniques are used to make the textile system more 몭lexible.

2.6.1. Machine learning‑based hybrid random‑fuzzy modeling framework for

antenna design It's a mixed‑machine learning approach for propagating
aleatory and epistemic uncertainties in antenna design. UQ (uncertainty
quanti몭ication) for antenna design follows statistical methods, and there are
some complex scenarios to design with that. In textiles, the UQ approach is
followed [79]. As it uses statistics, they use probability distribution functions
with some random variables, and, because of that, some represent variables
affected by epistemic uncertainty. In this paradigm, x is a real‑valued
parameter represented by a PD (Possible Distribution) (x), such that:

π: R → [0,1],∃ x ∈ R: π (x) = 1.

In this case, PDF represents the frequency of an event over a given time
interval, and PD represents the value that x is presumed to be. And 0 denotes
interval, and PD represents the value that x is presumed to be. And 0 denotes
an impossibility, whereas 1 denotes a possibility. As shown in Figure 13, PD
can be speci몭ied as rectangular or triangular. PDs are commonly used to
represent so‑called total ignorance, which occurs when no information about
a parameter's variability is known [80]. The epistemic variable in fuzzy sets
is x. It is distinguished by its cuts. Cuts are indicated by red lines in Figure 13.


Figure 13

(A) A triangular PD, π(x), and (B) the corresponding possibility Π (solid) and necessity N
(dashed) measures [78].

Due to various constraints, a machine learning strategy combining BO

(Bayesian Optimization) and the PC expansion method has been developed.
Figure 14 BO is mainly used for global optimization problems. Figure 15 is
the proposed hybrid algorithm which is combined with BO and PC.


Figure 14
 Flowchart of BO algorithm [78].


Figure 15

Proposed hybrid algorithm [78].

The proposed algorithm speeds up the standard of hybrid algorithms as it

uses the BO framework. The 몭inal hybrid UQ method presents better accuracy
and higher computational ef몭iciency.

2.6.2. On the use of knitted antennas and inductively coupled RFID tags for
wearable applications A knitted folded dipole antenna with an inductively
connected RFID chip was designed and tested. Wireless smart gadgets are
now employed in clothing. Physiological sensors and low‑power computing
units are integrated into these garments, enabling continuous biomedical
monitoring and activity tracking [82, 83]. RFID (radio‑frequency
identi몭ication) technology uses low‑power radio waves to collect data and
automatically identify items. It has been demonstrated that the backscattered
power (RSSI) transmitted by a passive RFID tag may be employed as a metric
for identifying material deformations for typical metal‑based tags [84]. They
used knitted‑based manufacturing techniques to create comfortable and
battery‑free wearable stain sensors, just as they did in this work. Comfortable
to wear, very stretchy, with impedance matching between chip and antenna,
and appropriate radiating properties to continue communication at various
and appropriate radiating properties to continue communication at various
levels of physical deformation, are all requirements for this sensor. A
wearable stain sensor is paired with a folded dipole antenna in this device.
SVM and Gaussian 몭ilters are two machine learning techniques that are used
to evaluate data. The goal of this system was to track the movements of the
body, hence knitted antennae implanted in the host garment with inductively
connected RFID tags were constructed.

2.6.3. ClothFace: a batteryless RFID‑Based textile platform for handwriting

recognition It's a ClothFace technology prototype based on UHF‑RFID for
handwriting recognition embedded in cotton fabric. Textile antennas and a
10 × 10 array of RFID ICs (integrated circuits) with a unique code were
employed in this. Human‑machine interaction is always reliant on touch or
body movement, and the most prevalent on‑body interfaces, such as
trackpads and tapping buttons [86, 87, 88, 89], are typically blended around
the arm to identify hand movement. Skin electronics [90] have recently been
proposed as a 몭lexible technology for on‑body touch and gesture recognition.
This project is an expanded version of [91], which showed a basic prototype
of ClothFace technology, a battery‑free textile‑based handwriting platform.
The upgraded work will be a real‑time recognition system that will be tested
in real‑world scenarios. It can recognize any number from 0 to 9 and can also
handle complex functions thanks to machine learning methods. The test
error rates ranged from 0.23 to 1.7 for picture identi몭ication using a machine
learning technology called CNN (Convolutional Neural Network). This
technology will enable us to turn the clothing and textiles we wear on a daily
basis into sophisticated user interfaces. It can help the user increase their
character recognition accuracy.

2.6.4. Surrogate‑based in몭ill optimization applied to electromagnetic problems

An overview of SBO (Surrogate Based Optimization) methods is presented.
Mostly, this paper focused on data‑driven approximation using several SBO
methods. There are many types of surrogate models, like SVM and Gaussian
Process (GP). Basically, SBO creates a mapping between input and output
parameters. It speeds up other optimization algorithms. Textile antennas are
made of a nonconductive textile substrate. To tackle the inverse problem of
textile antennas, the SUMO (Surrogate modelling) toolkit is employed. SBO is
a toolkit that is based on the EI (Expected Improvement) criterion and is
incorporated into the SUMO toolbox. To overcome the EM problem, they
employed a machine learning approach called SVM. SVM aids in the reduction
of errors between measured and simulated data.

2.7. Machine learning and deep learning for global positioning system
antenna design application

In order to get an accurate determination of geographical locations, the

Global Positioning System (GPS) has been developed for civil and military
use. Transmission of information by using satellites in Earth orbit allows us
to measure the distance between the user and the satellite. For many
applications, the use of machine learning and deep learning for GPS has been
described below.

2.7.1. In multipath environments a machine learning approach for GPS code

phase estimation A neural network (NN) dependent delay locked loop (DLL)
is built into Global Positioning System (GPS) receivers for multipath
reduction in [93]. The NN works on samples that are evenly spaced in the
autocorrelation function. A statistical distribution model takes into account
multipath time delay and power attenuation, and the NN is trained using that
model. Three additional solutions are compared to the recommended
solution. In high multipath situations, the NN based DLL produces less code
phase root mean squared errors than the three standard models.

Multipath interference in the Global Positioning System (GPS) has been

recognized as one of the critical error drivers. Multipath signals alter
autocorrelation functions in the phase locked loops (PLLs) and delay locked
loops (DLLs) of GPS receivers. This causes carrier phase estimates and biases
in the code and leads to errors in the navigation solution. Many techniques
have been developed to mitigate the effects of multipath. They are grouped
into two categories: one is signal processing techniques, and the other is
antenna techniques. The conventional early‑to‑late (E‑L) DLL was unable to
handle the aforementioned scenarios. They do not make use of multipath
signal statistical models and are still sensitive to multipath effects. A NN‑
based DLL (NNDLL) was created in [93] to alleviate positioning issues caused
by multipath by focusing on the autocorrelation function in the receiver's
evaluation. It provides more information about multipath signals than a
pseudo range measurement. As a result, the proposed method differs from
other machine learning‑based multipath mitigations. The NN's goal was to
determine the receiver's motion and the type of multipath environment in
order to adapt the receiver's tracking strategy. In [94], NN is in charge of
order to adapt the receiver's tracking strategy. In [94], NN is in charge of
sample processing for the autocorrelation function. This technique uses
samples from the autocorrelation function to create an estimate of the code
phase error. This occurs both in the presence and absence of multipath. The
proposed method does not necessitate any hardware modi몭ications; however
it does necessitate more autocorrelation function samples for comparison, as
with other traditional methods. From the results, it is clear that the given
NNDLL performs better than the conventional solution in high multipath
situations.

2.7.2. On unmanned aerial systems detection of GPS spoo몭ing attacks Many

civil and military software packages have gained an interest in unmanned
aerial systems (UAS). A machine learning method has been proposed to
detect GPS spoo몭ing signals based on an arti몭icial neural network. The
aftermath shows a low probability of incorrect alarms and a favorable
probability of detection.

Depending on various sensors, including the Global Positioning System (GPS),

the Unmanned Aerial System (UAS) does its operation. According to [96, 97],
and [98], GPS makes UAS an ideal system for tracking and navigation goals
with a precision of up to 3 m. For location tracking and time synchronization
in real time, GPS is used by several devices. Signals are received by at least
four satellites, which act as GPS receivers. Because of the unencrypted signals
of the satellite, the public GPS receiver is not secure and cyber‑attacks,
including GPS spoo몭ing, can happen through those GPS receivers. So cyber‑
attacks can happen at a higher power, and in the attack, fake signals like
satellite signals are transmitted by hackers. The hacker can later rebroadcast
the GPS signals that were previously stored. By changing the time delays and
information in the signals, an attacker can cause the receiver to calculate a
random position. This can be done with minimal software and hardware. In
[99], a method based on state approximation analysis using Support Vector
Machine was presented to detect GPS spoo몭ing attempts on UAS. However, in
the long run, this strategy degrades in performance. Another method in [100]
dubbed Crowd‑GPS‑Sec is used to locate and identify GPS spoo몭ing assaults
on aircraft and UAVs. However, this method is not feasible because it takes
몭ifteen minutes to achieve a localization accuracy of roughly 150 m. By this
point, hackers have already done their harm. Because traditional solutions
have shortcomings, this study proposes a new machine learning method
based on arti몭icial neural networks (NN). In this technology, an algorithm
analyzes actual or fake GPS signals and makes judgments about the existence
or absence of attacks.

2.7.3. Dethroning GPS in low power accurate 5G positioning system In this

study, a Deep Learning (DL) dependent millimeter wave (mmWave)
positioning solution's energy consumption is evaluated. Then, with the
advanced and accurate outdoor positioning systems, it was later
differentiated. With millimeter wave networks, the suggested method
reduces the energy requirement for precise pointing. So, for mobile devices,
the design provides ef몭icient and accurate positioning.

This research closes the loop on a formally proposed Beamformed

Fingerprint (BFF) placement approach for mmWaves by demonstrating
energy feasibility across several DL models. This approach achieves an
uneven perfection level in the appearance of non‑Line‑of‑Sight (NLOS) (one
order of magnitude better than the prior state of the art). This research
compares the energy consumption of locating and tracking systems that
work using mmWaves. Because 몭ingerprint positioning approaches can share
the same precise DL models, a large number of results can be transferred
between them. The developed system outperforms GNSS‑based systems in
terms of accuracy and feasibility.

3. Analysis Go to: ▸

For antenna design, machine learning has shown great results, but it also has
some issues. Choosing the perfect algorithm for any experiment is a major
challenge. Because all kinds of simulation data are not suitable for all
algorithms, an unsuitable algorithm may not be able to 몭ind the perfect result.
So datasets need to be checked 몭irst before using an algorithm. Before
starting work, we need to know very well what area of the problem we are
working on. Because valueless results can be found from wrong assumptions,
and it will be a waste of time and resources. Getting a clean dataset is very
hard, and getting an accurate result is also necessary. Several simulations
have to be done, so that proper training data can be found. Preprocessing
data is a dif몭icult task because the data needs normalization and feature
selection, and for large datasets, a huge amount of time is needed. Debugging
the algorithm for solving a problem is also an important task in the 몭ield of
machine learning.
To investigate the performance gap between the suggested design and the
traditional design, we employed Monte Carlo simulations, where the average
is derived using 100 channel realizations, and a total of 1000 symbols are
used for each channel realization. Furthermore, the desired or required BER
is set to 103 in these results. By simulating the proposed learning‑assisted
adaptation, we were able to show that it easily ful몭ills the required BER while
giving a much greater data rate than traditional link‑adaptation based on SNR
threshold values.

In the learning beam codebook, it is capable of optimizing the beam patterns,

but in more complicated scenarios it is not applicable. Map‑based mmWave
channel models have some hardware issues. That's why they are cost‑
effective. Long‑range gesture recognition is also costly for the radar, and the
model accuracy is not stable, nor is it tested in different types of scenarios.

In [26], a Reinforcement Learning (RL) technique for spatial beams was

developed to maximize the mean data throughput of a multi‑antenna
wireless system that implies hybrid beam‑forming in the millimeter wave
frequency band to speed up the selection process. From the result, it is clear
that only a fraction of the iterations are required for the RL‑based approach.
On the other hand, the compared brute force solution necessitates numerous
iterations. But there is no clear solution to data security and clean data
collection.

Locations, sizes of the receiver, and nearest vehicles were used to determine
the ideal beam pair index in [32], which took full advantage of machine
learning equipment with previous beam training information. Many levels of
situational acknowledgment are a key part of this study, and an inclusive
analogy of numerous classi몭ication schemes has been addressed. As a result,
it is stated that 86% alignment probability can be achieved, but there is
always a security concern. A lot of factors can go wrong, and that can lead to
missing vehicle locations in the feature. The noisy features are GPS
inaccuracy, location updating frequency, and penetration rate.

In [35], a massive multi‑input multi‑output system using millimeter wave

(mmWave) by multiple users for beam alignment has been presented. An
alignment approach with partial beams utilizing Machine Learning
(AMPBML) has been suggested without any prior information, such as user
location details. But there is always an issue of clear data for testing and
location details. But there is always an issue of clear data for testing and
training. The implementation is also time‑consuming.

Every speci몭ication of the pre‑coders used to obtain the optimum decoder is

regarded as a mapping relation in the deep neural network (DNN) in the [39]
solution. In this study, a deep learning assisted mmWave massive MIMO
architecture was employed for practical hybrid pre‑coding. However, because
Deep Learning lacks common sense, the system is vulnerable. When mistakes
are made, the results can be severe.

An energy‑ef몭icient hybrid pre‑coding approach utilizing one‑bit PSs has been

investigated in [41] for practical frequency‑selective wideband mmWave
massive MIMO systems. Furthermore, as the CEO algorithm for machine
learning advances, a cross‑entropy optimization (CEO) based hybrid pre‑
coding technique to optimize the earnable sum rate of the reviewed system
has been presented. But for this solution, a huge amount of time and
resources are needed. There is also high error susceptibility.

At this moment, the detection of COVID‑19 is primarily done via polymerase

chain reaction (PCR) tests. However, researchers are looking for
advantageous alternative solutions. Besides the detection of COVID‑19, there
is also a focus on carrying out antibody tests that can help determine a
previously infected person, resulting in a better understanding of the virus'
spread. Besides using the THz band in imaging for detecting viruses, THz
technology can also assist patients' remote operations during a pandemic.
For example, THz‑based wearable sensors or implants on the patient's body
can collect health data at high rates and forward it to the healthcare support
staff, where actions can be taken remotely to assist patients. The
development of any new technology always raises both positive and negative
feelings. Nowadays, the major focus is on 5G technology, and there is a strong
belief among a considerable number of people that 5G technology negatively
affects human health [57].

The auto encoder model was implemented in a MATLAB tool in this article,
and scalograms generated from Wi‑Fi signals were used for training,
validation, and testing.

They describe the use of next‑generation body‑centric communication for

occupancy monitoring, which can provide a cost‑effective and privacy‑
preserving solution for lowering energy usage and carbon footprint. They
used machine learning algorithms; in this section, they didn't give any
description of the K‑nearest algorithms. In other papers, there are no
con몭licts of interest.

In the THz antenna with a machine learning approach to 6G network, we can

see there are some hardware issues like antennas, so it will be costly. Again, if
we see beam selection, there are some assumptions in multi‑user uplink, so it
is not suitable in all situations. Using THz DL CT is very time‑consuming. THz
communication will be very 몭lexible in the future, but as we need 6G
networks and other hardware, it will become very cost‑effective.

A machine learning technique called Support Vector Machines (SVMs) was

used to develop and optimize re몭lect arrayantennas [89]. As a result,
computing time is reduced without a decrease in accuracy. The main
signi몭icance of [89] is that by using SVMs, cross polar isolation and cross
polar discrimination can be improved and the computing time remains the
same. But there is a problem with a lack of quality data, inadequate
infrastructure and resources. SVMs do not work well with large datasets
because the required time is higher.

In [90], to simplify feasible beam hopping (BH) in multibeam satellite

systems, a Deep Learning (DL) based path was developed. In the satellite
coverage area, to manage time variants and irregular traf몭ic requests, BH can
maintain 몭lexibility. A learning and optimization approach was used to create
a rapid, near‑optimal, and viable solution for BH scheduling. But there are
some problems with Deep Learning. The duration of development is long,
large and clean data is needed, and it is also computationally expensive.

A deep neural network model of tuning Optical Beamforming Networks

(OBFNs) was devised in [95]. Small and 몭lat Phased Array Antennas (PAAs)
must be tuned for OBFNs in order to receive signals from space, allowing
planes and satellites to communicate. But problems can be faced with
repairing and maintaining satellites. Also, much more data is required in
neural networks than in traditional machine learning algorithms.

An arti몭icial intelligence (AI) based pointing and tracking method for mobile
terminals and stations in satellite terrestrial networks was developed in
[108] to ensure that mobile stations and terminals experience minimal
communication interference and can access ideal antenna signals from other
stations or terminals. But there are also some limitations, and those require
supervision, cost, and maintenance. There is no one‑size‑몭its‑all solution.

In [115], the use of AI in satellite communication has been discussed. There is

no practical application for this study.

The research on combining unmanned aerial vehicles and machine learning

is still in its early stages. The current study discovered that research in this
area is inconsistent, with the majority of it relating to computer/wireless
networks, smart cities, the military, agriculture, mining, and statistical
analysis of wild life. In UAV, the random forest and support vectors have been
employed in different ways. It can be done on a model for detecting and
identifying unregistered consumer UAVs, and trained machine learning
models for recognizing objects in UAV and satellite imagery can also be
constructed.

Deep RL can also be used to adjust the speed of the UAV cloudlet (s)
dynamically in order to improve user performance. When the derived
solutions operate on data with different properties than the data used to
train the model, the performance of Machine learning techniques may be
reduced or unanticipated behaviors may occur. They can also obtain resource
location information more quickly and ef몭iciently using this method.

It should be noted that machine learning tools are frequently used in the
literature to solve problems that could be solved in a more simple and
deterministic manner, giving the impression that the need for machine
learning is not well justi몭ied, which could lead to machine learning
misapplication in many cases.

In this paper, they can also use machine learning algorithms and re‑
identi몭ication using the features of each target. In the future, the pattern of
the animal's skin and the shape of the animal's body can be considered.

In textile antenna design, all of the work is focused on the 몭lexibility of textile
areas. In the UQ method, some random variables are introduced, but not all
the variables are used to show the result. In cloth face technology, it is cost‑
effective as there are some hardware issues. The SBO method has not been
experimented with on other antennas.
In the simulation result of the proposed NNDL is compared to that of
traditional code phase tracking solutions such as E‑L DLL, HRC, and narrow
correlator. While the NN outperformed traditional approaches in a multipath
environment, the NNDLL was shown to be able to match the performance of
traditional code phase tracking approaches when multipath was not present.
With genuine GPS signals, the NNDLL was able to achieve nearly half the
RMSE of an E‑L DLL while being trained with simulated GPS signals. But
there are many problems with neural networks, and those are the large
amount of data needed, the lack of clean data, and the computationally
expensive nature.

Based on an arti몭icial neural network, proposes a machine learning method

for detecting GPS spoo몭ing signals. But there are some problems that can be
faced by this solution, and those are high error‑susceptibility and a lack of
skilled resources.

In the Deep Learning dependent millimeter wave positioning solution's

energy consumption is evaluated. But some issues can be faced with these
problems, such as the long duration of development, the huge amount of data
needed and the cost. DL systems are fragile and when errors are made, the
errors can be huge (see Tables 3, 4, 5, 6, 7, 8, and 9).

Table 3
Comparison of the different machine learning techniques used in the investigated papers
for Millimeter Wave.
Reference Antenna Algorithm Comparison to Result
No. Used Used

[1] MIMO Reinforcement Brute‑force search Reduction in the

antenna Learning (RL) number of
algorithm iterations required
to locate the most
suitable analog
beamformers and
digital precoders
for transmission,
without
compromising the
upper bound data
rate reached
through brute‑
force search.
[7] Millimeter Random forest Naive‑Bayes, The results
wave AdaBoost, RBF‑ suggest that given
antenna SVM and Gradient perfect
Boosting assumptions, we
may get up to 86
percent alignment
probability.
[8] MIMO Alignment Multi‑path In terms of total
antenna Method with decomposition training time slots
Partial Beams and recovery, as and spectral
using ML well as adaptive ef몭iciency, the
(AMPBML) compressed AMPBML
sensing and outperforms
hierarchical existing methods
search. such as adaptive
Open incompressed
a separate window

Table 4
Comparison of the different machine learning techniques used in the investigated papers
for body centric.
for body centric.

Reference Antenna Algorithms Compared Result

No. Used Used to

[32] THz supervised Use of the THz band for

antennas algorithms body‑ centric networks.
[33] UWB linear Absorption of
antenna regression electromagnetic radiation
by muscle tissues under
radiating near‑몭ield
circumstances.
[34] planar k‑nearest linear The system was built with
antenna algorithm, regression off‑the‑shelf, non‑wearable
support vector components.
machine.
[35] TX antenna classical deep In comparison to other
and Rx machine learning methods, using a wireless
antenna learning signal for standby emotion
detection is a better option.
[36] conventional ML Current position of the
antenna body‑ centric
communication networks.

Open in a separate window

Table 5
Comparison of the different machine learning techniques used in the investigated papers
for THz.
 Reference Antenna Used Algorithm Compared Result
No. Used to

[37] PCA THz DL‑CT THz CT It shows much superior

image quality.
[38] NB,Nu, NR RFC SVM Reduce the computational
complexity hybrid
beamforming
[39] Photoconductive CNN SVM Developing identi몭ication of
metamaterials in mixtures
[40] Multi‑mode DNN Demo of 6G mobile network
multiple
antenna
[41] UM‑MIMO DNN Plasmonic Future vision of THz
antennas, communication
PCA

Open in a separate window

Table 6
Comparison of the different machine learning techniques used in the investigated papers
for Satellite.

Reference Antenna Algorithm Used Compared to Result

No. Used

[46] Re몭lectarrays SVM MoM‑LP Accelerate

computing time
without
compromising
accuracy.
[47] Multibeam Branch‑and‑ Typical data‑ The optimization
antenna bound (B&B) and driven component can
(MBA) simplex strategies and ensure the
algorithms (SA) traditional solution's
are examples of iterative ef몭iciency and
Reference Antenna Algorithm Used Compared to Result
No. Used
are examples of iterative ef몭iciency and
DL‑based optimization increase overall
optimization approaches performance
(DBO) algorithms while speeding up
the method of BH
pattern selection
and allocation.
[52] Phased Deep neural Non‑linear Large‑scale OBFNs
Array network programming can be tuned for
Antennas any desired delay.
(PAAs)
[65] Satellite Reinforcement Traditional Ascertain that our
Antennas Learning (RL) satellite‑ mobile stations
algorithm terrestrial and terminals
networks receive the best
antenna signal
and are subjected
to the least
amount of
communication
Open ininterference
a separate window
from

Table 7
Comparison of the different machine learning techniques used in the investigated papers
for UAV.

Reference Antenna Algorithm Compared Result

No. Used Used to

[72] planar k‑neural support Synthesize the research on

networks vector unmanned aerial vehicles
machines (UAVs) based on a machine
learning environment.
[73] conventional reinforcement Why, how and which types of
learning algorithms are used in U‑
Reference Antenna Algorithm Compared Result
No. Used Used
learning to algorithms are used in U‑
RANS
[74] re몭lectarrays k‑nearest Localization as a classi몭ication
algorithms problem by using machine
learning
[75] mimo arti몭icial Get a detailed overview of the
antenna intelligence AI's potential applications in
UAV‑based networks.
[76] planar linear Grain yield and protein
regression content are predicted.

Open in a separate window

Table 8
Comparison of the different machine learning techniques used in the investigated papers
for Textile.

Reference Antenna Used Algorithm Compared Result

No. Used to

[78] Dual‑polarized PDF Hybrid machine learning‑

textile patch based framework
antenna
[81] Folded dipole SVM Knitted folded dipole
antenna, antenna design and
application
[85] Dipole antenna CNN CNN Clothface technology that
can recognize
handwriting
[92] Textile antenna SUMO Overview of SBO
toolbox

Open in a separate window

 Table 9
Comparison of the different machine learning techniques used in the investigated papers
for GPS.

Reference Antenna Algorithm Compared to Result

No. Used Used

[93] GPS Neural Conventional early‑ In high multipath

antenna network minus‑late DLL, situations, the NN‑based
(NN) narrow correlator, DLL generates lower code
and high resolution phase root mean squared
correlator. error than the three
traditional approaches
(standard early‑minus‑late
DLL, narrow correlator,
and high resolution
correlator).
[95] GPS SMA Neural Support Vector It has a high likelihood of
antenna network Machine and Crowd‑ detection and a low
GPS‑Sec likelihood of false alert.

Open in a separate window

4. Conclusion Go to: ▸

This paper provides an overview of the uses of machine learning, deep

learning, and arti몭icial intelligence in antenna design. A comprehensive study
was conducted on various antenna designs, and we found that the newly
developed methods of antenna design by machine learning, deep learning,
and arti몭icial intelligence give better results than conventional methods. We
explored the 몭ields of millimeter wave, UAV, THz, satellite, textile, body
centric, and GPS for antenna design. We found that the use of machine
learning, deep learning, and arti몭icial intelligence can save time and, with the
minimization of errors, can provide high accuracy for the above 몭ields of
minimization of errors, can provide high accuracy for the above 몭ields of
antenna design and also speed up the antenna design process. In this study,
fewer simulations, ef몭icient antenna behavior prediction, and less
computational time have been seen in the 몭ield of antenna design by machine
learning, deep learning, and arti몭icial intelligence.

Declarations Go to: ▸

Author contribution statement

All authors listed have signi몭icantly contributed to the development and the
writing of this article.

Funding statement

This research did not receive any speci몭ic grant from funding agencies in the
public, commercial, or not‑for‑pro몭it sectors.

Data availability statement

No data was used for the research described in the article.

Declaration of interests statement

The authors declare no con몭lict of interest.

Additional information

No additional information is available for this paper.

Acknowledgements Go to: ▸

The authors of this research would like to express their gratitude to North
South University's Electrical and Computer Engineering Department.

References Go to: ▸
1. Lizarraga E.M., Maggio G.N., Dowhuszko A.A. 2019 XVIII Workshop on Information
Processing and Control (RPIC) IEEE; 2019, September. Hybrid beamforming algorithm
using reinforcement learning for millimeter wave wireless systems; pp. 253–258. [Google
Scholar]

2. Dowhuszko A., H¨am¨al¨ainen J. Performance of transmit beamforming codebooks with

separate amplitude and phase quantization. IEEE Signal Process. Lett. July
2015;22(7):813–817. [Google Scholar]

3. Chen C. An iterative hybrid transceiver design algorithm for millimeter wave MIMO
systems. IEEE Wireless Commun. Letters. June 2015;4(3):285–288. [Google Scholar]

4. Ayach O., Rajagopal S., Abu‑Surra S., Pi Z., Heath R. Spatially sparse precoding in
millimeter wave MIMO systems. IEEE Trans. Wireless Commun. Mar. 2014;13(3):1499–
1513. [Google Scholar]

5. Moghadam N., Fodor G., Bengtsson M., Love D. On the energy ef몭iciency of MIMO hybrid
beamforming for millimeter‑wave systems with nonlinear power ampli몭iers. IEEE Trans.
Wireless Commun. Nov. 2018;17(11):7208–7221. [Google Scholar]

6. Dowhuszko A., Corral‑Briones G., H¨am¨al¨ainen J., Wichman R. Performance of

quantized random beamforming in delay‑tolerant machine‑type communication. IEEE
Trans. Wireless Commun. Aug. 2016;15(8):5664–5680. [Google Scholar]

7. Wang Y., Klautau A., Ribero M., Narasimha M., Heath R.W. 2018 IEEE Globecom
Workshops (GC Wkshps) IEEE; 2018, December. Mmwave vehicular beam training with
situational awareness by machine learning; pp. 1–6. [Google Scholar]

8. Ma W., Qi C., Li G.Y. Machine learning for beam alignment in millimeter wave massive
MIMO. IEEE Wireless Commun. Lett. 2020;9(6):875–878. [Google Scholar]

9. Alkhateeb A., El Ayach O., Leus G., Heath R.W. Channel estimation and hybrid precoding
for millimeter wave cellular systems. IEEE J. Sel. Top. Signal Process. Oct. 2014;8(5):831–
846. [Google Scholar]

10. Xiao Z., He T., Xia P., Xia X.‑G. Hierarchical codebook design for beamforming training
in millimeter‑wave communication. IEEE Trans. Wireless Commun. May
2016;15(5):3380–3392. [Google Scholar]

11. Xiao Z., Dong H., Bai L., Xia P., Xia X.‑G. Enhanced channel estimation and codebook
design for millimeter‑wave communication. IEEE Trans. Veh. Technol. Oct.
2018;67(10):9393–9405. [Google Scholar]

12. Huang H., Song Y., Yang J., Gui G., Adachi F. Deep‑learning‑based millimeter‑wave
massive MIMO for hybrid precoding. IEEE Trans. Veh. Technol. 2019;68(3):3027–3032.
[Google Scholar]

13. Ghosh A., et al. Millimeter‑wave enhanced local area systems: a highdata‑rate
approach for future wireless networks. IEEE J. Sel. Area. Commun. Jun. 2014;32(6):1152–
1163. [Google Scholar]

14. Mir T., Siddiqi M.Z., Mir U., Mackenzie R., Hao M. Machine learning inspired hybrid
precodingfor wideband millimeter‑wave massive MIMO systems. IEEE Access.
2019;7:62852–62864. [Google Scholar]

15. Alkhateeb A., Leus G., Heath R.W. Limited feedback hybrid precoding for multi‑user
millimeter wave systems. IEEE Trans. Wireless Commun. Nov. 2015;14(11):6481–6494.
[Google Scholar]

16. Han S., I. C.‑I., Xu Z., Rowell C. Large‑scale antenna systems with hybrid analog and
digital beamforming for millimeter wave 5G. IEEE Commun. Mag. Jan. 2015;53(1):186–
194. [Google Scholar]

17. Gao X., Dai L., Han S., Chih‑Lin I., Heath R.W. Energy‑ef몭icient hybrid analog and digital
precoding for MmWave MIMO systems with large antenna arrays. IEEE J. Sel. Area.
Commun. Apr. 2016;34(4):998–1009. [Google Scholar]

18. Alkhateeb A. 2019. DeepMIMO: A Generic Deep Learning Dataset for Millimeter Wave
and Massive MIMO Applications. arXiv preprint arXiv:1902.06435. [Google Scholar]

19. Satyanarayana K., El‑Hajjar M., Mourad A.A., Hanzo L. Multi‑user hybrid beamforming
relying on learning‑aided link‑adaptation for mmWave systems. IEEE Access.
2019;7:23197–23209. [Google Scholar]

20. Zhang Y., Alrabeiah M., Alkhateeb A. 2020 IEEE 21st International Workshop on Signal
Processing Advances in Wireless Communications (SPAWC) IEEE; 2020, May. Learning
beam codebooks with neural networks: towards environment‑aware mmWave MIMO; pp.
1–5. [Google Scholar]

21. Love D., Heath R., Lau V., Gesbert D., Rao B., Andrews M. An overview of limited
feedback in wireless communication systems. IEEE J. Sel. Area. Commun. Oct.
2008;26(8):1341–1365. [Google Scholar]

22. Alkhateeb A., El Ayach O., Leus G., Heath R. Channel estimation and hybrid precoding
for millimeter wave cellular systems. IEEE J. Selected Topics Signal Proc. Oct.
2014;8(5):831–846. [Google Scholar]

23. Hur S., Kim T., Love D., Krogmeier J., Thomas T., Ghosh A. Millimeter wave
beamforming for wireless backhaul and access in small cell networks. IEEE Trans.
beamforming for wireless backhaul and access in small cell networks. IEEE Trans.
Commun. Oct. 2013;61(10):4391–4403. [Google Scholar]

24. Mo J., Ng B.L., Chang S., Huang P., Kulkarni M.N., Alammouri A., Zhang J.C., Lee J., Choi
W. Beam codebook design for 5g mmwave terminals. IEEE Access. 2019;7:98387–98404.
[Google Scholar]

25. Lim Y.G., Cho Y.J., Sim M.S., Kim Y., Chae C.B., Valenzuela R.A. Map‑based millimeter‑
wave channel models: an overview, data for B5G evaluation and machine learning. IEEE
Wireless Commun. 2020;27(4):54–62. [Google Scholar]

26. Lim Y.‑G., et al. Waveform multiplexing for new radio: numerology management and
3D evaluation. IEEE Wireless Commun. Mag. Oct. 2018;25(5):86–94. [Google Scholar]

27. ICT‑317669 METIS Project Deliverable D1.4 v.3, METIS Channel Models. June 2015.
[Google Scholar]

28. Seidel S.Y., Rappaport T.S. Site‑speci몭ic propagation prediction for wireless in‑building
personal communication system design. IEEE Trans. Veh. Technol. 1994;43(4 Nov):879–
891. [Google Scholar]

29. Scalabrin M., Bielsa G., Loch A., Rossi M., Widmer J. Machine learning based network
analysis using millimeter‑wave narrow‑band energy traces. IEEE Trans. Mobile Comput.
2019;19(5):1138–1155. [Google Scholar]

30. GONGFA LI1,3, (Member, IEEE), HAO WU1, GUOZHANG JIANG2, 4, SHUANG XU1, 4
AND HONGHAI LIU5, 6, (Senior Member, IEEE), ”Dynamic gesture recognition in the
Internet of Things”, , IEEE Access

31. Naosekpam Veronica, Sharma Rupam Kumar. Machine learning in 3D space gesture
recognition. JurnalKejuruteraan. 2019;31(2):243–248. [Google Scholar]

32. Saeed N., Loukil M.H., Sarieddeen H., Al‑Naffouri T.Y., Alouini M.S. 2020. Body‑Centric
Terahertz Networks: Prospects and Challenges. [Google Scholar]

33. Vidhya S.S., Devi S.R., Shanthi K.G. Human muscle mass measurement through passive
몭lexible UWB‑myogram antenna sensor to diagnose sarcopenia. Microprocess. Microsyst.
2020;79:103284. [Google Scholar]

34. Aziz Shah S., Ahmad J., Tahir A., Ahmed F., Russel G., Shah S.Y.…Abbasi Q.H. Privacy‑
preserving non‑wearable occupancy monitoring system exploiting Wi‑Fi imaging for
next‑generation body centric communication. Micromachines. 2020;11(4):379. [PMC free
article] [PubMed] [Google Scholar]

35. Khan A.N., Ihalage A.A., Ma Y., Liu B., Liu Y., Hao Y. Deep learning framework for
subject‑independent emotion detection using wireless signals. PLoS One. 2021;16(2)
[PMC free article] [PubMed] [Google Scholar]

36. Hall P.S., Hao Y. 2006 First European Conference on Antennas and Propagation. IEEE;
2006, November. Antennas and propagation for body centric communications; pp. 1–7.
[Google Scholar]

37. Hung Y.C., Yang S.H. 2019 44th International Conference on Infrared, Millimeter, and
Terahertz Waves (IRMMW‑THz) IEEE; 2019, September. Terahertz deep learning
computed tomography; pp. 1–2. [Google Scholar]

38. Ma X., Chen Z., Li Z., Chen W., Liu K. 2019 IEEE International Conference on
Communications Workshops (ICC Workshops) IEEE; 2019, May. Low complexity beam
selection scheme for terahertz systems: a machine learning approach; pp. 1–6. [Google
Scholar]

39. Liu F., Zhang W., Sun Y., Liu J., Miao J., He F., Wu X. Secure deep learning for intelligent
terahertz metamaterial identi몭ication. Sensors. 2020;20(19):5673. [PMC free article]
[PubMed] [Google Scholar]

40. Yang P., Xiao Y., Xiao M., Li S. 6G wireless communications: vision and potential
techniques. IEEE Network. 2019;33(4):70–75. [Google Scholar]

41. Sarieddeen H., Saeed N., Al‑Naffouri T.Y., Alouini M.S. Next generation terahertz
communications: a rendezvous of sensing, imaging, and localization. IEEE Commun. Mag.
2020;58(5):69–75. [Google Scholar]

42. Sengupta K., Nagatsuma T., Mittleman D.M. Terahertz integrated electronic and hybrid
electronic‑photonic systems. Nature Electronics. 2018;1(12):622. [Google Scholar]

43. Jornet J.M., Akyildiz I.F. Channel modeling and capacity analysis for electromagnetic
wireless nanonetworks in the terahertz band. IEEE Trans. Wireless Commun. Oct.
2011;10(10):3211–3221. [Google Scholar]

44. Rappaport T.S., et al. Wireless communications and applications above 100 GHz:
opportunities and challenges for 6G and beyond. IEEE Access. 2019;7:78. 729–757.
[Google Scholar]

45. Jepsen P.U., Cooke D.G., Koch M. Terahertz spectroscopy and imaging — modern
techniques and applications. Laser Photon. Rev. 2011;5(1):124–166. [Google Scholar]

46. Prado D.R., Ló

pez‑Ferná
ndez J.A., Arrebola M., Goussetis G. Support vector regression
to accelerate design and crosspolar optimization of shaped‑beam re몭lectarray antennas
for space applications. IEEE Trans. Antenn. Propag. 2018;67(3):1659–1668. [Google
Scholar]

47. Lei L., Lagunas E., Yuan Y., Kibria M.G., Chatzinotas S., Ottersten B. 2020 IEEE 91st
Vehicular Technology Conference (VTC2020‑Spring. IEEE; 2020, May. Deep learning for
beam hopping in multibeam satellite systems; pp. 1–5. [Google Scholar]

48. Freedman A., Rainish D., Gat Y. Ka and Broadband Communication Conference, Bologna,
Italy. Oct. 2015. Beam hopping: how to make it possible. [Google Scholar]

49. Lei J., Vazquez‑Castro M. Proc. IEEE ICC. 2011. Multibeam satellite frequency/time
duality study and capacity optimization. [Google Scholar]

50. Alegre‑Godoy R., Alagha N., Vazquez‑Castro M. Proc. IEEE ICC. Jun. 2012. Offered
capacity optimization mechanisms for multibeam satellite systems. [Google Scholar]

52. Nugroho H., Wibowo W.K., Annisa A.R., Rosalinda H.M. Deep learning for tuning
optical beamforming networks. TELKOMNIKA Telecommun. Comput. Electr. Control.
2018;16(4):1607–1615. [Google Scholar]

53. Zhuang L. University of Twente; November 2010. Ring Resonator‑Based Broadband

Photonic Beamformer for Phased Array Antennas. Ph.D. dissertation. [Google Scholar]

54. Wilson T., Rohlfs K., Hü

ttemeister S. Springer Berlin Heidelberg; 2013. Tools of Radio
Astronomy, Ser. Astronomy and Astrophysics Library. [Google Scholar]

55. Balanis C. Wiley; 2008. Modern Antenna Handbook. [Google Scholar]

56. Hansen R.C. John Wiley & Sons; 2009. Phased Array Antennas; p. 213. [Google
Scholar]

57. Meijerink A., Roeloffzen C., Zhuang L., Marpaung D., Heideman R., Borreman A., van
Etten W. Proceedings of the IEEE Symposium on Communications and Vehicular Technology.
Liege; Belgium: 2006. Phased array antenna steering using a ring resonator‑based optical
beam forming network; pp. 7–12. [Google Scholar]

58. Meijerink A., Roeloffzen C., Meijerink R., Zhuang L., Marpaung D., Bentum M., Burla M.,
Verpoorte J., Jorna P., Hulzinga A., van Etten W. Novel ring resonator based integrated
photonic beamformer for broadband phased array receive antennas ‑ Part I: design and
performance analysis. J. Lightwave Technol. 2010;28(1):3–18. [Google Scholar]

59. Lenz G., Eggleton B., Madsen C.K., Slusher R. Optical delay lines based on optical
몭ilters. IEEE J. Quant. Electron. 2001;37(4):525–532. [Google Scholar]

60. Zhuang L., Roeloffzen C.G., Van Etten W. Proceedings of the IEEE Symposium on
Communications and Vehicular Technology, Twente, The Netherlands. November 2005.
Continuously tunable optical delay line; p. 23. [Google Scholar]
Continuously tunable optical delay line; p. 23. [Google Scholar]

61. Zhuang L. University of Twente; 2005. Time‑delay Properties of Optical Ring

Resonators. Master’s thesis. [Google Scholar]

62. Bazaraa M.S., Sherali H.D., Shetty C.M. John Wiley & Sons; 2013. Nonlinear
Programming: Theory and Algorithms. [Google Scholar]

63. Boot J.C., et al. North. Holland Publishing Company; Amsterdam: 1964. Quadratic
Programming: Algorithms, Anomalies, Applications, Ser. Studies in Mathematical and
Managerial Economics. [Google Scholar]

64. Garcı́
aGarcı́
a A., et al. University of Twente; Enschede: February 2009. Optical Phase
Synchronization in Coherent Optical Beamformers for Phased Array Receive Antennas.
Master’s thesis. [Google Scholar]

65. Liu Q., Yang J., Zhuang C., Barnawi A., Alzahrani B.A. Arti몭icial intelligence based
mobile tracking and antenna pointing in satellite‑terrestrial network. IEEE Access.
2019;7:177497–177503. [Google Scholar]

66. Hu M., Liu W., Peng K., Ma X., Cheng W., Liu J., Li B. Joint routing and scheduling for
vehicle‑assisted multidrone surveillance. IEEE Internet Things J. Apr. 2019;6(2):1781–
1790. [Google Scholar]

67. Hu M., Liu W., Lu J., Fu R., Peng K., Ma X., Liu J. On the joint design of routing and
scheduling for vehicle‑assisted multi‑UAV inspection. Future Generat. Comput. Syst. May
2019;94:214–223. [Google Scholar]

68. Mbyamm K.M.J., Wang L., Varus M.L. DSTP‑end to end based approach to optimize
data transmission for satellite communications. Proc. Int. Conf. Netw. Inf. Syst. Comput.,
Apr. 2016:67–70. [Google Scholar]

69. Li X., Huang X., Mathisen S., Letizia R., Paoloni C. Design of 71‑76 GHz double‑
corrugated waveguide traveling‑wave tube for satellite downlink. IEEE Trans. Electron.
Dev. Jun. 2018;65(6):2195–2200. [Google Scholar]

70. Chen M., Hao Y., Lai C., Wu D., Li Y., Hwang K. Opportunistic task scheduling over co‑
located clouds in mobile environment. IEEE Trans. Services Comput. May
2018;11(3):549–561. [Google Scholar]

72. Khan A.I., Al‑Mulla Y. Unmanned aerial vehicle in the machine learning environment.
Procedia Comput. Sci. 2019;160:46–53. [Google Scholar]

73. Kouhdaragh V., Verde F., Gelli G., Abouei J. On the application of machine learning to
the design of UAV‑based 5G radio access networks. Electronics. 2020;9(4):689. [Google
Scholar]

74. Won D., Park M.W., Chi S. 2018 IEEE International Conference on Industrial Engineering
and Engineering Management (IEEM) IEEE; 2018, December. Construction resource
localization based on UAV‑RFID platform using machine learning algorithm; pp. 1086–
1090. [Google Scholar]

75. Lahmeri M.A., Kishk M.A., Alouini M.S. 2020. Arti몭icial Intelligence for UAV‑Enabled
Wireless Networks: A Survey. arXiv preprint arXiv:2009.11522. [Google Scholar]

76. Zhou X., Kono Y., Win A., Matsui T., Tanaka T.S. Predicting within‑몭ield variability in
grain yield and protein content of winter wheat using UAV‑based multispectral imagery
and machine learning approaches. Plant Prod. Sci. 2020:1–15. [Google Scholar]

77. Shao W., Kawakami R., Yoshihashi R., You S., Kawase H., Naemura T. Cattle detection
and counting in UAV images based on convolutional neural networks. Int. J. Rem. Sens.
2020;41(1):31–52. [Google Scholar]

78. Kan D., De Ridder S., Spina D., Couckuyt I., Grassi F., Dhaene T.…Ginste D.V. 2020 14th
European Conference on Antennas and Propagation (EuCAP) IEEE; 2020, March. Machine
learning‑based hybrid random‑fuzzy modeling framework for antenna design; pp. 1–5.
[Google Scholar]

79. Kan D., Spina D., De Ridder S., Grassi F., Rogier H., VandeGinste D. IEEE Antennas and
Wireless Propagation Letters, Early Access. 2019. A machine‑learning based epistemic
modeling framework for textile antenna design. [Google Scholar]

80. Shafer G. Princeton University Press; 1976. A Mathematical Theory of Evidence.

[Google Scholar]

81. Patron D., Mongan W., Kurzweg T.P., Fontecchio A., Dion G., Anday E.K., Dandekar K.R.
On the use of knitted antennas and inductively coupled RFID tags for wearable
applications. IEEE Transact. Biomed. Circ. Syst. 2016;10(6):1047–1057. [PubMed] [Google
Scholar]

82. OM Signal [Online]. Available: http://www.omsignal.com.

83. The Mimo Smart Baby Monitor [Online]. Available: http://www.mimobaby.com.

84. Occhiuzzi C., Paggi C., Marrocco G. Passive RFID strain‑sensor based on meander‑line
antennas. IEEE Trans. Antenn. Propag. 2011;59(12):4836–4840. [Google Scholar]

85. He H., Chen X., Mehmood A., Raivio L., Huttunen H., Raumonen P., Virkki J. ClothFace: a
batteryless RFID‑based textile platform for handwriting recognition. Sensors.
2020;20(17):4878. [PMC free article] [PubMed] [Google Scholar]
86. Harrison C., Tan D., Morris D. Proceedings of the SIGCHI Conference on Human Factors
in Computing Systems, Atlanta, GA, USA. 10–15 April 2010. Skinput: Appropriating the
body as an input surface; pp. 453–462. [Google Scholar]

87. Laput G., Xiao R., Chen X.A., Hudson S.E., Harrison C. Proceedings of the 27th Annual
ACM Symposium on User Interface Software and Technology, Honolulu, HI, USA. 5–8
October 2014. Skin buttons: cheap, small, low‑powered and clickable 몭ixed‑icon laser
projectors; pp. 389–394. [Google Scholar]

88. Lin S.Y., Su C.H., Cheng K.Y., Liang R.H., Kuo T.H., Chen B.Y. Proceedings of the 24th
Annual ACM Symposium on User Interface Software and Technology, Santa Barbara, CA,
USA. 16–19 October 2014. Pub‑point upon body: exploring eyes‑free interaction and
methods on an arm; pp. 481–488. [Google Scholar]

89. Hamdan N., Kosuru R.K., Corsten C., Borchers J. Proceedings of the 2017 ACM
International Conference on Interactive Surfaces and Spaces, Brighton, UK. 17–20 October
2017. Run & Tap: investigation of on‑body tapping for runner; pp. 280–286. [Google
Scholar]

90. Weigel M., Nittala A.S., Olwal A., Steimle J. Proceedings of the 2017 CHI Conference on
Human Factors in Computing Systems, Denver, CO, USA. 6–11 May 2017. SkinMarks:
Enabling interactions on body landmarks using conformal skin electronics; pp. 3095–
3105. [Google Scholar]

91. He H., Chen X., Raivio L., Huttunen H., Virkki J. Proceedings of the 2020 14th European
Conference on Antennas and Propagation (EuCAP), Copenhagen, Denmark. 15–20 March
2020. Passive RFID‑based textile touchpad. [Google Scholar]

92. Couckuyt I., Declercq F., Dhaene T., Rogier H., Knockaert L. Surrogate‑based in몭ill
optimization applied to electromagnetic problems. Int. J. RF Microw. Computer‑Aided Eng.
2010;20(5):492–501. [Google Scholar]

93. Orabi M., Khalife J., Abdallah A.A., Kassas Z.M., Saab S.S. 2020 IEEE/ION Position,
Location and Navigation Symposium (PLANS) IEEE; 2020, April. A machine learning
approach for GPS code phase estimation in multipath environments; pp. 1224–1229.
[Google Scholar]

94. Sokhandan N., Ziedan N., Broumandan A., Lachapelle G. Context aware adaptive
multipath compensation based on channel pattern recognition for gnss receivers,”
NAVIGATION. J. Inst. Navig. September 2017;70(5):944–962. [Google Scholar]

95. Manesh M.R., Kenney J., Hu W.C., Devabhaktuni V.K., Kaabouch N. 2019 16th IEEE
Annual Consumer Communications & Networking Conference (CCNC) IEEE; 2019, January.
Detection of GPS spoo몭ing attacks on unmanned aerial systems; pp. 1–6. [Google Scholar]

96. Global Positioning System Standard Positioning Service Performance Standard. fourth
ed. U.S. Department of Defense; Sep. 2008. [Google Scholar]

97. RiahiManesh M., Mullins M., Foerster K., Kaabouch N. IEEE Aerospace Conference.
2018. A preliminary effort toward investigating the impacts of ADS‑B message injection
attack. [Google Scholar]

98. RiahiManesh M., Kaabouch N. Analysis of vulnerabilities, attacks, countermeasures

and overall risk of the automatic dependent surveillance‑broadcast (ADS‑B) system. Int. J.
Crit. Infrastruct. Protect. 2017 [Google Scholar]

99. Panice G., Luongo Salvatore, Gigante Gabriella, Pascarella Domenico, Di Benedetto
Carlo, Vozella Angela, Pescapè
A. A SVM‑based detection approach for GPS spoo몭ing
attacks to UAV. IEEE Int. Conf. Automat. Comput. (ICAC) 2017:1–11. [Google Scholar]

100. Jansen Kai, Schä

fer Matthias, Moser Daniel, Lenders Vincent, Pö
pper Christina,
Schmitt Jens. IEEE Symposium on Security and Privacy. 2018. Crowd‑GPS‑Sec: leveraging
crowdsourcing to detect and localize GPS spoo몭ing attacks; pp. 1–14. [Google Scholar]

102. Liu Y., Wang Y., Liu H., Zhou A., Liu J., Yang N. International Conference on Green,
Pervasive, and Cloud Computing. Springer; Cham: 2020, November. Long‑range gesture
recognition using millimeter wave radar; pp. 30–44. [Google Scholar]

103. Elbir A.M., Papazafeiropoulos A., Kourtessis P., Chatzinotas S. Deep channel learning
for large intelligent surfaces aided mm‑wave massive MIMO systems. IEEE Wireless
Communications Letters. 2020;9(9):1447–1451. [Google Scholar]

104. Dong P., Zhang H., Li G.Y., Gaspar I.S., Naderi Alizadeh N. Deep CNN‑based channel
estimation for mmWave massive MIMO systems. IEEE J. Sel. Topics Signal Process. Sep.
2019;13(5):989–1000. [Google Scholar]

105. Huang H., Yang J., Huang H., Song Y., Gui G. Deep learning for super‑resolution
channel estimation and DOA estimation based massive MIMO system. IEEE Trans. Veh.
Technol. Sept 2018;67(9):8549–8560. [Google Scholar]

106. Yang Y., Zhang S., Gao F., Xu C., Ma J., Dobre O.A. 2020 International Conference on
Wireless Communications and Signal Processing (WCSP) 2020. Deep learning based
antenna selection for channel extrapolation in FDD massive MIMO. [Google Scholar]

107. Rusek F., Persson D., Lau B.K., Larsson E.G., Marzetta T.L., Edfors O., Tufvesson F.
Scaling up MIMO: opportunities and challenges with very large arrays. IEEE Signal
Process. Mag. Jan. 2013;30(1):40–60. [Google Scholar]
108. Noh S., Zoltowski M.D., Love D.J. Training sequence design for feedback assisted
hybrid beamforming in massive MIMO systems. IEEE Trans. Commun. Jan.
2016;64(1):187–200. [Google Scholar]

109. Han Y., Hsu T., Wen C., Wong K., Jin S. Ef몭icient downlink channel reconstruction for
FDD multi‑antenna systems. IEEE Trans. Wireless Commun. Jun. 2019;18(6):3161–3176.
[Google Scholar]

110. Wen C., Shih W., Jin S. Deep learning for massive MIMO CSI feedback. IEEE Wireless
Commun. Lett. Oct. 2018;7(5):748–751. [Google Scholar]

111. Alrabeiah M., Alkhateeb A. Proc.53rd Asilomar Conference on Signals, Systems, and
Computers, Paci몭ic Grove, CA, USA. Nov. 2019. Deep learning for TDD and FDD massive
MIMO: mapping channels in space and frequency; pp. 1465–1470. [Google Scholar]

112. Yang Y., Gao F., Li G.Y., Jian M. Deep learning‑based down link channel prediction for
FDD massive MIMO system. IEEE Commun. Lett. Nov. 2019;23(11):1994–1998. [Google
Scholar]

113. Choi H., Choi J. Downlink extrapolation for FDD multiple antenna systems through
neural network using extracted uplink path gains. IEEE Access. Apr. 2020;8:67100–
67111. [Google Scholar]

114. Klautau A., Batista P., Gonzalez‑Prelcic N., Wang Y., Heath R.W. 2018 Information
Theory and Applications Workshop (ITA) 2018. 5G MIMO data for machine learning:
application to beam‑selection using deep learning. Published. [Google Scholar]

115. NYUSIM,” http://wireless.engineering.nyu.edu/nyusim, accessed: 2018‑01‑20.

116. Jaeckel S., Raschkowski L., Bo ¨rner K., Thiele L. QuaDRiGa: a 3‑ d multi‑cell channel
model with time evolution for enabling virtual 몭ield trials. IEEE Trans. Antenn. Propag.
June 2014;62(6):3242–3256. [Google Scholar]

Articles from Heliyon are provided here courtesy of Elsevier

FOLLOW NCBI
Connect with NLM

National Library of Medicine

8600 Rockville Pike
Bethesda, MD 20894

Web Policies
FOIA
HHS Vulnerability Disclosure

Help
Accessibility
Careers

NLM NIH HHS USA.gov