Content uploaded by Mahmoud Shawky
Author content
All content in this area was uploaded by Mahmoud Shawky on May 13, 2023
Content may be subject to copyright.
An Efficient Deep Learning-based Spectrum
Awareness Approach for Vehicular Communication
Basit A. Zaidi†, Mahmoud A. Shawky†, Ahmad Taha†, Qammer H. Abbasi†,
Muhammad Ali Imran†, and Shuja Ansari†
†James Watt School of Engineering, University of Glasgow, Glasgow, G12 8QQ, United Kingdom
Email: {s.zaidi.2, m.shawky.1}@research.gla.ac.uk,
{Qammer.Abbasi, Muhammad.Imran, Ahmad.Taha, Shuja.Ansari}@glasgow.ac.uk
AbstractÐIntelligent transportation systems require a reliable
exchange of information between network terminals in different
vehicular communication environments. Making effective use of
the dedicated spectrum is crucial to maximizing communication
performance. This requires optimising the modulation order
according to different channel conditions. This paper proposes
a lightweight spectrum awareness methodology that uses wide-
band spectrum monitoring and deep learning-based modulation
classification techniques to optimise the modulation order. We
introduce a channel quality indicator block in which the clas-
sifier’s accuracy of detection is used as a forward indicator for
the choice of the best modulation type for transmission. By using
a 3D stochastic vehicular channel, we evaluate the classification
performance at different channel parameter settings, including,
speed, variance, and signal-to-noise ratio in urban and rural
areas. The experimental analyses demonstrate the capability of
the proposed approach to supporting a high detection probability
for acceptable false decision-making ≤20%.
Index TermsÐDeep learning, Modulation classification, Spec-
trum monitoring, Vehicular communication.
I. INT ROD UC TI ON
The adoption of intelligent transportation systems con-
tributes to decreasing the number of road fatalities and improv-
ing transportation safety [1]. By allowing wireless communi-
cation between different vehicular ad-hoc network (VANET)
terminals, traffic-related messages are shared among vehicles
[2]. These messages include critical traffic information such
as location, speed, heading, etc. VANET is a mobile commu-
nication technology used in the vehicle domain that facilitates
vehicle-to-vehicle (V2V) and vehicle-to-infrastructure (V2I)
communications [1], [2]. Due to the recent influx of wire-
less technologies, the dedicated short-range communication
(DSRC) spectrum (5.85 to 5.925 GHz) becomes fully oc-
cupied, limiting the communication throughput, especially in
traffic congestion scenarios. This matter motivated researchers
to turn to the unoccupied frequency spectrum to increase the
network capacity [3]. Optimising communication performance
in VANET requires terminals to have a good observation of the
channel spectrum, referred to as ªspectrum awareness.º In gen-
eral, spectrum awareness includes interference environment
identification and modulation classification [4]. Accordingly,
it is imperative to implement reliable spectrum monitoring
techniques in order to ensure reliable communication.
Besides, successful classification of the modulation order
with high accuracy helps in identifying the best modulation
order. In other words, the classification accuracy can be used as
a forward channel quality indicator (CQI) engine to optimise
the modulation order, as shown in Fig. 1(a). One important
factor that can affect the transmission of data is the modulation
order, which refers to the number of possible states that the
carrier signal can take on. A higher modulation order allows
for more data to be transmitted, but it can also make the signal
more susceptible to interference. By using a CQI engine to
optimize the modulation order, it is possible to improve the
quality and reliability of data transmission over the forward
channel. The CQI index value is based on the accuracy of de-
tection rather than the traditional scale ranging from 0 (poorest
channel quality) to 15 (best channel quality). High detection
accuracy refers to better channel quality, thereby using the
same or higher classified modulation order at the transmission
and vice versa, as shown in Fig. 1(b). Due to the vehicular
channel quality fluctuation between high and low in urban and
rural areas, respectively, the modulation order must be opti-
mised between the communicating terminals for an acceptable
probability of error Pe. However, the channel’s unpredictable
behaviour (i.e., line-of-sight and non-line-of-sight variations)
and the hardware imperfections (i.e., carrier frequency offset
and the additive noise) lead to unexpected signal variations
which results in increasing Pe, posing a challenging scenario.
This makes the need for a reliable modulation classification
technique for fast and slow fading vehicular channels crucial.
The current state-of-the-art of modulation optimisation for
mobile and vehicular communication depends on channel
probing [5], which consumes high communication overhead.
Pilot-based channel estimation schemes are characterized by
a high amount of communication overhead and low spectrum
utilization. In this challenging scenario, this paper contributes
the following:
1) For effective observation of unoccupied channels, we
apply the spectrum aggregation-based ultra-wideband
spectrum monitoring method.
2) By using deep learning, this work develops a lightweight
feature-based modulation classification approach that
can detect modulation order with high probability of
detection for an acceptable Pe.
3) Based on a 3-dimension (3D) stochastic vehicular chan-
nel, the proposed approach was evaluated at various ve-
978-1-6654-9122-8/23/$31.00 ©2023 IEEE
2023 IEEE Wireless Communications and Networking Conference (WCNC) | 978-1-6654-9122-8/23/$31.00 ©2023 IEEE | DOI: 10.1109/WCNC55385.2023.10118615
Authorized licensed use limited to: University of Glasgow. Downloaded on May 13,2023 at 17:27:14 UTC from IEEE Xplore. Restrictions apply.
Image Extraction Subsystem
Image
Resizing
Processing
DNN
Classification
Modulation
CQI
Engine
Pre-
processing
Rx
Tx
(a) Block diagram of the modulation classification.
Accuracy
of
Classification
Lower
Modulation Order
Same or Higher
Modulation Order
Low
High
DNN
Classification
Block
Modulation
Block
(b) CQI optimisation engine.
Fig. 1: The proposed spectrum awareness approach.
hicle speeds, signal-to-noise ratios (SNRs), and channel
variations in urban and rural areas.
The paper is organised as follows. Section II introduces
related works. Section III presents the 3D stochastic V2V
channel model and deep learning model. Sections IV and V
discuss the proposed approach and performance evaluation,
respectively. Finally, Section VI concludes the paper.
II. RE LATE D WOR KS
In this section, we review the recent works related to
spectrum monitoring and modulation classification.
A. Spectrum monitoring techniques
According to the literature presented by Gupta et al. [4], two
major schemes are presented for spectrum monitoring, receiver
statistics-based, and energy ratio-based spectrum monitoring.
The former is used to detect the presence of primary users
by counting the bit error using a low-density parity check
code and comparing it with a threshold value. However, the
hardware impairments affects the bit error counting used for
spectrum monitoring. The latter observes the spectrum at the
transmitter end based on the subcarriers. For determining
the energy ratio, two same-level sliding windows are used
consecutively. Robert et al. [6] propose a real-time monitoring
system that can gather data from a three-axis antenna on three
synchronised receiving channels. Shiba et al. [7] introduce a
multi-frequency sampling network for designing a wideband
spectrum monitor in the internet of things applications. In
this study, we use the NI-LabView example block diagram
that utilizes the aggregate spectrum built up band-by-band for
spectrum monitoring; see ref. [8] for more information.
Modulation
Classification
Techniques
Feature-based:
- Instantaneous time domain.
- Transform domain.
- Statistical.
- Constellation shape.
- Zero-crossing features.
Likelihood-based:
- Gauss-Legendre rule.
- Gauss-Hermite quadrature rule.
Fig. 2: Techniques for modulation classification [13]
B. Modulation classification techniques
The two major categories of modulation classification tech-
niques are likelihood-based (LB) [9], [10] and feature-based
(FB) [11], [12], see Fig. 2. The former classifies modulation
as multiple hypotheses testing problems that lead to optimal
solutions but are computationally complex and requires pre-
known channel parameters. As for the latter, it uses features to
represent the signal, and if features and classifiers are chosen
properly, it can achieve nearly optimal performance with
reduced complexity. To reduce the complexity of the LB tech-
nique, Shi et al. [9] propose two approximate LB algorithms
to classify linearly modulated signals using Gauss±Legendre
and Gauss±Hermite quadrature rules. Zheng et al. [10] intro-
duce a maximum average likelihood algorithm for orthogonal
frequency division multiplexing (OFDM) system to determine
the modulation order. However, complexity has a trade-off
with the total number of possible active subcarrier patterns.
For FB technique, Lee et al. [11] converted the characteristic
values of wireless signals into 2D images. Afterwards, signals
are classified using a convolutional neural network (CNN).
Nevertheless, this method is not capable of detecting all forms
of signals. On the basis of the feature selection algorithm,
Zhang et al. [12] present a mixed recognition algorithm. A
tree-like feature structure is also used to develop a multi-layer
smooth support vector machine classifier (SVM). However,
most existing classification techniques only consider classi-
fication performance at different SNRs without considering
variation and instability in channels [14], [15]. To the best of
our knowledge, this study is unique in that it evaluates the
proposed approach in urban and rural areas under different
channel conditions.
III. PRELIMINARIES
This section reviews the V2V channel modeled in [16] and
discusses the designed deep learning model.
A. Review of the 3D stochastic V2V channel [16]
This study adopts the stochastic vehicular channel modeled
in [16]. Consider a scenario in which two vehicles, Alice
and Bob, wirelessly communicate at a central frequency fc.
In this case, Bob’s received signal is the combination of a
number of Lcomponents coming from different moving and
fixed scatterers, see Fig. 3. The lth multipath component has a
different fading coefficient and phase delay denoted by aland
Authorized licensed use limited to: University of Glasgow. Downloaded on May 13,2023 at 17:27:14 UTC from IEEE Xplore. Restrictions apply.
Alice
Bob
Scattered
Component
Fig. 3: Vehicular channel model representation for spectrum
awareness.
TABLE I: Channel parameter settings
Channel Parameters Value
Number of multipath components (L)Urban: 16, Rural: 5
Speed of Tx/Rx 10,20,30 m/s
Speed of the scatterer 30 m/s
Azimuth departure/arrival angles αA(B),lU[−π, π)
Departure/arrivals’ elevation angles βA(B),l U[−π, π)
Incident reflected scatterers’ angles α1(2),lU[0, π /3)
Weibull distribution - scale coefficient (ρ) 2.985
Weibull distribution - shape coefficient (a) 0.428
ϕl, respectively. Hence, the channel observation at the side of
Bob at time tis represented by
HB(t) =
L
X
l=1
|al|exp (jϕl) exp (j2πvlt)(1)
where the doppler parameter vlcombines Alice’s, Bob’s, and
the lscatterer’s doppler shifts, denoted by vA,l,vB,l , and vS,l,
respectively, as follows.
vl=vA,l +vB,l +vS,l (2)
where
vA(B),l =uA(B)max
fc
ccos αA(B),l cos βA(B),l
vS,l =uS
fc
c(cos α1,l + cos α2,l )
(3)
where uA(B)max is the Alice’s and Bob’s vehicles maximum
speeds, αA(B),l and βA(B),l are the Alice’s and Bob’s azimuth
and elevation angles of departure and arrivals, respectively,
and αS,l and βS,l are the scatterers’ incident and reflected
angles, respectively. According to ref. [16], the randomness of
scatterer’s speed uSfollows the Weibull distribution denoted
by
puS(uS) = wua−1
Sexp (−wua
S/a)(4)
where wand aare scale and shape parameters, respectively.
In this study, we modeled the V2V channel with parameter
settings listed in Table I.
Classification
SoftMax
Fully Connected
Outsize: 6
(6)
1,2
(5)
1,2
(4)
1,2
(3)
1,2
(2)
1,2
Max Pooling 2D (1)
Pool size: 1,2 & Stride: 1,2
(6)
(5)
(4)
(3)
(2)
Relu (1)
(6)
0.1
0.1
0.0001
(5)
0.1
0.1
0.0001
(4)
0.1
0.1
0.0001
(3)
0.1
0.1
0.0001
(2)
0.1
0.1
0.0001
Batch Normalization (1)
Mean Decay: 0.1
Variance Decay: 0.1
Epsilon: 0.00001
(6)
1,8
96
1,1
(5)
1,8
64
1,1
(4)
1,8
48
1,1
(3)
1,8
32
1,1
(2)
1,8
24
1,1
Convolution 2D (1)
Filter size: 1,8
No. Filters: 16
Stride: 1,1 & Dilation Fac tor:1,1
Input
Input size: 273,328,3
Fig. 4: The proposed deep learning model architecture.
B. The proposed image-based deep learning model
This subsection details the deep neural network (DNN)
classification block highlighted in yellow in Fig. 1(a). Fig.
4 presents the flowchart of the proposed image-based deep
learning model. The input layer is an RGB image layer
having dimensions 273 ×328 pixels followed by 6 con-
secutive 2D convolution layers with filter numbers equal to
{16,24,32,48,64,96}. Each convolution layer is followed by
batch normalization, ReLU activation, and 2D max pooling
layers. The last max pooling layer is connected to a fully
connected (FC) layer with 6 classes and weight and bias learn
rate factors equal 10. Finally, the FC layer is followed by
softmax and classification output layers. The total number
of layers is 28 layers. Training the network involves tuning
some of the parameters. Correctly adjusting these parameters
according to constellation image data input helps create a good
model in less computational time. For the training process, we
use the ºadamº solver, set the initial learn rate to 0.001, the
validation frequency to 50, max epochs to 30, and mini-batch
size to 64.
IV. SPE CT RUM AWARENESS APPROACH
This section discusses the employed spectrum awareness
approach, including spectrum monitoring and modulation clas-
sification techniques. For ultra-wideband spectrum monitoring,
we use the NI-LabView example given in [8] and the USRP
X310-1st RF channel. Then, we set the start and stop fre-
quencies to be 5.88 to 5.91 GHz, respectively. In addition,
we run the signal transmission at the 2nd channel of the
USRP at fc= 5.9GHz. Fig. 5 shows the received spectrum,
Authorized licensed use limited to: University of Glasgow. Downloaded on May 13,2023 at 17:27:14 UTC from IEEE Xplore. Restrictions apply.
Fig. 5: Spectrum monitoring from 5.88 to 5.91 GHz.
2-PSK
4-PSK
8-PSK
16-PSK
8-QAM
16-QAM
USRP
X310
Tx
USRP
X310
Rx
3D V2V
Channel
Components
AWGN
Pre-
processing
Wireless
Channel
Training Dataset Setting
Processing
DNN
Classification
Image
Resizing
Feature Extraction Subsystem
Testing Dataset Setting
Fig. 6: Implementation block diagram.
highlighting the USRP X310 transmitter channel. It can be
noted that the amplitude of the spectrum sensed at 5.9GHz is
higher than that of unoccupied channels. Then, we present the
modulation classification technique in a two-phase process as
follows.
A. The training phase
In this phase, the training dataset setting, depicted in Fig.
6, is adjusted for acquiring the dataset used for training the
DNN. It is noteworthy to mention that the training dataset
has the internal additive complex gaussian noise of the USRP
X310 receiver channel. We use an OFDM communication
system at fc= 5.9GHz for the DSRC, with 256 subcarriers,
64 cyclic-prefix, and 125 subcarriers holding the transmitted
data. Then, we use 2-PSK, 4-PSK, 8-PSK, 16-PSK, 8-QAM,
and 16-QAM modulation and demodulation processes at the
side of the transmitter and receiver, respectively. According
to the modulation order, we acquired 150 training images
with dimensions 273 ×328 pixels for each constellation
type. Training samples for different constellations in the polar
coordinates (in-phase and quadrature axes) are presented in
Fig. 7. Based on the obtained data, the total training time was
[74:13] minutes using Core-i7 CPU @ 2.7 GHz laptop with
16 GB RAM.
B. The classification phase
In this phase, using the testing dataset setting depicted in
Fig. 6, we simulated the 3D stochastic V2V channel reviewed
in subsection III(A). Using the channel parameter settings
listed in Table I and the complex additive gaussian noise block,
`
(a) 2-PSK
(b) 4-PSK
(c) 8-PSK
(d) 16-PSK
(e) 8-QAM
(f) 16-QAM
Fig. 7: Training samples of different constellations.
we evaluated the classifier’s performance at different SNRs,
vehicle speeds, and channel variations in urban and rural areas.
For urban and rural areas, we set Lto 16 and 5 multipath
components, respectively. These components are convoluted
with the received OFDM symbol at the receiver side. Based on
the power of the added complex gaussian noise, we evaluated
the probability of detection Pdat different SNRs.
V. PERFORMANCE EVALUATION
In this section, we examine how classification performance
is affected under different test parameters (i.e., channel vari-
ations var., speed, SNRs) in urban and rural areas. Four
experiments were conducted where three parameters were
kept constant while one parameter varied for each case. The
following are the experiments analyses.
A. Experiment 1: Pdat different channel variations var.
In this experiment, we set the maximum vehicle speeds
uA(B)max in (3) to 30 m/s and SNR value to 25 dB simulated
in an urban area (i.e, Lequals 16 multipath components).
Then, we adjusted the value of the channel variation var. to
0.1, 0.3, and 0.6. Table II shows Pdof the six modulation
constellations at the mentioned var. values. It can be noted that
Pdis inversely proportional to the increase in the var. value.
For example, for 8-PSK, the Pdequals 90.1% at var. =0.1.
While this value at var. =0.3 and 0.6 are 87.1% and 86.5%,
respectively. The reason for this fact is that the increment in
the var. value results in higher scattered constellations, leading
to lower Pd. According to the 16-QAM constellation, the Pd
is not affected by the var. as it has unique constellation points
compared to other classes.
B. Experiment 2: Pdat different vehicle speeds uA(B)max
In this experiment, we set the channel variation var. to 0.3
and SNR value to 25 dB simulated in an urban area. Then, we
adjusted the value of uA(B)max to 10, 20, and 30 m/s. Table
III shows Pdof the six modulation constellations at different
vehicle speeds. It can be noted that Pdis inversely proportional
to the increase in the uA(B)max value. For example, for 8-
PSK, the Pdequals 89.1% at uA(B)max =10 m/s. While this
Authorized licensed use limited to: University of Glasgow. Downloaded on May 13,2023 at 17:27:14 UTC from IEEE Xplore. Restrictions apply.
TABLE II: Pdat different channel variations
Modulation SNR = 25 dB, Speed = 30 m/s, Urban
order var. = 0.1var. = 0.3var. = 0.6
2-PSK 76.9% 76.2% 72.7%
4-PSK 98.7% 98.6% 98.1%
8-PSK 90.1% 87.1% 86.5%
16-PSK 87.5% 86.5% 85.9%
8-QAM 79.3% 78.7% 77.3%
16-QAM 100% 100% 100%
TABLE III: Pdat different vehicle speeds
Modulation SNR = 25 dB,var. = 0.3, Urban
order Speed Speed Speed
= 10 m/s = 20 m/s = 30 m/s
2-PSK 77.8% 76.9% 76.2%
4-PSK 98.9% 98.9% 98.6%
8-PSK 89.1% 88.6% 87.1%
16-PSK 88.7% 88.1% 86.5%
8-QAM 81.5% 79.5% 78.7%
16-QAM 100% 100% 100%
value at uA(B)max =20 and 30 m/s are 88.6% and 87.1%,
respectively.
C. Experiment 3: Pdat different SNRs
In this experiment, we set the channel variation var. to 0.3
and uA(B)max in (3) to 30 m/s simulated in an urban area.
Then, we adjusted the value of the SNR to 15, 20, and 25
dB. Table IV shows Pdof the six modulation constellations at
different SNRs. It can be noted that Pdis directly proportional
to the increase in the SNR. For example, for 8-PSK, the Pd
equals 87.1% at SNR = 25 dB. While this value at SNR = 20
and 15 dB are 84.5% and 82.3%, respectively.
D. Experiment 4: Pdin urban and rural areas
In this experiment, we set the channel variation var. to 0.3,
uA(B)max in (3) to 30 m/s, and SNR to 25 dB. Then, we
adjusted the value of the Lin (1) to 16 and 5 multipath
components for urban and rural areas, respectively. Table
V shows Pdof the six modulation constellations in both
scenarios. It can be noted that Pdin a rural area is better
than that of an urban area. For example, for 8-PSK, the Pd
equals 98.5% in a rural area. While this value equals 87.1%
in an urban area.
Finally, we summarise a case experiment in the form of a
confusion matrix, as presented in Table VI. These results are
obtained at var. =0.3, uA(B)max =30 m/s, and SNR = 25 dB
simulated in an urban area. As shown in the matrix, there are
two major confusing cases. In case 1, the network confuses 2-
PSK with 8-QAM, which only happens at low SNRs and high
uA(B)max and var.. In case 2, the network confuses 16-PSK
with 16 QAM due to the same reasons discussed in case 1,
leading to false detection probability.
VI. CONCLUSIONS
This paper introduces an efficient deep learning-based mod-
ulation optimisation order that saves significant communica-
TABLE IV: Pdat different signal-to-noise ratios
Modulation Speed= 30 m/s,var. = 0.3, Urban
order SNR = 15 dB SNR = 20 dB SNR = 25 dB
2-PSK 68.3% 72.4% 76.2%
4-PSK 91.7% 97.7% 98.6%
8-PSK 82.3% 84.5% 87.1%
16-PSK 84.9% 85.7% 86.5%
8-QAM 73.4% 74.7% 78.7%
16-QAM 100% 100% 100%
TABLE V: Pdin urban and rural areas
Modulation Speed= 30 m/s,var. = 0.3, SNR = 25 dB
order Urban Rural
2-PSK 76.2% 99.1%
4-PSK 98.6% 100%
8-PSK 87.1% 98.5%
16-PSK 86.5% 100%
8-QAM 78.7% 97.9%
16-QAM 100% 100%
TABLE VI: The confusion matrix at Speed= 30 m/s,var. =
0.3, and SNR = 25 dB in urban area
True
2-PSK 76.2 0 0 0 23.8 0
4-PSK 098.6 0 0 0 1.4
8-PSK 0 0 87.1 12.8 0 0.01
16-PSK 0 0 0.3 86.5 0 13.2
8-QAM 0 0 1.3 9.2 78.7 1.8
16-QAM 00000100
2-PSK
4-PSK
8-PSK
16 -PSK
8-QAM
16-QAM
Probability of detection (%)
tion overhead compared to channel probing-based approaches.
By designing a CQI engine block, we optimise the modulation
order for V2V channel. The evaluation process discussed the
effects of the channel variations, vehicle speeds, and SNR
values on the classifier’s detection probability in urban and
rural areas. Based on the experimental results, the proposed
classifier has sufficient detection probability for an acceptable
false detection ≤20%. In future work, we will explore the
possibility of testing the classifier on a realistic vehicular
wireless channel at varying terminal speeds.
REFERENCES
[1] M. A. Al-Shareeda, M. Anbar, and I. Hasbullah, ªSurvey of Authen-
tication and Privacy Schemes in Vehicular Ad Hoc Networksº, IEEE
Sensors Journal, vol. 21, no. 2, Jan. 2021.
[2] M. A. Shawky, Q. H. Abbasi, M. A. Imran, S. Ansari, and A. Taha,
ºCross-Layer Authentication based on Physical-Layer Signatures for Se-
cure Vehicular Communicationº, IEEE Intelligent Vehicles Symposium
(IV), pp. 1315-1320, 2022.
[3] G. Wu, and H. Chu, ºSpectrum Sharing with Vehicular Communication
in Cognitive Small-Cell Networksº, International Journal of Antennas
and Propagation, vol. 2020, Mar. 2020.
[4] M. S. Gupta, K. Kumar, ºProgression on spectrum sensing for cognitive
radio networks: A survey, classification, challenges and future research
issuesº, Journal of Network and Computer Applications, vol. 143, pp.
47-76, 2019.
Authorized licensed use limited to: University of Glasgow. Downloaded on May 13,2023 at 17:27:14 UTC from IEEE Xplore. Restrictions apply.
[5] C. MehlfQihrer, M. Wrulich, J.C. Ikuno, D. Bosanska, and M. Rupp,
ºSimulating the Long Term Evolution physical layerº, in Proc. of 17th
European Signal Processing Conference (EUSIPCO 2009) Glasgow,
Scotland, UK, pp. 1471-1478, 2009.
[6] H. Robert, B. Paul, M. Simona, and S. Annamaria, ºReal Time Broad-
band Electromagnetic Spectrum Monitoring System based on Software
Defined Radio Technologyº, 9th International Conference on Modern
Power Systems (MPS), pp. 1-6, 2021.
[7] T. Shiba, T. Furuichi, K. Akimoto, M. Motoyoshi, S. Kameda, and
N. Suematsu, ºReal-Time Wideband Spectrum Monitor Using Multiple
Sampling Frequency Direct RF Undersampling for Wireless IoTº, 2021
51st European Microwave Conference (EuMC), pp. 725-728, 2022.
[8] NI, ºWideband Spectrum Monitoring with NI USRPº, avail-
able at: https://forums.ni.com/t5/Example-Code/Wideband-Spectrum-
Monitoring-with-NI-USRP/ta-p/3533000.
[9] Q. Shi, and Y. Karasawa, ªAutomatic modulation identification based on
the probability density function of signal phaseº, IEEE Trans. Commun.,
vol. 60, no. 4, pp. 1033±1044, Apr. 2012.
[10] J. Zheng, and Y. Lv, ºLikelihood-Based Automatic Modulation Classi-
fication in OFDM With Index Modulationº, IEEE Trans. on Vehicular
Tech., vol. 67, no. 9, pp. 8192-8204, Sep. 2018.
[11] J. H. Lee, K. -Y. Kim and Y. Shin, ºFeature Image-Based Automatic
Modulation Classification Method Using CNN Algorithmº, 2019 In-
ternational Conference on Artificial Intelligence in Information and
Communication (ICAIIC), pp. 1-4, 2019.
[12] X. Zhang, J. Sun, and X. Zhang, ºAutomatic Modulation Classification
Based on Novel Feature Extraction Algorithmsº, IEEE Access, vol. 8,
pp. 16362-16371, 2020.
[13] S. Peng, S. Sun, and Y. -D. Yao, ºA Survey of Modulation Classification
Using Deep Learning: Signal Representation and Data Preprocessing,º
IEEE Transactions on Neural Networks and Learning Systems, pp. 1 -
19, Jun. 2021.
[14] N. Daldal, A. Sengur, K. Polat, and Z. CÈ
omert, ºA novel demodulation
system for base band digital modulation signals based on the deep long
short-term memory modelº, Appl. Acoust., vol. 166, no. 107346, Sep.
2020.
[15] Z. Wu, S. Zhou, Z. Yin, B. Ma, and Z. Yang, ºRobust Automatic Mod-
ulation Classification Under Varying Noise Conditionsº, IEEE Access,
vol. 5, pp. 19733-19741, 2017
[16] P. Karadimas, and D. Matolak, ªGeneric stochastic modeling of vehicle-
to-vehicle wireless channelsº, Vehicular Communications, vol. 1, no. 4,
pp. 153±167, 2014.
Authorized licensed use limited to: University of Glasgow. Downloaded on May 13,2023 at 17:27:14 UTC from IEEE Xplore. Restrictions apply.