RECEPTOR OFDM MEJORADO CON CNN:
LOGRAR BAJAS TASAS DE ERROR DE BITS
EN ENTORNOS DE ALTO RUIDO
CNN-ENHANCED OFDM RECEIVER:
ACHIEVING LOW BIT ERROR RATES IN HIGH-NOISE
ENVIRONMENTS
Bayron Patricio Sánchez Bustamante
Universidad Católica Andrés Bello UCAB, Venezuela
pág. 8259
DOI: https://doi.org/10.37811/cl_rcm.v9i4.19411
Receptor OFDM Mejorado con CNN:
Lograr Bajas Tasas de Error de Bits en Entornos de Alto Ruido
Bayron Patricio Sánchez Bustamante1
bpsanchez.24@est.ucab.edu.ec
https://orcid.org/0000-0001-6522-4353
Universidad Católica Andrés Bello UCAB
Venezuela
RESUMEN
Este estudio investiga la aplicación del aprendizaje profundo (DL) para mejorar los sistemas de
multiplexación por división de frecuencia ortogonal (OFDM), centrándose en la recuperación de señales
en entornos ruidosos. Se emplea una red neuronal convolucional (CNN) para reconstruir símbolos de
modulación por desplazamiento de fase en cuadratura (QPSK) a partir de datos recibidos corruptos. La
CNN aumenta su capacidad para extraer y aprender características espaciales, mitigando eficazmente
el ruido y las distorsiones inherentes a los canales de ruido gaussiano blanco aditivo (AWGN). Dentro
del marco MDFO, las señales transmitidas se dividen en subportadoras ortogonales, cada una
modificada a través de QPSK. La CNN procesa estos símbolos recibidos, identificando patrones
subyacentes para recuperar con precisión los datos originales a pesar de las deficiencias del canal. A
través de un entrenamiento extenso en grandes conjuntos de datos, el modelo demuestra capacidades
superiores de corrección de errores en comparación con los métodos tradicionales, logrando
reducciones significativas en las tasas de error de bits (BER). Los resultados experimentales confirman
que el enfoque basado en DL no solo mejora la calidad de la señal recibida, sino que también permite
un mayor rendimiento de datos. Estos hallazgos subrayan el potencial de las CNN para revolucionar las
estrategias de corrección de errores en los sistemas OFDM, ofreciendo una alternativa sólida a las
técnicas convencionales para las redes de comunicación de próxima generación.
Palabras clave: aprendizaje profundo, multiplexación ortogonal por división de frecuencia, red
neuronal convolucional
1 Autor principal.
Correspondencia: bpsanchez4@utpl.edu.ec
DOI: https://doi.org/10.37811/cl_rcm.v9i4.19411
Receptor OFDM Mejorado con CNN:
Lograr Bajas Tasas de Error de Bits en Entornos de Alto Ruido
Bayron Patricio Sánchez Bustamante1
bpsanchez.24@est.ucab.edu.ec
https://orcid.org/0000-0001-6522-4353
Universidad Católica Andrés Bello UCAB
Venezuela
RESUMEN
Este estudio investiga la aplicación del aprendizaje profundo (DL) para mejorar los sistemas de
multiplexación por división de frecuencia ortogonal (OFDM), centrándose en la recuperación de señales
en entornos ruidosos. Se emplea una red neuronal convolucional (CNN) para reconstruir símbolos de
modulación por desplazamiento de fase en cuadratura (QPSK) a partir de datos recibidos corruptos. La
CNN aumenta su capacidad para extraer y aprender características espaciales, mitigando eficazmente
el ruido y las distorsiones inherentes a los canales de ruido gaussiano blanco aditivo (AWGN). Dentro
del marco MDFO, las señales transmitidas se dividen en subportadoras ortogonales, cada una
modificada a través de QPSK. La CNN procesa estos símbolos recibidos, identificando patrones
subyacentes para recuperar con precisión los datos originales a pesar de las deficiencias del canal. A
través de un entrenamiento extenso en grandes conjuntos de datos, el modelo demuestra capacidades
superiores de corrección de errores en comparación con los métodos tradicionales, logrando
reducciones significativas en las tasas de error de bits (BER). Los resultados experimentales confirman
que el enfoque basado en DL no solo mejora la calidad de la señal recibida, sino que también permite
un mayor rendimiento de datos. Estos hallazgos subrayan el potencial de las CNN para revolucionar las
estrategias de corrección de errores en los sistemas OFDM, ofreciendo una alternativa sólida a las
técnicas convencionales para las redes de comunicación de próxima generación.
Palabras clave: aprendizaje profundo, multiplexación ortogonal por división de frecuencia, red
neuronal convolucional
1 Autor principal.
Correspondencia: bpsanchez4@utpl.edu.ec
pág. 8260
CNN-Enhanced OFDM Receiver:
Achieving Low Bit Error Rates in High-Noise Environments
ABSTRACT
This study investigates the application of deep learning (DL) to enhance Orthogonal Frequency-
Division Multiplexing (OFDM) systems, focusing on signal recovery in noisy environments. A
Convolutional Neural Network (CNN) is employed to reconstruct Quadrature Phase Shift Keying
(QPSK) symbols from corrupt received data. The CNN leverages its ability to extract and learn spatial
features, effectively mitigating noise and distortions inherent in additive white Gaussian noise (AWGN)
channels. Within the OFDM framework, transmitted signals are partitioned into orthogonal subcarriers,
each modulated via QPSK. The CNN processes these received symbols, identifying underlying patterns
to accurately recover the original data despite channel impairments. Through extensive training on large
datasets, the model demonstrates superior error correction capabilities compared to traditional methods,
achieving significant reductions in bit error rates (BER). Experimental results confirm that the DL-
based approach not only improves received signal quality but also enables higher data throughput.
These findings underscore the potential of CNNs to revolutionize error correction strategies in OFDM
systems, offering a robust alternative to conventional techniques for next-generation communication
networks.
Keywords: deep learning, orthogonal frequency-division multiplexing, convolutional neural network
Artículo recibido 20 julio 2025
Aceptado para publicación: 20 agosto 2025
CNN-Enhanced OFDM Receiver:
Achieving Low Bit Error Rates in High-Noise Environments
ABSTRACT
This study investigates the application of deep learning (DL) to enhance Orthogonal Frequency-
Division Multiplexing (OFDM) systems, focusing on signal recovery in noisy environments. A
Convolutional Neural Network (CNN) is employed to reconstruct Quadrature Phase Shift Keying
(QPSK) symbols from corrupt received data. The CNN leverages its ability to extract and learn spatial
features, effectively mitigating noise and distortions inherent in additive white Gaussian noise (AWGN)
channels. Within the OFDM framework, transmitted signals are partitioned into orthogonal subcarriers,
each modulated via QPSK. The CNN processes these received symbols, identifying underlying patterns
to accurately recover the original data despite channel impairments. Through extensive training on large
datasets, the model demonstrates superior error correction capabilities compared to traditional methods,
achieving significant reductions in bit error rates (BER). Experimental results confirm that the DL-
based approach not only improves received signal quality but also enables higher data throughput.
These findings underscore the potential of CNNs to revolutionize error correction strategies in OFDM
systems, offering a robust alternative to conventional techniques for next-generation communication
networks.
Keywords: deep learning, orthogonal frequency-division multiplexing, convolutional neural network
Artículo recibido 20 julio 2025
Aceptado para publicación: 20 agosto 2025
pág. 8261
INTRODUCTION
Orthogonal Frequency-Division Multiplexing (OFDM) is a cornerstone of modern telecommunications,
prized for its spectral efficiency and resilience to multipath interference (Wu, 2019). However, its
performance degrades significantly in high-noise environments, such as additive white Gaussian noise
(AWGN) channels, where traditional error-correction methods struggle to maintain low bit error rates
(BER) (Shi et al., 2020). This challenge is exacerbated in emerging applications like 5G, IoT, and
industrial wireless systems, where reliable communication under low signal-to-noise ratios (SNR) is
critical. To address this limitation, i propose a CNN-enhanced OFDM receiver that leverages deep
learning (DL) to reconstruct Quadrature Phase Shift Keying (QPSK) symbols from corrupted signals,
achieving superior BER performance in noisy conditions.
While OFDM is robust against frequency-selective fading, its reliance on conventional demodulation
techniques becomes a bottleneck in high-noise scenarios (Li et al., 2021). Recent advances in DL,
particularly Convolutional Neural Networks (CNNs), offer transformative potential for physical-layer
communication tasks, including channel equalization and signal recovery (Elbir & Papazafeiropoulos,
2020). Prior work by Wu (2019) and Shi et al. (2020) demonstrated CNNs’ ability to decode modulated
signals, while Li et al. (2021) and Elbir and Papazafeiropoulos (2020) validated their efficacy in MIMO-
OFDM systems. Yet, a gap remains in optimizing CNNs for single-carrier OFDM receivers to combat
AWGN-induced distortions—a challenge this study tackles.
The approach integrates a CNN into the OFDM receiver pipeline, exploiting its capacity to extract
spatial features from noisy time-domain signals. Unlike fixed modulation schemes, the CNN adapts
dynamically, learning to map corrupted QPSK symbols to their original constellations through
hierarchical feature abstraction (Fig. 1). This is achieved via a 7-layer architecture with ReLU-activated
convolutional layers and a sigmoid output, trained end-to-end using binary cross-entropy loss. The
model’s innovation lies in its joint optimization of feature extraction and symbol recovery, bypassing
the need for hand-engineered equalizers.
In the era of high-throughput, low-latency networks, the proposed system aligns with the demand for
adaptive receivers. Our primary objective is to demonstrate that a CNN-based receiver outperforms
traditional OFDM in BER reduction, especially at low SNR (3–10 dB).
INTRODUCTION
Orthogonal Frequency-Division Multiplexing (OFDM) is a cornerstone of modern telecommunications,
prized for its spectral efficiency and resilience to multipath interference (Wu, 2019). However, its
performance degrades significantly in high-noise environments, such as additive white Gaussian noise
(AWGN) channels, where traditional error-correction methods struggle to maintain low bit error rates
(BER) (Shi et al., 2020). This challenge is exacerbated in emerging applications like 5G, IoT, and
industrial wireless systems, where reliable communication under low signal-to-noise ratios (SNR) is
critical. To address this limitation, i propose a CNN-enhanced OFDM receiver that leverages deep
learning (DL) to reconstruct Quadrature Phase Shift Keying (QPSK) symbols from corrupted signals,
achieving superior BER performance in noisy conditions.
While OFDM is robust against frequency-selective fading, its reliance on conventional demodulation
techniques becomes a bottleneck in high-noise scenarios (Li et al., 2021). Recent advances in DL,
particularly Convolutional Neural Networks (CNNs), offer transformative potential for physical-layer
communication tasks, including channel equalization and signal recovery (Elbir & Papazafeiropoulos,
2020). Prior work by Wu (2019) and Shi et al. (2020) demonstrated CNNs’ ability to decode modulated
signals, while Li et al. (2021) and Elbir and Papazafeiropoulos (2020) validated their efficacy in MIMO-
OFDM systems. Yet, a gap remains in optimizing CNNs for single-carrier OFDM receivers to combat
AWGN-induced distortions—a challenge this study tackles.
The approach integrates a CNN into the OFDM receiver pipeline, exploiting its capacity to extract
spatial features from noisy time-domain signals. Unlike fixed modulation schemes, the CNN adapts
dynamically, learning to map corrupted QPSK symbols to their original constellations through
hierarchical feature abstraction (Fig. 1). This is achieved via a 7-layer architecture with ReLU-activated
convolutional layers and a sigmoid output, trained end-to-end using binary cross-entropy loss. The
model’s innovation lies in its joint optimization of feature extraction and symbol recovery, bypassing
the need for hand-engineered equalizers.
In the era of high-throughput, low-latency networks, the proposed system aligns with the demand for
adaptive receivers. Our primary objective is to demonstrate that a CNN-based receiver outperforms
traditional OFDM in BER reduction, especially at low SNR (3–10 dB).
pág. 8262
Experimental results confirm that the CNN achieves near-zero BER at 3 dB Eb/N0, surpassing
theoretical QPSK benchmarks (Fig. 6). This robustness stems from the model’s ability to generalize
across noise levels, a feat unattainable with static demodulation.
This study makes three key contributions to the field of robust communication systems: (1) a novel
lightweight CNN architecture seamlessly integrated into the OFDM receiver chain, specifically
optimized for real-time QPSK symbol recovery in noisy environments; (2) empirical validation
demonstrating a 40% reduction in BER compared to conventional OFDM systems under AWGN
conditions, with near-zero error rates achieved at 3 dB Eb/N0; and (3) a practical, scalable solution for
next-generation networks, demonstrated through extensive simulations and analysis of generalization
across varying SNR levels. These advances collectively bridge the gap between OFDM’s theoretical
robustness and the adaptive capabilities of deep learning, offering a deployable framework for high-
noise scenarios in 5G and IoT applications.
This research bridges OFDM’s theoretical strengths with DL’s adaptability, offering a blueprint for
noise-resilient communication systems. The remainder of the paper details the methodology (Section
2), results (Section 3), and implications for future research (Section 4).
METHODOLOGY
This study presents a comprehensive methodology for evaluating a CNN-enhanced OFDM receiver's
performance in high-noise environments. The approach combines quantitative experimental
design with deep learning techniques to systematically assess the proposed system's ability to reduce
bit error rates (BER) under varying noise conditions. Using computer simulations and a stratified dataset
of 10,000 OFDM signals, we compare the CNN-based receiver against conventional methods across
key metrics, including BER and symbol recovery accuracy. The methodology
emphasizes reproducibility through standardized parameters and publicly available code, while
acknowledging limitations in computational complexity and channel generalization. By integrating
signal processing theory with machine learning implementation, this framework provides a rigorous
foundation for developing noise-resilient communication systems.
Experimental results confirm that the CNN achieves near-zero BER at 3 dB Eb/N0, surpassing
theoretical QPSK benchmarks (Fig. 6). This robustness stems from the model’s ability to generalize
across noise levels, a feat unattainable with static demodulation.
This study makes three key contributions to the field of robust communication systems: (1) a novel
lightweight CNN architecture seamlessly integrated into the OFDM receiver chain, specifically
optimized for real-time QPSK symbol recovery in noisy environments; (2) empirical validation
demonstrating a 40% reduction in BER compared to conventional OFDM systems under AWGN
conditions, with near-zero error rates achieved at 3 dB Eb/N0; and (3) a practical, scalable solution for
next-generation networks, demonstrated through extensive simulations and analysis of generalization
across varying SNR levels. These advances collectively bridge the gap between OFDM’s theoretical
robustness and the adaptive capabilities of deep learning, offering a deployable framework for high-
noise scenarios in 5G and IoT applications.
This research bridges OFDM’s theoretical strengths with DL’s adaptability, offering a blueprint for
noise-resilient communication systems. The remainder of the paper details the methodology (Section
2), results (Section 3), and implications for future research (Section 4).
METHODOLOGY
This study presents a comprehensive methodology for evaluating a CNN-enhanced OFDM receiver's
performance in high-noise environments. The approach combines quantitative experimental
design with deep learning techniques to systematically assess the proposed system's ability to reduce
bit error rates (BER) under varying noise conditions. Using computer simulations and a stratified dataset
of 10,000 OFDM signals, we compare the CNN-based receiver against conventional methods across
key metrics, including BER and symbol recovery accuracy. The methodology
emphasizes reproducibility through standardized parameters and publicly available code, while
acknowledging limitations in computational complexity and channel generalization. By integrating
signal processing theory with machine learning implementation, this framework provides a rigorous
foundation for developing noise-resilient communication systems.
pág. 8263
Research Approach and Type
This study adopts a quantitative approach with an experimental design, focused on evaluating the
performance of a CNN-enhanced OFDM receiver in high-noise environments. The research
is applied in nature, aiming to implement and validate a practical solution for reducing bit error rate
(BER) in OFDM communication systems under adverse channel conditions.
Research Design
The study employs a comparative experimental design, contrasting the performance of a conventional
OFDM receiver with the proposed CNN-OFDM receiver. A cross-sectional approach is used, with data
collected and analyzed under controlled signal-to-noise ratio (SNR) conditions during a specific time
period.
Study Population and Sample
▪ Study population: QPSK-modulated OFDM signals transmitted through an additive white Gaussian
noise (AWGN) channel.
▪ Sample: A dataset of 10,000 simulated OFDM data blocks was generated, divided into training
(70%), validation (15%), and test (15%) sets.
▪ Sampling: Stratified random sampling ensured balanced SNR distribution (0-20 dB) to cover
diverse noise scenarios.
Data Collection Techniques and Instruments
▪ Computer simulation: MATLAB and Python were used to generate OFDM signals, add AWGN,
and process the data.
▪ CNN training:
• Input data: Real and imaginary components of demodulated QPSK symbols
• Preprocessing: Data normalization for training optimization
• CNN model: 7-layer architecture (convolutional, pooling, and dense layers) with ReLU and
sigmoid activation functions (Fig. 1)
• Loss and optimization: Binary cross-entropy loss function with Adam optimizer
▪ Evaluation metrics:
• BER: Comparison between transmitted and CNN-decoded bits
Research Approach and Type
This study adopts a quantitative approach with an experimental design, focused on evaluating the
performance of a CNN-enhanced OFDM receiver in high-noise environments. The research
is applied in nature, aiming to implement and validate a practical solution for reducing bit error rate
(BER) in OFDM communication systems under adverse channel conditions.
Research Design
The study employs a comparative experimental design, contrasting the performance of a conventional
OFDM receiver with the proposed CNN-OFDM receiver. A cross-sectional approach is used, with data
collected and analyzed under controlled signal-to-noise ratio (SNR) conditions during a specific time
period.
Study Population and Sample
▪ Study population: QPSK-modulated OFDM signals transmitted through an additive white Gaussian
noise (AWGN) channel.
▪ Sample: A dataset of 10,000 simulated OFDM data blocks was generated, divided into training
(70%), validation (15%), and test (15%) sets.
▪ Sampling: Stratified random sampling ensured balanced SNR distribution (0-20 dB) to cover
diverse noise scenarios.
Data Collection Techniques and Instruments
▪ Computer simulation: MATLAB and Python were used to generate OFDM signals, add AWGN,
and process the data.
▪ CNN training:
• Input data: Real and imaginary components of demodulated QPSK symbols
• Preprocessing: Data normalization for training optimization
• CNN model: 7-layer architecture (convolutional, pooling, and dense layers) with ReLU and
sigmoid activation functions (Fig. 1)
• Loss and optimization: Binary cross-entropy loss function with Adam optimizer
▪ Evaluation metrics:
• BER: Comparison between transmitted and CNN-decoded bits
pág. 8264
• Accuracy: Proportion of correctly recovered QPSK symbols
Ethical Considerations
As a simulation-based study, no human participation or sensitive data were involved. However, results
presentation transparency and method reproducibility were ensured.
Inclusion and Exclusion Criteria
Inclusion
• QPSK-modulated OFDM signals
• SNR range between 0 dB and 20 dB
Exclusion
• Signals with other modulation schemes (e.g., 16-QAM)
• Channels with non-Gaussian interference (e.g., multipath)
Limitations
▪ Generalizability: Results are limited to AWGN channels; performance in real-world channels (e.g.,
multipath) requires further validation.
▪ Computational complexity: CNN training demands significant resources, which may constrain
implementation in power-constrained devices.
Rigor and Reproducibility
To ensure validity, standard libraries (TensorFlow, NumPy) and reproducible parameters (batch size:
64, epochs: 200) were used. Synthetic datasets and code are publicly available to facilitate replication.
This methodology provides a clear framework for assessing the CNN-OFDM receiver's effectiveness,
combining experimental rigor with advanced deep learning techniques. The systematic approach
enables comprehensive evaluation of the proposed solution's potential for improving communication
reliability in noisy environments.
Deep learning has proven to be successful in various fields, achieving significant performance
improvements. Some examples include computer vision, natural language processing, speech
recognition, among others. Regarding communication systems, there have been numerous research
efforts, including the present study, which integrates deep learning into the physical layer of
communications. The employed DL model is convolutional neural networks (Van Luong et al., 2022).
• Accuracy: Proportion of correctly recovered QPSK symbols
Ethical Considerations
As a simulation-based study, no human participation or sensitive data were involved. However, results
presentation transparency and method reproducibility were ensured.
Inclusion and Exclusion Criteria
Inclusion
• QPSK-modulated OFDM signals
• SNR range between 0 dB and 20 dB
Exclusion
• Signals with other modulation schemes (e.g., 16-QAM)
• Channels with non-Gaussian interference (e.g., multipath)
Limitations
▪ Generalizability: Results are limited to AWGN channels; performance in real-world channels (e.g.,
multipath) requires further validation.
▪ Computational complexity: CNN training demands significant resources, which may constrain
implementation in power-constrained devices.
Rigor and Reproducibility
To ensure validity, standard libraries (TensorFlow, NumPy) and reproducible parameters (batch size:
64, epochs: 200) were used. Synthetic datasets and code are publicly available to facilitate replication.
This methodology provides a clear framework for assessing the CNN-OFDM receiver's effectiveness,
combining experimental rigor with advanced deep learning techniques. The systematic approach
enables comprehensive evaluation of the proposed solution's potential for improving communication
reliability in noisy environments.
Deep learning has proven to be successful in various fields, achieving significant performance
improvements. Some examples include computer vision, natural language processing, speech
recognition, among others. Regarding communication systems, there have been numerous research
efforts, including the present study, which integrates deep learning into the physical layer of
communications. The employed DL model is convolutional neural networks (Van Luong et al., 2022).
pág. 8265
Deep learning model
The structure of the CNN (Convolutional Neural Network) model, as seen in Figure 1, is a deep neural
network architecture commonly used for signal and image processing. It consists of multiple layers,
including hidden layers that perform feature extraction and output layers that generate final predictions.
Each hidden layer is composed of a set of neurons that apply linear and non-linear transformations to
the input data. The non-linear function used in the hidden layers can be the ReLU (Rectified Linear
Unit) activation function, which returns the input value if it is positive and zero otherwise. Hidden layers
often include convolutional and pooling layers for feature extraction. The weights of the neurons and
model parameters, denoted as θ, are adjustable values learned during the network's training to minimize
a loss function.
The number of layers, L, may vary depending on the model's design and the complexity of the problem
(Sandoval et al., 2017).
Figura 1 Architecture of the Convolutional Neural Network.
The operation of each layer of the network is as follows: The input layer receives the input data, which
corresponds to the demodulated QPSK signal. In the CNN model, real values of the real part of the
demodulated QPSK signal are used. The formula representing the input layer is:
𝐗 = [𝑥1𝑥2, … 𝑥𝑛] (1)
Where 𝑿 is the input vector containing the real values of the demodulated QPSK signal.
Deep learning model
The structure of the CNN (Convolutional Neural Network) model, as seen in Figure 1, is a deep neural
network architecture commonly used for signal and image processing. It consists of multiple layers,
including hidden layers that perform feature extraction and output layers that generate final predictions.
Each hidden layer is composed of a set of neurons that apply linear and non-linear transformations to
the input data. The non-linear function used in the hidden layers can be the ReLU (Rectified Linear
Unit) activation function, which returns the input value if it is positive and zero otherwise. Hidden layers
often include convolutional and pooling layers for feature extraction. The weights of the neurons and
model parameters, denoted as θ, are adjustable values learned during the network's training to minimize
a loss function.
The number of layers, L, may vary depending on the model's design and the complexity of the problem
(Sandoval et al., 2017).
Figura 1 Architecture of the Convolutional Neural Network.
The operation of each layer of the network is as follows: The input layer receives the input data, which
corresponds to the demodulated QPSK signal. In the CNN model, real values of the real part of the
demodulated QPSK signal are used. The formula representing the input layer is:
𝐗 = [𝑥1𝑥2, … 𝑥𝑛] (1)
Where 𝑿 is the input vector containing the real values of the demodulated QPSK signal.
pág. 8266
Next are the hidden layers of the CNN model, which are composed of convolutional layers, pooling
(aggregation) layers, and a fully connected layer. Each of these layers uses the ReLU activation function
to introduce non-linearity into the model. The ReLU activation function is defined as:
𝑹𝒆𝑳𝑼(𝒙) = 𝑚𝑎𝑥(0, 𝑥) (2)
The general formula for each hidden layer is:
𝐙 = 𝑅𝑒𝐿𝑈(𝑊 ∗ 𝑥_𝑝𝑟𝑒𝑣𝑖𝑜𝑢𝑠 + 𝑏) (3)
Where 𝒁 is the output vector of the hidden layer, 𝑊 is the weight matrix, 𝑥_𝑝𝑟𝑒𝑣𝑖𝑜𝑢𝑠 is the input vector
from the previous layer, and 𝑏 is the bias vector.
Finally, there is the output layer of the CNN model that uses the sigmoid activation function to produce
a probability-like output. The formula representing the output layer is:
𝐘 = 𝑠𝑖𝑔𝑚𝑜𝑖𝑑(𝑊 ∗ 𝑥_𝑝𝑟𝑒𝑣𝑖𝑜𝑢𝑠 + 𝑏) (4)
Where 𝒀 is the output vector representing the probability of the input signal belonging to each of the
classes (Perez, 2014).
Integrated system architecture. The The OFDM (Orthogonal Frequency Division Multiplexing) system
is a modulation technique used in wireless and wired communications to transmit data through multiple
orthogonal subcarriers. The flow in the transmitter of the OFDM system originates with the generation
of random data in the form of binary sequences, representing the information that is intended to be
transmitted (Daldal et al., 2022). Afterwards, the generated data is subjected to Quadrature Phase Shift
Keying (QPSK) modulation. In QPSK modulation, every two consecutive bits are mapped to a QPSK
symbol, which represents specific phase and amplitude. These QPSK symbols can be represented as a
complex signal. Next, the QPSK symbols are grouped into blocks and applied to an Orthogonal
Frequency Division Multiplexing (OFDM) system. In this system, the QPSK symbols are distributed
over multiple orthogonal subcarriers in the frequency domain. Then, the Inverse Discrete Fourier
Transform (IDFT) is used to convert the signal from the frequency domain to the time domain (Wei et
al., 2021). Following the generation of the OFDM signal, a cyclic prefix (CP) is inserted to mitigate
Inter-Symbol Interference (ISI). Subsequently, Additive White Gaussian Noise (AWGN) is added to
the signal. This noise represents the interference and imperfections of the transmission channel, and the
Next are the hidden layers of the CNN model, which are composed of convolutional layers, pooling
(aggregation) layers, and a fully connected layer. Each of these layers uses the ReLU activation function
to introduce non-linearity into the model. The ReLU activation function is defined as:
𝑹𝒆𝑳𝑼(𝒙) = 𝑚𝑎𝑥(0, 𝑥) (2)
The general formula for each hidden layer is:
𝐙 = 𝑅𝑒𝐿𝑈(𝑊 ∗ 𝑥_𝑝𝑟𝑒𝑣𝑖𝑜𝑢𝑠 + 𝑏) (3)
Where 𝒁 is the output vector of the hidden layer, 𝑊 is the weight matrix, 𝑥_𝑝𝑟𝑒𝑣𝑖𝑜𝑢𝑠 is the input vector
from the previous layer, and 𝑏 is the bias vector.
Finally, there is the output layer of the CNN model that uses the sigmoid activation function to produce
a probability-like output. The formula representing the output layer is:
𝐘 = 𝑠𝑖𝑔𝑚𝑜𝑖𝑑(𝑊 ∗ 𝑥_𝑝𝑟𝑒𝑣𝑖𝑜𝑢𝑠 + 𝑏) (4)
Where 𝒀 is the output vector representing the probability of the input signal belonging to each of the
classes (Perez, 2014).
Integrated system architecture. The The OFDM (Orthogonal Frequency Division Multiplexing) system
is a modulation technique used in wireless and wired communications to transmit data through multiple
orthogonal subcarriers. The flow in the transmitter of the OFDM system originates with the generation
of random data in the form of binary sequences, representing the information that is intended to be
transmitted (Daldal et al., 2022). Afterwards, the generated data is subjected to Quadrature Phase Shift
Keying (QPSK) modulation. In QPSK modulation, every two consecutive bits are mapped to a QPSK
symbol, which represents specific phase and amplitude. These QPSK symbols can be represented as a
complex signal. Next, the QPSK symbols are grouped into blocks and applied to an Orthogonal
Frequency Division Multiplexing (OFDM) system. In this system, the QPSK symbols are distributed
over multiple orthogonal subcarriers in the frequency domain. Then, the Inverse Discrete Fourier
Transform (IDFT) is used to convert the signal from the frequency domain to the time domain (Wei et
al., 2021). Following the generation of the OFDM signal, a cyclic prefix (CP) is inserted to mitigate
Inter-Symbol Interference (ISI). Subsequently, Additive White Gaussian Noise (AWGN) is added to
the signal. This noise represents the interference and imperfections of the transmission channel, and the
pág. 8267
power of the AWGN is determined based on the Signal-to-Noise Ratio (SNR) - the ratio of signal power
to noise power (Elsayed et al., 2020). The signal obtained is determined by the following equation:
𝑥(𝑛) = 1
√𝐾 ∑ 𝑋(𝑘)
𝐾−1
𝑘=0
𝑒𝑗2𝜋𝑘
𝐾𝑛, 𝑛=0,1,…,𝐾−1, (5)
The equation 4 represents the Discrete Fourier Transform (DFT), which is a mathematical formula used
to represent a discrete-time signal as a sum of complex exponentials. Where: 𝑥(𝑛) is the discrete-time
signal at time n, 𝑋(𝑘) is the Discrete Fourier Transform coefficient at frequency 𝑘, 𝐾 is the number of
samples in the signal, j is the imaginary unit and 𝑒𝑗2𝜋𝑘
𝐾𝑛 is a complex exponential function (Jhan &
Arteaga, 2015).
In the deep learning model section, the following process is followed. The OFDM signal with AWGN
noise is processed to obtain the data that will be fed into the CNN model. This involves performing
additional processing to adapt the data to the input dimensions of the deep learning model (Benítez
Jiménez, 2019). Subsequently, this data is fed into the model for classification. The CNN model uses
convolutional, pooling, and fully connected layers to learn and extract relevant features from the input
data. Figure 2 shows the integration of the CNN model in the OFDM system.
Figure 2 CNN model integrated in OFDM, adapted from [12].
power of the AWGN is determined based on the Signal-to-Noise Ratio (SNR) - the ratio of signal power
to noise power (Elsayed et al., 2020). The signal obtained is determined by the following equation:
𝑥(𝑛) = 1
√𝐾 ∑ 𝑋(𝑘)
𝐾−1
𝑘=0
𝑒𝑗2𝜋𝑘
𝐾𝑛, 𝑛=0,1,…,𝐾−1, (5)
The equation 4 represents the Discrete Fourier Transform (DFT), which is a mathematical formula used
to represent a discrete-time signal as a sum of complex exponentials. Where: 𝑥(𝑛) is the discrete-time
signal at time n, 𝑋(𝑘) is the Discrete Fourier Transform coefficient at frequency 𝑘, 𝐾 is the number of
samples in the signal, j is the imaginary unit and 𝑒𝑗2𝜋𝑘
𝐾𝑛 is a complex exponential function (Jhan &
Arteaga, 2015).
In the deep learning model section, the following process is followed. The OFDM signal with AWGN
noise is processed to obtain the data that will be fed into the CNN model. This involves performing
additional processing to adapt the data to the input dimensions of the deep learning model (Benítez
Jiménez, 2019). Subsequently, this data is fed into the model for classification. The CNN model uses
convolutional, pooling, and fully connected layers to learn and extract relevant features from the input
data. Figure 2 shows the integration of the CNN model in the OFDM system.
Figure 2 CNN model integrated in OFDM, adapted from [12].
pág. 8268
Finally, the model output is the prediction made by the model for the input data, in other words, the
neural network takes the received symbols as input and emits the estimated symbols as output (Ferreras
Extreme, 2021).
To achieve a higher-performance model, its complexity was increased, which involved modifications
to the architecture of the neural network. More convolutional layers, pooling layers, fully connected
layers, and dense layers were added. The training epoch was increased, and the training parameters
were also adjusted to minimize the loss function (Daldal et al., 2022).
CNN model training
The deep learning model consists of 7 layers, out of which 5 are hidden layers. The first layer is the
input layer of the CNN model, containing no neurons. The second layer is a 3x3 convolutional layer
with 32 neurons. The third layer is a 2x2 pooling layer, without neurons. The fourth layer is a 3x3
convolutional layer with 16 neurons. The fifth layer is a 2x2 pooling layer, without neurons. The sixth
layer is a dense layer with 128 neurons. Finally, the seventh layer is an output layer with a single neuron
and a sigmoid activation function. This layer produces classification predictions for the input signal
with the same number of features. The activation function used for all convolutional and dense layers
is ReLU (Rectified Linear Unit). The ReLU activation function is a non-linear function that helps
prevent the model from saturating (Ferreras Extreme, 2021).
The deep learning model was trained using simulation-generated data with the OFDM system, which
is divided into a tensor of dimensions (data_train, data_test), where data_train represents the training
data, and data_test represents the test data. During training, the model iteratively adjusts its parameters
(weights and biases) to minimize the loss function and improve its ability to make accurate predictions.
In each training iteration, new input data is generated, and the expected results are obtained. These data
are used to update the model's weights and enhance its performance. The model was trained with 200
epochs and a batch size of 64.
The epochs parameter specifies the number of times the model is trained on the entire training dataset,
and the batch_size parameter defines the number of samples used in each training step (Centeno Franco,
2019).
Finally, the model output is the prediction made by the model for the input data, in other words, the
neural network takes the received symbols as input and emits the estimated symbols as output (Ferreras
Extreme, 2021).
To achieve a higher-performance model, its complexity was increased, which involved modifications
to the architecture of the neural network. More convolutional layers, pooling layers, fully connected
layers, and dense layers were added. The training epoch was increased, and the training parameters
were also adjusted to minimize the loss function (Daldal et al., 2022).
CNN model training
The deep learning model consists of 7 layers, out of which 5 are hidden layers. The first layer is the
input layer of the CNN model, containing no neurons. The second layer is a 3x3 convolutional layer
with 32 neurons. The third layer is a 2x2 pooling layer, without neurons. The fourth layer is a 3x3
convolutional layer with 16 neurons. The fifth layer is a 2x2 pooling layer, without neurons. The sixth
layer is a dense layer with 128 neurons. Finally, the seventh layer is an output layer with a single neuron
and a sigmoid activation function. This layer produces classification predictions for the input signal
with the same number of features. The activation function used for all convolutional and dense layers
is ReLU (Rectified Linear Unit). The ReLU activation function is a non-linear function that helps
prevent the model from saturating (Ferreras Extreme, 2021).
The deep learning model was trained using simulation-generated data with the OFDM system, which
is divided into a tensor of dimensions (data_train, data_test), where data_train represents the training
data, and data_test represents the test data. During training, the model iteratively adjusts its parameters
(weights and biases) to minimize the loss function and improve its ability to make accurate predictions.
In each training iteration, new input data is generated, and the expected results are obtained. These data
are used to update the model's weights and enhance its performance. The model was trained with 200
epochs and a batch size of 64.
The epochs parameter specifies the number of times the model is trained on the entire training dataset,
and the batch_size parameter defines the number of samples used in each training step (Centeno Franco,
2019).
pág. 8269
The model's predictions are compared with the original transmitted data. This involves calculating
performance metrics, such as Bit Error Rate (BER) or accuracy and comparing them with the metrics
obtained in the OFDM simulation example. The Bit Error Rate compares the difference between the
transmitted bits and the bits predicted by the deep learning model (Shi et al., 2020).
The mathematical expression for the loss, '𝑙𝑜𝑠𝑠,' between the predictions '𝑦𝑝𝑟𝑒𝑑' of the CNN model and
the original transmitted data '𝑦_𝑡𝑟𝑢𝑒' can be computed using the binary cross-entropy function as
follows:
𝑙𝑜𝑠𝑠 = −[𝑦_𝑡𝑟𝑢𝑒 ∗ 𝑙𝑜𝑔(𝑦𝑝𝑟𝑒𝑑 ) + (1 − 𝑦𝑡𝑟𝑢𝑒) ∗ 𝑙𝑜𝑔(1 − 𝑦𝑝𝑟𝑒𝑑 ) ] (6)
The binary cross-entropy function calculates the loss for each element in the vectors 𝑦_𝑡𝑟𝑢𝑒 and 𝑦𝑝𝑟𝑒𝑑,
and then sums the losses to obtain an aggregated measure of discrepancy between the model's
predictions and the original transmitted data. This allowed for evaluating the performance and
generalization ability of the model in terms of binary classification (Wei et al., 2021).
The training process of the CNN model is computationally expensive, and the amount of time it takes
to train the model will depend on the size of the training data, the number of epochs, and the hardware
used for training (Elbir & Papazafeiropoulos, 2020).
RESULTS AND DISCUSSION
Many tests were carried out in different situations, reaching the following results.
Model Performance
To assess the model's performance, the test data is used, which is a dataset that was not used during
training. In Figure 3, the model's performance is observed as a function of accuracy and loss. As shown
in Figure 3b, the training accuracy increases as more epochs are performed, indicating that the model
is improving in classifying the training data. Similarly, the validation accuracy also increases or remains
stable as more epochs are performed, indicating that the model is generalizing correctly (Aguado López,
2020).
The model's predictions are compared with the original transmitted data. This involves calculating
performance metrics, such as Bit Error Rate (BER) or accuracy and comparing them with the metrics
obtained in the OFDM simulation example. The Bit Error Rate compares the difference between the
transmitted bits and the bits predicted by the deep learning model (Shi et al., 2020).
The mathematical expression for the loss, '𝑙𝑜𝑠𝑠,' between the predictions '𝑦𝑝𝑟𝑒𝑑' of the CNN model and
the original transmitted data '𝑦_𝑡𝑟𝑢𝑒' can be computed using the binary cross-entropy function as
follows:
𝑙𝑜𝑠𝑠 = −[𝑦_𝑡𝑟𝑢𝑒 ∗ 𝑙𝑜𝑔(𝑦𝑝𝑟𝑒𝑑 ) + (1 − 𝑦𝑡𝑟𝑢𝑒) ∗ 𝑙𝑜𝑔(1 − 𝑦𝑝𝑟𝑒𝑑 ) ] (6)
The binary cross-entropy function calculates the loss for each element in the vectors 𝑦_𝑡𝑟𝑢𝑒 and 𝑦𝑝𝑟𝑒𝑑,
and then sums the losses to obtain an aggregated measure of discrepancy between the model's
predictions and the original transmitted data. This allowed for evaluating the performance and
generalization ability of the model in terms of binary classification (Wei et al., 2021).
The training process of the CNN model is computationally expensive, and the amount of time it takes
to train the model will depend on the size of the training data, the number of epochs, and the hardware
used for training (Elbir & Papazafeiropoulos, 2020).
RESULTS AND DISCUSSION
Many tests were carried out in different situations, reaching the following results.
Model Performance
To assess the model's performance, the test data is used, which is a dataset that was not used during
training. In Figure 3, the model's performance is observed as a function of accuracy and loss. As shown
in Figure 3b, the training accuracy increases as more epochs are performed, indicating that the model
is improving in classifying the training data. Similarly, the validation accuracy also increases or remains
stable as more epochs are performed, indicating that the model is generalizing correctly (Aguado López,
2020).
pág. 8270
Figure 3 a) Loss in relation to the number of epochs and b) accuracy in relation to the number of epochs
As with precision, in Figure 3a), it can be observed that as more epochs are performed, the training loss
decreases, indicating that the model is learning and fitting better to the training data. At the same time,
the validation loss also decreases with epochs, indicating that the model is properly generalizing (Van
Luong et al., 2022).
The dashed lines in Figure 3a) and 3b) represent the New Test Loss and New Test Accuracy,
respectively, which demonstrate the model's performance on a new set of test data. Their proximity to
the training and validation curves indicates that the model generalizes well to new data. In other words,
the model can perform well with unseen data. This is a desirable property for a model, as it means the
model can be confidently used to make predictions on new data (Centeno Franco, 2019).
BER evolution during training
During the model training, as shown in Figure 4, when using a Signal-to-Noise Ratio (SNR) of 6 dB,
the deep learning model converges, and 200 epochs are sufficient to achieve optimal performance. As
the number of epochs increases, the Bit Error Rate (BER) decreases. This is because the model has
more opportunities to learn from the data and improve its accuracy. However, there will come a point
where the BER will stabilize, indicating that the model has reached its maximum performance (Li et
al., 2021).
a) b)
Figure 3 a) Loss in relation to the number of epochs and b) accuracy in relation to the number of epochs
As with precision, in Figure 3a), it can be observed that as more epochs are performed, the training loss
decreases, indicating that the model is learning and fitting better to the training data. At the same time,
the validation loss also decreases with epochs, indicating that the model is properly generalizing (Van
Luong et al., 2022).
The dashed lines in Figure 3a) and 3b) represent the New Test Loss and New Test Accuracy,
respectively, which demonstrate the model's performance on a new set of test data. Their proximity to
the training and validation curves indicates that the model generalizes well to new data. In other words,
the model can perform well with unseen data. This is a desirable property for a model, as it means the
model can be confidently used to make predictions on new data (Centeno Franco, 2019).
BER evolution during training
During the model training, as shown in Figure 4, when using a Signal-to-Noise Ratio (SNR) of 6 dB,
the deep learning model converges, and 200 epochs are sufficient to achieve optimal performance. As
the number of epochs increases, the Bit Error Rate (BER) decreases. This is because the model has
more opportunities to learn from the data and improve its accuracy. However, there will come a point
where the BER will stabilize, indicating that the model has reached its maximum performance (Li et
al., 2021).
a) b)
pág. 8271
Figure 4 BER in relation to the number of epochs
Accuracy of the model in relation to the power SNR.
The precision relationship of the deep learning model as a function of the signal-to-noise ratio (SNR)
helps determine if the model's accuracy is unaffected by the level of noise and if the model generalizes
well at different SNR power levels. Figure 5 presents a horizontal line displaying the accuracy obtained
on test data for each SNR power value. Accuracy represents the proportion of correct predictions made
by the model in relation to the total number of predictions (Elsayed et al., 2020).
Figure 5 Accuracy with respect to power SNR.
The model's accuracy increases as the signal-to-noise ratio (SNR) of the signal increases. This is
because a higher SNR indicates less noise in the signal, making it easier for the model to decode the
signal (Perez, 2014).
Figure 4 BER in relation to the number of epochs
Accuracy of the model in relation to the power SNR.
The precision relationship of the deep learning model as a function of the signal-to-noise ratio (SNR)
helps determine if the model's accuracy is unaffected by the level of noise and if the model generalizes
well at different SNR power levels. Figure 5 presents a horizontal line displaying the accuracy obtained
on test data for each SNR power value. Accuracy represents the proportion of correct predictions made
by the model in relation to the total number of predictions (Elsayed et al., 2020).
Figure 5 Accuracy with respect to power SNR.
The model's accuracy increases as the signal-to-noise ratio (SNR) of the signal increases. This is
because a higher SNR indicates less noise in the signal, making it easier for the model to decode the
signal (Perez, 2014).
pág. 8272
Performance metrics.
The model's robustness was evaluated using the Bit Error Rate (BER), which is calculated by comparing
the output bits of the network with the input bits. The lower the BER value, the better the model's
performance (Ferreras Extreme, 2021).
The curves in Figure 6 depict the evolution of the bit error rate as the signal-to-noise ratio (SNR) is
modified.
The red curve displays the performance of the trained deep learning model. As the Signal-to-Noise
Ratio (SNR) increases, a gradual decrease in the bit error rate is observed. This indicates that the CNN
OFDM receiver can enhance signal detection and estimation at the receiver, reducing bit errors as the
signal strengthens relative to the noise (Shi et al., 2020).
Figure 6 CNN OFDM receiver in relation to the Original OFDM receiver
On the other hand, the blue curve represents the theoretical bit error rate for the QPSK modulation
example. As the Signal-to-Noise Ratio (SNR) increases, a gradual decrease in the theoretical bit error
rate is observed. This curve illustrates how the bit error rate is expected to behave under ideal conditions
(Li et al., 2021).
When comparing both curves, the performance of the CNN OFDM receiver can be evaluated in relation
to the Original OFDM receiver example. If the red curve is below the blue curve, it indicates that the
CNN OFDM receiver model has succeeded in improving the bit error rate compared to the Original
Performance metrics.
The model's robustness was evaluated using the Bit Error Rate (BER), which is calculated by comparing
the output bits of the network with the input bits. The lower the BER value, the better the model's
performance (Ferreras Extreme, 2021).
The curves in Figure 6 depict the evolution of the bit error rate as the signal-to-noise ratio (SNR) is
modified.
The red curve displays the performance of the trained deep learning model. As the Signal-to-Noise
Ratio (SNR) increases, a gradual decrease in the bit error rate is observed. This indicates that the CNN
OFDM receiver can enhance signal detection and estimation at the receiver, reducing bit errors as the
signal strengthens relative to the noise (Shi et al., 2020).
Figure 6 CNN OFDM receiver in relation to the Original OFDM receiver
On the other hand, the blue curve represents the theoretical bit error rate for the QPSK modulation
example. As the Signal-to-Noise Ratio (SNR) increases, a gradual decrease in the theoretical bit error
rate is observed. This curve illustrates how the bit error rate is expected to behave under ideal conditions
(Li et al., 2021).
When comparing both curves, the performance of the CNN OFDM receiver can be evaluated in relation
to the Original OFDM receiver example. If the red curve is below the blue curve, it indicates that the
CNN OFDM receiver model has succeeded in improving the bit error rate compared to the Original
pág. 8273
OFDM receiver. This demonstrates the deep learning model's ability to leverage its learning capacity
and enhance transmission quality by reducing bit errors compared to a more basic approach like QPSK
Modulation (Wu, 2019).
When Eb/No (dB) = 3, it is observed that the red curve corresponding to the CNN OFDM receiver
becomes vertical on the graph. This is because the model has achieved an error rate close to zero (almost
negligible BER) at that point. In other words, the CNN OFDM receiver has achieved high accuracy and
the ability to recover transmitted bits correctly, even in the presence of noise. This indicates that the
CNN OFDM receiver has achieved outstanding performance, attaining an extremely low error rate
(Elbir & Papazafeiropoulos, 2020).
CONCLUSIONS
In this article, it has been observed that the Deep Learning-based algorithm demonstrates superior
performance compared to the OFDM system in terms of Bit Error Rate (BER) at different Signal-to-
Noise Ratio (SNR) values. The deep learning model's curve always remains below the QPSK example
curve, indicating higher resistance to noise and more precise signal recovery. Additionally, this model
exhibits greater robustness against channel noise and imperfections compared to the QPSK example.
Its BER curve shows slower growth as SNR decreases, suggesting higher resistance capabilities under
lower SNR conditions.
The CNN model's performance relies on its ability to learn and extract meaningful features from input
data. Unlike the QPSK example, which uses a fixed modulation scheme, the CNN OFDM receiver can
adapt and optimize its internal representations to enhance signal detection and decoding. It is essential
to note that the CNN model's performance is derived from training on a simulated dataset representing
the communication link. However, caution must be exercised when applying the model in real-world
scenarios or different channel conditions, as performance may vary. Additional evaluations and tests
are required to assess the model's generalization capabilities.
By combining the robustness and flexibility of OFDM systems with the learning and adaptive
capabilities of Convolutional Neural Networks (CNN), it is possible to significantly enhance the
performance of wireless communication systems in various scenarios and channel conditions.
OFDM receiver. This demonstrates the deep learning model's ability to leverage its learning capacity
and enhance transmission quality by reducing bit errors compared to a more basic approach like QPSK
Modulation (Wu, 2019).
When Eb/No (dB) = 3, it is observed that the red curve corresponding to the CNN OFDM receiver
becomes vertical on the graph. This is because the model has achieved an error rate close to zero (almost
negligible BER) at that point. In other words, the CNN OFDM receiver has achieved high accuracy and
the ability to recover transmitted bits correctly, even in the presence of noise. This indicates that the
CNN OFDM receiver has achieved outstanding performance, attaining an extremely low error rate
(Elbir & Papazafeiropoulos, 2020).
CONCLUSIONS
In this article, it has been observed that the Deep Learning-based algorithm demonstrates superior
performance compared to the OFDM system in terms of Bit Error Rate (BER) at different Signal-to-
Noise Ratio (SNR) values. The deep learning model's curve always remains below the QPSK example
curve, indicating higher resistance to noise and more precise signal recovery. Additionally, this model
exhibits greater robustness against channel noise and imperfections compared to the QPSK example.
Its BER curve shows slower growth as SNR decreases, suggesting higher resistance capabilities under
lower SNR conditions.
The CNN model's performance relies on its ability to learn and extract meaningful features from input
data. Unlike the QPSK example, which uses a fixed modulation scheme, the CNN OFDM receiver can
adapt and optimize its internal representations to enhance signal detection and decoding. It is essential
to note that the CNN model's performance is derived from training on a simulated dataset representing
the communication link. However, caution must be exercised when applying the model in real-world
scenarios or different channel conditions, as performance may vary. Additional evaluations and tests
are required to assess the model's generalization capabilities.
By combining the robustness and flexibility of OFDM systems with the learning and adaptive
capabilities of Convolutional Neural Networks (CNN), it is possible to significantly enhance the
performance of wireless communication systems in various scenarios and channel conditions.
pág. 8274
The use of CNN offers exciting opportunities for future studies and improvements, such as real-time
adaptation and learning, MIMO-OFDM systems, and CNN design optimization.
Finally, the integration of deep learning enables improvements in error correction capabilities, resulting
in better performance in terms of Bit Error Rate (BER). The CNN OFDM receiver can learn to extract
relevant features from received OFDM signals and make accurate predictions, leading to improved
decoding and error correction.
REFERENCES
Aguado López, I. (2020). *Degree in Mathematics Deep Learning: Convolutional Neural Networks in
R*. Universidad de Sevilla.
Benítez Jiménez, R. (2019). *Web-supervised learning based on multimodal image information*.
Universidad de Sevilla. https://bit.ly/3pCZcHe
Cabrera, M., Tarrés, F., Rey, F., & Ruiz, T. (2001). *Introduction to communications systems*.
Universidad Politécnica de Cataluña.
Centeno Franco, A. (2019). *Deep learning* [Unpublished Final Degree Project]. University of Seville.
Constantine, P. V. (2007). *Introduction to Telecommunication Systems*. Telecommunications
Systems, 43. https://bit.ly/44e0vLS
Daldal, N., Sezer, Z. A., Nour, M., Alhudhair, A., & Polat, K. (2022). A new generation communication
system based on deep learning methods for the process of modulation and demodulation from
the modulated images. *Journal of Electrical and Computer Engineering, 2022*, 9555598.
https://doi.org/10.1155/2022/9555598
Elbir, A. M., & Papazafeiropoulos, A. K. (2020). Hybrid precoding for multi-user mmWave massive
MIMO systems: A deep learning approach. *IEEE Transactions on Vehicular Technology,
69*(1), 552–563. https://doi.org/10.1109/TVT.2019.2951501
Elsayed, M., Shimotakahara, K., & Erol-Kantarci, M. (2020). Mitigation of inter-cell interference
between beams based on machine learning in mmWave. *IEEE International Conference on
Communications, 2020-June*. https://lc.cx/_33ucO
Ferreras Extreme, A. (2021). *Study of algorithms of convolutional neural networks in dataset of
medical images*. Universidad de Sevilla. https://bit.ly/3NT8U1x
The use of CNN offers exciting opportunities for future studies and improvements, such as real-time
adaptation and learning, MIMO-OFDM systems, and CNN design optimization.
Finally, the integration of deep learning enables improvements in error correction capabilities, resulting
in better performance in terms of Bit Error Rate (BER). The CNN OFDM receiver can learn to extract
relevant features from received OFDM signals and make accurate predictions, leading to improved
decoding and error correction.
REFERENCES
Aguado López, I. (2020). *Degree in Mathematics Deep Learning: Convolutional Neural Networks in
R*. Universidad de Sevilla.
Benítez Jiménez, R. (2019). *Web-supervised learning based on multimodal image information*.
Universidad de Sevilla. https://bit.ly/3pCZcHe
Cabrera, M., Tarrés, F., Rey, F., & Ruiz, T. (2001). *Introduction to communications systems*.
Universidad Politécnica de Cataluña.
Centeno Franco, A. (2019). *Deep learning* [Unpublished Final Degree Project]. University of Seville.
Constantine, P. V. (2007). *Introduction to Telecommunication Systems*. Telecommunications
Systems, 43. https://bit.ly/44e0vLS
Daldal, N., Sezer, Z. A., Nour, M., Alhudhair, A., & Polat, K. (2022). A new generation communication
system based on deep learning methods for the process of modulation and demodulation from
the modulated images. *Journal of Electrical and Computer Engineering, 2022*, 9555598.
https://doi.org/10.1155/2022/9555598
Elbir, A. M., & Papazafeiropoulos, A. K. (2020). Hybrid precoding for multi-user mmWave massive
MIMO systems: A deep learning approach. *IEEE Transactions on Vehicular Technology,
69*(1), 552–563. https://doi.org/10.1109/TVT.2019.2951501
Elsayed, M., Shimotakahara, K., & Erol-Kantarci, M. (2020). Mitigation of inter-cell interference
between beams based on machine learning in mmWave. *IEEE International Conference on
Communications, 2020-June*. https://lc.cx/_33ucO
Ferreras Extreme, A. (2021). *Study of algorithms of convolutional neural networks in dataset of
medical images*. Universidad de Sevilla. https://bit.ly/3NT8U1x
pág. 8275
Jhan, G., & Arteaga, P. R. (2015). *Application of deep learning to digital signal processing*.
Autonomous University of the West. https://bit.ly/3pHOMWO
Li, S., Zhou, J., Huang, Z., & Hu, X. (2021). CNN-based error-correcting code recognition with
blocking and embedding mechanisms. *Digital Signal Processing, 111*, 102986.
https://bit.ly/44zBL08
Li, Y., Chen, M., Yang, Y., Zhou, M. T., & Wang, C. (2017). Channel equalization based on
convolutional recurrent neural networks: An experimental study. *2017 23rd Asia-Pacific
Communications Conference (APCC)*, 1–6. https://lc.cx/b0gj1P
Perez, A. (2014). *Physical layer of the OSI model the physical layer: Communication signals*.
Whitakert Jim. https://bit.ly/3pmhw7y
Sandoval, F., Poitau, G., & Gagnon, F. (2017). Hybrid peak-to-average power ratio reduction
techniques: Performance review and comparison. *IEEE Access, 5*, 27145–27161.
https://lc.cx/vKiNvD
Shi, J., Hong, S., Cai, C., Wang, Y., Huang, H., & Gui, G. (2020). Automatic modulation recognition
method based on deep learning in the presence of phase compensation. *IEEE Access, 8*,
42831–42847. https://lc.cx/jbrPdw
Van Luong, T., Shlezinger, N., Xu, C., Hoang, T. M., Eldar, Y. C., & Hanzo, L. (2022). Rolling
interference cancellation based on deep learning for non-orthogonal downlink. *IEEE
Transactions on Vehicle Technology, 71*(11), 11876–11888. https://lc.cx/hmq6rh
Wei, X., Hu, C., & Dai, L. (2021). Deep learning for beamspace channel estimation in millimeter-wave
massive MIMO systems. *IEEE Transactions on Communications, 69*(1), 182–193.
https://lc.cx/1jVMj9
Wu, T. (2019). Deep learning methods based on CNN and RNN for the demodulation of digital signals.
*ACM International Conference Series of Proceedings, Part F147767*, 122–127.
https://lc.cx/ihYmq4
Jhan, G., & Arteaga, P. R. (2015). *Application of deep learning to digital signal processing*.
Autonomous University of the West. https://bit.ly/3pHOMWO
Li, S., Zhou, J., Huang, Z., & Hu, X. (2021). CNN-based error-correcting code recognition with
blocking and embedding mechanisms. *Digital Signal Processing, 111*, 102986.
https://bit.ly/44zBL08
Li, Y., Chen, M., Yang, Y., Zhou, M. T., & Wang, C. (2017). Channel equalization based on
convolutional recurrent neural networks: An experimental study. *2017 23rd Asia-Pacific
Communications Conference (APCC)*, 1–6. https://lc.cx/b0gj1P
Perez, A. (2014). *Physical layer of the OSI model the physical layer: Communication signals*.
Whitakert Jim. https://bit.ly/3pmhw7y
Sandoval, F., Poitau, G., & Gagnon, F. (2017). Hybrid peak-to-average power ratio reduction
techniques: Performance review and comparison. *IEEE Access, 5*, 27145–27161.
https://lc.cx/vKiNvD
Shi, J., Hong, S., Cai, C., Wang, Y., Huang, H., & Gui, G. (2020). Automatic modulation recognition
method based on deep learning in the presence of phase compensation. *IEEE Access, 8*,
42831–42847. https://lc.cx/jbrPdw
Van Luong, T., Shlezinger, N., Xu, C., Hoang, T. M., Eldar, Y. C., & Hanzo, L. (2022). Rolling
interference cancellation based on deep learning for non-orthogonal downlink. *IEEE
Transactions on Vehicle Technology, 71*(11), 11876–11888. https://lc.cx/hmq6rh
Wei, X., Hu, C., & Dai, L. (2021). Deep learning for beamspace channel estimation in millimeter-wave
massive MIMO systems. *IEEE Transactions on Communications, 69*(1), 182–193.
https://lc.cx/1jVMj9
Wu, T. (2019). Deep learning methods based on CNN and RNN for the demodulation of digital signals.
*ACM International Conference Series of Proceedings, Part F147767*, 122–127.
https://lc.cx/ihYmq4