Uploaded by sharma.d1603

Physical Layer Using Deep Learning

advertisement
PROJECT WORK PRESENTATION ON
AN INTRODUCTION TO DEEP LEARNING FOR
THE PHYSICAL LAYER
Submitted in partial fulfillment of the requirements for the subject PROJECT WORK in VII SEMESTER
OF
BACHELOR OF TECHNOLOGY IN ELECTRONICS & TELECOMMUNICATION ENGINEERING
BY
AANCHAL SHARMA - 20116001
DEEPSHIKHA SHARMA - 20116028
SMITA PRAJAPATI - 20116097
PALAK SINGH - 20116905
Under the guidance of
DR. SAIKAT MAJUMDER
ASSISTANT PROFESSOR
DEPARTMENT OF ELECTRONICS AND COMUNICATION ENGINEERING
NIT Raipur
CONTENTS
I.
II.
III.
IV.
V.
VI.
VII.
VIII.
IX.
X.
INTRODUCTION
LITERATURE REVIEW
PROBLEM STATEMENT AND METHODOLOGY
IMPLEMENTATION
RESULTS AND DISCUSSIONS
CONCLUSION
FUTURE SCOPE
ACKNOWLEDGEMENT
TABLE OF FIGURES
REFERENCES
ACKNOWLEDGEMENT
We extend our deepest gratitude to Dr. Saikat Majumdar, Assistant Professor, Department
of Electronics and Communication Engineering, National Institute of Technology,
Raipur, for his invaluable guidance and mentorship throughout our major project on "Introduction
to Physical Layer with Deep Learning." Additionally, we acknowledge the wholehearted support
and encouragement provided by the Department of Electronics and Communication
Engineering, NIT Raipur. Our sincere appreciation goes to each member of our team whose
relentless dedication and collaborative efforts have been instrumental in the successful completion of
this project. Together, we have strived to combine our skills and knowledge, contributing to the
culmination of this endeavor. We are immensely thankful for the opportunities, guidance, and support
extended by these entities, which have significantly enriched our learning experience.
01
INTRODUCTION
●
Deep learning revolutionizes system design in communications, offering a novel
approach beyond traditional modeling of channels and hardware imperfections for
reliable data transfer.
●
Demonstrating practicality, they train neural networks to model integrated
transmitter-receiver systems within a single autoencoder NN, optimizing for
diverse channels and applications without known optimal solutions.
●
Expanding to multiple transmitter-receiver pairs, they employ "radio transformer
networks" (RTNs) to enhance deep learning expertise, improving symbol detection
performance in interference channels, and demonstrate CNN superiority for
modulation classification on complex IQ samples.
DATASET
●
RML2016.10a dataset includes three types of digital modulation signals and eight
types of analog modulation signals. The SNR of various modulated signals ranges
from −20 dB to 18 dB, with an interval of 2 dB. There are 20 kinds of SNRs in total.
The size of each modulated signal is (2, 128); 128 corresponds to 128 sampling
points per signal, and 2 corresponds to the input signal being a quadrature IQ twoway signal. The RML2016.10a dataset has a total of 220,000 signals.
● The dataset used for implementation consists of 1.2 million sequences of 128
complex-valued baseband IQ samples corresponding to ten different digital and
analog single-carrier modulation schemes (AM, FM, PSK, QAM, etc.) that have
gone through a wireless channel with harsh realistic effects including multipath
fading, sample rate and center frequency offset.
02
LITERATURE
REVIEW
●
The article explores using deep learning for communication system design, framing it as an
end-to-end task. It extends to multi-transmitter networks and introduces radio transformer
networks, showcasing CNNs for modulation classification using raw IQ samples.
●
The literature highlights challenges of implementing DL in the physical layer, offering
solutions like data augmentation, robust training, and hardware optimization for real-time
communication systems.
●
The paper is an invaluable introductory guide for researchers and practitioners aiming to
use DL techniques to enhance wireless communication system performance and reliability
in the physical layer.
03 PROBLEM STATEMENT
● To implement the research paper on “An Introduction to Deep Learning for the Physical Layer” by IEEE members Timothy
O’Shea and Jakob Hoydis. The research paper aims to explore and redefine communication system design in the physical
layer using deep learning (DL) techniques. It seeks to address challenges and limitations within traditional approaches by
proposing an end-to-end reconstruction task that optimizes both transmitter and receiver components simultaneously.
● In the project, we have attempted to implement some applications of deep learning for the physical layer which have been
mentioned in the paper. These include:
i. Joint optimisation of transmitter and receiver components in a single process by interpreting the communications system
as an Autoencoder.
ii. Extension of the Autoencoder idea to networks of multiple transmitters and receivers.
iii. Introduction to the concept of Radio Transformer Networks (RTNs) for performance improvement in fading channels.
iv. Attaining competitive accuracy with respect to traditional schemes for modulation classification by the application of
Convolutional Neural Networks (CNNs) on raw IQ samples.
OBJECTIVES
RADIO TRANSFORMER NETWORKS
FOR
AUGMENTED
SIGNAL
PROCESSING ALGORITHMS
AUTOENCODERS FOR END-TOEND COMMUNICATION SYSTEMS
01
02
AUTOENCODERS FOR MULTIPLE
TRANSMITTERS AND RECEIVERS
03
04
CNN FOR CLASSIFICATION TASKS
04 METHODOLOGY AND IMPLEMENTATION
A. AUTOENCODERS FOR END-TO-END COMMUNICATIONS
SYSTEMS
Figure 1. A simple communications system consisting of a transmitter and a
receiver connected through a channel
•
From a DL point of view, this simple communications system can be seen
as a particular type of autoencoder.
Figure 2. A communications system over an AWGN channel represented as an
autoencoder
•
The transmitter consists of a feedforward NN with multiple dense layers followed by
a normalization layer. The receiver is also implemented as a feedforward NN. Its last
layer uses a softmax activation.
•
In our case, the purpose of the autoencoder is different. It seeks to learn
representations x of the messages s that are robust with respect to the channel
impairments mapping x to y.
•
The goal of an autoencoder is to find a low-dimensional representation of its input at
some intermediate layer which allows reconstruction at the output with minimal
error.
IMPLEMENTATION OUTPUT
Figure 2. BLER versus SNR for Autoencoder (2,2)
Figure 4. BLER versus SNR for Autoencoder (8,8)
Figure 3. BLER versus SNR for Autoencoder (4,4)
Figure 5. BLER versus SNR for Autoencoder (7,4)
B. AUTOENCODERS FOR MULTIPLE TRANSMITTERS AND
RECEIVERS
Figure 3. The two-user interference channel seen as a combination of two interfering
autoencoders that try to reconstruct their respective messages.
Transmitter 1 wants to communicate message 𝑠1 ∈ 𝑀
to Receiver 1 while Transmitter 2 wants to communicate
message 𝑠2 ∈ 𝑀 to Receiver 2.
•
Both transmitter-receiver pairs are implemented as NNs and the only difference is
that the transmitted messages 𝑥1, 𝑥2 ∈ ∁𝑛 now interfere at the receivers,
resulting in the noisy observations .
•
From a DL point of view, the training approach consists of minimizing a weighted
sum of both losses, i.e., 𝐿 = α𝐿1 + 1 − 𝛼 𝐿2 for some 𝛼 ∈ 0,1 .
•
There are many other possibilities to train such a system and there is no optimal
approach. However, in our experiments, the desired result is identical BLERs for both
transmitter-receiver pairs.
IMPLEMENTATION OUTPUT
Figure 4. Implementation output of autoencoder using multiple transmitters and receivers
C. RADIO TRANSFORMER NETWORKS FOR SIGNAL
PROCESSING ALGORITHMS
Figure 5. A radio receiver represented as an RTN
One way of augmenting DL models with expert propagation
domain knowledge but not signal specific assumptions is
through the use of an RTN.
An RTN consists of three parts: (i) a learned parameter estimator 𝑔𝜔: ℝ𝑛 ↦ ℝ𝑃
′
(ii) a parametric transform 𝑡: ℝ𝑛 × ℝ𝑃 ↦ ℝ𝑛 (iii) a learned discriminative
′
network 𝑔: ℝ𝑛 ↦ 𝑀.
•
•
The basic functioning of an RTN is best understood as a problem of phase offset
estimation and compensation.
•
The goal of 𝑔𝜔 is to estimate a scalar, that is close to the phase offset 𝜙, which is
then used by the parametric transform 𝑡 to compute the signal and then fed into the
discriminative network 𝑔 for further processing, such as classification.
Implementation Output
Figure 6. Autoencoder training loss with and without RTN
D. CNNS FOR CLASSIFICATION TASKS
•
We look at the well-known problem of modulation classification of
single carrier modulation schemes based on sampled radio frequency
time-series data, i.e., IQ samples.
•
We leverage a series of narrowing convolutional layers followed by
dense/fully connected layers and terminated with a dense softmax
layer for our classifier.
•
We compare the classification accuracy of the CNN against that of
extreme gradient boosting with 1000 estimators, as well as a single
scikit-learn tree.
Implementation Output
Figure 7. CNN versus SNR performance comparison
Figure 8. Confusion matrix of CNN at SNR = 10 dB
CONCLUSION
01
02
Here you could describe the
topic of the section
Here you could describe the
topic of the section
03
04
Here you could describe the
topic of the section
Here you could describe the
topic of the section
Project Schedule
Status Report
Project Timeline
Upcoming Report
06 CONCLUSION
•
•
•
•
We implemented Autoencoder – (2,2), (4,4) , (8,8) and (7,4).
The autoencoder achieves the same BLER as uncoded BPSK for (2,2),
and it outperforms the uncoded BPSK for (8,8) over the full range of
Eb/N0.
The trained Radio Transformer Network Model is to be implemented on
dataset to study its performance.
Confusion Matrix is to be plotted for different modulation schemes to
determine the modulation classification accuracy of CNN.
07 REFERENCES
● Immortal, “GitHub - immortal3/AutoEncoder-Based-Communication-System: Tensorflow Implementation and result of Autoencoder Based Communication System From Research Paper : ‘An Introduction to Deep Learning for the Physical Layer’
http://ieeexplore.ieee.org/document/8054694/,” GitHub. https://github.com/immortal3/AutoEncoder-BasedCommunication-System.
● Gram-Ai, “radio-transformer-networks/radio_transformer_networks.py at master · gram-ai/radio-transformernetworks,” GitHub. https://github.com/gram-ai/radio-transformernetworks/blob/master/radio_transformer_networks.py
● RML2016.10b—https://radioml.com/datasets/radioml-2016-10-dataset/
Download