Thesis

Deep learning for wireless communications : flexible architectures and multitask learning

Herunterladbarer Inhalt

PDF Herunterladen
Creator
Rights statement
Awarding institution
  • University of Strathclyde
Date of award
  • 2024
Thesis identifier
  • T17006
Person Identifier (Local)
  • 201563046
Qualification Level
Qualification Name
Department, School or Faculty
Abstract
  • Demand for wireless connectivity has never been higher and continues to grow rapidly. Connecting more devices requires mindfulness in managing the limited resources of energy and radio spectrum. The advent of Software Defined Radio (SDR) has enabled breathroughs in radio configurability, enabling dynamic spectrum access and physical layer optimizations at runtime. In recent years Machine Learning (ML) has been a key enabling technology of various innovations in the wireless communications domain, taking advantage of the newfound flexibility in SDR. The new ML-based signal processing models are no longer based entirely on Digital Signal Processing (DSP) expertise, but are developed in a data-driven approach. This paradigm shift in receiver design is recent, and appropriate architectures and best model training practices have yet to be established. This thesis explores multiple wireless communications tasks addressed with the toolbox of Deep Learning (DL), which is a subset of ML. Many existing DL solutions are hampered by the limitations of the chosen architectures, which limits their adoptability as drag-and-drop solutions by wireless system designers. Recurrent Neural Network (RNN) and Fully Convolutional Neural Network (FCN) architecture types are explored that enable the adaptability one would expect of classic DSP functions (like the filter). The field of wireless communications boasts a wealth of data, due to the mature and feature-rich simulation software ecosystem. In Radio Frequency Machine Learning (RFML) this is regularly leveraged to produce datasets for the new data-driven models. Techniques like Multitask Learning (MTL) can exploit this simulated data even further by allowing models to be trained on their primary task, like signal classification or demodulation, while simultaneously estimating the channel quality. The findings presented in this work show that fully convolutional architectures can be more appropriate for tasks like frame synchronization compared to commonly applied classification models. RNN-based autoencoders achieve good results as an end-to-end trainable receiver solution, however they can be challenging to apply to longer sequences. MTL is identified as an excellent technique not only for training unique models, capable of performing multiple tasks, but as a regularization technique in RFML.
Advisor / supervisor
  • Stewart, Robert W. (Electrical engineer)
  • Crockett, Louise H. (Louise Helen)
Resource Type
DOI

Beziehungen

Objekte

Miniaturansicht Titel Hochladedatum Sichtbarkeit Aktionen
file details PDF of thesis T17006 2024-06-26 Öffentlich Herunterladen