Thesis

Streaming-based convolutional neural networks for physical layer wireless communication receivers

Herunterladbarer Inhalt

PDF Herunterladen
Creator
Rights statement
Awarding institution
  • University of Strathclyde
Date of award
  • 2025
Thesis identifier
  • T17494
Person Identifier (Local)
  • 201885078
Qualification Level
Qualification Name
Department, School or Faculty
Abstract
  • With the rising demand for efficient wireless connectivity, Artificial Intelligence (AI) has become a key enabler for enhancing wireless radio system performance. Intelligent models support cognitive decision-making and real-time processing, enabling low-latency, edge-capable systems for Physical Layer (PHY) wireless communications tasks. While AI models perform well on Graphic Processing Unit (GPU)-based systems, edge deployment can significantly reduce latency and increase throughput, enabling operation in constrained environments. This thesis presents the development and evaluation of a streaming-based Convolutional Neural Network (CNN) accelerator for Software Defined Radio (SDR) receivers operating on live, real-time signals. Designed from first principles using a synchronous dataflow model, the architecture is purpose-built to align with the continuous dataflow of FPGA-based SDR pipelines, enabling per-sample processing without data loss. The accelerator is implemented on the AMD Zynq UltraScale+ Radio Frequency System-on-Chip (RFSoC) platform, demonstrating low-latency operation (29.6 µs) and high-throughput (34k classifications per second) real-time operation in PHY layer tasks. To support real-time deployment of CNN models on live signals captured by the ADC, this thesis introduces the DeepRFSoC dataset generation methodology, which enables training on realistic loopback data affected by simulated channel impairments and hardware-specific distortions. This methodology enables the accelerator to operate in real-time on the SDR platform processing live signals as they are captured. A quantisation investigation is presented, comparing Post-Training Quantisation (PTQ) and Quantisation-Aware Training (QAT) under real-time live signal reception conditions on the AMD RFSoC. Results show that 8-bit QAT models can outperform their floating-point counterparts by 3%, while 4-bit and 2-bit models maintain competitive accuracy with only a 2% reduction, demonstrating the viability of quantised CNN models for real-time PHY-layer inference on edge FPGA-based SDR platforms, contributing to the development of intelligent, low-latency SDR systems.
Advisor / supervisor
  • Stewart, Bob
  • Crockett, Louise H. (Louise Helen)
Resource Type
DOI

Beziehungen

Objekte

Miniaturansicht Titel Hochladedatum Sichtbarkeit Aktionen
file details PDF of thesis T17494 2025-11-04 Öffentlich Herunterladen