With the emergence of beyond 5G and 6G communication systems, the increasing demand for wireless connectivity has intensified the scarcity of available radio spectrum. Traditional fixed spectrum allocation policies grant exclusive access to licensed Primary Users (PUs) even when the spectrum remains idle, which leads to inefficient spectrum utilization. Dynamic spectrum management addresses this limitation by enabling Secondary Users (SUs) to opportunistically access underutilized frequency bands through spectrum sensing. In this context, Deep Learning (DL)-based spectrum sensing techniques have attracted increasing interest due to their ability to learn discriminative signal features directly from raw Radio Frequency (RF) data. These methods can outperform conventional sensing approaches that are prone to missed PU detection and traditional Machine Learning (ML) approaches that rely on handcrafted features. The first contribution of this thesis is a systematic analysis and benchmarking of DL-based spectrum sensing using publicly available RF datasets. Multiple Deep Neural Network (DNN) baselines and proposed architectures are evaluated under consistent experimental settings using RadioML2016.10A (RML16) and RML22-derived datasets. This study highlights key limitations in existing evaluation practices, including strong dependence on simulation-only datasets, limited investigation of signal representation effects, and increasing architectural complexity that can hinder efficient deployment. The second contribution is the design and evaluation of lightweight state-space-based spectrum sensing models. Two DNN architectures, termed MambaSenseLite and MambaSense, are proposed and tested across a wide Signal-to-Noise Ratio (SNR) range, modulation conditions, and two signal representations, namely In-phase/Quadrature (I/Q) and Amplitude/Phase (A/P). On the RML16 dataset, MambaSenseLite achieves up to 13.43% higher sensing accuracy than the best-performing baseline at -8 dB SNR. It also provides 96.72% lower inference latency than the best DNN baseline, with an inference time of 0.02 ms per sample. The results further show that signal representation can significantly affect sensing performance for weaker DNN models, where using the A/P format instead of I/Q improves average low-SNR sensing accuracy by up to 12.81%.The third contribution advances spectrum sensing evaluation beyond simulated data through the collection of hardware-based datasets using software-defined radios. Two datasets are developed using Universal Software Radio Peripheral (USRP) B210 devices, including an Over-the-Air (OTA) dataset that captures practical channel and hardware effects and a controlled Additive White Gaussian Noise (AWGN)-only dataset obtained via a wired configuration. A lightweight Convolutional Neural Network (CNN) incorporating dilated convolutions is proposed to efficiently extract temporal features from raw I/Q samples. On real OTA measurements at -6 dB SNR, the proposed CNN outperforms the second-best baseline by 19.35%, while being 40.5% smaller in model size and requiring 85.18% fewer floating-point operations. Cross-dataset analysis also shows that reliable sensing on real OTA data requires substantially higher SNR than synthetic benchmarks. Overall, this thesis provides a systematic analysis of how dataset characteristics, signal representation, and DNN design influence spectrum sensing performance. By combining efficient DL architectures with extensive evaluation on both public benchmarks and hardware-collected datasets, this work contributes practical insights towards reliable and deployable spectrum sensing for next-generation wireless communication systems.
| Date of Award | 16 Jun 2026 |
|---|
| Original language | English |
|---|
| Sponsors | European Union's Horizon 2020 Research & Innovation Programme |
|---|
| Supervisor | Pavlos Lazaridis (Main Supervisor) & Qasim Ahmed (Co-Supervisor) |
|---|