REPOSITORY > RESULTS

Abstract

Wireless communication networks have evolved significantly over the past decade, driven
by rapidly growing traffic demands and device diversity. This evolution has intensified pressure
on the radio-frequency spectrum, a limited medium where usage remains uneven and
time-varying, making spectrum optimization a highly relevant research direction. Reliable
optimization requires continuous awareness of the dynamic radio environment, encompassing
transmission detection, technology recognition, and usage pattern tracking, which can
enable closed-loop control of network parameters.
Deep learning-based models have shown remarkable signal recognition abilities over
the last decade, significantly outperforming classical approaches in classification tasks.
However, their deployment is constrained by the cost and scarcity of labeled data in realworld
conditions. The emergence of open radio access network (O-RAN) architectures
and open testbeds, exposing large-scale unlabeled spectrum data through standardized
interfaces, creates a timely opportunity for unsupervised deep learning approaches that can
learn from unlabeled data automatically, reducing or eliminating the necessity of expert
annotation.
This dissertation investigates unsupervised deep learning, with particular emphasis on
deep clustering, as a scalable and practically deployable approach for radio access technology
(RAT) recognition from unlabeled spectrum measurements. The main contributions
are: (i) a comprehensive review of unsupervised deep learning for wireless signal
recognition, linking recognition challenges to suitable unsupervised learning families and
identifying open research directions; (ii) adaptation and benchmarking of deep clustering
architecture for RAT recognition from spectrograms, based on established deep learning
modules; (iii) a label-free spectrum sensing workflow that requires no prior knowledge of
active RATs or operating conditions and uses post-hoc cluster-to-class mapping; and (iv)
a lightweight CNN integrated into the deep clustering framework and tailored to the fast
Fourier transform (FFT)-amplitude representation of RAT transmissions.
The results demonstrate that deep clustering consistently learns more discriminative
representations than baseline approaches. Systematic adaptation of image processing modules
for spectrogram processing yields reductions in model complexity of up to one order of
magnitude while preserving or improving clustering performance by up to 2.5 times. A dedicated
CNN module for FFT amplitudes further reduces the parameter count by two orders
of magnitude compared to the module for processing spectrograms, and by 22% compared
to the state of the art approach. At the same time it reduces computational cost by 50%
and improves feature separability and clustering quality by 35 percentage points relative to
an alternative architecture using the same modules. Validation across multiple real-world
datasets collected in diverse radio environments confirms the consistency and practical viability
of the proposed approaches, contributing to data-driven spectrum awareness as a
foundational capability for future AI-native wireless communication networks.