Enhancing Spectrum Prediction and Awareness with Deep Learning Approaches
No Thumbnail Available
Date
2024-05
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Addis Ababa University
Abstract
The growing demand for ubiquitous connectivity and real-time applications has placed significant
strain on the finite radio spectrum, making efficient utilization crucial. To address this,
static allocation methods must be replaced with real-time spectrum management that ensures
optimal sharing, utilization, and management of spectrum bands among multiple users. By
building comprehensive Spectrum Awareness (SA) through continuous monitoring and situational
understanding of the radio environment and network conditions, and by intelligently
predicting spectrum availability and usage through accurate Spectrum Prediction (SP), new
opportunities for wireless services can be unlocked, paving the way for the next generation of
connected technologies.
In that regard, this PhD thesis aims to enhance SP and SA tasks using Deep Learning (DL)
techniques, addressing three primary challenges: achieving accurate long-term spectrum
predictions with limited data, understanding spectrum usage across multiple dimensions, and
developing robust prediction models to improve accuracy. These challenges are tackled across
three network perspectives: Cognitive Radio Networks (CRNs), Mobile Network Operators
(MNOs), and Unmanned Aerial Vehicle (UAVs)-assisted networks.
First, long-term spectrum predictability and multi-dimensional analysis are explored for CRNs
using a distance-dependent central data fusion center and predictors based on Long Short-
Term Memory (LSTM) and Convolutional Long Short-Term Memory (ConvLSTM) networks.
These models analyze temporal and spatial dependencies to generate single- and multi-location
interpolated spectrum data. Evaluation of prediction performance showed, LSTM models
performing best for lower-frequency bands with deterministic Primary Users (PU) patterns,
achieving an improvement by 9.7% over the baseline model and a prediction error below 5
dBm for 2.5 hours, using only 75 minutes of past monitoring data. Conversely, ConvLSTM
models excel at higher-frequency bands by processing interpolated spectrum maps, yielding
approximately a 14% improvement in prediction accuracy. With limited PU knowledge and
sparse sensor deployment, increasing the cooperative region provides minimal accuracy gains,
with a maximum error probability reduction by 0.15 at 1500 meters coverage radius.
Second, the SP problem for MNOs is tackled using voice and data traffic from Global System for
Mobile (GSM) and Long-Term Evolution (LTE) networks as proxies for spectrum utilization. By
mapping traffic data to channel utilization and employing DL-based models, spectrum utilization
was predicted accurately without direct monitoring. Voice traffic, with its deterministic nature,
allowed for pre-mapping before prediction, while LTE data traffic required a post-prediction
mapping approach to prevent error propagation. For this prediction task, different DL models
based on Convolutional Neural Networks (CNNs) and LSTMwere proposed, and their prediction
accuracy was assessed using Root Mean Squared Error (RMSE) and Mean Absolute Error (MAE) voice channel utilization, improving performance by 26.9% at the cluster level and 36% at the
base station level due to its ability to capture temporal patterns. On the other hand, for LTE
traffic prediction, the CNN-LSTM hybrid model achieved a 40% improvement in RMSE at the
cluster level, effectively managing multivariate data and detecting peak-hour irregularities.
Lastly, for wireless networks with 3D architectures, such as UAV-assisted systems, Spectrum
Situational awareness (SSA) is achieved with Volumetric Radio Environment Maps (VREMs)
that represent propagation loss across both areal and altitude dimensions. This work introduces
two novel DL approaches for constructing the VREMs: the Volume-to-Volume (Vol2Vol) method
and the Sliced-VREM method. The Vol2Vol approach directly models 3D data using a 3D-GAN,
achieving high accuracy at greater altitudes, with a Structural Similarity Index (SSIM) of
0.9447. On the other hand, the Sliced-VREM method leverages stacked 2D environmental
maps and transmitters with an altitude-aware Spider-UNet to efficiently capture altitude
dependencies, making it more computationally practical for lower-altitude scenarios. Both
approaches demonstrate significant performance improvements over traditional 2D baselines,
with MAE reductions of up to 62.9%. Furthermore, preliminary spectrum occupancy maps
generated from these VREMs reveal spatial and altitudinal variations, even at low heights of
9 m or less, showcasing their potential to provide actionable insights for dynamic spectrum
sharing without relying on extensive monitoring infrastructure.
Description
Keywords
Spectrum prediction(SP), Spectrum awareness (SA), Cognitive Radio Networks (CRNs), Long-Term Prediction, Deep Learning, Volumetric Radio Environment Maps (VREM), Traffic-Driven Models, Spectrum Situational Awareness (SSA)