Federated learning empowered ultra-dense next-generation wireless networks

Al-Quraan, Mohammad Mahmoud Younes (2024) Federated learning empowered ultra-dense next-generation wireless networks. PhD thesis, University of Glasgow.

Abstract

The evolution of wireless networks, from first-generation (1G) to fifth generation (5G), has facilitated real-time services and intelligent applications powered by artificial intelligence (AI) and machine learning (ML). Nevertheless, prospective applications like autonomous driving and haptic communications necessitate the exploration of beyond fifth-generation (B5G) and sixth-generation (6G) networks, leveraging millimeter-wave (mmWave) and terahertz (THz) technologies. However, these high-frequency bands experience significant atmospheric attenuation, resulting in high signal propagation loss, which necessitates a fundamental reconfiguration of network architectures and paves the way for the emergence of ultra-dense networks (UDNs). Equipped with massive multiple-input multiple-output (mMIMO) and beamforming technologies, UDNs mitigate propagation losses by utilising narrow line-of-sight (LoS) beams to direct radio waves toward specific receiving points, thereby enhancing signal quality. Despite these advancements, UDNs face critical challenges, which include worsened mobility issues in dynamic UDNs due to the susceptibility of LoS links to blockages, data privacy concerns at the network edge when implementing centralised ML training, and power consumption challenges stemming from the deployment of dense small base stations (SBSs) and the integration of cutting edge techniques like edge learning. In this context, this thesis begins by investigating the prevailing issue of beam blockage in UDNs and introduces novel frameworks to address this emerging challenge. The main theme of the first three contributions is to tackle beam blockages and frequent handovers (HOs) through innovative sensing-aided wireless communications. This approach seeks to enhance the situational awareness of UDNs regarding their surroundings by using a variety of sensors commonly found in urban areas, such as vision and radar sensors. While all these contributions share the common goal of proposing sensing-aided proactive HO (PHO) frameworks that intelligently predict blockage events in advance and performs PHO, each of them presents distinctive framework features, contributing significantly to the improvement of UDN operations. To provide further details, the first contribution adhered to conventional centralised model training, while the other contributions employed federated learning (FL), a decentralised collaborative training approach primarily designed to safeguard data privacy. The utilisation of FL technology offers several advantages, including enhanced data privacy, scalability, and adaptability. Simulation results from all these frameworks have demonstrated the remarkable performance of the proposed latency-aware frameworks in improving UDNs’ reliability, maintaining user connectivity, and delivering high levels of quality of experience (QoE) and throughput when compared to existing reactive HO procedures lacking proactive blockage prediction. The fourth contribution is centred on optimising energy management in UDNs and introduces FedraTrees, a lightweight algorithm that integrates decision tree (DT)-based models into the FL setup. FedraTrees challenges the conventional belief that FL is exclusively suited for Neural Network (NN) models by enabling the incorporation of DT models within the FL context. While FedraTrees offers versatility across various applications, this thesis specifically applies it to energy forecasting tasks with the aim of achieving the energy efficiency requirement of UDNs. Simulation results demonstrate that FedraTrees performs remarkably in predicting short-term energy patterns and surpasses the state-of-the-art long short-term memory (LSTM)-based federated averaging (FedAvg) algorithm in terms of reducing computational and communication resources demands.

• Item Type: Thesis (PhD)
• Qualification Level: Doctoral
• Additional Information: Supported by funding from a UESTC scholarship.
• Subjects: T Technology > TK Electrical engineering. Electronics Nuclear engineering
• Colleges/Schools: College of Science and Engineering > School of Engineering
• Supervisor's Name: Mohjazi, Dr. Lina, Zoha, Dr. Ahmed, Centeno, Dr. Anthony and Imran, Professor Muhammad
• Date of Award: 2024
• Depositing User: Theses Team
• Unique ID: glathesis:2024-84124
• Copyright: Copyright of this thesis is held by the author.
• Date Deposited: 14 Mar 2024 17:22
• Last Modified: 14 Mar 2024 17:24
• Thesis DOI: 10.5525/gla.thesis.84124
• URI: https://theses.gla.ac.uk/id/eprint/84124

Related URLs

• Article DOI
• Article DOI
• Article DOI
• Article DOI
• Conference proceeding
• Conference proceeding
• Conference proceeding
• Conference proceeding