Logo image
Federated learning empowered interference management in 5G ultra-dense networks
Dissertation   Open access

Federated learning empowered interference management in 5G ultra-dense networks

Lucky Oghenechodja Daniel
Doctor of Philosophy (PHD), University of Johannesburg
Handle:
https://hdl.handle.net/10210/517210

Abstract

5G mobile communication systems Computer networks Neural networks (Computer science) Federated learning (Machine learning) Wireless Communication Systems Machine Learning
The emergence of Ultra-Dense Networks (UDNs), a technology that boosts the performance of fifth-generation (5G) telecommunication networks, has introduced significant throughput loss due to network interference. Numerous types of interference are largely reported in the literature. These include inter-cell interference (ICI) characterized by distinct user equipment (UEs) competing for the use of the same frequency bandwidths. In addition, there is also cell-edge interference (CEI) that results from UEs which simultaneously receive signals from 5G equipment at different frequency bandwidths. Co-channel interference (CCI) is another notable form of interference that is caused by the presence of dense materials in between satellite and UEs (CRI) that leads to two distinct points in the network using the same channels. Limitations in resource allocation can also cause cross-tier interference if the number of UEs that are anticipated to use the same resources grows significantly large. Co-tier interference (CTI) is yet another example in which excessive use of frequency spectrum through reuse causes interference among the supported UEs. This thesis presents distinct solutions to these various types of network interference. The first solution in the thesis is a hybridized model with a Long Short-Term Memory (LSTM) neural network component. This LSTM model inputs twenty one features namely: uplink times, downlink times, total online time, user displacement, user velocity, user position and many others, and returns lag one future values of only six features namely: uplink times, downlink times, total online time, user displacement, user velocity and user traffic load definition. Upon successful training of the LSTM model, the model is used to forecast future values of the 6 features with historic values of 21 features supplied as inputs to the model. This is performed using the first set of unseen data to avoid training bias. The estimated future values that are retrieved from the trained LSTM model are injected as inputs in a Generalized Linear Model (GLM) that has its target variable as UEs’ required frequency bandwidths. In this manner, the thesis has successfully introduced self-aware UDNs that can self-configure and restrict various UEs access to specific frequency bandwidths in a manner that prohibits network interference and enhances
pdf
Mr_Daniel_PhD_Thesis11.57 MBDownloadView
Open Access

Metrics

2 File views/ downloads
21 Record Views

Details

Logo image