Machine learning techniques for self-interference cancellation in full-duplex systems

Selim, Mohamed Elsayed Mohamed (2024) Machine learning techniques for self-interference cancellation in full-duplex systems. Doctoral (PhD) thesis, Memorial University of Newfoundland.

[English] PDF - Accepted Version Available under License - The author retains copyright ownership and moral rights in this thesis. Neither the thesis nor substantial extracts from it may be printed or otherwise reproduced without the author's permission. Download (8MB)

Abstract

Full-duplex (FD), enabling remote parties to transfer information simultaneously in both directions and in the same bandwidth, has been envisioned as an important technology for the next-generation wireless networks. This is due to the ability to leverage both time and frequency resources and theoretically double the spectral efficiency. Enabling the FD communications is, however, highly challenging due to the self-interference (SI), a leakage signal from the FD transmitter (Tx) to its own receiver (Rx). The power of the SI is significantly higher when compared with the signal of interest (SoI) from a remote node due to the proximity of the Tx to its co-located Rx. The SI signal is thus swamping the SoI and degrading the FD system's performance. Traditional self-interference cancellation (SIC) approaches, spanning the propagation, analog, and/or digital domains, have been explored to cancel the SI in FD transceivers. Particularly, digital domain cancellation is typically performed using model-driven approaches, which have proven to be effective for SIC; however, they could impose additional cost, hardware, memory, and/or computational requirements. Motivated by the aforementioned, this thesis aims to apply data-driven machine learning (ML)-assisted SIC approaches to cancel the SI in FD transceivers|in the digital domain|and address the extra requirements imposed by the traditional methods. Specifically, in Chapter 2, two grid-based neural network (NN) structures, referred to as ladder-wise grid structure and moving-window grid structure, are proposed to model the SI in FD transceivers with lower memory and computational requirements than the literature benchmarks. Further reduction in the computational complexity is provided in Chapter 3, where two hybrid-layers NN structures, referred to as hybrid-convolutional recurrent NN and hybrid-convolutional recurrent dense NN, are proposed to model the FD SI. The proposed hybrid NN structures exhibit lower computational requirements than the grid-based structures and without degradation in the SIC performance. In Chapter 4, an output-feedback NN structure, referred to as the dual neurons-` hidden layers NN, is designed to model the SI in FD transceivers with less memory and computational requirements than the grid-based and hybrid-layers NN structures and without any additional deterioration to the SIC performance. In Chapter 5, support vector regressors (SVRs), variants of support vector machines, are proposed to cancel the SI in FD transceivers. A case study to assess the performance of SVR-based approaches compared to the classical and other ML-based approaches, using different performance metrics and two different test setups, is also provided in this chapter. The SVR-based SIC approaches are able to reduce the training time compared to the NN-based approaches, which are, contrarily, shown to be more efficient in terms of SIC, especially when high transmit power levels are utilized. To further enhance the performance/complexity of the ML approaches provided in Chapter 5, two learning techniques are investigated in Chapters 6 and 7. Specifically, in Chapter 6, the concept of residual learning is exploited to develop an NN structure, referred to as residual real-valued time-delay NN, to model the FD SI with lower computational requirements than the benchmarks of Chapter 5. In Chapter 7, a fast and accurate learning algorithm, namely extreme learning machine, is proposed to suppress the SI in FD transceivers with a higher SIC performance and lower training overhead than the benchmarks of Chapter 5. Finally, in Chapter 8, the thesis conclusions are provided and the directions for future research are highlighted.

Actions (login required)

View Item