Neural Network Aided Self-interference Cancellation for True Full-Duplex Wireless Communications

Abstract

The tremendous demand for mobile data services drives communication engineers to improve many aspects of wireless communication to achieve higher data rates and spectrum efficiency. True full-duplex communication systems have gained recent attention because these could double spectral efficiency relative to conventional systems. True full duplex uses the same frequency for simultaneous transmission and reception, so the locally transmitted signal would strongly interfere with the reception of the signal of interest. Therefore, the main challenge of implementing true full-duplex lies in cancelling the self-interference. Though different cancellation techniques have been developed, some require either complex circuitry or considerable computation. Also, most literature on this topic is simulation-based. Accordingly, the research presented in this thesis experimentally investigates full-duplex selfinterference cancellation techniques for an orthogonal frequency-division multiplexing (OFDM)- based communication system. A thorough assessment of the three main contemporary techniques and the joint cancellation schemes is conducted regarding different transmit power levels and signal bandwidths. The performance of the digital self-interference cancellation technique is associated with the accuracy of modelling the non-linear distorted self-interference signal, so models and simulations of the dominant analog non-linear components in OFDM full-duplex systems are presented. Additionally, novel neural networks applied self-interference cancellation techniques for true full-duplex communication systems are proposed. In particular, the non-linear auto-regressive exogenous (NARX) network-based canceller can achieve the same performance as the conventional polynomial model, with significantly higher computational efficiency and less memory requirement.