Applications of deep learning for physical layer of wireless communication systems

• Main Author: Xie, Yihang
• Other Authors: Teh Kah Chan
• Format: Thesis-Doctor of Philosophy
• Language: English
• Published: Nanyang Technological University 2024
• Subjects: Computer and Information Science; AI-5G
• Online Access: https://hdl.handle.net/10356/177369
• Institution: Nanyang Technological University

The objective of the thesis is to exploit the recent advances in deep learning-based techniques for wireless communication systems. The proposed methods aim to improve the performance and robustness of communication systems in different scenarios. The first research work proposes a deep learning-based receiver for non-orthogonal multiple access (NOMA) joint signal detection. The receiver combines channel estimation, equalization, and demodulation into an end-to-end process, which results in improved performance and robustness compared to traditional methods. The proposed method is tested with a tapped-delay line (TDL) channel model, which is commonly used in fifth generation (5G) communication systems. The second research work presents a deep learning-based receiver for orthogonal frequency-division multiplexing (OFDM) systems. The proposed receiver, named one-dimensional transmit and recovery network (1D-TRNet), is designed to be robust to non-linear clipping distortion. The 1D-TRNet receiver outperforms traditional OFDM receivers and state-of-the-art generalized approximate message passing (GAMP) receivers in terms of bit-error rate (BER) performance. The third research work proposes a deep learning-based 5G receiver, named communication transformer network (Comm-Trans Net), which is designed to be robust for different sub-types of TDL channels. The proposed receiver uses an attention mechanism to compensate for the multi-path fading effect of different OFDM subcarriers. The proposed positional encoding method for each OFDM subcarrier offsets the deep fading effect, resulting in improved performance compared to traditional MMSE channel estimation with GAMP receivers and state-of-the-art deep learning-based receivers. The fourth research work introduces a novel communication system called the diffusion model-based generative semantic communication (DM-GSC) system. The system utilizes the diffusion model and semantic embedding to enable efficient and robust communication in wireless networks. The proposed DM-GSC system shows promising results in robustness for different channel conditions. Overall, the thesis demonstrates the potential of deep learning-based techniques in improving the performance and robustness of wireless communication systems. The proposed methods have the potential to enhance the quality of service and enable novel communication scenarios in the future.