Compressed Sensing and Deep Learning for Low-complexity Signal Detection in Communication Systems

• Ano de defesa: 2023
• Autor(a) principal: Souza de , Pedro
• Orientador(a): Mendes , Luciano , Souza, Rausley
• Banca de defesa: Mendes, Luciano , Souza , Rausley , Lopes, Estevan
• Tipo de documento: Tese
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Instituto Nacional de Telecomunica????es
• Programa de Pós-Graduação: Doutorado em Engenharia de Telecomunica????es
• Departamento: Instituto Nacional de Telecomunica????es
• País: Brasil
• Palavras-chave em Português: aprendizado de m??quina; redes neurais; sistemas de comunica????o; sensoriamento comprimido; detec????o de sinais; baixa complexidade.
• Área do conhecimento CNPq: Engenharia - Telecomunica????es
• Link de acesso: https://tede.inatel.br:8080/tede/handle/tede/239

The requirements involved in the new generation of mobile communications, combined with massive device connectivity, create complex heterogeneous scenarios. In these scenarios, classical models become limited, primarily due to the difficulty of modeling intractable mathematical relationships. Moreover, even if the optimum solution is available, it can be that its computational complexity is prohibitive in practice. Therefore, alternative ways of ap proaching these scenarios are desired, one promising approach being machine learning (ML) algorithms and neural networks (NNs). Harnessing NNs??? power for solving general optimization problems, for example, has an interesting appeal in such heterogeneous scenarios. On the other hand, compressed sensing was proposed as a technique to save storage and energy by compressing signals using simple linear transformations. Although compressed signals can be perfectly recovered, the complexity of the reconstruction operation is high. However, there are applications where compressive signals are processed directly in the compressed domain, with spectrum sensing being an example. This gave rise to an emerging concept, denoted as compressed learning (CL), that uses ML algorithms to extract information from ompressed signals. This work contribution is two pronged: (i) we investigate the CL concept applied to spectrum sensing for cognitive radios, where we propose a detector based on NNs to identify vacant channels from the compressed signal. For this, we assume perfect and imperfect channel state information and also dataset samples mismatch, where channel delay profile and statistics mismatches are considered; (ii) we moreover propose an architecture for deep NNs for multiple-input multiple-output (MIMO) systems, using the so-called deep unfolding concept. It is demonstrated that the proposed deep unfolding detector is orders-of-magnitude less complex, yet presenting no severe penalties in performance. Additionally, we propose a lattice reduction aided detector scheme for MIMO systems that achieves a similar diversity order to that of the optimum detector but also with significant less computational complexity.