Towards artificial intelligence in visible light communication systems

• Ano de defesa: 2023
• Autor(a) principal: Costa, Wesley da Silva
• Tipo de documento: Tese
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal do Espírito Santo; BR; Doutorado em Engenharia Elétrica; Centro Tecnológico; UFES; Programa de Pós-Graduação em Engenharia Elétrica
• Palavras-chave em Português: Eficiência de potência; Eficiência espectral; Equalização OFDM; Inteligência Artificial; Metaheuristicas; subject.br-rjbn; Engenharia Elétrica
• Link de acesso: http://repositorio.ufes.br/handle/10/17088

In recent years, the increasing applications of Internet of Things (IoT) and smart devices have accelerated the demand for signal bandwidth. However, radiofrequency (RF) wireless systems cannot meet this upcoming need because of spectrum congestion in urban areas and insufficient bandwidth, mainly in indoor environments. These facts pave the way for alternatives to reduce the pressure of the RF spectrum in such conditions and also ensure high data rates, low latency, reliability, and low cost. The advance of Light-Emitting Diode (LED) technology provided high energy efficiency lighting with a high-speed light intensity switching. These facts, along with the possible spectrum crunch, have given rise to research interests in Visible Light Communication (VLC), through which data are transmitted using the existing lighting infrastructure. VLC offers a complementary alternative to radio-based systems, with an unlicensed optical spectrum (approximately 400 THz), security at the physical layer, low power, high speed, and immunity to RF electromagnetic interference. A high data rate can be achieved by combining the broadband VLC channel and multicarrier modulation schemes. Orthogonal Frequency Division Multiplexing (OFDM) is largely studied because of its spectral efficiency promotion and capacity to deal with multipath fading. However, the nonlinearity introduced by the LED offers a challenge to the OFDM parameters settings, due to its high Peak-to-Average Power Ratio (PAPR). This work tackles the challenge of conveying OFDM signal considering the Intermodulation Distortion (IMD) produced by the LED nonlinearity. This Thesis addresses the nonlinear LED effects, VLC parameters, and its performance for conventional and constant-envelope OFDM. The VLC system is modeled and optimization algorithms are evaluated to achieve parameters that provide maximum power and spectral efficiencies, constrained by modulation figure of metrics: bit error rate and error vector magnitude. This work also presents the application of artificial neural networks in the physical layer of VLC systems. The Long Short-Term Memory (LSTM) neural network is applied to predict future positions, as well as channel gain, and also forecast optimized parameters. Additionally, an investigation of the OFDM equalization using deep learning architectures for a multipath single-input single-output VLC channel is proposed. Convolutional Neural Network (CNN) architectures are applied in a direct OFDM mapped symbols equalization, without channel estimation, interpolation, nor element-wised division, denominated Convolutional Neural Network Direct-Equalizer (CNN-DE). Results show that the optimization based on meta-heuristics was capable to determine the VLC parameters in order to satisfy the communication constraints. Additionally, they emphasize the trade-off between power and spectral efficiency in VLC: higher spectral efficiency is achieved with the increase of offset current (power) to deal with the IMD; in contrast, to achieve higher power efficiency, a lower spectral efficiency is obtained. The optimization results using constant-envelope outperforms the conventional OFDM with the proper choice of the phase modulation index. The LSTM showed as a powerful tool for routing forecasting and assessing the optimized parameters. The CNN-DE equalizer (regression) was capable of detecting the correct symbol for lower SNR (≤ 10 dB). Additionally, the classification version of the CNNDE was able to predict and classify the mapped symbols similarly to the least-square-based equalization.