Machine Learning Applied to 5G and Towards 6G Fiber/Wireless Systems Linearization

• Autor(a) principal: Pereira, Luiz Augusto
• Data de Publicação: 2023
• Tipo de documento: Tese
• Idioma: por
• Título da fonte: Biblioteca Digital de Teses e Dissertações da INATEL
• Texto Completo: https://tede.inatel.br:8080/tede/handle/tede/250
• topic: 5G; 6G; aprendizado de m??quina; p??s-distor????o; pr??-distor????o; redes neurais artificiais e r??dio sobre fibra anal??gico;; 55G; 6G; analog radio over fiber; ANN; machine learning, post-distortion and pre-distortion;; Engenharia - Telecomunica????es

The thesis main goals are proposing and evaluating machine learning (ML)-based linearization schemes for analog radio over fiber (A-RoF) systems from the fifth and sixth generations of mobile communications (5G and 6G). The performance analyses of the linearization techniques were based on the use of polynomial models to represent the non-linearities of the A-RoF system and artificial neural networks (ANNs) to perform pre-and/or post-distortion of the transmitted signal. The main A-RoF system non-linear components were taken into account. First, a linearization scheme was developed for the Mach-Zehnder modulator (MZM). It employs a multi-layer perceptron (MLP) ANN to perform the pre-and/or post-distortion of the transmitted signal. The electrical power amplifier (PA) non-linear effects were also evaluated. PA is commonly used in A-RoF systems to amplify the radio frequency (RF) signal before being radiated. In this case, it is required a recurrent neural network (RNN), which is able to deal with the power amplifier memory effect. RNN has an internal memory structure that can be properly designed to compensate for the non-linear degradation, considering the memoryless non-linear distortions introduced by the MZM and the memory non-linear distortion introduced by the PA. Finally, the chromatic dispersion (CD) introduced by the optical fiber was also taken into account, since the transport optical link needs to be extended to provide coverage in remote areas. The augmented real-valued time delay neural network (ARVTDNN) was used, since it allows simultaneously compensating all the aforementioned effects. All results were obtained by using Python simulations. This work also presents the application of ARVTDNN linearization scheme into a fiber/wireless (FiWi) system. Basically, the obtained results from this thesis demonstrate that ML-based linearization schemes represent potential solutions to maximize the performance of A-RoF systems from the 5G and 6G transport networks.