Redes neurais recorrentes para modelagem e linearização de amplificadores de potência com controle de ganho discreto
Resumo
Resumo : Nos sistemas de transmissão sem fio, um componente crítico e que vem sendo alvo de vários estudos é o amplificador de potência (PA). Ele tem o papel de transmitir potência da bateria até o sinal que vai ser irradiado pelo sistema. O que o torna tão importante é o fato dele ser o componente com maior gasto energético do sistema, além de ser uma grande fonte de distorções para o sinal. Na literatura, existe uma arquitetura de um PA com multimodo de ganho, permitindo que o gasto energético seja controlado dependendo da demanda do sistema. Porém, o comportamento de comutar entre modos de ganho, além da própria não linearidade do PA, causam distorções indesejáveis para a saída do sistema. Técnicas de processamento digital podem ser empregadas para obter um desempenho melhor do circuito de transmissão, mas para isso é necessário conhecer o modelo matemático do PA. Neste trabalho será feito o estudo da rede neural recorrente (RNN) para a modelagem do amplificador de potência com multimodo de ganho. Esses modelos também serão usados para obter os parâmetros do modelo da Pré-Distorção Digital (DPD), que tem comportamento inverso ao do circuito do PA. Quando o DPD e o PA são posicionados em cascata, reduzimos as distorções e aumentamos a linearidade do sistema. Após o treinamento das redes neurais, avaliamos o desempenho de cada PA e, por fim, analisamos a linearidade do sistema com o uso do DPD em cascata Abstract : In wireless transmission systems, a critical component that has been the subject of several studies is the power amplifier (PA). Its role is to transmit power from the battery to the signal that will be radiated by the system. What makes it so important is the fact that it's the component with the highest energy consumption in the whole system, futhermore, it's a major source of distortion to the signal. In the literature, there is an architecture of a PA with
multi-gain mode, allowing energy consumption to be controlled depending on the system demand. However, switching between gain modes, in addition to the nonlinearity of the PA itself, causes undesirable distortions to the system output. Digital processing techniques can be used to obtain better performance from the transmission circuit, but it is necessary to know an accurate PA mathematical model. In this work, the recurrent neural network (RNN) will be studied for the power amplifier with multimode gain model. They will be trained to obtain the parameters of the Digital Pre-Distortion (DPD) model, which has the opposite behavior of the PA. When the DPD and PA are put in cascade, the distortion effects decrease and the linearity of the system improves. After training the neural networks, we evaluated the performance of each PA and, finally, analyzed the linearity of the system using cascaded DPD
Collections
- Engenharia Elétrica [39]