Digital Predistortion of RF Power Amplifiers Using Recurrent Neural Networks

With the advance of wireless standards, the Fifth Generation (5G) communication systems are adopting higher carrier frequencies and wider signal bandwidth. While such an adaptation serves to high capacity communication, modern radio frequency (RF) systems inevitably face significant challenges in maintaining high linearity while minimizing power consumption. The RF power amplifiers (PAs) account for the majority of the power consumption and cause foremost nonlinear distortion in RF systems. This thesis presents several novel techniques based on Recurrent Neural Network (RNN)s to overcome some of the major challenges in Digital Predistortion (DPD) modeling for linearization of RF PAs in wideband applications of advanced wireless communication systems. The first part of the thesis focuses on adapting recurrent neural networks to DPD application to improve modeling accuracy by targeting low complexity. It starts with understanding the dynamics of RNNs and exploring the choice of an appropriate neural network (NN) to best model complex PA characteristics. And, it comes with a useful recurrent neural network structure that can be associated with a physically inspired PA model. Next, based on this network, we propose a new neural network structure that is particularly suitable for modeling the more complicated behavior of recent RF PAs. For this model, a novel signal pre-processing technique is developed to model the complex interaction between amplitude and phase information and then integrated into the selected gated recurrent structure. As a result, Phase-Gated JANET (PG-JANET) model can provide improved linearization performance with lower complexity in wideband applications compared to the existing models. The second part of the thesis presents another novel recurrent neural network based behavior model called Decomposed Vector Rotation-JANET (DVR-JANET) for digital predistortion of RF power amplifiers, again considering wideband scenarios. We propose here a new recurrent Decomposed Vector Rotation (DVR) scheme by representing the memory terms of the DVR model with recurrent states and redesigning its piecewise modeling according to the recurrent learning. To ensure a stable RNN modeling and better accuracy, we similarly integrate the recurrent DVR into the gated learning mechanism of modified Just Another NETwork (JANET), previously chosen network for PA modeling. The experimental results for the proposed model confirm that, this model provides much improved linearization performance with significantly reduced model complexity, compared to the recent DPDs available in the literature.