Augmented Phase-Normalized Recurrent Neural Network for RF Power Amplifier Linearization

In this article, we present a highly accurate recurrent neural network (RNN) for behavioral modeling and digital predistortion (DPD) of radio frequency (RF) power amplifiers (PAs). We describe a deep, residual recurrent unit (RRU) that minimizes the overhead of the recurrent operation. Phase normalization is incorporated with the proposed unit to allow for efficient processing of the baseband signal phase with the real-valued RNN structure. Furthermore, we augment the phase normalization concept with dedicated envelope cell states that support the mapping of RF envelope dominated distortions. Combination with a trainable, input-ended finite impulse response (FIR) filtering leads us to proposing the augmented phase-normalized RRU (APNRRU). Our experimental validation, including a detailed modeling study of the proposed concepts with three different GaN Doherty PA units, as well as several DPD linearization examples, shows that the APNRRU offers excellent linearization already with modest complexity of just 550 model parameters. In addition, the results demonstrate the ability to linearize also demanding wideband PA operation with noncontiguous multicarrier signals with 400-MHz composite bandwidth, outperforming the prior art solutions.