Advanced Neural Network Modeling Techniques for RF Power Amplifier Linearization

Abstract

Mobile communications form a backbone of modern society, with radio access networks serving as the wireless interface between mobile devices and the core network. With the evolution towards the 5G New Radio (NR) network, energy efficiency has emerged as a paramount concern alongside performance. The power amplifiers (PAs) in the wireless frontend significantly contribute to the network’s energy consumption due to a fundamental trade-off between PA efficiency and linearity. Digital predistortion (DPD) has become the standard approach for mitigating the PA’s inherent nonlinearity and allowing for an efficient PA operation. However, existing solutions face limitations in modeling the increasingly complex distortions occurring in PA broadband operation, e.g., in the 5G C-band (3.5 GHz), while meeting stringent computational constraints.

This thesis explores novel neural network (NN) approaches to improve the capability of DPD linearization. The research addresses several key challenges: efficiently expressing RF passband distortion using complex-valued baseband signals, modeling sophisticated dynamic distortions in wideband operation, compensating for long-term transient memory effects, and reducing computational complexity for practical implementation. To overcome these challenges, tailored NN modeling techniques are developed that achieve superior accuracy and lower complexity compared to the state-of-the-art.

Multiple new approaches are studied and proposed, resulting in several relevant contributions to the field. A novel phase normalization technique is proposed that aligns the NN processing with the physical origin of the PA distortions, demonstrating a 3 dB improvement in linearized adjacent channel leakage or reducing the complexity to one-third when compared to the relevant state-of-the-art reference. Building on this foundation, an optimized recurrent NN architecture is developed, which combines phase normalization with an efficient deep recurrent cell structure and time-delay processing. The proposed recurrent model achieves accurate linearization also in a challenging setting with a Gallium Nitride Doherty PA operating a non-contiguous multi-carrier signal with a 400MHz composite bandwidth in the 5G NR C-band. Furthermore, the thesis addresses the compensation of long-term transient effects occurring in a time division duplexing transmission, which has become the primary duplexing technique for new 5G frequency bands. A machine learning approach to model the PA gain transient is devised, allowing to fully compensate the gain transient while jointly predistorting short-term memory effects, achieving a consistently low error vector magnitude of below 0.5%. Finally, a mixture-of-experts NN approach is proposed that enables introducing processing sparsity by dynamically activating only a part of the NN. By exploiting this sparsity, a new tradeoff between runtime and model complexity is established. Associated DPD experiments demonstrate a runtime complexity reduction of about 50%, while maintaining a high linearization accuracy.

The contributions of this thesis confirm that properly tailored NN modeling techniques significantly advance the behavioral modeling capability of stateof- the-art DPD approaches. The conducted experimental RF measurement validation confirms the high suitability of the NN modeling approach for DPD linearization and its potential to offer a superior balance of performance and complexity. In conclusion, the proposed NN techniques are a promising technology for the DPD linearization systems of our next-generation radio networks.