Complexity-Reduced Neural Network for Behavioral Modeling and Digital Predistortion of RF Wireless Transmitters

Neural networks (NNs) are promising for behavioral modeling and compensation of complicated nonlinearities in 5G transmitters with broadband high-efficiency structural power amplifiers (PAs). The high computational complexity of NNs, however, poses a serious challenge to their practical implementation. In response to this challenge, a complexity-reduced NN (CR-NN) approach is proposed in this paper, which builds on the relationship between data features and model capacity. Considering the memory fading property of RF PAs, the CR-NN first utilizes post-filtering to significantly reduce the input features of the NN body. This is followed by the employment of adaptive non-uniform piecewise linear unit to improve the model capacity without increasing the complexity. In order to validate the proposed method, the experiments are carried out based on a two-stage Doherty PA and a GaN-based Doherty PA. Experimental results show that the proposed CR-NN method can suppress the strong nonlinearity of PA from -23.24/-24.34 dBc to -51.18/-50.72 dBc with a computational complexity comparable to that of the linear-in-parameters models, thus demonstrating that the proposed method can significantly improve the performance-complexity tradeoff of NN-DPDs and contribute to the linearization of 5G and future systems.