Digital Predistortion for Fully-Connected Hybrid Beamforming Massive MIMO Transmitters Under Dynamic Directions

In practical applications, the main beam direction of massive multi-input multi-output (mMIMO) transmitters varies with the user's location. However, different directions will result in different distortions in mMIMO transmitters, and this poses a big challenge to digital pre-distortion (DPD). This paper first analyzes the relationship between direction configuration and main beam distortion in the fully-connected hybrid beamforming (FC-HBF) transmitters. In order to transmit the signal correctly under dynamic directions, two DPD models using heterogeneous neural networks (HNNs) based on vector decomposition (VD) and block-oriented (BO) structure, called HNN-VD and HNN-BO, respectively, have been proposed. The direction information is used directly as the input of HNNs. Then, a gating-based neural network model is proposed called direction adaptive neural network (DANN), which is composed of a gating network and a backbone network, and its core idea is to dynamically adjust the hidden layer features of the backbone network by utilizing direction information through the gating network. The simulation indicates that all of the proposed models exhibit good performance, but DANN has the best performance and low complexity among models, which provides a good digital pre-distortion scheme for FC-HBF transmitters under dynamic directions.