Hybrid Digital Pre-Distortion for Active Phased Arrays Subject to Varied Power and Steering Angle

Abstract

This letter proposes a memory polynomial (MPM)-aided deep neural network (DNN) digital pre-distortion (MaD-DPD) method for active phased arrays (APAs) subject to varied input power and steering angle. This has been challenging for traditional array linearization methods using either MPM or DNN, which rely on the in-phase and quadrature-phase (I/Q) signal as input and output to derive model parameters. In comparison, the proposed method actively incorporates MPM and DNNs to achieve linearization. The model uses only two varied APA state parameters (input power and steering angle) as input and the MPM coefficients as regression target, eliminating the need for model parameter updating. The MaD-DPD method is validated using a four-by-four antenna array at 28 GHz with 21 input power levels and a broad range of steering angles from −78° to 78°, improving up to 13.16% in error vector magnitude (EVM) and 18.21 dBc in adjacent channel leakage ratio (ACLR).