A CNN-Based Digital Predistortion Method for Multi-Beam Phased-Array Transmitters

This paper presents a multi-beam digital predistorter (MB-DPD) for phased-array transmitters. MB-DPD uses a convolutional neural network (CNN) to compensate for nonlinear distortions, hardware impairments, and cross-beam coupling over wide ranges of multiple beams’ steering angles, without the need for interpolation. By encoding steering angles’ information alongside the I/Q baseband signals of the main and neighboring subarrays, the proposed MB-DPD scheme mitigates both in-band and out-of-band distortions for any set of beam angles’ values that fall within the training angles’ ranges. Furthermore, MB-DPD uses transfer learning to reduce the need for retraining of the CNN layer hyperparameters as the steering angles’ ranges are altered or extended. Experimental validation with a 15 MHz signal at 3.5 GHz on a 20-element phased-array demonstrates that MB-DPD improves the normalized mean square error (NMSE) by about 10 dB across the beams’ ranges and significantly improves the adjacent channel power ratio (ACPR) by about 15 dB, and it is benchmarked against beam-oriented DPD (BO-DPD), and deep neural networks (DNN) to show the robustness and distortion mitigation of MB-DPD in various multi-beam setting conditions without any need for interpolation.