Supervised Learning for Nulling Angle Prediction with Measured Self-Interference Channels for Full-Duplex Isolation Enhancement

This paper introduces a neural network (NN)-based nulling control beamforming scheme that leverages measured self-interference (SI) channel data to minimize mutual coupling (MC) in full-duplex large-scale antenna arrays. Two NN models using an optimized nulling angle dataset generated by Particle Swarm Optimization algorithms are trained, one model incorporates convolutional layers, while the other consists solely of fully connected layers. The results demonstrate that both models achieve significant performance improvements, with an average beam-level MC reduction of 9.3 dB, resulting in an average Tx-Rx isolation exceeding 74.7 dB. Specifically, over 92.6% of downlink and uplink user locations exhibit enhanced beam-level isolation, and 46.5% of test cases achieve isolation improvements greater than 10 dB. Additionally, the training time efficiency and accuracy between the two NN models are investigated.