Fast Electrical Balance Duplexer Tuning Using Neural Networks for RF Self-Interference Cancellation in In-Band Full-Duplex Systems

In in-band full-duplex systems, the self-interference of radio frequency poses a significant challenge that limits the operational efficiency of modern telecommunications infrastructure. This paper introduces a novel approach utilizing neural networks (NNs) to optimize impedance matching between the antenna and the balance network of an electrical balance duplexer. This method significantly enhanced the TX-RX isolation by adjusting the digitally controllable capacitor banks within the balance network. The proposed NN model efficiently navigates the extensive search space of these capacitor banks for estimating the probabilities of candidate configurations and selecting the top-k candidates to substantially reduce the search space, thereby improving the speed of the impedance matching process. The simulation results demonstrated that the NN approach achieved near-optimal isolation performance, thereby verifying it effectiveness. Specifically, the proposed method achieved near-optimal mean isolation within 0.18dB, whereas reducing the search space size from approximately 4.3 billion to 256 for a particular capacitor bank structure. This substantial reduction emphasizes the effectiveness of focusing the search exclusively within the candidate space predefined by the NN, thereby yielding much faster convergence.