Neural Network-Based Self-Interference Cancellation for Frequency Division Duplex

Abstract:
In frequency division duplex systems with simultaneous transmission and reception, receivers suffer from outof- band emissions by the wireless transmitter leaking into the receive band. This so-called self-interference can be cancelled by means of estimating the interfering signal from the known transmit signal and subtracting it from the received signal. Conventional cancellation approaches rely on a mathematical model of the self-interference generation. However, these cannot capture effects beyond their model assumptions. With neural networkbased cancellation, present interference-generating effects are captured during data-driven training. In this work, we combine such a cancellation method with a neural network-based receiver for the demodulation of the signal-of-interest. Simulation results with 5G NR-conform waveforms show that this approach is both well-performing and adaptive to many self-interference and radio channel scenarios. Our approach is able to cancel selfinterference that has undergone power amplifier nonlinearity, IQ imbalance, and a frequency-selective channel down to a signalto- interference ratio of 0.0 dB.