Neural Network Tuning for Analog-RF Self-Interference Cancellation

In-band full-duplex (IBFD) technology can empower multi-function wireless systems to simultaneously operate different modalities if the consequential self-interference is properly mitigated. Analog-RF cancellation represents one of the few techniques that allows for the additional suppression of transmit noise and nonlinearities before the receiver input, but has lacked the tuning speed required to enable advanced IBFD applications. This paper discusses a novel tuning approach based on machine learning principles, and introduces the optimal neural network configuration that drastically enhances tuning speed. The statistical performance of a prototype canceller was measured over a 20 MHz bandwidth centered at 2.5 GHz, and averaged more than 30 dB of cancellation with only 9 tuning iterations. These results are additionally compared to a survey of previously-reported cancellers, and reveal that this unique tuning approach provides the fastest convergence speed to date.