An Augmented Structure of Neural Networks-Based Cancelers With High Isolation for Self-Interference Cancellation in In-Band Full-Duplex Radios

Self-interference (SI) is regarded as a main challenge in in-band full-duplex (IBFD) radios. In this letter, an augmented structure for neural networks (NNs)-based cancelers is proposed to suppress the nonlinear SI. By enhancing the higher-order nonlinear representations and incorporating the connection information of grid structure into the augmented structure, the proposed NNs-based cancelers significantly improve SI cancellation performance. Compared to existing NNs, the proposed NNs-based cancelers require less complexity when realizing comparable cancellation performance to the polynomial-based canceler, and improve the SI cancellation ratio by about 1 dB. Moreover, the residual SI signal power can be reduced to only 1.73 dB above the noise floor with minimal increase in complexity.