Joint Detection and Self-Interference Cancellation in Full-Duplex Systems Using Machine Learning

The fundamental challenge in full-duplex (FD) communications is to cancel the strong self-interference (SI) signal. SI cancellation is usually implemented by a combination of passive SI cancellation, active analog SI cancellation, and active digital SI cancellation. A part of the SI cancellation needs to be carried out in the analog domain to avoid saturating the analog front-end of the receiver, but digital cancellation is generally easier to implement using DSP circuits. Polynomial models are often used in practice to model transceiver non-linearities, but they typically have a very large number of trainable parameters which translates into a high computational complexity. More recently, various machine learning methods have been successfully used to perform non-linear digital SI cancellation as a lower- complexity alternative to polynomial models. However, in all of the aforementioned works the SI cancellation and the signal-of- interest detection are treated independently. In this work, we propose to use a neural network for joint detection and nonlinear SI cancellation. Preliminary experimental results with a measured dataset from a full-duplex testbed show that this joint approach can provide bit-error rate improvements of 1 to 3 dB. We also discuss some limitations of our current method and we outline directions for future research.