Efficient nonlinear self-interference cancellation with uplink training
Pilot-Projected Nonlinear Self-Interference Estimation with Simultaneous Uplink Training
Information Theory
Summary
In radios that send and receive signals simultaneously on the same channel, part of the signal they send can interfere with the signal they want to receive, creating a problem called self-interference. This paper studies ways to remove that interference digitally, even when a special test signal (pilot) for the uplink channel is also present during training. The authors propose two new estimation methods that handle the interference and the uplink channel better, making the cancellation more reliable and efficient. Their methods show improved performance in simulations, reducing interference below noise levels with moderate training effort.
What this means in practice
- •For wireless hardware engineers: Improve self-interference cancellation in full-duplex radios during simultaneous uplink channel training, enabling better signal reception with moderate training resources.
- •For telecommunications system designers: Design full-duplex communication systems with adaptive nonlinear self-interference estimation to enhance uplink data quality in noisy environments.
Authors
Seyed Mohammad Azimi-Abarghouyi, Thomas Eriksson
Abstract
This paper considers nonlinear digital self-interference (SI) cancellation and uplink-channel estimation for single-antenna in-band full-duplex radios when a desired uplink pilot is present during SI training. In this setting, a rich nonlinear SI dictionary can absorb part of the structured pilot response, while its large and correlated set of basis functions can make reliable SI estimation difficult, particularly with limited training. During SI-coefficient estimation, we treat the unknown uplink channel as a nuisance parameter and eliminate it to obtain an equivalent reduced SI estimation problem on the orthogonal complement of the pilot convolution subspace. Since the projected nonlinear dictionary can remain ill conditioned, we develop two complementary full-dictionary estimators: a pilot-projected ridge estimator with Bayesian-information-criterion tuning (PP-RidgeBIC), which uses global ridge shrinkage, and a pilot-projected group-sparse Bayesian estimator (PP-GSBE), which learns branch-dependent Bayesian shrinkage from the data. These estimators offer two complementary operating points: PP-RidgeBIC favors computational efficiency, whereas PP-GSBE provides greater adaptivity to heterogeneous nonlinear SI components. Simulations show substantial gains in SI cancellation and uplink reception over non-projected and hard-selection alternatives; notably, PP-GSBE suppresses the residual SI below the noise floor with moderate training and consistently provides the strongest performance under nonlinear model mismatch.