Cross-Comparison of Neural Architectures and Data Sets for Digital Self-Interference Modeling

TL;DR

This paper evaluates and compares neural network models for digital self-interference cancellation in full-duplex communication, highlighting the effectiveness of simpler models and the benefits of a new Wiener-Hammerstein approach.

Contribution

It provides a cross-comparison of neural architectures and data sets, introducing a Wiener-Hammerstein model that improves generalization in self-interference modeling.

Findings

Hammerstein model performs well across diverse data sets.

Wiener-Hammerstein model enhances generalization.

Simpler models can be effective for self-interference modeling.

Abstract

Inband full-duplex communication requires accurate modeling and cancellation of self-interference, specifically in the digital domain. Neural networks are presently candidate models for capturing nonlinearity of the self-interference path. This work utilizes synthetic and real data from different sources to evaluate and cross-compare performances of previously proposed neural self-interference models from different sources. The relevance of the analysis consists in the mutual assessment of methods on data they were not specifically designed for. We find that our previously proposed Hammerstein model represents the range of data sets well, while being significantly smaller in terms of the number of parameters. A new Wiener-Hammerstein model further enhances the generalization performance.