On Neural-Network Representation of Wireless Self-Interference for Inband Full-Duplex Communications

Gerald Enzner,Aleksej Chinaev,Svantje Voit,Aydin Sezgin

2024-10-02

Abstract:Neural network modeling is a key technology of science and research and a platform for deployment of algorithms to systems. In wireless communications, system modeling plays a pivotal role for interference cancellation with specifically high requirements of accuracy regarding the elimination of self-interference in full-duplex relays. This paper hence investigates the potential of identification and representation of the self-interference channel by neural network architectures. The approach is promising for its ability to cope with nonlinear representations, but the variability of channel characteristics is a first obstacle in straightforward application of data-driven neural networks. We therefore propose architectures with a touch of "adaptivity" to accomplish a successful training. For reproducibility of results and further investigations with possibly stronger models and enhanced performance, we document and share our data.

Signal Processing

What problem does this paper attempt to address?

The problem that this paper attempts to solve is the self - interference (SI) problem in wireless co - frequency full - duplex communication. Specifically, the paper explores how to use neural network models to represent and eliminate self - interference channels in order to achieve more efficient self - interference cancellation.

### Problem Background

In an Inband Full - Duplex (IBFD) communication system, the transmitter and receiver use the same frequency resources for simultaneous transmission. This requires the system to be able to effectively eliminate the strong self - interference signal from the transmitter to the receiver. Traditional self - interference cancellation methods include:

- **Active or passive shielding in the propagation domain**
- **Adaptive or non - adaptive cancellation in the analog receiving unit**
- **Adaptive cancellation in the digital baseband part**

These methods need to be used in combination to meet the self - interference cancellation requirement of approximately 100 dB.

### Application of Neural Networks in Self - interference Cancellation

Although neural networks perform well in many fields, they are less applied in self - interference modeling. The main reasons include:

- **Challenge of time - varying self - interference characteristics**: The characteristics of self - interference channels are usually time - varying, which poses a challenge to data - driven neural network training.
- **Lack of public data sets**: There are limited data sets for comparative analysis and benchmarking.

### Main Contributions of the Paper

To address the above challenges, this paper proposes a neural network architecture with "adaptive" capabilities, which can be successfully trained when dealing with time - varying linear and nonlinear elements. Specifically:

1. **Propose an adaptive architecture**: By introducing a part of trainable weights to represent invariant characteristics and another part of "adaptive" weights to fit the variable characteristics of the data.
2. **Publish a data set**: To ensure the reproducibility of results and further research, the authors publish their data set.

### Conclusion

The paper demonstrates the potential of neural networks in wireless self - interference representation and points out that model design needs to consider the variability of self - interference channels and nonlinear functions. Although the neural network model performs well in some cases, further research and improvement are still required to achieve self - interference cancellation at the noise level.

### Formula Summary

The formulas involved in the paper are as follows:

- Non - linear function of the power amplifier (PA):
\[
PA(s) = F(|s|)\cdot e^{j\cdot\arg(s)}
\]
where,
\[
F(|s|)=\arctan(c_f\cdot|s|)
\]

- Non - linear function of the analog - to - digital converter (A/D):
\[
AD(x) = G(|x|)\cdot e^{j\cdot\arg(x)}
\]
where,
\[
G(|x|)=\begin{cases}
|x| & \text{if } |x| < c_g\\
c_g & \text{if } |x| \geq c_g
\end{cases}
\]

- Self - interference signal generation:
- For the "Hammerstein" system:
\[
y_H[k]=PA(s[k])*h_{SI}[k]+n[k]
\]
- For the "Wiener" system:
\[
y_W[k]=AD(z[k]*h_{SI}[k]+n[k])
\]

where, \( * \) represents discrete - time convolution, and \( n[k] \) represents the remote signal and received noise, whose intensity is - 90 dB of the self - interference signal.

Through these formulas and model architectures, the paper demonstrates the application potential of neural networks in self - interference cancellation and points out the direction of future research.