Hybrid-Layers Neural Network Architectures for Modeling the Self-Interference in Full-Duplex Systems

Mohamed Elsayed, Ahmad A. Aziz El-Banna, Octavia A. Dobre, Wanyi Shiu, Peiwei Wang

TL;DR

This paper introduces two novel hybrid neural network architectures—HCRNN and HCRDNN—designed to eliminate self-interference in full-duplex wireless systems. By combining convolutional, recurrent, and dense layers, these models achieve state-of-the-art interference cancellation while slashing computational complexity by over 50% compared to traditional polynomial-based methods.

What Problem Does This Paper Address?

In the quest for 5G-and-beyond data rates, Full-Duplex (FD) technology is a "holy grail" because it allows a device to transmit and receive data simultaneously on the exact same frequency. Theoretically, this doubles spectral efficiency.

The catch? Self-Interference (SI). The device’s own powerful transmission leaks into its sensitive receiver, drowning out the incoming signal from distant nodes. While analog filters help, they aren't perfect. Digital cancellation is required to mop up the residual interference. Traditionally, engineers use complex polynomial models to mimic the non-linear distortions caused by hardware (like power amplifiers and IQ mixers). However, these polynomials are computationally "expensive," making them difficult to implement on battery-constrained mobile devices.

Methods

The authors propose a "hybrid" approach, arguing that different neural network layers have different strengths. They introduce two specific architectures:

  1. HCRNN (Hybrid-Convolutional Recurrent NN):
    • Input: Signal data is arranged into a 2D grid of In-phase (I) and Quadrature (Q) components over time.
  2. Convolutional Layer: Acts as a feature extractor. It identifies patterns like the "memory effect" and non-linearities at a reduced scale, significantly lowering the number of parameters.
  3. Recurrent Layer: Processes the sequence of features extracted by the CNN to learn the temporal behavior of the signal.
  4. HCRDNN (Hybrid-Convolutional Recurrent Dense NN):
    • Adds an extra Dense (fully connected) layer after the recurrent layer. This provides an additional "degree of freedom" to fine-tune the balance between performance and speed.

By using the CNN to reduce the dimensionality of the data before it hits the RNN, the authors avoid the massive computational overhead typical of deep recurrent networks.

Datasets

The study utilizes a publicly available dataset generated from a realistic FD hardware testbed.

  • Signal Type: 10 MHz QPSK-modulated OFDM (standard for modern wireless).
  • Sampling: 20 MHz.
  • Size: 20,480 time-domain baseband samples.
  • Split: 90% for training, 10% for verification.
  • Hardware Context: The testbed provided 53 dB of passive analog suppression, leaving the NN to handle the remaining digital cancellation.

Key Insights and Ideas

  • CNNs aren't just for images: By treating I/Q signal samples as a 2D grid, the authors successfully applied convolutional filters to extract structural features of the interference, which is much more efficient than feeding raw data into a dense network.
  • The "Hybrid" Advantage: Standard NNs usually stick to one type of layer (e.g., all-dense or all-recurrent). This paper shows that "specialization" works: CNNs for feature extraction + RNNs for sequence modeling = high accuracy at a low "cost."
  • Weight Sharing: The weight-sharing nature of CNNs is the primary driver for the reduction in parameters, solving the memory-bloat issue found in previous real-valued time-delay neural networks (RV-TDNNs).

Key Results

The proposed models were compared against traditional polynomial models and several state-of-the-art NNs (like RNNs and CV-TDNNs).

  • Cancellation Performance: All models achieved roughly 44.4 dB of cancellation, effectively bringing the interference down to the receiver's noise floor.
  • Computational Efficiency:
    • HCRNN reduced the number of floating-point operations (FLOPs) by 52.18% compared to the polynomial baseline.
  • HCRDNN achieved up to a 55.07% reduction in FLOPs.
  • Memory Footprint: The HCRDNN 2 model reduced the number of required parameters by 28.53%, making it much more suitable for hardware implementation in mobile platforms.

Limitations and Future Work

  • SISO vs. MIMO: This work focused on Single-Input Single-Output (SISO) systems. Modern 5G/6G systems use Massive MIMO (multiple antennas), which introduces complex "cross-talk" between antennas. The authors note that generalizing these hybrid NNs to MIMO environments is a critical next step.
  • Other ML Techniques: The authors suggest exploring other machine learning approaches, such as Support Vector Machines (SVMs), though they warn that the computational complexity of SVMs might be a hurdle.
  • Real-time Deployment: While the inference stage is optimized, the paper primarily focuses on offline training. Future research could look into online learning to adapt to changing wireless environments in real-time.

Summary updated

Abstract

Full-duplex (FD) systems have been introduced to provide high data rates for beyond fifth-generation wireless networks through simultaneous transmission of information over the same frequency resources. However, the operation of FD systems is practically limited by the self-interference (SI), and efficient SI cancelers are sought to make the FD systems realizable. Typically, polynomial-based cancelers are employed to mitigate the SI; nevertheless, they suffer from high complexity. This article proposes two novel hybrid-layers neural network (NN) architectures to cancel the SI with low complexity. The first architecture is referred to as hybrid-convolutional recurrent NN (HCRNN), whereas the second is termed as hybrid-convolutional recurrent dense NN (HCRDNN). In contrast to the state-of-the-art NNs that employ dense or recurrent layers for SI modeling, the proposed NNs exploit, in a novel manner, a combination of different hidden layers (e.g., convolutional, recurrent, and/or dense) in order to model the SI with lower computational complexity than the polynomial and the state-of-the-art NN-based cancelers. The key idea behind using hybrid layers is to build an NN model, which makes use of the characteristics of the different layers employed in its architecture. More specifically, in the HCRNN, a convolutional layer is employed to extract the input data features using a reduced network scale. Moreover, a recurrent layer is then applied to assist in learning the temporal behavior of the input signal from the localized feature map of the convolutional layer. In the HCRDNN, an additional dense layer is exploited to add another degree of freedom for adapting the NN settings in order to achieve the best compromise between the cancellation performance and computational complexity. Complexity analysis and numerical simulations are provided to prove the superiority of the proposed architectures.
BibTeX
@misc{elsayed2021hybrid,
  author = {Mohamed Elsayed and Ahmad A. Aziz El-Banna and Octavia A. Dobre and Wanyi Shiu and Peiwei Wang},
  title = {{Hybrid-Layers Neural Network Architectures for Modeling the Self-Interference in Full-Duplex Systems}},
  year = {2021},
  eprint = {2110.09997},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2110.09997},
}
Figure