Deep Learning Based Detection for Spectrally Efficient FDM Systems

David Picard, Arsenia ChortiarXiv 2021 · DOI · arXiv

Abstract

In this study we present how to approach the problem of building efficient detectors for spectrally efficient frequency division multiplexing (SEFDM) systems. The superiority of residual convolution neural networks (CNNs) for these types of problems is demonstrated through experimentation with many different types of architectures.

Introduction

Modern wireless communication relies heavily on Orthogonal Frequency Division Multiplexing (OFDM), where sub-carriers are mathematically orthogonal to prevent interference. While effective, this orthogonality limits spectral efficiency. Spectrally Efficient Frequency Division Multiplexing (SEFDM) overcomes this by overlapping sub-carriers, thereby packing more data into the same bandwidth. However, this introduces significant Inter-Carrier Interference (ICI), warping the decision boundaries (Voronoi partitions) and making traditional optimal detection computationally prohibitive. This research investigates the efficacy of deep neural networks in learning to decode these complex, interference-heavy signals.

Methodology

The detection problem is framed as a classification task: predicting the original symbol class $z$ from a received signal vector $x$. Four neural architectures were evaluated:

  1. Multi-Layer Perceptron (MLP): Standard fully connected layers.
  2. Residual MLP (ResMLP): MLPs incorporating skip connections to facilitate gradient flow.
  3. Convolutional Neural Networks (CNN): Networks utilizing sliding filters to exploit the structured nature of communication signals.
  4. Residual CNN (ResCNN): A hybrid architecture combining convolutional layers within residual blocks.

Signals are processed as 2D real vectors (representing complex values) using ReLU activation. The experimental setup utilized $N=32$ sub-carriers with QPSK modulation and an overlap factor of $\alpha=0.1$. Models were trained on $2 \times 10^8$ synthetic symbols and evaluated on $4 \times 10^6$ symbols.

Results and Discussion

Inductive Bias and Architecture Performance

A primary finding is that CNNs significantly outperform MLPs. The local connectivity and parameter sharing of CNNs provide a beneficial inductive bias that matches the time-series-like structure of SEFDM signals. As shown in Figure 1, standard CNN architectures already provide a substantial improvement over linear baselines like Matching Filters (MF) and Gram-Schmidt (GS) orthogonalized filters.

Performance of a 5-layer CNN architecture compared to linear baselines.

The Role of Residual Connections and Depth

Residual connections proved essential for training deeper models. In SEFDM systems, the Gram matrix can have vanishing eigenvalues, leading to vanishing gradients in standard deep networks. Residual "highways" allow information to bypass these bottlenecks. Furthermore, while increasing width only benefited shallow MLPs, increasing depth consistently improved ResCNN performance.

The Residual 2-CNN architecture with 256 embeddings emerged as the superior model. As illustrated in Figure 2, this architecture nearly closes the performance gap, coming within approximately 1dB of the ideal non-overlapping QPSK baseline at high Signal-to-Noise Ratios (SNR).

Bit Error Rate (BER) performance of the Residual 2-CNN architecture with 256 embeddings, showing near-optimal detection.

Conclusion

This study demonstrates that Residual CNNs are highly effective for signal detection in SEFDM systems. By leveraging appropriate inductive biases and residual learning, these models can navigate complex ICI to achieve high spectral efficiency with manageable detection error rates. Future work will focus on optimizing these models for real-time hardware implementation and exploring their robustness across dynamic overlap factors and higher-order modulation schemes.

Summary updated

Abstract

In this study we present how to approach the problem of building efficient detectors for spectrally efficient frequency division multiplexing (SEFDM) systems. The superiority of residual convolution neural networks (CNNs) for these types of problems is demonstrated through experimentation with many different types of architectures.
BibTeX
@misc{picard2021deep,
  author = {David Picard and Arsenia Chorti},
  title = {{Deep Learning Based Detection for Spectrally Efficient FDM Systems}},
  year = {2021},
  eprint = {2103.11409},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2103.11409},
}
Figure