Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Novel Deep Neural OFDM Receiver Architectures for LLR Estimation

Erhan Karakoca, Hüseyin Çevik, İbrahim Hökelek, Ali Görçin

TL;DR

This work addresses lifting OFDM receivers into the neural domain by directly estimating soft LLRs from IQ-based OFDM grids. It introduces two architectures, the Dual Attention Transformer (DAT) and the Residual Dual Non-Local Attention Network (RDNLA), which exploit spatio-temporal and channel correlations to improve BER and BLER over traditional systems and prior neural models. Through a BCE-based training objective and comprehensive simulations in a SIMO uplink scenario, the authors show that DAT and RDNLA achieve superior BER/BLER performance, with mean inference times in the 11–13 ms range, while conventional LS estimation remains faster. The study highlights practical considerations like decoding error dispersion and proposes future directions for error-aware loss and decoder-in-the-loop feedback to further enhance block-level reliability in neural receivers.

Abstract

Neural receivers have recently become a popular topic, where the received signals can be directly decoded by data driven mechanisms such as machine learning and deep learning. In this paper, we propose two novel neural network based orthogonal frequency division multiplexing (OFDM) receivers performing channel estimation and equalization tasks and directly predicting log likelihood ratios (LLRs) from the received in phase and quadrature phase (IQ) signals. The first network, the Dual Attention Transformer (DAT), employs a state of the art (SOTA) transformer architecture with an attention mechanism. The second network, the Residual Dual Non Local Attention Network (RDNLA), utilizes a parallel residual architecture with a non local attention block. The bit error rate (BER) and block error rate (BLER) performance of various SOTA neural receiver architectures is compared with our proposed methods across different signal to noise ratio (SNR) levels. The simulation results show that DAT and RDNLA outperform both traditional communication systems and existing neural receiver models.

Novel Deep Neural OFDM Receiver Architectures for LLR Estimation

TL;DR

This work addresses lifting OFDM receivers into the neural domain by directly estimating soft LLRs from IQ-based OFDM grids. It introduces two architectures, the Dual Attention Transformer (DAT) and the Residual Dual Non-Local Attention Network (RDNLA), which exploit spatio-temporal and channel correlations to improve BER and BLER over traditional systems and prior neural models. Through a BCE-based training objective and comprehensive simulations in a SIMO uplink scenario, the authors show that DAT and RDNLA achieve superior BER/BLER performance, with mean inference times in the 11–13 ms range, while conventional LS estimation remains faster. The study highlights practical considerations like decoding error dispersion and proposes future directions for error-aware loss and decoder-in-the-loop feedback to further enhance block-level reliability in neural receivers.

Abstract

Neural receivers have recently become a popular topic, where the received signals can be directly decoded by data driven mechanisms such as machine learning and deep learning. In this paper, we propose two novel neural network based orthogonal frequency division multiplexing (OFDM) receivers performing channel estimation and equalization tasks and directly predicting log likelihood ratios (LLRs) from the received in phase and quadrature phase (IQ) signals. The first network, the Dual Attention Transformer (DAT), employs a state of the art (SOTA) transformer architecture with an attention mechanism. The second network, the Residual Dual Non Local Attention Network (RDNLA), utilizes a parallel residual architecture with a non local attention block. The bit error rate (BER) and block error rate (BLER) performance of various SOTA neural receiver architectures is compared with our proposed methods across different signal to noise ratio (SNR) levels. The simulation results show that DAT and RDNLA outperform both traditional communication systems and existing neural receiver models.

Paper Structure

This paper contains 13 sections, 35 equations, 3 figures, 1 table.

Table of Contents

  1. Introduction
  2. System Model and Background
  3. Proposed DL-Based LLR Estimation Methods
  4. Input Preprocessing and Embedding
  5. Dual Attention Transformer
  6. Multi-Head Self Attention
  7. Dual-Attention Mechanism
  8. Residual Dual Non-Local Attention Network
  9. Parallel Residual Block
  10. Dual Non‑Local Attention Block
  11. Model Training
  12. Simulation Experiments and Results
  13. Discussion and Future Work

Figures (3)

  • Figure 1: System Model of conventional and neural network based receiver.
  • Figure 2: The DAT (left) and RDNLA (right) LLR estimation architectures for neural receiver.
  • Figure 3: BER (a) and BLER (b) of the proposed neural receivers with SOTA models at varying noise levels.