[Paper Review] Blind Coherent Preamble Detection via Neural Networks

Jafar Mohammadi, Gerhard Schreiber+2 more|arXiv (Cornell University)|Sep 30, 2021

Blind Source Separation Techniques18 references4 citations

TL;DR

This paper proposes a neural network-based blind coherent preamble detection scheme that enhances signal-to-noise ratio by learning optimal coherent combining of signals across multiple antennas and time instances, leveraging Zadoff-Chu sequence properties and Kronecker-structured channel covariance models. The method, named HyNE, outperforms conventional matched filtering, especially in low SNR and high-mobility scenarios, without requiring explicit channel state knowledge or SNR-specific model retraining.

ABSTRACT

In wireless communications systems, the user equipment (UE) transmits a random access preamble sequence to the base station (BS) to be detected and synchronized. In standardized cellular communications systems Zadoff-Chu sequences has been proposed due to their constant amplitude zero autocorrelation (CAZAC) properties. The conventional approach is to use matched filters to detect the sequence. Sequences arrived from different antennas and time instances are summed up to reduce the noise variance. Since the knowledge of the channel is unknown at this stage, a coherent combining scheme would be very difficult to implement. In this work, we leverage the system design knowledge and propose a neural network (NN) sequence detector and timing advanced estimator. We do not replace the whole process of preamble detection by a NN. Instead, we propose to use NN only for extit{blind} coherent combining of the signals in the detector to compensate for the channel effect, thus maximize the signal to noise ratio. We have further reduced the problem's complexity using Kronecker approximation model for channel covariance matrices, thereby, reducing the size of required NN. The analysis on timing advanced estimation and sequences detection has been performed and compared with the matched filter baseline.

Motivation & Objective

• To address the challenge of preamble detection in low-SNR and high-mobility scenarios where conventional matched filtering fails due to unknown channel effects.
• To enable blind coherent combining of signals across multiple receive antennas and time instances without requiring explicit channel state information.
• To reduce computational complexity by exploiting Kronecker structure in channel covariance matrices for neural network design.
• To improve timing advance estimation accuracy in random access procedures, especially below -20 dB SNR.
• To develop a single, robust neural network model that generalizes across diverse channel conditions and SNRs, avoiding per-SNR model retraining.

Proposed method

• The method uses a neural network (HyNE) to learn optimal coherent combining weights for signals received across multiple antennas and time instances, maximizing SNR without prior channel estimation.
• The network is trained using a loss function derived from the MMSE estimator, enabling it to approximate optimal linear combining in a data-driven manner.
• A Kronecker approximation model is applied to the channel covariance matrix, reducing the number of trainable parameters and enabling efficient generalization across diverse channel conditions.
• The approach reformulates preamble detection as a channel estimation-like problem, allowing transfer of knowledge from well-studied channel estimation techniques to coherent combining.
• The network is designed to process multi-dimensional inputs (frequency, spatial, temporal), with theoretical analysis showing that frequency-domain correlation is negligible due to Zadoff-Chu CAZAC properties.
• The method preserves the conventional matched filter structure but replaces only the coherent combining stage with a learned neural network, ensuring compatibility with existing systems.

Experimental results

Research questions

• RQ1 Can a neural network effectively learn blind coherent combining of random access preamble signals across multiple antennas and time instances without prior channel state information?
• RQ2 How does the performance of the proposed neural network-based coherent combining (HyNE) compare to conventional matched filtering in terms of detection and timing estimation accuracy across varying SNR levels?
• RQ3 To what extent does exploiting Kronecker structure in channel covariance matrices reduce the complexity and improve generalization of the neural network model?
• RQ4 Does the proposed method maintain robustness across diverse channel models and SNR conditions without requiring per-SNR model retraining?
• RQ5 Why does the frequency domain contribute minimally to performance gain, and how does this align with the CAZAC properties of Zadoff-Chu sequences?

Key findings

• HyNE significantly outperforms conventional matched filtering in low-SNR regimes, particularly below -20 dB, where detection and timing estimation performance improves substantially.
• The Kronecker approximation model reduces the number of trainable parameters in the neural network, enabling efficient training and deployment while maintaining high performance.
• The method achieves robust performance across a wide range of SNRs and channel models without requiring separate models for different SNR conditions, unlike prior approaches.
• Theoretical analysis confirms that frequency-domain components contribute little to performance gain due to the zero cross-correlation properties of Zadoff-Chu sequences, validating the focus on spatial and temporal domains.
• Timing advance estimation shows notable improvement only at extremely low SNRs (< -20 dB), indicating that the main benefit of HyNE lies in preamble detection rather than timing estimation.
• The neural network-based coherent combining achieves performance close to the theoretical MMSE bound, demonstrating effective learning of optimal signal combining in the absence of channel state information.