Joint User Activity and Data Detection in Grant-Free NOMA using Generative Neural Networks

Yixuan Zou, Zhijin Qin, Yuanwei LiuICC 2021 · DOI · arXiv · OpenReview

TL;DR

This paper introduces GenMUD, a novel generative neural network-based approach for efficient multi-user detection (MUD) in grant-free NOMA systems. By leveraging generative models and a specialized network architecture, GenMUD achieves superior detection accuracy and speed compared to existing methods, while also proposing a practical sparsity estimator for improved system performance.

What Problem Does This Paper Address?

The rapid growth of the Internet of Things (IoT) necessitates massive connectivity, leading to the exploration of grant-free Non-Orthogonal Multiple Access (NOMA) schemes. In these systems, a key challenge is efficiently detecting which users are active and decoding their data, especially when the number of users (K) significantly exceeds the available resources (M). Traditional methods like compressed sensing (CS) rely on iterative algorithms that can be computationally expensive and slow, hindering real-time applications. Existing deep learning approaches often require multiple models for different system configurations, lacking flexibility. Furthermore, accurately estimating user activity (sparsity) without prior knowledge is crucial but difficult. This paper tackles these issues by proposing a single, flexible generative neural network model for MUD that is both accurate and fast, and by developing a practical sparsity estimation method.

Methods

The core of this paper is the GenMUD (Generative Multi-User Detection) framework. It utilizes a generative neural network (G) trained using Model-Agnostic Meta-Learning (MAML) [12]. Instead of directly solving the complex MUD optimization problem, GenMUD learns a mapping from a low-dimensional latent space (z) to the space of transmitted signals (X) via the generator: X = G(z). The MUD problem is then reformulated as finding the latent variable z that minimizes the reconstruction error: ||Y - HG(z)||^2.

Key methodological innovations include:

  • Generative Network Architecture: A custom, low-complexity network is designed. It uses 1x1 convolutional layers instead of computationally expensive fully-connected layers, and avoids large kernel sizes that are less suitable for the uncorrelated nature of communication data. This architecture is designed to handle varying input dimensions and user activity levels with a single trained model.
  • MAML for Fast Recovery: The network is trained using MAML, enabling it to recover signals in a minimal number of gradient descent steps (as few as three), significantly speeding up the detection process.
  • Sparsity Estimator: A closed-form, low-complexity sparsity estimator is derived. This estimator, based on the received signal power and SNR, allows for sparsity-blind MUD, removing the need for prior knowledge of user activity.

Datasets

The paper does not explicitly mention specific named datasets. Instead, it describes the generation of simulated data for training and evaluation. The training dataset is characterized by:

  • Number of users (K): 200
  • Number of active users (S): 40
  • Time frame length (J): 7
  • Signal-to-Noise Ratio (SNR): 20 dB
  • Number of subcarriers (M): 100
  • Modulation: Quadrature Phase Shift Keying (QPSK)
  • Channel model: Rayleigh fading

The evaluation is performed using simulated data across various SNRs and user sparsity levels, comparing the proposed method against established algorithms.

Key Insights and Ideas

  • Generative Models for MUD: The paper demonstrates that generative networks, typically used for image synthesis, can be effectively adapted for signal reconstruction tasks like MUD, offering flexibility and speed.
  • Low-Complexity Architecture is Key: The custom 1x1 convolutional architecture is crucial for scaling generative models to the high-dimensional inputs of communication systems without prohibitive computational costs.
  • MAML for Rapid Adaptation: Applying MAML significantly accelerates the recovery process, making generative models practical for latency-sensitive applications.
  • Practical Sparsity Estimation: The proposed closed-form sparsity estimator is a significant contribution, enabling sparsity-blind MUD without complex iterative approximations, making it broadly applicable.
  • Robustness to Noise and Sparsity: The GenMUD framework shows remarkable performance even under low SNR and varying sparsity, indicating its robustness.

Key Results

  • Superior SER: GenMUD significantly outperforms OMP, DyCS, and BPDN in terms of Symbol Error Rate (SER) across a wide range of SNRs. It achieves near-oracle performance at low SNRs.
  • High Detection Probability (Pd): GenMUD achieves the highest detection probability among all compared methods (except for the oracle LS), reaching nearly 100% at SNRs above 10 dB.
  • Low False Alarm Probability (Pfa): The proposed method exhibits the lowest false alarm probability, approaching zero as SNR increases, indicating accurate identification of active users.
  • Accurate Sparsity Estimation: The proposed sparsity estimator demonstrates high accuracy (low normalized error, En) and robustness to varying SNR and frame lengths.
  • Negligible Impact of Sparsity Estimation: When using the estimated sparsity, GenMUD's performance (Pd and Pfa) is nearly identical to using known sparsity, highlighting the effectiveness of the estimator.

Limitations and Future Work

  • Computational Cost of Latent Updates: While MAML speeds up recovery, the latent variable optimization still involves gradient descent steps, which can be computationally intensive compared to purely feed-forward networks.
  • Training Data Dependency: Like all deep learning models, GenMUD's performance is dependent on the training data. While the paper shows robustness, performance might degrade in scenarios significantly different from the training conditions.
  • Future Work: The paper suggests exploring further improvements in the generative network architecture and training strategies to enhance performance. Additionally, extending the framework to more complex communication scenarios, such as systems with multiple antennas or different modulation schemes, would be valuable. The sparsity estimator's applicability to other MUD algorithms is also a promising avenue.

Summary updated

Abstract

Grant-free non-orthogonal multiple access (NOMA) is considered as one of the supporting technology for massive connectivity for future networks. In the grant-free NOMA systems with a massive number of users, user activity detection is of great importance. Existing multi-user detection (MUD) techniques rely on complicated update steps which may cause latency in signal detection. In this paper, we propose a generative neural network-based MUD (GenMUD) framework to utilize low-complexity neural networks, which are trained to reconstruct signals in a small fixed number of steps. By exploiting the uncorrelated user behaviours, we design a network architecture to achieve higher recovery accuracy with a low computational cost. Experimental results show significant performance gains in detection accuracy compared to conventional solutions under different channel conditions and user sparsity levels. We also provide a sparsity estimator through extensive experiments. Simulation results of the sparsity estimator showed high estimation accuracy, strong robustness to channel variations and neglectable impact on support detection accuracy.
BibTeX
@inproceedings{zou2021joint,
  author = {Yixuan Zou and Zhijin Qin and Yuanwei Liu},
  title = {{Joint User Activity and Data Detection in Grant-Free NOMA using Generative Neural Networks}},
  booktitle = {ICC},
  year = {2021},
  doi = {10.48550/arxiv.2101.02324},
}
Figure