Cooperative Multi-Cell Massive Access With Temporally Correlated Activity

Xiaojun Yuan, Weifeng Zhu, Meixia Tao, Xiaojun Yuan 0002, Fan Xu, Yunfeng GuanIEEE Trans. Wireless Communications 2024 · DOI

TL;DR

This paper tackles the challenge of efficiently detecting and estimating user activity and channels in massive multi-cell networks, where user behavior is predictable over time. They introduce a novel algorithm, DCS-MMV-GAMP, that leverages both spatial cooperation between access points and temporal correlations in user activity to achieve superior performance. The work also proposes practical solutions for limited fronthaul capacity, offering insights into the trade-offs between different approaches.

What Problem Does This Paper Address?

The proliferation of Internet of Things (IoT) devices necessitates massive access capabilities in future cellular networks (like 6G). Traditional random access methods are too inefficient for this scale. Grant-free (GF) access offers a promising solution by allowing devices to transmit without explicit permission, but it faces a major hurdle: distinguishing active users from inactive ones (activity detection, AD) and estimating their communication channels (channel estimation, CE), especially when many users share non-orthogonal pilot signals. This problem is further complicated in multi-cell environments due to inter-cell interference and the practical limitations of fronthaul links connecting base stations to a central unit. Existing solutions often focus on single cells or ignore the temporal correlation of user activity, which is a significant factor as devices tend to remain active for extended periods. This paper aims to bridge this gap by developing a cooperative multi-cell AD and CE framework that effectively utilizes both spatial cooperation and temporal activity patterns.

Methods

The core of this paper is the DCS-MMV-GAMP (Dynamic Compressed Sensing-based Multiple Measurement Vector Generalized Approximate Message Passing) algorithm. Here's a breakdown of the key methodological contributions:

  • User-Centric AP Cooperation: Instead of involving all access points (APs) for every user, the system intelligently selects a subset of APs that are geographically close to a user to perform AD and CE. This significantly reduces computational complexity and communication overhead.
  • Generalized Sliding-Window Detection: This strategy allows the algorithm to process user activity over a window of consecutive frames, rather than frame-by-frame. This is crucial for exploiting the temporal correlation in user behavior. It encompasses existing frame-by-frame and block-based detection methods as special cases.
  • Bayesian Inference and Factor Graphs: The problem is framed as a Bayesian inference task. A factor graph is constructed to represent the statistical dependencies between received signals, user activities, and channels.
  • DCS-MMV-GAMP Algorithm: This algorithm combines the power of Generalized Approximate Message Passing (GAMP) for efficient channel estimation with message passing (MP) for activity refinement. It iteratively updates activity probabilities and channel estimates, leveraging both spatial (across APs) and temporal (across frames) information.
  • Finite Fronthaul Capacity Schemes (QF and DF): To address the practical constraint of limited bandwidth between APs and the central unit (CU), two schemes are proposed:
  • Quantize-and-Forward (QF): APs quantize their received signals and send them to the CU for centralized processing.
  • Detect-and-Forward (DF): APs perform local AD and then forward the detected activity information (in terms of log-likelihood ratios) to the CU.

Datasets

Not discussed in the paper. The evaluation is based on simulated data with specified parameters.

Key Insights and Ideas

  • Temporal Correlation is Key: The paper strongly emphasizes that exploiting the predictable nature of user activity over time is crucial for improving AD and CE performance, especially in massive access scenarios.
  • User-Centric Cooperation is Efficient: By intelligently selecting APs for each user, significant computational savings are achieved without sacrificing performance, making the system scalable.
  • Sliding Window for Temporal Exploitation: The generalized sliding window is a flexible and effective way to capture temporal dependencies, outperforming simpler frame-by-frame approaches.
  • QF vs. DF Trade-offs: The choice between QF and DF for limited fronthaul capacity depends critically on the number of antennas at the APs. QF is better with few antennas, while DF becomes superior with large-scale antenna arrays.
  • DCS-MMV-GAMP Balances Performance and Complexity: The proposed algorithm offers a good balance, achieving near-optimal performance while being computationally feasible for large-scale systems.

Key Results

  • The proposed DCS-MMV-GAMP algorithm significantly outperforms existing benchmarks (SI-aided MMV-AMP, CS-MMV-GAMP) in terms of both activity detection (lower error detection ratio) and channel estimation (lower normalized mean square error).
  • The performance gains are particularly pronounced when user activity is temporally correlated (higher transition probabilities).
  • The generalized sliding-window strategy shows improved performance compared to frame-by-frame methods, especially for the initial frames.
  • For finite fronthaul capacity:
  • QF-DCS can achieve near-optimal performance with sufficient capacity, even with a small number of antennas (M=2).
  • DF-DCS demonstrates superior performance over QF-DCS when the fronthaul capacity is limited and APs are equipped with a large number of antennas (M=10). This is because DF's performance is less sensitive to quantization resolution.

Limitations and Future Work

  • Assumed Knowledge: The paper assumes that parameters like transition probabilities (α, β, pa) and noise variance are known or can be accurately estimated.
  • Idealized Fronthaul: While QF and DF address finite capacity, the initial framework assumes infinite capacity for the core DCS-MMV-GAMP algorithm.
  • Uniform Quantization: The QF and DF schemes use simple uniform quantization. More advanced quantization techniques could further improve performance.
  • Future Work: The paper suggests that advanced lossy compression techniques (like sparse coding or deep learning) could be used in DF to further enhance compression efficiency. The theoretical analysis of convergence using state evolution is also highlighted as a valuable tool for system design.

Summary updated

BibTeX
@article{yuan2024cooperative,
  author = {Xiaojun Yuan and Weifeng Zhu and Meixia Tao and Xiaojun Yuan and Fan Xu and Yunfeng Guan},
  title = {{Cooperative Multi-Cell Massive Access With Temporally Correlated Activity}},
  journal = {IEEE Trans. Wireless Communications},
  year = {2024},
  doi = {10.1109/twc.2023.3317265},
}
Figure