Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Learning Beamforming in Cell-Free Massive MIMO ISAC Systems

Umut Demirhan, Ahmed Alkhateeb

TL;DR

This work tackles joint beamforming for ISAC in cell-free massive MIMO, where many distributed APs create complex coupling between communication and sensing. It introduces a heterogeneous GNN that models APs, UEs, and sensing targets as node types with two edge types to capture communication and sensing channels, enabling scalable beamforming predictions without full retraining when network size changes. The method uses unsupervised training with edge embeddings initialized from channel data and a power-normalized final output to maximize the objective $\sum_{u} R_u + \beta_s \log_2(1 + {\rm SNR}^{(s)})$, achieving near-optimal communication performance and favorable sensing trade-offs. Results demonstrate strong generalization to different numbers of APs and robustness to network changes, highlighting the approach as a practical, scalable ML-based solution for real-time JSC beamforming in cell-free ISAC systems.

Abstract

Beamforming design is critical for the efficient operation of integrated sensing and communication (ISAC) MIMO systems. ISAC beamforming design in cell-free massive MIMO systems, compared to colocated MIMO systems, is more challenging due to the additional complexity of the distributed large number of access points (APs). To address this problem, this paper first shows that graph neural networks (GNNs) are a suitable machine learning framework. Then, it develops a novel heterogeneous GNN model inspired by the specific characteristics of the cell-free ISAC MIMO systems. This model enables the low-complexity scaling of the cell-free ISAC system and does not require full retraining when additional APs are added or removed. Our results show that the proposed architecture can achieve near-optimal performance, and applies well to various network structures.

Learning Beamforming in Cell-Free Massive MIMO ISAC Systems

TL;DR

This work tackles joint beamforming for ISAC in cell-free massive MIMO, where many distributed APs create complex coupling between communication and sensing. It introduces a heterogeneous GNN that models APs, UEs, and sensing targets as node types with two edge types to capture communication and sensing channels, enabling scalable beamforming predictions without full retraining when network size changes. The method uses unsupervised training with edge embeddings initialized from channel data and a power-normalized final output to maximize the objective , achieving near-optimal communication performance and favorable sensing trade-offs. Results demonstrate strong generalization to different numbers of APs and robustness to network changes, highlighting the approach as a practical, scalable ML-based solution for real-time JSC beamforming in cell-free ISAC systems.

Abstract

Beamforming design is critical for the efficient operation of integrated sensing and communication (ISAC) MIMO systems. ISAC beamforming design in cell-free massive MIMO systems, compared to colocated MIMO systems, is more challenging due to the additional complexity of the distributed large number of access points (APs). To address this problem, this paper first shows that graph neural networks (GNNs) are a suitable machine learning framework. Then, it develops a novel heterogeneous GNN model inspired by the specific characteristics of the cell-free ISAC MIMO systems. This model enables the low-complexity scaling of the cell-free ISAC system and does not require full retraining when additional APs are added or removed. Our results show that the proposed architecture can achieve near-optimal performance, and applies well to various network structures.
Paper Structure (12 sections, 14 equations, 4 figures)

This paper contains 12 sections, 14 equations, 4 figures.

Table of Contents

  1. Introduction
  2. System Model
  3. Signal Model
  4. Communication Model
  5. Sensing Model
  6. Problem Definition
  7. Heterogenous Graph Neural Networks for Cell-free ISAC MIMO Systems
  8. Graph Definition
  9. Graph Neural Networks
  10. GNN Design for Cell-free ISAC MIMO Beamforming
  11. Simulations
  12. Conclusion

Figures (4)

  • Figure 1: The APs transmit signals with beamforming to jointly serve the UEs and sense the target. The signals reflected from the target are collected back at the APs for multistatic sensing.
  • Figure 2: The heterogeneous graph structure for the cell-free ISAC MIMO system is defined with three types of nodes (AP, UE, ST) and two types of edges (AP-UE and AP-ST).
  • Figure 3: The performance of the neural networks trained with $U=2$ and $M=5$ for different sensing weight values.
  • Figure 4: The proposed solution is trained with $U=2$ UEs and $M=5$ APs, and tested with different number of APs.