Robust Learning-Based Sparse Recovery for Device Activity Detection in Grant-Free Random Access Cell-Free Massive MIMO: Enhancing Resilience to Impairments

TL;DR

This work tackles device activity detection in grant-free random access for cell-free MIMO, where non-orthogonal pilots and impairments challenge reliable detection. It introduces a centralized deep multilayer perceptron (DMLP) detector that uses top-$T$ AP signals per device, processes them through multiple hidden layers, and outputs per-device activity with a threshold, trained via binary cross-entropy. Across simulations, the DMLP shows strong robustness to imperfect CSI and finite ADC resolution, outperforming traditional ML detectors under realistic conditions and demonstrating practical viability for dense mMTC in CF-mMIMO. The study highlights the importance of detector design that decouples clustering from inference and remains resilient to hardware and channel uncertainties in next-generation wireless networks.

Abstract

Massive MIMO is considered a key enabler to support massive machine-type communication (mMTC). While massive access schemes have been extensively analyzed for co-located massive MIMO arrays, this paper explores activity detection in grant-free random access for mMTC within the context of cell-free massive MIMO systems, employing distributed antenna arrays. This sparse support recovery of device activity status is performed by a finite cluster of access points (APs) from a large number of geographically distributed APs collaborating to serve a larger number of devices. Active devices transmit non-orthogonal pilot sequences to APs, which forward the received signals to a central processing unit (CPU) for collaborative activity detection. This paper proposes a simple and efficient data-driven algorithm tailored for device activity detection, implemented centrally at the CPU. Furthermore, the study assesses the algorithm's robustness to input perturbations and examines the effects of adopting fixed-point representation on its performance.

Figures (6)

• Figure 1: Distribution of devices and access points within the network.
• Figure 2: Architecture of the DMLP algorithm proposed for device activity detection.
• Figure 3: ROC curve analysis for different pilot sequence lengths $L$.
• Figure 4: ROC curve analysis for different sparsity degrees $\epsilon$.
• Figure 5: ROC performance variation of both device activity detectors in presence of imperfect channel state information.