UADFormer: A Transformer-Based Deep Learning Method for User Activity Detection in Cell-Free Massive MIMO Systems

Grant-free random access is a critical enabling technology for massive ultra-reliable and low-latency communications (mURLLC), and user activity detection (UAD), determining which users are active based on received signals, is indispensable in grant-free access. The cell-free massive multiple-input multiple-output (MIMO) system, providing macro diversity and supporting more users, is an architecture suitable for mURLLC. This paper investigates the UAD problem with changeable pilot sequences in cell-free massive MIMO systems, and proposes deep learning based UADFormer, where users are assigned to access points (APs) for detection using user grouping algorithm and transformer-based neural networks (NNs) in APs are employed to output user activity state by exploiting the correlation between received signals and pilot sequences. For users detected by multiple APs, a fusion NN based on transfer learning in the central processing unit is utilized to enhance UAD accuracy. Additionally, considering limited computational and storage resources in APs, sparse attention mechanism is used in NNs to reduce computational complexity and residual attention score is employed to improve NN performance. Simulation results show that UADFormer outperforms other schemes in various evaluation metrics and reduces computational complexity.