Distributed DISTA Network with Adaptive Inter-DU Cooperation for Massive MIMO Detection

To alleviate the computational burden of signal detection in uplink massive MIMO systems, the distributed baseband processing architecture has been developed. However, the quality of local channel state information (CSI) and received signal may vary considerably across different distributed units (DUs), and the information exchange among DUs is constrained by different interconnection topologies, making efficient distributed MIMO signal detection challenging. In this paper, we propose a distributed deep--unfolding detector based on the deep iterative shrinkage-thresholding algorithm (DISTA). In each layer of the proposed distributed DISTA network, every DU performs DISTA-based local signal detection in parallel using its own received signal and CSI, and the resulting local detection results are then exchanged and aggregated among DUs. For inter-DU information aggregation, a layer-wise learnable cooperation weight matrix is introduced to combine local detection results from different DUs, with its parameters optimized through end-to-end training under heterogeneous local observation conditions. Meanwhile, a topology mask is incorporated to enforce communication constraints imposed by various DU interconnection topologies. To improve the robustness of the proposed distributed DISTA network to different signal--to--noise ratio (SNR) levels, we further make the empirically determined momentum coefficient learnable and apply a weight-decay regularization to it in the training stage. Simulation results demonstrate that under heterogeneous SNR conditions, the proposed method outperforms the centralized MMSE detector in the low-to-medium SNR range (1–7 dB), while also maintaining robust detection performance with different numbers of DUs and interconnection topologies.