Model-Based Deep Learning for Massive Access in mmWave Cell-Free Massive MIMO System

In this paper, massive access in millimeter wave (mmWave) cell-free massive multiple-input multiple-output (CF-mMIMO) system is investigated. We propose a model-based deep learning algorithm to solve the joint active detection and channel estimation (JADCE) problem in grant-free random access. By exploiting structured sparsity and cluster sparsity, we unfold the vector approximate message propagation (VAMP) algorithm with Bernoulli Gaussian mixed distribution into a network and introduce a parameter estimation module to adapt different active ratios and noise variance based on the expectation maximization (EM) algorithm. The proposed network benefits from the param-eter learning ability of deep learning and the low computation complexity of the model-based method. Simulation results show that the proposed network achieves better detection performance and faster convergence rate than the considered state-of-the-art algorithms.