Graph Neural Network-Based Low-Complexity Channel Estimation for Massive MU-MIMO Systems

Massive Multiple-Input Multiple-Output (MIMO) technology constitutes a fundamental component of 5G networks. As the system scales with an increasing number of antennas and users, the dimensionality of the channel matrix expands, elevating the computational complexity of channel estimation algorithms as a critical concern. For the single-cell multi-user large-scale MIMO channel estimation problem in time-division duplex (TDD) mode with uniform planar array (UPA) antennas, we proposes a structured computational theoretical approach using the Toeplitz structure of the covariance matrix to greatly reduce the computational complexity of the large-dimensional matrix in the Linear Minimum Mean Squared Error (LMMSE) channel estimation algorithm. As for the estimation of the sample covariance matrix, we proposes a redundant averaging method to estimate the covariance matrix of the channel. However, since the redundant averaging method requires a sufficiently large number of samples to achieve a certain accuracy, a graph neural network channel estimator is proposed to optimize it. Simulation results show that the proposed algorithm can reduce the computational complexity while ensuring good performance.