Enhanced Deep Learning for Massive MIMO Detection Using Approximate Matrix Inversion

Massive multiple-input multiple-output (MIMO) is a crucial technology in fifth-generation (5G) and beyond 5G (B5G). However, the huge number of antennas used in massive MIMO systems causes a high computational complexity during signal detection. In this paper, we propose an efficient massive MIMO detection technique which is based on approximate matrix inversion methods and deep learning to enhance the system performance while keeping computational complexity low. Three approximate methods which are Gauss-Seidel (GS), successive over-relaxation (SOR), and conjugate gradient (CG) are exploited for the initialization of a modified version of the MM network (MMNet) algorithm. The performance of the proposed technique is validated under both Gaussian and realistic channel scenarios, i.e., Quadriga channels models. Simulation results show that the proposed technique outperforms MMNet, minimum mean square estimation (MMSE), detection network (DetNet), and orthogonal approximate message passing deep net (OAMP-Net) in terms of symbol error rate (SER) during offline training. It also provides a significant SER improvement of up to 87% when compared to MMNet in the online training scenario.