Interference cancellation assisted enhanced sparsely connected neural network for signal detection in massive MIMO systems

Abstract

In recent years, deep learning (DL) has become one of the potential solutions for massive MIMO signal detection. Considering that eliminating interference among the receive antennas at the base-station is intrinsically critical, we propose a method that combines DL and interference cancellation (IC) algorithms for uplink signal detection in massive MIMO systems. Firstly, by optimizing the conventional detection network (DetNet) and the sparsely connected neural network (ScNet) detection algorithms, we propose an enhanced version of ScNet, named EScNet, based on the convolutional neural networks (CNN). Secondly, an IC mechanism is employed, and its corresponding DNN layer structure is designed accordingly. Specifically, parallel and successive interference cancellation-aided EScNet algorithms, namely EScNet-PIC and EScNet-SIC, are proposed, respectively. The proposed algorithms are implemented with two stages on each DNN layer, where the first stage accounts for the proposed EScNet algorithm, which demodulates the received symbols as the input to the second stage for interference cancellation. Simulation results verify that our proposed EScNet-PIC and EScNet-SIC algorithms are particularly salient for massive MIMO signal detection compared to various existing algorithms, and they achieve an SNR gain of at least 0.5 dB at the BER level of $10^{-3}$ and up to 4dB for various antenna configurations. Moreover, the proposed algorithms also exhibit fast and stable convergence and relatively low complexity. With the capability of operating in both independent and correlated fading channel environments, they can serve as promising technical candidates for massive MIMO signal detection.