AI Receiver Design with Deep Learning Based Channel Estimation and MIMO Detection

Deep learning (DL) and artificial intelligence (AI) based channel estimation (CE) and multi-input and multi-output (MIMO) detection (MIMODet), have gained emerging interests in communication systems, aiming for leveraging the advantages of AI. However, it is a non-trivial task for AI based CE or MIMODet to outperform classical algorithms, especially when those are close to optimal. For instance, the conventional linear minimum mean-square-error (LMMSE) based CE, and QRdecomposition and M-algorithm (QRM) based MIMO detection, respectively. In this paper, we propose a novel AI receiver design that connects CE and MIMODet in a unified neural network (NN) architecture and adopts structural input features and offline-trainings. The CE utilizes a super-resolution based convolutional NN (SRCNN), and the MIMODet applies a QRMNet based on QRM detector enhanced graphical NN (GNN). Numerical results show that our AI receiver can achieve accurate CE and near-optimal block error rate (BLER) performance in tests of the fifth-generation new-radio ($5 \mathrm{G}-\mathrm{NR}$) system.