Abstract:
Millimeter wave (mmWave) massive multiple input multiple output (mMIMO) systems are at the core of fifth-generation (5G) and beyond-fifth-generation (B5G) technologies due to their ability to support error-free communication even at low signal-to-noise power ratio (SNR). In a multi-user system, mMIMO, along with robust beamforming methods, can serve with very low outage probability. However, in order to perform such precoding, the transmitter must know the channel state information (CSI) estimated by the receiver. This calls for low-latency, lowcomplexity and error-free channel estimation. Initial deployment of these systems focused on statistical estimation methods: least squares (LS), and minimum mean-squared error estimation (MMSE). The former is computationally less complex but suffers from high estimation error. The latter, on the other hand, has almost error-free estimation but requires second-order noise statistics, and is computationally very expensive. Recent literature has introduced deep learning for this task, which promises to be a low-error and low-latency scheme. In this work, we propose an adaptive convolution neural network (CNN) based approach for estimating the CSI at the receiver, followed by orthogonal matching pursuit (OMP) for hybrid beamforming. Based on the SNR, the receiver will deploy the suitable model. We test our approach at very low SNR conditions, and compare the performance with LS and MMSE estimation
