Design of an Efficient CSI Feedback Mechanism in Massive MIMO Systems: A Machine Learning Approach Using Empirical Data

Abstract

Massive MIMO regime reaps the benefits of spatial diversity and multiplexing gains, subject to precise channel state information (CSI) acquisition. In the current communication architecture, the downlink CSI is estimated by the user equipment (UE) via dedicated pilots and then fed back to the gNodeB (gNB). The feedback information is compressed to reduce overhead. This compression increases the inaccuracy of acquired CSI; thus, degrading precoding quality. Though various autoencoder-based CSI feedback mechanisms have been proposed in recent studies, their computational complexity is enormous. Motivated by these issues, this paper proposes a machine learning-based CSI feedback prediction network, CsiFB-PNet, which exploits twin channel predictors. The feedback at the UE is evaluated with respect to the predicted channel. Further, to enhance the performance, the UE trains the model and reports it to the gNB. CsiFB-PNet can work for both time-division and frequency-division duplex systems, reducing feedback overhead and effectively recovering the compressed CSI. We demonstrate the performance of CsiFB-PNet in a multi-carrier system using the empirical data recorded at the Nokia campus. Furthermore, we use a clustered delay line channel model of the 3GPP 5G new radio standard protocol to make a fair comparison with the benchmark scheme. Numerical results and the complexity analysis verify that the CsiFB-PNet outperforms existing autoencoder-based technique. In particular, CsiFB-PNet precisely recovers the compressed CSI at a reduced overhead cost.