Joint Uplink Channel Denoising and Feedback-Based Multi-Resolution Precoder Prediction for Cellular TDD MIMO Systems

Accurate downlink (DL) channel state information (CSI) is crucial for effective precoding in massive multiple-input multiple-output (MIMO) systems. Operating in time-division duplexing (TDD) mode enables both uplink (UL) and DL CSI to be utilized for DL precoding, while each alternative presents its own unique drawbacks. UL CSI experiences degradation for cell edge users due to their low signal-to-noise ratio (SNR), while DL CSI feedback suffers from low frequency resolution. Moreover, channel aging in high mobility scenarios necessitates frequent CSI updates. Although channel or precoder prediction is a promising solution, the inherent shortcomings of UL CSI and DL CSI feedback make accurate prediction challenging. In this paper, we propose artificial intelligence (AI)-based precoder prediction solutions that jointly exploit UL CSI and codebook-based precoder matrix indicator (PMI) feedback to address the aforementioned challenges. In our first solution, a PMI-based precoder sequence for a subband (SB) that covers many subcarriers is used as input to an AI model that generates a context vector. This context vector is then combined with multiple precoders obtained from noisy UL CSI, with each precoder corresponding to a different subset of subcarriers within the SB at the last time step, thereby yielding a precoder prediction with high frequency resolution. Although we study a recurrent neural network (RNN)-based approach, our framework can also be implemented with other architectures such as Transformers. Additionally, we introduce a UL CSI denoising stage to enhance performance and propose a multi-resolution prediction network that leverages both PMI and UL CSI-based precoder sequences with arbitrary time-frequency resolution.