Transformer-Driven Robust Recovery of Massive MIMO CSI Feedback With Temporal Information

To mitigate the growing overhead resulting from the increasing number of antennas in massive multiple-input multiple-output (MIMO) systems, various autoencoder-based schemes employing two-sided artificial intelligence (AI) models for channel state information (CSI) feedback-specifically, feedback of the precoding matrix-have been proposed. In these twosided AI models, the encoder and decoder are separately deployed at the user equipment (UE) and the base station (BS), posing significant challenges in maintaining synchronization between them in real-world scenarios with multiple UEs and BSs. In this paper, we propose a novel transformer-based temporaldomain multi-length optimized (TMO) decoder that effectively leverages temporal correlations in MIMO channels to enhance CSI reconstruction. Unlike existing schemes that require different models or additional signaling overhead for re-synchronization between the encoder and decoder under varying conditions, the transformer-based TMO decoder can robustly recover CSI by utilizing feedback of the precoding matrix with temporal information using a single fixed model for inputs of varying sequential time steps. Performance evaluations demonstrate that the proposed transformer-based TMO decoder achieves significantly higher squared generalized cosine similarity and average downlink throughput compared to state-of-the-art techniques.