An Autoregressive Model-Based Differential Framework With Learnable Regularization for CSI Feedback in Time-Varying Massive MIMO Systems

In frequency division duplex (FDD) mode, the substantial feedback overhead in massive multi-input multi-output (MIMO) systems needs to be mitigated. Existing channel feedback methods that utilize channel temporal correlation exhibit limited performance under low compression ratios (CRs) or high-speed user equipment (UE) in the outdoor scenario. To address these challenges, we propose an autoregressive (AR) model-based differential framework incorporating a regularization learning network (RE-LENet) for channel state information (CSI) feedback in time-varying massive MIMO systems. The proposed AR model-based differential framework can capture the channel temporal correlation more effectively, reducing the degradation of channel reconstruction performance over time. We also design a convolutional neural network (CNN)-based RE-LENet to enhance the reconstruction performance of both the channel differential terms and the initial channel simultaneously. Numerical results indicate that the proposed CSI feedback framework outperforms existing methods.