In this paper, we propose Temporal Progressive Refinement - Recurrent Concatenation Network (TPR-RCNet), a temporal progressive refinement architecture for Sounding Reference Signals (SRS) channel estimation in New-Radio (NR) systems. The proposed net...
In this paper, we propose Temporal Progressive Refinement - Recurrent Concatenation Network (TPR-RCNet), a temporal progressive refinement architecture for Sounding Reference Signals (SRS) channel estimation in New-Radio (NR) systems. The proposed network leverages both the previously estimated channel from a Neural Network (NN) and the current channel information to progressively enhance estimation accuracy at the time instance. In addition, the temporal residual term enables the model to learn the difference between consecutive channels, thereby improving its ability to capture temporal variation in the channel. Conventional approaches that exploit temporal properties often rely on multiple-symbol channel inputs, which significantly increase model complexity. For unit-symbol-level estimation, recurrent-based methods are inherently standalone architectures, and their end-to-end sequential nature reduces flexibility for integration with other models. By contrast, the proposed TPR-RCNet offers a flexible framework that can be readily extended with various NN architectures, providing adaptability to different design requirements. In our simulations, the proposed TPR-RCNet outperforms baseline models, including unit-symbol-based and window-size-based channel estimation. Furthermore, the advantage of TPR-RCNet becomes more pronounced as it is applied progressively over consecutive time instances.