Continual Learning-Aided Super-Resolution Scheme for Channel Reconstruction and Generalization in OFDM Systems

TL;DR

This work tackles OFDM channel estimation under limited pilots and shifting channel distributions by formulating channel reconstruction as a time-frequency super-resolution problem. It introduces a dual-attention SRNN (DA-SRNN) to exploit two channel correlations for accurate reconstruction from pilot data, and a continual-learning (CL) framework with Elastic Weight Consolidation (EWC) to maintain performance across distributions. The paper defines the SR objective $\mathbf{H}^{(d)} = \mathcal{F}_{SR}(\Theta; \hat{\mathbf{H}}_p^{(d)})$ and the CL objective $\mathcal{L}_{total} = \mathcal{L}_h + \alpha \mathcal{L}_{EWC}$, and demonstrates superior reconstruction and generalization on 3GPP TD-A/D models compared to baselines. Practically, the proposed CL-DA-SRNN enables robust, pilot-efficient OFDM channel estimation in non-stationary wireless environments.

Abstract

Channel reconstruction and generalization capability are of equal importance for developing channel estimation schemes within deep learning (DL) framework. In this paper, we exploit a novel DL-based scheme for efficient OFDM channel estimation where the neural networks for channel reconstruction and generalization are respectively designed. For the former, we propose a dual-attention-aided super-resolution neural network (DA-SRNN) to map the channels at pilot positions to the whole time-frequency channels. Specifically, the channel-spatial attention mechanism is first introduced to sequentially infer attention maps along two separate dimensions corresponding to two types of underlying channel correlations, and then the lightweight SR module is developed for efficient channel reconstruction. For the latter, we introduce continual learning (CL)-aided training strategies to make the neural network adapt to different channel distributions. Specifically, the elastic weight consolidation (EWC) is introduced as the regularization term in regard to loss function of channel reconstruction, which can constrain the direction and space of updating the important weights of neural networks among different channel distributions. Meanwhile, the corresponding training process is provided in detail. By evaluating under 3rd Generation Partnership Project (3GPP) channel models, numerical results verify the superiority of the proposed channel estimation scheme with significantly improved channel reconstruction and generalization performance over counterparts.

Figures (5)

• Figure 1: Illustration of time-frequency grids of OFDM subcarriers. (a) whole channel resources; (b) pilot resources.
• Figure 2: Basic blocks of the dual-attention-aided super-resolution architectures. $(c, h, w)$ denotes the dimension of feature map with $c$ channel, $h$ height and $w$ width. Max Pool and Avg Pool denote the maximum pooling operation and average pooling operation, respectively. Concat denotes the matrix concatenation operation. Conv2D denotes the 2D convolution operation.
• Figure 3: Illustration of structure diagram of CL-DA-SRNN scheme with considering two channel distributions.
• Figure 4: NMSE of channel reconstruction versus SNR. (a) TDL-A model; (b) TDL-D model.
• Figure 5: NMSE of channel generalization versus SNR.