Alternative Meta-Learning With 3D Dual-CNN for MIMO Channel Estimation

This study introduces a novel 3D dual Convolutional Neural Network (CNN) architecture with meta-learning enhancements for effective MIMO channel estimation in dynamic and noisy environments. Traditional 2D CNN-based methods often neglect the temporal correlations and Doppler effect in channel characteristics, leading to suboptimal performance. Addressing this, the proposed model incorporates spatial-frequency-time (SFT) and angle-delay-Doppler (ADD) domain estimators within a 3D framework, leveraging sparse feature representation through Fourier transformations. Furthermore, a meta-learning approach is utilized to optimize training efficiency and adaptability, especially for non-stationary channels with high variability. In addition, we propose a novel architecture, called alternative meta-learning for 3D dual CNN, by adopting the meta-learning, i.e., offline-learning, to only one CNN, having the feature of non-stationary channel, in dual CNN. Via simulation results, the robustness and superior performance of the proposed methods across various signal-to-noise ratios (SNRs) and Doppler shifts are verified.