Online Data Generation for MIMO-OFDM Channel Denoising: Transfer Learning vs. Meta Learning

Abstract: Channel denoising is a practical and effective technique for mitigating channel estimation errors in multiple-input multiple-output orthogonal frequency-division multiplexing (MIMO-OFDM) systems. However, adapting denoising techniques to varying channel conditions typically requires prior knowledge or incurs significant training overhead. To address these challenges, we propose a standard-compatible strategy for generating online training data that enables online adaptive channel denoising. The key idea is to leverage high-quality channel estimates obtained via data-aided channel estimation as practical substitutes for unavailable ground-truth channels. Our data-aided method exploits adjacent detected data symbols within a specific time-frequency neighborhood as virtual reference signals, and we analytically derive the optimal size of this neighborhood to minimize the mean squared error of the resulting estimates. By leveraging the proposed strategy, we devise two channel denoising approaches, one based on transfer learning, which fine-tunes a pre-trained denoising neural network, and the other based on meta learning, which rapidly adapts to new channel environments with minimal updates. Simulation results demonstrate that the proposed methods effectively adapt to dynamic channel conditions and significantly reduce channel estimation errors compared to conventional techniques.