Residual Channel Attention Network-Based Channel Interpolation Using Noise2Noise for Massive MIMO-OFDM Systems

In massive multi-input multi-output (MIMO) orthogonal frequency division multiplexing (OFDM) systems, traditional pilot-aided channel estimation methods usually need interpolation to obtain the channel response other than pilot subcarriers. Therefore, the accuracy of interpolation is crucial for channel estimation. In this paper, we propose a deep learning-based channel interpolation method that utilizes the sparsity and correlation of the massive MIMO-OFDM channel in the angular-frequency domain. Specifically, we first consider the angular domain least squares (LS) channel estimation matrix at pilot subcarriers as a low-resolution 2D image. Then, a residual channel attention network (RCAN) for channel interpolation is proposed to reconstruct the whole channel state information (CSI). Moreover, to make the approach more practical in real-world scenarios, we further introduce the Noise2Noise method, which can be used to restore channel response without clean labels. Finally, simulation results demonstrate the superiority of the proposed channel interpolation method in terms of performance and practicability.