Channel Estimation for High-speed Mobile MIMO-OFDM Systems Using Deep Learning

In high-speed mobile channel environment characterized by fast time-varying and Doppler frequency shifts, interference between subcarriers becomes increasingly prominent. In this case, channel estimation in multiple-input multiple-output orthogonal frequency division multiplexing (MIMO-OFDM) system faces significant challenges, including low time correlation, high interpolation errors, and Doppler frequency shifts. To address these issues, we propose a channel estimation framework that integrates attention mechanisms with an enhanced Super-Resolution Convolutional Neural Network (SRCNN). The whole network adopts the network structure of the autoencoder, embeds the channel attention module with powerful channel dynamic feature extraction ability at the encoder end, and reweights the channel features of the least squares (LS) estimation input with the learned attention map in turn, which can fully extract the features of the original data. Subsequently, the improved SRCNN reconstructs and refines the features of the dimensionally reduced data, exploiting its strong reconstruction and denoising capabilities to recover high-resolution characteristics. Finally, the decoder outputs the fully estimated channel matrix. Simulation results demonstrate that the proposed channel estimation method significantly outperforms the traditional least squares algorithm while exhibiting computational complexity slightly higher than LS but lower than the linear minimum mean square error (LMMSE) algorithm. Notably, as the number of user-transmitted symbols and subcarriers increases, the deep learning-based method shows greater advantages over the LMMSE algorithm, greatly reducing practical computational costs.