Channel Estimation for mmWave MIMO using Deep Residual Learning

In a massive multiple-input multiple-output (MIMO) system based on a time division duplex (TDD), the base station (BS) must accurately estimate the channel state information (CSI) for a user terminal (UT). However, due to the dynamic nature of the channel, the length of pilot signals is constrained, and there is a finite number of orthogonal pilot signals available. To fully leverage the benefits of spatial diversity and multiplexing gains in massive multiple-input multiple-output (MIMO) systems, it is essential to have precise downlink channel state information (CSI) at the base station (BS). By employing minimum mean square error (MMSE) channel estimation, it becomes possible to mitigate the impact of pilot contamination in a massive multiple-input multiple-output (MIMO) system. This is achieved by utilizing the fully known covariance matrix of channels for all user terminals (UTs) that share the same pilot signal. The paper explores the combined use of MMSE and NMSE channel estimation techniques in conjunction with a Convolution encoder for channel matrix prediction. Several existing methods either suffer from overfitting or exhibit complexity similar to the zero forcing algorithm. To address these challenges, the paper introduces a three-layer residual block that effectively mitigates overfitting and efficiently handles large datasets. Additionally, the paper proposes two residual models that effectively handle dataset size variations.