A CSI Entropy-Based Estimation Algorithm and Modified Recurrent Neural Networks for Time-Varying Non-Stationary Channels

Accurate channel state information (CSI) estimation is helpful for reliable transmission in high mobility communication scenarios, in which the channel being highly time-varying and non-stationary. Existing long short-term memory (LSTM) for channel estimation experiences accuracy degradation as the channel deteriorates, resulting in an unsatisfiable reliable requirement. In this paper, we propose a modified neural network architecture based on LSTM to exploit the channel's temporal correlation inherent for fast and accurate CSI acquisition, named CeLSTM. In the CeLSTM, we design a CSI entropy rectifier linear unit (CeRLU) for mapping the calculated CSI entropy into recurrent neural networks, which measures the temporal coherence of historically estimated CSI to facilitate the accurate estimation of current CSI. The CeLSTM learns nonlinear channel statistical correlation characteristics and conducts self-corrects from received signals by offline and online learning, respectively, to estimate CSI in dynamic environments. Simulation results show that CeLSTM improves the CSI estimation accuracy by 45% and 60% on average compared with the minimum mean square error (MMSE) and LSTM, respectively. Benefiting from CeRLU, the estimation time-consuming based on CeLSTM is 60% lower than LSTM and exhibits good robustness through MSE performance analysis under various time-varying channels.