Robust Machine Learning for Channel Estimation with Varying Delay and Doppler Shift Conditions

Accurate channel estimation at the receiver is critical for subsequent downlink data transmission in orthogonal frequency division multiplexing (OFDM) systems. However, tracking the rich variety of propagation environments is challenging. At the same time, the variation of the channel in time and frequency also makes channel estimation difficult. In this paper, we exploit machine learning techniques as an attractive option to learn improved solutions for channel estimation problem. Instead of training a single large neural network for all channel conditions, we design dedicated sub-networks for channels with varying delay and Doppler shift conditions. Then we propose a hard selection method to estimate the channel with the optimal sub-network which is determined by a channel condition classifier. We further propose a soft selection method to combine multiple sub-networks for better estimation performance. Simulation results show that compared with the conventional least squares (LS) channel estimation method, the proposed networks achieve approximately 7 dB normalized mean square error (NMSE) gain at SNR = 10 dB.