Deep Learning Based Channel Estimation for OFDM Systems With Doubly Selective Channel

Abstract

In wireless communication orthogonal frequency division multiplexing (OFDM) systems with high mobility, channel estimation becomes challenging due to the double selectivity of channels. To solve this problem, a deep learning based method which fully considers the channel features and noise effects is proposed in this letter. In the pilot-based channel estimation scheme, a compression-and-reconstruction channel estimation network (CRCENet) is designed to improve the channel estimation performance, which fuses not only the low-level features of the rapid changes between adjacent subcarriers in the short term and the high-level features of the overall channel changes under large time spans in the backbone, but also the features using the attention mechanism in the up-sampling branches. Meanwhile, considering that the samples with different signal-to-noise ratios (SNRs) have different credibility, loss weights based on credibility are assigned to data with different SNRs during network training, to further improve the estimation performance of the network. The experimental results revealed that CRCENet performed better than the traditional least squares estimation algorithm and state-of-the-art learning networks.