SE Block-Assisted ResNet for Channel Estimation in OFDM System

Channel estimation (CE) is an important part of wireless communication system, which has a significant impact on the quality of wireless communication. Considering a single-input single-output (SISO) downlink scenario, this paper proposes a “squeeze and excitation” (SE) block combined with residual neural network (SE-ResNet) method to improve the CE performance of the orthogonal frequency division multiplexing (OFDM) system. The SE-ResNet is inputted into a CSI matrix of the pilot position obtained by the least squares (LS) method, and a raw feature matrix is learned by the convolutional layer and a partial residual layer. And the global information of the original feature matrix in each channel is compressed into a descriptor through the squeeze operation in the “SE” block. Then attention map is obtained from information aggregated in the descriptors by an excitation operation which fully captures channel-wise dependencies. The attention map is multiplied by the original feature matrix to get a new feature matrix, which is resized by an interpolation layer to obtain CSI of entire frame. New feature matrix is helpful for interpolation layer to get more accurate complete CSI. To further improve CE performance and optimize network model, this paper adopts methods of network tailoring and Concrete Autoencoder (Concrete AE) for pilot design. Simulation results show that the proposed SE-ResNet is superior to the traditional LS and minimum mean squared error (MMSE) methods in various practical wireless environments. Network pruning will reduce the number of parameters of the network within an acceptable loss range, which will reduce computational costs. Using the pilot scheme designed by Concrete AE for CE will have better performance.