摘要: |
将信道估计视为低分辨率图像重建为高分辨率图像,借鉴图像超分辨重建思想,提出了一种基于快速超分辨重建及残差连接思想的信道估计方法——ResFSRNet。采用最小二乘法(Least Square,LS)计算单个正交频分复用(Orthogonal Frequency Division Multiplexing,OFDM)子帧中所有导频处的信道响应,将其视为小尺寸低分辨率“图像”作为神经网络输入,利用多个卷积层对其进行特征提取,且融入残差连接提升性能,最后通过转置卷积重构出完整OFDM子帧信道响应。在不同抽头延迟线信道环境中进行仿真,通过信道估计误差和链路误码率结果比较,表明ResFSRNet性能优于LS、实用信道估计及基于超分辨率重建的ChannelNet,且较ChannelNet在减少约99%计算量的前提下提高了约2 dB信道估计性能。 |
关键词: 信道估计 深度学习 超分辨率重建 残差连接 |
DOI:10.20079/j.issn.1001-893x.230224001 |
|
基金项目:重庆市自然科学基金项目(cstc2019jcyj-msxmX0079) |
|
Channel Estimation Based on Fast Super-resolution Reconstruction and Residual Connection |
HUANG Fengxiang,DUAN Hongguang,LIU Hexin |
(School of Communication and Information Engineering,Chongqing University of Posts andTelecommunications,Chongqing 400065,China) |
Abstract: |
By considering channel estimation as a low-resolution image reconstruction into a high-resolution image,a channel estimation method based on the idea of fast super-resolution reconstruction and residual connection(ResFSRNet) is proposed by referring to the idea of image super-resolution reconstruction.The least square(LS) method is used to calculate the channel response at all pilots in a single orthogonal frequency division multiplexing(OFDM) subframe.Then,the channel response is treated as a small size low resolution “image” as input to the neural network,and multiple convolutional layers are used to extract its features,incorporating residual connections to improve performance.Finally,the complete OFDM subframe channel response is reconstructed through transposed convolution.Simulations in different tapped delay line(TDL) channel environments show that the performance of ResFSRNet is better than that of LS,practical channel estimation(PCE) and ChannelNet based on super-resolution reconstruction by comparing the channel estimation error and link bit error rate results.Moreover,compared with ChannelNet,ResFSRNet improves the channel estimation performance by about 2 dB while reducing computational complexity by about 99%. |
Key words: channel estimation deep learning super-resolution reconstruction residual connection |