| 摘要: |
| 针对目前基于正交时频空调制系统中使用脉冲导频进行信道估计会引入高均峰比和信道估计困难的问题，提出了一种基于双向长短期记忆网络（Bidirectional Long Short-term Memory，BiLSTM）的信道估计算法。该算法首先将低功率导频序列放置于时频域中，再对发送信号添加功率放大器的非线性衰变模型，对接收信号导频部分的接收信号使用最小二乘法进行粗略信道的计算，之后通过BiLSTM结合卷积块注意力模块对完整信道进行拟合估计，从而得到时频域的完整信道矩阵。为进一步减少误差，通过神经网络对完整信道矩阵进行修正处理。仿真结果表明，该算法求得非线性衰变影响信号的信道矩阵的归一化均方误差提升了3~15 dB，均峰比下降了5~6 dB，误比特率性能提升3~8 dB。 |
| 关键词: 正交时频空调制(OTFS) 信道估计 双向长短记忆网络（BiLSTM） 深度学习 |
| DOI：10.20079/j.issn.1001-893x.241121001 |
|
| 基金项目:重庆市自然科学基金创新发展联合基金（中国星网）(CSTB2023NSCQ-LZX0114) |
|
| A BiLSTM-based Orthogonal Time Frequency Space(OTFS) Channel Estimation Algorithm |
| WANG Huahua,WEN Zichen,WEI Fanbo |
| (School of Communications and Information Engineering，Chongqing University of Posts and Telecommunications，Chongqing 400065，China) |
| Abstract: |
| In response to the issues of high peak-to-average power ratio(PAPR) and difficulties in channel estimation caused by the use of impulse pilots in current orthogonal time frequency space(OTFS) modulation systems，a channel estimation algorithm based on Bidirectional Long Short-term Memory(BiLSTM) networks is proposed.This algorithm first places low-power pilot sequences in the time-frequency domain，then applies a nonlinear distortion model of the power amplifier to the transmitted signal.The received signal at the pilot positions is used to compute a coarse channel estimate via the least squares method.Subsequently，the complete channel is estimated by fitting the data using a BiLSTM network combined with a convolutional block attention module，thereby obtaining the full channel matrix in the time-frequency domain.To further reduce errors，the complete channel matrix is refined through a neural network.Simulation results demonstrate that the proposed algorithm improves the normalized mean square error of the channel matrix for signals affected by nonlinear distortion by 3~15 dB，reduces the PAPR by 5~6 dB，and enhances the bit error rate performance by 3~8 dB. |
| Key words: orthogonal time frequency space(OTFS) channel estimation bidirectional long short-term memory(BiLSTM) deep learning |
《电讯技术》编辑部 