一种模型驱动的深度学习OFDM接收机

针对正交频分复用(Orthogonal Frequency Division Multiplexing, OFDM)接收机解调精度低和计算复杂度高的问题，采用深度学习方法构建了一种新的模型驱动的接收机模型，称为FBLTNet(Fully Connected, Bi-LSTM and Transformer-encoder Neural Network)。该模型分为信道估计和信号检测两个部分，其中信道估计以全连接神经网络(Fully Connected Deep Neural Network, FCDNN)替代线性插值，信号检测则使用深度自注意力网络编码器Transformer-encoder和双向长短期记忆网络(Bidirectional Long-Short Term Memory, Bi-LSTM)的组合网络，实现信号的解调和比特流的恢复。在瑞利衰落信道下测试了不同调制方式的接收机性能，结果表明FBLTNet与基于深度学习的接收机以及传统接收机相比，误比特率性能得到了显著的改善；与数据驱动的无线接收机算法相比，线下训练模型收敛时间和测试时间分别减少了33.0%和25%,网络结构参数...

A Model-driven OFDM Receiver Based on Deep Learning

To overcome the problems of low demodulation precision and high computational complexity in the orthogonal frequency division multiplexing(OFDM) wireless receiver,a new model-driven OFDM receiver model named FBLTNet（Fully Connected，Bi-LSTM and Transformer-encoder Neural Network） including channel estimation and signal detection is proposed on the deep learning technology.Fully Connected Deep Neural Network(FCDNN) is applied to replace linear interpolation in the channel estimation.In the signal detection part,the combination of Transformer-encoder and Bidirectional Long Short Term Memory(Bi-LSTM) is applied to achieve signal demodulation and bit stream recovery.The receiver performance in different modulation modes is tested in the Rayleigh fading channel.Simulation results show that the FBLTNet achieves a significant improvement performance in the bit rate compared with the deep learning based receivers as well as the conventional receivers.Compared with the data-driven wireless receiver algorithm,the offline training model has better performance by reducing the convergence time,testing time and network parameters 33.0%,25% and 29.5%,respectively.