降噪自编码器辅助的下行MIMO-SCMA编解码方法

Denoising autoencoder-aided downlink MIMO-SCMA codec method

为了改善稀疏码分多址系统在多天线应用中的误码率性能,将深度学习引入多输入多输出稀疏码分多址系统,提出了一种降噪自编码器辅助的编解码方法。发射端使用多个深度神经网络单元构建多天线稀疏码分多址编码器,通过神经网络的学习获得每个用户在不同发射天线上的码本,采用降噪自编码器的结构在输入端引入噪声层,使得编码器的输出为更具鲁棒性的特征表示;接收端设计了一个全链接的深度神经网络作为解码器,该解码器将多天线检测与多用户检测联合进行,一次解码即可获得用户数据;采用端到端的训练方式对编解码器进行训练,优化神经网络的结构与参数,使得神经网络能够快速收敛。实验结果表明,提出的编解码方法可以降低多输入多输出稀疏码分多址系统的误码率,同时减少接收端检测的时间。

Aiming to improve the bit error rate(BER)performance of sparse code multiple access(SCMA)systems in multi-antenna applications,deep learning is introduced in a MIMO-SCMA system and a denoising autoencoder-aided Codec method(DAE-MIMO-SCMA)is proposed.Multiple deep neural network(DNN)units are used by the transmitter to construct the MIMO-SCMA encoder.The codebook of each user on different transmitting antennas is obtained through neural network(NN)learning.Moreover,the noise layer is used at the transmitter so that the output of the encoder is more robust.At the receiver is designed a fully connected DNN as a decoder,which combines multi-antenna detection and multi-user detection to obtains the original data of all users at one time.An end-to-end training method is used to train the Codec,optimizing the structure and parameters of the NN,which improves the convergent rate.Experimental results show that the proposed Codec method can lower the BER of the MIMO-SCMA system and reduce the detection time at the receiver.