系统工程与电子技术
Systems Engineering and Electronics
收稿日期：2020-10-12；修回日期：2021-02-06。
基金项目：中国科学院空间科学战略性先导科技专项(XDA15060100)，中国科学院战略高技术创新基金(GQRC-19-14）资助课题
基于 Stacked-TCN 的空间混叠信号单通道盲源
分离方法
赵孟晨
1
，
2
，姚秀娟
2
，王 静
2
，董苏惠
1
，
2
（1. 中国科学院大学电子电气与通信工程学院，北京 100049；
2. 中国科学院国家空间科学中心，北京 100190；）
摘 要：针对空间互联网星地通信场景中的混叠信号分离精度不足问题，提出 了基于深
度学习的堆叠时域卷积网络(stacked time domain convolutional network，Stacked-TCN)分离方
法。首先，对混合信号提取编码特征表示。然后，通过时域卷积网络训练得到源信号的深层
特征掩模，将每个信号源的掩模与混合信号编码特征做 Hadamard 乘积，得到源信号的编码
特征表示。最后，使用 1-D 卷积，对源信号特征进行解码，得到原始波形。实验采用负的比
例不变信噪比(scale-invariant source to noise ratio, SISNR)作为网络训练的损失函数，即单通道
盲源分离性能的评价指标。结果表明，Stacked-TCN 方法与独立成分分析(independent
component analysis，ICA)、非负矩阵分解(non-negative matrix factorization，NMF)两种经典分
离算法和时域音频分离网络(time-domain audio separation network，TasNet)、Wave-U-Net 两种
深度学习分离方法相比，损失函数降低了 9.11~36.63dB。其中，8PSK_16QAM 分离损失为
-3.23dB@SNR=5dB，比其他 4 种方法降低了 12.77dB。Stacked-TCN 方法具有更好的分离精
度和噪声鲁棒性。
关键词：欠定盲源分离；同频干扰；单通道；时域卷积网络(TCN)
中图分类号：TN 911.7 文献标志码：A
Single-channel blind source separation of spatial aliasing signal
based on Stacked-TCN
ZHAO Mengchen
1
，
2
，YAO Xiujuan
2
，WANG Jing
2
，DONG Suhui
1
，
2
(1. School of Electronic and Communication Engineering, University of Chinese Academy of Sciences ,
Beijing 100049, China; 2. National Space Science Center; CAS, Beijing 100190, China)
Abstract: Aiming at the problem of insufficient separation accuracy of the aliased signals in
the satellite-to-ground communication scene, a Stacked-TCN method based on deep learning is
proposed. First, the coding feature of the mixture is extracted. Then, the mask of the source is
obtained through TCN. The Hadamard products is performed to the mask of source and the
mixture feature to obtain the source feature. Finally, 1-D convolution is used to decode the source
feature to obtain the waveform. SISNR is used as the loss function, that is, the evaluation index of
single-channel blind source separation performance. The results show that the Stacked-TCN
method reduce the loss by 9.11dB~ 36.63dB compared with algorithms of ICA, NMF, TasNet and
Wave-U-Net ,and reduce the loss by 12.77dB in 8PSK_16QAM mixture when SNR is 5dB in
which the loss is -3.23dB.This method has better separation accuracy and noise robustness.
Keywords: underdetermined blind source separation; co-frequency interference; single
channel; time-domain convolutional network (TCN)
网络首发时间：2021-03-02 11:45:33
网络首发地址：https://kns.cnki.net/kcms/detail/11.2422.tn.20210227.2208.006.html
评论0