基于卷积长短时记忆深度神经网络的带内全双工非线性数字自干扰消除

带内全双工(IBFD)技术能够有效提高无线通信系统的频谱效率，近年来引起了广泛关注。然而，同时发送和接收引起的线性和非线性自干扰给IBFD带来了巨大挑战。传统的非线性自干扰消除主要是基于多项式模型和深度神经网络(DNN)来实现。多项式模型方法存在模型失配导致自干扰效果恶化的风险，而DNN方法无法针对高维数据特有的空频相关性、时间相关性等特点进行处理。该文基于卷积长短时记忆深度神经网络(CLDNN)，通过在输入层中引入3维张量以及在卷积层设置复数卷积层结构，分别设计了两种重建自干扰信号的网络结构——2维CLDNN(2D-CLDNN)和复值CLDNN(CV-CLDNN)，充分利用卷积神经网络局部感知和权值共享的优势，在高维特征中学习到更抽象的低维特征，从而提高自干扰消除的效果。实际场景中获得数据的评估结果显示，当功率放大器记忆长度M和自干扰信道多径长度L满足M+L=13时，通过总共60次训练轮数，该文提出的结构比传统DNN方法在非线性自干扰消除方面可以实现至少26%的改进，训练轮数也有明显减少。

Driven Non-linear Digital Self Interference Cancellation for In-Band Full Duplex Systems Based on Convolution Long Short-term Memory Deep Neural Network

In-Band Full Duplex (IBFD) technology can effectively improve the spectral efficiency of wireless communication system, which has attracted extensive attention in recent years. However, the linear and nonlinear self interference caused by simultaneous transmission and reception brings great challenges to IBFD. The traditional nonlinear self interference cancellation is mainly based on polynomial model and Deep Neural Network(DNN). Polynomial-model-based method has the risk of deterioration of self-interference effect caused by model mismatch, while DNN-based method can not deal with the unique characteristics of space-frequency correlation and time correlation of high-dimensional data. Based on Convolution Long short-term memory Deep Neural Network(CLDNN), two network structures for reconstructing self-interference signals, Two-Dimensional CLDNN(2D-CLDNN) and Complex-Value-CLDNN(CV-CLDNN), are designed by introducing three-dimensional tensor in the input layer and setting complex convolution layer structure in the convolution layer, which makes full use of the advantages of local perception and weight sharing of convolutional neural network, so as to learn more abstract low-dimensional features from high-dimensional features, so as to improve the effect of self-interference cancellation. The evaluation results of the data obtained in the actual scene show that, when the memory length M of power amplifier and the multipath length L of self interference channel meet M+L=13, through a total of 60 training epochs, the structure proposed in this paper can achieve at least 26% improvement in nonlinear self-interference cancellation compared to the traditional DNN method, the training period is also significantly reduced.