基于数据驱动深度学习方法的无线信道均衡

The Wireless Channel Equalization Based on Data-Driven Deep Learning Method

无线信道均衡可以被看成将接收端符号恢复成发射符号集中某个符号的问题;而无线通信系统中的许多恢复过程可以被认为是通过学习一组具有良好的概率包络和相干时间的随机滤波器来克服信号的线性混合、旋转、时移、缩放以及卷积等特性。具体地,使用卷积神经网络(CNN)来学习这些滤波器,然后将学习到的滤波器送入后续的循环神经网络进行时域建模,最后对信号进行分类。实验显示:卷积-循环神经网络(CRNN)均衡器与传统的递归最小二乘滤波器(RLS)、多层感知机滤波器(MLP)在达到相同误码率(SER)情况下好2～4 d B。

Channel equalization can be viewed as a task that classifies or reconstructs the received signal as a symbol from the transmitting symbol set at the receiver. Many recovery processes in wireless communication systems can be considered to overcome linear mixing, rotation, time-shift, scaling and convolution by learning a set of random filters with good probabilistic envelope and coherent time.Concretely, convolutional neural network（CNN） is used to learn these filters, which are send into the subsequent recurrent neural network（RNN） for temporal modeling,and finally the signals are classified. Experimental results show that our convolutional recurrent neural network-based（CRNN） equalizer outperforms the recursive least square（RLS） and multi-layer perceptron network（MLP） equalizers by average 2 to4 d B with the same symbol error rate（SER）.