基于信道特征生成对抗网络的信道建模方法

基于生成对抗网络(generative adversarial networks, GAN)的数据生成特性，提出一种用于信道特征生成的GAN改进模型，即信道特征生成对抗网络(channel feature generative adversarial networks, CFGAN)。采用完全无监督学习信道特征方式，利用线性编码向量与生成信道之间的互信息关系和变分互信息最大化原理，实现编码向量与信道特征对应；采用实测室内电力线信道数据集训练CFGAN模型，训练完成的CFGAN能够学习到不同信道特征分布。仿真表明，在-80～-10 dB大动态衰减范围内，CFGAN可根据学习到的信道特征生成具有明显区别的4类信道模型，并且生成信道和实测信道的信道特征差异小于2%。

Channel modeling method based on channel feature generative adversarial networks

This paper proposes an improved model of generative adversarial networks (GAN) tailored for channel feature generation, named as channel feature generative adversarial networks (CFGAN). Using a completely unsupervised learning channel feature method, the model utilizes the mutual information relationship between the linear coding vector and the generated channel, alongside variational mutual information maximization principles, to establish a correspondence between the coding vector and channel characteristics. The CFGAN model is trained using a dataset of measured indoor power line channel data. The trained CFGAN can learn different channel feature distributions. Simulation shows that in a large dynamic range channel with an attenuation amplitude of -80~-10 dB, CFGAN can generate four types of channel models with significant differences based on the learned channel characteristics, and the difference in channel characteristics between the generated channel and the measured channel is less than 2%.