Deep Learning for joint channel estimation and feedback in massive MIMO systems(1)

1.Introduction

In this paper, a comprehensive research is carried out to establish a joint channel estimation and feedback framework for the FDD massive MIMO systems based on DL techniques. The main contributions of this paper are summarized as follows.

1. The DL based joint channel estimation and feedback framework of downlink channels in FDD massive MIMO systems is proposed in this paper, which is the first to the best of our knowledge. Two networks are constructed to perform explicit and implicit channel estimation and feedback, respectively. The channel estimation and feedback network (CEFnet) employs a lightweight CNN structure to explicitly obtain the refined estimate of channels and utilizes a denoising autoencoder structure (DAE) to compress and reconstruct the noisy channel matrices. The other pilot compression and feedback network (PFnet) compresses and sends back the pilot information directly to the BS without estimating the channels.
   第一次在FDD的模式下设计了基于DL的联合信道估计和信道信息反馈，之前的paper大多数只做feedback。提出了显式和隐式的网络架构，使用了轻量级的CNN架构，使用了去噪模块。
2. A multi-signal-to-noise-ratios (SNRs) training technique is proposed to cope with multiple SNR cases so that the construction of multiple individual models for each single SNR can be avoided, which significantly reduces the storage space and makes the trained network robust to channel noise. Moreover, quantization module is enrolled into the whole network to generate data-bearing bitstreams and observe the robustness of the two networks to the quantization distortion.
   使网络适应了不同信噪比的情况（这还是很重要的），同时测试了量化噪声的影响
3. Performance analysis of the proposed two networks is provided. Both proposed networks demonstrate excellent reconstruction capacity while the CEFnet works a little better than PFnet but PFnet generates less parameters that needs storing than CEFnet. Moreover, the two networks are also proved to be robust to the quantization errors and noise.
   提供了性能分析

2.System model

2.1 Massive MIMO system

主要考虑FDD的下行链路，第$i$根天线的下行链路
$$y_i=X_i\widetilde{h_i}$$其中，$X_i\in C^{\hat{k}\times \hat{k}}$是一个对角阵，$\widetilde{h_i}\in C^{\hat{K}\times 1}$是信道频域响应,这么建模的原因详见关于Y=HX的一些思考，看懂这篇这样写公式就好理解了。
总的建模可以写为
$$y=\sum _{i=1}^{N_t}{X}_i\widetilde{h_i}+n$$其中$\tilde{H}=[\widetilde{h_1},\widetilde{h_2},...,\widetilde{h_{N_t}}]$

2.2 Channel Estimation

使用LS做信道估计，作为输入DL网络的初始值

2.3 Channel Feedback

包含Encoder和Decoder，具体可能还会含有其他模块，诸如Denoise，Quantization模块