基于深度展开的大规模MIMO系统CSI反馈算法

doi:10.11959/j.issn.1000-436x.2022237

摘要/Abstract

摘要：

针对现阶段大规模MIMO系统中基于深度学习的信道状态信息（CSI）反馈算法待训练参数过多、可解释性不强的问题，提出了2种基于深度展开的CSI反馈算法。一种是基于可学习参数的近似消息传递（AMP）算法，该算法利用深度学习中的可学习参数将AMP算法中阈值函数的阈值和Onsager校正项的参数替换，增强了阈值函数在应对非严格稀疏数据时的非线性能力。另一种是基于卷积网络的 AMP 算法，该算法将阈值函数模块替换为卷积残差学习模块，利用该模块去除 AMP 算法中每轮迭代产生的高斯随机噪声。仿真分析表明，所提算法具有比AMP算法更好的CSI反馈表现，其中基于卷积网络的AMP算法具有比基于深度学习的代表性方法更优异的CSI重构性能。

关键词: CSI反馈, 深度学习, 深度展开, 近似消息传递, 可学习参数, 卷积网络

Abstract:

In order to solve the problem that the channel state information (CSI) feedback algorithm based on deep learning in massive MIMO systems at present had too many parameters to be trained and could not be explained well, two CSI feedback algorithms based on depth expansion were proposed.The first one was approximate message delivery (AMP) algorithm based on learnable parameters.The learnable parameters in deep learning were used to replace the threshold value of the threshold function in the AMP algorithm and the parameters of the Onsage correction term.The nonlinear ability of threshold function in dealing with non-strict sparse data was enhanced.The other was the AMP algorithm based on convolutional network, which replaced the threshold function module with the convolutional residual learning module, and used the module to remove the Gaussian random noise generated by each iteration of the AMP algorithm.Simulation results show that the proposed two algorithms have better CSI feedback performance than AMP algorithm, and the AMP algorithm based on convolutional network has better CSI reconstruction performance than the representative method based on deep learning.

Key words: CSI feedback, deep learning, deep unfolding, approximate message passing, learnable parameter, convolutional network

中图分类号:

TN929.5

廖勇, 程港, 李玉杰. 基于深度展开的大规模MIMO系统CSI反馈算法[J]. 通信学报, 2022, 43(12): 77-88.

Yong LIAO, Gang CHENG, Yujie LI. CSI feedback algorithm based on deep unfolding for massive MIMO systems[J]. Journal on Communications, 2022, 43(12): 77-88.

图/表 14

图1

图2

图3

图4

图5

图6

表1

图7

图8

表2

不同算法编码器和解码器的参数量"

压缩率	编/解码器		参数量/个
压缩率	编/解码器	CsiNet	CsiNetPlus	LP-AMP-3	CNN-AMP-5
$\frac{1}{4}$	编码器	1 048 612	1 049 264	1 048 576	1 048 576
$\frac{1}{4}$	解码器	1 051 780	1 052 486	1 054 723	1 052 901
$\frac{1}{8}$	编码器	524 324	524 986	524 288	524 288
$\frac{1}{8}$	解码器	527 492	527 548	530 435	528 613
$\frac{1}{16}$	编码器	262 180	262 424	262 144	262 144
$\frac{1}{16}$	解码器	265 384	265 846	268 291	266 469
$\frac{1}{32}$	编码器	131 108	131 342	131 072	131 072
$\frac{1}{32}$	解码器	134 312	134 462	137 219	135 397

表2

表3

不同算法的乘法次数"

压缩率	编/解码器			乘法次数/次
压缩率	编/解码器	AMP	CsiNet	CsiNetPlus	LP-AMP-3	CNN-AMP-5
$\frac{1}{4}$	编码器	1 048 576	1 085 440	1 089 642	1 048 576	1 048 576
$\frac{1}{4}$	解码器	210 917 376	4 329 472	4 338 284	7 270 912	15 960 576
$\frac{1}{8}$	编码器	524 288	561 152	567 254	524 288	524 288
$\frac{1}{8}$	解码器	105 458 688	3 805 184	3 823 128	3 672 320	10 192 128
$\frac{1}{16}$	编码器	262 144	299 008	308 238	262 144	262 144
$\frac{1}{16}$	解码器	52 729 344	3 543 040	3 583 462	1 836 160	7 307 904
$\frac{1}{32}$	编码器	131 072	167 936	179 928	131 072	131 072
$\frac{1}{32}$	解码器	26 364 672	3 411 968	3 472 842	918 080	5 865 792

表3

表4

图9

图10

参考文献 22

[1]	LI C G , LIU P , ZOU C ,et al. Spectral-efficient cellular communications with coexistent one- and two-hop transmissions[J]. IEEE Transactions on Vehicular Technology, 2016,65(8): 6765-6772.
[2]	BARRIAC G , MADHOW U . Space-time communication for OFDM with implicit channel feedback[J]. IEEE Transactions on Information Theory, 2004,50(12): 3111-3129.
[3]	RAO X B , LAU V K N . Distributed compressive CSIT estimation and feedback for FDD multi-user massive MIMO systems[J]. IEEE Transactions on Signal Processing, 2014,62(12): 3261-3271.
[4]	LIANG P Z , FAN J C , SHEN W H ,et al. Deep learning and compressive sensing-based CSI feedback in FDD massive MIMO systems[J]. IEEE Transactions on Vehicular Technology, 2020,69(8): 9217-9222.
[5]	WEN C K , SHIH W T , JIN S . Deep learning for massive MIMO CSI feedback[J]. IEEE Wireless Communications Letters, 2018,7(5): 748-751.
[6]	WANG Y , YAO H , ZHAO S . Auto-encoder based dimensionality reduction[J]. Neurocomputing, 2016,184: 232-242.
[7]	GUO J J , WEN C K , JIN S ,et al. Convolutional neural network-based multiple-rate compressive sensing for massive MIMO CSI feedback:design,simulation,and analysis[J]. IEEE Transactions on Wireless Communications, 2020,19(4): 2827-2840.
[8]	廖勇, 王帅, 孙宁 . 快时变FDD大规模MIMO系统智能CSI反馈方法[J]. 通信学报, 2021,42(7): 211-219.
	LIAO Y , WANG S , SUN N . Intelligent CSI feedback method for fast time-varying FDD massive MIMO system[J]. Journal on Communications, 2021,42(7): 211-219.
[9]	廖勇, 王世义 . 高速移动环境下基于 RM-Net 的大规模 MIMO CSI反馈算法[J]. 通信学报, 2022,43(5): 166-176.
	LIAO Y , WANG S Y . CSI feedback algorithm based on RM-Net for massive MIMO systems in high-speed mobile environment[J]. Journal on Communications, 2022,43(5): 166-176.
[10]	廖勇, 李玉杰 . 一种轻量化低复杂度的FDD大规模MIMO系统CSI反馈方法[J]. 电子学报, 2022,50(5): 1211-1217.
	LIAO Y , LI Y J . Lightweight and low complexity CSI feedback method for FDD massive MIMO systems[J]. Acta Electronica Sinica, 2022,50(5): 1211-1217.
[11]	JAGANNATH A , JAGANNATH J , MELODIA T . Redefining wireless communication for 6G:signal processing meets deep learning with deep unfolding[J]. IEEE Transactions on Artificial Intelligence, 2021,2(6): 528-536.
[12]	QING C J , CAI B , YANG Q Y ,et al. Deep learning for CSI feedback based on superimposed coding[J]. IEEE Access, 2019,7: 93723-93733.
[13]	LIU Z Y , ROSARIO M D , DING Z . A Markovian model-driven deep learning framework for massive MIMO CSI feedback[J]. IEEE Transactions on Wireless Communications, 2021,21(2): 1214-1228.
[14]	GUO J H , WANG L , LI F ,et al. CSI feedback with model-driven deep learning of massive MIMO systems[J]. IEEE Communications Letters, 2021,26(3): 547-551.
[15]	DONOHO D L , MALEKI A , MONTANARI A . Message-passing algorithms for compressed sensing[J]. Proceedings of the National Academy of Sciences, 2009,106(45): 18914-18919.
[16]	BALATSOUKAS-STIMMING A , STUDER C . Deep unfolding for communications systems:a survey and some new directions[C]// Proceedings of IEEE International Workshop on Signal Processing Systems (SiPS). Piscataway:IEEE Press, 2019: 266-271.
[17]	MONTANARI A , ELDAR Y C , KUTYNIOK G . Graphical models concepts in compressed sensing[M]. London: Cambridge University, 2012.
[18]	LOHIT S , KULKARNI K , KERVICHE R ,et al. Convolutional neural networks for noniterative re-construction of compressively sensed images[J]. IEEE Transactions on Computational Imaging, 2018,4(3): 326-340.
[19]	FRANOIS C , . Xception:deep learning with depthwise separable convolutions[C]// Proceedings of The IEEE International Conference on Computer Vision and Pattern Recognition:Piscataway:IEEE Press, 2017: 1251-1258.
[20]	ZHANG K , ZUO W M , CHEN Y J ,et al. Beyond a gaussian denoiser:residual learning of deep CNN for image denoising[J]. IEEE Transactions on Image Processing, 2017,26(7): 3142-3155.
[21]	LIU L F , OESTGES C , POUTANEN J ,et al. The COST 2100 MIMO channel model[J]. IEEE Wireless Communications, 2012,19(6): 92-99.
[22]	ZHANG Z J , . Improved Adam optimizer for deep neural networks[C]// Proceedings of IEEE/ACM 26th International Symposium on Quality of Service. Piscataway:IEEE Press, 2018: 1-2.

训练轮次	LP-AMP-3	LP-AMP-7	CNN-AMP-3	CNN-AMP-7
10	5.930×10^-5	5.931×10^-5	2.008×10^-5	1.015×10^-5
30	5.929×10^-5	5.716×10^-5	1.271×10^-5	7.029×10^-6
50	5.928×10^-5	5.599×10^-5	1.162×10^-5	5.993×10^-6
90	5.933×10^-5	5.580×10^-5	1.112×10^-5	5.309×10^-6
110	5.936×10^-5	5.553×10^-5	1.089×10^-5	5.032×10^-6
200	5.933×10^-5	5.561×10^-5	1.078×10^-5	5.118×10^-6
300	5.931×10^-5	5.559×10^-5	1.074×10^-5	5.125×10^-6

仿真参数	设置
双工模式	FDD
发射天线数	32
接收天线数	1
子载波数	1 024
信道模型	COST2100
调制方式	正交相移键控
室内场景频段/GHz	5.3
室外场景频段/MHz	300
系统带宽/MHz	20