高速移动环境下基于RM-Net的大规模MIMO CSI反馈算法

doi:10.11959/j.issn.1000-436x.2022097

摘要/Abstract

摘要：

针对高速移动环境信道特征复杂多变，同时存在加性噪声和非线性效应的影响，提出一种残差混合网络（RM-Net）的大规模MIMO CSI反馈算法。RM-Net通过学习高速移动信道的空间结构与时间相关性，具备去除大规模MIMO信道噪声的能力，能显著提高CSI压缩率与恢复质量。系统仿真结果表明，RM-Net可消除高速移动场景加性噪声的影响，学习并适应稀疏、双选衰落信道特征，在高压缩率与低信噪比条件下依然具有较好的性能表现，所提算法性能大幅优于其他基于压缩感知（CS）和深度学习（DL）的CSI反馈算法。

关键词: 高速移动, 大规模MIMO, CSI反馈, 深度学习, 去噪

Abstract:

Aiming at the complex and changeable channel characteristics in high-speed mobile environment, and the influence of additive noise and nonlinear effects, a residual mixing network (RM-Net) for massive MIMO CSI feedback was proposed.By learning the spatial structure and temporal correlation of high-speed mobile channel, the network was able to remove massive MIMO channel noise, and the CSI compression rate and recovery quality could be significantly improved.System simulation results show that RM-Net can eliminate the influence of additive noise in high-speed mobile scenarios, learn and adapt to the channel characteristics of sparse and double-selective fading channels, and still has good performance under the conditions of high compression rate and low signal-to-noise ratio.The proposed algorithm performance is much better than other CS-based and DL-based CSI feedback algorithms.

Key words: high-speed mobility, massive MIMO, CSI feedback, deep learning, denoising

中图分类号:

TN929.5

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.

Yong LIAO, Shiyi WANG. 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.

图/表 13

图1

图2

图3

表1

表2

图4

图5

图6

图7

图8

表3

不同算法时间复杂度对比"

算法	MACC次数	时间复杂度
DCT-OMP	$2 (\frac{N^{3}}{r}) + k ((N + 2)) {(\frac{N}{r})}^{8} + N^{3}$	$O (k {(\frac{N}{r})}^{8})$
FFT-OMP	$3 (\frac{N^{3}}{r}) + 2 k ((N + 2) {(\frac{N}{r})}^{2} + {(\frac{N}{r})}^{8}) + 2 N^{3}$	$O (k {(\frac{N}{r})}^{8})$
DCT-SP	$2 (\frac{N^{3}}{r}) + k (3 {(\frac{N}{r})}^{2} + {(\frac{N}{r})}^{8}) + N^{3}$	$O (k {(\frac{N}{r})}^{8})$
FFT-SP	$3 (\frac{N^{3}}{r}) + 2 k (3 {(\frac{N}{r})}^{2} + {(\frac{N}{r})}^{8}) + 2 N^{3}$	$O (k {(\frac{N}{r})}^{8})$
CsiNet	$63 N^{2} + \frac{N^{3}}{r} + \frac{N^{3}}{r^{2}} + 17 N$	$O ((\frac{N^{3}}{r}))$
RM-Net	$(81 + \frac{81}{8}) N^{2} + 36 N + 2 N^{3} + \frac{N^{3}}{r} + \frac{N^{3}}{r^{2}}$	$O ((\frac{N^{3}}{r}))$

表3

表4

速度为300 km/h时不同算法在不同压缩率下的系统耗时"

压缩率	DCT-OMP/s	DCT-SP/s	FFT-OMP/s	FFT-SP/s	CsiNet/s	RM-Net/s
$\frac{1}{2}$	0.046 111	0.113 320	0.354 618	0.509 470	0.000 041	0.000 050
$\frac{1}{4}$	0.029 846	0.090 926	0.089 616	0.210 245	0.000 039	0.000 058
$\frac{1}{8}$	0.019 217	0.077 054	0.053 133	0.166 253	0.000 038	0.000 670
$\frac{1}{16}$	0.013 105	0.051 884	0.032 966	0.110 745	0.000 037	0.000 059
$\frac{1}{32}$	0.010 936	0.002 812	0.025 075	0.005 312	0.000 037	0.000 620

表4

表5

速度为150 km/h时不同算法在不同压缩率下的系统耗时"

压缩率	DCT-OMP/s	DCT-SP/s	FFT-OMP/s	FFT-SP/s	CsiNet/s	RM-Net/s
$\frac{1}{2}$	0.058 612	0.139 869	0.385 385	0.553 002	0.000 029	0.000 053
$\frac{1}{4}$	0.036 871	0.107 178	0.110 933	0.243 348	0.000 027	0.000 068
$\frac{1}{8}$	0.024 685	0.095 483	0.066 412	0.198 248	0.000 021	0.000 044
$\frac{1}{16}$	0.017 630	0.065 788	0.042 496	0.127 908	0.000 025	0.000 058
$\frac{1}{32}$	0.014 217	0.073 280	0.029 348	0.046 163	0.000 026	0.000 051

表5

参考文献 25

[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]	BOCCARDI F , HEATH R W , LOZANO A ,et al. Five disruptive technology directions for 5G[J]. IEEE Communications Magazine, 2014,52(2): 74-80.
[3]	BARRIAC G , MADHOW U . Space-time communication for OFDM with implicit channel feedback[J]. IEEE Transactions on Information Theory, 2004,50(12): 3111-3129.
[4]	TSENG C C , WU J Y , LEE T S . Enhanced compressive downlink CSI recovery for FDD massive MIMO systems using weighted block-minimization[J]. IEEE Transactions on Communications, 2016,64(3): 1055-1067.
[5]	HE K M , ZHANG X Y , REN S Q ,et al. Identity mappings in deep residual networks[C]// Proceedings of the 14th European Conference on Computer Vision. Berlin:Springer, 2016: 630-645.
[6]	SZEGEDY C , IOFFE S , VANHOUCKE V . Inception-v4,inception-ResNet and the impact of the residual connections on learning[C]// Proceedings of the 31st AAAI Conference on Artificial Intelligence. Palo Alto:AAAI Press, 2017: 4278-4284.
[7]	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.
[8]	LIU Z Y , ZHANG L , DING Z . Exploiting bi-directional channel reciprocity in deep learning for low rate massive MIMO CSI feedback[J]. IEEE Wireless Communications Letters, 2019,8(3): 889-892.
[9]	YAO H T , DAI F , ZHANG S L ,et al. DR2-Net:deep residual reconstruction network for image compressive sensing[J]. Neurocomputing, 2019,359: 483-493.
[10]	WANG T Q , WEN C K , JIN S ,et al. Deep learning-based CSI feedback approach for time-varying massive MIMO channels[J]. IEEE Wireless Communications Letters, 2019,8(2): 416-419.
[11]	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.
[12]	SUN Q , WU Y Z , WANG J ,et al. CNN-based CSI acquisition for FDD massive MIMO with noisy feedback[J]. Electronics Letters, 2019,55(17): 963-965.
[13]	JANG Y , KONG G , JUNG M ,et al. Deep autoencoder based CSI feedback with feedback errors and feedback delay in FDD massive MIMO systems[J]. IEEE Wireless Communications Letters, 2019,8(3): 833-836.
[14]	LIU L F , OESTGES C , POUTANEN J ,et al. The COST 2100 MIMO channel model[J]. IEEE Wireless Communications, 2012,19(6): 92-99.
[15]	HUANG G , LIU S C , DER MAATEN L V ,et al. CondenseNet:an efficient DenseNet using learned group convolutions[C]// Proceedings of 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition. Piscataway:IEEE Press, 2018: 2752-2761.
[16]	ZHANG T , QI G J , XIAO B ,et al. Interleaved group convolutions[C]// Proceedings of 2017 IEEE International Conference on Computer Vision. Piscataway:IEEE Press, 2017: 4383-4392.
[17]	KAISER L , GOMEZ A N , CHOLLET F . Depthwise separable convolutions for neural machine translation[J]. Computer Science Computation and Language, 2017,34(4): 145-168.
[18]	VEIT A , WILBER M J , BELONGIE S . Residual networks behave like ensembles of relatively shallow networks[C]// Proceedings of the 2016 Conference on Advances in Neural Information Processing Systems. Barcelona:Neural Information Processing Systems Foundation, 2016: 550-558.
[19]	IOFFE S , SZEGEDY C . Batch normalization accelerating deep network training by reducing internal covariate shift[J]. Computer Science Machine Learning, 2015,15: 56-61.
[20]	KOLOMVAKIS N , MATTHAIOU M , COLDREY M . Massive MIMO in sparse channels[C]// Proceedings of 2014 IEEE 15th International Workshop on Signal Processing Advances in Wireless Communications. Piscataway:IEEE Press, 2014: 21-25.
[21]	BARANIUK R G . Compressive sensing lecture notes[J]. IEEE Signal Processing Magazine, 2007,24(4): 118-121.
[22]	KUO P H , KUNG H T , TING P A . Compressive sensing based channel feedback protocols for spatially-correlated massive antenna ar rays[C]// Proceedings of 2012 IEEE Wireless Communications and Networking Conference (WCNC). Piscataway:IEEE Press, 2012: 492-497.
[23]	IQBAL R , ABHAYAPALA T D , LAMAHEWA T A . Generalised Clarke model for mobile-radio reception[J]. IET Communications, 2009,3(4): 644-654.
[24]	FAN G H , SUN J L , GUI G ,et al. Fully convolutional neural network based CSI limited feedback for FDD massive MIMO systems[J]. IEEE Transactions on Cognitive Communications and Networking, 2021,PP(99): 1.
[25]	JUHA M , PEKKA K , TOMMI J ,et al. WINNER II channel models[M]. New Jersey: Wiley Publishing, 2008.

配置类型	配置参数
显卡	GeForce GTX 1660Ti
处理器	Intel Core^?i7-9750H
优化器	ADAM
损失函数	MSE
运行环境	Python 3.6.5
DL框架	Pytorch 1.2.0
训练集	10 000
验证集	3 000
测试集	1 000
epoch	200
batch	250
学习率	0.001
神经元数量	2.1×10⁶～3.1×10⁶

参数	数值
双工模式	FDD
发射天线/根	64
接收天线/根	4
载波频率/GHz	2.9
子载波数	84
系统带宽/MHz	1.4
观测矩阵	高斯随机矩阵
稀疏基	DCT、FFT
多径时延/μs	[0, 4, 8, 12, 16]
路径增益/dB	[0, -3, -6, -8, -17.2]
多普勒频移/kHz	[0.422, 0.845]
最大相对速度/(km·h^-1)	[150, 300]
信噪比/dB	[-10, -5, 0, 5, 10, 15, 20]