AI接收机使能的6G低密度DMRS性能研究

doi:10.12267/j.issn.2096-5931.2026.10.011

摘要/Abstract

摘要：

第六代移动通信技术(The 6th Generation Mobile Communication Technology, 6G)对频谱效率和连接密度提出了更高要求,解调参考信号(Demodulation Reference Signal, DMRS)导频开销成为制约系统性能的关键因素。降低DMRS密度可将更多资源释放给数据传输,同时为端口复用提供更大空间,但也使接收机面临信道估计精度下降与色噪声等性能挑战。面向6G,3GPP RAN1启动了“AI接收机赋能的低密度DMRS”专项研究,利用人工智能(Artificial Intelligence,AI)技术提升接收机在低密度DMRS场景下的信号处理能力,共提出了7种AI接收机架构。首先,从AI接收机技术分类角度,综述了近年来基于AI的信道估计、多输入多输出(Multiple-Input Multiple-Output, MIMO)检测、星座解调及多模块联合优化四个方向的研究进展;随后,系统梳理了7种AI接收机架构的定义、分类、优势以及局限;并在此基础上,对其中典型架构进行了链路级仿真验证。仿真结果表明,AI接收机在低密度DMRS下仍能保持优异的译码性能。

关键词: 6G, 导频开销, AI接收机, 低密度DMRS

Abstract:

The 6th Generation Mobile Communication Technology (6G) imposes high requirements on spectral efficiency and connection density, making Demodulation Reference Signal (DMRS) pilot overhead a critical factor limiting system performance. Reducing DMRS density releases resources for data transmission and provides great room for port multiplexing, but it also exposes the receiver to challenges such as degraded channel estimation accuracy and colored noise. Toward 6G, 3GPP RAN1 has initiated the “low-density DMRS with AI receiver” study, which leverages Artificial Intelligence (AI) to enhance the receiver’s signal processing capability under low-density DMRS and proposes seven AI receiver architectures in total. From the perspective of AI receiver taxonomy, this paper first reviews recent advances in four directions: AI-based channel estimation, MIMO detection, constellation demapping, and multi-module joint optimization. It then systematically summarizes the definitions, taxonomy, advantages, and limitations of the seven AI receiver architectures, and on this basis conducts link-level simulation verification for representative architectures. The simulation results show that AI receivers still maintain superior decoding performance under low-density DMRS.

Key words: 6G, pilot overhead, AI receiver, low-density DMRS

中图分类号:

TN929.5

贾承璐, 孙鹏. AI接收机使能的6G低密度DMRS性能研究[J]. 信息通信技术与政策, 2026, 52(10): 87-96.

JIA Chenglu, SUN Peng. Performance study of AI-receiver-enabled 6G low-density DMRS[J]. Information and Communications Technology and Policy, 2026, 52(10): 87-96.

导出引用管理器 EndNote|Ris|BibTeX

链接本文:

http://ictp.caict.ac.cn/CN/10.12267/j.issn.2096-5931.2026.10.011

http://ictp.caict.ac.cn/CN/Y2026/V52/I10/87

图/表 12

参考文献 24

[1]	3GPP WG 1. Chair notes RAN1#122b[R/OL]. 2025. https://www.3gpp.org/ftp/tsg_ran/WG1_RL1/TSGR1_120b/Inbox/Chair_notes.
[2]	Lagona L, Bockelmann C, et al. Transformer-based pilot-to-prediction for MIMO-OFDM systems[J]. IEEE Wireless Communications Letters, 2026, 15: 2333-2337. doi: 10.1109/LWC.2026.3676796 URL
[3]	Liao X, He X, Zhang Z, et al. Channel estimation for MIMO-OFDM based on adaptive transformer network in high-speed mobile scenarios[J]. Journal of Electronics & Information Technology, 2026, 48(8): 3604-3615.
[4]	Qiao Z, Xue J, Zhang J, et al. Adaptive implicit-based deep learning channel estimation for 6G communications[J]. IEEE Communications Magazine, 2025, 63(8): 64-70.
[5]	Huang C, Wang C X, Li J, et al. AI-driven 6G pervasive predictive wireless channel modeling for IEEE P1944 standardization: a comprehensive survey[J]. IEEE Journal of Selected Topics in Electromagnetics, Antennas and Propagation, 2026, 2: 255-275. doi: 10.1109/JSTEAP.2026.3708808 URL
[6]	Ai B, Lu Y, Fang Y, et al. 6G-enabled smart railways[J]. Proceedings of the IEEE, 2026.
[7]	Watanabe T, Takigawa M, Ohseki T, et al. Unified channel estimation framework for flexible sparse DMRS patterns via transformer-based point cloud neural operator[C/OL]// Proceedings of the IEEE International Conference on Communications (ICC). 2026, 1-6. https://ieeexplore.ieee.org/abstract/document/11587829.
[8]	Zhou X, Liang L, Ye H, et al. Reducing pilots in channel estimation with predictive foundation models[J]. IEEE Transactions on Communications, 2026, 74: 9360-9375. doi: 10.1109/TCOMM.2026.3698937 URL
[9]	Yamamoto H, Nakamura A, Uchida D, et al. Indoor experimental trial on AI-based pilotless transmission[J]. IEEE Access, 2026, 14: 64161-64170. doi: 10.1109/ACCESS.2026.3684876 URL
[10]	Ibarra-Hernández R F, Castillo-Soria F R, Gutiérrez C A, et al. Efficient deep learning-based detection scheme for MIMO communication systems[J]. Sensors, 2025, 25(3): 669. doi: 10.3390/s25030669 URL
[11]	Daha M Y, Khurshid K, Ashraf M I, et al. Optimizing signal detection in MIMO systems: AI vs approximate and linear detectors[J]. IEEE Communications Letters, 2024, 10: 2387-239.
[12]	Mahmoud H, Mi D, Ismail T, et al. Enhanced performance of signal detection using deep learning approach in massive MIMO systems[C/OL]// Proceedings of the IEEE International Conference. 2025, 192-197. https://ieeexplore.ieee.org/document/11006798.
[13]	Ali A K, Shin J, Lim Y, et al. Explore advanced hybrid deep learning for enhanced wireless signal detection in 5G OFDM systems[J]. Computer Modeling in Engineering & Sciences, 2025, 145(3): 4245-4278.
[14]	Arslan M, Tong M S. An innovative neural network technique for MIMO systems: iterative decoding and channel estimation for optimal performance[C]// Proceedings of the Photonics & Electromagnetics Research Symposium (PIERS). 2025, 192-197. https://ieeexplore.ieee.org/document/11394038.
[15]	Gansekoele A, Balatsoukas-Stimming A, et al. Joint demapping of QAM and APSK constellations using machine learning[J]. IEEE Open Journal of the Communications Society, 2025, 6: 1695-1709. doi: 10.1109/OJCOMS.2025.3547261 URL
[16]	Hebbar S A, Ankireddy S K, Athi H, et al. Deep-OFDM: neural modulation for high mobility[PP/OL]. arXiv(2026-04-16)[2026-06-02]. https://arxiv.org/abs/2506.17530.
[17]	Bray N, Hempel M, Sharif H. Towards fully DL-driven RF: a systematic survey of deep learning for wireless transceiver signal processing[J]. Applied Sciences, 2026, 16(4): 1878. doi: 10.3390/app16041878 URL
[18]	Roessler A, Lopez M M, Diab R, et al. AI/ML-driven 6G network solutions with energy efficiency considerations[J]. IEEE Access, 2026, 14: 39617-39645. doi: 10.1109/ACCESS.2026.3666644 URL
[19]	Doha S R, Abdelhadi A. Deep learning in wireless communication receivers: a survey[J]. IEEE Access, 2025, 13:113586-113605. doi: 10.1109/ACCESS.2025.3584000 URL
[20]	Wiesmayr R, Cammerer S, Aït Aoudia F, et al. Design of a standard-compliant real-time neural receiver for 5G NR[C/OL]// Proceedings of the IEEE International Conference on Machine Learning for Communications and Networking (ICMLCN). 2025, 1-6. https://ieeexplore.ieee.org/abstract/document/11140048.
[21]	Hashempoor H, Choi W. Deep learning-based data-assisted channel estimation and detection[J]. IEEE Transactions on Machine Learning in Communications and Networking, 2025, 3: 534-551. doi: 10.1109/TMLCN.2025.3559472 URL
[22]	Humse K K, Anjanappa C. Deep learning approaches for channel estimation and BER reduction in multipath fading OFDM systems[J]. SN Computer Science, 2025, 6(8): 930. doi: 10.1007/s42979-025-04485-1
[23]	3GPP WG 1. Chair notes RAN1#125[R/OL]. 2026. https://www.3gpp.org/ftp/tsg_ran/WG1_RL1/TSGR1_125/Inbox/Chair_notes.
[24]	3GPP TS 38. 901 Study on channel model for frequencies from 0.5 to 100 GHz V18.0.0[S/OL]. https://www.3gpp.org/ftp/Specs/archive/38_series/38.901.

架构编号	覆盖范围	功能
架构1	基于AI的DMRS导频去噪	AI模型仅对DMRS RE处的LS粗估计信道进行去噪处理,后续的信道插值、均衡和解调仍采用传统方法
架构2	基于AI的数据RE信道估计	AI模型以DMRS处的LS估计为输入,输出精炼后的目标RE信道响应
架构3	基于AI的目标RE信道预测	AI模型利用历史时隙的信道估计预测未来时隙的信道响应,适用于高速移动等快变信道场景
架构4	基于AI的联合信道估计、均衡和解调	单个AI模型同时替代信道估计、均衡和星座解调三个模块,实现端到端接收处理
架构5	基于AI的信道估计+基于AI的均衡和解调	两个AI模型分别处理信道估计、均衡和解调,模块间解耦程度高于架构4
架构6	信道估计+均衡+基于AI的解调	信道估计和均衡采用传统方法,仅将星座解调替换为AI模型
架构7	基于AI的信道估计+均衡+基于AI的解调	信道估计和解调分别采用AI模型,均衡保留传统MMSE干扰拒绝合并(Interference Rejection Combining, IRC)方法

架构编号	覆盖范围	功能
架构1	基于AI的DMRS导频去噪	AI模型仅对DMRS RE处的LS粗估计信道进行去噪处理,后续的信道插值、均衡和解调仍采用传统方法
架构2	基于AI的数据RE信道估计	AI模型以DMRS处的LS估计为输入,输出精炼后的目标RE信道响应
架构3	基于AI的目标RE信道预测	AI模型利用历史时隙的信道估计预测未来时隙的信道响应,适用于高速移动等快变信道场景
架构4	基于AI的联合信道估计、均衡和解调	单个AI模型同时替代信道估计、均衡和星座解调三个模块,实现端到端接收处理
架构5	基于AI的信道估计+基于AI的均衡和解调	两个AI模型分别处理信道估计、均衡和解调,模块间解耦程度高于架构4
架构6	信道估计+均衡+基于AI的解调	信道估计和均衡采用传统方法,仅将星座解调替换为AI模型
架构7	基于AI的信道估计+均衡+基于AI的解调	信道估计和解调分别采用AI模型,均衡保留传统MMSE干扰拒绝合并(Interference Rejection Combining, IRC)方法

参数	取值
频率范围/GHz	3.5
子载波间隔/kHz	30
波形	CP-OFDM
系统带宽(资源块数)	12
OFDM符号数	14
信道模型	CDLA~E, UMi
时延拓展	{30 ns, 100 ns, 300 ns}
运动速度/km/h	0~300
基站接收天线	32 RX, (M, N, P, Mg, Ng, Mp, Np) = (4, 4, 2, 1, 1, 1, 1)
用户发射天线	4 TX, (M, N, P, Mg, Ng, Mp, Np) = (4, 1, 1, 1, 1, 1, 1)
预编码粒度	宽带
链路自适应	固定调制编码方案
MIMO方案	单用户MIMO
DMRS频域密度	Comb 2/4/6/8/12
DMRS符号位置	3符号[2, 7, 11]; 2符号[2, 11]
MIMO层数	{1, 2}
传输块大小(Transport Block Size,TBS)	固定TBS

参数	取值
频率范围/GHz	3.5
子载波间隔/kHz	30
波形	CP-OFDM
系统带宽(资源块数)	12
OFDM符号数	14
信道模型	CDLA~E, UMi
时延拓展	{30 ns, 100 ns, 300 ns}
运动速度/km/h	0~300
基站接收天线	32 RX, (M, N, P, Mg, Ng, Mp, Np) = (4, 4, 2, 1, 1, 1, 1)
用户发射天线	4 TX, (M, N, P, Mg, Ng, Mp, Np) = (4, 1, 1, 1, 1, 1, 1)
预编码粒度	宽带
链路自适应	固定调制编码方案
MIMO方案	单用户MIMO
DMRS频域密度	Comb 2/4/6/8/12
DMRS符号位置	3符号[2, 7, 11]; 2符号[2, 11]
MIMO层数	{1, 2}
传输块大小(Transport Block Size,TBS)	固定TBS

参数	架构2	架构6
模型结构	Transformer	Transformer
模型输入	LS估计经线性插值后的DMRS信道估计结果	均衡后的星座符号
模型输出	精炼后的全RE信道频率响应	LLR
训练标签	理想信道估计	已知编码比特
训练方式	离线多场景/配置混合训练	离线多场景/配置混合训练
MIMO层数	1或2	1或2
总计算量	1.9 GFLOPs/14符号/12RB/32RX/传输层	140 MFLOPs/14符号/12RB/传输层
总参数量	5.8 M	0.02 M
基线对比方案	MMSE信道估计+MMSE-IRC均衡+ Maxlog解调	MMSE信道估计+MMSE-IRC均衡+ MAP解调
泛化策略	混合训练信道类型:CDL{A, B, C, D, E} 时延拓展:{30, 100, 300 ns} 运动速度:0~30或60~120 km/h SNR:-20~0dB或-10~10dB MIMO层:{1, 2}层	混合训练信道类型:UMi SNR:-20~0 dB或-10~10 dB MIMO层: {1, 2}层调制方式:16QAM, 64QAM