高速移动场景下基于自适应Transformer网络的MIMO-OFDM信道估计

廖希; 何香妮; 张哲; 王洋

doi:10.11999/JEIT260075

高速移动场景下基于自适应Transformer网络的MIMO-OFDM信道估计

doi: 10.11999/JEIT260075 cstr: 32379.14.JEIT260075

廖希^1,,
何香妮¹,
张哲^{1, 2, ,},
王洋^{1, 2}

1.
重庆邮电大学通信与信息工程学院重庆 400065
2.
中国人民解放军31401部队长春 130012

基金项目: 重庆市自然科学基金(No.CSTB2025YITP-QCRC0045)

详细信息

作者简介:
廖希：女，博士，教授，研究方向为面向6G的信道测量与建模、广域无线通信等，邮箱为　liaoxi@cqupt.edu.cn

何香妮：女，硕士生，研究方向为OFDM系统信道估计

张哲：男，硕士，工程师，研究方向为无线通信，邮箱为　19628069@qq.com

王洋：男，博士，教授，研究方向为第六代移动通信技术、毫米波太赫兹测量与建模、涡旋电磁波和智能反射面等

通讯作者:
张哲　19628069@qq.com

中图分类号: TN911.23
计量
- 文章访问数: 60
- HTML全文浏览量: 31
- PDF下载量: 2
- 被引次数: 0
出版历程
- 收稿日期: 2026-01-21
- 修回日期: 2026-07-13
- 录用日期: 2026-07-13
- 网络出版日期: 2026-07-24

Channel Estimation for MIMO-OFDM Based on Adaptive Transformer Network in High-Speed Mobile Scenarios

LIAO Xi^1
,,
HE Xiangni¹,
ZHANG Zhe^{1, 2
, ,},
WANG Yang^{1, 2}

1.
College of Communication and Information Engineering, Chongqing University of Posts and Telecommunications, Chongqing 400065, China
2.
PLA Unit 31401, Changchun 130012, China

Funds: Chongqing Natural Science Foundation (No.CSTB2025YITP-QCRC0045)

摘要

摘要: 高速移动场景下显著多普勒频移引发信道快速时变特性及子载波间干扰加剧，导致MIMO-OFDM系统信道估计精度下降及鲁棒性不足，本文提出一种基于特征线性调制(Feature-wise Linear Modulation, FiLM)的自适应Transformer信道估计模型—AdaFiT。首先，采用可分离二维线性上采样获取全时频域信道估计，并利用点卷积在通道维度对多天线复数信道的实部与虚部进行联合建模，结合卷积特征增强模块提取多尺度时频特征；然后，设计信道自适应特征调制模块，将信道参数作为先验条件信息，基于FiLM机制生成特征层面的缩放系数与偏移系数，仿射调制Transformer网络的块嵌入序列，实现对信道统计变化的条件自适应；最后，通过二维可学习位置编码的Transformer编码器全局建模块级特征，并结合残差重建结构融合局部与全局特征，完成高精度信道重建。结果表明，所提AdaFiT模型在CDL-C和CDL-A两种信道模型下均优于LS-Blinear、LMMSE及现有自适应方法，对不同信道模型具有较好的适应性。其中，在CDL-C信道模型下的动态信噪比、多普勒频移及不同时延扩展条件下，均方根误差最大可提升7 dB。
- MIMO-OFDM系统 /
- 信道估计 /
- Transformer网络 /
- 自适应特征调制 /
- 特征线性调制
Abstract: Objective Accurate channel state information (CSI) is essential for coherent detection, beamforming, and adaptive resource allocation in multiple-input multiple-output orthogonal frequency division multiplexing (MIMO-OFDM) systems. In high-mobility environments, large Doppler shifts and multipath propagation jointly cause doubly selective fading, destroy subcarrier orthogonality, and aggravate inter-carrier interference. Consequently, conventional least squares (LS) and linear minimum mean square error (LMMSE) estimators suffer substantial performance degradation. Existing deep learning estimators provide strong nonlinear modeling capability, but often show insufficient adaptability to variations in signal-to-noise ratio (SNR), delay spread, and Doppler shift, or inefficiently incorporate physical channel priors. To address these problems, a Transformer network with adaptive feature modulation, termed AdaFiT, is proposed for channel estimation in high-mobility MIMO-OFDM systems. Methods The proposed AdaFiT framework performs channel estimation by jointly exploiting local time-frequency correlations, global dependencies, and explicit channel-aware adaptation. It takes least squares (LS) estimates at pilot positions, together with signal-to-noise ratio (SNR), delay spread, and maximum Doppler shift, as inputs. A separable two-dimensional linear upsampling module first interpolates the sparse pilot estimates to the full OFDM time-frequency grid by independently processing the real and imaginary components along the frequency and time dimensions. A convolutional feature enhancement module then extracts robust local representations: a complex feature mixing layer fuses multi-antenna real and imaginary components, while multi-scale convolutional blocks with channel attention capture short-range time-frequency correlations and suppress noise. Next, a feature-wise linear modulation (FiLM)-based channel-adaptive module embeds the three channel parameters through independent multilayer perceptrons and combines them into a channel-state representation. This representation generates scaling and shifting coefficients to dynamically recalibrate the block-embedded sequences according to varying channel statistics. Finally, the modulated sequences are processed by a Transformer encoder with learnable two-dimensional positional encoding to model long-range dependencies across the time-frequency grid and antenna dimensions. A residual reconstruction structure fuses the global features with locally enhanced representations, yielding accurate channel estimates with global consistency and preserved local details. Results and Discussions Simulation analyses are conducted based on the CDL-C and CDL-A channel models. The LS-based bilinear interpolation method, the LMMSE method, the AdaFortiTran model, and the AdaFiT model without the adaptive module are selected as comparison schemes. The mean squared error (MSE) performance of these models is compared under different signal-to-noise ratio (SNR), maximum Doppler shift, and delay spread conditions. Under the CDL-C channel model, AdaFiT achieves the lowest MSE over the entire SNR range (Fig. 3). At an SNR of 0 dB, its MSE is approximately 2.5 dB lower than that of AdaFortiTran, and the performance gain increases to approximately 7 dB at an SNR of 30 dB. Compared with the AdaFiT model without the adaptive module, AdaFiT achieves a maximum MSE gain of approximately 2.1 dB, which verifies the effectiveness of the proposed channel-adaptive feature modulation module. When the maximum Doppler shift increases from 200 Hz to 1400 Hz, AdaFiT consistently maintains the lowest MSE (Fig. 4). In the high-Doppler range of 1000–1400 Hz, AdaFiT outperforms the AdaFiT model without the adaptive module and AdaFortiTran by approximately 2 dB and 3.5 dB, respectively, demonstrating improved robustness against rapid channel variations. For delay spreads ranging from 100 ns to 700 ns, AdaFiT also maintains the lowest MSE over the entire range (Fig. 5), achieving approximately 3 dB and 5.5 dB gains over the AdaFiT model without the adaptive module and AdaFortiTran, respectively. These results demonstrate that the proposed adaptive feature modulation mechanism effectively improves the adaptability of the model to time-selective and frequency-selective fading.To further evaluate the performance of the AdaFiT model under different channel models, simulation analysis is conducted based on the 3GPP CDL-A channel model. At SNRs of 0–5 dB, AdaFiT outperforms LMMSE by approximately 4–5 dB, while the performance gain reaches approximately 7 dB at SNRs of 20–30 dB (Fig. 6(a)). Compared with AdaFortiTran, AdaFiT achieves an MSE gain of approximately 2 dB under low-SNR conditions, and the gain increases to approximately 5.5 dB under high-SNR conditions. In the maximum Doppler shift experiment, the MSE of AdaFiT remains between approximately –38 dB and –37 dB over the range of 200–800 Hz and is approximately 5 dB lower than that of AdaFortiTran (Fig. 6(b)). Although the MSE of AdaFiT increases when the maximum Doppler shift exceeds 1000 Hz, it still achieves an approximately 6 dB gain over LMMSE at 1400 Hz and remains superior to AdaFortiTran. Over the entire delay spread range, AdaFiT achieves approximately 4–6 dB gain over LMMSE and approximately 4–5.5 dB gain over AdaFortiTran (Fig. 6(c)). Conclusions The proposed AdaFiT framework performs channel estimation by jointly exploiting local time-frequency correlations, global dependencies, and an explicit channel-aware adaptation mechanism. Simulation results under the CDL-C and CDL-A channel models show that AdaFiT consistently achieves lower MSE than the LS-based bilinear interpolation method, LMMSE, AdaFortiTran, and the AdaFiT model without the adaptive module under different SNR, maximum Doppler shift, and delay spread conditions. These results verify the effectiveness of the proposed adaptive feature modulation mechanism and demonstrate that AdaFiT maintains high estimation accuracy and stable performance under different channel models, indicating good adaptability to varying channel environments.
- MIMO-OFDM Systems /
- Channel estimation /
- Transformer network /
- Adaptive feature modulation /
- Feature-wise Linear Modulation(FiLM)

HTML全文

图 1 导频模式($ {N}_{\text{t}}=4,D=12,{N}_{\text{p}}=2 $)

下载: 全尺寸图片幻灯片

图 2 AdaFiT模型架构及子模块

下载: 全尺寸图片幻灯片

图 3 不同信噪比下各模型的MSE

下载: 全尺寸图片幻灯片

图 4 不同最大多普勒频移下各模型的MSE

下载: 全尺寸图片幻灯片

图 5 不同时延扩展下各模型的MSE

下载: 全尺寸图片幻灯片

图 6 CDL-A信道下各模型在不同信道条件下的MSE性能对比

下载: 全尺寸图片幻灯片

图 7 先验信息失配对AdaFiT信道估计性能的影响

下载: 全尺寸图片幻灯片

表 1 模型复杂度对比

模型	参数量	复杂度
LS-Bilinear	−	$ O\left({N}_{\text{tx}}{N}_{\text{rx}}{N}_{\text{f}}{N}_{\text{t}}\right) $
LMMSE	−	$ O\left({N}_{\text{tx}}{N}_{\text{rx}}{N}_{\text{f}}{}^{3}{N}_{\text{t}}\right) $
AdaFortiTran	2.738M	$ O\left({({{N}_{\text{tx}}}{{N}_{\text{rx}}})}^{2}{N}_{\text{f}}{N}_{\text{t}}+{S}^{2}(D+k)+S{(D+k)}^{2}\right) $
AdaFiT(无自适应模块)	0.268M	$ O\left({({{N}_{\text{tx}}}{{N}_{\text{rx}}})}^{2}{N}_{\text{f}}{N}_{\text{t}}+{S}^{2}D+S{D}^{2}\right) $
AdaFiT(本文模型)	0.958M	$ O\left({({{N}_{\text{tx}}}{{N}_{\text{rx}}})}^{2}{N}_{\text{f}}{N}_{\text{t}}+{S}^{2}D+S{D}^{2}+({N}_{\text{tx}}{N}_{\text{rx}}){N}_{\text{f}}{N}_{\text{t}}\right) $

下载: 导出CSV

表 2 MIMO-OFDM系统参数

参数	符号	数值	参数	符号	数值
发射天线数	$ {N}_{\text{tx}} $	4	接收天线数	$ {N}_{\text{rx}} $	4
载波频率	$ {f}_{\text{c}} $	3.5 GHz	信道模型	−	CDL-C; CDL-A
子载波间隔	$ {\Delta }f $	15 kHz	系统带宽	$ B $	5 MHz
OFDM符号数	$ {N}_{\text{t}} $	14	总子载波数	$ {N}_{\text{f}} $	300
单天线导频数	$ {N}_{\text{p}} $	50	导频间隔	$ D $	12
数据调制	−	QPSK	信噪比	$ s $	[0, 5, 10, 15, 20, 25, 30] dB
时延扩展	$ {\tau }_{\text{ds}} $	[50, 150,···, 650] ns	最大多普勒频移	$ {f}_{\text{d}} $	[100, 200,···, 1000] Hz

下载: 导出CSV

参考文献(19)

[1]	SINGH S, KUMAR S, MAJHI S, et al. Blind carrier frequency offset estimation techniques for next-generation multicarrier communication systems: Challenges, comparative analysis, and future prospects[J]. IEEE Communications Surveys & Tutorials, 2025, 27(1): 1–36. doi: 10.1109/COMST.2024.3472109.
[2]	GUO Jiajia, CHEN Tong, JIN Shi, et al. Deep learning for joint channel estimation and feedback in massive MIMO systems[J]. Digital Communications and Networks, 2024, 10(1): 83–93. doi: 10.1016/j.dcan.2023.01.011.
[3]	XU Jiarui, LI Lianjun, ZHENG Lizhong, et al. Learning to Estimate: A real-time online learning framework for MIMO-OFDM channel estimation[J]. IEEE Transactions on Wireless Communications, 2025, 24(4): 2634–2646. doi: 10.1109/TWC.2024.3487111.
[4]	CAI Jun, YIN Chuan, and DING Youwei. Optimization of resource allocation in FDD massive MIMO systems[J]. Digital Communications and Networks, 2024, 10(1): 117–125. doi: 10.1016/j.dcan.2022.11.017.
[5]	李一兵, 汤云鹤, 简鑫, 等. 面对高速移动场景的OTFS系统导频设计方法[J]. 电子与信息学报, 2025, 47(2): 490–497. doi: 10.11999/JEIT240349. LI Yibing, TANG Yunhe, JIAN Xin, et al. Pilot design method for OTFS system in high-speed mobile scenarios[J]. Journal of Electronics & Information Technology, 2025, 47(2): 490–497. doi: 10.11999/JEIT240349.
[6]	蒲旭敏, 刘雁翔, 宋米雪, 等. 基于模型驱动深度学习的OTFS信道估计[J]. 电子与信息学报, 2024, 46(2): 680–687. doi: 10.11999/JEIT230072. PU Xumin, LIU Yanxiang, SONG Mixue, et al. Orthogonal time frequency space channel estimation based on model-driven deep learning[J]. Journal of Electronics & Information Technology, 2024, 46(2): 680–687. doi: 10.11999/JEIT230072.
[7]	廖勇, 罗渝, 荆亚昊. 6G新型时延多普勒通信范式: OTFS的技术优势、设计挑战、应用与前景[J]. 电子与信息学报, 2024, 46(5): 1827–1842. doi: 10.11999/JEIT231133. LIAO Yong, LUO Yu, and JING Yahao. 6G new time-delay Doppler communication paradigm: Technical advantages, Design challenges, applications and prospects of OTFS[J]. Journal of Electronics & Information Technology, 2024, 46(5): 1827–1842. doi: 10.11999/JEIT231133.
[8]	KIM D, PARK S, KANG J, et al. Block-fading non-stationary channel estimation for MIMO-OFDM systems via meta-learning[J]. IEEE Communications Letters, 2022, 26(12): 2924–2928. doi: 10.1109/LCOMM.2022.3204763.
[9]	GHEREKHLOO S, ARDAH K, and HAARDT M. SALSA: A sequential alternating least squares approximation method for MIMO channel estimation[J]. IEEE Transactions on Vehicular Technology, 2024, 73(5): 7430–7435. doi: 10.1109/TVT.2023.3347290.
[10]	GIZZINI A K and CHAFII M. A survey on deep learning based channel estimation in doubly dispersive environments[J]. IEEE Access, 2022, 10: 70595–70619. doi: 10.1109/ACCESS.2022.3188111.
[11]	DOHA S R and ABDELHADI A. Deep learning in wireless communication receivers: A survey[J]. IEEE Access, 2025, 13: 113586–113605. doi: 10.1109/ACCESS.2025.3584000.
[12]	JIANG Peiwen, WEN Chaokai, JIN Shi, et al. Dual CNN-based channel estimation for MIMO-OFDM systems[J]. IEEE Transactions on Communications, 2021, 69(9): 5859–5872. doi: 10.1109/TCOMM.2021.3085895.
[13]	LEE J, AHN S, PARK S I, et al. Alternative meta-learning with 3D dual-CNN for MIMO channel estimation[J]. IEEE Wireless Communications Letters, 2025, 14(11): 3650–3654. doi: 10.1109/LWC.2025.3600061.
[14]	GAO Wei, ZHANG Wei, LIU Libin, et al. Deep residual learning with attention mechanism for OFDM channel estimation[J]. IEEE Wireless Communications Letters, 2025, 14(2): 250–254. doi: 10.1109/LWC.2022.3232378.
[15]	FOLA E, LUO Yang, and LUO Chunbo. AttenReEsNet: Attention-aided residual learning for effective model-driven channel estimation[J]. IEEE Communications Letters, 2024, 28(8): 1855–1859. doi: 10.1109/LCOMM.2024.3412802.
[16]	LI Jinbao and PENG Qi. Lightweight channel estimation networks for OFDM systems[J]. IEEE Wireless Communications Letters, 2022, 11(10): 2066–2070. doi: 10.1109/LWC.2022.3193199.
[17]	LIU Fangyu, ZHANG Jing, JIANG Peiwen, et al. CE-ViT: A robust channel estimator based on vision transformer for OFDM systems[C]. GLOBECOM 2023-2023 IEEE Global Communications Conference, Kuala Lumpur, Malaysia, 2023: 4798–4803. doi: 10.1109/GLOBECOM54140.2023.10436847.
[18]	GULER B and JAFARKHANI H. AdaFortiTran: An adaptive transformer model for robust OFDM channel estimation[C]. ICC 2025-IEEE International Conference on Communications, Montreal, Canada, 2025: 3797–3802. doi: 10.1109/ICC52391.2025.11160810.
[19]	TURKOGLU M O, BECKER A, GÜNDÜZ H A, et al. FiLM-Ensemble: Probabilistic deep learning via feature-wise linear modulation[C]. The 36th Annual Conference on Neural Information Processing Systems (NeurIPS 2022), New Orleans, USA, 2022: 22229–22242.