Bi-directional enhancement network with adaptive gradient optimization for mmWave beam alignment

doi:10.23919/JSEE.2025.000121

Abstract

Abstract:

To address the significant overhead of the beam training in millimeter wave (mmWave) wireless communications, we propose a balanced multimodal fusion network with bi-directional enhancement in the previous work. It can predict the optimal mmWave beam according to the mmWave wide beam and sub-6 GHz channel state information. Utilizing the interaction and complementarity between different modes can improve the accuracy and robustness of predictions compared to the single mode (mmWave only or sub-6 GHz only). However, the feature extraction of one modality can potentially impact other modalities via parameter backpropagation. Although multimodal models perform better than unimodal models, they may be underutilized. To solve the problem, the adaptive multimodal gradient (AMMG) optimization method is designed to update the gradient of per feature extraction branch adaptively in this paper. Furthermore, a multimodal bi-directional enhancement (MBdE) block is proposed to combine features of other modalities, enhancing the complementarity of mmWave modality and sub-6 GHz modality. As revealed by the numerical results, it is evident that the proposed scheme delivers more precise outcomes with lower overhead compared to conventional and modern deep learning methods.

Key words: beam prediction, multimodal fusion, complementarity, gradient optimization, bi-directional enhancement

Jiawen Liu, Xiaohui Li, Dandan Ma, Ruihao Guo. Bi-directional enhancement network with adaptive gradient optimization for mmWave beam alignment[J]. Journal of Systems Engineering and Electronics, 2026, 37(4): 1186-1195.

Figures/Tables 10

Fig 1

Fig 2

Table 1

Table 2

Fig 3

Fig 4

Table 3

Fig 5

Fig 6

Fig 7

References 20

1	Hong W The role of millimeter-wave technologies in 5G/6G wireless communications. IEEE Journal of Microwaves, 2021, 1 (1): 101. doi: 10.1109/JMW.2020.3035541
2	Sun S, Rappaport T S, Heath R W, et al MIMO for millimeter-wave wireless communications: beamforming, spatial multiplexing, or both. IEEE Communications Magazine, 2014, 52 (12): 110. doi: 10.1109/MCOM.2014.6979962
3	Hong S H, Park J, Kim S J, et al Hybrid beamforming for intelligent reflecting surface aided millimeter wave MIMO systems. IEEE Trans. on Wireless Communications, 2022, 21 (9): 7343. doi: 10.1109/TWC.2022.3157880
4	Zhang D Y, Shen S Y, She C Y, et al Training beam sequence design for mmWave tracking systems with and without environmental knowledge. IEEE Trans. on Wireless Communications, 2022, 21 (12): 10780. doi: 10.1109/TWC.2022.3187167
5	Hashemi M, Koksal C E, Shroff N B Out-of-band millimeter wave beamforming and communications to achieve low latency and high energy efficiency in 5G systems. IEEE Trans. on Communications, 2018, 66 (2): 875. doi: 10.1109/TCOMM.2017.2771308
6	Cheng M, Wang J B, Cheng J, et al Joint scheduling and precoding for mmWave and Sub-6 GHz dual-mode networks. IEEE Trans. on Vehicular Technology, 2020, 69 (11): 13098. doi: 10.1109/TVT.2020.3021012
7	Semiari O, Saad W, Bennis M Joint millimeter wave and microwave resources allocation in cellular networks with dual-mode base stations. IEEE Trans. on Wireless Communications, 2017, 16 (7): 4802. doi: 10.1109/TWC.2017.2703109
8	Wang Y, Wei Z Q, Feng Z Y Beam training and tracking in mmWave communication: a survey. China Communications, 2024, 21 (6): 1. doi: 10.23919/jcc.ea.2021-0873.202401
9	Jang T H, Lee A, Kim S, et al Wideband wide beamwidth modified angled dipole antenna for 5G millimeterwave IoT applications. IEEE Access, 2023, 11, 63324. doi: 10.1109/ACCESS.2023.3285619
10	Chiu S E, Ronquillo N, Javidi T Active learning and CSI acquisition for mmWave initial alignment. IEEE Journal on Selected Areas in Communications, 2019, 37 (11): 2474. doi: 10.1109/JSAC.2019.2933967
11	Qi C H, Wang Y J, Li G Y. Deep learning for beam training in millimeter wave massive MIMO systems[J]. IEEE Trans. on Wireless Communications, 2020. DOI: 10.1109/TWC.2020.3024279.
12	Ma K, He D X, Sun H C, et al Deep learning assisted calibrated beam training for millimeter-wave communication systems. IEEE Trans. on Communications, 2021, 69 (10): 6706. doi: 10.1109/TCOMM.2021.3098683
13	Wang W, Zhang W Joint beam training and positioning for intelligent reflecting surfaces assisted millimeter wave communications. IEEE Trans. on Wireless Communications, 2021, 20 (10): 6282. doi: 10.1109/TWC.2021.3073140
14	Rezaie S, Manchon C N, Carvalho E D. Location-and orientation-aided millimeter wave beam selection using deep learning[C]// Proc. of the IEEE International Conference on Communications, 2020. DOI: 10.1109/ICC40277.2020.9149272.
15	Klautau A, Gonzalez-Preclic N, Heath R W LIDAR data for deep learning-based mmWave beam-selection. IEEE Wireless Communications Letters, 2019, 8 (3): 909. doi: 10.1109/LWC.2019.2899571
16	Ma K, Zhao P Y, Wang Z C. Deep learning assisted beam prediction using out-of-band information[C]. Proc. of the IEEE 91st Vehicular Technology Conference, 2020. DOI: 10.1109/VTC2020-Spring48590.2020.9128825.
17	Alrabeiah M, Alkhateeb A Deep learning for mmWave beam and blockage prediction using Sub-6 GHz channels. IEEE Trans. on Communications, 2020, 68 (9): 5504. doi: 10.1109/TCOMM.2020.3003670
18	Sim M S, Lim Y G, Park S H, et al Deep learning-based mmWave beam selection for 5G NR/6G with sub-6 GHz channel information: algorithms and prototype validation. IEEE Access, 2020, 8, 51634. doi: 10.1109/ACCESS.2020.2980285
19	Liu J W, Li X H, Fan T, et al Multimodal fusion assisted mmWave beam training in dual-model networks. IEEE Trans. on Vehicular Technology, 2024, 73 (1): 995. doi: 10.1109/TVT.2023.3308093
20	Xu Z Y, Jiang W, Geng J Dual-branch dynamic modulation network for hyperspectral and LiDAR data classification. IEEE Trans. on Geoscience and Remote Sensing, 2023, 61, 5514813. doi: 10.1109/tgrs.2023.3285097

Parameter	Notation	Value
Center frequency /GHz	$ {f_\text{sub}},{f_m} $	3.5, 28
Cell radius /m	$ {d_\text{sub}},{d_m} $	100, 100
Transmit power /dBm	$ P_T^\text{sub},P_T^m $	15, 20
Bandwidth of subchannel /MHz	$ {\omega _\text{sub}},{\omega _m} $	10, 100
Number of transmit antennas	$ {N_t},{N_T} $	16, 64
Number of receive antennas	$ {N_r},{N_R} $	1, 8
Delay spread within one cluster /ns	$ \Delta {\tau _{{c_\text{sub}}}},\Delta {\tau _{{c_m}}} $	7, 5
Spread within one cluster	$ \Delta {\phi _{{c_\text{sub}}}},\Delta {\phi _{{c_m}}} $	$ {3^ \circ },{\text{ }}{2.4^ \circ } $
Number of clusters	$ N_{c1}^\text{sub},N_{c1}^m $	6, 3
Number of path per cluster	$ N_\text{ray}^\text{sub},N_\text{ray}^m $	10, 20
MmWave shadowing fading	$ \varepsilon _\text{NLOS}^m,\varepsilon _\text{LOS}^m $	8.7, 5.8
Sub-6 GHz shadowing fading	$ \varepsilon _\text{NLOS}^\text{sub},\varepsilon _\text{LOS}^\text{sub} $	6, 4
MmWave path loss constant offset	$ \alpha _\text{NLOS}^m,\alpha _\text{LOS}^m $	72, 61.4
MmWave attenuation constant	$ \beta _\text{NLOS}^m,\beta _\text{LOS}^m $	3.91, 2.2
Sub-6 GHz path loss constant offset	$ \alpha _\text{NLOS}^\text{sub},\alpha _\text{LOS}^\text{sub} $	13.5,36.3
Sub-6 GHz attenuation constant	$ \beta _\text{NLOS}^\text{sub},\beta _\text{LOS}^\text{sub} $	2.92, 2

Module	Layer	Parameter
Convolution module	Convolution	$ {f_{\mathrm{i}}}{\text{ }} = {\text{ }}2 $, $ {\text{ }}{f_{\mathrm{o}}}{\text{ }} = {\text{ }}64 $, (4, 4, 0), BatchNorm
	ReLU activation	$ {f_{\mathrm{i}}}{\text{ }} = {\text{ }}64,{\text{ }}{f_{\mathrm{o}}}{\text{ }} = {\text{ }}64 $
	Convolution	$ {f_{\mathrm{i}}}{\text{ }} = {\text{ }}64 $, $ {\text{ }}{f_{\mathrm{o}}}{\text{ }} = {\text{ }}256 $, (4, 4, 0), BatchNorm
	ReLU activation	$ {f_{\mathrm{i}}} = 256, {f_{\mathrm{o}}} = 256 $
	Pooling	$ {f_{\mathrm{i}}}{\text{ }} = {\text{ }}256,{\text{ }}{f_{\mathrm{o}}}{\text{ }} = {\text{ }}256 $, max-polling
Fusion module	Concatenate	$ {f_{\mathrm{i}}}{\text{ }} = {\text{ }}256,{\text{ }}{f_{\mathrm{o}}}{\text{ }} = {\text{ }}256 $
Fusion module	Feature fussion	concatation([x1, x2], axis=1)
Output module	Fully connected	$ {f_{\mathrm{i}}} = 256, {f_{\mathrm{o}}} = 64 $, dropout = 0.3
Output module	Softmax activation	$ {f_{\mathrm{i}}} = 64, {f_{\mathrm{o}}} = 64 $

Algorithm	Model size /MB	Number of FLOPs /M	Train time of per epoch /s	Execution time of per prediction/μs
MMF	2.89	0.907	10.125	29.178
CNN	3.20	1.235	11.357	31.213
AMMG-MMF	2.89	1.56	15.784	29.178
MBdE-AMMG- MMF	3.56	1.98	17.536	30.874

[1]	Bingdong LIU, Ruihang YU, Zhiming XIONG, Meiping WU. A lightweight pure visual BEV perception method based on dual distillation of spatial-temporal knowledge [J]. Journal of Systems Engineering and Electronics, 2026, 37(1): 36-44.
[2]	Jun TANG, Wanting QIN, Qingtao PAN, Songyang LAO. A deep multimodal fusion and multitasking trajectory prediction model for typhoon trajectory prediction to reduce flight scheduling cancellation [J]. Journal of Systems Engineering and Electronics, 2024, 35(3): 666-678.
[3]	Mingwang Zhang and Yanli Lü. New Mehrotra’s second order predictor-corrector algorithm for P_*(κ) linear complementarity problems [J]. Journal of Systems Engineering and Electronics, 2010, 21(4): 705-712.