Generalizable and Robust Beam Prediction for 6G Networks: An Deep-Learning Framework with Positioning Feature Fusion

TL;DR

This work tackles the overhead of beam training in 6G-style mmWave/mMIMO systems by introducing a position-aware deep learning framework. It employs dual-branch RegNet backbones to separately learn beam-domain and position-related features, fused through AutoFusion and GanFusion modules to produce robust beam predictions while supervised by coordinate information. The approach demonstrates consistent gains over strong baselines in both in-distribution and out-of-distribution urban scenarios, especially for high-resolution codebooks and low-SNR conditions, and remains feasible within sub-millisecond inference budgets. Overall, the method offers a scalable, robust pathway to efficient, information-guided beam management suitable for future wireless networks and ISAC-enabled architectures.

Abstract

Beamforming (BF) is essential for enhancing system capacity in fifth generation (5G) and beyond wireless networks, yet exhaustive beam training in ultra-massive multiple-input multiple-output (MIMO) systems incurs substantial overhead. To address this challenge, we propose a deep learning based framework that leverages position-aware features to improve beam prediction accuracy while reducing training costs. The proposed approach uses spatial coordinate labels to supervise a position extraction branch and integrates the resulting representations with beam-domain features through a feature fusion module. A dual-branch RegNet architecture is adopted to jointly learn location related and communication features for beam prediction. Two fusion strategies, namely adaptive fusion and adversarial fusion, are introduced to enable efficient feature integration. The proposed framework is evaluated on datasets generated by the DeepMIMO simulator across four urban scenarios at 3.5 GHz following 3GPP specifications, where both reference signal received power and user equipment location information are available. Simulation results under both in-distribution and out-of-distribution settings demonstrate that the proposed approach consistently outperforms traditional baselines and achieves more accurate and robust beam prediction by effectively incorporating positioning information.

Figures (5)

• Figure 1: Overall architecture of the proposed model. The pipeline begins with RSRP-based beam inputs derived from different codebooks. A dual-branch backbone extracts beam-specific and position-specific features. These are integrated via a feature fusion module and ultimately used to predict both beam prediction and positioning outputs across multiple regions.
• Figure 2: Structure of the AutoFusion module, adapted based on the original design in the reference. The module learns to combine two feature branches and generate a fused representation suitable for both loss supervision and target prediction.
• Figure 3: Structure of the GanFusion module. The generator learns to synthesize fused features from one domain, while the discriminator distinguishes them from features generated by the AutoFusion module. The adversarial training encourages the generator to produce task-consistent and domain-aligned representations.
• Figure 4: Base station and UE locations with LOS/NLOS distribution in different city scenarios. Green dots indicate LOS UEs, blue dots indicate NLOS UEs, and black dots represent base station positions.
• Figure 5: Training and validation loss convergence of the proposed models across different city scenarios. Solid lines denote the average loss, while shaded regions indicate the interquartile range (IQR).