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

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.

💡 Research Summary

The paper addresses the prohibitive training overhead associated with exhaustive beam sweeping in ultra‑massive MIMO systems envisioned for 5G‑Advanced and future 6G networks. To mitigate this, the authors propose a deep‑learning framework that fuses radio‑domain measurements with explicit positioning information. The core architecture consists of a dual‑branch RegNet: one branch (the Position Extraction Branch) is supervised by ground‑truth 3‑D UE coordinates, forcing the network to learn geometry‑aware embeddings; the other branch (the Beam Feature Branch) processes conventional radio features such as Reference Signal Received Power (RSRP).
A dedicated Feature Fusion Module integrates the two modality‑specific embeddings. Two fusion strategies are explored: (1) Adaptive Fusion, which employs an attention‑like weighting mechanism that dynamically emphasizes the more informative modality under varying channel conditions; and (2) Adversarial Fusion, which introduces a domain discriminator to align the distributions of the position‑derived and beam‑derived features, thereby producing a modality‑invariant fused representation.
The authors generate extensive datasets using the DeepMIMO simulator at 3.5 GHz, following 3GPP TR 38.843 specifications. Four urban scenarios (dense downtown, intersection, high‑rise, and low‑density) are created, each providing both RSRP maps and precise UE locations. Experiments are conducted under both in‑distribution (ID) and out‑of‑distribution (OOD) conditions to assess robustness.
Results show that the proposed model consistently outperforms baselines that rely solely on RSRP or that use position information without sophisticated fusion. Specifically, Top‑1 beam prediction accuracy improves by an average of 7.3 percentage points over an RSRP‑only RegNet and by 4.1 points over prior position‑assisted methods. The gains are most pronounced at low SNR, where the Adaptive Fusion effectively leverages spatial cues to compensate for weak radio signals. The Adversarial Fusion further enhances OOD generalization, maintaining high accuracy even when building layouts and channel parameters differ from the training set. Inference latency remains below 1 ms, satisfying real‑time beam management requirements.
The paper acknowledges limitations: reliance on accurate position labels during training, the need for large labeled datasets, and the current focus on static UE positions rather than dynamic trajectories. Future work is suggested in three directions: (i) self‑supervised or semi‑supervised learning to reduce labeling burden; (ii) incorporation of temporal trajectory data via recurrent or transformer models to enable predictive beam tracking for mobile users; and (iii) distributed multi‑BS or multi‑sensor fusion to further improve robustness in dense urban deployments.
Overall, the study demonstrates that integrating positioning features through carefully designed fusion mechanisms can substantially boost beam prediction performance and robustness, offering a viable path toward low‑overhead, high‑accuracy beam management in next‑generation massive MIMO systems.