Abstract
The problem of beam alignment and tracking in high-mobility scenarios such as high-speed railway (HSR) is increasingly challenging due to the large overhead and significant time delays associated with fast time-varying channel estimation. This paper presents a learning-aided beam prediction scheme designed to predict beam directions and channel amplitudes with fine time granularity. By transforming high-dimensional beam prediction into a two-stage task—parameter estimation followed by cascaded hybrid beamforming—the proposed method utilizes a data-driven fusion module and a learnable non-linear mapping to handle complex trajectories. Simulation results demonstrate that this approach achieves near-zero overhead and latency, significantly outperforming traditional beam management schemes.
Introduction
As high-speed railway (HSR) technology advances toward speeds of 500 km/h, the demand for high-capacity wireless communication for passengers and operational systems has surged. Millimeter-wave (mmWave) MU-MIMO systems are the primary candidates to meet these demands. However, mmWave communication relies on highly directional "pencil beams" to overcome significant path loss. In HSR environments, the extremely short "dwell time"—the duration a beam remains aligned with a fast-moving receiver—renders traditional beam tracking methods inefficient.
The Challenges of High-Mobility mmWave
Traditional beam management is reactive. It relies on frequent "handshakes" or pilot signals to estimate the channel and adjust beam direction. At 500 km/h, this approach faces two critical bottlenecks:
- Resource Overhead: In scenarios with many users, the system spends the majority of its time on beam training rather than data transmission.
- Feedback Latency: The time required to process pilot signals and feedback beam indices often exceeds the channel's coherence time. By the time a beam is selected, the train has already moved out of its path.
Proposed Methodology: Proactive Beam Prediction
To overcome these limitations, we move away from reactive tracking toward a proactive prediction framework that anticipates the train's movement and adjusts beams in advance. The system architecture is divided into two primary stages.
Stage 1: Parameter Estimation and Data Fusion
The first stage involves estimating the train's instantaneous location and velocity. We utilize two independent observation sources:
- Pilot-based Estimation: Derived from received reference signals.
- Physical Measurement Estimation: Derived from Doppler frequency shifts and communication delays.
Recognizing that the reliability of these sources varies with the train's distance from the base station (BS), we implement a Data-Driven Data Fusion Module. This lightweight neural network learns to weight the inputs from different estimators based on the train's position, producing a fused state estimate that is more robust and accurate than any single source.
Stage 2: Non-linear Track Mapping and Prediction
Real-world HSR tracks are not always straight lines. To account for curves and geometric complexities, we incorporate a Learnable Non-linear Mapping module. Unlike "black-box" deep learning models, this module utilizes a piece-wise linear function approach. This design ensures the model remains interpretable while providing the flexibility to fit non-linear trajectories with high precision.
Once the trajectory is mapped, the system predicts future beam directions and channel amplitudes. This information is then used to calculate optimal analog and digital precoders using a Minimum Mean Square Error (MMSE) criterion.
Simulation and Performance
The proposed scheme was evaluated using the 3GPP TR 38.901 Urban Macro (UMa) Line-of-Sight (LoS) channel model. The simulations considered speeds up to 500 km/h and various track geometries.
Key Findings
- Throughput Gains: For a carriage with 50 users, the proposed scheme achieved a 60% increase in mean effective throughput compared to standard 5G beam tracking.
- Overhead Reduction: In high-density scenarios (100 users), traditional beam training consumed 80.5% of resources. Our prediction scheme reduced this overhead to a mere 2.97%.
- Accuracy and Latency: The system achieved a beam prediction accuracy of 98.68%. Crucially, because the base station predicts the required beam parameters ahead of time, the system effectively operates with zero feedback delay.
Conclusion
By shifting from reactive tracking to proactive prediction, this learning-aided scheme addresses the fundamental bottlenecks of mmWave communication in high-mobility HSR scenarios. The combination of model-based priors and data-driven refinement provides a robust, interpretable, and highly efficient solution for the next generation of railway communications. Future work will explore the integration of Integrated Sensing and Communication (ISAC) to further refine location data without the need for dedicated pilots.
Summary updated