Abstract: Beamforming techniques are essential to compensate for severe path loss in millimeter-wave (mmWave) communications by adopting large antenna arrays and formulating narrow beams. However, performing accurate beam alignment over such narrow beams for efficient link configuration by traditional beam selection approaches typically imposes significant latency and computing overheads, which is often infeasible in highly dynamic vehicle-to-vehicle (V2V) scenarios. In contrast, utilizing out-of-band contextual information, such as vehicular position information, is a potential alternative. This paper presents a deep learning-based solution utilizing vehicular position information for predicting optimal beams. Experimental evaluation on real-world measured mmWave datasets shows that the solution can achieve up to 84.58% of the maximum received power on average, confirming it as a promising solution for 60 GHz enabled V2V communications.
Introduction
Millimeter-wave (mmWave) bands, particularly at 60 GHz, are pivotal for the high-bandwidth requirements of modern autonomous vehicles, such as sharing high-resolution 3D LiDAR point clouds. However, these frequencies face significant challenges due to extreme path loss and atmospheric absorption. To maintain a reliable link, beamforming is used to concentrate signal energy into narrow, directional beams.
The primary bottleneck in highly dynamic V2V environments is alignment. Traditional methods rely on exhaustive searching or complex channel state information (CSI), which are often too slow; by the time the best beam is found, the vehicle has moved. This research seeks to predict the optimal beam direction instantly using data that is already available: vehicle position.
Proposed Methodology
The authors propose a position-aware approach that leverages "out-of-band" information—specifically real-time GPS coordinates—to predict optimal beam indices. By treating beam selection as a classification problem, the system can proactively align beams before the signal degrades.
The core of the solution is a deep learning model designed to process normalized GPS trajectories. Instead of predicting a single beam, the model outputs a "Top-M" set of candidate beams, significantly reducing the search space for the hardware while ensuring high link quality.
Deep Learning Architecture
The proposed architecture utilizes a 1D Convolutional Neural Network (CNN) to extract spatial features from movement data.
The model consists of three convolutional blocks, each containing a 1D-CNN layer and a max-pooling layer. These are followed by fully connected layers and a softmax output layer that provides a probability distribution over 64 possible beam indices. This structure allows the model to capture the temporal and spatial dynamics of vehicle trajectories effectively.
Dataset and Experimental Setup
The study leverages the DeepSense6G dataset, specifically scenarios 36-39, which represent real-world urban and inter-city environments in Arizona (Tempe, Phoenix, Scottsdale, and Chandler). The data includes over 112,000 samples collected at 10 samples per second from two moving vehicles. The receiver (Unit 1) utilized a 64-element phased array, while the transmitter (Unit 2) used an omnidirectional antenna, providing a robust baseline for evaluating beam prediction accuracy in real-world dynamics.
Key Results and Performance
The deep learning approach demonstrates significant improvements over traditional baselines, such as inverse multipath fingerprinting:
- Accuracy Boost: In Scenario 36, the proposed model achieved a 19.67% increase in Top-1 accuracy over the baseline.
- Power Efficiency: The model achieved a received power ratio of 84.58% on average for Top-1 predictions. This means the system captures nearly 85% of the signal strength that an exhaustive, time-consuming search would find, but with a fraction of the overhead.
- Robustness: Results remained consistent across different urban and inter-city scenarios, proving the model's ability to generalize to various driving environments.
Conclusion and Future Work
This research highlights the potential of using easily accessible GPS data to solve complex mmWave communication challenges. By shifting from reactive to proactive beamforming, V2V systems can maintain stable, high-speed links in dynamic environments.
Future work will focus on multi-modal fusion—incorporating LiDAR, Radar, or camera data alongside GPS—to improve robustness in "urban canyons" where GPS signals may be weak or obstructed. Additionally, the authors aim to rigorously quantify end-to-end latency reduction in live hardware implementations.
Summary updated