TL;DR
This paper introduces a deep learning-based approach that uses real-time GPS coordinates to proactively predict optimal beamforming directions for 60 GHz millimeter-wave (mmWave) vehicle-to-vehicle (V2V) communications. By bypassing time-consuming traditional beam-sweeping methods, the proposed 1D Convolutional Neural Network (CNN) significantly reduces latency while capturing up to 84.58% of the maximum possible received power.
What Problem Does This Paper Address?
Millimeter-wave (mmWave) communications (operating at frequencies like 60 GHz) are essential for Connected and Autonomous Vehicles (CAVs) because they provide the massive bandwidth needed to exchange data-heavy files, such as 3D LiDAR point clouds. However, mmWave signals suffer from severe path loss and atmospheric attenuation. To combat this, systems use directional beamforming, focusing signals into narrow, high-gain beams.
The catch? For two moving vehicles to communicate, their narrow beams must be perfectly aligned. Traditional standards (like IEEE 802.11ad or 3GPP 5G NR) rely on exhaustive beam sweeping—testing every possible beam pair combination to find the best connection. This process can take around 10 milliseconds and must be repeated frequently as vehicles move. In highly dynamic V2V scenarios, this overhead creates a massive latency bottleneck, often leading to beam misalignment and dropped links.
The authors address this by exploring out-of-band contextual information. Instead of wasting network overhead on channel state estimation, can we use a vehicle's physical location data to predict the best beam direction ahead of time?
Methods
The authors propose a position-aware deep learning framework that translates GPS coordinates directly into optimal beam selections.
- Data Preprocessing: Raw GPS data (latitude and longitude) from both the transmitter ($v_2$) and receiver ($v_1$) are normalized and mapped into a Cartesian coordinate system to ensure stable neural network convergence.
- Model Architecture: The core of the solution is a custom 1D CNN designed for sequential coordinate data. It consists of:
- Three Convolutional Blocks: Each block contains 1D convolutional layers followed by a max-pooling layer to extract spatial features and prevent overfitting.
- Flatten and Fully Connected Layers: These layers process the high-level features to map positions to beam configurations.
- Softmax Layer: The final layer outputs a probability distribution over the available beam indices, allowing the system to recommend a subset of the "Top-$M$" best beams.
Instead of searching through all 64 possible beam directions, the receiver only sweeps through the top-$M$ beams predicted by the network, drastically narrowing the search space.
Datasets
The model was trained and evaluated using the DeepSense6G dataset (specifically scenarios 36, 37, 38, and 39). This is a real-world multi-modal dataset collected in Arizona (Tempe, Phoenix, Scottsdale, and Chandler) under various urban and inter-city conditions during both day and night.
The hardware setup featured:
- Transmitter (Tx): A vehicle equipped with a single omnidirectional antenna operating at 60 GHz.
- Receiver (Rx): A vehicle equipped with four mmWave phased arrays (16 elements each) facing right, left, back, and front, yielding a 64-beam codebook.
- Sensors: Both vehicles used GPS Real-Time Kinematics (RTK) sensors sampling at 10 samples/second.
Key Insights and Ideas
- Proactive vs. Reactive Alignment: Traditional beamforming reacts to channel degradation. This method is proactive, using physical topology (GPS) to predict channel characteristics before a single mmWave packet is sent.
- Reducing the Search Space via Top-$M$ Predictions: Rather than aiming for a perfect Top-1 prediction every time, predicting a small subset ($M \in \{1, 5, 9, 13\}$) provides a safety net for the system. It balances the speed of reduced overhead with the reliability of a quick local sweep.
- V2V Complexity over V2I: Unlike Vehicle-to-Infrastructure (V2I) setups where the base station is stationary, V2V requires tracking relative motion between two mobile nodes, making real-world dataset validation crucial compared to idealized ray-tracing simulations.
Key Results
The proposed deep learning model was compared against a baseline fingerprinting method (which indexes historical channel data into uniform location bins).
- Accuracy Improvement: For Top-1 beam prediction in Scenario 36, the deep learning model achieved a 19.67% increase in average accuracy over the baseline under identical beam search overheads.
- Received Power Ratio: In the best-case scenario (Scenario 36), using just the Top-1 predicted beam yielded 84.58% of the maximum possible received power on average compared to an exhaustive search.
- Across all tested scenarios (36–39), the deep learning approach consistently outperformed the baseline in both prediction accuracy and received power ratios across various values of $M$.
Limitations and Future Work
While promising, the paper notes a few areas for improvement:
- Single Modality Reliance: The system relies entirely on GPS data. GPS can suffer from inaccuracies in "urban canyons" (areas with tall buildings) or tunnels, which could degrade beam prediction performance.
- Future Directions: The authors plan to extend this work into a multimodal sensor fusion framework. Combining GPS data with other on-board non-RF sensors—such as cameras, RADAR, and LiDAR—could make the beam prediction model much more robust against sensor-specific failures.
Summary updated