TL;DR
Predicting 3D cellular network coverage for low-altitude drone corridors is notoriously difficult because base station antenna patterns are proprietary secrets and drone flight data is sparse. This paper introduces a physics-guided, disentangled representation learning framework that compresses base station operational parameters into balanced, physically meaningful features. By forcing the neural network architecture to mirror signal propagation equations, the framework cuts prediction errors by 7% over top machine learning baselines while using 66% fewer parameters than standard deep models.
What Problem Does This Paper Address?
As the low-altitude economy takes off—driven by unmanned aerial vehicles (UAVs) for logistics, surveillance, and passenger transport—designing safe "aerial corridors" requires highly accurate Low-Altitude Network Coverage (LANC) prediction. Drones rely on cellular networks (4G/5G) for command and control, meaning coverage gaps can be catastrophic.
Predicting low-altitude coverage is fundamentally different from predicting terrestrial ground coverage. Ground networks rely heavily on 2D models that account for multipath scattering from buildings. Low-altitude networks operate in 3D space where direct line-of-sight (LOS) paths dominate, making coverage highly sensitive to the complex, multi-beam configurations of Base Stations (BSs).
Engineers face two massive hurdles when trying to model this:
- Proprietary Data: Telecom equipment manufacturers keep their exact antenna multi-beam directionality diagrams strictly confidential.
- Data Scarcity and Imbalance: Collecting 3D signal strength data via drone road tests is incredibly expensive. This leaves researchers with sparse data points where high-dimensional base station configurations are barely sampled, but low-dimensional spatial locations vary wildly, causing standard machine learning models to severely overfit.
Methods
The authors tackle these challenges using a dual strategy: Expert-Knowledge-Based Feature Compression and a Model-Guided Disentangled Representation Network.
1. Feature Compression
Instead of feeding raw, high-dimensional base station parameters (like absolute latitude, longitude, mechanical tilt, and digital tilt) into a black-box model, the authors use wireless communication principles to simplify the feature space. They transform absolute coordinates and orientation angles into relative horizontal/vertical angles ($\Delta\theta^{(H)}_n, \Delta\theta^{(V)}_n$) and a single distance metric ($Dn$) relative to the drone's position. They also use a signal propagation model to mathematically decouple the impact of transmission power and bandwidth from the raw signal data.
2. Disentangled Network Architecture
Standard deep neural networks (DNNs) struggle with generalization when data is sparse. To fix this, the authors design a "prior-constrained" network. Instead of one massive fully connected network, they split the architecture into three independent, parallel Multi-Layer Perceptrons (MLPs):
- Distance Fading Network: Learns how signals decay over distance.
- Frequency Fading Network: Captures frequency-dependent attenuation.
- Antenna Gain Network: Learns the semantics of beam patterns based on relative angles and channel configurations.
Crucially, the outputs of these three networks are combined additively to predict the final Reference Signal Received Power (RSRP). This directly mirrors the log-scale structure of physical path loss equations ($Path Loss = Distance Fading + Frequency Fading + Antenna Gain$), preventing the model from learning physically impossible relationships.
Datasets
The framework was trained and evaluated on real-world low-altitude network data collected via UAV drive tests in China:
- Nanchang City Dataset: Collected across three distinct areas at flight altitudes of 150m, 300m, and 500m. It contains over 93,000 samples across 142 base stations. The training data primarily featured Huawei base stations.
- Ganzhou City Dataset: A separate cross-region dataset featuring base stations manufactured entirely by ZTE. This was used to test the model's ability to generalize across different equipment manufacturers without explicit antenna type data.
Key Insights and Ideas
- Physics as a Regularizer: Forcing a neural network's architecture to structurally mimic a physical law (addition in the log-domain for path loss) acts as an incredibly powerful regularizer. It drastically reduces the model's search space without sacrificing capacity.
- 66% Parameter Reduction: The authors mathematically prove that by decoupling spatial processing into parallel subnetworks, their architecture reduces the number of hidden layer parameters by approximately 66% compared to a standard MLP of equivalent representation capacity.
- Bypassing Proprietary Secrets: The framework demonstrates that vendor-specific antenna beam patterns can be implicitly learned and generalized across different manufacturers (e.g., learning on Huawei data and predicting on ZTE hardware) simply by conditioning the network on basic channel counts and coverage scenarios.
Key Results
- Baseline Outperformance: The proposed framework consistently beat traditional machine learning models (KNN, Lasso, Random Forest, XGBoost) across all data sampling rates, achieving an average Mean Absolute Error (MAE) reduction of ~0.5 dB (a 7% error reduction over the best baseline).
- Ablation Success: Compared to a standard MLP using raw features, the disentangled framework reduced MAE by roughly 20%.
- Real-World Reliability: In practical, cross-regional testing in an completely unseen area (Xinyue Lake), the model achieved a highly practical MAE of 4.74 dB. Even in the zero-shot cross-manufacturer test in Ganzhou, it maintained a robust MAE of 6.26 dB.
Limitations and Future Work
The primary limitation is that the model assumes a dominant line-of-sight (LOS) path. While this assumption holds well above 50 meters, it ignores ground-level environmental shadowing and scattering from complex city architecture. The authors note that incorporating 3D building maps could further improve prediction accuracy by 0.5 to 2 dB.
Future work will focus on integrating large language models (LLMs) to refine network coverage predictions and incorporating detailed 3D environmental data directly into the neural network architecture.
Summary updated