Federated Learning-driven Beam Management in LEO 6G Non-Terrestrial Networks

This paper proposes a Federated Learning framework for LEO 6G Non-Terrestrial Networks that leverages High-Altitude Platform Stations to distribute beam selection tasks, demonstrating that a Graph Neural Network model outperforms a Multi-Layer Perceptron in prediction accuracy and stability, especially at low elevation angles.

The Gist

Imagine the future of the internet as a giant, invisible web of satellites zooming around the Earth, like a massive swarm of bees. These are Low Earth Orbit (LEO) satellites, and they are the backbone of the upcoming 6G network. Their job is to beam high-speed data down to your phone, no matter where you are on the planet.

But here's the problem: These satellites are moving incredibly fast, and the air between them and your phone is messy. Sometimes clouds block the signal, sometimes the angle is bad, and sometimes the signal gets weak. To keep the connection strong, the satellites need to constantly aim their "flashlights" (called beams) perfectly at you.

If they aim wrong, your video call freezes. If they aim right, you get super-fast internet.

The Old Way: The Exhaustive Search

Traditionally, figuring out the best angle is like trying to find the perfect spot to shine a flashlight in a dark room by testing every single direction one by one. The satellite has to ask, "Is this angle good? Is this one better?" This takes too much time and uses up a lot of battery and bandwidth. It's slow and clunky.

The New Idea: The "Team of Local Experts"

This paper proposes a smarter way using something called Federated Learning (FL).

Think of the satellites not as a single giant brain, but as a team of local experts.

• The Problem: You can't send all the raw data (like every single photo of the sky and signal strength) from thousands of satellites to a central server to learn. It would clog the network, like trying to mail a million letters at once.
• The Solution: Instead, each satellite (or group of satellites in the same "lane" or orbital plane) learns on its own. They figure out the best flashlight angles based on their local experience.
• The Collaboration: They don't share the messy raw data; they only share the lessons they learned (the "rules" or patterns). A central controller then combines these lessons to create a "Super-Brain" that knows how to aim beams perfectly for everyone.

The Two Contenders: The Generalist vs. The Neighbor-Knowing Expert

The researchers tested two different types of "AI brains" to see which one learns best:

1. The MLP (Multi-Layer Perceptron): Imagine a Generalist. This AI looks at your location and the satellite's position and makes a guess based on a checklist. It's fast and lightweight, like a quick calculator. It treats every beam direction as a separate, unrelated option.
2. The GNN (Graph Neural Network): Imagine a Neighbor-Knowing Expert. This AI understands that beams are connected. If the beam pointing "North" is good, the beam pointing "North-North-East" is probably also good. It looks at the whole map of beams as a connected web (a graph) and learns how they relate to each other.

The Results: Who Won?

The researchers ran a simulation with 1,000 different scenarios over two hours. Here is what happened:

• The GNN (The Expert) Crushed It: It was much better at predicting the perfect beam. It got the right answer 96% of the time, compared to the MLP's 88%.
• Stability Matters: The real test was at the "horizon" (low angles), where the signal is tricky. The MLP got confused and kept switching beams back and forth (like a flashlight flickering). The GNN stayed steady, knowing exactly which beam to hold onto.
• The Trade-off: The GNN is slightly bigger and takes a bit longer to train (like a more complex brain), but it's still small enough to fit on a satellite. The extra smarts are worth the tiny cost.

The Big Picture

This paper shows that by letting satellites learn together without sharing private data, and by using an AI that understands how things are connected (the GNN), we can make the future 6G satellite internet faster, more reliable, and less prone to dropping calls.

In a nutshell: Instead of one giant brain trying to control the whole sky, we have a team of smart, local experts sharing their wisdom to keep your connection strong, even when the satellites are zooming by at 17,000 mph.

Technical Summary

Here is a detailed technical summary of the paper "Federated Learning-driven Beam Management in LEO 6G Non-Terrestrial Networks."

1. Problem Statement

The rapid evolution of Sixth-generation (6G) systems demands global, hyper-reliable, and low-latency connectivity, necessitating the integration of Non-Terrestrial Networks (NTNs) with terrestrial infrastructure. Specifically, Low Earth Orbit (LEO) satellite constellations face significant challenges in beam management due to:

• Dynamic Propagation Conditions: High mobility and rapidly changing channel states.
• Limitations of Conventional Methods: Traditional Channel State Information (CSI)-based beam selection suffers from excessive computational complexity, high signaling overhead, and poor adaptability to dynamic environments.
• Data Privacy and Latency: Centralized learning approaches in LEO networks incur prohibitive signaling overhead and latency, while raw data exchange raises privacy concerns.

The core challenge is to develop a scalable, intelligent, and lightweight beam selection mechanism that maximizes Signal-to-Noise Ratio (SNR) without relying on centralized data aggregation.

2. Methodology

The authors propose a Hierarchical Federated Learning (FL) framework tailored for LEO mega-constellations, utilizing High-Altitude Platform Stations (HAPS) as edge aggregators.

System Model

• Network Topology: A LEO constellation operating at the S-band (2 GHz) with altitudes between 1,015–1,325 km. The system includes multiple orbital planes, satellites equipped with $4\times4$ Uniform Planar Array (UPA) antennas, and static User Equipments (UEs) with small UPAs.
• Data Generation: A simulation over 2 hours generates 1,000 snapshots. For each snapshot, the system evaluates all candidate beams for visible links (where elevation angle $\theta > \theta_{min}$).
• Target Variable: The ground truth is determined via exhaustive search to find the beam $b^*$ that maximizes the received SNR, accounting for array gain, path loss, Rician fading, and shadowing.

Federated Learning Architecture

The framework is structured hierarchically to handle the distributed heterogeneity of orbital planes:

1. Local Level (Intra-Plane): Each orbital plane acts as a logical FL client. Satellites within a plane coordinate via a HAPS (acting as an edge aggregator) to train a local model using local data.
2. Global Level (Inter-Plane): After 200 local training epochs, model parameters are sent to a ground-based Network Control Center (NCC). The NCC performs Federated Averaging (FedAvg) to create a global model, which is then broadcast back to all clients.
3. Privacy & Efficiency: This approach ensures data locality, reduces communication overhead, and accommodates non-IID (Independent and Identically Distributed) data caused by distinct orbital geometries.

Machine Learning Models

Two architectures are evaluated for beam prediction:

• Multi-Layer Perceptron (MLP): A lightweight, fully connected network that evaluates candidate beams independently based on geometric features.
• Graph Neural Network (GNN): Models beams as graph nodes connected to neighbors. It uses graph convolutions to exchange information across beams, learning relative beam quality and local consistency, which is crucial for handling spatial correlations.

3. Key Contributions

• Novel FL Framework: Proposes a hierarchical FL architecture for LEO NTNs that leverages HAPS for intra-plane aggregation and ground stations for inter-plane coordination, addressing the unique constraints of satellite mobility and intermittent connectivity.
• Geometric-Aware Beam Prediction: Integrates geometric information (satellite/UE azimuth, elevation, and relative positions) directly into the ML input to enable robust beam selection without raw channel data exchange.
• Comparative Analysis of ML Architectures: Provides a rigorous comparison between MLP and GNN in the context of NTN beam management, demonstrating the superiority of graph-based approaches in capturing spatial relationships.

4. Performance Results

The models were evaluated on a shared test set using metrics including Top-1/Top-3 accuracy, stability, and model complexity.

Metric	MLP	GNN	Improvement
Top-1 Accuracy	88.41%	96.14%	+7.73%
Top-3 Accuracy	98.03%	99.52%	+1.49%
Mean Accuracy (Across Clients)	89.15%	95.97%	+6.82%
Training Time	6.27 sec	19.23 sec	(Higher for GNN)
Model Size	38.40 KB	71.43 KB	(Larger for GNN)

Key Findings:

• Superior Accuracy: The GNN significantly outperforms the MLP, achieving 96.14% Top-1 accuracy compared to 88.41%.
• Low Elevation Performance: The GNN demonstrates superior stability and accuracy at low elevation angles, where channel dynamics, fading, and shadowing are most severe.
• Beam Switching Stability: The GNN closely follows the "oracle" switching trend (ideal beam selection), whereas the MLP exhibits higher, less stable switching rates at low elevations.
• Feasibility: Despite the GNN's higher complexity (approx. 2x model size and 3x training time), both models remain lightweight (under 72 KB) and computationally feasible for NTN deployment.

5. Significance and Conclusion

This work validates that Federated Learning is a viable solution for scalable beam management in 6G LEO networks, effectively balancing privacy, latency, and performance.

• GNN Dominance: The study conclusively shows that modeling beam relationships as a graph (GNN) is critical for NTN environments. By capturing inter-beam dependencies, GNNs provide more robust predictions in dynamic, non-IID scenarios compared to independent evaluation methods (MLP).
• Practical Deployment: The framework proves that intelligent beam management can be achieved with minimal model size and training time, making it suitable for resource-constrained satellite hardware.
• Future Outlook: The authors suggest extending this framework to fully hierarchical architectures and incorporating interference-aware beam management with joint user-satellite optimization, paving the way for fully autonomous 6G NTNs.