Graph Neural Networks Empower UAV Networks: Link Quality Prediction and Intelligent Routing Support System

The team from Vietnam National University, Ho Chi Minh City proposes a Graph Neural Network (GNN)-based link quality prediction scheme for UAV networks. By jointly modeling network topology, node features, and link features, it achieves intelligent routing decision support in dynamic topology environments. This scheme aims to address problems such as drastic topology changes and unstable link quality in UAV ad-hoc networks caused by high-speed node movement, providing forward-looking guidance for routing protocols and improving network transmission performance.

UAV ad-hoc networks have great potential in emergency rescue, environmental monitoring, military communications, and other fields. However, high-speed 3D movement of nodes leads to drastic topology changes, bringing unique challenges:

1. Wireless link quality is prone to attenuation and interruption; traditional routing protocols (e.g., AODV, OLSR) rely on local information and struggle to capture global spatiotemporal patterns;
2. 3D movement causes link quality to be affected by multiple factors such as multipath effects, Doppler shift, and occlusion, showing nonlinearity and time-varying characteristics. Predicting the trend of link quality changes in advance has become the key to improving transmission performance.

The core innovation lies in introducing GNNs into link prediction and jointly modeling three types of information:

1. Network Topology: Abstracted as a dynamic graph where UAVs are vertices and communication links are edges. GNN message passing aggregates neighbor information to achieve multi-hop topology awareness;
2. Node Features: 3D coordinates, flight speed, acceleration, remaining battery, etc., are mapped to high-dimensional vectors via an embedding layer, reflecting node status and contribution to connectivity;
3. Link Features: Edge features such as signal strength, signal-to-noise ratio (SNR), delay, and packet loss rate are combined with node features to learn complex mapping relationships.

The system architecture forms a closed loop from data collection to routing decision:

• Perception Layer: Collects GPS, IMU, and RF data via sensors/communication modules, preprocesses and synchronizes them to construct a time-series graph input;
• Prediction Layer: Uses temporal GNNs, combining graph convolution, edge feature fusion, and gated temporal updates to output the link quality score for the next moment;
• Decision Layer: Converts prediction results into signals understandable by routing protocols (link cost, availability probability, routing suggestions) and seamlessly integrates with mainstream protocols.

Compared to traditional routing methods, this scheme achieves a paradigm shift:

• Predictive Advantage: Identifies the downward trend of link quality in advance, reserves adjustment time windows, and avoids packet loss/delay;
• Global Awareness: GNN message passing allows nodes to obtain multi-hop network status, overcoming local topology limitations;
• Adaptive Learning: End-to-end training requires no manual feature engineering, can be fine-tuned for different scenarios, and achieves scenario adaptability.

Application scenarios are wide-ranging:

• Emergency Rescue: Improves the communication reliability of UAV swarms in disaster scenarios;
• Logistics Delivery: Helps avoid urban communication blind spots;
• Military Tactics: Enhances real-time communication and anti-destruction capabilities in high-mobility scenarios;
• Aerial Base Stations: Improves the backhaul link quality of 5G/6G space-air-ground integrated networks.

This project demonstrates the advantages of GNNs in solving the dynamic challenges of UAV networks. It achieves accurate link quality prediction by uniformly modeling dynamic graph structures and converts it into routing support signals. This 'prediction-decision' linkage idea provides a reference for the design of intelligent UAV networks. As applications expand, AI-empowered network technologies will play an increasingly important role.