Original title: Adaptive Beam Hopping and Power Control for Dual-Layer Over-the-Air Online Federated Learning in LEO Satellite Networks
Zhendong Li, Shaojie Wang, Zhou Su, Zihao Zhang +4 · September 2026
Training artificial intelligence directly across low-Earth orbit satellites and ground devices is challenging. Devices constantly gather new data that quickly goes stale if not used right away. Satellites also have strict limits on radio beams and power. To fix this, researchers designed a dual-layer system using over-the-air computation, a method where signals combine naturally as radio waves travel through the air. First, ground devices beam their local model updates directly to passing satellites. Next, those satellites send the combined signals to a central processing station. Because satellites constantly move and data arrives continuously, deciding where to point beams and how much power to use is tough. The team built an artificial intelligence control framework using deep reinforcement learning to make these split-second choices. Their method outperformed other benchmark strategies, achieving higher long-term data utilization and faster learning convergence while keeping signal distortion within required limits.
Imagine managing traffic in a city where construction crews are constantly moving, blocking different streets. Each driver has a new destination, and if they can't get through quickly, their trip becomes outdated. In this paper, ground devices are like drivers sending updates, satellites are like moving construction crews that can block or allow data flow, and the system must choose which satellite to connect to and how strongly to transmit to keep the learning process going without too much delay or error.
This helps ground devices in remote or moving locations get faster, more accurate AI updates from satellite networks. It affects anyone relying on real-time data processing where traditional fixed connections fail.
The abstract indicates results are from numerical simulations, so real-world performance in actual satellite systems is not proven. The work is specific to dual-layer OTA aggregation in LEO networks and does not address scalability to larger networks or exhaustive comparisons with all alternative methods.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
Adaptive Beam Hopping and Power Control for Dual-Layer Over-the-Air Online Federated Learning in LEO Satellite Networks — Zhendong Li, Shaojie Wang, Zhou Su et al.
Abstract licence: CC0
Explanation generated by Synthesis from the abstract — it is not the original paper. This is a preprint, so it hasn't necessarily been peer reviewed.