Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Federated Learning in Satellite Constellations

Bho Matthiesen, Nasrin Razmi, Israel Leyva-Mayorga, Armin Dekorsy, Petar Popovski

TL;DR

A software-defined management architecture for UxV networks to achieve wider coverage, reduced latency, and higher delivery ratio is proposed and preliminary experiment results confirm its superiority and potentials.

Abstract

Federated learning (FL) has recently emerged as a distributed machine learning paradigm for systems with limited and intermittent connectivity. This paper presents the new context brought to FL by satellite constellations, where the connectivity patterns are significantly different from the ones observed in conventional terrestrial FL. The focus is on large constellations in low Earth orbit (LEO), where each satellites participates in a data-driven FL task using a locally stored dataset. This scenario is motivated by the trend towards mega constellations of interconnected small satellites in LEO and the integration of artificial intelligence in satellites. We propose a classification of satellite FL based on the communication capabilities of the satellites, the constellation design, and the location of the parameter server. A comprehensive overview of the current state-of-the-art in this field is provided and the unique challenges and opportunities of satellite FL are discussed. Finally, we outline several open research directions for FL in satellite constellations and present some future perspectives on this topic.

Federated Learning in Satellite Constellations

TL;DR

A software-defined management architecture for UxV networks to achieve wider coverage, reduced latency, and higher delivery ratio is proposed and preliminary experiment results confirm its superiority and potentials.

Abstract

Federated learning (FL) has recently emerged as a distributed machine learning paradigm for systems with limited and intermittent connectivity. This paper presents the new context brought to FL by satellite constellations, where the connectivity patterns are significantly different from the ones observed in conventional terrestrial FL. The focus is on large constellations in low Earth orbit (LEO), where each satellites participates in a data-driven FL task using a locally stored dataset. This scenario is motivated by the trend towards mega constellations of interconnected small satellites in LEO and the integration of artificial intelligence in satellites. We propose a classification of satellite FL based on the communication capabilities of the satellites, the constellation design, and the location of the parameter server. A comprehensive overview of the current state-of-the-art in this field is provided and the unique challenges and opportunities of satellite FL are discussed. Finally, we outline several open research directions for FL in satellite constellations and present some future perspectives on this topic.
Paper Structure (13 sections, 6 figures)

This paper contains 13 sections, 6 figures.

Table of Contents

  1. Introduction
  2. Fundamentals
  3. A Classification of Satellite FL Scenarios
  4. Sporadic Direct Connection to PS
  5. Near-Persistent Direct Connection to PS
  6. Multi-hop Connection to PS via Inter-Cluster Connectivity
  7. Conclusions

Figures (6)

  • Figure 1: FL in a satellite constellation (a) without and (b) with ISLs. In the former case, the satellites must wait until the next pass to send their local update and to receive the updated global model. In the latter case, the satellites can update the models if at least one of them has connection to the PS. In both cases, the connectivity of the satellites to the PS is predictable.
  • Figure 2: Illustration of a 80°: 40/5/1 Walker star and a 60°: 40/5/1 Walker delta constellation, both at an altitude of 500 km. The frontal view is for an observer in the equatorial plane at 0° longitude and the top view from the polar plane towards the North pole with 0° longitude pointing down.
  • Figure 3: Connectivity pattern from satellites in a Walker star constellation with five orbits at 80° inclination and 500 km altitude towards a GS in Bremen, Germany, and a satellite in polar orbit at 2000 km altitude, respectively. In addition, cluster connectivity towards the GS is shown for a 80°: 40/5/1 Walker star constellation, that is, the same constellation as before but with eight satellites per orbit. Connectivity of individual satellites within each orbital plane towards the GS is shown in grey below the cluster connectivity. In all cases, the connectivity towards the out-of-constellation node is sporadic.
  • Figure 4: Connectivity pattern from five client clusters towards a satellite in equatorial orbit at 2000 km altitude. Each client cluster corresponds to an orbital plane in a 80°: 40/5/1 Walker star constellation where satellites are equipped with intra-orbit ISLs. Connectivity of individual satellites within each orbital plane is shown in grey below the cluster connectivity. While the cluster connectivity towards the out-of-constellation satellite is close to persistent, the connectivity of the individual satellites is sporadic.
  • Figure 5: Spatio-temporal graph models for FL in satellite constellations for different connectivity patterns. Three snapshots of the time graph are displayed for each scenario. Client clusters are denoted as $\mathcal{C}_i$, while the PS is labeled as $\mathcal{P}$.
  • ...and 1 more figures