Decentralized estimator for dynamical multi-agent systems with an application to LEO satellite formations

João Marafuz Gaspar

Key information

Authors:

João Marafuz Gaspar (João Marafuz Gaspar)

Supervisors:

João Pedro Gomes (João Pedro Castilho Pereira Santos Gomes); João Xavier (João Manuel de Freitas Xavier)

Published in

07/01/2025

Abstract

The emergence of satellite formations has intensified the demand for scalable, accurate, and resource-efficient onboard localization techniques. Traditional centralized orbit estimation frameworks suffer from severe scalability limitations due to communication bottlenecks, computational burden, and reduced fault tolerance. This thesis presents a novel decentralized estimation framework based on moving horizon estimation, tailored for dynamical multi-agent systems with application to LEO satellite formations. The estimation problem is posed for networks with decoupled nonlinear deterministic dynamics and arbitrarily coupled stochastic outputs. To solve the estimation problem, a second-order optimization algorithm based on Newton’s method is developed by exploiting the structure of the global cost function, whose Hessian matrix enables exact computation of the Newton direction in a decentralized fashion via a structured message-passing protocol over tree-structured networks. This formulation permits fully decentralized implementation using only local communication, with memory, computational, and communication requirements that do not scale with the dimension of the network, and with each agent estimating only its state exclusively. For tree-structured output coupling topologies, the Newton direction is exactly decoupled across the network. For general output coupling topologies, several algorithmic variants are introduced, including majorization-minimization and preconditioned conjugate gradient methods, preserving decentralization while approximating the Newton step. The methodology is validated through extensive simulations using a custom orbital dynamics simulator under varying observation models and time-invariant output coupling topologies. The proposed estimators are benchmarked against centralized and consensus-based frameworks. Results confirm that the decentralized Newton-based framework provides superior scalability and competitive estimation accuracy, particularly when enhanced with bearing measurements.