Machine Learning Applications in Spacecraft State and Environment Estimation

Abstract

There are some problems in spacecraft systems engineering with highly non-linear characteristics and noise where traditional nonlinear estimation techniques fail to yield accurate results. In this thesis, we consider approaching two such problems using kernel methods in machine learning. First, we present a novel formulation and solution to orbit determination of spacecraft and spacecraft groups which can be applied with very weakly observable and highly noisy scenarios. We present a ground station network architecture that can perform orbit determination using Doppler-only observations over the network. Second, we present a machine learning solution to the spacecraft magnetic field interference cancellation problem using distributed magnetometers paving the way for space magnetometry with boom-less CubeSats.

We present an approach to orbit determination under very broad conditions that are satisfied for n-body problems. We show that domain generalization and distribution regression techniques can learn to estimate orbits of a group of satellites and identify individual satellites especially with prior understanding of correlations between orbits and provide asymptotic convergence conditions. The approach presented requires only observability of the dynamical system and visibility of the spacecraft and is particularly useful for autonomous spacecraft operations using low-cost ground stations or sensors. With the absence of linear region constraints in the proposed method, we are able to identify orbits that are 800 km apart and reduce orbit uncertainty by 92.5% to under 60 km with noisy Doppler-only measurements.

We present an architecture for collaborative orbit determination using networked ground stations. We focus on clusters of satellites deployed in low Earth orbit and measurements of their Doppler-shifted transmissions made by low-gain antenna systems in a software-defined federated ground station network. We develop a network architecture enabling scheduling and tracking with uncertain orbit information. For the proposed network, we also present scheduling and coordinated tracking algorithms for tracking with the purpose of generating measurements for orbit determination. We validate our algorithms and architecture with its application to high fidelity simulations of different networked orbit determination scenarios. We demonstrate how these low-cost ground stations can be used to provide accurate and timely orbital tracking information for large satellite deployments, which is something that remains a challenge for current tracking systems.

Last, we present a novel approach and algorithm to the problem of magnetic field interference cancellation of time-varying interference using distributed magnetometers and spacecraft telemetry with particular emphasis on the computational and power requirements of CubeSats. The spacecraft magnetic field interference cancellation problem involves estimation of noise when the number of interfering sources far exceed the number of sensors required to decouple the noise from the signal. The proposed approach models this as a contextual bandit learning problem and the proposed algorithm learns to identify the optimal low-noise combination of distributed magnetometers based on indirect information gained on spacecraft currents through telemetry. Experimental results based on on-orbit spacecraft telemetry shows a 50% reduction in interference compared to the best magnetometer.