Deep-Learning Based Multiple-Model Bayesian Architecture for Spacecraft Fault Estimation

Abstract

This thesis presents recent findings regarding the performance of an intelligent architecture designed for spacecraft fault estimation. The approach incorporates a collection of systematically organized autoencoders within a Bayesian framework, enabling early detection and classification of various spacecraft faults such as reaction-wheel damage, sensor faults, and power system degradation. To assess the effectiveness of this architecture, a range of performance metrics is employed. Through extensive numerical simulations and in-lab experimental testing utilizing a dedicated spacecraft testbed, the capabilities and accuracy of the proposed intelligent architecture are analyzed. These evaluations provide valuable insights into the architecture’s ability to detect and classify different types of faults in a spacecraft system. The study has successfully implemented an intelligent architecture for detecting and classifying faults in spacecraft. The architecture was analyzed through numerical simulations and experimental tests, demonstrating enhanced early detection capabilities. The incorporation of autoencoders and Bayesian methods proved to be a powerful combination, allowing the architecture to effectively capture and learn from complex spacecraft system dynamics and detect various types of faults. This research presents an advanced and reliable approach to early fault detection and classification in spacecraft systems, highlighting the potential of the intelligent architecture and paving the way for future developments in the field.