Date of Award:

5-2015

Document Type:

Thesis

Degree Name:

Master of Science (MS)

Department:

Mechanical and Aerospace Engineering

Committee Chair(s)

Rees Fullmer

Committee

Rees Fullmer

Abstract

CubeSats are employed in a variety of missions as scientific platforms, low-cost technology demonstrators, and in the future, they will conduct service missions as part of larger satellite constellations. CubeSat ADCS designers today are being tasked with designing to ever increasing accuracy requirements. The ability to hold those requirements rests with the attitude control system being robust enough and able to sufficiently respond to changes in the control environment. The need for greater control autonomy is then evident in the need for these systems to be able to react independently to changes in the system dynamics and to identify and accommodate system faults. A key ability of a fault tolerant control system then is the capability to successfully detect and isolate faults as it provides a method for early detection and diagnosis of unforeseen faults, which in turn provides a spacecraft with a greater degree of autonomy.

The focus of this thesis was to determine whether there is enough information coming from a pattern of residuals defined by the attitude determination and control system to allow a neural network to discern whether or not a fault has occurred. To do this, a set of residuals was defined based on a comparison of a spacecraft's ADCS telemetry to estimated state values for both nominal and fault states. This set of residuals served as the training set for a series of neural networks that were trained using either the resilient back-propagation, Levenberg-Marquardt, or Levenberg-Marquardt with Bayesian regularization training algorithms. After the networks were trained their outputs were calculated for both reapplication of the training data as well as for novel data of which they had no a priori knowledge.

The performance of the neural networks in detecting fault with this scheme leaves much to interpretation. Though all of the networks were trained from the same example set, significant differences exist in the ability of the networks to positively detect and isolate the faults with consistency. Where one network may excel in detecting the faults in a certain components, it may fare poorly at another. In general, the networks were better able to detect and isolate faults in the components of the attitude control and guidance subsystems, and with few exceptions less able to isolate faults of the attitude determination sensors.

Checksum

c68711e6d750c0ae9e84701106fc6876

Share

COinS