Refereed Conference Meeting Proceeding
It is impossible to pre-define a controller for every fault an autonomous system can experience as some faults are unknown at design time. Current fault tolerant control (FTC) architectures switch control to a pre-defined fault controller when a known fault is identified. Blended control implements a controller that is composed of multiple individual controllers instead of discretely switching between them. In this article we present a novel fault tolerant control architecture based on blended control that uses a high-level deep learning agent to learn the optimal blending proportions between low-level controllers for unknown faults. Faults are abstracted to the effect they have on the performance of a task while removing the inherent fault identification delays experienced by existing FTC architectures. The presented architecture is validated on a quadcopter trajectory tracking task and trained to tolerate abrupt rotor loss of effectiveness. We compare our approach against a switched architecture with the same underlying controllers and show its ability to learn unknown fault tolerance.
Principles of Diagnostics
Digital Object Identifer (DOI):
National University of Ireland, Cork (UCC)
Open access repository: