@phdthesis {pub2577,
	title = {Data-driven Fault Detection for Component Based Robotic Systems},
	author = {Raphael Golombek},
	year = {2014},
	month = {February},
	abstract = {Advancements in the field of robotics enable the creation of systems with cognitive
abilities which are capable of close interaction with humans in real world scenarios.
These systems may take over jobs previously executed by humans like house cleaning
and cooking or they can be supportive and act as an elderly caretaker. One
consequence of this progress is the increased need for dependable and fault tolerant
behavior of today{\textquoteright}s robotic systems because they share the same spaces with
humans and operate in close proximity to them. Undependable and fault behavior
may frustrate users or even endanger them resulting in poor acceptance of robotic
systems.
The contribution of this thesis is a fault detection approach called AuCom. Fault detection
is a basis element for fault tolerant system behavior which is the ability of a system
to autonomously cope with occurring faults while it is engaged in interaction. The approach
is designed to tackle the specific needs of cognitive robotic systems which
feature a component based hardware and software structure and are characterized
by frequent changes due to research and development efforts as well as uncertain
and variant behavior resulting from the interaction in real world environments.
The solution presented in this thesis belongs to the class of data-driven faultdetection
approaches. These type of approaches assume that fault relevant information can be
directly derived from data gathered in the robotic system. The data exploited in this
work for fault detection is the communication between the system{\textquoteright}s components and it
is represented with features which are common to all elements of the communication
(i.e., they are generic). In particular, the approach assumes that the current element
of the communication can be estimated from history and that a deviation from the
expected estimation indicates a fault. This assumption is encoded in the model terms
of a novel representation of the communication as a time-series of temporal dynamic
features.
A concrete integration of the approach into a real system is exemplified on our robotic
platform BIRON. In addition, exemplary integration solutions for robotic frameworks
currently prominent in literature are discussed in this thesis. The actual capability of
the approach to report faults is evaluated for several artificial systems in simulation
and on BIRON in an off-line and on-line manner. The performance is compared to a
histogram-based baseline approach.},
	publisher = {University of Bielefeld},
	booktitle = {University of Bielefeld}
}