Abstract:
UAS(Unmanned Aerial Systems) mishaps are high, and their pilots face many
control challenges. The reliability of UAS has been seen as a dominant mishap
cause but in several instances the aircraft could have been saved if the health state
of the aircraft had been understood at an earlier point by the pilot. Manned and
unmanned aircraft pilots both benefit from the use of their own experience in the
detection and mitigation of faults during flight. However it has been suggested that
pilots within a GCS(Ground Control Station) face difficulties in maintaining their
situational awareness due to the nature of remote control. The use of a cognitive
framework as a basis for case based reasoning is suggested as a way to integrate
through life learning into the Safety Management System. The population of the
case base for such a system would require a large investment of time to create.
The use of machine learning is suggested and evaluated to address this issue by
generating cases for CBR. This has seen some success and even the use of an
AIS(Artificial Immune System) in this thesis. An AIS was used in order to try
to address the problem of cost and time caused by high pre-processing required
by common machine learning methods. A simulation of the Aerosonde UAS was
created and multiple flights simulated to build up a set of representative set flight
data. Several fault cases were included in the simulated flights of varying severities.
Different machine learning schemes were evaluated using the data set and their
effectiveness compared in order to evaluate the ability of the algorithm to learn
from flight data without extensive pre-processing. The complex dataset made the
problem difficult but in analysis the AIS performed slightly better than the neural
network with which it was compared. In due time and with development it's
computational cost could be reduced and its effectiveness increased. The benefit
of an automated method to learn from aircraft incidents and mishaps can readily
be seen in a fleet scenario where it would be uneconomical to analyse flight data
of unmanned aircraft in the same way that it would be done for manned aircraft.
This semi-supervised approach reduces personnel requirements and enhances the
ability of operators to learn from mishaps by relating mishap cases to the current
situation and being transparent in their alerting criteria.