Aerospace and Electronic Systems Magazine September 2017 - 12

Graceful Degradation and Recovery from GNSS Interruptions

Figure 10.

Time history of the position error (green square) and position blender's covariance (red triangle).

Finally, to ensure that the error traces shown in Figure 8 are
not overly optimistic, a covariance analysis can be performed. The
time history of position errors from Figure 10 and the one-sigma
bounds derived from the position blender's covariance matrix
are plotted in Figure 10. As expected, the errors grow during the
GPS outage but the covariance matrix appears to be conservative
bounds on this error.

HANDLING A GPS RECEIVER FAILURE
The second test data analyzed came from a flight in which an unplanned but fortuitous fault in the GPS receiver occurred. While
the exact cause of the failure has not been determined, the candidate culprits are either a loose antenna connection or an outdated
receiver firmware. Nevertheless, the fault provided useful data for
validating the decentralized architecture in an actual failure scenario.
The GPS failure manifested itself as large excursions in position and velocity estimates. This can be seen in Figure 11, which
shows the altitude measurements output by the GPS receiver during the flight test. At t = 44 s, a large and oscillatory altitude excursion signals the onset of the failure. Similar errors were present in
the other measurements, such as velocities, latitude, and longitude.
During this flight, the system was initialized and operated with
a functional GPS receiver the first 44 s. Between t = 44 and 132 s,
the GPS sensor output contained large errors. This period coincides
with the taxi, takeoff, and initial climb-out phase of flight. While
the reference GNSS/INS, which employs a limited fault detection
or isolation algorithm, was running in real time for this flight test,
the UAV was not using this solution for control. Instead, the UAV
12

Figure 11.

GPS sensor errors during a fault.

was being operated remotely by a human pilot using visual cues.
If the operation was autonomous, the FCS would have been using
this faulted solution and probably led to a loss of the vehicle. The
output of GNSS/INS filter is shown by the gray traces in Figure 12.
This solution diverged at t = 44 s and never converged.
The dark traces in Figure 12 show the performance of the decentralized filtering using blenders. The fault tolerance built into
the decentralized system was able to ignore the invalid measurements. Although no aid is provided until after the GPS sensor converges back to a stable state, the system built upon the decentralized
approach provides an accurate-enough solution for vehicle control.
The bird's eye view of the ground track of the UAV in Figure 13
shows a similar trend. This plot shows the solutions from each of
the position estimation filters during a 60-s segment of the test
flight. The non-fault-tolerant GNSS/INS provides an unstable position solution, as well as the unstable attitude solution shown in

IEEE A&E SYSTEMS MAGAZINE

SEPTEMBER 2017

Table of Contents for the Digital Edition of Aerospace and Electronic Systems Magazine September 2017