Image: Christine Daniloff |
The
Federal Aviation Administration (FAA) has mandated that by 2020, all commercial
aircraft must be equipped with a new tracking system that broadcasts GPS data,
providing more accurate location information than ground-based radar. In
anticipation of the deadline, the FAA has also charged Massachusetts Institute of Technology (MIT) researchers with
leading an investigation of the system’s limits and capacities.
At
the 30th Digital Avionics Systems Conference, MIT researchers will present an
early result of that investigation, a new algorithm that uses data from the
tracking system to predict and prevent collisions between small aircraft. In
the last 10 years alone, 112 small planes have been involved in midair
collisions, and thousands more have reported close calls.
The
chief challenge in designing a collision-detection algorithm, says Maxime
Gariel, a postdoc in MIT’s International
Center for Air Transportation
and lead author on the paper, is limiting false alarms. “If half the time it’s
a false alert,” Gariel says, “[people] are not going to listen to it, or
they’ll turn it off.” At the same time, the algorithm has to have some room for
error: While GPS is more accurate than radar tracking, it’s not perfect; nor are
the communications channels that planes would use to exchange location
information. Moreover, any prediction of a plane’s future position can be
thrown off by unexpected changes of trajectory.
Puckish predictions
Much of the work on the new algorithm involved optimizing the trade-off between
error tolerance and false alarms. Gariel and his collaborators—John Hansman,
the T. Wilson (1953) Professor of Aeronautics and Astronautics and Engineering
Systems, and Emilio Frazzoli, an associate professor of aeronautics and
astronautics—adopted a two-tiered system of alerts: A moderate alert would warn
pilots that their trajectories are converging, and a high alert would indicate
a severe risk of collision.
Associated
with each alert is a volume of space around each plane, which Gariel describes
as a “hockey puck,” that describes the plane’s probable position given a
certain GPS reading. (The volume is puck-shaped because there’s less likelihood
of error in the vertical direction than in any horizontal direction.) The
hockey puck that corresponds to the high alert is smaller and of fixed size.
The hockey puck that corresponds to the moderate alert is larger and fluctuates
according to planes’ trajectories.
For
instance, if two planes are headed in the same direction, their moderate-alert
hockey pucks are relatively small; but if they’re headed toward each other,
their hockey pucks are larger, since they’ll have much less time to react to an
impending collision. If an extrapolation from two planes’ recent trajectories
suggests that either set of hockey pucks will intersect, the system issues the
corresponding alert.
To
calculate the optimal puck sizes, Gariel used six months’ worth of data from
airports in the San Francisco
area. But in testing the algorithm’s utility, the researchers had the advantage
of a very accurate computer model of air traffic created by researchers at
MIT’s Lincoln Laboratory. Based on more than eight months of data from all the
aviation radar systems in the United
States, the Lincoln Lab model generates
random trajectories for hypothetical aircraft that accord very well with
real-world statistics. Working together with Fabrice Kunzi, a graduate student
in Hansman’s group, Gariel and his colleagues tested their algorithm against
the Lincoln Lab model and found that, indeed, it had a low false-alarm rate.
Model behavior
David Gray, the FAA’s lead on the project, explains that while the agency will
require small aircraft to broadcast their GPS coordinates by 2020, it hasn’t
yet mandated that they install equipment for receiving and processing such
broadcasts. “One of the key things that we want to provide as part of this
system is additional value to the general-aviation [small-plane] pilot,” Gray
says. “We hope it adds value and tips the scale in the direction of saying, ‘Yes, this is something that I want.'”
Gray
has not yet had the opportunity to review the MIT researchers’ results in
detail, but says that “from the limited data I’ve seen, it seems that the
algorithms that they’re looking at are performing better than the algorithms
that are in existing systems that can be bought today.” He points out, however,
that the Lincoln Lab air-traffic model is based on radar data, and that small
planes often fly below radar—particularly near airports, where nearly 60% of
midair collisions take place. “They’re using the model for the scenarios that
it’s applicable for,” Gray says, “and I think that’s going to be great. But for
the scenarios that it’s not applicable for, they’re going to have to develop
other scenarios for us to assess.”
Indeed,
Gariel and Kunzi are working to develop a new computer model that takes into
account the standard flight paths that small aircraft tend to fall into near
airports, to see if the collision-detection algorithm still performs as well.
They’re also hoping to begin testing the algorithm on real planes.