CRIPPLED combat aircraft should soon stand a better chance of making it back
to base, thanks to a novel neural network that has just passed its first major
test.
When the control surfaces of an aeroplane, such as the rudder or ailerons,
are damaged or malfunctioning there is often very little a pilot can do to avert
a disaster. But in a series of tests that finished last week, NASA engineers
have shown that smart software can keep aircraft flying even if some of their
control surfaces are disabled.
The tests, at NASA鈥檚 Dryden Flight Research Center in Edwards, California,
used a modified F-15 aircraft similar to those flying combat missions against
Serbia. The software was tested in high performance manoeuvres, such as tracking
a target or performing a 360 degree roll. The neural net managed to keep
disabled planes under control even at supersonic speeds. This is the first time
such a system has helped control a piloted aircraft.
Advertisement
The software was developed at NASA鈥檚 Ames Research Center in Moffett Field,
California. It is designed to kick in when the flight control system detects a
mismatch between data on the plane鈥檚 airspeed, bearing and the forces on its
wings, tail and fuselage, and a computer model showing what data should be
received if the plane were flying normally.
Neural networks are programs that learn to perform a certain task by trial
and error, rather like the human brain. The Ames team trained theirs to control
an F-15 on a flight simulator, before letting it loose on the real thing. In an
emergency, the network reassesses flight data six times every second to work out
the best way of using the available control surfaces and the plane鈥檚 engines to
maintain normal flight.
The pilot would be warned of any failures, such as the loss of his ailerons.
But in most cases, he shouldn鈥檛 notice that the plane is handling differently
from normal. 鈥淭he computer would determine that the ailerons aren鈥檛 available
and would have another scheme in mind,鈥� says Mike Thomson, the engineer at
Dryden who supervised the tests. 鈥淚t would learn to use another control surface
颈苍蝉迟别补诲.鈥�
Thomson believes commercial airliners could also benefit from the system,
although it would be restricted to modern fly-by-wire planes. When an airliner
loses its control surfaces, the results can be horrific: at Sioux City, Iowa, in
1989, a full hydraulic failure left an aircrew trying to land a DC-10 using only
the throttles of its two remaining engines. It crash-landed, killing 110
passengers.
However, civil authorities may be less willing than the military to embrace
the system. They will demand standardised tests of the system鈥檚 performance,
says Tom Anderson, an expert on software reliability at Newcastle University.
But such tests will be hard to devise, as the performance of neural networks is
inherently difficult to predict.