NASA HAS TWO EXPERIMENTAL PROGRAMS under way using neural net software to help control damaged planes as well as help them land. This generic control software may eventually be used in many commercial industries.
Neural net software learns by observing pairs of related patterns in the real world and learns to perform different tasks in response to different patterns: it discovers relationships by observation. A neural network consists of many interconnected processors (or nodes) using computational principles derived from neurons in the brain. Each node assigns a value to the input from each of its counterparts. As these values change, the network can adjust the way it responds.
The new software, under joint development by scientists at NASA's Ames Research Center, in Moffett Field, California, and McDonnell Douglas Corporation in St. Louis, Missouri, may allow a damaged aircraft's computer to "relearn" to fly and to land a plane safely after a major equipment failure or explosion. Airplane sensors can send speed, direction and force data to the computer program. The aircraft's computer compares the pattern of what is actually happening to the aircraft with a pattern showing how the airplane should fly. If there is a mismatch, the neural net, which learned the core of a dozen basic aeronautical equations that define how airplanes fly, finds a way to make the plane work with a new pattern, if mechanically possible.
"If sensor data show that a pilot command is being improperly responded to, and the airplane is turning too abruptly, the airplane's neural network can rapidly learn to assist the pilot in use of the stick, engines, flaps, rudders and other control surfaces in ways that may be very unconventional, but possibly successful," said Ames computer scientist Charles (Chuck) Jorgensen.
Preliminary tests used an earlier version of the new software installed in a modified F-15 jet fighter. In the next phase, Ames and McDonnell Douglas will validate the software in very high fidelity simulations. After the simulation test, the software will be flight-tested in ACTIVE, a specially modified F-15 at NASA's Dryden Flight Research Center in Edwards, California, in 1997.
"Once we prove neural net software can rapidly learn to fly a crippled jet fighter and help pilots land it safely, then engineers will be more likely to use the intelligent software in power plants, automobiles and other less complicated systems to avoid disasters after equipment failures," Jorgensen said.
Another project using neural net software is the Low-Observable Flight Test Experiment (LoFLYTE), an 8-foot by 4-inch aircraft, developed by Accurate Automation Corporation in Chattanooga, Tennessee, for NASA and the Air Force. The model is a Mach 5 waverider design-a futuristic hypersonic aircraft configuration that actually cruises on top of its own shockwave. LoFLYTE represents the first known flying waverider vehicle configurationThe remotely piloted aircraft has been designed to demonstrate that neural network flight controls are superior to conventional flight controls. The aircraft's flight controller consists of a network of multiple-instruction, multiple-data neural chips. The network will be able to continually alter the aircraft's control laws to optimize flight performance and take the pilot's responses into consideration. Over time, the neural network system could be trained to control the aircraft. The use of neural networks in flight would help pilots fly in quick-decision situations and help damaged aircraft land safely. Using this type of control system at high speeds can be a big advantage because things happen so quickly that the pilot cannot control the aircraft as easily as at subsonic speeds. In addition to demonstrating a flight control system that learns, the flight of the model also is key as a low-speed demonstration of a hypersonic vehicle. "We're very interested in both outcomes, both the neural net technology and the flight characteristics," said Robert Pegg of the Hypersonic Vehicles Office at NASA's Langley Research Center in Hampton, Virginia.
The waverider was chosen as the testbed for the neural networks because the configuration has an inherently high hypersonic lift-to-drag ratio. If neural networks can control this "worst-case scenario" configuration, then they should be able to handle virtually any other configuration. The waverider configuration also was chosen because it allows for long hypersonic cruise ranges of up to 8,000 miles. At an altitude of 90,000 feet, a Mach 5 waverider would fly at a rate of one mile per second.
LoFLYTE is a project under the Small Business Innovative Research (SBIR) program, administered through Langley and the Air Force Wright Laboratory in Dayton, Ohio. Accurate Automation Corporation won the contract to integrate the neural network technology into the Langley design
Neural net software learns by observing pairs of related patterns in the real world and learns to perform different tasks in response to different patterns: it discovers relationships by observation. A neural network consists of many interconnected processors (or nodes) using computational principles derived from neurons in the brain. Each node assigns a value to the input from each of its counterparts. As these values change, the network can adjust the way it responds.
The new software, under joint development by scientists at NASA's Ames Research Center, in Moffett Field, California, and McDonnell Douglas Corporation in St. Louis, Missouri, may allow a damaged aircraft's computer to "relearn" to fly and to land a plane safely after a major equipment failure or explosion. Airplane sensors can send speed, direction and force data to the computer program. The aircraft's computer compares the pattern of what is actually happening to the aircraft with a pattern showing how the airplane should fly. If there is a mismatch, the neural net, which learned the core of a dozen basic aeronautical equations that define how airplanes fly, finds a way to make the plane work with a new pattern, if mechanically possible.
"If sensor data show that a pilot command is being improperly responded to, and the airplane is turning too abruptly, the airplane's neural network can rapidly learn to assist the pilot in use of the stick, engines, flaps, rudders and other control surfaces in ways that may be very unconventional, but possibly successful," said Ames computer scientist Charles (Chuck) Jorgensen.
Preliminary tests used an earlier version of the new software installed in a modified F-15 jet fighter. In the next phase, Ames and McDonnell Douglas will validate the software in very high fidelity simulations. After the simulation test, the software will be flight-tested in ACTIVE, a specially modified F-15 at NASA's Dryden Flight Research Center in Edwards, California, in 1997.
"Once we prove neural net software can rapidly learn to fly a crippled jet fighter and help pilots land it safely, then engineers will be more likely to use the intelligent software in power plants, automobiles and other less complicated systems to avoid disasters after equipment failures," Jorgensen said.
Another project using neural net software is the Low-Observable Flight Test Experiment (LoFLYTE), an 8-foot by 4-inch aircraft, developed by Accurate Automation Corporation in Chattanooga, Tennessee, for NASA and the Air Force. The model is a Mach 5 waverider design-a futuristic hypersonic aircraft configuration that actually cruises on top of its own shockwave. LoFLYTE represents the first known flying waverider vehicle configurationThe remotely piloted aircraft has been designed to demonstrate that neural network flight controls are superior to conventional flight controls. The aircraft's flight controller consists of a network of multiple-instruction, multiple-data neural chips. The network will be able to continually alter the aircraft's control laws to optimize flight performance and take the pilot's responses into consideration. Over time, the neural network system could be trained to control the aircraft. The use of neural networks in flight would help pilots fly in quick-decision situations and help damaged aircraft land safely. Using this type of control system at high speeds can be a big advantage because things happen so quickly that the pilot cannot control the aircraft as easily as at subsonic speeds. In addition to demonstrating a flight control system that learns, the flight of the model also is key as a low-speed demonstration of a hypersonic vehicle. "We're very interested in both outcomes, both the neural net technology and the flight characteristics," said Robert Pegg of the Hypersonic Vehicles Office at NASA's Langley Research Center in Hampton, Virginia.
The waverider was chosen as the testbed for the neural networks because the configuration has an inherently high hypersonic lift-to-drag ratio. If neural networks can control this "worst-case scenario" configuration, then they should be able to handle virtually any other configuration. The waverider configuration also was chosen because it allows for long hypersonic cruise ranges of up to 8,000 miles. At an altitude of 90,000 feet, a Mach 5 waverider would fly at a rate of one mile per second.
LoFLYTE is a project under the Small Business Innovative Research (SBIR) program, administered through Langley and the Air Force Wright Laboratory in Dayton, Ohio. Accurate Automation Corporation won the contract to integrate the neural network technology into the Langley design
