|
|
Better than Human Flight Control Systems
Re-engineered neural networks are dramatically changing By Rick Robinson |
It's reassuring to know the jetliner you're boarding is equipped with numerous automatic systems, which support the big machine itself as well the pilots that fly it. Such systems allow the plane to fly unattended at times, and they kick in quickly if a wind blast should upset the airplane or the pilot should fly too close to the terrain or other traffic.
|
But who, or what, is watching the automatic systems? What happens if an automatic system, or any of the vital flight surfaces controlled by such systems, alters or fails at a critical moment, requiring major adjustments throughout the plane's flight-control system?
In the future, the answer may be "don't worry, the neural network will take care of it." Much like the human brain, artificial neural networks consist of collections of processing elements that are highly interconnected. Each collection transforms a set of inputs to a set of desired outputs.
Artificial neural networks, already used for such applications as pattern recognition and process control, are being re-engineered by Georgia Institute of Technology scientists. The redesigned networks will perform a host of monitoring and adjustment functions in aircraft — what those in the field call "adaptive control."
"We're pursuing adaptive control using neural networks on a whole spectrum of aircraft," says Dr. Anthony J. Calise, professor of
aerospace engineering at Georgia Tech.
Calise and his colleagues are engaged in seven aircraft-related, neural-network projects, funded variously by NASA, the Air Force, industry sponsors and Georgia Tech. Fighter and tiltrotor aircraft, which can be tricky to control and need complex flight-control systems, have been the first candidates fitted with neural networks. But commercial aircraft such as jetliners are likely to benefit, too, Calise says, as are guided missiles, guided munitions and almost anything else that flies.
Meanwhile, some of these projects are already well advanced. Calise, his colleague Dr. J.V.R. Prasad and their Georgia Tech research team — working closely with Boeing Phantom Works — installed a neural network onboard an experimental aircraft, the X-36, a one-fifth-size unmanned jet research aircraft that simulates conditions on a full-size vehicle. The X-36, developed under NASA auspices by Boeing, was tested recently at Edwards Air Force Base, and its neural network functioned quite successfully, Calise says. But the tests had to be cut short because of hardware problems.
And in Georgia Tech's Uninhabited Aerial Vehicle Lab, Calise, post-doctoral researcher Dr. Rolf Rysdyk and a team of graduate research assistants are working with Guided Systems Technology Inc. to test a neural network installed in a remotely piloted helicopter.
Similar, But Better
This process of pilot recovery is made more tricky by the complexities of modern flight control systems. Many of today's aircraft, including military jets, have intrinsically unstable designs and cannot be flown without so-called "feedback systems" — highly sophisticated flight control systems that can sense aircraft performance and angular rates and then feed back a corrective signal that automatically helps stabilize the vehicle.
Neural networks may be able to dramatically alter this process. By focusing on two main areas — how a vehicle flies and the potential failure scenarios for that vehicle — researchers are trying to train the neural network to deal with any system failure.
"A neural network is working all the time," Calise says. "When something goes wrong, it adapts just like a human being might adapt to a failure." But, he adds, with one potentially crucial difference: The neural network has the potential to deal with failure more quickly and accurately than any human pilot.
In fact, ideally the recovery process would happen so rapidly it would be virtually transparent to the pilot. A failure would occur, perhaps accompanied by minor airframe bumps, but the plane would continue to fly and respond to commands normally. In the critical first few seconds after the failure, the pilot would probably know something had happened, but not what. Only when detailed reports began to filter in would the nature of the failure become clear.
The neural network system is sufficiently sophisticated that it can even deal with damage to the airframe, Calise says. "The neural network is there to maintain the handling qualities of a well-functioning flight control system — even if you get part of a wing shot off, which is just one of the failures we've mimicked in simulation."
Imitating the Brain
Neural neworks are typically organized in layers. Layers are made up of a number of interconnected "nodes" that contain an "activation function." Patterns are presented to the network via the "input layer," which communicates to one or more "hidden layers" where the actual processing is done via a system of weighted "connections." The hidden layers then link to an "output layer" where the answer is output as shown in this graphic.
(Click on the image to see larger 38k version).
Neural networks, too, have "neurons" that receive signals from other neurons, process them and then output a signal. These "neurons" are units of computer code that are mathematical representations of a human neuron's input-output functions. They are linked together by connections called "weights"; changes in these "weights" allow the network to learn by recognizing patterns or configurations.
One well-known approach to neural network training involves "back propagation," used for such applications as pattern recognition. In back propagation, the system is programmed to learn from its errors; repeated exposure to different objects and patterns is combined with retraining of the weights whenever there is a mistake. This approach is analogous to the human experience of avoiding the same mistake, thus learning from experience.
But back propagation involves many computations and experiments, Calise notes. Much of the work is done off line in a tedious, time-consuming process, then put on line when the system works correctly.
"In flight control we don't have that luxury," Calise says. Such a system must learn on line in real time as it is exposed to new environments - and be ready to deal with unpredictable failures.
To train aircraft networks in real time, the Georgia Tech team has had to bring new tools to bear. Even formulations used in other cutting-edge neural network research, such as robotics applications and automation, don't apply to aircraft neural networks.
A Software Add-on
"That ability, to add on rather than replace, has been a key to our being able to get to where we are today," Calise says. This integration of neural network function with flight control promises to improve aircraft design and testing, along with enhancing flight and flight safety.
Today, flight control systems are designed for many different operating conditions. As an aircraft moves from one operating condition to another, a flight control system must be adjusted to compensate — a process called "gain scheduling."
In today's digital flight control systems, which are implemented in the form of computer code, "gain" means multiplication of an error signal by a number. Gains are varied, either automatically or manually, for speed, altitude, dynamic pressure, payload and numerous other factors.
Designing a sound gain schedule for a given aircraft is laborious and expensive. Extensive wind tunnel and other tests are necessary to gather data, followed by demonstration of the gain schedule in flight.
Neural networks can alter the entire process by simply removing gain-schedule design, Calise says. Instead, scheduling is done automatically in real time. The neural-networked controller adapts as it flies, automatically rescheduling itself to current flight conditions. For instance, changes in the aircraft's load would not require rescheduling the controller; rather, the neural network would automatically learn the shift in the centers of gravity and the system would reconfigure itself.
The result: top in-flight performance along with big savings in flight control system engineering costs.
Calise sums up: "Even in the absence of a failure, we gain large advantages from using a learning system because it can adapt to configuration changes transparently."
Tiltrotor aircraft offer a good example of the advantages of automatic gain scheduling. Such a vehicle flies like both a helicopter and a conventional aircraft, demanding a highly complex flight control system.
"We've demonstrated that without scheduling and with a single set of gains — and with the neural network present in the flight control system — we're able to fly the aircraft throughout its flight envelope," he says. "We can go through hover, transition, forward flight, different altitudes, different airspeeds and maintain even better performance than the gain-scheduled flight control system does."
Under Calise's direction, Rysdyk is working on a neural network that functions on a civilian tiltrotor aircraft.The research team performed a piloted full-motion flight simulation test of this flight control system at the NASA Ames Research Center last spring.
Guided munitions is another area with many different configurations and designs. Each munition type has to be wind tunnel tested as part of its autopilot design, an expensive process. Neural networks hold out the promise of a single bolt-on tail assembly that could fly, guide and stabilize an entire munitions class better than conventional autopilot designs — and do so without the need of extensive wind tunnel testing.
A Profitable Sabbatical
"That research led to a very productive effort in the control area" that in turn led to the current crop of adaptive control projects, he says. "It was a fruitful sabbatical."
Ultimately, he says, neural network development may even allow a one-size-fits-all approach throughout entire aviation areas. A flight control system for one aircraft could be moved to a different aircraft type, and it would reschedule for that vehicle automatically.
"That's kind of like the dream of adaptive control — to be able to do something like that," Calise says. "And I think we're very close to realizing that dream."
For obvious safety reasons, all modern aircraft have redundant controls — backup controls in case a primary control conks out, Calise explains. But redundancy alone can't compensate automatically for a major system loss; pilots must be trained to manually compensate if, for instance, an engine should sputter or a rudder fail. During such an emergency, the pilot must explore the remaining controls and learn what actions will restore control. In essence, the pilot is swiftly retraining himself or herself to fly a degraded vehicle.
Neural networks essentially consist of an algorithm — basically a computer code. In an aircraft-based system, the neural-network algorithm becomes part of, and works cooperatively with, the flight control system computer. Like other software, such algorithms have to be tested and refined until they do the job right. As Calise puts it: "The key to using a neural network is to be able to develop efficient algorithms for training neural networks."
A neural network tries, in so far as possible, to mimic human brain function, Calise explains. The brain's active ingredients, so to speak, are its neurons, an interconnected biological system in which each neuron has mechanisms for receiving and sending signals to other neurons.
Rather than strip out existing flight control software and replace it with a new system, the Georgia Tech scientists chose to insert the neural network into existing flight control systems. Thus the network becomes an addition to, rather than a replacement for, an existing flight control system.
Calise likes to stress the roots of his team's impressive project roster. The seven current projects can be traced back to a 1990 sabbatical that Calise was able to spend at Drexel University in Philadelphia, Pa. (an institution where he himself once taught). Taking a needed break after 15 straight years of teaching gave him time to pursue full-time research into neural networks, working with a colleague in the signal processing and image processing areas.
Contents
Research Horizons GT
Research News GTRI
Georgia Tech
Last updated: February 10, 2000