View Full Version : Scientific American article: crippled not crashed

2nd Aug 2004, 06:24
Dear moderator(s), Please move to appropriate spot.


July 26, 2004

Crippled but Not Crashed

Neural networks can help pilots land damaged planes

By Mike Corder

On July 19, 1989, as United Airlines flight 232 cruised over Iowa, the fan disk of the tail engine on the DC-10 broke apart, and the debris cut through all three of the plane's hydraulic lines. Because the pilots could not move any of the jet's control surfaces--the ailerons on the wings and the elevators and rudder on the tail--a horrific crash seemed inevitable. But by carefully adjusting power to the two remaining engines, the crew managed to maneuver the plane to the Sioux City airport. Although the jet flipped over and caught fire after hitting the runway, 184 of the 296 passengers and crew members survived.

The pilots of flight 232 proved that it was possible to control a modern airliner using only the engines. And this discovery led some innovative engineers to wonder if they could program flight computers to achieve the same feat, making it easier for a crew to safely land a heavily damaged aircraft. This research has been gradually progressing over the past 15 years, and the technology could be incorporated into commercial and military planes in the not too distant future. To judge how well these computer-controlled flight systems perform, I decided to see if they could enable a moderately experienced pilot like myself to fly a crippled jet.

But first, a little background. On early aircraft, the control stick and rudder pedals were directly connected to the control surfaces with wires or rods or cables. But as planes got faster and larger, pilots found it hard to move the stick. So engineers added "power steering," connecting the cables to hydraulic servos that amplify the pilot's efforts. Then, with the advent of the digital age, aircraft makers developed control systems that feed the input from pilots into a computer. This so-called fly-by-wire system can greatly improve an airplane's performance. For example, a fighter jet may fly well when lightly loaded but not so well when it carries bombs on its wings. With a computer in the loop, the control rules can be modified to make the plane behave more consistently. Fly-by-wire also allows the creation of safeguards: if a pilot tries to do something that would cause the aircraft to break apart or plummet to the ground, the computer can ignore the inputs and take the plane only to the edge of the flight envelope.

Shortly after the crash of flight 232, Frank W. (Bill) Burcham, Jr., then chief propulsion engineer at the NASA Dryden Flight Research Center in Edwards, Calif., began an effort to develop software that would enable jet engines to compensate for damage to a plane's control surfaces. Initially the research was considered too far-out to be funded, but a few engineers at Dryden volunteered their spare time. The project, which became known as Propulsion Controlled Aircraft (PCA), eventually received a small budget and proceeded to flight tests with an MD-11 jet. On August 29, 1995, the PCA team brought the plane in for a smooth landing at Edwards Air Force Base using only the computer-controlled engines to maneuver the craft. The NASA engineers felt they had demonstrated that airliner safety could be significantly enhanced just by modifying a plane's software. Unfortunately, none of the aircraft manufacturers chose to adopt the technology.

A few years later researchers in the Intelligent Flight Control (IFC) group at the NASA Ames Research Center in Mountain View, Calif., followed up on the PCA work by developing a system that would allow the computer-controlled engines of a damaged aircraft to work together with any control surfaces that remain functional. The system is based on neural-network software, which mimics the behavior of the human brain by learning from experience--the network's connections strengthen with use and weaken with disuse. The neural networks in the IFC system compare the way the plane should be flying with the way it actually is flying. Differences may be caused by inaccuracies in the reference model, normal wear and tear on the plane, or damage to the aircraft's physical structure. The networks monitor these differences and attempt to minimize them.

For example, if you want to make an undamaged airplane climb, you pull back on the control stick, which raises the elevators. But if the elevators are not working, the IFC system will raise both ailerons to lift the airplane's nose. (Ailerons typically move asymmetrically, with one rising as the other falls.) If this maneuver does not correct the error or if it reaches the limits imposed to prevent the aircraft from rolling over, the IFC system uses the thrust of the engines to achieve the desired pitch.

The Ames researchers tested their system by inviting professional airline pilots and NASA test pilots to fly in the lab's simulator. First, the pilots operated the simulated aircraft under normal conditions. Then the researchers mimicked a variety of failures and observed how the pilots reacted using different types of control systems. In almost every case, the IFC system performed better than a conventional fly-by-wire control system. When the engineers simulated the failure of all tail controls, only half the pilots could safely land the plane using the fly-by-wire system, but all of them made it back to the runway using IFC.

So what's it like to fly a plane equipped with neural networks? At the invitation of Karen Gundy-Burlet, head of the IFC group, I recently spent several hours in its lab to see the system firsthand. I am a private pilot with no experience flying larger aircraft. The IFC simulator was set up to represent a very big plane: the U.S. Air Force's four-engine C-17 transport jet. The simulator features a large wraparound screen to show the animated landscape and a mockup of a glass cockpit, which replaces the traditional flight gauges with flat-panel color displays.

Gundy-Burlet set me up on a 12-mile final approach to the San Francisco airport and let me embarrass myself trying to get an undamaged plane to the ground. Don Bryant, a retired U.S. Navy fighter pilot who works with the IFC group, was polite enough not to openly laugh at my ham-handed attempts to control the craft. My biggest problem was my unfamiliarity with the glass cockpit, which is only now starting to appear in private planes. I spent more time staring at the simulated display trying to find familiar values such as airspeed and altitude than I did actually flying the aircraft. That said, I got a basic feel for how the undamaged plane flew.

Then Gundy-Burlet reset the simulator to the initial location and said, "Captain, I'm sorry, but you've lost all the control surfaces on the tail." Both the elevators and rudders were inoperative, which would probably be a death sentence for an amateur pilot in the real world. But I was pleasantly surprised to find that the simulated aircraft was pretty controllable. I made a few gentle turns to get a feel for the plane while also trying to stay on the right heading. The damaged jet was sluggish in roll and pitch, but its behavior seemed more natural once I slowed down my steering. This change was undoubtedly facilitated by the neural networks, which were training themselves to compensate for the damage. As the networks adjusted to the new conditions, the plane kept getting easier to fly. Within a few minutes, I was able to safely land the simulated craft, although it did stray from the runway.

The overall experience was fairly tame, almost ordinary. It was only later that I recognized the true magnitude of this advance. A private pilot who had never flown a large aircraft was able to land a heavily damaged four-engine jet without killing anybody (in a simulation, at least).

How quickly might this technology see actual use? NASA researchers plan to flight-test the IFC system on F-15 fighter jets and C-17 transport craft over the next two years. The earliest adopters will most likely be the makers of military aircraft. Damage-compensating flight controls should be particularly useful to pilots who fly aircraft that get shot at from time to time.

Mike Corder is a freelance writer in Santa Cruz, Calif., who is building a Van's Aircraft RV-7A plane in his spare time.

2nd Aug 2004, 06:42
A fascinating article. Having done some work with Neural Networks myself I can see tremedous potential in this. Not to replace the pilot (not in our lifetimes anyway), but to help in abnormal situations.

Having said that, I wonder how a Neural Net can control the pitch of the aircraft? Much easier with a DC-10 or MD-11 with the engine high up on the tail, but what about the standard twin-jet config?

Might we see the high engine making a comeback?

2nd Aug 2004, 06:54
Well, it's interesting that they use the Sioux City DC 10 crash as an example, as I recently commented that, to my recollection, NOBODY managed to land that plane safely in a simulator (and I include the last minute cartwheel as a "safe" landing under the circumstances...)

2nd Aug 2004, 08:38
The limiting factor inherent in neural nets is the relevance of the 'training set' to the actual circumstance.

Training a neural net is basically a matter of superimposing a new set of relationships and values on the base set or sets. This idea might be helpful if a certain type of fundamental failure - not already anticipated in the aircraft design & redundancy principles - could be 'trained' into the NN controls for future availability in the event of that or a very similar contingency. But there must be a fairly short list of such contingencies that are both flyable and not already accounted for. The trick of steering with engines-only is a good one...probably wouldn't be a bad choice to include it somewhere down the line in production aircraft, but then folks will have to periodically test it, and use it in sim training, and maintain it, etc...

The bottom line is that every new feature / function has an initial cost, a training cost, a support cost, etc. For things considered to be extremely low probability, the weight of trade-offs often countervails the putative benefit.

Neural nets are probably a good way to implement aircraft controls altogether. Accountability is a problem. They are, by definition, non-deterministic automata that don't always give the desired answer to the 'standard' set of constraints - just like people. Probably NN controls are a good intermediate step on the way to truly intelligent controls and systems. We can build those today that work pretty well. We just can't PROVE that they will always work when needed. Because they won't 'always' do any specific thing in the exact same way.... just like people....but it is much harder to look them in the eye to see if they can be trusted.

2nd Aug 2004, 13:13
Having said that, I wonder how a Neural Net can control the pitch of the aircraft? Much easier with a DC-10 or MD-11 with the engine high up on the tail, but what about the standard twin-jet config?

You'll recall that the Sioux City DC-10 lost it's center engine. So whatever control they achieved, they achieved just with the two underwing engines.

Want to increase pitch? Increase power on both engines.

2nd Aug 2004, 13:56

One of the reasons that no one could replicate souix city has to do with the limitations of similulators. When ever anyone jumps in a sim the invariable complaint is that it is too pitch sensitive. That is because the sim can't replicated the feel of pitch (increased or decreased G loading) It can replicate the feel of acceleration and deceleeration, uncordinated flight and some turbulence.

But the phugoid occilations that Soiux city experienced are beyond the ability of any sim to simulate, and that had a huge part in the handling of the aircraft.

Also, Soiux City DID NOT CARTWHEEL. They struck the runway pretty much wings level at about 1200 feet per minute. What then happened is that the wing broke (A Known problem with the DC-10/md-11 failure mode is that the wing breaks, look the the EWR Fedex MD-11, Memphis MD-10, Hongkong MD-11, When one wing breaks then the other wing flips the remaining parts of the aircraft over)


3rd Aug 2004, 07:25
You'll recall that the Sioux City DC-10 lost it's center engine. So whatever control they achieved, they achieved just with the two underwing engines. I had in fact forgotten that point when I wrote my post yesterday morning. I was thinking purely in terms of lost hydraulics.

I think it was the "insufficient coffee factor" :zzz:

Thanks for pointing that out! :D