Air France 447: How scientists found a needle in a haystack


The cockpit voice recorder from Air France 447, as it was found at the bottom of the Atlantic Ocean.

Last weekend, investigators announced that they had recovered the flight data recorder from the wreckage of Air France 447—a jetliner that crashed in the deep Atlantic two years ago. But, while the discovery of the data recorder is recent, the story of how Flight 447 was found goes back a month.

This year's search was the fourth attempt to find the wreckage of Flight 447, and it probably would have been the last, even if the plane hadn't been found. Previous searches had been done by boat, mini-sub, and—back when there was still a chance of catching the audio signal from the plane's black boxes—underwater acoustic sensors. In 2010, scientists from the Woods Hole Oceanographic Institute were brought in to search for the crash site using autonomous robot subs. Still nothing had been found.

On March 22, 2011, the Woods Hole team set out from Brazil to try again. They'd barely been at the search location for a week when they found what they were looking for. On April 3, researchers spotted the plane's debris field, 13,000 feet down, smack in the middle of a massive underwater mountain range.

The success was astounding, but I wanted to know ... what made this search different from the others? What could the team from Woods Hole do that other groups could not, and how did their system work? To find out, I spoke with Mike Purcell, senior engineer with Woods Hole, and the chief of sea search operations for the mission.

Maggie Koerth-Baker: Your team found Flight 447 with the help of an autonomous submarine called the Remus 6000. Can you tell me a little about the history of that sub? What could the Remus 6000 do that previous systems couldn't?

Mike Purcell: The first one was developed in 2001. Really, they have a greater depth limit. There are no other deep water subs that can go to 6000 meters. That's one way it's unique. Also, between the six Remus 6000's that exist out in the world right now, there's probably been more missions done with a Remus 6000 than any other deep water AUV.

To do a search, the Remus 6000 gets a mission program, a track line to swim. It goes into the water and uses various navigation techniques to swim the track line. There isn't anybody actively controlling it. But it's also not as smart as you might think. It's not making decisions based on terrain, other than staying some fixed altitude off the bottom. It can't go around things or avoid stuff that might be in front of it. It does go up over mountain ridges, but the Remus 6000s do sometimes run into things, too. They don't have the full sensor capacity and independent thinking to make decisions that some totally autonomous robot might. One reason that's the case—it's just harder to do that in the water than in the air. We're really limited to one kind of sensor, acoustic sensors, underwater.

MKB: What kind of research do Remus 6000 subs normally work on? Was this search different in any way, from a technological or logistics perspective?

MP: Our lab ... we've been involved in the development of AUVs. We've been making the newer and better ones over the last 15 years. It was only in about 2008 that we started getting involved in operations. We purchased a couple Remus 6000s and we're the operators. They were involved in search for Amelia Earhart's airplane. We did some localization of deep corals in Gulf Stream off of Florida. We mapped the Titanic site with AUV's last year. And then we've now worked on the Air France survey twice, once in 2010 and once in 2011.

Even when we've done these searches for the airplanes there's been a tremendous amount of data collected, and that's been made available, or will be made available in the future, to the science community. What kinds of things can people do with seafloor data? I'm not a geologist, so I'm not totally sure what they might do. But a lot of the seafloor is totally unexplored. We've got about 1500 square miles mapped. And I think there's a lot of interesting geography there in the Mid-Atlantic Ridge where we did this search.

MKB: How many people involved in running one of these searches, and what do they do?

MP: We had three vehicles out there. When we're running three vehicles we have 12 people, working in two 12-hour shifts. There's six people on each shift. And they're doing things like getting the vehicles in and out of the water. Reprogramming the vehicles. Tracking the vehicles. There's usually two AUVs in the water at all times. And there's a guy who is dedicated to processing the data.

MKB: The Remus 6000s had previously been involved in the search for Air France 447, but hadn't found it before. Was there a major location change, or some other shift in how the search was done this time around? Were you involved in deciding where the search would happen?

MP: We were out there first in 2010, and there'd been a pretty big modeling study that guided the search then. Of the entire area, which is 17,000 square kilometers, 7,000 had been what we were search going into this year. The plan was to search it all. There was one difference, we just decided to start close to the last known position of the plane, instead of further away from it. The BEA [Ed: Bureau d'Enquêtes et d'Analyses, the French air safety investigators] identified three search zones, big areas that they wanted us to do in order, and then, from there, we sort of had the freedom to decide where we start in those areas. So we started out based on where we left off last year.

MKB: The mid-ocean ridge, where the search was conducted, has been described as something like an underwater Himalayan mountain range. A lot of reports I've read on it made it sound very foreboding and not like a place where it would be possible to find anything. But WHOI has been doing research on the mid-ocean ridge for decades. Is the scary reputation deserved? What challenges do you face doing research in that location, as an organization that has experience with it?

MP: So, I think this mission was different for us in that we were trying to search such a huge area. We needed our vehicles to swim up and down those mountains. The water out there was 4000 meters meters deep at the deepest spot and very close to that was where we found the wreckage. But just a few miles away it was only 2000 meters meters deep. There are some very steep mountains.

We had a pretty good contour map of the seafloor, but you always learn some new things when you put the vehicle down there. Occasionally it would run into a mountain or cliff. Sometimes, when that happens, it stops the mission. We have a drop weight on the vehicle and it will float back to the surface. Other times a crash leads to problems a couple of missions later&mash;the AUV will suddenly develop an electrical problem. Sometimes, the cliffs fall away so fast that we can't swim down the slope as fast as its falling off, so we don't get good data and have to go back and do it again swimming up the slope. Compared to past experience here, the fact that we were trying to cover so much area was the new challenge.

We can do 100 square kilometers in a day with two vehicles ... but that really depends on geography. If the bottom is flat we can do more. We've done 180 square kilometers in a day with really flat terrain.

MKB: How does a search with the Remus 6000 work, before the point where you know you've found what you're looking for? Are you back on the boat watching live video from the sub, or is it a bit more "blind" than that, so to speak?

MP: Fairly more blind. There's no data transmission back to the ship while the vehicles are in the water. We get status messages—acoustic messages that come in periodically and tell us how deep it is, latitude and longitude, just a status check to tell us whether there's problems. When the AUV gets back, it takes 45 minutes to download the data and then another half hour to process and get a good look at it. During that time the other team is switching out the battery and getting the vehicle ready to go back in. The ideal is that a vehicle is only out of the water for three hours, while somebody is looking at the data to decide where we go from here and are there things we want to look at again. When we're running three vehicles we get a data dump three times a day.

What we get is images of the acoustic data. The vehicle sends out an acoustic ping every second, it goes out 700 meters on every side of the vehicle. That gets logged and turned into an image. It'll do that for about 20 hours at a stretch.

MKB: When your crew first saw what turned out to be the wreckage, what would they have been looking at? Does the image look anything like a wrecked plane, or is this a much more abstract sort of thing?

MP: It's a lot more abstract. It just looks like a bunch of bright returns on an area of no return. We're looking at a 24 inch monitor and we've 1400 meters of space that is being imaged on that monitor. Basically we were looking for a lot of small returns over a big area.

I think it takes a considerable amount of time to get to be good at identifying the data. Lots of people can look at a flat seafloor and see what sticks out. But once you have to start dealing with terrain changes, being able to pick wreckage out of that takes a lot more experience. It's good to have an experienced person looking at this stuff. Our guys probably had 15 years.


Bright spots on a dark field. How the wreckage looked to Woods Hole data processors.

MKB: How do you tell the difference between stuff that's supposed to be on the sea floor and what you're looking for?

MP: Sometimes it's the strength of the return. Sometimes ... you might have a ridge line that's mostly sediment but some exposed rocks. So you see there's six bright spots coming out of the ridge, but you can also tell there's a ridge there. So you figure it's probably natural. You get a feel for what's natural and what's not natural. Metal tends to have sharp or square edges, That sends back a stronger signal than rounded rocks that have been down there for a million years or more. In this case, the airplane was out in the middle of a very flat area with no natural camouflage that would make it hard to find.

MKB: Did the capabilities of your machinery make the most difference, or did it have more to do with the experience that your operators have conducting this kind of seafloor search?

MP: A lot of people would have found the wreckage where it was, with other vehicles or even towed systems. But when you look at the big picture, the whole area that had to be searched, the Remus vehicles were the best tool able to work in that terrain. And the fact that we had three of them working around the clock. That just increases our search rate. That factor there is both a vehicle thing and it's also the people out there doing the operating.

MKB: During the search you sent progress reports back to victim's families. How do you write a progress report like that? How much different is it from the way you frame findings, and wording you use if you're writing a progress report for something less emotionally charged?

MP: I sort of lucked out. I never had to send a progress report. It was supposed to go out every Tuesday. They gave us a waiver on the first Tuesday because we'd just gotten to the site, and then we found the wreckage on Sunday. We had a BEA representative on board and I talked with him every day. A lot of people had a lot at stake in that search. They've been working at it for a long time. That was more of a factor than the fact that it was a crashed airplane. That sort of changes the attitude. People were serious about the job. They took it seriously and wanted to complete it.

Images: BEA, via Reuters


  1. It still boggles my mind… I’d be curious to know a little more about the modeling study that he alluded to and how its conclusions for a search area were arrived at… Hopefully the FDR and CVR, both of which appeared to be in decent condition, will reveal quite a bit more information on the circumstances that brought her down.

  2. Thanks for this Maggie; fascinating stuff. It’s hard to imagine seeing that image and thinking “there’s the airplane!” They must be very patient and very good.

  3. Can someone knowledgable explain why air safety investigators have to rely on a hardware black box?

    Why can’t this data be streamed?

    Compare with the telemetry readings available to the Columbia investigation team, which practically pointed a flashing neon sign at the cause of the breakup.

    1. I’m not knowledgeable in the matter but if I were to guess, the cost of implementing a data collection policy is enormous and with this find, investigators still have a 100% recovery rate of blackboxes.

    2. Well, for one thing, the possibility that the phenomenon that causes the crash will also stop the data stream early enough to render the system useless.

    3. Why can’t this data be streamed?

      No doubt it will be, but development cycles in aviation are very long. It takes time to design stuff when the design has to be exactly right. Its not like the work is going to be trusted to Amazon EC2.

    4. Wally,
      Recording to tape is much cheaper than streaming through a sat. connection. I live aboard a boat, when not connected to a GSM/3G/Edge signal, i use Inmarsat BGAN (the only Worldwide sat broadband provider) to retrieve my emails and reads news on the net. The cost is more than usd6 per meg !!!
      Air France is a company with no money, that crash is a direct result of bad maintenance. This is due to stupid work regulations in France : working hours are limited to 35hours per week. If you work more, it’s at overtime rates. So Air France can’t actually maintain it’s fleet properly because it would cost them too much in salaries … It’s a well known issue with people working at Air France, but the french “government” will never acknowledge such a fact : ” Sacrebleu, how can you say we are broke, this is France you are talking about !!!”
      BenBen, french national (but i saved my life and now live in asia)

      1. So tired overworked people would catch more problems than a fresh pair of eyes? Or is this accompanied by a law that sets a maximum number of employees at Air France? Is there only the one guy or what?

        Also, what about pilots? Don’t they work long and odd hours?

    5. Why can the Space Shuttle stream info and commercial airliners don’t?

      Simple: it’s the money.

      Governments will spend the money, while commercial aviation won’t. If the airlines did, you’d see your ticket prices go up.

      Realistically, commercial aircraft seldom crash, and when they do most often it is on land and the ‘black boxes’ are recoverable. It’s only the extremely few flights which crash in the water where recovery of the black boxes becomes problematic.

      In the middle of the ocean the only way you’re going get reliable data transmission is via satellite. Say the FDR and the CVR generate an average data stream of 2Mbps — now go figure out what INMARSAT will charge for that over a several hour period mid-ocean. Now multiply that by the number of flights an airline has over the oceans at any point in time.

    6. Streaming Flight Data is nearly impossible. Flight Data is a comprehensive set of information records that includes telemetric lectures, voice, condition metrics and measuring, and detailed time synchronization tracks. Streaming that kind of information is too difficult.

      Is not like having a 3G internet connection or something. This is not like having an mp3 file and a txt codified note. Is a complex system of information which really needs an specialized data protocol that nowadays is not supported by any type of wireless conection.

      In the other hand there are thousands of flights daily. There is no infrastructure or financial capacity to assume something like that.

  4. Amazing find! Now hopefully investigators will be able to determine the root cause of the of the crash. A tip of the figurative hat to Woods Hole.

  5. Streaming flight data would require monitoring of pilots at work in real time. There is some professional resistance to this kind of oversight. I sort of understand the reluctance to being monitored every second at work but in the case of pilots, I think they should get the hell over it because safety should be the #1 concern.

    1. While that is certainly a cause for concern from pilots and a hot topic right now as there is evidently to be far greater audio and video (so I’ve read but not confirmed) recording of the cockpit in the 787… the data streaming question that came up because of 447 in particular is usually around greater systems status and telemetry data and not simply recording the flight crew actions (although to the point about the pitot tubes it is indeed massively important to know how they reacted to an alternate law state). I think the fact that there were ACARS messages at all is what has even raised the question of real time monitoring in the first place. We got a taste of it and want more…

      Anywho, implementation, monitoring and subsidizing of such a presumably standardized system across hundreds of thousands of flights across the globe is no small task either logistically or financially. I think there is likely a solution out there, but its unlikely that it is immediately around the corner for the vast majority of aircraft.

      Not to forget: part of what makes AF447 of such interest is that it is thankfully an incredibly rare occurrence for a plane to fall out of the air without report, and basically vanish with little trace. Does this incredibly rare incident immediately mandate a reworking of how planes report their location data? Would the associated cost be worth it once its made its way into your ticket price?

  6. There’s also the matter of the quantity of data due to the number of flights. NASA has a bit bigger budget per-flight than an airline. There are thousands upon thousands of flights ever day, and streaming the data along the entire flight would require a pretty hefty connection that’s persistent across thousands of miles of unpopulated areas. (I.e. the middle of the ocean in this instance.) The only solution then would be a satellite connection and using that constantly to stream flight data would be really expensive. Of course, the search for this black box was also really expensive, but $30mm wouldn’t go a long way when talking about a persistent satellite network connection for thousands of flights every single day. I imagine more than $30mm would be spent on data every day in that case.

    1. It would not be necessary to stream every second of every flight.

      If streaming were to be turned on when the plane deviated significantly from the flightplan or normal flight parameters, the volume streamed would be less by a couple of orders of magnitude at least.

      1. Yeah, I don’t think that would be useful. Mostly for the reason that Antinous suggest. You would need to negotiate a successful TCP connection over a satellite link and then send the data, by the time the connection was negotiated the plane would be in the ocean.

  7. They make those where I work. It’s kind of weird seeing one on the bottom of the ocean carrying data that can help explain how a plane full of people lost their lives.

  8. Very interesting interview.

    I’ll have a question, though.
    The picture shows the black box as it was when it was first screened by the search teams.
    How come after two years that heavy thing is not (at least a bit) buried into the sand, or covered with naturally generated sea rubbish?

    1. The ammount of sediment that accumulates on the seabed is highly dependent on the specific locality. Places near large river deltas would obviously accumulate very quickly, as well as very biologically productive areas (all those little critters die and snow down). It is not unusual, however, for the accumulation on the seabed in the far open ocean to be measured in millimeters per century, or even less.

      Still, it is amazing that the CVR is just sitting there like it is instead of buried in a mangled pile of wreckage. That plane must have hit the water very, very fast.

    2. I’ll guess: water very deep, very dark, very still, very cold, little oxygen, and the box is built to resist corrosion. Give it another 98 years and it’ll look like something off the Titanic.

      Cockpit video recorders could help out a lot in accident investigations, because many accidents are caused by pilot error, and even when it is not a factor it would help immensely to see what the pilots did and when. The NTSB has wanted such recorders for years, and the technology is well within reach- but pilots don’t want them, and airlines won’t lift a finger unless the FAA makes them, so we don’t have them.

  9. I think I’ve figured it out. Video of the cockpit is all but irrelevant. Video of the entire outside skin of the plane could be doable for recorders, with some value.

    As for streaming data, surely you could define a range for ‘normal’ — only recorded. Then a range for ‘abnormal’ — streamed, but 95% of the time only for a few minutes.

    Finally, if ‘abnormal’ is followed by ‘nonexistent’, you have all the (relevant) data. The unexplored skill of the information age is intelligent filtering.

    The only problem is making sure that executives know the real value of transient abnormal. Not usually relevant but occasionally instructive, and especially in accident situations.

  10. The next thing that needs to be figured out is why they still use tapes and other sensitive recording materials in those boxes?

    And the endless loop audio recorder.. useless. Record the whole damn flight, that CAN be done with today’s modern solid state gear.

  11. Just a quick point: It seems it was an engineer and not a scientist that found this. The team lead is credited as an engineer and this is very much a task in engineering.

  12. Thanks for an informative article. Does anyone know whether the actual wreckage location had been previously searched back in 2009 when the locator beacons may still have been active? Given the subsea terrain at the wreckage location, and assuming that the beacons operated as designed, I’m curious to know whether their signals could have been detected, or whether the initial search was simply conducted in the wrong area? From the photos, it looks like the beacon is still attached to the CVR but possibly missing from the flight data recorder.

  13. Damn, the folks who found the wreck of the Titanic have blown us all clean away again!

  14. Why don’t they make the flight recorders capable of floating? At least if they were to be floating on the surface of the sea it would be relatively easier to locate them.

    1. I think it should be easier to the plane debris and then the recorders than to find a small box floating at the ocean. Besides that, it’s not very common that a plane crashes in such kind off ocean mountain area as happened to this air france plane.

Comments are closed.