Anybody running a Butterfly Vario with XCSoar?

On Thursday, July 18, 2013 6:59:42 AM UTC+2, jfitch wrote:
How do you know a wind vector is exactly accurate? Genuinely interested.....

After 20 seconds, I get a result that is consistent with the following calculations. Adding more data will not change the wind vector (unless the wind really changes). The EKF wind matches the CirclingWind result, but the CirclingWind is more volatile (inaccurate) and needs much more time for the initial estimate.

That's not a proof, but it strongly suggests that our EKF wind is very accurate and quick.

On Thursday, July 18, 2013 6:59:42 AM UTC+2, jfitch wrote:
Also I can't claim to understand the EKF filter, is it possible to describe in layman's terms how the vector is derived from straight flight observations? Component velocity obviously, but the vector?

I forgot to respond to that part of your post. I can't because I don't understand this filter either. Kalman and EKF are pure magic to me. It was John Wharington who engineered this. If you're interested, check out the source code, XCSoar is open source!

http://git.xcsoar.org/cgit/master/xc...nd/WindEKF.cpp

How do you know a wind vector is exactly accurate? Genuinely interested.....

Im some sense you don't, truthing that kind of thing isn't easy or cheap. You can get some confidence in it through simulation and otherwise you fly with it a bunch and it consistently seems "about right"

Also I can't claim to understand the EKF filter, is it possible to describe in layman's terms how the vector is derived from straight flight observations? Component velocity obviously, but the vector?

The EKF is a generalization of a kalman filter which is used to estimate a value that isn't measurable directly but is related to some quantity that you can measure. Basically it runs a simulation of your system. From this simulation you get the value you are interested in (we like to call them "states"), you also get out a prediction of what you expect your next measurement to be. By comparing this measurement to what you actually get from a sensor you can correct errors in your simulation. There's some linear algebra in between this description and the thing actually working, but it works so well that it seems like magic.

I poked through the source code, and the way it works in XCSoar is pretty cute. The "states" used here are the wind speeds and a term to take care of density change with altitude (and unit scaling). The "simulation" predicts that the wind speed should be constant all the time. When it is time to correct the model, the wind is subtracted from the GPS velocity, giving a guess at the airspeed vector which is then used to compute a guess at the dynamic pressure. By comparing the dynamic pressure measured to what it was expected to be, the EKF can start to drive out errors. After a few seconds of chunking away at this, the wind estimate should be pretty good.

HTH

I think that passing airspeed to XCSoar (V7 approach) is advantageous compared to passing the wind vector (Butterfly approach) to XCSoar.

On Thursday, July 18, 2013 12:59:42 AM UTC-4, jfitch wrote:
How do you know a wind vector is exactly accurate?

If you're relying on software to get you somewhere safely, the bottom line is that you want to know that the implementation is robust and correct under a variety of scenarios. You might be able to formally prove that the algorithm is correct, but that tells you nothing about the correctness of the implemented code.

To raise confidence in the implementation you need an "Oracle", a way of knowing the correct answer so that you can compare the expected answer with the computed result. One could write a program that would generate a synthetic GPS and airspeed trace for a known wind and heading, feed that to XCSoar (on the ground in a testing harness), and compare the actual result to the expected result. To raise confidence you would need to test as many non-equivalent scenarios as possible. To some extent these scenarios could be generated painstakingly by hand. To go one level further, one could write a program that would generate a large number of test scenarios (1000s) randomly and automatically run the scenarios through XCSoar. That would gradually raise confidence in the implementation. You also need to verify the correctness of the test program, and so all of this is rather time-consuming, and so we generally rely on the warm and fuzzy feeling that we get when software "seems to work fine". It is difficult to generate exhaustive scenario based testing and for that reason XCSoar's rather unique unit tests are a smart and efficient way to raise confidence. Some programs rely more heavily on scenario and user testing.

The thing that I like about the way that XCSoar takes the raw airspeed data from the V7 is that the V7 independently computes a wind vector and you can compare that vector to the vector computed by XCSoar. If XCSoar took the wind vector from V7, you would see the same, possibly wrong vector displayed in two places. So it is possible to do a "poor man's verification" by comparing the wind vectors computed by the V7 with the vector computed by XCSoar, in flight and under a variety of conditions. You cannot make a meaningful comparison when XCSoar takes the wind vector from the Butterfly because there is no independent computation of the vector. That's said, it is entirely possible for two (or three) completely independent implementations to make exactly the same mistake and end up agreeing on the wrong answer. The prudent pilot does a sanity-check on anything that a computer tells him in the air.

....
is it possible to describe in layman's terms how the vector is derived from straight flight observations? Component velocity obviously, but the vector?


Oops, forgot a bit...you are right that you can't get the wind vector from straight-line observation (at least with the current formulation, with more sensors you can do it). This will only allow you to predict the wind along the direction you are currently flying. As soon as you turn at all however, you'll have information about the wind in another direction. After a bit or circling or even just some turns back and forth the filter will be able to figure things out. If you fly for a long time in one direction the wind estimate might start to drift a bit though.

On Wednesday, July 17, 2013 11:56:21 AM UTC+2, Max Kellermann wrote:
On Wednesday, July 17, 2013 2:28:21 AM UTC+2, son_of_flubber wrote:

Does anyone know if XCSoar correctly interprets the wind direction information that it gets from a V7?



This question cannot be answered, because the V7 does not send a wind estimate over the wire.

Latest version of Butterfly Vario Firmware (more accurate: Butterfly NMEA Interface Unit firmware contained in the latest release) rotates output on LXWP 180° so that it matches XCSoars assumption.

Still don't know if this makes sense, yet it will work for now and I guess its a good idea to share the XCSoar teams assumption...

Cheers
Marc

On Thursday, July 18, 2013 6:10:05 AM UTC-7, jjbird wrote:
...

is it possible to describe in layman's terms how the vector is derived from straight flight observations? Component velocity obviously, but the vector?





Oops, forgot a bit...you are right that you can't get the wind vector from straight-line observation (at least with the current formulation, with more sensors you can do it). This will only allow you to predict the wind along the direction you are currently flying. As soon as you turn at all however, you'll have information about the wind in another direction. After a bit or circling or even just some turns back and forth the filter will be able to figure things out. If you fly for a long time in one direction the wind estimate might start to drift a bit though.

OK, I agree that wind velocity component in the direction of flight is derivable this way. Must the be engineer in me, "vector" means velocity and direction. My rather limited understanding of the EKF suggests that it is a very elegant way to get a quick reading for velocity out of the jittery GPS data.

On XCSoar then, if I check Both in the wind preferences and turn off the "prefer external" then I will get the XCSoar calculation based on circling or EKF filtered straight flight data? I am interested in comparing this with the Butterfly inertial derived wind vector.

Could ask this on the XCSoar forum, but for some reason I am banned from the site....

And thanks, Marc, for a quick resolution to this.