[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: [Orekit Developers] status on second order derivatives



Evan Ward <evan.ward@nrl.navy.mil> a écrit :


On 11/10/2014 11:51 AM, MAISONOBE Luc wrote:
Evan Ward <evan.ward@nrl.navy.mil> a écrit :


On 11/01/2014 04:23 AM, MAISONOBE Luc wrote:
Hi Evan,
Evan Ward <evan.ward@nrl.navy.mil> a écrit :


On 10/30/2014 07:10 AM, Luc Maisonobe wrote:
Hi Evan,

Le 29/10/2014 21:34, Evan Ward a écrit :
On 10/29/2014 03:14 PM, MAISONOBE Luc wrote:
Hi Evan,
Evan Ward <evan.ward@nrl.navy.mil> a écrit :

Hi Luc, Paul,

If I understand correctly this issue only affects propagators
that do
not use Cartesian representation of the state. When a
non-Cartesian
representation is used the propagator's position is combined with
the
Keplerian velocity and the resulting combination is not locally
tangent
(osculating) to the propagator's orbit. Am I understanding the
issue
correctly?
Yes, this is exactly what happens.

Could this issue be solved by using the propagator's complete
state to
determine the instantaneous P/V and use that to build a
CartesianOrbit?
Not exactly. The numericalPropagator once only integrated
CartesianOrbit and it was considered a limitation. Since a few
years,
we have added the possibility to directly integrate the orbit type
specifed by user. Moving back to CartesianOrbit only would be a
regression IMHO.

However what you suggest is also almost what I have in mind
since we
really use the propagator complete state. In fact, for the
numerical
propagator the conversion between acceleration and orbit is not
done
at orbit level after integration, but rather at force level as
we use
the Jacobian of the (P,V) vs orbital parameters conversion to
directly
get parameters derivatives. When you look at it, the Gauss
equations
are really the last columns of the Jacobian. This is done with the
TimeDerivatives interface, which for NumericalPropagator is
implemented using the Jacobian.

So as far as N?umericalPropagator is involved, I think
consistency is
automatically preserved, since we do integrated something that is
computed from the Jacobian.

The DSST propagator on the other hand propagated directly in
equinoctial parameters and uses another dedicated implementation of
the TimeDerivatives interface.
I guess I don't understand why we need more than 6 parameters to
specify
the osculating Keplerian orbit.
For Keplerian orbit, 6 parameters are enough and the relationship
between (a, e, i, pa, raan, M) and (x, y, z, vx, vy, vz) is a clear
bijection. If we restrict ourselves to Keplerian motion, this
relationship holds not only at t0 but also at t0+dt for every
value we
want. If we use Keplerian parameters evolution (i.e. only M evolving)
and recompute at several steps the corresponding Cartesian, we
will see
vx, vy, vz are really the derivatives or x, y, z, because the
relationship does take Kepler laws into account.

The problem arise when we introduce non-Keplerian parts. Then if we
blindly apply the same relationship for converting between the two
sets
of parameters, we see that they are not consistent anymore.

What happens is that the mapping:

  (a, e, i, pa, raan, M)     <---> (x, y, z, vx, vy, vz)

is such that when da/dt = 0, de/dt = 0, ... dM/dt = n, then the
computed
vx is dx/dt, vy is dy/dt and vz is dz/dt.

If now we decide to set arbitrarily da/dt to some made up non null
value
and so on for all parameters, we still compute the same 6 values
(x, y,
z, vx, vy, vz) as before (the initial value of a, e, i ... did not
change), but now differentiating the mapping shows that vx will
not be
equal to dx/dt anymore, vy will not be dy/dt anymore and vz will
not be
dz/dt anymore. The made up da/dt, de/dt will contribute to the
results.
This is what happens for Eckstein-Hechler as it computes the left
part
of the mapping and we (Orekit team) compute the right part using the
Keplerian-based mapping. If we want proper computation of vx, vy,
vz, we
need the additional parameters.

I agree that we need to account for the element rate terms to
correctly
compute the velocity. I think we could do the (a, e, i, pa, raan, M,
rates, other propagator state) --> (p, v) mapping within the
Propagator
and then return a Orbit built from the correctly computed PV. This way
the returned orbit would be the osculating Keplerian orbit that
matches
the propagator's position and velocity. If the user then translates
the
osculating PV to osculating Keplerian elements they will be different
from the propagator's elements, but that is to be expected since the
propagator includes perturbations and osculating elements don't.
(Maybe
this is the source of my confusion; when hear "osculating orbit" I
think
of a conic orbit that matches the instantaneous PV at a single instant
in time.)

OK, I think this time I understood your concerns.
I'll give this a try, were the only modification to CircularOrbit will
be to add a constructor that accepts both the circular and Cartesian
parameters, hence the class will not do the conversion by itself
anymore and only the propagator need to be aware of the circular
derivatives.

My current setting was to use either circular or Cartesian but never
build an orbit from both, so the conversion could be done only by the
orbit class and therefore it should have known about the circular
derivatives. With your proposal, it is simpler. Thanks for the
suggestion.

Why not just return a CartesianOrbit from the EHPropagator? Something
like:

//in EHPropagator.propagateOrbit
PVCoordinates pv = ...; //correctly compute instantaneous PV
return new CartesianOrbit(pv, frame, date, mu);

This way no modifications are necessary for the Orbit classes, and the
user gets the correct PV at the requested time. One caveat is that if
the user needs circular parameters they will have to convert using the
CircularOrbit(orbit) constructor, and the circular parameters under the
new scheme would not match those returned by the current implementation.

This looked very interesting, so I tried it ... see below.


Maybe this is a good time to add some documentation to the
Propagator/SpacecraftState classes to specify exactly how the returned
orbit is defined. The way I see it we have three options:

1. The current state in EH, where the EH parameters are used to create a
Keplerian orbit. The user sees the EH parameters, but the velocity and
local extrapolation is incorrect. This is what we are trying to replace.

Yes. It was wrong and it should be fixed.


2. Create and return a conic orbit computed from the propagator's
position and velocity. This matches the definition of osculating orbit
provided several authors. (e.g. Battin, Montenbruck & Gill) I think this
approach is the simplest for the user to understand and the simplest to
implement. (This is the option I outlined above.)

I tried it. It is very simple to implement and removes lots of code
with respect to the third option below.

The problems I got after making this change is that the orbit returned
by the propagator seems different from the initial orbit used to
create the propagator (or used in resetInitialState). Typically, the
unit tests sameDateKeplerian, sameDateCartesian and
propagatedEquinoctial in EcksteinHechlerPropagatorTest fail with semi
major axis being a few hundreds of meters away from expected value,
despite position was the same at 2.5e-8m. Of course this can be
expected as we know velocity is inconsistent.

However, as the input osculating orbit should be considered perfect I
don't know how to handle this.

I first tried to force the mean parameters computation to converge to
the same initial parameters, which worked great for the specific tests
outlined above but created very wrong propagated orbits. The
propagation comparison with numerical propagators
HolmesFeatherstoneAttractionModelTest.testEcksteinHechlerReference
failed with big errors (a few thousand meters after 12h and diverging,
when the expected error is periodic and stable at roughly 100m).

This result surprised me since I would expect that two propagators that
purport to implement the same model would produce the same ephemeris
given the same initial position and velocity. After experimenting with
your code for a while I think that the 100 meter differences we see with
options 1 and 3 are the cause of the 9km differences we see with option
2. When the EH propagator determines its mean orbit it either fits EH
p/v to the initial p/v (option 2) or it fits the EH elements to the
initial Keplerian elements (option 3). Option 3 seems to fit the orbit
as a whole better, but the EH velocity and the initial velocity are
different by a small amount, where as option 2 forces the velocity to
match exactly even when the EH velocity is in error by a small amount.
For example in  testEcksteinHechlerReference the velocity difference
between the two options at the initial epoch is ~0.15 m/s which seems
large enough to produce the observed km of difference over 50 000 s.

To check that the initial orbit matching is causing the errors, I found
that if the initial orbit is shifted by a quarter period (or -0.75
period) that the EH ephemeris stays within 70 m of the numerical propagator.

For better fitting, we can always use the PropagatorConverter implementations : they adjust the initial parameters on a complete sample instead of only one point.


So IMHO we can either fit the velocity exactly at the epoch or fit the
orbit parameters and tolerate a small velocity discrepancy at the epoch
in return for an orbit closer to the numerical propagator. Both can work
with option 2.

I definitely think option 2 is much simpler, so if you can find a way to meke it work, I would be very happy.

For now, I have pushed the two branches I use for experiments. One is called EHOrbit and implements option 3, the other is called Eckstein-Hechler-Cartesian and implements option 2. In both cases, I have updated the tests as much as possible to match the accuracy achieved. For option 3, now all tests without exception do pass. Interpolation is very good but shiftedBy behaves strangely (I feel the error increases too fast for my taste). For option 2, four tests in EcksteinHechlerPropagatorTest fail, as explained in my previous post. If we could get them working, I think this branch would be a much better solution to merge in master.


Do you have access the reference publication for the EH propagator? I
think it may help to see how the propagator is initialized there.

Unfortunately, no. The implementation was created from an appendix in the CNES 1995 book, not from the original paper. This book is not available anymore as far as I know. I will look again at it but don't think there is anything about initialization. The initialization process we used is a very simple fixed-point iteration as the model creates osculating parameters from mean parameters : so we iterate until the produced osculating match the initial orbit, currently in terms of circular parameters.

best regards,
Luc


Best Regards,
Evan


So up to now, I was not able to have both initial orbit, propagated
orbit and velocity correct. In all my tests, I had only two correct
among the three.


3. Create a combined orbit that has both the EH elements and the
position and velocity. This may be a bit confusing for the users since
the EH orbital elements are only define w.r.t. the EH propagation model.

In fact, as seen in the previous case, the orbital elements are also
visible to the user as they are used to initialize the propagator from
an osculating orbit (notwithstanding velocity). The hidden parameters
are the mean parameters from wich EH computes the osculating ones.

In other words, if a user writes a method to use a Propagator (the base
interface, not one of its sub classes) then the user wouldn't be able to
use the elements (a, e, i, ...) in the returned orbit since they
wouldn't know how the elements are defined. On the other hand, if the
user needs the EH parameters then I think this is the definition we
should choose. (IIUC this is what Luc is describing.)

I tried it for the last few days (before switching to your suggestion
number 2 above). It almost works but is quite complicated and I did
not yet manage to solve a last problem.

The way I've done it was to create an internal EHOrbit class extending
CircularOrbit and overriding a few methods. At construction time it
preserve the value of the circular osculating parameters (so initial
orbit is preserved when propagating to same date), and it does not
modify the mean parameters computation loop (so propagated orbit is
consistent with other numerical propagators), and it additionally
computes Cartesian parameters knowing about the derivatives (so it
solves the initial problem). It's a bunch of code, but is quite
straightforward thanks to the DerivativeStructure class which greatly
simplifies derivatives computation.

This however is not enough, and a few methods should be overriden too:
shitfetBy and interpolate. Failing to overriding thme creates another
type of inconsistency: the velocity of shifted or interpolated
instances computed directly by the base CircularOrbit class is wrong.
Overriding shiftedBy is trivial and works well. Overriding interpolate
is difficult as it could be fed from several types of orbit in the
sample, so I tried to rebuild the derivatives if the sample did not
include them. This is a nightmare and I utterly failed. The problem is
interpolate is automatically used for example when one creates an
ephemeris from an EcksteinHechler propagator. It is a feature shared
among all propagators and widely used.

So I am confused now.

I could attempt to reset the specialized EHOrbit class that knows
about non-Keplerian effect when converting back and forth between
circular and Cartesian, but I have yet to find how interpolate should
work. Perhaps we could simply say that the specific implementation of
interpolate in EHOrbit works *only* when the sample contains only
EHOrbit instances and throws an exception otherwise. It seems fair to
me as we cannot rebuild missing derivatives, and it would still work
correctly in the regular case where people use consistently EHOrbit
for a sample. In fact, it could probably be considered a general
contract of the interpolate method to use only one type of orbits at a
time and not to convert anything under the hood.

Any thought ?
Luc




The problem may not hold for NumericalPropagator (I still have to
check), because basically we do the computation the other way
round. We
start from x, y, z, vx, vy, vz and deduce a, e, ... from the
mapping. So
as long as our vx=dx/dt, vy=dy/dt, vz=dz/dt, the initial mapping will
still hold.

It seems that if we add rates to all the elements then the resulting
class could be classified as a Propagator.
It is only a way to get a consistent (P, V), and as a propagator it
would be really limited as it is a Taylor expansion. I would simply
qualify it as a perturbed orbit allowing local expansion.

Extending the Taylor series expansion to higher order derivatives
makes
sense if we need to return the propagator's internal state, and the
satellite's PV in one consistent object. Is there a use case for
providing the propagator's internal state? I don't know, but I can't
think of any.

If we need to provide the user with acceleration as well as PV then I
agree that we need a bigger container class to hold the extra
information. Maybe the new PVACoordinates would be the right choice

I agree this is a good place. However, I did not create an extra class
for this either, I simply added a third vector to PVCoordinates. We
can discuss this later, when the merge of the branch will be attempted.

instead of a PerturbedOrbit class since I don't think we would get the
same extrapolation from a PerturbedOrbit stored in Keplerian vs
Cartesian elements. To put it another way, If I converted a
PerturbedKeplerianOrbit to a PerturbedCartesianOrbit would they both
follow the same path?

At second order level, in the neighborhood of the matching date, yes
they should. Differences should appear at the first ignored derivative
level and build up from there.


I'm not against the PerturbedOrbit classes; I just want to make
sure we
don't choose a complex solution to a simple problem.

You are right. I have a clear tendency to over-engineer things, this
is why I ask for advice from time to time. This is the great force of
collaborative development : people don't get stuck in their own errors
too long.

I think these constructive discussions combine the best ideas from
everyone.


So to be clear, here is the consensus :

 - I will simplify what I have done to limit as much as possible
   the changes to CircularOrbit.
 - In this setting, orbits will not know about non-Keplerian effects.
 - As this knowledge is needed for accurate and more importantly
   consistent velocity computation, the mapping between orbital
parameters
   and Cartesian parameters will be precomputed by the propagators
 - If the mapping is not precomputed and orbits are built from one
   set of parameters only, then the mapping will be Keplerian only
   and will be performed byt orbit classes just as it is now


If we decide to define the orbit returned from a Propagator according to
option 3 above then I think this is a good plan to implement it.

Best Regards,
Evan

best regards,
Luc


There might be a case for
this intermediate level of fidelity/speed, but is it worth the added
complexity? Especially since we already have one that is fast (the
Orbit
class) and one that is precise (the Propagator classes). Maybe I'm
still
not understanding your proposal.
I detected the problem when checking some pointing attitude modes
that
are related to spacecraft velocity (alignment of the spacecraft axis
with ground drift for Earth observation). This mode transforms
something
that is a derivative (velocity) into a regular non-differentiated
value
(an angle). So we go up one order of derivation. For such modes, a
wrong
velocity induces a wrong angle and a wrong acceleration induces a
wrong
angular rate. As the angular rate is used for time shift, I needed
it to
be accurate.

You also mentioned issues with Doppler which would also occur.

We do need accurate velocity, and we do need a velocity that is
consistent with the derivative of the position.

Definitely agree that getting the velocity correct is important.
Thanks
again for all the work you've put into solving this bug.

Best Regards,
Evan


best regards,
Luc

Best Regards,
Evan

Then we wouldn't have to modify the existing set of Orbit
classes, and
the user would see the correct osculating P/V. (This might be
equivalent
to your second approach.)

As far as where to put the code, it seems like the conversion
code would
be specific to the internal state representation used by the
propagator,
so I think it makes sense to keep the code as private to the
propagator.
Though if you think there would be other uses for the
conversion, I
think a public static factory method would work well.
Yes, the conversion code is propagator dependent. The first
implementation I played with is Eckstein-Hechler propagator, and it
definitely is Eckstein-Hechler specific (it's a simple
derivation of
the original equations, so each time we did compute something like
e =
a * c + b * d, now we also have another statement to compute eDot =
aDot * c + a * dCot + bDot * d + b * dDot). For numerical
propagator
we already have the derivatives since we start from the derivatives
and integrate them, so its even simpler. For DSST, this will be
a mix
as the mean elements are integrated and the short periodics
terms are
computed from Fourier coefficients which are straightforward to
differentiate. For ephemeris-based propagator, we will need to
compute
the derivatives of the underlying polynomials, which is also
straightforward.

Thanks Luc for finding this issue and doing the analysis. I can
see how
this would be an issue when computing the Doppler as well as time
shifting.
your welcome

best regards,
Luc

Best Regards,
Evan

On 10/29/2014 06:57 AM, MAISONOBE Luc wrote:
Hi Paul,

paulcefo <paulcefo@buffalo.edu> a écrit :

Luc,

Do I correctly understand that your concern is that Keplerian
transformations do apply outside the osculating space?
The problem I had was that we did use Keplerian-only expression
to set up local Taylor expansions around the current point (a few
seconds away). This was slightly wrong when all the parameters
were
time-dependent and not only the anomaly was time-dependent. Of
course,
the error increasing with the time offset with respect to the
central
date at which the Taylor expansion is built. The fix was
simply to
not forget the derivatives of these other parameters.

This Taylor expansion feature is a built-in feature available in
all
Orekit orbits, it typically allows to do computation in the
vicinity of
an already computed point without needed to trigger a complete
propagator.
It can even be used for some computation inside the run of a
propagator,
as for example when the higher level propagator takes care of
the long
term propagation and at each step we need some additional points
surrounding the current step to compute attitude evolution in
some
specific modes.

My concern was how to implement this fix in our current
architecture,
and more precisely were to put the code: in an existing class
or in
a dedicated class which would be used only by propagators.

best regards,
Luc

Paul

--
Dr. Paul J. Cefola
Consultant in Aerospace Systems, Spaceflight Mechanics, &
Astrodynamics
Adjunct Faculty, Dept. of Mechanical and Aerospace Engineering,
University at Buffalo (SUNY)

4 Moonstone Way
Vineyard Haven, MA 02568
USA

508-696-1884 (phone on Martha's Vineyard)
978-201-1393 (cell)

paulcefo@buffalo.edu
paul.cefola@gmail.com

On 10/29/2014 6:02 am, MAISONOBE Luc wrote:
Hello,

As some of you may be aware, I have been working for a few
months on
second order derivatives in the git branch
position-velocity-acceleration. This work is still ongoing but
I hope
to finish it for 7.0 and merge the branch back to master soon.
For
now, there are still failing tests so I can't do it.

This change should allow us to reach several goals :

- improved accuracy in shiftedBy methods
- improved accuracy in interpolators (with user-defined
  choices to use or not first and second derivatives
  from the sample)
- improved accuracy in attitude
- removal of ugly hidden finite differences in some classes
  (most notably attitude modes) with hard-coded steps
- hopefully faster Earth transforms, by replacing Hermite
  interpolation with single point extrapolation
- availability of non-Keplerian acceleration everywhere
- availability of angular acceleration in attitude and frames
- proper composition of dynamics in frames
- possibility to propagate orbits in non-inertial frames
- possibility to propagate orbits without a central body
  (interplanetary missions, Lagrange point missions, ...)

There is one point that bothers me right now. As I removed
some of
the
ugly finite differences, some non-regression tests started to
fail. I
finally found the raw cause of these failures and was
surprised to
discover an old bug in the way we use the osculating orbits
produced
by the Eckstein-Hechler analytical propagator. This propagator
takes
zonal terms into account, and produces directly circular
parameters a,
ex, ey, ... When we compute anything related to geometry, we
compute
Cartesian coordinates using the Orbit getPVCoordinates method.
As the
Orbit classes do not know anything about the perturbation, the
(P, V)
pair does in fact implicitly relies on Keplerian-only
expressions. So
the velocity part is *not* consistent with the derivative of
the
position. The real derivative of the position takes the
non-Keplerian
effects into account which are ignored by getPVCoordinates. The
difference is small, but as the tests threshold were
deliberatly very
tight, the tests started to fail when the various pointing
directions
were not computed anymore from finite differences mainly
involving
position and when they relied on the computed velocity. So the
problem
already happens in the master branch, it is not specific to the
introduction of acceleration (it was just detected here during
testing).

The solution is in fact quite simple. If an orbit has been
produced by
a non-Keplerian propagator, the propagator already knows about
the
derivatives of the orbital elements (which are circular in the
Eckstein-Hechler model case but can be any kind of
parameters for
other propagators). The propagator should therefore provide
these
derivatives to the orbit so they can be used in the
PVCoordinates
conversion. The code is very simple and straightforward. I have
checked this and got very interesting results with
Eckstein-Hechler/Circular, as for example a simple
interpolation over
a 900s arc with proper velocity/acceleration has a 88m error
with two
base points now whereas it was 5162 m before (and 0.02m vs
650m for 3
points, 1.0e-5m vs 259m for 4 points).

Here is what bothers me:

Should we create specialized classes for perturbed orbits or
should we
simply add a constructor to the existing orbits with the
parameters
derivatives and set them to 0 when they are not known?

For my tests, I created PerturbedCircularOrbit which extends
CircularOrbit and override the protected initPVCoordinates
method and
the public shiftedBy and interpolate methods. I could also have
simply
moved everything into CircularOrbit with a new constructor.

I do not like much the PerturbedXxxxOrbit approach, as it
forces to
create also additional entries in the OrbitType enum with
additional
converters and it becomes awkward if for example a user
configures a
NumericalPropagator to generate XxxxOrbit, despite this
propagator
will in fact really generate PerturbedXxxOrbit because it is
what a
Numerical propagator is for. So there should be either an
internal
modification of the user setting from OrbitType.XXXX to
OrbitType.PERTURBED_XXXX or an error triggered which would
invalidate
*all* current user code as it would become forbiddent to
generate
XXXX
orbits now.

On the other hand, the drawback of modifying the existing
classes to
hold the non-Keplerian derivatives is that they will consume
more
memory. I don't think it is a problem with current computers.

In any case, initial orbits created directly from user code
or by
reading files would not include the derivatives and therefore
will be
built as usual (by calling the unmodified classes in the first
approach, or by using the already existing constructors in the
second
approach, assuming these constructors will automatically set
the
derivatives to Keplerian-only values). In any case, full-blown
perturbed orbits will be created internally by Orekit
propagators,
which can easily be modified to provide the derivatives they
know (by
creating instances of the new derived classes in the first
approach,
or by using new constructors with additional parameters in the
second
approach).

My humble opinion would be to use the second approach to solve
this
bug. I will probably do this in the
position-velocity-acceleration
branch so it will include accelerations right from the start
and will
be merged to master at the same time as the rest of the
branch. Of
course, this will be a dedicated commits (Git branches are
great!).

What do you think ?

best regards,
Luc

----------------------------------------------------------------

This message was sent using IMP, the Internet Messaging
Program.



----------------------------------------------------------------
This message was sent using IMP, the Internet Messaging Program.





----------------------------------------------------------------
This message was sent using IMP, the Internet Messaging Program.







----------------------------------------------------------------
This message was sent using IMP, the Internet Messaging Program.







----------------------------------------------------------------
This message was sent using IMP, the Internet Messaging Program.







----------------------------------------------------------------
This message was sent using IMP, the Internet Messaging Program.