YouSub
Untitled

en English

Untitled English subtitles

0 downloads · 669 lines · added Aug 2, 2015

Report a problem
What's wrong?
Advertisement

English transcript

Hello everybody, we will continue our lecture series on LQR design. Last two lectures we
have seen this foundations of LQR followed by several and things like that and we will
follow up with further ideas on LQR design in this lecture as well.
So, primarily two things I want to discuss in this lecture, one is robust control design
through LQR augmentation ok. So, it is possible to have some sort of a robust control design
and we will see what the kinds are. When we talk about robust control, the robust control
can be from several factors actually, robustness with respect to let us say external noise,
robustness with respect to modeling and certainty, robustness is part of that is robustness with
respect to parameter variation something like that actually. Because no matter how much
information we can have or how much modeling you can do, they will be some sort of on model
dynamics of the time, which includes parameter uncertainties as well actually.
So, that kind of things can be handled through LQR and when we talk about external noise
into the system, we will see later that it is still possible to use LQR under the assumption
of kalman, I mean under the augmentation of kalman filtering basically. We know that means,
that essentially leads to this LQG design short of ideas actually, Linear Quadratic
Gaussian I think, you we did not talk about that here yet; we will talk about robustness
with respect to parameter of variation or some sort of a forcing information of sort
of thing actually, we will see that ok. And the second thing what you want discuss
here is solution of LQR problems using this state transition matrix approach sort of ideas.
So, what happens here is if you follow this record equation approach, you essentially
land up with N by N non-linear, either differential equation or algebraic equation in terms of
the recovery matrix actually, ok. So, solution becomes kind of and involve and many times
we will end up with numerical solution. So, that part of the computational basically,
so we do not have to do that one, you follow this state transition matrix approach, the
drawback here is that at the problem dimensionality increases from N to 2 N actually.
So, we consider this X dot and lambda dot something like together sort of ideas actually,
remember X is n dimension means lambda is also N dimensional lambda actually. So, you
increase the dimensionality, but kind of reduce the complexity sort of thing, so that is the
whole idea there. So, it is a different approach and I thought it is good to know that actually.
Under this umbrella, we will talk about soft constrained problems first, what we have discussing
so far and it is also possible to discuss hard constrained problem. That means, you
really want some part of the states to go to a desired value, that is in equality sense
and all that, it is possible to do that, there are certain drawbacks also doing that way,
but we will talk about that as we go along actually. So, let us get going first is robust
control design for systems with parametric inaccuracies through LQR augmentation that
is what we are discussing.
So, what we are discussing, normally write that X dot equal to f of X U for non-linear
system, assuming that, that set of parameter values are kind of constrained actually, they
do not vary or they are known basically. So, they do not play any significant role sort
of thing, but in general we can also write this X dot equal to f of X U P, where P contains
the parameter, I mean parameter vector actually. And for example, if you have a moving mass,
then X the dynamic equation is m X double dot equal to f; that means m is a parameter
of the system. Similarly, if a rotating body, moment of inertia becomes a parameter, if
it is aero aerodynamic controlled aircraft, then aerodynamic forces and moments there
are several coefficients, which can be thought about as parameter of the system dynamics,
think like that actually ok. So, those are the values that we are talking
about as part of the system dynamics actually and we have a non-linear system like this,
X dot equal to f of X U P and this what you are talking here is we already know a U star,
some method is there, I do not know we do not want to do alone on that, we assume that
U star is some control which is already designed, assuming there is a parameter vector P star.
That means, that is that is the knowledge which is known to us actually, in other words
if the mass of the vehicle something like 102 kg and we assume that it is 100 kg, that
means, that P star turns out be 100, the 2 we really do not know whether it is 100 102
105 or 98 whatever actually. So, that part is uncertain actually, but p
star is largely known, I mean exactly known the other, if you if you take p star is a
parameter vector which is completely known, then you can design a new star based on and
that will result in a next star trajectory. So, that means for a known set of parameter
vectors, we already have the controller ready basically, it can come from optimal control,
it can come from any other controller as well actually, we are not bothered about that at
this point of time actually. And one possibilities certainly optimal control,
non what nonlinearly optimal control that trajectory optimal ideas that we discussed
before actually alright, so this is what it is actually. So, now the, but the reality
is the mass is not 100 kg and it can be 100 200 598 whatever it is. So, we consider that
part is something like delta P, so P is actually a combination of P star, which is already
known to us and an unknown factor which is delta P actually ok. So, remember we are not
having any sensor of anything to measure that online, we just have to leave that error actually.
So, that is the concept, can we design something like some control or which will which will
actually.
So, what is the objective here? Objective can be described something like this, what
you assume first is delta P is small we do not have to talk about large uncertainty and
all that, we talk about small uncertainty, that too that is what I just told about actually,
if mass is 100kg because consider 100 to 105 like that, it is kind of plus or minus 5 percent
in accuracy or may be 110 kg or 10 percent accuracy like that. We do not talk about something
like 50 percent, 100 percent, 200 percent accuracy like that accuracy; that means, that
information is a different class of problem altogether and we do not want to talk about
that here, ok. But, largely it is like that, in a in a normal
situation well there is a small comment, large uncertainties are not unrealities, they are
also realities when they talk about now people talk about something call reconfigurable control.
If you have large uncertainties the system plant dynamics is largely different from what
we discuss actually, if for example, if your aircraft is flying and you have a vital damage,
that means half of the wing is gone or something lost and things like that, that kinds of uncertainties
are large uncertainties, we do not talk about that. What we are talking here is delta p,
that means, some parameter inaccurate information; that means, wind tunnel data some coefficients
have been computed through polynomial think like that, that polynomial numbers can be
inaccurate, but that inaccuracy is not suppose to go beyond 5, 10, 20 percent actually ok.
Similarly, mass moment of inertia cannot go much more than 5 10 percent actually. So,
that kind of inaccuracy we are talking about, anyway coming back what we are talking is
parameter vector P is can be described something like this, P star plus delta P, where P star
is known, delta P is unknown actually. So, delta P by assumptions remains small and it
will also remains constant, it does not vary with time as well, we assume that it remains
constant, we just do not know that part, but it does not vary with time, if mass is instead
of 100 it is 105, it is 105 it does not keep on varying 105 to 107 to 110 and think like
that on line actually, so that assumption is this.
So, what happens now if you have P instead of P star, your state will develop as X naught
X star, so we consider that our real state, I mean state variable X of t is something
like X star plus delta X actually. What is objective? The goal is to come up with an
extra control delta U such that the controller also needs to be a part of U should be equal
to U star plus delta U basically. So, remember what we have done here is we
have a plant, here we have a U star already designed for parameter vector P star, which
will result in X star. Now, P is not P star and which will result X is not X star and
hence we will need some other U not U star to compensate for all that actually. That
means, what we really require is to compute a delta U such that, U star plus delta U will
make sure that X if it is X star, so X goes to X star goes as soon as possible actually.
And why because X star is already available to us, we know that everything good about
that, because that is a nominal control design already done actually. So, in essentially
if we can design some delta U, so that u star plus delta will enforce this, the actual X
2 to follow X star then our job is done actually. In other words, we want this delta X 2 to
go to 0 actually that is the problem, so how do you do that? Remember delta p is not known
actually, now using the Taylor series expansion about X star u star p star and all that, you
can you can also write this actually right, if starting from this non-linear system dynamics
following the same principle of linearization of systems and again that is a very standard
procedure and some of you do not know how to linearise system of non-linear system,
why actually have a small discussion about that in the very beginning class, several
lecture in my other course as well as actually. Anyway so this is what it is, we can go back
to this thing and we have a nominal X star, nominal U star and nominal P star available,
so using those values we can talk about a linearization about those values actually.
So, we will end up with this kind of thing, where A star, B star and E star can be computed
this way, these are matrices can be computed that way ok. Just see the Jacobin matrices
and all evaluated at this nominal value actually. So, we have some sort of a linearism expression
all there alright. So, what is the idea here? Idea is to cancel
the effect of delta P basically, so if you have to cancel the effect of delta P, also
you need some sort of delta U steady state because remember P is a parameter, that delta
P is a constant value. So, delta X dot is there, so to cancel that effect, we need a
constant bias control sort of thing, that delta let us the talk that is some steady
state control actually, delta steady state bias control, the delta U ss.
So, assuming that is available which is not 0 anyway, so what we really need to have a
cost function is something like this, delta X should go to 0, so that means that part
is a quadratic function and remember delta u should not go to 0, but delta u minus delta
U ss should go to 0, because that this steady state control we need to nullify that we are
aware actually. So, this difference should go to 0 that means,
you can construct a cost function with this difference being a quadratic function actually
and ultimately remember as long as you talk about 0 to infinities sort of thing, the integrant
values should ultimately go to 0, otherwise the cost function will remain unbounded, it
will go to infinity and all that and which cannot be minimize actually ok.
So, that is a incompatible formulation a very quick check is whatever you are putting it
as a quadratic function here, that at some point of time it should go to 0 and it should
remain 0 actually. So, you put it quick and then tell this cost function can be minimize
actually; however, this turns out to be an I mean it is it is ok because this has to
go, but the, but the problem here is we do not know this value, delta U ss we do not
know actually. So, even if you compute a difference like the delta U minus delta U ss equal to
some express and think like that, it is virtually of no use because we cannot recover delta
U, because we do not know delta U ss actually; so that is the problem there, so solve that
is explained here actually ok.
Since delta U ss is unknown and hence the cost I mean cost function cannot be used in
practice, so let us find out some alternative ideas actually. So, as an alternative U what
will happen in steady state? Steady state delta X is going to 0 that is our objective
actually and delta X dot should also go to 0, this expression what you have here this
should go to 0 and this should also go to 0. So, let us put it back 0 equal to 0 plus
all that and hence delta U S can you can think about I can compute it basically, because
this is there and it is not a very good solution is a pseudo inverse involving and all that,
but still it in approximate sense I can compute it and as long as my control number of dimension
of the control variable are more than the dimensional parameter, which is sometimes
true sometimes not true whatever; if that is true this pseudo inverse is not good actually,
I mean pseudo inverse is not bad. So, then you can think about I can compute it and hence
I can use this cost function, but again wait a second delta P is fundamental nature of
the problem, the delta P is not known, ok. So, even though this expression is true in
general, we cannot compute it because the value for delta P is not known actually, so
this idea also fails basically. So, now it will look out for some other approach actually,
so let us see what this is. Now, for to further algebra will define some of these variables
which is typically non-linear systems also, we redefine this state variable, control variable
and all that, X is redefine as delta X and U is redefine as delta U like that. So, do
not have to keep on talking delta X delta U delta P all that actually.
So, we define it and we remember that when we talk about X, U, P and we are actually
talking about delta X, delta Y, delta P here, it typically done in linearization I mean,
linearization procedure actually, ok.
And similarly we define A, B, E as something like A star, B star, E star so that we do
not have to keep on talking about this star values also, we do not have to keep on writing
star unnecessarily, alright. And after this redefinition, we also define Z as a new variable
which is B U plus E P part, remember this is now X dot equal to A X plus B U plus E
P in our new definition and all. So, this part what we have here we define
it is Z actually, Z is B U plus E P total actually and we also define some auxiliary
control variable V, which is U dot actually and with these redefinitions, the system dynamics
turns out to be something like this, X dot equal to A X plus Z plus Z and Z dot is nothing
but Z is defined like that; remember P is a constant thing parameter and then E is constant,
this derivative will not be there that is 0, Z dot is B is also a constant matrix, so
Z dot equal to B times U dot actually, ok. So, but U dot by definition is V, so you can
put it has B V. So, X dot equal to A X plus Z and Z dot equal to B V actually; so, this
essentially leads to this, you have this X dot and Z dot is nothing but, this kind of
ideas. You can put this system dynamics together and tell A is there and then this part is
I, Z and X and Z are there and then you have the Z dot, Z dot is B also 0 0 and you have
a B actually.
So, X dot and Z dot is nothing but, A dot X plus P dot B, so that is this. So, that
is once I define this capital X now, this thing is forgotten to put that and this is
obvious actually, X dot is defined here anyway, so this capital X is this vector X and Z.
So, this capital X dot is something this one, this X dot is nothing but, A head times X
plus B head times V, again the similar to what we have discussed in the last class,
it turns out that if the pair A B is controllable, then this pair A head X also controllable
actually, ok.
So, that gives us a hope to proceed further with the LQR design. So, we now formulate
this kind of a thing, we have this cost function X transpose Q 1 X plus Z transpose Q 2 Z plus
V transpose R at V, because V is a control variable now, but remember actually V is nothing
but, U dot. So, we are forced to write because, this formulation new formulation V is the
control variable and R has to be there because R cannot be 0, it is an initial positive definite.
So, this R term is there, but physically it means that it take actually minimize we attempt
to minimize the control derivative also and this is also true in a physically meaning
full sense, because whenever we have a disturbance storm system and all, we do not want to kind
of make the control variables since of that too much actually.
So, we want to have this control rate minimization as well also, that is a philosophical argument,
but mathematical argument is because V is a control variable, we need to have some term
for V transpose V also and that is whatever it results in the physical interpretation
actually. So, we have X transpose Q 1 X and Z transpose Q 2 X, this is just because we
have to we can split this and interpret that way, this part is Q matrix largely and this
was a R matrix actually ok. Now, what happens in a steady state situation,
now is remember you go back to this, in a steady state situation, this entire term which
is now in a in a new notation, it is B times U plus E times B, B U plus E P is has go to
0 there is no doubt about that, that has to happen; just that here we are not able to
solve for delta U ss because delta P is not known, over in steady state this condition
should happen. So, this condition this is what is written here, in steady state Z has
to go to 0 ok. So, if we define this Q head something like
this, partition matrix Q 1 0 and 0 Q 2 like that and by definition this capital X and
Z, the cost function turns out to be like this. So, you have a cost function which is
quadratic in terms of capital X and quadratic in terms of V. So, we have this standard quadratic
cost functions and we need to see this standard linear dynamics also basically. So, with respect
to this standard the equation and this cost function, now it is problem is comfortable
actually. So, we can go ahead and solve it actually
ok, but also just remember this is a formulation which is quite similarly to the case, where
the intension was to minimize the derivative of control. We have discussed all that in
the in the previous lecture actually, you can some of you forget that, you can probably
revise that as well as actually, very similar to that the development happens to be somewhat
very similar to that. Anyway this is the cost function and this are the steady state linear
and this is a I mean cost function, which is quadratic. So, we can go head and use our
LQR theory so to get this one.
So, what ultimately my control variable is V here, so it is V equal to minus R inverse
R is a R head here, R inverse B transpose B is nothing but this vector, this vector
0 B transpose, R inverse B transverse this one, very knows B transpose record matrix
solution P head times capital X, capital X is AX and Z. So, my V is nothing but U dot
know, so I can substitute that is U dot something like this and remember Z is not computable
again, because we do not know this delta P, the P is nothing but delta P, remember that
actually ok. So, this is P is not known, so Z is also not
compatible actually. So, we cannot implement this and also remember this is a dynamic controller
U dot and think like that, so we do not want that to implement the actually. So, how do
you go about that? Now, the idea here is we know Z, Z is something like this, so what
is X dot? X dot equal to let me let me write it here probably. So, we have this X dot equal
to A X plus B U plus E P right, so that a
…

Need subtitles from another video?