1 Controllability and Observability: MAE 280 B 5 Maur Icio de Oliveira
1 Controllability and Observability: MAE 280 B 5 Maur Icio de Oliveira
1.1
State-space: x = Ax + Bu, y = Cx + Du. Dimensions: x Rn , Notation u Rm , A B C D H(s) = C(sI A)1B + D Note that H(s) is always proper! Similarity transformation: T 1AT T 1B A B CT D C D Similarity does not change transfer function H(s) = CT (sI T 1AT )1T 1 B + D = C(sI A)1B + D System response: Y (s) = H(s)U (s) + C(sI A)1x(0) Input MIMO comes for free! Initial conditions y Rp . x(0) = x0,
Transfer function:
MAE 280 B
1.2
X(t) =
0
eA BB T eA d.
Observability Matrix:
Observability Gramian:
C CA O(A, C) = . . . . CAn1
t
Y (t) =
0
eA C T CeA d.
MAE 280 B
1.3
Controllability
Problem: Given x(0) = 0 and any x, can one compute u(t) such that x(t) = x for some t > 0? Theorem: The following are equivalent a) The pair (A, B) is controllable; b) The Controllability Matrix C(A, B) has full-row rank; c) There exists no z = 0 such that z A = z , z B = 0;
d) The Controllability Gramian X(t) is positive denite for some t 0. 1.4 Observability
Problem: Given y(t) over t [0, ] with t > 0 can one compute x(t) for all t t [0, t]? Theorem: The following are equivalent a) The pair (A, C) is observable; b) The Observability Matrix O(A, C) has full-column rank; c) There exists no x = 0 such that Ax = x, Cx = 0;
d) The Observability Gramian Y = Y (t) is positive denite for some t 0. 1.5 Things you should already know
1. Why a) and b) are equivalent. 2. Why can we stop C(A, B) at An1B and O(A, C) at CAn1? 3. Kalman canonical forms. E.g. if (A, C) is not observable then A o 0 Bo A B Aoo Ao Bo C 0 Co 0 0 where Ao Rrr and (Ao , Co ) is observable.
7 MAE 280 B Maur de Oliveira cio
1.6
Theorem: The pair (A, C) is observable if and only if there exists no x = 0 such that Ax = x, Cx = 0. (1) Proof: Suciency: Assume there exists x = 0 such that (1) holds. Then CAx = Cx = 0, CA2x = CAx = 0, . . . CAn1x = CAn2x = 0 so that O(A, C)x = 0, which implies that the pair (A, C) is not observable. Necessity: Assume that (A, C) is not observable. Then transform it into the equivalent non observable realization where Ao 0 A= , Aoo Ao Chose x = 0 such that Ao x = x. Then Ao 0 Aoo Ao 0 x = 0 , x Co 0 0 x = 0. C = Co 0 .
MAE 280 B
1.7
Controllability Gramian
Problem: Given x(0) = 0 and any x, compute u(t) such that x(t) = x for > 0. some t Solution: We know that
t
x = x(t) =
0
(tt)
x=
0
eA(t ) BB T eA
t
(t )
z d, d z, = t
=
0 t
eA(t ) BB T eA
T
(t )
=
0
eA BB T eA d
z,
and
t 1
z=
0
eA BB T eA d
x, u(t) = B e
T A (tt)
T
t 0
eA BB T eA d
X(t) :=
0
eA BB T eA d
MAE 280 B
1.8
Stabilizability
Problem: Given any x(0) = x can one compute u(t) such that x(t) = 0 for some t > 0? Theorem: The following are equivalent a) The pair (A, B) is stabilizable; b) There exists no z = 0 and such that z A = z , + 0. 1.9 Detectability z B = 0 with
Problem: Given y(t) over t [0, ] with t > 0 can one compute x(t)? t Theorem: The following are equivalent a) The pair (A, C) is detectable; b) There exists no x = 0 and such that Ax = x, + 0. Cx = 0 with
MAE 280 B
10
1.10
Satellite of mass m with thrust in the radial direction u1 and in the tangential direction u2. From Skelton, DSC, p. 101. Newtons law
mr = u1 + u2 + fg ,
Using cylindrical coordinates e1 = we have r = re1 , u1 = u1 e1 , u2 = u2e2 , fg = km e1 . r2 cos , sin e2 = sin , cos
MAE 280 B
11
We need to compute
2 = d (re1 ) = d (re1 + re1 ) = re1 + 2re1 + re1 , r dt2 dt
so that Newtons law can be rewritten as km m( r2)e1 + m(2r + r)e2 = u1e1 + u2 e2 2 e1 , r r or, equivalently, as the two scalar dierential equations km m( r2) = u1 2 , r r m(2r + r) = u2 . In state space r r 0 0 x3 0 u1 x4 0 + , x = = x= u2 . r x1x2 k/x2 1/m r 0 4 1 2x3x4/x1 0 1/(mx1) This is a nonlinear system and we look for equilibrium ( = = 0) when r u1 = u2 = 0. This can be stated as x1x2 k/x2 = 0, 4 1 2x3x4/x1 = 0. The second condition implies x3 = r and/or x4 = must be zero. We choose x3 = r = 0 which implies x1 = r = r constant and x4 = = k = x3 1 k = r3 k = r32 .
This nonlinear system is in the form x = f (x, t) + g(x)u. We will linearize f (x, t) and g(x)u around the equilibrium point (, u) to x obtain the linearized system x = (fx)T [x(t) x(t)] + (gx )T [x(t) x(t)] + g()u. u x For this problem r 0 0 x3 t x 0 , g(x) = 0 , x(t) = , u = 0, f (x, t) = 2 4 0 1/m x1x4 k/x2 0 1 2x3x4/x1 0 1/(mx1) and 0 0 (fx )T = 2 k/3 + 2 r 0 0 1 0 0 0 0 0 1 = 0 0 2 r 3 2 0 2 / 0 r 0 0 1 0 0 0 1 . 0 0 2 r 0 2 / 0 r
If we looking at the satellite (from the earth) we can say that we can observe r and (how?), that is y= Questions: Can we estimate the state of the satellite is by measuring only r? Can we estimate the state of the satellite is by measuring only ? Can we estimate the state of the satellite by measuring r and ? Can the system be controlled to remain in circular orbit using radial thrusting (u1) alone? 5) Can the system be controlled using tangential thrusting (u2) alone? 1) 2) 3) 4) 1 0 0 0 x. 0 0 0 1
MAE 280 B
13
Question: Can we estimate the state of the satellite by measuring only r? Answer: Is the system observable when C = 1 0 0 0 ? Compute the observability matrix C CA O(A, C) = CA2 , CA3 1 0 0 0 0 0 1 0 = 2 3 0 0 2 r 0 0 2 0 Physical interpretation: measuring r does not give any information on or ! Note that if we that know the satellite is in equilibrium and measure k then == But we still do not know since (t) = (0) + t, and we do not know (0)! k . r3
MAE 280 B
14
Question: Can we estimate the state of the satellite by observing only? Answer: Is the system observable when observability matrix C CA O(A, C) = CA2 , CA3 0 0 = 6 3/ r 0 C = 0 0 0 1 ? Compute the
0 0 1 0 2 / r 0 0 0 4 2 0 2 3/ r 0
Physical interpretation: again, if we try to reconstruct from we still need to know at some t! From that point on
t
= (t) +
( )d.
MAE 280 B
15
Question: Can we estimate the state of the satellite by measuring r and ? Answer: Is the system observable when C = observability matrix C CA O(A, C) = CA2 , CA3 1 0 0 0 = 3 2 6 3/ r 0 0 1 0 0 0 ? Compute the 0 0 0 1
0 0 0 0 0 1 0 1 0 0 2 / r 0 0 0 2 r 0 0 4 2 2 0 0 0 2 3/ r 0
MAE 280 B
16
Question: Can the system be controlled to remain in circular orbit using radial thrusting (u1) alone? 0 0 Answer: Is the system controllable when B = 1/m? Compute the 0 controllability matrix O(A, C) = B AB A2B A3 B , 1 0 2 0 1 0 0 2 / r 0 = 0 2 0 m 1 0 2 / r 0 2 3/ r Note that 1 2 0 0 = 2 0 0 2 / r 2 3/ r
which implies that the system is not controllable from u1 ! Physical interpretation: there must be a change in the angular velocity if one changes the radius!
MAE 280 B
17
Question: Can the system be controlled using tangential thrusting (u2) alone? 0 0 Answer: Is the system controllable when B = 0 ? Compute the 1/(m) r controllability matrix O(A, C) = B AB A2B A3 B , 2 r 0 0 0 1 0 1 0 4 2 = 0 2 r 0 2 3 r m r 2 0 1 0 4
The above matrix has full rank, so the system is controllable from u2! Physical interpretation: we can change the radius by changing the angular velocity!
MAE 280 B
18
1.11
More on Gramians
X(t) =
0
eA BB T eA d,
is the solution to the dierential equation d X(t) = AX(t) + X(t)AT + BB T . dt If X = limt X(t) exists then AX + XAT + BB T = 0. Theorem: The Observability Gramian
t
Y (t) =
0
eA C T CeA d,
is the solution to the dierential equation d Y (t) = AT Y (t) + Y (t)A + C T C. dt If Y = limt X(t) exists then AT Y + Y A + C T C = 0.
MAE 280 B
19
e BB e
T AT
d d = dt
t 0
eA(t ) BB T eA
(t )
d, ,
=
0
d A(t ) T T e BB T eA (t ) + eA(t ) BB T eA (t ) dt
t
=t
=A
0
eA(t ) BB T eA
t
(t )
d
T
+
0
eA(t ) BB T eA
(t )
AT + BB T ,
= AX(t) + X(t)AT + BB T . For the second part, use the fact that X(t) is smooth and therefore
t
lim X(t) = X
d X(t) = 0. t dt lim
MAE 280 B
20
Lemma: Consider the Lyapunov Equation AT X + XA + C T C = 0 where A Cnn and C Cmn . 1. A solution X Cnn exists and is unique if and only if j (A) + (A) = 0 i for all i, j = 1, . . . , n. Furthermore X is symmetric. 2. If A is Hurwitz then X is positive semidenite. 3. If (A, C) is detectable and X is positive semidenite then A is Hurwitz. 4. If (A, C) is observable and A is Hurwitz then X is positive denite. Proof: Item 1. The Lyapunov Equation is a linear equation and it has a unique solution if and only if the homogeneous equation associated with the Lyapunov equation admits only the trivial solution. Assume it does not, that is, there X = 0 such that AT X + XA = 0 Then, multiplication of the above on the right by x = 0, the ith eigenvector of i A and on the right by x = 0 yields j 0 = xAT Xxj + xXAxj = [j (A) + (A)] xXxj . i i i i Since i (A) + j (B) = 0 by hypothesis we must have x Xxj = 0 for all i, j. i One can show that this indeed implies X = 0, establishing a contradiction. That X is symmetric follows from uniqueness since 0 = (AT X + XA + C T C)T (AT X + XA + C T C) = AT (X T X) + (X T X)A so that X T X = 0. Item 2. If A is Hurwitz then limt eAt = 0. But
X=
0
eA t C T CeAt dt
0
0
and
A X + XA = lim
d AT t T At T e C Ce dt = eA t C T CeAt dt
21
= C T C.
MAE 280 B
Item 3. Assume (A, C) is detectable, X 0 and that A is not Hurwitz. Then there exists and x = 0 such that Ax = x with + 0. But if X solves the Lyapunov equation xC T Cx = xAT Xx + xXAx = ( + ) xXx 0 which implies Cx = 0, hence (A, C) not detectable. Item 4. Assume that (A, C) is observable and A is Hurwitz. From Item 2. X 0. Assume X is not positive denite, that is, there exists x = 0 such that X x = 0. It follows that
0 = x Xx =
0
xe
AT t
C Ce x dt =
0
At
y (t)y(t) dt
which implies that response y(t) = CeAt x = 0 to a non null initial condition x(0) = x is null, which contradicts the hypothesis that (A, C) is observable.
MAE 280 B
22