You are on page 1of 5

1.

a. The matrix A has dimensions 2 $ 2, so there are 2 states. The matrix B has
two columns, so there are 2 inputs. The matrix C has two rows, so there
are 2 outputs.
b. The poles are given by the eigenvalues λ of the A-matrix, computed as
det(λ I − A) = 0
which gives us
det(λ I − A) = det 

λ +1

−1
λ +1

0 

= (λ + 1)2 = 0.

The system has a double pole in s = −1.

The zeros of the system are given by det G (s) = 0, where
G (s) = C(sI − A)−1 B + D =
We have
det G (s) =

1
s+ 1
1
s+ 1

1
(s+1)2
s+ 2
(s+1)2

!

.

s+2
1
1

=
,
(s + 1)3 (s + 1)3
(s + 1)2

and we can see that the system does not have any zeros.
c. The controllability matrix is defined as
Wc = ( B

AB ) = 

1

0

0 α

−1 α
0 −α 

For α ,= 0 the two rows of Wc are linearly independent which implies that
the system is controllable. For α = 0, the system is not controllable.
2.
a. The constraint can be rewritten as

p S(iω ) F −1 (iω )p ≤ 1 ∀ω
i.e.
sup p F −1 (iω ) S(iω )p ≤ 1
ω

so we have W (s) =

F −1 (s)

=

s+ 2
2s .

b. A constraint of the type
sup p Wa (iω ) S(iω )p ≤ 1
ω

+a
where Wa (s) = s2s
is impossible to satisfy unless p Wa ( z)p ≤ 1 for all unstable
zeros s = z of the process, i.e. a ≤ z for all such zeros. Here we have a = 2,
and there is an unstable zero at s = 1. Therefore, it is not possible to satisfy
the constraint on S(s) with any stabilizing controller.

1

3.
a. Step responses A and B have static errors in the load disturbance response,
which is caused by not having an integrator in the controller. If there is no
integrator, the low-frequency phase of L(s) should be the same as that for
P(s), i.e. −90○ as in controller 2 and 6. Step response B is faster than A,
which corresponds to a higher cross-over frequency as in controller 6.
Thus, B → 6, A → 2.
Step response D is unstable, which corresponds to a negative phase margin.
The phase of the open-loop system for controller 1 is below −180○ at the
cross-over frequency, so the closed-loop system cannot be stable.
Thus, D → 1.
Step response F has a slow response to the load disturbance, which corresponds to a small low-frequency gain. Of the remaining controllers 3, 4, 5,
the low-frequency gain is smallest for controller 5.
Thus, F → 5.
The remaining controllers 3 and 4 have cross-over frequencies at approximately ω c3 = 2.5 and ω c4 = 5.5. Higher cross-over frequency gives a faster
response, so controller 4 corresponds to step response E and controller 3 to
step response C.
Thus, E → 4, C → 3.
b. A pre-filter F (s) (typically a low-pass filter) applied to the reference signal could reduce the over-shoot. The pre-filter has no effect on the load
disturbance response.
c. Low values of the observability or controllability Gramians of a balanced
realization means that the corresponding states have little effect on the
input-output behavior of the system. We can see that some states have
very little effect compared to other states so it is likely that we could do
model-reduction on the controller to get a simpler lower-order controller
with almost the same properties.
4.

This system has two inputs and two outputs. A common approach to control
such systems is to pair each input to one of the outputs. All the three given
controllers use this approach. To determine a suitable input-output-pairing
the Relative Gain Array RG A( P(s)) is computed.

T

 0.01
4



(
s
+
2
)(
s
+
4
)
s
+
1
s
+
2






 =
.∗


8
32
0
s+ 4
s
+
4
 


32
0
0 1
(s + 2)(s + 4) 
(s+2)(s+4) 







=




32
32
1
0
0
(s+2)(s+4)


 0

RG A( P(s)) = P. ∗ ( P−1 )T = 
 8

−4
s+ 2
0.01
s+ 1

From the anti-diagonal structure of RG A( P(s)) we can conclude that we
should use u1 to control y2 and u2 to control y1 . The controller should then
have the structure represented in C2 and C3 .
To determine which of the two controllers C2 and C3 to use, we can examine
the poles of the closed-loop system. The closed-loop transfer function from
r to y is given by
G (s) = ( I + PC)−1 PC
2

The poles of this transfer function are given by the zeros of the matrix
I + PC, i.e. the values of s where this matrix looses rank. These values are
most easily determined from the equation
det( I + PC) = 0
Using C2 we obtain
det( I + P(s) C2 ) =

(s − 2)(s + 12)
(s + 2)(s + 4)

This equation has a zero for s = 2, i.e. if we use this controller G (s) will
have a pole in s = 2 and the system will become unstable.
Using C3 we obtain

det( I + P(s) C3 ) =

(s + 6)(s + 12)
(s + 2)(s + 4)

We see that the closed-loop system will have its poles at s = −6 and s = −12,
and it will then be stable. C3 is therefore the most suitable controller for
this process.
5.
a. In the first simulation, y1 and y2 are close to the point (3, 1), and in the
second simulation to (4, −1). Relating these points to the level curve plot,
we see that (4, −1) corresponds to lower cost. Thus, simulation 2 results in
lower emissions.
b. The system is given on the form
x˙ = Ax + Bu
y = Cx

and the cost function can be rewritten as
Z ∞
Z ∞
T
T
J=
y Q y y + u Qu udt =
x T C T Q y C x + uT Qu udt.
|{z}
| {z }
0
0
Q1

Q2

The optimal feedback gain L is found by finding the positive definite solution S to the Riccati equation
AT S + SA + C T Q y C − SB Qu−1 B T S = 0

and computing L as
L = Qu−1 B T S.

3

6.
a. We have

r

w =  Tout  u = ( h ) z =
n  

 

e
e
y=
h
Tout

b. We need to express all elements of the vectors z and y in terms of elements
of the vectors w and u.

e = r − Tin = r − n − G p h − G p Gv Tout = ( 1

Tout = Tout = ( 0

h = h = (0

0

1

r

− G p Gv

r

−1 )  Tout  − G p h
n

0 )  Tout  + 0 ⋅ h

r

n

0 )  Tout  + 1 ⋅ h
n

We can now write z and y as 

    

e
1 − G p Gv − 1
−Gp
z=
=
w+
u
h
0
0
0
1 
    

1 − G p Gv − 1
e
−Gp
=
y=
w+
u
0
1
0
0
Tout
and we have 

− G p Gv
Pzw =
0
0  

−Gp
Pzu =
1 

1 − G p Gv
Pyw =
0
1  

−Gp
Pyu =
0
1

−1
0 

−1
0 

7.
a. The optimal filter gain K is given by
1
K = ( PC T + N R12 ) R−
2

where P is the positive definite solution to the Riccati equation
1
T
T
T
AP + PAT − ( PC T + N R12 ) R−
2 ( PC + N R 12 ) + N R 1 N = 0.

4

We have A = −1, C = 1, N = 1, R1 = 10, R2 = 1, R12 = 0, which gives the
equation in P
P2 + 2P − 10 = 0
with solutions
P = −1 ±

11

with the unique positive solution P = −1 + 11 ( 2.32.

Thus, K = P = 2.32. The variance of the estimation error for the optimal
Kalman filter is equal to P, so var( x˜ ) = 2.32. 

y1
b. With y =
, the system can be written as
y2
x˙ = − x + u + v1 
  

1
v21
y=
+
1
v22 

  

1 0
1
The difference from the system in (a) is that C =
, R2 =
.A
1
0 1
standard Kalman filter can be used to estimate x from y, and the optimal
gain K is computed as in (a) with the modified C and R2 matrices.
c. Option 1:
With R2 = 0.5, the Riccati equation in P becomes
P2 /0.5 + 2P − 10 = 0,
which gives P1 = 1.79.
Option 2: 

  

1
1 0
With C =
, R2 =
, the Riccati equation in P becomes
1
0 1
2P2 + 2P − 10 = 0,
which is the same equation as in Option 1, thus giving P2 = P1 = 1.79.

Since the estimation error variance is equal to P it is the same for the two
options. Option 2 is much cheaper than option 1 and is probably the best
choice.

5