Department of

AUTOMATIC CONTROL

FRTN10 Multivariable Control

Exam 2016-10-25, 08:00–13:00

Points and grades

All answers must include a clear motivation and a well-formulated answer. Answers may
be given in English or Swedish. The total number of points is 25. The maximum number
of points is specified for each subproblem.

Accepted aid
The textbook Glad & Ljung, standard mathematical tables like TEFYMA, an authorized
“Formelsamling i Reglerteknik”/”Collection of Formulas” and a pocket calculator. Hand-
outs of lecture notes and lecture slides (including markings/notes) are also allowed.

Results
The result of the exam will be entered into LADOK. The result as well as solutions will
be available on the course home page:
http://www.control.lth.se/course/FRTN10

1
Solution to Exam in FRTN10 Multivariable Control 2016-10-25
1. Let
s −2(s + 1)
 
s+1 s+2 
G(s) =  
 1 
3
s+2
a. Compute the poles of G(s). Also state their multiplicity. (1 p)
b. Compute the (transmission) zeros of G(s). Do they impose any fundamental perfor-
mance limitations? (1 p)

Solution
a. The 1 × 1 minors of G(s) are
s −2(s + 1) 1
, , 3,
s+1 s+2 s+2
and the 2 × 2 determinant is
s 1 −2(s + 1) 6s2 + 13s + 6
· −3· =
s+1 s+2 s+2 (s + 1)(s + 2)
The least common denominator all subdeterminants is (s + 1)(s + 2), thus the poles
of G(s) are −1 and −2, both with multiplicity 1.
6s2 + 13s + 6
b. The maximal minor is the 2 × 2 determinant . The zeros are given by
(s + 1)(s + 2)
the roots of the numerator polynomial, which are
13 5
s=− ±
12 12
giving the zeros −3/2 and −2/3. The zeros are not in the RHP and hence do not
impose any fundamental limitations.

2. Consider the stable system

! !
−1 0 6 2
ẋ = x+ u
−0.1 −5 0 1
y = Cx

a. Calculate the controllability Gramian Sx of the system. Is the system controllable?

Which state is more difficult to control? (2 p)
b. The observability Gramian of the system is
!
1 0
Ox =
0 10
Calculate the Hankel singular values of the system. Which state would you truncate
and how large could the error
||y − yr ||2
||u||2
become if we were to keep the static properties of the system, but reduce it to first
order? Here, yr denotes the output signal from the reduced order system. (2 p)

2
Solution

a. The controllability Gramian is the symmetric 2 × 2 matrix Sx that solves the Lya-
punov equation ASx + Sx AT + BB T = 0. Straightforward calculations give
!
20 0
Sx =
0 0.1

The system is controllables since the controllability Gramian is non-singular. The

second state is much harder to control than the first one, the second eigenvalue of
Sx being 200 times smaller than the first one.

b. We have !
20 0
Sx Ox =
0 1
Since Sx Ox is diagonal, the squared Hankel
√ singular values are immediately given
by the diagonal elements, implying σ1 = 20 and σ2 = 1. You would keep the state
corresponding to the larger singular value, i.e. x1 . The error bound is

||y − yr ||2
≤ 2σ2 = 2
||u||2

if we reduce the system to first order.

3. A commonly used multivariable model of a distillation column was derived by Wood

and Berry in 1973. The model is

12.8e−s −18.9e−3s
 
 1 + 16.7s 1 + 21s 
G(s) = 
 6.6e−7s

−19.4e−3s 
1 + 10.9s 1 + 14.4s
a. Calculate the Relative Gain Array in stationarity and decide how the input–output
pairing should be done for decentralized control. Will there be visible interaction
between the two control loops? (2 p)

b. Find suitable matrices W1 and W2 such that G̃ = W2 GW1 is decoupled in stationar-

ity. (1 p)

c. For each of the four statements below, explain whether it is true or false. (2 p)

i. After the decoupling in subproblem b, there will be no visible cross-coupling in

the closed-loop step responses.
ii. By finding suitable decoupling matrices W1 and W2 we modify our physical
process so that it becomes diagonal.
iii. The fundamental limitations imposed by the process deadtimes can be removed
by properly chosen decoupling matrices.
iv. The decoupling approach allows us to design SISO controllers while still taking
the MIMO process information into account.

Solution

3
 a. The static gain matrix is !
12.8 −18.9
G(0) =
6.6 −19.4
from which the RGA in stationarity becomes
!
2.0094 −1.0094
G(0) .* G(0) −T
=
−1.0094 2.0094

From this we can conclude that we should couple u1 ↔ y1 and u2 ↔ y2 . However,

the RGA is far from an identity matrix, so there will be visible interation.

b. The matrices could for instance be chosen as W1 = I and W2 = G(0)−1 .

c. i. False! G̃ is only decoupled in stationarity, so all dynamic behavior will still

give rise to cross-couplings. Also, the outputs of the closed loop system G is
not necessarily decoupled just because the outputs of G̃ are. This we saw in
for instance lab 2, where the tank levels were still coupled, but the sum and
difference of them were decoupled.
ii. False! With the decoupler we do not change the physical process at all, just the
control structure.
iii. False! Fundamental limitations cannot be removed by any controller structure.
iv. True. The decentralized structure means that we design SISO controllers, and
the decoupler takes the MIMO cross-coupling in stationarity into account.

d n

r e u v z y
Σ C(s) Σ P (s) Σ

−I

Figure 1 MIMO control system in Problem 4.

4. Consider the MIMO control system in Figure 1, where P and C are the (matrix)
transfer functions of the process and the controller, respectively.
!
r
a. Calculate the (matrix) transfer function from to u in terms of P and C. (1 p)
d

b. Suppose that P (s) has 2 inputs and 3 outputs. What dimensions must then r, d and
C(s) have? (1 p)

c. The singular value (sigma) plots of P and C are shown in Figure 2. Can stability of
the closed-loop system be guaranteed using the Small Gain Theorem? (1 p)

Solution

4
 Singular values of P Singular values of C
1 1
Singular Values (abs) 10 10

Singular Values (abs)

0 0
10 10

-1 -1
10 10

-2 -2
10 10

-3 -3
10 10
-1 0 1 2 -1 0 1 2
10 10 10 10 10 10 10 10
Frequency (rad/s) Frequency (rad/s)

Figure 2 Sigma plots in Problem 4c.

a. From the block diagram, setting n = 0, we see that

U = C(R − P (D + U ))

Solving for U we obtain

  R

U =  (I + CP )−1 C −(I + CP )−1 CP


 

 
D


b. C(s) must have 3 inputs and 2 outputs, r must be a vector of size 3 and d must be
a vector of size 2.

c. No. The gain of each system is given by the maximum of the largest singular value.
The gain of P is larger than 0.7 and the gain of C is larger than 3. The loop gain is
hence larger than 1 and stability can not be asserted using the Small Gain Theorem.

5. A young student who has only taken a basic course in control has attempted to design
a controller for the process

2−s
P (s) =
s(s2 + 5s + 12)

The controller C(s) was designed as a state feedback from an observer. The control
poles were placed in −7 ± i and −8 and the observer poles in −14 ± 2i and −16.
The student proudly proclaims: “Look at these poles! I have designed a very fast and
well-damped closed-loop system!”.
You become suspicious and ask the student to plot the magnitude of the sensitivity
function S. The result in shown in Figure 3.

a. The student has forgotten all about the sensitivity function and its interpretation.
Explain to him/her why this sensitivity function is a sign of very poor robustness.
Also explain how you could immediately realize that it should not be possible to
design a very fast closed-loop system for this plant. (2 p)

5
 |S|
10 1

Magnitude (abs)
10 0

10 -1
10 -1 10 0 10 1 10 2 10 3
Frequency (rad/s)

Figure 3 Magnitude plot of the sensitivity function in Problem 5.

b. Using the sensitivity weighting function

s + Ms ω0
Ws (s) = , Ms , ω0 > 0
Ms s

show that the specification

|S(iω)| ≤ |Ws−1 (iω)|, ∀ω

is impossible to fulfill for any value of Ms if ω0 > 2. For what values of ω0 is it

impossible to fulfill Ms ≤ 1.4? (2 p)

Solution

a. The inverse of the maximum sensitivity, 1/Ms , measures the minimum distance be-
tween the Nyquist curve and the critical point −1. Here, Ms ≈ 6, implying that the
robustness is poor. The guaranteed amplitude margin is only Am = MMs −1
s
= 1.2.
The plant has a non-minimum-phase (NMP) zero in 2. The rule of thumb for unstable
zeros then says that the bandwidth of the closed-loop system cannot be faster than
2 rad/s (and should not be faster than 1 rad/s to ensure Ms ≤ 2).

b. The Maximum Modulus Theorem implies that the specification kWs Sk∞ ≤ 1 is
impossible to fulfill if |Ws (z)| > 1, where z is the location of the NMP zero. We have

2 + Ms ω0
|Ws (2)| =
2Ms

which is always greater than 1 if ω0 ≥ 2.

With Ms = 1.4 it is impossible to fulfill the specification if

2 + 1.4ω0
|Ws (2)| = >1 ⇒ ω0 > 0.57
2 · 1.4

6
6. A controller derived from the standard LQG framework will not automatically fea-
ture integral action. One way of approximating it is to add a noise model to the
Kalman filter where the noise is assumed to have a very high spectral density for low
frequencies. Assuming that the initial model of the system is

ẋ(t) = Ax(t) + Bu(t) + v1 (t)

z(t) = Cx(t)
y(t) = Cx(t) + v2 (t)

we would like to extend it to

ẋ(t) = Ax(t) + Bu(t) + v1 (t)

z(t) = Cx(t) + w(t)
y(t) = Cx(t) + v2 (t) + w(t)

where w is noise with high spectral density for low frequencies. We model w as white
1
noise n filtered through H(s) = s+δ , i.e.

w = Hn

Here, δ > 0 is some small number.

a. Explain why we do not use a pure integrator, i.e. H(s) = 1s , in our noise model when
designing an LQG controller. (1 p)
b. Extend the process model with the noise model, such that it attains the form

ẋe (t) = Ae xe (t) + Be u(t) + v1e (t)

z(t) = Ce xe (t)
y(t) = Ce xe (t) + v2 (t)

where v1e = [ v1T n ]T . Express Ae , Be and Ce in A, B and C. (1 p)

c. Now assume that A = −1, B = 1, C = 1 and δ = 0.001. Which one of the three
controllers A, B or C in Figure 4 could be the transfer function of an LQG controller
based on the extended model? You can assume that the low- and high-frequency
asymptotes of the controllers are visible in the plot. (1 p)

Solution

a. We would ideally like to have a noise model which has an infinite amplification for
static signals, i.e. a pure integrator to achieve a true integral action in our LQG
controller. However, the extended model will not be stabilizable (we can’t affect
the noise model integrator state with the control signal) with a pure integrator. To
compute an LQG controller we require that our system model is stabilizable. If we
1
instead use H = s+δ , the system model is stabilizable since the noise model now is
asymptotically stable.
b. A state space representation of the noise model is:

ẋw (t) = −δxw (t) + n(t)

w(t) = xw (t)

7
 4
10
Controller A
Controller B
3 Controller C
10

2
Magnitude 10

1
10

0
10

−1
10

−2
10
−4 −2 0 2 4
10 10 10 10 10
Frequency [rad/s]

Figure 4 Bode diagram of three possible candidates for the LQG controller in Problem 6.

Inserting this into the initial model yields the extended model:
" # " #" # " # " #
ẋ(t) A 0 x(t) B v1 (t)
= + u(t) +
ẋw (t) 0 −δ xw (t) 0 n(t)
" #
x(t)
z(t) = [ C 1]
xw (t)
" #
x(t)
y(t) = [ C 1] + v2 (t)
xw (t)

That is
" # " #
A 0 B
Ae = Be = Ce = [ C 1]
0 δ 0

c. The three controllers only differ for low frequencies. With the addition of the noise
model we expect an LQG controller which has a high gain for low frequencies, and
thus controller C can’t be an LQG controller based on the extended model.
With true integral action the controller’s static gain would be infinite. However, we
can’t achieve true integral action with this method, although we can get arbitrarily
close by letting δ approach zero. Thus the controller gain has to level out eventually
for low frequencies, at some large (but not infinite) gain. Since controller B does not
level out, only controller A can be a possible LQG controller based on the extended
model.

7. Consider the closed-loop system in Figure 5, where P0 is a SISO system. A controller

should be designed to attenuate the effect of input disturbances d on the process
output x, while keeping the effect of measurement noise n on the control signal u to
a minimum.

8
 d

y u x
−C(s) Σ P0 (s)

n
Σ

Figure 5 Block diagram of the closed-loop system in Problem 7.

a. Define z = [ x u ]T and w = [ d n ]T , and rewrite the system in Figure 5 to the

more general form in Figure 6. Provide the expressions for Pzw , Pzu , Pyw and Pyu .
(1 p)
b. Determine the closed-loop system from w to z in terms of P0 and C. (1 p)
1 1
c. Now assume P0 (s) = s+2 . Two controllers, C1 (s) = s(s+3) and C2 (s) = 2, have been
found to achieve stable closed-loop systems. Determine the Q-parameterization for
each controller, where the parameter Q(s) should be a stable transfer function. The
resulting closed-loop system should be linear in Q(s). (1 p)
d. The response of the closed-loop system is tested for a step input in d and an impulse
in n for each controller. Also, the L2 -gain of the closed-loop system is computed for
each case. The results are:

Evaluation C1 C2
Maximum value of x(t) after unit step in d. 0.44 0.25
Minimum value of u(t) after unit impulse in n. −0.14 −2.0
L2 -gain of closed-loop transfer function from n to u. 1.0 0.5

Determine a controller which achieves x(t) ≤ 0.4 after a unit step in d and u(t) ≥ −0.8
after an impulse in n. Additionally, the maximum allowable L2 -gain of the transfer
function from n to u is 0.95.

Note: You can express your controller in terms of the Q-parameters Q1 (s) and
Q2 (s) for the controllers C1 (s) and C2 (s) respectively. (1 p)

Solution
a. With the given definition of z and u, by analyzing the block diagram we obtain

Pxd Pxn Pxu P0 0 P0

   
" #
Pzw Pzu
P = =  Pud Pun Puu  =  0 0 1 
   
Pyw Pyu
Pyd Pyn Pyu P0 1 P0

where
" # " #
P0 0 P0
Pzw = , Pzu = , Pyw = [ P0 1], Pyu = P0
0 0 1

9
 z w
" #✛
Pzw (s) Pzu (s)
✛

Pyw (s) Pyu (s) ✛

y u
✲ −C(s)

Figure 6 General closed-loop system in Problem 7.

b. The closed-loop system H from w to z is:

" # " #
P0 0 C P0
H = Pzw − Pzu C(1 + Pyu C) −1
Pyw = − [ P0 1]
0 0 1 + P0 C 1
" P0 P0 C #
1+P0 C − 1+P 0C
= P0 C C
− 1+P 0C
− 1+P 0C

c. Every controller C(s) which stabilizes the system can be parameterized with a stable
transfer function Q(s) as:
Q(s)
C(s) =
1 − Q(s)P0 (s)
We can thus calculate Q(s) for a given controller C(s):
C(s)
Q(s) =
1 + P0 (s)C(s)
This parametrization results in a closed-loop system which is linear in Q(s). For C1
we get:
C1 (s) s+2
Q1 (s) = = 3
1 + P0 (s)C1 (s) s + 5s2 + 6s + 1
and C2 :
C2 (s) 2(s + 2)
Q2 (s) = =
1 + P0 (s)C2 (s) s+4

d. All of the specifications are convex in Q(s), and either Q1 (s) or Q2 (s) (parameterizing
C1 (s) and C2 (s) respectively) fulfills either of the specifications. Thus we can search
for a convex combination of Q1 (s) and Q2 (s) which fulfills all of the specifications.
One such convex combination is Q3 (s) = 0.7Q1 (s) + 0.3Q2 (s) since it will result in:

max x(t) ≤ 0.7 · 0.44 + 0.3 · 0.25 = 0.383 ≤ 0.4

min u(t) ≥ 0.7 · (−0.14) + 0.3 · (−2.0) = −0.698 ≥ −0.8
P0 C3

1 + P C ≤ 0.7 · 1.0 + 0.7 · 0.5 = 0.85 ≤ 0.95

0 3 ∞

10
With this Q3 (s) we have the resulting controller

Q3 (s) 0.7Q1 + 0.3Q2

C3 (s) = =
1 − Q3 (s)P0 1 − (0.7Q1 + 0.3Q2 )P0

11