You are on page 1of 13

IN 227 Control Systems Design

Lecture 12

Instructor: G R Jayanth
Department of Instrumentation and Applied Physics
Ph: 22933197
E-mail: jayanth@isu.iisc.ernet.in

Drawbacks of 1 DOF control

We have spent all the time thus far studying analysis and design of 1 DOF controllers,
d
where we have just one transfer function to tune. We studied how general
u
x
controllers can be designed to reject a known disturbance or track a given input , and r
P( j)
C ( j)
how specific controllers improve performance when we can model the disturbance or - e
the plant well. Indeed a lot can be achieved with just one transfer function C(j) that
is available for tuning.
As engineers we seek simplicity and clarity and need to avoid adding any more
controllers than the barest minimum that gets the job done. However, there are very
practical reasons why just one controller may be inadequate to achieve the desired
d
control performance. Some of these reasons are discussed below.
+ x
r
u
Measurement noise: Since the use of a sensor is central to feedback, one of the
P( j)
C ( j)
unintended consequences of feedback control is to bring measurement noise into the - e
system and cause it to affect the output. The closed-loop transfer function on the
+
right shows that the input-output relationship for reference and noise are the same.
n
1
1
R
N
Thus, our desired goal of tracking the reference perfectly also leads to the undesired X
1
1
1
1
result of letting in a lot of noise. If the noise is too large, there would be tradeoff
C ( j) P( j)
C ( j) P( j)
between control performance and output precision.
Dynamic performance under plant variations: If the plant parameters vary over a
wide range, we would need to specify large gain and phase-margins so that the
closed-loop system does not get destabilized for some parameter values. Such large
margins will slow down the closed-loop system. Thus, we would need to make
tradeoffs between the closed-loop response and stability.

Two degree of freedom control

The goal of two degree of freedom control is to introduce another compensator into
the feedback control structure, so that we have one more degree of freedom left
after designing for input tracking and rejection of one disturbance. Thus, it is an
attempt to alleviate some of the drawbacks of one degree of freedom control.
r
u
One possible way to do this is shown on the right: introduce another transfer
C ( s)
function F(s) in the feedback path. We now see that we have two distinct closed- e
loop transfer functions that we can tune: that between r(t) and x(t) and between
F ( s)
n(t) and x(t). The former is given by X/R=Gcl1=[C(s)P(s)]/[1+C(s)P(s)F(s)] and the
latter is given by X/N=Gcl2=[C(s)Ps)F(s)]/[1+C(s)P(s)F(s)]. The third transfer function,
between d(t) and x(t) is just 1-Gcl2.
However, the above configuration is not the only way to achieve two degrees of
u
freedom. Some of the other configurations are given on the right and many more r F (s)
C ( s)
have been proposed (See Ref. for other examples). The last configuration has a
- e
special name-conditional feedback because the feedback loop is active only if there
is error between the nominal and the actual plant models. Thus, one may get the
impression that by developing a sufficiently clever arrangement of the two
C (s) Gcl (s) / P(s)
compensators, we may be able to achieve performance that may ordinarily not be
r
C ( s)
possible.
u
However, mathematically, all of the different block arrangements enable tuning just
F ( s)
two distinct transfer functions. Thus, there is no difference between one or the
e
other and it is possible to achieve the exact same performance with one
Gcl ( s) +
configuration that we may achieve with another, more exotic, one.
For the same reasons, adding more controllers that can be tuned is also useless: no
matter how many more are added, the overall response would be decided by just
two distinct transfer functions, each of which contain the lumped version of all the
multitude of controllers that we have added.
More generally, if a particular plant has n access points, we can, in general, tune n
distinct transfer functions and no more. How we arrange our controllers to achieve
the tuning is a matter of choice and convenience.

d
P( s )

+
n

P( s )

d
x

+
n

P( s )

d
x

2 DOF Control applied to reject noise

Having pointed out that all the different 2DOF structures achieve the exact same
thing, we decide to stick with one of them. This is shown on the right. It can be
Feedback
d
thought to have two blocks-a feedback block and a pre-filter block. We will spend Pre-filter
r F ( s)
x
u P( s ) +
the rest of the lecture examining how the two can be designed.
C ( s)
- e
Example 1: One of the most striking problems with 1 DOF control configuration is
that the transmission of noise is the same as that of the reference. It is required to
+
n
suppress the effect of noise without affecting the transmission of the reference.
Solution: Let us start out by focusing only on noise and how its effect may be
reduced. Assuming that the noise is white, we note that the only way to reduce its
effect on the output is to limit the bandwidth of the closed-loop control system:
20log |CP|
that way, only the fraction of noise power that is within the pass band of the control
system appears at the output. We know from previous lectures that a sure-shot way
gain cross-over with
integrator
to reduce bandwidth is to use pure integral control. The Bode plot on the right,
Original plant
pulled from lecture#7, compares the control bandwidth of the original plant (in gray
dashes) with the integrator-controlled system. What we desire is to pull down the Desired
Log()
bandwidth even further, as shown on the right so as to limit the effect of noise. This
is achieved by reducing the gain KI of the integrator. When the control bandwidth is
much less than that of the plant, the plant can be modeled by just its DC gain K0.
u
x
Thus, the desired feedback block reduces to the one shown on the right. The closedK
K /s
e
loop transfer function is T1 (s) 1 (s K1 1) where K1=K0KI <<1.
+
While this has successfully addressed the issue of noise, it has dramatically reduced
n
the tracking performance of fast reference signals because of the ultra low closed1
T1 ( s)
loop bandwidth. If this was a 1DOF system, we can do no better: we either have to
s K1 1
suppress noise as above or increase the bandwidth to track faster reference signals.
However, with 2DOF control, we can do both. we now utilize the pre filter F(s) to
cancel the slow closed-loop dynamics: F (s) (s K1 1) (s K2 1) where K2>>K1. Now
1
1
the transmission from r(t) to x(t) is given by X/R=1/(s/K2+1). This can be made
X
R
N
s
K

1
s
K

1
2
1
arbitrarily fast by choosing high K2. The transmission between noise and output
Due to feedback
Due to pre-filter
remains X/N=1/(s/K1+1) which is of ultra-low bandwidth.
I

2 DOF Control applied to achieve insensitivity to plant gain variations

Root Locus
0.8
0.6
0.4

Imaginary Axis

Example 2: Having observed how the pre-filter and feedback combination can be utilized,
we now put it to a different use: to minimize the effect of variation in plant properties on
the overall response. For the sake of illustration, we choose a simple plant:
P(s)=10/(s+10). We desire the closed-loop dynamics to have a single dominant pole at 20rad/s and zero steady state error to DC reference. Suppose, also, that the gain of the
plant varies from 2 to 50, i.e., by a factor of 25 (!), we desire that the overall response of
the system remains unchanged in the face of this variation.
Solution: Before we decide that we need 2DOF control configuration to solve this
problem, we inspect whether 1DOF control can already solve it. To achieve zero steady
state error we add a pole at s=0. To get the root-locus to pass through s=-20, we cancel
the plant pole at -10 with a zero at the same point and add another zero a bit to the left
of s=-20, say, at s=-25. Finally, to ensure causality we add a pole far away, at s=-50. Thus
C(s)=K(s+10)(s+25)/[s(s+50)]. The root-locus of CP is shown on the right. In order to
ensure that the closed loop pole is at s=20, we find that K=12. The step response of the
closed-loop system is given on the right for the nominal plant and at the extreme values
of its DC gain. While the nominal performance meets our requirements, we notice wide
variation in dynamics as the plant model changes. Thus, 1DOF control is inadequate to
control the variation of the closed-loop poles due to that of the open-loop poles.

0.2
0
-0.2
-0.4
-0.6
-0.8
-50

-40

-30

-20

-10

Real Axis

K0=50
Step Response
1
0.9
0.8
0.7

Amplitude

0.6

Nominal, K0=10

K0=2

0.5
0.4
0.3
0.2
0.1
0

0.05

0.1

0.15
Time (sec)

0.2

0.25

0.3

There is one way we can reduce the variation of the closed-loop pole due to changes in plant
gain, and this insight is obtained primarily from the root-locus: since the root-loci end in
zeros, we notice that as we increase the open-loop gain of a system, the closed-loop pole
tends closer and closer to a zero. At high enough gains, even if the gain varies, the position
of the closed-loop pole relative to the zero does not change much. At theoretically infinite
gain, the pole is exactly at the zero and any variation in plant gain, no matter how big, does
not vary the closed-loop pole at all. So, one possible approach to minimize the variation of
the closed-loop pole due to plant variation is to place an open-loop zero close to the desired
closed-loop pole location and increase the open-loop gain to as high a value as possible.
So, in the present case, we place an open loop zero at s=-20. To ensure zero steady state
error and to guarantee causality, we add a pole at s=0 and a pole far away at s=-50. As
before we choose to cancel the plant pole at s=-10. So, our feedback controller is of the form
C(s)=K(s+20)(s+10)/s(s+50), where K is a large constant. We see from the root locus below
that when the gain is large, one of the closed-loop poles close to s=-20 and the other is far
away, beyond s=-50. Thus, the closed-loop denominator is of the form (s+20-)(s+50+),
where ~0 and >>1. The dominant dynamics is that of the slowest pole, namely s=-20+ . If
K is high enough, the variation of the plant gain would result in small changes in , so the
dominant dynamics would remain essentially the same.
Although this strategy appears to solve the problem of fixing the dominant dynamics, there
is one big issue: since Gcl=CP/(1+CP) , the open-loop zeros of CG are also the zeros of Gcl, i.e.,
of the closed-loop system. Thus Gcl is of the form Gcl =(s+20)K/[(s+20-)(s+50+)]. Since ~0,
we see that the numerator zero at s=-20 almost perfectly cancels the dominant dynamics of
the closed-loop system. This renders all our effort in designing the desired closed-loop
dynamics useless. Thus, zero placement strategy does not work with a 1DOF control system.
However, with a 2DOF control system, we can choose to cancel the root of the problem,
namely, the closed-loop zero (s+20), by means of the pre-filter, by setting F(s)=20/(s+20). So,
the transfer function from the reference r to the output x is given by X/R= K/[(s+20)(s+50+)]. This retains the desired dominant dynamics. The step response (next slide)
shows that despite plant variation, the system response barely changes.
Imaginary Axis

2 DOF Control applied to achieve insensitivity to plant gain variations

Root Locus
0.5
0.4
0.3
0.2

0.1

-0.1
-0.2
-0.3

Closed loop poles

-0.4
-0.5

-70

-60

-50

-40

-30

Real Axis

-20

-10

Feedback

Pre-filter

C ( s)

P( s )

d
x

Step Response

K0=50

0.9

K0=2

0.8

0.7

Amplitude

F ( s)

( s 10)( s 20)
20
K=200
F ( s)
s( s 50)
s 20

0.6

Plant gain K0=10

0.5
0.4
0.3
0.2
0.1
0

0.05

0.1

0.15

0.2

0.25

0.3

Time (sec)

K=2000
Step Response
1
0.9

K0=50

0.8

K0=2

0.7

Amplitude

C ( s) K

0.6
0.5

Plant gain K0=10

0.4
0.3
0.2
0.1
0

0.05

0.1

0.15
Time (sec)

0.2

0.25

0.3

2 DOF Control applied to achieve insensitivity to plant gain variations

We saw from the previous example that placement of a controller zero near the desired
closed-loop pole position along with the use of a pre filter guarantees robustness against
variations in the plant gain. However, we did not quantify how high a gain we need to
choose to achieve the desired robustness. We shall take that up now.
P0cl 1.5 2 j
Example 3: Let the plant be given by: P(s)=1/[s(s+1)(s+2)].
Here we assume that the gain of the plant K0, which is nominally =1, varies by a factor of
z
10, from 0.33 to 3. We wish to design a controller such that, for the nominal plant, the
dominant dynamics of the plant is determined by closed-loop poles at s=-1.52i. We also
desire that when the gain of the plant varies by a factor of 10, the closed-loop pole varies at
P5 P4
P3
P1
P
most by 0.4.
X 2
X
X
X X
(far away) -3
-2
-1
Solution: We know from our previous experience that in order to minimize the variation of
a dominant pole, we need to place a controller zero close to it and crank up the loop gain to
a high enough value that the closed-loop poles actually sit at the designated spot. Thus we
need a pair of zeros near s=-1.52i. However, we do not yet know how close to place them
z
or in what direction. These will be the subjects of the design procedure. Firstly we note that
for the sake of causality, our controller needs at least two poles. So, we add two poles far
z
cl
away. The location of all the open-loop poles and zeros is shown on the right. The location
Pmax
P0cl
of the nominal closed-loop pole at Pcl0 =-1.52i has also been indicated.
cl
Locus of Pcl : When the plant gain changes, the position of the point Pcl will also change. In
Pmin
order to study the variation of the closed-loop pole position Pcl , we use the fact that
cl
cl
cl
zP0cl .zP0cl
1+KG(s)C(s)=0 at the points Pcl . Thus, we have K 1 C (s)G(s) PP
1 0 .P2 P0 ...P5 P0
Let us call the open-loop gain for the nominal plant as K0, the open-loop gain when the
plant gain is maximum as Kmax and when it is minimum as Kmin. Let us assume that the
closed-loop poles vary correspondingly to points Pclmax and Pclmin. We note that if the gain K
is large, Pclmax and Pclmin are close to Pcl0 . Thus, we would have PiPcl0 ~ PiPclmin~PiPclmax (i=1..5).
K max zP0cl
K min zP0cl
cl
cl
cl
cl
cl ,
cl
Likewise zP0 ~ zPmax ~ zPmin . The only distance that varies considerably is zP . Thus, we see
K0
zPmax
K0
zPmin
cl
Since the distances zP are in general complex numbers, and the ratio of these complex
numbers is real, we infer that the three complex numbers are collinear. Thus, our analysis
concludes that when K varies, the closed loop poles vary along a straight line.

2 DOF Control applied to achieve insensitivity to plant gain variations

Let us call zPcl0 =x. Then, zPclmin =3.33x and zPclmax =0.33x. Thus we have the range of variation
of the poles to be zPclmin -zPclmax =3x. We wish this variation to be within 0.4. Thus we get
x=0.4/3. This stipulates the distance of the zero from the desired location of the closed loop
pole at -1.5+2i. However, we still do not know in what direction to locate the zero.
Since the far off poles are expected to be really far off, we assume that the angle they
subtend at the nominal pole position is close to zero. This would imply that any deficit angle
necessary for the root-locus to pass through the nominal pole position should be supplied
primarily by the zero. Under this assumption, the zero would have to subtend an angle of 36
in order for the root-locus to pass through the pole position. Equivalently, it should make an
angle of 216 with respect to the closed-loop pole (figure on the right).
i 216 /180
Thus the numerator of the controller would be (s z)(s z ) ,where, z 1.5 2i [0.4 3]e .
Where to place the far off poles? The number of far off poles is a design choice. We need at
least two to ensure causality and, for now, we shall just stick with two poles. Further we shall
assume that the two poles are coincident, i.e., P4=P5. Their existence results in further
branches of the root locus, some of which may escape into the right half plane when the
controller gain is set high enough to ensure that the closed-loop pole movement is within the
prescribed bounds. This escape has to be prevented.
Since we have decided to place the far off poles really far off, the complex plane on the right
shows a more accurate rendering of the different open loop pole positions. We note that for a
point Q on the root locus, its distance to the three poles and the two zeros near the origin are
approximately the same and the angles subtended by them are also approximately the same
P2Q PQ
zQ zQ
in magnitude, i.e., PQ
1
3
PQ
1 .P2Q.PQ
3 .P4Q.P5Q
PQ
Therefore, KQ
1 .P4Q.P5Q . Thus, for points such as Q which are very far
zQ.zQ

away ,the entire collection of 3 poles and 2 zeros behaves as though it was composed of just a
single pole. Thus, when considering for off points on the root-locus, the entire collection near
the origin can be replaced by (3-2), i.e., 1 pole near the origin (last figure on the right).

cl
Pmax

3.33x

P0cl
x

0.33x

cl
Pmin

x=0.4/3
= 216

P0cl

z
Root locus
Q

P4, P5

P1, P2, P3, z, z

P4, P5

2 DOF Control applied to achieve insensitivity to plant gain variations

With the useful approximation made in the previous slide, the open-loop system, at large distances
from the origin, looks as though it was composed of 3 poles, viz., the lumped pole at P and the poles
P4, P5 as shown in the figure on the right. The root locus of the system is also shown on the right. If Q is
the point it crosses the imaginary axis, then the sum of the angles of the 3 poles at Q should be -180.
As shown in the figure, the pole P adds approximately -90. Thus the two poles P4 and P5 should
together add the remaining, i.e., each should add -45. Thus in the figure on the right, =45.

The gain at which the root locus crosses the imaginary axis is given byKQ PQ.P4Q.PQ
5
P4, P5
Assuming the distance of P4 and P5 from the origin is , and =45 ,we have PQ= and P4,5Q=2 .
Thus, KQ .( 2 )2 2 3
The gain at which the root locus crosses the imaginary axis should be more than the maximum gain

assumed by the open-loop system when the plant gain varies. Clearly the gain of the open-loop system
is maximum when the closed-loop pole Pcl is closest to the open-loop zero z, i.e., when Pcl = Pclmax. This
is given by K max

The terms within the bracket above are all known since Pclmax has been determined. The last two terms
cl
cl
are not known. However, since Pclmax is close to the origin, we may assume that P4 Pmax P5 Pmax
cl
cl
cl
2
1 max .P2 Pmax .P3 Pmax
To ensure that all closed-loop poles are stable we need KQ Kmax , i.e., 2 3 PP

cl
cl
cl
cl
cl
zPmax .zPmax

PP
.
P
P
.
P
P
1
1
max
2
max
3
max
or,

cl
cl
2

cl
cl
cl
PP

cl
cl
1 max .P2 Pmax .P3 Pmax

P4 Pmax P5 Pmax
cl
cl
zPmax .zPmax

zPmax .zPmax

30
For the specific problem at hand, evaluating the above expression gives
cl
cl
cl
PP

.
P
P
.
P
cl
cl
1 0
2 0
3 P0
We set =30 and evaluate the nominal control gain using K0
P4 P0 P5 P0
cl
cl
zP0 .zP0

As the final step, we need to design the pre filter F(s). We note that the
pre filter
should possess unit
DC gain and should cancel the two zeros of the closed-loop reference-to-output transfer function. Thus,
F(s) should be of the form F (s) zz [(s z)(s z )] . The final root locus and step responses are shown in
the next slide.

Root
locus

90
P

P( s )

K0
s( s 1)( s 2)

Root Locus
80
60

s 2 3.216s 6.278
C ( s) 17930
( s 30)2

F ( s)

Imaginary Axis

40

6.278
s 2 3.216s 6.278

20
0
-20
-40
-60
-80
-70

-60

-50

-40

-30

-20

-10

10

20

30

Real Axis

Comparison of step responses for 1 DOF and 2 DOF control designs

1StepDOF
Response

2.5

1.4

K=3.3

1.2

1.2

Amplitude

K=1

0.8

Root Locus

K=0.33
0.6

K=0.33
K=1

0.8

K=3.3

0.6

0.4

0.4

0.2

0.2

Range of variation: 0.4

1.5

Imaginary Axis

1.4

2
DOF
Step Response

1
0.5
0
-0.5
-1
-1.5

2
Time (sec)

-2
-2.5
0

2
Time (sec)

-2

-1.5

-1
Real Axis

-0.5

Two degree of freedom control-some notes

How does it work? It appears, from the previous demonstrations, that the use of two degree of freedom control
somehow makes the variation of all the closed-loop poles miraculously go away. In practice however, we can see
how the system works if we trace the trajectory of one of the faster poles as the plant gain varies. This is shown in the
root-locus plot below. We see that when the plant gain varies, the chosen fast pole swings by wide amounts. Thus, in
minimizing the variation of the slow poles, we have translated their variation into the large variation of fast poles.
In order to control the variation of closed-loop poles at -1.52i, we had to introduce poles very far away at -30. Thus,
our controller has to possess much higher bandwidth than the actual system it is trying to control (Bode plot below):
this excessive gain-bandwidth requirement is the price we pay for having to control the motion of a few dominant
poles.
We note that the insight we obtained using root-locus to minimize variation of closed-loop poles cannot be obtained
readily using Bode plots. So, design of robust control systems is one example where root locus trumps over Bode plots.
However, we also note that the use of root-locus depends on the concepts of dominant poles. If we do not desire such
a dominant response, design using root-locus becomes much harder. Here frequency domain based design, although
not very intuitive, can do a better job and may even yield a controller with lesser gain-bandwidth.

Root Locus
80

Bode plot of the controller C(s)


Bode Diagram

60

90
80

Magnitude (dB)

K0=3.33

20

K0=1
0

70
60
50

K0=0.33

40
135

-20
-40

Phase (deg)

Imaginary Axis

40

-60

90

45

-80
-70

-60

-50

-40

-30

-20
Real Axis

-10

10

20

30

0
-1
10

10

10

Frequency (rad/sec)

10

10

Reference
(1) I M Horowitz, Synthesis of Feedback Systems, Academic Press (1963)

You might also like