You are on page 1of 12

IN 227 Control Systems Design

Lecture 20

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

Control of unstable plants

20log |L|
-20 dB/decade

Unstable plants possess one or more poles in the right half of the complex plane (RHP).
A plant that possesses one unstable pole at s=a can be written as PU(s)=P1(s)/(s-a),
where, P1(s) does not contain any pole in the RHP.
The way an unstable pole would affect the closed-loop dynamics can be seen by writing
PU(s)=[(s+a)/(s-a)][PMP(s)] where, PMP(s)=P1(s)/(a+s), is a minimum phase well behaved
plant model and (s+a)/(s-a) is a Blaschke product. Let us now assume the a well behaved
controller C(s), i.e., a controller with no right half plane poles, is used to control the
plant and the resulting loop gain is LU(s)=C(s)PU(s). For reference, we shall also define the
minimum phase counterpart LMP(s)=C(s)PMP(s). Thus, LU(s)=[(s+a)/(s-a)]LMP(s).
We can easily draw the bode plot of the minimum phase part, viz., LMP(j). We assume,
for the sake of illustration, that LMP(j) contains an integrator so that the phase of
LMP(j) is -90 and the gain roll-off is -20dB/decade at low frequencies. The gain
characteristics of LU(j) is the same as that of LMP(j) since |(a+j)/(a-j)|=1 at all
frequencies (Fig. top right). The gain crosses over at gc. The phase of LMP(s) gradually
decreases further and in general crosses - at a certain frequency. The Blaschke product
(a-j)/(a+j) possesses a phase lag of -+2tan-1(/a). Thus, the overall phase
characteristic starts at -3/2, increase initially, cross-over at pcl, continue increasing
until the phase lag of LMP(s) begins to dominate. At this point the phase crosses over
again at a higher frequency pcu, and continues to decrease beyond pcu (Fig. top right).
In contrast with the general phase characteristics of minimum phase plants, the phase
characteristics above appear unfamiliar. For example, how can we define the gain margin
when we have two phase cross-over frequencies? Does the conventional phase margin,
defined as the phase at gc, carry the same significance as in minimum phase systems?
How does the position of gc relative to pcl and pcu affect stability?
Bode plots seldom answer questions related to stability independently: they are merely
tools for achieving control performance. We need to transfer the open-loop system into
the Nyquist domain and examine its stability.

20log |LU| (=20log |LMP|)


log
gc
log

-
-3/2

pcl

pcu

LU LMP 2 tan 1 ( / a)

Closed loop stability of unstable plants

s
Before answering the questions posed in the previous slide, we first need to develop the condition under
R
which an open-loop system with one unstable pole would result in a stable closed loop system: The
sa
denominator transfer function of the overall closed-loop system is
( s a) DMP (s a) N MP
sa
a
0
1 LU 1
LMP
, where, LMP=NMP/DMP.

sa
( s a) DMP
We note that 1+LU has a pole in the RHP. Thus, when the s-variable is taken round the contour C once in
the clock-wise direction, the term 1/(s-a) encircles the critical point -1+j0 once in the counter-clockwise
direction. Therefore, if the numerator polynomial of 1+LU does not possess any zeros on the RHP, i.e., the
closed-loop system is stable, then 1+LU should encircle the critical point in the counter clockwise direction
exactly once. If the net encirclement is found to be zero, then 1+LU has to possess a zero within C whose
C
clockwise encirclement cancels that of 1/(s-a). More generally, number of unstable poles=number of
j
clockwise encirclements+1.
s
In analyzing the stability for LU, we have to look at four cases: (a) gc <pcl (b) pcl < gc < pcu (c) gc >pcu
(d) the phase lag does not cross radians at any frequency. The Nyquist contour C is shown on the right.
a
0
Since we have an integrator, and hence a singularity at the origin, we have to introduce an infinitesimally
small kink in the contour at s=0 to avoid the singularity (viz., a semicircle of radius r0).
We see from the Nyquist plots of LU (figs. below) that for cases (a), (c) and (d), the critical point is encircled
once in the clockwise direction, and thus, the closed loop system would possesses two unstable poles.
Only for case (b), i.e., when the grain crossover frequency is between the two phase crossover frequencies
(pcl < gc < pcu ), is the critical point encircled in the counter clockwise direction once and thus, the
closed-loop system would be stable.

Case (a)

-1+j0

Case (b)

Case (d)

Case (c)

-1+j0

-1+j0

-1+j0

Theoretical minimum gain cross-over frequency for specified PM

We shall now work out the minimum gain cross-over frequency necessary to
achieve a specified phase margin PM.
Let LU cross 0 dB at gc and let the slope of the magnitude characteristic in the
neighborhood of gc be -40 db/decade. This is also the slope of LMP. Bodes gainphase relationship tells us that the phase of LMP at any frequency is determined
predominantly by the slope of the magnitude characteristic near alone,
regardless of the actual shape of the magnitude characteristic of LMP at other
frequencies. Thus, the phase of LMP at gc is - rad. Consequently, the net phase of
the loop gain LU at gc is ( jgc a) ( jgc a) 2 tan 1 (gc / a)
To satisfy the prescribed phase margin PM, we note that LU (gc ) PM .
Equating these two we get gc tan PM
a

Since tan() is an increasing function of , we note from the equation above that
for a given PM, the gain cross-over frequency is an increasing function of . Thus,
the lowest gain cross-over frequency is obtained for the smallest value of .
However, since the low frequency loop gain needs to be greater than 0dB, we
require that >0 in order for the magnitude characteristic to cross the 0dB line.
Thus, for the given phase margin PM, the absolute maximum gain cross-over
frequency is obtained when ~0 and is given by gc a tan PM
2

We see, therefore, that the minimum possible value of the gain cross-over
frequency gc is theoretically limited to a.tan(PM/2) due to the RHP pole at s=a.

20log |L|

-40 dB/decade
log
gc
log

-3/2

pcl

PM
pcu

Theoretical minimum gain cross-over frequency to satisfy the specified


GM and PM

Since we have two phase cross-over frequencies, we have two corresponding gain
margins GML=20log|LU(jpcl)| and GMU=-20log|LU(jpcu)|. For stability, we require
GML,GMU>0dB.
Since pcl < gc, the minimum gain cross-over frequency necessary is decided by the
specified GML. We shall therefore focus on achievement of the specified GML and
PM.
By definition, LU (pcl ) . Assuming that the slope of LU remains approximately
-40 db/decade between gc and pcl, and thus, the phase of LMP remains at at

pcl,we have, 2 tan 1 ( pcl a) or equivalently, pcl tan( )


a
2
Since the slope of LU remains approximately -40 db/decade between gc and pc,
(fig. below right), we have GML 40 log gc pcl .Using the fact that pcl a tan( 2)
tan ( PM ) 2
and gc a tan ( PM ) 2 , we obtain, GML
40

log

tan( 2)

20log |L|

GML

-40 dB/decade
log
gc
GMU
log

PM
pcl

pcu

-3/2

The last equation above is a transcendental equation in . It is solved numerically to


obtain and subsequently used to evaluate gc. The figure below sketches the
minimum value of gc (normalized with respect to a) for different phase and gain
margin requirements.
PM=30

-40 dB/decade
20log |LU|
GML

log

gc

40

gc

20

log

LU

2
tan( PM 2)
5dB

10dB
GML(dB)

pcl

Consequences of the theoretical limits

The theoretical limits to the minimum gain crossover frequency necessary to achieve the specified stability
margins pose corresponding challenges to the control of unstable systems. Some of them are highlighted below.
(1) Enhanced effect of measurement noise: The fact that the gain cross-over frequency has a lower limit
indicates that even if we desire the benefits of feedback only to a relatively small frequency relative to the
value of the unstable pole a, we have to keep the magnitude of the loop gain LU high up to significantly higher
frequencies (usually several multiples of a). Thus, bandwidth of the transmission function relating the sensor
noise to the output (-LU/(1+LU)) also needs to be typically several times a. Thus, we have to let in much more
measurement noise, when compared to minimum phase plants, to reap the same benefits of feedback.
(2) Plants with RHP pole and RHP zero: Suppose the plant possesses a RHP zero in addition to an RHP pole, i.e.,
the plant transfer function is of the form PNMP=(a-s)P1/(s-b), where P1 is a minimum phase transfer function,
then, if we use a minimum phase controller C, we can write the resulting non minimum phase loop gain
LNMP=CPNMP as LNMP

CP1
as sb
is the minimum phase counterpart of the non
.
LMP , where LMP
a s s b
( s a)( s b)

minimum phase loop gain. As we saw in the previous lecture, for the specified gain and phase margins, the term
imposes an upper limit on the gain crossover frequency gc to a value around a. Likewise, for the same gain and
phase margins, the term

sb
, imposes a lower limit on the gain cross-over frequency gc to a value around b.
s b

Thus, if b>a, then it is impossible to design a controller that satisfies the specified gain and phase margins.

Example: Suppose we desire a gain margin of 10dB and a phase margin of 30: based on the graphs provided in
the previous lecture, we see that the upper limit to the gain cross-over frequency is gc0.2a. To achieve the
same specifications, we see from the sketch in the previous slide that the lower limit should be gc5b. Thus, if
5b>0.2a or b>0.04a, it is impossible to achieve the specified gain and phase margin requirements.

as
as

Bode Sensitivity Integrals

u
e
r
The third consequence is that an unstable plant, even without a RHP zero, imposes
C p (s)
Ce ( s)
fundamental limits on the gain and phase margins in practical control systems. The
- Controller C(s)
controller in practical control systems possess two subsystems: the electronic subsystem
Ce where the control law is implemented, and the physical subsystem Cp, also called the
actuator (ex. motors, solenoid valves, etc), which transforms electrical energy into
changes in physical state. While the electronic subsystem can typically be designed to
possess very high bandwidth, the physical subsystem is fundamentally limited in the
frequency range over which it operates.
We shall now explore the limits to achievable stability margins imposed by a very real
practical constraint, viz., limited actuator bandwidth. This can be done in a very
transparent manner by means of Bode sensitivity integrals.
Bode sensitivity integrals: All the benefits of feedback are tied to achieving a high loop
gain L(j). The consequence of high L(j) is the reduction in the sensitivity function
S(j)=1/(1+L(j)), with resulting good rejection of disturbances and good robustness to
variations in plant parameters. As control engineers, therefore, we desire zero sensitivity
at all frequencies. However, Bode sensitivity integrals impose constraints on the
achievable reduction in sensitivity. In particular, if the loop gain L is minimum phase,
then the sensitivity function satisfies the first equation below. If L is open-loop unstable,
then the corresponding sensitivity function satisfies the second equation below. Both of
them can be derived from Cauchys theorem.

Bode sensitivity integrals

ln S ( j) d 0

For minimum phase L

ln S ( j) d Re( p )
i

For open-loop unstable L


(pi are the open loop poles in the RHP)

PNMP

The sensitivity dirt

A1: Area of
reduced sensitivity
A1+A2=constant

Unstable L
Minimum phase L
0

Tall heap?

Log|S|

Log|S|

Bode sensitivity integrals are conservation laws: they state that the net area under the
log|S| curve is conserved. Thus, no matter what controller is used, it is impossible to
reduce sensitivity at all frequencies. If our controller reduces sensitivity in a certain part of
the frequency domain (< 0 in the fig. right), then it will inevitably increase sensitivity in
other parts of the frequency domain. Thus, log|S| is informally called sensitivity dirt (Ref.
2): we would like to have as little of it as possible, but if we dig a trench to remove it
somewhere, the removed dirt has to be piled up or spread around somewhere else.
In the case of minimum phase systems, the amount of dirt that gets piled up elsewhere is
the same as the amount of dirt that got removed somewhere. However, in the case of
open loop unstable systems, the amount of dirt that needs to be piled up is greater by the
amount pi than the amount of dirt that was removed (Fig. on the right). If the loop gain
LU possesses a single unstable pole at s=a, then the excess sensitivity dirt that needs to be
distributed is a.
Once we have removed sensitivity dirt up to a certain frequency 0 in the interest of
reaping the benefits of feedback, we need to decide on the strategy to distribute this dirt
plus the excess dirt a: do we distribute them uniformly over the reminder of the
frequency domain or do we make a tall heap, i.e., greatly increase sensitivity, around some
frequency 1?
To decide this, we need to understand the result of increased sensitivity: if we choose to
make a tall heap at some frequency 1, we note that log|S(j1)|>>1. Thus, 1 [1 L( j1 )] K ( 1)
or equivalently 1 L( j1 ) 1 K . We note from the Nyquist plot shown on the right that for
any loop gain L(j), 1+L(j) is the complex number starting at the critical point (-1+j0) and
ending at the tip of the complex number L. Thus,1 L( j1 ) const. 1 K , describes a circle
of radius 1/K centered at the critical point.
As seen from the schematic on the right, if the loop gain assumes values on the circle
1 L( j1 ) 1 K , then the maximum phase margin achievable is PMmax~2sin-1(1/2K) and the
maximum gain margin achievable is GMmax=1/(1-1/K).
Thus, if K>>1, then PMmax~sin-1(1/K)~0 and GMmax~0dB. Thus, high sensitivity is
synonymous with small phase and gain margins.
Log|S|

A2: Area of
heightened sensitivity

Thin layer?
0 1

Radius 1/K

Im(L)

-1+j0
1+L(j)

L(j)

Re(L)

Unit circle
1/GMmax
-1+j0
1/K

(1-1/K)
PMmax

Distributing the sensitivity dirt

We conclude, from the previous slide, that in the interest of stability, we need to minimize the
height of sensitivity dirt in the region >0. Since we have an infinitely large region from 0 to
, to spread the dirt, we may be tempted to spread it out as an infinitesimally thin layer at all
frequencies, and thus, ensure that sensitivity is small everywhere. This procedure, however,
assumes that it is possible to control the thickness of the dirt layer over all frequencies from 0
to .In practice, though, all actuators possess a bandwidth, i.e., a frequency a beyond
which it is not possible to control their dynamic performance. Indeed their gain attenuates
greatly for all larger frequencies and causes the loop gain L to tend to zero.

Log|S|

Loop shape cannot be


controlled beyond this
frequency

A1: Area of
reduced sensitivity

Thus, the shape of log|S| cannot be controlled for >a, and consequently, the area enclosed

by it is a constant (), i.e., ln S ( j) d . Since ln S ( j) d ln S ( j) d ln S ( j) d , we


a

conclude that ln S ( j) d const. a . In other words, contrary to the theoretical

Log|Smin|

Since we would be reducing the dirt for frequencies between 0 and 0 by the amount A1, we
1

note that

ln S ( j) d A1. Thus, we see that


0

ln S ( j) d a A . In other words, the


1

amount of dirt that we need to spread between 0 and a is a-+A1. This should be done in
such a way that the maximum height of the dirt pile is minimized. A moments thought reveals
that the only way to do this is to distribute the dirt equally over the entire frequency range,
i.e., log|S(j)|=constant for 0<<a as shown on the right.

Thus the minimum height of the dirt pile is ln Smin

a A1

a A1

, or equivalently, Smin e a 0
a 0

Log|S|

expectations, the frequency region available to spread the dirt is limited to a.

A2= a-+A1

0
A1

Maximum achievable stability margins

a A1
specifies fundamental limits to the minimum
a 0

The final equation on the previous slide, viz., Smin exp

achievable sensitivity in the region 0<<a and thus sets upper limits to the achievable gain and phase margins for
the specified actuator bandwidth a and the specified amount of feedback necessary (A1) and the frequency 0 up to
which it is necessary.

We note from this equation that the absolute least sensitivity |Smin| is reached when the expected benefits of
feedback and the frequency range of expectation are both small relative to a and a, i.e., A1/a, 0/a<<1
a

and is given by Smin exp

For the sake of illustration, let us assume that the amount of sensitivity dirt for >a is negligible, i.e., ~0. Thus, we

a
. Suppose the speed of the actuator used to control the system is comparable to the rate at which
a
the open-loop plant response diverges, i.e., a a, then we note that Smin exp 23.14. For this case, the maximum

get Smin exp

achievable phase margin is 2.5 and the maximum possible gain margin is 0.86dB!

Suppose we pick an actuator that is half as fast as the unstable plant, i.e., a a 2 , the we note that Smin exp 2 534
Thus, for this case, the maximum phase margin possible is 0.1 and the maximum gain margin possible is 0.04dB!

It is worth noting the startlingly low values of stability margins for both the cases above: if the actuator is not fast
enough, the resulting control system is, for all practical purposes, unstable. It is also worth noting that, in our estimates
above, nowhere did we make any assumption about the specific dynamics of the plant or the structure of the controller.
All we used was just one little piece of information: the ratio of the speed of the unstable pole to that of the actuator,
i.e., a / a . This was sufficient, with the help of Bode integrals, to work out the best stability margins possible,
regardless of the specific controller we might choose to use.

Some comments on Bode integrals

Bode integrals are of fundamental importance to control engineers because of their generality: they
enable us to work out the best possible stability margins, with the barest minimum information about
the system to be controlled. Thus, they can quickly reveal the limits of achievable performance and
stability even before the plant has been fully identified and a specific controller is designed.
Any practical controller can only do as well as the best possible. Since the shape of the sensitivity function
that guarantees minimum sensitivity for 0<<a cannot be readily constructed using simple pole-zero
patterns for the controller, the practical controller, constructed with just a few poles and zeros in the
interest of minimizing complexity, will probably do much worse than the best possible estimate.
The actuator bandwidth is a fundamental practical limit, if not a theoretical one. Thus, the integrals
provide a way to incorporate this limit into calculations of the best possible stability specifications and the
desired benefits of feedback possible for a given actuator.
In some sense, the Bode Sensitivity Integrals, in combination with the Bode gain-phase relationship and
ideal Bode characteristic occupy the same place in control theory as Newtons laws do in mechanics:
they specify the laws that the control systems cannot disobey, and are therefore invaluable in computing
the limits to the benefits of feedback and the minimum price that we should pay for enjoying these
benefits.
Controllers that are designed without paying adequate attention to these fundamental limits, especially
for controlling unstable plants, are likely to have disastrous consequences (see Ref. 2). One of them
(Chernobyl nuclear meltdown) is one of the worst industrial disasters in the history of humanity. If other
such disasters are to be avoided, it is only prudent for control engineers to develop intimate working
knowledge and familiarity with Bodes laws.

References
(1) Marcel Sidi, Design of Robust Control systems From Classical to Modern Practical Approaches, Kreiger
Publishing Co. FL, USA, (2001)
(2) Gunter Stein, Respect the Unstable, IEEE Control Systems Magazine, pp. 12-25 (August 2003)

You might also like