Topic #2

16.30/31 Feedback Control Systems

Basic Root Locus

Basic aircraft control concepts

Basic control approaches

Fall 2010

16.30/31 22

Aircraft Longitudinal Control

Consider the short period approximate model of an 747 aircraft.
x sp = Aspxsp + Bspe
where e is the elevator input, and

U0
w
Zw /m
xsp =
, Asp =
1
1
(Mw + Mw Zw /m) Iyy
(Mq + Mw U0)
Iyy
q

Ze /m
Bsp =
1
Iyy
(Me + Mw Ze /m)
Add that = q, so s = q
Take the output as , input is e, then form the transfer function1

1 q(s)
1
(s)
=
=
0 1 (sI Asp)1Bsp
e(s) s e(s) s
For the 747 (40Kft, M = 0.8) this reduces to:

(s)
1.1569s + 0.3435
= 2
Ge (s)
e(s)
s(s + 0.7410s + 0.9272)

so that the dominant roots have a frequency of approximately 1

rad/sec and damping of about 0.4
PoleZero Map
1

0.68

0.54

0.42

0.3

0.2

0.09
0.8

0.8

0.6

0.84

0.6

0.4

0.4

Imaginary Axis

0.95
0.2

0.2

0.2

0.2
0.95

0.4

0.6

0.4

0.6

0.84

0.8

1
1

0.8
0.68
0.9

0.54
0.8

0.7

0.6

0.42
0.5
0.4
Real Axis

0.3

0.2

0.3

0.2

0.09
0.1

10

0.1

Fig. 1: Note - this is the Pole-zero map for Gqe

1 Much

more on how to do this part later

Fall 2010

16.30/31 23

Basic problem is that there are vast quantities of empirical data to

show that pilots do not like the ying qualities of an aircraft with this
combination of frequency and damping
What is preferred?
Undamped natural frequency s rad/sec

7
6
5
POOR
ACCEPTABLE

4
3

SATISFACTORY

2
UNACCEPTABLE

1
0
0.1

0.2

0.4

0.6 0.8 1

Damping ratio s

Image by MIT OpenCourseWare.

Fig. 2: Thumb Print criterion

This criterion was developed in 1950s, and more recent data is pro
vided in MILSPEC8785C
Based on this plot, a good target: frequency 3 rad/sec and
damping of about 0.6

Problem is that the short period dynamics are no where near these
numbers, so we must modify them.
Could do it by redesigning the aircraft, but it is a bit late for
that. . .

Fall 2010

16.30/31 24

First Short Period Autopilot

First attempt to control the vehicle response: measure and feed it

back to the elevator command e.
Unfortunately the actuator is slow, so there is an apparent lag in
the response that we must model
ec

4
s+4

ea

Ge (s)

Dynamics: ea is the actual elevator deection, ec is the actuator

command created by our controller
4
= Ge (s)ea; ea = H(s)ec; H(s) =
s+4
The control is just basic proportional feedback
ec = k ( c)
which gives that
= Ge (s)H(s)k ( c)
or that

(s)
Ge (s)H(s)k
=
c(s) 1 + Ge (s)H(s)k

Looks good, but how do we analyze what is going on?

Need to be able to predict where the poles are going as a function
of k Root Locus

Fall 2010

16.30/31 25

Root Locus Basics

r

Gc(s)

Gp(s)

Assume that the plant transfer function is of the form

npz
Np
(s zpi)
Gp = Kp
= Kp ni pp
Dp
i (s ppi )
and the controller transfer function is
ncz
Nc
(s zci)
Gc(s) = Kc
= Kc nicp
Dc
i (s pci )
2

Assume that npp > npz and ncp > ncz

Signals are:
u
y
r
e

control commands
output/measurements
reference input
response error

Unity feedback form. We could add the controller Gc in the feedback

path without changing the pole locations.
Will discuss performance and add disturbances later, but for now just
focus on the pole locations

2 Errata:

Added n values for the number of poles and zeros

Fall 2010

16.30/31 26

Basic questions:

Analysis: Given Nc and Dc, where do the closed loop poles go

as a function of Kc?

Synthesis: Given Kp, Np and Dp, how should we chose Kc, Nc, Dc
to put the closed loop poles in the desired locations?

Block diagram analysis: Since y = GpGce and e = r y, then

easy to show that
y
Gc Gp
=
Gcl (s)
r 1 + Gc Gp
where

Kc Kp N c N p
DcDp + KcKpNcNp
is the closed loop transfer function
Gcl (s) =

Denominator called characteristic equation c(s) and the roots

of c(s) = 0 are called the closed-loop poles (CLP).
The CLP are clearly functions of Kc for a given Kp, Np, Dp, Nc, Dc
a locus of roots [Evans, 1948]

Fall 2010

16.30/31 27

Root Locus Analysis

General root locus is hard to determine by hand and requires Matlab

tools such as rlocus(num,den) to obtain full result, but we can get
some important insights by developing a short set of plotting rules.
Full rules in FPE, page 279 (4th edition).
Basic questions:
1. What points are on the root locus?
2. Where does the root locus start?
3. Where does the root locus end?
4. When/where is the locus on the real line?
5. Given that s0 is found to be on the locus, what gain is need for
that to become the closed-loop pole location?
6. What are the departure and arrival angles?
7. Where are the multiple points on the locus?

Question #1: is point s0 on the root locus? Assume that Nc and

Dc are known, let
Nc Np
Ld =
and K = KcKp
Dc Dp
c(s) = 1 + KLd(s) = 0

So values of s for which Ld(s) = 1/K, with K real are on the RL.
For K positive, s0 is on the root locus if
Ld(s0) = 180 l 360,

l = 0, 1, . . .

If K negative, s0 is on the root locus if

[0 locus]

Ld(s0) = 0 l 360,

l = 0, 1, . . .

These are known as the phase conditions.

Fall 2010

16.30/31 28

Question #2: Where does the root locus start?

c = 1 + K

Nc N p
=0
DcDp

DcDp + KNcNp = 0
So if K 0, then locus starts at solutions of DcDp = 0 which are
the poles of the plant and compensator.

Question #3: Where does the root locus end?

Already shown that for s0 to be on the locus, must have

1
Ld(s0) =
K

So if K , the poles must satisfy:

Ld =

NcNp

=0

DcDp

There are several possibilities:

1. Poles are located at values of s for which NcNp = 0, which are
the zeros of the plant and the compensator

2. If Loop Ld(s) has more poles than zeros

As |s| , |Ld(s)| 0, but we must ensure that the phase
condition is still satised.

Fall 2010

16.30/31 29

More details as K :
Assume there are n zeros and p poles of Ld(s)
Then for large |s|,

Ld(s)

1
(s )pn

So the root locus degenerates to:

1
1+
=0
(s )pn
So n poles head to the zeros of Ld(s)

Remaining p n poles head to |s| = along asymptotes

dened by the radial lines
180 + 360 (l 1)
l =
pn

l = 1, 2, . . .

so that the number of asymptotes is governed by the number of

poles compared to the number of zeros (relative degree).
If zi are the zeros if Ld and pj are the poles, then the centroid
of the asymptotes is given by:
=

pj

pn

zi

Fall 2010

16.30/31 210

Example: L(s) = s4
Im

Re

Number of asymptotes and ?

Example G(s) =

s+1
s2(s + 4)
Im

Re

Number of asymptotes and ?

Example G(s) =

s1
s2(s 4)
Im

Number of asymptotes and ?

Re

Fall 2010

16.30/31 211

Question #4: When/where is the locus on the real line?

Locus points on the real line are to the left of an odd number
of real axis poles and zeros [K positive].
Follows from the phase condition and the fact that the phase
contribution of the complex poles/zeros cancels out

Question #5: Given that s0 is found to be on the locus, what gain

is needed for that to become the closed-loop pole location?
Need

Dp(s0)Dc(s0)
1

K
=
|Ld(s0)|
Np(s0)Nc(s0)

Since K = KpKc, sign of Kc depends on sign of Kp

e.g., assume that Ld(s0) = 180, then need Kc and Kp to be

same sign so that K > 0

Fall 2010

16.30/31 212

Root Locus Examples

Im

Re

Fig. 3: Basic

Im

Re

Fig. 4: Two poles

Im

Re

Fig. 5: Add zero

Examples similar to control design process: add compensator dynam

ics to modify root locus and then chose gain to place CLP at desired
location on the locus.

Fall 2010

16.30/31 213

Im

Re

Fig. 6: Three poles

Im

Re

Fig. 7: Add a zero again

Im

Re

Fig. 8: Complex Case

Im

Fig. 9: Very Complex Case

Re

Fall 2010

16.30/31 214

Performance Issues
Interested in knowing how well our closed loop system can track var
ious inputs
Steps, ramps, parabolas
Both transient and steady state
For perfect steady state tracking want error to approach zero
lim e(t) = 0

Can determine this using the closed-loop transfer function and the
nal value theorem
lim e(t) = lim se(s)

s0

So for a step input r(t) = 1(t) r(s) = 1/s

y(s)
Gc(s)Gp(s)
=
r(s)
1 + Gc(s)Gp(s)
e(s)
1
=
r(s)
1 + Gc(s)Gp(s)

y(s)
= Gc(s)Gp(s)
e(s)

so in the case of a step input, we have

r(s)
1/s
e(s) =
=
1 + Gc(s)Gp(s) 1 + Gc(s)Gp(s)
1/s
1
lim se(s) = lim s
=
e()
s0
s0 1 + Gc (s)Gp (s)
1 + Gc(0)Gp(0)
So the steady state error to a step is given by
1
ess =
1 + Gc(0)Gp(0)
To make the error small, we need to make one (or both) of Gc(0),
Gp(0) very large

Fall 2010

16.30/31 215

Clearly if Gp(s) has a free integrator (or two) so that it resembles

1
sn (s+)m with n 1, then
lim Gp(s)

s0

ess 0

Can continue this discussion by looking at various input types (step,

ramp, parabola) with systems that have a dierent number of free
integrators (type), but the summary is this:
step

ramp parabola

type 0

1
1 + Kp

type 1

1
Kv

type 2

1
Ka

where
Kp = lim Gc(s)Gp(s)

P
osition Error Constant

Kv = lim sGc(s)Gp(s)

Velocity Error Constant

s0

s0

Ka = lim s2Gc(s)Gp(s)
s0

Acceleration Error Constant

which are a good simple way to keep track of how well your system
is doing in terms of steady state tracking performance.

Fall 2010

16.30/31 216

Dynamic Compensation

For a given plant, can draw a root locus versus K. But if desired
pole locations are not on that locus, then need to modify it using
dynamic compensation.
Basic root locus plots give us an indication of the eect of adding
compensator dynamics. But need to know what to add to place
the poles where we want them.
New questions:
What type of compensation is required?
How do we determine where to put the additional dynamics?
There are three classic types of controllers u = Gc(s)e

1. Proportional feedback: Gc Kg a gain, so that Nc = Dc = 1

Same case we have been looking at.

2. Integral feedback:
u(t) = Ki

t
0

e( )d Gc(s) =

Used to reduce/eliminate steady-state error

Ki
s

If e( ) is approximately constant, then u(t) will grow to be very

large and thus hopefully correct the error.

Fall 2010

16.30/31 217

Consider error response of Gp(s) = 1/(s + a)(s + b)

(a > 0, b > 0) to a step,
r(t) = 1(t) r(s) = 1/s
where
e
1
=
= S(s)
r 1 + Gc G p

e(s) =

r(s)
(1 + GcGp)

where S(s) is the Sensitivity Transfer Function for the

closed-loop system

To analyze error, use FVT limt e(t) = lims0 se(s)

so that with proportional control,
s
1
1
lim ess = lim
=
K
t
s0 s 1 + Kg Gp (s)
1 + ab
g
so can make ess small, but only with a very large Kg

With integral control, lims0 Gc(s) = , so ess 0

Integral control improves the steady state, but this is at

the expense of the transient response
Typically gets worse because the system is less well damped

Fall 2010

16.30/31 218

, add integral feedback to

(s + a)(s + b)
improve the steady state response.

Example #1: G(s) =

Im

Re

Fig. 10: RL after adding integral FB

Increasing Ki to increase speed of the response pushes the poles

towards the imaginary axis more oscillatory response.

Now combine proportional and integral (PI) feedback:

K2 K1 s + K 2
=
s
s
which introduces a pole at the origin and zero at s = K2/K1
Gc = K1 +

PI solves many of the problems with just integral control

Im

Re

Fig. 11: RL with proportional and integral FB

Fall 2010

16.30/31 219

3. Derivative Feedback: u = Kde so that Gc(s) = Kds

Does not help with the steady state

Provides feedback on the rate of change of e(t) so that the

control can anticipate future errors.
1
Example # 2: G(s) =
, (a > 0, b > 0)
(s a)(s b)
with Gc(s) = Kds
Im

Re

Fig. 12: RL with derivative FB

Derivative feedback is very useful for pulling the root locus into
the LHP - increases damping and more stable response.

Typically used in combination with proportional feedback to form

proportional-derivative feedback PD
Gc(s) = K1 + K2s
which moves the zero from the origin.

Unfortunately pure PD is not realizable in the lab as pure dierenti

ation of a measured signal is typically a bad idea
Typically use band-limited dierentiation instead, by rolling-o the
PD control with a high-frequency pole (or two).

Fall 2010

16.30/31 220

Controller Synthesis
First determine where the poles should be located
Will proportional feedback do the job?
What types of dynamics need to be added? Use main building block
GB (s) = Kc

(s + z)
(s + p)

Looks like various controllers, depending how Kc, p, and z picked

If pick z > p, with p small, then
Im

GB (s) Kc

(s + z)
s

Re

which is essentially a PI compensator, called a lag.

If pick p z, then at low frequency, the impact of p/(s + p) is
small, so
Im

Re

GB (s) Kc(s + z)
which is essentially PD compensator, called a lead.

Various algorithms exist to design the components of the lead and lag
compensators

Fall 2010

16.30/31 221

Classic Root Locus Approach

Consider a simple system Gp = s2 for which we want the closed

loop poles to be at 1 2j
Will proportional control be sucient? no
So use compensator with 1 pole.
Gc = K

(s + z)
(s + p)

So there are 3 CLP.

To determine how to pick the p, z, and k, we must use the phase
and magnitude conditions of the RL
To proceed, evaluate the phase of the loop
s+z
Ld(s) =
(s + p)s2
at s0 = 1 + 2j. Since we want s0 to be on the new locus, we know
that Ld(s0) = 180 360l
Im

(-1,2j)

Re

Fig. 13: Phase Condition

As shown in the gure, there are four terms in Ld(s0) the two
poles at the origin contribute 117 each
Given the assumed location of the compensator pole/zero, can
work out their contribution as well

Fall 2010

16.30/31 222

Geometry for the real zero: tan =

2
p1

2
z1

and for the real pole: tan =

Since we expect the zero to be closer to the origin, put it rst on

the negative real line, and then assume that p = z, where typically
5 10 is a good ratio.

So the phase condition gives:

2(117) + = 180

2
2
arctan
= 53
arctan
z1
10z 1

but recall that

tan(A B) =
so

tan(A) tan(B)
1 + tan(A) tan(B)

( z2 1 ) ( 10z21 )

1 + ( z2 1 )( 10z21 )

= 1.33

which give z = 2.2253, p = 22.2531, kc = 45.5062

1
2
3
4
5
6

7
8
9
10
11
12
13

% RL design using angles

clear all

target = 1+2*j;

phi origin= 180atan(imag(target)/real(target))*180/pi;

syms z M; ratio=10;

phi z=(imag(target)/(z+real(target)));
phi p=(imag(target)/(ratio*z+real(target)));
M=(phi zphi p)/(1+phi z * phi p);
test=solve(Mtan(pi/180*(2*phi origin180)));
Z=eval(test(1));
P=ratio*Z;
K=1/abs((target+Z)/(target2*(target+P)));
[Z P K]

Fall 2010

16.30/31 223

Pole Placement
Another option for simple systems is called pole placement.
Know that the desired characteristic equation is

d(s) = (s2 + 2s + 5)(s + ) = 0

Actual closed loop poles solve:

c(s) = 1 + GpGc = 0
s2(s + p) + K(s + z) = 0
s3 + s2p + Ks + Kz = 0
Clearly need to pull the poles at the origin into the LHP, so need a
lead compensator Rule of thumb:3 take p = (5 10)z.
Compare the characteristic equations:

c(s) = s2 + 10zs2 + Ks + Kz = 0

d(s) = (s2 + 2s + 5)(s + )

= s3 + s2( + 2) + s(2 + 5) + 5 = 0
gives
s2 + 2=10z
s 2 + 5=K
s0
5=zK
solve for , z, K
25
5z
; =
5 2z
5 2z
z = 2.23, = 20.25, K = 45.5
K=

3 Errata:

changed the rule of the pole zero ratio.

Fall 2010

16.30/31 224

Fig. 14: CLP with pole placement

Code: Pole Placement

1
2
3
4
5
6
7
8
9
10
11

% Fall 2009

close all

figure(1);clf

set(gcf,'DefaultLineLineWidth',2)

set(gcf,'DefaultlineMarkerSize',10)

set(gcf,'DefaultlineMarkerFace','b')

clear all;%close all;

set(0, 'DefaultAxesFontSize', 14, 'DefaultAxesFontWeight','demi')

set(0, 'DefaultTextFontSize', 14, 'DefaultTextFontWeight','demi')

12
13
14
15
16
17
18
19
20
21
22

%Example: G(s)=1/22

%Design Gc(s) to put the clp poles at 1 + 2j

z=roots([20 49 10]);z=max(z),k=25/(52*z),alpha=5*z/(52*z),

num=1;den=[1 0 0];

knum=k*[1 z];kden=[1 10*z];

rlocus(conv(num,knum),conv(den,kden));

hold;plot(alpha+eps*j,'d');plot([1+2*j,12*j],'d');hold off

r=rlocus(conv(num,knum),conv(den,kden),1)'

axis([25 5 15 15])

print dpng r300 rl pp.png

Fall 2010

16.30/31 225

Observations

In a root locus design it is easy to see the pole locations, and thus
we can relatively easily identify the dominant time response
Caveat is that near pole/zero cancelation complicates the process
of determining which set of poles will dominate
Some of the performance specications are given in the frequency
response, and it is dicult to determine those (and the corresponding
system error gains) in the RL plot
Easy for low-order systems, very dicult / time consuming for higher
order ones
As we will see, extremely dicult to identify the robustness margins
using a RL plot
A good approach for a fast/rough initial design

Matlab tool called sisotool provides a great interface for designing

and analyzing controllers

MIT OpenCourseWare
http://ocw.mit.edu

16.30 / 16.31 Feedback Control Systems

Fall 2010

For information about citing these materials or our Terms of Use, visit: http://ocw.mit.edu/terms.