You are on page 1of 70

Goodwin, Graebe, Salgado

©
, Prentice Hall 2000
Chapter 3
Chapter 1
Modeling for Control
Topics to be covered include:
 How to select the appropriate model complexity
 How to build models for a given plant
 How to describe model errors.
 How to linearize nonlinear models
It also provides a brief introduction to certain commonly
used models, including
 State space models
 High order differential and high order difference equation models
Goodwin, Graebe, Salgado
©
, Prentice Hall 2000
Chapter 3
The power of a mathematical model lies in the fact
that it can be simulated in hypothetical situations, be
subject to states that would be dangerous in reality,
and it can be used as a basis for synthesizing controllers.
The Raison d'être for Models
Goodwin, Graebe, Salgado
©
, Prentice Hall 2000
Chapter 3
Model Complexity
- In building a model, it is important to bear in mind
that all real processes are complex and hence any
attempt to build an exact description of the plant is
usually an impossible goal.
- Fortunately, feedback is usually very forgiving and
hence, in the context of control system design, one
can usually get away with rather simple models,
provided they capture the essential features of the
problem.
Goodwin, Graebe, Salgado
©
, Prentice Hall 2000
Chapter 3
We introduce several terms:
 Nominal model. This is an approximate description of
the plant used for control system design.
 Calibration model. This is a more comprehensive
description of the plant. It includes other features not used
for control system design but which have a direct bearing
on the achieved performance.
 Model error. This is the difference between the nominal
model and the calibration model. Details of this error may
be unknown but various bounds may be available for it.
Goodwin, Graebe, Salgado
©
, Prentice Hall 2000
Chapter 3
Building Models
A first possible approach to building a plant model is to postulate a
specific model structure and to use what is known as a black box
approach to modeling.
In this approach one varies, either by trial and error or by an algorithm, the
model parameters until the dynamic behavior of model and plant match
sufficiently well.
An alternative approach for dealing with the modeling problem is to
use physical laws (such as conservation of mass, energy and
momentum) to construct the model.
In this approach one uses the fact that, in any real system, there are basic
phenomenological laws which determine the relationships between all the
signals in the system.
In practice, it is common to combine both black box and
phenomenological ideas to building a model.
Goodwin, Graebe, Salgado
©
, Prentice Hall 2000
Chapter 3
Control relevant models are often quite simple
compared to the true process and usually combine
physical reasoning with experimental data.
Building Models
Goodwin, Graebe, Salgado
©
, Prentice Hall 2000
Chapter 3
Mathematical Models
For continuous time systems
Goodwin, Graebe, Salgado
©
, Prentice Hall 2000
Chapter 3
Linearization
Although almost every real system includes
nonlinear features, many systems can be reasonably
described, at least within certain operating ranges, by
linear models.
Goodwin, Graebe, Salgado
©
, Prentice Hall 2000
Chapter 3
Thus consider

Say that {x
Q
(t), u
Q
(t), y
Q
(t); t e } is a given set of
trajectories that satisfy the above equations, i.e.
Goodwin, Graebe, Salgado
©
, Prentice Hall 2000
Chapter 3
Goodwin, Graebe, Salgado
©
, Prentice Hall 2000
Chapter 3
Example: Non-linear system
Consider a continuous time system with true model
given by


Assume that the input u(t) fluctuates around u = 2.
Find an operating point with u
Q
= 2 and a linearized
model around it.
Goodwin, Graebe, Salgado
©
, Prentice Hall 2000
Chapter 3
Figure 3.4: Nonlinear system output, y
nl
(t), and linearized
system output, y
l
(t), for a square wave input of increasing
amplitude, u(t).
Goodwin, Graebe, Salgado
©
, Prentice Hall 2000
Chapter 3
Example: Inverted pendulum
Figure 3.5: Inverted pendulum
In Figure 3.5, we have used the following notation:
y(t) - distance from some reference point
u(t) - angle of pendulum
M - mass of cart
m - mass of pendulum (assumed concentrated at tip)
 - length of pendulum
f(t) - forces applied to pendulum
Goodwin, Graebe, Salgado
©
, Prentice Hall 2000
Chapter 3
Example of an Inverted Pendulum
Goodwin, Graebe, Salgado
©
, Prentice Hall 2000
Chapter 3
Application of Newtonian physics to this system
leads to the following model:




where ì
m
= (M/m)
Goodwin, Graebe, Salgado
©
, Prentice Hall 2000
Chapter 3
This is a linear state space model in which A, B and C are:
Goodwin, Graebe, Salgado
©
, Prentice Hall 2000
Chapter 3
Summary
 In order to systematically design a controller for a
particular system, one needs a formal - though possibly
simple - description of the system. Such a description is
called a model.
 A model is a set of mathematical equations that are
intended to capture the effect of certain system variables
on certain other system variables.
Goodwin, Graebe, Salgado
©
, Prentice Hall 2000
Chapter 3
Linear Continuous Time Models
The linear form of this model is:


Introducing the Heaviside, or differential, operator µ():
We obtain:
Goodwin, Graebe, Salgado
©
, Prentice Hall 2000
Chapter 3
Laplace Transforms
The study of differential equations of the type
described above is a rich and interesting subject. Of
all the methods available for studying linear
differential equations, one particularly useful tool is
provided by Laplace Transforms.
Goodwin, Graebe, Salgado
©
, Prentice Hall 2000
Chapter 3
Definition of the Transform
Consider a continuous time signal y(t); 0 s t < ·.
The Laplace transform pair associated with y(t) is
defined as
Goodwin, Graebe, Salgado
©
, Prentice Hall 2000
Chapter 3
A key result concerns the transform of the derivative
of a function:
Goodwin, Graebe, Salgado
©
, Prentice Hall 2000
Chapter 3
Table 4.1: Laplace transform table
Goodwin, Graebe, Salgado
©
, Prentice Hall 2000
Chapter 3
Table 4.2: Laplace transform properties. Note that
| | | | { } . 0 0 ) )( ) ( , ,... 3 , 2 , 1 , ) ( ) ( , ) ( ) (
2 1
< ¬ = = e = = t t f t f k t y s Y t f s F
i i
Goodwin, Graebe, Salgado
©
, Prentice Hall 2000
Chapter 3
Transfer Functions
Taking Laplace Transforms converts the differential
equation into the following algebraic equation



where

and

G(s) is called the transfer function.
This can be expressed as
s
n
Y(s) + a
n 1
s
n 1
Y(s) + : : : + a
0
Y(s)
= b
n 1
s
n 1
U(s) + : : : + b
0
U(s) + f (s; x
o
)
-
-
-
-
Goodwin, Graebe, Salgado
©
, Prentice Hall 2000
Chapter 3
Linear State Space Models
Goodwin, Graebe, Salgado
©
, Prentice Hall 2000
Chapter 3
Transfer Functions for Continuous
Time State Space Models
Taking Laplace transform in the state space model
equations yields


and hence






G(s) is the system transfer function.
Goodwin, Graebe, Salgado
©
, Prentice Hall 2000
Chapter 3
From Transfer Function to State
Space Representation
We have seen above how to go from a state space
description to the corresponding transfer function.
The converse operation leads to the following
question:
Given a transfer function G(s), how can
a state representation for this system be
obtained?
Goodwin, Graebe, Salgado
©
, Prentice Hall 2000
Chapter 3
Development
Consider a transfer function G(s) =
B(s)
/
A(s).
We can
then write


We note from the above definitions that
Goodwin, Graebe, Salgado
©
, Prentice Hall 2000
Chapter 3
We can then choose, as state variables, x
i
(t) = v
i
(t), which
lead to the following state space model for the system.





The above model has a special form. We will see later
that any completely controllable system can be expressed
in this way. Before we do this, we need to introduce the
idea of controllability.
Goodwin, Graebe, Salgado
©
, Prentice Hall 2000
Chapter 3
Example 3.3
Consider the simple electrical network shown in
Figure 3.1. Assume we want to model the voltage
v(t)


On applying fundamental network laws we obtain
the following equations:
Figure 3.1: Electrical
network. State space model.
Goodwin, Graebe, Salgado
©
, Prentice Hall 2000
Chapter 3
These equations can be rearranged as follows:




We have a linear state space model with
Goodwin, Graebe, Salgado
©
, Prentice Hall 2000
Chapter 3
Example 1
Consider a separately excited d.c. motor. Let v
a
(t)
denote the armature voltage, u(t) the output angle. A
simplied schematic diagram of this system is shown
in Figure.
Figure : Simplified model of a d.c. motor
Goodwin, Graebe, Salgado
©
, Prentice Hall 2000
Chapter 3
A laboratory servo kit
Goodwin, Graebe, Salgado
©
, Prentice Hall 2000
Chapter 3
A demonstration robot containing
several servo motors
Goodwin, Graebe, Salgado
©
, Prentice Hall 2000
Chapter 3
Let
J - be the inertia of the shaft
t
e
(t) - the electrical torque
i
a
(t) - the armature current
k
1
; k
2
- constants
R - the armature resistance
Application of well known principles of physics tells
us that the various variables are related by:

Goodwin, Graebe, Salgado
©
, Prentice Hall 2000
Chapter 3
Goodwin, Graebe, Salgado
©
, Prentice Hall 2000
Chapter 3
Solution of Continuous Time
State Space Models
A key quantity in determining solutions to state
equations is the matrix exponential defined as



The explicit solution to the linear state equation is
then given by
Goodwin, Graebe, Salgado
©
, Prentice Hall 2000
Chapter 3
Controllability and Stabilizability
-An important question that lies at the heart of
control using state space models is whether we can
steer the state via the control input to certain
locations in the state space.
-Technically, this property is called controllability or
reachability.
-A closely related issue is that of stabilizability.
-We will begin with controllability.
Goodwin, Graebe, Salgado
©
, Prentice Hall 2000
Chapter 3
Controllability
The issue of controllability concerns whether a given
initial state x
0
can be steered to the origin in finite
time using the input u(t).
Formally, we have the following:
Definition : A state x
0
is said to be controllable if
there exists a finite interval [0, T] and an input {u(t),
t e [0, T]} such that x(T) = 0. If all states are
controllable, then the system is said to be completely
controllable.
Goodwin, Graebe, Salgado
©
, Prentice Hall 2000
Chapter 3
Reachability
A related concept is that of reachability. This
concept is sometimes used in discrete-time systems.
It is formally defined as follows:
Definition : A state is said to be
reachable (from the origin) if, given x(0) = 0, there
exist a finite time interval [0, T] and an input {u(t),
t e [0, T]} such that If all states are
reachable, the system is said to be completely
reachable.
0 = x
. ) ( x T x =
Goodwin, Graebe, Salgado
©
, Prentice Hall 2000
Chapter 3
Test for Controllability
Theorem : Consider the state space model


(i) The set of all controllable states is the range space of
the controllability matrix I
c
[A, B], where


(ii) The model is completely controllable if and only if
where I
c
[A, B] has full row rank.

Proof: Uses Cayley-Hamilton Theorem - see book.

Goodwin, Graebe, Salgado
©
, Prentice Hall 2000
Chapter 3
Example 1
Consider the state space model


The controllability matrix is given by


Clearly, rank I
c
[A, B] = 2; thus, the system is completely
controllable.
Goodwin, Graebe, Salgado
©
, Prentice Hall 2000
Chapter 3
Example 2
For


The controllability matrix is given by:


Rank I
c
[A, B] = 1 < 2; thus, the system is not
completely controllable.
Goodwin, Graebe, Salgado
©
, Prentice Hall 2000
Chapter 3
Stabilizability
A state space model is said to be stabilizable if its
uncontrollable subspace is stable.
Goodwin, Graebe, Salgado
©
, Prentice Hall 2000
Chapter 3
A fact that we will find useful in what follows is
that, if the system is completely controllable, there
exist similarity transformations that convert it into
special forms, known as canonical forms. This is
established in the following two lemmas.
Goodwin, Graebe, Salgado
©
, Prentice Hall 2000
Chapter 3
Controllability Canonical Form
Lemma : Consider a completely controllable state space
model for a SISO system. Then there exists a similarity
transformation that converts the state space model into the
following controllability-canonical form:



where ì
n
+o
n-1
ì
n-1
+ …+ o
1
ì+o
0
= det(ìI - A) is the
characteristic polynomial of A.
Proof: See the book.
Goodwin, Graebe, Salgado
©
, Prentice Hall 2000
Chapter 3
Controller - Canonical Form
Lemma: Consider a completely controllable state space
model for a SISO system. Then there exists a similarity
transformation that converts the state space model into the
following controller-canonical form:



where ì
n
+o
n-1
ì
n-1
+ …+ o
1
ì+o
0
= det(ìI - A) is the
characteristic polynomial of A.
Proof: See the book.

Goodwin, Graebe, Salgado
©
, Prentice Hall 2000
Chapter 3
Finally, we remark that it is very common indeed to
employ uncontrollable models in control-system
design. This is because they are a convenient way of
describing various commonly occurring disturbances.
For example, a constant disturbance can be modeled by
the following state space model:

which is readily seen to be uncontrollable and, indeed,
nonstabilizable.
Goodwin, Graebe, Salgado
©
, Prentice Hall 2000
Chapter 3
Observability and Detectability
Consider again the state space model


In general, the dimension of the observed output, y, can
be less than the dimension of the state, x. However, one
might conjecture that, if one observed the output over
some nonvanishing time interval, then this might tell us
something about the state. The associated properties are
called observability (or reconstructability). A related
issue is that of detectability. We begin with
observability.
Goodwin, Graebe, Salgado
©
, Prentice Hall 2000
Chapter 3
Observability
Observability is concerned with the issue of what can
be said about the state when one is given
measurements of the plant output.
A formal definition is as follows:
Definition 17.6: The state x
0
= 0 is said to be
unobservable if, given x(0) = x
0
, and u[k] = 0 for k > 0,
then y[k] = 0 for k > 0. The system is said to be
completely observable if there exists no nonzero
initial state that it is unobservable.
Goodwin, Graebe, Salgado
©
, Prentice Hall 2000
Chapter 3
Reconstructability
A concept related to observability is that of
reconstructability. This concept is sometimes used
in discrete-time systems. Reconstructability is
concerned with what can be said about x(T), on the
basis of the past values of the output, i.e., y[k] for
0 s k s T. For linear time-invariant continuous-time
systems, the distinction between observability and
reconstructability is unnecessary. However, the
following example illustrates that, in discrete time,
the two concepts are different.
Goodwin, Graebe, Salgado
©
, Prentice Hall 2000
Chapter 3
Consider


this system is clearly reconstructable for all T > 1,
because we know for certain that x[T] = 0 for T > 1.
However, it is completely unobservable, because
y[k] = 0 ¬k, irrespective of the value of x
0
.
Goodwin, Graebe, Salgado
©
, Prentice Hall 2000
Chapter 3
In view of the subtle difference between observability
and reconstructability, we will use the term
observability in the sequel to cover the stronger of the
two concepts.

Goodwin, Graebe, Salgado
©
, Prentice Hall 2000
Chapter 3
Test for Observability
A test for observability of a system is established in
the following theorem.
Theorem : Consider the state model

(i) The set of all unobservable states is equal to the null
space of the observability matrix I
0
[A, C], where
Goodwin, Graebe, Salgado
©
, Prentice Hall 2000
Chapter 3
(ii) The system is completely observable if and only if
I
0
[A, C], has full column rank n.

Proof: See the book.
Goodwin, Graebe, Salgado
©
, Prentice Hall 2000
Chapter 3
Example 1
Consider the following state space model:


Then


Hence, rank I
0
[A, C] = 2, and the system is
completely observable.
Goodwin, Graebe, Salgado
©
, Prentice Hall 2000
Chapter 3
Example 2
Consider


Here


Hence, rank I
0
[A, C] = 1 < 2, and the system is not
completely observable.
Goodwin, Graebe, Salgado
©
, Prentice Hall 2000
Chapter 3
Duality
We see a remarkable similarity between the
controlability and observability results. We can
formalize this as follows:
Theorem (Duality). Consider a state space model
described by the 4-tuple (A, B, C, D). Then the
system is completely controllable if and only if the
dual system (A
T
, C
T
, B
T
, D
T
) is completely
observable.
Goodwin, Graebe, Salgado
©
, Prentice Hall 2000
Chapter 3
Observer Canonical Form
There are also duals of the canonical forms given in
Lemmas 17.2 and 17.3. For example the dual of
Lemma 17.3 is
Lemma : Consider a completely observable SISO
system given by

Then there exists a similarity transformation that
converts the model to the observer-canonical form
Goodwin, Graebe, Salgado
©
, Prentice Hall 2000
Chapter 3
Goodwin, Graebe, Salgado
©
, Prentice Hall 2000
Chapter 3
Canonical Decomposition
Further insight into the structure of linear dynamical
systems is obtained by considering those systems
that are only partially observable or controllable.
These systems can be separated into completely
observable and completely controllable systems.
The two results of previous lemmas can be combined
as on the next slide.
Goodwin, Graebe, Salgado
©
, Prentice Hall 2000
Chapter 3
Canonical Decomposition Theorem
Theorem : (Canonical Decomposition Theorem).
Consider a system described in state space form. Then,
there always exists a similarity transformation T such
that the transformed model for takes the form x x
1 ÷
= T
Goodwin, Graebe, Salgado
©
, Prentice Hall 2000
Chapter 3
Where
(i) The subsystem is both completely
controllable and completely observable and has the
same transfer function as the original system.
] , , [
1 1 0
C B A
c
Goodwin, Graebe, Salgado
©
, Prentice Hall 2000
Chapter 3
(ii) The subsystem



is completely controllable.
Goodwin, Graebe, Salgado
©
, Prentice Hall 2000
Chapter 3
(iii) The subsystem



is completely observable.
Goodwin, Graebe, Salgado
©
, Prentice Hall 2000
Chapter 3
The canonical decomposition described above leads to
Lemma : Consider the transfer-function matrix H(s)
satisfying

Then

where and correspond to the observable
and controllable part of the model.
, ,
0 1 c
A C
1
B
Goodwin, Graebe, Salgado
©
, Prentice Hall 2000
Chapter 3
Previous lemma shows that the uncontrollable and
the unobservable parts of a linear system do not
appear in the transfer function.
Conversely, given a transfer function, it is possible
to generate a state space description that is both
completely controllable and observable.
We then say that this state description is a minimal
realization of the transfer function.
As mentioned earlier, nonminimal models are
frequently used in control-system design to include
disturbances.
Goodwin, Graebe, Salgado
©
, Prentice Hall 2000
Chapter 3
Controllability depends on the structure of the input
ports: where, in the system, the manipulable inputs
are applied.
Thus the states of a given subsystem might be
uncontrollable for one given input but completely
controllable for another.
This distinction is of fundamental importance in
control-system design, because not all plant inputs
can be manipulated (consider, for example,
disturbances) to steer the plant to reach certain
states.
Goodwin, Graebe, Salgado
©
, Prentice Hall 2000
Chapter 3
Similarly, observability depends on which outputs are being
considered.
Certain states may be unobservable from a given output, but
they may be completely observable from some other output.
This also has a significant impact on output-feedback control
systems, because some states might not appear in the plant
output being measured and fed back.
However, they could appear in crucial internal variables and
thus be important to the control problem.
Goodwin, Graebe, Salgado
©
, Prentice Hall 2000
Chapter 3
Pole-Zero Cancellation and
System Properties
The system properties described above are also
intimately related to issues of pole-zero
cancellations. To facilitate the subsequent
development, we introduce the following test, which
is useful for studying issues of controllability and
observability.