1
Course information
Recommended references
R. L. Williams II & D. A. Lawrence, Linear State-Space
Control Systems Wiley, 2007.
Chi-Tsong Chen, Linear System Theory and Design
2
Foundations in Control Engineering
In this course, we learn a basic theory of
automatic feedback control systems.
What is control?
What is automatic control? And why?
What is feedback control? And why?
3
What is Control?
Make some object (system, plant) behave as
we desire.
Imagine control around you!
Room temperature control
Car/bicycle driving
Balance of bank account
Control (move) the position of the pointer
Laundry machine
Automobile direction and speed control
Airplane, rocket, satellite control
4
What is Automatic Control?
Not manual!
Why do we need automatic control?
Convenient (room temperature control, laundry
machine)
Dangerous (hot/cold places, space, bomb removal)
Impossible for human (nanometer scale precision
positioning, work inside the small space that human
cannot enter)
It exists in nature. (human body temperature control)
Low cost, high efficiency, etc.
Many examples of automatic control around us
5
Open-loop & closed-loop control
Open-loop Disturbance
Sensor
Sensor
Sensor
4. Implemenation 1. Modeling
2. Analysis
3. Design
8
Goal of classical control course
To learn basics of feedback control systems
Modeling as a transfer function and a block diagram
Laplace transform (Mathematics!)
Mechanical, electrical, electromechanical systems
Analysis
Time response, frequency response
Stability: Routh-Hurwitz criterion, Nyquist criterion
Design
Root locus technique, frequency response technique,
PID control, lead/lag compensator
9
Weak points of classical control
Classical control theory deals with only
Single-input single-output (SISO) systems
Linear time-invariant (LTI) systems
No optimality concept in controller design
No robustness consideration, except gain
margin and phase margin
10
Brief history of Control Engineering
Classical control (-1950)
Transfer function
Frequency domain
Modern control (1960-) (contents in this course)
State-space model
Time domain
Optimality
Post-modern control (1980-)
Robust control
Hybrid control, etc.
11
Linear state-space models
Continuous-time Discrete-time
x : state vector
u : input vector
y : output vector
System System
u y u y
12
Advantages of state-space models
State-space models can represent fairly general
systems, such as
Multivariable systems
Time-varying systems
Nonlinear systems
Under certain conditions, variables inside the
system (=states) can be
controlled without actuators,
observed without sensors.
Analysis and design can be done in a
numerically reliable manner.
13
Goals of the course
To learn control
theory with
theorylinear
withstate -space
linear statemodels
Modeling as a state-space model
Differential or difference equation (instead of TF)
Linear algebra (instead of Laplace transform)
Analysis
Stability, controllability, observability
Realization, minimality
Design
State
Statefeedback,
feedback,observer
observer
Matlab simulation
14
Course plan
Topics CT DT
Modeling
Stability
Controllability/observability
Realization
State feedback/observer
15
Summary
Introduction
This course gives foundations for advanced
control theory.
nonlinear control
robust and optimal control
adaptive control
digital control, sampled-data control
hybrid control
system identification
Next, model classification
16
Model classification
17
Controller design procedure (review)
Disturbance
Ref. Input Output
Controller Actuator plant
Sensor
4. Implemenation 1. Modeling
Controller Mathematical model
2. Analysis
3. Design
Input
Physical Output
system
Modeling
Model
19
Important remarks on models
No mathematical model exactly represents a physical
system.
20
Model classification
Continuous-time and discrete-time
Memoryless, causal and noncausal
Lumped and distributed
Time-invariant and time-varying
Linear and nonlinear
21
Model classification
Continuous-time and discrete-time
Continuous-time system
Input/output vectors are continuous-time signals.
System
Discrete-time system
Input/output vectors are discrete-time signals.
22
Example
Continuous-time system K B
Mass-spring-damper system M
f(t) y(t)
RLC circuit
v(t)
i(t) R L
C
Discrete-time system
Digital computer
Daily balance of a bank account y[k] : balance at k-th day
u[k] : deposit/withdrawal
a : interest rate
23
Model classification
Continuous-time and discrete-time
Memoryless, causal and noncausal
Lumped and distributed
Time-invariant and time-varying
Linear and nonlinear
24
Model classification
Memoryless, causal, and noncausal
Memoryless system : Current output depends on only
current input.
Causal system : Current output depends on current
and past input.
Noncausal system : Current output depends on future
input.
26
Model classification
Continuous-time and discrete-time
Memoryless, causal and noncausal
Lumped and distributed
Time-invariant and time-varying
Linear and nonlinear
27
Model classification
Lumped and distributed
For a causal system,
System
Distributed system
Unit time delay system y(t)=u(t-1)
29
Model classification
Continuous-time and discrete-time
Memoryless, causal and noncausal
Lumped and distributed
Time-invariant and time-varying
Linear and nonlinear
30
Model classification
Time-invariant and time-varying
For a causal system,
System
31
Example
Car, rocket etc.
M M(t)
If we regard M to be If we regard M to be
constant (even though M changing (due to, e.g., fuel
changes very slowly), then consumption), then this
this system is time-invariant. system is time-varying.
32
Model classification
Continuous-time and discrete-time
Memoryless, causal and noncausal
Lumped and distributed
Time-invariant and time-varying
Linear and nonlinear
33
Model classification
Linear and nonlinear
For a causal system,
34
Remarks
All systems in real world are nonlinear.
This linear relation holds
only for small y(t) and f(t).
35
Summary
Model classification
Continuous-time and discrete-time
Memoryless, causal and noncausal
Lumped and distributed
Time-invariant and time-varying
Linear and nonlinear
In this course, we will not consider noncausal,
distributed, nonlinear systems in this course.
Next, state-space models
36
State-space models
37
Review and todays topic
Last lecture was about model classification
Continuous-time and discrete-time
Memoryless, causal and noncausal
Lumped and distributed
Time-invariant and time-varying
Linear and nonlinear
Today, we introduce linear state-space models
to describe causal lumped linear systems.
38
Linear state-space models
Continuous-time Discrete-time
x : state vector
u : input vector
y : output vector
System System
u y u y
39
Remarks
The first equation, called state equation, is a first
order ordinary differential (CT case) and
difference (DT case) equation.
The second equation, called output equation, is
an algebraic equation.
Two equations are called state-space model.
If a system is time-invariant, the matrices A, B,
C, D are constant (independent of time).
Pay attention to sizes of matrices and vectors.
They must be always compatible!
40
Examples
Mass with a driving force
Mass-spring-damper system
RLC circuit
DC motor
41
Mass with a driving force
By Newtons law, we have
My(t)=u(t) M
u: input force,
y: output position
Define state variables
Then,
42
Mass-spring-damper system
By Newtons law
K B
M
43
RLC circuit
u(t): input voltage u(t)
i(t) R L
y(t): output voltage
C y(t)
By Kirchhoffs voltage law
Current for inductor
Voltage for capacitor
44
Model of DC motor
La Output
ia(t) Ra
ea(t) eb(t) Jm
Input Bm
45
Modeling of DC motor
Armature circuit
Mechanical load
Angular position
46
DC motor: output = speed
By substitution
47
DC motor: output = position
By substitution
48
DC motor: two outputs
Output vector
49
Summary
Linear state-space models
Examples
Mechanical systems
Electrical systems
DC motor
Next, linearization
50
Linearization
51
Why linearization?
Real systems are inherently nonlinear. (Linear
systems do not exist!) Ex. f(t)=Kx(t), v(t)=Ri(t)
Nonlinear systems are difficult to deal with
mathematically.
Many control analysis/design techniques are
available for linear systems.
Linear approximation is often good enough for
control system analysis and design purposes.
How to linearize nonlinear systems?
52
Todays outline
Examples of nonlinear systems
Linearization of simple functions (1-dim, 2-dim)
Linearization of nonlinear systems
Examples revisited
53
A pendulum
Input torque
Motion of a pendulum
Nonlinear!
54
Water level control
Mass flow rate equation
55
Todays outline
Examples of nonlinear systems
Linearization of simple functions (1-dim, 2-dim)
Linearization of nonlinear systems
Examples revisited
56
Linearization: 1-dim case
Linearize a function y=f(x) around x=x0 (scalar)
Consider a solution (x0,y0)
57
Example of 1-dim case
Linearize a function around x=1
58
Linearization: 2-dim case
Linearize a function y=f(x) around x=x0
Consider a solution
Jacobian
59
Example of 2-dim case
Linearize a function below around
Jacobian computation
60
Todays outline
Examples of nonlinear systems
Linearization of simple functions (1-dim, 2-dim)
Linearization of nonlinear systems
Examples revisited
61
Linearization of nonlinear system
Nonlinear system (possibly time-varying)
62
Linearization of state equation
A(t) B(t)
C(t) D(t)
65
A pendulum revisited
Nonlinear model
Linearization around
Note that this function satisfies state-space model.
66
Water level control revisited
Nonlinear model
Linearization around
Note that this function satisfies state-space model.
67
Summary
Linearization of nonlinear systems
Examples
Pendulum
Water level control
Inverted pendulum (Appendix)
Next, solution to state-space models
68
Homework
1-3: Consider a state-space model (see next
page) for a satellite:
Satellite (m)
Earth
69
Homework (contd)
Nonlinear state-space model
i-th element
70
Homework (contd)
(a) Verify that the following is a solution to the
state-space equation.
71
Another nonlinear example
Cart with an inverted pendulum
Motion of a pendulum
Nonlinear!
Explicit form
omitted here
72
Solution to continuous-time
LTI systems
73
Some acronyms and notation
SS : State-space
CT : Continuous-Time
DT : Discrete-Time
LTI : Linear Time-Invariant
LTV : Linear Time-Varying
A:=B : A is defined by B.
: Set of real numbers
: Set of complex numbers
74
Todays topic
How to solve state-space model equation
75
Solution to CT LTI SS model
CT LTI state-space model
Memorize this!
Solution
where
76
Remarks
We can see the linearity of the system from this
equation.
Verification :
77
Matrix exponential
Definition
Property
78
1. Definition of matrix exponential
Nilpotent matrix
=0
Diagonal matrix
79
2. Laplace transform
Formula
cf. scalar case: Partial fraction expansion
Ex:
Ex
80
3. Diagonal form
Suppose that we have distinct eigenvalues and
corresponding eigenvectors as
Then,
Similarity transformation
81
Diagonal form: proof
Suppose that we have eigenvalues/vectors as
If T is nonsingular
By definition,
82
Diagonal form: example
Ex
Eigenvalue computation
Eigenvector computation
Important:
in Matlab, use expm for
matrix exponential, and 83
not just exp
Summary
Solution to CT LTI systems
Computation of matrix exponential
By definition
By Laplace transform
By diagonal form or Jordan form
By Cayley-Hamilton Theorem
Next,
Discretization
Solution to discrete-time LTI systems
84
Homework
1-4. Solve the state equation
85
Jordan form
For any matrix A, there is a nonsingular T s.t.
Matrix exponential
86
Cayley-Hamilton Theorem
For a matrix A, the following holds:
88
Discretization
Solution to DT LTI systems
89
What is discretization?
Approximation of a CT system by a DT system.
CT system
DT system
90
Why discretization?
Digital control : For realization of a controller in a
digital computer, we need a DT controller.
Design
CT system CT controller
Discretization Discretization
DT system DT controller
Design
91
Digital control system
DT controller
D/A Actuator CT system
Stabilization,
Disturbance
and noise
rejection, A/D Sensor
Regulation,
Tracking etc.
92
Advantages of digital control
Reduced cost
A single digital computer can replace numerous
analog controllers. ( Reduction in cost!)
Flexibility in response to design changes
Any modifications that are required in the future can
be implemented with simple software changes rather
than expensive hardware modifications.
High reliability
93
Analog controller inflexibility
Lead compensator using operational amplifiers
To modify the controller,
we need to physically replace
C1 C2 electrical elements.
R4
R2
R1
R3
- -
vi(t) + + vo(t)
94
Discretization by ZOH
Given CT system & sample T, find DT system:
DT sys
95
ZOH by state-space model
Continuous-time system
where
(C&D: unchanged!)
96
An example
Mass with a driving force
M=1
97
Simulation with Matlab
Code for discretization 0.7
Step responses
StepResponse
for
CT & DT systems
0.6
Amplitude
Sampling time
Discretization 0.3
0.2
0.1
0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Time(sec)
98
Derivation (optional)
Solution to state equation
x[k+1]
Solution
k=0:
k=1:
k=2:
By induction
102
Summary
Discretization (Zero-Order Hold)
Digital control
Formula of discretized state-space model
Example with Matlab simulation
Solution to DT LTI SS models
103
1-5. Discretize the following continuous-time
system by ZOH with a sampling period T=1:
104
BIBO stability
105
Course plan
Topics CT DT
Modeling 9 9
Stability
Controllability/observability
Realization
State feedback/observer
LQR/Kalman filter
106
Stability
Utmost important specification in control design!
After stability, performance (tracking,
disturbance/noise attenuation, etc.)
Unstable systems to be stabilized by feedback
Unstable closed-loop systems are useless.
Unstable system may hit mechanical/electrical
stops (saturation), may break down or burn out.
In this course, we learn two types of stability:
BIBO stability
Internal stability
107
Bounded signal
Definition: A signal u(t) is called bounded if there
exists a um such that
u(t) u(t)
bounded unbounded
um
-um
108
BIBO stability
Definition: A system is called BIBO stable if
every bounded (possibly vector) input u(t)
excites bounded (possibly vector) output y(t).
u(t) y(t)
ym
um
-um -ym
System
109
BIBO stability for CT LTI systems
Im
An CT LTI system G(s) is BIBO stable
if and only if
all the poles of G(s) are in the open Re
left half of the complex plane.
111
Simple criteria for BIBO stability
1st order polynomial
112
Examples
G(s) stable?
Yes / No
Yes / No
Yes / No
Yes / No
Yes / No
113
BIBO stability for DT LTI systems
Im
An DT LTI system G(z) is BIBO stable
j
if and only if
1
all the poles of G(z) are in the open Re
unit disc of the complex plane.
114
From SS to TF (Continuous-time)
CT LTI SS model
Memorize this!
115
BIBO stability for CT LTI SS model
Example
SS model
TF
Pole is s=-1
(After pole/zero cancellation) BIBO stable!
116
Important remark
In this example, CT LTI system is BIBO stable.
However, from the state equation,
System
Internal stability
Does x(t) converge to 0?
x(0): nonzero y(t)
u(t)=0 System
118
Summary
BIBO stability
Definition
Conditions for both CT and DT LTI systems
Examples
Transform from state-space model to transfer
function model
Next, internal stability
119
Homework
2-1. Consider a feedback system:
r(t) y(t)
k
120
Z-transform (review)
Definition: For a sequence {f[k]: k=0,1,2,.}
Shift property
121
From SS to TF (Discrete-time)
DT LTI SS model
Memorize this!
122
Internal stability
123
Course plan
Topics CT DT
Modeling 9 9
Stability
Controllability/observability
Realization
State feedback/observer
LQR/Kalman filter
124
Review & todays topic
Last lecture was about BIBO stability.
BIBO stability cannot deal with stability for state-
space model with nonzero initial states.
Today, we study internal stability for
LTI CT:
LTI DT:
125
Issue in BIBO stability
Consider three BIBO stable open-loop systems:
126
Issue in BIBO stability (contd)
Why is unstable cancellation by series
connection not allowed?
Plant model is never exact! (and hence cancellation
will not occur in real world.)
Neither nonzero initial condition nor input disturbance
is allowed. (very fragile)
Some internal signal may go unbounded, even if
output is bounded.
BIBO stability cannot detect such cancellation,
but internal stability can!
127
Internal stability: Mechanical example
Imagine what will happen if we start near x=0.
128
Phase plot
Plot of state trajectory in state-space
129
Definition of internal stability
Consider system (no input): (DT case is analogous.)
130
Eigenvalue criteria
Eigenvalue of A :
Stability CT DT
Asymptotically
stable
Marginally
stable For For
131
Idea of stability condition
CT DT
Examples Examples
System is unstable.
Asymptotically stable
Marginally stable Unstable
134
Examples
Are systems with the following A-matrices
asymptotically stable, marginally stable, or
unstable?
CT DT CT DT CT DT
135
Remark on stability for LTV system
Example:
137
Homework
2-2: Are systems with the following A-matrices
asymptotically stable, marginally stable, or
unstable?
CT DT CT DT CT DT
138
Lyapunov Theorem
139
Review & examples
Last time, eigenvalue criteria for internal stability.
CT : asymptotically stable
DT : asymptotically stable
Pendulum Inverted
pendulum
Marginally
Unstable
stable
140
Todays topic & Outline
We will study another condition for internal
stability, called Lyapunov Theorem.
Outline
Positive definite matrix
Lyapunov Theorem
Theorem
Example
Idea of the theorem
141
Positive definite matrix
A symmetric matrix P (n-by-n) is called positive
definite if
Example
142
Facts on positive definite matrix
Fact 1:
Fact 2 (Sylvesters criterion):
Ex.
Ex.
145
Lyapunov Theorem (CT case)
All the eigenvalues of a matrix A have negative
real parts (in the open left-half plane)
if and only if
solution P of the following Lyapunov equation
147
Remarks
Normally, Q is taken to be the identity matrix.
For LTI SS model, Lyapunov Theorem has no
advantage over eigenvalue criteria.
In this course, we want to understand the idea of
the theorem, which will be useful in studying
advanced control.
148
Outline
Positive definite matrix
Lyapunov Theorem
Theorem
Example
Idea of the theorem
149
Example
CT case
Not unique!
NOT all eigs. are in the unit disc!
150
Outline
Positive definite matrix
Lyapunov Theorem
Theorem
Example
Idea of the theorem
151
Idea of Lyapunov Theorem
1. For a fixed P>0, define
a (Lyapunov) function
V(x)
152
Idea of Lyapunov Theorem (contd)
2. Consider the trajectory
V(x(t)), where the
behavior of x(t) is
subject to
V(x(t))
(Altitude of
your position)
x(t) 153
Idea of Lyapunov Theorem (contd)
3. Take a derivative of
V(x(t)) w.r.t. t:
In DT case, take
x(t) 154
Idea of Lyapunov Theorem (contd)
4. If AP+PA=:-Q<0, then
x(t) 155
Idea of Lyapunov Theorem (contd)
Solving the Lyapunov
equation w.r.t. P
is equivalent to finding a
(Lyapunov) function V(x)
s.t.
x(t) 156
Summary
Lyapunov Theorem
Positive definite matrix
Lyapunov equation
Main idea of the theorem
157
Appendix: Who is Lyapunov?
Aleksandr Lyapunov (1857-1918)
Russian mathematician
University of Saint Petersburg
Stability analysis
Year: 1892
Lyapunov method
Applicable to nonlinear, time-
varying systems, too.
(Normally taught in
Nonlinear Control Course.)
158
Controllability
159
Controllability & observability
Consider a system with state vector:
x1: position
Car1
x2: velocity Intuitively, we cannot
Control x3-x6
x3: position
Car2 Observe x1-x4
x4: velocity
x5: position
Car3
x6: velocity
161
Model of three car example
State-space model
162
Why controllability & observability
important?
Clarify essential/redundant sensors/actuators
Clarify the structure of the system
Which state is controllable, uncontrollable,
observable, unobservable?
Which state is redundant from input-output viewpoint?
(Minimality of a realization)
Clarify possibilities and limitations in
Control (state-feedback, linear quadratic regulator)
Estimation (observer, Kalman filter)
163
A mechanical example
Mass-spring-damper
164
Some questions
If I want to transfer any initial positions &
velocities to any final positions & velocities in a
specified time, is it possible?
What if we use only one actuator? (because the other
actuator broke down)
What if we set some spring and/or damper to be
zero?
Can you answer these questions by intuition?
If yes, how about the system that has 100 (instead of
two) mass-spring-damper subsystems?
165
Controllability for LTI system
Consider a state equation
166
Necessary and sufficient condition
for controllability
Controllability matrix
has full row rank, i.e.,
Rank = 1 or 2 ???
167
Examples
Size of A-matrix
Ex
Uncontrollable!
Ex
Controllable!
Ex
Uncontrollable!
168
Three car example
State-space model (m1=1)
Controllability matrix
Uncontrollable!
170
Controllability test: results
b1 b2 k1 k2 Rboth Ru1 Ru2
1 1 1 1 4 4 4
0 1 1 1 4 4 4
1 0 1 1 4 4 4
1 1 0 1 4 4 3
1 1 1 0 4 3 4
0 1 1 0 4 4 4
1 0 0 1 4 4 3
171
Uncontrollable cases
k1=0, u1=0
k2=0, u2=0
b2=k1=0, u1=0
172
Summary
Controllability
Definition
Necessary and sufficient condition
Rank computation
Mechanical example
Next,
when (A,B) is controllable, minimum energy control
when (A,B) is uncontrollable, what can we do and
what cannot we do?
173
Homework
2-3: Read and understand Example 3-8, pp 86-
88 of Ogatas book.
Derive a SS model for inverted pendulum from (3.55)
& (3.56), with friction of the cart (with coefficient b),
1 inputs: u
4 output: Position & velocity of the cart, angle and
angular velocity of the pendulum.
Parameter values are given in the next slide.
Using Matlab, check the controllablity.
Attach Matlab code in your report.
174
Homework (contd)
Parameter values
175
EXAMPLE 3-7 Obtain the transferfunctionsX(s)/U(s) and X2$)/U(s) of the mechanicalsystemshownin
unT;"1lnr"tions
of motion for the systemshownin Figure3-19 are
mrir: - k t x r - k r l * , * x 2 )- b \ * y - i 2 ) + u
Simplifying,we obtain
mri, + bit + (kt * k)x1 : b*z * k2x2I u
xr(s) m r s z+ b s + k r + k ,
(346)
u(s) ( m r s 2+ b s + k r + k 2 ) ( m 2 s* 2b s * k 2 + h ) - ( u s + *r)2
XzG) bs+k2
(347)
u(s) + bs + k2 + t < r ) - ( o s + * r ) 2
(m,sz + bs * k1 + kr)(m2s2
Equations(3-46)and (347) arethe transferfunctionsX 1$)/U (s) andX r(s)/ U (s), respectively.
Figure F19
Mechanicalsystem.
EXA,IVIPLE 3-8 An inverted pendulum mounted on a motor-driven cart is shown in Figure 3-20(a).This is a model
of the attitude control of a space booster on takeoff. (The objective of the attitude control prob-
lem is to keep the space booster in a vertical position.) The inverted pendulum is unstable in that
it may fall over any time in any direction unless a suitable control force is applied. Here we consider
.1)
r5'l
get
(a)
-46)
\4'7)
.ivelY.
figure 3-20
r a) Inverted
g'endulumsystem;
r h ) free-bodY
(b)
Jiagram.
^ * O +' t s i n o ) = u (34e)
dt'
d2 .,
rn*(tcos9):V-mg (3-50)
M4::: U_ H (3-s1)
dt'
Sincewe mustkeepthe invertedpendulumvertical,we canassume thatg(r) ande(r) are
suchthatsin0 = 0,cosd -- L, and002: 0.Then,Equations
smallquantities (3+gj througrrid_sol
canbelinearized.The linearizedequationsare
I(i:Vt|-Ht (3-s2)
m ( i + t 6 1= n (3-s3)
0:V-mg (3-s4)
From Equations(3-51) and (3-53),we obtain
(M+m)i+ml(i :u (3-ss)
From Equations (3-52), (3-53), and (3-54), we have
I(): mgl| - Ht
: mgl| - l(mi + mtd)
or
(t+mt2)6+mti:msto (3-s6)
Equations (3-55) and (3-56) describethe motion of the inverted-pendulum-on-the-cartsystem.
They constitutea mathematicalmodel of the system.
179
Review
Controllability
System
(A,B) is controllable if, for any x0 and any x1,
there is u(t) which transfers from x0 to x1 in a
(specified) finite time.
Nec. and suf. condition for controllability
Controllability matrix has full row rank, i.e.,
180
Todays topics
If (A,B) is controllable
Minimum energy control
If (A,B) is uncontrollable
Decomposition into controllable part and
uncontrollable part
181
Example
M
Consider a controllable system:
182
Nec & suf conditions for controllability
Controllability matrix has full row rank, i.e.,
Ans.
(Proof given
in Appendix)
184
Example revisited
System: (M=1)
185
Example (contd)
1
position
0.5
x1
0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
t
2
1
velocity
x2
0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
t
10
0
acceleration
u
-10
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
t
186
Remarks
Minimum energy control is an open-loop control.
System
187
Todays topics
If (A,B) is controllable
Minimum energy control
If (A,B) is uncontrollable
Decomposition into controllable part and
uncontrollable part
188
Coordinate transformation
System
Coordinate transformation
190
Decomposition
If (A,B) is not controllable with
then there exists a coordinate transformation
(i.e., nonsingular T) that decomposes states into
controllable part and the uncontrollable part:
191
How to find T? (Check ctrbf.m)
We use image space of controllability matrix.
Ex.
Uncontrollable!
192
Image space
For a matrix M (q-by-p):
Im M
A basis of Im M
193
Summary
Controllability
If controllable, minimum energy control
If not controllable, decomposition
Next, observability
194
Homework
2-4: Consider a controllable system
196
Appendix (contd)
Proof of minimality
A
Suppose that other u satisfies the state transition.
B
A-B
Cauchy-Schwartz inequality
197
Decomposition for controllability
198
Coordinate transformation (review)
System
Coordinate transformation
200
Why is decomposition important?
We can see what is not possible by using control
input u. (We cannot affect at all the
uncontrollable part.)
If uncontrollable part is unstable, then we cannot
stabilize the system by feedback. (We will learn
this in more detail later.)
This may suggest addition of actuators, or
change of plant parameters and actuator
locations.
201
Image space
For a matrix M (q-by-p):
Im M
A basis of Im M
202
How to find T? (proof omitted)
We use image space of controllability matrix.
Ex.
Uncontrollable!
203
Another example
Ex
Controllability matrix
Uncontrollable!
Transformation matrix
204
Matlab command ctrbf.m
205
Matlab command (contd)
Example
206
Mechanical example: revisited
Mass-spring-damper with k1=u1=0
207
Mechanical example: contd
Controllability matrix
208
Remarks
Rank of controllability matrix indicates the
number of controllable new states (z).
In the example, z4 is the uncontrollable state. In
fact, z4 is constant for any input, since
209
Exercise
Using Matlab, for the following two systems,
compute T which decomposes states into
controllable and uncontrollable parts, and figure
out which state is not controllable (Set all the
parameters (k,b,m) to be one.)
210
Summary
Decomposition for controllability
Image space
Examples
Matlab command for decomposition
Next, observability
211
Observability
212
Review & todays topic
Controllability: How much can we control x by
controlling u? (Done)
Nec. & suf. condition
Minimum energy control
Decomposition
Observability: How much can we observe x by
measuring y? (Todays topic)
213
Very simple example: revisited
Three cars with one input and one output
x1: position
Car1
x2: velocity Intuitively, we cannot
Control x3-x6
x3: position
Car2 Observe x1-x4
x4: velocity
x5: position
Car3
x6: velocity
214
Model of three car example
State-space model
215
Observability for LTI system
System equations (no input)
System
216
Remarks
If a system is observable, then we can
determine x(0), and therefore, {x(t), t>0}.
However, this is possible after future time (t1)
comes, and it is not practical.
In practice, we want to estimate x(t).
Later in this course, we will learn design of state
estimator, called observer.
Observability is necessary to construct a
successful observer.
217
Necessary and sufficient condition
for observability
Observability matrix
Unobservable!
Ex.
Observable!
219
Three car example: revisited
Model
Observability matrix
Unobservable!
221
Duality
Control Estimation
System
222
Mathematical duality
(A,B) is controllable (A,B) is observable.
(A,B) is observable
223
Decomposition
If (A,C) is not observable with
then there exists a coordinate transformation
(i.e., nonsingular T) that decomposes states into
observable part and the unobservable part:
224
Kernel space
For a matrix M (q-by-p):
ker M
A basis of ker M
225
How to find T? (proof omitted)
We use kernel space of observability matrix.
Ex. Unobservable!
226
Summary
Observability
Definition
Condition by observability matrix
Duality between controllability and observability
Decomposition (Matlab command obsvf.m)
(We can use duality for decomposition.)
Next, Kalman decomposition
227
2-5: Check the controllability and observability
1. System
2. System
228
Kalman decomposition
229
Review & Todays topic
So far, we have learned controllability and
observability:
Condition
Decomposition
Duality
Today, we will study the combination of
decompositions for controllability and
observability, called Kalman decomposition.
230
Rudolf Emil Kalman (1930-)
American-Hungarian
Mathematical system theorist
During 1950s & 60s, he developed
state-space control theory
Controllability, observability
Linear quadratic regulator
Kalman filter
etc.
Most of the theory in this course
have been established by Kalman!
231
Two decompositions
SS model
System
Decomposition for Decomposition for
controllability observability
232
Kalman decomposition (idea)
SS model
System
233
Kalman decomposition (idea)
Conceptual figure (Not block diagram!)
System
234
Kalman decomposition
Every SS model can be transformed by z=Tx for
some appropriate T into a canonical form:
235
Remarks
(Aco,Bco): controllable & (Aco,Cco): observable
Transfer function is determined by ONLY
controllable & observable parts.
236
How to find T? (review)
For controllability, we used image space of
controllability matrix.
237
How to find T?
For Kalman decomposition,
238
Very simple example: revisited
Three cars with one input and two outputs
x1: position
Car1
x2: velocity
x5: position
Car3
x6: velocity
239
Model of three car example
State-space model
240
Derivation of T
Controllability matrix
Observability matrix
241
Example (contd)
After transformation,
242
Matlab command minreal.m
243
Summary
Kalman decomposition
Next, controllability and observability for
discrete-time systems
244
Homework
3-1. Using Matlab, do Kalman decomposition of the
following systems. Write explicitly which states are
controllable / uncontrollable and observable /
unobservable. If possible, do hand-calculations too.
245
Decomposition (review)
If (A,C) is not observable with
then there exists a coordinate transformation
(i.e., nonsingular T) that decomposes states into
observable part and the unobservable part:
246
Kernel space
For a matrix M (q-by-p):
ker M
A basis of ker M
247
How to find T? (proof omitted)
We use kernel space of observability matrix.
Ex. Unobservable!
248
Controllability and observability
Discrete-time case
249
Why contollability & observability?
Definitions of controllabiltiy (open-loop) and
observability (estimation of past) are not so
practical.
They are important properties in many feedback
controller and state estimator design techniques.
The lack of these properties limits achievable
performance. (later in this course.)
Note that stability, controllability, and
observability are independent properties.
250
Review & todays topic
So far, for CT systems, we learned controllability
and observability:
Definitions
Conditions
Duality
Kalman decomposition
Today, we study discrete-time counterpart
(which is almost the same as CT results; actually
easier than CT cases!).
251
An example
Mass-spring-damper
252
Controllability for DT LTI system
Consider a state equation
253
Observability for DT LTI system
System equations (no input)
System
254
Conditions
255
Derivation for controllability condition
Solve recursively
256
Simple examples
Uncontrollable case
No solution.
Controllable case
There is a solution.
257
Derivation for observability condition
Solve recursively
given
258
Simple examples
(Unknown)
Unobservable case
Non-unique solutions.
Observable case
Unique solution.
259
Minimum energy control
If controllable, find the input with minimum
energy (least-squares sum), i.e., solve
LS solution
260
Remark
Recall that, for CT systems, if (A,B) is
controllable, then any state transfer is possible in
any (epsilon) time.
However, this is NOT the case for DT systems.
Ex.
261
An example: revisited
Set m=1, k=4, b=0.
262
An example (contd)
Is the discretized system controllable?
263
Summary
DT controllability and observability (very similar
to CT results.).
Duality & Kalman decomposition apply to DT
systems, too.
264
Realization
265
Motivation
Now, we have learned two ways to describe LTI
systems, i.e., TF and SS models.
TF: SS:
Memorize this!
267
Realization: From TF to SS
Given a rational proper transfer matrix
find matrices (A,B,C,D) s.t.
Rationality: (polynomial)/(polynomial)
Rational Non-rational
Properness: deg(num)<=deg(den)
Proper Non-proper
268
Remarks
For one SS model, there is a unique TF.
For one TF model, there are infinitely many SS!
unique
(A,B,C,D) G(s)
infinitely many Coordinate transformation T
Order difference
269
First step for realization
Always extract D-matrix first!
Ex
270
Controllable canonical form
SISO example
or equivalently
271
Companion matrix
The following form (and its transpose) of a
matrix is called companion matrix (form):
272
SISO example
Ex.1
Ex.2
273
Controllable canonical form
MIMO case
274
MIMO example
Transfer matrix
275
Remark
Controllable canonical realization is always
controllable (but not always observable). Why?
276
Derivation of
controllable canonical form
TF
We have
277
Derivation (contd)
Introduce state variables
Then,
278
Summary
Realization (In Matlab, tf2ss.m)
Controllable canonical form
Next,
Observable canonical form
Diagonal form
279
Homework
3-2. Obtain controllable canonical realizations
for the following transfer matrices. By hand-
calculation, check the correctness of your
answers.
280
Realization II
281
Realization (review)
Given a rational proper transfer matrix
find matrices (A,B,C,D) s.t.
282
Review
In the last lecture, we learned a realization,
called a controllable canonical form.
or equivalently
284
Companion matrix (review)
The following form (and its transpose) of a
matrix is called companion matrix (form):
285
SISO example
Ex.1
Ex.2
286
Observable canonical form
MIMO case
287
MIMO example
Transfer matrix
289
Derivation of
observable canonical form
TF
290
Derivation (contd)
Draw block-diagram
291
Parallel connection of SS models
Formula
292
Series connection of SS models
Formula
293
An example
Series connection
295
Homework
3-3: Obtain realizations in observable canonical
forms for the following transfer matrices.
296
Minimal realization
297
Review
In the last lecture, we considered the realizations
of the following transfer matrix:
298
Todays topic
Realization of the smallest order
How to characterize such realization?
How to obtain such realization?
Some terminologies:
Minimal realization of G(s):
Realization (A,B,C,D) of G that has the smallest
dimension of A-matrix.
McMillan degree of G(s):
The dimension of A of the minimal realization. (This
indicates the complexity of a system G(s).)
299
Why minimal realization?
Easy to
Analyze (understand) the system
Design a controller
Implement a controller
Computationally less demanding in both design
and implementation
Higher reliability
Few parts to go wrong in the hardware
Few bugs to fix in the software
300
Two important facts on minimal
realization
Fact 1: A realization (A,B,C,D) is minimal
if and only if
(A,B) is controllable and (A,C) is observable.
301
Nonminimal realization examples
Ex
302
How to obtain minimal realization
SISO (Single-Input-Single-Output) case
Remove common factors from numerator and
denominator of G(s). Then, realize G in a controllable
(or observable) canonical form.
C.C.F.
303
How to obtain minimal realization
SIMO case (G(s) is a column vector)
Use the controllable canonical form.
304
How to obtain minimal realization
MIMO (Multiple-Input-Multiple-Output) case
1. Realize G(s) in some non-minimal canonical form.
2. Use the Kalman decomposition to remove
uncontrollable/unobservable parts. (minreal.m)
305
Kalman decomposition (review)
Every SS model can be transformed by z=Tx for
some appropriate T into a canonical form:
306
Kalman decomposition (review)
Conceptual figure (Not block diagram!)
System
307
Example
TF
Order 6
308
Example (contd)
Minimal realization
Controllable!
Observable!
Order 3 Minimal!
309
Example (contd)
After realization, check
the correctness by
recovering the original (1,1)
transfer matrix.
(2,1)
(1,2)
(2,2)
310
Summary
Minimal realization
Realization for DT systems is exactly the same
as that for CT systems.
311
3-5: By both hand-calculation and Matlab, obtain
minimal realization for:
312
Appendix: Proof of Fact 1
only if-part : If (A,B) is not controllable or (A,C) is not
observable, due to Kalman decomposition theorem,
there is a realization with less state dimension.
if-part : Suppose (A,B,C) controllable & observable.
Contradiction!
313
Appendix: Proof of Fact 2
Given a rational proper transfer matrix, assume two
minimal realizations:
Define T as (by )
Then, due to
314
Rank computation
Rank of a matrix can be computed by Singular
Value Decomposition (SVD): For any real matrix
M (m-by-n), we can decompose M as
Singular values
315
State feedback
316
Controller design process (review)
Disturbance
Reference Input Output
Controller Actuator Plant
Sensor
4. Implemenation 1. Modeling
2. Analysis
3. Design
317
Design of control systems
Objective: Find a feedback control so that
The feedback system is internally stable, and
The plant output y(t) follows as closely as possible to
the reference r(t).
? plant
318
Example: Cruise control
http://www.engin.umich.edu/group/ctm/examples/
Modeling Force
(input)
friction
Speed
(output)
319
Cruise control (contd)
Specifications: For initial velocity 10m/s,
r(t)=0
Rise time < 5sec 10
Velocity (m/s)
6
Open-loop system 4
Pole: -0.05
2
Takes too much time.
0
0 20 40 60 80 100
Time (sec)
321
State feedback (contd)
Block diagram Plant
322
Assumptions
D=0
Just for simplicity
This assumption holds in most practical problems.
State vector x is available.
This assumption does not hold in many applications,
and will be removed later.
State estimator (observer and Kalman filter)
Reference r
Regulation: r=0 (will be considered during this week.)
Tracking or servo: nonzero r (will be considered later.)
323
Pole placement theorem
If (A,B) is controllable, the eigenvalues of (A-BK)
can be placed arbitrarily (provided that they are
symmetric with respect to the real axis).
Im
K
Re
325
Direct method (contd)
Desired characteristic polynomial
326
Feedback cruise control
State equation
A B
327
Feedback cruise control (contd)
pole = -0.5 pole = -1 pole = -1.5
10 10 10
Velocity (m/s)
Velocity (m/s)
Velocity (m/s)
5 5 5
0 0 0
0 2 4 6 8 10 0 2 4 6 8 10 0 2 4 6 8 10
Time(sec) Time(sec) Time(sec)
0 0 0
Control input (kN)
-5 -5 -5
328
Exercise
Try the simulation by yourselves! (Matlab code is
on the course homepage.) Change the pole
location, and get a feeling how responses are
affected by the pole location.
For the system below, design a state feedback
u=-Kx so that the closed-loop system has -1 and
-2 as its eigenvalues, by using the direct
method.
329
Summary
State feedback
Pole placement theorem
Direct method for calculating state feedback gain
Suitable only for problems with n=1,2,3.
Cruise control example
As the pole is moved to the left, the convergence
becomes faster, at the price of large control input.
Next, canonical form method for designing the
state feedback gain
330
State feedback design
Canonical form method
331
Review & todays topic
In the last lecture
State feedback
Arbitrary pole placement is possible by a state
feedback u=-Kx if and only if (A,B) is controllable
A direct method to compute the feedback gain K
(applicable to only problems of small sizes)
In todays lecture
A canonical form method for state feedback with a
scalar input (Extension to multi-input cases is
possible, but complicated.)
332
DC motor speed control
http://www.engin.umich.edu/group/ctm/examples/
333
DC motor speed control (contd)
State-space model
334
DC motor speed control (contd)
Specifications: For initial condition
r(t)=0
1
Settling time < 1 sec
Overshoot < 5% 0.8
336
Pole placement theorem (review)
If (A,B) is controllable, the eigenvalues of (A-BK)
can be placed arbitrarily (provided that they are
symmetric w.r.t. the real axis).
Im
K
Re
Step 2: Set
338
Canonical form method (contd)
Step 3: Specify the desired closed-loop poles, i.e.,
the desired characteristic polynomial
339
Idea of canonical form method
If (A,B) is in a controllable canonical form
340
Idea (contd)
If (A,B) is not in a controllable canonical form,
the matrix T in Step 2 transforms A & B as
Note that
341
DC motor speed control: revisited
SS model
Step 4:
342
Speed control (contd)
pole = -5+j, -5-j pole = -5+5j, -5-5j
1 1
Ang. velo. (rad/s)
0 0
-0.5 -0.5
0 0.5 1 1.5 2 2.5 3 0 0.5 1 1.5 2 2.5 3
Time (sec) Time (sec)
10 10
0 0
-10 -10
-20 -20
0 0.5 1 1.5 2 2.5 3 0 0.5 1 1.5 2 2.5 3
Time (sec) Time (sec)
343
Exercise
344
Multi input example
Matrices
BK =
A-BK =
345
Summary
State feedback
Canonical form method to design state feedback gain
place.m in Matlab
DC motor speed control example
As the pole is moved away from the real axis, the
overshoot becomes larger.
Next,
Lyapunov method to design state feedback gain
How to select desired pole locations
Stabilizability
346
State feedback design
Lyapunov equation method
347
Review & todays topics
In the last two lectures
State feedback
Pole placement theorem: Arbitrary pole placement is
possible by a state feedback u=-Kx if and only if (A,B)
is controllable
Direct method & canonical form method to compute
the feedback gain K (place.m)
Todays topics
Lyapunov equation method
Where to place closed-loop poles
Stabilizability
348
State feedback (review)
Block diagram Plant
349
Pole placement theorem (review)
If (A,B) is controllable, the eigenvalues of (A-BK)
can be placed arbitrarily (provided that they are
symmetric w.r.t. the real axis).
Im
K
Re
351
Lyapunov equation method
Step 0: Check if (A,B) is controllable. If it is, go to
Step 1.
Step 1: Select a matrix F with desired eigenvalues
s.t.
Step 2: Select a matrix
Step 3: Solve the Lyapunov equation w.r.t. T
353
Idea of Lyapunov eq. method
From Steps 3 & 5,
If T is nonsingular
Conditions
are necessary for the Lyapunov equation to
have a nonsingular solution. (Proof omitted.)
354
An example
SS model
CL poles
Step 1:
Step 2:
Step 3:
Step 5:
355
DC motor position control
http://www.engin.umich.edu/group/ctm/examples/
356
DC motor position control (contd)
State-space model
357
DC motor position control (contd)
Specifications: For
r(t)=0 1.02
Position (rad)
Zero steady state error
1.01
Open-loop system
y(t) does not
Poles = 0, -59.226 1.005
go to zero!
-1.4545E+6
Not asymptotically stable! 1
0 0.02 0.04 0.06 0.08 0.1
Time (sec)
Position (rad)
Position (rad)
0.5 0.5
0 0
-0.5 -0.5
0 0.02 0.04 0.06 0.08 0.1 0 0.02 0.04 0.06 0.08 0.1
Time (sec) Time (sec)
10 10
0 0
-10 As the pole is moved away from-10the real axis, the overshoot
0 0.02 0.04 0.06 0.08 0.1 0 0.02 0.04 0.06 0.08 0.1
becomes larger
Time (sec)
and the response becomes faster.
Time (sec)
359
Exercise
360
Review & todays topics
In the last two lectures
State feedback
Pole placement theorem: Arbitrary pole placement is
possible by a state feedback u=-Kx if and only if (A,B)
is controllable
Direct method & canonical form method to compute
the feedback gain K (place.m)
Todays topics
Lyapunov equation method
Where to place closed-loop poles
Stabilizability
361
Rules of thumbs
Move the poles for improvement of stability and
performance.
Do not try to move poles farther than necessary.
Control input should not be too large.
363
Review & todays topics
In the last two lectures
State feedback
Pole placement theorem: Arbitrary pole placement is
possible by a state feedback u=-Kx if and only if (A,B)
is controllable
Direct method & canonical form method to compute
the feedback gain K (place.m)
Todays topics
Lyapunov equation method
Where to place closed-loop poles
Stabilizability
364
Stabilizability
Suppose that (A,B) is NOT controllable.
If the uncontrollable part of A-matrix is stable,
then the system is called stabilizable.
Controllable pair
365
Summary
State feedback
Lyapunov eq. method to design state feedback gain
Where to place closed-loop poles (one always needs
trial-and-error)
Rules of thumbs
Butterworth configuration
Stabilizability
Next,
Tracking (servo) control
366
Servo control
367
Review & todays topic
State feedback
Pole placement theorem
Methods to compute state feedback gain K
Regulation problem, i.e., r(t)=0
Today
Servo (tracking) problem, i.e., nonzero r(t)
We also consider disturbance rejection.
For simplicity, we only deal with SISO cases (but
there are similar results for MIMO cases.)
368
DC motor position control
http://www.engin.umich.edu/group/ctm/examples/
Torque
disturbance
369
DC motor position control (contd)
State-space model
370
DC motor position control (contd)
Specifications: For zero initial condition
r(t)=1 rad
Settling time < 40 ms
Overshoot < 16%
Zero steady state error for
Step reference
Step disturbance
Open-loop system
Poles = 0, -59.226, -1.4545E+6
Not asymptotically stable!
371
DC motor position control (contd)
Standard state feedback structure does not
work, as shown in the next slide.
Plant
372
DC motor position control (contd)
Closed-loop poles:
Response to step r(t) Response to step w(t)
800 35
700 30
600
25
Position (rad)
Position (rad)
500
20
400
15
300
10
200
5
100
0 0
0 0.01 0.02 0.03 0.04 0.05 0 0.01 0.02 0.03 0.04 0.05
Time (sec) Time (sec)
Not satisfactory!
373
Robust tracking
Given
design a feedback control law s.t.
feedback system is internally stable
y(t) will track asymptotically any step r(t) even
with small plant parameter variations
with step disturbance w(t) of unknown magnitude
374
State feedback with an integrator
Block diagram
Plant
375
Closed-loop system
SS model
Closed-loop A-matrix
8
0.8
Position (rad)
Position (rad)
6
0.6
0.4 4
0.2
2
0
0
0 0.01 0.02 0.03 0.04 0.05 0 0.01 0.02 0.03 0.04 0.05
Time (sec)
Satisfactory! Time (sec)
377
Why no steady state error to any
step disturbance?
Transfer function from w to y:
380
Transfer function derivation (contd)
Total system
381
Internal model principle
For robust tracking (y tracks r even with w &
plant perturbation), the controller must contain
the dynamics of the exogenous system.
Controller Plant
382
Summary
Servo control
State feedback with an integrator
Reduction to standard pole placement technique
Internal model principle
383