Tutmukp01 Talk

Tutorial on Control Theory
Stefan Simrock, ITER
ICALEPCS, WTC Grenoble, France, Oct. 10-14, 2011 1

Stefan Simrock, “Tutorial on Control Theory” , ICAELEPCS, Grenoble, France, Oct. 10-14, 2011
Outline
• Introduction to feedback control
• Model of dynamic systems
• State space
• Transfer functions
• Stability
• Feedback analysis
• Controller Design
• Matlab / Simulink Tools
• Example beam control

• Example Plasma Control
ICALEPCS, WTC Grenoble, France, Oct. 10-14, 2011 2

1.Control Theory
Objective:
The course on control theory is concerned with the analysis and design of closed loop
control systems.
Analysis:
Closed loop system is given determine characteristics or behavior
Design:
Desired system characteristics or behavior are specified configure or synthesize closed
loop system
Input Variable
Plant
Measurement of
Variable Variable
sensor
Control-system components
3
1.Introduction
Definition:
A closed-loop system is a system in which certain forces (we call these inputs) are
determined, at least in part, by certain responses of the system (we call these outputs).
System System
O O
inputs outputs
Closed loop system
4
1.Introduction
Definitions:
The system for measurement of a variable (or signal) is called a sensor.
A plant of a control system is the part of the system to be controlled.
The compensator (or controller or simply filter) provides satisfactory
characteristics for the total system.
System Manipulated System

input Error variable output
+ Compensator Plant
Sensor
Closed loop control system

Two types of control systems:
A regulator maintains a physical variable at some constant value in the

presence of perturbances.
A servomechanism describes a control system in which a physical variable is
required to follow, or track some desired time function (originally applied in order
to control a mechanical position or motion).
5
1.Introduction
Example 1: RF control system
Goal:
Maintain stable gradient and phase.
Solution:
Feedback for gradient amplitude and phase.
Phase amplitude
controller controller Klystron cavity
~
~
Controller
Gradient
set point
+
-
Phase detector
6
1.Introduction
Model:
Mathematical description of input-output relation of components combined with block
diagram.
Amplitude loop (general form):
Klystron
Reference controller
error cavity
input RF power output
_+
amplifier plant
amplifier
Monitoring
transducer
Gradient detector
7
1.Introduction
RF control model using “transfer functions”
controller Klystron
Control input U(s)

Reference input cavity Output
Error E(s) Y(s)
R(s)
H c (s )
_+
K(s) P(s)
M(s)
Gradient detector
A transfer function of a linear system is defined as the ratio of the Laplace
transform of the output and the Laplace transform of the input with I. C .’s =zero.
Input-Output Relations
Input Output Transfer Function
U(s) Y(s) G(s) = P(s)K(s)
E(s) Y(s) L(s) = G(s)Hc(s)
R(s) Y(s) T(s) = ( 1 + L(s)M(s)) −1 L(s)

8
1.Introduction
Example2: Electrical circuit
i(t) R1
V1(t) R2 V2(t)
C
Differential equations: t
1
R1 i(t) + R2 i(t) + ∫ i (τ ) dτ = V1(t)
t
C0
1
R2 i(t) + ∫ i(τ ) dτ = V2(t)
C0
Laplace Transform:
1
R1 I(s) + R2 I(s) + I(s) = V1(s)
s ⋅C
1
R2 I(s) + I(s) = V2(s)
s ⋅C
Transfer function:
V2(s) R2 ⋅ C ⋅ s + 1
G(s) = =
V1(s) (R1 + R2 )C ⋅ s + 1
Input V1 ,output V2
9
1.Introduction
Example 3: Circuit with operational amplifier
R2 C
i1 R1
-
+
Vi Vo
.
Vi(s) = R1 I 1(s) and Vo(s) = − R2 +
1 
 I 1(s)
 s ⋅C 
V (s) R ⋅C ⋅ s + 1
G(s) = 0 = − 2
Vi(s) R1 ⋅ C ⋅ s
It is convenient to derive a transfer function for a circuit with a single operational
amplifier that contains input and feedback impedance:
Z f (s)
I(s)
Z i(s) -
+
Vi(s) Vo(s)
.
Vo(s) Z f (s)
Vi(s) = Z i(s) I(s) and Vo(s) = − Z f (s) I(s) G(s) = =−
Vi(s) Z i(s)
10
Model of Dynamic System
We will study the following dynamic system:
Parameters:
k : spring constant
γ k γ : damping constant
u(t) : force
Quantity of interest:
m=1 y(t) y(t) : displacement from equilibrium
u(t)
Differential equation: Newton’s third law (m = 1)
&y&(t ) = ∑ Fext = − k y (t ) − γ y& (t ) + u (t )

&y&(t ) + γ y& (t ) + k y (t ) = u (t )
y (0 ) = y0 , y& (0 ) = y& 0
2
Equation is linear (i.e. no y& like terms).
∂ ∂
Ordinary (as opposed to partial e.g. = f (x,t ) = 0 )
∂x ∂t
All coefficients constant: k (t ) = κ ,γ(t ) = γ for all t
11
Model of Dynamic System
Stop calculating, let’s paint!!!
Picture to visualize differential equation
1.Express highest order term (put it to one side)
&y&(t ) = − k y (t ) − γ y& (t ) + u (t )
2.Putt adder in front

u (t ) &y&(t )
+
− γ y& (t )
− k y(t )
3.Synthesize all other terms using integrators!

Block diagram
u (t ) + &y&(t ) y& (t ) y (t )
- ∫ ∫
-
k 12
2.1 Linear Ordinary Differential Equation (LODE)
General form of LODE:
y (n ) (t ) + an −1 y (n −1) (t ) + ... + a1 y& (t ) + a0 y (t ) = bm u (m ) (t ) + ... + b1 u& (t ) + b0 u (t )
m ,n Positive integers, m ≤ n ; coefficients a0 ,a1 ,...,a n −1 , b0 ,...,bm real numbers.
Mathematical solution: hopefully you know it
Solution of LODE: y(t ) = yh (t ) + yp (t ),

Sum of homogeneous solution y h (t ) (natural response) solving
y (n ) (t ) + a n − 1 y (n − 1 ) (t ) + ... + a 1 y& (t ) + a 0 y (t ) = 0
And particular solution y p (t ) .
How to get natural response yh (t) ? Characteristic polynomial

χ ( λ ) = λ n + an −1 λ n −1 + a1 λ + a0 = 0
(λ − λ1 ) r ⋅ (λ − λr +1 ) ⋅ ... ⋅ (λ − λn ) = 0
y h (t ) = (c1 + c2 t + ... + cr t r −1 ) e λ t + cr +1e λ
1
+ ... + cn e λn t
r +1 t
Determination of y p (t ) relatively simple, if input u (t ) yields only a finite number of

independent derivatives. E.g.: u (t ) ≅ e , β r t .
ξt r
13
2.1Linear Ordinary Differential Equation (LODE)
Most important for control system/feedback design:
y (n) (t ) + a n −1 y (n −1 ) (t ) + ... + a1 y& (t ) + a0 y(t) = bm u ( m ) (t ) + ... + b1 u& (t ) + b0 u (t )
In general: given any linear time invariant system described by LODE can be
realized/simulated/easily visualized in a block diagram (n = 2 , m = 2 )
Control-canonical form
b2
b1
+ +
u (t ) + x2 x1 y (t )
-
∫ ∫ b0 +
-
a1
a0
Very useful to visualize interaction between variables!
What are x1 and x2 ????
More explanation later, for now: please simply accept it!
14
2.2 State Space Equation
Any system which can be presented by LODE can be represented in State space
form (matrix differential equation).
What do we have to do ???
Let’s go back to our first example (Newton’s law):
&y&(t ) + γ y& (t ) + k y (t ) = u (t )
1. STEP: Deduce set off first order differential equation in variables
x j (t ) (so-called states of system)
x1 (t ) ≅ Position : y (t )
x2 (t ) ≅ Velocity : y& (t ) :
x&1 (t ) = y& (t ) = x2 (t )
x&2 (t ) = &y&(t ) = −k y(t ) − γ y& (t ) + u(t )

= −k x1 (t ) − γ x2 (t ) + u(t )
One LODE of order n transformed into n LODEs of order 1
15
2. STEP:
Put everything together in a matrix differential equation:
 x&1 (t ) 0 1  x1 (t ) 0 
=
 x& (t )  -k - γ   x (t ) + 1  u (t )
 2     2   
x& (t ) = A x (t ) + B u (t )
State equation
 x (t )
y (t ) = [1 0 ]  1 
 x 2 (t )
y (t ) = C x(t ) + D u (t )
Measurement equation
Definition:
The system state x of a system at any time t0 is the “amount of information” that,
together with all inputs for t ≥ t0 , uniquely determines the behaviour of the system
for all t ≥ t0 .
16
The linear time-invariant (LTI) analog system is described via
Standard form of the State Space Equation
x& (t ) = A x(t ) + B u (t ) State equation

y (t ) = C x(t ) + D u (t ) State equation
 x1 (t )
Where x& (t ) is the time derivative of the vector x(t ) =  ⋅ ⋅ ⋅  . And starting conditions x (t0 )
 xn (t )
System completely described by state space matrixes A, B, C, D ( in the most cases D =0 ).
Declaration of variables
Variable Dimension Name
X (t ) n×1 state vector
n×n system matrix
B n× r input matrix
u (t ) r ×1 input vector
y (t ) p ×1 output vector
C p×n output matrix
D p×r matrix representing direct coupling
between input and output
17
Why all this work with state space equation? Why bother with?
BECAUSE: Given any system of the LODE form
y (n ) (t ) + a n −1 y (n −1 ) (t ) + ... + a1 y& (t ) + a0 y (t ) = bm u (m ) (t ) + ... + b1 u& (t ) + b0 u (t )

Can be represented as
x& (t ) = A x (t ) + B u (t )
y (t ) = C x (t ) + D u (t )
with e.g. Control-Canonical Form (case n = 3 ,m = 3 ):

 0 1 0 0 
A =  0 0 1  , B = 0  , C = [b0 b1 b2 ] , D = b3
 − a0 − a1 − a2  1 
or Observer-Canonical Form:
0 0 − a0  b0 
A = 1 0 − a1  ,B = b1  ,C = [0 0 1] ,D = b3
0 1 − a2  b2 
Notation is very compact, But: not unique!!!
Computers love state space equation! (Trust us!)
Modern control (1960-now) uses state space equation.
General (vector) block diagram for easy visualization.
18
Block diagrams:
Control-canonical Form:
+
b2 b1
+ x2 x1 + y (t )
- ∫ ∫ b0
u (t ) -
a1 a0
+
Observer-Canonical Form:
u (t )
b0 b1 b2
+ +
x1 + x2 + + y(t)
∫ ∫
- -
a0 a1
19
Now: Solution of State Space Equation in the time domain. Out of the hat…et voila:
t
x(t ) = Φ (t ) x(0 ) + ∫ Φ (τ ) B u (t − τ ) dτ
0
Natural Response + Particular Solution

y (t ) = C x(t ) + D u (t )
t
= C Φ(t ) x(0 ) + C ∫ Φ(τ ) B u (t − τ ) dτ + D u (t )
0
With the state transition matrix

A2 2 A3 3
Φ (t ) = I + At + t + t + ... = e A t
2! 3!
Exponential series in the matrix A (time evolution operator) properties of Φ (t ) (state transition matrix).
dΦ (t )
1. = A Φ (t )
dt
2 .Φ (0 ) = I
3.Φ (t1 + t 2 ) = Φ (t1 ) ⋅ Φ (t 2 )
4 .Φ −1 (t ) = Φ (− t )
Example:
0 1  0 0  1 t 
A=  ⇒ A 2
=   , Φ (t ) = I + At =   = eAt
0 0  0 0  0 1 
Matrix A is a nilpotent matrix.
20
2.3 Examples
Example:
It is given the following differential equation:
d2 d
y (t ) + 4 y (t ) + 3 y (t ) = 2 u (t )
dt 2 dt
State equations of differential equation:
Let x1 (t ) = y (t ) and x2 (t ) = y& (t ) . It is:
x&1 (t ) = y& (t ) = x2 (t )
x&2 (t ) + 4 x2 (t ) + 3 x1 (t ) = 2 u(t )
x&2 (t ) = −3 x1 (t ) − 4 x2 (t ) + 2 u(t )
Write the state equations in matrix form:
 x1 (t )
Define system state x(t ) =   . Then it follows:
x
 2  (t )
0 1  0 
x& (t ) =   x (t ) + 2  u (t )
 - 3 − 4   
y (t ) = [1 0 ] x(t )
21
2.3 Cavity Model
Ig circulator Conductor Ib
Conductor
. Zo . Zo Ro Ib
Last
I g ~ Generator Zo L ~ Beam-Current
C
Resonator
. . .
Coupler 1:m
I 'g Ir Ib
Equivalent circuit:
Generator
I 'g ~ Rext L C ~ Ib
Ro
Resonator
1 & 1
C ⋅ U&& + ⋅ U + ⋅U = I&g′ + I&b
RL L
1 ω
ω1 / 2: = = 0
2 RL C 2QL
2 
U&& + 2ω1 / 2 ⋅ U& + ω02 ⋅ U = 2 RL ω1 / 2 ⋅  I&g + I&b 
m  22
2.3 Cavity Model
Only envelope of rf (real and imaginary part) is of interest:
U (t ) = (U r (t ) + i U i (t )) ⋅ exp (i ωHF t)
I g (t ) = (I gr (t ) + i I gi (t ))⋅ exp(i ωHF t )
I b (t ) = (I bω r (t ) + i I bω i (t )) ⋅ exp(i ωHF t ) = 2(I b0 r (t ) + i I b0 i (t )) ⋅ exp(i ωHF t )
Neglect small terms in derivatives for U and I
U&&r + iU&&i (t ) << ωHF

2
(Ur (t ) + iUi (t ))
2ω (U& + iU& (t )) << ω2 (U (t)+ iU (t))
1/ 2 r r HF r i
t2 t2
∫ (I& (t ) + iI& (t )) dt << ∫ ω (I (t ) + iI (t )) dt

t1
r i
t1
HF r i
Envelope equations for real and imaginary component.

r 1 
U& r (t ) + ω1/ 2 ⋅Ur + ∆ω⋅Ui = ωHF   ⋅  I gr + Ib0r 
Q  m 
r 1 
U& i (t ) + ω1/ 2 ⋅Ui − ∆ω⋅Ur = ωHF   ⋅  I gi + Ib0i 
Q  m  23
2.3 Cavity Model
Matrix equations:
1 
U& r (t ) − ω1 / 2 − ∆ω  U r (t )  r  1
I
0   m gr
(t ) + I b0 r (t )
& =  ⋅ U (t )  + ωHF   ⋅ 0 ⋅
1  1

U i (t )   ∆ ∆ω − ω1/ 2   i  Q  I gi (t ) + I b0 i (t ) 
 m 
With system Matrices:
− ω1 / 2 − ∆ω   r  1 0 
A=  B = ωHF   ⋅  
 ∆ω − ω1/ 2    
Q 0 1
1 
r U r (t ) r
I
 m gr (t ) + I b0 r (t )
x (t ) =   u (t ) =  
U i (t )  1
 I (t ) + I (t ) 
 m gi b0 i

General Form:
r r r
x& (t ) = A ⋅ x (t ) + B ⋅ u (t )
24
2.3 Cavity Model
Solution: t
r r r
x (t ) = Φ (t ) ⋅ x (0 ) + ∫ Φ (t − t' ) ⋅ B ⋅ u (t ′) dt ′
0
− ω1 / 2 t cos ( ∆ωt ) − sin ( ∆ωt )

Φ(t) = e  sin ( ∆ωt ) 
 cos ( ∆ωt )
Special Case:
1 
r
I
 m gr (t ) + I b0 r (t ) I r 
u (t ) =   = : 
 I gi (t ) + I b0 i (t )   I i 
1
m 
r
ωHF  
U r (t )  Q  ⋅ ω1 / 2 − ∆ω  ⋅ 1 − cos( ∆ωt ) − sin( ∆ωt )  e −ω1 / 2 t  ⋅  I r 
=
U (t )  ω 2 + ∆ω 2  ∆ω    sin( ∆ωt )    
 i  1/ 2  ω1/ 2    cos ( ∆ωt )  Ii 
25
2.3 Cavity Model
Integrator
+ 1 1 Scope
2
s s
Step - - Integrator 1
Gain
Gain 1 Harmonic oscillator
4
Gain 2
Harmonic oscillator
x′ = Ax + Bu Scope
y = Cx + Du
Step
State space
Step
cavity x′ = Ax + Bu
Load Data
y = Cx + Du
State space Scope
Step 26
2.3 Cavity Model
Integrator
+ 1
k s
Step - -
Gain
Gain 2
w12
dw
Gain 4
Load Data dw
Scope
Gain 5
+ 1
s
Step 1 - + Integrator 1
Gain 3
w12
27
2.4 Masons Rule
Mason’s Rule is a simple formula for reducing block diagrams. Works on continuous and
discrete. In its most general form it is messy, but For special case when all path touch
H(s) =
∑ ( forward path gains )
1-∑ (loop path gains )
Two path are said to touch if they have a component in common, e.g. an adder.
10 11
H5
U 1
H1
2 3
H2 4 5
H3 6 Y
9 8 7
H4
Forward path: F1: 1 - 10 - 11 - 5 – 6
F2 : 1 - 2 - 3 - 4 - 5 – 6 G( f1 ) = H5 H3
G( f2 ) = H1 H2 H3
Loop path : I1: 3 - 4 - 5 - 8 – 9
G(I1 ) = H2 H4
I2 : 5 - 6 - 7
G(I2 ) = H3
Check: all path touch (contain adder between 4 and 5)
G( f1 ) + G( f 2 ) H 5 H 3 + H1 H 2 H 3 H 3 (H 5 + H1 H 2 )
By Mason’s rule: H= = =
1 − G(l1 ) − G(l2 ) 1 − H 2 H 4 − H 3 1 − H 2 H4 − H3 28
2.5 Transfer Function G (s)
Continuous-time state space model
x& (t ) = A x(t ) + B u (t ) State equation
y (t ) = C x(t ) + D u (t ) Measurement equation
Transfer function describes input-output relation of system.
U(s) System Y(s)
s X (s) − x(0) = A X (s ) + B U(s)
X (s ) = (sI − A) x (0 ) + (sI − A) B U (s )
−1 −1
= φ(s ) x (0 ) + φ(s ) B U (s )
Y (s ) = C X (s ) + D U (s )
= C[ (sI − A) ]x(0 ) + [c(sI − A) B + D]U (s )
−1 −1
= C φ(s ) x(0 ) + C φ(s ) B U (s ) + D U (s )
Transfer function G (s ) ( pxr ) (case: x(0)=0):
G (s ) = C (sI − A) B + D = C φ(s ) B + D
−1
29
2.5 Transfer Function
Transfer function of TESLA cavity including 8/9-pi mode
H cont (s ) ≈ H cav (s ) = H π (s ) + H 8 (s )
π
9
(ω1 / 2 )π  s + (ω1 / 2 )π − ∆ωπ 

π − mode H π (s ) = 
2 

∆ωπ + (s + (ω1 / 2 )π )  − ∆ωπ
2
s + (ω1 / 2 )π 
(ω1 / 2 )8 π  s + (ω1 / 2 )8 π − ∆ω 8 
8  9
9
π 
2  
9
π − mode H 8 (s) = −
∆ ω + ( )
+  s + (ω1 / 2 )8 π  
9 π s ω 8
π 
9
∆ω 82 8
π
1/ 2
9

9
π  9 
9
30
2.5 Transfer Function of a Closed Loop System
R (s ) E (s ) U (s ) Y(s)
H c (s ) G(s )
-
M (s)
We can deduce for the output of the system.
Y (s) = G(s) U (s) = G(s) Hc (s) E(s)

= G(s) Hc (s) [R(s) − M (s ) Y (s)]
= L(s) R(s) − L(s) M (s) Y (s)
With L(s ) the transfer function of the open loop system (controller plus plant).
(I + L(s ) M (s )) Y (s ) = L(s ) R(s )

Y (s ) = (I + L(s ) M (s )) L(s ) R(s )
−1
= T (s ) R(s )
T (s ) is called : Reference Transfer Function

31
2.5 Sensitivity
The ratio of change in Transferfunction T(s) by the parameter b can be defined as:
System characteristics change with system parameter variations
∆T(s) b The sensitivity function is defined as:

S=
T(s) ∆b ∆T(s) b ∂T(s) b
SbT = lim =
∆b →0 ∆b T(s) ∂b T(s)
Or in General sensitivity function of a characteristics W with respect to the parameter b:
∂W b
SbW =
∂b W
K
Example: plant with proportional feedback given by Gc (s ) = K p G p (s ) =
s + 0 .1
K p G p (s ) |S|
Reference transfer function T(s): T (s ) =
1 + K p G p (s )H k
Kp=10
− K p G p ( jω )H k − 0.25K p
S HT ( jω ) = = Kp=1
1 + K p G p ( jω )H k 0.1 + 0.25 K p + jω
Increase of H results in decrease of T
System cant be insensitive to both H,T omega
32
2.5 Disturbance Rejection
Disturbances are system influences we do not control and want to minimize its
impact on the system.
G c (s) ⋅ G p (s) G d (s)
C (s ) = R(s) + D(s)
1 + G c (s) ⋅ G p (s) ⋅ H(s) 1 + G c (s) ⋅ G p (s) ⋅ H(s)
= T(s) ⋅ R(s) + T d (s) ⋅ D(s)
D(s)
Gd (s)
To Reject disturbances, make Td ⋅ D(s ) small!
R(s) Gc (s) Gp (s) C(s)
Using frequency response approach to

Plant
investigate disturbance rejection H(s)
In general Td ( jω ) cant be small for all -ω
Design Td ( jω ) small for significant
portion of system bandwidth
Reduce the Gain Gd ( jω ) between dist. Input and output

Increase the loop gain Gc( jω )Gp( jω ) without increasing the gain Gd ( jω ).Usually
accomplished by the compensator Gc( jω ) choice
Reduce the disturbance magnitude d (t ) should always be attempted if reasonable
Use feed forward compensation, if disturbance can be measured.
33
2.6 Stability
Now what do we know:
The impulse response tells us everything about the system response to any arbitrary
input signal u(t) .
what we have not learnt:

If we know the transfer function G(s), how can we deduce the systems behavior?
What can we say e.g. about the system stability?
Definition:
A linear time invariant system is called to be BIBO stable (Bounded-input-bounded-output)

For all bounded inputs u (t ) ≤ M 1 (for all t) exists a boundary for the output signal M 2 ,
So that y (t ) ≤ M 2 . (for all t) with M 1 and M 2 , positive real numbers.
Input never exceeds M 1 and output never exceeds M 2 , then we have BIBO
stability!
Note: it has to be valid for ALL bounded input signals!
34
2.6 Stability
1
Example: Y (s ) = G (s ) U (s ), integrator G (s ) =
s
1.Case
u (t ) = δ (t ), U (s ) = 1
1
y (t ) = L−1 [Y (s )] = L−1   = 1
s
The bounded input signal causes a bounded output signal.
2.Case
1
u (t ) = 1 , U (s ) =
s
 1 
y (t ) = L− 1 [Y (s )] = L− 1  2  = t
s 
BIBO-stability has to be shown/proved for any input. Is is not sufficient to show

its validity for a single input signal!
35
2.6 Stability
Condition for BIBO stability:
We start from the input-output relation
Y (s ) = G(s ) U (s )
By means of the convolution theorem we get
t t ∞
y (t ) = ∫ g (τ ) u (t − τ ) dτ
0
≤ ∫ g (τ ) u (t − τ )
0
dτ ≤ M 1 ∫
0
g (τ ) dτ ≤ M 2
Therefore it follows immediately:
If the impulse response is absolutely integrable

∞
∫
0
g (t ) dt < ∞
Then the system is BIBO-stable.
36
2.7 Poles and Zeroes
Can stability be determined if we know the TF of a system?
[sI − A] adj
G (s ) = C Φ (s ) B + D = C B+D
χ (s )
Coefficients of Transfer function G(s) are rational functions in the complex variables
∏ mk =1 (s − zk ) N ij (s )
g ij (s ) = α ⋅ n =
∏ l =1 (s − pl ) Dij (s )
z k Zeroes.pl Poles, α real constant, and it is m ≤ n (we assume common factors have
already been canceled!)
What do we know about the zeros and the poles?
Since numerator N (s ) and denominator D(s ) are polynomials with real coefficients,
Poles and zeroes must be real numbers or must arise as complex conjugated pairs!
37
Stability directly from state-space
Re call : H (s ) = C (sI − A) B + D
−1
Assuming D=0 (D could change zeros but not poles)

Cadj (sI − A)B b(s )
H (s ) = =
det (sI − A) a (s )
Assuming there are no common factors between the poly Cadj (sI − A)B and det (sI − A)
i.e. no pole-zero cancellations (usually true, system called “ minimal” ) then we can identify
and b(s) = Cadj (sI − A) B
a(s) = det (sI − A)
i.e. poles are root of det (sI − A)

th
Let λ i be the i eigenvalue of A
if Re {λi } ≤ 0 forall i => System stable

So with computer, with eigenvalue solver, can determine system stability directly from coupling matrix A.
38
2.8 Stability Criteria
A system is BIBO stable if, for every bounded input, the output remains bounded with
Increasing time.
For a LTI system, this definition requires that all poles of the closed-loop transfer-function
(all roots of the system characteristic equation) lie in the left half of the complex plane.
Several methods are available for stability analysis:
1.Routh Hurwitz criterion
2.Calculation of exact locations of roots

a. Root locus technique
b. Nyquist criterion
c. Bode plot
3.Simulation (only general procedures for nonlinear systems)
While the first criterion proofs whether a feedback system is stable or unstable,
the second Method also provides information about the setting time (damping term).
39
Pole locations tell us about impulse response i.e. also stability:
Medium oscillation Im (s) = ω Fast oscillation

Medium decay No growth
X
X X
Medium oscillation
Medium growth
X X X
Re (s) = σ
X X
No Oscillation No oscillation No oscillation
Fast Decay No growth Fast growth
40
Furthermore: Keep in mind the following picture and facts!
Complex pole pair: Oscillation with growth or decay.
Real pole: exponential growth or decay.
Poles are the Eigenvalues of the matrix A.
Position of zeros goes into the size of c j ....
In general a complex root must have a corresponding conjugate root ( N(s), D(S) polynomials
with real coefficients.
41
2.8 Bode Diagram
dB
ω1
ω2 ω
Gm Gain Margin
00 ω2 ω1 ω
− 900
0
Phase Margin φm
− 180
The closed loop is stable if the phase of the unity crossover frequency of the OPEN LOOP
Is larger than-180 degrees.
42
2.8 Root Locus Analysis
Definition: A root locus of a system is a plot of the roots of the system characteristic
equation (the poles of the closed-loop transfer function) while some parameter of the
system (usually the feedback gain) is varied.
X X X
p3 p2 p1
K
K H (s ) =
(s − p1 ) (s − p 2 ) (s − p 3 )
R (s ) + Y(s)
K H(s)
-
K H (s )
G CL (s ) = roots at 1+ KH (s ) = 0 .
1 + K H (s )
How do we move the poles by varying the constant gain K? 43

2.8 Root Locus Examples
1 1
s − p1 (s − p1 )(s − p2 )
X X X
p1 p2 p1
(a) (b)
s − z1 s − z1
(s − p1 )(s − p2 ) (s − p1 )(s − p2 )
X O X O X X
p2 z1 p1 z1 p2 p1
(d)
(c)
44
2.8 Root Locus Examples (Cnt’d)
p2
1 X 1
(s − p1 )(s − p2 )(s − p3 ) (s − p1 )(s − p2 )(s − p3 )
X X X X
p3 p2 p1 p1
X
p3
(e) (f)
p2 s − z1
1
X
(s − p1 )(s − p2 )(s − p3 ) (s − p1 )(s − p2 )(s − p3 )
X X X O X
p1 p3 p2 z1 p1
X
p3
(h)
(g)
45
3.Feedback
The idea:
Suppose we have a system or “plant”
“open loop”
plant
We want to improve some aspect of plant’s performance by observing the output

and applying a appropriate “correction” signal. This is feedback
r
plant
“closed loop”
Ufeedback
Question: What should this be?
46
3.Feedback
Open loop gain:
U G(s)
Y
−1
 
G O.L
(s ) = G(s ) =  u 
 y
Closed-loop gain:
U G(s)
Y
U fb “closed loop”
H (s )
G(s)
G C.L(s) =
1 + G(s) H(s)
Pr oof: y = G (u − u fb )
= G u − G u fb ⇒ y +G Hy = G u
y G
= G u − G Hy ⇒ =
u (1 + G H )
47
3.1 Feedback-Example 1
Consider S.H.O with feedback proportional to x i.e.:
Where &x& + γ x& + ωn2 x = u + u fb
u fb (t ) = −α x (t )
U &x& 1 x& 1 x y
+ - S s
- -
ω n2
α
Then &x& + γ x& + ω n2 x = u − α x

==> &x& + γ x& + (ωn2 + α ) x = u
Same as before, except that new “natural” frequency ωn2 + α

48
1
Now the closed loop T.F. is: G C.L. (s ) =
(
s 2 + γ s + ωn2 + α )
G O.L. (iω)
dB
G C.L. (iω)
1
ω)
log(ω
ωn2 log ωn log ωn2 + α
1
ω n2 + α
DC response: s=0
So the effect of the proportional feedback in this case is to increase the bandwidth
of the system
(and reduce gain slightly, but this can easily be compensated by adding a constant gain in front…)
49
In S.H.O. suppose we use integral feedback:
t
u fb (t ) = −α ∫ x(τ ) dτ
0
t
i.e &x& + γ x& + ωn2 x = u − α ∫ x(τ ) dτ
0
&x& x& x y
U 1 1
+ - S S
- -
γ
ω n2
α
s
Differentiating once more yields: &x&& + γ &x& + ωn2 x& + α x = u&
No longer just simple S.H.O., add another state

50
1
s + γs + ω n2
2
G (s ) =
C.L.
 α  1 
1 +   2 
2
(
 s  s + γs + ω n + α ) 
s
=
( )
s s + γs + ω n2 + α
2
Observe that
1. G C.L. (0 = 0 )
2. For large s (and hence for large ω )
dB 1
G C.L. (s ) ≈ ≈ G O.L. (s )
G O.L. (iω)
(2 2
s + γ s + ωn )
1
ω)
log(ω
ω n2
G C.L. (iω)
So integral feedback has killed DC gain
i.e system rejects constant disturbances
51
Suppose S.H.O now apply differential feedback i.e.
u fb (t ) = −α x&(t )
&x& 1 x& 1 x
+ - S S
- -
γ
ω n2
x
αS
α x&
Now have
&x& + (γ + α ) x& + ωn2 x = u
So effect off differential feedback is to increase damping
52
1
Now G C.L. (s ) =
s 2 + (γ + α ) s + ω n2
dB
G O.L. (iω)
1
ω)
log(ω
ω n2
G C.L. (iω)
So the effect of differential feedback here is to “flatten the resonance” i.e. damping is increased.
Note: Differentiators can never be built exactly, only approximately.
53
3.1 PID controller
(1) The latter 3 examples of feedback can all be combined to form a
P.I.D. controller (prop.-integral-diff).
u x= y
+ S.H.O
-
P.I.D controller
K p + K D s + K l /s
ufb =up +ud +u l
(2) In example above S.H.O. was a very simple system and it was clear what
physical interpretation of P. or I. or D. did. But for large complex systems not
obvious
Require arbitrary “ tweaking”

That’s what we’re trying to avoid
54
3.1 PID controller
For example, if you are so smart let’s see you do this with your P.I.D. controller:
6th order system

G 3 resonant poles
3 complex pairs
6 poles
Damp this mode, but leave the other two modes undamped, just as they are.
This could turn out to be a tweaking nightmare that’ll get you nowhere fast!
We’ll see how this problem can be solved easily.
55
3.2 Full State Control
Suppose we have system
x& (t ) = A x (t ) + B u (t )
y (t ) = C x (t )
Since the state vector x(t) contains all current information about the system the
most general feedback makes use of all the state info.
u fb = − k1 x1 − ..... − k n xn
= -k x
Where k = [k1......k n ] (row matrix)
Where example: In S.H.O. examples
[ ]
Proportional fbk : up = −k p x = − k p 0
Differential fbk : ud = −kd x& = − [0 kd ]

56
Theorem: If there are no poles cancellations in
b (s )
G O.L. (s ) = = C (sI − A ) B
−1
a (s )
Then can move eigen values of A − BK anywhere we want using full state feedback.
Proof:
Given any system as L.O.D.E. or state space it can be written as:
AO.L. B
 x1   0 1 ... 0   x1  0 
 ...   0 ... ... .. .   ...  0 
 =   +  u
 ...   0 ... ... 1   ...   ...
       
 x n   -a 0 ... .. . -a n- 1   xn  1 
 x1 
 ... 
y = [b0 ... ... bn- 1 ]  
 ... 
 
 xn 
Where
bn − 1 s n − 1 + ... + b0
= C (sI − A )
O.L. −1
G B= n
s + a n − 1 s n − 1 + ... + a0
57
O.L.
i.e. first row of A Gives the coefficients of the denominator
( )
a O.L.(s ) = det sI − AO.L. = s n + an−1s n−1 + ... + a0
Now
AC.L. = AO.L. − BK
 0 1 ... 0  0 
 0 ... ... ...  0 
=  −   [k ... ... k ]
 0 ... ... 1  ... 0 n-1
   
-a
 0 ... ... -a n-1  1 
 0 1 ... 0 
 0 ... ... ... 
= 
 0 ... ... 1 
 
-(a
 0 0+ k ) ... ... -(an-1 + k )
n −1 
So closed loop denominator
(
a C.L. (s ) = det sI − A C.L. )
= s n + (a 0 + k 0 )s n − 1 + ... + (a n − 1 + k n − 1 )
Using u = − Kx have direct control over every closed-loop denominator coefficient
can place root anywhere we want in s-plane.
58
Example: Detailed block diagram of S.H.O with full-scale feedback
u &x& 1 x& 1 x y
+
- - S S
-
γ
ω n2
k2 x&
+
x
k1
Of course this assumes we have access to the x& state, which we actually
Don’t in practice.
However, let’s ignore that “ minor” practical detail for now.

( Kalman filter will show us how to get x& from x ).
59
With full state feedback have (assume D=0)
u x& 1 x y
+ B + C
s
-
u fb = −kx A
K
So x& = A x + B[u + u fb ]
= Ax + Bu + BKufb
x& = ( A − BK ) x + B u
u fb = −Kx
y = Cx
With full state feedback, get new closed loop matrix
(
AC.L. = AO.L. − BK )
Now all stability info is now given by the eigen values of new A matrix
60
3.3 Controllability and Observability
The linear time-invariant system
x& = Ax + Bu
y = Cx
Is said to be controllable if it is possible to find some input u(t) that will transfer the
initial state x(0) to the origin of state-space, x(t0 ) = 0,with t0 finite
The solution of the state equation is:

t
x (t ) = φ (t )x (0 ) + ∫ φ (τ )B u (t − τ ) d τ
0
For the system to be controllable, a function u(t) must exist that satisfies the equation:
t0
0 = φ (t 0 )x (0 ) + ∫ φ (τ )Bu (t 0 − τ ) dτ
0
With t 0 finite. It can be shown that this condition is satisfied if the controllability matrix
CM = [B AB A2 B ... An-1 B]
Has inverse. This is equivalent to the matrix C M having full rank (rank n for an n- th
order differential equation).
61
3.3 Controllability and Observability
Observable:
The linear time-invariant system is said to be observable if the initial conditions x(0)
Can be determined from the output function y(t), 0 ≤ t ≤ t1 where t1 is finite. With
t
y (t ) = Cx = C φ (t )x 0 + C ∫ φ (τ )Bu (t − τ ) d τ
0
The system is observable if this equation can be solved for x(0). It can be shown that
the system is observable if the matrix:
 C 
 CA 
OM =  
 ... 
 n-1 
CA 
Has inverse. This is equivalent to the matrix O M having full rank (rank n for an n-th
Order differential equation).
62
4.Discrete Systems
Where do discrete systems arise?
Typical control engineering example:
Digitized
Continuous system sample
u(k) uc (t ) yc (t ) y(k)
DAC h(t ) ADC
t t t t
“Digitized” “Zero-order-hold” “continuous” “Digitized”
Computer controller
Assume the DAC+ADC are clocked at sampling period T.
Continued… 63
4. Discrete Systems
Then u(t) is given by:
u (k ) ≡ u c (t ); kT ≤ t < (k + 1 )T
y (k ) ≡ y c (kT ); k = 0 ,1,2 ,...
Suppose: time continuous system is given by state-Space
x& c (t ) = A x c (t ) + B u c (t ); x c (0 ) = x0
y c (t ) = C x c (t ) + D u c (t )
Can we obtain direct relationship between u(k) and y(k)? i.e. want
Equivalent discrete system:
u( k ) y( k )
DAC h( t ) ADC
u( k ) y( k )
h( k )
64
4. Discrete Systems
Yes! We can obtain equivalent discrete system.
t
Recall x c (t ) = e x c (0 ) + ∫ e Aτ .Bu c (t − τ) dτ
At
0
T
From this x c (kT + T ) = e AT
x c (kT ) + ∫ e Aτ .Bu c (kT − τ ) d τ
0
Observe that u (kT + T- τ ) = u (kT ) for τ ∈ [ 0 ,T]

i.e. u (kT + T-τ ) is constant u (kT ) over τ ∈ [ 0 ,T]
i.e. can pull out of integral.
 t Aτ 
==> x c (kT + T ) = e x c (kT ) +  ∫ e ⋅ B d τ  u c (kT
At
)
0 
x(k + 1 ) = Ad x(k ) + Bd u (k )
y(k ) = Cd x(k ) + Dd u (k )
x(0 ) = xc (0 )
T
So Ad = e , Bd = ∫ e Aτ .B dτB dd = C, Dd = D
AT
So we have an exact (note: x(k + 1) = x(k ) + x& (k ) T + O(.)) discrete time equivalent to the time
Continuous system at sample times t=kT- no numerical approximation!
65
4.1 Linear Ordinary Difference Equation
A linear ordinary difference equation looks similar to a LODE
y (k + n ) + an-1 y (k + n − 1) + ... + a1 y (k + 1) + a0 y (k ) = bm u (k + m ) + ... + b1 u (k + 1) + b0 u (k )
n ≥ m; Assumes initial values y (n-1), ...., y(1),y (0 ) = 0 .
Z-Transform of the LODE yields (linearity of Z-Transform):
z n Y ( z ) + z n −1an −1 Y (z ) + ... + za1 Y (z ) + a0 Y (z ) = z mbm U ( z ) + ... + zb1 U ( z ) + bo U (z )
It follows the input-output relation:
(z n
) ( )
+ z n −1an −1 + ... + za1 + a0 Y (z ) = z mbm + .... + zb1 + b0 U(z)
z mbm + .... + zb1 + b0
Y (z ) = n U (z )
z + ... + za1 + a0
Y (z ) = G(z ) U (z )
Once again:
if U ( z ) = 1, (u (k ) = δ (k )), then Y (z ) = G (z ).
Transfer Function of system is the Z-Transform of its pulse response!
66
4.1 z-Transform of Discrete State Space Equation
x(k + 1) = Ad x(k ) + Bd u (k )
y(k ) = C x(k ) + D u (k )
Applying z-Transform on first equation:

z ⋅ X (z )-z x(0 ) = Ad X (z ) + Bd U (z )
(zI-Ad ) X (z ) = z x(0 ) + B U (z )
X ( z ) = (zI-Ad ) z x(0 ) + (zI − Ad ) B U (z )
−1 −1
Homogeneous solution
NOW: Particular solution

Y ( z ) = CX (z ) + D U ( z )
( )
= C ( zI-Ad ) z x(0 ) + C ( zI − Ad ) B + D U (z )
−1 −1
If x(0)=0 then we get the input-output relation:
Y (z ) = G (z ) U (z ) with
G ( z ) = C ( zI-A d )− 1 B + D
Exactly like for the continuous systems!!!!!!!
67
4.2 Frequency Domain/z-Transform
For analyzing discrete-time systems:
z-Transform
(analogue to Laplace Transform for time-continuous system)
It converts linear ordinary difference equation into algebraic equations: easier to find
a solution of the system!
It gives the frequency response for free!
z-Transform ==generalized discrete-time Fourier Transform
Given any sequence f (k ) the discrete Fourier transform is

∞
~
F (ω ) = ∑
k = −∞
f (k ) e −iωk
1
with ω = 2 πf, f = the sampling frequency in Hz,
T
T difference / Time between two samples.
∞
In the same spirit: F ( z ) = Z[f (k ) ] = ∑ f (k ) z
k =0
-k
.
With z a complex variable
Note: if f (k ) = 0 for k = -1 ,- 2 ,...... th en F~ (ω ) = F(z = e iω ).
68
4.3 Stability (z-domain)
A discrete LTI system is BIBO stable if
u (k ) < M; ∀k = y (k ) < K; ∀k
Condition for BIBO stability:
k k k ∞
y (k ) = ∑ u (k − i ) h(i ) ≤ ∑ u(k − i ) h(i )
0 0
≤ M ∑ h(i ) ≤ M ∑ h(i )
0 0
∞
∴ ∑ h (i ) < ∞ BIBO stable.
0
For L.O.D.E State space system:
∏ i = 1 (z − z i ) k
H ( z ) = α.
∏ i = 1 (z − p i )
n
= ∑i=1
β i T i (z )
With partial fraction of the rational function:

Once again pole locations tell a lot about shape of pulse response.
Zeros determine the size of β i

69
4.3 Stability (z- domain)
Im{z}
unit circle
X
X .. . .
X
Growing
Constant . ..
X
. Damping
Damping
X X X
Re{z}
X Damping
X
Damping
X
X Growing
z-Plane 70
4.3 Stability (z- domain)
In General
Complex pair oscillatory growth / damping
Real pole exponential growth / decay but maybe oscillatory too (e.g: r n1(n) wherer < 0)
The farther inside unit circle poles are
The faster the damping the higher stability
i . e p i ≤ 1 system stable
71
4.3 stability (z-domain)
Stability directly from State Space:
Exactly as for cts systems, assuming no pole-zero cancellations and D=0
b(z )
H (z ) = = C ( zI − Ad ) Bd
−1
a (z )
Cadj ( zI − Ad )Bd
=
det ( zI − Ad )
b ( z ) = Cadj ( zI − Ad )B d
a ( z ) = det ( zI − Ad )
Poles are eigenvalues of Ad

So check stability, use eigenvalue solver to get e-values of the matrix Ad , then
If λ i < 1 for all i system stable
Where λi is the ith e-value of A d .
72
4.4 Discrete Cavity Model
Converting the transfer function from the continuous cavity model to the discrete model:
ω12  s + ω12 -∆ 
H (s ) =  ∆ω 
∆ω 2 + (s + ω12 )
2
 s + ω12 
The discretization of the model is represented by the z-transform:
 1   H (s )  z − 1  −1  H(s)
H (z ) =  1 −  Z  = ⋅ Z L  |t =kTs
 z  s  z   s 
ω12 ω − ∆ω  ω12 z −1 

H (z ) = ⋅  12 −  ⋅ 
∆ω 2 + ω122  ∆ω ω12   ∆ω 2 + ω122 z 2 − 2 ze ω12Ts ⋅ cos( ∆ωT )s + e 2 ω12Ts 
 ω -∆   ω12Ts  ∆ω ω12  
( )
⋅  z − e ω12Ts ⋅ cos( ∆ωTs ) ⋅  12  -e ⋅ sin( ∆ωTs ) ⋅  
  ∆ω ω12  −
 12ω ∆  
73
4.5 Linear Quadratic Regulator
Given: x (k + 1 ) = A x (k ) + B u (k )
z (k ) = C x ( x )
(Assume D=0 for simplicity)
Suppose the system is unstable or almost unstable.We want to find ufb(k) which will
bring x(k) to Zero, quickly, from any Initial condition.
i.e.
X
+ {A,B,C}
u fb (k ) = ?
74
4.5 Trade Off
Z Z
K K
Ufb
Ufb
K K
(1) “Bad“ damping (1) “Good“ damping

Large Output excursions Small Output excursions
(2) But “Cheap“ control i.e ufb Small (2) But “expensive control i.e ufb large.
75
4.5 Quadratic Forms
A quadratic form is a quadratic function of the components of a vector:
 x1 
x =   ∈ R2
 x2 
f ( x ) = f (x 1 , x 2 )
= ax 12 + bx 1 x 2 + cx 1 + dx 22
 1 
 a b
2  x1   x1 
[ ]
= x1 x 2  [
  + c 0  ]
 b d   x2 
1  x2 
 2  PT
Q
f (x ) = x T Qx + P T x + e
Quadratic Part Linear Part Constant
76
4.5 Quadratic Cost for Regulator
What do we mean by “bad“ damping and “cheap“ control? We now define precisely
what we mean. Consider:
∞
J ≡ ∑ i i i R ui }
{x T
i =0
Q x + u T
The first term penalizes large state excursions, the second penalizes large control.
Q ≥ 0 ,R > 0
Can tradeoff between state excursions and control by varying Q and R.
Large Q “good“ damping important
Large R actuator effort “expensive“
77
4.5 LQR Problem Statement
(Linear quadratic regulator)
Given: xi + 1 = Ax i + Bu i ; x 0 given:
Find control sequence {u0 ,u1,u2 ,...} such that

∞
J = ∑ x {Q x
i =0
T
i i + u iT R u i }
Answer: = minimum
The optimal control sequence is a state feedback sequence {ui }0∞
u i = − K opt xi
(
K opt = R + B T SB ) −1
B T SA
S = AT SA + Q − A T AB R + B T SB ( ) −1
B T SA
Algebraic Riccati Equation (A.R.E) for discrete-time systems.
Note: Since ui = state feedback, it works for any initial state x 0

78
4.5 LQR Problem Statement
Remarks:
(1) So optimal control, ui = − K opt xi is state feedback! This is why we are

interested in state feedbck.
(2) Equation A.R.E. is matrix quadratic equation. Looks pretty intimidating but
Computer can solve in a second.
(3) No tweaking ! Just specify {A,B,C,D} and Q and R, press return button, LQR
Routine Spits out Kopt- done
(Of course picking Q and R is tricky sometimes but that‘s another story).
(4) Design is guaranteed optimal in the sense that it minimizes.

∞
( ) {
J lqr x0 , {u } = ∑ xiT Q xi + uiT R ui
∞
i 0 }
i =0
(Of course that doesn‘t mean its “best“ in the absolute sense .-)
79
4.5 LQR Problem Statement - Remarks
(5) As vary Q/R Ratio we get whole family of Klqr‘s, i.e. can Trade-off between state
excursion (Damping) Vs actuator effort (Control)
State
excursions
Achievable
∞ J z1
Jz = ∑x
i =0
T
i Qxi
= ∑x T
i C T pCx i
= ρ ∑ z iT z i optimal ∞
Ju = ∑u
i=0
T
i Ru i
J u1 Actuator effort
80
4.6 Optimal Linear Estimation
Our optimal control has the form u opt (k ) = − K (k ) xopt (k )
This assumes that we have complete state Information xopt k ( )

-not actually true!.
e.g: in SHO, we might have only a Position sensor but Not a velocity sensor.
How can be obtain “good“ estimates of the velocity state from just observing
the position state?
Furthermore the Sensors may be noisy and the plant itself maybe subject to
outside disturbances (process noise) i.e. we are looking for this:
Process Cx (k )
X
noise
+ {A,B,C}
w(k ) û = K x̂(x|k−1) Noise

+ sensor
Amazing box which
Calculates “good“ estimate v (k )
K Of x(k) from
y (k )
x̂( x|k − 1) y(0),……y(k-1)
81
4.6 Problem Statement :
x (k + 1) = A x (k ) + B w (k )
z (k ) = C x ( x )
y (k ) = C x (k ) + v (k )
Process w(k ) z (k )
X
noise + {A,B,C}
û = K x̂(x|k−1) sensor
Noise v (k )
+
x̂ (x|k − 1)
K Estimator
y (k )
Assume also x 0 is Random & Gaussian and that x (k ), w(k ) + V (k )

( )
are all mutually Independent for all k.
Find : x̂ (k|k−1) Optimal estimate of x(k) given y0 ,.., yk −1

Such that “mean squared error“
E [ x (k ) − x̂ (k|k − 1 )
2
2
]= minimal
( )
Fact from statistics: x̂ k k − 1 = E x(k ) [ ( y0 ,..., yk −1 )]
82
4.6 Kalman Filter
The Kalman filter is an efficient algorithm that computes the new x̂i +1 i ( the linear-least-mean
{ }
( square estimate) of the system state vector xi +1 , given y0 ,..., yi ,by updating the old estimate
x̂i i −1 and old ~
xi i −1 (error) .
x̂i i −1 Kalman x̂i +1 i

(old estimate) Filter (new estimate)
(step i)
pi i −1 pi +1 i
(old error variance) (new error variance)
2
p i i−1 = ~
xi i−1 yi
2 (new measurement)
The Kalman Filter produces x̂i +1 i from x̂i i −1 ( rather than x̂i i ), because it “tracks” the system
“dynamics”. By the time we compute x̂i i from x̂i i −1 , the system state has changed from
xi to xi +1 = Axi + Bwi
83
4.6 Kalman Filter
The Kalman Filter algorithm can be divided in a measurement update and a time update:
Kalman Filter
x̂i i −1 x̂i i x̂i +1 i
Measure. Time
update update
pi i −1 pi i pi +1 i
yi
Measurement update (M.U.):
(
x̂i i = x̂i i −1 + p i i −1C T Cp i i −1C T + V ) (y − Cx̂ )
−1
i i i −1
C (Cp + V ) Cp
T T −1
pi i = pi i −1 − pi i −1 i i −1
C i i −1
x̂i + 1 i = A x̂i i
Time Update (T.U.):
pi + 1 i = Ap i i AT + BWB T
With initial conditions: x̂0 −1 = 0
p̂0 −1 = X 0 84
4.6 Kalman Filter
By pluggin M.U. equations into T.U. equations. One can do both steps at once:
x̂i +1 i = Ax̂i i
(
= Ax̂i i −1 + Api i −1C T Cp i i −1C T + V ) (y − Cx̂ )
−1
i i i −1
(
x̂i +1 i = Ax̂i i −1 + Li yi − C x̂i i −1 )
where (
Li ≡ A pi i −1C T (Cp i i −1
CT +V ) ) −1
pi +1 i = Api i AT + BWB T
[ (
= A pi i −1 − pi i −1C T Cp i i −1C T + V ) −1
]
Cp i i −1 AT + BWB T
(
pi +1 i = Api i AT + BWB T − Api i −1C T Cp i i −1C T + V ) (Cp
−1
i i −1
)
− 1 AT
Known as discrete time Riccati Equation

85
4.6 Picture of Kalman Filter
wi xi +1 −1 xi zi
B + Z 1 C
zi
A vi
yi +
Kalman Filter
x̂i ŷi
x̂i +1 i −1 i −1
+ -
−1
Z 1 C - +
+
A ei
Li
Time varying gain
86
Continued..
4.6 Picture of Kalman Filter
Plant Equations:
xi +1 = Axi + Bui
yi = Cxi + vi
Kalman Filter:
x̂i +1 i = Ax̂i i −1
(
+ Li yi − ŷi i −1
)
yi i −1
= Cx̂i i −1
If v=w=0=> Kalman filter can estimate the state precisely in a finite number of steps.
87
4.6 Kalman Filter
Remarks:
(1) Since y i = Cx i + v i and ŷ i i − 1 = C x̂ i can write estimator equation as
x̂i +1 i = A x̂i i −1
(
+ Li C xi + vi − C x̂i i −1
)
= ( A − Li C ) x̂i i −1
+ Li C xi + vi
can combine this with equation for xi +1

 xi +1   A 0   xi  B 0   wi 
 x̂  =    x̂ +
 i +1 i   iL C A − L i 
C  i i −1 
 0 1 vi 
 zi  C 0   xi 
 ŷ  =    x̂ 
 i i −1  0 C   i i −1 
(2) In practice, the Riccati equation reaches steady state in a few steps. People
Often run with steady-state K.F.i.e
Lss = Apss C T (CPss C T + V)−1
Where
pss = Apss AT + BWBT − Apss C T (CPss C T + V)−1CPss A
88
4.7 LQG Problem
Now we are finally ready to solve the full control problem.

i.e.
Wk zk
X
p(z)
{A,B,C}
U fb
+
-
+ Vk
Xc
H( z)
{Ac ,B c ,C c ,D c } yk
Given: x = Ax + Bu + B w
k +1 k k w k
z = Cx
k k
y = Cx + v
k k k
w ,w = Wδ , v ,v = Vδ
i j ij i j ij
w ,v =0
i j
w k ,v k both Gaussian
For Gaussian, K.L. gives the absolute best estimate 89
4.7 LQG problem
Separation principle: ( we won’t prove)
The separation principle states that the LGQ optimal controller is obtained by:
(1) Using Kalman filter to obtain least squares optimal estimate of the plant state,
i.e.: Let xc(k) = x̂k k −1
(2) Feedback estimated LQR- optimal state feedback

u(k) = − K x (k)
LQR c
= -K x̂
LQR k k − 1
i.e. can treat problems of

- optimal feedback and
- state estimate separately.
90
4.7 Picture of LQG Regulator
Wk B
xk +1
xk zk
+ B + Z −1 1 C
-
zk vk
+
A yk
Kx̂k k −1
K
x̂k k −1 ŷ k k −1
- x̂k +1
+ + Z −1 1 C +
-
+ -
A ek
91
4.7 LQG Regulator
x = A x + ( − Bu ) + B w
k +1 k k w k
Plant z =Cx
k k
y =Cx +v
k k k
 
x̂ = A x̂ + Bu + L  y − C x̂ 
LQG Controller k + 1k k k −1 k  k k k − 1 
u = − K x̂
k k k −1
−1 −1
k = −  R + BT SB 
+ S = AT SA + Q − AT SB  R + BT SB  BT SA
   
−1
T 
L = APC V + CPC T  +P
 
−1
T T T 
= APA + BWB − APC V + CPC T  CPC T
 
92
4.7 Problem Statement (in English)
Want a controller which takes as input noisy measurements, y, and produces as output a
Feedback signal ,u, which will minimize excursions of the regulated plant outputs (if no pole
-zero cancellation, then this is equivalent to minimizing state excursions.)
Also want to achieve “regulation” with as little actuator effort ,u, as possible.
Problem statement (Mathematically)

Find: Controller H ( z ) = C zI − A
c c
( ) −1
B +D
c c
Xc
Uk H( z)
yk
{Ac ,B c ,C c ,D c }
Controller: xc (k + 1) = Ac x(k + 1) + Bc y(k )

yc (k ) = Cc xc (k )
Which will minimize the cost
Limit
 T T 
J = +
Ex Q x u R u 
Where LQG  k k k k
xk +1 = Axk + (− Buk ) + Bwwk k → ∞
Rms “state” Rms “actuator”
Plant zk = C xk excursions effort
yk = C xk + vk
93
4.7 Problem Statement
Remarks:
(1). Q and R are weigthing matrices that allow trading off of rms u and rms x.
T
(2) if Q = C ρ C; ρ > 0 then trade off rms z VS rms u
(3). In the stochastic LQR case, the only difference is that now we don’t have complete state
information yi = Cxi + vi we have only noisy observations
i .e can’t use full state feedback.
Idea: Could we use estimated state Feedback? (i.e. -K x̂ )

k k- 1
94
4.7 Problem Statement
(5) We can let Q/R ratio vary and we’ll obtain family of LQG controllers. Can
Plot rms z vs rms u for each one
Trade-Off curves
rms Z
ACHIEVABLE
LQG, Q/R=0.01
Zrms(1)
X other
LQG, Q/R=100
Zrms(2)
Urms(1) Urms(2) rms U

So by specifying (1) system model, (2) noise variances, (3) optimally criterion
J LQG , and plotting trade off curve completely specifies limit of performance of
System i. e which combinations of (Zrms,Urms) are achievable by any controller

-good “benchmark curve”.
95

Tutmukp01 Talk

Uploaded by

Document Information

Original Description:

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Tutmukp01 Talk

Uploaded by

Copyright:

Available Formats

Tutorial on Control Theory

Stefan Simrock, ITER

ICALEPCS, WTC Grenoble, France, Oct. 10-14, 2011 1

• Matlab / Simulink Tools

• Example beam control

ICALEPCS, WTC Grenoble, France, Oct. 10-14, 2011 2

Closed loop system

System Manipulated System

Closed loop control system

A regulator maintains a physical variable at some constant value in the

Amplitude loop (general form):

Control input U(s)

U(s) Y(s) G(s) = P(s)K(s)

E(s) Y(s) L(s) = G(s)Hc(s)

R(s) Y(s) T(s) = ( 1 + L(s)M(s)) −1 L(s)

&y&(t ) = ∑ Fext = − k y (t ) − γ y& (t ) + u (t )

2.Putt adder in front

3.Synthesize all other terms using integrators!

Mathematical solution: hopefully you know it

Solution of LODE: y(t ) = yh (t ) + yp (t ),

How to get natural response yh (t) ? Characteristic polynomial

Determination of y p (t ) relatively simple, if input u (t ) yields only a finite number of

x j (t ) (so-called states of system)

x&2 (t ) = &y&(t ) = −k y(t ) − γ y& (t ) + u(t )

x& (t ) = A x(t ) + B u (t ) State equation

y (n ) (t ) + a n −1 y (n −1 ) (t ) + ... + a1 y& (t ) + a0 y (t ) = bm u (m ) (t ) + ... + b1 u& (t ) + b0 u (t )

with e.g. Control-Canonical Form (case n = 3 ,m = 3 ):

Natural Response + Particular Solution

With the state transition matrix

Let x1 (t ) = y (t ) and x2 (t ) = y& (t ) . It is:

Write the state equations in matrix form:

Neglect small terms in derivatives for U and I

U&&r + iU&&i (t ) << ωHF

∫ (I& (t ) + iI& (t )) dt << ∫ ω (I (t ) + iI (t )) dt

Envelope equations for real and imaginary component.

− ω1 / 2 t cos ( ∆ωt ) − sin ( ∆ωt )

Transfer function describes input-output relation of system.

U(s) System Y(s)

s X (s) − x(0) = A X (s ) + B U(s)

= C φ(s ) x(0 ) + C φ(s ) B U (s ) + D U (s )

Transfer function G (s ) ( pxr ) (case: x(0)=0):

(ω1 / 2 )π  s + (ω1 / 2 )π − ∆ωπ 

We can deduce for the output of the system.

Y (s) = G(s) U (s) = G(s) Hc (s) E(s)

(I + L(s ) M (s )) Y (s ) = L(s ) R(s )

T (s ) is called : Reference Transfer Function

System characteristics change with system parameter variations

∆T(s) b The sensitivity function is defined as:

Using frequency response approach to

Reduce the Gain Gd ( jω ) between dist. Input and output

what we have not learnt:

A linear time invariant system is called to be BIBO stable (Bounded-input-bounded-output)

Note: it has to be valid for ALL bounded input signals!

The bounded input signal causes a bounded output signal.

BIBO-stability has to be shown/proved for any input. Is is not sufficient to show

Condition for BIBO stability:

We start from the input-output relation

By means of the convolution theorem we get

Therefore it follows immediately:

If the impulse response is absolutely integrable

Can stability be determined if we know the TF of a system?

What do we know about the zeros and the poles?

Assuming D=0 (D could change zeros but not poles)

Require arbitrary “ tweaking”

Poles are eigenvalues of Ad

If λ i < 1 for all i system stable

Large Q “good“ damping important

Large R actuator effort “expensive“