Lecture - Control System Design

Control System Design:
An Introduction
ElectricalMechanical Electro- Electronics
Engineer Mechanics Engineer
Modeling &
Simulation
Sensors
Actuators
Computer Embedded Controls

Systems
Control
Engineer
Engineer
Mechatronic System Design

Control System Design Introduction
K. Craig
Topics
Control System Design Overview
Fundamental Concepts
System Inputs
Step and Sine Inputs
Transfer Functions and Analogies
Poles and Zeros of Transfer Functions
Block Diagrams and Loading Effects
Time Domain and Frequency Domain
State-Space Representation
Linearization of Nonlinear Effects
K. Craig
Open-Loop Control
Basic and Feedforward Control
Closed-Loop Control
Stability and Performance
Sensitivity Analysis
Feedback Control Design Procedure
PID Control and Digital Implementation
Pulse Width Modulation
Parasitic Effects
Sensor Fusion
Observers for Measurement and Control
Advanced Control: Adaptive, Fuzzy Logic
Trade-Offs & Control Design Performance Limitations
K. Craig
Control System Design Overview

Classical Control Design (root-locus and frequency
response analysis and design, i.e., transform methods) is
applicable to linear, time-invariant, single-input, singleoutput systems. This is a complex frequency-domain
approach. The transfer function relates the input to output
and does not show internal system behavior.
Modern Control Design (state-space analysis and

design) is applicable to linear or nonlinear, time-varying or
time-invariant, multiple-input, multiple-output systems.
This is a time-domain approach. This state-space system
description provides a complete internal description of the
system, including the flow of internal energy.
K. Craig
The aim of both techniques is to find a compensation

Gc(s) that satisfies the design specifications.
D(s)
V(s)
A(s)
M(s)
E(s)
R(s)
Gc(s)
C(s)
G(s)
B(s)
H(s)
N(s)
+
Feedback Control System

Knowledge of both approaches, modern and
classical, is essential to produce the best designs.
K. Craig
Fundamental Concepts
System Inputs
Transfer Functions and Analogies
Block Diagrams and Loading Effects
Time Domain and Frequency Domain
State-Space Representation
K. Craig
System Inputs
K. Craig
System Inputs
Initial Energy
Storage
Kinetic
External Driving
Potential
Deterministic
Random
Stationary
Input / System / Output

Concept:
Classification
of
System Inputs
Transient
Periodic
Sinusoidal
Unstationary
"Almost
Periodic"
NonSinusoidal
K. Craig
Input some agency which can cause a system to respond.

Initial energy storage refers to a situation in which a system,
at time = 0, is put into a state different from some reference
equilibrium state and then released, free of external driving
agencies, to respond in its characteristic way. Initial energy
storage can take the form of either kinetic energy or
potential energy.
External driving agencies are physical quantities which vary
with time and pass from the external environment, through
the system interface or boundary, into the system, and
cause it to respond.
We often choose to study the system response to an
assumed ideal source, which is unaffected by the system to
which it is coupled, with the view that practical situations
will closely correspond to this idealized model.
K. Craig
External inputs can be broadly classified as deterministic or

random, recognizing that there is always some element of
randomness and unpredictability in all real-world inputs.
Deterministic input models are those whose complete time
history is explicitly given, as by mathematical formula or a
table of numerical values. This can be further divided into:
transient input model: one having any desired shape, but
existing only for a certain time interval, being constant before
the beginning of the interval and after its end.
periodic input model: one that repeats a certain wave form
over and over, ideally forever, and is further classified as
either sinusoidal or non-sinusoidal.
almost periodic input model: continuing functions which are
completely predictable but do not exhibit a strict periodicity,
e.g., amplitude-modulated input.
K. Craig
10
Random input models are the most realistic input models

and have time histories which cannot be predicted before
the input actually occurs, although statistical properties of
the input can be specified.
When working with random inputs, there is never any
hope of predicting a specific time history before it
occurs, but statistical predictions can be made that
have practical usefulness.
If the statistical properties are time-invariant, then the
input is called a stationary random input. Unstationary
random inputs have time-varying statistical properties.
These are often modeled as stationary over restricted
periods of time.
K. Craig
11

Engineers typically use two inputs to evaluate
dynamic systems: a step input and a sinusoidal input.
Step Input
By a step input of any variable, we will always
mean a situation where the system is at rest at
time t = 0 and we instantly change the input
quantity, from wherever it was just before t = 0, by
a given amount, either positive or negative, and
then keep the input constant at this new value
forever. This leads to a transient response called
the step response of the system.
K. Craig
12
Sine Input
When the input to the system is a sine wave, the
steady-state response of the system, after all the
transients have died away, is called the frequency
response of the system.
These two input types lead to the two views of dynamic
system response: time response and frequency response.
Why only use these two types of input to evaluate a
dynamic system?
The practical difficulty is that precise mathematical
functions for actual real-world inputs will not generally
be known in practice. Therefore the random nature of
many practical inputs makes difficult the development
of performance criteria based on the actual inputs
experienced by real system.
K. Craig
13
It is thus much more common to base

performance evaluation on system response to
simple "standard" inputs step input and sine
wave input. This approach has been successful
for several reasons:
Experience with the actual performance of various
classes of systems has established a good correlation
between the response of systems to these standard
inputs and the capability of the systems to accomplish
their required tasks.
Design is much concerned with comparison of
competitive systems. This comparison can often be
made nearly as well in terms of standard inputs as for
real inputs.
Simplicity of form of standard inputs facilitates
mathematical analysis and experimental verifications.
K. Craig
14
Transfer Functions & Analogies

Definition and Comments
The transfer function of a linear, time-invariant,
differential equation system is defined as the ratio of
the Laplace transform of the output (response
function) to the Laplace transform of the input (driving
function) under the assumption that all initial
conditions are zero.
By using the concept of transfer function, it is
possible to represent system dynamics by algebraic
equations in the Laplace variable s, or the differential
operator D. The highest power of s or D in the
denominator determines the order of the system.
K. Craig
15
Dx
x
D
dx
dt
Dx
x
D2
(x) dt
Differential Operator D
d2x
dt 2
Ds
(x) dt dt
Mass-Spring-Damper
Physical Model
Laplace Variable s
Kx
B(dx/dt)
B
K
M
+x
Newtons 2nd Law
F(t)
2
dx
Fx M dt 2 Mx
F(t) Bx Kx Mx
Mx Bx Kx F(t)
+x
F(t)
Physical Model
Free-Body Diagram
(x is measured from the
static equilibrium
position)
K. Craig
16
Using the differential operator D we can transform the

differential equation to an algebraic equation and then write
the transfer function for the system.
Mx Bx Kx F(t)
Differential Equation
d x
dx
MD x M 2 =Mx
BDx B Bx
dt
dt
MD 2 x MDx Kx F(t) Algebraic Equation
2
(MD 2 BD K)x F(t)

x
1
Transfer Function
2
F MD BD K
K. Craig
17
The transfer function is a property of a system itself,

independent of the magnitude and nature of the input
or driving function.
The transfer function gives a full description of the
dynamic characteristics of the system.
The transfer function does not provide any information
concerning the physical structure of the system; the
transfer functions of many physically different systems
can be identical.
If the transfer function of a system is known, the
output or response can be studied for various forms of
inputs with a view toward understanding the nature of
the system.
If the transfer function of a system is unknown, it may
be established experimentally by introducing known
inputs and studying the output of the system.
K. Craig
18
iin
iout
R
ein
eout
RC Low-Pass Filter
Basic Component
Equations
(Constitutive Equations)
ein eout iR
Kirchhoffs Current Node Law
i R i C i out
iR iC 0
iR iC
ein eout
deout
C
R
dt
deout
iC
dt
deout
RC
eout ein
dt
deout
eout Kein
dt
Deout eout Kein
eout
K
ein D 1
K 1
RC
K. Craig
19
Tank
(Area A)
Large Reservoir
Constant Height H
Analogies
supply pressure
pS
fluid density
g acceleration due to gravity
B df o
fo fi
K dt
fi
p tank gh
RC
fo
+v
q = volume flow rate

pS constant gH (bottom of reservoir)
K
B
Flow
Resistance
R
A
g
dh
h H
dt
deout
eout Kein
dt
K. Craig
20
Simulation Block Diagram

deout
eout Kein
dt
deout 1
Kein eout
dt
1st Order System Block Diagram

RC Electrical System
RC
Input
Block
External
Voltage
ein
Gain Block
K 1
Summation
Block
Gain Block
deout
dt
Integration
Block
eout
Voltage
eout
Output
Block
1
Gain Block
K. Craig
21
The Three Basic Element Input-Output Relationships
Resistor
Damper
Inductor
Mass
Capacitor
Spring
K. Craig
22
qin = i, v
qout = e, f
Resistor, Damper
1
1
i e v f
R
B
e Ri f Bv
Capacitor, Spring
de
i C CDe
dt
1
e
i
CD
Inductor, Mass
di
e L LDi
dt
1
i
e
LD
1 df 1
v
Df
K dt K
K
f v
D
dv
f M
MDv
dt
1
v
f
MD
K. Craig
23
Step Response and Impulse Response

The integral of a step input is a ramp and the derivative of a
step input is an impulse.
An impulse has an infinite magnitude and zero duration and
is mathematical fiction and does not occur in physical
systems.
If, however, the magnitude of a pulse input to a system is
very large and its duration is very short compared to the
systems speed of response, then we can approximate the
pulse input by an impulse function. The impulse input
supplies energy to the system in an infinitesimal time.
The step response of a component or system is the time
response to a step input of some magnitude. The impulse
response of a system is the derivative of the step response
and is the time response to an impulse input of some
strength.
K. Craig
24
The impulse function is explained by the figure,

where we approximate the step function by a
terminated ramp and then let the rise time of
the ramp approach zero. As we let the ramp
get steeper and steeper, the magnitude of
de/dt approaches infinity, and its duration
approaches zero, but the area under it will
always be es. If es = 1 (a unit step function), its
derivative is called the unit impulse function
with an area or strength equal to one unit. The
step function is the integral of the impulse
function, or conversely, the impulse function is
the derivative of the step function. When we
multiply the impulse function by some number,
we increase the strength of the impulse, but
strength now means area, not height as it
does for ordinary functions.
K. Craig
25
Step Responses
of the
Three Basic Elements
K. Craig
26
Frequency Response
If the input to a linear system is a sine wave, the
steady-state output (after the transients have died
out) is also a sine wave with the same frequency,
but with a different amplitude and phase angle.
Both amplitude ratio and phase angle change with
frequency.
The following plots show the frequency response
of the three basic elements.
Note that a decibel dB = 20 log10 (amplitude ratio).
0 dB is an amplitude ratio of 1
+ 6 dB is an amplitude ratio of 2
- 6 dB is an amplitude ration of
+ 20 dB is an amplitude ratio of 10
- 20 dB is an amplitude ratio of 1/10.
K. Craig
27
Frequency Response
K. Craig
28
Frequency Response
K. Craig
29
1
KD
qin
qout
t
q out
q in A sin t
1
1
q in q in dt q out initial
KD
K0
1 t
A sin t
K 0
A
A
q out initial
cos t
K
K
A
A
q out initial
sin t
K
2 K
q out q out initial

q out
Frequency Response
K. Craig
30
Analogies
Analogies Give Engineers Insight!
Insight based on fundamentals is the key to
innovative multidisciplinary problem solving.
A person trying to explain a difficult concept will often say
Well, the analogy is The use of analogies in everyday
life aids in understanding and makes everyone better
communicators. Mechatronic systems depend on the
interactions among mechanical, electrical, magnetic, fluid,
thermal, and chemical elements, and most likely
combinations of these. They are truly multidisciplinary and
the designers of mechatronic systems are from diverse
backgrounds. Knowledge of physical system analogies can
give design teams a significant competitive advantage.
K. Craig
31
Electrical Mechanical Analogies

A signal, element, or system which exhibits mathematical
behavior identical to that of another, but physically
different, signal, element, or system is called an
analogous quantity or analog.
Lets explore the common electrical-mechanical analogy.
These systems are modeled using combinations of pure (only
have the characteristic for which they are named) and ideal
(linear in behavior) elements: resistor (R), capacitor (C), and
inductor (L) for electrical systems and damper (B), spring (K), and
mass (M) for mechanical systems. The variables of interest are
voltage (e) and current (i) for electrical systems and force (f) and
velocity (v) for mechanical systems.
K. Craig
32
Force causes velocity, just as voltage causes current.

A damper dissipates mechanical energy into heat, just as
a resistor dissipates electrical energy into heat.
Springs and masses store energy in two different ways
(potential energy and kinetic energy), just as capacitors
and inductors store energy in two different ways (electric
field and magnetic field).
Spring
Potential Energy
Mass
Kinetic Energy
1 2 1 (Kx) 2 1 f 2
Kx
2
2 K
2K
1
Mv 2
1 2 1 q2
Ce
2
2C
1 2
Li
2
Capacitor
Electric Field
Energy
Inductor
Magnetic Field
Energy
The product (f)(v) represents instantaneous mechanical

power; (e)(i) represents instantaneous electrical power.
K. Craig
33
K. Craig
34
K. Craig
35
force f
velocity v
damper B
spring K
mass M
voltage e
current i
resistor R
capacitor 1/C
inductor L
Resistor e Ri
di
Inductor e L
dt
1
Capacitor e idt
C
Electrical Mechanical
Analogies
Damper f Bv
dv
Mass f M
dt
Spring f K vdt
K. Craig
36
RC Electrical System
Spring-Damper Mechanical System
i
ein
eout
R
C
fo
fi f B f K 0
+v
f i Bv Kx 0
ein e R eC 0
f i Bv f o 0
ein iR eout 0
f
fi B o f o 0
K

B
fo fo fi
K
fo
1
B
B
fi
K
D 1
K
K. Craig
de
ein C out R eout 0
dt
de
RC out eout ein
dt
eout
1
RC
ein RCD 1
fi
37
LR Electrical System
Mass-Damper Mechanical System
i
ein
eout
fo
ein e L e R 0
fi f B f M 0
di
ein L eout 0
dt
d eout
ein L
eout 0
dt R
L deout
eout ein
R dt
L
eout
1
R
ein L D 1
R
f i Bv M v 0
M
+v
fi
f
fi f o M o 0
B

M
fo fo fi
B
M
fo
1
B
fi M D 1
B
K. Craig
38
LRC Electrical System

K
fo
ein
eout
Mass-Spring-Damper
Mechanical System
ein e L e R eC 0
fi f K f B f M 0
di
Ri eout 0
dt
d de
de
ein L C out R C out eout 0
dt dt
dt
d 2 eout
de
LC 2 RCdt out eout ein
dt
dt
eout
KS
1
=
ein LCD 2 RCD 1 1 D 2 2 D 1
2n
n
f i Kx Bv M v 0
ein L
1
R C
KS 1
LC
2 L
M
+v
fi
f
f
fi f o B o M o 0
K
K

M
B
fo fo fo fi
K
K
fo
KS
1
=
f i M D 2 B D 1 1 D 2 2 D 1
K
K
2n
n
K
B 1
KS 1
M
2 KM
K. Craig
39
We can use this analogy to explain the flow of current and the
changes in voltages in a LC (inductor-capacitor) electrical circuit
difficult to envision for most mechanical engineers and even
for some electrical engineers by comparing it to a spring-mass
mechanical system.
The diagrams on the next two slides are color-coded: green, blue,
purple, and orange diagrams for each system correspond to each
other, as do the vertical lines on the graph indicating capacitor
voltage and inductor current at the four specific instances. By
comparing the motion of the mass its changing potential energy
corresponding to energy stored in the electric field of the capacitor
and its changing kinetic energy corresponding to energy stored in
the magnetic field of the inductor one can better understand how
electrical capacitors and inductors function.
For enhanced multidisciplinary engineering system design and

better communication and insight among the design team
members, the use of analogies is a powerful addition to an
engineers toolbox.
K. Craig
40
Inductor-Capacitor (LC) Mass-Spring (MK) Oscillations

i=0
i=0
L
eL
eC
eC
eL
eC
imax
L
imax
eC = 0
eL = 0
eL
eC = 0
i
L
eL = 0
v=0
x = +max
eL
v=0
x = -max
eC
i
L
eL
eC e
L
v = max
x=0
v = max
x=0
eC
K. Craig
41
i=0
i=0
L
eL
eC
eL
eC
i
L
eC
eL
eC
imax
eC = 0
eL = 0
eL = 0
eL
eC = 0
i
L
eL
imax
L
eC e
L
eC
K. Craig
42

Definition of Poles and Zeros
A pole of a transfer function G(s) is a value of s
(real, imaginary, or complex) that makes the
denominator of G(s) equal to zero.
A zero of a transfer function G(s) is a value of s
(real, imaginary, or complex) that makes the
numerator of G(s) equal to zero.
K(s 2)(s 10)
For Example:
G(s)
s(s 1)(s 5)(s 15) 2
Poles: 0, -1, -5, -15 (order 2)
Zeros: -2, -10, (order 3)
K. Craig
43
Colocated Control System

All energy storage elements that exist in the
system exist outside of the control loop.
For purely mechanical systems, separation
between sensor and actuator is at most a rigid
link.
Noncolocated Control System
At least one storage element exists inside the
control loop.
For purely mechanical systems, separating link
between sensor and actuator is flexible.
K. Craig
44
x0
x1
m0
x 0 (s)
m1s 2 K
G 0 (s)
F(s) m t s 2 (m es 2 K)
x1 (s)
K
G1 (s)
F(s) m t s 2 (m es 2 K)
m1
Frictionless Surface
m t m0 m1
1
1
me

m0 m1
Open-Loop Poles
Open-Loop Zeros
G0(s) Colocated System

G1(s) Noncolocated System
s0
s0
s i
K
me
Rigid Body Mode

Flexible Mode
K
Colocated System: s i
m1
Noncolocated System: No Zeros
K. Craig
45
Colocated Transfer Function
x 2 (s)
Ms 2 2K
G(s)
F(s) (Ms 2 3K)(Ms 2 K)
i1
K
1
M
i3
3K
3
M
i2
2K
1
M
Complex Conjugate Poles
Complex Conjugate Zeros

x1
x2
F(t)
K
Frictionless Surface
2
K. Craig
46
Mode Shapes
undeflected
K
K
1
M
3K
3
M
node
fixed
K
2K
2
M
K. Craig
47
Physical Interpretation of Poles and Zeros

Complex Poles
Complex Poles of a colocated control system and
those of a noncolocated control system are identical.
Complex Poles represent the resonant frequencies
associated with the energy storage characteristics of
the entire system.
Complex Poles, which are the natural frequencies of
the system, are independent of the locations of
sensors and actuators.
Complex Poles correspond to the frequencies where
the system behaves as an energy reservoir. Energy
can freely transfer back and forth between the
various internal energy storage elements of the
system.
K. Craig
48
Complex Zeros
Complex Zeros of the two control systems are
quite different and they represent the resonant
frequencies associated with the energy
storage characteristics of a sub-portion of the
system defined by artificial constraints
imposed by the sensors and actuators.
Complex Zeros correspond to the frequencies
where the system behaves as an energy sink.
Complex Zeros represent frequencies at which
energy being applied by the input is
completely trapped in the energy storage
elements of a sub-portion of the original
system such that no output can ever be
detected at the point of measurement.
K. Craig
49
Block Diagrams & Loading Effects

A block diagram of a system is a pictorial representation
of the functions performed by each component and of the
flow of signals. It depicts the interrelationships that exist
among the various components.
It is easy to form the overall block diagram for the entire
system by merely connecting the blocks of the
components according to the signal flow. It is then
possible to evaluate the contribution of each component
to the overall system performance.
A block diagram contains information concerning dynamic
behavior, but it does not include any information on the
physical construction of the system.
K. Craig
50
Many dissimilar and unrelated systems can be

represented by the same block diagram.
A block diagram of a given system is not unique. A
number of different block diagrams can be drawn for
a system, depending on the point of view of the
analysis.
Blocks can be connected in series only if the output
of one block is not affected by the next following
block. If there are any loading effects between
components, it is necessary to combine these
components into a single block.
K. Craig
51
Some Rules of Block Diagram Algebra
K. Craig
52
The unloaded transfer function is an incomplete

component description.
To properly account for interconnection effects one
must know three component characteristics:
the unloaded transfer function of the upstream
component
the output impedance of the upstream component
the input impedance of the downstream
component
Only when the ratio of output impedance Zo over
input impedance Zi is small compared to 1.0, over the
frequency range of interest, does the unloaded
transfer function give an accurate description of
interconnected system behavior.
K. Craig
53
G1(s) and G2(s) are Unloaded Transfer Functions

u
G1(s)
G2(s)
Y(s)
1
G1 (s)
G 2 (s)
Zo1
U(s)
1
Zi2
Zo1
1
Zi2
Only if this is true for the frequency

range of interest will loading effects
be negligible.
K. Craig
54
In general, loading effects occur because when

analyzing an isolated component (one with no other
component connected at its output), we assume no
power is being drawn at this output location.
When we later decide to attach another component to
the output of the first, this second component does
withdraw some power, violating our earlier
assumption and thereby invalidating the analysis
(transfer function) based on this assumption.
When we model chains of components by simple
multiplication of their individual transfer functions, we
assume that loading effects are either not present,
have been proven negligible, or have been made
negligible by the use of buffer amplifiers.
K. Craig
55
Loading Effects Example iin

Passive
RC Low-Pass Filter
iout
R
ein
eout
ein RCs 1 R eout

i Cs
i
1
out
in
eout
1
1
when iout 0
ein RCs 1 s 1
K. Craig
56
iin
iout
R
ein
R
C
eout
2 RC Low-Pass Filters in Series

eout
1 1
G(s)1unloaded G(s) 2unloaded
ein
RCs 1 RCs 1
eout
1
Analysis of Complete Circuit:
ein RCs 12 RCs

K. Craig
57
iin
iout
R
ein
Zout
eout
i out
ein
Zin
iin
ein 0
iout 0
RCs 1
RCs 1
Cs
eout
Output Impedance
Input Impedance
K. Craig
58
eout
G(s)1loaded G(s) 2unloaded
ein
1
1
RCs 1 1 Zout 1
Zin 2
2
RCs 1 RCs
RCs 1
Only if Zout-1 << Zin-2

for the frequency range of interest
will loading effects be negligible.
K. Craig
59
Time & Frequency Domains

We have all witnessed how engineers from different
backgrounds describe the same concepts using different
language and different points of view which often can lead
to confusion and ultimately design errors. Being able to
describe concepts, with clarity and insight, in a variety of
ways is essential.
Time domain and frequency domain are two ways of
looking at the same dynamic system. They are
interchangeable, i.e., no information is lost in changing
from one domain to another. They are complementary
points of view that lead to a complete, clear
understanding of the behavior of a dynamic engineering
system.
K. Craig
60
The time domain is a record of the response of a dynamic

system, as indicated by some measured parameter, as a
function of time. This is the traditional way of observing
the output of a dynamic system.
An example of time response is the displacement of
the mass of the spring-mass-damper system versus
time in response to the sudden placement of an
additional mass (here 50% of the attached mass) on
the attached mass. The resulting response is the step
response of the system due to the sudden application
of a constant force to the attached mass equal to the
weight of the additional mass. Typically when we
investigate the performance of a dynamic system we
use as the input to the system a step input.
K. Craig
61
Time Domain
Physical Model Step Response
Physical System Model
K. Craig
62
Over one hundred years ago, Jean Baptiste Fourier

showed that any waveform that exists in the real world
can be generated by adding up sine waves. By picking
the amplitudes, frequencies, and phases of these sine
waves, one can generate a waveform identical to the
desired signal. While the situation presented below is
contrived, it does illustrate the idea. On the left is a realworld signal and on the right are three signals, the sum of
which is the same as the real-world signal.
Real-World Signal
Three Component Signals

K. Craig
63
Frequency Domain
Real-World Signal
Three Component Signals
Any real-world signal can be broken down into a sum of sine waves
and this combination of sine waves is unique.
Every dynamic signal has a frequency spectrum and if we can
compute this spectrum and properly combine it with the system
frequency response, we can calculate the system time response.
K. Craig
64
Most signals and processes involve both fast and slow

components happening at the same time. In the time
domain (temporal) we measure how long something
takes, whereas in the frequency domain (spectral) we
measure how fast or slow it is. No one domain is always
the best answer, so the ability to easily change domains is
quite valuable and aids in communicating with other team
members.
To show how the time and frequency domains are the
same, the figure on the next slide shows three axes: time,
amplitude, and frequency. The time and amplitude axes
are familiar from the time domain. The third axis,
frequency, allows us to visually separate the sine waves
that add to give us the complex waveform. Note that
phase information is not represented here.
K. Craig
65
Relationship between
Time & Frequency Domains
K. Craig
66
State Space Representation

A state-determined system is a special class of
lumped-parameter dynamic system such that: (i)
specification of a finite set of n independent
parameters, state variables, at time t = t0 and (ii)
specification of the system inputs for all time t t0 are
necessary and sufficient to uniquely determine the
response of the system for all time t t0.
The state is the minimum amount of information
needed about the system at time t0 such that its future
behavior can be determined without reference to any
input before t0.
K. Craig
67
The state variables are independent variables capable

of defining the state from which one can completely
describe the system behavior. These variables
completely describe the effect of the past history of the
system on its response in the future.
Choice of state variables is not unique and they are
often, but not necessarily, physical variables of the
system. They are usually related to the energy stored
in each of the system's energy-storing elements, since
any energy initially stored in these elements can affect
the response of the system at a later time.
State variables do not have to be physical or
measurable quantities, but practically they should be
chosen as such since optimal control laws will require
the feedback of all state variables.
K. Craig
68
The state-space is a conceptual n-dimensional

space formed by the n components of the state
vector. At any time t the state of the system may be
described as a point in the state space and the time
response as a trajectory in the state space.
The number of elements in the state vector is
unique, and is known as the order of the system.
Since integrators in a continuous-time dynamic
system serve as memory devices, the outputs of
integrators can be considered as state variables that
define the internal state of the dynamic system.
Thus the outputs of integrators can serve as state
variables.
K. Craig
69
The state-variable equations are a coupled set of first-order ordinary

differential equations. The derivative of each state variable is expressed as
an algebraic function of state variables, inputs, and possibly time.
x(t) A(t)x(t) B(t)u(t)

y(t) C(t)x(t) D(t)u(t)
Linear, Time-Varying
D(t)
Direct Transmission Matrix
Input Matrix
B(t)
u(t)
Inputs
Output Matrix
x(t)
dt
x(t)
y(t)
C(t)
+
Outputs
State Matrix
A(t)
K. Craig
70
Mx Bx Kx F t
x
1
F MD 2 BD K
xv
1
v F Bv Kx
M
1
0
0
x
x

v K B v 1 F

M
M
M
1
Kx 2 Spring Potential Energy
2
1
Mv 2 Mass Kinetic Energy
2
K. Craig
71
iR
iout = 0
L
iL
R
ein
i iR iL iC
iC
eout eC
di
ein Ri L eout
dt
1
eout i dt
C
deout i
dt
C
eout
d 2 eout
deout
LC 2 RC
eout ein
dt
dt
eout
1
ein LCD2 RCD 1

1 2 Capacitor Electric
CeC
Field Energy
2
1 2 Inductor Magnetic
Li L
2
Field Energy
i L

eC 1
C
1

1
i
L
L ein
eC
0
0
K. Craig
72

Many real-world nonlinearities involve a smooth
curvilinear relation between an independent variable x
and a dependent variable y: y = f(x)
A linear approximation to the curve, accurate in the
neighborhood of a selected operating point, is the
tangent line to the curve at this point. This approximation
is given conveniently by the first two terms of the Taylor
series expansion of f(x):
df
d 2f
(x x) 2
y f (x)
(x x) 2
dx x x
dx x x
2!
df
yy
(x x)
df
dx x x
y y
(x x)
dx x x
y Kx
K. Craig
73
Often a dependent variable y is related nonlinearly to

several independent variables x1, x2, x3, etc. according to
the relation: y=f(x1, x2, x3, ).
We may linearize this relation using the multivariable form
of the Taylor series:
f
f
y f (x1 , x 2 , x 3 , )
(x1 x1 )
(x 2 x 2 )
x1 x ,x ,x ,
x 2 x ,x ,x ,
1
x 3
(x 3 x 3 )
x1 ,x 2 ,x 3 ,
f
y y
x1
x1 ,x 2 ,x 3 ,
f
x 1
x 2
x1 ,x 2 ,x 3 ,
f
x 2
x 3
x 3
x1 ,x 2 ,x 3 ,
y K1x 1 K 2 x 2 K 3 x 3
The partial derivatives can be thought of as the sensitivity of the dependent
variable to small changes in that independent variable.
K. Craig
74
Example
Magnetic Levitation System
Applications
include magnetic
bearings for
vacuum pumps,
conveyor systems
in clean rooms,
high-speed
levitated trains,
and
electromagnetic
automotive valve
actuators.
Electromagnet
Phototransistor
Infrared LED
Levitated Ball
K. Craig
75
Magnetic Levitation System
Electromagnet
i2
f x,i C 2
x
+x
Ball (mass m)
Linearization:
mg
2
2
2
2i
i
i
2
i
Equation of Motion:
C 2 C 2 C 3 x C 2 i
x
x
x
x
i2
mx mg C 2
i2
2i 2
2 i
x
mx mg C 2 C 3 x C 2 i
At Equilibrium:
x
x
x
i2
mg C 2
x
2i 2
2i
mx C 3 x C 2
x
x
K. Craig
76
Use of Experimental Testing in Multivariable Linearization
f m f (i, y)
f
f
f m f i0 , y0
y y0
y i0 ,y0
i
i i0
i0 ,y0
K. Craig
77
Control System Types

Process or Plant
Process Inputs
Manipulated Inputs
Disturbance Inputs
Response Variables
Disturbance
Inputs
Manipulated
Inputs
Control systems are an integral part

of the overall system and not
after-thought add-ons!
The earlier the issues of control are
introduced into the design process, the
better!
Plant
Response
Variables
Why Controls?
Command Following
Disturbance Rejection
Parameter Variations
Everything Needs Controls
for Optimum Functioning!
K. Craig
78
Classification of Control System Types

Open-Loop
Basic
Input-Compensated Feedforward
Disturbance-Compensated
Command-Compensated
Closed-Loop (Feedback)
Classical
Root-Locus
Frequency Response
Modern (State-Space)
Advanced
e.g., Adaptive, Fuzzy Logic
K. Craig
79
Basic Open-Loop Control System

Flow of Energy
and/or Material
Control
Effector
Plant Disturbance Input
Plant
Manipulated
Input
Controlled
Variable
Plant
Control
Director
Satisfactory if:
disturbances are not too great
changes in the desire value are not too severe
performance specifications are not too stringent
Desired Value
of
Controlled Variable
K. Craig
80
Open-Loop Input-Compensated Feedforward Control:

Disturbance-Compensated
Disturbance
Compensation
Flow of Energy
and/or Material
Control
Effector
Disturbance
Sensor
Controlled
Variable
Plant
Plant
Manipulated
Input
Control
Director
Measure the disturbance

Estimate the effect of the disturbance on the
controlled variable and compensate for it
Desired Value
of
Controlled Variable
K. Craig
81
Disturbance-Compensated Feedforward
Control
Basic Idea: Measure important load variables and
take corrective action before they upset the
process.
In contrast, a feedback controller, as we will see,
does not take corrective action until after the
disturbance has upset the process and generated
an error signal.
There are several disadvantages to disturbancecompensated feedforward control:
The load disturbances must be measured on
line. In many applications, this is not feasible.
K. Craig
82
The quality of the feedforward control

depends on the accuracy of the process
model; one needs to know how the
controlled variable responds to changes in
both the load and manipulated variables.
Ideal feedforward controllers that are
theoretically capable of achieving perfect
control may not be physically realizable.
Fortunately, practical approximations of
these ideal controllers often provide very
effective control.
K. Craig
83
Open-Loop Input-Compensated Feedforward Control:

Command-Compensated
Flow of Energy
and/or Material
Control
Effector
Based on the
knowledge of plant
characteristics, the
desired value input is
augmented by the
command
compensator to
produce improved
performance.
Plant
Manipulated
Input
Controlled
Variable
Plant
Control
Director
Command
Compensator
Desired Value
of
Controlled Variable
K. Craig
84
Comments:
Open-loop systems without disturbance or command
compensation are generally the simplest, cheapest,
and most reliable control schemes. These should be
considered first for any control task.
If specifications cannot be met, disturbance and/or
command compensation should be considered next.
When conscientious implementation of open-loop
techniques by a knowledgeable designer fails to yield
a workable solution, the more powerful feedback
methods should be considered.
K. Craig
85
Closed-Loop (Feedback)
Control System
Flow of Energy
and/or Material
Control
Effector
Plant
Manipulated
Input
Controlled
Variable
Plant
Controlled
Variable
Sensor
Open-Loop Control System

is converted to a
Closed-Loop Control System
by adding:
Control
Director
Desired Value
of
Controlled Variable
measurement of the controlled variable

comparison of the measured and desired values of the
controlled variable
K. Craig
86
Basic Benefits of Feedback Control

Cause the controlled variable to accurately follow the
desired variable; corrective action occurs as soon as
the controlled variable deviates from the command.
Greatly reduces the effect on the controlled variable
of all external disturbances in the forward path. It is
ineffective in reducing the effect of disturbances in
the feedback path (e.g., those associated with the
sensor), and disturbances outside the loop (e.g.,
those associated with the reference input element).
Is tolerant of variations (due to wear, aging,
environmental effects, etc.) in hardware parameters
of components in the forward path, but not those in
the feedback path (e.g., sensor) or outside the loop
(e.g., reference input element).
K. Craig
87
Can give a closed-loop response speed much

greater than that of the components from which
they are constructed.
Inherent Disadvantages of Feedback Control

No corrective action is taken until after a deviation
in the controlled variable occurs. Thus, perfect
control, where the controlled variable does not
deviate from the set point during load or set-point
changes, is theoretically impossible.
It does not provide predictive control action to
compensate for the effects of known or
measurable disturbances.
K. Craig
88
It may not be satisfactory for processes with large

time constants and/or long time delays. If large
and frequent disturbances occur, the process may
operate continually in a transient state and never
attain the desired steady state.
In some applications, the controlled variable
cannot be measured on line and, consequently,
feedback control is not feasible.
For situations in which feedback control by itself is
not satisfactory, significant improvements in control
can be achieved by adding feedforward control.
K. Craig
89
Feedback Control System Block Diagram

D(s)
V(s)
A(s)
M(s)
E(s)
R(s)
Gc(s)
B(s)
G c (s)G(s)
C(s)
R(s) 1 G c (s)G(s)H(s)
C(s)
G(s)
D(s) 1 G c (s)G(s)H(s)
C(s) G c (s)G(s)H(s)
N(s) 1 G c (s)G(s)H(s)
C(s)
G(s)
+
H(s)
Closed
Loop
N(s)
+
B(s)
G c (s)G(s)H(s)
E(s)
Open Loop
C(s)
G c (s)G(s)
E(s)
Feedforward
K. Craig
90
Stability and Performance

If a system in equilibrium is momentarily excited by command
and/or disturbance inputs and those inputs are then removed,
the system must return to equilibrium if it is to be called
absolutely stable.
If action persists indefinitely after excitation is removed, the
system is judged absolutely unstable.
If a system is stable, how close is it to becoming unstable?
Relative stability indicators are gain margin and phase margin.
If we want to make valid stability predictions, we must include
enough dynamics in the system model so that the closed-loop
system differential equation is at least third order.
An exception to this rule involves systems with dead times,
where instability can occur when the dynamics (other than
the dead time itself) are zero, first, or second order.
K. Craig
91
The analytical study of

stability becomes a
study of the stability of
the solutions of the
closed-loop systems
differential equations.
A complete and general
stability theory is based
on the locations in the
complex plane of the
roots of the closed-loop
system characteristic
equation, stable
systems having all of
their roots in the LHP.
K. Craig
92
Results of practical use to engineers are mainly limited to

linear systems with constant coefficients, where an exact
and complete stability theory has been known for a long
time.
Exact, general results for linear time-variant and
nonlinear systems are nonexistant. Fortunately, the
linear time-invariant theory is adequate for many practical
systems.
For nonlinear systems, an approximation technique
called the describing function technique has a good
record of success.
Digital simulation is always an option and, while no
general results are possible, one can explore enough
typical inputs and system parameter values to gain a high
degree of confidence in stability for any specific system.
K. Craig
93
Two general methods of determining the presence of

unstable roots without actually finding their numerical
values are:
Routh Stability Criterion
This method works with the closed-loop
system characteristic equation in an algebraic
fashion.
Nyquist Stability Criterion
This method is a graphical technique based on
the open-loop frequency response polar plot.
Both methods give the same results, a statement of
the number (but not the specific numerical values) of
unstable roots. This information is generally
adequate for design purposes.
K. Craig
94
This theory predicts excursions of infinite magnitude for

unstable systems. Since infinite motions, voltages,
temperatures, etc., require infinite power supplies, no
real-world system can conform to such a mathematical
prediction, casting possible doubt on the validity of our
linear stability criterion since it predicts an impossible
occurrence.
What actually happens is that oscillations, if they are to
occur, start small, under conditions favorable to and
accurately predicted by the linear stability theory. They
then start to grow, again following the exponential trend
predicted by the linear model. Gradually, however, the
amplitudes leave the region of accurate linearization,
and the linearized model, together with all its
mathematical predictions, loses validity.
K. Craig
95
Since solutions of the now nonlinear equations are usually

not possible analytically, we must now rely on experience
with real systems and/or nonlinear computer simulations
when explaining what really happens as unstable
oscillations build up.
First, practical systems often include over-range alarms
and safety shut-offs that automatically shut down
operation when certain limits are exceeded. If certain
safety features are not provided, the system may destroy
itself, again leading to a shut-down condition. If safe or
destructive shut-down does not occur, the system usually
goes into a limit-cycle oscillation, an ongoing,
nonsinusoidal oscillation of fixed amplitude. The wave
form, frequency, and amplitude of limit cycles is governed
by nonlinear math models that are usually analytically
unsolvable.
K. Craig
96
Most of our discussion of performance will involve

rather specific mathematical performance criteria
whereas the ultimate success of a controlled process
generally rests on economic considerations which are
difficult to calculate.
This rather nebulous connection between the
technical criteria used for system design and the
overall economic performance of the manufacturing
unit results in the need for much exercise of judgment
and experience in decision making at the higher
management levels.
Control system designers must be cognizant of these
higher-level considerations but they usually employ
rather specific and relatively simple performance
criteria when evaluating their designs.
K. Craig
97
Control System Objective

C follow desired value V and ignore disturbances
Technical performance criteria must have to do
with how well these two objectives are attained
Performance depends both on system characteristics
and the nature of V, D, and N.
D(s)
V(s)
A(s)
M(s)
E(s)
R(s)
Gc(s)
B(s)
C(s)
G(s)
+
H(s)
Basic Linear Feedback System

K. Craig
N(s)
+
98
The practical difficulty is that precise mathematical functions

for V, D, and N will not generally be known in practice.
Therefore the random nature of many practical commands
and disturbances makes difficult the development of
performance criteria based on the actual V, D, and N
experienced by real system.
It is thus much more common to base performance
evaluation on system response to simple "standard" inputs
such as steps, ramps, and sine waves.
This approach has been successful for several reasons:
In many areas, experience with the actual performance of
various classes of control systems has established a
good correlation between the response of systems to
standard inputs and the capability of the systems to
accomplish their required tasks.
K. Craig
99
Design is much concerned with comparison of

competitive systems. This comparison can often
be made nearly as well in terms of standard
inputs as for real inputs.
Simplicity of form of standard inputs facilitates
mathematical analysis and experimental
verifications.
For linear systems with constant coefficients,
theory shows that the response to a standard
input of frequency content adequate to exercise
all significant system dynamics can then be used
to find mathematically the response to any form of
input.
K. Craig
100
Standard performance criteria may be classified as

falling into two categories:
Time-Domain Specifications: Response to steps,
ramps, and the like, e.g., step response criteria
rise time, peak time, percent overshoot, settling
time, decay ratio, and steady-state error.
Frequency-Domain Specifications: Concerned
with certain characteristics of the system
frequency response, e.g., bandwidth, peak
amplitude ratio, gain margin, and phase margin.
K. Craig
101
Both time-domain and frequency-domain design

criteria generally are intended to specify one or the
other of:
speed of response
relative stability
steady-state errors
Both types of specifications are often applied to the
same system to ensure that certain behavior
characteristics will be obtained.
All performance specifications are meaningless
unless the system is absolutely stable.
K. Craig
102
It is important to realize that, because of model

uncertainties, it is not merely sufficient for a system to
be stable, but rather it must have adequate stability
margins.
Stable systems with low stability margins work only
on paper; when implemented in real time, they are
frequently unstable.
The way uncertainty has been quantified in classical
control is to assume that either gain changes or
phase changes occur. Typically, systems are
destabilized when either gain exceeds certain limits
or if there is too much phase lag (i.e., negative phase
associated with unmodeled poles or time delays).
The tolerances of gain or phase uncertainty are the
gain margin and phase margin.
K. Craig
103
Consider the following design problem:

Given a plant transfer function G2(s), find a
compensator transfer function G1(s) which yields
the following:
stable closed-loop system
good command following
good disturbance rejection
insensitivity of command following to modeling
errors (performance robustness)
stability robustness with unmodeled dynamics
sensor noise rejection
K. Craig
104
Without closed-loop stability, a discussion of

performance is meaningless. It is critically important
to realize that the compensator is actually designed
to stabilize a nominal open-loop plant. Unfortunately,
the true plant is different from the nominal plant due
to unavoidable modeling errors.
Knowledge of modeling errors should influence the
design of the compensator.
We assume here that the actual closed-loop system
is absolutely stable.
K. Craig
105
Open-Loop Shaping
Desired Shape for Open-Loop
Transfer Function
Smooth transition from the low to highfrequency range, i.e., -20 dB/decade
slope near the gain crossover frequency
Gain above this level
at low frequencies
stable closed-loop system

good command following
good disturbance rejection
insensitivity of command
following to modeling
errors (performance
robustness)
stability robustness with
unmodeled dynamics
sensor noise rejection
Gain below this level

at high frequencies
Frequencies for good

command following,
disturbance reduction,
sensitivity reduction
Sensor noise,
unmodeled highfrequency dynamics
are significant here.
Linear, Time-Invariant Systems

K. Craig
106
Instability in Feedback Control Systems

All feedback systems can become unstable if
improperly designed.
In all real-world components there is some kind of
lagging behavior between the input and output,
characterized by s and ns.
Instantaneous response is impossible in the real
world!
Instability in a feedback control system results from
an improper balance between the strength of the
corrective action and the system dynamic lags.
K. Craig
107
+M
-M
Example
Liquid level C in a tank is
manipulated by controlling
the volume flow rate M by
means of a three-position
on/off controller with error
dead space EDS.
Transfer function 1/As
between M and C
represents conservation of
volume between volume
flow rate and liquid level.
Liquid-level sensor
measures C perfectly but
with a data transmission
delay dt.
EDS
C(t)
Area A
Tank Liquid-Level Feedback Control System
K. Craig
108
Tank Level Feedback Control System

M
Three-Position
On-Off Controller
M
M
Step Input
Sum
Dead Zone
Sign
Flow Rate
1/A
s
Plant
C
C
Transport Delay
B
B
M = 3, A = 2, tau_dt = 0.1 : stable

M = 5, A = 2, tau_dt = 0.2 : unstable
MatLab / Simulink Block Diagram

Strength of corrective action is represented by M (also by 1/A).
System dynamic lag is represented by dt.
K. Craig
109
Stable Behavior of the Tank Liquid-Level

3.5
3
2.5
signal C: solid
signal B: dotted
signal 0.1*M: dashed
1.5
1
0.5
0
0.5
1.5
time (sec)
2.5
K. Craig
110
Unstable Behavior of the Tank Liquid-Level

3.5
3
2.5
1.5
signal C: solid
signal B: dotted
signal 0.1*M: dashed
1
0.5
0
-0.5
0.5
1.5
time (sec)
2.5
K. Craig
111
Sensitivity Analysis
Consider the function y = f(x). If the parameter x
changes by an amount x, then y changes by the
amount y. If x is small, y can be estimated from
the slope dy/dx as follows:
dy
y
x
dx
The relative or percent change in y is y/y. It is

related to the relative change in x as follows:
y dy x x dy x
y dx y y dx x
K. Craig
112
The sensitivity of y with respect to changes in x is

given by:
x dy dy / y d(ln y)
S
y dx dx / x d(ln x)
y
x
Thus
y
y x
Sx
y
x
Usually the sensitivity is not constant. For example,

the function y = sin(x) has the sensitivity function:
x dy x
x cos(x)
x
S
cos(x)
y dx y
sin(x)
tan(x)
y
x
K. Craig
113
Sensitivity of Control Systems to Parameter

Variation and Parameter Uncertainty
A process, represented by the transfer function G(s),
is subject to a changing environment, aging,
ignorance of the exact values of the process
parameters, and other natural factors that affect a
control process.
In the open-loop system, all these errors and
changes result in a changing and inaccurate output.
However, a closed-loop system senses the change
in the output due to the process changes and
attempts to correct the output.
The sensitivity of a control system to parameter
variations is of prime importance.
K. Craig
114
Accuracy of a measurement system is affected by

parameter changes in the control system
components and by the influence of external
disturbances.
A primary advantage of a closed-loop feedback
control system is its ability to reduce the systems
sensitivity.
Consider the closed-loop system shown. Let the
disturbance D(s) = 0.
D(s)
R(s)
+
M(s)
E(s)
Gc(s)
C(s)
G(s)
B(s)
H(s)
K. Craig
115
An open-loop systems block diagram is given by:
R(s)
C(s)
Gc(s)
G(s)
The system sensitivity is defined as the ratio of

the percentage change in the system transfer
function T(s) to the percentage change in the
process transfer function G(s) (or parameter) for a
small incremental change:
C(s)
T(s)
R(s)
T / T T G
T
SG
G / G G T
K. Craig
116
For the open-loop system

C(s)
T(s)
G c (s)G(s)
R(s)
T / T T G
G(s)
T
SG
G c (s)
1
G / G G T
G c (s)G(s)
For the closed-loop system
G c (s)G(s)
C(s)
T(s)
T / T T G
S
G / G G T
1
G
1
(1 G c GH) 2 G c G
G c 1 G c GH
1 G c GH
T
G
K. Craig
117
The sensitivity of the system may be reduced

below that of the open-loop system by increasing
GcGH(s) over the frequency range of interest.
The sensitivity of the closed-loop system to
changes in the feedback element H(s) is:
G c (s)G(s)
C(s)
T(s)
T / T T H
S
H / H H T
(G c G) 2
G c GH
H
2
G cG
(1 G c GH)
1 G cGH
1 G c GH
T
H
K. Craig
118
When GcGH is large, the sensitivity approaches

unity and the changes in H(s) directly affect the
output response. Use feedback components that
will not vary with environmental changes or can
be maintained constant.
As the gain of the loop (GcGH) is increased, the
sensitivity of the control system to changes in the
plant and controller decreases, but the sensitivity
to changes in the feedback system (measurement
system) becomes -1.
Also the effect of the disturbance input can be
reduced by increasing the gain GcH since:
C(s)
G(s)
D(s)
1 G c (s)G(s)H(s)
K. Craig
119
Therefore:
Make the measurement system very accurate and
stable.
Increase the loop gain to reduce sensitivity of the
control system to changes in plant and controller.
Increase gain GcH to reduce the influence of
external disturbances.
In practice:
G is usually fixed and cannot be altered.
H is essentially fixed once an accurate
measurement system is chosen.
Most of the design freedom is available with
respect to Gc only.
K. Craig
120
It is virtually impossible to achieve all the design

requirements simply by increasing the gain of Gc.
The dynamics of Gc also have to be properly
designed in order to obtain the desired performance
of the control system.
Very often we seek to determine the sensitivity of the
closed-loop system to changes in a parameter
within the transfer function of the system G(s). Using
the chain rule we find:
T
T G
S SGS
Very often the transfer function T(s) is a fraction of

the form:
N(s, )
T(s, )
D(s, )
K. Craig
121
Then the sensitivity to (0 is the nominal value)

is given by:
ST
T / T ln T ln N
ln D
SN SD
/ ln ln 0 ln 0
K. Craig
122
Feedback Control System Design

Procedure
Control Engineering is an important part of the design
process of most dynamic systems.
The deliberate use of feedback can:
Stabilize an otherwise unstable system
Reduce the error due to disturbance inputs
Reduce the tracking error while following a command
input
Reduce the sensitivity of a closed-loop transfer
function to small variations in internal system
parameters
K. Craig
123
Remember that the purpose of control is to aid the

product or process the mechanism, the robot, the
chemical plant, the aircraft, or whatever to do its
job.
Engineers must take into account early in their plans
the contribution of control to the design process!
More and more systems are being designed so that
they will not work without feedback!
Control system design begins with a proposed
product or process whose satisfactory dynamic
performance depends on feedback for:
Stability
Disturbance Regulation
Tracking Accuracy
Reduction of the Effects of Parameter Variations
K. Craig
124
Having a general step-by-step approach for feedback

control system design serves two purposes:
It provides a useful starting point for any realworld controls problem.
It provides meaningful checkpoints once the
design process is underway.
System
Design
System Dynamics
&
Control Structure
K. Craig
125
Modern
Multidisciplinary
Engineering
System
Other Components
Communications
Operator
Interface
Human Factors
Computation
Software, Electronics
Actuation
Power Modulation
Energy Conversion
Simultaneous
Optimization
of all
System Components
Instrumentation
Energy Conversion
Signal Processing
Physical System
Mechanical, Fluid, Thermal,
Chemical, Electrical,
Biomedical, Civil, Mixed
K. Craig
126
Multidisciplinary Engineering System Design Team
Social Scientists
&
Non-Technical Experts
ElectricalMechanical Electro- Electronics

Engineer Mechanics Engineer
Modeling &
Simulation
Business
Experts
ProblemSpecific
Engineers
Sensors
Actuators
Computer Embedded Controls

Systems
Control
Engineer
Engineer
Physicists, Chemists,
Mathematicians, &
Computer Scientists
Multidisciplinary System Design

K. Craig
127
Sequence of Steps for Feedback Control System

Design
1. Understand the process and translate dynamic
performance requirements into time, frequency, or polezero specifications.
What is the system and what is it supposed to do?
The importance of understanding the process cannot
be overemphasized!
Do not confuse the approximation with the reality!
You must be able to:
Use the simplified model for its intended purpose
Return to an accurate model or the actual physical
system to really verify the design performance
K. Craig
128
Examples of Dynamic Performance Requirements
Time Response
Frequency Response
Pole-Zero
K. Craig
129
2. Select the types and number of sensors considering

location, technology, functional performance,
physical properties, quality factors, and cost.
If you cant observe it, you cant control it!
Which variables are important to control?
Which variables can physically be measured?
Select sensors that indirectly allow a good
estimate to be made of the critical unmeasurable
variables.
It is important to consider sensors for the
disturbances, e.g., in chemical processes, it is
often beneficial to sense a load disturbance
directly because improved performance can be
obtained if this information is fed forward to the
controller.
K. Craig
130
3. Select the types and number of actuators

considering location, technology, functional
performance, physical properties, quality factors,
and cost.
In order to control a dynamic system, you must be
able to influence the response. The actuator does
this.
Before choosing a specific actuator, consider
which variables can be influenced.
The actuators must be capable of driving the
system so as to meet the required performance
specifications.
K. Craig
131
4. Make a linear model of the process, actuator, and

sensor.
Take the best choice for process, actuator, and sensor.
Identify the equilibrium point of interest.
Construct a small-signal dynamic model valid over the
range of frequencies included in the performance
specifications.
Validate this model with experimental measurements
where possible.
Express the model in many forms: state-variable, polezero, and frequency-response forms.
Simplify and reduce the order of the model, if necessary.
Quantify model uncertainty.
K. Craig
132
5. Make a simple trial design based on concepts of

lead-lag compensation or PID control.
To form an initial estimate of the complexity of the
design problem, sketch a frequency-response
(Bode) plot and a root-locus plot with respect to
plant gain.
If the plant-actuator-sensor model is stable and
minimum phase, the Bode plot will probably be
the most useful; otherwise, the root locus shows
very important information with respect to
behavior in the right-half plane.
Try to meet specifications with a simple controller
of the lead-lag, PID variety.
Do not overlook feedforward of the disturbances.
Consider the effect of sensor noise.
K. Craig
133
6. Consider modifying the plant itself for improved closedloop control.

Based on the simple control design, evaluate the source
of the undesirable characteristics of system
performance.
Reevaluate the specifications, the physical configuration
of the process, and the actuator and sensor selections in
light of the preliminary design. Return to step 1 if
improvement seems necessary or feasible.
It may be much easier to meet specifications by altering
the process than to meet them by control strategies
alone!
Consider all parts of the design, not only the control
logic, to meet the specifications in the most costeffective way.
K. Craig
134
7. Make a trial pole-placement design based on optimal

control or other criteria.
If the trial-and-error compensators do not give entirely
satisfactory performance, consider a design based on
optimal control.
Select the location for your control poles that balance
system performance and control effort.
Select the location for the estimator poles that
represent a compromise between sensor and process
noise.
Plot the corresponding open-loop frequency response
and the root locus to evaluate the stability margins of
this design and its robustness to parameter changes.
Compare this optimal design with the transformmethod design and select the better of the two.
K. Craig
135
8. Build a computer model and simulate the performance of

the design.
After reaching the best compromise among process
modification, actuator and sensor selection, and controller
design choice, run a simulation of the system.
Include important nonlinearities, parasitic effects, and
parameter variations you expect to find during operation.
Design iterations should continue until the simulation
confirms acceptable stability and robustness.
As the design progresses, more complete and detailed
models (truth models) will be used.
Digital control implementation should be taken into account.
If the performance is not satisfactory, return to step 1 and
repeat. Consider modifying the plant itself for improved
closed-loop control.
K. Craig
136
9. Build a prototype and test it.

The proof of the pudding is in the eating!
Simulation without experimental verification is at
best questionable and at worst useless!
At this point you verify the quality of the model,
discover unexpected effects, and consider ways to
improve the design.
Implement the controller using an embedded
software/hardware.
Tune the controller, if necessary.
After these tests, you may want to reconsider the
sensor, actuator, and process and return to step 1.
K. Craig
137
This outline is an approximation of good practice.

One very important consideration (Step 6) was for
changing the plant itself to make the control problem
easier and provide maximum closed-loop
performance.
In many cases, proper plant modifications can
provide additional damping or increase the
stiffness, change in mode shapes, reduction of
system response to disturbances, reduction of
Coulomb friction, change in thermal capacity or
conductivity, and so on.
Designing the system and throwing it over the
wall to the control group is inefficient and flawed!
System design and control design must be done
simultaneously!
K. Craig
138
Digital Implementation of PID Control

Information Domain
Digital Set Point
Sampled &
Quantized
Measurement
A/D
Converter
Digital
Computer
Sampled & Quantized

Control Signal
Sampling
System
D/A
Converter
Sampling
Switch
Anti-Aliasing
Filter
Sensor
Plant /
Process
Actuator
Power Domain
K. Craig
139
Advantages of Digital Control

The current trend toward using dedicated,
microprocessor-based, and often decentralized
(distributed) digital control systems in industrial
applications can be rationalized in terms of the major
advantages of digital control:
Digital control is less susceptible to noise or
parameter variation in instrumentation because
data can be represented, generated, transmitted,
and processed as binary words, with bits
possessing two identifiable states.
K. Craig
140
Very high accuracy and speed are possible

through digital processing. Hardware
implementation is usually faster than software
implementation.
Digital control can handle repetitive tasks
extremely well, through programming.
Complex control laws and signal conditioning
methods that might be impractical to
implement using analog devices can be
programmed.
High reliability can be achieved by minimizing
analog hardware components and through
decentralization using dedicated
microprocessors for various control tasks.
K. Craig
141
Large amounts of data can be stored using

compact, high-density data storage methods.
Data can be stored or maintained for very long
periods of time without drift and without being
affected by adverse environmental conditions.
Fast data transmission is possible over long
distances without introducing dynamic delays,
as in analog systems.
Digital control has easy and fast data retrieval
capabilities.
Digital processing uses low operational
voltages (e.g., 0 - 12 V DC).
Digital control has low overall cost.
K. Craig
142
Digital Signals are:

discrete in time
quantized in amplitude
You must understand the effects of:

sample period
quantization size
Discrete
in
Time
Continuous
in
Time
Discrete
in
Amplitude
D-D D-C
Continuous
in
Amplitude
C-D
C-C
K. Craig
143
In a real sense, the problems of analysis and design

of digital control systems are concerned with taking
into account the effects of the sampling period, T,
and the quantization size, q.
If both T and q are extremely small (i.e., sampling
frequency 50 or more times the system bandwidth
with a 16-bit word size), digital signals are nearly
continuous, and continuous methods of analysis and
design can be used.
It is most important to understand the effects of all
sample rates, fast and slow, and the effects of
quantization for large and small word sizes.
It is worthy to note that the single most important
impact of implementing a control system digitally is
the delay associated with the D/A converter, i.e., T/2.
K. Craig
144
Simulation of Continuous and Quantized Signal

1.4
amplitude: continuous and quantized
1.2
0.8
0.6
0.4
0.2
0.1
0.2
0.3
0.4
0.5
time (sec)
0.6
0.7
0.8
0.9
K. Craig
145
Continuous Output and D/A Output

1
0.8
0.6
0.4
amplitude
0.2
0
-0.2
-0.4
-0.6
-0.8
-1
0.1
0.2
0.3
0.4
0.5
time (sec)
0.6
0.7
0.8
0.9
K. Craig
146
Aliasing
The analog feedback signal coming from the sensor
contains useful information related to controllable
disturbances (relatively low frequency), but also may
often include higher frequency "noise" due to
uncontrollable disturbances (too fast for control
system correction), measurement noise, and stray
electrical pickup. Such noise signals cause
difficulties in analog systems and low-pass filtering is
often needed to allow good control performance.
K. Craig
147
In digital systems, a phenomenon called aliasing

introduces some new aspects to the area of noise
problems. If a signal containing high frequencies is
sampled too infrequently, the output signal of the
sampler contains low-frequency ("aliased")
components not present in the signal before
sampling. If we base our control actions on these
false low-frequency components, they will, of course,
result in poor control. The theoretical absolute
minimum sampling rate to prevent aliasing is 2
samples per cycle; however, in practice, rates of
about 10 are more commonly used. A highfrequency signal, inadequately sampled, can produce
a reconstructed function of a much lower frequency,
which can not be distinguished from that produced by
adequate sampling of a low-frequency function.
K. Craig
148
Simulation of Continuous and Sampled Signal

0.5
0.4
amplitude: analog and sampled signals
0.3
0.2
0.1
0
-0.1
-0.2
-0.3
-0.4
-0.5
0.2
0.4
0.6
0.8
1
time (sec)
1.2
1.4
1.6
1.8
K. Craig
149
PID Control & Digital Implementation

A digital controller differs from an analog controller in that
the signals must be sampled and quantized.
A signal to be used in digital logic needs to be sampled
first; then the samples need to be converted by an analogto-digital (A/D) converter into a quantized digital number.
Once the digital computer has calculated the proper next
control signal value, this value needs to be converted into a
voltage and held constant or otherwise extrapolated by a
digital-to-analog converted (D/A) in order to be applied to
the actuator of the process.
The control signal is not changed until the next sampling
period.
K. Craig
150
As a result of the sampling, there are more strict limits on

the speed or bandwidth of a digital controller than on
analog devices.
A reasonable rule of thumb for selecting the sampling
period is that during the rise time of the response to a
step, the input to the discrete controller should be
sampled approximately 6 times. By adjusting the
controller for the effects of sampling, the sampling can
be adjusted to 2 to 3 times per rise time. This
corresponds to a sampling frequency that is 10 to 20
times the systems closed-loop bandwidth.
The quantization of the controller signals introduces an
equivalent extra noise into the system, and to keep this
interference at an acceptable level, the A/D converter
usually has an accuracy of 10 to 12 bits.
K. Craig
151
We will consider a simplified technique for finding a

discrete (sampled, but not quantized) equivalent to a
given continuous controller.
The method depends on the sampling period Ts being
short enough that the reconstructed control signal is
close to the signal that the original analog controller
would have produced.
We also assume that the numbers used in the digital
logic have enough accurate bits so that the quantization
implied in the A/D and D/A processes can be ignored.
Finding a discrete equivalent to a given analog controller
is equivalent to finding a recurrence equation for the
samples of the control which will approximate the
differential equation of the controller.
K. Craig
152
The assumption is that we have the transfer function of an

analog controller and wish to replace it with a discrete
controller that will accept samples of the controller input
e(kTs), from a sampler and, using past values of the control
signal, u(kTs), and present and past values of the input,
e(kTs), will compute the next control signal to be sent to the
actuator.
Lets consider the PID controller, as an example. The
proportional-integral-derivative (PID) controller is the most
widely used controller in use today. It can stabilize a
system, increase the speed of response of a system, and
reduce steady-state errors of a system.
t
de(t)
u(t) K P e(t) K I e( ) d K D
0
dt
KI
U(s) K P
K Ds E(s)
s
K. Craig
153
Proportional Control
Virtually all controllers have a large proportional gain.
While we will see that derivative gain can provide
incremental improvements at high frequencies, and
integral gain improves performance at lower
frequencies, the proportional gain is the primary actor
across the entire frequency range of operation.
Here the manipulating variable U is directly
proportional to the actuating signal E.
The corrective effort is made proportional to system
"error"; large errors engender a stronger response
than do small ones. We can vary in a continuous
fashion the energy and/or material sent to the
controlled process.
K. Craig
154
Proportional control exhibits nonzero steady-state

errors for even the least-demanding commands and
disturbances.
Why is this so? Suppose for an initial equilibrium
operating point xc = xv and steady-state error is
zero. Now ask xc to go to a new value xvs. It
takes a different value for the manipulated input U
to reach equilibrium at the new xc. When the
manipulated input U is proportional to the
actuating signal E, a new U can only be achieved
if E is different from zero which requires xc xv;
thus, there must be a steady-state error.
u P (t) K P e(t)
u P (kTs Ts ) K P e(kTs Ts )
K. Craig
155
Integral Control
When a proportional controller can use large loop
gain and preserve good relative stability, system
performance, including those on steady-state error,
may often be met.
However, if difficult process dynamics such as
significant dead times prevent use of large gains,
steady-state error performance may be
unacceptable.
When human process operators notice the existence
of steady-state errors due to changes in desired
value and/or disturbance they can correct for these
by changing the desired value ("set point") or the
controller output bias until the error disappears. This
is called manual reset.
K. Craig
156
Integral control is a means of removing steady-state

errors without the need for manual reset. It is
sometimes called automatic reset.
du(t)
K I e(t)
dt
U(s) K I
E(s)
s
u(t) K I e() d
0
If the value of e(t) is doubled, then the value of u(t)

varies twice as fast.
For e(t) = 0, u(t) remains stationary.
We have seen why proportional control suffers from
steady-state errors. We need a control that can
provide any needed steady output (within its design
range, of course) when its input (system error) is
zero.
K. Craig
157
Proportional vs. Integral Control
Integral control has the undesirable side effects of

reducing response speed and degrading stability.
K. Craig
158
Although integral control is very useful for removing

or reducing steady-state errors, it has the undesirable
side effect of reducing response speed and
degrading stability.
Why? Reduction in speed is most readily seen in the
time domain, where a step input (a sudden change)
to an integrator causes a ramp output, a much more
gradual change.
Stability degradation is most apparent in the
frequency domain (Nyquist Criterion) where the
integrator reduces the phase margin by giving an
additional 90 degrees of phase lag at every
frequency, rotating the (B/E)(i) curve toward the
unstable region near the -1 point.
K. Craig
159
Occasionally an integrating effect will naturally

appear in a system element (actuator, process, etc.)
other than the controller.
These gratuitous integrators can be effective in
reducing steady-state errors. Although controllers
with a single integrator are most common, double
(and occasionally triple) integrators are useful for the
more difficult steady-state error problems, although
they require careful stability augmentation.
Conventionally, the number of integrators between E
and C in the forward path has been called the system
type number.
K. Craig
160
In addition to the number of

integrators, their location (relative
to disturbance injection points)
determines their effectiveness in
removing steady-state errors.
In Figure (a) the integrator gives
zero steady-state error for a step
command but not for a step
disturbance.
By relocating the integrator as in
Figure (b), either or both step
inputs Vs and Us can be
"canceled" by M without requiring
E to be nonzero.
Integrators must be located upstream

from disturbance-injection points if they
are to be effective in removing steadystate errors due to disturbances.
Location is not significant for steadystate errors caused by commands.
K. Craig
161
Integral control can be used by itself or in combination

with other control modes. Proportional + Integral (PI)
Control is the most common mode.
Integral gain provides DC and low-frequency stiffness.
When a DC error occurs, the integral gain will move to
correct it. The higher the gain, the faster the correction.
Fast correction implies a stiffer system.
Dont confuse DC stiffness with dynamic stiffness. A
system can be quite stiff at DC and not stiff at all at high
frequencies! Higher integral gains will provide higher
DC stiffness but will not substantially improve stiffness
at or above the loop bandwidth.
K. Craig
162
PI controllers are more complicated to implement than P

controllers. Saturation becomes more complicated, as
integral wind-up must be avoided. In analog controllers,
clamping diodes must be added, and in digital
controllers, saturation algorithms must be coded.
Integral gain can cause instability. In the open loop, the
integral, with its 90 phase lag, reduces phase margin.
In the time domain, the common result of adding integral
gain is overshoot and ringing. As a result, larger
integral gains usually reduce bandwidth.
K. Craig
163
u I (kTs Ts ) K I
kTs Ts
KI
kTs
e() d
e() d K I
kTs Ts
kTs
e() d
u I (kTs ) K I [area under e( ) over one period]

Ts
u I (kTs ) K I e(kTs Ts ) e(kTs )
2
Graphical Interpretation
of Numerical
Integration:
Area of Trapezoid
K. Craig
164
Derivative Control
Proportional and integral control actions can be used
as the sole effect in a practical controller.
But the various derivative control modes are always
used in combination with some more basic control law.
This is because the derivative mode produces no
corrective effect for any constant error, no matter how
large, and therefore would allow uncontrolled steadystate errors.
One of the most important contributions of derivative
control is in system stability augmentation. If absolute
or relative stability is the problem, a suitable derivative
control mode is often the answer.
The stabilization or "damping" aspect can easily be
understood qualitatively from the following discussion.
K. Craig
165
Invention of integral control may have been

stimulated by the human process operators desire to
automate their task of manual reset. Derivative
control hardware may first have been devised as a
mimicking of human response to changing error
signals. Suppose a human process operator is given
a display of system error E and has the task of
changing manipulated variable M (say with a control
dial) so as to keep E close to zero.
K. Craig
166
If you were the operator, would you produce the same

value of M at t1 as at t2? A proportional controller would
do exactly that.
A stronger corrective effect seems appropriate at t1 and
a lesser one at t2 since at t1 the error E is E1,2 and
increasing, whereas at t2 it is also E1,2 but decreasing.
The human eye and brain senses not only the ordinate
of the curve but also its trend or slope. Slope is clearly
dE/dt, so to mechanize this desirable human response
we need a controller sensitive to error derivative.
Such a control can, however, not be used alone since it
does not oppose steady errors of any size, as at t3, thus
a combination of proportional + derivative control, for
example, makes sense.
K. Craig
167
The relation of the general concept of derivative

control to the specific effect of viscous damping in
mechanical systems can be appreciated from the
figure below.
Here an applied torque T tries to control position of
an inertia J. The damper torque on J behaves
exactly like a derivative control mode in that it always
opposes velocity d/dt with a strength proportional to
d/dt making motion less oscillatory.
K. Craig
168
Derivatives of E, C, and almost any available signal

in the system are candidates for a useful derivative
control mode.
First derivatives are most common and easiest to
implement.
The noise-accentuating characteristics of derivative
operations may often require use of approximate
(low-pass filtered) derivative signals.
Derivative signals can sometimes be realized better
with sensors directly responsive to the desired value,
rather than trying to differentiate an available signal.
In addition to stability augmentation, derivative
modes may also offer improvements in speed of
response and steady-state errors.
K. Craig
169
The derivative gain advances the phase of the loop

by virtue of the 90 phase lead of a derivative. Using
derivative gain will usually allow the system
responsiveness to increase, allowing the bandwidth
to nearly double in some cases.
Derivative gain has high gain at high frequencies. So
while some derivative gain does help the phase
margin, too much hurts the gain margin by adding
gain at the phase crossover frequency, typically a
high frequency. This makes the derivative gain
difficult to tune. The designer sees overshoot
improve because of increased phase margin, but a
high-frequency oscillation, which comes from
reduced gain margin, becomes apparent.
K. Craig
170
Derivatives are also very sensitive to noise. The

derivative gain needs to followed by a low-pass filter
to reduce noise content. However, the lower break
frequency of the filter, the less benefit can be gained
from the derivative gain.
Proportional + Derivative Control
de(t)
u(t) K P e(t) K D
dt
U(s)
K P K Ds
E(s)
Derivative control has an anticipatory character,
however, it can never anticipate any action that has
not yet taken place.
Derivative control amplifies noise signals and may
cause a saturation effect in the actuator.
K. Craig
171
In the derivative term, the roles of u and e are

reversed from integration and a consistent
approximation can be written down at once.
u I (kTs Ts ) u I (kTs ) K I
Ts
e(kTs Ts ) e(kTs )
2
Integration
Ts
u D (kTs Ts ) u D (kTs ) K D e(kTs Ts ) e(kTs )
2
Differentiation
K. Craig
172
Z Operator
The Laplace Transform variable s is a differential
operator. The Z Transform variable z is a prediction
operator or a forward-shift operator.
U(z)
is the transfrom of u(kTs )
zU(z) is the transform of
u(kTs Ts )
Consider the integral term.

Ts
u I (kTs Ts ) u I (kTs ) K I e(kTs Ts ) e(kTs )
2
Ts
zU I (z) U I (z) K I zE(z) E(z)
2
T z 1
U I (z) K I s
E(z)
2 z 1
K. Craig
173
The derivative term is the inverse of the integral term.

2 z 1
U D (z) K D
E(z)
Ts z 1
The complete discrete PID controller is thus
described by:
Ts z 1
2 z 1
U(z) K P K I
KD
E(z)
2 z 1
Ts z 1
The effect of the discrete approximation in the zdomain is as if everywhere in the analog transfer
function the operator s has been replaced by the
composite operator 2 z 1
Ts z 1
2 z 1
The discrete equivalent to Da(s) is Dd (z) Da
T
z
1
s
K. Craig
174
Example Problem
Y
45
U (s 9)(s 5)
U
s6
D s 1.4
E
s
G s
Plant Transfer Function

PI Controller Transfer Function
The closed-loop system has a rise time of about 0.2

seconds and an overshoot of about 20%.
What is the discrete equivalent of this controller?
Compare the step responses and control signals of
the two systems. Consider a sample period of 0.07
seconds (about three samples per rise time) and a
sample period of 0.035 seconds (about 6 samples
per rise time).
K. Craig
175
1.21z 0.79
Dd (z) 1.4
for Ts 0.07
z 1
u(k 1) u(k) 1.4 1.21e(k 1) 0.79e(k)
1.105z 0.895
Dd (z) 1.4
for Ts 0.035
z 1
u(k 1) u(k) 1.4 1.105e(k 1) 0.895e(k)
K. Craig
176
Output Response
Control Signals
K. Craig
177
Pulse Width Modulation

Pulse width modulation (PWM) is a technique in which a
series of digital pulses is used to control an analog
circuit. The length and frequency of these pulses
determines the total power delivered to the circuit. PWM
signals are most commonly used to control DC motors,
but have many other applications ranging from controlling
valves or pumps to adjusting the brightness of an LED.
The digital pulse train that makes up a PWM signal has a
fixed frequency and varies the pulse width to alter the
average power of the signal. The ratio of the pulse width
to the period is referred to as the duty cycle of the signal.
K. Craig
178
For example, if a PWM signal has a 10 ms period

and its pulses are 2 ms long, that signal is said to
have a 20 percent duty cycle.
PWM can be used to reduce the total amount of
power delivered to a load without losses normally
incurred when a power source is limited by resistive
means. This is because the average power delivered
is proportional to the modulation duty cycle. With a
sufficiently high modulation rate, passive electronic
filters can be used to smooth the pulse train and
recover an average analog waveform.
K. Craig
179
High frequency PWM power control systems are easily

realizable with semiconductor switches. The discrete
on/off states of the modulation are used to control the
state of the switch(es) which correspondingly control
the voltage across or current through the load.
The major advantage of this system is the switches
are either off and not conducting any current, or on
and have (ideally) no voltage drop across them. The
product of the current and the voltage at any given
time defines the power dissipated by the switch, thus
(ideally) no power is dissipated by the switch.
Realistically, semiconductor switches are non-ideal
switches, but high efficiency controllers can still be
built.
K. Craig
180
A PWM signal is generated by comparing a triangle wave

signal with a DC signal.
This 3-Op-Amp Circuit produces a triangular wave and a
variable-pulse-width output.
U1A and U1B form a triangle-wave generator. U1B is a
comparator.
K. Craig
181
Waveforms created
by the 3-Op-Amp
Circuit
U1A is configured as an integrator and U1B as a

comparator with hysteresis. At power up, the
comparators output voltage is assumed to be zero.
K. Craig
182
Using PWM to Generate Analog Output

PWM can be used as an inexpensive digital-toanalog (D/A) converter. A wide variety of
microcontroller applications exist that need analog
output but do not require high-resolution D/A
converters.
Conversion of PWM waveforms to analog signals
involves the use of analog low-pass filters.
In a typical PWM signal, the frequency is constant,
but the pulse width (duty cycle) is a variable. The
pulse width is directly proportional to the amplitude of
the original unmodulated signal.
K. Craig
183
Typical PWM Waveform
Frequency Spectrum
of a
PWM Signal
External Low-Pass Filter

FPWM K FBW
K 1
K. Craig
184
Signal BW
4 kHZ
FPWM
PWM
K=5
K FBW 20kHz
dB 20log10
f 20 kHz
1
2f
Choose the -3 dB point

at 4 kHz
1
RC
3.98(105 )
2f
R 4.0 k C 0.01F
RC 1
2
14.2dB
If the -14 dB attenuation will not suffice, a higher-order active lowpass filter may be necessary or a higher PWM frequency.
K. Craig
185
Simulink Simulation
PWM Low -Pass Filter Demonstration
PWM
1
Amplitude
200
Amplitude
Frequency (Hz)
Frequency (hz)
PWM Output
Start Time (s)
1
0.0159s+1
Analog
Start
Transfer Fcn
RC Low-Pass Filter
Duty %
50
PWM
Duty (%)
1/tau = 10 Hz = 62.8 rad/s

PWM Frequency K = 5, 10, 15, 20
K. Craig
186
Response of Low-Pass Filter (1/tau = 10 Hz) to PWM Signal (50% Duty, 50 Hz)
0.7
0.6
amplitude
0.5
0.4
0.3
0.2
0.1
0.05
0.1
0.15
0.2
0.7
0.25
time (sec)
0.6
amplitude
0.5
0.4
0.3
0.2
0.1
0.05
0.1
0.15
0.2
0.25
time (sec)
K. Craig
187
0.7
0.6
amplitude
0.5
0.4
0.3
0.2
0.1
0.05
0.1
0.15
0.2
0.25
0.7
time (sec)
0.6
amplitude
0.5
0.4
0.3
0.2
0.1
0.05
0.1
0.15
0.2
0.25
time (sec)
K. Craig
188
Simulink Simulation
PWM LR Circuit Demonstration
PWM
1
Amplitude
30000
Amplitude
Frequency (Hz)
Frequency (hz)
1/.04
PWM Output
Current_PWM
0.00175s+1
Start Time (s)
Start
Duty %
50
PWM
Duty (%)
Transfer Fcn
LR Circuit
R = 0.04 Ohms
L = 70 micro-henries
1/tau = 91 Hz = 571 rad/s
1/.04
0.00175s+1
Step
50% of 1 V
Current_Step
Transfer Fcn
LR Circuit
R = 0.04 Ohms
L = 70 micro-henries
K. Craig
189
Response of LR Circuit (tau = 0.00175) to Step Input 0.5 volts

14
12
current (amps)
10
0
0
0.002 0.004 0.006 0.008 0.01 0.012 0.014 0.016 0.018

time (sec)
0.02
K. Craig
190
What should the PWM switching frequency be so that the

current waveform is within P% of the step response?
R
I
estep
Ie
T
R
2

L
i Ie
Ie
1
I
100
Rt
L
Ie
Ie
T
R
2

L
T
R
2

L
T
R
2

L
1
I
100
1
I
100
P
2L
P
e
1
T
ln 1
100
R
100
R
f
P
2L ln 1
100
Percentage
Frequency
28.4 kHz
5.6 kHz
10
2.7 kHz
20
1.3 kHz
50
412 Hz
K. Craig
191
Response of LR Circuit (tau = 0.00175) to Step Input 0.5 volts and PWM 2.7 kHz
14
12
current (amps)
10
0
0
0.002 0.004 0.006 0.008 0.01 0.012 0.014 0.016 0.018

time (sec)
0.02
K. Craig
192
Response of LR Circuit (tau = 0.00175) to Step Input 0.5 volts and PWM 28.4 kHz
14
12
current (amps)
10
0
0
0.002 0.004 0.006 0.008 0.01 0.012 0.014 0.016 0.018

time (sec)
0.02
K. Craig
193
Parasitic Effects
Parasitic effects are present in all real-world systems
and are troublesome to account for when the
systems are designed. They are rarely disabling
alone, but are debilitating if not dealt with effectively.
These effects include:
Coulomb Friction
Time Delay
Saturation
Compliance / Resonance
Backlash
Nonlinearity
Noise
Quantization
K. Craig
194
Questions:
Are they significant?
While individually they may not be debilitating, in
combination they might be.
Also, implementing solutions to any of these
effects might exacerbate other effects.
What to do about them?
Approaches:
Ignore them and hope for the best! Murphys Law says
ignore them at your own peril.
Include the parasitic effects that you think may be
troublesome in the truth model of the plant and run
simulations to determine if they are negligible.
If they are not negligible and can adversely affect your
system, you need to do something but what?
K. Craig
195
General Remedies:
Alter the design to reduce the effective loop gain
of the controller, especially at high frequencies
where the effects of parasitics are often
predominant. This generally entails sacrifice in
performance.
Techniques specifically intended to enhance
robustness of the design are also likely to be
effective, but may entail use of a more
complicated control algorithm.
K. Craig
196
Sensor Fusion
When measuring a particular variable, a single type
of sensor for that variable may not be able to meet all
the required performance specifications.
We sometimes combine several sensors into a
measurement system that utilizes the best qualities of
each individual device.
Thus, sensors complement each other, giving rise to
the name complementary filtering. Another name is
sensor fusion and a more advanced version of a
similar idea is called Kalman filtering.
K. Craig
197
Basic Concept
If a time-varying signal is applied to both a low-pass
filter and a high-pass filter, and if the two filter output
signals are summed, the summed output signal is
exactly equal to the input signal.
Low-Pass Filter
1
s 1
qi
s
s 1
qi
High-Pass Filter
K. Craig
198
Bode Diagram
0
Magnitude (dB)
-10
-20
-30
-40
-50
-60
90
Phase (deg)
45
0
-45
-90
-2
10
-1
10
10
10
10
Frequency (rad/sec)
K. Craig
199
K. Craig
200
The high-pass filter and the low-pass filter do not

have to be the simple filters shown. An example of
stronger filters would be:
3s 2 3s 1
Low Pass Filter
3
2
s 3s 3s 1
s3
High Pass Filter
3
2
s 3s 3s 1
Mechatronics Example: Absolute Angle

Measurement
The two basic sensors used are a micro-electromechanical (MEMS) rate gyro using piezoelectric
tuning forks (no spinning wheel) and an inclinometer.
K. Craig
201
The inclinometer measures tilt angle relative to gravity

vertical by immersing two circular sector capacitance
plates in a dielectric liquid. Angular tilting causes one
pair of plates to increase capacitance and the other to
decrease. These capacitance changes cause a
frequency change in an oscillator, which is then
converted to a pulse-width-modulated (PWM) signal.
By low-pass filtering the PWM signal, a DC voltage
proportional to tilt angle is obtained.
A rate gyro gives a DC voltage output proportional to
angular velocity, with a flat frequency response to
about 50 Hz. Op-amp analog integration would give us
angular position, but the bias error in the rate gyro,
when integrated, quickly gives an unacceptable, everincreasing drift of the position signal.
K. Craig
202
The inclinometer does not suffer from a drift problem

(no integration is involved) and can thus be used to
correct for the gyro drift problem. It cannot, however,
be used by itself for angle measurements in
applications that require a fast response (like
measuring vehicle or robotic motions) since it is a
first-order instrument with low bandwidth, typically 0.5
Hz to 6 Hz, too slow for many applications.
The two sensors are thus good candidates for a
complementary-filtering application, giving both
angular position and angular velocity data over about
a 50-Hz bandwidth with negligible drift.
K. Craig
203
While the configuration of the separate high-pass and

low-pass filters is most useful for explaining the basic
concept of complementary filtering, the practical
implementation uses instead a feedback type of
configuration that produces identical differential
equations and transfer functions.
Also, realistic sensor models should be used for
analysis and simulation purposes. The inclinometer is
modeled as a first-order system (e.g., Ki =1, time
constant = 0.3). The rate gyro is modeled as a secondorder system (e.g., Kg = 1, damping ratio = 0.5, and
natural frequency = 50 Hz).
sensor
Ki
Inclinometer
actual i s 1
K gs
sensor
2
Rate Gyro
2 g s
s
actual
1
2
ng ng
K. Craig
204
The gyro bias error is taken as a constant (e.g., 0.005

rad/s) and the inclinometer noise is taken as a small
random signal.
The complementary filter has two adjustments: n
which we take to be 0.2 rad/s and which we take to
be 0.7. The major effect is that of n; larger values
correct bias effects more quickly but filter noise
effects less effectively.
To test out this algorithm, we will take the input angle
to be zero for the first 20 seconds to see how the
system fights out the gyro bias and attenuates the
inclinometer noise. At 20 seconds, the input angle
steps up to 1.0 radian, so we can see the response to
sudden changes.
K. Craig
205
K. Craig
206
Analyzing this block diagram results in the following

equation:
s2
2s
1
2
n
n
m 2
inc inc _ n
rg
rg _ b
2
s
2s
s
2 s
1
2
2
n n
n n
High-Pass Filter
Low-Pass Filter
This is how the Watson Vertical Reference System is
implemented. The description of that system is
shown on the next page.
K. Craig
207
K. Craig
208
Simulink Simulation
gyro_bias
Sensor Fusion
1/s
Constant
1/s
theta_rg
Uncorrected
Gyro Angle
Num_rg(s)
theta_m
Den_rg(s)
Rate Gyro TF
Measured
Angle
1/s
Step
Input
Angle
2*zeta*omega_n
(omega_n)^2
Manual Switch
Sine Wave
Num_inc(s)
Den_inc(s)
theta_actual
Inclinometer TF
theta_inc
Noisy
Inclinometer
Angle
Actual
Angle
Band-Limited
White Noise
K. Craig
209
Comparison: Integrated Rate-Gyro Angle and Corrected Angle (Noisy Inclinometer Angle Not Shown)
1.4
1.2
angle (radians)
1
0.8
0.6
Corrected Angle
Uncorrected Angle
0.4
0.2
0
-0.2
10
15
20
25
time (sec)
30
35
40
45
50
K. Craig
210
Observers
Controls and Sensors
Sensors measure the quantity under control.
We often assume the availability of near-perfect
feedback signals. This assumption is often
invalid.
Four Common Problems Caused By Sensors

Sensors and associated cabling are expensive.
Sensors plus associated cabling reduce control
system reliability.
K. Craig
211
Some signals are impractical to measure or are

inaccessible.
Sensors usually induce significant errors, e.g.,
noise, cyclical errors, limited responsiveness.
Observers
Augment or replace sensors in a control system.
They are algorithms that combine sensed signals
with other knowledge of the control system to
produce observed signals which can be more
accurate, less expensive to produce, and more
reliable than sensed signals.
They are an alternative to adding new sensors or
upgrading existing ones.
K. Craig
212
Observed Signals, compared to sensed signals, can be:

More accurate
Less expensive to produce
More reliable
Observer Advantages:
Remove sensors which reduces cost and improves
reliability.
Improve the quality of signals that come from sensors
allowing performance enhancement.
Observer Disadvantages:
Can be complicated to implement.
Expend computational resources.
Because observers form software control loops, they
can become unstable under certain conditions.
K. Craig
213
Design Issues
Select observer technique for a given system
How much will observer improve performance?
Additional cost?
Limitations of observers?
Will observer be useful and what are required
resources?
Implementation Issues
Installation of observers
Tune an observer
Recognize effects of changing system parameters
on observer performance
K. Craig
214
Observers can combine knowledge of plant, power

converter device output, and feedback device to
extract a feedback signal that is superior to that
which can be obtained by using a feedback device
alone.
Principle of an Observer
By combining a measured feedback signal with
knowledge of the control-system components
(plant + feedback system) the behavior of the
plant can be known with greater accuracy and
precision than by using the feedback alone.
K. Craig
215
Role of an Observer in a Control System

Observer augments sensor output and provides a
feedback signal.
K. Craig
216
Observer can be used to enhance system performance

More accurate than sensor
Reduce phase lag inherent in sensor
Can provide observed disturbance signals which can
be used to improve disturbance response
Can reduce system cost by augmenting performance of
low-cost sensor
Can eliminate a sensor altogether
Observers
Not a panacea
Add complexity to a system
Requires computational resources
May be less robust than physical sensors especially
when parameters change substantially during operation
K. Craig
217
An observer is a mathematical structure that

combines sensor output and plant excitation
signals with models of the plant and sensor.
Physical System
+_
Controller
Power
Converter
PC
Plant
Sensor
Y
+
Observer
Controller
Modeled System
Plant
Model
CO
Sensor
Model
YO
Observer
K. Craig
218
An observer provides feedback signals that are

superior to the sensor output alone.
There are 5 Elements of an Observer:
Sensor output
Power converter output (plant excitation)
Model (estimation) of the plant
Model of the sensor
PI or PID observer compensator
K. Craig
219
Key Guidelines for using an Observer in a Motion

System
Performance Requirements
Machines that demand rapid response to command
changes, stiff response to disturbances, or both will
likely benefit from an observer.
The observer can reduce phase lag in the servo
loop, allowing higher gains, which improve command
and disturbance response.
Available Computational Resources
Observers almost universally rely on digital control.
If the actual or planned control system is executed
on a high-speed processor, an observer can be
added without significant cost burden.
K. Craig
220
If digital control techniques are already

employed, the additional design effort to
implement an observer is relatively small.
Controls Expertise in the User Base
Observers require some level of controls
expertise for installation and configuration.
The user base must be capable of
understanding the features of an observer if it
is to provide benefit.
Sensor Noise
Observers are most effective when the
position sensor produces limited noise.
Sensor noise is often a problem in motioncontrol systems.
K. Craig
221
Noise in servo systems comes from two major

sources: EMI generated by power converters and
transmitted to the control section of the servo
system, and resolution limitations in sensors,
especially in the feedback sensor.
EMI can be reduced through appropriate wiring
practices and through the selection of
components that limit noise generation.
Noise form sensors is difficult to deal with as
observers often exacerbate sensor-noise
problems. Lowering observer bandwidth will
reduce noise susceptibility, but it also reduces the
ability of the observer to improve the system, e.g.,
reducing observer bandwidth reduces the
accuracy of the observed disturbance signal.
K. Craig
222
The availability of high-resolution feedback sensors

raises the likelihood that an observer will
substantially improve system performance.
Phase Lag in Motion Control Systems
The two predominant sensors in motion-control
systems are incremental encoders and resolvers.
Incremental encoders respond to position change
without substantial phase lag.
Resolver signals are commonly processed with a
tracking loop, which generates substantial phase
lag, their presence makes it more likely that an
observer will substantially improve system
performance. The same applies to a sine encoder.
K. Craig
223
Independent of a feedback sensor, most motioncontrol systems generate phase lag in the control
loop when they derive velocity from position.
Velocity is commonly derived from position using
simple differences. It is well known to inject a
phase lag of half the sample time. The phase lag
also provides an opportunity for the observer to
improve system performance.
Summary: Five Key Guidelines for using an Observer
in a Motion System
Need for high performance in the application
Availability of computational resources in the
controller
K. Craig
224
Ability of the average user to install and configure

the system
Availability of highly resolved position feedback
signal
Presence of phase lag in the position or velocity
feedback signals
The more of these guidelines that an application
meets, the more likely the observer can substantially
improve system performance.
K. Craig
225
Adaptive Control
Normal Control Design Procedure
Develop a hierarchy of models (both physical and
mathematical) of the plant, ranging from a truth model, the
most realistic model developed, to a design model, one
simple enough for design purposes while still capturing the
essential characteristics of the actual system.
Validate the models through comparisons of predicted
responses with actual measured responses.
Design a controller on the basis of the plant design model.
Test the control design by simulation on both the design
model and the truth model.
Implement the control design on the actual plant.
Tune the controller after installation.
K. Craig
226
Robust Control System

When we use a model of the plant as the basis of a control
system design, we are assuming tacitly that this model is a
reasonable representation of the plant.
Although the design model always differs from the true
plant in some details, we are confident that these details
are not important enough to invalidate the design.
Also, ordinary feedback attempts to reduce the effects of
plant uncertainty and disturbances.
A two-degrees-of-freedom control system (feedforward +
feedback) results in a robust control system where the
following key issues can be addressed effectively.
command tracking and disturbance rejection

insensitivity to modeling errors and to sensor noise
insensitivity to unmodeled high frequency dynamics
stability margins
K. Craig
227
GC2(s)
R(s)
E(s)
+
GC1(s)
D(s)
+
GP(s)
Y(s)
U(s)
+
B(s)
H(s)
Two-Degrees-of-Freedom
Control System
N(s)
K. Craig
228
Need for Adaptation

There are many applications, however, for which
a design model cannot be developed with any
reasonable degree of confidence. For example:
Processes in which the underlying physical
principles are not understood well enough for
physical and mathematical modeling.
Processes for which the physical principles are
understood but which have parameters that cannot
be measured or accurately estimated.
Moreover, most dynamic processes change with

time. Parameters may vary because of plant
load, normal wear, aging, breakdown, and
changes in the environment in which the process
operates.
K. Craig
229
For example, the mass of the object being moved by a

robot manipulator will have a considerable effect on the
dynamics of the closed-loop system, and will mean that
a controller which is well tuned for an intermediate value
of the mass will be less well tuned if an extreme value is
used, and may even result in an unstable system.
The feedback mechanism provides some degree of

immunity to discrepancies between the physical plant
and the model that is used for the design of the control
system.
But sometimes that is not enough. A control system
designed on the basis of a nominal design model may
not behave as well as expected, because the design
model does not adequately represent the process in
its operating environment.
K. Craig
230
How can one deal with processes that are prone to

large changes, or for which adequate design models
are not available?
One approach is brute force, i.e., high loop gain: as the
loop gain becomes infinite, the output of the process
tracks the input with vanishing error. Brute force rarely
works, however, for well-known reasons: dynamic
instability, control saturation, and susceptibility to noise
and other extraneous inputs.
Modern robust control design techniques tolerate
substantial variation in one or more parameters, but
often with a sacrifice in performance. Moreover, these
techniques are not readily applicable to processes for
which no design model is available.
K. Craig
231
Adaptive Control
Adaptive control may provide a solution to the
problem. The basic idea is to have the control law
adapt its own behavior, as it learns about the
process it is designed to control, or as the process
changes with its environment.
An adaptive controller is a controller that can
modify its behavior in response to changes in
process dynamics and disturbance
characteristics; it is a controller with adjustable
parameters and a mechanism for adjusting the
parameters.
The controller becomes nonlinear because of the
parameter-adjustment mechanism, but with a
special structure.
K. Craig
232
Parameter
Adjustment
Controller Parameters
Command
Signal
Controller
Control
Signal
Plant
Output
Adaptive Control System
An adaptive control system can be thought of as having

two loops:
normal feedback loop with the process and the
controller
parameter-adjustment loop, which is often slower than
the normal feedback loop
K. Craig
233
Note that there are a number of solutions to the

problem of keeping a controller in tune, as the
parameters of the system it is controlling vary.
They range from common-sense approaches to
much more mathematical ones.
The extra complexity of the more mathematical
approaches is often justified by lesser hardware
requirements and more reliable operation.
However, it is very difficult to prove the stability
properties of controllers whose parameters can
vary as time passes, therefore only fairly
restricted adaptation may be allowed in some
applications.
K. Craig
234
Two basic viewpoints on adaptive control have

emerged over the years.
The first assumes that a design model is available,
but that the parameters of the model are either not
known or subject to a wide variation.
Gain Scheduling
Self Tuning
The second assumes that no such design model is
available.
Model Reference
Direct Methods (gain scheduling and modelreference) change control parameters directly while
Indirect Methods (self tuning) change control
parameters based on a solution to a design problem.
K. Craig
235
First Viewpoint: Design Model is Available

The control system design depends explicitly upon the
parameters that are subject to variation, and then uses
measurements or estimates of these parameters as inputs
to the control system for the purpose of tuning it during the
course of its operation.
If the parameters can be measured directly, the control
system gains can be scheduled for these measured
parameters. This approach is often called gain scheduling.
If it is not possible to measure the uncertain parameters
directly, the relationships between these parameters and
the other measurable quantities in the system might
feasibly be exploited to estimate the parameters. A design
based on estimates of parameters may be called a selftuning controller, since it uses its own operating data to
estimate parameters in real time and thereby tune its own
behavior.
K. Craig
236
Gain
Schedule
Controller
Parameters
Operating
Condition
Command
Signal
Controller
Control
Signal
Plant
Output

Gain Scheduling
In gain scheduling there is an inner loop composed of the process and
the controller and an outer loop that adjusts the controller parameters on
the basis of the operating conditions. Gain scheduling can be regarded
as a mapping from process parameters to controller parameters. It can
be implemented as a function or a table lookup. Gain scheduling is thus
a very useful technique for reducing the effects of parameter variations.
K. Craig
237
Process
Parameters
Specification
Controller
Design
Command
Signal
Estimation
Controller
Control
Signal
Plant
Output

Self-Tuning
In self-tuning, the inner loop consists of the process and an ordinary
feedback controller. The parameters of the controller are adjusted in real
time by the outer loop, which is composed of an estimator and a design
calculation. The process model and the control design are updated at each
sampling period. The block labeled controller design represents an online solution to a design problem for a system with known parameters.
K. Craig
238
Second Viewpoint: Design Model is Unavailable

If the process is not amenable to being modeled with any
reasonable accuracy, the control law must be designed
with little or no knowledge of the process. Modelreference adaptive control (MRAC) takes this approach.
Known only is how the closed-loop system is required to
behave in response to the command signal. The desired
behavior is represented by an ideal or reference model.
The key problem with MRAC is to determine the
adjustment mechanism so that a stable system, which
brings the performance error to zero, is obtained. This
parameter vector is typically a set of controller gains and
has nothing to do with the physical parameters of the
process being controlled. In principle, the operation of
the tuner is indifferent to the dynamics of the plant.
K. Craig
239
Model
Adjustment
Mechanism
Command
Signal
Controller
Control
Signal
Plant
Output

Model-Reference Adaptive
In MRAC there is an inner feedback loop composed of the process and the
controller and an outer loop that adjusts the controller parameters. The reference
input is applied to both the real closed-loop plant and the ideal model. If the
performance error (the difference between the process output and the model
output) is sufficiently small, the closed-loop system is operating as desired and left
alone. But if the error is appreciable, it becomes the input to an adjustment
mechanism that adjusts controller parameters.
K. Craig
240
Summary
In adaptive control, the process is controlled by a controller
that has adjustable gains. It is assumed that there exists
some kind of design procedure that makes it possible to
determine a controller that satisfies some design criteria if
the process and its environment are known. This is called
the underlying design problem. The adaptive control
problem is then to find a method of adjusting the controller
when the characteristics of the process and its environment
are unknown or changing.
Gain scheduling and model-reference adaptive control are
called direct methods, because the controller parameters
are changed directly without the characteristics of the
process and its disturbances first being determined. The
self-tuning controller is called an indirect method, as the
controller parameters are obtained from a solution to a
design problem using the estimated process parameters
and possibly disturbance characteristics.
K. Craig
241
The construction of an adaptive controller thus contains

the following steps:
Characterize the desired behavior of the closedloop system.
Determine a suitable control law with adjustable
parameters.
Find a mechanism for adjusting the parameters.
Implement the control law.
The key factors in choosing adaptive control are:
Variations in process dynamics
Variations in the character of the disturbances
Engineering efficiency and ease of use
Use of an adaptive controller will not replace good

process knowledge, which is still needed to choose
specifications, the structure of the controller, and the
design method.
K. Craig
242
Fuzzy Logic Control

Intelligent control is the discipline in which control
algorithms are developed by emulating certain
characteristics of intelligent biological systems. The
emergence of intelligent control has been fueled by
advancements in computing technology. Examples
include:
Expert Systems, computer programs that emulate
the actions of a human who is proficient at some
task, are being used to construct expert
controllers that seek to automate the actions of a
human operator who controls a system.
K. Craig
243
Fuzzy Systems, rule-based systems that use fuzzy

logic for knowledge representation and inference,
are being used to automate the perceptual,
cognitive, and action-taking characteristics of
humans who perform control tasks.
Artificial Neural Networks emulate biological neural
networks and have been used to learn how to control
systems by observing the way that a human
performs a control task and to learn in an on-line
fashion how best to control a system by taking
control actions, rating the quality of the responses
achieved when these actions are used, and then
adjusting the recipe used for generating control
actions so that the response of the system improves.
K. Craig
244
Genetic Algorithms, either on line or off line, are

being used to evolve controllers by maintaining a
population of controllers and using survival of the
fittest principles where fittest is defined by the
quality of the response achieved by the controller.
The trend in the field of control is to integrate the
functions of intelligent systems with conventional control
systems to form highly autonomous systems that have
the capability to perform complex control tasks
independently with a high degree of success.
These intelligent controllers are not mystical; they are
simply nonlinear, often adaptive controllers, and there
seems to be an existing conventional control approach
that is analogous to every new intelligent control
approach that has been introduced.
K. Craig
245
Fuzzy Control
A fuzzy controller can be designed to roughly
emulate the human deductive process (i.e., the
process whereby we successfully infer
conclusions from our knowledge). As shown in
the figure on the next slide, the fuzzy controller
consists of four main parts.
The rule base holds a set of if-then rules that
are quantified via fuzzy logic and used to
represent the knowledge that human experts
may have about how to solve a problem in
their domain of expertise.
K. Craig
246
Reference
Input
r(t)
Fuzzification
Fuzzy Controller
Inference
Mechanism
Inputs
Process
Rule
Base
u(t)
Outputs
y(t)
Fuzzy Controller Architecture

K. Craig
247
The fuzzy inference mechanism successively decides

what rules are the most relevant to the current
situation and applies the actions indicted by the rules.
The fuzzification interface converts numeric inputs
into a form that the fuzzy inference mechanism can
use to determine which knowledge in the rule base is
most relevant at the current time.
The defuzzification interface combines the
conclusions reached by the fuzzy inference
mechanism and provides a numeric value as an
output.
Overall, the fuzzy control design methodology, which
primarily involves the specification of the rule base,
provides a heuristic technique to construct nonlinear
controllers.
K. Craig
248
Trade-Offs & Performance Limitations

Feedback control is at the heart of every mechatronic system.
Changes cannot be effected instantaneously in a dynamical
system, and a correct control decision applied at the wrong
time could result in catastrophe.
Control systems must be safe and robust, and guaranteed to
be so, before any thoughts of performance are considered.
Robustness achieved through feedback control is subject to
limits; there is a robustness tradeoff present in all feedback
systems.
An understanding of fundamental limitations, in practical,
physical terms rather than abstract, mathematical terms, is an
essential element in all engineering.
K. Craig
249
The ability to identify performance trade-offs and

fundamental performance limitations (at both the
component and system level), their sources, and quantify
their impact on performance is essential in mechatronics
design and control.
The ability of most mechatronic systems to deliver
exceptional performance relies mainly on the dynamic
interaction between its components and on the
performance of its control system.
Mechatronic systems are typically composed of
components such as sensors, actuators, communication
hardware, electronics, and signal conversion hardware.
Each of these components or subsystems have their own
performance characteristics, trade-offs, and limitations that
would impact the overall performance of the system.
K. Craig
250
D(s)
V(s)
+
K(s)
K(s)G(s)
C(s)
T(s)
1 K(s)G(s) V(s)
1
C(s)
S(s)
1 K(s)G(s) D(s)
G(s)
C(s)
T(s) S(s) 1
K. Craig
251
Fundamental Limitations in Feedback Control

Systems
These fundamental limitations typically take three
forms.
Some are in the form of algebraic equations, e.g.,
T(s) + S(s) = 1 holds at all frequencies, where
T(s) and S(s) are the complementary and
sensitivity transfer functions, respectively.
This relation can be regarded as a constraint
on design, preventing independent choices
being made in regard to command-following
and disturbance-rejection performance.
K. Craig
252
Other constraints take the form of a frequency-domain

integral on a closed-loop transfer function such as S(s).
Hendrik Bode observed that there is a fundamental
limitation on the achievable sensitivity function S(s)
(sensitivity to disturbances and modeling errors) for a
feedback system.
The log of the magnitude of the sensitivity function of a LTI
(linear, time-invariant), SISO (single-input, single-output)
feedback system, integrated over frequency, is conserved
under the action of feedback it is zero for stable plant /
compensator pairs and is some fixed positive value for
unstable ones.
log S(i) d 0
Sensitivity improvements in one frequency range must be

paid for with sensitivity deteriorations in another frequency
range, and the price is higher if the plant is open-loop
unstable.
K. Craig
253
Sensitivity reduction at low frequency unavoidably leads

to sensitivity increase at higher frequencies.
K. Craig
254
However, physical systems do not exhibit good

frequency response fidelity beyond a certain
bandwidth. This is due to uncertain or unmodeled
dynamics in the plant, to digital control
implementations, to power limits, and to
nonlinearities, for example. So the integration is
performed over a finite frequency range, a
constraint imposed by the physical hardware we
use in the control loop. All the action of the
feedback design, the sensitivity improvements as
well as the sensitivity deteriorations, must occur
within a finite frequency range. So reducing the
sensitivity of a system to disturbances at one range
of frequencies by feedback control will amplify
transients and oscillations at other frequencies.
K. Craig
255
Example: Stick-Balancing (Inverted Pendulum

Problem)
The difference between balancing long sticks and
short sticks has to do with the location of the
unstable mode. As gets smaller, the unstable
pole magnitude gets larger.
1 2
s
2
s
m 2 Mm
F
2 M
s s
g
2
3 12
g
s2
X
3
2
s
m 2 Mm
F
2 M
s s
g
2
3 12
K. Craig
256
A long rigid stick is easy to balance, but as it

becomes shorter, balancing it becomes more
difficult. Handicaps to human control include
reaction time, neuromuscular lags, limb
inertias, and other uncertainties. So while the
stick is rigid, the compensator frequency range
might be good for a frequency range up to
about 2 Hz. The control strategy is to keep the
sensitivity as small as possible over that
range. However, a dramatic increase in
sensitivity occurs as the stick becomes shorter
and even minor imperfections in the
implementation will cause instability. This is
the reason humans have trouble balancing
short sticks.
K. Craig
257
The third type takes the form of time-domain

integral constraints on a system signal such as
the feedback error.
If the open-loop transfer function has two
poles at the origin (a double integrator) and
the closed-loop system is stable, then the error
e(t) following the application of a unit step
input applied at t = 0 must satisfy the relation:
e(t)dt 0
Equal areas of positive and negative error

must result.
K. Craig
258

Lecture - Control System Design

Încărcat de

Informații document

Titlu original

Drepturi de autor

Formate disponibile

Partajați acest document

Partajați sau inserați document

Opțiuni de partajare

Vi se pare util acest document?

Este necorespunzător acest conținut?

Drepturi de autor:

Formate disponibile

Lecture - Control System Design

Încărcat de

Drepturi de autor:

Formate disponibile

Control System Design:

Computer Embedded Controls

Mechatronic System Design

Control System Design Introduction

Control System Design Overview

Modern Control Design (state-space analysis and

The aim of both techniques is to find a compensation

Feedback Control System

Control System Design Introduction

Control System Design Introduction

Input / System / Output

Control System Design Introduction

Input some agency which can cause a system to respond.

External inputs can be broadly classified as deterministic or

Random input models are the most realistic input models

Step and Sine Inputs

It is thus much more common to base

Transfer Functions & Analogies

Newtons 2nd Law

Using the differential operator D we can transform the

(MD 2 BD K)x F(t)

The transfer function is a property of a system itself,

q = volume flow rate

Simulation Block Diagram

1st Order System Block Diagram

Control System Design Introduction

The Three Basic Element Input-Output Relationships

Control System Design Introduction

Control System Design Introduction

Step Response and Impulse Response

The impulse function is explained by the figure,

Control System Design Introduction

Control System Design Introduction

Control System Design Introduction

Control System Design Introduction

q out q out initial

Control System Design Introduction

Electrical Mechanical Analogies

Force causes velocity, just as voltage causes current.

The product (f)(v) represents instantaneous mechanical

Control System Design Introduction

Control System Design Introduction

Spring-Damper Mechanical System

Mass-Damper Mechanical System

Control System Design Introduction

LRC Electrical System

Control System Design Introduction

For enhanced multidisciplinary engineering system design and

Inductor-Capacitor (LC) Mass-Spring (MK) Oscillations

Control System Design Introduction

Control System Design Introduction

Poles and Zeros of Transfer Functions

Colocated Control System

Control System Design Introduction

G0(s) Colocated System

Rigid Body Mode

Control System Design Introduction

Colocated Transfer Function

F(s) (Ms 2 3K)(Ms 2 K)

Complex Conjugate Poles

Complex Conjugate Zeros

Control System Design Introduction

Control System Design Introduction