FLUX AND SPEED ESTIMATION TECHNIQUES FOR

SENSORLESS CONTROL OF INDUCTION MOTORS

DISSERTATION


Presented in Partial Fulfillment of the Requirements for

The Degree Doctor of Philosophy in the

Graduate School of The Ohio State University


By

Mihai Comanescu, M.S.

* * * * *

The Ohio State University

2005


Disertation Committee: Approved by:

Dr. Longya Xu, Adviser ____________________
Adviser
Dr. Vadim Utkin Graduate Program in Electrical
Engineering
Dr. Donald G. Kasten



ABSTRACT



The focus of this research is the development of novel techniques for estimation
and control of sensorless induction motor drives. The concepts of Direct Field
Orientation (DFO) and Indirect Field Orientation (IFO) related to vector control are
explained, however, only DFO has been used. The V/Hz method is briefly reviewed and
the field-oriented sensorless scheme for speed control is presented.
In a sensorless drive, the speed must be estimated from the system measurements.
Depending on the objective of the control (speed or torque control), the speed estimate
must be used in one or more areas of the control scheme. This idea and the main
techniques for speed estimation are explored.
The dissertation investigates the issues related to low-speed flux estimation when
a Voltage Model observer is used. Pure integration cannot be implemented due to offsets
in the measured signals and integrators must be replaced by low pass filters. At low
speed, the flux estimates are incorrect in both magnitude and angle; consequently, the
rotor position obtained by the DFO method is incorrect. An improved Voltage Model
observer that corrects the errors is developed based on a Programmable Low Pass Filter
ii
and a vector rotator. The method requires estimation of the stator frequency and this is
done by a Phase Locked Loop synchronized with the voltage vector.
The traditional rotor flux MRAS method can be used for speed estimation,
however, under non-ideal integration the dynamics of the speed estimate exhibits right-
hand side plane zeros. Additionally, system tuning is difficult and may yield under
damped responses. Two novel Sliding Mode MRAS observers are designed and
implemented and their features are used for speed estimation. Simulations and
implementation show that they yield accurate estimates, good dynamic behavior and
convenient design.
The d-q rotational frame currents of an induction machine are not decoupled.
Decoupling can be achieved by canceling the cross-coupled terms in the equations of the
synchronous frame currents. This approach is both inconvenient and inaccurate.
A novel approach for decoupling is presented: an Integral Sliding Mode controller
complements a traditional controller that acts on a simulated plant. The use of the Integral
SM controller guarantees that the currents in the real motor will track those of the
simulated model. The additional controller compensates for the cross-terms and for
variations of the machine parameters. The method is also valuable for allowing fast and
efficient tuning of the current controllers.

iii










Dedicated to Alex and Florina





























iv




ACKNOWLEDGMENTS



First, I must express my thanks to The Ohio State University, The Graduate
School and The Department of Electrical Engineering for having me as a graduate
student. I have spent six enjoyable years that have made me a better person and a better
engineer.
Second, I thank my adviser, Dr. Longya Xu, for conducting my education, for the
intellectual support, enthusiasm and patience. I have always felt encouraged by his
friendly attitude and appreciation of my efforts. In the future, I hope we will have the
chance to continue and develop our relationship.
I need to thank Dr. Vadim Utkin for his teaching and for taking the time to
discuss with me on several topics.
Many thanks to Dr. Stephen Sebo and Dr. Donald Kasten for their suggestions on
this material.
I thank the members of The Power Electronics and Electric Machines research
group, especially to Mr. Jinchuan Li and Mongkol Konghirun for getting me started with
the DSP software.
Finally, I am thankful to my friend Bogdan Proca for the patience to explain to me
what he already knew.
v





VITA


August 6, 1968.Born, Bucharest, Romania
May 1992.Politehnica University, Bucharest, Romania
Bachelor of Science in Electrical Engineering
September 1992-September 1993Energy Research and Modernizing Institute,
Research Engineer

September 1993 September 1996 .Cet Vest Power Plant, Engineer

September 1996 - July 1999 .. Time&Co International, Engineer

September 1999 June 2000 The Ohio State University, Graduate Teaching
and Research Associate

June 2000 September 2000..Delphi Automotive Systems, Intern

June 2001 September 2001..Solidstate Controls Inc, Intern

June 2002 September 2002..Solidstate Controls Inc, Intern

June 2004 September 2004..Smiths Aerospace, Intern



PUBLICATIONS


M. Comanescu, L.Xu: A Modified Voltage Model Observer with PLL-based
Frequency Estimator for Sensorless Control of Induction Motors. IEEE
Transactions on Industrial Electronics, Special Issue on Sensorless Control, 2004.
M. Comanescu, A. Keyhani, M.Dai: Design and Analysis of 42V Permanent Magnet
Generator for Automotive Applications. IEEE Transactions on Energy Conversion,
2003.
vi
M. Comanescu, A. Keyhani: Design of Permanent Magnet Synchronous Generator
for Automotive Applications, Mechatronics 2000 Conference, Georgia Institute of
Technology, 2000.



FIELDS OF STUDY


Major Field: Electrical Engineering
Specialization: Power Electronics, Electric Machines and Control Systems


vii



TABLE OF CONTENTS



Page

Abstract......ii

Dedication.......iv

Acknowledgments...v

Vita.vi

List of Figures.x

Nomenclature....xiii

Abbreviations.xv

1. Introduction..1

2. Induction Machine Dynamics and Control......4
2.1 Three-Phase Quantities, Space Vector Definition and Transformations...4
2.2 Dynamic Equations of the Induction Machine..8
2.3 Induction Machine Control......13
2.3.1 V/Hz Control..13
2.3.2 Field-Oriented or Vector Control...15
2.3.3 Direct Field Orientation (DFO).....16
2.3.4 Indirect Field Orientation (IFO).....17
2.4 Experimental Setup....20

3. Sensorless Control and Issues at Low-Speed Operation22
3.1 Overview of Induction Machine Sensorless Control.23
3.2 Speed Estimation Methods ...24
3.3 Low-Speed Sensorless Control..30
3.3.1 Reasons for instability at low-speed..30
viii
3.3.2 Limits of stable operation at low-speed.32

4. An Improved Flux Observer Based on PLL Frequency Estimator for Induction
Machine Sensorless Control .37
4.1 The Voltage Model Observer.....37
4.2 The Effect of Incorrect Field Orientation on the IM Drive with Constant
Field Current......42
4.3 A Programmable Low Pass Filter Based on PLL Frequency Estimation..45
4.3.1 Description of the Programmable LPF..46
4.3.2 Description of the Phase Locked Loop (PLL)...49
4.4 Control System Description...51
4.5 Simulation Results.....53
4.6 Experimental Results.....56
4.7 Conclusions........62

5. Classic and Sliding Mode MRAS Speed Estimators for Sensorless Induction
Motor Control ... 63
5.1 Dynamics of Classical MRAS.......64
5.2 MRAS Single-Manifold Sliding Mode Observer..69
5.3 MRAS Single-Manifold Discrete Time Sliding Mode Observer..74
5.4 MRAS Double-Manifold Sliding Mode Observer77
5.5 Experimental Results.81
5.6 Conclusions86

6. Robust Current Control by Integral Sliding Mode87
6.1 Problem Description..88
6.2 Decoupled Current Control by Integral Sliding Mode..93
6.3 Simulation Results.96
6.4 Experimental Results.99
6.5 Conclusions..102

7. Conclusions..103
7.1 Summary..103
7.2 Future Research Suggestions...105

Bibliography106
ix



LIST OF FIGURES



Page

Figure 2.1 Equivalence of space vector and abc components..5
Figure 2.2 Stationary reference frame ()...6
Figure 2.3 Rotating reference frame (dq)..7
Figure 2.4 Induction machine d equivalent circuit in arbitrary reference frame.10
Figure 2.5 Induction machine q equivalent circuit in arbitrary reference frame.10
Figure 2.6 V/Hz control scheme..14
Figure 2.7 Principle of Direct Field Orientation (DFO)..16
Figure 2.8 Block diagram of sensorless DFO induction motor control..18
Figure 2.9 Block diagram of the experimental setup..21
Figure 3.1 Speed estimation by rotor flux MRAS method.27
Figure 4.1 Real and VM estimated fluxes and rotor angle.41
Figure 4.2 Torque production under incorrect field orientation.43
Figure 4.3 Command current under incorrect rotor position.44
*
q
i
Figure 4.4 Stator current under incorrect rotor position.45
Figure 4.5 Block diagram of the programmable LPF.46
x
Figure 4.6 Locus of vector as a function of k.47
1
s
Figure 4.7 Proposed PLL for frequency estimation in IM drives...50
Figure 4.8 Structure of the control system for study of the PLL based observer52
Figure 4.9 Real and estimated voltage angle at startup..53
Figure 4.10 Estimated frequency and V
q
at startup..55
Figure 4.11 Real and estimated rotor fluxes.55
Figure 4.12 Command and estimated frequency, 9-27-9 Hz step57
Figure 4.13 Command and estimated frequency, 9-45-9 Hz step57
Figure 4.14 Back-EMF and stator flux by programmable LPF, f=18 Hz, k=1.59
Figure 4.15 Back-EMF and stator fluxes by VM observer and PLL programmable
LPF, f=2.1 Hz, k=159
Figure 4.16 Stator flux estimation from VM to PLL programmable LPF during
transition 60
Figure 4.17 Stator fluxes during 100-700 rpm step change..61
Figure 5.1 Speed estimation dynamics of classical MRAS under ideal integration...65
Figure 5.2 Root locus of sped estimation for classical MRAS under ideal
integration ....65
Figure 5.3 Real and MRAS estimated speed under non-ideal integration..68
Figure 5.4 Real and MRAS estimated fluxes under non-ideal integration.68
Figure 5.5 Real and SM MRAS estimated speed under ideal integration..72
Figure 5.6 Real and SM MRAS estimated speed under non-ideal integration...73
Figure 5.7 VM and SM MRAS estimated fluxes under non-ideal integration73
Figure 5.8 Real and DTSM MRAS estimated speed under non-ideal integration..75
xi
Figure 5.9 Real and DTSM MRAS estimated speed under non-ideal integration..76
Figure 5.10 Real and estimated speed by two-dimensional SM MRAS under non-ideal
integration..79
Figure 5.11 Reference and estimated fluxes by two-dimensional SM MRAS under
non- ideal integration.79
Figure 5.12 Equivalent controls ..80
eq eq

,
Figure 5.13 Fluxes of single manifold SM MRAS...82
Figure 5.14 Fluxes and speed at acceleration; single manifold SM MRAS..83
Figure 5.15 Fluxes and speed at deceleration; single manifold SMMRAS..83
Figure 5.16 Fluxes of double manifold SM MRAS..84
Figure 5.17 Equivalent controls for double manifold SM MRAS....85
Figure 5.18 Fluxes and speed of double manifold SM MRAS...85
Figure 6.1 Scheme for induction machine decoupled current control......91
Figure 6.2 ISM controller structure for the d axis current....95
Figure 6.3 Dynamics of d-q currents with no compensation voltages......97
Figure 6.4 Real and model currents with ISM controllers....98
Figure 6.5 Real and model currents with ISM controllers under detuning...99
Figure 6.6 Real currents with no compensation voltages...100
Figure 6.7 Real currents under ISM control.......101
Figure 6.8 Real currents under ISM control.......102

xii



NOMENCLATURE



, , , arbitrary, synchronous, rotor and slip speed
e r sl
dt
d
p, differential operator
ds
v , d,q stator voltages (arbitrary reference frame)
qs
v
dr
v , v d,q rotor voltages (arbitrary reference frame)
qr
ds
i , i d,q stator currents (arbitrary reference frame)
qs
dr
i , i d,q rotor currents (arbitrary reference frame)
qr
ds
, d,q stator fluxes (arbitrary reference frame)
qs

dr
, d,q rotor fluxes (arbitrary reference frame)
qr

ls
L , stator, rotor leakage inductances
lr
L
m
L magnetizing inductance
m ls s
L L L + = stator self inductance
m lr r
L L L + = rotor self inductance
J , motor inertia and friction constant B
s
, , stator and rotor flux vectors, stator current vector
r

s
I
xiii
r
, rotor fluxes in the stationary reference frame
r

i , i stator currents in the stationary reference frame

V ,V stator voltages in the stationary reference frame

P number of poles
2
P
n
p
= number of pole pairs
s
R , stator, rotor resistance
r
R
r s
m
L L
L
2
1 = total leakage reactance
r
r
r
R
L
T = rotor time constant
r
T
1
= inverse of rotor time constant
r s
m
L L
L

= notation
|
|
.
|

\
|
+ =
s r
r
m
s
R R
L
L
L
2
2
1

notation
r
m
p
L
L
J
n
2
3
= torque constant
d
V
,
V
stator voltages in the rotational reference frame
q
d
i , i stator currents in the rotational reference frame
q
r
, magnitude and angle of the rotor flux vector
e
*, ref reference values
^ estimated values
xiv



ABBREVIATIONS

IM Induction Machine or Induction Motor
EMF Electromotive Force
DSP Digital Signal Processor
PWM Pulse Width Modulation
A/D Analog to Digital
D/A Digital to Analog
DFO Direct Field Orientation
IFO Indirect Field Orientation
PI Proportional Integral Controller
LPF Low Pass Filter
PLL Phase Locked Loop
VM Voltage Model Observer
CM Current Model Observer
MRAS Model Reference Adaptive System
SM Sliding Mode
ISM Integral Sliding Mode
SM MRAS Sliding Mode Model Reference Adaptive System
DTSM MRAS Discrete Time Sliding Mode Model Reference Adaptive System
xv



CHAPTER 1
INTRODUCTION



The control and estimation of induction motor drives is a vast subject.
Traditionally, the induction machine has been used with constant frequency sources and
the squirrel-cage machine is currently utilized in many industrial applications, from
chemical plants and wind generation to locomotives and electric vehicles. Its main
advantages are the mechanical and electrical simplicity and ruggedness, the lack of
rotating contacts (brushes) and its capability to produce torque over the entire speed
range.
With the development and maturing of the field-oriented or vector control theory
that started about three decades ago, researchers have more and more considered the
induction machine a good candidate for variable speed and servo applications. The
objectives are related to either process control or energy savings.
Generally, control and estimation of induction machine drives is more difficult
than that of dc drives. The main reasons are the complex dynamic behavior and the need
to execute relatively complicated calculations for estimation and control using
microprocessors with limited cost, speed and accuracy.
1
Recently, sensorless vector control has received a lot of attention. The elimination
of the speed sensor reduces the hardware complexity, size, and cost and increases the
reliability of the drive. The requirement is to have a high-performance four-quadrant
drive that can deliver controllable torque over the entire speed range. While for medium
and high speeds the problem has been solved, stable operation and good performance at
low speeds has not been robustly achieved.
The objective of this dissertation is to develop improved control and estimation
algorithms for the induction machine. Some of the material focuses on the improvement
of low-speed operation while a subsequent part deals with advanced speed estimation
methods based on MRAS. The dissertation is organized as follows:
Chapter 2 presents the general theory of three-phase quantities and their
equivalence with space vectors, stationary and rotating reference frames, Clarke and Park
transformations. The models of the induction machine used for estimation and control,
the principles of Direct and Indirect Field Orientation and the control scheme used in this
research are presented. Finally, the experimental setup is described.
Chapter 3 gives an overview of sensorless control. Basically, a sensorless drive
does not have a speed sensor and this brings the problem of estimating the rotor speed by
using the system measurements. Speed estimation techniques are briefly reviewed,
however, some of them are overly complex and have not been pursued. The chapter
analyzes the factors that compromise the robustness of the drive at low-speed. Practically,
estimation algorithms (for either fluxes or speed) that have been developed under ideal
conditions do not work well or must be modified to work on a real system. Imperfections
2
in signal acquisition, the presence of noise and dc offsets, the inability to implement pure
integration and DSP quantization decrease the performance of the drive.
Chapter 4 presents the estimation of rotor fluxes using the Voltage Model
observer. It is shown that the flux estimates are erroneous at low speed due to non-ideal
integration and the effect of incorrect field orientation on the motor drive is studied. An
improved Voltage Model observer based on a Programmable Low Pass filter is developed
and implemented.
Chapter 5 discusses speed estimation using MRAS methods. The classical MRAS
works well under ideal conditions (pure integration) but exhibits inappropriate dynamics
if real integrators are used. The chapter develops two speed estimation methods that
combine the MRAS principle with Sliding Mode theory. Thus, two MRAS Sliding Mode
observers are proposed and implemented. The method provides improved dynamics of
the speed estimate and convenient design.
Chapter 6 discusses synchronous current decoupling and tuning of controllers for
a sensorless IM drive. A novel method for current control is proposed based on Integral
Sliding Mode controllers. The strategy provides decoupled control, desired current
dynamics and convenient tuning.


3



CHAPTER 2

INDUCTION MACHINE DYNAMICS AND CONTROL



In this chapter, the dynamic equations of the induction machine are presented.
Based on the machines mathematical models or considering the physics of the machine,
the available methods for induction machine control are briefly reviewed. The basics of
field-oriented or vector control are explained. Finally, the experimental setup used in this
research is presented.


2.1 Three-Phase Quantities, Space Vector Definition and
Transformations

A set of balanced and symmetric three-phase quantities is described by:
( t f f
a
cos = ) (2.1)
|
.
|

\
|
=
3
2
cos

t f f
b
(2.2)
|
.
|

\
|
=
3
4
cos

t f f
c
(2.3)
4
Note that they also satisfy:
0 = + +
c b a
f f f (2.4)
The magnitude, frequency and phase fully define the three sinusoidal waveforms. The
same information is completely described by a vector that has magnitude f and is rotating
with angular speed in a two-dimensional plane. This vector is called a space-vector and
will be used to describe three-phase quantities. Knowledge of the vectors magnitude,
position and speed is equivalent to knowing the components given by (2.1)-(2.3). Figure
2.1 shows that if the space vector is known, the abc components are obtained by simply
taking its projection on a set of axes that are 120 apart.


a
b
c
a
f
b
f
c
f

f

Figure 2.1 Equivalence of space vector and abc components


5
Since the vector is rotating in a plane, one out of the three abc axes is redundant. The
vector is still fully described if its projected on a system of coordinates with only two
axes. This coordinate system is called stationary reference frame, its axes , are
orthogonal and coincides with a (Figure 2.2).The quantities in the abc reference frame
are physical quantities and, therefore, can be measured. It is therefore important to
establish the relationship between the abc and quantities. If abc quantities are known
and quantities must be found, this is denoted as the Clarke transformation (2.5)-(2.6).

f
a
f f =

(2.5)
(
b a
f f f 2
3
1
+ =

) (2.6)


a
b
c

Figure 2.2 Stationary reference frame ()
6
The inverse Clarke transformation is used when the abc quantities must be found using
the components:

f f
a
= (2.7)

f f f
b
2
3
2
1
+ = (2.8)

f f f
c
2
3
2
1
= (2.9)
Another reference frame used with sinusoidal three-phase quantities is similar to the one
in Figure 2.2 except that its axes are rotating with the same angular frequency as the
space vector. This is denoted as rotational reference frame or the dq reference frame.
Very often, it is of interest to select this reference frame such that it is aligned with a
certain space vector. Figure 2.3 depicts this situation.


d

Figure 2.3 Rotating reference frame (dq)

7
The quantities in the dq reference frame can be computed using the components; this
is also known as the Park transformation (2.10)-(2.11).


sin cos f f f
d
+ = (2.10)


sin cos f f f
q
= (2.11)
The equations of the inverse Park transform are:

sin cos
q d
f f f = (2.12)

sin cos
d q
f f f + = (2.13)


2.2 Dynamic Equations of the Induction Machine

An induction machine can be described in an arbitrary reference frame that is
rotating with angular speed by equations (2.14)-(2.23):
qs ds ds s ds
p i R v + = (2.14)
ds qs qs s qs
p i R v + = (2.15)
( )
qr r dr dr r dr
p i R v + = (2.16)
( )
dr r qr qr r qr
p i R v + + = (2.17)
dr m ds s ds
i L i L + = (2.18)
qr m qs s qs
i L i L + = (2.19)
dr r ds m dr
i L i L + = (2.20)
qr r qs m qr
i L i L + = (2.21)
8
(
ds qs qs ds p e
i i n T =
2
3
) (2.22)
L r r e
T B Jp T + + = (2.23)
The equations of the voltages and fluxes can be used to draw two circuit models as in
Figures 2.4 and 2.5. In the general form presented, the above equations are not very
helpful for either estimation or control of the machine. However, by particularizing =0,
the induction machine representation in the stationary reference frame is obtained.
The voltage equations of the induction machine in the stationary reference frame are:
ds ds s ds
p i R v + = (2.24)
qs qs s qs
p i R v + = (2.25)
qr r dr dr r dr
p i R v + + = (2.26)
dr r qr qr r qr
p i R v + = (2.27)
Equations (2.24)-(2.27) describe the machine behavior as seen from the stationary
reference frame. The flux equations are unchanged. The stator voltage equations are
especially useful since they allow computation of the stator fluxes. This model will be
used in the subsequent chapters.

9
ds
v
ds

s
R
r
R
qs

ls
L
lr
L
( )
qr r

m
L
dr

ds
i dr
i
dr
v

Figure 2.4 Induction machine d equivalent circuit in arbitrary reference frame


s
R
r
R
ls
L
l r
L
m
L
ds

( )
dr r

qs
v
qr
v
qs

qr

qs
i qr
i

Figure 2.5 Induction machine q equivalent circuit in arbitrary reference frame


The next objective is to manipulate the equations in the frame such that the un-
measurable rotor currents are eliminated. The desired set of states consists of the rotor
fluxes and the stator currents. The model obtained is useful for estimation since it allows
computation of the rotor fluxes. These are needed if direct field orientation is employed.
10
Also, the stator currents are desired since they are measured quantities. The mathematical
manipulation is relatively long and will be omitted; it can be found in [1]. The induction
machine equations (with revised notations) in the reference frame are:

i L n p
m r p r r
+ = (2.28)

i L n p
m r p r r
+ + = (2.29)

V
L
n i pi
s
r r p r
1
+ + + = (2.30)

V
L
n i pi
s
r r p r
1
+ + = (2.31)
( )
L r r r r
T
J J
B
i i p
1
=

(2.32)
Expressions (2.28) and (2.29) should especially be noted since they can serve to
compute the rotor fluxes. Note that the two sets of equations (2.24)-(2.27) and (2.28)-
(2.32) are equivalent, either one can be used for estimation depending on the approach
chosen.
To derive a model that is useful for control, (2.28)-(2.32) are transformed into the
rotational reference frame that is aligned with the rotor flux vector. For that, the angle
e

of this vector with respect to the axis must be known. The desired state variables of the
new state-space model are .
r e q d r
i i , , , ,
The rotor flux vector is oriented along the direction of the d axis of the rotational
reference frame and only its magnitude
r
is a state in the model. The two currents i
d
and
i
q
are the projection of the current space vector onto this rotating frame. Finally,
r
is the
rotor mechanical speed. The induction machine state-space model in the rotational
reference frame aligned with the rotor flux vector is:
11
d
s r
q
m q r p r d d
V
L
i
L i n i pi


1
2
+ + + + = (2.33)
q
s r
d q
m d r p r r p q q
V
L
i i
L i n n i pi


1
+ = (2.34)
d m r r
i L p + = (2.35)
r
q
m r p e
i
L n p

+ = (2.36)
J
T
J
B
i p
L
r q r r
= (2.37)
The equations in this reference frame show the following:
The magnitude of the flux vector is controlled only by the current i
d
. This is also
known as the flux current. If the flux magnitude is kept constant, the steady-
state relationship between the flux and current i
d
becomes .
d m r
i L =
The second term in the torque expression has simplified since
q
=0. The torque is
the product between the flux magnitude and the i
q
current (also known as the
torque current). The expression of the electromagnetic torque T shows
that the machines torque can be controlled the same as in a DC machine.
q r e
i =
The transformation angle
e
can be found by integrating the sum of the rotor
electrical speed and the slip speed. This is the base for Indirect Field Orientation
(Equation (2.36)).




12
2.3 Induction Machine Control

2.3.1 V/Hz Control
The open-loop Volts/Hz control was by far the most popular method to control
the machine speed, especially because of its simplicity. The method is based on the
control of the stator frequency. The objective is to control the machine speed while
keeping constant the magnitude of the stator flux. As a result, the machine retains its
torque/ampere capability at any speed. By neglecting the stator resistance drop, the stator
flux is kept constant if
e
s
s
V

=
and the name of the method comes from this equation.
Early approaches assume that the rotor speed
r
is approximately equal to the
synchronous speed
e
(slip speed is neglected). For speed control, the speed reference
r

is set and the stator voltage V
s
is computed to maintain the desired stator flux. Integration
of the reference speed gives the angle of the stator voltage (
e
). Finally, the space vector
described by V
s
and
e
is used as command voltage for the three-phase inverter that
powers the machine.
The approach can be improved by considering the effect of the non-zero stator
resistance and of the slip speed. To compensate for the stator resistance, a boost voltage
can be added to the stator voltage. This is especially useful at low speed since the stator
resistance absorbs the major amount of the stator voltage. The slip speed is also different
from zero. To compensate, the slip is estimated using model equations and is added in the
integration of the voltage angle. The control scheme is presented in Figure 2.6.

13
IM
PI
Estimator
*
r

^
r

*
e

*
sl

- +
Inverter
*
s
V

Figure 2.6 V/Hz control scheme


A detailed discussion regarding the V/Hz method and other scalar control schemes can be
found in [2]. Generally, the method is very robust and works well even at very low stator
frequencies. The robustness of the method stems from the approach used in computation
of the voltage angle
e
. A good-quality
e
waveform (straight and undistorted) is obtained
and this results in smooth shaft motion without ripple torque or oscillations.
The major disadvantage of the V/Hz method is its sluggish dynamic response
since the method disregards the inherent machine coupling. A step change in the speed
command produces a slow torque response. During the transient, the magnitude of the
stator flux is not maintained (the magnitude decreases) and the machines torque response
is not sufficiently fast. In addition, there is some amount of under damping in the
machines flux and torque responses that increases at lower frequencies. In some
operating regions, the system may become unstable.

14
2.3.2 Field-Oriented or Vector Control
The modern approach for induction machine control is based on vector or field-
oriented control. The invention of vector-control in early 1970s brought a renaissance in
the high-performance control of ac drives. This is because, in spite of the coupled and
nonlinear machine model, the induction machine can be controlled similarly to a
separately excited dc machine.
In a dc machine, neglecting the armature reaction and saturation, the expression of
the torque is:
a f e
I K T = (2.38)
where K is a constant. The construction of a dc machine is such that the field flux
f
is
proportional with the field current I
f
and is orthogonal with the flux produced by the
armature current I
a
. These two fluxes are kept orthogonal by the collector-brush
assembly. They are decoupled and stationary in space. This means that when torque is
controlled by controlling the current I
a
, the field flux is not affected and vice-versa.
The induction machine exhibits the same behavior when viewed in the rotational
reference frame aligned with the rotor flux. The field current (i
d
) and the torque current
(i
q
) appear as dc quantities in steady state and are orthogonal and decoupled. Therefore,
the flux and torque currents can be independently controlled to obtain torque production
in the same manner as in a dc machine. Accurate torque control has two prerequisites:
Accurate control of the currents i
d
and i
q
with no steady-state error and desired
dynamics.
Accurate estimation of the rotor flux angle
e
that allows transformation of the
variables from the stationary to the rotating reference frame.
15
Generally, the first condition is easy to achieve. Traditional PI controllers are the most
common solution and they yield good experimental results; other types of controllers
have also been used (fuzzy-logic, sliding mode etc).
Estimation of the angle
e
based on the equations of the model can be done in two
ways: by DFO or by IFO.

2.3.3 Direct Field Orientation (DFO)
This method computes the rotor flux angle based on the projections of the flux
vector on the stationary reference frame (Figure 2.7). The flux components and
are needed. Several model-based observers can be used to estimate them. The difficulties
of the method come from the various problems associated with implementation of the
observers (integration, dependency on machine parameters etc).
r

Figure 2.7 Principle of Direct Field Orientation (DFO)

16

With the rotor fluxes known, the angle of the flux vector (also called rotor position) is
given by:
r
r
e

1
tan

= (2.39)
However, since sin and are really needed for the Park transformation, these can
be found directly by:
e

e
cos
2 2
sin
r r
r
e

+
= (2.40)
2 2
cos
r r
r
e

+
= (2.41)

2.3.4 Indirect Field Orientation (IFO)
An alternative approach to obtain the rotor flux angle is by Indirect Field
Orientation (IFO). The method is based on Equation (2.36). The right-hand side terms in
(2.36) are the rotor electrical speed and the slip speed. Their integration provides the
desired angle. For accurate estimation, the method requires correct values for both the
rotor speed and the rotor time constant. Also, the integration can produce incorrect rotor
angle estimates if the initial condition is not properly chosen. Generally, IFO works best
for drives that use a speed sensor to measure the rotor speed (sensored drives). In
sensorless drives, the speed information is not available by direct measurement and it
must be estimated. The accuracy of the angle produced by IFO depends heavily on the
speed estimate.
17
Many speed estimators are available, however, steady-state errors, floating of the
machine parameters, estimation delay due to low pass filtering are common problems that
influence the speed estimation bandwidth and accuracy. An incorrect speed estimate can
generally be acceptable for speed feedback but has severe effects on the drive stability
and performance if the transformation angle is found by IFO.
Various topics regarding sensorless control are presented in the subsequent
chapters. The basic IM control scheme used in this research is shown in Figure 2.8.


3-Phase
Inverter
IM
PWM
*

V
*

V
*
a
V
*
b
V
*
c
V

abc
abc
abc
T
dc
V
dq
dq
*
d
V
*
q
V
PI
PI
Speed
and
Flux
Estimator
dc
V
abc
T
a
i
b
i
- -
-
PI
m
L
1
*
r

*
d
i
*
q
i
*
r

r
r

1
tan

d
i
q
i

i
+
+
comp
dq
V

Figure 2.8 Block diagram of sensorless DFO induction motor control


18
The control scheme presented has the following characteristics:
PI controllers are used to regulate the currents and the speed.
The angle
e
is computed using (2.39) for convenience of computations.
Various flux and speed estimators have been used and they are not specifically
shown in the control scheme. Details will be presented in the subsequent chapters.
For easy implementation, the flux magnitude
r
is not regulated directly. Instead,
the desired flux level is obtained by setting the reference d-axis current. The
approach significantly simplifies the software and is equivalent to flux regulation
as long as the magnetizing inductance L
m
does not saturate.
Note the block that adds two compensation voltages to the outputs of the current
controllers. Generally, these are used in order to completely decouple the
dynamics of the d and q-axis currents. Their expressions are:
|
|
.
|

\
|
+ =
r
q
m q r p s
comp
d
i
L i n L V

2
(2.42)
|
|
.
|

\
|
+ + =
r
d q
m d r p r r p s
comp
q
i i
L i n n L V

(2.43)
The addition of the two decoupling voltages is not mandatory. For the motor
under study, it was found through simulation that the addition of the terms in
(2.42)-(2.43) may not significantly improve the dynamics. The calculation of the
above terms is computationally intensive and requires both the speed and the
magnitude of the rotor flux. In the work related to Chapters 4 and 5, the
compensation voltages have been omitted.


19
2.4 Experimental Setup

The experimental setup used in this research consists of the following:
Dayton model 2N863M 3-phase induction motor. The motor plate data and rated
parameters are presented in Table 1.


Rating hp Pole # 4
Speed 1725 rpm Voltage 220 V
R
s
10.98
L
ls
,L
lr
0.0149 H
L
m
0.297 H
R
r 5.572

Table 1 Rated parameters of Dayton motor model 2N863M


Spectrum Digital DMC 1500 3-phase inverter. The rated values for the inverter
are: 350V dc bus maximum value; 5 A continuous current; 10 A peak current.
Current sensor interface: this current acquisition board was built in the laboratory.
The board has 2 LEM current sensors model HY 5-P (maximum peak current: 5
A) and two differential op-amp structures to allow signal interfacing with the DSP
board.
20
Spectrum Digital TMS 320F2407-PGEA Evaluation Module. This is the
controller board of the setup. It contains the Texas Instruments 16 bit, fixed-point
Digital Signal Processor as well as analog interfaces and emulator port. The board
has a digital to analog converter (D/A) with 4 channels that has been used to
display the waveforms of interest.
A block diagram of the experimental setup is shown in Figure 2.9.


Spectrum Digital
3-Phase
Inverter
Dayton
2N863M
D/A
LEM a
LEM b
Spectrum Digital
TMS320F2407
Evaluation Board
PWM
Signals
Analog
Signals
V
dc
,i
a
,i
b
Digital
Scope
Currents i
a
,i
b
350 V (max)
5 V 12 V

Figure 2.9 Block diagram of the experimental setup
21



CHAPTER 3

SENSORLESS CONTROL AND ISSUES AT LOW-SPEED
OPERATION



This chapter presents the problem of sensorless control of the induction machine.
Vector controlled IM drives without a speed sensor have become an attractive technology
in the past few years. The absence of the mechanical speed sensor improves the
ruggedness of the drive and reduces the cost and the volume of the assembly. The
problem was largely studied by researchers and has commercial applications especially
for motors running at medium or high speeds. At low-speed, however, it is still
problematic to achieve robust control.
First, the idea of sensorless control is presented. Later in the chapter, practical
problems regarding estimation and control of IM drives in the low speed area are
examined.





22
3.1 Overview of Induction Machine Sensorless Control

Traditionally, a sensored IM drive uses several measurements to achieve closed-
loop control. The machine stator voltages, phase currents and the rotor speed are
measured and used for feedback.
A sensorless drive is one where the speed sensor has been eliminated. Since the
rotor speed must still be used in the control scheme, this variable must be estimated
instead of measured.
In a sensorless IM control system, the speed estimate:
Must be known for certain operations.
In a drive whose objective is to control the speed, the speed estimate is required for
speed feedback.
May be needed for certain operations.
The speed estimate is needed if IFO is used to estimate the rotor flux angle. However,
if DFO is used, no speed estimate is needed.
If the voltage compensation terms in Equations (2.42), (2.43) are used, the speed
estimate is required for their computation.
May not be needed at all
If the objective of the drive is to control the torque, the reference torque current is
available (for example, it is produced by an acceleration pedal). The whole speed
feedback loop is discarded. The speed information may not be needed at all.
In this research, Direct Field Orientation (DFO) was used. For the motor studied, the
addition of the compensation voltages shown in Figure 2.8 may not have a significant
23
effect on the system dynamics if high-gain controllers can be implemented. Generally,
the speed estimate has only been used for speed feedback.


3.2 Speed Estimation Methods

There are several methods available for speed estimation of an induction motor.
The estimation is usually complex and is dependent on the machine parameters. The
methods can be classified as:
Slip calculation
Direct synthesis from state equations
Model reference adaptive system (MRAS)
Sliding Mode observers
Speed adaptive flux observer (Luenberger observer)
Extended Kalman filter
Slot harmonics
The slip calculation method is based on Equation. (3.1):
) (
1
r
q
m e
p
r
i
L
n
= (3.1)
As presented, the method uses the rotating reference frame torque current i
q
to compute
the slip speed and depends on the inverse of the rotor time constant that varies with
temperature. For simplicity, the dynamics of the flux magnitude
r
can be disregarded.
The synchronous speed
e
can be computed from the model of the machine as:
24
( ) ( )
2 2
s s
s s s s
e
i R V i R V

+

= (3.2)
A second method is available for the computation of the synchronous speed: if the DFO
method is used and if the rotor flux angle
e
is computed using (2.39). In this case,
e
can
be computed by taking the derivative of the rotor angle, i.e:
( ) ( ) 1 ( ) (
1
= k k
T
k
e e
s
e
) (3.3)
Generally, the calculation in Equation (3.2) depends on the stator resistance R
s
and on the
flux estimates
s
and
s
which may be distorted or may have dc offsets. Also, the terms
of the numerator become very small at low speed and the method tends to be inaccurate.
The second approach is simpler and easier to implement, however, imperfections in the
waveform of the rotor angle
e
usually require additional filtering.
The speed signal can be synthesized directly from the model equations using
stationary reference frame variables [2].
The expression of the speed estimate is:
(
(

|
.
|

\
|

+
=

i i L
r r m r r
r
r
r r
r
2 2
1
) (3.4)
This speed synthesis method uses the derivatives of the fluxes and the machine
parameters and will tend to give poor accuracy of estimation.
Speed estimation using Model Reference Adaptive System (MRAS) is based on
comparison of the outputs of two models (one called reference model and a second one
called adjustable model). The parameter of one model (speed, in this case) is adapted
such that the error between the model outputs is equal to zero. For the induction machine,
25
the MRAS method can be applied for models that yield either the rotor fluxes as in [3]-
[5] or the back-EMFs as in [6]-[7].
The stationary reference frame fluxes can be simultaneously computed using two
sets of equations:
Stator equations (3.5) and (3.6) which are speed independent.
( )

= = dt e dt I R V
s s s s s
(3.5)
(
s s s
m
r ref
I L
L
L

= ) (3.6)
Where
s
,
ref

are the stator flux vector and the rotor flux (reference) vector.
Rotor equations (which are speed dependent).
An observer can be constructed in the form (3.7), (3.8):

i L
dt
d
m
r r
r
r
+ =
^ ^ ^
^
(3.7)

i L
dt
d
m
r r
r
r
+ =
^ ^ ^
^
(3.8)
An error term in the form (3.9) is created and is used as input into a PI controller.
r
ref
r
r
ref
r

^ ^
= (3.9)
The speed estimation scheme is shown in Figure 3.1.
26
Stator equations
(Reference Model)
Rotor equations
(Adjustable Model)

V

V

i
X
X
+
-
PI
^
r

ref
r

ref
r

^
r

^
r

Figure 3.1 Speed estimation by rotor flux MRAS method


The method has been studied by several researchers and gives good accuracy with
acceptable dynamics under ideal conditions, see [3]-[5]. For implementation, several
changes must be considered for both the reference and the adjustable model especially
because ideal integration cannot be easily achieved. As a result, the quality of the speed
estimate decreases. Implementation issues stemming from non-ideal integration will be
detailed in Chapter 5.
Sliding mode (SM) is another technique that can be used for rotor speed
estimation. A sliding mode flux and speed observer is described in [8]. The fluxes
produced by the SM observer are used for field orientation and the speed estimate enters
the speed feedback loop. The sliding mode method is based on high-gain discontinuous
feedback and is attractive because of order-reduction and uncertainty rejection.
27
Additionally, stable observers can be designed and the sliding mode gains are easy to
select. At very low speed, however, robustness and speed estimation accuracy may
deteriorate. Several other SM observers have been reported and details can be found in
[9]-[11]. The sliding mode technique has been successfully employed in this research.
Chapter 5 presents the development of novel SM observers for speed estimation.
A Luenberger observer can be designed to yield the speed estimate. The technique
was first described in [2]. Consider the induction machine state-space representation
(3.10), (3.11) where X=[i

,i

,
r
,
r
] and Y=[i

,i

].
s
BV AX X + =

(3.10)
CX Y = (3.11)
Matrix A depends on the rotor speed
r
0 and the output Y consists of the two stationary
frame currents.
A speed adaptive full-order observer can be designed using a constant matrix G in the
form:
|
.
|

\
|
+ + =

s s s
i i G BV X A X
^ ^ ^ ^
(3.12)
^ ^
X C Y = (3.13)
The speed adaptation mechanism is derived upon the condition that the derivative of the
Lyapunov function consisting of the speed and flux mismatches be negative definite . The
speed adaptation law uses a PI controller and is given by (3.14):

|
.
|

\
|
+
|
.
|

\
|
= dt e e K e e K
I P r

^ ^ ^ ^ ^
(3.14)
where e , e .
^

i i =
^

i i =
28
This observer uses four equations and is computationally intensive. The speed estimation
depends on a large set of machine parameters; stator, rotor resistances and model
inductances contribute to the elements of the matrix A. The estimation error tends to be
more dominant in the low speed region. Because of the disadvantages explained above,
this method has not been pursued.
The Extended Kalman Filter method [2],[12] appends a fifth state (the motor
speed
r
) to the model described by the Luenberger observer. The method is derived
under the assumption that (speed is constant). The motor speed is both a
parameter and a state in the model and is therefore estimated directly. The matrix G of
the previous observer (which was a constant matrix) is replaced by a time-varying matrix
K such that the error between the model and the observer vanishes. Because of its high
complexity this method has not been pursued. The computation time is long even with a
very fast DSP and there are also parameter variation problems.
0 =

r
The Slot Harmonics method is based on the rotor slots effect: they produce space
harmonics in the air gap flux, thus causing ripple in the induced stator voltages. The
magnitude and frequency of the ripple is proportional to the rotor speed and this allows
speed estimation using a signal processing circuit. However, off the shelf machines are
usually designed for small reluctance variation due to slots and the method becomes less
precise.




29
3.3 Low-Speed Sensorless Control

3.3.1 Reasons for instability at low-speed
Generally, sensorless IM drives work well in the medium and high-speed region.
At low speed and particularly at zero speed, the drive may become unstable, the torque
capability is diminished or lost and speed regulation may be inaccurate.
The physical explanation for this phenomenon is that all estimation methods,
directly or indirectly, rely on the effect of the rotor-induced voltage. The induced voltage
becomes very small at low frequency and disappears at zero frequency. However, this is
only the theoretical side of the explanation. There are several practical issues that
contribute to the low-speed instability problem.
In a sensorless IM drive, the instability at low speed is due to one or a
combination of the factors below:
Incorrect field orientation.
If the DFO method is used, the stationary reference fluxes must be estimated using one of
the available observers (estimators). If the estimated fluxes are erroneous (magnitude
and/or phase is incorrect), the Park transformation angle is incorrect and field orientation
is lost. This causes coupling between the d and q axis of the control system and the torque
response is compromised. Torque pulsations, ripple or instability are common effects of
incorrect field orientation.
Distortions in the rotor flux angle
e

With the DFO method, the flux angle is estimated by (2.39) or (2.40)-(2.41). The quality
of
e
depends on the quality of the flux waveforms; ideally, they should be offset-free,
30
pure sinusoidal signals. In implementation, dc offsets and distortions can frequently be
seen in the flux waveforms; they are caused by small errors in the analog portion of
signal acquisition circuits. They cause distortions in the transformation angle that can
produce instability at low speed.
Note that this is a resonant chain: non-ideal input signals results in non-ideal flux
estimates and distorted rotor position. The result is a control voltage vector that does not
rotate smoothly. The control system applies this voltage at the motor terminals and, since
it is not smooth, this causes distorted currents. These currents are sampled and used as
input. Consequently, the chain repeats.
Incorrect speed estimate
In the low speed area, it can happen that the speed estimate swings between positive
and negative values. The speed controller reacts to this and can produce an oscillatory
torque that leads to instability.
Delay due to filtering
Several estimation methods use algebraic calculations or dynamic systems to obtain
the speed estimate. Very often, simulations show that a smooth estimate is obtained and
no filtering is needed. In reality, with real signals that have noise, distortions or offsets,
the speed estimate is not smooth and must be filtered. The use of a LPF adds a delay in
the feedback portion of the control system. If excessive, filtering leads to instability,
especially at low speeds.
DSP Quantization
Very often, control and estimation algorithms are verified through simulation. It
should be noted that a PC generally represents numbers in a 64-bit IEEE floating point
31
format, thus allowing both very small and very large numbers to be represented with
excellent accuracy. For modern DSPs and especially for fixed-point types, the dynamic
range is much smaller. The number of bits used for number representation is fixed and
quantization (dropped bits or dropped decimals) occurs.
With the 16-bit DSP used in this research, the quantization effect is different for
positive and negative numbers. Thus, if a positive number is truncated, the result is
positive and smaller than the un-truncated number (this is a rounding towards zero).
When a negative number is truncated, because of its 2s complement representation, the
result is more negative than the original number (this is a rounding towards minus
infinity). If hundreds of operations are executed, the above effect is hard to quantify.
Instabilities and limit-cycle are known to appear [13],[14]. This decreases the overall
robustness of the drive, especially at low speed.

3.3.2 Limits of stable operation at low-speed
A sensorless IM drive must satisfy certain performance requirements and must also
be cost-effective. This limits the number and type of sensors that can be used. The
established practice is that only the stator voltages and currents are directly measured.
The rest of the variables of interest (fluxes, speed, motor parameters etc) must be
estimated using the above measurements.
Since estimation methods are very often quite complex, the accurate acquisition of
the voltage and current signals is a concern.
Voltage acquisition
32
The voltage applied at the terminals of the machine is a high-frequency PWM
switching voltage and cannot be measured directly. The terminal voltages must first be
processed through hardware low pass filters in order to extract the fundamental
components. An additional signal conditioning circuit is needed to allow correct
interfacing with the DSP board. Usually, this is problematic because of the high-
bandwidth needed for the LPF. The LPF may reduce or eliminate the high-frequency
components but will also attenuate and delay the fundamental (the amount of attenuation
and delay varies with frequency). The attempt to compensate the voltage signals by
software operations is generally laborious and may be inaccurate since the signals
frequency must be known.
An attractive alternative is to compute the stator-applied voltages using the
measurement of the inverter dc bus voltage and the PWM duty-cycles that are already
available in the DSP environment.
The line-neutral voltages are:
(
c b a
dc
an
T T T
V
V = 2
3
) (3.15)
(
c b a
dc
bn
T T T
V
V + = 2
3
) (3.16)
where T
a
,T
b
,T
c
are duty-cycles of the PWM inverter.
The stator voltages in the stationary reference frame are:
an
V V =

(3.17)
(
bn an
V V V 2
3
1
+ =

) (3.18)
33
The voltage signals obtained in this way are free of harmonics and offsets. Generally,
they are not exactly equal to the true voltages applied on the motor because of inverter
nonlinearity. The inverter is a quasi-nonlinear device. The blanking time needed to avoid
shoot-through of the switches and the voltage drops on the semiconductor devices make
it nonlinear. The nonlinearity is negligible if a modulation index greater than 0.7-0.8 is
used, however, the difference between command and produced voltage increases as the
modulation index decreases. This is precisely the case of an induction machine running at
low speed/frequency where a small stator voltage is applied. A method to compensate
this effect was presented in [26],[15]. The method shows that a non-circular command
voltage vector must be used in order to obtain a circular voltage at the motor terminals.
For good accuracy, the typical voltage drops of the diodes and IGBTs in the inverter
bridge must be known and some tuning is needed. Good experimental results have been
reported, however, implementation is computationally intensive and a powerful processor
must be used.
Current acquisition
Errors in the measured currents appear frequently and the causes are mainly in the analog
portion of the signal conditioning circuits. Software operations may be attempted in order
to alleviate the effect. One problem stems from the interfacing of sensor-generated
signals with the A/D input. Usual DSPs accept analog input signals in the range 0-3.3V
and a dc offset (equal to half the A/D range) must be added to the ac signal produced by
the sensor. It is common that the dc offsets at the A/D input differ by 1% pu due to
different voltage drops on wires or connectors.
34
Software routines can be used to estimate the offset for each current acquisition
channel. The signal is sampled 20-30,000 times before motor startup and the no-load
offset for each channel is obtained. Practice shows that the current channel offsets change
when the motor is powered and offsets would need to be recomputed in real time. This is
a difficult operation and our experimental attempts only produced slightly better results at
the expense of a great computational burden.
Current imbalance may appear in the digital system as a result of A/D (sampling)
operations that do not match the systems hardware. Usually, a modern DSP
accommodates 8 or 16 analog input channels; however, it only has one A/D converter
block. A software multiplexer controls which analog input will be passed next to the A/D
block. Under the mutiplexers command, the analog input selected is charging the A/D
input capacitor. When the next channel is commanded for A/D conversion, the incoming
signal will be applied on the same capacitor. Parasitic resistances (from wires,
connectors) in the signal path help create a classical RC low pass filter. Its time constant
is very small but increases with the increase in R. If the charge/discharge time of the
capacitor is of the same order of magnitude with the time interval between sampling of
the signals, the capacitor exhibits memory. If a fast sampling sequence is desired (as is
most often the case), the acquired currents look distorted. The problem can be alleviated
by keeping the wires as short as possible, using op-amps with high slew-rate and by
increasing the time interval between current samples or by repeated sampling. Distorted
or offseted current signals lead to several effects.
First, since they contribute to flux estimation, there is a decrease in the quality of the
flux waveforms. The presence of offsets eliminates the possibility of using a pure
35
integrator for estimator equations. Consequently, compromises or compensations
regarding integration must be made.
Offsets in the line currents become ac signals when the physical currents are
represented in the dq rotational reference frame. The ac ripple influences the behavior of
the current controllers. In [17], it is shown that the dc offsets in the motor currents
produce parasitic torques and speed oscillations of the fundamental and double the
fundamental frequency
.
The authors propose closed loop compensation based on the
resulting speed oscillations. The method only works for steady state and is not
sufficiently accurate for flux estimation during transient regimes.

36



CHAPTER 4

AN IMPROVED FLUX OBSERVER BASED ON PLL
FREQUENCY ESTIMATOR FOR INDUCTION MACHINE
SENSORLESS CONTROL



This chapter presents an improved method for flux estimation for sensorless
vector control of induction motors based on a programmable low pass filter (LPF) and a
vector rotator. A Phase Locked Loop (PLL) synchronized with the voltage vector is used
for estimation of the stator frequency. The performance of both the PLL and the flux
estimator is studied. The method proposed is designed to improve the flux estimation in
the low-speed area. First, the traditional Voltage Model observer and the effect of
erroneous field orientation on the control system shown in Figure 3.1 are discussed.


4.1 The Voltage Model Observer

Estimation of the fluxes using the Voltage Model observer is a convenient method
for IM sensorless vector control. Techniques based on rotor equations require knowledge
37
of the mechanical speed and depend heavily on the accuracy of the rotor time constant.
Full order observers are more difficult to implement and rely on an even larger set of
machine parameters [18]-[24].
The Voltage Model (VM) observer is both attractive and important for several
reasons: First, if necessary, the critical parameter R
s
can be estimated in real-time by
using stator-mounted temperature sensors [25] and accurate stator fluxes can be obtained.
Second, the coefficients in the equations linking the stator and rotor fluxes do not vary
significantly with operating conditions [26], thus, good estimation of the rotor fluxes is
possible and the decoupled rotor Direct Field Orientation (DFO) method can be used.
Third, dual-flux MRAS observers use the VM as reference in order to estimate the motor
speed. Finally, the estimate of the flux angle (in either stator or rotor DFO method) is
crucial for the stability and performance of a sensorless drive.
Many researchers have addressed the issues related to flux observers and the main
practical difficulty is well known: the DC offsets and drifts present in the motor back-
EMFs make pure integration very difficult. To solve this problem, most methods
described in the literature avoid the pure integration and are based on low pass filters
with either fixed or variable cutoff frequency [19]-[25]. An offset compensation method
and pure integration have been reported in [15],[16]. Excellent experimental results are
shown at low frequency; however, the complexity, computational burden and accuracy
issues almost prohibit its implementation on a low-cost, fixed-point processor.
The Voltage Model observer estimates the stator and rotor fluxes according to
(4.1) and (4.2):
( )

= = dt e dt I R V
s s s s s
(4.1)
38
(
s s s
m
r
r
I L
L
L
= ) (4.2)
Note that the stator fluxes are obtained by integrating the stationary axis back-
EMFs and the rotor flux results from an algebraic equation. In (4.1), V
s
and I
s
are both
available since they are the systems measured quantities.
It was explained in Chapter 3 that pure integration of real sinusoidal DSP signals
can not be achieved directly because the dc offsets drive the output of the integrator to
plus or minus infinity. This is exactly what happens if an attempt is made to directly
integrate the stator equations (4.1).
In order to avoid the divergence of the output, pure integrators are replaced by
low pass filters with fixed cutoff frequency. The transfer function of a pure integrator
|
.
|

\
|
s
1
is slightly altered by adding a small constant term in the denominator
|
|
.
|
\
+
c
s
1

|
. Thus,
the dc gain of the LPF is finite and the output will not diverge when the inputs have dc
offsets. However, the LPF emulates a pure integrator only if the input frequency is much
higher than
c
. This effect is discussed in detail later in this chapter.
Synthesis of the stator fluxes using a fixed-cutoff LPF is easy to implement.
Selection of an appropriate cutoff frequency can be done by experimental trials (usual
values are 2-4 Hz); generally, this value depends on the hardware setup, quality of
voltage and current feedback signals, A/D conversion precision, system noise etc. With
the above LPF replacing the pure integrator, the dc offset problem can be largely
alleviated and the flux estimates are close to the real ones if the stator frequency is five
times or higher than the filter cutoff. However, magnitude and angle errors are significant
in the low frequency range.
39
With the pure integrator replaced by a LPF with the cutoff frequency
c
to
minimize the dc offset effects, the relationship between the stator flux and back-EMF
becomes:
c s
s
s e

+
=
1
(4.3)
The magnitude and phase errors between the flux vector obtained by (4.3) and the
one that would result from pure integration are:
2 2
c s
s
mag

+
= (4.4)
c
s
phase

1
tan
2

= (4.5)
At low values of
s
, the estimator yields a flux vector whose magnitude is smaller
than real and whose phase lags to the back-EMF by an angle less than 90 degrees. As
s
increases,
mag
tends to 1 while
phase
tends to 0 and the estimate is close to the real flux.
Figure 4.1 shows the real motor fluxes and the fluxes produced by the VM observer that
was implemented with a 3.2 Hz bandwidth LPF instead of a pure integrator.
40

Figure 4.1 Real and VM estimated fluxes and rotor angle, 0.1 s/div


For the real and estimated rotor fluxes shown in Figure 4.1, the magnitude and angle
errors are obvious. The real and estimated rotor flux angles differ by about 30 electrical
degrees. The waveforms are obtained at a shaft speed of 100 rpm (approximately 5 Hz
stator frequency) with i
d
=0.6 pu and T
L
=0.2 Nm. At lower speed, the errors are even more
pronounced.
Several attempts have been made to achieve pure integration and/or to eliminate
the above magnitude/angle errors. Three new integration algorithms are presented in [27]
where fixed cutoff filters and feedback are used to emulate pure integration. The method
most suited for ac machine flux estimation exploits the orthogonality of the back-EMF
and stator flux vectors. From a practical perspective, the approach uses very small
41
quantities (which are prone to quantization error) to drive the input of a PI controller that
is used for feedback. Additionally, the method is computationally intensive.
Programmable LPFs are used in [19]-[21] and [25] for stator flux estimation. The
axis back-EMFs are fed into variable-cutoff filters and the outputs are corrected by a
vector rotator block. Filter constants and vector rotator parameters are functions of the
stator frequency. With expression (3.2) used to estimate the stator frequency however, the
accuracy deteriorates especially at lower speeds this is because in practice, both the
motor fluxes and back-EMFs are corrupted by dc offsets. Other techniques are presented
in [28]-[33].


4.2 The Effect of Incorrect Field Orientation on the IM Drive with
Constant Field Current

As explained, a VM observer implemented with LPF will produce errors in the
rotor position. Since the angle is used for field orientation, it is clear that the Park
transformation is done with an incorrect angle.
The immediate consequence is that the torque dynamics (which is supposed to be
similar to a dc machine) is compromised. The dq currents are not correctly aligned to the
real rotor flux vector. The attempt to regulate one of the currents will change the other
one and vice-versa (decoupling is lost). The torque produced suffers and the speed
dynamics may show oscillations or even instability depending on how large the angle
error is.
42
This section briefly explores the effect of incorrect rotor position for the specific
control scheme used in this research. It was mentioned that the scheme does not employ
flux regulation but rather uses a predetermined d axis current to establish the field. The d
axis current controller will insist on achieving this current, however, this field current is
wrongly aligned. As a result, the behavior of the q axis current controller is affected. The
situation is explained in Figure 4.2. The d-q axes correspond to the real rotational
reference frame (aligned with the rotor flux vector). However, the controller is using
another set of axes (d
est
and q
est
) and will force i on this estimated direction. To generate
a certain torque (in steady state), the q axis current controller will synthesize a value for
such that the desired torque is produced. The angle is the error angle between the
real and estimated reference frames.
*
d
*
q
i


d
q
d
est
q
est

*
d
i
*
q
i

Figure 4.2 Torque production under incorrect field orientation


43
In Figure 4.2, the torque is produced by the projections of the two reference currents on
the real dq axis. Thus, the torque expression is:
( )( ) cos sin sin cos
* * * *
q d q d e
i i i i K T + = (4.6)
For a given load torque and neglecting other power losses, (4.6) turns into a 2
nd
order
equation that allows computation of i as a function of , T
*
q
L
and . Simulation results
in Figures 4.3 and 4.4 show the steady-state effect of incorrect field orientation: an
unnecessary increase in the stator current. Under the same torque production
requirements (0.5 pu), the stator current increases with the error in rotor position.
Additionally, if the angle error is big, the current synthesized by the controller can be
negative even when positive torque is produced.
*
d
i
*
q
i



Figure 4.3 Command current under incorrect rotor position
*
q
i
44

Figure 4.4 Stator current under incorrect rotor position


4.3 A Programmable Low Pass Filter Based on PLL Frequency
Estimation

A new method for flux estimation is proposed in this section. The method uses a
programmable low pass filter and a vector rotator to synthesize the stator fluxes. The
stator frequency is estimated by a Phase Locked Loop (PLL). In essence, the flux
estimation is still in the category of programmable LPF with vector rotator correction.
The PLL synthesizes a rotating reference frame that always tries to
instantaneously align with the voltage vector. The stator frequency is only a by-product
of the PLL process. Implementation is much easier than other sophisticated methods.
45
Simulations and experimental results show that, with proper tuning, a PLL can estimate
the stator frequency with sufficient bandwidth and the method can be used for sensorless
field orientation control of induction motors with improved performance.

4.3.1 Description of the Programmable LPF
Figure 4.5 shows the block diagram of the variable-cutoff LPF used.


s
k s +
1
s
e
PLL
s

2
1 k +
( ) )
1
tan
2
(
1
k
sign j
s
e

( )
s
sign
s

1
s

Figure 4.5. Block diagram of the programmable LPF


The components of the back-EMF vector are fed into the two LPFs whose
cutoff frequency is k times the estimated stator frequency. The LPFs output vector is
compensated in magnitude and phase to obtain the flux vector that would result from
analytical integration of (4.1). The overall structure of the estimator is designed to
emulate the frequency function of a pure integrator and to avoid output saturation. Figure
4.6 shows a vector diagram and the locus of the vector as a function of k.
1
s

46

s
e
s

) 1 (
1
= k
s

0 >
s

o
45
o
45

s
e
) 1 (
1
< k
s

0 >
s

o
45
o
45
s

) 1 (
1
> k
s

2
1
=
s
s

( )
o
45 ,
1
=
s s


Figure 4.6 Locus of vector as a function of k
1
s

For k=1, the intermediate vector lags the back-EMF by 45
1
s

and its magnitude is

2 times smaller than that of the vector . As k increases, lags the back-EMF by an
angle less than 45
s

1
s

and its magnitude decreases.

In implementation, the components of are available at every sampling time.
For
1
s

s
>0, the pair of stator fluxes is constructed as:
s s
,
( )

sin cos
1 1
s s s
M + = (4.7)
( )

cos sin
1 1
s s s
M + = (4.8)
with M and given in (4.9) and (4.10).
2
1 k M + = (4.9)
47
|
.
|

\
|
=

k
1
tan
2
1

(4.10)
Since k is a constant, the compensation gain M and the sin/cos of the compensation angle
can be computed offline. The performance of this estimator depends mainly on the
accuracy of the
s
estimate but also on the choice of k. Thus, at high values of k, the
cutoff frequency of the LPF is high and this helps attenuate the offsets present in the
back-EMFs. However, M is also high and the compensation block re-amplifies the
offsets. A steady-state offset analysis can be performed to estimate the worst-case offsets
that appear in the flux components, . If the offsets at the LPF inputs are denoted
as e , the maximum offsets at the compensator output are:
s s
and
dc dc
e

and
(
dc dc
s
dc dc
e e
k
k

+
+
= =
1 1
2
max , max ,
) (4.11)
The expression above shows that output offsets are frequency-dependent and
behavior may worsen if
s
is very small. Comparison to the offsets yielded by a fixed-
cutoff LPF, assuming k=1 and
dc dc
e e

= , shows that the programmable LPF is better for
frequencies
c s
2 2 . An additional feature of the estimator is the existence of an
optimal k; k=1 minimizes the first factor in (4.11).
The biggest disadvantage of this method comes from the requirement of the
vector rotator to know the sign of the stator frequency. For low and especially near zero
frequency, the frequency estimate may temporarily change sign or oscillate around zero.
The uncertainty in the sign of
s
seriously upsets the angle compensation.


48
4.3.2 Description of the Phase Locked Loop (PLL)
The idea of using a PLL for frequency estimation comes from Uninterruptible
Power Supply (UPS) systems and was inspired by [34]. A typical three-phase UPS uses
two power feeders. Input voltages of the first feeder are rectified and feed the DC bus of
an inverter. The second feeder (also called bypass) is used as backup and is supposed to
take over the loads if any fault develops in the power conditioning circuit. This brings the
problem of synchronizing the output voltages of the inverter with those of the bypass.
Synchronization is required only if the frequency of the bypass is within acceptable limits
(59-61 Hz). A PLL can be used to slowly lock the inverter output voltage vector to the
bypass vector from any initial condition. The stability and dynamics of the
synchronization are relatively easy to achieve because the bypass voltage has almost
constant magnitude and its frequency varies slowly in the 60 Hz vicinity.
Conditions are much tougher for a PLL intended for frequency estimation in an
IM drive. This is because the reference vector likely changes its frequency (and its
magnitude) much faster during transients. Additionally, estimation for a wider range and
for both positive and negative frequencies must be achieved. The block diagram of the
PLL proposed in this paper is shown in Figure 4.7:
For the PLL shown, the stator voltage is used as a reference vector. The PLL
synthesizes the voltage angle ; this is the angle of a rotational reference frame that is
aligned with vector V . At any sampling time, the correct angle is found if either (4.12) or
(4.13) is true.
v
s
s d
V V = (4.12)
0 =
q
V (4.13)
49

V
Park
q
V
-
0
PI 1 ( ) ( ) ( )
s s v v
T k k k + = + 1
s

( )
v
sin
( )
v
cos

Figure 4.7 Proposed PLL for frequency estimation in IM drives


The projection of the voltage vector on the q axis of this reference frame is used
as the error signal in order to enforce condition (4.13). Voltage components, V ,
are transformed into the rotational reference frame given by and is fed into a PI
controller. The PI output is the frequency estimate and is used in the integration of .

V and
v
v q
V
An important feature of this PLL is that the accuracy of frequency estimation
depends only on the quality of the reference vector. In a sensorless controlled IM drive,
the stator voltage is the cleanest vector available (compared to current, flux etc). Also, if
stator voltages are constructed using the switching states of the inverter, noise and offsets
can be kept to a minimum. Additionally, the voltage vector has a considerable magnitude
and produces a consistent, large-enough error signal at the input of the PI block in Figure
4.7.
In previously published work of LPF estimators, the synchronous frequency was
determined by (4.14):
50
( ) ( )
2 2
s s
s s s s
e
i R V i R V

+

= (4.14)
The offsets and distortions in the fluxes and back-EMFs, the dependence on stator
resistance and the decreasing magnitude of the numerator at low speeds deteriorate the
s

estimate.


4.4 Control System Description

The overall structure of the control system used in this section is presented in
Figure 4.8. The system uses rotor direct field orientation control. Both a fixed-cutoff LPF
and a programmable LPF have been used for stator flux estimation and their outputs are
compared.
The angle of the rotor flux is computed using
r
r
e

1
tan

=
(4.15)
The rotor speed is estimated as:
slip e r
= (4.16)
The stator voltages (line-neutral) are obtained using the status of the inverter switches:
(
c b a
dc
an
T T T
V
V = 2
3
) (4.17)
(
c b a
dc
bn
T T T
V
V + = 2
3
) (4.18)
51
The system uses two stator flux estimators that run in parallel and one rotor flux
calculator. To focus on evaluating the PLL based flux estimation, there is no active flux
control loop included.


*
r

-
c
s +
1 s

( )
e
sin
( )
e
cos
Inverter
with
Control
Induction
Motor
Progr.
LPF
Rotor
Flux
Calc.
s
V
s
I
s
V
s
I
PLL
s

PLL
s

PI
*
q
i
*
d
i
r

Figure 4.8 Structure of the control system for study of the PLL based observer







52
4.5 Simulation Results

The proposed control system has been simulated using Simulink. A sensorless IM
drive simulation was set up. The motor is started using a traditional VM observer. The
speed reference is set at 500 rpm. At t=0.3s, the speed reference is changed to 1000 rpm.
The dynamic performance of the PLL and the accuracy of the frequency
estimation process are analyzed. The PLL is operated in parallel with the rest of the
control system and the stator frequency, angle and voltage V are monitored.
v q



Figure 4.9 Real and estimated voltage angle at startup


Figure 4.9 shows the real angle of the stator voltage vector and the estimated one.
After about 80 ms from startup, the angle estimated by the PLL has converged to the real
one. The speed reference change at t=0.3s produces very little disturbance in the angle
estimation.
53
Figure 4.10 shows the frequency estimate by both the PLL and the classical
method (using Equation 4.14). The projection of the voltage vector on the q axis of the
PLL reference frame can also be seen. In the simulation, the model adds noise and offsets
(0.5% p.u.) to the measured voltages/currents in order to simulate the behavior of the
signal acquisition hardware. In a real system, the projections of the voltage vector have
superimposed ac components; this causes the estimation ripple in Figure 4.10. At the end
of the transient, the estimated frequency has stabilized and the PLL holds V
q
close to
zero. When the step change at t=0.3s is applied, the frequency estimate locks to the new
value very quickly. On the other hand, it can be seen that the frequency estimated by the
traditional method has more ripple than by PLL method (this is because the offsets in the
current channels intervene). Also, implementation of Equation 4.14 produces some
steady state error even in simulation; estimation only gets worse if the back-EMFs or
fluxes are distorted.
Note that the final goal of using PLL is for flux estimation improvement. Figure
4.11 compares the fluxes estimated by the PLL based algorithm to the real motor fluxes.
As evidenced by the results, the estimated fluxes are very close to the real ones. The
PLL-based programmable LPF successfully overcomes the transient of the
speed/frequency at 0.3s and the estimated flux waveforms track the motor real fluxes
very well. It is also noticeable that for the initial several cycles, however, the flux
estimates are not quite right. This can be attributed to the initial convergence time needed
by the PLL. It can be suspected that this estimator might not be able to provide flux
information for motor startup in the sensorless vector control mode.

54

Figure 4.10 Estimated frequency and V
q
at startup


Figure 4.11 Real and estimated rotor fluxes
55
4.6 Experimental Results

The software of the controller is organized in two interrupts. A fast interrupt (50
s) processes feedback signals, runs the classical VM observer and the PLL based
programmable LPF, computes the rotor flux angle and regulates the currents through two
PI controllers. A slow interrupt (100 s) estimates and regulates the rotor speed, executes
the PLL algorithm and outputs the PWM commands. The inverter switching frequency is
10 kHz. The phase currents are measured through Hall sensors. Stator voltages are
computed using (4.17)-(4.18) without voltage sensors.
The initial investigation is focused on the dynamic performance of the PLL
frequency estimator. To study that, a simple V/Hz algorithm was used to run the motor. A
step change in frequency is applied and the dynamics of the PLL output is recorded.
Figure 4.12 shows the command and estimated frequency for a 9-27-9 Hz step change.
Figure 4.13 shows the same waveforms for a 9-45-9 Hz change. It can be seen that, with
a proper tuning, the PLL can estimate the stator frequency with high bandwidth. The
waveforms of f
command
in Figures 4.12 and 4.13 should be noise-free step functions,
however, all signals are shown noisy due to D/A and probe noise.


56

Figure 4.12 Command and estimated frequency, 9-27-9 Hz step, 50ms/div



Figure 4.13 Command and estimated frequency, 9-45-9 Hz step, 50ms/div


The second investigation is aimed at the stator flux waveforms produced by the
PLL programmable LPF. The results are compared with those estimated by the classical
observer. Both stator flux estimators run in parallel but only one rotor flux calculator is
used due to DSP time constraints. At a convenient stator frequency, the system is
57
transitioned and field orientation is supported by the PLL based programmable LPF. For
comparison, the classical VM observer is implemented at the same time with a cutoff
frequency of 19.98 rad/s (3.18 Hz); this was selected experimentally. Values smaller than
3.18 Hz were tried but the obtained flux waveforms corresponding to Equation 4.3 had
significant offsets.
As discussed in the theory section, the PLL programmable LPF is expected to
produce stator fluxes very close to those of the fixed-cutoff LPF for frequencies greater
than 5
c
(17 Hz). For lower frequencies, the classical observers output is lagging the
back-EMFs by an angle less than 90 degrees and leading the stator flux obtained by
programmable LPF.
Figure 4.14 shows the axis signals obtained by the proposed PLL method. The
back-EMF (magnified 128 times for convenience of inspection), the intermediate flux
and the final estimated stator flux
1
s

s
s

are shown. The stator frequency is about 18 Hz

and k=1. For k=1, the stator flux lags
1
s
by 45 degrees and its magnitude is
approximately 2 bigger.
To verify the behavior of the PLL based flux observer, Figure 4.15 was obtained
at low frequency (approx. 2.1 Hz) where the difference between the two estimators is
significant. The back-EMF is very small and noisy in this case. The output of the VM
observer is incorrect in both magnitude and phase. The flux given by the PLL based
programmable LPF lags the back-EMF by 90 degrees.

58

Figure 4.14 Back-EMF and stator flux by programmable LPF, f=18 Hz, k=1, 20ms/div



Figure 4.15.Back-EMF and stator fluxes by VM observer and PLL programmable LPF,
f=2.1 Hz, k=1, 0.2 sec/div


The flux estimation is also tested in the field orientation control mode for the
induction motor and the results are shown in Figure 4.16. In the testing, the field
59
orientation control was supported by the flux estimation obtained from the VM observer
first and then the PLL based programmable LPF took over flux estimation for the field
orientation control. The test is done at a frequency of 8 Hz. The transition moment from
VM observer to the PLL based programmable LPF was recorded in the figure; note the
difference in magnitude and phase angle between the two estimators. The signal that
commands the transition is on Channel 2.



Figure 4.16 Stator flux estimation from VM to PLL programmable LPF during transition,
200 ms/div


The behavior in terms of offsets can also be observed in Figure 4.16. The fluxes
given by the VM appear to have significant offsets and the waveforms are shifted - this is
60
due to the current signals. Under the same circumstances, the programmable LPF yields
flux waveforms that are more symmetric because of the higher cutoff frequency used.
Figure 4.17 shows the flux waveforms for a 100-700 rpm step change in the speed
command. It can be seen that the flux maintains its correct angle and magnitude during
the speed transient.



Figure 4.17 Stator fluxes during 100-700 rpm step change, 100 ms/div





61
4.7. Conclusions

The chapter discusses the problems associated with stator flux estimation in a
sensorless vector control of induction motor drives. It is shown analytically and
experimentally that the traditional voltage model observer produces significant estimation
errors in the low frequency range. If the stator frequency is known, a programmable LPF
and a vector rotator can be used to correct the problem. Estimation of the stator frequency
by a PLL is proposed and investigated. It is shown that frequency estimation can be
achieved with high bandwidth. Unlike the classical approach where frequency is
estimated analytically, the PLL relies only on the stator voltage vector and is less
influenced by offsets, distortions or variation of the stator resistance.
62



CHAPTER 5

CLASSIC AND SLIDING MODE MRAS SPEED
ESTIMATORS FOR SENSORLESS INDUCTION MOTOR
CONTROL



This chapter starts with a brief analytical discussion and a simulation study of the
classical MRAS method. The method is used to estimate the speed in a sensorless IM
drive.
Later, two novel sliding mode (SM) MRAS speed observers are developed and
their features are compared with the classical method. The SM methods proposed use the
fluxes yielded by the Voltage Model observer as reference and construct sliding mode
flux observers that allow speed estimation. Stability and dynamics of the proposed
estimators is discussed. The proposed estimators are found to be very robust and are easy
to design and implement. Unlike the classical MRAS, the speed estimation process is
based on algebraic calculations and it does not exhibit under damped poles or right hand
plane zeros. Experimental results confirm the validity of the approach.



63
5.1 Dynamics of Classical MRAS

The classical rotor flux MRAS method was first reported in [3]; it is also treated
in [4]-[6]. The approach is based on parameter adaptation in a speed-dependent flux
observer such that the observer outputs match a speed-independent reference model. The
back-EMF MRAS method is based on the same principle but uses the stationary frame
EMF components as reference [7].
First, the behavior, estimation dynamics and selection of the controller parameters
for the classical rotor flux MRAS method under ideal and non- ideal integration are
briefly analyzed. It was reported that the back-EMF MRAS generally has inferior
performance ([5],[6]) and will not be treated here.
The rotor flux MRAS method presented in [3] uses the Voltage Model (VM) as
reference and the Current Model (CM) as adjustable model to estimate the motor speed.
The technique was presented in Chapter 3 (Equations (3.3)-(3.6) and Figure 3.1).
The performance and the dynamics of the MRAS speed estimator were studied in
[3]. First, it is assumed that ideal integration is used in both the reference and the
adjustable models. The transfer function from the change is speed ( ) to the change in
estimated speed (
r

r
^
) is of the third order (Figure 5.1).
The transfer function G
1
(s) is given by (5.1) and
sl
, and
0
are the slip
frequency, inverse of the rotor time constant and flux magnitude.
( )
2 2 1
) (
sl
s
s
s G


+ +
+
= (5.1)
64
-
) (
1
s G
s
K
K
I
P
+
2
0

r

r
^

+

Figure 5.1 Speed estimation dynamics of classical MRAS under ideal integration


The root locus of the classical MRAS speed estimation for the motor parameters
used in this research is shown in Figure 5.2. A zoomed picture of the area near the origin
is included.



Figure 5.2 Root locus of sped estimation for classical MRAS under ideal integration
65
Figure 5.2 shows that gains K
p
, K
i
can be selected to obtain two critically or over
damped poles. The third pole cannot be made faster than , however, this dynamics
may be acceptable.
In [4], the simplifying assumption
sl
=0 is used since for most motors the slip is
small. Indeed, in this case, the factor (s+) can be simplified from the expression of
and the closed loop transfer function obtained is of the second order. As a result,
algebraic expressions for the design of the PI gains are obtained as a function of the
desired poles.
( ) s G
1
Despite the assumption (which is generally not true), simulation studies suggests
that the method in [4] is useful. The PI gains can be designed following this approach and
the third order system in Figure 5.1 behaves the same as the reduced order system used
for design.
Generally, the classical MRAS works very well if ideal integration is used in both
models. The fluxes of the adjustable model tend to the reference fluxes; the speed
estimate has small or no steady state error and acceptable dynamics can be designed.
However, in a real system, pure integration is difficult to implement and LPFs are used
instead of integrators. In this remainder of this chapter, this technique will be referred to
as non-ideal integration.
When LPFs are used in the classical MRAS, the model in Figure 5.1 is augmented
with additional dynamics. The consequence is that right-hand plane zeros appear in the
speed estimation loop [3]. The effect of the right-hand plane zeros is unacceptable: when
a change in the real speed occurs, the speed estimate changes in the opposite direction
during the initial transient. In a sensorless vector controlled drive, this can lead to
66
oscillations or can cause a disproportionate response of the speed controller. The
introduction of High Pass Filters as in [5] to compensate the effect of non-ideal
integration may not correct this behavior. Additional problems appear if high PI gains are
designed under noise conditions they must be reduced and an under damped response is
obtained.
Simulation results of the classical MRAS are shown in Figures 5.3 and 5.4; these
will be later compared with our proposed speed estimation technique. The PI gains of the
MRAS estimator have been designed for critically damped poles using the approach in
Error! Reference source not found.; the values are K
p
= 5505, K
i
= 2.77x10
6
. Low pass
filters with 3.18 Hz cutoff frequency are used in both the VM and CM observers to avoid
ideal integration. The flux magnitude is
0
=0.4 Wb, = 17.86, T
load
=0.2 pu. The motor is
started to a 500 rpm reference speed, at t=0.3 the reference speed is changed to 750 rpm.
Figure 5.3 shows the dynamics of the speed estimate. The right-hand side zeros in the
speed estimation loop cause the dip at t=0.3s.
Also, under non-ideal integration, the fluxes of the adjustable model do not match
the references (Figure 5.4).
67

Figure 5.3 Real and MRAS estimated speed under non-ideal integration


Figure 5.4 Real and MRAS estimated fluxes under non-ideal integration
68
5.2. MRAS Single-Manifold Sliding Mode Observer

For the speed estimation method proposed below, the rotor fluxes , are
obtained by the VM observer. It is assumed that these are the real motor fluxes and they
simultaneously satisfy the rotor equations. Thus, expressions (5.2)-(5.3) are obtained.
The development under ideal integration is presented first; later, the effect of non-ideal
conditions is discussed.
ref

ref

i L
dt
d
m
ref ref
r
ref
+ = (5.2)

i L
dt
d
m
ref ref
r
ref
+ = (5.3)
The single manifold sliding mode observer is designed as:

I L
dt
d
m r
+ =
^ ^ ^
^
(5.4)

I L
dt
d
m r
+ =
^ ^ ^
^
(5.5)
The speed estimate and the manifold s are:
) (
^
s sign M r = (5.6)


^ ^
ref ref
s = (5.7)
Note that the speed estimate is a discontinuous function of the manifold and M is
a positive constant. To show that sliding mode can be enforced in the manifold s=0, we
69
should show that there exists M sufficiently high such that the manifold is attractive, i.e.
(5.8):
0 <

s s (5.8)
After differentiating (5.7) and replacing the derivatives of the fluxes from (5.2)-
(5.5), the following expression is obtained:
) ( ) ( ) , , , , (
^ ^ ^
s sign M L f s
ref ref
m r
ref

+ =

(5.9)
where f is a function of the reference and estimates fluxes, speed and motor parameters.
Since the term ( is greater than zero when the motor is excited and f has a
positive upper estimate, it is clear from (5.9) that sufficiently high M can be selected such
that condition (5.8) is fulfilled. Thus, sliding mode is enforced in the manifold s and after
sliding mode begins we have s=0.
)
^ ^
ref ref

+
The boundary layer method described in [8] is used to find the equivalent control

eq
. Once sliding mode occurs, we can also assume along with . The
expression of the equivalent control becomes:
0 =

s 0 = s
) (
) ( ) (
^ ^
^ ^
,
ref ref
ref
m
ref
m
r eq r
I L I L
r

+
+
+ = (5.10)
From (5.10), when the estimated fluxes tend to the reference fluxes, the
equivalent speed tends to the real speed. The equivalent speed represents the low
frequency component of the discontinuous term (5.6). Thus, while the high frequency
switching function is fed into the observer, its low frequency component can be obtained
by low pass filtering and represents the speed estimate.
70
The continuous time sliding mode concept in [8] is developed with the
assumption that the switching frequency of the control is infinite. For implementation
with a digital signal processor, the sampling frequency is finite and the observer
equations (5.4)-(5.5) must be discretized.
A few choices are available for the switching term (5.6). With a very small
sampling time (<50 s), expression (5.11) can be used and the chattering in the flux
waveforms will be kept at an acceptable level:
) ( ) (
^
k
r s sign M k = (5.11)
For larger sampling times, a continuous high-gain approximation of the sign function as
in (5.12) should be used to reduce the discretization chatter ( is a small positive
constant).

>
=

k k
k k
r
s if s
M
s if s sign M
k
) (
) (
^
(5.12)
Generally, (5.12) is more difficult to implement and requires the tuning of the additional
parameter . The advantage is that this approach will reduce or may completely eliminate
the need for filtering to get the speed estimate.
A third approach would be to design a Discrete Time Sliding Mode observer
this is based on a different technique and will be discussed later.
The proposed observer is simulated in discrete time under the same conditions as
the classical MRAS. The approach given by (5.11) has been used. Figure 5.5 shows the
real and estimated speed under ideal integration; this was obtained through a 15Hz
71
bandwidth LPF. For lower ripple, the bandwidth can be reduced. Under ideal integration
in the VM observer, the estimated fluxes match the references exactly and are not shown.



Figure 5.5 Real and SM MRAS estimated speed under ideal integration


Figures 5.6 and 5.7 show the behavior under non-ideal integration- a 3.18 Hz LPF
is used in the VM observer. The waveform of the speed estimate in Figure 5.6 has about
10 rpm (2%) steady state error, however, the dynamics at step change is much improved
compared to the classical MRAS. The fluxes estimated by the SM MRAS method differ
slightly from the reference fluxes -their amplitude is about 7% bigger. Compared to the
fluxes in Figure 5.4, this is an improvement.


72

Figure 5.6 Real and SM MRAS estimated speed under non-ideal integration



Figure 5.7 VM and SM MRAS estimated fluxes under non-ideal integration




73
5.3 MRAS Single-Manifold Discrete Time Sliding Mode Observer

The development of the MRAS Discrete Time Sliding Mode observer is based on
the theory in [8]. Discrete time sliding mode is defined as the motion of the trajectory of a
discrete time system on a manifold s such that at sampling times k=1,2 ..n. ( ) 0 = k s
The first step in design is to select the system manifold. After that, at sampling
time k the control term in the observer equations should be computed such that
. The method is based on the forward (Euler) approximation of the integral to
allow computation of the terms in the manifold at sampling time k+1. For the observer
proposed, the manifold is:
k r,
^
( ) 0 1 = + k s
k
ref
k
k
ref
k
k s ,
^
,
,
^
,
) (

= (5.13)
The general method cannot be directly applied here. The dynamic system
described by (3.3)-(3.4) is not in the standard form since computation of the
rotor fluxes is done by an algebraic and not a differential equation. The attempt to
predictively compute the rotor fluxes for the next sampling time (Euler method) by
Equation (3.4) requires the values of the currents that have not been sampled yet.
Bu Ax x + =

To overcome this difficulty, the standard method has been slightly changed and
will give a speed estimate with a delay of one sampling time. The discrete time observer
equations are:
s k m
k k
k r
k k T i L + + =

] [
1 ,
1 ,
^
1 ,
^ ^
1 ,
1 ,
^
,
^

(5.14)
s k m
k k
k r
k k T i L + + =

] [
1 ,
1 ,
^
1 ,
^ ^
1 ,
1 ,
^
,
^

(5.15)
74
The above right hand sides and the known quantities , are replaced in the
expression of the manifold (5.13) and condition is enforced. The speed estimate
can be computed algebraically. Note that at sampling time k, the method computes
a speed estimate for sampling time k-1. This is only a small inconvenience as long as the
mechanical speed varies much slower than the electrical variables. The expression
obtained is:
ref
k ,

ref
k ,

( ) 0 = k s
1 ,
^
k r
ref
k
k
ref
k
k
k
ref
k k
ref
k
m
ref
k
k
ref
k
k
ref
k
k k
ref
k
k r
i i
L
Ts
Ts
,
1 ,
^
,
1 ,
^
1 , , 1 , ,
,
1 ,
^
,
1 ,
^
,
1 ,
^
1 ,
^
,
1 ,
^
1

+
+

=

(5.16)
Figure 5.8 shows the real and estimated speed by DTSM MRAS under non-ideal
integration in the reference model. Except for the initial transient when the system
trajectory approaches the manifold, the estimate matches the real speed with very small
steady state error and good dynamics. The estimated speed in Figure 5.8 was filtered
through a 15Hz LPF.



Figure 5.8 Real and DTSM MRAS estimated speed under non-ideal integration

75
Generally, this method is more difficult to implement, especially on low cost,
fixed-point processors. Equation (5.16) is relatively complicated and requires a real-time
division. Additionally, it was found by simulation that the first term in (5.16) is dominant.
This term is obtained by multiplying a very large quantity |
.
|

\
|
Ts
Ts 1
with a very small
quantity (the term that depends on fluxes). With a fixed-point processor, the digital
quantization alone will likely reduce the estimation accuracy.
Figure 5.9 shows the reference and DTSM estimated fluxes. After the observer
has converged, the error between those is very small.



Figure 5.9 Real and DTSM MRAS estimated speed under non-ideal integration

76
5.4 MRAS Double-Manifold Sliding Mode Observer

A second MRAS Sliding Mode observer is proposed under the assumption that
reference fluxes and motor parameters are known. The observer equations are:

i L s sign U
dt
d
m
+ = ) (
1 0
^
(5.17)

i L s sign U
dt
d
m
+ = ) (
2 0
^
(5.18)
Where manifolds s
1
, s
2
are:
ref
s

=
^
1
(5.19)
ref
s

=
^
2
(5.20)
After subtracting (5.2-5.3) from (5.17-5.18), the result is:
(

(
(

=
(
(

) (
) (
1 0
0 1
) , , , (
) , , , (
2
1
0
2
1
2
1
s sign
s sign
U
f
f
s
s
r
ref ref
r
ref ref




(5.21)
Since f
1
and f
2
both have positive upper estimates and the identity matrix is
positive definite, form (5.21) satisfies the sufficient condition for sliding mode to exist as
it was formulated in [8]. Consequently, the positive constant U
0
can be selected high
enough such that sliding mode is enforced on the intersection of the two manifolds. After
sliding mode occurs, s
1
=0, s
2
=0 identically and that the observed fluxes equal the
reference fluxes.
If the switching functions s
1
and s
2
are filtered, the equivalent controls are
obtained. We will denote these as . The equivalent controls represent the low
eq eq

,
77
frequency components of the switching functions. Comparison of the original equations
(5.2), (5.3) with the observer equations yields the following result:
(

=
(

(
(

=
(
(

r r
ref ref
ref ref
eq
eq

(5.22)
If the motor is excited, matrix is nonsingular, ( ).
Therefore, (5.22) has unique solution and the rotor speed estimate is:
0 ) det(
2 2
> + =
ref ref


2 2
ref ref
eq ref eq ref
r

+

= (5.23)
Note that the equivalent controls are needed to compute the speed estimate. These
are sinusoidal signals of the stator frequency. If the discretization described by (5.11) is
used to implement the switching terms, low pass filtering is needed. Unlike the previous
SM observer where the LPF bandwidth was only influencing the output ripple, in this
case the LPF attenuates the high frequencies but also affects the magnitude/phase of the
low frequency sinusoidal signal. This makes selection of the LPF time constant more
difficult. If the stator frequency increases towards the LPF bandwidth, the attenuation in
is significant and will produce errors in (5.23). On the other hand, if a high
bandwidth is chosen, little filtering is achieved at low frequency. To avoid this, the
switching manifolds are designed using (5.12). In this case, the sliding mode gains are
reduced as the observer trajectories approach the manifolds. Simulation shows that no
filtering is needed. Figure 5.10 shows the speed estimate in the same conditions as before
and with non-ideal integration in the reference model.
eq eq

,


78

Figure 5.10 Real and estimated speed by two-dimensional SM MRAS under non-ideal
integration


Figure 5.11 shows the reference and estimated fluxes; they coincide.



Figure 5.11 Reference and estimated fluxes by two-dimensional SM MRAS under non-
ideal integration

79
The waveforms of the two equivalent controls used for speed estimation are
shown in Figure 5.12. There is no need to filter them.



Figure 5.12 Equivalent controls
eq eq

,


When the speed increases at t=0.3 s, their magnitude also increases as it was
expected from (5.22). System (5.22) contains a second equation and this may offer the
illusion that the inverse of the rotor time constant can also be estimated along with the
rotor speed. This is not true. Note that the system mentioned results from the subtraction
of the original equations from those of the observer. Thus, the result is valid if and only if
used in the observer equations is equal to the real that appears in the rotor equations.
Simulation studies confirms that under incorrect , the speed estimate has steady-state
80
error. The motors rotor resistance was increased by 50% and the SM observer was run
with the rated value. The error at small loading (0.2 pu load torque) is about 10% and is
as high as 60% at high loading (1 pu load torque).
Next question is if a combined
r
and observer could be designed based on the
same approach. One can attempt to design a system as in (5.24)-(5.25). Equations are the
same as before except the estimate of is used.

I L s sign U
dt
d
m
^
1 0
^
) ( + = (5.24)

I L s sign U
dt
d
m
^
2 0
^
) ( + = (5.25)
Using the same procedure, after some algebra, the new matrix is obtained. This
matrix (which has time varying coefficients) is not invertible at all times and because of
that the approach does not work. A flux, speed and rotor time constant sliding mode
observer can be found in [9]. A complete proof of convergence is presented in [10].


5.5. Experimental Results

For the software used to study the SM observers, two interrupts are used. A fast
interrupt (50 s) processes the feedback signals, runs the classical VM observer and the
SM observer. The VM produced fluxes are used for direct field orientation. A slow
interrupt (115 s) regulates the rotor speed and outputs the PWM commands. The
inverter switching frequency is 8.7 kHz. The phase currents are measured through Hall
81
sensors. Stator voltages are computed using the switching states of the inverter. Since the
setup does not have a speed sensor, the sliding mode speed estimates are compared with
the speed estimate given by the slip method.
Figure 5.13 shows the reference and observed fluxes by the single-manifold SM
MRAS at a rotor speed of approximately 250 rpm and no load. The flux waveforms are
very similar in both magnitude and phase.
The reference flux, the observed flux of the single manifold SM MRAS, the
speed estimated by the slip method and by the proposed method are shown in Figure 5.14
(100-350 rpm acceleration) and Figure 5.15 (350-100 rpm deceleration). It can be seen
that the speed dynamics has no dip or under damped behavior and the magnitudes of the
estimates are close.



Figure 5.13 Fluxes of single manifold SM MRAS, 0.1 Wb/div, 20ms/div

82

Figure 5.14 Fluxes and speed at acceleration; single manifold SM MRAS, 0.1 Wb/div,
225 rpm/div, 500ms/div


Figure 5.15 Fluxes and speed at deceleration; single manifold SM MRAS, 0.1 Wb/div,
225 rpm/div, 500ms/div.
83
Figure 5.16 shows the fluxes of the double manifold SM MRAS. In this case, the
selection of the manifold guarantees that estimated fluxes match the references.
Equivalent controls at steady state are shown in Figure 5.17. Figure 5.18 shows
the fluxes and speed estimates for a 100-350 rpm acceleration.
eq eq

,



Figure 5.16 Fluxes of double manifold SM MRAS, 0.1 Wb/div, 20ms/div


84

Figure 5.17 Equivalent controls for double manifold SM MRAS, 0.1 Wb/div, 50ms/div



Figure 5.18 Fluxes and speed of double manifold SM MRAS, 0.1 Wb/div, 225 rpm/div,
500ms/div


85
Experimentally, it was found that the both methods are very robust. For the
single-manifold method, the sliding mode gains and the time constant of the LPF are easy
to design and the estimate is quite accurate. For the double-manifold observer, the
additional design parameter needs to be selected and it must be checked that the
reduction of the sliding mode gain according to (5.12) takes place. For a fixed-point
processor, the latter method is more computationally intensive, especially due to the real-
time division in (5.23). Also, for very accurate estimation, 32x16 bit multiplications are
needed.


5.6. Conclusions

The chapter discusses some of the problems associated with the classical flux
MRAS method under non-ideal integration and proposes two novel sliding mode MRAS
speed estimators. The speed estimate is to be used in the feedback loop of a sensorless
induction motor drive. The methods proposed are relatively easy to implement and only
require the appropriate selection of the sliding mode gains. Unlike the classical MRAS,
the speed is estimated by either low pass filtering or by algebraic equations. The dynamic
of the speed estimate obtained by the sliding mode technique does not exhibit an under
damped response or right hand plane zeros. Simulations and experiments prove that good
accuracy and improved dynamics can be obtained, even on a low-cost DSP.
86



CHAPTER 6

ROBUST CURRENT CONTROL BY INTEGRAL SLIDING
MODE



This chapter discusses the problems of current decoupling and controller tuning
associated with sensorless vector-controlled induction motor drives.
In field-oriented control, the synchronous frame currents of the induction machine
are regulated in order to obtain a torque dynamic that should resemble that of a dc
machine. However, the synchronous currents are not naturally decoupled and decoupling
compensators must be designed. Controller tuning is an additional problem since
controller gains obtained by theoretical methods or simulation quite often do not work
well on the real system.
A novel approach for current control is proposed. The method uses Integral
Sliding Mode (ISM) controllers to achieve decoupling and robust dynamic response. The
synchronous frame control voltages are synthesized as the sum of two controller outputs:
one is a traditional controller (PI, linear etc) that acts on a simulated plant model; a
second one is an ISM controller. The ISM controller accounts for the coupling effects and
compensates the parameter variations in the current loops of the machine. The method
87
allows for robust current control and efficient tuning. Simulations and experimental
results confirm the theoretical analysis.


6.1 Problem Description

The objective of field-oriented control is to obtain a torque dynamic of the
induction machine similar to that of a dc machine. When classical rotor field orientation
is used and considering the machines synchronous frame model, the torque is the
product between the rotor flux magnitude and current i
q
- these are spatially orthogonal
vectors. The magnitude of the rotor flux is controlled by current i
d
.
Unlike the field and armature currents in a dc machine, the synchronous frame
currents i
d
, i
q
are not naturally decoupled. Equations of the currents in the synchronous
reference frame contain cross-coupled terms. Without decoupling compensators, the
effect is that a change in one of the currents produces a transient disturbance in the other.
The coupling problem has been extensively studied, thus, [35] discusses coupling and
compensator design for RL loads and extends the analysis to the current controllers used
for an induction machine. Two additional PI controllers are proposed in [36] to achieve
decoupled control. The Internal Model Control method presented in [37] proposes an
alternative way for controller design while still using a traditional decoupling
compensator. The feedforward decoupling terms are also used by the method in [38]
where the control laws are synthesized using the state-space approach.
88
Theoretically, the effect of d-q axis current coupling can be alleviated by using
high-gain controllers or it can be completely eliminated by using decoupling
(feedforward) compensators.
A second problem related to synchronous current control is the design and/or
tuning of the controllers. Proportional Integral (PI) controllers are the most commonly
used and are easy to implement. Selection of the controller gains can be done analytically
or by simulation. Very often, the sets of gains obtained by either method do not work
well in implementation. There are several causes for this mismatch: floating of the
machine parameters under operating conditions, saturation effects, signal acquisition
offsets and noise, improper field orientation. As a result, the system is almost always
tuned by empirical methods or by trial and error.
This chapter proposes a novel approach for decoupled current control in which the
cross-coupled terms of the dq current dynamics are treated as unknown disturbances.
Integral Sliding Mode controllers are used for each axis to reject these disturbances and
to compensate for parameter variation.
It will be shown that the proposed approach for current control exhibits the
following desirable properties:
The estimates of rotor speed and rotor flux magnitude are not needed for
decoupling.
Decoupling is achieved for the entire speed/torque range of the drive and is not
influenced by parameter variation.
The dynamics of the dq currents can be easily designed and is robust to
disturbances and parameter variation.
89
The method provides fast and convenient tuning.
The method can be applied to all AC machines or three phase systems with
synchronous current control.
The equations of the induction machine currents in the synchronous reference
frame aligned with the rotor flux are given by (6.1), (6.2):
d
s r
q
m q r p r d d
V
L
i
L i n i pi


1
2
+ + + + = (6.1)
q
s r
d q
m d r p r r p q q
V
L
i i
L i n n i pi


1
+ = (6.2)
Examination of (6.1) and (6.2) shows that the currents i
d
, i
q
depend on their
respective control components (V
d
, V
q
) but also on additional terms. The extra terms are
functions of the machine parameters, flux magnitude, speed and currents. The result is
that the dynamics of i
d
, for example, is not independent of i
q
and vice-versa. To obtain dc
machine-like behavior, the control strategy must provide both regulation and dynamic
decoupling. Decoupling can be done by adding feedforward control components that
attempt to cancel the cross coupled terms like in [35]-[38]. The scheme for decoupled
current control is presented in Figure 6.1.
Under the assumption that the dynamics of the flux is much slower than that of
the current, equation (6.1) can be simplified using the steady-state expression of the
flux .
d m r
i L =
The scheme in Figure 6.1 uses PI controllers that have proportional gains K
d
, K
q

and integral times T
d
, T
q
. The terms added by the decoupling compensator (feedforward
terms) are given by (6.3) and (6.4) they attempt to cancel the cross terms in (6.1) and
(6.2). The dq axis control voltages are obtained by adding these to the PI outputs.
90
|
|
.
|

\
|
+ =
r
q
m q r p s
comp
d
i
L i n L V

2
(6.3)
|
|
.
|

\
|
+ + =
r
d q
m d r p r r p s
comp
q
i i
L i n n L V

(6.4)


( )
m
s
L s
L

+
1
+
+

+ s
L
s
1
+
+
d
i
q
i
ref
d
i
ref
q
i
comp
d
V
comp
q
V
d
V
q
V
-
-
|
|
.
|

\
|
+
r
q
m q r p s
i
L i n L

2
|
|
.
|

\
|
+ +
r
d q
m d r p r r p s
i i
L i n n L

s
T
s
k
d
d
1
+
s
T
s
k
q
q
1
+

Figure 6.1 Scheme for induction machine decoupled current control


If the cancellation takes place, the dynamics is simplified and it is easy to design suitable
controllers. When PIs are used, the zero of the controller is usually selected to cancel the
plants pole. The resulting closed-loop transfer functions of the currents (6.5), (6.6) are of
the first order. Ideally, poles can be made as fast as desired by increasing the proportional
gains of the controllers; in reality, these gains are limited by noise.
91
( )
s
d
s
d
d
L
k
s
L
k
s G

+
= (6.5)
( )
s
q
s
q
q
L
k
s
L
k
s G

+
= (6.6)
The effect of coupling is not the same in all IM drives. The magnitude and
variation of the cross terms depend on operating conditions, speed/torque range, field flux
and on the values of the parameters. It can be shown analytically that the effect of the
cross terms can be reduced by designing a controller with a high ratio k
d
/T
d
(or k
q
/T
q
).
Since the value of T
d
, (T
q
) is fixed by the desired pole/zero cancellation, this can be easily
achieved by increasing the proportional gain of the controller. If the gains verified by
simulation work well on the real system the decoupling compensators are not needed.
However, noise, offsets or other imperfections in the real system very often negate the
use of high gains and the compensation terms are needed.
Synthesis of the decoupling terms can significantly increase the complexity of the
control algorithm. First, the division in (6.1), (6.2) needs to be done in real-time and this
is time consuming if fixed-point processors are used. Second, the speed estimate is
needed. While in speed-controlled sensorless IM drives this estimate is already available
(and is used for speed feedback), torque-controlled drives (e.g. electric vehicle) would
need to add a speed estimator only as a prerequisite for decoupling. For the IM drives that
use a predetermined i
d
(and no flux regulation loop), the flux magnitude must also be
calculated since it is used for decoupling. Finally, even if all needed estimates are
available and there is no constraint on the computation time, parameter variations
92
(especially L
m
and ) generally compromise both the decoupling and the controller-plant
pole-zero cancellation.


6.2 Decoupled Current Control by Integral Sliding Mode

The idea of using an Integral Sliding Mode controller for robust current control
was inspired by the theory in [8] and [39]. The general method will be presented here for
clarity:
Consider a dynamic system in the form:
u x B x f x ) ( ) ( + =

(6.7)
and suppose there exists a feedback control law u
0
(x) such that the feedback system has
desired properties. We will denote this ideal closed loop system as:
0 0 0
0 ) ( ) ( u x B x f x + =

(6.8)
Under real conditions however, system (6.8) operates under uncertainty
conditions that occur due to parameter variations, unmodeled dynamics or external
disturbances. As a result, the real dynamics is:
) , ( ) ( ) ( t x h u x B x f x + + =

(6.9)
where h(x,t) is a bounded disturbance that can be rejected by the control. The problem is
to find a control u such that the trajectories of system (6.9) satisfy from the
initial time instant.
) ( ) (
0
t x t x
93
To design the control u, the idea is to add to u
0
a second control component u
1

whose action will reject the unknown disturbance h(x,t). If u
1
is a discontinuous control,
the condition is:
) , ( ) (
1
t x h u x B
eq
= (6.10)
where u
1eq
is the equivalent control of u
1
. Following the derivation in [8], the additional
control is designed as:
) (
1
s sign M u = (6.11)
z x s s + = ) (
0
(6.12)
|
0
0
) ( ) ( u x B x f
x
s
z +

| (6.13)
)) 0 ( ( ) 0 (
0
x s z = (6.14)
Manifold s contains two terms: s
0
(x) is a conventional manifold that can be
selected as a combination of the system states; z is the integral term given by (6.13).
For the current loops in Figure 6.1, consider that the decoupling voltages perfectly
cancel the cross terms. In this case, the entire dynamics consists only of the plant and PI
controller transfer functions. In both loops, the plant parameters are considered constant.
The controllers can be designed for pole-zero cancellation and desired closed-loop pole
locations. This system will be denoted the ideal system, similar to the one described by
(6.8).
The design of the ISM controllers is done separately for the d and q axis currents.
For example, for the d axis, consider now that the cross terms are part of the dynamics
and they act like disturbance. An ISM controller can be added to the control of the ideal
94
system to reject this disturbance. The control of the ideal system (u
0
) must also be found
and a model of the d axis dynamics is used for this. The scheme is shown in Figure 6.2:
The equations of the d axis ISM controller are:
) (
,
s sign M u
d SM
= (6.15)
z i P s
d
+ =
1
(6.16)
( )
|
|
.
|

\
|
+ =

d
s
d m
u
L
i L P z
, 0 1
1

(6.17)
0 ) 0 ( = z (6.18)
where P
1
and M are design parameters. Note that this controller must use a model of the d
axis ideal plant in order to synthesize the control u
0,d
. The output of the simulated model
is and is used as feedback by the PI controller. The real motor current (i
sim
d
i
d
) is used by
the ISM controller along with the ideal control term u
0,d
. The control voltage (V
d
) that
will be passed as command to the PWM inverter is the sum of the two controls.


( )
m
s
L s
L

+
1
+
- s
T
s
k
d
d
1
+ ref
d
i
d
V
sim
d
i
ISM
Controller
d
i
d
u
, 0
d SM
u
,

Figure 6.2 ISM controller structure for the d axis current



95
The controller for the q axis current is designed in similar manner. However, its equations
are different since the q axis current time constant is different.
The ISM controller replaces the traditional decoupling compensator, however, the
structure not only compensates the cross-coupled terms but also the variations of
parameters that change the time constants of the current loops in Figure 6.1.


6.3. Simulation Results

A sensorless induction motor drive with direct field orientation is simulated in
current control mode. There are no flux or speed controllers. The angle of the rotor flux is
found using direct field orientation (DFO). In order to avoid any interference caused by
improper field orientation, the simulation does not use a flux observer. The model uses
the real motor fluxes (instead of estimated fluxes) to compute the rotor position. Thus, it
is guaranteed that any disturbances seen in the dq currents are due to coupling and not
due to incorrect rotor position.
The shape of the reference currents is chosen for the most common situation that
occurs in the constant torque region of an IM drive: constant i
d
and variable i
q
.
The first objective is to demonstrate the coupling between the synchronous
currents when no compensation voltages are used. Figure 6.3 shows the waveforms of the
dq currents when i
q
reference is stepped between 0.8 Amps and -0.8 Amps. The d
reference current is set at 0.4 Amps. The load torque is 0.01 Nm. Controller parameters
96
are: T
d
=0.0026, k
d
= 8.72, T
q
=0.0018, k
q
= 14.54. They correspond to closed loop poles
at -300 and -500 respectively.
Results show that the step in i
q
produces a dip in the waveform of i
d
. Note also the
under damped behavior at startup. It should be explained that for this induction motor in
particular, this disturbance does not have a devastating effect on the torque waveform.
However, for motors with different parameters or other three phase loads, the magnitude
of the disturbance can be unacceptable.
It can be easily verified that under the same references and tuning, i
d
is
undisturbed if a traditional decoupling compensator is used.



Figure 6.3 Dynamics of d-q currents with no compensation voltages


97
Figure 6.4 shows the real currents and the model currents when ISM controllers
are used. The PIs acting on the current models have the same tuning. It can be seen that
the currents are decoupled and the dip in i
d
is eliminated.
The simulation is carried out at rated machine parameters. The controller zeros
cancel the poles of the plant model. For both d and q axes: P
1
=0.1, M=150. It is clear that
the extra controller compensates the cross terms of the real current dynamics and makes
the real current loops behave like the models.



Figure 6.4 Real and model currents with ISM controllers, 50 ms/div


The simulation in Figure 6.5 was carried out with detuned machine parameters. In
the induction machine model, L
m
is reduced to 50% of the rated value and R
r
is increased
by 70%. The current models still use the rated parameters. The PIs of the current
simulators are designed for pole/zero cancellation and closed loop poles at -300 and -100.
Simulation results show that the dynamics of the real currents matches that of the
simulated currents. The currents are still decoupled and there is no dip in i
d
. The transient
time of about 50 ms confirms the time constant of 10 ms for the q axis current
98
corresponding to the desired pole at -100. Thus, the proposed controllers provide both
decoupling and compensate the parameter mismatch.
In conclusion, in order to obtain decoupling and desired dynamics of the dq
currents it is sufficient to adequately design the model PIs and to properly select the
parameters of the ISM controllers.



Figure 6.5 Real and model currents with ISM controllers under detuning, 20 ms/div


6.4. Experimental Results

The experimental results are obtained using the 32-bit fixed-point TMS320-F2812
DSP. The software of the controller is entirely executed in one 66.6 interrupt. The
inverter switching frequency is 15 kHz. The phase currents are measured through Hall
sensors. Stator voltages are computed using the duty cycles of the inverter and the sensed
dc bus voltage. Field orientation is obtained by the DFO method using the flux observer
presented in [40]. Clearly, the implementation cannot satisfy the assumption that rotor
s
99
position is perfect, in fact, it was found experimentally that this observer does not behave
well enough through the zero speed zone at fast speed reversal. Speed reversal is avoided
since oscillations in the rotor position can easily be mistaken for coupling and they falsify
the conclusions of the method proposed.
The experimental waveforms are obtained using the second PWM state machine
on the F2812 DSP chip. The variables to be displayed are truncated to 16 bits and used as
duty cycles. The 30kHz PWM pulses are filtered through hardware low pass filters and
scoped.
Figure 6.6 shows the dq currents when no compensation voltages are used. The
reference i
q
is step changed between 0.01 pu and 0.15 pu and from 0.01 pu to 0.15 pu
(1 pu = 4.5 amps). The reference i
d
is 0.1 pu. The command voltages are obtained using
discrete time PI controllers - their parameters correspond to the simulation in Figure 6.3.
Note that after the step is applied, the motor accelerates until the dc bus is exhausted and
at the end of acceleration i
q
cannot be maintained any more. There is no load torque
applied on the shaft. The dip in i
d
confirms the coupling effect.



Figure 6.6 Real currents with no compensation voltages, 100 ms/div, 0.675 A/div
100
The waveforms in Figure 6.7 were obtained using the ISM controllers. The PIs of
the current models are tuned for pole/zero cancellation and closed poles at -300 and -500
for the d, q axis respectively. The current simulated by the q axis controller is also shown.
Note that until i
q
saturates, their shape is very similar. The currents are decoupled and the
step in i
q
does not produce any disturbance in i
d
.



Figure 6.7 Real currents under ISM control, 100 ms/div, 0.675 A/div


Figure 6.8 shows a zoom of the transient when i
q
is stepped from 0.01 pu to 0.15
pu. The closed loop pole of the model is placed at 500.
It can be seen that the dynamics of the real current matches the simulated current
in terms of shape and rise time. The 10 ms transient confirms the 2 ms desired time
constant of the current loop.
101

Figure 6.8 Real currents under ISM control, 10 ms/div, 0.675 A/div


6.5. Conclusions

The chapter discusses the problem of synchronous current decoupling and
controller tuning for a sensorless induction motor drive. A novel method is proposed for
current decoupling based on ISM controllers.
The ISM controller achieves decoupling and compensates for the variation of time
constants in the current loops. The dynamics of the real current matches the dynamics
designed for the current models. Simulation and experiments prove that the ISM
controllers eliminate the cross coupling and achieve robust control.
The proposed controller is easy to tune and the algorithm is not computationally
intensive.
102



CHAPTER 7

CONCLUSIONS



This section summarizes the contributions of this dissertation and provides ideas
for future research in this area.


7.1 Summary

There are several factors responsible for the instability of sensorless induction
motor drives at low speed. One theoretical factor is the incorrect field orientation
supplied by the Voltage Model Observer.
An improved Voltage Model Observer based on a programmable low pass filter
was designed and studied. The stator frequency is estimated using a phase locked
loop - this provides improved estimation compared to machine model equations.
The new observer corrects the deficiency of the traditional method and provides
proper field orientation.
103
The traditional rotor flux MRAS method can be used for speed estimation in a
sensorless induction machine drive. When real integration is used in the reference
model, the dynamics of the speed estimate exhibits right hand side zeros and is
difficult to tune.
Two MRAS sliding mode observers for speed estimation were proposed and
studied. In the strategy proposed, the speed estimate is obtained by either filtering
or by algebraic calculations. The estimate obtained exhibits superior properties
compared to the traditional MRAS and the observers are easier to design.
A novel method for synchronous current decoupling based on Integral Sliding
Mode controllers was proposed for sensorless induction motor drives. Unlike the
traditional approach, the new method does not use the speed and flux estimates
and decoupling is not affected by parameter variations of the machine.
The decoupling method allows decoupled control and desired dynamics of the
synchronous reference frame currents. It can be extended to other AC machine or
three phase loads.


7.2 Future Research Suggestions

For the programmable low pass filter proposed in Chapter 4, an experimental
study regarding the performance of the observer as a function of the ratio between
the cutoff frequency of the filter to the stator frequency may be required.
104
For the double manifold sliding mode MRAS speed observer in Chapter 5, it
should be investigated if the concept can be extended to estimate both the speed
and the rotor time constant.
The effectiveness of the novel current control scheme proposed in Chapter 6
should be studied under constraints related the limited dc bus voltage
105



BIBLIOGRAPHY




[1] M.N. Marwali, An integrated digital signal processor based system for speed
sensorless vector control of induction machines MS Thesis, The Ohio State
University, 1997.

[2] B.K. Bose, Modern Power Electronics and AC Drives, Prentice Hall, 2002.

[3] C. Schauder, Adaptive Speed Identification for Vector Control of Induction
Motors Without Rotational Transducers, IEEE Trans. on Industry Applications,
Vol. 28, No.5, September/October 1992,pp.1054-1061.

[4] H. Tajima, Y. Hori Speed Sensorless Field Orientation Control of Induction
Machine, IEEE Trans. on Industry Applications, Vol. 29, No.1, January/Febuary
1993, pp.175-180.

[5] R. Blasco-Gimenez, G.M. Asher, M. Sumner, K.J. Bradley Dynamic
Performance Limitations for MRAS Based Sensorless Induction Motor Drives.
Part1: Stability Analysis of the Closed Loop Drive, IEE Proceedings Electric
Power Applications, Vol.143, No.2, March 1996, pp. 113-122.

[6] M.N. Marwali, A. Keyhani, A Comparative Study of Rotor Flux Based MRAS
and Back EMF Based MRAS Speed Estimators for Speed Sensorless Vector
Control of Induction Machines, IEEE Industry Applications Society, Annual
Meeting, New Orleans, 5-9 October 1997, pp.160-166.

[7] F.Z. Peng, T. Fukao, Robust Speed Identification for Speed-Sensorless Vector
Control of Induction Motors, IEEE Trans. on Industry Applications, Vol. 30, No.
5, Sept-Oct 1994, pp. 1234-1240.

[8] V.I. Utkin, J.G. Guldner, J. Shi, Sliding Mode Control in Electromechanical
Systems, Taylor and Francis, 1999.

106
[9] Z. Yan, V. Utkin, Sliding Mode Observers for Electric Machines An
Overview, IEEE 28
th
Annual Conference of the Industrial Electronics Society, 5-
8 Nov. 2002, Vol. 3, pp.1842-1847.

[10] Z. Yan, Control and Observation of Electric Machines by Sliding Modes, Ph.D.
Thesis, The Ohio State University, 2002, pp. 97-103.

[11] A. Derdiyok, M.K. Guven, H. Rehman, N. Inanc, L. Xu Design and
Implementation of a New Sliding-Mode Observer for Speed Sensorless Control of
Induction Machine, IEEE Trans on Industrial Electronics, Vol. 49, No. 5, Oct
2002, pp. 1177-1182.

[12] Y.R. Kim, S.K. Sul, M.H. Park, Speed sensorless vector control of induction
motor using extended Kalman filter, IEEE Trans. on Industry Applications, Vol.
30, Sept/Oct 1994, pp. 1225-1233.

[13] G.F. Franklin, J.D.Powell, M.L.Workman, Digital Control of Dynamic Systems,
Prentice Hall, 1997.

[14] M. Konghirun, Fast-Transient Current Control Strategy and Other Issues for
Vector-Controlled AC Drives, PhD Thesis, The Ohio State University, 2003.

[15] J. Holtz, Sensorless control of induction motor drives, Proceedings of the IEEE,
Vol. 90, No. 8, Aug. 2002, pp. 13591383.

[16] J. Holtz, J. Quan, Drift and parameter compensated flux estimator for persistent
zero stator frequency operation of sensorless controlled induction motors, IEEE
Trans. on Industry Applications, Vol. 39, No. 4, July-Aug. 2003, pp. 1052-1060.

[17] D.W. Chung, S.K. Sul, D.C. Lee, Analysis and compensation of current
measurement error in vector controlled AC motor drives, Industry Applications
Conference, Vol.1, 6-10 Oct. 1996, pp. 388 393.

[18] H. Kobuta, K. Matsuse, T. Nakano, DSP based speed adaptive flux observer of
induction motor, IEEE Trans. On Industry Applications, Vol. 29, No. 2, March-
April 1993, pp. 344-348.

[19] M.H. Shin, D.S. Hyun, S.B. Cho, S.Y. Choe, An improved stator flux estimation
for speed sensorless control of induction motors, IEEE Power Electronics
Specialists Conference, 1998, vol 2, pp. 1581-1586.

[20] M.H. Shin, D.S. Hyun, S.B. Cho, S.Y. Choe, An improved stator flux estimation
for speed sensorless stator flux orientation control of induction motors, IEEE
Trans. on Power Electronics, 2000, Vol 15, No.2, pp. 312-317.

107
[21] M. Hinkkanen, J. Luomi, Modified integrator for voltage model flux estimation
of induction motors, IEEE 27
th
Annual Conference on Industrial Electronics,
2001, Vol. 2, pp. 1339-1343.

[22] N.R.N. Idris, A.H.M. Yatim, An improved stator flux estimation in steady-state
operation for direct torque control of induction machines, IEEE Trans. on
Industry Applications, Vol. 38, No.1, January/Febuary 2002.

[23] D. Hurst, T.G. Habetler, G. Griva, F. Profumo, Zero-speed tacholess IM torque
control: simply a matter of voltage integration, IEEE Trans. on Industry
Applications, Vol. 34, No. 4, July-Aug 1998, pp. 790-795.

[24] X. Xu, D.W. Novotny, Implementation of direct stator flux orientation control on
a versatile DSP based system, IEEE Trans on Industry Applications, Vol. 27,
No. 4, July-Aug 1991, pp. 694-700.

[25] B.K. Bose, M.G. Simoes, D.R Crecelius, K. Rajashekara, R. Martin, Speed
sensorless hybrid vector controlled induction motor drive, Thirtieth IAS Annual
Meeting, IAS '95, Vol.1, 8-12 Oct. 1995 pp.137 143.

[26] P.L. Jansen, R.D. Lorenz, D.W. Novotny, Observer-based direct field
orientation: analysis and comparison of alternative methods, IEEE Trans on
Industry Applications, Vol. 30, No. 4, July-Aug 1994, pp. 945-953.

[27] J. Hu, B. Wu, New integration algorithms for estimating motor flux over a wide
speed range, IEEE Trans. on Power Electronics, Vol. 13, No. 5, Sept. 1998, pp.
969 977.

[28] P. Vranka, G. Griva, F. Profumo, Practical improvement of a simple V-I flux
estimator for sensorless F.O. controllers operating in the low speed region,
IECON '98, Proceedings of the 24th Annual Conference of the IEEE, Vol.3, 31
Aug. - 4 Sept. 1998, pp.16151620.

[29] X. Xu, R.De Doncker, D.W. Novotny, A stator flux oriented induction machine
drive, Power Electronics Specialists Conference, vol.2, April 1988, pp. 870-876.

[30] R.D. Lorenz, T.A. Lipo, D.W. Novotny, Motion control with induction motors,
Proceedings of the IEEE, Vol. 82, No. 8, Aug. 1994, pp. 1215-1240.

[31] P.L. Jansen, R.D. Lorenz, A physically insightful approach to the design and
accuracy assessment of flux observers for filed oriented induction machine
drives, IEEE Trans on Industry Applications, Vol. 30, No. 1, Jan-Feb 1994, pp.
101-110.

108
[32] J. Holtz, J. Quan, Sensorless vector control of induction motors at very low
speed using a nonlinear inverter model and parameter identification, IEEE Trans.
on Industry Applications, Vol. 38, No. 4, July-Aug. 2002, pp. 1087-1095.

[33] H. Kobuta, K. Matsuse, The improvement of performance at low speed by offset
compensation of stator voltage in sensorless vector controlled induction
machines, Thirty-First IAS Annual Meeting, IEEE, 6-10 Oct. 1996, Vol.1, pp.
257-261.

[34] V. Kaura, V. Blasko, Operation of a phase locked loop system under distorted
utility conditions, IEEE Trans. On Industry Applications, Vol. 33, No. 1, Jan.-
Feb. 1997, pp. 58 63.

[35] F. Briz, M.W. Degner, R.D. Lorenz Analysis and design of current regulators
using complex vectors, IEEE Trans. on Industry Applications, Vol. 36, No.3,
May-June 2000, pp. 817-825.

[36] J. Jung, S. Lim, K. Nam PI type decoupling control scheme for high speed
operation of induction motors, Power Electronics Specialists Conference, PESC
'97, 22-27 June 1997, Vol. 2, pp.1082-1085.

[37] L. Harnefors, H.P. Nee Model-based current control of AC machines using the
internal model control method, IEEE Trans. on Industry Applications, Vol. 34,
No.1, Jan-Feb 1998, pp. 133-141.

[38] D.C. Lee, S.K. Sul, M.H. Park High performance current regulator for a field-
oriented controlled induction motor drive, IEEE Trans. on Industry Applications,
Vol. 30, No. 5, Sept-Oct 1994, pp. 1247-1257.

[39] V.I. Utkin, J. Shi Integral sliding mode in systems operating under uncertainty
conditions, 35
th
IEEE Conference on Decision and Control, Kobe, Japan, Vol.4,
11-13 Dec. 1996.

[40] C. Lascu, I. Boldea, F. Blaabjerg, A modified direct torque control for induction
motor sensorless drive IEEE Trans. on Industry Applications, Vol. 36, No. 1,
Jan-Feb 2000, pp. 122-130.

109