Characterization and Modeling of

Hysteresis in Ferromagnetic Material

Charakterisieren und Modellieren der

ferromagnetischen Hysterese

Der Technischen Fakultat

der Friedrich-Alexander-Universitat
Erlangen-Nurnberg
zur
Erlangung des Doktorgrades Dr.-Ing.

vorgelegt von

Shasha Bi
aus Henan, China
Als Dissertation genehmigt
von der Technischen Fakultat
der Friedrich-Alexander-Universitat Erlangen-Nurnberg
Tag der mundlichen Prufung: 20. 11. 2014

Vorsitzende des Promotionsorgans: Prof. Dr.-Ing. habil. Marion Merklein

Gutachter: Prof. Dr.-Ing. Reinhard Lerch
Prof. Dr.-Ing. Peter Wellmann
Acknowledgements
This thesis was finished during my work as an academic assistant at the Chair of Sensor
Technology, University of Erlangen-Nuremberg.
In particular, I would like to express my sincere gratitude to my first supervisor, Prof. Dr.-
Ing. Reinhard Lerch for the opportunity to work at his chair. It is due to his persistent sup-
port and patient supervision that I have finally finished my dissertation at the University of
Erlangen-Nuremberg. I am also grateful for his recommendation for the IEEE Magnetics
Society student travel grant and the academic summer school.
I am also indebted to Dr. Alexander Sutor, who has guided me since I was an exchange
master student in 2007. He was always willing to teach me experimental techniques, take
time to add new ideas to my project and share his experience. His valuable suggestions and
encouragement have been indispensable to my work.
I am sincerely grateful to my secondary supervisor Prof. Dr.-Ing. Peter Wellmann for his
professional advice in material science and his careful review of my dissertation.
Furthermore, I am obliged to Hon. Prof. Dr.-phil. Manfred Gross and Prof. Dr.-Ing. Dr.-
Ing. habil. Robert Weigel for their efforts in the exchange students project between University
of Erlangen-Nuremberg and Tongji University. It was with this project that I began my studies
in Germany.
I would like to thank all my colleagues and friends at the Chair of Sensor Technology. In
particular, I would also like to thank Dr. Thorsten Albach, Dr. Stefan Rupitsch, Dr. Felix
Wolf, Dr. Lizhuo Chen, Jurgen Ilg, Dr. Fabian Wein, Jens Grabinger, Andreas Hauck, Adrian
Volk, etc. I gained a lot of inspiration from your wise advice and our meaningful discus-
sions. Without the help of Michael Gunther, Ute Bollert and Manfred Pelz, I could not have
achieved good measurement results. Many thanks to Cornelia Salley-Sippel, Bettina Melberg
and Christine Peter for your excellent organization efforts. The good and friendly working
atmosphere at this department inspired my creativity and efficiency, and I will never forget it.
I would like to thank Dr. Ning Li from the Chair of Materials for Electronics and Energy
Technology for the access to optical microscope. I would also like to thank Da Li from ZAE
Bayern for providing the experimental samples with laser technology.
Finally, and most importantly, I would like to thank my family, especially my parents for
their unrestricted support. I am also extremely grateful to my husband for his love and support.

Erlangen, March 2014 Shasha Bi

Contents

Notations and Abbreviations III

Abstract V

Kurzfassung VII

1 Introduction 1
1.1 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 State of the Art . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.3 Objectives and Achievements in Thesis . . . . . . . . . . . . . . . . . . . . 6
1.4 Structure of Thesis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7

2 Theoretical Background 9
2.1 Ferromagnetism . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
2.1.1 Classification of Magnetism . . . . . . . . . . . . . . . . . . . . . . 9
2.1.2 Magnetization and Domain Distribution . . . . . . . . . . . . . . . . 11
2.1.3 Hysteresis Curves and Magnetic Properties . . . . . . . . . . . . . . 12
2.1.4 Magnetic Anisotropy . . . . . . . . . . . . . . . . . . . . . . . . . . 16
2.2 Preisach Model of Hysteresis . . . . . . . . . . . . . . . . . . . . . . . . . . 17
2.2.1 The Relay Switch Operator . . . . . . . . . . . . . . . . . . . . . . . 17
2.2.2 The Classic Preisach Model . . . . . . . . . . . . . . . . . . . . . . 17
2.2.3 Geometric Interpretation of the Preisach Model . . . . . . . . . . . . 19
2.3 Finite Element Analysis of Magnetic Field . . . . . . . . . . . . . . . . . . . 21
2.3.1 Partial Differential Equations of Electromagnetic Analysis . . . . . . 21
2.3.2 Nonlinear Electromagnetic Field Computation . . . . . . . . . . . . 22

3 The Inverted Preisach Model 25

3.1 The Pseudo-inverse of the Preisach Model . . . . . . . . . . . . . . . . . . . 25
3.1.1 The Weight Function in the Pseudo-inverse . . . . . . . . . . . . . . 26
3.1.2 The Inverted Everett Function . . . . . . . . . . . . . . . . . . . . . 28
3.2 The Inverted Model with Modified Switch Operator . . . . . . . . . . . . . . 30
3.2.1 The Modified Switch Function . . . . . . . . . . . . . . . . . . 30
3.2.2 Modeling of the Inverted Weight Function . . . . . . . . . . . . . . . 34

4 Experimental Techniques 39
4.1 Epstein Frame Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
4.1.1 Frequency Dependence of Hysteresis Loops . . . . . . . . . . . . . . 42

I
Contents

4.1.2 Temperature Dependence of Hysteresis Loops . . . . . . . . . . . . . 45

4.1.3 Measurements of Grain-oriented Steels . . . . . . . . . . . . . . . . 47
4.2 Double C-yoke SST . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
4.2.1 Plastic Deformation Effects . . . . . . . . . . . . . . . . . . . . . . 51
4.3 Vibrating Sample Magnetometer . . . . . . . . . . . . . . . . . . . . . . . . 53
4.3.1 Scalar Characterization with VSM . . . . . . . . . . . . . . . . . . . 55
4.3.2 Vector Vibrating Sample Magnetometer . . . . . . . . . . . . . . . . 55
4.3.3 Calibration Method . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
4.3.4 Vector Characterization with VSM . . . . . . . . . . . . . . . . . . . 58
4.3.5 Vector Characterization in Rotational Field . . . . . . . . . . . . . . 60

5 Model Identification and Calculation 63

5.1 Identification with the Inverted Everett Function . . . . . . . . . . . . . . . . 63
5.2 Identification of the Inverted Preisach Model . . . . . . . . . . . . . . . . . . 65
5.2.1 Simulation of the Forced Magnetization Effect . . . . . . . . . . . . 66
5.2.2 Parameter Determination . . . . . . . . . . . . . . . . . . . . . . . . 68
5.3 Calculation of the Inverted Preisach Model . . . . . . . . . . . . . . . . . . . 69
5.3.1 Precalculation of V Matrix . . . . . . . . . . . . . . . . . . . . . . . 69
5.3.2 Calculation Procedure . . . . . . . . . . . . . . . . . . . . . . . . . 70
5.4 Validation of the Inverted Preisach Model . . . . . . . . . . . . . . . . . . . 72
5.4.1 Simulation of Major Loops . . . . . . . . . . . . . . . . . . . . . . . 72
5.4.2 Simulation of Asymmetric Hysteresis Loops . . . . . . . . . . . . . 76
5.5 Comparison of Identification Methods . . . . . . . . . . . . . . . . . . . . . 80

6 Calculation Optimization 83
6.1 Discrete Form of the Inverted Preisach Model . . . . . . . . . . . . . . . . . 83
6.2 Precalculation of the F Matrix . . . . . . . . . . . . . . . . . . . . . . . . . 86
6.3 Updating of the Staircase Interface G(t) . . . . . . . . . . . . . . . . . . . . 88
6.4 Evaluation of the Optimized Algorithm . . . . . . . . . . . . . . . . . . . . 90

7 Conclusions and Outlook 93

7.1 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
7.2 Outlook . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97

Bibliography 99

List of Figures 105

List of Tables 109

List of Publications 111

II
Notations and Abbreviations
Notations
A (Wb/m) magnetic vector potential
B, B (T) magnetic flux density or magnetic induction
Br (T) retentivity or residual induction
Bs (T) saturation value of magnetic flux density
(volume) magnetic susceptibility
D (As/m2 ) electric flux density
(m) skin depth
relative error between measured and simulated data
mean relative error
E (V/m) electric field strength or intensity
e unit vector
(As/Vm) electric permittivity
0 (As/Vm) dielectric constant, 0 = 8.854 1012 As/Vm
E(, ) inverted Everett function
E inverted Everett matrix
F(, ) new defined function for the inverted Preisach model
F F function matrix
f (Hz) frequency
J (A/m2 ) current density
H, H (A/m) magnetic field strength or intensity
Hc (A/m) coercivity or coercive field strength
H increasing magnetic field input of the forward Preisach operator
H decreasing magnetic field input of the forward Preisach operator
H 1 inverse processing of the hysteresis models
J (T) magnetic polarization
L (H) electric inductance
lm (m) length of magnetic circuit
M, M (A/m) magnetization
M increasing magnetization input of the inverted Preisach operator
M decreasing magnetization input of the inverted Preisach operator
Ms (A/m) saturation value of magnetization
(Vs/Am) permeability
0 (Vs/Am) permeability of vacuum, 0 = 4 107 Vs/Am
r relative permeability

III
Notations and Abbreviations

n discretization size of the Preisach triangle

N size of magnetization input
Nd demagnetization factor
(Am/Vs) magnetic reluctivity
nabla operator
curl
divergence
gradient
Pc (W/kg) core loss
(Vs) magnetic flux
Qk trapezoids divided in the S+ region in Preisach plane
(1/m) electric conductivity
(kg/m3 ) material density
S switch matrix
t (s) time
T time period or one cycle of the magnetization process
Tc (K) Curie temperature
(A) magnetic ampere turns

b forward Preisach operator

b inverted Preisach operator

forward (relay) switch operator
inverted switch operator
(, ) forward weight function
(, ) inverted weight function
V weight function matrix

Abbreviations
AC alternating current
CPM classical Preisach model
DC direct current
EML effective magnetic path length
FE finite element
FORCs first-order reversal curves
GO grain-oriented
HT heat treatment
MSE mean squared error
RD rolling direction
SST single sheet tester
TD transverse direction
VSM vibrating sample magnetometer
VVSM vector vibrating sample magnetometer

IV
Abstract
This thesis proposes a new inverted hysteresis model for including the ferromagnetic hystere-
sis in electromagnetic field calculations.
Numerical techniques have been commonly utilized to solve electromagnetic field prob-
lems. In finite element (FE) formulations, the magnetic vector potential is introduced as un-
known, by which the magnetic flux can be directly obtained. The nonlinear hysteretic relation
is usually approximated by a single-valued magnetization curve. This neglect of the hystere-
sis effect may lead to serious inaccuracies in calculation. To solve this problem, hysteresis
model is required to be implemented in FE codes. However, most hysteresis models employ
the magnetic field as the input variable and take the magnetic flux (or magnetization) as the
output. Models have to be inverted for implementation. As the local inversion with iterative
methods costs considerable computation, in recent years, the possibility of developping in-
verted hysteresis model with the roles of input and output being exchanged has gained great
attention.
The inverted hysteresis model proposed in this thesis is based on the classic Preisach model.
The inversion is realized by introducing a modified switch operator and the corresponding
weight function. By using this new switch operator, the inverted model remains the wiping
out and congruency properties. It also guarantees the total positiveness of weight function in
the Preisach plane. This makes it possible and easy to model the weight function distribution.
A differentiable weight function is proposed for the inverted model. To determine the func-
tion parameters, a quasi-Newton method algorithm is applied to optimize the mean squared
error (MSE) between the measured and simulated data. This inverted model can be identified
directly from major loops, which are more accessible than a set of first-order reversal curves
required in the pseudo-inverse approach. Besides major hysteresis loops, minor loops and re-
versal curves can also be simulated. The simulation results show good approximation to the
measurement data.
The inverted model was verified for both soft and hard magnetic materials. The hysteresis
loops of steel sheets and thin films samples were measured by applying three different char-
acterization methods, the Epstein frame method, double C-yoke SST and vibrating sample
magnetometer. The measurements can be carried out to evaluate different influence factors
on hysteresis loops, such as frequency, temperature, anisotropy and plastic deformation. The
experimental data can be used for further expansion of the model.
In order to improve the calculation efficiency, the double integrations of the weight func-
tion were numerically approximated in precalculation stage. The integral expression of the
inverted model was rewritten in a discrete form. In this way, it cost low memory usage and
short calculation time during the calculation procedure.

Keywords: inverted hysteresis model, differentiable weight function, finite element method, numerical

V
Abstract

discrete formulation, vibrating sample magnetometer, double C-yoke, Epstein frame

VI
Kurzfassung
In der vorliegenden Arbeit wird ein neues inverses Hysterese-Modell entwickelt, um die fer-
romagnetische Hysterese in den elektromagnetischen Feldberechnungen zu beinhalten.
Numerische Techniken bieten oft die Losungsansatze fur die Probleme in der elektromagne-
tischen Feld. In den Finite-Elemente (FE) Formulierungen ist das magnetische Vektorpoten-
tial als ein unbekannte Parameter eingefuhrt, dadurch der magnetische Fluss direkt erhalten
werden kann. Das nichtlineare Hystereseverhaltnis wird ublicherweise durch eine einwerti-
ge Magnetisierungskurve approximiert. Die Vernachlassigung des Hysterese-Effekts fuhrt zu
den groen Ungenauigkeiten bei der Berechnung. Um dieses Problem zu beheben, ist ein in
FE-Codes implementiertes Hysterese-Modell notwendig. Allerdings verwenden die meisten
Hysterese-Modelle das Magnetfeld als Eingangsgroe und den magnetischen Fluss (oder Ma-
gnetisierung) als Ausgangsgroe. Die Modelle mussen fur die Umsetzung invertiert werden,
weil die lokale Inversion mit den iterativen Verfahren einen hohen Rechenaufwand erfordert.
In den letzten Jahren versucht man die Moglichkeit, ein inverses Hysterese-Modell unter Aus-
tauschung zwischen Eingang- und Ausgangsgroe zu entwickeln.
Das inverse Hysterese-Modell in dieser Arbeit basiert auf dem klassischen Preisach-Modell.
Die Inversion wird durch einen modifizierten Schalteroperator und die entsprechenden Ge-
wichtungsfunktion realisiert. Unter Verwendung von dem neuen Schalteroperator erhaltet das
inverse Modell die Ausloschung und Kongruenzeigenschaften. Es gewahrleistet eine positiv
Distribution der Gewichtungsfunktion in der Preisach-Ebene, die ermoglicht eine einfache
Approximation zur Gewichtungsfunktion. Eine differenzierbare Gewichtungssfunktion wird
fur das inverse Modell eingesetzt. Um die Funktionsparameter zu bestimmen, wird ein Quasi-
Newton-Verfahren zur Optimierung der mittlere quadratische Abweichung zwischen den ge-
messenen und simulierten Daten angewandt. Das inverse Modell kann direkt von der aueren
Hystereseschleifen identifiziert werden. Die auere Hystereseschleifen sind mehr erreichen als
die ersten Ordnung Umkehrung Kurven, die in der pseudo-inverse Methode benotigt werden.
Mit der Modell konnen die auere Hystereseschleifen, auch die kleinere Loops und die Um-
kehrung Kurven simuliert werden. Die Simulationsergebnisse zeigen eine gute Annaherung
an die Messdaten.
Das inverse Modell ist fur beide weichen und harten magnetischen Materialien verifiziert.
Die Hystereseschleifen von Stahlbleche und Dunnschichten werden durch drei verschiedenen
Charakterisierungsmethoden, Epstein-Rahmen, Doppel-C-Joch SST und Vibrating-Sample-
Magnetometer gemessen. Die Messungen wertet die verschiedenen Einflussfaktoren wie z.B.
Frequenz, Temperatur, Anisotropie und plastische Verformung auf Hystereseschleifen aus.
Die experimentellen Daten konnen fur die weitere Expansion des Modells verwendet werden.
Um die Effizienz zu verbessern, werden die doppelten Integrationen der Gewichtungsfunk-
tion im Vorausberechnungsprozess numerisch approximiert. Die integrale Form des invertier-
ten Modell wird in einer diskreten Form umgeschrieben. Auf diese Art und Weise kostet es

VII
Kurzfassung

geringe Speicherauslastung und Rechenzeit wahrend des Berechnungsprozesses.

Schlagworter: inverses Hysterese-Modell, differenzierbare Gewichtungssfunktion, Finite-Elemente-

Methode, numerische diskrete Formulierung, Vibrating-Sample-Magnetometer, Doppel-C-Joch, Epstein-
Rahmen

VIII
1 Introduction
1.1 Motivation
Hysteresis is the central feature of ferromagnetic materials. It is present in much electromag-
netic equipment, such as transformers, motors, actuators and magnetic recording, etc. With
the development of computers, numerical techniques, such as the finite element (FE) method,
have commonly been used to solve electromagnetic field problems. However, hysteresis is
disregarded in most electromagnetic field calculations, the nonlinear relation is usually ap-
proximated by a single-valued magnetization curve [1, 2]. This neglect of the hysteresis effect
may lead to serious inaccuracies in calculations involving hysteresis losses, especially when
semi-hard or hard magnetic materials are involved in magnetic systems [3, 4].
In FE formulations the magnetic vector potential is introduced as an unknown, by which
the magnetic flux can be directly obtained. However, most hysteresis models, including the
Preisach model, employ the magnetic field as the input variable and take the magnetization
or magnetic flux as the output quantity (forward hysteresis models). Therefore, in order to
include hysteresis models in electromagnetic field calculations, the hysteresis models must be
inverted to incorporate them into FE formulations.
Iterative methods are usually used in the local inversion of forward hysteresis models. Such
methods require considerable computation to obtain converged solutions [5]. It is possible
to circumvent this costly iterative procedure by applying an inverted hysteresis model, which
employs the magnetization or magnetic flux as input. The magnetic field is then calculated
directly without iteration.
This thesis presents an inverted model, which is accurate, effective and can be simply iden-
tified from less experimental data. The inverted model is based on the Preisach model, which
is widely applied in electromagnetic computation. The identification can be simplified by
proposing a differentiable weight function, whose parameters are determined from major hys-
teresis loops. With reference to the Everett function method, the integration of the weight
function is approximated in a precalculation procedure, thus improving the efficiency of the
calculation process.

1.2 State of the Art

In the past few years, the inclusion of hysteresis in FE formulations has attracted great interest,
and much research effort has been put into this research field. Several techniques have been
proposed to solve this problem [1, 6, 7, 5, 8]. However, these techniques have generally not
been adopted in most FE software packages.

1
1 Introduction

In electromagnetic field computations the material capabilities are usually limited to the
linear case (using a constant permeability) or approximated by a single-valued magnetization
curve. The comparisons can be seen in Fig. 1.1. The inclusion of hysteresis is mostly used
for prediction of power losses in electrical devices, which is related to the area surrounded
in hysteresis loops [3, 4]. When the simulation involves soft magnetic materials with a small
coercive field, the calculation error is acceptable when applying the single-valued nonlinear
approximation. But in the case of hard or semi-hard materials, the nonhysteretic model may
lead to considerable errors [9].

Figure 1.1: Comparisons of the material capabilities approximated in calculation. (a) linear
(constant permeability), (b) single-valued nonlinear, (c) inclusion of hysteresis.

To incorporate hysteresis models in FE codes, several points need to be considered. A stable

hysteresis model which can accurately describe hysteresis behaviour of materials is required.
Furthermore, the inclusion method should be considered from a computational point of view.
An efficient procedure for the handling of hysteresis nonlinearity is preferable. For FE users
or industry designers, the identification process is another important factor, which seriously
affects the calculation accuracy. If this process is too complex, interest in modelling hysteresis
will decrease.

1.2.1 Models of Hysteresis

Until now, many hysteresis models have been proposed and applied in electromagnetic com-
putation. Comparisons and evaluations have been made by Ivanyi in [10].
Attributing to accuracy, efficiency and robustness, the Preisach model of hysteresis is the
most common choice in electromagnetic analysis [1]. Therefore, the inverted model proposed
in this work is based on this model. It was first published by F. Preisach in 1935 and attracted
great research interest during the next decades [11, 12, 13]. The basic properties of this model,
such as return-point memory and wiping out property were stated in Madelung rules [14]. M.
Kranoselskii proposed mathematical expression for this model and made it available for many
physical hysteresis phenomena [15]. I. D. Mayergoyz researched this model in depth in [16].
He proposed the discrete formulation of the Preisach model and the identification method

2
 1.2 State of the Art

with the Everett function. This work has made a great contribution and has benefitted the
application of the Preisach model in numerical computation [17].
The classic Preisach model is a static scalar model. In the past years, many extensions
and modifications attempts have been made and published in papers. These models are called
Preisach type models [18]. Della Torre included the dynamic effect and proposed a mov-
ing Preisach model [19]. Zirka researched congruency problems and proposed a history-
dependent model [20]. Adly introduced the artificial neural network technology to the iden-
tification of the Preisach model [21, 2]. The scalar model has also been extended to a vector
one [16, 22, 6, 23].
The Preisach model is usually considered as a phenomenological and mathematical model
rather than a truly physical one [18]. However, from calculation point of view, the application
of the Preisach model provides great advantages in computation [3, 2].
Another frequently used hysteresis model is the Jiles-Atherton model (JA). Compared with
the Preisach model, the JA model is considered to be based more closely on the physics of
magnetism [24]. As the domain wall motion and spinning is taken into account, the magne-
tization process is determined from the energy analysis [25]. An important comparison was
made between the JA model and the Preisach model by Philips in [26].
Other important models, such as the Stoner and Wohlfarth model and the Hodgdon model,
are not as popular as the Preisach model. This may be due to their limitations of generality or
computation speed.

1.2.2 Inclusion of Preisach Model

As has been introduced in the motivation, hysteresis models must be inverted to include them
in a FE scheme. In FE formulations, the magnetic flux density B can be directly obtained from
the magnetic vector potential, so that it is considered to be B-oriented in some publications [5].
On the other hand, like most of the hysteresis models, the Preisach model is considered to be
H-based [5]. As it employs the magnetic field intensity H as input, it takes the magnetization
M or magnetic flux density B as the output quantity. Thus, the H-based Preisach model has to
be inverted for it to be coupled with FE codes.

Iteration Method

To deal with this nonlinear problem, an iterative method is usually used to locally inverse the
hysteresis models. Saitz and Dlala have made important reviews of this method, especially for
the FE analysis of the electromagnetic field in electrical machines [27, 3]. The most commonly
applied are the fixed-point method and the Newton-Raphson method [27, 8, 3, 6]. The fixed-
point method converges stably but slowly. The Newton-Raphson method uses derivatives
to accelerate its convergence, hence, the existing continuous derivative is crucial to obtain
converged solutions. To ensure accuracy and convergence of the results, a large amount of
computation is always required in FE analysis, which may reach 90% of the total CPU time
[28, 5].

3
1 Introduction

Inverted Preisach Model

In order to circumvent the iterative procedure, Takahashi et al. proposed applying an inverse
weight function method [29]. Essentially, it is an attempt to construct an inverted (B-based)
Preisach model, which employs the magnetic flux density B as the input. The magnetic field
can be directly calculated as output quantity. In this work the inverted model still uses the
relay switch operator in the forward Preisach model. The inverted Preisach weight function
is calculated from measured reversal curves. A large amount of experimental data and cal-
culation is required for this approach. To simplify this method, an attempt was made to find
suitable analytical weight function for the inverted Preisach model. However, it is a difficult
task because the distribution of the weight function is quite complicated [1].
Dlala offered a definition for the inverted Preisach model in [5]. In some publications, the
role of H and B has been exchanged in the inverted Preisach model [5, 7, 1]. However, this
approach is called the pseudo-inverse of the Preisach model, because it used the same Preisach
operator and retains all the basic properties of the Preisach model [3]. This feature offers the
pseudo-inverse approach great advantages in calculation. The problem regarding the weight
function can be solved by applying the Everett function method, which is commonly applied
in the identification of the forward Preisach model [16]. The pseudo-inverse of the Preisach
model has been extended to a vector model in [5, 1].
To evaluate the computation efficiency, Dlala made a comparison between the iteration
method and the pseudo-inverse approach. As no iteration is required in FE simulation incor-
porating with the B-based model, the calculation time was 10 times shorter than the simulation
with iteration method [5].

1.2.3 Identification of the Preisach Model

The identification of the model plays a significant role in the accuracy of the results. The
weight function of the H-based Preisach model has been analytically approximated by both
Gaussian [30] and Lorentzian functions [31]. However, the results produced by applying
analytical weight functions do not approximate well to the measurement data [32]. The Everett
function method is usually applied in identification [16]. It requires a family of first-order
reversal curves as experimental data [see Fig. 1.2]. However, the measurement of FORCs is
difficult and tedious [33], and this method is criticized because experimental errors may lead
to inaccuracy of simulation results [1].
The high requirement for experimental data may lead to a lack of attraction for FE users.
To simplify the identification process, Naidu proposed an approach applying a major loop
to identify the forward Preisach model [34]. Dlala proposed an algorithm to identify the B-
based model [1]. The major loop was firstly approximated by material data written in catalog,
such as coercivity and saturated magnetic flux. By calculation, FORCs were generated with
reversal points controlled by B. This identification method uses less experimental data but the
calculation procedure is relatively complex.
In this thesis, reference was made to the approach proposed by Sutor in [35, 23]. A contin-
uous analytical weight function was proposed for a forward Preisach model. Limited amount
of parameters in the function were identified by a quasi-Newton method algorithm to optimize

4
 1.2 State of the Art

2 2

Magn. polarization J (T)

Magn. polarization J (T)

0 0

2 2
60 0 60 60 0 60
Magn. field H (kA/m) Magn. field H (kA/m)
(a) (b)

Figure 1.2: Experimental data for identification. (a) FORCs (b) Major hysteresis loop.

the mean squared error (MSE) between the measured and simulated hysteresis loops.

1.2.4 Characterization Methods for Magnetic Hysteresis

The identification of hysteresis models requires reliable experimental data. There are many
different characterization methods. As magnetic performance is strongly shape dependent and
could be influenced by different measuring conditions [24], the applied methods and sample
shapes should be indicated for validation.
The Epstein frame is a standardized method for determining the magnetic properties of
electrical steels. According to standards (e.g., IEC 60404-2) this method shows excellent
repeatability and reliability. To decrease the sample amount, a 25cm Epstein frame is applied
instead of the classic one (50cm frame) [36]. However, this method is frequently criticized
because the effective magnetic path length (EML) is approximately determined for all kinds
of materials. When the steels under test are grain oriented, the difference to the real EML can
be significant [24].
Nowadays the single sheet tester (SST) method plays a more important role in the charac-
terization of electrical steels. Antonelli conducted numerical analysis on the field distribution
in Epstein frame and SST [37]. The simulation results proved that the field distribution is more
uniform in SST. The magnetic properties determined with the Epstein frame are considered to
be the average magnetic properties of the sample material.
Vibrating sample magnetometer (VSM) is applied to sensitively detect the magnetic mo-
ment. It was developed by Van Oosterhout and Foner in 1950s [38, 39]. The approach has
been greatly improved by applying Mallinson pickup coils and lock-in amplifier [40, 41].
Nowadays, VSM has become a standard instrument and is most widely used in professional
magnetism laboratories. In the modern magnetic materials research field, especially in mag-
netic recording technology, the investigation of magnetization vectors is attracting more and
more attention. By adding more pickup coils systems, VSM can be extended to 2D or 3D
vector VSM. Whats more, rotational field measurements can be carried out with VSM by

5
1 Introduction

mechanically rotating the tested sample in magnetic field [24, 39].

1.3 Objectives and Achievements in Thesis

The objective of this thesis is to propose a new hysteresis model which can be implemented in
FE formulations. The model should be capable of accurately describing hysteresis phenomena
in ferromagnetic materials, and efficiency is a further important factor. The model should be
constructed and optimized from the view of computation and application.

Inverted Preisach Model

Iterations can be circumvented by directly applying the inverted model in FE analysis. The
inverted model was constructed by exchanging the input and output relation in the classic
Preisach model (CPM). Unlike the pseudo inverse of the Preisach model, the proposed inverted
hysteresis model utilizes a new switch operator, which was modified from the relay operator
in the CPM.
By applying the modified switch operator, the inverted model retain the wiping out and con-
gruency properties. The corresponding weight function has a completely positive distribution
in the Preisach plane. This allows the weight function distribution to be modelled. A differen-
tiable weight function was proposed for the inverted model, which can be identified from less
experimental data.

Simple Identification Process

The inverted model is identified from major loops. Based on the quasi-Newton method, an
algorithm is applied to optimize the mean squared error (MSE) between the measured and
simulated data. With this method, parameters in the differentiable weight function can be
rapidly determined. The inverted model was verified for both soft and hard magnetic materials.
Besides symmetric hysteresis loops, asymmetric hysteresis curves, such as minor loops or
FORCs can also be simulated. By comparison, the simulation results produced by the inverted
hysteresis model show good approximation to the measurement data.

Efficient Calculation Procedure

The calculation procedure has been optimized by applying a numerical approximation method,
by which the double integration of the weight function can be avoided.
Similar to the calculation in pseudo inverse method, the Preisach triangle is discretized by
uniform grid in preprocessing stage. Integrations of the weight function over grid elements are
approximated and stored in a table. The integral expression of the inverted model is rewritten
in a discrete form, according to which the output is calculated by adding or subtracting the
values interpolated in the precalculated table. During calculation it is not necessary to store or
update the state of all switch operators in the Preisach plane. In this way, much memory space
is saved and the calculation efficiency is improved.

6
 1.4 Structure of Thesis

Reliable Experimental Data

Characterization of the electrical steels was made by means of an Epstein frame apparatus.
The obtained results show good consistency with the standard material data. Measurements
can be carried out at different temperatures and frequencies. To evaluate the influence caused
by plastic deformation, local magnetic properties were determined with a double C-yoke SST.
The characterization of thin films was realized by applying VSM. In addition to scalar deter-
mination, magnetization components both parallel and perpendicular to the applied field were
determined by applying another pickup set. Furthermore, vector investigation was carried out
in rotational field measurements by mechanically rotating the tested sample.

1.4 Structure of Thesis

In the whole thesis, the SI system of units is applied. The thesis is structured as follows,

In chapter 2, the theoretical background of ferromagnetism is generally introduced. In

this thesis, most research efforts were focused on modeling and the characterization of
ferromagnetic materials. As the symbol of magnetism, hysteresis is discussed from its
physical origin, which is the foundation for many important hysteresis models, such as
the Preisach model. To include the Preisach model in FE analysis, the FE formulations
for electromagnetic field calculation are shown in the last section.

In Chapter 3, the proposed inverted Preisach model is described with details, e.g. the
modification of the relay operator and the analysis of the weight function distribution.
For comparison, the pseudo inverse is introduced in this chapter.

In Chapter 4, three applied characterization methods, the Epstein frame, double C-yoke
SST and VSM, are separately introduced. In addition to their main measurement prin-
ciple, various measurement results are shown to illustrate the sample properties.

In Chapter 5, the identification method and procedure are described. The accuracy of the
model can be evaluated from comparisons between measurement and simulation results.

In Chapter 6, calculation of the inverted model is discussed from a computational point

of view. The applied numerical method and procedures are introduced in this chapter.
The efficiency improvement is evaluated on the optimized procedure.

In Chapter 7, a summary and outlook is made with regard to this thesis.

7
2 Theoretical Background
2.1 Ferromagnetism
Ferromagnetic materials are widely applied attributing to their strong magnetism. To char-
acterize and simulate the magnetic behavior of ferromagnetic materials, it is necessary to
understand the physical background of ferromagnetism.

2.1.1 Classification of Magnetism

In physics, according to different magnetic behavior, substances are classified as diamagnetic,
paramagnetic, or ferromagnetic, etc [42]. As shown in Fig. 2.1, magnetic orders and temper-
ature dependency vary in different types of substances. In comparison with strong ferromag-
netism and ferrimagnetism, paramagnetism and diamagnetism exhibit a weak response to an
external magnetic field, which can only be detected by sensitive instruments.

Figure 2.1: Classification of magnetism. Magnetic orders are shown on the left, temperature
dependence of magnetic behavior is shown on the right. T is temperature, is
the magnetic susceptibility, Ms is the spontaneous magnetization. In (c) and (d)
substances are in zero applied field within a single domain (below Tc ), and in a
group of atoms (above Tc ) [42].

9
2 Theoretical Background

Table 2.1: Curie Temperatures for Some Ferromagnetic and Ferrimagnetic Materials
Ferromagnetic Materials Ferrimagnetic Materials
Substance Tc (K) Substance Tc (K)
Ni 627 MnFe2 O4 573
Fe 1043 NiFe2 O4 858
Co 1388 BaOFe2 O3 723
Gd 292 Y3 Fe5 O12 523
Dy 88 FeFe2 O4 858

Diamagnetic materials exhibit negative magnetism, which is independent of temperature

[see Fig. 2.1(a)]. As another weak magnetism, paramagnetism exhibits relatively strong tem-
perature dependency [see Fig. 2.1(b)]. According to the Curie law,
M C
= = , (2.1)
H T
where C is the Curie constant per gram [42]. The distribution of magnetic moments are dis-
ordered, which is due to the disturbance of thermal energy [24]. When temperature increases,
the susceptibility decreases linearly.
Compared with paramagnetic material, ferromagnetic material exhibits strong magnetism
even in the absence of an external magnetic field. It is due to spontaneous magnetization result-
ing from interatomic exchange interaction [43, 24]. As is shown in Fig. 2.1(c), spontaneous
magnetization (Ms ) decreases with the increase of temperature. Above a certain temperature
(Tc ), ferromagnetic materials begin to exhibit paramagnetic behavior, where the susceptibility
of ferromagnetic material is described in Curie-Weiss law
C
= , (2.2)
T Tc
where Tc is the boundary temperature and called Curie temperature. In Table. 2.1 the Curie
temperature of some ferro- and ferrimagnetic materials are listed. It is an important parameter
in design of equipment and metal processing.
In Fig. 2.1(d), it can be seen that like ferromagnets, ferrimagnetic materials exhibit sponta-
neous magnetization below the Curie temperature, but become paramagnetic above it. How-
ever, there are two different sublattices in ferrimagnetic materials shown in magnetic orders of
Fig. 2.1(d). On different sublattices the magnetic moments are opposed but not equal, which
leads to a spontaneous magnetization in ferrimagnetic materials.
In industry, ferrits are important materials alongside ferromagnets. In addition to their
economic advantage, they are mostly applied in high-frequency conditions due to their high
resistivity.

10
 2.1 Ferromagnetism

2.1.2 Magnetization and Domain Distribution

Magnetic Domains
As mentioned in the previous chapter, ferromagnetic materials still exhibit magnetization even
in the absence of an external magnetic field. However, ferromagnetic materials can be demag-
netized despite the internal molecular field [44, 24]. To explain this phenomenon, Pierre
Weiss assumed that material is divided into many small regions (about 1-100 m), which
exhibits spontaneous magnetization in different directions. The small regions are known as
domains, and the boundaries between different domains are called domain walls. When ma-
terial is demagnetized, multiple domains are formed, which results in zero net magnetization
[see Fig. 2.2(c)]. In saturation state, most of the domains are oriented along the external field
direction.

Figure 2.2: Diagram of magnetic domains with minimization of magnetostatic energy. Hs in

(a) is the stray field caused by spontaneous magnetization in magnet, (b) the forma-
tion of domain structure reduces the magnetic energy, (c) with multiple domains
the whole magnetization is zero, the magnet is in minimum energy state [45].

When a ferromagnet or a ferrimagnetic material is heated above its Curie temperature, the
uniform magnetization within a domain spontaneously disappears. Each atom has its own
direction of magnetic moment, which leads to paramagnetic behavior.
In 1935, it was mathematically proved that domains form to ensure the lowest magnetostatic
energy [46]. In 1949 the domain structure in silicon-iron single crystals was published by H.
J. Williams, R. M. Bozorth, and W. Shockley [47]. This was considered as direct experimental
evidence of the domain theory. Since that time, domain theory has been essential in explaining
the shape of a magnetization curve or the mechanism of magnetic hysteresis [42, 24].

Magnetization and Domain Distribution

When a ferro- or ferrimagnetic material is placed in an increasing magnetic field, its net mag-
netization is increased up to a limited value called the saturation magnetization. This process
is called technical magnetization as the increase in net magnetization is not the same as a

11
2 Theoretical Background

change in intensity of the spontaneous magnetization in Fig. 2.1(c) and (d). During the tech-
nical magnetization the net magnetization is changed by domain wall displacement and by
rotation of domain magnetization [43].
Fig. 2.3 presents domain arrangements in different states of magnetization. Suppose the
magnetization starts from a fully demagnetized state, then it is at the origin 0 in Fig. 2.3.
When a small positive magnetic field is applied, as shown at points 1 and 2, the domains
whose direction of spontaneous magnetization is close to the direction of the applied field
start to grow. At the same time, the other domains are gradually eliminated. When the applied
magnetic field is further increased, most of magnetizations are lined up almost in parallel as
shown at point 3. Finally the net magnetization changes to a lesser degree with the increase of
an applied field. The magnetization is realized by rotation of the direction of magnetization.
As shown at point 4, it is at saturation state. The curve from the demagnetized state to satu-
ration (path 0 4) is called the primary magnetization curve [24] or normal induction curve
[42].

Figure 2.3: Domain arrangements for various status of magnetization [24]. The material is in
demagnetized state at 0. When the applied magnetic field increases, the magneti-
zation increases along the path 0-1-2-3-4 until saturation state.

2.1.3 Hysteresis Curves and Magnetic Properties

Major Hysteresis Loops

After saturation has been reached in Fig. 2.3, if the applied magnetic field is removed, the
magnetization will not return back to zero. Instead, there is still some magnetization left in the
material. To drive the magnetization back to zero, a field in the opposite direction should be
applied to the material. This behavior is called magnetic hysteresis. It is considered as typical
symbol of magnetism and an important character of magnetic properties [18].

12
 2.1 Ferromagnetism

A hysteresis loop is a plot showing the variation of magnetization with magnetic field. In
material catalogues it is normally in terms of the flux density B (or the polarization J) and the
magnetic field strength H. According to (2.3), H forms part of B, after saturation, a constant
Ms is reached, but B continues to increase with magnetic field H [42].

B = 0 (H + M ) (2.3)

B = 0H + J (2.4)
where 0 is the permeability of free space.
As shown in Fig. 2.4, the magnetic flux density increases from the demagnetized state to
saturation along the path 0-1. It is similar to the magnetization curve shown in Fig. 2.3. By
measuring the primary magnetization curve the permeability of material can be obtained by
B
(H) = 0 r (H) = . (2.5)
H
where r is known as relative permeability.
If we start decreasing the magnetic field strength, due to an irreversible change of domain
walls locations, the return path is along the path 1-2 but not along the path 1-0. At zero
magnetic field (point 2), there is a retentivity or residual induction Br in the magnetic material.
To decrease the flux density back to zero (point 3), a magnetic field is applied in the opposite
direction. This field is called coercivity Hc . If we continue increasing the magnetic field in
a negative direction, the negative saturation Bs will be reached (point 4). If the magnetic
field returns back to zero then increases to reach the positive saturation state (point 1), the path
1-2-3-4-5-6 forms a closed loop known as the major hysteresis loop.

Minor Hysteresis Loops and FORCs

In the primary magnetization process, if a reverse happens at some intermediate point before
saturation, such as point a in fig. 1.5(a), a small closed loop an be obtained inside the major
loop. This can be achieved by symmetrically reversing the corresponding field. As shown in
Fig. 2.5.(a), the small loop whose tip is at point a is called a minor hysteresis loop. On this
minor loop, the rest induction at point b is called remanence. The magnetic field at point c
is called coercive field. They are named differently to the retentivity and coercivity of major
loops [42]. Besides symmetrical (symmetrical to origin) minor hysteresis loops (loop abc),
asymmetrical minor loops can also be formed at points d and e inside the major loop in Fig.
2.5(a).
In identification of hysteresis models (e.g. Preisach hysteresis model), as shown in Fig.
2.5(b), first-order reversal curves (FORCs) are usually measured by decreasing the field at
some points on the ascending major loop (or increasing the field at some points on the de-
scending major loop) [48].
In AC field measurements, primary magnetization curves are produced by joining tips of
minor loops. To achieve the demagnetized state, material is firstly magnetized to saturation.
By applying a series of alternating fields with slowly decreasing amplitude, the induction of

13
2 Theoretical Background

Figure 2.4: A typical major hysteresis loop. From demagnetized state at 0 to the saturation
state Bs at 1, the 01 path forms as a primary magnetization curve. When H = 0,
the magnetic flux at 2 is the remanent flux density Br . To decrease the magnetic
flux back to zero, coercive field strength Hc in negative direction is applied. The
path 1-2-3-4-5-6-1 forms a major hysteresis loop.

the material becomes smaller and smaller. Another demagnetization method is called thermal
demagnetization. As mentioned in Chapter 2.1.1, when ferro- or ferrimagnetic material is
heated above its Curie point, it becomes paramagnetic. After cooling it in the absence of a
magnetic field, the material can be demagnetized [42].

Magnetic Properties Determined from Hysteresis curves

The characterization of magnetic materials is usually performed by measuring hysteresis loops,
from which much information can be extracted regarding the magnetic properties. E.g. the
coercivity, retentivity, permeability, core loss, etc. According to these parameters, magnetic
materials are applied in various fields. Normally these parameters are presented in material
catalogues.
Of these parameters, the core loss Pc is particularly important, especially for electrical steels.
It describes the energy irreversibly transformed into heat during one magnetization cycle.
According to engineering concept, hysteresis loop areas are applied to estimate the loss per
unit volume V injected into a magnet in a magnetization cycle [18]. Normally, the core loss Pc
is determined as specific power loss P related to the mass m of the sample [24]. By introducing
the material density , the core loss Pc can be determined by
P V 1
I I
Pc = = H dB = H dB, (2.6)
m Tm T
where T is one cycle of the magnetization process. There are various reasons for such

14
 2.1 Ferromagnetism

Figure 2.5: Minor loops (a) and first order reversal curves (FORCs) (b) in major loops.

losses, eddy current is a significant factor. Therefore, core losses is usually specified at a
specified frequency and maximum flux density.
According to different magnetic properties, ferromagnetic materials are divided into three
major application categories. As shown in Fig. 2.6, hard magnetic material has a relatively
high coercive field and residual magnetization. It is desirable for permanent magnets because
it is difficult to magnetize (or demagnetized) and it remains a largely saturated magnetization
when the driving field is removed. In contrast, soft magnetic materials have a very low coer-
cive field but high permeability. They are mostly used in AC applications such as magnetic
cores of transformers, motors and inductors, as low hysteresis loss is preferred and required in
order to minimize energy dissipation. Occupying the middle ground between the former two
materials, semi-hard magnetic materials have a moderately high coercive field and residual
magnetization. For this reason, they are mostly applied in magnetic recording.

Figure 2.6: Major hysteresis loops of Soft, semi-hard and hard magnetic materials.

The limit to distinguish between the three kinds of materials is not so rigid. Typically,
the materials with coercivity less than 1000A/m belong to soft magnetic materials. For hard

15
2 Theoretical Background

magnetic materials, coercivity is greater than 50000A/m. Sometimes the same material may
be either magnetically soft or hard, depending on its physical condition [43, 42].

2.1.4 Magnetic Anisotropy

There are several factors which affect the shape of magnetization curves or hysteresis loops, in-
cluding temperature and mechanical stress. One strong factor is known as magnetic anisotropy,
which means the magnetic properties depend on the direction of magnetization.
Consider a single crystal of iron with body centered cubic structure. As shown in Fig.
2.7, magnetization is applied along [1 0 0], [1 1 0] and [1 1 1] crystallographic directions. The
shapes of the magnetization curves obviously differ according to the three crystallographic
directions. Along the [1 0 0] axis the magnetization is easiest, in other words, the saturation
can be reached with a low magnetic field. This direction is called easy axis. In contrast,
the [1 1 1] axis is called the hard axis because along this axis the magnetization is most
difficult. Such dependence of magnetic behavior on crystallographic directions is known as
magnetocrystalline anisotropy [42]. It is intrinsic to material and assumed to be the result of
spin-orbit-lattice interaction [24].

Figure 2.7: Magnetization curves of single crystal of iron [24].

From an energy point of view, the magnetization or demagnetization along the easy axis
cost less energy than along noneasy axis. This can be applied to minimize the energy loss
of electrical steel in transformers. As it is expensive to produce the single crystal alloys, the
practical way is to develop anisotropy in polycrystalline alloys. Goss patented an invention of
textured silicon-iron electrical steels [24]. By applying rolling treatments to steels, one easy
axis of the crystals is positioned close to the rolling direction (RD). In transformers magnetic
field is applied parallel to the RD of the electrical steels.
Furthermore, anisotropy can be inducted by plastic deformation, annealing, irradiation, etc
[42]. For magnetic thin films, uniaxial anisotropy can be induced by applying a magnetic field
during the deposition process. If the magnetic behavior is the same in all directions, the mate-
rial is characterized as isotropic. However, due to the demagnetization field, isotropic material

16
 2.2 Preisach Model of Hysteresis

with different sample shapes may also exhibit anisotropy phenomenon, which is called shape
anisotropy. This is the reason why sample shapes should be indicated in magnetic characteri-
zation.

2.2 Preisach Model of Hysteresis

Various models have been proposed for simulating hysteresis behavior [10, 16, 48]. Some of
them are generally based on real physics of the magnetization process, while some are purely
mathematical or phenomenographic models. Many hysteresis models are evaluated by Ivanyi
in [10].
Among these hysteresis models, the Preisach model is widely applied, due to its accuracy
and generality. The classic Preisach model is a static (rate-independent) scalar model, ferro-
magnetic material is assumed to be assembly of independent domains. The model can trace
minor loops and other complex magnetizing processes [48] but it is usually described as a
phenomenological rather than physics-based model.

2.2.1 The Relay Switch Operator

The classic Preisach model was initially proposed by F. Preisach in 1935 [11]. It assumes that
the magnets are composed of separate magnetic particles (or elementary domains) . These
particles are assumed to exhibit simplest hysteresis behavior. As shown in Fig. 2.8(a), these
particles are driven by the same external magnetic field H and have rectangular hysteresis
loops. Each hysteresis loop has the same saturation magnetic moment ms but different switch
magnetic fields H and H . The switching happens when
(
+ms H > H ,
m(H , H ) = (2.7)
ms H < H .

When H is between H and H , the state of the magnetic moment depends on the history
of applied field H. To describe this simplest hysteresis behavior mathematically, the relay
switch operator is applied as an elementary hysteresis operator. It can be represented by a
rectangular loop shown in Fig. 2.8(b). The switching happens at and values of the input.
The relay switch operator is defined as

+1
if H(t) ,

[H](t) = [H](t ) if < H(t) < , (2.8)

1 if H(t) .

2.2.2 The Classic Preisach Model

The classic Preisach model was generalized in a mathematical idea by Krasnoselskii in the
1970s. It can be applied as a mathematical tool to any physical hysteresis phenomenon.

17
2 Theoretical Background

Figure 2.8: (a) is the rectangular hysteresis loop of each Preisach particle, (b) is the diagram
of hysteresis switch operator [H].

It was significantly improved by Mayergoyz in [16]. The Preisach model is defined as a

superposition of relay switch operators to calculate the overall magnetization. Each relay
operator is statistically weighted by a weight function (, ) dependent on and . The
Preisach model is expressed as:
ZZ
M(t) = [H](t)
b = (, ) [H](t)dd , (2.9)

where b is applied as a notation of the Preisach hysteresis operator, which takes the magnetic
field H as input, the magnetization M as the output1 .
The mechanism of the Preisach model can be simply illustrated by a diagram shown in Fig.
2.9(a). Relay operators are connected in parallel. Depending on the input H(t) at a certain
time t, some relay hysterons switch up (+1), while others switch down (1). Each output is
multiplied by their own weight (, ). Finally, these individual outputs are integrated over
the whole valid domain.
Each
 elementary 2
unit can be represented by a point (, ) in the Preisach plane
hysteresis
P = (, ) R : [16, 49]. Depending on the output of the relay operator at different
points, the Preisach plane can be separated into two subsets S+ (t) and S (t) for hysteresis units
in the up or down states [see Fig. 2.9 (b)].
(
S+ (t) if (, ) P [H](t) = +1,
(, ) (2.10)
S (t) if (, ) P [H](t) = 1.

1
b is distinguished from the inverted Preisach hysteresis operator ,
b which takes the magnetization M as input,
the magnetic field H as the output. The inverted Preisach model is introduced in the following chapters.

18
 2.2 Preisach Model of Hysteresis

Figure 2.9: Schematic interpretation of the Preisach model. (a) is the block diagram describing
the Preisach model in (2.9). (b) shows the Preisach plane.

In calculation, (2.9) can be expressed as

ZZ ZZ
[H](t)
b = (, )dd (, )dd . (2.11)
S+ (t) S (t)

As shown in Fig. 2.9 (b), the subsets S+ (t) and S (t) are divided by an interface L(t) At a
certain time t 0 , the interface L(t 0 ) reflects the state of the Preisach operator [H](t
b 0 ). During a

period of time t [t0 , tn ], the movement of L(t) reflects the system dynamics [49].

2.2.3 Geometric Interpretation of the Preisach Model

The Preisach model can be geometrically interpreted with an example shown in Fig. 2.10.
The input H(t) starts with a value H0 less than Hs [see Fig. 2.10 (a)]. All relay operators are
equal to 1. In other words, S (t) occupies the whole Preisach plane S. This corresponds to
the negative saturation state [16]. The input H(t) is subdivided into four segments with local
extrema {H1 , H2 , H3 , H4 }.
S-I. When H(t) increases from H0 to H1 , according to (2.19), the relay operators in area
H1 switch up [see S+ in Fig. 2.10 (c)- I ]. At the same time, the operators in area
H1 switch down. In the area < H1 < , which can be called the memory area,
the relay operators keep the old value at previous instants of time. As it starts from a
negative saturation state, the operators in memory area keep 1 state. The interface L(t)
is horizontal and moves upwards as H(t) is increased.
S-II. When H(t) decreases from H1 to H2 , in the memory area < H2 < , the relay op-
erators should keep the state in Fig. 2.10 (c)- I ]. Geometrically, this decrease forms
a vertical link of L(t) and it moves from right to left. In this way, the area of S+ (t)
becomes smaller, the reduction area is occupied by S (t) area. This results in a reversal
curve at H = H1 [see Fig. 2.10 (b)- II ].

19
2 Theoretical Background

Figure 2.10: Geometric interpretation of the Preisach model. (a) shows the process of input
H(t), the whole process is subdivided into four segments I-IV. (b) shows the
corresponding output [H](t).
b (c) are the Preisach planes in different segments
I-IV, where L(t) is the interface between S+ and S .

S-III. In this segment, H(t) increases from H2 to H3 . The mechanism of memory formation
is the same as mentioned above. With the increase of H(t), on the right of dashed line
= H2 , the horizontal interface moves upwards until = H3 . The interface L(t) is a
staircase line whose vertices have and coordinate as (H1 , Hs ), (H1 , H2 ), (H3 , H2 )
and (H3 , H3 ). They are local extrema of the input history. By this method, the Preisach
model keeps and updates the memory (local extrema) in the interface line L(t).

S-IV. In this segment H(t) continues to increase until H(t) = H4 . This monotonic increase
of input H(t) results in a horizontal final link of L(t) which moves upwards until H4
is reached. As H4 > H1 > H3 > H2 , all vertices whose -coordinates are below H4
have been wiped out [see Fig. 2.10 (c)- IV]. In other words, the memory kept in L(t) is
erased. This wiping out behavior also occurs in the monotonically decreasing condition.
It is known as the wiping out property of the Preisach model.

20
 2.3 Finite Element Analysis of Magnetic Field

As can be seen in Fig. 2.10 (b)- IV, a minor loop is formed at H = H1 . It is assumed
that the minor loops developed by back and forth variations between two input values
should be congruent, which is considered as the congruency property of the Preisach
model.

The above-mentioned wiping out and congruency properties are the necessary and sufficient
conditions for a hysteresis nonlinearity to be described by the Preisach model [16].

2.3 Finite Element Analysis of Magnetic Field

Accurate analysis of magnetic field is essential for the evaluation and design of electrical
devices. The finite element (FE) method is one of the most widely used numerical techniques
in computational electromagnetism. It has been successfully applied to solve magnetic fields
in complicated geometries such as electrical machines and transformers [1].
In the first part of this section, the partial differential equation (PDE) of electromagnetic
fields is introduced. In the second part, common methods for the approximation of nonlinear
hysteretic relation are presented. The description provided indicates why hysteresis models
need to be inverted when it is included into the FE analysis.

2.3.1 Partial Differential Equations of Electromagnetic Analysis

The problem of numerical analysis on a macroscopic level is solving partial differential equa-
tions (PDE) under prescribed boundary conditions [2]. The FE method is widely applied
to approximate the solution of the partial differential equations. The PDEs for magnetic field
analysis are derived from Maxwells equations, which were proposed by James Clerk Maxwell
in 1862. Based on the work and experiments of Ampere, Faraday and Gauss, Maxwells equa-
tions state the relationships between the fundamental electromagnetic quantities [50].
In static and quasistatic (eddy current) magnetic fields, the displacement current can be
neglected. The governing parts of Maxwells equations are expressed as follows

H = J, (2.12)
B
E = , (2.13)
t
B = 0, (2.14)

where E is the electric field intensity. The electric current density J is applied here as
the source of electromagnetic field. In addition, the constitutive relations are appended to
Maxwells equations to describe the properties of the media.

J = E = (Ei + Es ), (2.15)
B = H = 0 (H + M ). (2.16)

21
2 Theoretical Background

Here it is assumed there are no moving bodies in the simulation area. Ei is the impressed (or
irrotational) electric field intensity, Es is the solenoidal electric field intensity. Normally, the
magnetic reluctivity = 1/ is applied instead of the permeability in calculation. is the
electric conductivity of simulated media.
By introducing scalar or vector magnetic potentials, the Maxwells equations can be trans-
formed into the desired PDEs [2]. As commonly be applied in FE method, the vector magnetic
potential A is defined by
B = A. (2.17)
Applying (2.16) and (2.17) to the first part of the above mentioned Maxwells equations,
the PDE for the quasistatic electromagnetic fields is

A
+ A = Ji . (2.18)
t
In electromagnetic modeling, Ji refers to impressed current density due to a given electric
potential difference (the current or voltage loaded coil) [50]. It can be obtained by Ji = Ei .

2.3.2 Nonlinear Electromagnetic Field Computation

In the PDE applied in FE analysis (2.18), the magnetic vector potential is employed as the
unknown. According to (2.17), the magnetic flux density can be directly obtained as an out-
put quantity. In order to calculate the magnetic field strength, the constitutive relationship is
required for application.
Most paramagnetic and diamagnetic materials, such as air, aluminum and copper, are con-
sidered as magnetically linear mediums. The permeability or reluctivity is set to be constant.
However, ferromagnetic materials are usually involved in electromagnetic simulation, the cal-
culation become more complex because of their nonlinear characteristics. Normally the non-
linearity is approximated by a single-valued magnetization curve [see Fig. 2.11 (a)] in the FE
analysis. As shown in Fig. 2.11 (b), the nonlinear relation (b) can be obtained from this
magnetization curve for simulation purposes. With this method, the hysteresis characteristic
is completely neglected. To achieve more accurate results, hysteresis must be taken into con-
sideration, especially in calculating the hysteresis losses in electrical machines. This problem
can be solved by including hysteresis models in the FE analysis.
Most hysteresis models, such as the Preisach model and the Jiles-Atherton model, employ
magnetic field H as an input variable and take magnetic flux B (or magnetization M) as the
output quantity. These models can be called H-based or forward hysteresis models. To include
hysteresis models in B-oriented FE analysis, hysteresis models need to be inverted before they
are implemented in FE formulations. Here we introduce a operator H 1 (B) to represent the
inverse processing of the hysteresis models.
According to the mentioned conditions, the constitutive relationship is summarized as

B
in magnetically linear medium,
H = (B)B in magnetically nonlinear medium, hysteresis is ignored, (2.19)
H (B) in magnetically nonlinear medium, hysteresis is considered.
1

22
 2.3 Finite Element Analysis of Magnetic Field

1.6 1400

1.4 1200

Reluctivity (Am/Vs)
Magn. flux B (T)
1.2 1000

1 800

0.8 600

0.6 400

0.4 200

0.2 0
0 500 1000 1500 2000 0 0.4 0.8 1.2 1.6
Magn. field H (A/m) Magn. flux B (T)
(a) (b)

Figure 2.11: The single-valued magnetization curve (a) and obtained reluctivity curve (b).
Measurement was carried out on electrical steel M33035A with Epstein frame
method at 50Hz.

The block diagram in Fig. 2.12 illustrates the necessary and two main approaches to inverse
hysteresis models. Iterative methods are typically employed in the local inversion of forward
hysteresis models. The fixed point method and the Newton-Raphson method are the most
commonly applied techniques. However iteration methods always require much computation
to obtain converged solutions, especially in transient simulations. It is more efficient to find a
inverted hysteresis model, which applies B as input to calculate H. This inverted model can
be directly implemented in FE analysis without iterations.

Figure 2.12: Block diagram of including hysteresis models into electromagnetic FE analysis.

In this chapter, general introductions have been made to the physical origin of ferromag-
netic hysteresis, definition of the classic Preisach model and the FE analysis of electromagnetic
field. These theories have been applied in the characterization and modeling of the ferromag-
netic materials in this work. In the next chapter, an inverted Preisach model, which is based
on the classic Preisach model, is proposed and introduced with details.

23
3 The Inverted Preisach Model
Most hysteresis models employ the magnetic field as the input variable and take the magne-
tization or magnetic flux as the output quantity. These models are considered as forward (or
H-based) hysteresis models. The previous chapter explained why (forward) hysteresis models
need to be inverted in order to include them in the B-oriented FE analysis [1, 5]. Usually it-
erative methods are utilized in the local inversion of forward hysteresis models [5]. However,
such methods require much computation for converged solutions.
In order to circumvent the iterative procedure, much research effort has been made to find
a hysteresis model, which takes the magnetization or magnetic flux as input. With this model
the magnetic field can be directly calculated without iteration. In comparison with the forward
hysteresis models, the roles of input and output are exchanged in this model. Therefore, this
type of model is considered as inverted hysteresis model.
As a result of its accuracy, simplicity and efficiency, the classical Preisach model (CPM)
has been widely applied within FE applications. The inverse of the CPM is defined by Dlala
[5] in a simplified manned as

H(t) = [M](t)
b b1 [H](t),
= b (3.1)

where b is the CPM operator in (2.9) introduced in section 2.2.

b is defined as the inverted
Preisach operator. Mathematically, b can be considered as the inverse operator b1 . This
chapter introduces two different inverted Preisach models. One uses the relay switch operator
of CPM and known as the pseudo-inverse approach. The other one is a new inverted Preisach
model using a modified switch operator1 .

3.1 The Pseudo-inverse of the Preisach Model

In several published inversion works [29, 54, 1], the inverted model is still considered as
a superposition of infinite relay operators . As shown in (3.2), the magnetic field H is
calculated from the magnetization M. Except the exchanged roles of input and output, the
expression looks similar to the forward Preisach model in (2.9), the only difference is it uses
a different weight function (, ) to realize the inverse operation.
ZZ
H(t) = [M](t)
b = (, ) [M](t)dd , (3.2)

1 The inverted Preisach model introduced in this chapter has been published in [51, 52, 53].

25
3 The Inverted Preisach Model

where the relay switch operator works in the same way as the one in the forward Preisach
model (2.7). The integration is also fulfilled in the Preisach plane P = (, ) R2 : .


+1
if M(t) > ,
[M](t) = [M](t ) if < M(t) < , (3.3)

1 if M(t) < .

3.1.1 The Weight Function in the Pseudo-inverse

By using the relay switch operator , the pseudo-inverse of the Preisach model maintains
the basic properties of the Preisach model, such as the wiping out and congruency properties.
The past extrema values of input are also stored in memory. These properties provide a strong
advantage in numerical calculation. However, the modeling of the inverted weight function
(, ) is difficult, as the weight of each relay operator is distributed in a complicated
manner in the Preisach plane.
The distribution of weight function can be analyzed with hysteresis reversal curves. To
analyze the inverted weight function (, ), a H(M) reversal curve is used and shown in
Fig. 3.1(a). The input M(t) increases from the negative saturation to 1 then to 2 . The
corresponding outputs at 1 and 2 are H1 and H2 . At the point (2 , H2 ), a reversal curve
is formed by decreasing the input. Finally, the input decreases to 1 (1 = 1 ). As the output
at 1 is on the reversal curve from H2 , it is labeled as H2 1 .

Figure 3.1: The distribution of inverted weight function in the Preisach plane. (a) is a H(M)
reversal curve, (b) is the corresponding Preisach plane. The Inputs M(t) firstly
increase from negative saturation to 1 , then decrease from 2 to 1 (1 = 1 ).

According to the work principle of the relay switch operator described in (3.3), when
input increases from 1 to 2 , the change of the output can be calculated by the integration of
the weight function in the Preisach plane.
ZZ
H2 H1 = 2 (, )dd > 0. (3.4)
S1+S2

26
 3.1 The Pseudo-inverse of the Preisach Model

When input decreases from 2 to 1 , the operators in triangle S1 switch from +1 to

1. The output decreases from H2 to H2 1 . As 1 = 1 and H2 1 < H1 , the integration of
(, ) in area S2 can be calculated as
ZZ ZZ ZZ
(, )dd = (, )dd (, )dd
S2 S1+S2 S1
H2 H1 H2 H2 1
=
2 2
H2 1 H1
= < 0. (3.5)
2
The integration of (, ) in area S2 is negative, as the integration of (, ) in area S1 + S2
is proved to be positive in (3.4), it can be assumed that the integration in area S1 should be
quite large.
The weight function of Preisach model is supposed to be symmetric in the Preisach plane
[16]. From the above analyzed example, we can deduce that the integration of the weight
function has a sharp increase over the area near the line = . In the area close to the point
(0 , 0 ), the weight function has a negative distribution.

60
Magn. field H (kA/m)

Norm. appro. (, )

1
30

0 0

30 1
1
1
60 0 0
2 1 0 1 2
Magn. polarization J (T) 1 1
(a) (b)

Figure 3.2: (a) shows the six H(J) reversal curves measured with VSM on 1.9 m Fe49 Co49 V2
thin film sample. (b) shows the calculated weight function distribution from mea-
sured reversal curves in (a).

According to the analysis method applied above, six measured H(J) reversal curves shown
in Fig. 3.2(a) are used to calculate the weight function distribution. The resulted weight func-
tion distribution is presented in Fig. 3.2(b). It is obvious that the inverted weight function
(, ) is partly positive and partly negative in the Preisach plane. The complicated distri-
bution makes it difficult to find an analytical weight function. Improper modeling may cause
inaccuracy and nonsmoothness of the results. As a result, it is necessary to find an alternative
method for calculation with the pseudo-inverse of the Preisach model.

27
3 The Inverted Preisach Model

3.1.2 The Inverted Everett Function

The Everett function method has been widely applied in numerical implementation of the
forward Preisach models [16, 1]. With this method, it is not necessary to determine the weight
function nor to calculate the double integrals in (2.9). By introducing this approach to the
pseudo-inverse of the Preisach model, the problems of the inverted weight function can be
solved. To distinguish it from the Everett function for forward models, the Everett function
for inverted models is called the inverted Everett function and is defined as
1
ZZ
E(, ) = (H H ) = (, ) dd . (3.6)
2
T (, )

Similarly to the integral over S1 in Fig. 3.1, E(, ) is equal to the integral over a triangle
T (, ), which is related to the output increments along reversal curves.
In numerical calculation, the integral expression of the Preisach model is usually rewritten in
a discrete form. Mayergoyz has proposed a discrete expression of the forward Preisach model,
which is calculated with the Everett functions [16]. As the pseudo-inverse of the Preisach
model remains the basic properties of the forward Preisach model, its discrete expression is
formed with reference to the Mayergoyzs method.
As shown in Fig. 3.3, the Preisach plane can be separated into two subsets S+ (t) and S (t)
by a staircase interface L(t). The relay switch operators in the S+ (t) area are equal to +1,
while the ones in the S (t) area are equal to 1. Therefore, the pseudo-inverse of the Preisach

Figure 3.3: Geometric interpretation of the trapezoid areas in a Preisach plane. L(t) is the
staircase interface dividing S+ and S regions. (a) The input M(t) is decreasing, it
forms a vertical final link at last vertex of L(t). (b) The input M(t) is increasing, it
forms a horizontal final link at last vertex of L(t).

28
 3.1 The Pseudo-inverse of the Preisach Model

model in (3.2) can be expressed as

ZZ ZZ
H(t) = (, )dd (, )dd
S+ (t) S (t)
ZZ ZZ
= (, )dd + 2 (, )dd . (3.7)
T S+ (t)

The integration of (, ) in the whole Preisach triangle T can be calculated by an inverted

Everett function ZZ
(, )dd = E(0 , 0 ) (3.8)
T

To calculate the integration of (, ) in S+ (t) area by the inverted Everett functions, as

shown in Fig. 3.3, the S+ (t) area can be subdivided into trapezoids (and a triangle) [16]. All
the vertices on the interface L(t) are extrema of the input. Their coordinates in the Preisach
plane are represented by Mk and mk . The notation Mk represents the input maximums and mk
represents the input minimums. When the input M(t) is decreasing, as shown in Fig. 3.3(a),
it forms a vertical final link at last vertex of L(t). The S+ (t) area can be subdivided into
n(t) trapezoids Qk . When the input M(t) is increasing, as shown in Fig. 3.3(b), it forms a
horizontal final link at last vertex of L(t). The S+ area can be divided into n(t) 1 trapezoids
and 1 triangle.
When M(t) is decreasing, the integration of (, ) in the S+ (t) area can be expressed as
ZZ n(t) ZZ
(, )dd = (, )dd . (3.9)
k=1 Q
S+ (t) k

When the input M(t) is increasing, the integration of (, ) in the S+ (t) area can be ex-
pressed as
ZZ n(t)1 ZZ ZZ
(, )dd = (, )dd + (, )dd . (3.10)
k=1
S+ (t) Qk Tn(t)

Each trapezoid Qk is geometrically a difference of two triangles T (Mk , mk1 ) and T (Mk , mk ).
As the integral over a triangle can be calculated by the inverted Everett function (3.6). The
integral over a trapezoid Qk can be calculated as
ZZ ZZ ZZ
(, )dd = (, )dd (, )dd
Qk T (Mk ,mk1 ) T (Mk ,mk )
= E(Mk , mk1 ) E(Mk , mk ). (3.11)

Combine (3.7), (3.8), (3.9) and (3.11), in the decreasing case, the output H(t) can be calcu-

29
3 The Inverted Preisach Model

lated as
n(t)
H(t) = E(0 , 0 ) + 2 [E(Mk , mk1 ) E(Mk , mk )]. (3.12)
k=1

In the increasing case, the output H(t) can be calculated as

n(t)1
H(t) = E(0 , 0 ) + 2 [E(Mk , mk1 ) E(Mk , mk )] + 2E(M(t), mn(t)1 ). (3.13)
k=1

In FE applications, the inverted Everett functions can be precalculated at all discrete points
of the Preisach plane. All E(, ) results are saved as an inverted Everett matrix. In the
calculation procedure, (3.12) and (3.13) are calculated by interpolating in the inverted Everett
matrix. It is relatively fast and saves considerable memory space.
From a computational point of view, this approach is an efficient way of including the
Preisach model in FE calculation. However, the inverted model with relay switch operator is
not considered as an exact inverse, it is called pseudo-inverse of the Preisach model in [1, 7].
To calculate the inverted Everett matrix, a group of FORCs are required for identification2 .
The measurement of FORCs is difficult and tedious. The errors caused in measurements can
lead to inaccuracy of the simulation.

3.2 The Inverted Model with Modified Switch Operator

As analyzed in the previous section, using the relay switch operator in the inverted model
leads to a complicated weight function. In this section, a modified switch operator is
introduced to the inverted model in (3.14).
ZZ
H(t) = [M](t)
b = (, ) [M](t)dd . (3.14)

By means of the new switch operator, the weight function is positive in the whole Preisach
plane and it can be modeled with an analytical weight function. Furthermore, the inverted
model also keeps the wiping out and congruency properties of the Preisach model. More
details are explained in the following subsections.

3.2.1 The Modified Switch Function

In the forward Preisach model, H is applied as input, the overall magnetization M is calculated
by superposition of infinite relay operators. Fig. 3.4(a) illustrates a typical M(H) loop. The
relay switch operator is applied to model the rectangular hysteresis loop of elementary
domains. As shown in Fig. 3.4(b), and correspond to up and down switching values
2 More FORCs, which are uniformly distributed inside the major loop, can improve the accuracy of simulation
results. In this work at least 20 FORCs are measured to calculate the inverted Everett matrix.

30
 3.2 The Inverted Model with Modified Switch Operator

[16]. As the switch up value is always greater

than the switch down value , the
2
Preisach plane is in area P = (, ) R : shown in Fig. 3.4(c).
In the inverted Preisach model, M is applied as input, similarly, the overall magnetization H
is calculated by superposition of infinite switch operators. Fig. 3.4(d) shows a typical H(M)
loop. Compared with the M(H) loop, the H(M) loop has different ascending and descending
paths. A new switch operator is presented to model the H(M) loop. As shown in Fig.
3.4(e), in the rectangular loop of , the switch up value is smaller than the switch
down value . This switch behavior is similar to the ascending and  descending path of
0 2

H(M) loop. The corresponding Preisach plane is located in area P = (, ) R :
shown in Fig. 3.4(f).

Figure 3.4: Forward and inverted Preisach models. (a), (c) and (e) are the general M(H) curve,
switch operator and Preisach plane of the forward hysteresis model. (b), (d)
and (f) are the general H(M) curve, switch operator and Preisach plane of the
inverted hysteresis model.

The switch operator can be mathematically expressed as

+1
if M(t) ,

[M](t) = 0 0 [M](t ) if < M(t) < , (3.15)

1 if M(t) .

It can be noticed that is equal to the negative value of the 0 0 . Here a coordinate con-
version is required.

0 = Mmax + M(t),
0 = M(t) Mmax , (3.16)

where Mmax is the maximum of the input (saturation magnetization). As derives from the
modification of , in this work it is called the modified switch operator.

31
3 The Inverted Preisach Model

The work principle of can be explained clearer with the geometrical interpretation
shown in Fig. 3.5. In the region M(t) , is equal to +1, which is represented with dark
color in the Preisach plane shown in Fig. 3.5. In the region M(t) , is equal to 1.
which is represented with light color in the Preisach plane. In the region < M(t) < in the
Preisach plane, this region is called memory area. The output of depends on the history
of 0 0 . As shown in Fig. 3.5, the regions with skew lines are the memory areas for switch
operators ( : < M(t) < ; : < M(t) < ).

Figure 3.5: Geometrical interpretation of in the Preisach plane. In (a) the input M(t)
increases from negative saturation to M1 . In (b) the input M(t) decreases from M1
to M2 . The regions with skew lines are the memory areas.

When the input M(t) increases from negative saturation to M1 , as shown in Fig. 3.5(a), all
the in the memory area remain 1 state. According to (3.15), all the in memory area
switch to +1.
When the input M(t) decreases from M1 to M2 , as shown in Fig. 3.5(b), in the memory
area keeps the old state value as in Fig. 3.5(a), whereas negates the state value of 0 0
according to (3.16).
Like the interpretation of the classic Preisach model, here an input M(t) with local extrema
{M1 , M2 , M3 , M4 } is applied to the inverted model, the corresponding output H(t) and the
state of the switch operator b in the Preisach plane are shown in Fig. 3.6. The input starts with
a value less than Ms . The whole process is subdivided into four subsections (I-IV). As the
value of in the memory area is obtained according to 0 0 , the state of the switch operator
is also shown as a reference in Fig. 3.6(d).
In all subsections, the work principle of is the same as introduced in the classic Preisach
model (Fig. 2.10). The interface G(t) keeps geometrically reversed from L(t). In the last sub-
section IV, when M(t) increases to M4 , a minor loop is formed at M(t) = M1 [see Fig. 3.6 (b)-
IV]. The minor loops are assumed to be congruent, which is similar to the congruency prop-
erty described in the Preisach model. Furthermore, as M4 > M1 > M3 > M2 , in the Preisach
plane shown in Fig. 3.6(d)- IV the past extrema M1 , M2 and M3 are wiped out from G(t). This
is consistent with the wiping out property of the Preisach model.
From the interpretation, it can be concluded, by means of the modified switch operator
, that the inverted model remains the congruency and wiping out properties of the Preisach

32
 3.2 The Inverted Model with Modified Switch Operator

Figure 3.6: Geometric interpretation of the inverted Preisach model. (a) shows the process of
input M(t), the whole process is subdivided into four segments I-IV. (b) shows
the corresponding output [M](t).
b (c) are in the Preisach planes, where the
magenta line G(t) is the interface between S+ and S . (d) are in the Preisach
planes, where the green line L(t) is the interface. G(t) and L(t) geometrically
related.

model. By means of coordinate conversion in the Preisach plane, the staircase interface G(t)
can be obtained from the past input extrema. These properties can be utilized in the numerical
implementation of inverted model in an FE calculation process.
It is simpler and clearer to use the staircase interface to obtain the output of . As

33
3 The Inverted Preisach Model

shown is Fig. 3.7, for , the interface L(t) starts from the top-left point3 and ends at the
= = M(t). For , the interface G(t) starts from the = = M(t) and ends at the
bottom-right point. Geometrically L(t) and G(t) are completely related to each other. When
the shape of L(t) is known, it is easy to obtain G(t).

Figure 3.7: The geometrical relation between L(t) and G(t). (a) L(t) is the S+ (t) and S (t)
interface by applying the relay switch operator . (b) G(t) is the S+ (t) and S (t)
interface by applying the relay switch operator . (c) illustrates the geometrical
relation between the two interfaces.

By applying , the extrema of input history are saved in the vertices coordinates of L(t).
As the vertices of G(t) can be obtained from L(t) by coordinate conversion in (3.16), the
modified switch operator also keeps the input history in memory. The past extrema values
of input shape the staircase interface G(t).

3.2.2 Modeling of the Inverted Weight Function

In order to find a suitable weight function for the inverted model, analysis was made by means
of the H(M) reversal curves. As shown in Fig. 3.8, the magnetization between Ms and Ms is
uniformly discretized into n parts with interval M. The Preisach plane shown in Fig. 3.8(b)
can be discretized into n cells in both and directions. In each cell, k = k = Ms + kM.
Here a simple example n = 4 is explained.
Fig. 3.8(a) shows the H(M) reversal curves. Hk is the output when M increases from Ms
to k . All Hk are distributed along the ascending branch of the H(M) major loop. Hk j
( j < k) is the output when M decreases from k to j . They are distributed along the reversal
curves starting at Hk .
The change of output H can be applied to analyze the integral of the weight function in
the Preisach plane. Owing to the symmetry of the hysteresis loops, the weight function is
supposed to be symmetric with respect to the line = . In Fig. 3.8(b), the weight function
integrals in cells S1 -S6 are analyzed.
3 It is different from the interface introduced in the classic Preisach model in section 2.2. The interface L(t) in
the CMP separates the S+ (t) and S (t) subsets in the Preisach plane. Here the interface L(t) is applied to
obtain G(t). It always starts from the top-left point of the Preiach plane.

34
 3.2 The Inverted Model with Modified Switch Operator

Figure 3.8: Analysis of weight function with H(M) reversal curves. The input M is uniformly
discretized into 4 parts. (a) shows the H(M) reversal curves, analysis required H
outputs are noted on the curves. The Preisach plane shown in (b) is discretized into
10 cells. The weight function integral over cells S1 -S6 are analyzed. The integrals
over red cells are greater than the integrals over the gray cells.

When M decreases from 4 to 1 , the output is H4 1 . According to the work principle of

the inverted switch operator, only cell S1 is included in the subset S+ of the Preisach plane.
If M keeps decreasing until Ms , all the switch operators switch to 1. The weight function
integral over S1 can be calculated as
ZZ
H4 1 (Hs ) = 2 (, )dd . (3.17)
S1

When M decreases from 3 to 1 , the output is H3 1 . In the Preisach plane cells S1 and S2
are included in the subset S+ . The weight function integral over S1 + S2 can be calculated as
ZZ
H3 1 (Hs ) = 2 (, )dd . (3.18)
S1 +S2

By subtracting (3.17) from (3.18), the weight function integral over S2 can be calculated as
1
ZZ ZZ ZZ
(, )dd = (, )dd (, )dd = (H H4 1 ). (3.19)
2 3 1
S2 S1 +S2 S1

Using this method, the weight function integrals over different cells can be calculated with
the output values on reversal curves. The calculations of the weight function integrals over
cells S1 -S6 are listed in Table 3.1. With reference to the H(M) reversal curves shown in Fig.
3.8(a), the weight function integrals over cells S1 and S4 are relatively greater than other cells.
As weight function is symmetric with respect to the line = , the cells with red color in

35
3 The Inverted Preisach Model

Table 3.1: Integrals of (, ) in regions S1 -S6 in Fig. 3.8

Regions Integrals

(, ) dd = 21 (H4 1 + Hs )
RR
S1
S1

(, ) dd = 21 (H3 1 H4 1 )
RR
S2
S2

(, ) dd = 21 (H2 1 H3 1 )
RR
S3
S3

(, ) dd = 21 (H1 H2 1 )
RR
S4
S4

(, ) dd = 21 (H4 2 H3 1 )
RR
S5
S5

(, ) dd = 21 (H3 2 H4 2 H2 1 + H3 )
RR
S6
S6

Fig. 3.8(b) represent greater integral results.

According to the analysis example shown in Fig. 3.8, it can be assumed that the weight
function (, ) becomes very large in areas near vertices of the Preisach triangle. From the
vertices to the center of the Preisach triangle, (, ) gradually decreases to a minimum. Fig.
3.9(b) presents the approximated weight function (, ) calculated from the H(J) reversal
curves shown in Fig. 3.9(a). According to the shape of this surface, the weight function
(, ) can be modeled based on a hyperbolic paraboloid function.

60
Norm. appro. (, )

1
Magn. field H (kA/m)

30
0.5
0

30 0
1
0
60 1
2 1 0 1 2 1 1 0
Magn. polarization J (T)
(a) (b)

Figure 3.9: (a) shows six H(J) reversal curves measured with VSM on 1.9 m Fe49 Co49 V2
thin film sample. (b) shows the calculated weight function distribution from mea-
sured reversal curves in (a).

36
 3.2 The Inverted Model with Modified Switch Operator

To find an analytical function to approximate the inverted weight function, it is necessary to

make reference to the forward Preisach weight function. Until now, many analytical functions,
such as the Gaussian and Lorentzian functions, have been applied to approximate the forward
Preisach weight function [1].
An analytical weight function has been proposed by Sutor for a forward Preisach model,
which can be identified by a major hysteresis loop [35, 55]. This analytical weight function
derives from the fact that the general shape of major M(H) loops are often similar to the arc
tangent function, the forward weight function is designed to be similar to its derivative [35].
A
(, ) = n o (3.20)
1 + [( + ) ]2 + [( h) ]2

In the same way, as the general shape of major H(M) loops are often similar in shape to the
tangent function, the form of the weight function should be similar to its derivative [51]. The
differentiable weight function (, ) is proposed as
 2
(, ) = A [( + 1)1 ] [( h)2 ] . (3.21)

1
1
0.8
(, )

0.5
0.6
0 0

1 0.4

0 = 0.2
1 1
1 0 1 0 1
1
(a) (b)
1
1 =
0.8
(, )

0.5
0.6
0 0

1 0.4

0 0.2
1 1
1 0 1 0 1
1
(c) (d)

Figure 3.10: Normalized weight function examples. The forward weight function (, ) is
shown in (a) (b). The inverted weight function (, ) is shown in (c) (d).

37
3 The Inverted Preisach Model

0.5 0.044

Derivative dM/dH
0.25 0.033
Norm. M

0 0.022

0.25 0.011

0.5 0
0.5 0.25 0 0.25 0.5 0.5 .25 0 0.25 0.5
Norm. H Norm. H
(a) M(H) curve
0.5 0.08

Derivative dH/dM
0.25 0.06
Norm. H

0 0.04

0.25 0.02

0.5 0
0.5 0.25 0 0.25 0.5 0.5 .25 0 0.25 0.5
Norm. M Norm. M
(b) H(M) curve

Figure 3.11: (a) shows the ascending branch of a M(H) major loop and its derivative dM/dH.
(b) shows the ascending branch of a H(M) major loop and its derivative dH/dM.
Both M and H are normalized for comparison. The M(H) major loop was mea-
sured on a 1.9 m FeCo thin film by vibrating sample magnetometer.

Fig. 3.10 shows the shapes of (, ) and (,

). The forward weight function (, ) in
2

the Preisach plane P = (, ) R :  is shown in Fig. 3.10(a). The inverted weight
function (, ) in Preisach plane P0 = (, ) R2 : is shown in Fig. 3.10(c). By
applying the modified switch operator, (, ) remains positive in the whole Preisach plane.
As analyzed in Fig. 3.8(b), (, ) becomes large near the vertices of the Preisach triangle.
It can be observed in contour maps shown in Fig. 3.10(b) and (d), that both the forward and
inverted weight functions are symmetric with respect to the line = .
The forward weight function (, ) has an obvious peak near the center of the Preisach
triangle. This is due to the big slopes of M(H) curves near the coercivity Hc . Fig. 3.11(a)
shows the ascending branch of a normalized M(H) major loop. The slope can be observed by
means of derivatives dM/dH in Fig. 3.11(b). The fast change of the magnetization M results
in a maximum slope near coercivity. The identification of (, ) near coercivity should be
carefully concerned. Compared with M(H) curves, the magnetic field H changes smoothly
with respect to M. It can be seen in Fig. 3.11(d), that the derivative dH/dM derives from the
ascending branch of a H(M) major loop shown in Fig. 3.11(c). The smooth distribution offers
a great advantage in approximating the inverted weight function.
This Chapter describes the definition of this inverted Preisach model. To identify this model,
hysteresis curves are required as experimental data. Three characterization methods have been
applied in this work and they are introduced with details in the next chapter.

38
4 Experimental Techniques
In order to precisely model the magnetic hysteresis behavior, reliable experimental data is
required for identification. As introduced in the section on ferromagnetism theory, the mag-
netic properties of various materials may differ greatly. The difference can be reflected in the
shape of hysteresis curves. Even for the same magnetic material, the shape of the hysteresis
curves can be greatly influenced by many factors, such as frequency, temperature, anisotropy,
deformation or even the shape of the samples.
Until now, many characterization methods have been proposed for the determination of the
magnetic properties. Some of these methods have been described by international standards,
e.g., the Epstein frame and the single sheet tester (SST) methods. Some of them have been
developed as commercial instruments and widely applied in professional laboratories, such
as vibrating sample magnetometer (VSM). These methods apply different measurement tech-
niques. Magnetic materials are processed in different shapes, measurements are carried out
with DC or AC excitation, in closed or open magnetic circuits, etc. Therefore, it is advisable to
mention the characterization method and sample shape in measurement results. In this work,
three characterization methods were applied to determine the magnetic properties in different
conditions.

4.1 Epstein Frame Method

The Epstein frame is the most commonly used method for characterization of electrical steels.
The measurement results show great repeatability and reliability according to standards (e.g.,
IEC 60404-2). As shown in Fig. 4.1, a frame is formed by four coil sets. Each coil set
comprises an outer winding and an inner winding. The outer windings are connected in series
as primary winding (N1 is the number of turns). The magnetizing current Iprim is applied to the
primary winding for magnetizing specimens. The magnetic field H is calculated from Iprim as

N1 Iprim
H= , (4.1)
lm
where lm is the effective magnetic path length (EML). The inner windings are connected in
series as the secondary winding (N2 is the number of turns). The voltage induced in the
secondary winding Usec is measured to obtain the flux density by
dB Usec
= , (4.2)
dt N2 A
where A is the cross sectional area of the specimens, it is defined from the weight of specimens.

39
4 Experimental Techniques

Figure 4.1: 25cm Epstein frame constructed according to standard IEC 60404-2.

Specimens are cut as strips from electrical steels. They are placed inside the coil sets and
overlapped at corners. Due to the overlapping, the effective magnetic path length (EML) is
approximately determined as 0.94m for all kinds of materials. This is criticized because when
the steels under test are grain oriented, the difference to the real EML can be significant [24].
Fig. 4.2 shows an automated measurement system with Epstein frame1 . During the mea-
surement process, the excitation signal is fed to a power amplifier to ensure the required power
for magnetizing. the magnetizing current Iprim is measured through the voltage of a shunt re-
sistor UR , which is connected in series with the primary winding of the Epstein frame. The
voltages UR and Usec are measured using an oscilloscope. To compensate for the air flux
generated in the secondary winding, a mutual inductor M is applied here. The adjustment of
the mutual inductor is made in absence of the specimen. Electrical steels are widely applied

Figure 4.2: Overview of the automated measurement system with Epstein frame.

1 The automation method is introduced in the published paper [56]

40
 4.1 Epstein Frame Method

in transformers or electric motors. In order to reduce energy loss caused by the eddy cur-
rent, electrical steels are usually made of SiFe alloys and produced as thin lamination. The
addition of silicon can increase the electrical resistivity of the steel. However, high silicon
content makes the steels hard and brittle, which leads to difficulties in mechanical treatments.
Normally the standard steels are produced with 3 to 4.5% silicon [24, 57].
Based on the measured parameters, electrical steels are classified according to international
standards. Table. 4.1 presents the classification of non-oriented electrical steels according to
European standard EN 10106. The steel names are designated with the nominal thickness and
core loss is measured at 1.5T, 50Hz. The letter M represents electrical steel and the letter A
stands for the annealing process.

Table 4.1: Classification of Non-oriented Electrical Steel, Standard EN 10106 (2007)

Nominal Core loss at Polarization Polarization Conventional

Steel name Thickness 1.5T, 50Hz for 5000A/m for 10000A/m density
(mm) (W/kg) (T) (T) kg/dm3
M330-35A 0.35 3.30 1.60 1.70 7.65
M350-50A 0.50 3.00 1.60 1.70 7.65
M600-50A 0.50 6.00 1.66 1.76 7.75
M800-100A 1.00 8.00 1.66 1.75 7.70
M1300-100A 1.00 13.00 1.7 1.78 7.80

The parameters specified in the standard can be applied to evaluate the instrument accuracy.
Table. 4.2 presents the material parameters measured with the Epstein frame setup shown in
Fig. 4.2. For specimens with thickness not more than 0.50 mm , the measured core losses are
quite close to the standard ones. The average relative difference is about 8.5%.

Table 4.2: Results of Measurements with Epstein Frame

Nominal Core loss at Polarization Polarization Conventional

Steel name Thickness 1.5T, 50Hz for 5000A/m for 10000A/m density
(mm) (W/kg) (T) (T) kg/dm3
M330-35A 0.35 2.856 1.631 1.746 7.650
M350-50A 0.50 3.091 1.651 1.760 7.650
M600-50A 0.50 5.490 1.744 1.851 7.850
M800-100A 0.98 6.090 1.667 1.773 7.714
M1300-100A 0.98 8.248 1.740 1.836 7.795

41
4 Experimental Techniques

4.1.1 Frequency Dependence of Hysteresis Loops

AC measurement is carried out here with a frequency up to 400Hz. Specimens are magnetized
with an alternating sinusoidal magnetic field, whose peak value is gradually increased. A
family of hysteresis loops are measured as shown in Fig.4.3, the primary magnetization curves
are obtained by joining tips of these loops.

Figure 4.3: A family of hysteresis loops and the primary magnetization curve obtained by
connecting the tips of these loops.

The frequency dependency of the hysteresis loops can be described by permeability and
core loss curves at different frequencies. The relative permeability r is calculated at each tip
point of the primary magnetization curves according to (2.5). Fig. 4.4 shows the permeability
r relative to the magnetic field H measured at different frequencies. It can be seen that the
permeability decreases with the increasing of the frequency. When the specimen is magnetized
to saturation, the relative permeability r tends to be one. At higher frequencies, as more
power is required for magnetization, measurement is limited by instrument range.
The core loss Pc introduced in (2.6) can be calculated from the area of the hysteresis loops.
f
I
Pc = H dB. (4.3)

As shown in Fig. 4.5, the core loss increases with the increasing of the frequency. The
thicker the specimen is, the more obvious this tendency is. E.g. at the same magnetic flux
B = 0.4T, compare the core loss at 10Hz and 400Hz, the core loss of the specimen M1300-
100A (thickness 1 mm) increases from 0.11W/kg to 15.32W/kg, the core loss of the specimen
M330-35A (thickness 0.35 mm) increases from 0.05W/kg to 3.93W/kg.
In an electromagnetic field with high frequency, the loss caused by eddy currents plays an
important role in the total core loss in specimen. Based on skin effect [50], the penetration
depth in the specimen can be evaluated as
1
= , (4.4)
f
where is the electric conductibility of the specimen, and is the absolute permeability of

42
 4.1 Epstein Frame Method

7400
10Hz
M33035A
3700
400Hz
0

7800
10Hz
M35050A
3900
400Hz
0
Rela. permeability r

4800
10Hz M60050A
2400
400Hz
0

4600
10Hz
M800100A
2300
400Hz
0

4800
10Hz
M1300100A
2400
400Hz
0 1 2 3 4
10 10 10 10
Magn. field H (A/m)

10Hz 50Hz 100Hz 200Hz 400Hz

Figure 4.4: Comparison of permeability curves at different frequencies (10Hz 400Hz). Mea-
surements are carried out at room temperature.

the material. The eddy currents can be reduced by decreasing of lamination thickness and
increasing of the resistivity of the material.
The frequency dependency can be analyzed with hysteresis loops. The hysteresis loops
of three specimens M330-35A, M600-50A and M1300-100A are shown in Fig. 4.6. In Fig.
4.6 (a)(c)(e), hysteresis loops were measured with the same maximal magnetic field H. With
the increase of frequency, the magnetic flux B at the tip point decreases, which leads to a
decrease in permeability. In other words, to magnetize the materials at high frequencies,

43
4 Experimental Techniques

26

M33035A
400Hz
13
10Hz
0

22

M35050A
400Hz
11
10Hz
0
Core losses Pc (W/kg)

42
400Hz
21 M60050A

10Hz
0

32
M800100A
400Hz
16
10Hz
0

42
400Hz
21 M1300100A

10Hz
0
0 0.4 0.8 1.2 1.6 2
Magn. flux B (T)
10Hz 50Hz 100Hz 200Hz 400Hz

Figure 4.5: Comparison of core loss curves at different frequencies (10Hz 400Hz). Mea-
surements are carried out at room temperature.

greater magnetic field is required.

It can be more clearly observed in Fig. 4.6 (b)(d)(f), where hysteresis loops are shown with
the same maximum magnetic flux B. With the increase in frequency, greater magnetic field is
required for magnetization. To change the magnetic flux back to zero, a greater coercive field
Hc in the opposite direction is needed. The hysteresis loops measured at high frequencies have
an oval form, which leads to greater loop areas and an increase of the core loss. This can be
utilized to explain the frequency dependency of the core loss shown in Fig. 4.5.

44
 4.1 Epstein Frame Method

1 1
Magn. flux B (T)

Magn. flux B (T)

0.5 0.5
400Hz 10Hz 400Hz 10Hz
0 0

0.5 0.5

1 1
100 50 0 50 100 120 60 0 60 120
Magn. field H (A/m) Magn. field H (A/m)
(a) (b)
1 1
Magn. flux B (T)

Magn. flux B (T)

0.5 0.5
400Hz 400Hz 10Hz
0 10Hz 0

0.5 0.5

1 1
160 80 0 80 160 300 150 0 150 300
Magn. field H (A/m) Magn. field H (A/m)
(c) (d)
0.4 0.4
Magn. flux B (T)

Magn. flux B (T)

0.2 0.2
400Hz
0 400Hz 0 10Hz

0.2 0.2
10Hz
0.4 0.4
100 50 0 50 100 200 100 0 100 200
Magn. field H (A/m) Magn. field H (A/m)
(e) (f)

10Hz 25Hz 50Hz 75Hz 100Hz 200Hz 300Hz 400Hz

Figure 4.6: Comparison of hysteresis curves at different frequencies (10Hz 400Hz). (a)(b)
M330-35A, steel thickness is 0.35 mm (c)(d) M600-50A, steel thickness is
0.50 mm, (e)(f) M1300-100A, steel thickness is 1.00 mm.

4.1.2 Temperature Dependence of Hysteresis Loops

Temperature is another influencing factor on the shape of hysteresis loops. As introduced
in Fig. 2.1, above the Curie temperature ferromagnetic materials behave in a paramagnetic
manner. Below the Curie temperature the spontaneous magnetization of ferromagnetic ma-
terials decreases with an increase in temperature. In this work the measurements are carried

45
4 Experimental Techniques

out at - 40 C140 C, which is much lower than the Curie Temperature of the electrical steels
(appr. 740 C). The temperature has little influence on the saturation magnetization of electri-
cal steels. But the temperature dependency can be reflected by the core loss curves.
Fig. 4.7 shows the core loss curves measured at 50Hz. The core loss decreases with an
increase in temperature. For comparison, hysteresis loops with the same magnetic field are
shown in Fig. 4.8. When the temperature increases, both the magnetic flux and the coercive
field decreases. This may be due to the magnetic orderings. At high temperatures, random
thermal motion makes it more difficult for electrons to maintain alignment [58]. This results
in smaller loop areas and core loss.
Another possible explanation, may be the decrease of the electric conductivity at high tem-
peratures. The eddy currents are reduced in the materials with lower electric conductivity,
resulting in a decrease in core loss caused by eddy currents.

5 10
M35050A M60050A
Core losses Pc (W/kg)

Core losses Pc (W/kg)

40C 40C

2.5 5
140C 140C

0 1 2 3 4
0 1 2 3 4
10 10 10 10 10 10 10 10
Magn. field H (A/m) Magn. field H (A/m)
(a) (b)
10 15
M800100A M1300100A
Core losses Pc (W/kg)

Core losses Pc (W/kg)

40C
40C

5 7.5
140C 140C

0 1 2 3 4
0 1 2 3 4
10 10 10 10 10 10 10 10
Magn. field H (A/m) Magn. field H (A/m)
(c) (d)

40C 20C 60C 100C 140C

Figure 4.7: Comparison of core loss curves at different temperatures (- 40 C 140 C). (a)
M350-50A, (b) M600-50A, (c) M800-100A, (d) M1300-100A.

46
 4.1 Epstein Frame Method

1
40C 1.4 40C
Magn. flux B (T)

Magn. flux B (T)

140C 140C

0.5 0.7

140C 40C
140C 40C
0 0
0 50 100 0 180 360
Magn. field H (A/m) Magn. field H (A/m)
(a) (b)

1.6 40C
1 40C
Magn. flux B (T)

Magn. flux B (T)

140C
140C

0.5 0.8

140C 40C 140C 40C

0 0
0 100 200 0 250 500
Magn. field H (A/m) Magn. field H (A/m)
(c) (d)

40C 20C 20C 60C 100C 140C

Figure 4.8: Comparison of hysteresis curves at different temperatures (- 40 C 140 C). (a)
M350-50A, (b) M600-50A, (c) M800-100A, (d) M1300-100A.

4.1.3 Measurements of Grain-oriented Steels

As mentioned in Section 2.1.4, anisotropy can be developed by applying rolling treatments to
the electrical steels. Most crystals of these steels are positioned with one easy axis close to
the rolling direction (RD). As the magnetization along the easy axis is easiest, the optimum
magnetic properties are developed in the RD. Relatively, the direction perpendicular to RD is
called the transverse direction (TD). The electrical steels manufactured with this texture are
named grain-oriented (GO) electrical steel.
As shown in Fig. 4.9, a specimen is cut from GO M33035 steel along RD. Fig. 4.9(a) is the
optical microscope image of the specimen surface. The photo is made by means of an optical
microscope keyence vhx500. The texture is clearer in Fig. 4.9(b) and (c), which are surface
topographies made by a confocal microscope Nanofocus Surf explorer. The height of the
surface in focus is detected and shown in colors.

47
4 Experimental Techniques

Figure 4.9: The surface images of GO steel M330-35A in RD. (a) is the optical microscope
image of the specimen surface. (b) and (c) are the 2D and 3D surface topographies.

The difference in the magnetic properties in the RD and TD can be evaluated using mea-
surements with the Epstein frame method. Specimens are individually cut from GO steels in
the RD or TD. As the RD is close to one easy axis of the material, it is easier to magnetize
the RD specimen. Therefore, when the RD specimen is under test, the effective magnetic path

2 2
Magn. flux B (T)

Magn. flux B (T)

1 1
35TW400RD 35Z155RD
35TW400TD 35Z155TD
0 1 2 3 4
0 1 2 3 4
10 10 10 10 10 10 10 10
Magn. field H (A/m) Magn. field H (A/m)
(a) (b)
Core losses Pc (W/kg)

Core losses Pc (W/kg)

35TW400RD 35Z155RD
4 4
35TW400TD 35Z155TD

2 2

0 0
0 0.5 1 1.5 0 0.5 1 1.5
Magn. flux B (T) Magn. flux B (T)
(c) (d)

Figure 4.10: Measurements of GO steels 35TW400 and 35Z155 at 50Hz at room temperature.
(a)(b) are primary magnetization curves. (c)(d) are core loss curves.

48
 4.1 Epstein Frame Method

length (EML) should be shorter than the TD specimen case. As the EML in the Epstein frame
is approximated to be the same for all kinds of materials, it should be stated that the magnetic
properties characterized here are considered as the average magnetic properties.
Fig. 4.10 shows the primary magnetization curves and core loss curves of GO steels
35TW400 and 35Z155. From the primary magnetization curves shown in Fig. 4.10(a) and
(b), it can be concluded that with the same magnetic field, greater magnetic flux is detected
in RD specimen. To saturate the specimens, a smaller magnetic field is required for the RD
specimens. Therefore, the magnetization of TD specimens cost more energy. As shown in
Fig. 4.10(c) and (d), at the same magnetic flux, the core loss of TD specimen is higher than
the RD specimen. The difference is more obvious in the measurements of 35Z155 specimens.
This can be analyzed with the measured hysteresis loops. If the same magnetic field is
applied to RD and TD specimens, as shown in Fig. 4.11(a) and (b), great differences of the
magnetic flux exist at the tips of the hysteresis loops. As a result, higher permeability can
be achieved in the RD specimen. Comparing the hysteresis loops with the same maximum
magnetic flux, as shown in Fig. 4.11(c) and (d), the TD loop area may be more than double
the RD loop area, which results in a smaller core loss in RD specimens. Therefore, it is more
efficient to apply GO SiFe electrical steels in transformers. Normally the GO steels are more
expensive than non-oriented steels.
1.2 1.8
Magn. flux B (T)
Magn. flux B (T)

35TW400RD 35Z155RD
0.6 0.9
35TW400TD 35Z155TD

0 0
0 40 80 0 40 80
Magn. field H (A/m) Magn. field H (A/m)
(a) (b)
1.2 1.2
Magn. flux B (T)

Magn. flux B (T)

0.6 0.6

35TW400RD 35Z155RD
35TW400TD 35Z155TD
0 0
0 120 240 0 150 300
Magn. field H (A/m) Magn. field H (A/m)
(c) (d)

Figure 4.11: Hysteresis loops of GO steels 35TW400 and 35Z155 measured at 50Hz at room
temperature. (a)(b) are hysteresis loops with the same maximum magnetic field,
(c)(d) are hysteresis loops with the same maximum magnetic flux.

49
4 Experimental Techniques

4.2 Double C-yoke SST

In recent years single sheet testers (SST) have been widely applied for characterization of
magnetic materials. In contrast to the Epstein frame method, the SST method does not have
the problem of effective magnetic path length as the magnetic field is directly measured by
applying a H-coil. According to numerical analysis, the field distribution is more uniform in
the SST [37]. Normally less specimens are needed for measurements with SST. However, to
carry out comparable determination, samples need to be produced in a large area, which is not
appropriate for electrical sheets.
The double C-yoke measurement system is shown in Fig. 4.12. Steel samples are placed
between two symmetrically furnished C-yokes. The excitation signal is fed to two excita-
tion coils surrounding each C-yoke and connected in series. In this way, two parallel closed
magnetic circuits are formed symmetrically. The eddy current can be reduced in AC measure-
ments.
In the middle of the steel sample, a small H-coil is placed close to the specimen. Based on
the Maxwells equations, near the interface of two adjacent media, the tangential component
of magnetic fields in two media should be equal to each other [50, 24]. Therefore, the H-coil
is applied to measure the tangential field component in the air. In the middle of the sample, it
can be expressed as

Hair k = Hsample k Hsample . (4.5)

The magnetic field in the steel sample can be directly obtained from Hair k . The H-coil in this
measurement system is wound around a 50 mm 15 mm 1.6 mm polyvinyl chloride core
(PVC), which has a constant permeability equal to 0 .

(a) (b)

Figure 4.12: Overview of the double C-yoke measurement system. (a) is the diagram of the
measurement system, (b) shows the double C-yoke SST constructed in laboratory.

50
 4.2 Double C-yoke SST

By applying a B-coil, the magnetic flux of the sample is calculated from the induced voltage
by (4.2). The B-coil is placed around the steel sample and the H-coil. Before measurements,
the compensation for the air flux is made by applying numerical method.
To realize a small dimension, the H-coil is made of thin wires with a small number of
windings. The voltage signal over the H-coil is quite weak, therefore an amplifier is utilized
here for enlargement. At frequencies greater than 25Hz, there is a deformation in measured
hysteresis curves. The measurements here are limited to 1Hz 25Hz. Because of the low-
frequency excitation signal, much effort has to be put into signal conditioning in frequency
domain [59].
According to the Epstein frame method, the measured magnetic properties are the average
magnetic properties of the samples. In the double C-yoke SST, the magnetic properties are
determined in a specific part of the sample. Therefore, the magnetic properties character-
ized with the double C-yoke SST are the local magnetic properties. These can be applied to
evaluate the plastic deformation influence on the local magnetic properties.

4.2.1 Plastic Deformation Effects

The plastic deformation may cause local magnetic degradation of the electrical steels. This is
attributed to the generation of dislocations acting as pinning sites for the domain walls [60].
To observe the plastic deformation influence over local magnetic properties2 , cold plastic
training is applied to the specimen. As shown in Fig. 4.13, the steel sample [Fig. 4.13(a)] was
folded in the middle [Fig. 4.13(b)] and attended to be inserted into the yoke. The H-coil was
put on the surface of the deformed section of the specimen [Fig. 4.13(c)] for measurement.

Figure 4.13: Cold plastic training on electrical steel M33035A samples. (a) shows the steel
sample before deformation, (b) shows the cold plastic training on the steel sam-
ple, (c) shows the deformed part of the steel sample.

With the double C-yoke SST shown in Fig. 4.12, local magnetic properties were measured
in both the deformed and the undeformed sections of the sample. Fig. 4.14 shows the de-
termined permeability curves. Compared with the results measured before deformation, the
permeability clearly decreases in the deformed section. The local magnetic property in the
2 This section as been published in [51, 61]

51
 4 Experimental Techniques

undeformed part is also influenced. These experiments were carried out both on RD and TD
samples. The results indicate that the magnetic properties of soft magnetic materials are quite
sensitive to plastic straining. Generally, plastic deformation causes changes in the dislocation
density, which in turn affects the pinning strength of domain walls and consequently their
movement [62].

7000 3400
Rela. permeability r

r
Rela. permeability
3500 1700

0 1 2 3
0 1 2 3
10 10 10 10 10 10
Magn. field H (A/m) Magn. field H (A/m)
(a) (b)
Before deformation Deformed part Undeformed part

Figure 4.14: Permeability curves of steel sample M33035A at 1Hz at room temperature. (a)
RD steel sample, (b) TD steel sample.

For further observation, heat treatment (HT) in a nitrogen environment was carried out on
the plastically deformed sample. After annealing in nitrogen environment, the local magnetic
properties were again measured in the deformed and undeformed sections. It can be seen from
the comparisons in Fig. 4.15, that the local magnetic properties exhibit an obvious recovery
after thermal treatment. This may be due to the reduction of distortion within the particles
during the heat treatment [63].

3400
Rela. permeability r

Before deformation
1700
Deformed part
Undeformed part
Deformed part (HT 800C)
Undeformed part (HT 800C)
0 1 2 3
10 10 10
Magn. field H (A/m)

Figure 4.15: Comparisons of permeability curves. TD steel sample M33035A was tested after
HT 800 C at 25Hz at room temperature.

52
 4.3 Vibrating Sample Magnetometer

4.3 Vibrating Sample Magnetometer

Since a thin film magneto resistor was applied by Hunt as a magnetic reading head in 1970 [24,
64], ferromagnetic thin films have attracted growing interest both from fundamental scientific
researchers and in the computer industry [36]. The ferromagnetic thin films are commonly
characterized in open circuit. In this case, magnetic moment is usually detected with sensitive
magnetometers [65]. The vibrating sample magnetometer (VSM) is widely applied, which
was developed by Van Oosterhout and Foner in 1950s [38, 39]. Nowadays, VSM has become
a standard instrument and is most widely used in professional magnetism laboratories.
Fig. 4.16 presents an overview of the vibrating sample magnetometer measurement system.
The sample is held by a rod and inserted in the gap of an electromagnet, which is applied to
generate a static magnetic field. The rod is connected with a shaker to vibrate the sample with
a specific amplitude and frequency.

Figure 4.16: Diagram of the vibrating sample magnetometer measurement system.

The small magnetized sample can be considered as a vibrating dipole. Suppose the magnetic
moment is directed along the axis of the pickup coils, then the induced voltage detected in the
pickup coils is proportional to the magnetic moment m:

d (t) d z(t)
u(t) = = m G(z) . (4.6)
dt dt

53
4 Experimental Techniques

(a) (b)

Figure 4.17: Overview of the VSM setup constructed in laboratory. (a) shows the electro-
magnet, the part included by a dashed square is enlarged and shown in (b). The
sample and pickup coils system are located between the electromagnet poles.

where G(z) is the geometry factor depending on the pickup coil geometry. In measurement, a
sinusoidal vibration z(t) = z0 sin t is applied to the sample. The induced voltage is

u(t) = m G(z) z0 cos t. (4.7)

Appropriate pickup coil configurations can cause the induced voltage to be insensitive to
sample position [41]. As shown in Fig. 4.17(b), a Mallinson four-coil system [40] is con-
structed between the electromagnet poles [see Fig. 4.17(a)]. The two pairs of coils are con-
nected in series opposition, their axes are parallel to the magnetic field and perpendicular to
the direction of sample vibration.
Due to the small sample size (film thickness less than 3 m), the induced signal is very
small. A lock-in amplifier is applied to detect the small voltage signal. In the measurement
system, a reference signal is generated by the lock-in amplifier. The signal has a specific
frequency (80Hz in this work) and fed to the shaker. By means of the accelerometer, the
vibrating amplitude z0 is monitored to keep it stable. The output of the lock-in amplifier,
which is the RMS (root mean square) value of u(t), is followed by

Urms = K m, (4.8)

where K = G(z) z0 , which can be calibrated by a standard sample with known magnetic
moment.
In the VSM, the sample is magnetized in an open magnetic circuit. A Hall sensor is placed
near the sample to measure the magnetic field. In the open magnetic circuit, the demagnetiza-
tion field of sample Hd should be considered. If a magnetizing field H is applied to the sample,
the internal field of the sample Hin is calculated as

Hin = H Hd = H Nd M , (4.9)

54
 4.3 Vibrating Sample Magnetometer

where Nd is the demagnetization factor, which depends on the shape of the sample. As shown
in Fig. 4.18, the ferromagnetic thin films are in the form of a disk. The magnetic field is
applied parallel to the sample surface. As the film thickness (less than 3 m) is much smaller
than the diameter of samples (10 mm), the demagnetization factor Nd in (4.9) can be considered
to be zero.

Figure 4.18: The sample for VSM measurement. Ferromagnetic thin films (material
Fe49 Co49 V2 and Ni81 Fe19 ) were produced by RF magnetron sputtering. Using
the etching technology, the sputtered films were made in the form of a disk.

The demagnetization field can be neglected here. Samples are assumed to be uniformly
magnetized [66], the magnetization M can be obtained with the sample volume V as
1 m
M= mi = . (4.10)
V V

4.3.1 Scalar Characterization with VSM

In the VSM setup, the axis of pickup coils is parallel to the magnetic field. As the magnetic
field H and magnetization M are vector quantities, it is considered as scalar characterization.
The magnetization component parallel to the magnetic field is determined.
In this work, DC measurement is carried out with a VSM system. Besides symmetric
hysteresis loops, asymmetric hysteresis loops, such as FORCs and minor loops, can be deter-
mined. Fig. 4.19 shows the FORCs and minor loops measured on a 1.9 m Fe49 Co49 V2 thin
film. As mentioned in Section 3.1, the FORCs can be applied for identification of Preisach
hysteresis models.

4.3.2 Vector Vibrating Sample Magnetometer

Scalar VSM characterization presents only the magnetization component in the direction of
the applied field. In the modern field of magnetic materials research, the full description
of magnetization vectors is attracting more and more attention. Especially for anisotropic
materials, which have different magnetic properties depending on the magnetization direction.
By applying another pickup coils system, which is perpendicular to the applied field, the
VSM can be extended to a vector VSM (VVSM) for 2D vector characterization. This can
be seen in Fig. 4.20(a), the two pickup coils systems are orthogonal to each other (biaxial
pickup coils). In this work they are called longitudinal coils (LC) and transversal coils (TC).
The parallel and perpendicular components of the magnetic moment are proportional to the

55
Magn. fi

Magn. fi
60
4 Experimental Techniques 60
Input (b)

2 2
Magn. polarization J (T)

Magn. polarization J (T)

1 1

0 0

1 1

2 2
60 30 0 30 60 60 30 0 30 60
Magn. field H (kA/m) Magn. field H (kA/m)
(a) (b)

Figure 4.19: Scalar characterizations of 1.9 m Fe49 Co49 V2 thin film. (a) First order reversal
curves (FORCs). (b) Minor loops.

voltages induced in LC and TC. The relation described in (4.8) can be written as

m k = Klong Urms,long ,
m = K tran Urms,tran . (4.11)

As shown in Fig. 4.20(b), the angle between the two components is calculated by
!  
m K tran Urms,tran
= arctan = arctan . (4.12)
mk Klong Urms,long

4.3.3 Calibration Method

Accurate characterizations with VSM or VVSM require careful calibrations before measure-
ment. Specifically, the factors Klong and K tran in (4.11) need to be obtained using a reasonable
calibration method.
The longitudinal pickup coils are calibrated by Mu-metal film (thickness of 25 m) with
known saturated magnetic polarization Js = 0.8T. As a reference sample, the Mu-metal film is
cut into a shape similar to the samples applied in measurements. When the reference sample is
magnetized to saturation, the magnetic moment vector can be considered to align completely
parallel to the field, which implies
Js
m k ms = V Ms = V . (4.13)
0

56
 4.3 Vibrating Sample Magnetometer

(a) (b)

Figure 4.20: The longitudinal coils (LC) and transversal coils (TC) for vector characterization.

By combining (4.13) and (4.11), Klong can be obtained with Js of the reference sample.

V Js
Klong = (4.14)
0 Urms,long

The calibration of the transversal pickup coils is carried out by measuring the angular de-
pendency of the pickup coils. A hard magnetic material with high coercivity and remanent
magnetization is applied as the reference sample. The sample is magnetized to saturation then
rotated in a zero field. The rotation is realized and controlled by means of a rotation stage.
During rotation the induced voltages in LC and TC are measured in every angular increment.
The remanent magnetization remains stable in the zero field, which is calculated by
q
mr = m k 2 + m 2 . (4.15)

After calibration, the resultant parallel and perpendicular components of magnetic moment
should be orthogonal as shown in Fig. 4.21.

Figure 4.21: The parallel and perpendicular components of magnetic moment after calibration.

57
4 Experimental Techniques

4.3.4 Vector Characterization with VSM

The vector characterization includes the magnetization component perpendicular to the ap-
plied field. This provides more information than the scalar characterization.
Fig. 4.22 presents the major hysteresis loops obtained by vector characterization. The
isotropic samples were firstly magnetized to saturation then rotated in zero field. During
rotation the magnitude of the remanent magnetization does not change but the direction rotates
with the sample. After 90 rotation ( = 90 ), the remanent magnetization aligns close to the
perpendicular direction. Major loops were measured both parallel and perpendicular to the
applied field. As measurements start at the zero field, the start value of the perpendicular
component J is approximated to the remanent magnetic polarization Jr . When the field
increases, the total magnetization increases and gradually aligns parallel to the field, while the
perpendicular component decreases to a minimum.

2 1
Magn. polarization J (T)

Magn. polarization J (T)

1 0.5

0 0

1 J|| 0.5 J||

J J
2 1
60 30 0 30 60 10 5 0 5 10
Magn. field H (kA/m) Magn. field H (kA/m)
(a) (b)

Figure 4.22: Major hysteresis loops measured both parallel and perpendicular to the applied
fields. Samples: (a) 1.9 m Fe49 Co49 V2 thin film, (b) 1.7 m Ni81 Fe19 thin film.

The angular dependence of vector characterization is illustrated in Fig. 4.23. For compar-
ison, the experiments described above were repeated with different rotation angles . The
correspongding primary magnetization curves start at different values. The start values of J
and J k are approximated to the projections of the remanent magnetic polarization Jr as

J Jr sin , J k Jr cos . (4.16)

Compared with isotropic materials, the anisotropic materials exhibit more obvious angular
dependence of the magnetic properties. As shown in Fig. 4.24(a), an anisotropic sample3
3 The sputtered thin film was magnetized at high temperature. The film is anisotropic along the film surface.
Circular samples were made by applying laser cutting technology.

58
 4.3 Vibrating Sample Magnetometer

0.8 2
Magn. polarization J (T)

Magn. polarization J (T)

J J||
= 60
0.4 1.5
|| = 0
0 1

0.4 0.5
= 90
|| = 90
0.8 0
0 20 40 60 0 20 40 60
Magn. field H (kA/m) Magn. field H (kA/m)
(a) (b)
= 60 = 45 = 0 = 30 = 90

Figure 4.23: Primary magnetization curves measured both perpendicular (a) and parallel (b)
to the applied field. Before measurements the sample was rotated with degree
from original position. Sample: 1.9 m Fe49 Co49 V2 thin film.

was magnetized along the sample surface in three different directions. Major loops were
measured parallel to the applied field. It can be seen from the comparison, the sample can
reach saturation with the lowest magnetic field along = 0 , while with the highest magnetic
field along = 90 . It can be assumed that the easy axis of the material is close to the = 0
direction, and the hard axis is close to the = 90 direction.

1 1
Magn. polarization J (T)

Magn. polarization J (T)

0.5 0.5

0 0

= 0
0.5 0.5 J||
= 60
= 90 J
1 1
10 5 0 5 10 10 5 0 5 10
Magn. field H (kA/m) Magn. field H (kA/m)
(a) (b)

Figure 4.24: Characterization of anisotropic sample. (a) scalar characterization along different
directions. (b) vector characterization. Sample: 400 nm Ni81 Fe19 thin film.

59
4 Experimental Techniques

Fig. 4.24(b) shows the vector characterization when the anisotropic sample was magnetized
along the hard axis. The polarization component parallel to the applied field J k is close to the
hard axis, while the polarization component perpendicular to the applied field J is close to
the easy axis. When the magnetic field is zero, J is much greater than J k . The reason can be
analyzed according to the scalar measurements shown in Fig. 4.24(a). The anisotropic sample
can get saturated with very low magnetic field along the easy axis. As a result, J shows an
obvious increase near the zero field.

4.3.5 Vector Characterization in Rotational Field

The VVSM can be applied for vector characterizations in a rotational field. As the electro-
magnet cannot be moved, the rotational field is realized by mechanically rotating the tested
sample. [24, 65]. By adjusting the rotation angles and the magnitude of the field, the rotational
field can be applied with different trajectories. In the characterization results shown here, both
the magnetic polarization and magnetic field intensity are normalized for comparison.
The vector characterizations shown in Fig. 4.25 were carried out in a rotational field with
circular trajectory. The sample was firstly magnetized in the 0 direction and the magnetic
field intensity was then rotated with a constant magnitude from 0 to 360 . In the measure-
ment shown in Fig. 4.25(a), the sample was firstly magnetized to negative saturation, then
returned close to zero polarization by applying a positive coercive field (about 9kA/m). The
measurement was made when the field started to rotate. It can be seen from the polar plot,
during rotation the magnetic polarization vector rotates mostly in the negative direction. By
increasing the rotation magnitude of the magnetic field to 25kA/m, the results become quite

90 90
120 60 norm. H 120 60
norm. J 25kA/m
150 9kA/m 30 150 30

180 0 180 0

210 330 210 330

240 300 240 300

270 270
(a) (b)

Figure 4.25: Vector characterizations of an isotropic sample in rotational field with circular
trajectory. (a) the magnetic field rotates with a radium close to the coercive field
9kA/m. (b) the magnetic field rotates with radium of 25kA/m. Sample: 1.9 m
Fe49 Co49 V2 thin film.

60
 4.3 Vibrating Sample Magnetometer

different. As shown in Fig. 4.25(b), the magnetic polarization vector starts with a magnitude
in the 0 direction, then rotates closely together with the magnetic field.
The magnetic field can also be adjusted to rotate with ellipse trajectories. The measurement
shown in Fig. 4.26(a) started from zero polarization at positive coercive field 9kA/m. The
magnitude of the magnetic field reaches maximum (15kA/m) at transversal directions (90
and 270 ). In comparison with the result shown in Fig. 4.25(a), the magnetic polarization
vector rotates closer to the magnetic field near transversal directions.
The measurement shown in Fig. 4.25(b) started at a positive remanent polarization in a zero
field. The magnetic field component along 0 direction changes in a small range between zero
and 10kA/m, the resultant polarization vector rotates close to the remanent polarization.

90 90
120 60 norm. H 120 60
norm. J
10kA/m
150 30 150 30

9kA/m
180 0 180 0
10kA/m

210 330 210 330

15kA/m
240 300 240 300
270 270
(a) (b)

Figure 4.26: Vector characterizations of an isotropic sample in rotational field with ellipse
trajectory. (a) the magnetic polarization vector rotates from zero. (b) the mag-
netic polarization vector rotates from remanent polarization. Sample: 1.9 m
Fe49 Co49 V2 thin film.

The above mentioned characterizations are applied to an isotropic sample. If anisotropic

material is magnetized in a rotational field, due to its direction dependent magnetic properties,
the polarization vector behaves differently. Fig. 4.27 shows the vector characterizations of an
anisotropic sample. The magnetic field intensity rotates with a constant magnitude.
In Fig. 4.27(a), the rotation magnitude of the magnetic field is 2kA/m. With reference to the
major loops shown in Fig. 4.24(a), at 2kA/m, obvious difference of polarization exists along
different directions. In the rotational field shown in Fig. 4.27(a), the magnetic field reach
2kA/m in the 0 direction then rotates from 0 to 360 . The resultant polarization rotates with
an ellipse trajectory. From the shape of the ellipse, it can be assumed that the 0 direction is
not completely parallel to the easy axis.
In Fig. 4.27(b), the rotation magnitude of the magnetic field is increased to 4kA/m. Ac-
cording to the measurement shown in Fig. 4.24(a), when the magnetic field reaches 4kA/m,

61
4 Experimental Techniques

90 90
120 60 norm. H 120 60
2kA/m norm. J 4kA/m
150 30 150 30

180 0 180 0

210 330 210 330

240 300 240 300

270 270
(a) (b)

Figure 4.27: Vector characterizations of an anisotropic sample in rotational field with circular
trajectory. (a) the magnetic field rotates with a radium of 2kA/m. (b) the mag-
netic field rotates with radium of 4kA/m. Anisotropic sample: 400 nm Ni81 Fe19
thin film.

the sample is almost saturated along every direction. The polarizations display no obvious
difference. Therefore, in the rotational field with radium of 4kA/m shown in Fig. 4.27(b), the
rotation trajectory of the polarization vector is almost circular. During rotation the polarization
vector rotates close to the magnetic field.
In this chapter three different characterization methods have been introduced. By means
of the Epstein frame method, determination of the average magnetic properties can be carried
out at different frequencies or temperatures. By applying the double C-yoke SST, the influ-
ence of plastic deformation can be evaluated using local magnetic property measurements. In
order to characterize anisotropic materials, Epstein frame samples are cut from grain oriented
steels along different directions. This is made by applying a VSM, because the sample can be
mechanically rotated in the magnetic field. The VSM is widely utilized to detect the magnetic
moment of small magnetic samples. By applying orthogonal pickup coils systems, vector
characterization can be carried out in both undirectional and rotational magnetic field.
Reliable characterization methods provide accurate experimental data, which can be applied
in identification of hysteresis models. The identification method and the numerical calculation
procedure of the inverted Preisach model are introduced in the next chapter.

62
5 Model Identification and Calculation
This chapter introduces the identification and numerical calculation method of the inverted
Preisach model. The identification method plays an important role in the accuracy of the
models. For comparison, the inverted Everett function, which is applied as the identification
method for the pseudo-inverse of the Preisach model, is described in the first section. In or-
der to calculate the inverted Everett matrix, a set of first order reversal curves (FORCs) are
required as the experimental data for identification. By means of the inverted Everett func-
tion, the problem of determining weight function can be solved. However, this identification
method requires a large amount of experimental data and careful work in measuring.
In comparison to the pseudo-inverse of the Preisach model, the proposed inverted hysteresis
model is identified with less experimental data. Only major hysteresis loops are required to
determine the parameters in the weight function. The determination of parameters is fulfilled
by a numerical optimization method. A quasi-Newton method algorithm is applied to optimize
the mean squared error (MSE) between the measured and simulated data. More details are
introduced in Section 5.2.
In Section 5.3, the numerical calculation method of the inverted Preisach model is intro-
duced. The double integral of the weight function is calculated by applying a numerical
approximation method. The model is verified with hysteresis loops of both hard and soft
materials. Simulations have been made on major loops, minor loops and reversal curves. The
simulation results show good approximation to the measurement data.
In the last section, a short summary comparing the two identification methods is presented.

5.1 Identification with the Inverted Everett Function

As introduced in Section 3.1, with the pseudo-inverse of the Preisach model, the output is
calculated by adding or subtracting the Everett function values at the vertices of the staircase
interface. Therefore, to identify the pseudo-inverse of the Preisach model, the inverted Everett
functions E(, ) have to be calculated at all nodes of the Preisach plane. These values
are arranged according to the coordinates of the nodes and form a so-called inverted Everett
matrix. Normally, the Preisach plane is uniformly discretized by a n n square mesh in the
precalculation stage. The mesh values of E(, ) form a triangular matrix Enn as the inverted
Everett matrix.
E1 ,1 E1 ,2 E1 ,n1 E1 ,n

E2 ,2 E2 ,n1 E2 ,n
E= . . .
. .
. . (5.1)

. . .

En1 ,n1 En1 ,n
0

En ,n

63
5 Model Identification and Calculation

Figure 5.1: Identification of the pseudo-inverse of the Preisach model. On the left is the
Preisach plane discretized by n (here n = 9) mesh. In the middle is the input
signal for measurement of FORCs. On the right are the FORCs for identification.
All of the inverted Everett function results on these mesh nodes form the inverted
Everett matrix.

The experimental data for identification is a discretized set of FORCs. The Everett function
E(, ) is calculated according to (3.6). As the example illustrates in Fig. 5.1, the correspon-
dency between the values of E(i , j ) and the points on the FORCs is described in the plot.
E.g., E(4 , 4 ) is calculated from half of the difference H4 H4 4 . To obtain H4 and H4 4 ,
the magnetization M increases from negative saturation to 4 , then decreases back to 4 . The
calculated E(4 , 4 ) can be indexed with (4 , 4 ) in the inverted Everett matrix.

Fig. 5.2(b) shows the inverted Everett matrix calculated from the FORCs shown in Fig.
5.2(a). During calculation, the Everett function values of vertices on the staircase interface
are obtained from the inverted Everett matrix using an interpolation method. In this work the
bilinear interpolation method is applied.

The measurement of FORCs is usually carried out point by point with the DC method. In
order to obtain an abundant inverted Everett matrix for calculation, as shown in Fig. 5.2(a),
FORCs distributed intensively inside the major loop should be measured. Therefore, an enor-
mous amount of measuring points are required, which is difficult and time-consuming. Fur-
thermore, measurement errors, e.g. the errors caused by noise in the measurement of the
magnetic field, are carried over into the identification results [1, 5]. The simulation results are
close to the experimental data, but the experimental errors may lead to the discontinuity and
non-smoothness of the simulated curves.

64
 5.2 Identification of the Inverted Preisach Model

1.8

Magn.polarization J (T) 0.9

0.9

1.8
60 30 0 30 60
Magn. field H (A/m)
(a)
1.66

60
E(, ) (kA/m)

40

B (T)
0
30
20
1.66
0
1.66 0 B (T)
0 1.66 0
1.66 1.66 1.66 0 1.66
B (T) B (T)
(b) (c)

Figure 5.2: The calculated inverted Everett matrix. (a) illustrates the measured FORCs for
identification (60 reversal curves). (b) is the surface of the inverted Everett matrix.
(c) is the corresponding contour plot. Measurement of FORCs was carried out
with VSM on 1.9 m Fe49 Co49 V2 thin film sample.

5.2 Identification of the Inverted Preisach Model

The identification of the inverted weight function is accomplished with major loops. Com-
pared with FORCs, it is easier to measure with both AC and DC methods. To decrease the
experimental errors, mean experimental data can be obtained by repeating measurements. The
advantage of identification with major loops becomes more obvious in the case of evaluating
influence factors on hysteresis properties, such as frequency, temperature, and anisotropy.
Measurements of identification data have to be carried out in different conditions. Therefore,
the identification with major loops may attract more FE users.
During the identification process, the experimental data for identification is normalized for
simplification as
M(t) H(t)
e(t) = , fmeas (t) = . (5.2)
Mmax Hmax

65
5 Model Identification and Calculation

Consequently, the simulated magnetic field Hsimu (t) is calculated as

Hsimu (t) = Hmax fsimu (t), (5.3)

5.2.1 Simulation of the Forced Magnetization Effect

Some of the major loops are measured in intensive saturation. As shown in Fig. 5.3(a), the
applied external magnetic field is further increased when the material is completely saturated.
Theoretically, the magnetization retains an almost constant value and increases no further.
This value is called the saturation magnetization. In practice, due to the forced magnetization
effect, the magnetization (or polarization) keeps increasing with the applied magnetic field
beyond actual saturation [42, 35]. The major loops with and without the forced magnetization
effect are shown in Fig. 5.3(a).

Figure 5.3: (a) The M(H) loops with (black line) and without (red line) the forced magneti-
zation effect. (b) The normalized H(M) loops with (black line) and without (red
line) the forced magnetization effect.

For identification of the inverted Preisach model, the M(H) loop is inverted and normalized
to f (e) loops. As shown in Fig. 5.3(b), the curves at e = 1 (saturation magnetization) are
almost vertical straight lines, which are especially obvious in the f (e) loops obtained from the
M(H) loops measured in intensive saturation. It is difficult to model these curves and this may
lead to numerical problems in simulation. Therefore, in the inverted modeling the vertical
straight section should be removed from major loops as far as possible.
To take the forced magnetization effect into account, an analytical tangent function was pro-
posed with reference to the forward Preisach model proposed by Sutor in [35]. The function
is expressed as  

g[e](t) = b tan a e(t) , (5.4)
2

66
 5.2 Identification of the Inverted Preisach Model

Figure 5.4: (a) The graphic relation between fsimu and (e).
b (b) Comparison of the resulted
analytical tangent function g(e) according to different factor a.

where the parameter a in (5.4) is limited to [0, 1), because the range of the tangent function is
limited to (1, 1). The reason for choosing the tangent function as a basis is due to the fact
that most H(M) loops bear a similarity to the tangent function.
By adding the analytical tangent function in (5.4) to the inverted Preisach model, the simu-
lation result fsimu (t) is calculated as

fsimu (t) = [e](t)

b + g[e](t). (5.5)

The graphic relation between [e](t)

b and fsimu is shown in Fig. 5.4(a). When the input
e = 0,
b = Hc ,
fsimu = (5.6)
Hmax
where Hc is the coercivity of the magnetic materials. This important parameter is usually
presented in material catalogues. The difference between the measured and simulated coer-
civities will affect the accuracy of the electromagnetic simulation involving hysteresis effect.
[e](t)
b can be considered as a hysteresis loop which shares the same coercivity with fsimu .
The hysteresis loop [e](t)
b is an ideal major loop without the forced magnetization effect.
The constant part in saturation is shown with a red dashed line in Fig. 5.4(a).
Depending on the influence exerted by the forced magnetization effect, the slopes of the
simulated H(M) loop can be controlled with the a factor in the analytical tangent function in
(5.4). As shown in Fig. 5.4(b), with the increase of the a factor, the slope becomes steeper.
To simulate the H(M) loop with less forced magnetization effect, which retains an almost
constant value beyond saturation, the factor a can be set with a large value in the analytical
tangent function in (5.4).

67
5 Model Identification and Calculation

5.2.2 Parameter Determination

According to the weight function (see equ 3.21) introduced in section 3.2.2, the five parameters
in the function, A, 1 , 2 , and h are required to be determined. As the forced magnetization
effect is taken into account in this inverted Preisach model, the parameters a and b in the
tangent function (5.4) are also involved as the identification target.
The identification of the inverted model is fulfilled with measured major loops. The seven
parameters are determined by a quasi-Newton method algorithm by optimization of the mean
squared error (MSE) between the measured and simulated data. The MSE is calculated as
NT
1
MSE( f ) =
NT [ fmeas(tk ) fsimu(tk ) ]2 , (5.7)
k=1

where NT is the size of the measured data fmeas (t).

In this work a LabVIEW optimization VI based on the Broyden quasi-Newton method was
applied to minimize the mean squared error (MSE). In the algorithm, unknown parameters are
determined with an inverse method. As should always be an even number, it is manually
set as a constant in the initialization of the algorithm. The MSE calculated by (5.7) is taken as
the optimization target of the scheme. Relative change factors of the other six parameters are
taken as a variable vector X. In the initialization part of the algorithm, as shown in (5.8), the
six parameters are set with initial values, all the elements in variable vector X are set to 1.

1,ini 1
2,ini 1

Aini 1
Pini =
, Xini = .
1 (5.8)
h ini


aini 1
bini 1

In the algorithm, as shown in (5.9), new parameters Pk are calculated by the Hadamard
product of Pini and the updated Xk at every iteration.

1,ini x1,k 1,ini x1,k 1,k
2,ini x2,k 2,ini x2,k 2,k

Aini x3,k Aini x3,k Ak
Pk = Pini Xk = hini x4,k = hini x4,k = hk ,
(5.9)

aini x5,k aini x5,k ak
bini x6,k bini x6,k bk

where k is the index for the iterations. The new parameters Pk are applied in the inverted
Preisach model to calculate the MSE. The minimal MSE is iteratively determined until the
stopping criteria is reached.
This quasi-Newton method algorithm is efficient (iteration steps k < 1000). However, in
solving nonlinear problems, this method cannot guarantee convergence of solutions, especially
in approximation of the H(M) loops, which contain large vertical straight parts.

68
 5.3 Calculation of the Inverted Preisach Model

5.3 Calculation of the Inverted Preisach Model

The inverted Preisach model should be numerically implemented according to (3.14) for com-
putation of the output. Depending on the value (+1 or 1) of the switch operators , the
whole Preisach plane can be divided into two regions S+ (t) and S (t), where are equal
to +1 in S+ (t) and they are equal to 1 in the S (t) area. Therefore, (3.14) can be written as
ZZ ZZ
[e](t)
b = (, )dd (, )dd . (5.10)
S+ (t) S (t)

The calculation in (5.10) involves double integrals of the weight function (, ) over the
Preisach plane. Although a differentiable weight function (, ) has been proposed in 3.21,
it is very difficult to find its antiderivative function. Furthermore, it is easier to apply a numeri-
cal method to compute a numerical approximation of the double integrals. In calculation of the
inverted Preisach model, the approximation work is carried out in a precalculation procedure
before the calculation stage.

5.3.1 Precalculation of V Matrix

In the precalculation stage, the Preisach plane is uniformly discretized by creating a n n
1
square mesh. As shown in Fig. 5.5(a), the square mesh covers the whole Preisach triangle. The
middle points of every square are taken as the computation nodes. The black nodes (i , j ) are
0.9
located inside or on the diagonal of the Preisach triangle P0 = (, ) R2 : i . The
hollow nodes represent the nodes outside the Preisach triangle.
0.8

0.7

0.6
Norm. appro. (, )

1
0.5
0.5
0.4
0 1
1 0.3
0 0
0.2
1 1
(a) (b) 0.1

Figure 5.5: (a) The square mesh for discretization of the Preisach plane in the precalculation
0
stage. The black nodes () are in the Preisach triangle, the hollow nodes () are
outside the Preisach triangle. (b) shows the approximated (, ) surface.

As shown in Fig. 5.5(b), if n is large enough, the integral of the weight function over the

69
5 Model Identification and Calculation

square segment s can be approximated by the volume of cubes as

ZZ
(, )dd s (i , j ), (5.11)
s

where (i , j ) is the center point of the square segment s. As shown in (5.12), the integral
of the weight function over a specified area S in the Preisach plane can be approximated by a
sum of the cubic volume, which is included in S.
ZZ
(, )dd s (i , j ) (5.12)
S (i , j )S

Because the hollow nodes are outside the Preisach triangle, the integrals over these segments
including hollow nodes are set to be zero. For the square segments including the nodes (i , j )
on the diagonal (i = j ), only half of the square segment is included in the Preisach plane.
The integral over these segments s(i = j ) is approximated as

1
ZZ
(, )dd s (i , j ). (5.13)
2
s(i = j )

The approximated integrals over every unit segment of the mesh shown in Fig. 5.5(a) can
be calculated as

s (i , j ) if i < j ,
1

i , j = s (i , i ) if i = j , (5.14)

2
0 if i > j .

All of i , j can form a n n lower triangular matrix V as

1 ,1 1 ,2 1 ,n1 1 ,n

2 ,2 2 ,n1 2 ,n

V= . .. .
.. ..
. (5.15)

.
n1 ,n1 n1 ,n
0

n ,n

The precalculated weight function matrix V is used as a constant in the calculation procedure.
As an interpolation method is applied in updating the S+ (t) and S (t) regions in the Preisach
plane, the mesh size, which is controlled by n in discretization, plays an important role in the
accuracy of the calculation results. In this work, n is set to be greater than 100.

5.3.2 Calculation Procedure

According to (5.10), to calculate the output, it is critical to update the S+ (t) and S (t) regions
of the Preisach plane. As introduced in (3.15) in Section 3.2.1, in the memory area of the

70
 5.3 Calculation of the Inverted Preisach Model

Preisach plane, the modified switch operator , [M](t) is calculated from the history value
of the relay switch operator 0 0 [M](t ). Therefore, the history value of 0 0 in area P =
(, ) R2 : should also be saved in calculation. All of the switch values of and


are saved in a n n matrix S, which has the same size as the precalculated matrix V.

1 ,1 1 ,2 1 ,n1 1 ,n

2 ,1 2 ,2 2 ,n1 2 ,n

.. .
. .
. .
.
S= . (5.16)

. . .

n1 ,1 n1 ,2 n1 ,n1 n1 ,n
n ,1 n ,2 n ,n1 n ,n

The matrix S is called the switch matrix here. As shown in (5.16), the value of is saved
in the lower triangular part of matrix S, while the value of is stored in the upper triangular
part. During calculation, according to the current input e(t), a linear interpolation method is
applied to update the interface between S+ (t) and S (t) regions in the Preisach plane, which
are the +1 and 1 regions in the matrix S.
The production of the weight function and the corresponding switch value is fulfilled by
applying the Hadamard product of the S matrix and the V matrix as

1 ,1 1 ,1 1 ,2 1 ,2 1 ,n1 1 ,n1 1 ,n 1 ,n

2 ,2 2 ,2 2 ,n1 2 ,n1 2 ,n 2 ,n
SV= .. .. ..
.

. . .

n1 ,n1 n1 ,n1 n1 ,n n1 ,n
0

n ,n n ,n

The integral of the weight function over the whole Preisach plane is numerically realized
by the sum of all elements in the Hadamard product matrix as
ZZ
[e](t)
b = (, ) [e](t)dd

n n
i, j i, j
i=1 j=1
= (S V) . (5.17)

As the weight function matrix V is a lower triangular matrix, by applying the Hadamard
product to the S matrix and V matrix, the resultant matrix in (5.3.2) is also a lower triangular
matrix. The value of the relay switch operator i , j becomes zero by the Hadamard product.
In this way, the values of relay switch operators in the S matrix are saved and updated for
calculation of . but they are not involved in the integral approximation.
Finally, the resultant [e](t)
b is applied in (5.6) and (5.3) for computation of the output
magnetic field Hsimu (t).

71
5 Model Identification and Calculation

5.4 Validation of the Inverted Preisach Model

To evaluate the accuracy of the inverted model, relative errors between the measured and
simulated magnetic field are calculated and expressed as
 
 Hmeas [k] Hsimu [k] 
[k] =  . (5.18)
Hmax 

For an input with a size of N, the accuracy of the model is evaluated with the mean relative
error .
In this work, simulations were performed on hysteresis curves of several different magnetic
materials, which have been mentioned in Chapter 4. The inverted model was verified for both
soft and hard magnetic materials. Besides major hysteresis loops, minor loops and FORCs
can also be simulated. By comparison, the simulation results show good approximation to the
measurement data, which can be demonstrated and evaluated with relative errors.

5.4.1 Simulation of Major Loops

The proposed inverted hysteresis model is identified with major loops. The agreement between
simulated and measured major loops demonstrates the accuracy of the identification method.
The following presents the simulated major loops of six materials. Table 5.1 lists the re-
sultant parameters determined by applying the quasi-Newton method algorithm. It can be
seen that the parameter for the soft material M330-35A ( = 10) is much greater than
applied for the other materials. The electrical steel M330-35A can easily be magnetized
with a low magnetic field. The magnetic permeability increases rapidly near the coercivity
(53.73 A/m) and decreases when the material reaches saturation. In the case of the inverted
hysteresis model, the change of the output relative to input is analyzed with the reluctivity,
which increases rapidly when the input reaches the maximum (saturation magnetization). As
parameter is the exponent in the weight function expression 3.21, the branch slope of major
loops can be controlled with the value of . The major loops with a sharp slope near the
maximum of the input can be approximated with a large value of .

Table 5.1: Identified parameters for weight function

1 2 A h b a
M330-35A 1.22 2.15 2.05 0.438 10 0.981 0.0068
M1300-100A 1.07 2.12 4.43 0.399 4 0.977 0.0118
Fe49 Co49 V2 0.596 0.823 2.02 -0.178 2 0.87 0.0873
Ni81 Fe19 0.903 0.542 1.83 -0.608 2 0.77 0.175
35Z155-RD 0.996 1.89 32 0.405 6 0 0
35Z155-TD 0.828 1.33 40.5 0.464 2 0 0

72
 5.4 Validation of the Inverted Preisach Model

500 1000
Measured Measured
Simulated Simulated
(e) Hmax (e) H
250 500 max
Magn. field H (A/m)

Magn. field H (A/m)

0 0

250 500

500 1000
1.5 0.75 0 0.75 1.5 1.6 0.8 0 0.8 1.6
Magn. polarization J (T) Magn. polarization J (T)
(a) (b)
14 10
Rela. error (%)

Rela. error (%)

7 5

0 0
1.5 0.75 0 0.75 1.5 1.6 0.8 0 0.8 1.6
Magn. polarization J (T) Magn. polarization J (T)
(c) (d)

Figure 5.6: Simulated major loops of electrical steels M330-35A and M1300-100A. Measure-
ments were carried out with the Epstein frame method at 50Hz at room tempera-
ture. (a) shows the measured and simulated major loops of M330-35A samples. (c)
shows the relative errors between the measured and simulated major loops in (a).
The measured and simulated major loops of M1300-100A samples are presented
in (b). The relative errors are presented in (d).

The simulated major loops of electrical steels M330-35A and M1300-100A are shown in
b Hmax is presented with the measured and simulated major loops for
Fig. 5.6(a) and (b). (e)
b Hmax can be considered as a major loop, which shares the same coercivity
comparison. (e)
and saturation magnetic polarization with the measured major loop. It can be observed in Fig.
5.6(c) and (d) that the relative errors of both simulation results are under 13%. However,
large errors are distributed near the maximum of the input magnitude (12.75% for M330-35A
and 8.89% for M1300-100A). This may be due to the difficulty in defining the saturation
magnetization and the asymmetry of the measured major loops. But the simulated coercivity,
which is usually taken as an important material parameter, has a reasonably small relative error
(less than 1% for both results).

73
5 Model Identification and Calculation

4
x 10
6 3000
Measured Measured
Simulated Simulated
(e) Hmax (e) H
max
3 1500
Magn. field H (A/m)

Magn. field H (A/m)

0 0

3 1500

6 3000
1.8 0.9 0 0.9 1.8 0.8 0.4 0 0.4 0.8
Magn. polarization J (T) Magn. polarization J (T)
(a) (b)
8 12
Rela. error (%)
Rela. error (%)

4 6

0 0
1.8 0.9 0 0.9 1.8 0.8 0.4 0 0.4 0.8
Magn. polarization J (T) Magn. polarization J (T)
(c) (d)

Figure 5.7: Simulated major loops of Fe49 Co49 V2 and Ni81 Fe19 thin film samples. Measure-
ments were carried out with VSM at room temperature. (a) presents the measured
and simulated major loops of a 1.9 m Fe49 Co49 V2 thin film sample. (c) shows
the relative errors between the measured and simulated major loops in (a). The
measured and simulated major loops of a 1.7 m Ni81 Fe19 thin film sample are
presented in (b). The relative errors are shown in (d).

The simulated major loops of Fe49 Co49 V2 and Ni81 Fe19 thin film samples are shown in
Fig. 5.7(a) and (b). Compared with the results of Fe49 Co49 V2 , the simulated major loop of
Ni81 Fe19 is not as close to the measurement data. As shown in Fig. 5.7(b), the magnetic field
increases sharply near the maximal polarization 0.8T. This loop can hardly be approximated.
It has a maximal relative error 10.56% near the remanent polarization 0.32T.
The oriented electrical steel 35Z155 has excellent magnetic property in the rolling direction
(RD) compared with the transverse direction (TD). As shown in Fig. 5.8(a) and (b), the
35Z155-RD has a very low coercivity (24.83 A/m), it can be saturated by applying a magnetic
field of 50 A/m. The measured major loop of 35Z155-TD has an abnormal shape. Both of the

74
 5.4 Validation of the Inverted Preisach Model

50 260
Measured Measured
Simulated Simulated

25 130
Magn. field H (A/m)

Magn. field H (A/m)

0 0

25 130

50 260
1.5 0.75 0 0.75 1.5 1.2 0.6 0 0.6 1.2
Magn. polarization J (T) Magn. polarization J (T)
(a) (b)
40 16
Rela. error (%)

Rela. error (%)

20 8

0 0
1.5 0.75 0 0.75 1.5 1.2 0.6 0 0.6 1.2
Magn. polarization J (T) Magn. polarization J (T)
(c) (d)

Figure 5.8: Simulated major loops of oriented electrical steels 35Z155 in rolling (RD) and
transverse directions (TD). Measurements were carried out with the Epstein frame
method at 50Hz at room temperature. (a) presents the measured and simulated
major loops of 35Z155-RD, (c) shows the relative errors between the measured
and simulated major loops in (a). The measured and simulated major loops of
35Z155-TD are presented in (b). The relative errors are shown in (d).

major loops are difficult to approximate. By applying the inverted Preisach model, the fitted
loops have acceptable relative errors. The mean relative error is 6.43% for 35Z155-RD and
3.57% for 35Z155-TD.
The results presented above demonstrate that the simulated major loops have good agree-
ment with measurement data, especially for the major loops of non-oriented steels. The asym-
metry of the measured major loops and the improper definition of the saturation magnetization
can lead to large relative errors. In order to decrease relative errors, the major loops for iden-
tification should contain as little of the constant part as possible. Aside from the errors caused
by the experimental data, the discretization size of the Preisach triangle n plays another im-
portant role in the accuracy of the simulated results.

75
5 Model Identification and Calculation

15 90
M33035A M33035A

Maximal relative error max (%)

M1300100A M1300100A
Mean relative error (%)

Fe49Co49V2 Fe49Co49V2
Ni81Fe19 Ni81Fe19
10 60
35Z155RD 35Z155RD
35Z155TD 35Z155TD

5 30

0 0
0 50 100 150 200 0 50 100 150 200
Discretization size n Discretization size n
(a) (b)

Figure 5.9: Relative errors of the simulations performed with different discretization size n.
(a) shows the mean relative errors to different discretization size n. (b) shows
the maximal relative errors max to different discretization size n.

As shown in Fig. 5.9(a) and (b), relative errors decrease obviously when the discretization
size n increases from 10 to 100. This change is not as significant when n is larger than 100.
Therefore, most simulations in this work were performed with a discretization size of 100.

5.4.2 Simulation of Asymmetric Hysteresis Loops

The inverted Preisach model keeps the wiping out and congruency properties, which enables
it to be applied in the simulation of asymmetric hysteresis loops, such as FORCs (see Fig.
5.10), hysteresis curves with minor loops (see Fig. 5.11) and the magnetizing hysteresis loops
(see Fig. 5.12). The simulated reversal curves, asymmetric and symmetric minor loops are
guaranteed to be located inside the major loops. Before calculation started, the inverted model
was identified with the outer major loops. Parameters of the weight function were determined
by applying the quasi-Newton method algorithm. In calculation the measured magnetic po-
larization was taken as the input. The magnetic field was computed and compared with the
experimental data.
As shown in Fig. 5.10, Fig. 5.11 and Fig. 5.12, the simulated asymmetric hysteresis loops
are quite close to the experimental data. The mean relative error is 1.42% for the FORCs
shown in Fig. 5.10. is about 2.19% for the hysteresis curves with minor loops shown in Fig.
5.11. is about 1.61% for the magnetizing hysteresis loops shown in Fig. 5.12.

76
 5.4 Validation of the Inverted Preisach Model

60
(a)

Magn. field H (kA/m)

30

30
Measured
Simulated
60
1.8 0.9 0 0.9 1.8
Magn. polarization J (T)
Magn. polarization

1.8
(b)
J (T)

1.8
0 100 200 300 400
60
(c) Measured
30
Magn. field

Simulated
H (kA/m)

30

60
0 100 200 300 400
8
Rela. error

(d)
(%)

0
0 100 200 300 400
Data index

Figure 5.10: Simulation results of FORCs. Measurement was carried out with VSM on a
1.9 m Fe49 Co49 V2 thin film sample at room temperature. (a) presents the
measured and simulated FORCs. (b) shows the input magnetic polarization.
(c) presents the measured and simulated magnetic field to the data index. (d)
shows the relative errors. Reversal points on the FORCs are at 0 kA/m, 5 kA/m,
10 kA/m, 15 kA/m, 20 kA/m, 30 kA/m, 40 kA/m and 60 kA/m.

77
5 Model Identification and Calculation

60
(a)

Magn. field H (kA/m)

30

30
Measured
Simulated
60
1.8 0.9 0 0.9 1.8
Magn. polarization J (T)
Magn. polarization

1.8
(b)
J (T)

1.8
0 100 200 300
60
(c) Measured
30
Magn. field

Simulated
H (kA/m)

30

60
0 100 200 300
10
Rela. error

(d)
(%)

0
0 100 200 300
Data index

Figure 5.11: Simulation results of minor loops. Measurement was carried out with VSM on
a 1.9 m Fe49 Co49 V2 thin film sample at room temperature. (a) presents the
measured and simulated hysteresis curves with minor loops. (b) shows the input
magnetic polarization. (c) presents the measured and simulated magnetic field to
the data index. (d) shows the relative errors. Minor loops are located at 0 kA/m,
10 kA/m, 20 kA/m and 30 kA/m.

78
 5.4 Validation of the Inverted Preisach Model

500
(a)

250
Magn. field
H (A/m)

250
Measured
Simulated
500
1.4 0.7 0 0.7 1.4
Magn. polarization J (T)
Magn. polarization

1.4
(b)
J (T)

1.4
0 1000 2000 3000
500
(c) Measured
250
Magn. field

Simulated
H (A/m)

250

500
0 1000 2000 3000
12
Rela. error

(d)
(%)

0
0 1000 2000 3000
Data index

Figure 5.12: Simulation results of magnetizing hysteresis loops. Measurement was carried
out with the Epstein frame method on M330-35A steel samples at 10Hz at room
temperature. (a) presents the measured and simulated magnetizing hysteresis
loops. (b) shows the input magnetic polarization. (c) presents the measured and
simulated magnetic field to the data index. (d) shows the relative errors. The
discretization size n = 200.

79
5 Model Identification and Calculation

5.5 Comparison of Identification Methods

The identification method of a hysteresis model plays an important role in enhancing the
modeling accuracy. In this section, the identification method of the inverted Preisach model
is compared with the Everett function method applied in the pseudo-inverse approach. Table.
5.2 shows the comparison details. When the inverted Preisach model is applied, only a major
loop is required to determine the 7 parameters in the weight function. With the pseudo-inverse
approach, a set of FORCs are measured to calculate the inverted Everett matrix Enn . The
measurement of major loops is easier and faster than the measurement of FORCs. Moreover,
the experimental errors in measuring major loops can be reduced by averaging the testing
values of repeated measurements. But this is difficult in the measurement of FORCs. As the
inverted Everett matrix is directly calculated from the FORCs, the experimental errors may be
carried over into the simulation results.
To compare the identification accuracy, simulations of FORCs were performed both with
the inverted Preisach model and the pseudo-inverse approach. Simulation results are shown
in Fig. 5.13 and Fig.5.14. The inverted Everett matrix Enn is calculated by applying the
bilinear interpolation method. The discretization size of the inverted Preisach triangle is 100.
As shown in Table. 5.2, by applying the pseudo-inverse approach, the mean relative error is
2.67% for 40 FORCs and 1.97% for 60 FORCs. Adding 20 more reversal curves for identi-
fication decreases the mean relative error by about 26%. However, this gain is offset by the
considerable measurement effort required. Compared with the pseudo-inverse approach, the
inverted Preisach model requires only a major loop for identification. The identification pro-
cess is simpler and faster than the pseudo-inverse approach. The simulation results display
good agreement with the experimental data. These will be great advantages in attracting FE
users.

Table 5.2: Comparison of identification methods

Models Inverted Preisach Model Pseudo-inverse Approach
Less More
Experimental data
A major loop A set of FORCs
7 parameters Inverted Everett matrix
Determined parameters
in weight function Enn
Easy and fast Hard and slow
Measurement methods
Both AC or DC methods DC method
Mean relative error
2.37% 2.67%
(40 FORCs)
Mean relative error
2.09% 1.97%
(60 FORCs)

80
 5.5 Comparison of Identification Methods

Magn. polarization J (T) 1.8

1.8
0 200 400 600 800
Data index
(a)
60 60
Measured Measured
Simulated Simulated

30 30
Magn. field H (kA/m)

0 Magn. field H (kA/m) 0

30 30

60 60
1.8 0.9 0 0.9 1.8 1.8 0.9 0 0.9 1.8
Magn. polarization J (T) Magn. polarization J (T)
(b) (c)
18 14
Rela. error (%)

Rela. error (%)

9 7

0 0
0 200 400 600 800 0 200 400 600 800
Data index Data index
(d) (e)

Figure 5.13: Simulation results of FORCs with 40 reversal curves. The FORCs were mea-
sured with VSM on a 1.9 m Fe49 Co49 V2 thin film sample. (a) shows the input
magnetic polarization. (b) presents the simulation results computed by applying
the inverted Preisach model. (d) shows the relative errors of the simulation results
in (b). The simulation results computed by applying the pseudo-inverse approach
are shown in (c). The relative errors are presented in (e).

81
5 Model Identification and Calculation
Magn. polarization J (T)

1.8

1.8
0 500 1000 1500 2000
Data index
(a)
60 60
Measured Measured
Simulated Simulated
30 30
Magn. field H (kA/m)

Magn. field H (kA/m)

0 0

30 30

60 60
1.8 0.9 0 0.9 1.8 1.8 0.9 0 0.9 1.8
Magn. polarization J (T) Magn. polarization J (T)
(b) (c)
10 14
Rela. error (%)

Rela. error (%)

5 7

0 0
0 500 1000 1500 2000 0 500 1000 1500 2000
Data index Data index
(d) (e)

Figure 5.14: Simulation results of FORCs with 60 reversal curves. The FORCs were mea-
sured with VSM on a 1.9 m Fe49 Co49 V2 thin film sample. (a) shows the input
magnetic polarization. (b) presents the simulation results computed by applying
the inverted Preisach model. (d) shows the relative errors of the simulation results
in (b). The simulation results computed by applying the pseudo-inverse approach
are shown in (c). The relative errors are presented in (e).

82
6 Calculation Optimization
The inverted Preisach model introduced in this work can be identified with less experimental
data. With small relative errors it can be applied to describe the hysteresis behavior of both
hard and soft magnetic materials. To include the inverted model into FE codes, the calculation
efficiency is another important factor.
In comparison to the inverted Preisach model, the pseudo-inverse approach needs much
more experimental data for identification. However, the integral expression of the pseudo-
inverse approach can be rewritten in a discrete form (see equations in 3.12 or 3.13). By means
of the discrete form, the output is calculated by adding or subtracting the inverted Everett
function values, which are computed on the vertex of the staircase line. In the precalculation
stage, the inverted Everett function values are calculated on each node of the mesh covering
the whole Preisach triangle. All these Everett function values are saved in the inverted Everett
matrix as constant. During calculation, the Everett function values on the vertices of the
staircase line are obtained by interpolation in this inverted Everett matrix. In this way, it costs
little memory space and it is very efficient in calculation.
As introduced in Section 5.3.2, the inverted Preisach model is computed by applying a
numerical approximation method. A square mesh is generated to cover the whole Preisach
triangle. In each mesh cell the switch value and weight function value are calculated and
saved in two matrices. During the calculation process, all the elements in the switch matrix
have to be updated according to the current input. This costs both much computation time and
considerable memory usage.
To improve the calculation efficiency, research was conducted on the discrete form of the
inverted Preisach model. In this chapter the new algorithm utilizing the discrete form is in-
troduced. This algorithm was evaluated by comparing it with the previous algorithm and the
pseudo-inverse approach1 .

6.1 Discrete Form of the Inverted Preisach Model

In this work, the discrete form of the inverted Preisach model is proposed with reference to the
pseudo-inverse approach. As shown in Fig. 6.1, the inverted Preisach triangle T is separated
into two subsets S+ (t) and S (t) by a staircase interface G(t). The modified switch operators
in the S+ (t) area are equal to +1, while the ones in the S (t) area are equal to 1. The

1 This discrete form and the algorithm evaluation introduced in this chapter has been published in [67, 68]

83
6 Calculation Optimization

inverted Preisach model in (5.10) can be written as

ZZ ZZ
[e](t)
b = (, )dd + 2 (, )dd . (6.1)
T S+ (t)

Once the model has been identified, the integral of the weight function over the whole Preisach
triangle T can be considered as a constant. Therefore, the output [e](t)
b in (6.1) depends only
+
on the integral of the weight function over the positive region S (t).
To find the discrete form of the inverted Preisach model in (3.14), geometric analysis of the
Preisach plane was performed and this is illustrated in Fig.
6.1. The Preisach plane of the
inverted Preisach model is in P0 = (, ) R2 : . The staircase line G(t) separates
the S+ (t) and S (t) subsets in the Preisach plane. The points (Mk , mk ) are the vertices of G(t).
Mk and mk are the and coordinates of these vertices. The values of Mk and mk can be
obtained from the extrema of input history.

Figure 6.1: Geometric interpretation of the rectangular areas in the S+ (t) subset of the inverted
Preisach plane. G(t) is the staircase interface dividing S+ and S subsets. (a) The
magnetization process starts from the positive saturation state. (b) The magnetiza-
tion process starts from the negative saturation state.

As shown in Fig. 6.1, the S+ (t) subset can be divided into several rectangles and a triangle.
The rectangle R in Fig. 6.1 can be the difference between two trapezoids, e.g., the rectangle R
in Fig. 6.1(a). In another case, the rectangle can be the difference between a trapezoid and a

84
 6.1 Discrete Form of the Inverted Preisach Model

triangle, such as the rectangle R in 6.1(b). The points (Mk , mk ) are the upper right vertices of
these trapezoids or triangles.
In this work, an area S( p , q ) is defined as S(p , q ) = (, ) P0 , p , q .


As shown in Fig. 6.2, the area S( p , q ) always starts from the bottom left vertex of the
Preisach triangle (1, 1) and takes the point ( p , q ) as the upper right vertex. When the
point ( p , q ) is located on the line = , as shown in Fig. 6.2(a), the area S( p , q ) is a
triangle. Otherwise the area S( p , q ) is a trapezoid [see Fig. 6.2 (b)].

Figure 6.2: Geometric interpretation of the area S( p , q ) in the inverted Preisach plane. (a)
the triangle case. (b) the trapezoid case.

Here a function F( p , q ) is defined as the integral of the weight function over the area
S( p , q ). It is expressed as
ZZ
F( p , q ) = (, ) dd . (6.2)
S( p ,q )

A rectangle Rk shown in Fig. 6.1 can be considered as the difference between S(Mk , mk ) and
S(Mk , mk1 ). The integral of the weight function over this rectangle Rk is calculated as
ZZ ZZ ZZ
(, )dd = (, )dd (, )dd
Rk (t) S(Mk ,mk ) S(Mk ,mk1 )
= F(Mk , mk ) F(Mk , mk1 ). (6.3)

As shown in Fig. 6.1, besides these rectangles in the S+ (t) subset, there is always an
extra triangle. The upper right vertex of this triangle is on the point (e(t), e(t)). The integral
of the weight function over this triangle can be calculated by F (e(t), e(t)). Similarly, the
integral of the weight function over the whole Preisach triangle T can be calculated by F(1, 1).
According to the geometric interpretation shown in Fig. 6.1, the integral expression in (6.1)

85
6 Calculation Optimization

can be rewritten in a discrete form as

n(t)
[e](t)
b = F(1, 1) + 2F (e(t), e(t)) + 2 [F(Mk , mk ) F(Mk , mk1 )], (6.4)
k=1

where n(t) is the number of rectangles that S+ can be divided into.

In this work, an optimized algorithm is designed to calculate the output with this discrete
form in (6.4). Similarly to the pseudo-inverse approach, the integral computation can be re-
placed by adding or subtracting the F function values on the vertices of staircase interface.

6.2 Precalculation of the F Matrix

In the optimized algorithm, a precalculation procedure is required to calculate F function
values in the mesh cells. As introduced in Section 5.3.1, a n n square mesh is generated
to uniformly discretize the whole Preisach triangle. The integral of the weight function over
each mesh cell is numerically approximated as i , j . As expressed in (5.15 ), all i , j values
calculated in the mesh can form a n n lower triangular matrix, which is called the weight
function matrix and labeled with V.
As introduced in (6.2), F( p , q ) is the integral of the weight function over area S( p , q ).
In calculation, it can be numerically approximated by summing all the unit integrals i , j
covered inside the area S( p , q ). Mathematically F( p , q ) is calculated by selecting a sub-
matrix of V and summing all the i , j elements in this submatrix. This can be expressed
as ZZ p q
F( p , q ) = (, ) dd i, j = F p,q . (6.5)
i=1 j=1
S( p ,q )

All F p ,q calculated on nodes ( p , q ) can form a n n lower triangular matrix F where

F1 ,1 F1 ,2 F1 ,n1 F1 ,n

F2 ,2 F2 ,n1 F2 ,n

F= .. .
.. ..
. (6.6)

. .

Fn1 ,n1 Fn1 ,n
0

Fn ,n

Fig. 6.3 presents an example to explain how F p ,q is calculated from i , j . The nodes of
the mesh 2 represent the nonzero elements in the V and F matrices. The elements i , j with
(i > j) are zero as they are out of the Preisach plane. As shown in Fig. 6.3, the element F4 ,4
is an approximated integral over a triangle, the i , j included inside the dashed square are
selected as a submatrix. F4 ,4 is calculated by summing all the elements in this submatrix. In
the same way, (4 , 7 ) is the upper right vertex of a trapezoid, F4 ,7 is calculated by summing
2 The black nodes in Fig. 6.3 represent the selected elements in V and F matrices. The zero elements in the
triangular part of V and F matrices are not shown here.

86
 6.2 Precalculation of the F Matrix

Figure 6.3: Calculation of F p ,q from i , j . The Preisach plane is discretized by a n n (here

n = 9) square mesh.

all the elements included inside the dashed rectangle.

In the precalculation stage of the optimized algorithm, parameters of the weight function
are determined from the measured major loops. Integrals of the weight function over mesh
cells are numerically approximated and saved in a weight function matrix V. According to
(6.5) and (6.6), F function values can be calculated from the elements in V matrix. These F
function values form a F matrix.
Fig. 6.4 presents the V matrix and the F matrix for simulating the hysteresis curves of
an Fe49 Co49 V2 thin film sample and M330-35A steel samples. The the V matrix shown in
Fig. 6.4(c) has a distinct increase near the vertices of the Preisach triangle. This can be
deduced from the shape of major loop measured on M330-35A steel samples. With a very
low coercivity the magnetic field changes sharply relative to the magnetic polarization near
saturation.
The matrix F is precalculated and applied as a constant matrix in the calculation procedure.
According to the discrete form of the inverted model in (6.4), the output is calculated by adding
or subtracting the F function values. During calculation, vertices of the staircase interface G(t)
are updated according to the current input. The values of F(, ) on the vertices are obtained
by applying the interpolation method in the F matrix.

87
6 Calculation Optimization

Norm. F(, )
1
Norm. (, )

0.5 0.5

0 1 0 1
1 1
0 0 0 0

1 1
1 1
(a) (b)

Norm. F(, )
1
Norm. (, )

0.5 0.5

0 1 0 1
1 1
0 0 0 0

1 1
1 1
(c) (d)

Figure 6.4: The weight function matrix V and the calculated F matrix. All of these matrices
are normalized. (a) is the weight function matrix V identified from the major loop
of an Fe49 Co49 V2 thin film sample. (b) is the F matrix calculated from the V
matrix in (a). (c) is the weight function matrix V identified from the major loop
of M330-35A steel samples at 10Hz at room temperature. (d) is the F matrix
calculated from the V matrix in (c).

6.3 Updating of the Staircase Interface G(t)

In the calculation procedure of the optimized algorithm, it is not necessary to save or update
the entire switch matrix S according to the current input. Instead, only the vertices of the
staircase interface G(t) are required to be saved and updated with a matrix G where
 
M0 M1 Mk
G= . (6.7)
m0 m1 mk

The and coordinates of vertices are saved in the two rows of this matrix. k represents the
number of vertices on the staircase interface G(t).
As introduced in Section 3.2.1, by applying different switch operators and , the
resultant staircase interfaces L(t) and G(t) have geometrical correlation in the Preisach plane
(see Fig. 3.7). The coordinates of vertices on the staircase interface G(t) can be obtained by

88
 6.3 Updating of the Staircase Interface G(t)

means of its geometrical relation with L(t). In calculation, a matrix L is applied to save the
coordinates of vertices on the staircase interface L(t) as
 0
M0 M10 Mk0

L= . (6.8)
m00 m01 m0k

The coordinates of vertices on the staircase interface L(t) directly refer to the extrema of
input history. According to the geometrical correlation between L(t) and G(t), the vertices of
G(t) can be obtained from the vertices L(t) by a coordinate conversion

mi = 1 Mi0 + e(t),
Mi = e(t) m0i 1, (6.9)

where i = 1, 2, ..., k. e(t) is the normalized input and 1 is the maximal input after normaliza-
tion.

Figure 6.5: Updating the vertices in staircase interface L(t) according to the input extrema.

Based on the wiping out property of the Preisach model, the vertices in the staircase in-
terface L(t) are updated according to the local input extrema. As shown in Fig. 6.5, in the
increasing case, the precious local input maxima which are not greater than the current input
e will be wiped out, such as the local maximum M2 e. In the staircase interface L(t) of
the Preisach plane, the vertices whose coordinate are equal to these wiped maxima will be
deleted.
This occurs in a similar manner for monotonically decreasing input. The precious local
input minima which are not smaller than the current input e will be wiped out, such as the
local minimum m3 e. In the staircase interface L(t) of the Preisach plane, the vertices
whose coordinate are equal to these wiped minima will be deleted.

89
6 Calculation Optimization

In calculation, the corresponding columns including these wiped out extrema will be deleted
in the L matrix. The updated L matrix is converted to the G matrix by the coordinate con-
version described in (6.9). According to the vertices coordinates saved in the G matrix, the
corresponding F function values are obtained by bilinear interpolation in the F matrix. These
F function values are applied to calculate the output with the discrete form in (6.4).

6.4 Evaluation of the Optimized Algorithm

To evaluate the efficiency of the optimized algorithm, the average memory usage and execution
time of the program are calculated by applying the profile performance and memory tool
in LabVIEW. The two evaluation factors are compared with the previous algorithm and the
pseudo-inverse approach.
As shown in Fig. 6.6(a), the average memory usage is roughly linear, proportional to the
size of input. It clearly shows that a large amount of memory is in use when the previous
algorithm is applied for calculation. The average memory used in the previous algorithm
is about 1.5 times the average memory usage required in the pseudo-inverse approach. The
optimized algorithm and the pseudo-inverse approach uses nearly the same amount of memory
in calculation. In the previous algorithm a n n switch matrix S is applied to save the state
value of switch operators in the whole Preisach plane.
According to the definition of in (3.15), in the memory area of the Preisach plane the
value of switch operators depend upon their history value. Therefore, the switch matrix S cal-

4
380 10
Average memory usage (kBytes)

Average execution time (ms)

330 Pseudo inverse

3
10 Previous algorithm
Pseudo inverse Optimized algorithm
280 Previous algorithm
Optimized algorithm
2
10
230

1
180 10
1000 3000 5000 7000 9000 1000 3000 5000 7000 9000
Input size N Input size N
(a) (b)

Figure 6.6: Efficiency comparison of the three algorithms with average memory usage (a) and
average execution time (b). N is the size of the input M[k]. The discretization size
n is set to be 100 in the simulation.

90
 6.4 Evaluation of the Optimized Algorithm

culated in the last step has to be saved in a register to calculate the current S matrix. However,
for the pseudo-inverse approach and the optimized algorithm, only the vertices coordinates
of the staircase interface are updated in calculation. In this way, lots of memory usage is
economized in the calculation procedure.
As shown in Fig. 6.6(b), the previous algorithm costs the most execution time. The opti-
mized algorithm is about 28 times faster than the previous algorithm. Comparing the pseudo-
inverse approach and the new algorithm, there is no significant difference in the calculation
efficiency between the two algorithms. This demonstrates that the coordinate conversion from
the L matrix to the G matrix costs little execution time. Compared with the previous algo-
rithm, which needs to calculate the n2 switch operators in every step, the optimized algorithm
performs more efficiently in calculation.
The comparisons in Fig. 6.6 demonstrate that the calculation efficiency of the inverted
Preisach model has clearly been improved by applying the discrete form in computation. The
efficiency of the optimized algorithm is very close to the pseudo-inverse approach. In this way,
the inverted Preisach model proposed in this thesis not only has an advantage in identification,
but it is also efficient in calculation.

91
7 Conclusions and Outlook
7.1 Conclusions
In this thesis an inverted hysteresis model is proposed to include ferromagnetic hysteresis
in the electromagnetic field calculations. The model is based on the classic Preisach model,
which has been widely applied within FE applications, since it approaches the requirements of
accuracy, simplicity, and efficiency. The proposed inverted model applies a modified switch
operator, which makes it possible to approximate the weight function distribution with a differ-
entiable function. This inverted Preisach model with analytic weight function has advantages
in identification. A limited number of parameters in the weight function can be identified
from a major hysteresis loop. Compared with a set of FORCs required in the pseudo-inverse
approach, a major hysteresis loop contains less experimental data and the measurement is
more accessible. In order to evaluate the accuracy of this model, simulations were made on
both symmetric and asymmetric hysteresis curves of magnetic materials. The mean relative
errors of the simulation results are less than 7%. From an application point of view, accurate
modeling with less experimental data may attract more FE users.
Besides the advantage in identification, the efficiency in calculation plays another important
role in the implementation work. As this inverted model keeps the basic properties of the
Preisach model, its integral expression can be rewritten in a discrete form. The application of
this discrete form has obviously improved the calculation efficiency. The execution time of
the optimized algorithm is about 28 times faster than the previous algorithm. Comparing the
optimized algorithm with the pseudo-inverse approach, there is no significant difference in the
execution time and memory usage.

Inverted Preisach Model with Differentiable Weight Function

The inverted Preisach model proposed here introduces the magnetization M (or magnetic flux
B) as input and takes the magnetic field strength H as output. Without an iterative proce-
dure, the inverted model can be directly implemented into the FE formulations, in which the
magnetic vector potential is taken as an unknown.
In this thesis analysis concentrated firstly on the pseudo-inverse of the Preisach model,
which is most commonly applied to realize the Preisach inversion. The pseudo-inverse keeps
the switch function of the forward Preisach model and simply exchanges the roles of the input
and output. The advantage of the pseudo-inverse approach is the possibility of applying the
Everett function method. In this way, it is not necessary to find a suitable weight function for
the inverted model. However, the inverted Everett functions have to be identified from a set of
FORCs, which contain a large amount of experimental data. The measurement of FORCs is

93
7 Conclusions and Outlook

tedious and difficult. Furthermore, the experimental errors may be taken into the simulation
results.
To solve this problem, an attempt was made to find a suitable analytic weight function,
which could be identified with less experimental data. With the help of FORCs, the dis-
tribution of the weight function was calculated by applying numerical approximation. This
demonstrates that the weight function is partly positive and partly negative in the Preisach
plane (see Fig. 3.2). It is difficult to find a suitable function to approximate this distribution.
In the Preisach integral, the weight function is multiplied by the switch operator , which
can be represented by rectangular hysteresis loops. The Preisach model is constructed as a
superposition of an infinite number of the operators [16]. Fig. 7.1(a) and (c) present the
M(H) loop (forward major loop) and the elementary rectangular loop of . As shown in
Fig. 7.1(a), the magnetization output increases along the lower branch and decreases along the
upper branch. The elementary rectangular loop in Fig. 7.1(c) has corresponding ascending and
descending paths. In contrast, in the H(M) loop (inverted major loop) shown in Fig. 7.1(b),
the magnetic field increases along the upper branch and decreases along the lower branch. The
pseudo-inverse of the Preisach model still keeps the switch operator , whose rectangular
loop has opposite paths to the inverted major loop. This may be the leading cause of the
complicated weight function distribution.

Figure 7.1: Comparison of the forward (a) and inverted (b) major loops. Comparison of the
forward (c) and inverted (d) elementary rectangular loops.

According to this analysis, modification was made on the switch operator . A modified
switch operator was proposed and applied in the new inverted Preisach model. As shown
in Fig. 7.1(d), the rectangular hysteresis loop of has similar ascending and descending
paths to the inverted major loop shown in Fig. 7.1(b). By means of the modified switch
operator , the approximated weight function distribution is almost completely positive in
the Preisach plane (see Fig. 3.9). According to the shape of the approximated distribution,

94
 7.1 Conclusions

a differentiable analytic expression was proposed as the weight function (, ) of the in-
verted Preisach model. The application of analytic weight function can improve the flexibility
of modeling. Furthermore, the new inverted model retains the wiping out and congruency
properties, which can offer great advantage in calculation.

Simple Identification from Major Hysteresis Loops

To identify the inverted Preisach model, only major loops are required as experimental data.
It is also possible to use minor loops to improve the simulation accuracy. Seven parameters
are determined from measured major loops. A quasi-Newton method algorithm is applied to
optimize the mean squared error (MSE) between the measured and simulated data.
The inverted model was verified for both soft and hard magnetic materials. Simulations
were made on hysteresis loops of magnetic thin films, oriented and non-oriented electrical
steels sheets. Besides symmetric hysteresis loops, asymmetric hysteresis curves, such as minor
loops or FORCs can also be simulated. By comparison, the simulation results produced by the
inverted hysteresis model show good approximation to the measurement data. The accuracy
was evaluated with relative errors. The maximal relative errors mainly distributed near the
saturation magnetization input. The main possible reasons are

It is difficult to define an exact magnetization value in saturation. Due to the forced

magnetization effect, the magnetization may keep increasing with the magnetic field
after technical saturation is reached.

It is hard to choose the maximal magnetic field in major loop. In the inverted modeling,
the vertical straight part should be removed from H(M) loops as far as possible. The
vertical straight part will increase the difficulty and errors in the identification with major
loops.

The measured major loop is asymmetric. According to the definition of the Preisach
model, the major loops are supposed to be symmetric.

The discretization size of the Preisach triangle also affects the relative errors of simulation.
Both the maximal and the mean relative errors are clearly reduced by increasing the discretiza-
tion size n. Most simulations in this work were performed with a discretization size of 100.
The mean relative errors of these simulation results are less than 7%.

Efficient Calculation with Numerical Discrete Formulation

In the numerical computation of the inverted Preisach model, the double integral of weight
function over the whole Preisach plane is calculated by applying numerical approximation:

1. A n n square mesh is generated to cover the whole Preisach triangle.

2. The elementary square area s is multiplied to the (i , j ) value calculated at the mid-
dle points of each mesh square. The product s (i , j ) is taken as the approximated
double integral of the weight function over mesh square.

95
7 Conclusions and Outlook

3. All the resulted products s (i , j ) can be saved in a weight function matrix Vnn .

4. A switch matrix Snn is applied to save the current switch operator values at the middle
points of each mesh square.

5. Multiplication of the weight function and the switch value is fulfilled by applying the
Hadamard product of the S matrix and the V matrix.

6. The integral of the weight function over the whole Preisach plane is approximated by
the sum of all elements in the Hadamard product matrix S V

This method requires the S matrix to be constantly updated according to the current input.
It costs much memory usage and increases the computation time.
To make the calculation procedure comparable to the pseudo-inverse approach, the Everett
function method is taken for reference. A function F( p , q ) is defined to be equal to the
integral of the weight function over area S( p , q ). The area S( p , q ) always starts from the
bottom left vertex of the Preisach triangle and takes the point ( p , q ) as the upper right vertex.
The positive set S+ of the Preisach plane can be subdivided into segments. Each segment can
be calculated with F(, ) functions. Hence, the integral expression of the inverted model
has been rewritten in a numerical discrete form. During calculation, vertices of the staircase
interface are updated according to the current input. The output is calculated by adding or
subtracting the function values of F(, ). In comparison to updating the S matrix, much
memory space is saved and the calculation time decreases.
To evaluate the calculation efficiency, a comparison of memory usage and execution time
was made between the previous algorithm, the pseudo-inverse approach and optimized algo-
rithm. This demonstrates that: the previous algorithm utilizing the S matrix costs the most
memory usage; the new algorithm utilizing the numerical discrete formulation is about 28
times faster than the previous algorithm; comparing the pseudo-inverse approach and the new
algorithm, there is no significant difference in the execution time and memory usage of the
two algorithms.

Three Different Characterization Methods for Various Experimental Data

In this work three different characterization methods were applied: the Epstein frame method,
the double C-yoke SST and the vibrating sample magnetometer.

By means of the Epstein frame method, the average magnetic properties of electrical
steel samples were characterized. The obtained results show good consistency with the
standard data. Measurements can be carried out at different frequencies (10Hz 400Hz)
and at different temperatures (- 40 C 140 ).

By applying the double C-yoke SST, the local magnetic properties of electrical steel
samples were characterized. The change of the local magnetic properties can be applied
to evaluate the influence of plastic deformation.

96
 7.2 Outlook

By utilizing a VSM, the magnetic moment of small magnetic samples, such as the mag-
netic thin films, can be detected. By adding an orthogonal pickup coils systems, vector
characterization can be carried out in both undirectional and rotational magnetic field.

These measurement methods provide experimental data for identification of the inverted
Preisach model. The measurement results in different conditions, such as frequency, tempera-
ture, orientation, can be applied for model expansion in further research.

7.2 Outlook
The inverted model proposed in this thesis is based on the classic Preisach model, which is
mostly considered as a mathematical approach rather than a truly physical model. Hence, this
model can be applied mostly in the computation including hysteresis effect, especially in the
FE analysis.
A FE computation kernel has been developed in the Chair of Sensor Technology, University
of Erlangen- Nuremberg. It is possible to include this model in the FE code and apply it to
solve electromagnetic problems. In this way, the accuracy and calculation efficiency can be
compared with the iteration methods and other inverted hysteresis models.
The proposed inverted Preisach model is a rate-independent scalar model. As ferromagnetic
hysteresis could be influenced by many factors, this inverted model can be extended to take
the following effects into account.

Dynamic effect

Thermal effect

Stress effect

Effect caused by plastic deformation

The expansion could be accomplished with reference to the expansion of the forward Preisach
models [35, 55, 69]. Moreover, by introduction of an additional rotation operator, the inverted
hysteresis model can be applied to solve numerical problems where rotational electromagnetic
fields are involved [23, 70, 71, 72, 73, 74, 75].

97
Bibliography
[1] E. Dlala, Efficient algorithms for the inclusion of the preisach hysteresis model in non-
linear finite-element methods, IEEE Trans. Magn., vol. 47, no. 2, pp. 395408, 2011.

[2] M. Kuczmann and A. Ivanyi, The finite element method in magnetics. Akademiai Kiado,
2008.

[3] E. A. Dlala, Magnetodynamic vector hysteresis models for steel laminations of rotating
electrical machines. PhD thesis, Helsinki University of Technology, 2008.

[4] M. Kuczmann and A. Ivanyi, Neural network based vector hysteresis model and the non-
destructive testing method. PhD thesis, Ph. D. thesis, Budapest University of Technology
and Economics, 2004.

[5] E. Dlala, J. Saitz, and A. Arkkio, Inverted and forward preisach models for numerical
analysis of electromagnetic field problems, IEEE Trans. Magn., vol. 42, no. 8, pp. 1963
1973, 2006.

[6] M. Kuczmann, Measurement and simulation of vector hysteresis characteristics, IEEE

Trans. Magn., vol. 45, no. 11, pp. 51885191, 2009.

[7] D. Davino, A. Giustiniani, and C. Visone, Fast inverse preisach models in algorithms for
static and quasistatic magnetic-field computations, IEEE Trans. Magn., vol. 44, no. 6,
pp. 862865, 2008.

[8] J. Saitz, Magnetic field analysis of induction motors combining preisach hysteresis
modeling and finite element techniques, IEEE Trans. Magn., vol. 37, no. 5, pp. 3693
3697, 2001.

[9] H.-K. Kim, S.-K. Hong, and H.-K. Jung, An improved finite element analysis of
magnetic system considering magnetic hysteresis, IEEE Trans. Magn., vol. 36, no. 4,
pp. 689692, 2000.

[10] A. Ivany, Hysteresis models in electromagnetic computation, Akademiai Kiado, Bu-

dapest, 1997.

[11] F. Preisach, Uber die magnetische nachwirkung, Zeitschrift fur physik, vol. 94, no. 5-6,
pp. 277302, 1935.

[12] A. Visintin, Differential models of hysteresis. Springer Berlin, 1994.

99
Bibliography

[13] D. Everett and W. Whitton, A general approach to hysteresis, Transactions of the Fara-
day Society, vol. 48, pp. 749757, 1952.

[14] E. Madelung, On the magnetization by fast currents and an operation principle of the
magnetodetectors of rutherford-marconi, Ann. Phys, vol. 17, no. 5, pp. 861890, 1905.

[15] M. K. A. Pokrovskii, Systems with hysteresis, Moscow, Nauka, 1983.

[16] I. Mayergoyz, Mathematical models of hysteresis and their applications. Academic

Press, 2003.

[17] I. Mayergoyz and G. Friedman, Generalized preisach model of hysteresis, IEEE Trans.
Magn., vol. 24, no. 1, pp. 212217, 1988.

[18] G. Bertotti, Hysteresis in Magnetism: For Physicists, Materials Scientists, and Engi-
neers. Academic Press Series in Electromagnetism, Academic Press, 1998.

[19] E. Torre, Effect of interaction on the magnetization of single-domain particles, Audio

and Electroacoustics, IEEE Transactions on, vol. 14, no. 2, pp. 8692, 1966.

[20] S. E. Zirka, Y. I. Moroz, P. Marketos, and A. J. Moses, Congruency-based hysteresis

models for transient simulation, IEEE Trans. Magn., vol. 40, no. 2, pp. 390399, 2004.

[21] A. Adly and S. Abd-El-Hafiz, Using neural networks in the identification of preisach-
type hysteresis models, IEEE Trans. Magn., vol. 34, no. 3, pp. 629635, 1998.

[22] A. Adly and I. Mayergoyz, A new vector preisach-type model of hysteresis, Journal of
applied physics, vol. 73, no. 10, pp. 58245826, 1993.

[23] A. Sutor, J. Kallwies, and R. Lerch, An efficient vector preisach hysteresis model based
on a novel rotational operator, Journal of Applied Physics, vol. 111, no. 7, pp. 07D106
07D106, 2012.

[24] S. Tumanski, Handbook of Magnetic Measurements. Taylor & Francis, 2011.

[25] D. Jiles and D. Atherton, Theory of ferromagnetic hysteresis, Journal of magnetism

and magnetic materials, vol. 61, no. 1, pp. 4860, 1986.

[26] D. A. Philips, L. R. Dupre, and J. A. Melkebeek, Comparison of jiles and preisach

hysteresis models in magnetodynamics, IEEE Trans. Magn., vol. 31, no. 6, pp. 3551
3553, 1995.

[27] J. Saitz et al., Magnetic field analysis of electric machines taking ferromagnetic hystere-
sis into account. PhD thesis, Helsinki University of Technology, 2001.

[28] J. Saitz, Computation of the core loss in an induction motor using the vector preisach
hysteresis model incorporated in finite element analysis, IEEE Trans. Magn., vol. 36,
no. 4, pp. 769773, 2000.

100
 Bibliography

[29] N. Takahashi, S. Miyabara, and K. Fujiwara, Problems in practical finite element analy-
sis using preisach hysteresis model, IEEE Trans. Magn., vol. 35, no. 3, pp. 12431246,
1999.

[30] G. Consolo, G. Finocchio, M. Carpentieri, E. Cardelli, and B. Azzerboni, About iden-

tification of scalar preisach functions of soft magnetic materials, IEEE Trans. Magn.,
vol. 42, no. 4, pp. 923926, 2006.

[31] B. Azzerboni, E. Cardelli, G. Finocchio, and F. La Foresta, Remarks about preisach

function approximation using lorentzian function and its identification for nonoriented
steels, IEEE Trans. Magn., vol. 39, no. 5, pp. 30283030, 2003.

[32] O. Henze and W. M. Rucker, Identification procedures of preisach model, IEEE Trans.
Magn., vol. 38, no. 2, pp. 833836, 2002.

[33] J. Fuzi, Analytical approximation of preisach distribution functions, IEEE Trans.

Magn., vol. 39, no. 3, pp. 13571360, 2003.

[34] S. Naidu, Simulation of the hysteresis phenomenon using preisachs theory, IEE Pro-
ceedings A (Physical Science, Measurement and Instrumentation, Management and Ed-
ucation), vol. 137, no. 2, pp. 7379, 1990.

[35] A. Sutor, S. Rupitsch, and R. Lerch, A preisach-based hysteresis model for magnetic and
ferroelectric hysteresis, Applied Physics A: Materials Science & Processing, vol. 100,
no. 2, pp. 425430, 2010.

[36] H. Czichos, T. Saiato, and L. L. R. Smith, Springer handbook of materials measurement

methods. Springer, 2006.

[37] E. Antonelli, E. Cardelli, and A. Faba, Epstein frame: How and when it can be really
representative about the magnetic behavior of laminated magnetic steels, IEEE Trans.
Magn., vol. 41, no. 5, pp. 15161519, 2005.

[38] S. Foner, Vibrating sample magnetometer, Review of Scientific Instruments, vol. 27,
no. 7, pp. 548548, 1956.

[39] L. Benito, J. Arnaudas, and A. Del Moral, High-sensitivity vector magnetometer

for measuring magnetic torque at low temperatures, Review of scientific instruments,
vol. 77, no. 2, pp. 025101025101, 2006.

[40] J. Mallinson, Magnetometer coils and reciprocity, Journal of Applied Physics, vol. 37,
no. 6, pp. 25142515, 1966.

[41] S. Foner, The vibrating sample magnetometer: Experiences of a volunteer, Journal of

applied physics, vol. 79, no. 8, pp. 47404745, 1996.

[42] B. D. Cullity and C. D. Graham, Introduction to magnetic materials. Wiley-IEEE Press,

2011.

101
Bibliography

[43] S. Chikazumi and C. Graham Jr, Physics of Ferromagnetism, vol. 94. Oxford University
Press, 2009.
[44] P. Weiss, Hypothesis of the molecular field and ferromagnetic properties, J. Phys,
vol. 6, no. 4, pp. 661690, 1907.
[45] C. Hamann, Magnetische Hybridschichten-magnetische Eigenschaften lokal aus-
tauschgekoppelter NiFe-IrMn-Schichten. PhD thesis, Dresden, Techn. Univ., Diss., 2010,
2010.
[46] L. D. Landau and E. Lifshitz, On the theory of the dispersion of magnetic permeability
in ferromagnetic bodies, Phys. Z. Sowjetunion, vol. 8, no. 153, pp. 101114, 1935.
[47] H. Williams, R. Bozorth, and W. Shockley, Magnetic domain patterns on single crystals
of silicon iron, Physical Review, vol. 75, no. 1, p. 155, 1949.
[48] E. Della Torre, Magnetic hysteresis. IEEE press Piscataway, NJ, 1999.
[49] S. Schubert, Stochastic and Temperature-related Aspects of the Preisach Model of Hys-
teresis. PhD thesis, Chemnitz University of Technology, 2011.
[50] M. Kaltenbacher, Numerical simulation of mechatronic sensors and actuators. Springer
Verlag, 2007.
[51] S. Bi, A. Sutor, R. Lerch, and Y. Xiao, An efficient inverted hysteresis model with mod-
ified switch operator and differentiable weight function, IEEE Trans. Magn., vol. 49,
no. 7, pp. 31753178, 2013.
[52] S. Bi, A. Sutor, R. Lerch, and Y. Xiao, An efficient inverted hysteresis model with
modified switch operator and differentiable weight function, in Proc. 12th Joint MMM-
INTERMAG Conference, (Chicago, USA), p. 125, Jan 2013.
[53] S. Bi, A. Sutor, R. Lerch, and Y. Xiao, An efficient inverted hysteresis model with
modified switch operator and differentiable weight function, in Proc. IEEE Magnetics
Summer School, (Assisi, Italy), p. 24, Jun 2013.
[54] M. Kozek and B. Gross, Identification and inversion of magnetic hysteresis for sinu-
soidal magnetization, International Journal of Online Engineering (iJOE), vol. 1, no. 1,
2005.
[55] A. Sutor, S. J. Rupitsch, S. Bi, and R. Lerch, A modified preisach hysteresis operator for
the modeling of temperature dependent magnetic material behavior, Journal of Applied
Physics, vol. 109, no. 7, p. 07D338, 2011.
[56] S. Bi, A. Sutor, and J. Yue, Optimization of a measurement system for the hysteretic
characterization of high permeable materials, pp. 2230, 2008.
[57] H. Kettler, Measurement setups for the hysteretic characterization of high permeable
materials, Masters thesis, University of Erlangen Nuernberg, 2007.

102
 Bibliography

[58] G. Hornyak, H. Tibbals, and J. Dutta, Introduction to nanoscience and nanotechnology.

CRC Press, 2009.
[59] R. Lerch, Elektrische Messtechnik: Analoge, Digitale Und Computergestutzte Verfahren.
Springer DE, 2012.
[60] V. Iordache, E. Hug, and N. Buiron, Magnetic behaviour versus tensile deformation
mechanisms in a non-oriented fe(3 wt.%) si steel, Materials Science and Engineering:
A, vol. 359, no. 1, pp. 6274, 2003.
[61] S. Bi, A. Sutor, R. Lerch, and Y. Xiao, A measurement setup for acquiring the local
magnetic properties of plastically deformed soft magnetic materials, in Proc. IEEE 55th
MMM Conference, (Atlanta, USA), p. 176, Nov 2010.
[62] L. Piotrowski, B. Augustyniak, M. Chmielewski, F. Landgraf, and M. Sablik, Impact of
plastic deformation on magnetoacoustic properties of Fe-2% Si alloy, NDT & E Inter-
national, vol. 42, no. 2, pp. 9296, 2009.
[63] S. Wu, A. Sun, F. Zhai, J. Wang, Q. Zhang, W. Xu, P. Logan, and A. A. Volinsky,
Annealing effects on magnetic properties of silicone-coated iron-based soft magnetic
composites, Journal of Magnetism and Magnetic Materials, vol. 324, no. 5, pp. 818
822, 2012.
[64] R. P. Hunt, Magnetoresistive head, Feb. 3 1970. US Patent 3,493,694.
[65] H. Nalwa, Handbook of Thin Films: Deposition and Processing. Elsevier Science &
Technology Books, 2001.
[66] S. Hoon, An inexpensive, sensitive vibrating sample magnetometer, European Journal
of Physics, vol. 4, no. 2, p. 61, 1983.
[67] S. Bi, F. Wolf, R. Lerch, and A. Sutor, An inverted preisach model with analytical
weight function and its numerical discrete formulation, IEEE Trans. Magn., vol. 50,
no. 11, 2014.
[68] S. Bi, F. Wolf, R. Lerch, and A. Sutor, An inverted preisach model with analytical weight
function and its numerical discrete formulation, in Proc. IEEE INTERMAG Conference
2014, (Dresden, Germany), pp. 14621463, May 2014.
[69] A. Sutor, S. J. Rupitsch, S. Bi, and R. Lerch, A modified preisach hysteresis operator
for the modeling of temperature dependent magnetic material behavior, in Proc. IEEE
55th MMM Conference, (Atlanta, USA), p. 170, Nov 2010.
[70] A. Sutor, J. Kallwies, and R. Lerch, Vektorielles messen und modellieren von magnetis-
cher hysterese, Issues, vol. 79, p. 0, 2012.
[71] A. Sutor, S. Bi, and R. Lerch, Validation of the rotational vector preisach model with
measurements and simulations of vectorial minor loops, Applied Physics A, vol. 112,
no. 2, pp. 269273, 2013.

103
Bibliography

[72] A. Sutor, S. Bi, and R. Lerch, Validation of the rotational vector preisach model with
measurements and simulations of vectorial minor loops, in Proc. 12th Joint MMM-
INTERMAG Conference, (Chicago, USA), p. 140, Jan 2013.

[73] A. Sutor, S. Bi, and R. Lerch, Identification and verification of a preisach based vector
model for ferromagnetic materials, Applied Physics A, 2014.

[74] A. Sutor, S. Bi, and R. Lerch, Identification and verification of a preisach based vec-
tor model for ferromagnetic materials, in Proc. IEEE INTERMAG Conference 2014,
(Dresden, Germany), pp. 22542256, May 2014.

[75] A. Sutor, S. Bi, and R. Lerch, Identification and verification of a preisach based vector
model for ferromagnetic materials, in Proc. ITG-Fachbericht Sensoren und Messsys-
teme 2014, (Nurnberg, Germany), Jun 2014.

104
List of Figures

1.1 Comparisons of the material capabilities approximated in calculation . . . . . 2

1.2 Experimental data for identification . . . . . . . . . . . . . . . . . . . . . . 5

2.1 Classification of magnetism . . . . . . . . . . . . . . . . . . . . . . . . . . . 9

2.2 Diagram of magnetic domains with minimization of magnetostatic energy . . 11
2.3 Domain arrangements for various status of magnetization . . . . . . . . . . . 12
2.4 Typical major hysteresis loop . . . . . . . . . . . . . . . . . . . . . . . . . . 14
2.5 Minor loops and first order reversal curves . . . . . . . . . . . . . . . . . . . 15
2.6 Major hysteresis loops of Soft, semi-hard and hard magnetic materials . . . . 15
2.7 Magnetization curves of single crystal of iron . . . . . . . . . . . . . . . . . 16
2.8 Rectangular hysteresis loop of each Preisach particle and diagram of hysteresis
switch operator . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
2.9 Schematic interpretation of the Preisach model . . . . . . . . . . . . . . . . 19
2.10 Geometric interpretation of the Preisach model . . . . . . . . . . . . . . . . 20
2.11 Single-valued magnetization curve and obtained reluctivity curve . . . . . . . 23
2.12 Block diagram of including hysteresis models into electromagnetic FE analysis 23

3.1 Distribution of inverted weight function in the Preisach plane . . . . . . . . . 26

3.2 Calculated weight function (, ) by applying the relay switch operator 27
3.3 Geometric interpretation of the trapezoid areas in a Preisach plane . . . . . . 28
3.4 Forward and inverted Preisach models . . . . . . . . . . . . . . . . . . . . . 31
3.5 Geometrical interpretation of in the Preisach plane . . . . . . . . . . . . 32
3.6 Geometric interpretation of the inverted Preisach model . . . . . . . . . . . . 33
3.7 Geometrical relation between L(t) and G(t) . . . . . . . . . . . . . . . . . . 34
3.8 Analysis of weight function with H(M) reversal curves . . . . . . . . . . . . 35
3.9 Calculated weight function (, ) by applying the modified switch operator
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
3.10 Normalized weight function examples . . . . . . . . . . . . . . . . . . . . . 37
3.11 Ascending branches of major loops and their derivatives . . . . . . . . . . . 38

4.1 25cm Epstein frame . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40

4.2 Overview of the automated measurement system with Epstein frame . . . . . 40
4.3 A family of hysteresis loops and the primary magnetization curve . . . . . . 42
4.4 Comparison of permeability curves at different frequencies (10Hz 400Hz) . 43
4.5 Comparison of core loss curves at different frequencies (10Hz 400Hz) . . . 44
4.6 Comparison of hysteresis curves at different frequencies (10Hz 400Hz) . . . 45
4.7 Comparison of core loss curves at different temperatures (- 40 C 140 C) . . 46

105
List of Figures

4.8 Comparison of hysteresis curves at different temperatures (- 40 C 140 C) . 47

4.9 Surface images of GO steel M330-35A in RD . . . . . . . . . . . . . . . . . 48
4.10 Measurements of GO steels 35TW400 and 35Z155 at 50Hz at room temperature 48
4.11 Hysteresis loops of GO steels 35TW400 and 35Z155 measured at 50Hz at
room temperature . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
4.12 Overview of the double C-yoke measurement system . . . . . . . . . . . . . 50
4.13 Cold plastic training on electrical steel M33035A samples . . . . . . . . . . 51
4.14 Permeability curves of steel sample M33035A at 1Hz at room temperature . . 52
4.15 Permeability curves of TD steel sample M33035A after HT 800 C at 25Hz at
room temperature . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
4.16 Diagram of the vibrating sample magnetometer measurement system . . . . . 53
4.17 Overview of the VSM setup constructed in laboratory . . . . . . . . . . . . . 54
4.18 Sample for VSM measurement . . . . . . . . . . . . . . . . . . . . . . . . . 55
4.19 Scalar characterizations of 1.9 m Fe49 Co49 V2 thin film . . . . . . . . . . . . 56
4.20 Longitudinal coils (LC) and transversal coils (TC) for vector characterization 57
4.21 Parallel and perpendicular components of magnetic moment after calibration 57
4.22 Major hysteresis loops measured both parallel and perpendicular to the applied
fields . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
4.23 Primary magnetization curves measured both perpendicular and parallel to the
applied field . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
4.24 VVSM Characterization of anisotropic sample . . . . . . . . . . . . . . . . . 59
4.25 Vector characterizations of an isotropic sample in rotational field with circular
trajectory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
4.26 Vector characterizations of an isotropic sample in rotational field with ellipse
trajectory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
4.27 Vector characterizations of an anisotropic sample in rotational field with cir-
cular trajectory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62

5.1 Identification of the pseudo-inverse of the Preisach model . . . . . . . . . . . 64

5.2 Calculated inverted Everett matrix . . . . . . . . . . . . . . . . . . . . . . . 65
5.3 M(H) loops and the normalized H(M) loops with and without the forced mag-
netization effect . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
5.4 Graphic relation between fsimu and (e).
b Resulted analytical tangent function
g(e) according to different factor a . . . . . . . . . . . . . . . . . . . . . . . 67
5.5 Square mesh for discretization of the Preisach plane and the approximated
(, ) surface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
5.6 Simulated major loops of electrical steels M330-35A and M1300-100A . . . 73
5.7 Simulated major loops of Fe49 Co49 V2 and Ni81 Fe19 thin film samples . . . . 74
5.8 Simulated major loops of oriented electrical steels 35Z155 in rolling (RD) and
transverse directions (TD) . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
5.9 Relative errors of the simulations performed with different discretization size n 76
5.10 Simulation results of FORCs . . . . . . . . . . . . . . . . . . . . . . . . . . 77
5.11 Simulation results of minor loops . . . . . . . . . . . . . . . . . . . . . . . . 78
5.12 Simulation results of magnetizing hysteresis loops . . . . . . . . . . . . . . . 79

106
 List of Figures

5.13 Simulation results of FORCs with 40 reversal curves . . . . . . . . . . . . . 81

5.14 Simulation results of FORCs with 60 reversal curves . . . . . . . . . . . . . 82

6.1 Geometric interpretation of the rectangular areas in the S+ (t) subset of the
inverted Preisach plane . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
6.2 Geometric interpretation of the area S( p , q ) in the inverted Preisach plane . 85
6.3 Calculation of F p ,q from i , j . . . . . . . . . . . . . . . . . . . . . . . . 87
6.4 Weight function matrix V and the calculated F matrix . . . . . . . . . . . . . 88
6.5 Updating the vertices in staircase interface L(t) according to the input extrema. 89
6.6 Efficiency comparison of the three algorithms . . . . . . . . . . . . . . . . . 90

7.1 Comparison of the forward and inverted major loops. Comparison of the for-
ward and inverted elementary rectangular loops . . . . . . . . . . . . . . . . 94

107
List of Tables
2.1 Curie Temperatures for Some Ferromagnetic and Ferrimagnetic Materials . . 10

3.1 Integrals of (, ) in regions S1 -S6 in Fig. 3.8 . . . . . . . . . . . . . . . . 36

4.1 Classification of Non-oriented Electrical Steel . . . . . . . . . . . . . . . . . 41

4.2 Results of Measurements with Epstein Frame . . . . . . . . . . . . . . . . . 41

5.1 Identified parameters for weight function . . . . . . . . . . . . . . . . . . . . 72

5.2 Comparison of identification methods . . . . . . . . . . . . . . . . . . . . . 80

109
List of Publications
Publications
Journals
1. S. Bi, F. Wolf, R. Lerch and A. Sutor, An inverted Preisach model with analytical
weight function and its numerical discrete formulation, IEEE Transaction on Magnet-
ics, 50(11), 2014.

2. S. Bi, A. Sutor, R. Lerch and Y. Xiao, An efficient inverted hysteresis model with mod-
ified switch operator and differentiable weight function, IEEE Transaction on Magnet-
ics, 49(7), 2013.

3. S. Bi, A. Sutor, R. Lerch and Y. Xiao, A measurement setup for acquiring the local
magnetic properties of plastically deformed soft magnetic materials, J. Appl. Phys.,
109(7), E306, 2011.

4. A. Sutor, S. Bi and R. Lerch, Identification and verification of a Preisach based vector

model for ferromagnetic materials, Applied Physics A, 2014.

5. A. Sutor, S. Bi and R. Lerch, Validation of the rotational vector Preisach model with
measurements and simulations of vectorial minor loops, Applied Physics A, 112(2),
pages 269-273, 2013.

6. A. Sutor, S. Rupitsch, S. Bi and R. Lerch, A modified Preisach hysteresis operator for

the modeling of temperature dependent magnetic material behavior, J. Appl. Phys.,
109(7), D338, 2011.

Proceedings
7. S. Bi, F. Wolf, R. Lerch and A. Sutor, An inverted Preisach model with analytical
weight function and its numerical discrete formulation, In Proceedings of IEEE IN-
TERMAG Conference, pages 1462-1463, 2014. Dresden, 04.-08.05.2014.

8. S. Bi, A. Sutor, R. Lerch and Y. Xiao, An efficient inverted hysteresis model with
modified switch operator and differentiable weight function, In Proceedings of IEEE
Magnetics Summer School, page 24, 2013. Assisi, 9.-14.06.2013.

111
List of Publications

9. S. Bi, A. Sutor, R. Lerch and Y. Xiao, An efficient inverted hysteresis model with mod-
ified switch operator and differentiable weight function, In Proceedings of 12th Joint
MMM-INTERMAG Conference, page 125, CU-07, 2013. Chicago, 14.-18.01.2013.
10. S. Bi, A. Sutor, R. Lerch and Y. Xiao, A measurement setup for acquiring the local
magnetic properties of plastically deformed soft magnetic materials, page 176, ER-12,
In Proceedings of 55th MMM Conference, page 2010. Atlanta, 14.-18.11.2010.
11. A. Sutor, S. Bi and R. Lerch, Identification and verification of a Preisach based vector
model for ferromagnetic materials, In Proceedings of IEEE INTERMAG Conference,
pages 2254-2256, 2014. Dresden, 04.-08.05.2014.
12. A. Sutor, S. Bi, R. Lerch. Identification and Verification of a Vector Model for Ferroic
Materials, In ITG-Fachbericht Sensoren und Messsysteme 2014, 2014. 17. ITG / GMA
Fachtagung Sensoren und Messsysteme, Nurnberg, 03.-04.06.2014.
13. A. Sutor, S. Bi and R. Lerch, Validation of the rotational vector Preisach model with
measurements and simulations of vectorial minor loops, In Proceedings of 12th Joint
MMM-INTERMAG Conference, page 140, DH-08, 2013. Chicago, 14.-18.01.2013.
14. A. Sutor, S. Rupitsch, S. Bi and R. Lerch, A modified Preisach hysteresis operator for
the modeling of temperature dependent magnetic material behavior, In Proceedings of
55th MMM Conference, page 170, EH-04, 2010. Atlanta, 14.-18.11.2010.
15. S. Bi, A. Sutor, R. Lerch and J. Yue, Optimization of a measurement system for the
hysteretic characterization of high permeable materials, pages 22-30, Huethig Verlag,
Heidelberg, 2008. National Instruments VIP 2008.

Presentations
1. 2014 May, IEEE INTERMAG Conference Dresden (poster), An inverted Preisach
model with analytical weight function and its numerical discrete formulation.
Nominated for the best poster prize.
2. 2013 January, 12th Joint MMM/Intermag Conference Chicago (poster), An efficient in-
verted hysteresis model with modified switch operator and derivative weight function.
The IEEE Magnetic Society Student Travel Grant.
3. 2013 June, IEEE Magnetics Society Summer School Assisi (poster), A new inverted
hysteresis model with modified switch operator and differentiable weight function.
4. 2010 November, 55th MMM Conference Atlanta (poster), A measurement setup for
acquiring the local magnetic properties of plastically deformed soft magnetic materials.
5. 2008 October, National Instruments VIP 2008 Conference Munich (speech), Optimiza-
tion of a measurement system for the hysteretic characterization of high permeable ma-
terials.

112