Download as pdf or txt
Download as pdf or txt
You are on page 1of 5

M.

Kuczmann / Journal of Advanced Research in Physics 1(1), 011003 (2010) 1

Dynamic Preisach hysteresis model

Miklós Kuczmann*

Laboratory of Electromagnetic Field, Department of Telecommunication, “Széchenyi István”


University, Egyetem tér 1. H-9026 Győr, Hungary

 parameter of soft magnetic materials, and it must be


Abstract — The paper presents a scalar hysteresis analyzed while developing or designing an electrical device.
measurement system which can be used to prepare measured Rate-independent models in FEM can not predict losses
data for identifying the static Preisach hysteresis model as well accurately at higher frequencies. However Maxwell’s
as the dynamic one. The model identification is based on the equations contain the term of eddy currents and their
Everett function and the measured concentric minor loops. A
appropriate losses. These losses are smaller than the
fast and memory efficient implementation of the model will be
shown, which is very useful in numerical field analysis
measured ones. This is the reason why frequency
techniques. The dynamic extension of the static model is based dependence must be taken into account in hysteresis models
on the loss separation formula, and the measured excess applied in computer aided design applications.
magnetic field has been approximated by a feed-forward type The absent losses can be taken into consideration
neural network. The developed model has been inserted into a according to the loss separation theory [6]. An extra
finite element procedure to simulate a specimen with toroidal magnetic field strength term can be added to the one
shape. The nonlinear problem, based on the magnetic vector simulated by the rate-independent model, and the loop area,
potential, has been solved iteratively by the fixed point method. i.e. the loss can be increased as necessary [8]. According to
The comparisons between measured and simulated results
experiments, this extra magnetic field strength is depending
show the applicability of the developed model.
on the magnetic flux density B and its time variation
Keywords — Finite element methods, Magnetic hysteresis, dB / dt . This complicated term, H ( B, dB / dt ) , can be
Magnetic losses, Neural networks. described efficiently by a feed-forward type neural network
[5, 9] with two inputs and one output, resulting a simple
I. INTRODUCTION rate-dependent model.
The paper presents a measurement system, containing a
From electrical engineering point of view, the aim of toroidal shape core, which can be used to measure static and
measuring the hysteresis characteristics of ferromagnetic dynamic scalar hysteresis loops, and a scalar Preisach model
materials is to perform appropriate input data for numerical to predict these curves as well as the losses. The developed
implementation and identification of hysteresis models. The model has been inserted into a two dimensional FEM
realized models can be inserted into the electromagnetic procedure to analyze the measurement system.
field simulation procedures based on the Finite Element
Method (FEM). The electrical devices (e.g. motors and II. MEASURING HYSTERESIS LOOPS
transformers) may have various shapes which can be
modeled efficiently by FEM, but the constitutive relation The block diagram of the measurement set up can be seen
between the magnetic field strength H and the in Fig. 1 [1]. The magnetic field strength inside the analyzed
magnetization M or the magnetic flux density B must be toroidal shape core has been generated by the current i (t )
described accurately by hysteresis models. The connection flowing in the primary coil with N p  170 turns. The
between FEM procedures and hysteresis models can be
realized by iterative techniques. One of them is the so-called arbitrary waveform of current has been built up by applying
fixed point scheme [1-3]. some LabVIEW functions. The time variation of the
The rate-independent hysteresis models, such as the magnetic field strength results in a time varying magnetic
original scalar Preisach model, can simulate B-H flux    (t ) inside the specimen, which can be measured
relationships at very low frequencies [1, 2]. It is known by the secondary coil wound on the core. The secondary
from experiments that the enclosed area of the hysteresis coil has N s  182 turns, output of which is the induced
loop, which is the energy loss per cycle, is increasing voltage u (t ) .
rapidly by the increase of the frequency of the exciting field The magnetic field strength can be measured by the well
[4-8]. The energy loss is a very important technical known formulation
Manuscript received June 30, 2009.
*
Corresponding author (miklos.kuczmann@yahoo.com)
2 M. Kuczmann / Journal of Advanced Research in Physics 1(1), 011003 (2010)

output of the model can be calculated by a sum as [1, 2, 10]

Fig. 1. The block diagram of the measurement system.

N p i(t )
H (t )  , (1)
2 r
where r is the mean radius of the toroidal shape core
( 2r  0.176 m ). The magnetic flux density can be
calculated by integrating the measured induced voltage, i.e.
t

 u( ) d  B
1
B (t )  0 . (2)
NsS
0

Here B0 is an integration constant, and S  2  105 m 2 is Fig. 2. The measured concentric minor loops and the Everett function.
the cross section area of the specimen.
A Fourier transform based digital filtering technique, and K

a feedback proportional controller has been used to y t    E  0 ,  0   2  E  , 


k 1
k k 1   E  k ,  k  , (4)
eliminate measurement noise, and to generate sinusoidal
magnetic flux inside the specimen, respectively [1]. The where  k and  k are the increasing and the decreasing
waveform generation, the measurements and post sequences of the normalized magnetic field strength (or the
processing of measured signals have been worked out by magnetic flux density) applied to the sample and stored by
using LabVIEW functions. Current generation and using the staircase line Lt  . Moreover  0 and  0 denote
measurements are performed through the National the first turning point in the first magnetization curve and K
Instruments Data Acquisition Card, NI PCI-6251, installed is the number of stairs in the staircase line.
on the PC. Another practical advantage of the Everett function is the
easy measurement of it [1, 10]. The Everett function is
usually obtained from the measured first order reversal
III. PREISACH MODEL OF STATIC HYSTERESIS curves, however the concentric minor loops are applied
here. The Everett function can be calculated as
In the followings, xt  and y t  denote the input and the 
E  ,    y  y / 2 ,  (5)
output of the model. If xt   H t  and y t   Bt  , then the where y is the reversal output value of the hysteresis
model is called direct model, and it is referred to as inverse operator on the first magnetization curve, corresponding to
model in the reverse fashion. Here the inverse one is used. the input, value equal to  . The point  , y  defines the
The output of the Preisach model can be calculated by a peak value of the input as well as the output. The output
double integral defined over the Preisach triangle [1, 2],
y is according to the input x   on the concentric
y t       ,    ,  , xt  d d . (3) minor loop, starting from the point  , y  . The Everett

where   ,   is the distribution function determined from function has been approximated by a tensor product
measurements, and   ,  , xt  is the characteristics of one representation of piecewise cubic Hermite interpolating
splines [11]. The measured concentric minor loops and the
elementary hysteron. The variables  and  are called corresponding Everett function can be seen in Fig. 2.
switching fields of the rectangular hysteron. The aim is to develop a fast Preisach model because of its
The implementation of (3) results in a time consuming application in numerical field analysis. If the model is
algorithm because of the evaluation of the double integral. inserted into an FEM code, than one model is assigned to
A much faster method can be proposed using the Everett one finite element or to one Gaussian integration point of
function E  ,   , since it is the double integral of the the finite element. It can result in a huge number of Preisach
distribution function. According to this definition, the models to run simultaneously. The representation of the
M. Kuczmann / Journal of Advanced Research in Physics 1(1), 011003 (2010) 3

staircase line is a key issue of memory requirements and


speed. Only the value of  of the turning points are stored,
and an appropriate sequence of them is used in (4).
Let us suppose that the input is increasing to the value of
x1 , then it is decreasing (the illustration is plotted in Fig. 3).
This generates the turning point 1. The input signal is
decreasing till the value x2 , then it is increased and then
decreasing until x3 and x4 , respectively. This input
sequence is resulting in the turning points 2, 3, and 4. This
generates the sequence L   2 ,  4 ,  3 , 1 . If the input is
decreasing further from point 4 and x  x2 , then the points
2 and 3 must be erased from the sequence, but  4 is not a
turning point, of course. However, if the input is increasing
Fig. 4. Comparison between measured and simulated curves of a C19
from point 4 and x  x3 , then points 3 and 4 must be wiped stainless steel.
out from the staircase line. If the jump of input results in
x  x1 in one step, then all the four points must be cleared. IV. FREQUENCY DEPENDENT MODEL OF HYSTERESIS

Finally, it can be concluded that even number of staircase The investigated frequency dependent scalar Preisach
points must be cleared while a wiping out process. If the model is based on the theory that the total loss under
input is increasing, the number of staircase points is an even sinusoidal magnetization can be decomposed into three
number, and if the input is decreasing then the number of terms [5-8]: W total  W hysteresis  Wclassical  Wexcess , where
staircase points is an odd number. The sum in (4) can be
calculated in a fast and efficient way, moreover this W hysteresis , Wclassical , and Wexcess are the hysteresis loss, the
algorithm can be realized in a parallel mode using the classical eddy current loss, and the excess loss. The
elementary functions of Matlab. hysteresis loss is the area of the hysteresis loop at very low
The static Preisach model can simulate the measured frequency and it is supposed to be independent of the
curves accurately, as it is illustrated in Fig. 4. frequency, the classical eddy current loss is generated by the
eddy currents simulated by Maxwell’s equations, assuming
perfectly homogeneous conducting material. The second
term is proportional to the frequency. According to the
experiments, the total loss is usually larger than the sum of
these two loss terms. This is why an excess loss term must
be added to the sum of the hysteresis loss and the eddy
current loss. The last term is proportional to the square root
of the frequency, and it is depending on microstructural
features of the material. It can be very difficult to simulate
the excess loss term in a general mathematical way. The
total loss can be written as [5-8]
T dB

Wtotal  H dB  H
B 0 dt
dt . (6)

If the magnetic flux density is supposed to be sinusoidal,


then the magnetic field strength can be decomposed into
three terms as H  H hysteresis  H classical  H excess [5-8].
Here H hysteresis is the magnetic field strength calculated by
the static inverse Preisach hysteresis model. The second
term is the classical eddy current field according to the
quasi-static Maxwell’s equations, and the last term is the
absent magnetic field strength, which can be identified from
measurements. In numerical field analysis with static curve,
the first two terms are available, but their sum is usually
Fig. 3. The representation of the staircase line. smaller than the measured one. The excess magnetic field is
used to eliminate this difference. The excess magnetic field
is depending on the magnetic flux density and the speed of
its variation, i.e. a function H excess  H B, dB / dt  must
be approximated and added to the sum of
H hysteresis  H classical .
4 M. Kuczmann / Journal of Advanced Research in Physics 1(1), 011003 (2010)

Fig. 5. The excess magnetic field term can be simulated by neural


networks.

The excess magnetic field term is a complicated function


as it is illustrated in Fig. 5. Here the excess magnetic field is
calculated from static and dynamic hysteresis curves
according to sinusoidal magnetic flux density with a peak
value of 1.4 T, and at the frequency of 1 Hz, 5 Hz, 10 Hz,
20 Hz, 50 Hz, 100 Hz, and 200 Hz.
The excess magnetic field is a function of the magnetic
flux density and the variation of the magnetic flux density,
which can be approximated efficiently by a feed-forward
type neural network with two inputs (B and dB/dt) and one
output H excess . The training sequence is a database with a
large number of measured data. It has been prepared by Fig. 6. The measured and simulated hysteresis curves at f=50Hz.
comparing measured static loops and measured dynamic
loops. The excess magnetic field term can be obtained from VI. CONCLUSION
these curves. The layers and the number of neurons in the
layers have been set after some trials. The neural network
which can approximate the excess magnetic field term A possible generalization of the static inverse Preisach
appropriately is a network with three layers containing ten model has been realized based on measurements. The absent
neurons in the first, ten in the second, and one in the output magnetic field term according to the excess loss has been
layer. approached by a neural network. It results in a general
solution, because the neural network can simply
V. SIMULATION OF THE TOROIDAL SHAPE CORE approximate measured data. The main advantage of this
generalization is that the direct hysteresis model can also be
The analyzed problem consists of a toroidal shape core, used, because the absent magnetic flux density term can be
which can be simulated as a two dimensional, axial calculated from measured data, and it can be simulated in
symmetry problem [1]. Here, the magnetic vector potential the same way.
with two components, Ar , z   Ar r , z er  Az r , z e z , has In the future, dynamic minor loops will be studied at
been used, from which the radial component of the magnetic different frequencies, and the applicability of this method in
flux density B r , z  can be calculated and the developed vector modeling.
inverse scalar hysteresis model can be applied,
 
H r , z   B B r , z  . The fixed point method has been ACKNOWLEDGMENT

used to solve the resulting system of nonlinear equations. This paper was supported by the János Bolyai Research
First, the static hysteresis model has been implemented in Scholarship of the Hungarian Academy of Sciences
the FEM procedure, but eddy currents have been taken into (BO/00064/06), by “Széchenyi István” University (15-
account. The simulated hysteresis curves are much thinner 3210-02), by the Hungarian Scientific Research Fund
than the measured ones. It is significantly true when (OTKA PD 73242), and by the Hungarian-Romanian
increasing the frequency. The simulation results can be Bilateral partnership (RO-46/2007)
improved by using the dynamic hysteresis model. Fig. 6
shows a comparison between measured and simulated REFERENCES
concentric minor loops. Taking the excess magnetic field [1] M. Kuczmann, A. Iványi, The Finite Element Method in Magnetics,
Budapest, Academic Press, 2008.
into account results in much better results. [2] E. D. Torre, Magnetic hystereis, IEEE Press, New York, 1999.
M. Kuczmann / Journal of Advanced Research in Physics 1(1), 011003 (2010) 5

[3] E. Dlala, Magnetodynamic vector hysteresis models for steel


laminations of rotating electrical machines, Ph.D. dissertation,
Helsinki University of Technology, 2008.
[4] J. Füzi, “Computationally efficient rate dependent hysteresis model,”
COMPEL, 18, 1999, pp. 44-54.
[5] D. Makaveev, L. Dupré, M. De Wulf, J. Melkebeek, “Dynamic
hysteresis modeling using feed-forward neural networks,” J. Magn.
Magn. Mater., 254-255, 2003, pp. 256-258.
[6] G. Bertotti, “General properties of power losses in soft ferromagnetic
materials,” IEEE Trans. on Magn., 24, 1988, pp. 621-630.
[7] E. Dlala, “Comparison of Models for Estimating Magnetic Core
Losses in Electrical Machines Using the Finite-Element Method,”
IEEE Trans. on Magn., 45, 2009, pp. 716-725.
[8] S. E. Zirka, Y. I. Moroz, P. Marketos, A. J. Moses, D. C. Jiles, T.
Matsuo, “Generalization of the classical method for calculating
dynamic hysteresis loops in grain oriented electrical steels,” IEEE
Trans. Magn., 44, 2008, pp. 2113-2126.
[9] M. Kuczmann, A. Iványi, “A new neural-network-based scalar
hysteresis model,” IEEE Trans. on Magn., 38, 2002, pp. 857-860.
[10] E. Dlala, J. Saitz, A. Arkkio, “Inverted and forward Preisach models
for numerical analysis of electromagnetic field problems,” IEEE
Trans. on Magn., 42, 2006, pp. 1963-1973.
[11] Fritsch, F. N. and R. E. Carlson, “Monotone Piecewise Cubic
Interpolation,” SIAM J. Numerical Analysis, 17, 1980, pp. 238-246.

You might also like