Sie sind auf Seite 1von 5

Math/ Compur. Modding, Vol. II. pp. 197-201, 1988 0895.7177/88 $3.00 + 0.

00
Printed in Great Britain Pergamon Press plc

THE EXTRACTION OF ENGINEERING PARAMETERS FROM MATHEMATICAL


MODELS OF ELECTROMAGNETIC DEVICES

D.A.Lowther

Electrical Engineering Department, McGill University,


3480 University Street, Montreal, Quebec, H3A 2A7, Canada

Abstract. Tne mam phases of a ftnite element based computer


simulation program include preprocessing, solving and
postprocesslng. Of these the postprocesslng presents many
problems since It IS, essentially, an open ended process. The
underlying prlnclples, however, may be easily expressed.
Software structures and databases for providing flexible
result evaluation are described.

Keywords. Electromagnetics, Finite elements, Post-processtng,


NumerIcal analysts.

INTRODUCTION performing the flnal, ‘result


evaluation’ or ‘postprocessing’ phase of
the simulation.
The development of finite element based
computer software, coupled with the
rapld decrease in the cost Of computing FROM MATHEMATICAL SOLUTIONS
systems, has provided the design TO ENGINEERING RESULTS
engineer with the capability Of
simulating complex electromagnetic
devices on a desktop workstatlon. The The solution phase of any computer based
design of a usable system requgres the simulation system sets up and solves a
merging of several dlscipllnes; namely, set of equations which describe the
numerical analysis, database management mathematical model by whtch the physical
and man-machlne Interface design. device 13 being represented. If a
finite element approach is being used to
model the device, this results in a set
The process of simulation may be of linear equations which relate the
expressed as a sequence of magnetic field variations over a set of
transformation processes. Ftrst, the small patches which cover the device, to
real devtce has t0 be mapped onto ( the excltatlon and boundary condltlons,
approximated by ) a device which can be (Sllvester, 1983). The resulttng field
described to the slmulatlon program. solution IS expressed as a set of nodal
Second, the approximate model 1.3 values. However, considerably more
converted into a mathematical information is avallable from the
description of the problem. Third, the solutton because the set of trial
mathematical problem IS solved. functions used for the solution are also
Finally, results Of englneerlng Known. The comblnatlon of these trial
relevance are extracted from the functions and the sizes and locations of
mathematical solution. 1 ypically, these the surface patches determines how
results may include Impedances, forces, ‘good’ the solutton 1s In terms of
energies, flux distributtons, etc. In approximating the phystcal device. If
fact, the list of results which may be the solution of the equations is exact
desired IS almost endless - the result (no error), then the ‘goodness of the
extractton process IS essentially open- results depends only on how well the
ended and is limited only by the physlcal model has been approximated.
mathematical basis used for solving the
problem and tne valldlty of the results
In terms of the orlginal approxlmatlons. The accuracy of the Information about
the orlginal device which may now be
obtained has two maJOr components: the
The intentcon of this paper 15 to fit-st 15 the accuracy of modelling, the
discuss the processes tnvolved and the second the accuracy of solution. Until
problems which may be encountered In the advent Of adaptive meshing
Proc. 6th Int. Conf. on Mathematical Modding

techniques (Pinchuk, 19851,


(Blddlecombe, 19861, modelling accuracy
magnetcc core
required experience both with the class
Of device to be simulated and with the

r
modelling technique. Solution accuracy,
on the other hand, is dependent on the
algorithm used and its physical
Implementation, and IS largely under the
control Of the software
designer/programmer.
wlndlng

A third error component in the derived


lnformatlon is that due to the
algorithms used to extract it from the
solution. This error can take many
forms. For example, the distribution of atr
a specified quantity over a part of the LIE
solution space may be affected by the
sampling points; if too few are used
large changes such as singularities may
be undectected. In a second example F& 1: Simple I= Cm Inductance.
consider the implementation Of an
algorithm which requu-es the subtraction
Of two calculated quantities, such as however, depends on the shape of the B-H
might be encountered in a force curve. Thus the two definitions can
calculation. Under many circumstances result in very different values, often
the subtraction might be quite differing by an order of magnitude! The
reasonable, in some it might lead to problem arises because inductance is, in
large numerical errors due to the fact, only correctly defined in a linear
subtraction Of similar magnitude system. The two values may be seen to
quantities, (McFee, 1987). correspond to two methods of measuring
the inductance in the laboratory, i.e.
either by a change in flux as the
In addltlon to the computational errors current 1s reversed ( flux lInKage
mentloned above, there are others which method 1, or by measuring the terminal
are ‘procedural’ rather than current and voltage and using the
‘computatlonal’ and mght or might not fundamental components to generate an
be consldered as errors. TypIcally these Impedance value ( equivalent to an
are encountered by asKing for a piece of energy based method ). These two
lnformatlon In an ambiguous manner and measurements can show a difference m
may well result from a designer’s value slmllar to that encountered In the
InexperIence with the simulation system calculations. A problem arises when the
ana the mathematical techniques used, or designer computes the value using one
a lack of Knowledge about the device approach and measures It wtth the other
being designed. - the result appears to be an error in
the calculation process but IS actually
a IacK of understandlng of the physics
AS an example, consider the simple involved.
Inductance shown In figure 1. This
device may be analysed ustng a first
order vector potenttal formulation, Often, the dlf f erence may be hidden
providing It can be described In terms because of the structure of the devices
of a carteslan coordinate system and IS analysed. For Instance, if an air gap
essentcally two-dlmensional In nature, is Introduced Into the magnetic circuit,
i.e. the depth dmenslon 1s consIderably It Will dominate the device performance
greater than the other two. In and most of the energy WIII be stored In
addltlon, the device is assumed to be the air, I.e. In a Wear material. If
constructed from material having a non- the two methods are applied for this
ltnear B-H characteristic. The problem they may only dffer by a few
derlvatlon of Inductance IS purely a percent, and this IS easily written Off
postprocessing operation. to ‘numerical error’. This situation
may hold for many designs over a period
Of time until a device 1s encountered in
Two, generally used, deflnitlons of which the gap no longer dominates.
inductance may be consldered to compute Suddenly, the program develops a
the parameter for the speclf kc device. peculiar and lnexpltcable ‘bug’!
The first Involves the concept of ‘flux
lInKage per unit current’, the second is
based on the total energy stored by the A general rule for avoldlng the apparent
device for a given current. In linear error 1s to consider the computer based
materials these two definitions produce system to be the equivalent of a
the same value; for a non-llnear laboratory based experlmental system,
maternal, flux ltnkage per unit current and to only simulate experiments which
IS Independent of the path by which the can be physically constructed I” the
flux IinKage was set up, i.e. the shape laboratory. Thts latter approach can
Of the B-H characteristic is clarlf y many of the errors which belong
unimportant. The energy stored, to the ‘procedural’ class.
Proc. 6th Int. Conf. on Mathematical Modelling

RESULT EVALUATION PROCESSES Thus the ‘expansion’ 15 actually a


‘contractlon’ data as would be
Since the solution to the field problem expected. ThiAn result naturally leads
1s usually obtalned in terms of either to the concept of Ulfferentlatlon as an
one of the field quantatIes directly, or error Increasing process; since fewer
a potential function from which the values are requlrecl to store the result,
fteld quanttties may be derived, the they are stored at fewer points over the
rules for the manlpulatlon of the data problem, I.e. they could be consldered
are well defined. Namely, they are to be stored at the centroids of the
those basic mathematical operattons elements. Hence the density of sampling
which are valid for scalar and vector of the field over each element has been
fields, i.e. those of vector and scalar reduced, leadlng to an increase in
calculus and artthmetic. In terms of error.
problems analysed by the use of flnlte
elements, almost all these operations
can be reduced to relatively simple The changes In the volume of data to be
matrix operations performed on an stored during any parttcular process
element by element basis, (Silvester, depend heavily on the algorithm used for
1978). the process. Essentially, many steps in
a process, or the matrix operations
involved, may be concatenated.
However, the successive appltcatlon of Referring to the above example, the
several operations over all the elements magnitude of flux density requtres a
In the problem can lead to a data translation from nodal storage to
explosion. As a simple example consider element storage, two directional
a two-dimensional, first-order dlfferentiattons ( matrix mult8plies 1,
triangular mesh contanning N nodal and a magnitude calculation. If all
points and E elements (N < E <= 2N). these steps are concatenated, I.e.
Assume that a solution has been obtained performed on one element before moving
in terms of the magnetic vector on to the next, only a small buffer is
potential. The data generated by the needed for intermediate results and
solver contains one value per node, each these are overwritten as each element 19
node being a part of several elements. constdered. An alternate approach is to
Thus tnere are N f loatcng point values design a system in which all
to be stored. If the flux denstty 1s intermedlate results are stored, I.e.
wanted, then a derivative should be all data 1s converted to element based,
ta)ien of the shape function over each then the differentiation IS performed
element in two orthogonal directlons. over every element, then the magnitude
Thus the result is two zero-order 1s computed over every element. The end
functions over each element, I.e. 2E result 19 the same, however the sequence
floating point values, requlrlng up to 4 Of steps differs - the loop over all the
times the storage required for the elements has been placed at the centre
original field! If the magnitude of the Of the process rather than around the
flux density, rather than the lndtvtdual basic set of steps for an element. The
components, IS now required, then the first approach minimises data storage
number of values drops to one but either dtscards, or never expllcttly
per
element, i.e. only E floating point computes, IntermedIate values Which may
values are required. The lmpltcation 1s be of interest, whilst the second
that the maJor Implementation problem is retains all intermedtate values and the
the management of widely varying voltimes components of flux density are avallable
Of data. for use In other processes. The
difference may seem small, but the
impllcatlons of the different structures
The Uifflculty expressed above is one are fundamental to the system.
frequently encountered in designing a
software pacKage. Of course, although
the data expansion 1s real enough and In the problem of inductance calculation
requires careful database design In described above, if the first approach
order to handle It, the example given IS used, the device designer is provided
seems to contradict the assumption that with a command for Inductance; the
a differentiation process snould command has built into It assumptions (
generate less data than It IS given. or ‘Knowledge’ ) about the calculation
The apparent expansion in the data taltes method, coordinate system, and the
place because a change occurs tn the potential function. If any one of these
form of data storage; from nodal to assumptions does not match the
element based. In the orIgInal form, speciflcatlons for Inductance
nodal storage was used and the baste calculation for a particular device, the
shape functions for each element were program appears to have a ‘bug’. The
Known implicitly, i.e. built into the solution IS to create another Inductance
differentiation algorithm. If, Instead, routine and this can ultimately lead to
the data had been stored on an element an explosion in code, with a separate
by element basis, the shape functions routine for every possiblllty.
could have been held explicitly and 3E
floating point values would have been
required. In this format, each element The alternate approach provides a set of
IS handled on Its own and no assumptions prtmitive operations which are dependent
are made about field contlnulty from only on the form of the approximating
element to element. polynomials and not on the coordinate
200 Proc. 6th Ini. Cmf. on Mathematicrrl Modelling

system or potential function. Thus a dtf f erent coordinate systems, different


procedure for inductance calculation can basis functions, different potentials,
be defined to match the particular etc., tends to create large programs.
device speclflcatlons. AS described The alternatlve 19 to produce many
above, this mlnimises code but maxlmcses special purpose systems, each targetted
data. to a speclf IC tasK.

Software structures In which the PROGRAMMING LANGUAGES


algorithms have ‘Knowledge of the data
properties are relatively conventional,
structures in which data propertles and The UnderlyIng mathemattcal modelllng
information are removed from the processes used, coupled with the
algorithm are those frequently physlcal equations describing the
encountered In expert system and devtce, provide all the necessary basis
artlficml mtelllgence software for the set of permlsstble operations on
(Saldanha, 1986). The latter are well the solution. The effect of the physics
Known for theIt- capability to Increase 1s to constraln the valid domaln of
their data storage requlrements as they operations. The mathematical operations
execute.
whtch may be performed are not InfInIte.
In the case of flnlte element models,
they are actually quite restrlctlve in
SIMPLICITY VERSUS FLEXIBILITY that simple shape functtons are used
over sample geometrlc entItles. For
example, dlfferentlatton IS stmply
The two software approaches can result defined In term3 of the polynomtal used
In very different postprocessing as a shape function, stmllarly for
systems; an algorlthmlc approach leads Integration; the more complex vector
to a system which 1s essentially simple functions such as curl, div and grad may
to use, basic functions such as flux be built up from these basic operations.
density magnitude, force, Impedance, All of the basic arithmetic functions
etc., may be Implemented as simple may be equally easily defined, provtded
routines whtch mlnimtse the data only that all the operands of any
expansion problem and 0ptlmle.e operation are constructed on the same
processing speed. The data driven bas!s functions.
approach results In systems which are
much harder to use but are conslderably
more flexible, (Lowther, 1963). For These functions form the basic
example, force calculation In an ‘InstructIon’ set Of a processing
algorlthmlc system might be Implemented device. The next step IS to construct
by way Of Maxwell stresses; this may functions which generate results of use
worK very well for many of the designs to a designer. These functions form the
analysed, but can have problems In next hogher level of ‘Instructions’,
certain speclf Ic devices and the user I.e. the ones the end user is IlKely to
has no way of altering the computatton want. Each addltional layer of
procedure. The data driven approach, In InstructIons is created by adding
contrast, may Know nothing about force assumptions about the physical system,
calaculatton but can allow a stress thus slmpllfylng the user interface ( as
computation to be bullt up from basic well as restrtcting Its capablllties ).
derivatives. In the latter scenario,
the user can alter the computation
method in those situations where errors The Interface to these basic
occur. In other words, the data driven InstructIons may be through a structured
system can provide a programming programmlng language allowlng the user
environment which allows a user to access not only to speclflc
choose the most appropriate calculation postprocesslng operations but also
for the problem at hand. provldtng program control constructs
such as looping and conditional
branching. Thls language then provides
This dilemma, simplicity Of use against a relatively simple method Of
flexibility Of use, 13. one that 13 manipulating complete field solutions.
constantly encountered in the design of
high level programming languages. The
problems Of trading speed and RESULT PRESENTATION
compactness for flexlblllty and
generality are not unique to result
evaluation, they are possibly at their NumerIcal results, other than single
most obvtous In hardware design where numbers, are often difficult to

the counterparts are special purpose and Interpret and part of the requirement

general purpose computers. for a post-processing system 13 the


effictent communication of Information.
Frequently, the most effective method IS
Functtons can always be added to the a graphlcal display, i.e. the
special purpose software to Increase Its presentation of flux dlstrlbuttons.

CapabIlIty. However, these tend to be losses, temperature dlstrlbutlons,

added in a non-structured way to the graphs of field values against position,


overall detriment of the original etc. As users gain experience with a
design. Thus expansion to include system, however, It should allow them an
Proc. 6th Int. Conf. on Mathematical Modelling 201

open environment. This means access to aslted. For a result to be meaningful,


the raw or semi-processed data in a form the process for deriving, and the
which can be used by other utilities. tmpllcatlons of, the operations must be
Generally, this can be handled by well understood. Too often thid )S
providing a sequential file output ignored and a process which generates
structure. good answers for one class of problems
may well generate very bad results for
another.
However, as wlttl the mathematical
processes, the form of results display
needed may well depend on the problem However, In any result extraction
betng analysed and the ultimate use of process based on a numerically generated
the data. Thus a system whtch can only solution, the user should be aware of
display equipotentlal plots may be the basic errors Involved in the
extremely useful for examtning field solution and postprocessing processes
distrtbutlons but cannot be used to and needs to taKe them Into account If
examine loss distrlbutlons. Agaln, as any reliance IS to be placed on answers
with the mathematical operattons, a obtalned.
limited set of dtsplay options can be
deflned based only on the properties of
the data, I.e. whether It has spatial ACKNOWLEDGMENTS
variations, whether It is one, two or
more dimenslonal, whether it 1s global (
for example Inductance ) or specific to The author would llxe to acknowledge the
a point ( flux denstty value 1, whether financial support of the Natural Science
the approxtmating functions are and Engineering Research Council of
continuous, or whether it 1s vector or Canada and lnfolytlca Corporation for
scalar. the work described In thts paper.

Agaln, this set of display primitives REFERENCES


can be embedded in a programming
language to allow the creation of
specific displays relevant to a Biddlecombe, C. S., J. Slmxin, and C. W.
particular devtce. Trowbrldge (1986). Error Analysts
In finite element models Of
electromagnetic fields. IEEE Trans
CONCLUSIONS on Magnetics, HAG-22
-1 8x13.
Lowther, D. A., and B. Forghani (1983).
InteractIve post-processing
Some of the problems involved in techniques for electromagnetic
creatfng a system capable Of field analysis. IEEE Trans g
transforming mathematical data into Magnetics, MAG-19
----_---I 2168-2170.
engineering informatlon have been McFee, S., D. A. Lowther (1987). Towards
considered. These problems arise from accurate and consistent force
two sources: first, the ‘open-endedness’ calculatton in finite element based
Of any result extraction process creates computatlonal magnetostatics. IEEE
difficulties in designtng a system which Trans
-- on Hagnetlcs, MAG-23,
IS both f lexlble and simple and second, accepted for publication.
the more capable the system, the easier Plnchux, A. R., and P. P. Sllvester
it is to create results which are either (1985). Error estimation for
non-physical, or do not simulate the automatic adaptive flnlte element
correct environment for the questions mesh generatlon. IEEE Trans 03
being asKed. Magnetics, MAG-21, 2251-2254.
Saldanha, C. M., and D. A. Lowther
(1986). Automating the design
Nonetheless, it IS possible to create a process for electromagnetic
variety of useful postprocessing systems devices. I!E& Computer Aided
_
and to structure them such that Engineering Journal, 3, 173-179.
computational power IS available to the Sllvester, P. P. (1978). Construction of
user if, and when, it 1s needed. triangular finite element universal
matrices. International Journal for
Numerical Methods E Engineering,
Once a suitable ‘toolKIt’ Of operations 12, 237-244.
is provided, together with a flexible Sflvester, P. P., and R. L. Ferrari
result display system, the problem of (1983). Finite elements for
result extraction becomes, largely, one electrical engineers, CambrIdge
Of understanding the questions being University Press, Cambridge, U.K.

Das könnte Ihnen auch gefallen