Experimental design theory and structural optimization :
design of a major-third church bell
Citation for published version (APA):
Schoofs, A. J. G., van Asperen, F., Maas, P. J. J., & Lehr, A. (1985). Experimental design theory and structural optimization : design of a major-third church bell. (DCT rapporten; Vol. 1985.043). Technische Hogeschool Eindhoven.
Document status and date: Published: 01/01/1985 Document Version:
Publisher’s PDF, also known as Version of Record (includes final page, issue and volume numbers) Please check the document version of this publication:
• A submitted manuscript is the version of the article upon submission and before peer-review. There can be important differences between the submitted version and the official published version of record. People interested in the research are advised to contact the author for the final version of the publication, or visit the DOI to the publisher's website.
• The final author version and the galley proof are versions of the publication after peer review.
• The final published version features the final layout of the paper including the volume, issue and page numbers.
Link to publication
General rights
Copyright and moral rights for the publications made accessible in the public portal are retained by the authors and/or other copyright owners and it is a condition of accessing publications that users recognise and abide by the legal requirements associated with these rights. • Users may download and print one copy of any publication from the public portal for the purpose of private study or research. • You may not further distribute the material or use it for any profit-making activity or commercial gain
• You may freely distribute the URL identifying the publication in the public portal.
If the publication is distributed under the terms of Article 25fa of the Dutch Copyright Act, indicated by the “Taverne” license above, please follow below link for the End User Agreement:
www.tue.nl/taverne Take down policy
If you believe that this document breaches copyright please contact us at: openaccess@tue.nl
providing details and we will investigate your claim.
WFW 85.043 -1-
EXPERIMENTAL DESIGN THEORY AND STXUCTURAL OPTIMIZATION; DESIGN O F A MAJOR-THIRD CHURCH BELL
THEORIE DER VERSUCHSPLANUNG UND OPTIMIERUNG VON KON- CTRUKTIONEN; ENTWURF EINET? GROSSE-TERZ GLOCKE
R. Schoof s
,
F. van Asperen, P. Maas, Eindhoven Universi-t y of Technology.
A . Lelir, Koninklijke Ej.-jsbouts, Bell Foundry, Asten, Netherlands.
ABSTRACT
After a short introduction, Section 2 describes the
i t e r a t i v e procedure f o r coniputer aided structural op-
tj mization. The general op%imj.zation problem i s defined
and the algorithm t o solve i t i s presented. The use of the f i n i t e element method, together w i t h calculation of s e n s i t i v i t i e s , i s discussed.
I n Section 3 the extended experimental design theory i s
treated. The extension concerns the application o f sen-
s i t i v i t i e s . The mentioned theory i s used t o derive re-
duced mathemtical niodels e f a structure.
Section 4 contains some ideas f o r the application o f
such matheniatical models and argunients are given for the
integrati on of the experimental desi gn theory and %he
i t e r a t i v e optimization procedure.
I n Section 5 t h e proposed method i s applied i n order t o
design the geometry o f a so-called niajor-third church
b e l l .
F i n a l l y some conclusions are given.
I. INTRODUCTION
A widely used method for automated structural optimiza-
tion consists o f the combination o f the f ini.te element method and non-linear prograiiiming. The f i n i t e element method is used because o f i t s modelling f a c i l i t i e s and i t s analysi s accuracy. Non-linear programmi n y is used t o
-2-
the objective function (to be minimized) and of the con-
straints with respect to the design variables and the structural behaviour. Generally optimization problems
are solved iteratively, oft.en using gradient search nie-
Experimental design techniques are well suit.ed to esta-
blish eff kient program f o r physical experiments and to
derive from the results of the experiments a mathemati-
cal description cif the physical phenomenon under consi- derakj on.
Experimental design proves to be also applicable to nu-
merical experimentation t especially in the case o f
structural optimization, where the "measurements" con-
sist of finite element analyses.
thodfi.
2 . THE ITERATIVE PROCEDURE IN STRUCTURAL OPTIMIZATION We first present the general optimization problem and the algorithm to solve it. Then we discuss the re-
commendation of the finite element method as an analy-
sing module and the relation between a finite element
model and the design variables.
Finally some relations are given for the calculation of
sensitivities o f the objective function and the con-
straints.
GENERAL OPTIMIZATION PROBLEM
Assumhg that the structure under consideration can be described uniquelly by a column matrix 5:
XT
= (xl,..-,x,f( 1 1
xiIi.=lI,*.,n: so-callecl design variables
w e can write the constrained optimization problem as follows.
Minimize: F($ objective function (23
-3- g . . ( x )
i
O j = l 1 . . . , m inequality constraints ( 3 ) J h k ( x ) = O k = l , . . . I l equality constraints ( 4 ) 1 x ii
x i xy i = l,
.
.. ,
n side constraints ( 5 )I n the optimization problem the design variables are va- ried i n order t o obtain the best possible structure by
minimizing the objective function. Those variables may
consist o f :
- rnechanj.cal and physical properties of the material.
- the topology, i . e . the pattern of connections i n the
-
the gecmetry (the shape 1 and cross-sectional dimen-The objective functi.oii P i s related t o the most impor-
tant property of the structure (the structural weight is
often used), or a weighted sum of several properties.
The inequality constraints g- express the limitations
J
w i t h respect t o the behaviour of the structure, such iis allowable stress, minimum b u c k l i n g load, lowest eigen- frequency, e t c .
The equality constraints hk c m consist of a s e t o f ex- p l i c i t relations between the desi yn variables (and they can be used then t o reduce the number of the design va- r i a b l e s j ; inostiy they consist of a s e t of iinplicit ana- l y s i s equations which are used t o calculate the beha- viour of the structure
the range of variation of tiie design variables. They can
be derived from various considerations such as functio-
n a l i t y * fabricitifin o r aesthetics ~
The objective function F and the constraints c j j and hk niay be linear or nonlinear functions of the desiyn va- ri ables .
Most optimization algorithms require that a n i n i t k t 1 s e t o € design variables,
zo,
i s specified. Starting fromt h i s i n i t i a l s e t , the design I s updated i t e r a t i v e l y . Probably the most common form of t h i s i t e r a t i v e proce- dure i.s given by
structure.
::ions o€ the structure.
-4-
where q i s the iteration number and C i s a vector search
direction i n the design space. The scalar quantity Q
defines the distance that we wish t o move i.n the direc- t i o n C . So the iteration s t e p ( 6 ) consists of two parts:
1. F i n d i n g a convenient search direction S. For that purpose the most e f f e c t i v e optimization algorithms
require gradients t o be calculated of the objective
function and of the constraints w i t h respect t o the
design variables.
*
2.
Findi.ng the scalar Q 50 that, moving i n the direc-t i o n S I the objective function i s miniitii.zed. A great
number o f matheniatical programming techniques i linear
and nonlinear, can be used for the execution of
it.erati.on step ( 6 )
.
Several numerical software Libra- r i e s provide many appropriate subroutines as b u i l d i n g blocks for optj.nii.zation methods. A wj-dly used method f o r the solution of structural optimization problemsi s the solution of a sequence of linear programming
problem [ S1,P-nieZ:hod) . I n this method the optimiza-
tion problem i s 1j.nearj.zed about the current solution
5"' w i t h i n a restricted area, specified by so-called "move limits" f o r the design variables.
The f i n i t e element method has the following d i s t i n c t ad- vantages
f o r numerical optimization.
*
as an analysing module i.n computer progranis
1.
2 .
A great v a r i e t y of structures can be modelled i n a
general and f l e x i b l e way due t o the great number o f
element types available for different theories i n ap-
plied mechanics; s o very r e a l i s t i c engineering models
The numerical accuracy of the analysis can be con-
t r o l l e d e a s i l y by an appropriate modification of the f i n i t e element model. I n optimization problems one is
often interested i n only s l i g h t improvements of a structure, c o a r e a l i s t i c model and accurate analysis
results are of great. importance.
C.ZQ he formulated and an;.i?ye- clJr,d.
A major problem i n the optimization procedure i s the l i n k between the s e t of design variables and the f i n i t e element model. The desj.gn variables are user-oriented whereas the f i n i t e element model i s computer oriented.
I n our opt.imj.zation program DYNOPT [ 1 , 2 ] for: insi:ance, t h i s link i s realized €or two-dimensional problems. On the other hand, i n general purpose f i n i t e element
packages the incorporation o€ the design variable con-
cept has just s t a r t e d with the most simple types of va- r i a b l e s , such as cross-sectional dimensions.
I n the f i e l d o f computer graphics the rapid development o f s o l i d modelling programs can become of great iinpor-
tance for the solution of this shortcoming, becau:ie the
i n p u t needed f o r an advanced s o l i d modelling prograiii i s much more user oriented t h a n a f i n i t e element model, and the l i n k between a s o l i d model and a f i n i t e element mo- del h a s already been made.
By s e n s i t i v i t i e s of a c e r t a i n function we understand i n
this paper the gradj-ents o f the function w i t h respect t o the design variables.
The s e n s i t i v i t i e s o f the objective function can usually
be calculated straight-forwardly by perturbation o f each
o f the design variables. W i t h
zo
as the current design variables .it follows t h a t :where &xi i s a smaii perturbation of
xp.
The s e n s i t i v i t i e s o f the c o n s t r a h t s usually require the intermediate calculation of the s e n s i t i v i t i e s o f d i s -
placements since most commonly constraints (i. e .
,
s t r e s s limits) are d i r e c t l y related t o the displace- ment s
.
We f i r s t consider a linear f i n i t e element problem w i t h sti3t.i.c loading. T h i s problem is described by the matrix equation
Kg = p ( 8 )
where K i s the s t i f f n e s s matri.x of the structure g i s a coluitin matrix o f unknown displacements
-6-
We d i f f e r e n t i a t e ( 8 ) w i t h respect t o design vari.able xi
Rearranging and multi.plyiizg by K-' g.i.ves
The s e n s i t i v i t i e s on the right-hand-side of ( I O ) can e a s i l y be calculated by the perturbation method (as i n
( 7 ) ) . Assuming that the matrix K was decomposed during a pievious analysis
,
K-' i s e € € e c t j vely available,
so thecalculation o f
h/axi.
isFor a linear e l a s t i c dynamically loaded struckure the s e n s i t i v i t i e s o f the eigenfrequencies w.. can be derived
J
i n f a c t , straightforward.
:i
where K is the stiffness matrix, M is the mass matrix
and y . 7 i s the eigenvector coupled w i t h the ei.genfre- qüeiicy w Again the sensj.ti.vi.ti.es of K and M can be calculated by a perturbation method. I f we assuine that the eigenfrequencies and eigenvectors are available from a previous analysis, then the evaluation of ( î l ) is again straightforward.
I f t h e l i n k between design variables ( i n a broad sense] and the f i n i t e e l f t n ~ n t n!odel i s incorpoiated ia f i i i i t e
element packages, it w i l l n o t be too diffj.cult t o calcu- l a t e the mentioned s e n s i t i v i t i e s as a standard o u t p u t option of the package. T h i s w i l l briny structural o p t i - mization by mans o f s e n s i t i v i t y analysis much closer t o
a great many of the engineering people.
j .
3. EXPERIMEMTAL DESIGN THEORY
The development of the experimental design theory (EDT)
- 7 -
evaluation of physical experiments and for the control
of production q u a l i t y [3j. We are using techniques based
on this theory i n together w i t h the i t e r a t i v e structural optimization procedure. T h i s application can be seen as.
the p l a n n i n g and evaluation of numerical experiments,
i,. e . f i n i t e element analyses. I n contravention of the
expeririiental design l i t e r a t u r e , we describe EDT i n ternis of structural optimization according t o the notation i n section
2.
Together w i t h t.he description o f the common theory, we w i l l discuss the extension we have made, namely the use of s e n s i t i v i t i e s i n EDT. Also some impllcatj.ons w i l l be discussed of the f a c t t h a t the numerical process i s de- terministic.
When a structure can be described by n design variables
3 according t o ( I ) , we may search f o r functions
y j = y j ( g ) j = l r . . . i m i n a certain limited area
( 1 2 )
The functions y : have t o do w i t h several properties of f
the structure such as weight
,
displacements o r stressesa t certain points. I n other words, they represent the
objective function and the constraints i n a structural
optiinization problem. I n the following we w i l l comider only OE5 prsperty ?' y and f û ï bxevity we ömit t h e index
j .
LINEAR NODEL
To f i n d the r e l a t i o n
Y = Y($ ( 1 4 3
- 8 -
i n which :
Bi, i = l c . . . f k ; the unknown parameters we are looking
f i , i = ? ,
. . . ,
k ; t h e known functions of the design varia-for, the model i s l h e a r i n the B i s .
bles; we can choose both linear and nonlinear func-
t i o n s f o r them.
e ; a variable which i s a stochastic error at physical
experiments, or a deterministic model error a t de-
terministi c f i n i t e eleinent analyses. EXPERIMENTAL DESIGN
The formulation of an experimental design (ED) implies
the iiiakiny of a programme of measurements. I n a nurtieri- c a l experimental design a "measurement" may consist of a f i n i t e element analy:.; is. More concrete the formulation
of an ED implj-es:
1 .
The choice of discrete values ( l e v e l s ) f o r a l i design2.
The choice of certain combinatims o f l e v e l s «€ clif- variables x i .f erent
xi
sexperiment o r Each one f i n i t e element analysis. combination leads
t o une physical
ESTIMATION OF THE PARAMETERS B,
I f we have a number of p combinations of levels i n the
ED, speci.fied by the p column matrices
Y "p
_- X' ,"I i " L f " ' Y -
and we analyse the structure a t these p s e t s of design
variable values, then according t o ( 1 5 ) we write
I n matrix notation
2 = X @ 3-
-9-
where: X i s a p * k inatrix consisting of the components: y is a column matxix w i t h p observations o f y
e
is a column matrix w i t h p errors@ i s a column inatri.x w i t h k unknown parameters.
Xt,i = f i ( x t ) , t=l, . . . , p ; i=I,...,k
Generally, the number o f observations p exceeds the num-
ber of parameters k and a least-syu+res technique should
be applied.. I n t h a t case estimates @ f o r the parameters @ are calculated i n such a way t h a t the residual sum of square.; E?, i s winimuin, w i t h
. L
3KS
Setting 3 = O for .i,-1 I . .
.
, k39-i
results i n a s e t o f k linear eyuatj-ons
from which the estimates p can be calculated i f t.he na-
T T
t r i x (X X) i s not singular. A regular matrix (X X ) can
be obtclined by a n appropriate formulation o f the ED. Assuniing (XTX) i s regular it follows from
(20)
thatS u b s t i t u t i n g ( 1 8 1 i n ( 2 1 1 gives :
The estimates
i
are unbiassed f o r a stochastic experi--10-
and consequently the expected value of $ j.5:
On the other hand the estimates @ are biassed i n the
case of deterininistic observations, such as t.he results
o f f i n i t e element analy:jes, since equation (23) does not; hold for model errors, see ( I ? ) . Nevertheless i t proves
t o be successful t o estimate @ in t h i s way, and t o achieve a linear model t o predict values f o r y y ( ? ) which areA su f f i c i e n t l y a c p r a t e
.
U:.;ing the estimated pa- rameters&
the estimate y f o r the function y cas be calculated from :THE USE O F SENSITIVITIES TO ESTIMATE @
Differentiation of the natiiematical model ( I 5 ) w i t h ïes- peck t o the design variable xpi gives
i n Section 2 we saw that for several s i t u a t i o n s sensiti- v i t i e s can be calculated r e l a t i v e l y easy, because the matrices needed f o r the calculation of the result: y can be used, see (10) and ( 1 1 ) . I n our experience the coinpu- tational e f f o r t for the calculation of the n s e n s i t i v i - t i e s of y i s only a sinall fraction of the e f f o r t for the calculation of the function value of y i t s e l f . Conside- r i n g t h i s , ( 2 6 ) can be used advantageously, together w i t h ( 1 5 ) , t o estimate the parameters fl. For t h a t pur- pose we extend the s e t of equations (17) w i t h t:he f o l -
-1 1- I n niatrix notation C o n b i n a t i o n of ( 1 8 ) and ( 2 8 ) gives De f i n i n g and
x*T
= [X X'] ii leas-i-squares solution OF*
f ( 3 0 ) y= x
@can be used t o e s t h a t e the parameters [3 as, see ( 2 1 3
i
= (X*T x * ) - l X f TY
*
( 3 1 )The f u n c t i o n s f i , ( x ) t h a t appear i n ( 1 5 ) are mostly o f a form l i k e
where exp may have any value i n the r a n g e j
expj = O,
1 , .
..
, m-12-
( 3 3 ) provided a l l functj.ons f i ( x ) are different.
Model ( 1 5 ) then be(:oIae:j a polynomial of the order m. Ge- nerally the highest exponent i s EOS: the came for a l l de- sign variables i n the inodel. We t r y t o attune these ex- ponents t o the expected behaviour o f the structure By
means of explorative calculations, The collection o f
terms i n the model w i t h only one and the same design va- r i a b l e i s c a l l e d a main e f f e c t . For instance
can be the main e f f e c t of x j up t o the order 3 . General- l y the model a l s o contains so-called i.n.teractions i n the form o f products o f two or more different design varia-
bles. The highest exponent of a certain design variable
i n the interaction terms is derived from the order: o f .the concerning main e f f e c t ; niostly we r e s t r i c t the model
t;o f i r s t and second order interaction terms, such as
The formu1at:i.m of the ED implies the choice of sets o f
di.screte values for the design variables w i t h which %he
structure w i l l be ana1y:ied. We proceed as follows. An estimation o f the range of interest i s made for the
individual design variables; experience and explorative
calculations a r e used for t h i s purpose. With each design variable the range i s divided i n one or more intervals,
m o s t l y of the same length, resulting i n a certain nuniber
o f l e v e l s of the design variable values. T h a t number has
%o be matched t o the order o f the concerning main ef- f e c t ; it: M i l l be clear that it i s important whether or not s e n s i t i v i t i e s s h a l l be used in the Ell. Appropriate numbers o f l e v e l s are a requirement t o obtain a regular
niatrix (X X
1
~I n io so-called complete ED a l l possible combinations o f
levels appear. Generally this results i.n a very large
-13-
nuìtiber of analyses t o be carried out. A so-called frac-
ber of analyses needed for a complete ER. A fractional
ED can be formulated I n such a way that it gives a mini- mum loss of accuracy w i t h respect t o t h e complete ED. For the case no sensj.tivj.ti.es are used i n the ED, tables for fractional ED'S can be found i n the literature [ 6 ] . tional ED only needs a 2 , 1 3, 1 q , . . . 1 fraction of the nurn-
REDUCTION OF TEE MODEL
A
After the parameters @ have been estimated by ( 3 1 ) we t r y t o reduce the number of paraineters i n the matheinati- cal model by considering the r e l a t i v e influence of the individual terms i n the inodel. For that purpose the va-
riance-covariance matrix Y(@) i s calculated, where
V j . j , i # j : co-variance between B i and p j
It can be shown that
: variance of 8 .
'ii 3.
where KS,: d u f :
residual sum of squares, see ( 1 9 )
Y;l;mb2ï û2 degrees cif freedom, which i s tile number of rows i n the 1natrj.x X rni.nus the i i a ~ b e r k of the estiïiiates
@ .
*
To reduce the mathematical model we range i t s terms w i t h respect t o the value of
I i = l t . . . s k ( 3 5 )
Pi
i. i
The term j.n the model a t whj.ch ( 3 5 ) proves t o be the *i-
niliium, may be expected as the least important: one. Sub- sequently we omit one o r inore of least important terms and make new estimates f o r the reduced number of pasaine-
ters [s. The reduction of the model i s controlled by a graph 02 the residual sun of squares RC, against the number k of parameters i n the model, see Figure 1.
- 1 4 -
I
I
IFiguïe 1. Residual sum of squares vexsus number of
parameters
e .
The procedure mentioned above for reduckion of the model
i s repeated until a rapid growth o f KS, occurs. Just be- fore this point the model has the lowest nuznber of
terms, giving the highest predictional, c a p a b i l i t y . I n terms of the theory o f experiniental design this reduced niodel is the b e s t one.
4. APCSICATICNS GF REDuLED Î;IOUELs : INTEGRATION WITH THE
ITERATIVE OPTIMIZATION PROCEDURE
The procedure described i n Section 3 i s suitable for the derivation of reduced inathemati cal models f o r the beha- viour o f structures. The rnodels, based on a sei: of f j
-
n i t e element analyses! can be used in a number o f ways:1. d i r e c t l y , as an accurate and easy t o evaluate niathe-
aiiatical model o f the structure. These iiiodels can be implemented on a microcomputer a t a design o f f i c e
.
For that purpose a l i b r a r y o f models can be b u i l t up.
The models can also be considered as systematic para-
meter studies.
2.
as a f a s t analysing module i n an optimization pro- gramme, operating according t o the i t e r a t i v e proce-dure as descrjbed i n Section
2.
The resulting optimum-15-
3 .
or can serve as a h i g h q u a l i t y starting point: for the i t e r a t i v e procedure i n which a f i n i t e element package i s the analysing module. Because a d i r e c t f i n i t e ele-
ment analysis will be more accurate t h a n reduced mo-
del evaluations, an improvement of the optinia1 struc-
tural design i s t o be expected.
as a fast analysing module i n a computer programme
for simulation of the behaviour of a structure I
The integration o f the i t e r a t i v e optiiflj.zation procedure w i t . h a numerical experiinental design proves t o be very
successful. The reasons are that both procedures have
much i n common. I n each procedure
- a considerable number o f f i n i t e element analyses o f s l i g h t l y modified structures has t o be carried out. - the concept o f design variables i s used; a l i n k be-
tween these variables and the f i n i t e element model. i s needed.
- the s e n s i t i v i t i e s of the structural behaviour have t o be calculated and can be used w i t h great advantage.
On the other hand both procedures complement each other
i n a rather nice way due t o :
- the explorative parameter s t u d y capability of the re- duced model by which a good starting point f o r the i t e r a t i v e procedure can be found.
- the p o s s i b i l i t y t o obtain h i g h q u a l i t y optimal solu- tions by means o f the i t e r a t i v e procedure based on d i - r e c t f i n i t e element: analyses.
-
the p o s s i b i l i t y t o carry out a fast optinlization run, i f a reduced model i.s used as analysing module i n tiie i t e r a t i v e procedure.5. SOME APPLICATIONS: DESIGN O F A MAJOR-THIRD BELL
I
We have derived several reduced niathematical models and
have applied them i n structural optimization problems
such as:
- the behaviour o f penhole j o i n t f o r the connection of a
-16-
- shape opt.intization o f a l i n k of a chain for a conti- - op t i m i zat i on o f the c L' o s s - s e c ti ona
1
d j. inen s j. ons o f a1
u -nuous variable transmission. minum beains .
recent and very nice application of numerical experi-
mental design concerns the design o f a :m-called major-
t h i r d b e l l , and w i l l be discussed i n d e t a i l [ 1 , 2 ] .
PROBLEM DEFINITION
Figure
2.
Church b e l l and i t s cross section. These b e l l s a r e characterized by a range o f higher liar- inoni c s forming a ininor-third chord. Especially the b e l l sof a c a r i l l o n a r e tuned very precisely i n order t o achieve a consonant sound; the t u n i n g i s carried out by s l i g h t modifications o f the inner contour of the b e l l by means of inetal cxt.ting processes. Froin a musical point o f view it would be interesting t o design bells which are characterized by a major-third chord i n t h e i r range
of higher harmonics. Figure 3 shows a qualitatAve pic-
ture o f the vibration modes of the b e l l i n cross-sec-
tions through, respectively perpendicular t o the rota- tional a x i s .
-17- 2 5 P 3 6 4
Figure 3 . Vibration riiodes of a b e l l .
Table
1
shows t h e r a t i o s of the most iniportant eigen- frequencies €or both a minor- and a major-third b e l l .The code indicates the relevant vibration mode as a com-
bination of an a x i a l and a transversal mode according t o Figure 3 . F o r p r a c t i c a l reasons the frequencies are nor- rtialized w i t h respect t o the octave. I n the practice of the b e l l foundry the frequency r a t i o are measured i n so- called cents, a measure which i s defined as
-18-
1
200
*LOC ( 4 f / f 0 )freq. rati.0 i n cents = LOC(2) ( 3 6 1 where f o : frequency of the octave
f : considered frequency 1 t F r e qu en cy rat i o 3cocf;ave / f -
1
[cents] 4.0000 2.0000 1 .E818 1 . 5 8 7 4 1.3348 1 .o000 O. 6674 0 1200 1500 1600 1900 2 400 3100 0.5000I
3 6 0 0Table I . Most important frequency ratias o f a b e l l . The reader is referred t o [ 4 , 5 ] for iiiore inforinatj.on
about the musi.cal aspects of b e l l s .
A minor-third and a major-third b e l l only d i f f e r i n t h e t h i r d toner as can be seen i n Table 1; the other fre-
quency r a t i o s have t o be just t h e same.
Experienced b e l l founders have tried t o design major- t h i r d b e l l s by means o f a t r i a l and error approach. I n s p i t e of considerable empirical e f f o r t needed for shape variations, they d j d not succeed.
We solved the problem as a structural shape optimization
problem by mearis of a mathematical inodel as described i n Section 3 . For that purpose the b e l l is modelled as an
axisymmetric structure subjected t o a non-axisymmetric
-19-
GEOMETRIC MODET, OF THE BELL
The geometry of the b e l l can be described i n a flexi.ble
way be nieans of so-called cubical B-splines
[2]
and i s shown i n Figure 4 .Figure 4 . Geometric definition of the b e l l .
The r, z-coordinates a r e specified for a certain number
of points P i . Subsequently a cubical B-spline function
is used t o f i t . a smooth curve, t h e so-called p r o f i l e
curve!, going through these points. An other coordinate
s, along the p r o f i l e curve, i s defined. Next the wall thickness f:. of the b e l l i s defined i n a number of
points on t i e p r o f i l e curve; these points need not be
pers6 t h e same point.; Pi f o r which r and z were speci- f i e d . The wall thickness i s taken perpendicular t o the p r o f i l e curve. In intermediate points of the curve the wall. thickness i s again defined by means of a cubical B-
s n l i n o ;= *I-.,. - , . - . - : 1 : - - -
coordinates of the points P i can be taken on constant values withoui: l o s s of generality of the geometric mo- d e l . The specified r-coordinates r i together w i t h the specified wall thickness t j
r i a b l e s i n a shape optimization paoblein.
LUll\rIrLwll 1 1 1 L . J i c L . U L V J . L L L I ~ ~ ~ L coordinate S . The Z-
b u i l d up a s e t of design vc1-
FINITE ELEMENT MESH : EXPLORATIVE CALCULATIONS
isTe analysed the b e l l using the optiuiization program DYNOPT [l]. I n DYNOPT the f i n i t e element method i s used
as an analysing inodule . W i t h this programme eiyenfre-
quencies and vibration modes can be calculated o f an
axisymnietric structure w i t h non-axisy1timetri.c deforma- ti on. For t.hi s purpose two appropriate f i n i t e elements
-20-
are implemented, naniely a 6-node and a 8-node iaopara-
metric ring element w i t h curved element edges.
I n DYNOPT the SLP-niethod, mentioned i n Section
2
i s used as an optimization iiiodule.The optimization problem can be specified i n the input: f j.le of DYNOPT by the design variables
and the objective function. I n addition i n the i n p u t f j2.e a combinati on of s e t s of design variable values
,
according t o (161, can be specified. T h i s i n p u t definesthe f i n i t e element analyses t o be carried out i n a nu-
merical experimental design as described i n Section 3 .
For shape o p t i mization purposes DSTNOPT incorporates the p o s s i b i l i k y t:o specify an user supplied subroutine as a l i n k between the design varj.afsles and the f i n i t e element model. For the b e l l S U C h a user supplied subroutine i s
based on the geometric ;Rodel described i n the previous
section [ 2 ] .
F i n a l l y i n DYNOPT sensitivltj.ec of the eigenfrequencies can be calculated.
We made some explorative calculations f o r the b e l l . It
has been proved t h a t the eiyenfreyuencies can be calcu- lated s u f f i c i e n t l y accurate brhen 15 8-node elements are
the constraints
Figure 5 . a .
Figure 5 . (a) Nesh and ( b ) f i r s t estimate f o r the geomtry of a major t h i r d b e l l .
With the explorative calculations we a l s o considered the s e n s i t i v i t i e s of the eigenfrequencies. Based on those s e n s i t i v i t i e s we made a f i r s t estimate a t the geoiiietry of a major-third b e l l , see Figure 5 . b . Notable i s the s l i g h t bump a t h a l f the height of the b e l l . The fre- quency r a t i o s of this b e l l are not close enough t o the
-21 -
desired values such that the major-third b e l l cannot yet be realized by tuning the b e l l .
NUMERICAL EXPERIMENTAL DESIGN
I n order t o proceed w i t h the search for the major t h i r d b e l l we formulated an experimental design about the geo- metry according t o Figure 5 . b . With t h a t design 5 radii
and
2
thicknesses were chosen as the design variables;they a r e indicated i n Figure 5 . b.
We varied a l l design variables on two l e v e l s , namely 5% higher and 5% lower than the nominal values. S e n s i t i v i t y values were a l s o used t o estimate the parameters, so a
mathematical model of order 3 could be chosen.
For all frequency r a t i o s , except the octave which i s the reference (see Table I ) , the same model containing 48 unknown parameters @ i s postulated. O f course the e s t i - mated parameters @ €or t.he individual r a t i o s are diffe- rent.
We used a complete experimental design f o r the 7 design variables varied on
2
l e v e l s , so2
exp7 = 128 f i n i t eelement analyses had t o be carried out. Subsequently the
parameters $ have been estiinated.
The resulting mathematical models proved t o be very ac-
curate and they are suitable t o predict the results o€
d i r e c t f i n i t e element analyses very c l o s e l y .
OPTJMIZATION USING THE DERIVED ~ ~ ~ H ~ M A T I C A L MODELS I n order t o f i n d the geometry o€ a major-third b e l l , a
zero-order optimization was carried out, using the de-
rived mathematical nioGeis. we proceeäeä as follows. For
each frequency r a t i o a s u f f i c i e n t l y narrow band about the desired value i s defined
<
ri (ideal)i
r!f i-1,...,
7“i.
-
( 3 7 )Tiie range of variation o€ each design varj.able i.:i de- vj.ded i n a rather large number of equl d i stant intervals. Thus a narrow spaced 7--diinensional g r i d i s specified on the design variable space. Subsequently the following
-23-
w i t h , cast a bel1 according the proposed geometry. T h i s b e l l , w i t h a greatest: diazrieter o f about
1
m.,
could betuned exactly t o the frequency rati os characterizing t h e major-third b e l l and it can be considered as the first; r e a l major-third bell i n the world.
6 . CONCLUSIONS
The experimental design theory, extetidec2 w j t h the use of
s e n s i t i v i t i e s t o estimate the model parameters
,
proves to be a successful t o o l far t h e planning and evaluationof numerical experiments. Especially the integrati.on, w i t h the i t e r a t i v e opti rnl zation procedure
,
provides a useful optimization method for ;il a r g e class o f structu-ral optimization problems.
The f i r s t r e a l major-third b e l l has been realized.
LITERATURE
[ I ] F. van Asperen ( 1 9 8 4 ) , Het optiinaliseren van de
eigenfrequenties van axiaalsynimetrische construc- t i e s toegepast op een luid- of carrilionklok, Re-
port WFW-84. O 12, Techn
.
Univ. Eindhoven ( i n Dutch).
[2]
P. Maas (19851, Onderzoek naar de geometrie van eengrote-terts klok, Report WFW-85.027, Techn. Univ.
Eindhoven C j.n Dutch)
.
[ 3 ] U . A . Cox (19581, Planning of Experiments, John Wiley 6r Sons.
[4] And16 Lehr 119761, Leerboek der campanologie, Na-
tj onaal Beiaardmuseum, Asten, Holland ( i n Dutch)
.
[SI R. Perrin, T. Charnley, J. De Pont, Noma1 modes of
the inodern English church bell, Journal o f Sound
[6] W.G. Cochran and G.M. Cox, Experimental designs,
and ?ihLUtj.G:: ( 1983 1 3û 1 ;
,
29-49. John Wiley 6r Sons, 1957.-22-
Take a gridpoint: and evaluate the first frequency r a t i o . I f the side constraints ( 3 7 ) a r e not violated then pro- ceed w i t h the next frequency ratio; otherwise pass t o
the next g r i d point. I f all frequency r a t i o s for a cer-
tain g r i d point satisfy ( 3 7 ) , then note this point as a possible solution f o r the geometry of A major t h i r d b e l l .
Executing t h i s procedure several possible geometries
were found. Verified by means of d i r e c t f i n i t e element analyses, they proved t o be indeed very close t o the
ideal. Figure 6 shows some of the resulting geometries.
Note t h a t a i l the geometries show a inore o r l e s s pro-
nounced bump a t h a l f the height o f the b e l l .
.4
. -675
Figure 6 . Same possible geometries f o r a major- .
t h i r d b e l l .
,
After some s l i g h t modifications of the most promising geometry the b e l l foundry, that we axe co-operating