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

K. M.

Ragsdell Optimal Design of a Class of


Assist. Professor,
School of Mechanical Engineering
Mem. ASME Welded Structures Using
D. T. Phillips
Associate Professor,
Geometric Programming
School of Industrial Engineering.
Geometric Programming is a new technique developed to solve nonlinear engineering de-
sign problems including linear or nonlinear constraints: This paper illustrates the use of
Purdue University,
West Lafayette, Ind.
Geometric Programming in obtaining optimal design parameters for a class of welded
beam structures. The procedure is illustrated through the solution of a particular welded
beam design formulation. In G/P format the problem solved consists of 9 nonlinear con-
straints, 24 terms, 7 variables, with 16 degrees of difficulty and a nonlinear objective
function. Geometric Programming is compared to several other solution techniques, and
found to be very efficient. Computational experience suggests that other problems of this
class may be solved with similar efficiency. The welded beam problem given is a real
world design situation typical of many encountered in actual practice. The solution is
given for the first time in this paper.

Introduction viously decided that bar "A" will be welded to member "B". Out-
side considerations fix the material of the bar "A", F, and L.
Many very practical design opportunities can be formulated as
optimization problems, in nonlinear programming form. A high F = 6000 lb
percentage of these nonlinear programs can be stated in algebraic L = Win.
form. This is particularly true in the design of machine elements. BAR"A" = 1010 steel.
Geometric Programming is an optimization technique invented The objective is to find a feasible combination of h, I, t and b such
by Richard Duffin and Clarence Zener [l], 1 which is enjoying wide that the total cost is a minimum.
attention, both in theory and application. The attention currently Objective Function: Cost. We must now relate the total sys-
given to the technique is justified in large part by the vast number tem cost to the design variables;
of practical design applications. Very many problems in the struc-
tural design of machine elements can be placed in the proper for- x = [xu x2, x3, xt, ]T = [h, I, I, b]T (1)
mat for Geometric Programming. Such a practical problem origi-
The major cost components of such a weldment assembly are: (a)
nally suggested, but not solved by Harold Keith [2] is reformulated
set-up labor cost, (b) welding labor cost, and (c) material cost.
and solved using several methods including Geometric Program-
ming. This problem is really one of a family of structural optimiza- (2)
tion problems, which can be similarly formulated.
c„ + c, + c,
where
The W e l d e d B e a m F(x) = objective function
Consider the welded structure shown in Fig. 1. This welded as- Co = set-up cost
sembly is being considered for mass production. It has been pre- Ci = welding labor cost
C2 = material cost.
(a) Set-up Cost: Co- The company has chosen to make this
component a weldment, because of the existence of a welding as-
1
Numbers in brackets designate References at end of paper. sembly line. Furthermore, we assume that fixtures for set-up and
Contributed by the Design Engineering Division and presented at the De- holding of the bar during welding are readily available. We will
sign Engineering Technical Conference, Washington, D. C, September 17- therefore ignore Co in this particular total cost model.
19, 1975, of THE AMERICAN SOCIETY OF MECHANICAL ENGI-
NEERS. Manuscript received at ASME Headquarters June 6, 1975. Paper (b) Welding Labor Cost: cy. Let us assume that a man will do
No. 75-DET-86. the welding, and that his total wage is $10 per hour, including

Journal of Engineering for Industry AUGUST 1976 / 1021


Copyright © 1976 by ASME

Downloaded From: http://manufacturingscience.asmedigitalcollection.asme.org/ on 01/29/2016 Terms of Use: http://www.asme.org/about-asme/terms-of-use


0,(X) = T„ - "Kx) > 0 (10)
Bgzi.

Fig. 1
4 -ibfe
overhead, vacation, benefits, etc. Furthermore, let us assume that
<p2(x) = at - o-(x) > 0

<£3(x) = xt - xx > 0

tf>4(x) = x2 & 0

*5W 0
(11)

(12)

(13)

(14)

he can consistently lay down a cubic inch of weld in 6 min. So the 0 e (x) = P c (x) - F > 0 (15)
labor cost is:
<£7(x) = xx - . 125 > 0 (16)

c (10 (3) ' 8 (x) = .25 - DEL(x) > 0 (17)


» = h?)(60r^)(6"in^) = 1{
^)V«
where where
Vw = weld v o l u m e , i n . 3 ra = design shear stress of weld
T(X) = maximum shear stress in weld; a function of x
(c) Material Cost: C2- aa '= design normal stress for beam material
<r(x) = maximum normal stress in beam; a function of x
c
2 = c,Vw + c4AVV,
B (4) Pc(x) = bar buckling load; a function of x
DEL(x) = bar end deflection, a function of x
where
c 3 = $/volume of weld material = (.37) (.283) ($/in.' It is therefore necessary to model the important stress states.
d = $/volume of bar stock = (.17) (.283) ($/in.) (a) Weld stress: T(X). After Shigley [3], the weld shear stress
VB = volume of bar "A," (in.) has two components, T' and T"; where r' is the primary stress act-
ing over the weld throat area, and T" is a secondary torsional
Therefore, stress.
T' = F/{i2x^ (18)
Vw = 2(|ft 2 Z) (5)
n and
T" = MR/J
and (19)
with,
VB = tb(L + T) (6)
M-= F[L + (x 2 /2)] (20)
so:
c 2 = c3h2l + cttb(L + V) (7) R = {(x^/4)+[(x3 + Xl)/2fy^ (21)
2 2
Therefore, we see that the objective function becomes: J = 2{.707# 1 % 2 [(x 2 /12) + (x 3 + x t ) / 2 ) ] } (22)
F(x) = h2l + c3hH + e4tf>(Z, + t) (8) where:
and if: xi = h, X2 = I, xa = t, x 4= b M = moment of F about center of gravity of weld
group.
-F(x) = (1 + c 3 )x 1 2 x 2 + c 4 x 3 % 4 (£ + x2) (9) J = polar moment of inertia of weld group.
Functional Constraints. Not all combinations of x\, x% X3 Therefore, the weld stress T becomes:
and Xi can be allowed; if the structure is to support the load re-
quired. We can state several functional relationships between the T = [(T')2 + 2 T ' T " cos0 + (T")2]1/2 (23)
design variables which delimit the region of feasibility. The rela- where,
tionships are in the form of inequality constraints. xJ2R (24)

.Nomenclature.
F — applied cantilever load, lb Td = weld design shear stress, psi G = bar material shearing modulus, psi
L = fixed distance from load to support, in. T(X) = maximum shear stress in weld, psi y m (x) = m t h generalized polynomial
x = vector of design variables or primal ad - design normal stress for beam materi- "mt = sign term in primal problem, mth
variables al, psi polynomial, tth term
h = weld thickness variable, in. <j(x) = maximum normal stress in bar, psi Cmt = coefficient term in primal problem;
/ = weld length variable, in. P c ( x ) = bar buckling load, lb mth polynomial, £ th term
t = bar thickness, in. DEL(x) = bar end deflection, in. "m = mth inequality constraint value
6 = bar breadth, in. T' = primary weld stress, psi Tm = number of product terms in mlh
F(x) - objective function, $ T" = secondary weld stress, psi polynomial
co = set-up cost, 1972$ M = effective moment from F on weld T = total number of terms in primal prob-
ci = welding labor cost, 1972$ group, in.-lbs lem
C2 = material cost, 1972$ J = polar moment of inertia of weld, in.4 d(amt) = dual function
C3 = specific cost of weld material, 1972$/ 8 = angle between primary and secondary o>mt - dual variables
cu. in. weld stress vectors, radians d*(o>mt) = maximum value of dual function
C4 = specific cost of bar stock, 1972$/cu. in. E = Young's Modulus for bar stock, psi yo*(x) = minimum value of primal func-
VB = volume of bar stock, cu. in. / = area moment of inertia for bar, in. 4 tion
Vw ~ volume of weld material, cu. in. a = torsional rigidity for a narrow rectan- x* = optimal design variables
<fe(x) = inequality constraint functions, k gular beam, in. 2 -lb ^ mt = optimal dual variables
= 1,2, 3 , . . . , 8

1022 / AUGUST 1976 Transactions of the ASME

Downloaded From: http://manufacturingscience.asmedigitalcollection.asme.org/ on 01/29/2016 Terms of Use: http://www.asme.org/about-asme/terms-of-use


(b) Bar Bending Stress: a (x). The maximum bending stress is where
seen to be: a = ±1 (36)
2
a = 6FL/{xiX3 ) (25) and N orthogonality conditions;

(c) Bar Buckling Load: Pc (x). If the ratio (t/b) = (X3/X4) grows
large, there is a tendency for the bar to buckle. We must disallow
those combinations of X3 and X4 which will cause this buckling to
S S V „ A = 0 ; H = 1,2,..„W (37)
m=0 (=1
occur. It has been shown [4] that for narrow rectangular bars, a
good approximation to the buckling load is: The solution to this dual problem is found subject to T nonnegati-
vity restraints;
p.W = M 1 | ^ tl A\
V
EI]J (26)
2L;V a m = 0,1,2, . . . ,M
where "mt s
0; (38).
e
t = 1,2,3, . . . , T
E = Y o u n g ' s m o d u l u s = 30- 10 p s i

4,
(10) x x
zi
and M linear inequality constraints,
r
a = (g-)G«3x4 3 w m0 = amE omtwmt a 0; m = 1, 2, . . . , M (39)

G = s h e a r i n g m o d u l u s = 12* 10G p s i where

(d) Bar deflection: DEL (x). Assume the bar to be a cantilever T=HTm (40)
of length L to calculate the deflection. m=0

DEL(x) = 4FL3/(Ex33xt) (27) Using (35), (37), (38), and (39) the dual function can be derived
and is:
(e) Other Constraints. $3 states that it would be foolish to have
the weld thickness greater than the bar thickness. $4 and $5 are
nonnegativity restraints. Ndte that restraints on xi and X4 are d(0)mt) = 0 - [ £ I I (SjS&dymtUmt)]* (41)
w
given by other constraints. 07 is of real practical interest. This con- m-0 t=l • mt
straint introduces the fact that it is not physically possible to pro- where
duce an extremely small weld. a = ± 1 , and h a s the s a m e s i g n as 3>o*(x)>
Let: Td = 13,600 psi
and <Td = 30,000 psi and woo = 1, if by convention:
to complete the formulation of this welded beam problem as a
nonlinear program. (42)
Geometric Programming
Many nonlinear programs can be expressed as a set of General- It can be shown [5] that if a minimum exists for the primal func-
ized Polynomials, y m (x): tion, yo(x) and if all the srgnum functions are positive, any set of
dual variables wmt making d(wmt) a global maximum will also give
ym{x) = T} omtcmt I I xn%' 0,12, .,M. (28) the global primal minimum, yo*(x). So:
t=1 n=l
where
d*{umt) = V t o (43)
= ±1 t = 1,2,3, m = 0, 1,2, . . , M. (29)
and Once the optimal dual variables w*m( and a are known, the values
> 0. t= 1 , 2 , 3 , . . .,Tm, m = 0,1,2, .,M. (30) for the optimal primal variables are easily determined from the
following equations:
The primal problem is to minimize the objective function in N
variables, yo(x) where; Cot tl x
n " = w
ot a :Vo*( x ) = umod*(ii>mt) t = 1, 2, . . . , T 0
m»l
x — [xi, x%, x3, . . ., xn\ (31) (44)
and
subject to M inequality constraints
„=i ">m0 m — i , &, 0, . . , m
ym(x) sam; in = 1, 2, 3, . . . , M (32)
where
Although both equations (44) and (45) are nonlinear, they are non-
om = ± 1 , in = 1 , 2 , 3 , . . . , M (33)
linear in only one term, so that the primal values can easily be re-
and N nonnegativity restraints; covered by taking logarithms and solving the resulting simulta-
neous equations which are linear in log xn; n = 1, 2, 3 , . . . , N.
x„ > 0; n = 1,2,3, . . .,N (34) In order to avoid unnecessary confusion, a word is in order re-
garding notation. Note that all variables and constants are doubly
Geometric Programming does not solve the primal problem di- subscripted except for the exponents of the primal variables,
rectly, but instead solves an equivalent dual problem in T vari- which contain three subscripts. Observe the following symbology:
ables, given by a generalized normality condition:
Subscripts
T
o 1 s t subscript: the equation
a = 2jo-otw0f (35) 2 n d subscript: the term
3 r d subscript: the variable

Journal of Engineering for Industry AUGUST 1976 / 1023

Downloaded From: http://manufacturingscience.asmedigitalcollection.asme.org/ on 01/29/2016 Terms of Use: http://www.asme.org/about-asme/terms-of-use


For example; x + y
*• - * i (** j)2
*G ~ 12 4 (55)

tl ifth t e r m x 7 a ^2_ + ^
4
3 +

4
x
i) (56)
mth e q u a t i o n .
Using (53), (54), and (56), we now obtain an equivalent G/P con-
(m = 0 r e p r e s e n t s t h e o b j e c t i v e function;
straint formulation for (10):
m > 0 represents a constraint).

y6(x) = ^ V V V 1 + f J*£*rVVV
tit variable.
- jfth t e r m
-mth e q u a t i o n .
+ g - F ^ X i ' V V 1 * ? ^ " 2 + g i ^ x f W 2 * ? s 1 (57)
Hence, the position of every variable within the problem formula-
tion is easily determined through its subscripts. i'M) = jx^xfl + ^x^x1-i+^x1x3xfl + l ^ x f 1 s 1 (58)
Degrees of Difficulty. The difference between the number of
variables and the number of terms involved (including the con- ya(x) = 12x e x 2 - 2 - 3x 3 2 x 2 - 2 - 6xjX 3 x 2 - 2 - 3x! 2 x 2 - 2 < 1 (59)
straint terms) is called the degrees of difficulty. Duffin and Zener
y 9 (x) = 7 . 3 5 2 9 - 1 0 - 5 x 5 1 / 2 < 1
suggested this metaphor, because the set of equations given in (35)
and (37) do not admit a unique solution for the <o*mt values. How- And the usual nonnegativity restraints are assumed;
ever, N + 1 equations are still independent and can be solved in
terms of any T — (N + 1) of the remaining wmt variables. The pri-
mary difficulty arises in solving equation (41) for the optimal x„ > 0 w = 1 , 2 , 3 , . . . , 7 (60)
value,, d*(umt); since d*(umt) cannot be calculated until all o>*mt
The problem is now in correct Geometric Programming format.
values are uniquely determined.
Note that there are 7 variables, 24 terms, and 9 constraints; so the
Duffin has shown that if equation (41) is written in terms of T — problem has 16 degrees of difficulty.
(N + 1) of the variables, the optimal value of the unknown omegas
can be determined by maximizing this substituted dual function,
G / P Solution
and that this maximum value is equal to the minimum cost,
yo*(x). This desirable property is only true when all signum func- Using the dual approach associated with Geometric Program-
tions are positive. Solution techniques also exist for negative sig- ming the following problem is formulated:
num functions. In general, Duffin proved that a sufficient condi- MAXIMIZE:
tion for the primal function to be minimized is that the dual func-
tion in equation (41) be maximized with respect to the u>mt, subject d(a, ) = [-(1-10471 ^ 0 1 ( / 3 7 3 5 ) . 0 2 ( / J 4 8 1 1 ) a , 0 3 ( 1 6 8 ) 1 o 1 1
to the conditions given by equations (35) and (37). Once the dual w 01 w02 _ w 03
variables have been determined, the solution procedure for the op-
31 41 O 9 4 2 8
timal primal values is the same as before. x ( l ) " » ( .125)" ( .908)" c ^ 0 ) " 5 i ( .O2776co 50)M52
w»'51 0),'52

G / P F o r m u l a t i o n of W e l d e d B e a m x (18 10ea»8y,1(2.52-108a)l,0)Mgii(9'10Ba>,n.WM
Minimize: Ww61
d W
u e,
62 Wo,
u
63
9 8
3.528- 10 wfi()NWfi4
J\™64 (
. 2 . 5 2 - 10 wfi&L\"65
n ^ , 4 . 5 - 106w,
!m\u66
F(x) = 1.10471 xfx2 + 0.6735 x 3 x 4 + 0.04811 x 2 x 3 x 4 (46) •) S 5
( -

Subject to: / •25o)70. u 71 . .25w 7 0 . m72 . 50w 7 0 . c 7 3 . .25w 7fl n 74 - ^ w ^ . <d81


x
v
f r o m (11) yt(x) = 16.8 x4-lx3'2 s 1 (47) w71 w72 w73 ' w74 co81

f r o m (12) y2(x) = xxx^ s 1 (48) x ( 1 5 ^ - 8 2 (6^M)-»8S ( 3 0 ^ — M (7>3529)«91] (61)

f r o m (16) y 3 (x) = 0.125 x / s 1 (49)


f r o m (17) y 4 (x) = 9.08 x 3 " 3 x 4 < 1 (50) SUBJECT TO:
3
f r o m (15) y 5 (x) = .09428 x^xf + .02776 x 3 £ 1 (51)
a = w„, + w02 + w03 = 1.0 (62)
The remaining constraints are generated from equation (10) in
2w 01 + w21 - w31 - 2w 61 - 2w 62 - 2w 63 - 2w 64 - 2wG5 - 2coG6
the following manner.
From equation (10): + w83 + 2w 84 = 0 (63)
w
(52) oi + <% - 2w e i - w62 - 2w 64 - w 65 = 0 (64)
r/Td £ 1
w + w - 2w 3w w + w 2co w _ 2cO
02 03 ll ~ Sl ~ 51 52 + T2 + f3 02
and from (23)
- a » 8 3 = 0 (65)
r-^ J) -2
W + W W W
02 03 ~ ll ~ 21 + W
41 ~ 3w
51 = ° (^^

" . .1
W
G1 ~ W
62 ~ W
63 ~ (!4 ~ G5 ~ ^ 6 + .5 S i = 0
w W (67)
-fix 2(.707)x 1 * 2 [(x 2 Yl2) + ((*, + X0V4)] - 62w _ w
63 ~ 2wG4 - 2a)G5 - 2wGe + wB1 = 0 (68)
+ F L
^ +
V 2 3 ' + W/A +
(*3 + * l ) V * ] 11/2 , „ , w
64 + w
65 + w
66 "" w 72 _ w
73 ~ w
74 = ° (69)
2 * 1 x 2 [ x 2 7 l 2 ) + (x s + x*,)74]
,)V4] * [
' W
50 - W51 - 0) 5 2 = 0 (70)
Now l e t ;
^eo - w61 - co62 - a)63 - a)G4 - wG5 - wG6 = 0 (71)
W _ W — w w _ w
+ 1
9 9 . 70 71 ?2 ^ 73 74 = 0 (72/
&Ao £t%\ An A g w«o
u
80 _- ^Wo,
8 1 T+ alw„,
82 + w „ + Wo, = 0 (73)

1024 / AUGUST 1976 Transactions of the ASME

Downloaded From: http://manufacturingscience.asmedigitalcollection.asme.org/ on 01/29/2016 Terms of Use: http://www.asme.org/about-asme/terms-of-use


Using condensation [6] the following solution is obtained: SIMPLX: Simplex method and penalty function.
RANDOM: Richardson's random method.
w*0, = .1730 w
*52 = .0450 w* 72 = .2040 W*10 = .1630 The best point found by each of the methods is given in the fol-
.0121 * .0120 .2060 lowing table. The starting value used in each case is
w*02 = .5730 W*61 = W
*20 =
O) 73 =
w*03 = .2540 W*e2 = .0493 w* 74 = .0002 ^*50 = .1980 x0 = [ l , 7 , 4 , 2 F (80)
w*n = .1630 "*G3 = .0109 W*81 = .4920 W
*G0 = .2880
Xl* x2* x3* Xi* F(x*)
w*21 = .2060 W*64 = .1447 <A Z = .3940 OJ*70 = .3300
ADRANS .2393 6.3280 8.3695 .2458 2.41
, w* 31 = 0.000 "*G5 = .0640 W
*83 = .0230 W
*80 = .0740 APPROX .2444 6.2189 8.2915 .2444 2.38
w*41 = 0.000 W .0071 .0003 W .0740 DAVID .2434 6.2552 8.2915 .2444 2.38
*GG = W*84 = *90 =
MEMGRD .2441 6.2273 8.2927 .2444 2.38
w* 51 = .1520 OJ* 7 1 = .1140 OJ* 9 1 = .5760 SEEK1 .3918 4.6154 6.2309 .4327 3.20
SEEK3 .2454 6.2148 8.2546 .2466 2.39
Hence, from (61) SIMPLX .2792 7.7512 2.53
5.6256 .2796
d*(oomt) = y 0 * ( x ) = 2 . 3 8 6 $ . (74) RANDOM .4575 4.7313 5.0853 .6600 4.12

Recall that the only variables of interest are %\, x% xs and Xi, Closure
therefore any four dual variables can be used via equation (44) and The welded beam problem considered, proved to be a real test
(45) to recover the optimal values. Hence; for the Opti-Sep programs. The results reported in the previous
section were obtained only after very careful attention to the vari-
1.10471 xi2x2 = 2.386(w* 0 1 ) (75) ous program dependent parameters. The Geometric Programming
approach seemed to have no real difficulty with the problem.
0.6735 * 3 * 4 = 2.386(co* 02 ) (76)
The welded beam is a member of a large class of similar struc-
0.04811 x2x3xt = 2.386(co* 03 ) (77) tural design problems which can be similarly posed. One would ex-
pect favorable results using Geometric Programming for these
16.8*4-V = 1
(78) other problems. In general the authors computational experience
As stated earlier, these equations are linear in the logarithms and indicates that if the design problem is in nonlinear polynomial
can be easily solved to yield:' form, G/P is a superior solution algorithm. The authors are pres-
ently performing comprehensive comparisons on a large class of
x*t = h* = 0.2455 i n . engineering design problems in order to compare available solution
techniques.
x* 2 = I* = 6.1960 i n .
(79)
x*3 = t* = 8.2730 i n . Acknowledgment
x*t = b* = 0.2455 i n . The authors would like to thankfully acknowledge the program-
ming assistance of several graduate students, in particular Mr.
Comparison to O t h e r Methods Gary Gabriele.
The welded beam problem has been attempted using all of the
"Opti-Sep" [7] methods. Opti-Sep is a user-qriented package of References
optimization subroutines, ideally suited to design applications. 1 Wilde, D. J., and Beightler, C. S., Foundations of Optimization, Pren-
One of the major assets of the package is the small problem prepa- tice-Hall, 1967, Chapter 4, p. 99.
ration time required. A broad range of methods is included in the 2 Keith, Harold D., "Optimization Techniques in Design," ASME, No.
69-DE-14,1969.
package. The reader is referred to Professor Siddall's book [8], or 3 Shigley, J. E., Mechanical Engineering Design, McGraw-Hill, 1963, p.
the Opti-Sep users manual for a complete description. For the sake 271.
of the present discussion the following brief description of the sub- 4 Timoshenko, and Gere, Theory of Elastic Stability, McGraw-Hill,
routines employed is given: 1961, p. 257.
5 Wilde, D. J., and Beightler, C. S., Foundations of Optimization, Pren-
ADRANS: Gall's adaptive random search and penalty function. tice-Hall, 1967.
APPROX: Griffith and Stewart's successive linear approximation. 6 Phillips, D. T., and Beightler, C. S., Applied Geometric Programming,
DAVID: Davidon-Fletcher-Powell and penalty function. John Wiley, to appear.
MEMGRD: Miele's memory gradient and penalty function. 7 Siddall, J. N., "Opti-Sep: Designer's Optimization Subroutines,"
McMaster University, Canada, 1971.
SEEK1: Hooke and Jeeves and penalty function. 8 Siddall, J. N., Analytical Decision-Making in Engineering Design,
SEEK3: Hooke and Jeeves and different penalty function. Prentice-Hall, 1972.

Journal of Engineering for Industry AUGUST 1976 / 1025

Downloaded From: http://manufacturingscience.asmedigitalcollection.asme.org/ on 01/29/2016 Terms of Use: http://www.asme.org/about-asme/terms-of-use

You might also like