Escolar Documentos
Profissional Documentos
Cultura Documentos
upper bounds of the response variability of stochastic systems. The structural sys-
tems examined consist of linearly elastic trusses and frames subjected to static
loads. The computation of these bounds is achieved by extending the notion of
the "variability-response function" of a stochastic system to trusses and frames
analyzed by the finite element method. The variability-response function presents
many similarities to the frequency-response function used in random-vibration
analysis. Specifically, the variance of a specific response quantity is calculated as
the integral of the product of the power-spectral-density function describing the
stochastic properties of the system multiplied by the variability-response function
of the response quantity. Of equal importance is that this work provides insight
into the underlying mechanisms controlling the response variability of stochastic
truss and frame structures and an analytical basis on which the analysis can be
extended to two-dimensional structures such as plates and shells.
INTRODUCTION
Truss Element
Consider the truss element shown in Fig. 1(a) with two degrees of free-
dom. Assume that the elastic modulus of the element varies randomly along
its length according to the following form:
EM(x) = Etf[l +fie\x)] (1)
(e)
where Eff = the mean value of the elastic modulus; and f (x) = a one-
dimensional, univariate, zero-mean, homogeneous stochastic field. To avoid
the possibility of obtaining nonpositive values of the elastic modulus, the
stochastic field /<<0(x) is assumed to be bounded as follows:
- 1 +T,S/W(X) - tl (2a)
0 s x < L(e) . . . . (2b)
Ui u
x= 0 L<e
(a)
M3j-1
A"3t-1
u
"3t-2 3j-2
U3i
(b)
z = 0
566
and the element stiffness matrix can be written in the following form:
K w = K^e> + Xtf • AK&° .. (4)
where
... A^Etf r i - I "
*V = -&r-l-\
A<!
1J <5)
*T>(e) Eo 1 -1 ...
AK = (6)
° ~W'[-i IJ
XQe) is a random variable defined as
BEAM-COLUMN ELEMENT
Consider the beam-column element shown in Fig. 1(b) with six degrees
of freedom. Assume that the elastic modulus of the element varies randomly
along its length according to Eqs. 1 and 2. Using again standard finite ele-
ment analysis methodology based on the principle of stationary potential en-
ergy, the element nodal-displacement vector is denoted by
u w = [a3,-2 ua-i a3i a3j-2 "3,-1 u3J\T (8)
and the element stiffness matrix can be written in the following form:
K w = K&° + Xtf • AK&° + Xf • AK$e) + X? • AK£> ; (9)
where
~AE0/L 0 0 -(AEa/L) 0 0
0 12(E0I/L3) 6L(.E0I/L3) 0 -12(£„//L3) 6L(E0I/L3)
0 6L(E0I/L3) 4L2(E0I/L3) 0 -6L(E0I/L3) 2L2(E0I/L3)
-(AEJL) 0 0 AE0/L 0 0
0 -12(E0I/L3) -6i(£ 0 //L 3 ) 0 12(£0//L3) -6L{EaI/L3)
0 6L(E0I/L3) 2L\E0I/L3) 0 -6L(£0//L3) 4L\E0I/L3) _
567
0 o 0 0 0
0 -144L(£ 0 //L 6 ) -84L2(£:o//Z.6) 144i(£,,//L 6 ) -60L2(£0//i6)
0 -84i2(£0//i6) -48L 3 (£ 0 //i 6 ) S4L2(E0I/L6) -36L\E0I/L6)
AK(," = (12)
0 0 0 0 0
0 144L(£ 0 //L 6 ) 84L 2 (E 0 //i 6 ) -144i(£0//i )6 60L\E0I/L6)
0 -60L\E0I/L6) -36L 3 (£„//L 6 ) 0 6QL2(E0I/LS) -24L 3 (£ 0 //L 6 )
"0 0 0 0 0 0
o 144(£„//L6) 12L(E0I/L6) 0 -144(£0//£-6) 72L(E0I/L6)
o 12L(EJ/L") 36L\E0I/L6) 0 -72L(E0I/L6) 36L\E0I/L6) (13)
AK?
o 0 0 0 0 0
o -144(£0//L6) -72L(E0I/L6) 0 144(£ 0 //i 6 ) -12L{E0I/L6)
o 12L(E0I/L6) 36L2(E„//Z,6) 0 -12L(EJ/L6) 36L\E0I/L6)
Xie), and X20 are random variables defined as
/<e)(x)dc (14)
Jo
fJ
X?' = xfe\x)dx (15)
Jo (16)
and EQ, L, A, and / stand for E0'\ LM, Al'\ and Iie\ respectively, for the
sake ofJosimplicity in the notation. Lie), A{e), and I(e) = length, cross-sectional
area, and cross-sectional moment of inertia of the element, respectively.
Note again that matrix Koe) = mean value of matrix K<e) since random
variables X0e\ X('\ and X1? have mean value equal to zero. Equivalently, it
can be said that matrices Koe> and Xf • AK0<° + X? • AK?° 4- Xf A K ? '/'"" =
deterministic and stochastic parts of the element stiffness matrix K(e) re-
spectively.
The stochastic element stiffness matrix of the truss element (Eq. 4) and
the corresponding one of the beam-column element (Eq. 9) have both been
derived in a local coordinate system. To assemble the global-stiffness matrix
of the structure, both matrices have to be transformed into a global coor-
dinate system.
Truss Element
Fig. 2(a) shows the local coordinate system (x,y) of the truss element, the
global coordinate system (x, v), and the corresponding two sets of displace-
568
«2i-l
"3,-1
«3,-l"
«3;-2
"3i, U3i
FIG. 2. The Two Sets of Nodal Displacements (in Local and Global Coordinates)
for (a) the Truss Element; and (b) the Beam-Column Element
ments. Note that the element is oriented at an angle 8 from the jc-axis. The
element nodal-displacement vector in the global-coordinate system is given
by
U (e) = [«2,--i «2/ Uy-i U2j\T (17)
and the element stiffness matrix in the global coordinate system is expressed
as
K<«> = K « + z w . AK#> (18)
where
Ugo = Twjj(e)T(«) (19)
and
rji(c) cos 6 sin 8 0 0
(21)
0 0 cos 8 sin I
569
Beam-Column Element
Fig. 2(b) shows the local coordinate system (x,y) of the beam-column
element, the global coordinate system (x, y), and the corresponding two sets
Downloaded from ascelibrary.org by Universidad De Sevilla on 06/23/15. Copyright ASCE. For personal use only; all rights reserved.
where
K<f> = TCe)rK^Tw (24)
AKg° = TCe)rAKg°Tw (25)
(
AK,<° = T ^ A K ^ T ^ (26)
AKg° = T (e)r AK^T w (27)
and
cos e sin 9 0 0 0 0
—sin 0 cos 9 0 0 0 0
rW 0 0 1 0 0 0
(28)
0 0 0 cos 9 sin 9 0
0 0 0 -sin 6 cos 9 0
0 0 0 0 0 1
Note that Xg°, Xf\ Xf, Kg0, AKg°, AK^, and AKg° have been defined in
Eqs. 14, 15, 16, 10, 11, 12, and 13, respectively.
Global-Stiffness Matrix
Using the standard finite element analysis methodology, the global-stiff-
ness matrix K is assembled as follows:
Ne
K
= E K(e) (29)
570
x'~Ye\x)dx (32)
Jo
where
e= 1,2, ...,Ne (33)
»= l.JV, (34)
[ l for trusses]
A (i:>)
' [3 for frames]
From Eq. 32 it is obvious that the mean values of random variables Xel;
e = 1,2, ...,Ne;i= 1, N„ denoted by Xei; e = 1, 2, ..., Ne; and i = 1,
Nt are all equal to zero
Xel = 0 (36a)
e = 1, 2, ..., Ne (36b)
i = 1, N, (36c)
Another obvious conclusion drawn from Eq. 31 is that the global nodal-
displacement vector U will also be a function of random variables Xel; e =
1, 2, ..., Ne; and i = 1, N,
U = \J(Xel; e = 1,2,,..,Ne;i = 1,N,) (37)
Denoting now by U0 and Ko the values of the global nodal-displacement
vector and global stiffness matrix evaluated at Xei; e = 1, 2, ..., Ne; and i
= 1, Nt, respectively, the following relation holds between U0 and Ko:
U0 = Ko-'P (38)
Consider now the first-order approximation of the Taylor expansion of
function U around the mean values of random variables Xel; e = 1, 2, ...,
Ne; and i = 1, Nt
u s u
o + 2 2 ex* - XJ dXe, (39)
Substituting now Eq. 40 into Eq. 39, the following result is obtained:
K Ni
u u 3K
= o- EE v Un " Xei (41)
Hence, the first-order approximation of the mean value and the covatiance
matrix of U are now easily evaluated as
€[U] = U0 (42)
r
Cov [U,U] = e[(U - Uo)(U - U0) ]
Ne N, Nt Ni
' dK " 9K
= 2 2 2 2 Ko1
e i = l e 2 = l i'i=l r'2=l
_OA,, 1 / 1 _ £
U 0 UJ
_dXe2i2_
(KQ ) • elXe^-XeyJ (43)
Using Eq. 32, the expected value shown in Eq. 43 can be written
J o Jo
fi1"1^"1*^! - &W62 (47)
Substituting now Eq. 49 into Eq. 43, the following expression is obtained
for the covariance matrix of U:
572
o Jo
Then, the variance vector of U consisting of the diagonal elements of the
covariance matrix of U is found as
/•" AT, JV,- Nt I
dia dK dK
Var[U] = SJ(K) • 2 E 2 S( ^ U0 K?1 U0
dX.i,
J-« e = l i i = l i'2=l V
J o
rli'1
|T V*^, •
Jo
,-£<«>
fr'e-'^dK (51)
where
^^^ / . raxi ,
V(K) = ESEdiag(Ko-' U0
e = l i'i = l i'2=l
SK
•Ko- g'rv*^, • gf-V"^ (53)
dXei o Jo
The number of components of vector V(K) is obviously equal to the total
number of degrees of freedom of the structure. Each component of V(K) is
interpreted as the first-order approximation of the variability-response func-
tion (Deodatis and Shinozuka 1989) at the corresponding degree of freedom
of the structure. It is a straightforward task to show that all components of
vector V(K) are real and even functions of K.
Trusses
Using Eqs. 29 and 18 and the fact that JV, = 1 for trusses, vector V(K)
shown in Eq. 53 can be written, after some algebra, in the following form
for a truss:
N,
d 2(1 - cos KDe>)
V(K) = 2 iag (Ko-'AK^Uo) • K^AKJ?>U0 (54)
1
gei = - • sin KLM (56)
K
and AKQ', AK^, and AK2e) — extended versions of the matrices defined in
Eqs. 25, 26, and 27, respectively.
574
negativity is possibly observed for larger values of the wave number K where
|V,(K)|2; I = 1,2, ..., Nd approach zero and where the effect of the system
stochasticity on the response variability becomes increasingly less signifi-
cant, as demonstrated by Shinozuka (1987). In the following, |V,(K)|2; i =
1, 2, ..., Nd are assumed to be nonnegative for any value of K. If any one
of them is indeed found to be negative for some values of K, it is set equal
to zero at those values.
For any form of the power spectral density Sff(K), the following relation
holds:
2
f
Jo
Sff(K)dK = <j2ff
ff
coefficient of variation of the elastic modulus. Using now Eq. 62, it is a
straightforward task to show that the variance of the rth nodal displacement
assumes its maximum value for the following form of the power spectral
density 5^(K):
where 8( ) is Dirac's delta function; and K,* = wave number at which |VJ(K)|2
takes its maximum value
|y,<Kf )|2 > |VXK)| 2 0 < K < O O (66)
The maximum value of the variance of the ith nodal displacement Var[w,];
i = 1, 2, ..., Nd is then obtained as
maximum {Var[w,]} = <r}- |V;(i<f)|2; i = 1, 2, ..., Nd (67)
The expression shown in Eq. 67 can be interpreted as an upper bound for
the variance
Var[«,] < <£• |y,(Kf)|2; i = 1, 2 Nd (68)
or the standard deviation
Std[«,] < aff- |V,(K?)|; i=l,2,...,Nd (69)
for any form of the power-spectral-density function 5 # (K). Obviously, it fol-
lows from Eq. 69 that the upper bound of the coefficient of variation of the
ith nodal displacement COV [«,] with nonzero expected value is given by
Std[«J |V,(K,*)|
COV [«,] = — i - f < % - T 7 1 T ; i=U2,...,Nd (70)
|e{«;}| |e{u,}|
575
Nt
N, N,
+ 22 W" dX„
* A,,.- (80)
Hence, the first-order approximation of the mean value and the covariance
matrix of S^' are now easily evaluated as
etS^] = 8jf) (81)
v
Cov [§P,§ff>] = e[(S > - S W > ~ %Pf]
= 2 2 AK^, T^WW'WAKf^teJ
11=1 i 2 = l
dV^
+ 2 2 A^.TVW T«V.e{Xfl<Xfi2}
l'l = l 12=1 L«yr2.
+ E E Koa>T(/)
ii=l i'2=l .Mfn.
T
"fltf0"
N, N< N,
+ 2 2 2 s ^. dX , _ e E . sxel _ E
e=l /|=1 '2=1
= 2i + 2 2 + £ 3 + £ 4 (82)
The four terms of the expression for Cov [S V) , S^'], S 1 ; 22> 2 3 , and 2 4
will be calculated now using the expressions given in Eqs. 40 and 49
J-°= / r l i2=l
222 = - f sff(K) • f E A K ^ ^ W I K O - 1 9X
Uo
J-=» i,=l i,= l I A2- E J©
,Ltf) itf)
_,
• T^Kf • ef e**^i • ^"'e-^fedK (84)
Jo Jo
23 = ^(K)-ff Ko^Ko 1 dK
Uo [Continued]
> i!=l ij=l LdXft. £ J ©
577
dK'
^4= I ^ W - E i E ^ ' T ^ f K o 1 dX„;
U0
> e=l ( , = 1 i'2=l ( E J®
Downloaded from ascelibrary.org by Universidad De Sevilla on 06/23/15. Copyright ASCE. For personal use only; all rights reserved.
7, LW L<
r raKi i r r "
• Ko"1 — - U0 T^ESf? • #-V1*^, • ?r 'e-'^^K (86)
In Eqs. 83-86, symbol { }@ denotes the following: Although the number
of components of vector { } is obviously equal to the total number of degrees
of freedom of the structure, { }© is a vector having only the components of
{ } corresponding to the degrees of freedom of element (/).
Then, the variance vector of S a ) consisting of the diagonal elements of
the covariance matrix of S^' can be eventually written
where
il = l ( 2 = 1
o Jo
(V; Nt r
dia
- 2 2 S ( A K ^ T ^ W ^ T ^ Ko" U0
!| = 1 12=1 •fil. E J©
o Jo
0 Jo
Ne N, N,
+ EIEdiag(Ko/,T^K0- dK
'* n Uol W'T^
e = l (,= 1 ( 2 = 1 V L^«(,JB ©
Ko" (88)
dXei - £ J © Jo Jo
The number of components of vector V ^ ' ( K ) is obviously equal to the
number of degrees of freedom of finite element ( / ) . Each component of
V£°(K) is interpreted as the first-order approximation of the variability-re-
sponse function of the corresponding internal nodal force of finite element
( / ) . It is a straightforward task to show that all components of vector
V¥\K) are real and even functions of K.
578
+ E diag (KPTV^m^QWlPT*"
where AKo0 or AKJ^ is the extended version of the matrix defined in Eq.
20.
Frames
Using Eqs. 29 and 23 and the fact that AT, = 3 for frames, vector
V7'(K) shown in Eq. 88 can be written, after some algebra, in the following
form for a frame:
dia
VKV) = 2 S S ( A K ^ T ^ U ^ A K ^ T ^ W ' • {QfilQfn + WfllWfl2)
'1 = 1 (2=1
3 3
dia
- E E S (AKJP.iyW^KSr'T^^Ko-'AK^.Uaf© • (j^.fi^ + W ^
(l=l i2=l
3 3
dia
- S E 8 (Koa,T(/){Ko-IAK^1Uo}0)AK^1Tv)Uov) • ( e A 2 A + W ^ )
'1 = 1 ' 2 = 1
AT, Ni Ni
+E E E <Kag(Ky>l^{K0-1AK?L1Uo}©)K{pIV)-{K0-,AKgi1Uafa>
•(Gw,fi* + WeilWe,-2) (90)
where the gs and Ws were defined in Eqs. 56-61 and AK&° or AK^,
AK(!e) or AK^' and AK£° or AK^> are the extended versions of the matrices
defined in Eqs. 25, 26, and 27, respectively.
579
where the effect of the system stochasticity on the response variability be-
comes increasingly less significant, as demonstrated by Shinozuka (1987).
In the following, I V ^ K ) ! 2 ; i = 1, 2, . . . , Nf are assumed to be nonnegative
for any value of K. If any one of them is indeed found to be negative for
some values of K, it is set equal to zero at those values.
Following the same procedure as the one used for obtaining the spectral-
distribution-free bounds of the response variability of nodal displacements,
the following expressions are established for the upper bound of the variance
of the ith internal nodal force of finite element (/), Var^K']; i = 1, 2, . . . ,
N/.
VarfSY'] == <£. |y<f>(K,*)|2; i = 1, 2, ..., Nf (92)
and for the upper bound of the standard deviation:
Std[SK>] ss <jr \VT(*?)\; i=l,2,...,Nf (93)
for any form of the power spectral density 5 # (K). K,* is the wave number at
which l^rf^K)!2 takes its maximum value:
|V<P(K?)|2 a |Vjf>(K)|2 0< K< oo (94)
Obviously, it follows from Eq. 93 that the upper bound of the coefficient
of variation of the fth internal nodal force of finite element (/), COV
[S^] with nonzero expected value is given by
St(
rvwra/h
a ) V [ f l =W ] V IVW)!
; . , „
M ' W '=1.2.-.^ (95)
It is emphasized again that the upper bounds given in Eqs. 92, 93, and
95 do not depend on the power-spectral-density or autocorrelation function
of the stochastic field f(x) and in this sense they are "spectral-distribution-
free" or "correlation-function-free" bounds.
For later purposes, COV [ ^ ' ( K ) ] is defined as
with the spatial variability of the elastic modulus described by the following
power-spectral-density function, for every value of the wave number K ap-
pearing in Eq. 96:
NUMERICAL EXAMPLE
® 15 (21
Downloaded from ascelibrary.org by Universidad De Sevilla on 06/23/15. Copyright ASCE. For personal use only; all rights reserved.
© 7 ai 11 17 15
© a4 20
© © ic © 1<
© © © ©
1 5 13
ffmrmp mmmr VWffff» immm
4.0 m 4.0 m 4.0 m
E
1.2 m
tet25 'm
Elements: 1, 7, 13, 19 —• 0.55 m x 0.55 m (cross-section)
Elements: 2, 8, 14, 20 - • 0.50 m x 0.50 m (cross-section)
Elements: 3, 9, 15, 21 -> 0.45 m X 0.45 m (cross-section)
FIG. 4. (a) |V 19 (K)| • 0.362• 101 or COV [K, 9 (K)]; (b) |V 20 (K)| • 0.754• 103 or COV [H 20 (K)];
and (c) |v2l(K)| • 0.562 • 102 or COV [« 2 I (K)]
solute value of the variability-response function |V,(K)| and COV [W,(K)] are
plotted as a function of wave number K in Fig. 4 for degrees of freedom i
= 19, 20, 21.
Then, using Eqs. 90, 91, and 96, the first-order approximation of the
absolute value of the variability response function of internal nodal forces
|V$P(K)| and COV [S^V)]; i = 1, 2, 3 are plotted as a function of wave
number K in Fig. 5 for element 3 ( / = 3).
In order now to calculate the spectral-distribution-free upper bounds of the
response variability of the structure (Eqs. 70 and 95), the values of K* (i =
1,2, ...,Nd for Eq. 70 and / = 1, 2, . . . , Nf for Eq. 95) have to be cal-
culated. Since this calculation cannot be performed analytically, the values
of K,* are computed graphically from Figs. 4 and 5. It should be pointed out
that it is not possible that K,* be larger than 10 rad/m (the maximum value
of the wave number plotted in Figs. 4 and 5), since it is very easy to show
that both V(K) and V^'(K) possess the property
lim V(K) = 0; lim V / V ) = 0 (98)
©
Downloaded from ascelibrary.org by Universidad De Sevilla on 06/23/15. Copyright ASCE. For personal use only; all rights reserved.
0.07
0.06
0.05
0.04
0.03
0.02
0.01 -
0.00
0.00 2.00 4.00 6.00 8.00 10.00
Wave Number K (rad/m)
Sff(\<.) and the influence of the correlation distance of stochastic field f(x) on
the response variability of a stochastic system [e.g. Shinozuka and Deodatis
(1988)]. Finally, it should be pointed out that no Gaussian assumption is
made for stochastic field f(x).
CONCLUSIONS
ACKNOWLEDGMENTS
This work was supported by the National Science Foundation under Grant
No. CES-8813923 with Dr. S.-C. Liu as Program Director. The writer wishes
to profoundly thank Professor Masanobu Shinozuka for his valuable com-
ments on this work.
APPENDIX I. REFERENCES
Segerlind, L. J. (1984). Applied finite element analysis. John Wiley and Sons, New
York, N.Y.
Shinozuka, M. (1987). "Structural response variability." J. Engrg. Mech., ASCE,
113(6), 825-842.
Shinozuka, M., and Deodatis, G. (1988). "Response variability of stochastic finite
element systems." / . Engrg. Mech., ASCE, 114(3), 499-519.
Vanmarcke, E., et al. (1986). "Random fields and stochastic finite elements." J.
Struct. Safety, 3(3 + 4), 143-166.
Yamazaki, F., Shinozuka, M., and Dasgupta, G. (1988). "Neumann expansion for
stochastic finite element analysis." J. Engrg. Mech., ASCE, 114(8), 1335-1354.
A = cross-sectional area;
E = elastic modulus;
/ = cross-sectional moment of inertia;
K = stiffness matrix;
P = force vector;
Rjf = autocorrelation function;
S = internal nodal-force vector;
Sff = power-spectral-density function;
T = transformation matrix;
U = nodal-displacement vector;
V = variability-response function (nodal displacements);
Yp — variability-response function (internal nodal forces); and
e = expectation.
585