Professional Documents
Culture Documents
Identification of Finite Element Models in Structural Dynamics 1993
Identification of Finite Element Models in Structural Dynamics 1993
Identification of Finite Element Models in Structural Dynamics 1993
models considered here are held to be sufficiently complete for any uncertainty to be restricted to the values of
parameters No account is taken of the problems arising
from the use of models unable to accurately describe the
behavlour of the structure
The choice of modal or physical models thus depends
on whether it is more reasonable to formulate a reahsUc
analyUcal model with accurate values for the
geometrical and mechanical charactertsUcs or whether it
is possible to develop a sufficiently broad series of
experimental tests to allow adequate measurement of a
sufficient number of response quantmes
In this paper some problems of structural ldenUficanon procedures are discussed. Theoretical aspects are
examined and a numerical study concerning tdentlflcanon of the numerical model of a building which uses
pseudo-experimental data is developed. A fimte element
model is used in which some mechanical characteristics
are treated as parameters. Model updating Is performed
m the frequency domain and m a Bayesian context An
opumal procedure for the selection of parameters and
the location of sensors Is proposed and the reliablhty of
the estimation techmque ~,, evaluated Several cases of
parameter ldennflcatlon are considered showing the
influence of the number of measurements on the
accuracy of the result,,
Methodology
For large and complicated structural systems refined
models that can reliably predict the response under
Ltd
z = h(x) + n
(2)
~_~ y,(x,)G,~
+ S F J] J [HIS,, ~ [ z - h ( x ~ ) ]
+ S,-~(Xo - x ~ ) ]
22
(3)
and
G.; : ,?Jz.(x)/(x' l, ,
(4)
Both matrices H and G can be evaluated explicitly due
to the polynomial nature of h(x)
After the optimal values 2 are determined, it is worth
evaluating the dispersion of the estimated parameters as
measured by the a posterlorl covarlance matrix S
Assuming a Gausslan distribution for the estimate el x
in the nelghbourhood o f . f . S is given bv ~
S=(HIS,,
JH+S,
(5)
II i
(6)
with a zero
reducUon is
may not be
vector t~ can
on the basis
(7)
(8)
(9)
(10)
(11)
A, tr(S-') = E
1/a~, (Oh,/Ox,) 2
(12)
(13)
I d e n t i f i c a t i o n o f the s t r u c t u r e
The general aspects previously presented and others
more closely related to the parameter estimation are
examined m detail here by reference to a particular
structural problem_ Pseudo-experimental data IS consldered m the Identification procedure, this has allowed
different s~tuattons to be mvesugated
23
y~
p
4.80 . ~ - ~
/.////'././/.//././//'/'////.,z/.////A
. q ~ - - 4.20 - ~ - ~
I
3.00
AI
A2
4.20
A1
J-
A3
2.80
Y./z'./Z V//-"////////;
Illllll
I IllllIJ
7.00
3.00
,!
7.///.~
A3
3.00
A1
5.50
AI
A2
f////J//.////////J///////~
"]y////.///////.//////.,4
3.00
1.40
El
M1
M2
M3
M4
M5
M6
Figure 2
M7
M9
M10
24
M8
given. Some modes are mainly translational in the ydirection (lst, 4th, 7th, and 9th) and m the x-direction
(2nd, 5th and 8th), while the others are mainly rotational
(3rd, 6th and 10th). In particular, m the higher modes
(8th, 9th and 10th) the presence of deformable soil is
evident.
To reduce the computational effort m the minimization of l(x), the exact relationship h(x) furnished by the
fimte element model is approximated by an analytical
function (h(x), following the approach reported in
Mode
10
Frequency
g,
g,
2 O0
0 46
7 73
2 26
7 83
0 62
2 93
1 28
0 94
588
0 25
3 51
6 76
3 62
0 37
8 19
087
063
939
0 16
2 65
106
3 19
031
120
0 31
3 57
126
064
044
References 10, 11 and 13 Each selected response quantity is described by a polynomial form
t~,(x) = c, + E
a,~(x~ - x )
1,
+ b,~(x k - x ) 2 i = 1, 2, .
, m
(14)
(15)
Forty displacement components are considered as candidates to be instrumented, the components of the first
10 modes and related frequencies are taken as a set of
410 observable quantities. By ranking the 40 components according to their decreasing contribution to
S -~, the 15 shown in Figure 3 are selected The importance of the different modes was also analysed; on
average the first modes are the most important, although
for certain parameters (x~ and x2 for example) the
higher modes are also significant, since they activated
the soil springs (Figure 2). While in almost all of the
cases studied the measured components are those
selected, the number of modes considered is changed to
modify the global number of experimental quantities
used
Several cases are considered in the investigation performed as shown in Table 2, they differ mostly in the
Figure 3
25
Table 3
10 2
1
10
25
101
10 2
10 3
10 4
10 5
10 6
10 7
45102
1 1 104
1 2 107
1 9 105
5 2 10 4
1 4 107
1 1 10 4
1
2
3
4
10 5
1
1
16
10 6
1
3
11
10 7
2
0
30
6 1
5
10
85
6 2
1
10
25
6 4
0
6
6
P~
P2
/93
p4
t%
P6
27
15 7
290 1
117 4
37 1
41 2
28 34
19
16 0
53 0
91 2
28 8
54 8
22 56
73
34 3
1637 3
257 4
44 3
106 3
34 35
154
70 9
3391 1
2961 4
100 9
350 7
96 44
65
31 5
736 9
356 3
40 1
328 6
35 21
150
33 7
1187 4
1926 1
58 3
255 3
55 14
6
3
9
6
9
8
5
7
t~
10 ~
102
10 3
10 S
2
6
9
124
p2
60
47
99
50
1
3
4
45
80
75
32
50
26
6 3
0
10
10
2 1
5
10
85
2 2
l
10
25
2 3
0
10
10
2 4
0
6
6
2 b
~)
Z
2
! 6
1
0
15
2 7
I
I
16
P7
58
19 2
877 5
641 8
86 2
2594 2
47 54
#)8
P9
[)10
109
36 1
9257 4
2971 5
70 2
1407 1
60 33
64
56 9
702 8
13564 6
427 7
9284 4
65 34
83
102 4
751 6
189 2 0 8
469 0
5852 1
238 54
Case
6
6
6
6
10 4
1
0
15
Case
Table 4
10 3
0
10
10
Eng. Struct.
1993,
Vol.
15, No
p3
6
30
104
1131
P5
23
85
88
O0
5
44
45
937
p7
80
94
72
55
404
51 38
63 72
282 13
,O9
3
47
17
801
44
33
26
25
PtIF(d/?leler estlmaHo/1
and a( curacv
(16)
Lmm = 1 . 8 q E - 0 4
Lma x = 1.76
DL = 0,03
0.5
<~
0.5
<
o.o -
-0.5
(0.85,1.13)
I
-0.5
0.5
0.0
-0.5
('I-X"1)
Figure 4
0.0 -
0.0
0.5
(I-~)
Since the interpretative model (14) used in the identification is approximated, it ~s not possible to match exactly
the data z; m this sense the situauon is similar to the real
one, where analytical and experimental errors occur.
The optimal values of the parameters are obtamed by
minimizing the objecttve function which, according to
the assumptions previously introduced, has the
slmphfied form
m
t(x)= ~ 1/o~,,[z,-fi,(x)]-"
(17)
(18)
Eng. Struct.
1993,
Vol. 15, No 1
27
FE m o d e l s m s t r u c t u r a l d y n a m i c s
Capecchl and F
Vestrom
Tahle 5
Case
0 80
1 12
0 80
1 15
0 75
1 20
1 15
1 20
0 75
0 80
e~(ol
~J Iu,,)
101
10 2
10 3
104
10 5
10 6
10 7
087
0 81
0 77
084
0 83
0 84
0 83
110
1 16
0 97
112
1 14
1 13
1 14
079
0 81
0 58
078
0 79
0 79
0 79
116
1 14
1 03
103
1 17
1 17
1 12
075
0 75
4 91
073
0 75
0 74
0 74
121
1 26
6 87
111
1 22
1 22
1 20
114
1 12
3 68
120
1 18
1 18
1 14
117
1 13
1 78
104
1 22
1 20
1 18
077
0 84
0 80
030
0 90
0 80
0 77
079
0 91
0 67
161
0 54
0 62
0 76
317
6 30
2 0 0 O0
3770
12 2 0
7 80
2 45
1410
8110
3110
1610
3410
6910
1610
Table 6
Case
0 80
1 12
0 80
0 75
1 15
0 75
e,(%)
6 1
6 2
63
64
087
0 83
084
085
1
1
1
1
080
0 81
1 01
057
075
0 76
073
092
1 13
1 10
097
1 16
079
086
086
089
436
6 50
14 11
1697
Table 7
10
12
12
10
4
:
e,(%)
1410
8610
9110
1710
T
2
2
5
Case
0 80
1 15
0 75
1 20
1 15
1 20
0 75
0 80
81
82
8 3
080
079
1 18
1 16
1 15
0 59
075
076
0 66
1 22
1 23
1 41
1 16
1 15
0 91
1 18
1 17
1 53
074
076
0 81
078
079
0 72
- -
X 1
--
X 5
------
X2
--
--
X7
......
X3
m.~
1.5
1.5
X9
1.0
0 S
_.
-. . . .
-_;
._
. .
_.
o.5
o
10
e,(%)
1 29
1 54
28 24
7 0 10
2 3 10
9 5 10
----
X5
------
x2
--._
x7
. .
x3
---~
x9
- \ ? ' ,~
~1.0
n,r
Figure5
e~(%)
X1
\;'.~.
\\
\
28
2
4
I0
hi(
2
2
3
10
(20)
S = ]/10
(21)
Conclusions
The accuracy of a model can be improved by using
experimental data referred to an ident~ficauon procedure. When a complicated system is being considered
and the experimental information is limited, it is convenient to adopt a finite element model which is potentially
able to correctly describe the behaviour of the structure,
Table 8
Case
101
xm
102
xm
p,/pn
p,/pn
physical quantities are assumed as parameters to be ~denufied. This allows most of the a priori knowledge of the
structural behaviour to be retained
The use of a fimte element model m identification procedures IS discussed herein. The problem of the selection of parameters and measures was tackled first For
the former, spectral analysis of the Fisher matrix furnlshed a set of independent parameters, ranking them in
decreasing order of importance The optimal set of
measures was then determined as that which minimizes
the trace of the covarlance a posteriorl matrix of
parameters.
When finite element models are used and mechanical
constants are assumed as parameters, it is generally
impossible to estabhsh a priori ldentlfiabillty conditions.
In this context useful reformation is obtained by the
covarlance matrix S evaluated in the initial estimate x0.
A numerical study to identify the model of a framed
spatial structure using different sets of pseudoexperimental data was developed To reduce the
onerousness of the ident~ficauon procedure an approximate polynomial model that fits the finite element
model in the netghbourhood of the base parameter
values was adopted Model identification has been performed in several cases with a different number of
parameters and measurements. The results were always
satisfactory with a high number of data and no notable
alteration IS observed due to the approximation of the
polynomial model used The goodness of estimation
improves when the sensitivity of the observed quantities
with respect to the different parameters is of the same
order.
When the number of data is only strictly sufficient for
identifiabmlity, the accuracy of the estimation is very
poor, mainly in cases with a higher number of
parameters_ This is because the class of interpretative
models does not contain the true model of a structure,
which is to a certain extent a typical real condition.
The estimation accuracy was derived from the
covarlance matrix S of parameters that showed a high
dispersion of results when the number of experimental
data was reduced A simulation process was therefore
performed to investigate th~s important aspect. The
results obtained are in accord with the conclusions
derived from the covarlance matrix, the identification
problem is likely to be really ill condmoned and only a
large quantity of experimental data can guarantee a
satisfactory level of accuracy
Acknowledgment
This research was partially supported by Italian Mlmstry
of Education, funds MPI-40% 1988. The collaboration
of Dr G. Sllvano IS gratefully acknowledged
0800
316
1 119
728
0799
616
1 155
11 48
0750
700
1 193
1740
1 147
9_05
1 195
11 78
0750
678
0799
1005
0802
1752
1 119
2804
0804
3596
1 151
6533
0753
1396
1 198
2227
1 133
5096
1.205
3768
0759
3427
0787
95.36
29
References
I
2
3
4
5
6
8
9
10
11
12
13
14
15
30
Appendix
Let normahzed measures z' and sensitlwty matrix H ' be
introduced
Z' = S,~-I/2 [z - h(x0)]
H' = S,?I'2H
Smce
H ' r H ' = A = V d i a g [;~,]V r
Xo) = F ~
according to which the component o f z' along the direction o f vector U, depends only on one parameter c~,