Академический Документы
Профессиональный Документы
Культура Документы
Model Updating
John E. Mottershead, Michael Link, Tiago A.N. Silva, Yves Govers
and Hamed Haddad Khodaparast
Abstract Probabilistic and interval model updating methods are described, with
particular attention paid to variability in nominally identical test structures due, for
example, to the effect of accumulated manufacturing tolerances, or degradation of
performance caused by wear of engineering components. In such cases the updating
parameter distributions are meaningful physically either as PDFs or as intervals.
Stochastic model updating is an inverse problem, generally requiring multiple
forward solutions, which may be carried out very efciently by the use of surrogates, in place of full FE models. The procedure is illustrated by experimental
examples, including model updating of (i) a frame structure with uncertain locations
of two internal beams and (ii) the DLR AIRMOD structure, which displays
vibration characteristics very similar to those of a real aircraft.
Keywords Model updating
1 Introduction
Deterministic model updating of nite element models [14] has become a mature
technology. It is a classical inverse problem in the sense that a measurable output is
used to correct a set of analytical parameters that are themselves inaccessible to
J.E. Mottershead (&)
Centre for Engineering Dynamics, University of Liverpool, Liverpool L69 3GH, UK
e-mail: j.e.mottershead@liverpool.ac.uk
M. Link
Institute for Statics and Dynamics, University of Kassel, 34109 Kassel, Germany
T.A.N. Silva
LAETA, IDMEC, Instituto Superior Tcnico, Universidade de Lisboa, Lisboa, Portugal
Y. Govers
German Aerospace Centre (DLR), Institute of Aeroelasticity, 37073 Gttingen, Germany
H.H. Khodaparast
College of Engineering, University of Swansea, Swansea SA2 8PP, UK
Springer International Publishing Switzerland 2015
J.K. Sinha (ed.), Vibration Engineering and Technology of Machinery,
Mechanisms and Machine Science 23, DOI 10.1007/978-3-319-09918-7_5
65
66
measurement. As with most inverse problems the resulting system of equations are
ill-posed and require regularisation, discussed in full by Ahmadian et al. [5].
Variability in performance has become one of the major challenges facing
researchers and engineering scientists concerned with the robust reliability of engineering systems. This has resulted in the emergence of powerful probabilistic and
non-probabilistic (interval, fuzzy etc.) methods combined with modern computer
systems and codes. An aspect of uncertainty of particular interest is the inherent
variability that arises from the effect of accumulated manufacturing tolerances or
from wear so that nominally identical engineering systems behave differently.
The problem of variability in the dynamics of nominally identical test structures
was addressed by Mares et al. [6, 7] using a multivariate gradient-regression
approach combined with a minimum variance estimator. Haddad Khodaparast et al.
[8] developed a rst-order perturbation approach whereas the method of Hua et al.
[9] required the computation of second-order sensitivities. Govers and Link [10]
dened an objective function for the identication of updating-parameter covariances with forward propagation of parameters at each iteration. This approach was
modied by Rui et al. [11] who used a polynomial chaos expansion (PCE) as a
surrogate for the full nite element model. Fang et al. [12] used the analysis of
variance (ANOVA) for the selection of updating parameters together with MonteCarlo simulation (MSC). Adhikari and Friswell [13] used a sensitivity approach and
the Karhunen-Love expansion to represent distributed parameters.
Non-probabilistic model updating methods were developed by Haddad
Khodaparast et al. [14] and Erdogan and Bakir [15] using interval and fuzzyarithmetic-based procedures. In the former, a Kriging predictor is used as a surrogate and, in the latter, the membership functions of each updating parameter is
determined by minimising an objective function using a genetic algorithm.
This paper gives a brief overview of probabilistic and interval updating methods
based on rst-order output sensitivities. This is followed by experimental example
problems designed to illustrate the application of stochastic model updating
methods, including the updating of geometrical conguration, made possible by the
use of a trained surrogate model. Finally the example of probabilistic and interval
updating of the DLR AIRMOD structure is presented using modal data obtained by
multiple disassembly and reassembly of bolted jointsresulting in variability due
principally to slight differences in joint tightness in each of the separate tests.
67
model with parameters hj and j denotes the iteration index. The updated vector of
parameters is then given by,
hj1 hj Tj zm zj
where the transformation matrix Tj is generally the weighted pseudo inverse of the
sensitivity matrix Sj .
In the following analysis we consider a cloud of data points in multi-dimensional
space so that each point represents a separate test. There is a corresponding cloud of
predicted points and the objective is to adjust the statistics of the parameters so that the
position and orientation of the cloud, determined according to the mean values and
covariances, or alternatively the upper and lower bounds, of the predictions, is made to
agree with the data cloud.
where the overbar and D denote the mean and variability on the mean respectively.
Then by separating the zeroth-order and rst-order terms, Khodaparast et al. [8]
obtained the following two expressions:
O D0 :
O D1 :
hj1 hj T
j zm zj
j Dzm Dzj DTj zm zj
Dhj1 Dhj T
4
5
Equation (4) is the updating equation for the parameter mean values, exactly the
same as Eq. (12) given by Govers and Link [10].
Using assumptions of (i) small parameter variability, (ii) neglecting the third
right-hand-side term of Eq. (5) and (iii) statistical independence of the updated
parameters Dhj (and hence the predictions Dzj ) from the measurements Dzm , Silva
et al. [16] showed three previously developed expressions for the updated covariance Cov Dhj1 ; Dhj1 to be equivalent. The rst of the three expressions, due to
Khodaparast et al. [8] is expressed as,
68
T
Cov Dhj1 ; Dhj1 Cov Dhj ; Dhj Cov Dhj ; Dzj T
j
T
T
j CovDzm ; Dzm T
T
Tj Cov Dzj ; Dhj Tj Cov Dzj ; Dzj T
j
j
6
The second expression, developed by Govers and Link [10] using the Frobenius
norm for the minimisation of the difference between measured and analytical output
covariances, is given by,
T
j CovDzm ; Dzm T
Cov Dhj1 ; Dhj1 Cov Dhj ; Dhj T
j
T
j Cov Dzj ; Dzj T
T
in both Eqs. (6) and
;
Dz
Forward propagation
is
required
to
determine
Cov
Dz
j
j
(7), and Cov Dhj ; Dzj and Cov Dzj ; Dhj in Eq. (6). This is achieved by application
of the relationships,
T
j Cov Dhj ; Dhj S
Cov Dzj ; Dzj S
j
T
Cov Dhj ; Dzj Cov Dhj ; Dhj S
j
j Cov Dhj ; Dhj
Cov Dzj ; Dhj S
10
However, direct substitution of Eqs. (8), (9) and (10) into either Eqs. (6) or (7)
leads immediately to the third expression, given for example by Hart [17], page 47
(for the general case of linearly related sets of random variables),
T
j CovDzm ; Dzm T
Cov Dhj1 ; Dhj1 T
j
11
69
terms of upper and lower bounds on the updating parameters. The interval model
updating problem is formally described as,
~j ~hj1 ~hj
~zm ~zj S
12
where ~ represents an interval vector or matrix. An advantage of the meta-model
approach, developed by Haddad Khodaparast et al. [14] is that it avoids the use of
interval arithmetic, which is known to be unduly conservative. The accuracy of the
technique depends on the type of meta-model, the sampling used and the behaviour
of the outputs within the range of parameter variation. The Kriging meta-model was
used due to its excellent performance in the prediction of FE model behaviour
(particularly with non-smooth behaviour) and also the provision of a probabilistic
interpretation of the residual between the outputs of the FE model and the surrogate
itself. The interval model updating procedure is illustrated in Fig. 1. Firstly, the
mean values of updating parameters are found by deterministic model updating
using the mean values of measured data. If the solution is unique, the vector of
updated parameter mean values can be represented by a point in the parameter
space. In the next step, an initial hypercube around this point is constructed and the
meta-model (the Kriging predictor in the present case) is then used to map the space
of the initial hypercube of updating parameters to the space of outputs. Care should
be taken to make sure that the mapping is good enough to represent the relationship
between inputs and outputs with sufcient accuracy. This can be achieved by
increasing the number of training samples until the Mean Squared Error (MSE) at
an unsampled point falls below a threshold, when the Kriging model is deemed to
be accurate enough. Equation 6 may then be solved by taking all the measured data,
represented by circles in Fig. 1, and mapping them in the reverse direction (using
the Kriging predictor) to nd the vertices of the updated parameter hypercube. The
Fig. 1 Interval model updating using the Kriging predictor (Reprinted from [14] with permission
from Elsevier)
70
need for a highly computationally efcient meta-model is illustrated by the fact that
a large number of inversions are required for interval model updating. Further
details on the mathematical development of the method are provided by Haddad
Khodaparast et al. [14].
5 Experimental Examples
Stochastic model updating procedures and their effectiveness are illustrated by two
experimental examples, both of which are designed to reproduce the conditions that
would be experienced in the case of multiple nominally-identical structures each with
variability arising from, for example, manufacturing tolerances or degradation due to
wear. The objective is to estimate the spread of uncertainty in parameters deemed to
be responsible for observed variability in dynamic behaviour from modal tests.
In the rst example, a frame structure is tested nine times to determine the
natural frequencies. In each of the nine tests, two internal beams were placed in
different congurations and intervals were determined on the extreme positions of
the two beams. The use of a meta-model enabled the parameterization of beam
locations, which could not have been done by conventional model updating without
re-meshing of the FE model at each iteration.
The second example is the DLR AIRMOD structure, a replica of the GARTEUR
SM-AG19 benchmark described by Balmes [18]. Probabilistic and interval model
updating was carried out using data selected from 130 disassemblies and reassemblies of the structure with a modal test carried out at each stage. Details of the
test procedure, carried out at DLR Gttingen, can be found in [19, 20].
71
Fig. 2 Test structure and FE model (Reprinted from [14] with permission from Elsevier). a Test
structure, b Beam locations, c FE Model
1.0
1.0
60
60
3.7
2.0
1.0
2.0
60
20
0.2
7.6
1.0
3.0
60
20
0.2
2.8
2.0
1.0
20
60
1.8
9.8
2.0
2.0
20
20
6.5
0.1
2.0
3.0
20
20
2.4
3.0
3.0
1.0
20
60
0.6
11.0
3.0
2.0
20
20
0.3
8.4
3.0
3.0
20
20
2.2
0.6
1
2
3
4
5
6
Measured
Initial FE
Updated FE
Initial FE error %
Updated FE error %
22.524.3
24.427.8
47.149.9
74.481.1
219.5256.4
299.7312.4
21.624.6
23.735.5
43.767.6
71.182.5
224.1267.3
300.3339.7
22.624.6
23.927.5
45.150.6
74.081.4
224.1259.5
303.6317.2
[4.1, 1.1]
[3.0, 27.6]
[7.2, 35.6]
[4.4, 1.6]
[2.1, 4.3]
[0.2, 8.7]
[0.1, 1.1]
[2.1, 1.3]
[4.2, 1.4]
[0.5, 0.2]
[2.1, 1.2]
[1.3, 1.5]
72
Fig. 3 Space of rst and second natural frequencies (Reprinted from [14] with permission from
Elsevier). a Initial space, b Updated space
Fig. 4 AIRMOD structure and FE model. a Structure, b FE Model (Reprinted from [21] with
permission from Elsevier)
(a)
(b)
VTP/HTP
VTP/Fuselage
(c)
(d)
Wing/Fuselage
Winglet
73
0.23
fFE_ini/Hz
0.17
fFE_ini/%
fFE_upd/%
0.01
0.01
RBM Roll
0.65
0.56
RBM pitch
0.83
0.82
1.00
0.00
RBM Heave
2.17
2.14
1.62
0.02
2nWingBending
5.50
5.65
2.78
0.40
0.01
3nWingBending
14.91
15.11
1.35
WingTorsionAnti
31.96
33.31
4.25
0.25
WingTorsionSym
32.33
33.62
3.98
0.42
VtpBending
34.38
35.39
2.94
1.14
4nWingBending
43.89
44.66
1.77
0.08
1nWingForeAft
46.71
47.21
1.08
0.02
2nWingForeAft
51.88
52.91
1.99
0.05
5nWingBending
58.59
60.59
3.43
1.29
VtpTorsion
65.93
67.69
2.67
0.05
2nFuseLat
100.05
102.59
2.55
2.07
2nVtpBending
124.56
128.62
3.26
1.48
6nWingBending
129.38
132.08
2.08
0.14
7nWingBending
141.47
145.91
3.14
0.76
2nHtpBending
205.59
206.73
0.56
0.04
HtpForeAft
219.07
225.73
3.04
0.11
WingBendingRight
254.73
261.53
2.67
0.02
WingBendingLeft
255.02
262.64
2.99
0.32
3nWingForeAft
272.08
278.71
2.44
1.44
WingletBendingLeft
303.96
320.15
5.33
2.24
WingletBendingRight
304.32
321.64
5.69
2.35
3nFuseLat
313.68
324.12
3.33
2.39
WingTorsionSym2
328.55
336.31
2.36
0.60
WingTorsionAnti2
331.18
341.15
3.01
0.71
4nWingForeAft
336.21
343.55
2.18
0.15
2nFuseVert
348.68
359.54
3.12
1.77
74
Fig. 6 Updated parameter intervals: Changes with respect to the initial nite element model
(Reprinted from [21] with permission from Elsevier)
0.46 m. The total weight of the structure is 44 kg. Two additional masses of 167
grams each are installed at the forward tips of the winglets to ensure better excitation
of the wing torsion modes. The complete FE model shown in Fig. 4b, including the
main structure, cables, sensors and bungee cords consists of 1440 CHEXA, 6
CPENTA and 561 CELAS1, 55 CMASS1, 18 CONM2 and 3 CROD elements. The
structural behaviour of AIRMOD shows similarities to the dynamical characteristics
of a real aircraft, including mode shapes that show either symmetric or antisymmetric
deformations and closely space modes are found at 32, 255 and 330 Hz.
Modal testing of the AIRMOD structure and assessment of sources of uncertainty is discussed in detail in a series of papers by Govers and Link [10, 19, 20].
Model updating using both probabilistic and interval methods is described in full by
Govers et al. [21]. The means of the parameters were estimated by deterministic
model updating of 18 parameters using both the eigenvalue and mode-shape
residuals of the 14 active modes shown in Table 3. It is seen from in the table that
the mean frequency values of the updated model match perfectly with the test data
for the active mode set. Also, the passive modes (not included in the residual) are
signicantly improved, thereby providing validation of the updated FE model.
The parameter covariances and intervals were estimated using the natural frequencies of the same 14 modes (the mode shapes were not included) and results are
shown in Fig. 6. Three standard deviations are shown in the case of covariance
updating.
Updated covariance and interval results for the 14 natural frequencies are presented in Fig. 7. The size and orientation of the 105 frequency clouds show a very
good agreement between test (red) and updated analysis clouds (green) for the active
range. The interval-updated regions (black) are generally greater than those calculated by the covariance method. This is to be expected because the number of terms
that dene the multi-variate probability distributions on the updating parameters (the
means and fully populated covariance matrix) is greater than the set of upper- and
lower-bounds that dene the interval model. The probabilistic method starts with the
diagonal covariance matrix at the rst iteration (assuming uncorrelated updated
parameters), but becomes a fully populated matrix after convergence. The interval
75
Fig. 7 Frequency distributions from updated models (Reprinted from [21] with permission from
Elsevier)
76
method makes no assumption about the correlation of the updating parameters, which
are therefore conned to a hypercube, whereas the updated fully-populated covariance matrix denes a space of hyper-elliptic parameters. The hypercube creates a
greater sample space than that of the hyper-ellipse. Another observation from Fig. 6 is
that the degree of overestimation of the interval results is not of the same order for all
the frequencies, as can be seen by comparing plane f8-f10 to plane f1-f2. One may
relate this to the sensitivity of the natural frequencies with respect to the parameters
which are located outside the hyper-ellipse but inside the hypercube. In other words,
the greater the sensitivities of the frequencies to parameter variation, the higher will
be the degree of over-estimation seen in the results.
6 Conclusions
An overview of probabilistic and interval methods in nite element model updating
is presented. The study is focused on the problem of multiple nominally-identical
test structures where the problem is to quantify the certainty in parameters deemed
responsible for observed variability in dynamic behavior. The overview is
accompanied by two experimental example problems designed to demonstrate the
working of the techniques involved.
References
1. Natke HG (1992) Einfhrung in Theorie und Praxis der Zeitreihen und Modalanalyse. Vieweg
Verlag, Braunschweig/Wiesbaden, Germany
2. Mottershead JE, Friswell MI (1993) Model updating in structural dynamics: a survey. J Sound
Vib 167(2):347375
3. Friswell MI, Mottershead JE (1995) Finite element model updating in structural dynamics.
Kluwer Academic Publishers, Dordrecht
4. Mottershead JE, Link M, Friswell MI (2011) The sensitivity method in nite element model
updating: a tutorial. Mech Syst Signal Process 25(7):22752296
5. Ahmadian H, Mottershead JE, Friswell MI (1998) Regularisation methods for nite element
model updating. Mech Syst Signal Process 12(1):4764
6. Mares C, Mottershead JE, Friswell MI (2006) Stochastic model updating: part 1- theory and
simulated examples. Mech Syst Signal Process 20(7):16741695
7. Mottershead JE, Mares C, James S, Friswell MI (2006) Stochastic model updating: part 2application to a set of physical structures. Mech Syst Signal Process 20(8):21712185
8. Haddad Khodaparast H, Mottershead JE, Friswell MI (2008) Perturbation methods for the
estimation of parameter variability in stochastic model updating. Mech Syst Signal Proc 22
(8):17511773
9. Hua XG, Ni YQ, Chen ZQ, Ko JM (2008) An improved perturbation method for stochastic
nite element model updating. Int J Numer Meth Eng 73:18451864
10. Govers Y, Link M (2010) Stochastic model updatingcovariance matrix adjustment from
uncertain experimental modal data. Mech Syst Signal Process 24(3):696706
77
11. Rui Q, Ouyang H, Wang HY (2013) An efcient statistically equivalent reduced method on
stochastic model updating. Appl Math Model 37:60796096
12. Fang S-E, Ren W-X, Perera R (2012) A stochastic model updating method for parameter
variability quantication based on response surface models and Monte-Carlo simulation. Mech
Syst Signal Proc 33:8396
13. Adhikari S, Friswell MI (2010) Distributed parameter model updating using the KarhunenLove expansion. Mech Syst Signal Process 24:326339
14. Haddad Khodaparast H, Mottershead JE, Badcock KJ (2011) Interval model updating with
irreducible uncertainty using the Kriging predictor. Mech Syst Signal Proc 25(4):12041226
15. Erdogan YS, Bakir PG (2013) Inverse propagation of uncertainties in nite element model
updating through use of fuzzy arithmetic. Eng Appl Artif Intell 26:357367
16. Silva TAN, Mottershead JE, Link M, Maia NMM Updating the parameter covariance matrix
of a nite element model, in preparation
17. Hart GC (1982) Uncertainty analysis, loads and safety in structural engineering. Prentice-Hall,
Englewood Cliffs
18. Balmes E (1997) Garteur group on ground vibration testing: results from the test of a single
structure by 12 laboratories in Europe, proceedings of international modal analysis conference,
1997, pp 13461352
19. Govers Y, Link M (2010) Stochastic model updating of an aircraft like structure by parameter
covariance matrix adjustment, proceedings of international conference on noise and vibration
engineering, Leuven, Belgium, 2010, pp 26392656
20. Govers Y, Link M (2012) Using stochastic experimental modal data for identifying stochastic
nite element parameters of the AIRMOD benchmark structure, proceedings of the
international conference on noise and vibration engineering, USD2012, Leuven, Belgium,
2012, pp 46974715
21. Govers Y, Haddad Khodaparast H, Link M, Mottershead JE (2014) A comparison of two
stochastic model updating methods using the DLR AIRMOD test structure, mechanical
systems and signal processing, http://dx.doi.org/10.1016/j.ymssp.2014.06.003