Вы находитесь на странице: 1из 18

Mechanical Systems and Signal Processing 32 (2012) 135–152

Contents lists available at SciVerse ScienceDirect

Mechanical Systems and Signal Processing


journal homepage: www.elsevier.com/locate/ymssp

Parameter selection and stochastic model updating using


perturbation methods with parameter weighting matrix assignment
Nurulakmar Abu Husain a,n, Hamed Haddad Khodaparast b, Huajiang Ouyang b
a
Transportation Research Alliance, Universiti Teknologi Malaysia, 81310 UTM Johor Bahru, Johor, Malaysia
b
School of Engineering, University of Liverpool, Liverpool L69 3GH, United Kingdom

a r t i c l e i n f o abstract

Article history: Parameterisation in stochastic problems is a major issue in real applications.


Received 16 June 2011 In addition, complexity of test structures (for example, those assembled through laser
Received in revised form spot welds) is another challenge. The objective of this paper is two-fold: (1) stochastic
9 March 2012
uncertainty in two sets of different structures (i.e., simple flat plates, and more
Accepted 1 April 2012
Available online 20 April 2012
complicated formed structures) is investigated to observe how updating can be
adequately performed using the perturbation method, and (2) stochastic uncertainty
Keywords: in a set of welded structures is studied by using two parameter weighting matrix
Uncertainty approaches. Different combinations of parameters are explored in the first part; it is
Variability
found that geometrical features alone cannot converge the predicted outputs to the
Stochastic model updating
measured counterparts, hence material properties must be included in the updating
Parameterisation
Spot weld process. In the second part, statistical properties of experimental data are considered
Welded structure and updating parameters are treated as random variables. Two weighting approaches
are compared; results from one of the approaches are in very good agreement with the
experimental data and excellent correlation between the predicted and measured
covariances of the outputs is achieved. It is concluded that proper selection of
parameters in solving stochastic updating problems is crucial. Furthermore, appropriate
weighting must be used in order to obtain excellent convergence between the predicted
mean natural frequencies and their measured data.
& 2012 Published by Elsevier Ltd.

1. Introduction

Interest in uncertainty in engineering structures is growing, owing to the fact that structural properties are normally
uncertain and therefore uncertainty also exists in the dynamic response. Properties of an individual structure normally
change with time due to environmental erosion and damage [1,2], and also when the structure is being reassembled. It is
also unavoidable to have manufacturing variability [3] that exists among nominally identical structures, built in the same
way from the same materials, such as in a mass production of automotive body-in-white (BIW).
Uncertainty can originate from many sources, such as lack of knowledge, physical randomness of structures, inaccurate
information, etc. It may be classified into two categories [4,5]: epistemic and aleatory uncertainties, based on whether the
source of uncertainties is reducible or not. Quantification of uncertainty in outputs of a model with uncertain inputs is
generally referred to as uncertainty propagation, which can be categorised into three main categories: sampling methods
[6,7], response surface approximation (RSA) methods [8], and convex methods [9]. The Monte Carlo simulation (MCS)

n
Corresponding author. Tel.: þ60 7 5534665; fax: þ 60 7 5566159.
E-mail address: nurul@fkm.utm.my (N. Abu Husain).

0888-3270/$ - see front matter & 2012 Published by Elsevier Ltd.


http://dx.doi.org/10.1016/j.ymssp.2012.04.001
136 N. Abu Husain et al. / Mechanical Systems and Signal Processing 32 (2012) 135–152

method and its variants are examples of the sampling methods, while the perturbation and the interval methods belong to
the RSA and convex approaches, respectively. Polynomial chaos (PC) methods (which are a kind of RSA method but require
sampling in order to find the coefficients of polynomial and hence lead to reduced-order models) have also been widely
used in uncertainty quantification over the past decade [10,11].
Model updating of non-deterministic problems arising from uncertainties in the system requires iteration on forward
propagation method. Selection of the propagation method is mostly dependent on the representation of the uncertainty
(using various mathematical tools, including probability theory, interval arithmetic and fuzzy set theory, just to name a
few [12–15]). Moens and Vandepitte [16] discussed general theoretical and practical aspects of emerging non-probabilistic
approaches for uncertainty treatment in FE analysis, while Worden et al. [17] discussed some issues relating to uncertainty
analysis methods of nonlinear systems. Adhikari and Friswell [18] proposed two methods, namely (a) optimal point
expansion method, and (b) asymptotic moment method, to determine the moments and probability density functions of
eigenvalues in discrete linear stochastic systems. In this paper, the inverse problem of non-deterministic (or stochastic)
updating is formulated based on the mean-centred perturbation theory.
In deterministic problems, all parameters are set to their nominal values and it is assumed that no variability exists;
researchers such as Ahmadian et al. [19], Kim and Park [20], and Mottershead et al. [21] for example published works on
the parameterisation in deterministic model updating. In this paper, investigations on parameter selections for stochastic
updating are performed instead. The stochastic updating work is conducted on two sets of very different structures:
(1) simple flat plates, and (2) more complicated formed structures. It is recognised that many works have been published on
developing suitable methods for stochastic updating, see Section 2 below. In the latter part of this paper, the uncertainty
problem in a set of nominally identical laser welded structures is investigated, where two approaches are studied to
efficiently represent and propagate uncertainty due to manufacturing variability in the welded structures. This paper is a
further development of the perturbation method developed by Haddad Khodaparast et al. [22] for complicated structures in
which parameterisation is a major issue. The stochastic model updating formulation is included and the results are
discussed.

2. Stochastic model updating for uncertainty in structural dynamics

Structural parameters are normally assumed to be known in forward problems, but unknown in inverse problems such
as in model updating [23]. In dynamic predictions of structures, FE model of the structure is usually characterised by a
number of parameters that include elastic, viscoelastic and/or geometrical properties of the structure [24], which can be
categorised into two groups: (1) those that can be directly measured (e.g., thickness and other geometrical data), and
(2) those that cannot be measured directly (e.g., structural joints stiffness, damping, etc.). For the second group of
parameters, it is often easier to measure the variability in structural dynamic characteristics (such as modal properties and
frequency response functions (FRFs)) than measuring the uncertainty in this type of structural parameters. The statistical
estimates from the response measurements may then be used to deduce the statistical estimates of the parameters; this
inverse problem of estimating the parameter distribution from the measured response distribution is called uncertainty
identification or quantification.
Application of the FE model updating is well established for deterministic problems, but due to uncertainties in real test
structures, non-deterministic (or stochastic) model updating has attracted much research recently. The stochastic model
updating method allows for manufacturing variability and modelling uncertainty to be incorporated so that numerical
models with randomised parameters can be updated to match their experimental counterparts [25]. In addition, variability
in measurements can also be considered, as presented by Luczak et al. [26] where different test configurations for dynamic
predictions of fibre reinforced composite structures are taken into account and non-deterministic model updating is
carried out.
Non-deterministic model updating allows for robust and credible models to be produced which in turn increase
confidence in design and analysis of structures. Because of the significant benefits, many papers are published covering the
non-deterministic (or stochastic) model updating approach. A review of the stochastic FE method in various applications
including model updating issues was presented by Stefanou [27], with 196 references cited. Beck and Au [28] treated the
challenging task of model updating as a Bayesian statistical inference. Some other references [29–31] also cover the
stochastic model updating approach, amongst many other publications.
Nevertheless, stochastic model updating is computationally expensive, mainly due to the randomised parameters
involved in the procedure. A few decades ago, sensitivity analysis became a successful approach in assessing the effect of
parameter variations in theoretical models [32,33]. Nevertheless, due to time penalty and high computational effort in
sensitivity analysis, especially when working with randomised multi-parameter investigations and high order derivatives,
it must be performed and utilised efficiently. As an example, a perturbation method reported by Hua et al. [30] was
developed for estimating the statistical properties (i.e., means and covariances) of structural parameters on the basis of the
measured modal parameters with uncertainty. However, the method requires the determination of second-order
sensitivities in order to converge the predicted mean values and covariances upon the measured values. In order to avoid
large computational effort, Haddad Khodaparast and Mottershead [34] investigated two efficient methods in stochastic
model updating: (1) a perturbation method, and (2) a method based upon the minimisation of an objective function. The
first method was shown to be viable and the need to compute second order sensitivities (as reported by Hua et al. [30])
N. Abu Husain et al. / Mechanical Systems and Signal Processing 32 (2012) 135–152 137

was removed leading to a considerable reduction in computational effort in practical engineering applications.
Nevertheless, its range of applications is limited to small degree of uncertainties. Conventional deterministic model updating
is also subjected to this restriction, i.e., the initial starting estimate should be close enough to the true value. Both versions of
the updating method, deterministic and stochastic, are applied iteratively so that the restriction applies within each
iterative step. Moreover, the response should be sensitive to the updating parameters so that only small changes are
encountered. Another study using the perturbation method was presented by Govers and Link [35]. They demonstrated a
method to adjust parameter means and covariance matrix from multiple sets of experimental modal data. The method was
performed by updating mean values of parameters to minimise the difference between the measured and predicted
outputs, followed by updating of parameter covariance matrix, where the difference between the measured and analytical
output covariance matrices is minimised. Other methods, such as the Kriging method [36] and a combination of Latin
Hypercube Sampling and Neural Networks method [37], have also been studied by researchers in order to reduce the
computational effort of stochastic updating.
The perturbation method presented by Haddad Khodaparast et al. [22,34] is employed in this research for the stochastic model
updating, as explained further in the following sections. The main challenge is the application of the method to fairly complicated,
assembled structures. Stochastic updating of structures of such complexity has not been reported in the available literature.

2.1. Formulation of the perturbation method for stochastic uncertainty

The stochastic model updating equation is represented as follows [22,34]:


h^ j þ 1 þ Dhj þ 1 ¼ h^ j þ Dhj þðT^ j þ DTj Þðz^ m þ Dzm z^ j Dzj Þ ð1Þ

where h is the vector of parameters, z is the vector of output data, ^ denotes the mean values, while subscripts j for h, z and
T denote the parameters, predicted output data and transformation matrix at jth iteration. Similarly, subscripts (j þ1) for h,
z and T denote the parameters, predicted output data and transformation matrix at (j þ1)th iteration. Subscript m indicates
measured data and D denotes random terms.
The transformation matrix T may be expressed as
T T
T^ j ¼ ðS^ j Wee S^ j þWyy Þ1 S^ j Wee ð2Þ

which is comparable to the transformation used in deterministic model updating, as presented in Ref. [21]. S^ is the
sensitivity matrix evaluated at mean structural parameters, while Wee and Wyy are positive definite weighting matrices of
measured data and parameters, respectively. Wee is usually a diagonal matrix with elements equal to the reciprocals of
measurements variance. If the measured data are to be reproduced exactly, Wee is unity. On the other hand, some
parameters are usually estimated more accurately than others, especially in a finite element (FE) analysis. Hence, Wyy
needs to be introduced so that parameters that are estimated more accurately (i.e., less uncertain) do not change as much
as parameters with high uncertainty. In practice, this weighting matrix is usually chosen to be a diagonal matrix whose
diagonal elements are given by the reciprocals of estimated variances of the corresponding parameters. It must be
understood that it may be difficult to estimate the variances quantitatively, thus some engineering insight is required.
Alternatively, the parameters may be assumed to be equally uncertain.
Separating the zeroth-order and first-order terms from Eq. (1) gives
D0 : h^ j þ 1 ¼ h^ j þ T^ j ðz^ m z^ j Þ ð3Þ

D1 : Dhj þ 1 ¼ Dhj þ T^ j ðDzm Dzj Þ þ DTj ðz^ m z^ j Þ ð4Þ

Eqs. (3) and (4) are used to determine the parameter means and the parameter covariance matrix, respectively, in the
perturbation method. Assuming that measurements and parameters are uncorrelated, the parameter covariance matrix
equation can then be written as
T T T
Cyyj þ 1 ¼ Cyyj CyZj T^ j þ T^ j CEE T^ j T^ j CZyj þ T^ j CZZ j T^ j ð5Þ

where Cyy is the parameters covariance matrix, CEE is the covariance matrix of the measured outputs, CZZ is the covariance
matrix of the predicted outputs, CyZ is the covariance matrix of the parameters and the predicted outputs, and CZy is the
covariance matrix of the predicted outputs and the parameters. The covariance between two random variables is
computed using the mean-centred first order perturbation method [18], CEE is determined from the measured data, CZZ is
T
obtained by S^  Cyy  S^ where the assignment of Cyy relies on the engineering insight of the analyst. It is important to note
that the random variables for the input parameters are independent of one another in the first iteration, but they become
correlated in the consecutive iterations. Nevertheless, the dependencies of the parameters after convergence do not have
any physical meaning, thus only the diagonal terms of the updated covariance matrix are used in the algorithm.
A procedure outlined by Haddad Khodaparast et al. [22] is followed to determine the statistical data (i.e., means and
standard deviations) of the structural parameters that converge: (1) the mean predicted modal data (z^ ) on the mean
measured modal data (z^ m ), and (2) the covariance matrix of the predicted modal data (CZZ) on the covariance matrix of the
measured modal data (CEE). As mentioned earlier, a significant advantage of the perturbation method used here over
138 N. Abu Husain et al. / Mechanical Systems and Signal Processing 32 (2012) 135–152

another similar perturbation method by Hua et al. [30] is that only the first-order sensitivity matrix is needed in Eq. (5),
hence a big reduction in terms of computational effort is achieved. The Monte Carlo Simulation method with a multivariate
normal distribution of the input parameters is used to generate scatter plots of analytical output variables, before and after
updating is performed. However, it is important to note that the assumption of multivariate normal distribution is not a
requirement for the method and any other assumptions may be used. In the perturbation method used in this work (i.e., a
mean-centred perturbation method), only the first two statistical moments—the mean and standard deviations are
involved in the algorithm. If higher statistical moments (e.g. skewness and kurtosis of non-symmetrical distributions) are
to be computed in the updating process, then higher order perturbation methods (e.g. quadratic form) must be used.

2.2. Parameter selection for stochastic model updating

This section explains how parameter selections can be sufficiently made for the stochastic problems. Two studies of
stochastic uncertainty in (simple) flat plates and (more complicated) formed hat-shaped shells are performed to
investigate how updating can be adequately conducted with different combinations of parameters. Experimental data
from 33 samples of both the flat plates and the hat-shaped shells are utilised and the stochastic model updating approach
(Section 2.1) is used for the work. Different combinations of (geometrical and material properties) parameters are tested,
and important findings are then used as a guideline when selecting the parameters for the actual stochastic model
updating of the welded structures (Section 3).

2.2.1. Case 1: parameter selection for stochastic model updating in simple flat plates
Selection of parameters in the stochastic updating is investigated first by using simple structures such as the flat plate
illustrated in Fig. 1. The exercise is performed by using the geometrical and material properties as parameters to determine
the best combination for carrying out the stochastic model updating of such simple structures. Based on existing
measurement data, the covariance matrix of the measured outputs for the first five natural frequencies of the flat plates is
determined as follows:
2 3
0:01 0:03 0:05 0:05 0:09
6 0:06 0:11 0:11 0:19 7
6 7
6 7
CEEplate ¼ 6
6 0:32 0:17 0:54 7
7
6 7
4 sym: 0:18 0:31 5
0:95

Firstly, only the geometrical features of the structures, i.e., the thickness of the plates are studied. Three FE models are
developed for the flat plates as listed in the following:

1. FE model of flat plates with one thickness (i.e., original model) (FP1)
2. FE model of flat plates with three thicknesses (FP3)
3. FE model of flat plates with six thicknesses (FP6)

Initial values of 1.45 mm and standard deviations of 1.45  10  2 mm are assigned for the thickness of each region in all
three models. Next, the material properties of the flat plates are included as the updating parameters to see the effect of
the parameter selections. The stochastic model updating is then conducted for each model and their results are discussed
in the following sections. Scatter clouds of measured and predicted outputs are generated by the Monte Carlo simulation
following the Gaussian distribution to show convergence of the stochastic updating procedure. Preliminary investigation
was carried out beforehand in order to identify the number of samples needed in the Monte Carlo simulation; it was found
that samples of 500 and more give nearly identical results. Therefore, 500 numerical samples are used in plotting the
scatter clouds throughout this investigation.

Fig. 1. The flat plate.


N. Abu Husain et al. / Mechanical Systems and Signal Processing 32 (2012) 135–152 139

2.2.1.1. Stochastic model updating using FP1. The stochastic updating is conducted on the original model of the flat plates
(as depicted in Fig. 2), and the updated mean natural frequencies and parameters upon convergence are tabulated in
Tables 1 and 2 against their initial values. It can be seen from Table 1 that the updated mean natural frequencies obtained
using the FP1 model deviate from the mean measured natural frequencies by 8.96% in total; the updating procedure fails to
correct the initial FE model as the error after updating is bigger than the initial error. On the other hand, the updated
parameter is brought very close to the initial value, while the estimated standard deviation is one order smaller than the
initial value. Please note that, identification of standard deviations of parameters in the presence of measurement noise is
more difficult than the mean values of parameters; hence, bigger errors can often be found in the identified standard
deviations than the errors in the identified mean values.
Using the identified parameters, the covariance matrix of the predicted outputs (CZZFP1) is determined. Its
percentage of errors to the measured data is illustrated in Fig. 3(a), while Fig. 4 shows the scatter plots of the
measured and predicted outputs before and after the stochastic model updating process. Generally, the errors
between the outputs covariances are very big (up to about 130%) when using the FP1 model, hence it can be concluded
that the approach fails to produce a good estimate of the CZZ. Furthermore, the scatter plots illustrate that the updating
procedure fails to converge the predicted outputs to the measured data, although the size of the updated scatter cloud
is better estimated than the initial.

2.2.1.2. Stochastic model updating using FP3. For this exercise, the FE model of the flat plates is divided into three regions
(Fig. 5), with initial mean natural frequencies and parameters as tabulated in Tables 3 and 4. Upon convergence, the total
error of the mean natural frequencies is reduced to 7.43% as compared to 8.96% when using only one thickness as updating
parameter. Moreover, the identified parameters and their corresponding standard deviations are included in Table 4, with
all updated estimates identified close to their corresponding initial values apart from the standard deviations of T2 which is
one order smaller than the initial.
Using the identified parameters, the covariance matrix of the predicted outputs (CZZFP3) is determined and the
percentage of errors to the measured data is shown in Fig. 3(b). It can be observed that having more thickness parameters
reduces the overall error of the output covariance matrix, however the maximum error of the covariance for the FP3 model
is still very big (i.e., up to 60%). In addition to that, the scatter plots of the measured and predicted outputs before and after
the stochastic model updating process also fail to converge, as depicted in Fig. 6.

Fig. 2. The FP1 model.

Table 1
Mean natural frequencies (in Hz) estimated using FP1.

Mode I II III
Experiment Initial Updated

FE Error (%) FE Error (%)

1 24.12 24.27 0.62 24.63 2.13


2 66.92 67.24 0.48 68.24 1.97
3 77.65 75.31 3.01 76.42 1.58
4 131.97 132.51 0.41 134.47 1.89
5 158.80 154.31 2.83 156.59 1.39

Total error 7.35 8.96


140 N. Abu Husain et al. / Mechanical Systems and Signal Processing 32 (2012) 135–152

Table 2
Identified mean parameter for FP1.

Parameter Initial value (mm) Identified value (mm)

T 1.450 1.472

2
Std. (T) 1.45  10 7.18  10  3

Fig. 3. Scatter plots of the FP1 model before and after the stochastic model updating.

Fig. 4. The FP3 model.

Fig. 5. Scatter plots of the FP3 model before and after the stochastic model updating.
N. Abu Husain et al. / Mechanical Systems and Signal Processing 32 (2012) 135–152 141

Table 3
Mean natural frequencies (in Hz) estimated using FP3.

Mode I II III
Experiment Initial Updated

FE Error (%) FE Error (%)

1 24.12 24.27 0.62 24.27 0.61


2 66.92 67.24 0.48 67.32 0.60
3 77.65 75.31 3.01 75.29 3.03
4 131.97 132.51 0.41 132.68 0.53
5 158.80 154.31 2.83 154.57 2.66

Total error 7.35 7.43

Table 4
Identified mean parameter for FP3.

Parameter Initial value (mm) Identified value (mm)

T1 1.450 1.453
T2 1.450 1.449
T3 1.450 1.453

Std. (T1) 1.45  10  2 1.13  10  2


Std. (T2) 1.45  10  2 7.31  10  3
Std. (T3) 1.45  10  2 1.13  10  2

Fig. 6. The FP6 model.

Fig. 7. Scatter plots of the FP6 model before and after the stochastic model updating.
142 N. Abu Husain et al. / Mechanical Systems and Signal Processing 32 (2012) 135–152

2.2.1.3. Stochastic model updating using FP6. The FE model of the flat plates is divided into six regions (see Fig. 7) to
investigate if having more parameters would improve the predicted results. The initial and updated values of the mean
natural frequencies and parameters are presented in Tables 5 and 6, which do not show much improvement from the FP3
results. The mean natural frequencies and the identified mean parameters are exactly the same as predicted by the FP3
model, while their corresponding standard deviations are quite close to their initial values, except for T2 and T5.
Using the identified parameters, the covariance matrix of the predicted outputs (CZZFP6) is calculated and the percentage
of errors to the measured data is shown in Fig. 3(c). The scatter clouds are illustrated in Fig. 8. From the results, it may be
concluded that having more thickness parameters would not improve the stochastic updating results at all. It is believed
the main reason is that the actual deviation and variation of thickness from the mean values in the tested specimens are
very small and the discrepancy between the predicted frequencies and measured frequencies is not due to thickness.

Table 5
Mean natural frequencies (in Hz) estimated using FP6.

Mode I II III
Experiment Initial Updated

FE Error (%) FE Error (%)

1 24.12 24.27 0.62 24.27 0.61


2 66.92 67.24 0.48 67.32 0.60
3 77.65 75.31 3.01 75.29 3.03
4 131.97 132.51 0.41 132.68 0.53
5 158.80 154.31 2.83 154.57 2.66

Total error 7.35 7.43

Table 6
Identified mean parameter for FP6.

Parameter Initial value (mm) Identified value (mm)

T1 1.450 1.453
T2 1.450 1.449
T3 1.450 1.453
T4 1.450 1.453
T5 1.450 1.449
T6 1.450 1.453

Std. (T1) 1.45  10  2 1.35  10  2


Std. (T2) 1.45  10  2 1.26  10  3
Std. (T3) 1.45  10  2 1.35  10  2
Std. (T4) 1.45  10  2 1.35  10  2
Std. (T5) 1.45  10  2 1.26  10  3
Std. (T6) 1.45  10  2 1.35  10  2

Fig. 8. Scatter plots of the FPEG model before and after the stochastic model updating.
N. Abu Husain et al. / Mechanical Systems and Signal Processing 32 (2012) 135–152 143

Hence, different sets of parameters are explored by including the material properties in the stochastic updating procedure
as explained next.

2.2.1.4. Stochastic model updating using material properties. In this exercise, the material properties (i.e., the Young’s
modulus (E) and shear modulus (G)) are used for the stochastic updating. The FE model of the flat plates is assumed to have
uniform thickness of 1.45 mm with uniform material properties across the plate, as in the FP1 model (Fig. 2). The initial
mean values of 210 GPa and 81 GPa are assigned for the Young’s modulus and shear modulus respectively, with initial
standard deviations of 1% of their initial mean values. The updated results upon convergence are given in Tables 7 and 8,
which present significant improvement in terms of the natural frequencies with the total error of only 3.17%. The
identified parameters are compared with the updated parameters from the deterministic updating and they are found to
be quite close, indicating that the stochastic updating for the flat plates has been conducted well.
Fig. 9 presents the convergence of the predicted outputs over the measured outputs, while the percentage of errors to
the measured data is shown in Fig. 3(d), which is considerably lower than the errors produced using the thickness as
parameters. As demonstrated by the results, using the material properties as the updating parameters produce reasonable
parameter estimates that consequently converges the outputs successfully. Although the results seem promising, they
need to be confirmed in the next case using the more complicated formed hat-shaped shells.

2.2.2. Case 2: parameter selection for stochastic model updating in complicated formed shells
The findings obtained from the stochastic model updating of simple plates are tested using more complicated
structures, i.e., the formed hat-shaped shells. The formed shells are updated firstly by using only the thicknesses as the
updating parameters and secondly by using a combination of thickness and the material properties as the parameters. It is
considered that uncertainty in the fold radius is incorporated in the thickness of the folds, thus there is no need to include
the fold radius as one of the updating parameters. The existing measurement data of the hat-shaped shells is utilised and
the covariance matrix of the measured outputs for the first five natural frequencies of the hat-shaped shells is determined
as follows:
2 3
0:25 0:12 0:49 0:23 0:60
6 3:28 2:83 0:37 0:65 7
6 7
6 7
6
CEEhat ¼ 6 4:45 1:23 1:48 7 7
6 7
4 sym: 1:23 0:89 5
2:88

Firstly in this exercise, an FE model of the hat-shaped shells with four thicknesses (HS4) is employed. Initial values of
1.45 mm and standard deviations of 1.45  10  2 mm are assigned for the thickness of each region. Similar to the flat
plates, the initial standard deviations are deliberately set at 1% of the initial parameter values. Next, the thicknesses and
material properties of the hat-shaped shells are combined (i.e., HS2E) as updating parameters and results from both HS4
and HS2E are discussed.

Table 7
Mean natural frequencies (in Hz) estimated using FPEG.

Mode I II III
Experiment Initial Updated

FE Error (%) FE Error (%)

1 24.12 24.27 0.62 24.23 0.47


2 66.92 67.24 0.48 67.04 0.18
3 77.65 75.31 3.01 76.67 1.25
4 131.97 132.51 0.41 131.93 0.03
5 158.80 154.31 2.83 156.86 1.22

Total error 7.35 3.17

Table 8
Identified mean parameter for FPEG.

Parameter Initial value (GPa) Identified value (GPa)

E 210.0 209.6
G 81.0 83.8

Std. (E) 2.10 1.60


Std. (G) 0.81 1.23
144 N. Abu Husain et al. / Mechanical Systems and Signal Processing 32 (2012) 135–152

2.2.2.1. Stochastic model updating using HS4. The FE model of the hat-shaped shells is divided into four regions as
illustrated in Fig. 10, with T1 to T4 being the thickness of the folds, thickness of the flanges, thickness of the sidewalls and
thickness of the top, respectively. The initial values of the mean natural frequencies predicted by the model are tabulated
in Table 9. Upon convergence, the mean natural frequencies of the HS4 model are found to be closer to the measured
values with a total error of 5.44%. The updated parameters are shown in Table 10, which indicate that the thicknesses of
the folds and sidewalls are much smaller than the initial estimates, while the thicknesses of the flanges and top are slightly
bigger than the initial values.
Using these updated mean parameters, the covariance matrix of the predicted outputs CZZHS4 is computed and the
difference between the predicted outputs covariance over the measured outputs covariance is illustrated in Fig. 11(a).
Despite considerable errors of up to approximately 85% in the predicted outputs covariance, the updated scatter cloud of
the natural frequencies seems to converge successfully to the measured outputs as can be seen in Fig. 12.

2.2.2.2. Stochastic model updating using HS2E. In this exercise, the FE model of the hat-shaped shells (shown in Fig. 13) is
updated by using the combination of thicknesses and material properties, where T1 and T2 are the thickness of the fold

Fig. 9. Error after updating between the measured and predicted outputs co-variances for the flat plates.

Fig. 10. The HS4 model.


N. Abu Husain et al. / Mechanical Systems and Signal Processing 32 (2012) 135–152 145

Table 9
Mean natural frequencies (in Hz) estimated using HS4.

Mode I II III
Experiment Initial Updated

FE Error (%) FE Error (%)

1 70.11 67.28 4.03 69.12 1.41


2 273.70 256.98 6.11 268.11 2.04
3 287.92 273.47 5.02 283.29 1.61
4 334.73 334.41 0.10 333.44 0.38
5 395.43 386.35 2.30 391.66 0.95

Total error 15.26 5.44

Table 10
Identified mean parameter for HS4.

Parameter Initial value (mm) Identified value (mm)

T1 1.45 1.40
T2 1.45 1.48
T3 1.45 1.32
T4 1.45 1.56

Std. (T1) 1.45  10  2 4.76  10  2


Std. (T2) 1.45  10  2 4.81  10  2
Std. (T3) 1.45  10  2 5.24  10  2
Std. (T4) 1.45  10  2 1.81  10  2

Fig. 11. Scatter plots of the HS4 model before and after the stochastic model updating.

Fig. 12. The HS2E model.


146 N. Abu Husain et al. / Mechanical Systems and Signal Processing 32 (2012) 135–152

Fig. 13. Error after updating between the measured and predicted outputs co-variances for the hat-shaped shells.

Table 11
Identified mean parameter for HS2E.

Parameter Initial value Identified value

T1 (mm) 1.45 1.31


T2 (mm) 1.45 1.54
E (GPa) 210 216

Std. (T1) (mm) 1.45  10  2 6.12  10  2


Std. (T2) (mm) 1.45  10  2 1.14  10  2
Std. (E) (GPa) 2.10 2.65

Table 12
Mean natural frequencies (in Hz) estimated using HS2E.

Mode I II III
Experiment Initial Updated

FE Error (%) FE Error (%)

1 70.11 67.28 4.03 70.34 0.32


2 273.70 256.98 6.11 273.40 0.11
3 287.92 273.47 5.02 289.77 0.64
4 334.73 334.41 0.10 337.82 0.92
5 395.43 386.35 2.30 401.70 1.59

Total error 15.26 2.00

regions and the flat regions, while E is the Young’s modulus of the material used. The initial mean parameters and their
standard deviations are tabulated in Table 11, while the initial mean natural frequencies are given in Table 12. Upon
convergence, the total error of the mean natural frequencies against the measured data reduces significantly to 2%.
The thickness of the fold regions appears to be smaller, while the flat regions are thicker than the initial estimates.
The standard deviations estimated by the stochastic model updating are in the same order as the initial estimates.
The covariance matrix of the predicted outputs CZZHS2E predicted using these updated mean parameters is computed
and the difference between the predicted outputs covariance over the measured outputs covariance is illustrated in
Fig.11(b). It can be observed from the figure that the errors are generally smaller than the errors produced using only the
thickness as parameters. Furthermore, very good convergence is obtained from the updated scatter cloud of the natural
frequencies to the measured outputs, as illustrated in Fig. 14.
It may be concluded that the selection of parameters should be made not only to bring the mean outputs closer to the
measured outputs, but also to reach good convergence between the scatter plots of the predicted and measured data. The
findings presented in this section indicate that selecting some of the material properties as the updating parameters
provides better convergence than selecting only geometric parameters. Furthermore, the error between the covariance
matrix of the predicted and measured outputs is also reduced when including both geometrical and material properties in
the updating procedure, rather than choosing a number of the geometrical properties alone, as has been demonstrated in
N. Abu Husain et al. / Mechanical Systems and Signal Processing 32 (2012) 135–152 147

Fig. 14. Scatter plots of the HS2E model before and after the stochastic model updating.

Fig. 15. The laser spot welded structure.

this section. The findings are followed as guidelines when selecting the updating parameters in the main work, i.e.,
the stochastic model updating of the welded structures.

3. Stochastic model updating of the welded structures

Nine pairs of flat plates and hat-shaped shells were selected from the one used in the previous sections and assembled
together to make a set of nine laser spot welded structures (see Fig. 15). Modal testing was performed to determine the
first five natural frequencies of the nine structures, together with their means and standard deviations, as tabulated in
Table 13. The stochastic model updating is then carried out on the FE model of the laser welded structures based on the
measured data. It is important to ensure that the initial estimates of parameters is not far away from the ‘actual’ values, as
required by the perturbation method. Therefore, the deterministic model updating work [38] has to be carried out prior to
conducting the stochastic model updating. The deterministically-updated values of parameters (as in Table 14) are then
used as the initial inputs for the stochastic model updating work presented in this section.
Variability that exists in the experimental data of the welded structures is investigated by carrying out the following
steps simultaneously:

Step 1: Adjustment of the mean parameters, by minimising the difference between the measured and predicted outputs.
This is performed by using Eq. (3).
Step 2: Adjustment of the parameter covariance matrix, by minimising the difference between the measured and
analytical output covariance matrix. This is performed by using Eq. (5).

The perturbation formulation is applied by using the reciprocals of the measurement variance as the measurement
weighting matrix (Wee) and applying two different approaches for the parameters weighting matrix (Wyy) based on the
following assumptions.

Assumption 1. The substructures are assumed to remain exactly the same except in those small welded regions after the
welding process, i.e., no deformations are introduced on the flat plate and the hat-shaped shell due to welding. Thus, the
only source of variability in the measurements comes from the laser spot weld joints.
148 N. Abu Husain et al. / Mechanical Systems and Signal Processing 32 (2012) 135–152

Table 13
First five measured natural frequencies for the welded structures.

Sample Frequencies (Hz)

1 2 3 4 5

1 509.33 557.16 578.25 634.42 646.65


2 511.08 554.14 577.00 626.59 640.17
3 508.67 554.53 575.92 626.59 645.54
4 509.03 555.26 577.79 628.64 644.45
5 512.18 558.39 580.89 630.84 646.44
6 509.32 552.90 578.38 627.65 646.08
7 507.04 550.40 572.83 625.47 643.49
8 508.03 558.13 573.71 630.04 645.74
9 506.15 556.29 573.74 628.78 644.40

Mean 508.12 553.69 575.39 627.45 643.66


Std. 3.15 5.33 3.87 4.88 4.00

Table 14
Initial values of material and geometrical properties of the laser weld joints.

Properties Values

Diameter of welds (dweld) 5.5 mm


Young’s modulus of welds (Eweld) 220 GPa
Young’s modulus of patch (Epatch) 650 GPa

Assumption 2. The substructures are assumed to experience some deformations during the welding process. Therefore,
the variability in the measurements could be due to the deformation of the substructures and also the uncertainties
associated with the laser spot weld joints.

The first assumption is adopted in Approach 1, while the second assumption is applied in Approach 2. Both approaches
are explained as follows.

 Approach 1
In this approach, the variability of the measured data is assumed to come solely from the weld parameters (i.e., d, Eweld
and Epatch). A weighting matrix in the form of
Wyy1 ¼ lr diagð1,1,1Þ ð6Þ
is assigned for both steps of the stochastic model updating when estimating the means and covariances of the
parameters.
The transformation matrix T for Approach 1 may now be expressed as
T T
T^ j ¼ ðS^ j Wee S^ j þWyy1 Þ1 S^ j Wee ð7Þ

which will be applied in both Eqs. (3) and (5).


 Approach 2
In contrast to Approach 1, Approach 2 includes eight parameters: five from the substructures (i.e., E and G of the flat
plates, and T1, T2 and E of the hat-shaped shells) and three from the spot weld joints (i.e., d, Eweld and Epatch). As
explained earlier, the parameters of the substructures are included to incorporate some uncertainties in the
substructures following the welding process. However, these parameters are not expected to change much in
comparison with the weld parameters, hence different weighting coefficients are applied when updating the mean
parameters (step 1), as follows:
Wyy21 ¼ lr diagð1000,1000,1000,1000,1000,1,1,1Þ ð8Þ

where bigger weighting is given to the substructures parameters to limit their changes during updating. The weld
parameters are not weighted as much to allow them to change more.
On the other hand, the variances of all these eight parameters are highly uncertain; hence smaller weighting (i.e.,
similar weighting as in Approach 1) is used to allow for the parameter covariances to change freely in the second updating
N. Abu Husain et al. / Mechanical Systems and Signal Processing 32 (2012) 135–152 149

step (i.e., to update the parameter covariance matrix).


Wyy22 ¼ lr diagð1,1,1,1,1,1,1,1Þ ð9Þ
The covariances of the parameters are taken to possess more uncertainties than the mean values as there is no prior
information in the distribution of these parameters due to difficulties in measuring the actual parameters values from the
physical structures.
The transformation matrix T for Approach 2 may now be expressed in two ways
T T
T^ 21j ¼ ðS^ j Wee S^ j þWyy21 Þ1 S^ j Wee ð10Þ

and
T T
T^ 22j ¼ ðS^ j Wee S^ j þWyy22 Þ1 S^ j Wee ð11Þ

Assigning different weighting matrices for each updating step (i.e., Eq. (3) for step 1, and Eq. (5) for step 2) will ensure that
the mean parameter values reflect the physical parameters of the structures after the welding process, while allowing for
uncertainties in the distribution of parameter values to be considered. A regularisation parameter (lr) of 400 is selected
based on the L-curve [39] and employed in both approaches.
Stochastic model updating using Approach 1 is discussed first, followed by the second approach. The updating
procedure is conducted based on the measured mean natural frequencies of the welded structures obtained from modal
testing. From the measured data, the covariance matrix of the measured outputs (CEE) is given as
2 3
9:95 13:31 11:53 11:51 8:70
6 28:42 15:71 24:32 18:47 7
6 7
6 7
CEE ¼ 6
6 15:01 14:75 11:52 7
7
6 7
4 sym: 23:85 17:30 5
16:00

It can be seen from Table 15 that the updated mean natural frequencies obtained using the first approach are close to
the mean measured frequencies, and the total error is improved from the initial value. The initial means and standard
deviations of all weld parameters are shown in Table 16. As mentioned earlier, the initial mean values are approximated
from the deterministically-updated values, while the initial standard deviations are deliberately set at 1% of the mean
values. As can be observed from the table, the results of the updated means and standard deviations using Approach 1 are
not in good agreement with the initial values. Considerable changes can be seen in the mean parameters, especially for
Epatch, while the estimated standard deviations are generally bigger than the initial estimates. Using these identified
estimates, the covariance matrix of the predicted outputs (CZZ1) is computed and the percentage of error between CEE and
CZZ1 is shown in Fig. 16(a). The errors are very big (i.e., up to approximately 90%) when using the first approach, hence it
can be concluded that the approach fails to produce a good estimate of CZZ.

Table 15
Mean natural frequencies (in Hz) estimated for the welded structures.

Mode I II III IV
Experiment Initial Approach 1 Approach 2

FE Error (%) FE Error (%) FE Error (%)

1 508.12 494.01 2.78 500.68 1.46 498.10 1.97


2 553.69 552.87 0.15 567.61 2.51 568.14 2.61
3 575.39 564.72 1.85 573.42 0.34 570.35 0.88
4 627.45 615.33 1.93 631.56 0.66 632.85 0.86
5 643.66 620.32 3.63 635.44 1.28 636.40 1.13

Total error 10.34 6.25 7.43

Table 16
Identified mean parameters estimated for the welded structures.

Parameter Initial Approach 1 Approach 2

dweld (mm) 5.5 6.0 5.6


Eweld (GPa) 220 230 222
Epatch (GPa) 650 1595 737
Std. (dweld) (mm) 0.055 0.16 0.06
Std. (Eweld) (GPa) 2.20 5.57 2.65
Std. (Epatch) (GPa) 6.50 10.49 3.17
150 N. Abu Husain et al. / Mechanical Systems and Signal Processing 32 (2012) 135–152

Fig. 16. Error after updating between the measured and predicted outputs co-variances for the welded structures.

700 700

650 650

600 600

550 550
f2 (Hz)
f (Hz)
2

500 500

450 450

predicted outputs 400 predicted outputs


400
measured outputs measured outputs
350 350
350 400 450 500 550 600 650 700 350 400 450 500 550 600 650 700
f1 (Hz) f (Hz)
1

700 700

650 650

600 600

550 550
f3 (Hz)

f (Hz)
3

500 500

450 450

400 predicted outputs 400 predicted outputs


measured outputs measured outputs
350 350
350 400 450 500 550 600 650 700 350 400 450 500 550 600 650 700
f (Hz) f (Hz)
1 1

700
700
650
650
600
600

550
f3 (Hz)

550
f3 (Hz)

500 500

450 450

400 predicted outputs 400 predicted outputs


measured outputs measured outputs
350 350
350 400 450 500 550 600 650 700 350 400 450 500 550 600 650 700
f (Hz) f2 (Hz)
2

Fig. 17. Initial and updated scatter plots for the first, second and third natural frequencies.
N. Abu Husain et al. / Mechanical Systems and Signal Processing 32 (2012) 135–152 151

700
700 650
650
600
600
550

f3 (Hz)
f3 (Hz) 550
500 500
450 450
predicted outputs
400 400 predicted outputs
measured outputs
350 measured outputs
700 350
400 700
600 400 600
500 500
f ( 600 500 500
1 Hz 400 Hz) f (H 600 400 z)
) 700 f 2( 1 z) 700 f2 (H

Fig. 18. Initial and updated scatter plots for the first three natural frequencies.

The updated mean natural frequencies using the second approach are also tabulated in Table 15. It can be seen from the
table that the identified natural frequencies achieved by using the second approach and the measured frequencies are in
good agreement, although the total error produced by Approach 2 is slightly bigger than the total error of Approach 1. The
identified means and standard deviations using the second approach are given in Table 16. As can be observed from the
table, the changes in the identified means and standard deviations of the parameters with respect to the initial estimates
are reasonable.
CZZ2 is determined using the identified estimates of Approach 2, and compared with the covariance matrix of the
measured outputs (CEE). The difference between both matrices is illustrated in Fig. 16(b). The error of each element
appears significantly smaller than the one produced by the first approach. This shows that the estimated means and
standard deviations of the second approach are reasonable and the modification made to the weighting matrices is valid.
Figs. 17 and 18 show the convergence of the predictions upon the experimental data in the space of the first three
natural frequencies using the second approach. In these scatter clouds, five hundred samples are generated by the Monte
Carlo simulation following the Gaussian distribution. It can be seen from the figures that the general trends of the updated
results are similar to that of the initial results. The findings demonstrate that modification to the weighting assignment is
capable of bringing the numerical results to convergence. This could be due to the introduction of uncertainties in the
substructures following the welding process in the stochastic analysis.

4. Conclusions

Stochastic model updating using the perturbation method is explained in this paper. Then, parameter selection for two
different sets of structures (i.e., simple plates and complicated hat-shaped shells) is studied for stochastic model updating
procedure. Various options of updating parameters (i.e., geometrical properties, material properties, and their combina-
tions) are investigated in order to demonstrate the importance of right parameterisation. The findings are used as a
guideline in selecting the updating parameters for quantifying variability in the dynamics of the laser welded structures.
Information on manufacturing variability on the test structures are obtained by conducting modal testing on each
individual structure.
Selection of proper updating parameters is a key to the success of model updating. It requires experience and
engineering judgement (or physical insight). As some general guidelines, the selection of parameters should be made by
choosing the most sensitive parameters that affect the outputs of the system, which can be easily achieved by carrying out
a simple sensitivity analysis. The selection of parameters should also be made so that the mean outputs are closer to the
measured outputs, and convergence between the scatter plots of the predicted and measured outputs can be obtained.
This can be acquired by including both geometrical and material properties in the updating procedure, rather than
choosing a number of geometrical properties alone.
The stochastic model updating problem for the welded structures are performed by employing a mean-centred
perturbation method, where only the first two statistical moments—the mean and standard deviations are involved in the
algorithm. Two approaches of parameters weighting matrices are introduced; the first approach considers only the main
uncertain parameters that come from the laser spot weld joints, while more parameters are accounted for in the second
approach (i.e., by incorporating uncertainties associated with the laser spot weld joints and also uncertainties associated
with the substructures) due to possible variation in the geometry following the welding process. Both approaches are
tested to demonstrate the feasibility of the parameters weighting matrices assignment for estimating variability in welded
structures. It is found that the second approach has significantly improved the results over the first approach.
Scatter clouds of the measured and predicted outputs are generated by the Monte Carlo simulation following the
Gaussian distribution to show convergence of the stochastic updating procedure, and as demonstrated in this paper, the
predicted space has converged upon the measured space very well. The findings show that selection of parameters must be
152 N. Abu Husain et al. / Mechanical Systems and Signal Processing 32 (2012) 135–152

done appropriately based on uncertainty and sensitivity level of the parameters, as considered in this paper. Additionally,
the number of samples (i.e., five hundred) used in plotting the scatter clouds are clearly enough to give an accurate
estimate of the parameter variability using the perturbation method.

Acknowledgement

This work was carried out as part of the Ph.D. research project of the first author at the School of Engineering, University
of Liverpool under the supervision of the third author. The first author would like to thank the Malaysian Ministry of
Higher Education (MOHE) and the Universiti Teknologi Malaysia (UTM) for their support. All the authors are grateful to the
support from Prof. John E. Mottershead, Dr. Simon James and Mr Andy Snaylam.

References

[1] B.R. Mace, K. Worden, G. Manson, Uncertainty in structural dynamics, J. Sound Vib. 288 (2005) 423–429.
[2] Y. Xia, H. Hao, J.M.W. Brownjohn, P.-Q. Xia, Damage identification of structures with uncertain frequency and mode shape data, Earthquake Eng.
Struct. Dyn. 31 (2002) 1053–1066.
[3] J.E. Mottershead, C. Mares, S. James, M.I. Friswell, Stochastic model updating: part 2—application to a set of physical structures, Mech. Syst. Signal
Process. 20 (2006) 2171–2185.
[4] H.C. Frey, D.E. Burmaster, Methods for characterizing variability and uncertainty: comparison of bootstrap simulation and likelihood-based
approaches, Risk Anal. 19 (1999) 109–130.
[5] W.L. Oberkampf, J.C. Helton, C.A. Joslyn, S.F. Wojtkiewicz, S. Ferson, Challenge problems: uncertainty in system response given uncertain parameters,
Reliab. Eng. Syst. Saf. 85 (2004) 11–19.
[6] R.Y. Rubinstein, D.P. Kroese, Simulation and the Monte Carlo Method, John Wiley & Sons, Inc., New Jersey, 2008.
[7] K. Binder, D.W. Heermann, Monte Carlo Simulation in Statistical Physics: An Introduction, Springer, Germany, 2002.
[8] R.H. Myers, D.C. Montgomery, C.M. Anderson-Cook, Response Surface Methodology: Process and Product Optimization Using Designed Experiments,
John Wiley & Sons, Inc., New Jersey, 2009.
[9] Y. Ben-Haim, Convex models of uncertainty: applications and implications, Erkenntnis 41 (1994) 139–156.
[10] H.N. Najm, Uncertainty quantification and polynomial chaos techniques in computational fluid dynamics, Annu. Rev. Fluid Mech. 41 (2009) 35–52.
[11] H. Cheng, A. Sandu, Efficient uncertainty quantification with the polynomial chaos method for stiff systems, Math. Comput. Simulation 79 (2009)
3278–3295.
[12] J.R. Fonseca, M.I. Friswell, J.E. Mottershead, A.W. Lees, Uncertainty identification by the maximum likelihood method, J. Sound Vib. 288 (2005)
587–599.
[13] T. Haag, J. Herrmann, M. Hanss, Identification procedure for epistemic uncertainties using inverse fuzzy arithmetic, Mech. Syst. Signal Process. 24
(2010) 2021–2034.
[14] M. Guedri, A.M.G. Lima, N. Bouhaddi, D.A. Rade, Robust design of viscoelastic structures based on stochastic finite element models, Mech. Syst.
Signal Process. 24 (2010) 59–77.
[15] R.G. Ghanem, A. Doostan, On the construction and analysis of stochastic models: characterization and propagation of the errors associated with
limited data, J. Comput. Phys. 217 (2006) 6381.
[16] D. Moens, D. Vandepitte, A survey of non-probabilistic uncertainty treatment in finite element analysis, Comput. Methods Appl. Mech. Eng. 194
(2005) 1527–1555.
[17] K. Worden, G. Manson, T.M. Lord, M.I. Friswell, Some observations on uncertainty propagation through a simple nonlinear system, J. Sound Vib. 288
(2005) 601621.
[18] S. Adhikari, M.I. Friswell, Random matrix eigenvalue problems in structural dynamics, Int. J. Numer. Methods Eng. 69 (2007) 562–591.
[19] H. Ahmadian, G.M.L. Gladwell, F. Ismail, Parameter selection strategies in finite element model updating, ASME J. Vib. Acoust. 119 (1997) 37–45.
[20] G. Kim, Y. Park, An improved updating parameter selection method and finite element model update using multiobjective optimisation technique,
Mech. Syst. Signal Process. 18 (2004) 59–78.
[21] J. Mottershead, C. Mares, M.I. Friswell, S. James, Selection and updating of parameters for an aluminium space-frame model, Mech. Syst. Signal
Process. 14 (2000) 923–944.
[22] H. Haddad Khodaparast, J.E. Mottershead, M.I. Friswell, Perturbation methods for the estimation of parameter variability in stochastic model
updating, Mech. Syst. Signal Process. 22 (2008) 1751–1773.
[23] M.I. Friswell, J.E. Mottershead, Finite Element Model Updating in Structural Dynamics, Kluwer Academic Publishers, Dordrecht, The Netherlands,
1995.
[24] E. Balmes, Parametric families of reduced finite element models. theory and applications, Mech. Syst. Signal Process. 10 (1996) 381–394.
[25] C. Mares, J.E. Mottershead, M.I. Friswell, Stochastic model updating: part 1—theory and simulated example, Mech. Syst. Signal Process. 20 (2006) 1674–1695.
[26] M. Luczak, A. Vecchio, B. Peeters, L. Gielen, H. Van der Auweraer, Uncertain parameter numerical model updating according to variable modal test
data in application of large composite fuselage panel, Shock Vib. 17 (2010) 445–459.
[27] G. Stefanou, The stochastic finite element method: past, present and future, Comput. Methods Appl. Mech. Eng. 198 (2009) 1031–1051.
[28] J.L. Beck, S.-K. Au, Bayesian updating of structural models and reliability using Markov chain Monte Carlo simulation, J. Eng. Mech. 128 (2002) 380391.
[29] J.D. Collins, G.C. Hart, T.K. Hasselman, B. Kennedy, Statistical identification of structures, AIAA J. 12 (1974) 185–190.
[30] X. Hua, Y.Q. Ni, Z.Q. Chen, J.M. Ko, An improved perturbation method for stochastic finite element model updating, Int. J. Numer. Methods Eng. 73
(2008) 1845–1864.
[31] B. Faverjon, P. Ladeveze, F. Louf, Validation of stochastic linear structural dynamics models, Comput. Struct. 87 (2009) 829–837.
[32] R.M. Lin, H. Du, J.H. Ong, Sensitivity based method for structural dynamic model improvement, Comput. Struct. 47 (1993) 349–369.
[33] M. Friswell, Calculation of second and higher order eigenvector derivatives, J. Guidance 18 (1994) 919–921.
[34] H. Haddad Khodaparast, J.E. Mottershead, Efficient methods in stochastic model updating, in: Proceeding of ISMA 2008, Leuven, Belgium, 2008.
[35] Y. Govers, M. Link, Stochastic model updating—covariance matrix adjustment from uncertain experimental modal data, Mech. Syst. Signal Process.
24 (2010) 696–706.
[36] H. Haddad Khodaparast, J.E. Mottershead, K.J. Badcock, Interval model updating with irreducible uncertainty using the Kriging predictor, Mech. Syst.
Signal Process. 25 (2010) 1204–1226.
[37] M. Brehm, V. Zabel, J.F. Unger, Stochastic model updating using perturbation methods in combination with neural network estimations, in:
Proceedings of the IMAC XXVII A Conference and Exposition on Structural Dynamics, February 9–12, 2009, Orlando, Florida, USA.
[38] N. Abu Husain, H. Haddad Khodaparast, A. Snaylam, S. James, H. Ouyang, Finite element modelling and updating of laser spot weld joints in a top-hat
structure for dynamic analysis, Proc. IMechE Part C: J. Mech. Eng. Sci. 224 (2009) 851–861.
[39] P.C. Hansen, Analysis of discrete ill-posed problems by means of the L-curve, SIAM Rev. 34 (1992) 561–580.

Вам также может понравиться