Documentos de Académico
Documentos de Profesional
Documentos de Cultura
The manuscript was received on 31 March 2008 and was accepted after revision for publication on 18 June 2008.
DOI: 10.1243/09544070JAUTO866
Abstract: Today’s automotive industry uses finite element analysis (FEA) in a huge variety of
applications in order to optimize structures and processes before hardware is produced.
Efficiencies can be enhanced and margins are reduced because the external loads and
structural properties are identified with higher confidence. The accuracy of FEA predictions has
become increasingly important and directly influences the competitiveness of a product on the
market. Because automotive structures are under dynamic environments, the correlation on
the basis of static deformations independent of the mass and damping parameters do not
provide a valuable reference from the view of the dynamic characteristics. In this paper, by
systematically comparing the results from analytical and experimental analysis techniques,
finite element (FE) models can be validated by the deterministic and robust design on the basis
of each tolerance of design parameters, and improved so that they can be used with more
confidence in further analysis. Making use of different types of test datum, a recommended
procedure is to use a sequence of analysis in which mass, stiffness, damping, and external
loading are validated and, if necessary, updated.
JAUTO866 F IMechE 2008 Proc. IMechE Vol. 222 Part D: J. Automobile Engineering
1782 D-C Lee and C-S Han
design optimization techniques to minimize the dif- Test results show the feasibility of this approach,
ferences between test and FEA results. The problem and its practicality.
here has been the difficulty in obtaining sensitivity
derivatives efficiently for the case of many eigenvec-
tors or response solutions. This has led to a mixture 2 THEORETICAL CONSIDERATIONS FOR THE
of data comparisons, including combinations of DYNAMIC CORRELATION
modal and forced-response samples. In the absence
of prototypes, analytical methods such as FEA are The automotive body structure made by the sheet-
very useful in resolving noise and vibration pro- metal-forming process influences the junction boun-
blems, by predicting the dynamic behaviour of the daries and gradual variation shapes on the panels
automotive components and systems. FEA is a owing to the deformed shapes or strain distribu-
simulation technique and involves making assump- tions, which in turn influence the thinning effect of
tions that affect analytical results. Acceptance and panels or the geometric weak points. These material
use of these results are greatly enhanced through test and physical variations can change the structural
validation. The automotive development process performances [7–11]. When taking into account these
using computer-aided engineering (CAE) is shown parameters, the simulation model can give the rel-
in Fig. 1. As the stages from the concept verification iabilities of the prediction of structural responses.
plan to development verification plan are preceded, In this paper, the parameter estimation method using
the dependences on the CAE become larger and the the frequency response is presented. The explana-
model correlations with test become more impor- tion of the feasibility of parameter estimation on
tant. the frequency domain can be explained as follows.
The aim of the model-updating technique pre- Starting with a global FE model and using the
sented by this paper is to reduce the errors in the direct frequency response, the resulting steady-state
FE results by predicting changes to selected struc- matrix equations are
tural properties. This approach can bypass many of
the previous difficulties by using the following meth- {v2 M a zjvBa zK a fug~fF g ð1Þ
ods.
or, in terms of the dynamic stiffness matrix [Z], as
1. The error measures are defined directly from the
solution vectors to avoid large, complicated, sym- ½Z a ðvÞ~ {v2 M a zjvBa zK a
bolic equation entries and manually transcribed
data. where [Ma], [Ba], and [Ka] are the mass, damping,
2. Frequency response solutions are used to avoid and stiffness matrices, respectively, for the analytical
the difficult task of calculating eigenvector deri- model. Note that these matrices may be complex
vatives. and unsymmetric. The analytical solutions to the
3. Constraint equations are built into the solution to loads, Fi 5 F(vi), are
enforce test responses and produce faster con-
vergence. ½Z a ðvÞ uai ~fFi g ð2Þ
Proc. IMechE Vol. 222 Part D: J. Automobile Engineering JAUTO866 F IMechE 2008
A frequency response updating technique for automotive structures 1783
The objective of this procedure is to find the mod- of the magnitude of the dynamic characteristics for
ifications to the FE properties that will produce selected points and selected frequencies. Indirect
a new set of impedances [Za(v)], such that for each constraints are available to limit the changes in the
selected frequency design parameters for each iteration. An important
aspect of the optimization procedure is the linear-
½Z x ðvÞ uxi &fFi g ð3Þ ized iteration method using the sensitivity matrices.
The complete stiffness matrix [K] for frequency
where the arbitrary vector uxi at each frequency response analysis consists of a superposition from
contains results that match the test data. The change various sources according to
in impedance can be defined as X
½K ~½K1 zjg ½K1 z gS ½KS z½K2 ð7Þ
½DZ ðvi Þ~½Z x ðvi Þ{½Z a ðvi Þ ð4Þ
where [K1] is the structural stiffness matrix and [K2]
The matrix changes in [Ma], [Ba], and [Ka] which is the direct matrix input at nodes. g is the uniform
form the matrix [Za(v)], in turn, may be defined by structural damping coefficient. gs is the structu-
structural parameters, such as geometric dimen- ral damping coefficient of elements. Updating the
sions, material properties, or non-structural mass stiffness modelling with the displacement, the error
densities. Note that the potential number of para- function
meters could grow to a large number if every element
in the analytical system could change independ- err~DuT Ce DuzDpT Cp Dp ð8Þ
ently. However, a unique solution is impossible.
The relationship to the Lin–Ewins approach can should be minimized. In equation (8), Ce and Cp are
be shown by solving equation (3) by subtracting the diagonal weighting matrices for the selected
equation (2) and substituting equation (4) to ob- updating of the performance and the updating
tain parameters, respectively. The subscripts e and p
indicate the elastic and plastic performance indices,
Zia ðvi Þ uxi {uai &{½DZi ðvi Þ uxi ð5Þ respectively. Du is the difference between experi-
mental and analytical displacements: Du~
By inverting [Za(v)] at each frequency, a basis for n o n o
j j
L uj Lpk Dp~ uexp { uana . DRk is the differ-
iteration is obtained as
ence between updated and originally estimated
{1
uxi {uai &{ Zia ðvi Þ ½DZi ðvi Þ uxi ð6Þ parameters: Dpk 5 pku 2 pka.
[K] are known as the functions of {Du} relative to
In other words, the changes in the responses are res- the structural rigidities due to the geometric dimen-
tricted by the constant analytical structural proper- sions, and, therefore, {DF} can also be calculated in
terms of geometric dimensions. It can be assumed
ties in Zia ðvi Þ . The advantage of this method is that
the rows corresponding to the measured points that the tangent stiffness matrix, incremental defor-
may be solved simultaneously with the unmeas- mation, and load unbalance are functions of dimen-
sional parameters according to
ured points. The unmeasured points in the uxi
vectors are updated with each estimate of [DZi(vi)].
½K ~½K ðpÞ
The errors in the tested degrees of freedom become
the measure of convergence in the iterations. The fDug~fDuðpÞg
design parameters are the structural parameters to fDF g~fDF ðpÞg
be varied. Options include FE property data, scalar ð9Þ
element coefficients, and grid point locations. They
define the effects of the matrices [DK], [DC], and For design optimization of non-linear structural
[DM] indirectly through the FE sensitivity matrices. performance, the sensitivity analysis is given by
The objective function may include a variety of equation (7).
parameters, including the response errors, the limits LL[K]/LLpk is the derivative of the system stiffness
on the weight, and the modal properties. Current matrix with respect to the parameter and is given by
tests have used a minimization of a simple sum over
key frequencies of the squared errors at important ½K ðpk zDpk Þ{½K ðpk Þ
points. Optimization constraints are the upper limits L½K =Lpk ~ ð10Þ
pk
JAUTO866 F IMechE 2008 Proc. IMechE Vol. 222 Part D: J. Automobile Engineering
1784 D-C Lee and C-S Han
Proc. IMechE Vol. 222 Part D: J. Automobile Engineering JAUTO866 F IMechE 2008
A frequency response updating technique for automotive structures 1785
Step 3. Compare the results of the analysis with objective-function-solving suboptimization step is
such requirements as allowable elastic or plastic set to maximize or minimize a performance corres-
specifications. ponding to the substructure. The probability con-
Step 4. If the requirements are not met, perform straint related to the reliability of the performance
the optimization routine in order to set dp. function is introduced. The Monte Carlo simulation
Step 5. Correct dpi + 1 based on the fully stressed examines the variance of performance function and
design with b 5 0.9 according to constraint satisfaction probability at the optimal
points. The probability constraint has a suboptimi-
b zation problem in terms of state parameters, design
iz1
iz1
Upart
dp new
~ dp old
ð16Þ parameters, and random parameters.
Uavg
Let m 5 {m1 m2 … mn}T and p 5 {p1 p2 … pn}T be the
distributional random and deterministic design
where
parameters, respectively. The RBSO problem can be
formulated as
dpiz1 old
~G i a
! !
+Y i 2 dY i
i i{1
~ {+Y z S
T Minimize Wðm, pÞ
+Y i{1 2 +Y i Gi
subject to Zi ðGk ðm, pÞÞ¡ZiU , i~1, . . . , m
The objective DRi(P) can be approximated for each mLj ¡mj ¡mU L U
j , pj ¡pj ¡pj , j~1, . . . , n
design P(i) using the series expansion ð18Þ
X n
LðDY Þ where W is the structural objective function and
DY i ~DY i P ðiÞ z Dpj ð17Þ Zi(Gk(m, p)) is the probability of the kth structural
LPj
j~1 performance. ZiU is the required upper bound of
probability. mLj and mUj are the lower and upper
bounds, respectively, of the jth distributional ran-
The gradient L(DY)i/LPj can be obtained directly
dom design parameters; pLj and pU j are the lower
from the results of FEA. If the gradient is known,
and upper bounds, respectively, of the jth determi-
the search direction Dp can be obtained from the nistic design parameters. The reliability constraints
solution of an approximate optimization problem. are assumed to be independent. To calculate the
Change the design parameters using p i + 1 5 probability of Zi(Gk(m, r)), the first-order reliability
pi + dpi + 1. method (FORM) is used. At each design iteration,
Step 6. If the requirements are satisfied, perform the FORM needs to determine a reasonable design
the discrete design for the design parameters with assessment for each structural objective function
consideration of manufacturability. Otherwise, go for the next design iteration. To perform the
to step 2. gradient-based optimization iteration, the sensitiv-
ity of objective function and constraints would be
The deterministic structural optimization pro- needed. A suboptimization problem can be given
blem based on sequential quadratic programming by
can be formulated as a non-linear constrained
optimization problem. Similarly, the general relia-
X LG
bility-based structural optimization (RBSO) prob- Minimize or maximize G ðu, m, pÞ~Gzk Dp
Lp
lem can also be formulated as a non-linear
constrained optimization problem in which the subject to Qðu, m, pÞ¢0
constraint functions are reliability-measuring func- H ðu, m, pÞ~0
tions. The reliability constraints ensure a more
ð19Þ
feasible design guideline in the structural concept
design. Generally, the RBSO model includes two
types of design parameter: distributional random where k is the penalty factor with respect to each
parameters m and deterministic design parame- constraint and determined by the user.
ters p. In this paper, a new formulation that Let u* and m* denote the optimal state parameter
considers the mean value and variation in the and random parameter, respectively, of the subopti-
performance function sequentially is proposed. The mization problem, then the Kuhn–Tucker necessary
JAUTO866 F IMechE 2008 Proc. IMechE Vol. 222 Part D: J. Automobile Engineering
1786 D-C Lee and C-S Han
LG LH LQ
dG 1~ drzlT drzgT dr ð21Þ
Lp Lp Lp Fig. 2 Design procedure of automotive structure
through the model correlation (CAD, compu-
4 SIMULATION ter-assisted design)
Proc. IMechE Vol. 222 Part D: J. Automobile Engineering JAUTO866 F IMechE 2008
A frequency response updating technique for automotive structures 1787
X X
4 GjL ¡configuration vectorfG g¡Gju , j~1, . . . , n
Minimize Y ðpÞ ~ f t {wi f a ðpÞ for wi ~1
i i
i~1
where {G} is the move vector of the bead shape.
subject to 0:95ðti Þ0 ¡ti ¡1:05ðti Þ0 The CTBA is modelled using shell and solid elements
ð22Þ with appropriate thickness values along the beam
length. The body mounting points are modelled using
The geometrical constraints of the open sector shape rigid body elements and spring elements. The results of
JAUTO866 F IMechE 2008 Proc. IMechE Vol. 222 Part D: J. Automobile Engineering
1788 D-C Lee and C-S Han
point mobility are shown in Fig. 7(a), which explains If the solution was linear with respect to the design
the mobility trends on the selections of design parameters, unique solutions could be expected. The
parameters in the validation phase (ALT 3, with number of design parameters is the number of
consideration of three thinning effect zones; ALT 2, unknown parameters. The potential number of
with consideration of two thinning effects; ALT 1, with known coefficients is equal to the number of test
consideration of one thinning effect). Figure 7(b) degrees of freedom multiplied by the number of
shows the results between the selection of thinning sample frequencies multiplied by 2. It is recom-
domain and thinness of torsion beam through the mended that an over-determined system is used to
model-updating scheme. The dynamic test set-up and allow for possible redundant data. Of course, a large
durability simulations are shown in Fig. 8. Table 2 number of frequencies in a small range with only a
show the deterministic optimum and robust optimum few active modes will not provide much indepen-
on the penalty factor, k 5 1.1. dent information. However, the use of many sample
The errors can be generated by the residual forces test points on the structure and the use of multiple
of the constraint conditions. These values of three
measured points over all three frequencies were the
functions chosen to be minimized by the optimizer.
Some of many test variations were as follows.
Proc. IMechE Vol. 222 Part D: J. Automobile Engineering JAUTO866 F IMechE 2008
A frequency response updating technique for automotive structures 1789
load cases will help. If the design changes are large, static and dynamic specifications. The advantage of
the non-linearities may produce several different the presented procedure increases such accuracies of
reasonable solutions. durability designs as the damage locations and fatigue
life. The correlation method presented in this paper
5 CONCLUSION offers an efficient approach to robust design by
defining the robustness of the objective and the
This procedure is an efficient and accurate tool for constraint functions. The optimum evaluated from
reliable prediction of structural performance including robust design is the optimum intensive to the
JAUTO866 F IMechE 2008 Proc. IMechE Vol. 222 Part D: J. Automobile Engineering
1790 D-C Lee and C-S Han
Fig. 8 (a) CTBA test set-up and (b) results of durability simulation
variations in design variables satisfying the imposed design parameters and the robustness of objective
constraints. The design process is practical for struc- function is enhanced by decreasing the weighting
tural designs. The robust optimum is evaluated factor, while the robustness of the constraint function
through the sequential stage optimization which is enhanced by increasing the penalty factor. The
follows the deterministic optimization. Robust design factors should be selected depending on the char-
has to be performed considering the tolerances of the acteristics of the problem.
Proc. IMechE Vol. 222 Part D: J. Automobile Engineering JAUTO866 F IMechE 2008
A frequency response updating technique for automotive structures 1791
JAUTO866 F IMechE 2008 Proc. IMechE Vol. 222 Part D: J. Automobile Engineering