See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/273903885

Virtual SEA: Mid-Frequency Structure-Borne Noise Modeling Based on Finite

Element Analysis

Conference Paper · May 2003

DOI: 10.4271/2003-01-1555

CITATIONS READS

15 497

1 author:

Laurent Gagliardini
Groupe PSA
80 PUBLICATIONS   748 CITATIONS   

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Modeling engineering problems - Some transdiciplinar applied works View project

All content following this page was uploaded by Laurent Gagliardini on 20 February 2017.

The user has requested enhancement of the downloaded file.

 Virtual SEA: mid-frequency structure-borne noise modeling
based on Finite Element Analysis
L. Gagliardini, L. Houillon
PSA - Peugeot Citroen – Route de Gisy- F78943 VELIZY-VILACOUBLAY Cedex

L. Petrinelli
GECI France-2, Esplanade Grand Siècle-F78000- Versailles

G. Borello
INTERAC-10, Impasse Borde-Basse, ZA la Violette-F31240- L’UNION, [email protected], www.interac.fr

ABSTRACT out dedicated computational techniques that remained at

a research stage until now.
The virtual SEA is similar to experimental SEA, but
based on Frequency Response Functions computed Thus, the industrial design process led us to develop a
using a FE model of the studied structure. The method which main feature is a good adaptability to the
knowledge of the internal loss factors (defined by the current design process, including reduced operating
user) as well as numerous observation and excitation time, lowered specific expertise of users. More than a
points leads to a consistent data set that may be used to real innovation [6-7] the proposed method, called Virtual
properly identify a SEA model. This process has been SEA, pretends to allow a car body modeling that
improved by developing an original automatic sub- overpasses traditional SEA limitations by an extensive
structuring technique which guarantees an optimized use of numerical simulation.
model construction and consequently a robust
identification. This last feature allows a lower expertise This method and its application to the floor of a minivan
level of SEA users. (Peugeot 806) will be presented. Preliminary, a brief
presentation of the background of the method is
Virtual SEA has been applied to the floor of a minivan. proposed.
Convincing results are obtained when compared to
experiment. EXTENSION OF THE FEA PROCESS TO
MEDIUM FREQUENCIES
INTRODUCTION
Due to crash analysis, refined meshes such as the one
Improvement in vehicles acoustic performances at low of figure 1 are available, at least for the body in white,
frequencies (booming noise) and high frequencies from the early design stage of any vehicle project.
(insulation) let perceived medium frequencies (200 Hz-
1kHz) be the most critical band regarding passengers The mesh size -a few centimeters- allows computations
comfort vs. power-train or road excitations. At the same up to 1 kHz with no major difficulty. Because of
time, new design processes -such as systems insufficient memory size problems, a direct solver is used
engineering- reducing development time and prototypes instead of the common modal solution. The resulting
availability emphases computational methods, able to increase of the CPU time has not been seen as an
predict vehicle performances in any technical field: CFD obstacle, considering a number of recent improvements
and thermal comfort, durability, crash analysis, CEM, in numerical solvers for dynamic problems [8-10].
vehicle dynamics and NVH. Sound transmission at
medium frequencies and especially structure-borne
transmission is one of the last vibro-acoustic sub-
domains that are not covered by any computational
method, that is to say that cannot be handled properly
during the design phase of a project.

The main reason for this state of things is that this

frequency range is part of the physical medium
frequencies, where the response of the structure
involves global as well as local behaviors; therefore,
neither the ‘modal’ behavior –dominant at low
frequencies- nor the ‘statistical’ behavior –dominant at
high frequencies- can alone represent the medium
frequency behavior. Numerous authors [1-5] are working Figure 1. Body In White of a 307 SW car
Once it is admitted that FE computations are available at This quantity is preferred to the squared transfer of
rather high frequencies, we could be asked: ”why not mobility function, because its variance among input
using FEA directly?” points is much smaller, meaning it is more informative
about the dynamics of the system. Its use in the Virtual
The main objections to the direct use of FEA are: SEA Process appears explicitly in Appendix.

• Increasing frequency lets the structures be hyper- Figure2: Front central part of a minivan floor structure defining 4
sensitive to small uncertainties in material subsystems.
properties, geometrical details. Such a sensitivity,
inherent to mass production objects such as cars,
cannot be ignored in the medium frequency range.
In this case only space and frequency averaged
responses (i.e. energetic responses) can be Figure 3: Matrix of Energetic Transfer Function between the 4 sub-
systems as defined on figure 2, computed (red- 100 Hz-2 kHz) and
predicted with a convenient precision.
measured (blue- 20Hz-10 kHz).

• The FE modeling in itself does not provide any

understanding scheme, necessary to guide the
project improvement. The modal understanding of
the structure, used at low frequencies, is not
It may be noticed that, although observation as well as
relevant any more due to the high number of modes
overlapping to produce the observed responses.
The modal behavior provides clear design
information through the locations of kinetic and Observed subsystem i
elastic energies respectively indicating where mass
and stiffness or damping modification are sensitive.
At higher frequencies, when modes start Excited subsystem j
overlapping, such information is not available any
more. In the asymptotic case referred as “high
frequencies”, it is considered that kinetic and elastic
energies are equal at any point of the structure.

Figure 3 shows the comparison of computed and

measured space and frequency averaged transfer
functions between different parts of the minivan floor
shown on figure 2. The plot shows the matrix of excitation points do not have the same location in both
energetic transfer functions defined as : cases, measurements and computations lead to very
similar results in terms of spectral shape and spread
2 among excitation points… except for part 1, where
v calculations show an overestimated value that was later

{ }
∆f , Ω i
H ij = +
(1) related to a reinforcing plate, unfortunately weld to the
Re v F j j prototype. This problematic is shown on purpose, in
∆f
order to emphasis the difficulties inherent to a design
process: the body design is possibly changing locally,
where indicates a frequency average, and a resulting in the fact that any modeling, even the most

{ }
∆f Ω accurate, remains uncertain regarding the design
process. Work is in progress in order to account for such
space average, and Pi = Re v +j F j is the input
modeling uncertainties.
∆f
power at the excitation point (subscript + indicates the
conjugate transposed vector or matrix). LIMITATIONS OF CLASSICAL SEA

Looking at vehicle project issues (trim design mainly but

1 also vibration source strength specification), SEA
appears to be a very convenient tool for the design stage
of a project. Its wide use for air-borne transmission of
sound is a kind of a proof. But going to structure-borne
2
sound, some unexpected difficulties are arising: most of
3 car body parts do not look like conventional structures
such as flat plates, cylinders or beams. Moreover, when
a candidate subsystem appears, its limits are difficult to
draw, due to wired 3D-geometries. This early feeling is
4 unfortunately confirmed when practicing.
For additional reasons, experimental SEA or analytical Structure damping is modeled through a uniform
SEA are not able –at the moment- to produce a robust structural damping, which could later be used as a SEA
car body SEA model in the target frequency range [200- Damping Loss Factor. This damping shall be close to the
1kHz]. Experimental SEA is limited by the lack of actual damping in case that later CLF evaluation would
consistence of the measured data, but also because the depend on damping, as suggested by many authors.
sub-structuring is not controlled, leading to contradictory
behaviors in-between points of the same subsystem. Then, each node of the refined mesh, corresponding to
This sub-structuring problem is also encountered while the “NVH mesh” nodes, is successively excited by a unit
building analytical SEA models. With analytical models, force, normal to the structure and the transfer functions
the user also has to tune the model to account for are computed for all the nodes of the “NVH mesh”. This
heterogeneity (corrugations, small hollow bodies, simulation is performed using the Nastran V2001 direct
variable beam cross-section). solver on an IBM SP2 computer. The result of the
simulation is a rectangular matrix of transfer functions
Overcoming those difficulties requires a level of expertise between nodal normal unit loads, and 3 velocity DOFs at
that cannot be easily transferred to a large number of the observation points. This database is then
engineers, as required by today design process of a car. compressed by a FORTRAN program to compute the
input mobility and the Energetic Transfer Functions (1).
Before going further on, it seems important to recall that This new database is averaged over third octave bands.
analytical or parametric models are not rejected at all:
they would be the most helpful, when they are available. AUTOMATIC SUB-STRUCTURING PROCESS

VIRTUAL SEA As soon as a database of point to point Energetic

Transfer Functions is available, SEA analysis requires a
Considering the above mentioned difficulties, it was compression of this information over a limited number of
decided: subsystems. This process is called data reduction and is
the most critical phase while creating the SEA model.
To use finite elements to simulate a fully Past experimental SEA applications to car bodies have
consistent energetic transfer function matrix brought to the fore the critical issue of a “good” sub-
measurement, and associated input powers. structuring scheme.
To look for the optimal gathering of
observation/excitation points (sub-structuring process), When the database is compressed over an inappropriate
To use an existing experimental SEA process to sub-structuring scheme, the inverse problem that follows
identify a SEA model. (see next paragraph and appendix) may result in non-
physical solution. These discrepancies are mainly due to
Other interesting issues such as the method described in some abusive weak coupling assumption between SEA
[11], have been considered but rejected at this stage, subsystems. Unfortunately, working on a car body, this
because it relies on an a priori knowledge of the sub- “weak coupling assumption” is not intuitive and cannot
structuring. Even when the method -currently referenced be a priori predicted, until a clear definition exists.
as Energy Influence Coefficients- provides an exact
evaluation of the subsystems energy, the lack of control From a practical point of view, FE generates a
on the sub-structuring was considered as the most tremendous amount of data and dividing the data sets
critical. into subsystems could become a very tedious and nearly
impossible task without computer help. It was thus
Lets now consider each step of the proposed method. decided to automate this division process following the
here after principles:
ENERGETIC TRANSFER FUNCTIONS
COMPUTATION • A subsystem is an energetic entity that must exhibit
at the same time significant energy level difference
The first step is a numerical experiment using the Finite with other subsystems, and minimized internal
Element Analysis, in order to compute frequency heterogeneity.
response functions.
• For any partition scheme, averaged responses and
This approach relies on two different meshes. The first internal variances are used to compute the distance
one is a refined mesh allowing Finite Element to the above ideal situation.
computations up to 2 kHz: it is obtained without difficulty
since it is similar to the meshes used for crashworthiness • The partition scheme is optimal when the distance is
studies. The second one, called “NVH mesh”, is a coarse minimal.
mesh extracted from the refined mesh providing a
relevant sampling of the vibratory behavior of the An iterative algorithm has thus been developed.
structure. It contains several hundreds nodes and will be
used to estimate subsystems energies. Initialization
The initializing sequence creates a first set of As the iterative algorithm relies on energetic transfer
subsystems as follows. functions in a given frequency band (here 1/3rd octave),
the sub-structuring process can be performed in each of
1. An arbitrary excitation point is selected and a first set the studied bands, leading to different sub-structuring
of points whose response is lower than a given schemes. It is remarkable that the number of
energy threshold is retained. subsystems is increasing with frequency as suggested
2. These points are used as excitation points to create by SEA assumptions on modal densities (increasing with
a new series of sets. frequency) or weak coupling conditions (also improving
3. Reciprocity is then invoked to find strongly linked with frequency). From a practical point of view, it seems
points. that the model range of validity, is centered a few 1/3
4. These points are then isolated from the whole set of octave lower than the model optimization frequency.
points; a new subsystem is then created.
5. The above sequence is reiterated and ends up when INVERSE SEA PROCESS
all points have been affected to a subsystem.
As soon as an appropriate data reduction has been
performed over the now defined subsystems, the
At this stage a series of subsystem sets is available but coupling scheme and associated SEA parameters of the
the partition still needs to be improved. model can be identified [13-16]. The model identification
is detailed in appendix. The experimental SEA engine of
Optimization the SEA-XP software [14] was used for this process.
th
Different techniques have been tested to perform an This software developed by InterAC in early 90 to
optimization of the sub-structuring scheme: predict energy transfer in rocket engine at lift-off [17], has
been widely used up to now by automotive industry to
• Statistical entropy techniques tries to minimize an validate SEA models of car bodies and train [18-19]. ²It
entropy function expressed -as defined in the includes specific features required to take over the
Shannon’s theory of information- under the generic structural complexity of a car structure such as a Monte
following form: Carlo matrix inversion or connection scheme design
tool..

In practice, most of the car body structure is far from the

homogeneous conditions ideally handled by theoretical
SEA as input mobilities and transfer functions exhibit
where pn is the probability of an event and n the high degree of spatial variance. The parametric
number of events. The events can be the point identification of these high-variance SEA systems, is
membership of a given set. thus performed in SEA-XP by a dedicated matrix solver;
the Monte Carlo solver.
• Attraction force techniques rely on an attraction force
between a point and a set of points; each point The Monte Carlo solver randomly perturbs energy
should belong to the most attractive set. The matrices –according to the variance of spaced and
attraction force is expressed from the energetic frequency averaged Energetic Transfer Functions- and
transfer functions. filtered out unwanted sets of equivalent mass and
Coupling Loss Factors (non-physical negative values).
In statistical entropy techniques, a point is moved This improves the identification process by considering
randomly from one set to another until the entropy possible data sets, slightly different from the original –
reaches a minimum; another point is then randomly uncertain- one.
chosen and the process is reiterated as long as
necessary. The connection scheme design consists in the iteration
of two steps: selecting connection involving power flow
The attraction process works in a similar way but is higher than a given threshold, and checking the quality of
faster as a point can be moved from one set to another the associated identification process.
following maximum line forces. Unfortunately, this
optimization process may never converge, some of the Finally, the resulting SEA model is described in a neutral
points remaining definitely unstable. Further work is file, including sub-structuring and connection schemes,
required to identify robust convergence criteria. equivalent masses and modal densities of subsystems,
Coupling Loss Factors; the Damping Loss Factor is
In practice, the iterative algorithms are improving the uniform and constant as it was in the computations.
sub-structuring very rapidly and a subsystem set can be Actual values of damping are generally tuned and
identified within a few minutes. Providing such an modified later as in any SEA application. SEA-XP was
optimized sub-structuring scheme, independent of the found quite convenient as it can export created
user expertise, is one of the main advantage of the experimental models to AutoSEA in both version 1 and 2
proposed technique. format.
 system was used in order to avoid sensors added mass
perturbations at high frequencies (here above 1 kHz
APPLICATION TO THE MINIVAN FLOOR using standard equipment). At least 20 discrete points
were measured on each panel, including stiffeners when
PRINCIPLES OF THE APPLICATION they are part of the subsystems. The input mobility was
also measured in order to be able to compute the
Virtual SEA has been applied to the floor structure of a Energetic Transfer Functions.
minivan (See figure 4). The numerical results have been
compared to measurements and an AutoSEA2 model. Figure 5 shows the measured velocity maps in the third
Moreover , the effect of the trim on the floor structure has octave 1 kHz, for the bare (a) and trimmed (b) floor.
been model by modifying the DLF of the SEA These measurements show how the trim influence
subsystems touched by the trim. The results are also decreases with frequency: this is interpreted as an effect
compared to the experiments. of the decreasing coupling between the structure and the
foam layer of the insulating complex.
The refined mesh of the minivan floor contains about 600
000 degrees of freedom, whereas the “NVH mesh” only
consists in 500 nodes. The Transfer Functions have
been computed at about 300 frequencies between 180
Hz and 2250 Hz, covering 1/3 octave bands between 200
Hz and 2 kHz.

(a)

(b)
Figure 5: 1kHz third octave averaged velocities map measured over
(a) the floor; bare (a) and trimmed (b).

FLOOR STRUCTURE

As it was described in a previous paragraph, the Virtual

SEA prototype software allows the user to perform the
sub-structuring process in a particular third octave band.
In the present case, two Virtual SEA models have been
computed to cover the 200 – 2000 Hz range: the first one
for the 500 Hz third octave band and the second one for
the 2000 Hz third octave. The two models respectively
cover the 200 – 800 Hz range and the 500 - 2000 Hz
(b) range.

Figure 4. Top (a) and bottom(b) views of a minivan floor structure Few iterations of the Virtual SEA process are necessary
mesh to compute a robust sub-structuring. Figures 6 and 8
show these results on the “NVH meshes”. A different
color marks each SEA subsystem. Figures 7 and 9
present the respective AutoSEA1 model deduced from
the two sub-structuring analysis and associated coupling
MEASUREMENTS
schemes.
Experiments have been carried out with an artificial
At 500 Hz the numerical tool predicts a structure divided
excitation (electro-dynamic shaker) applied to the back
into 13 SEA subsystems, whereas 14 subsystems are
hook supporting the exhaust line. Only the panels
detected to model the minivan floor at 2000 Hz: the
vibration were measured. A LASER vibrometer scanning
differences of sub-structuring clearly appear. Moreover it
confirms the number of subsystems increases with the
frequency.

Figure 9. AutoSEA1 model: result of the 2000 Hz sub-structuring

Figure 6. Sub-structuring at 500 Hz

Before comparing experiments and Virtual SEA results,
an analytical AutoSEA2 model was built from the initial
finite element model, by the most experimented of the
authors, and with the knowledge of the optimal sub-
structuring scheme. This model is shown on figure 10,
providing a clear view of the subsystem position.

Figure 7. AutoSEA1 model: result of the 500 Hz sub-structuring

Figure 10. Analytical AutoSEA2 model

The same DLFs have been used for both analytical and
Virtual SEA modeling. Only two distinct DLF were
considered: one for the frame components and one for
the panels. These DLFs both decrease as a constant
power of the frequency.

The figures 11, 12, 13 and 14 present the comparisons

between the measurements, the AutoSEA2 results and
the Virtual SEA results, for the 4 panels indicated on
figure 10.

Above 500 Hz, the deviation between the measurements

and Virtual SEA results does not exceed 2dB, whereas
Figure 8. Sub-structuring at 2000 Hz the AutoSEA2 results overestimate the measurements
as soon as the frequency is low and subsystems are far
away from the excitation point.

Only based on the results of a numerical simulation, the

Virtual SEA approach predicts the experimental data with
differences that are acceptable.
 Figure 14. Comparisons of velocity for subsystem S58

TRIMMED FLOOR

The second case consists in evaluating the effect of a

trim on this minivan floor structure vibration. The
objective of this test is to show that nearly the same
structural SEA model can be used in both cases. The
sound insulation covering only modifies the DLF of the
panel. All other SEA parameters (equivalent masses,
CLF…) are kept constant.

Figure 11. Comparisons of velocity for subsystem S47 Figures 15, 16, 17 and 18 present the comparisons
between the experimental data and the Virtual SEA
predictions. Again, the effect of the trim on the vibratory
behavior is well predicted by the Virtual SEA technique.

Figure 12. Comparisons of velocity for subsystem S53

Figure 15. Reduction velocity for subsystem S47

Figure 13. Comparisons of velocity for subsystem S56

Figure 16. Reduction velocity for subsystem S53
Figure 17. Reduction velocity for subsystem S56 stochastic input data in order to provide probabilistic
results.

To end with, it is mentioned that associated acoustic

radiation problems have not been considered yet. It has
been assumed that in any case, radiation problems will
end up with a radiation efficiency determination.
Numerous papers deal with radiation efficiencies; from
FE computations to a simplified wave approaches; many
solutions seem well fitted to complete the proposed
approach.

REFERENCES

1. Maxit, L and Guyader, J-L. “Estimation of SEA

Coupling Loss Factors Using A Dual Formulation
Figure 18. Reduction velocity for subsystem S58
and FEM Modal Information, Part I: Theory and Part
II: Numerical Applications” Jnl. Of Sound & Vibration,
(2001) 239 (5) pp.907-948.
CONCLUSION 2. O. Bouthier, R. Bernhard, “Models of space
averaged energetics of plates”, 1992, American
The suggested method, called Virtual SEA, permits to Institute of Aeronautics and Astronautics Journal, 30
build a SEA model based on FEA results. A large
616-623.
energetic transfer matrix is computed, providing a wide
3. R. Langley, and P. Bremner, "A hybrid method for
observation of the vibratory behavior of the studied
structure. This matrix is then processed similarly to the vibration analysis of complex structural-acoustic
experimental SEA, in order to produce an “optimized” systems", Journal of the Acoustical Society of
SEA model of the studied structure. Such an SEA model America, 105(3), 1999, 1657-1671.
allows some understanding of the medium/high 4. A. Le Bot, “A vibroacoustic model for higj frequency
frequencies vibrations in terms of vibratory power flows analysis” Jnl. Of Sound & Vibration, (1998) 211 (4)
between subsystems. pp.537-554.
5. X.Zhao,N. Vlahopoulos,”A Hybrid Finite Element
The main advantage of Virtual SEA is the quasi- Formulation for Mid-Frequency Analysis of Systems
automatic generation of SEA subsystems and Coupling with excitation applied on shord Members”, Journal
Loss Factors. Virtual SEA does not require the a priori of sound and vibration,200, Vol 237(2), pp181-202.
knowledge of the sub-structuring, which considerably 6. C. Simmons, “Structure-borne sound transmission
reduces the required expertise and the associated through plate junctions and estimates of SEA
manpower. coupling loss factors using the FE method”, Journal
of Sound and Vibration, 144,215-227, 1991
The method was assessed in the case of a minivan floor. 7. S Finnveden, “Finite element techniques for
Computations and measurements were carried out evaluation of energy flow parameters”,
separately and never interfered, except for the damping Proceedings of NOVEM 2000, 2000.
estimation. An analytical SEA model was finally build in 8. Soize C., “Reduced models in the medium frequency
order to reflect today’s practice. The Virtual SEA results range for general dissipative structural-dynamics
are globally much better than analytical SEA results.
systems”, European Journal of Mechanics –A/Solids,
They follow the global trends even at the lowest
Vol 17, N° 4, 1998
frequencies and for subsystems distant to the source.
The trimmed structure has also been modeled using the 9. Meerbergen K., “The solution of parametrized
same SEA network, by increasing the damping of the symmetric linear systems and the connection with
panels. The prediction of the trim’s effect on the floor model reduction”, Technical report KM-2000-2, Free
vibrations is in good agreement with measurements. Field Technologies, 16, place de l’Université, 1348
Louvain-la-Neuve, Belgium. Revised.
This paper presents the current state of achievement of 10. Bennighof J.K., Kaplan M.F., Muller M.B. and Kim
the method, but numerous improvements can be M., “Meeting the NVH Computational Challenge:
expected. First, numerical simulation does not stop Automated Multi-Level Substructuring”, Proc
providing better and quicker results. Then, the automatic International Modal Analysis Conference XVIII, San
sub-structuring that is not yet fixed, will progress. Finally, Antonio, February 2000.
stochastic computations could be naturally introduced, 11. Mace B.R and Shorter P.J., “Energy Flow models
since the model identification process already considers from Finite Element Analysis”, Journal of Sound and
input data uncertainties. This last improvement will be Vibration (2000), 233(3), 369-389.
effective when SEA software environment will support 12. R. H. Lyon, “Statistical Energy Analysis of Dynamical
Systems”, MIT press, 1975.
13. Lalor, N. “The Experimental Determination of subsystems has to be equal to the power injected. In a
Vibrational Energy Balance in Complex Structures” frequency band centered on pulsation ωo, this energy
Paper 108429 Proc. SIRA Conf. on Stress & Vibn., balance result is expressed as:
London 1989.
14. SEA-XP User Guide Manual, InterAC Sarl, Version N
2.5 , May 2004. Pi = ∑η ω E ji (A2)
15. M. Rosen, G. Borello, “Damping and coupling loss j =1
factors estimation in SEA method: what is really
measured ?”, Internoise 96. where,
16. Cimerman B.P., Bharj T. and Borello G, “Overview of
Experimental Approach to Statistical Energy Pi designates the power injected at an excitation point in
Analysis” Paper 97NV169, Proc. SAE Noise & Vibn. the sub-structure i (computed),
Conf. Traverse City 1997.
17. –G. Borello, Identification of SEA coupling loss η the internal loss factor (input data)
factors on a liquid rocket engine, INTERNOISE -
1991 – Sydney, Australia E ji is the subsystem j energy when subsystem i is
18. –G. Borello, System model of a High Speed Train
excited according to the formula:
passenger coach using SEA and prediction in
working conditions, INTERNOISE - 1995 - Newport 2
Beach, California E ji = M j v ji (A3)
19. –G. Borello, Prediction and control of structure borne Ωj
noise transfers in vehicules using SEA,
EURONOISE – 4-7 October, 1998 – Munich, where the quadratic velocity is estimated as:
Germany
1
∑v
2
v ji ≅ 2
jik
Ωj Nk k

APPENDIX: EXPERIMENTAL/VIRTUAL SEA

MODEL IDENTIFICATION When each subsystem is excited, the following linear
system can be derived from the balance equation (A2),
The “experimental” SEA process described in [13, 14, invoking the uniform DLF, η, use in the computations
15], is shortly recalled here. Moreover, its application in and definition (1) of the energetic transfer functions:
the case of Virtual SEA that slightly differs from the
previous will be shown.
 1   H 11 ... H N1   M1 
The theoretical basis for the data reduction performed      
during the SEA model identification is described in the  ... ...  ... 
overview paper by Cimerman, Bharj and Borello [16]. 1    
 M  (A4)
1 =  ... H ji ... 
ηω      i
The virtually-derived SEA parameters are:  ...  ...   ... 
    
 1   ... ... H NN  MN 
· Subsystem “equivalent” mass (volume) M j [16]
· Junction coupling loss factors η jk and η kj
The vector of the equivalent mass is obtained by
Compared with experimental SEA, it may be noticed that inverting matrix H:
the Damping Loss Factor is known (input data) and that ++
the input power is exactly computed.  M1  H11 ... H N1  1
     
In a condensed form, the balance equation of the overall  ...   ...  ...
structure can be written as:  M  = 1  ... H ji ...   1  (A5)
 i  ηω    
LE=P (A1)
 ...   ...  ...
   ...  
MN   ... H NN  1
where L is the loss matrix, E the energy vector
where ++ indicated a pseudo-inverse provided matrix H
and P the input power vector. is generally rectangular, numerous excitation points
being used for each subsystem.
Equivalent mass determination
It should be noticed that the modal mass can be
By considering the energy balance of the total studied calculated without any assumption about the subsystems
system, the sum of the energies dissipated in each of the connection scheme. In addition, the energetic transfer
matrix is well conditioned: the diagonal terms are indeed The final expression of the CLF, describing the coupling
predominant. between subsystem i and the Ni-1 connected
subsystems appears as:
In practice, equation (A5) can lead to negative equivalent −1
masses in some 1/3 octave bands. This is the sign of an E E E N iα E Nii   Pi 
 αα − αi ... −   
inappropriate sub-structuring regarding the assumed
 ...   Eiα Eii E iα Eii   ω 0 Eii 
energetic balance. In other terms, it means it does not    Ekk Eki   Pi 
exist any physical SEA model based on the proposed η ki  = ... − ...   
 ...   Eik Eii   ω 0 Eii 
sub-structuring. Compared to previous application, the   k∈{α , N }, k ≠i  EαN i Eαi E N i N i E N ii   Pi 
use of the automatic sub-structuring process in Virtual E − E ... −
Eii 
 
SEA almost provides positive answers. Nevertheless,  iN i ii EiN i  ω 0 Eii 
specific inversion protocols are employed to avoid such (A8)
non-physical results (Monte-Carlo methods [14]). This problem has to be solved for each SEA subsystem,
in order to get the total set of CLF’s needed by the
Coupling loss factors (CLF) determination proposed coupling scheme.

Contrary to experimental SEA, where CLF can be This system of equations can produce a few negative
computed independently from DLF, Virtual SEA imposes CLF values in some frequency bands. This indicates that
the equivalent masses to be computed before the CLF. the proposed model –including coupling scheme,
These masses are used to compute the Energy Transfer estimated subsystems energy and SEA assumptions- is
matrix using expression (A3). not physical. This difficulty is overpass first by taking
advantage of the multiple excitation of each subsystem,
The CLF calculation is performed from an analysis of and formulating a least square problem, and second by
exchanges between subsystems. The balance equation using a Monte-Carlo inverse process as mentioned for
for subsystem i, subjected to excitation, is expressed as: the equivalent masses determination.

Again the use of the automatic sub-structuring technique

1  Ni  Ni
= η + ∑η ij  M i H ii − ∑η ji M j H ji (A6)
improves a lot the identification process.
ω 0  j ≠i

 j ≠i

For subsystem k, where no excitation applies, the

balance equation is:
 Ni  Nk
0 = η k + ∑ η kj M k H kk − ∑ η jk M j H jk
  (A7)
 
 j ≠k  j≠k

Ni and Nk represent the total number of couplings

between element i and element k as define in the
coupling scheme.
These balance equations consist in a set of N linear
equations, for each excited sub-system. Applying a load
successively on each subsystem, leads to a total of N*N
equations, involving all CLF.
This linear system is solved using the condensation
technique proposed by Lalor [13]. This technique
transforms the coupling coefficient calculation into the
successive resolution of smaller linear systems which
size is Ni-1 where Ni is the number of subsystems that
are coupled to subsystem i, according to the coupling
scheme.

View publication stats