���������������������������������������������

�����������������������������������������������

“Time-Frequency” domain analysis in nonlinear dynamic soil-structure

interaction
C. Obrembski & D. Clouteau
Laboratoire de Mécanique des Sols, Structures et Matériaux, École Centrale Paris, France
N. Greffet
Électricité de France R&D AMA, Clamart, France

ABSTRACT: Nonlinear dynamic soil-structure interaction problems have to be analysed in the time domain
while most effective linear models are in the frequency domain. In this paper, an approach to nonlinear inter-
action problems is elaborated coupling time and frequency domain. It allows the numerical computation of the
soil stiffness in frequency domain and the residual term due to nonlinear forces in the time domain. The results
given here validate the method and demonstrate its efficiency. Moreover, the importance of the partial uplift of
foundation considering soil-structure interaction is illustrated.

1 INTRODUCTION
In this paper, the method is applied to nonlinear
Due to the increasing demands of French nuclear effect such as uplift and convergence criteria of
safety agency (DGSNR), seismic vulnerability of the method are investigated. Numerical results are
nuclear plants is one of the major topics of research presented and the efficiency of the algorithm is
of Electricite de France (EDF). EDF is looking for checked on a first implementation.
an accurate evaluation of safety margins to assess
the resistance of buildings to earthquakes stronger 2 FORMULATION OF THE PROBLEM
than the design ones. To this end, the associated nu-
merical modelling has to take into account nonlinear In this part, the linear formulation of substructure
behaviour related to both the soil and the structure method coupling an half space and a finite domain
(contact, non elastic material behaviour...) in the time is described within a general concept. Then uplift be-
domain and has to consider the frequency dependent tween both domains is included in the formulation.
dynamic stiffness of soil. Finally, some existing solving methods are described.
A first approach coupling computations in frequency
and time domains has been developed by Kawamoto
(Kawamoto 1983), and applied by Darbre and Wolf 2.1 Linear model
(Darbre and Wolf 1988, Wolf 1988) for simple cases.
The linear soil-structure modelling (Fig 1) is based on
This method, called Hybrid Frequency-Time-Domain
equilibrium’s equation (1) expressed in the structure
method (HFTD), computes a linear response of the
domain and in the soil domain .
structure in the frequency domain and evaluates non-
linear effects in the time domain through the addition Div x f ẍ (1)
of pseudo forces updated at each computation itera-
tion. More recently, Clouteau and Devesa (Clouteau with or for the soil or the structure domain.
and Devesa 2002) have developed another algorithm Moreover, kinematical compatibility conditions (2)
based on similar ideas. However, the solution is not on soil-structure interface have to be fulfilled.
evaluated at each iteration, as only correction terms
are computed. Moreover, the pseudo linear system xs xb on
matrix is evaluated after each iteration to be as close
as possible to the actual tangent matrix in the time
window. xs n xb n on (2)

1255
 F represents internal nonlinear forces in the do-
mains. On the interface, this force is either equal to
the soil reaction force or vanishes where uplift takes
place.

2.3 Review of some existing solution methods

Nonlinear problems are considered as transient
problems, and it is usually assumed that incremental
technics in time domain have to be used.
Nowadays, different methods are available to evaluate
solution of an uplift problem. First, soil domain can
be approximated putting springs and dashpots under
the building foundation (Betbeder-Matibet 2003).
Figure 1. subdomains defi nition This method is not appropriated when soil stratifi-
cation is complex because this stratification cannot
with n the out normal of interface . be taken into account. Moreover, it is a frequency
When the model is discretized with finite elements independent model. It is used in engineering design
for the structure and boundary elements for the half thanks to its simplicity.
space, the vector x of DOF approximating the dis- Another method consists in convoluating the fre-
placements field satisfies formally equation (3) in the quency dependent stiffness with soil ’s displacements
time domain, leading to equation (4) in the frequency to obtain contact forces between soil and foundation.
domain: This method is explained in Wolf (Wolf 1988) and
has been applied by Wolf and Obernhuber (Wolf and
MẌ C Ẋ K X Fext (3) Obernhuber 1984) and by Savidis and Bode (Savidis
and Bode 2002). This method is an incremental
M C K X Fext (4) method and has to keep in memory all the displace-
where M is the mass matrix, C is the damping ma- ments that have been calculated before, but nonlinear
trix and K is the stiffness matrix. This matrix include effect due to uplift can be easily treated step by step.
dynamic impedance of the soil and is thus frequency The main cost of this method is concentrated in the
dependent impling the use of convolution product convolution computation.
in (3).
t
K t X K X (5) 3 TIME-FREQUENCY ALGORITHM
0

The time domain solution is deduced by inverse To solve the considered nonlinear problem, Clouteau
Fourier transform from the frequency solution. and Devesa (Clouteau and Devesa 2002) have elab-
orated a new algorithm called “Time-Frequency” al-
X X X (6) gorithm. Their approach is separated in two comple-
mentary parts. The first one is a classic linear solver
and the second one is a residual and updating evalu-
2.2 Uplift model ation. As in the HFTD methods, stability conditions
are of paramount importance.
Fondation uplift is considered as a slow disturbance
of the main structural dynamic motion. The uplift x
between soil and structure is inserted in the displace- 3.1 Linear initial solution
ment compatibility conditions equation. It is assumed
that x is small enough to keep each contact point in th First of all, solution of the nonlinear problem is ap-
evicinity of the same contact point than at initial time. proximated by a linearized system.

xs xb x on (7) F X Ẋ C0 Ẋ K0 t X (9)
The nonlinear force created by this uplift phe-
nomenon is included in balance equation (3) and a This solution is evaluated with the initial values (K0 ,
new equation has to be solved: C0 ) of the linearized model. In the uplift model of
structure on half-space, C0 and K0 are the damping
MẌ F X Ẋ F (8) and the stiffness matrix considering initial static state

1256 �����������������������������������������������
of equilibrium of the system. The balance equation of
the discretized linearized problem is:
MẌ C0 Ẋ K0 t X Fext (10)
This linear problem is solved in the frequency domain
as:
2
M j C0 K0 X Fext (11)
The first solution obtained is the solution of the lin-
earized problem with a frequency dependent stiffness
matrix (damping can also be frequency dependent)
and with a perfect contact between the soil and the
foundation. This solution is transformed in the time
domain by the classical inverse fast Fourier transform.
X1 X (12)
To obtain the transient solution, the system has to be
initialy in a state of equilibrium and has to be damped
to return to this state. Indeed, using fast Fourier trans-
form the method is limited to periodic solution with
bounded spectrum, leading to solutions vanishing at
the beginning and at the end of the time window. The
first solution corresponding to the linearized system Figure 2. Algorithm of the Time-Frequency Method
is X1 . This solution can be kept in mind to compute
several different nonlinear conditions of the system. K K with (16)

3.2 Residual and updating methods As the sum is limited on the interval , the com-
putation of tangent matrix K is compulsory only on
Given the displacement X on the time window, this time interval.
internal forces can be expressed with the associ-
ated nonlinear behaviour law as in a time integra- 3.3 Algorithm
tion scheme. These internal forces are thus known on
the entire time interval. With these internal nonlinear On figure (2), the algorithm scheme is presented. The
forces, a residual can be expressed, a first part cor- first step is the initialisation. H0 represents the
responding to the frequency dependent system and a transfert matrix where C0 and K0 are accounted for:
second part correponding to the time dependent sys-
tem characteristics: H0 M C0 K0 (17)
The second step is the resolution of the linear prob-
r K C X (13)
lem dX H0 R for each . In the first
loop, dX corresponds to the linear solution X1 . The
R F MẌ Fnl X Ẋ r (14) time residual correction includes the computation of
nl
the nonlinear forces associated with Xi . Finally, the
F is expressed from the solution at last iteration residual term is segmented. The first segment which
and r corresponds to the unbounded soil reaction’s contains non zero value is kept and the other are set to
force. zero.
Because residual term is correct just on its first ap-
pearance, residual term is considered just on some
time step, after it is putted on zero. This cutting of 4 NUMERICAL EXPERIMENTS AND RESULTS
the residue is called windowing. With this residue, a
corrective term is evaluated. The correction can be as- In this section, results on several different systems are
sessed with the initial transfer matrix H0 or with presented. Howerer, no frequency dependent system
an average matrix H computed with the av- has been studied yet. Results demonstrate the effi-
erage value of the stiffness matrix in the same time ciency of the method in accurately approximating the
window as the residual evaluation. nonlinear solution even when severe nonlinearities,
such as impact, appears. These results show some re-
dX H0 R (15) strictions and conditions to the use of this method.

���������������������������������������������� 1257
 5 1 4 1
x 10 x 10 5 10 10
4 x 10
1.5
2
8000
3
1.5
1
6000
2
1

Amplitude du residu fenetre

4000

Amplitude du residu fenetre

Amplitude du residu total

Amplitude du residu total

0.5
1
0.5 2000

0 0 0 0

−1 −0.5 −2000
−0.5

−4000
−2 −1

−1 −6000
−1.5
−3
−8000
−1.5 −2
−4
0 2 4 6 0 2 4 6 0 2 4 6 0 2 4 6
temps (s) temps (s) temps (s) temps (s)

4 30 30
5 20 5 20 x 10
x 10 x 10

1 3
2.5
0.03

2 0.8
2
0.6 0.02
1.5

Amplitude du residu fenetre

Amplitude du residu fenetre

Amplitude du residu total

0.4

Amplitude du residu total

1
1

Figure 3. system description

0.01
0.5 0.2

0 0 0 0

−0.5 −0.2
−1 −0.01
Linear displacements of the 2 masses −1 −0.4

5 −1.5 −0.6 −2
−0.02

−2
−0.8
−0.03

4 −2.5
−1
−3

0 2 4 6 0 2 4 6 0 2 4 6 0 2 4 6
temps (s) temps (s) temps (s) temps (s)

3 40 40 x 10
−5 60 x 10
−5 60

5 5

2 100 100
4 4
displacements [m]

3 3

Amplitude du residu fenetre

1 2 2

Amplitude du residu fenetre

Amplitude du residu total

50 50

Amplitude du residu total

1 1

0 0 0 0 0

−1 −1

−1 −50 −50 −2 −2

−3 −3

−100 −100 −4 −4
−2
−5 −5

0 2 4 6 0 2 4 6 0 2 4 6 0 2 4 6
temps (s)
−3 temps (s) temps (s) temps (s)

−4
Figure 5. Residue for different step of iterations

−5
0 1 2 3 4 5 6 7 8 9 10
time [s]
of the first one. That’s why it is not necessary, and
Figure 4. Linear displacements of the two masses
even recommended to windowed the residue, because
correcting a residue that will be missing in a win-
4.1 Impact of two mass-spring systems dowed scheme could caused longer time computation
The first system studied is composed of two masses or divergence. On figure (6), the “Time-Frequency”
each being supported on a spring and a dashpot, as algorithm has converged after 65 iterations giving the
shown on figure (3). The mass has an initial veloc- same results than as a classical time analysis solu-
ity whereas is at rest. The result in (Fig 4) is the tion. A necessary condition to reached convergence is
linear solution where mass 1 can cross through mass a large time window for the computation. In this ex-
2. Contact is modelised with a penalty technique. The ample, the time window necessary to obtain a conver-
penalty parameters is three hundred times greater gence is . If this interval is reduced to
than the stiffness of the spring. for example, the convergence is not reached.

for 5
Non linear solution

for 3

2
is the gap between the two masses in the state of
displacements [m]

equilibrium. The only source of nonlinearity is the 1

impact between the two masses. From the complete 0

solution, the residual is evaluated over the total time −1

interval. Figure (5) shows the residual for different it- −2

erations numbers. For the same iteration number, on
the left is the not filtered residual, and on the right −3

the filtered residual. It is important to notice that, not −4

only residual changes in time but also in amplitude. −5

0 1 2 3 4 5 6 7 8 9 10
It is a significant characteristic of this algorithm. An time [s]

existing residual term can be corrected just with the Figure 6. Nonlinear displacements of the two masses after con-
correction of a previous existing term. Inversely, the vergence
correction can create a residual, after the occurence

1258 �����������������������������������������������
4.2 Bidimensionnal model

#NOM DE LA FONCTION: depyp3

0.01
lineaire
non lineaire
Aster
0.005
deplacement vertical m

−0.005

−0.01

−0.015 Figure 9. Linear and nonlinear response of the system for a step
of deformation
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
temps s
ure (9), results without uplift are given in white and
Figure 7. Comparison between two methods in displacements.
- the linear solution, o the nonlinear response with the “Time- in grey with uplift. As expected, uplift induces larger
Frequency” method, + the nonlinear response with a classical horizontal and vertical displacements of the structure.
time domain method.
5 CONCLUSION
#NOM DE LA FONCTION: accxp3

4
lineaire “Time-Frequency” method have been shown to be
non lineaire
Aster able to resolve dynamic nonlinear problems with lo-
3
cal nonlinearities. Results given here show the good
2 quality of this method compared to classical time do-
main methods. Numerical efficiency of this technic
Acceleration horizontal m

1
still has to be evaluated more precisely. In the future,
0
the frequency dependency in the soil dynamic stiff-
−1 ness model and the use of a modal decomposition to
−2
reduce the problem size will be accounted for.
−3

ACKNOWLEDGEMENTS
−4

0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5

This research is supported by EDF and Ecole Centrale
temps s Paris. The authors would like to thank Rateau G. for
Figure 8. Comparison between two methods in acceleration. - his participation into the algorithm development.
the linear solution, o the nonlinear response with the “Time-
Frequency” method, + the nonlinear response with a classical
time domain method.
REFERENCES
Betbeder-Matibet, J 2003. L’approximation des ressorts
In this section, the dynamic response of a 2D model périphériques pour l’étude du décollement avec un
under Ricker acceleration is calculated using the modèle de ressorts couplés. VIème Colloque National
AFPS 2003, Génie Parasismique: Aspects Dynamique et
“Time-Frequency” method. The model is composed Vibratoire en Génie Civil.
of four elements in the horizontal direction and eigth
Clouteau, D. & Devesa, G. 2002. Décollement des fondations
elements in the vertical direction. This bidimensional sous séisme : méthodes temporelles et temps/fréquence.
model is resting on five unilateral springs and dash- 45–57.
pots. Thus uplift of the base can occur. The horizontal Darbre, G. R. & Wolf, J P 1988. Criterion of stability and
base displacements are locked, so that sliding is pre- implementation issues of hybrid frequency-time-domain
cluded. The “Time-Frequency” solution is compared procedures for non-linear dynamic analysis. Earthquake
to a classical Newmark-Newton scheme executed in Enginnering and structural dynamics: Volume 16: 569–
581.
Code Aster (EDF ). Both are plotted on figure (7)
EDF. manuel utilisateur code aster. website:
showing the accuracy of the “Time-Frequency” algo- http://www.code-aster.org, software under GPL li-
rithm for the displacements. Horizontal accelerations cense.
(fig 8) are also very similar. Kawamoto, J D 1983. Solution of nonlinear dynamic struc-
To obtain a good convergence, 100 iterations are tural systems by a hybrid frequency-time domain ap-
needed for the “Time-Frequency” algorithm. On fig- proach. Research report R 83-5, Department of Civil En-

���������������������������������������������� 1259
 gineering, Massachussets Institute of Technology, Cam-
bridge, MA.
Savidis, A. & Bode, C. 2002. Soil structure interaction due to
moving loads. Wave 2002: 47–59. Chouw and Schmid.
Wolf, J.P. 1988. Soil-Structure Interaction Analysis in Time
Domain. Prentice Hall. 4eme édition.
Wolf, J P & Obernhuber, P 1984. Non linear soil structure
interaction analysis using dynamic flexibility of soil for
impulse forces. Eighth World Conference on Earthquake
Engineering, San Francisco: 969–976, Volume 3.

1260 �����������������������������������������������