Professional Documents
Culture Documents
ks
.
ks
is the Kronecker delta:
ks
= 1 if k = s, and
ks
= 0 if
k = s.
Most materials involved in geomechanical applications
behave nonlinearly [12]. Even for a nite element formu-
lation, resulting in a discretized system of linear constitutive
equations, the nonlinearities will arise when solving inverse
problems. In general, the constitutive relations depend non-
linearly on the material and model parameters. The demand of
our estimation task, the inverse formulation, is to nd material
parameters incorporated within the constitutive laws from e.g.
the known displacements. The nite element formulation for a
general static problem is given by a relation between applied
loads and displacements:
K(x) u = f, (2)
where:
K is the stiffness matrix, dimension l l;
x is the state vector associated with the material
model used, as in equation (1) x has dimension
n 1;
u is the displacement vector, dimension l 1;
f is the load vector, dimension l 1.
The model response is observed at a nite number of
nodal points within the nite element model. Measurement
data in our case can involve displacements, stresses, pore water
pressures, etc. The choice of the type of measurements depends
on the particular geotechnical problem and the available mea-
surement tools. In this paper, displacements at a nite number
of nodal positions are chosen to be compared with measure-
ment data. The reason for this choice is that displacement
(settlement) is the quantity of interest in many geotechnical
problems and hence due to the well developed equipment for
measurement of settlements it would be possible to obtain real
data as input for the model parameter identication procedure.
In agreement with the measurement data, we take the
similar quantities from the FEA as observation data. The
observation vector at observed positions at time t
k
is related
with the displacement vector in the following manner
y
k
= Tu
k
+ v
k
, (3)
where:
y
k
is the observation vector (measurement), dimen-
sion m1;
T is the transfer matrix, dimension ml;
v
k
is measurement noise vector caused by uncertain-
ties in measurement equipment, dimension m1.
Observation noise is also white noise having
zero mean E(v
k
) = 0 and a covariance matrix
E(v
k
v
T
s
) = R
k
ks
.
From equation (2) we have u
k
= K
1
(x
k
)f. Thus,
equation (3) becomes
y
k
= TK
1
(x
k
)f + v
k
. (4)
The stiffness matrix, K(x
k
), is generally a non-linear
function of the material parameters x
k
and for a compact
presentation here we express the equation (4) in a more general
form:
y
k
= h(x
k
) + v
k
. (5)
For simplicity, the system noise w
k
and the observation
noise v
k
are assumed uncorrelated. Furthermore, it is assumed
that the parameter uncertainty is independent. The same ap-
plies to observation uncertainties. Thus the covariance matrices
Q
k
and R
k
are diagonal.
The state equation (1) and the observation equation (5)
are the basic equations representing the state-space system.
Our goal is to estimate the state vector of the nonlinear
system based on the knowledge about the modelled system
and observation availability which are both contaminated by
uncertainties. If the system is linear, the standard linear Kalman
lter is the best estimator to be used. For our case of
highly nonlinear system (geomechanical structure), a nonlinear
Kalman lter (extended Kalman lter) is required. Unlike the
standard Kalman lter, the extended Kalman lter linearizes
the system at the current state. The extended Kalman lter is
presented in the next section based on the state-space system
formulation given in this section.
II. EXTENDED KALMAN FILTER (EKF) - LOCAL
ITERATION PROCEDURE INCORPORATED WITH FINITE
ELEMENT ANALYSIS
First we introduce denition of a priori estimate and a
posteriori estimate. A priori estimate is the estimate of the
state before observation data are available. Each time the data
observation occurs, a priori estimate will be corrected to the
new state, which is called a posteriori estimate, that better
represents the system. In mathematical expressions, a priori
state estimate (estimate) and its covariance matrix at time t
k
are estimated as follows given the observation data up to time
t
k1
:
x
k
= E [x
k
|y
1
, y
2
, , y
k1
] ;
P
k
= E
(x
k
x
k
)(x
k
x
k
)
T
.
Whenever observation data at time t
k
are available, a posteriori
estimate and its covariance matrix can be calculated as follows:
x
+
k
= E [x
k
|y
1
, y
2
, , y
k
] ; P
+
k
= E
(x
k
x
+
k
)(x
k
x
+
k
)
T
.
The estimation procedure should be initiated by best initial
knowledge of the considered geomechanical model. As initial
condition, the initial state estimate and the covariance of the
estimation error should be available:
x
+
0
= E(x
0
); P
+
0
= E
(x
0
x
+
0
)(x
0
x
+
0
)
T
.
The initial state, which consists of model parameters, is chosen
based on engineering experience or preliminary examination
of the structure. The closer the initial parameter values to the
true ones are, the better it is to begin the EKF. But this is
not necessary. The initial covariance matrix is assigned rather
arbitrarily large due to the lack of condence in the initial
choice of the state vector.
Before observation data of the model are available, the EKF
propagates the mean and error covariance of the state through
time. The time update equations of the mean and covariance
of the state are calculated in the following way:
x
k
= x
+
k1
; P
k
= P
+
k1
+ Q, (6)
where x
k
is the mean of state, x
k
= E(x
k
), and
P
k
is the estimation error covariance matrix, P
k
=
E
(x
k
x
k
)(x
k
x
k
)
T
at time t
k
. The superscript
denotes a priori estimate, and the superscript + denotes a
posteriori estimate.
As soon as the observation data calculated by FEA are
available, observation update procedure is applied. We apply
here the local iteration scheme for the observation updates in
order to continuously get better a posteriori estimate without
calculating time update. This local iteration observation update
can be repeated as many times as desired, but for most systems
signicant improvement of the state estimate is obtained by
only one iteration [14]. At time t
k
in the local iterative loop
i of the EKF - local iteration procedure, the Kalman gain
K
k,i
, observation update of the state x
+
k,i+1
and estimation
error covariance matrix P
+
k,i+1
are calculated according to the
following equations:
H
k,i
=
h
x
x
+
k,i
K
g
k,i
= P
k
H
T
k,i
(H
k,i
P
k
H
T
k,i
+ R)
1
(7)
P
+
k,i+1
= (I K
g
k,i
)P
k
x
+
k,i+1
= x
k
+ K
g
k,i
y
exp
h( x
+
k,i
, 0) H
k,i
( x
k
x
+
k,i
)
x,k
)
= E[Tr(
x,k
x,k
)
T
]
= Tr(P
k
) (8)
In a stable system, the EKF always attempts to minimize
the cost function J. The state error covariance matrix P is
enlarged at observation step k, which means that an amount
of pseudo-error is added to the lter. This means we make EKF
think that its current state estimate x
k
is too far away from
the true state x
k
. The new Kalman gain K
k
, equation (7), will
also be enlarged with enlarging P
k
. The updated observation
equation in turn uses K
k
to scale up the innovation (y
k
h( x
+
k,i
, 0)H
k,i
( x
k
x
+
k,i
)) and adds it to the a priori estimate
x
k
. Thus, state variable estimate x changes more quickly, and
this makes convergence rate of the lter faster.
An independent treatment of the FEA program provides
a signicant advantage to the proposed method, in the sense
that any FEA program can be incorporated into the process
after the communicative interface between the main algorithm
ow and the FEA program is created. In this paper, the main
algorithm ow and the communicative interface are realised
via MATLAB computing language and Python scripts, whereas
the ABAQUS FEA software is used for performing the forward
simulations.
In order to begin the lter process, an initial value x
0
for
the state vector along with the corresponding estimation error
covariance matrix P
0
have to be assigned. Usually, the initial
state values are taken to be the most likely parameter values
that can be inferred from prior knowledge or preliminary
geomechanical tests. Values of the corresponding estimation
error covariance matrix explains how much condence the
lter designer may have on the geotechnical model under
consideration.
The effects of the process noise w
k
and the observation
noise v
k
are omitted from system equation and observation
equation for the sake of simplicity. However, their correspond-
ing covariance matrices Q and R remain in the equations of
time update and observation update of the EKF process. These
covariance matrices are written without subscript k, which
indicates that they remain unchanged over iterations.
The calculation of observation vector derivatives with re-
spect to the state values at the current state estimate, H
k,i
,
is crucial as the observation derivative matrix determines the
direction in which the state estimate x
k,i+1
progresses. Cal-
culation of the observation derivative matrix is approximated
by forward 2-points nite difference employing the forward
solution of FEA. Sensitivity matrix H is calculated for each
global iteration k and local iteration i.
The iterative procedure explained consists of a local it-
erative loop for calculating observation update and a global
iterative loop for the ltering process. The local iterative loop
iterates a certain amount of times according to user setting and
the global iterative process will continue until a stop criterion
is met. Possible stop criteria are: 1) cost function achieves
Layer 1
Layer 2
Layer 3
S
y
m
m
e
t
r
y
l
i
n
e
Pressure
Fig. 1. Finite element mesh and observation positions for the plane strain
footing structure
TABLE I. TRUE ELASTIC CHARACTERISTIC VALUES FOR THE THREE
SOIL LAYERS
Parameter Layer 1 Layer 2 Layer 3
E[N/m
2
] 2.0e7 1.0e7 0.5e7
0.3 0.2 0.4
a predened minimum value; 2) absolute difference of two
consecutive estimated state vectors is less than a predened
tolerance; 3) the specied number of iterations is exceeded.
In this paper, the third stop criterion is selected since the
number of iterations is usually small and it also serves the
purpose of examining the lter process at some later time after
convergence has been reached.
III. RESULTS AND DISCUSSIONS
A. Strip footing load on elastic soil - Identication of elastic
constants
The rst geotechnical problem to be used as an example is
the strip footing on an isotropic linear elastic soil. The plane
strain model with three layers is considered. The model takes
its horizontal limit of 2 m and the vertical limit is 0.5 m depth.
A strip load of 2 MPa is applied on the soil surface over
the length of 10 cm. Due to the symmetry of the problem
only a half of the structure is considered for the nite element
analysis and a symmetry boundary condition is imposed along
the symmetry line as shown in Figure 1.
The soil is assumed to obey an isotropic linear elastic
material law. A set of predened model parameters serves
the purpose of synthetic generation of measurement data by
numerical simulation. This set is considered as the exact and
true soil model parameter set and it is used as a basis for
verication of the nal estimated parameter values. The true
elastic soil parameters are given in Table I.
The observation positions are selected at the dotted points,
as shown in Figure 1, and they are grouped into observation
group 1 and observation group 2. Observation group 1 consists
of 4 (round-dotted) points located along the model symmetry
line. Observation group 2 consists of 4 (square-dotted) points
located at the same depths as the observation points of group 1
and 10 cm apart from them in horizontal direction. At points
from observation group 1, because of the symmetry boundary
condition, only vertical displacements are used. At points from
group 2, both vertical and horizontal displacement components
are observed.
The observation positions are intentionally chosen at depth
levels that are distributed to all the three soil layers. Thus we
may evaluate the sensitivity of observation data with respect
to model parameters involved in modeling the behavior of
the different soil materials composing the examined structure.
Different combinations of observation data will be selected to
test the inuence of the number of measurement or observation
points and their location within the model on the convergence
rate of the EKF method and its effectiveness.
Successful identication of the total six material model
parameters should prove the capability of the EKF to identify
a large number of parameters. The assessment of the capacity
of the EKF to model identication for this example is done
by means of three different tests:
test 1: all observation data from all groups are taken into
account, i.e. four vertical displacements of observation group 1
and four vertical and four horizontal displacements of obser-
vation group 2 (total number of observation data is twelve);
test 2: only observation data from group 1 and the vertical
displacement components from observation group 2 are taken
into account, (total number of observation data is eight);
test 3: only observation data from group 1 is taken into
account, (total number of observation data is four).
We begin with test 1, in which the largest number of
observation data is considered. We want to give the easiest
condition for the EKF to work by feeding it with as much
information as the measurement set-up can offer. Following,
in test 2 and test 3, we constrain the EKF by providing less
and less number of observation data to it. As a result, we
expect to assess the capability of the EKF to identify elastic
soil parameters when the amount of measurement data vary.
When varying the number of observation data we still keep
the strategy of spreading the observation positions to all the
three elastic layers.
The state value vector in this example reads:
x = [ E1 1 E2 2 E3 3 ] ,
where E1 and 1 are Youngs modulus and Poissons ratio of
elastic soil layer 1, E2 and 2 are those of elastic soil layer 2
and E3 and 3 are those of elastic soil layer 3.
The EKF in test 1 converges to a solution after about 15
iterations and the rate of convergence is examined for three
different sets of initial values for the state values. In order
to cover wide range of initial state values for the EKF, the
sets of an extreme initial state values are considered. In the
rst case the EKF begins with very small Youngs moduli
(E1 = E2 = E3 = 0.5 10
7
N/m
2
) and very large Poissons
ratios (1 = 2 = 3 = 0.45), whereas in the second case very
large Youngs moduli (E1 = E2 = E3 = 3.0 10
7
N/m
2
)
and very small Poissons ratios (1 = 2 = 3 = 0.1) are
considered. In the third case, a random predicted initial set
of parameters in between the two extreme cases is chosen
to initialize the KF process, namely (E1 = 1.0 10
7
, 1 =
0.1, E2 = 2.0 10
7
, 2 = 0.3, E3 = 1.5 10
7
, 3 = 0.2).
Fig. 2. KFiteration result - Test 1 (number of observations 12), initial
state value case 1
Fig. 3. KFiteration result - Test 1 (number of observations 12), initial
state value case 2
Figures 2 and 3 show EKF converging behavior in test 1
for two different cases of initial state values. In both cases,
the elastic soil parameters are identied after 16 iterations.
Parameters of the upper soil layer converge faster to their
true values compared to those of the lower one. It means,
more information may be captured observing the upper surface
of the structure and this is reasonable because settlement
measurements at places close to the surface of the strip footing
are most sensitive to the soil material characteristics.
In test 2, we use only the initial state value case 2 and run
the KF with eight observations. The progressing of the EKF
parameter identication process depicted in Figure 4 shows
that convergence rate is much slower than it is in test 1. For the
worst converging parameter, Poissons ratio of the rst layer
1, over 40 iterations are required for the EKFprocess to
converge. The same initial state parameters are taken for test 3,
where only four vertical displacements are taken as observation
data. The state value estimates converge to wrong values as can
be seen in Figure 5 and Table II where the deviation of the
identied parameter values from the true ones is also given.
The results from the three tests of EKF application to
linearelastic strip footing problem have proved that the more
observations are used, the better the convergence is and more
accurate the estimated state is. However, to collect more mea-
surement data more equipment and human effort are required
for a real construction site. The authors recommend that one
should take the number of measurement a few more than
the number of identied parameters for easy convergence.
Choosing observation locations where displacement responses
are sensitive also helps to reduce the number of observation
locations. Murakami et al. [6], [15] proposed a method for
computing sensitivity distribution of the observed displacement
for a linear analysis of an elastic soil medium.
TABLE II. EKFCONVERGENCE PERFORMANCE AND QUALITY OF
THE SOLUTION TEST 3
Parameter Initial state Estimated state Deviation from true state
E1 3.0 10
7
2.0891 10
7
4.45%
1 0.1 0.0246 91.8%
E2 3.0 10
7
0.9842 10
7
1.58%
2 0.1 0.2283 14.15%
E3 3.0 10
7
0.6457 10
7
29.14%
3 0.1 0.2751 31.22%
B. Retaining wall - Identication of Mohr-Coulomb elasto-
plastic constitutive parameters
An identication of Mohr-Coulomb constitutive model
parameters is carried out employing an example of a 4 m
high retaining wall structure as depicted in Figure 6. A 30 kPa
pressure and a 500 kPa pressure are applied on the surface of
the backll over the lengths of 6 m and 3 m and the problem is
considered as 2D plane strain problem. The backll structure
dened in this way is similar to the one discussed in [16], but
with modied geometry, materials and loads.
The Mohr-Coulomb constitutive model of the soil backll
is assumed to have a friction angle = 37
, a dilation angle
= 5
, where
friction angle and dilation angle are assumed as known
in order to generate the synthetic measurement data. Friction
angle and dilation angle are 37
and 5
respectively.
The EKF estimation process is run with only one local
iteration. One single setting for noise covariance matrices Q
and R and weight factor W for both initial state cases has
Fig. 5. EKFiteration result - Test 3 (number of observations 4), initial
state value case 2
Fig. 6. Soil backll supported by a concrete wall
shown difculty in obtaining fast convergence. Thus, these
settings are adjusted according to each initial state case.
The EKFconvergence performance was tested using dif-
ferent initial conditions sets. Converged parameter estimates
for initial case 1 are obtained after 15 iterations and as it
can be seen from Table III they show quite large deviation
from the true parameter values. The identied parameter values
for initial case 2 indicate faster EKFconvergence and less
discrepancy from the true parameter values. In table IV the
converged results in case 2 are obtained after 6 iterations.
TABLE III. EKFCONVERGED PARAMETER ESTIMATES AFTER 15
ITERATIONS INITIAL CASE 1
Parameter Initial state Estimated state Deviation from true state
40 37.32 0.32
10 4.66 0.34
Fig. 7. Loading history for the retaining wall example
Fig. 8. Displacements at observed points over simulation time
IV. CONCLUSIONS
Good convergence of the parameter estimation process
demonstrated through various geotechnical problems presented
in this work has proved that the EKF incorporated with
modern FEA program has prospects to be a reliable method
for parameter identication in geotechnical engineering. The
important advantageous feature of the EKF method, namely
the utilisation of the measurement uncertainties, which can-
not be avoided in large scale geotechnical structures, makes
this method very attractive and powerful. In addition, EKF-
procedure performs well even in the environment corrupted by
noise and due to the underlying characteristic of the algorithm,
the state-value vector estimate (i.e. identied parameters)
represents the values that best minimize the estimation error
covariance.
The implementation of the EKF incorporated with FEA
TABLE IV. EKFCONVERGED PARAMETER ESTIMATES AFTER 6
ITERATIONS INITIAL CASE 2
Parameter Initial state Estimated state Deviation from true state
40 36.98 0.02
10 5.02 0.02
with regard to implementation efforts against the resulting
outcome is relatively simple but still presents a very powerful
tool since using modern programming languages the EKF
algorithm can be realized within several code lines including
vector and matrix manipulations. However, the tuning and ad-
justing the EKFconguration plays a decisive role regarding
its successful implementation and therefore requires a special
attention. Good understanding of the model and the inuence
of each parameter on the model is crucial for the conguration
of the EKF in order to achieve its successful convergence.
ACKNOWLEDGMENT
The authors gratefully acknowledge the funding by the
German Research Foundation (DFG) within Collaborative Re-
search Center under grant SFB-837.
REFERENCES
[1] Gioda G and Sakurai S. Back analysis procedures for the interpretation
of eld measurements in geomechanics. Int. J. Numer. Anal. Meth.
Geomech. 1987; 11:555583.
[2] Gioda G and Locatelli L. Back analysis of the measurements performed
during the excavation of a shallow tunnel in sand. Int. J. Numer. Anal.
Meth. Geomech. 1999; 23:14071425.
[3] Schanz T, Zimmerer MM, Datcheva M, Meier J. Identication
of constitutive parameters for numerical models via inverse approach.
Felsbau - Rock and Soil Engineering 2006; 24 (2) :1121.
[4] Kalman RE. A new approach to linear ltering and prediction problems.
Trans. ASME Ser. D: J. Basic Eng. 1960; 82 (1) :3545.
[5] Masaru Hoshiya and Atsushi Suto. Kalman lter - nite element method
in identication. Journal of Engineering Mechanics 1993; 119 (2) :197
210.
[6] Murakami A. Studies on the application of Kalman ltering to some
geomechanical problems related to safety assessment. Doctoral disserta-
tion. Kyoto University, 1991.
[7] Wang Yang. Extended Kalman lter method for parameter identication
and its application in tunneling. Doctoral dissertation. Universit at
Innsbruck, 2004.
[8] Cheng Xiang Yang, Yong Hong Wu, Tung Hon and Xia-Ting Feng.
Application of extended Kalman lter to back analysis of the natural
stress state accounting for measuring uncertainties. Int. J. Numer. Anal.
Meth. Geomech. 2011; 35:694712.
[9] Gelb A (ed.). Applied optimal estimation. MIT Press, Cambridge,
Massachusetts, 1974.
[10] Nestorovi c-Trajkov T, K oppe H, Gabbert U. Active vibration control
using optimal LQ tracking system with additional dynamics. Interna-
tional Journal of Control 2005; 78 (15) :11821197.
[11] Nestorovi c Trajkov T, K oppe H, Gabbert U. Vibration control of a
funnel-shaped shell structure with distributed piezoelectric actuators and
sensors. Smart Materials and Structures 2006; 15:11191132.
[12] Potts DM, Zdravkovic L. Finite element analysis in geotechnical
engineering. Thomas Telford, 1999.
[13] Jazwinski AH. Stochastic processes and ltering theory. Academic
Press, 1968.
[14] Dan Simon. Optimal state estimation - Kalman, H and nonlinear
approaches. John Wiley & Sons, 2006.
[15] Akira Murakami and Takashi Hasegawa. Back analysis using Kalman
lter-nite elements and optial location of observed points Proceedings
of the sixth international conference on Numerical Methods in Geome-
chanics 1988, Innsbruck, 20512058.
[16] Sam Helwany. Applied soil mechanics with ABAQUS applications.
John Wiley & Sons, Inc., 2007.