Вы находитесь на странице: 1из 11

50th ASEP Anniversary International Convention & Exposition, “Safer, Smarter & Greener”

Technical Proceedings ISSN 1656-7757, September 28-30, 2011 Makati City, Philippines

ESTIMATING THE LIMIT STATE EXCEEDING PROBABILITY OF A


DETERIORATING STRUCTURE USING THE KALMAN FILTER,
EXTENDED KALMAN FILTER, UNSCENTED KALMAN FILTER AND
THE SEQUENTIAL MONTE CARLO SIMULATION
Lessandro Estelito Garciano and Ikumasa Yoshida
ABSTRACT: This paper focuses on determining the limit state exceeding probability of a
deteriorating model using optimal and sub-optimal Bayesian algorithms. Specifically the
Kalman filter (for a linear system), extended Kalman filter, unscented Kalman filter and the
sequential Monte Carlo simulation (for non-linear systems) are used to approximate the present
state of a deteriorating system given measurements tainted with noise of the system output. In
addition to the comprehensive discussion of the theory, numerical implementation and
comparison of the results through numerical examples are shown.

KEYWORDS: Kalman filter, extended Kalman filter, unscented Kalman filter, sequential Monte
Carlo simulation, limit state exceeding probability

1. INTRODUCTION

There are many instances in civil engineering practice when it is required to estimate the present
state a structure or its component, especially deteriorating ones (Yoshida et al 2009, Garciano
and Yoshida, 2010, 2008). This paper proposes to assess its present state by approximating its
probability of failure just after construction and then updating it using inspection data gathered
from the structure. This scenario of updating a dynamic system using information from new
measurements is ideal for a Bayesian updating approach. When certain conditions hold (linear
and Gaussian) the Kalman Filter (Kalman 1960) is an optimal solution. Often in real-world
problems, non-linearity and non-Gaussian distributions arise so that other methods like the
extended Kalman filter (Anderson and Moore 1979), unscented Kalman filter (Julier and
Uhlmann 1997) or the sequential Monte Carlo simulation (Gordon 1993, Kitagawa 1996) are
used to approximate the posterior density of the state. This paper therefore focuses on
determining the present state of a deteriorating system (by approximating the limit state
exceeding probability) using observation data, tainted with noise, obtained from the structure or
component. The paper is organized into several sections. Section 2 is the formulation of the
stochastic filtering problem, the theoretical Bayesian solution and its optimal and sub-optimal
algorithms that approximate the Bayesian solution. Section 3 illustrates how the method works
through an illustrative example and the discussion of the results and Section 4 are the
conclusions.
2. STOCHASTIC FILTERING PROBLEM
2.1 Optimal Bayesian Solution

Consider the following system model with transition function F and process noise w .
x k = F ( x k −1 , w k ) (1)
At discrete times, measurements become available and are related to the state vector through
the observation equation
z k = H ( xk , v k ) (2)
with measurement function H and measurement noise v. The pdf or probability density functions
p(wk)and p(vk) are assumed known and independent. There is no restriction on the form of the
functions F and H. The prior distribution of the state at k = 0 is assumed known and distributed

137
50th ASEP Anniversary International Convention & Exposition, “Safer, Smarter & Greener”
Technical Proceedings ISSN 1656-7757, September 28-30, 2011 Makati City, Philippines

according to the density p( x0 ) ≡ p( x0 | z0 ) . Two assumptions are imposed: (i) the state of the
system X k = { x0 ,…, xk } follow a 1st-order Markov process and
p( xk X k −1 ) = p( xk xk −1 ) (3)
(ii) the observations Z k = { z0 ,…, zk } , are conditionally independent given the states.
k
p( Z k X k ) = ∏ p( zk xk ) (4)
i=0
To obtain an optimum estimate of the state x k at time step k given the observations
Z k = {z1, z2 ,…, zk } , two conditional probabilities are required. The first is the time updating
process wherein the state p( xk | Z k −1 ) at time step k using data up to time step k–1 is determined.
The equation below shows how this can be calculated.
p( xk | Z k −1 ) = ∫ ∫ p( xk | xk −1, wk ) p(wk ) p( xk −1 | Z k −1 )dwk dxk −1 (5)
The second is the observation updating process where the posterior pdf of the state p( xk Zk )
using measurements Z k is determined. In this case, Bayes’ rule is used as shown below.
p( Z k x k ) p( x k )
p( x k Z k ) = (6)
p( Z k )
By mathematical manipulation the above equation can be expressed as
p( z k | x k ) p( x k | Z k −1 )
p( x k | Z k ) = ⋅ (7)
∫ k k k k −1 k
p ( z | x ) p ( x | Z ) d x
This is the optimal Bayesian solution to the filtering problem. When certain conditions hold, e.g.
Linear, Gaussian, the Kalman filter is optimal solution. However when non-linearities and non-
Gaussian distribution arise, the solution is intractable and approximation based methods like the
extended Kalman filter (EKF), unscented Kalman filter (UKF) and the sequential Monte Carlo
simulation (SMCS) can be used to approximate the posterior distribution.

2.2 Kalman Filter (KF)

The celebrated Kalman Filter is a special case of Bayesian solution to the filtering problem for
certain class of models (Chen 2003). If the state and observation equation are all linear and the
system and observation noises are Gaussian, the KF is the optimal solution to the filtering
problem. Consider Eqs. (1) and (2) which are reduced to the following equations
x k = Fx k −1 + w k (8)

z k = Hx k + v k (9)
with covariance Qk and Rk for wk and vk respectively. The noises wk and vk also have zero mean
and are statistically independent. The Kalman filter algorithm is as follows (Arumpalan et. al
2001):
(
p( x k −1 Z k −1 ) = N x k −1 ; x k −1 k −1 , Pk −1 k −1 ) (10)

p( x k Z k −1 ) = N (x ; x
k k k −1
, Pk k −1 ) (11)

138
50th ASEP Anniversary International Convention & Exposition, “Safer, Smarter & Greener”
Technical Proceedings ISSN 1656-7757, September 28-30, 2011 Makati City, Philippines

(
p( x k Z k ) = N x k ; x k k , Pk k ) (12)

where

x k k −1 = Fk x k −1 k −1 (13)

Pk k −1 = Qk −1 + Fk Pk −1 k −1 FkT (14)

(
x k k = x k k −1 + K k z k − H k x k k −1 ) (15)

Pk k = Pk k −1 − K k H k Pk k −1 (16)
The notation N( ) means a Gaussian density.
S k = H k Pk k −1 H kT + Rk (17)

K k = Pk k −1 H kT S k−1 (18)
Eq. 17 is the measurement prediction covariance of the innovation term given in Eq. 15, while
Eq. 18 is the Kalman gain equation. The KF is the optimal solution to the filtering problem if the
restrictions stated earlier hold.

2.3 Extended Kalman Filter (EKF)

In many cases the assumptions of linearity, Gaussianity and stationarity do not hold true. For this
large class of models a closed form solution might be difficult if not impossible to obtain. The
most common approach is to simplify the model or make some approximations. One such
approach is the extend Kalman filter wherein the state (or observation model) is linearized
around the current state (predicted observation). However the linearization does not necessarily
preserve the distributions of the state or observation models so that some well-known
disadvantages are encountered. Disregarding these disadvantages, the KF can be applied directly
to the linearized model. To implement EKF, F and H which are non-linear functions are
linearized locally as shown below.
dF ( x )
fˆ = x = mk −1 / k −1 , (19)
dx

dH ( x )
hˆ = x = mk / k −1
dx
Then an approximate solution of the filtering problem can be obtained by following Eqs. 10 to
18. However the EKF remains a sub-optimal filter because it always approximates the posterior
pdf of the state to be Gaussian. However if the true density is non-Gaussian, then it cannot be
truly described by a Gaussian. In such cases other algorithms like the unscented Kalman filter
(UKF) or the sequential Monte Carlo simulation will have a better performance compared to the
EKF.

2.4 Unscented Kalman Filter (UKF)

139
50th ASEP Anniversary International Convention & Exposition, “Safer, Smarter & Greener”
Technical Proceedings ISSN 1656-7757, September 28-30, 2011 Makati City, Philippines

The basic difference between the EKF and the UKF is the manner in which the state distribution
is approximated and propagated through the nonlinear system. Both methods approximate the
state distribution with a Gaussian but the UKF achieves better performance by carefully
choosing sample points which can completely capture the true mean and covariance of the
Gaussian distribution (Julier and Uhlmann 1997). When these points are propagated through the
nonlinear function, it can accurately determine the posterior mean and covariance of any
nonlinear system. The following is the algorithm of the UKF omitting some theoretical
consideration. The filter is initialized using the a priori mean and covariance of the state. The
sigma points are then calculated using.
χ k( 0−)1 = xk −1
χ k(i−)1 = xk −1 + ( (L + λ )Pk −1 )
i
i = 1… L (20)

−( (L + λ )Pk −1 )
i
χ k(i−)1 = xk −1 i = L + 1… 2 L

i
where ( L + λ ) Pk −1 is the ith column of the matrix square root, L is the dimension of the state
vector and is defined by

λ = α 2 (L + κ ) − L (21)
where is a scaling parameter which determines the spread of the sigma points and is a
secondary scaling parameter (usually set to 0). The sigma points are then propagated through the
nonlinear system using
χ k(i ) = F ( χ k(i−)1 ) i = 0...2 L (22)

The posterior mean and covariance are then determined from the statistics of the propagated
sigma points as follows
2L
x = ∑Wi ( m ) χ k(i ) , (23)
i =0

2L
Pk = ∑ Wi c [χ k(i ) − x k ][χ k(i ) − x k ]T + Q k (24)
i =0

where the weights defined by Wi (m ) are weights defined by

λ
W0( m ) =
(L + λ)
1
Wi ( m ) = Wi ( c ) = , i = 1 2 L (25)
2( L + λ )

Woc =
L+λ
λ
(
+ 1−α 2 + β )

140
50th ASEP Anniversary International Convention & Exposition, “Safer, Smarter & Greener”
Technical Proceedings ISSN 1656-7757, September 28-30, 2011 Makati City, Philippines

and is a parameter that incorporates any prior knowledge of the distribution of x. The
estimated measurement matrix is calculated by transforming the sigma points through the
nonlinear measurement model,
Y ki k −1 = H ( χ k(ik) −1 ), i = 0...2 L (26)
The mean and covariance of the observation are determined the statistics of the transformed
sigma points,
2L
z k = ∑ Wi mYk(/i )k −1 , (27)
i =0

2L
Pz k z k = ∑ Wi c [Yk(i ) − zk ][Yk(i ) − zk ]T + Rk (28)
i =0

where Rk is the measurement noise covariance matrix. The cross-correlation covariance and the
Kalman gain matrix is calculated using the equations shown below.
2L
Pxk z k = ∑ Wi c [χ k(i ) − xk ][Yk(i ) − zk ]T (29)
i =0

K k = Pxk zk Pz−k 1zk (30)


The observation update equations used to determine the mean and covariance of the filtered state
are shown below.
x k / k = x k / k −1 + K k ( z − z k ) (31)

Pk / k = Pk / k −1 − K k Pzk zk K kT (32)
2.5 Sequential Monte Carlo simulation (SMCS)

The algorithm is initiated by drawing a sample size n from the prior distribution
x k(i−)1 / k −1 ~ p( x o ) i = 1, …, n
(33)
wk(i ) ~ p( wk )
which are assumed known. It is assumed that x k(i−)1 / k −1 are i.i.d. samples from p(xk-1|Zk-1). As such,
xk is determined deterministically by
p( x k | x k −1 , wk ) = δ ( x k − F ( x k −1 , wk )) (34)
where ( ) is the delta function. Using the above equation, Eq. 5 can be reformulated into
1 n 1 n
p( x k | Z k −1 ) = ∫ ∫ δ ( x k − F ( x k(i−)1 , w k ) ∑ δ ( w k − w k(i ) ) ∑ δ ( x k −1 − x k(i−)1 / k −1 ). (35)
n i =1 n i =1
Therefore if we define
x k(i )k −1 = F ( x k(i−)1 k −1 , w k(i ) ) (36)
then the samples generated from this equation are independent realizations of the one step ahead
prediction density which is valid even for non-linear functions (Kitagawa 1996). The posterior
pdf of the state is determine as shown in the equation below

141
50th ASEP Anniversary International Convention & Exposition, “Safer, Smarter & Greener”
Technical Proceedings ISSN 1656-7757, September 28-30, 2011 Makati City, Philippines

n
p( xk | Z k ) = ∑ ak( i )δ ( xk − xk(i/)k −1 ) ⋅ (37)
i =1
where
n
a k(i ) = q k( i ) / ∑ q k( i ) (38)
i =1

is the normalized weight of each particle generated and q k( i ) = p (z k | x (ki /) k −1 ) is the likelihood.
This was the original estimation idea but leads to degeneracy (Gordon et. al., 1993). To handle
this problem the sampling importance re-sampling was introduced. There are several re-sampling
algorithms but this paper follows the method of Kitagawa, 1996 as follows:
1. Sort the particles x k(i )k −1 ,…, x k( nk)−1 according to the magnitude of its normalized weight a k(i ) .
2. For j = 1 to n, repeat
a. generate a uniform random number U.
b. if a k(i −1) < U ≤ a k(i ) is true, pick the particle/s corresponding to a k(i ) .
The resulting particles form an approximate sample from p( x k | Z k ) .

3. ILLUSTRATIVE EXAMPLE

3.1 Limit state exceeding probability

Consider a time-dependent reliability problem where S is the load and R is the resistance which
deteriorates over time t. The limit state function in this case is
g (t ) = R(t ) − S (39)
and the limit state exceeding probability is a follows:
p f = P( g (t ) < 0) (40)
To illustrate the application of KF, EKF, UKF and SMCS to this reliability problem, we consider
a limit state function of the form
g = ( x1 − x2 t ) − x3 . (41)
The components x1, x2 and x3 are random variables and the first component has some process
noise given as
x1 = x1 + w (42)
Assume further that inspection data z, can be obtained from the structure and its relationship to
the state variables is as follows:
z = x1 − 2tx2 + v (43)
for a linear case, and
z = x12 − 28 x1 + 208 − 2tx2 + v (44)
for a non-linear case. If the statistics of the random variables, state and observation noise are
known then the limit state exceeding probability can be determined KF, EKF, UKF or the SMCS.

3.2 Linear case

To start the analysis the statistics of the random variables are assumed as shown in Table 1.
These values apply to both linear and non-linear case. Three cases of observation data are also
assumed available at t = 20. The observation data can be interpreted as not so severe (z = 13),
severe (z = 11), very severe (z = 8). Subsequently the Pf is determined at time t = 0 (just after
construction) and when t = 20 using KF, UKF and SMCS as shown in Table 2. At t = 20
observation data is available so that x1 and x2 were updated and so are the Pfs. The results can be
found in Table 2, along the row t = 20 years. After updating

142
50th ASEP Anniversary International Convention & Exposition, “Safer, Smarter & Greener”
Technical Proceedings ISSN 1656-7757, September 28-30, 2011 Makati City, Philippines

Table 1. Statistics of the random variables


Statistic x1 x2 x3 v w1
Mean 15 0.1 10 0 0
Standard deviation 1.0 0.1 2.0 1.0 1.0

Table 2. KF, UKF and SMCS results for a linear observation model
Time (yrs) Observation Limit state exceeding probability, Pf
value, z KF UKF SMCS
0 0.0127 0.0127 0.0127
0.1587 0.1587 0.1587
20 8 0.2398 0.2398 0.2329
11 0.0786 0.0786 0.0850
13 0.0297 0.0297 0.0339
0.2965 0.2965 0.2965
30 0.5309 0.5309 0.5103
0.1758 0.1758 0.1890
0.0544 0.0544 0.0659

the Pfs when t = 30 (time updating process) can also be calculated. The results are again found in
Table 2 along the row t = 30 years. With regard to SMCS, 100 000 sample realizations for all
random variables were generated and some results of the simulation are shown in Figs. 1 and 2.
The top plots shown in Fig. 1 basically show independence between the random variables x1, x2
and x3, while the bottom part is a plot of the load against resistance. Fig. 2 compares the S against
R plots at t = 20 (without update) and updated using the observation values. The updated plots
also show particle degeneracy (or loss of diversity) as particles with larger weights have higher
probability of being selected. This is a limitation of the SMCS algorithm but will not be
discussed in this paper.

Figure 1. Sample realizations at t = 0

143
50th ASEP Anniversary International Convention & Exposition, “Safer, Smarter & Greener”
Technical Proceedings ISSN 1656-7757, September 28-30, 2011 Makati City, Philippines

Figure 2. Results with t = 20 and z = 13, 11 & 8

The succeeding figures show that KF/UKF and SMCS have identical results for the Pf when the
system and observation models are linear.

Figure 3. Pf results using SMCS

Figure 4. Pf results using KF / UKF

144
50th ASEP Anniversary International Convention & Exposition, “Safer, Smarter & Greener”
Technical Proceedings ISSN 1656-7757, September 28-30, 2011 Makati City, Philippines

3.3 Non-linear observation model

The system model is the same for both scenario therefore the values of Pf when t = 0 and when
t = 20 are the same with the linear model. However the Pfs can be updated using the non-linear
observation model in Eq. 44 coupled with the following data: not so severe (z = 13), severe (z =
11), very severe (z = 8). The figure below shows the non-linear relationship when x1 and x2 are
updated when z = 13.

Figure 5. Sample realizations at t = 20 and with z = 13

The top plot in Fig. 6 shows the R and S plot when t = 20 (without update) and when R and S are
the updated when observation data are available.

Figure 6. Results with t = 20 and z = 6, 11 and 13

145
50th ASEP Anniversary International Convention & Exposition, “Safer, Smarter & Greener”
Technical Proceedings ISSN 1656-7757, September 28-30, 2011 Makati City, Philippines

Table 3. EKF, UKF and SMCS results for a non-linear observation model
Time (yrs) Observation value, z Limit state exceeding probability, Pf
EKF UKF SMCS
0 0.0127 0.0127 0.0127
0.1587 0.1587 0.1587
20 6 0.2226 0.1626 0.2700
11 0.0274 0.0642 0.0574
13 0.0086 0.0302 0.0250
0.2965 0.2965 0.2965
30 0.5000 0.4070 0.5674
0.0614 0.2469 0.1241
0.0154 0.1627 0.0513

Table 3, Figs. 7 and 8 show the approximate Pfs from EKF, UKF and SMCS using the non-linear
observation model and the observation values.

Fig. 7. Pf results using SMCS

Figure 8. Pf results using EKF / UKF

4 CONCLUSION

Structures deteriorate over time and as such it becomes necessary to assess its present state. In
this paper we assess the state of a deteriorating structure by estimating the limit state exceeding

146
50th ASEP Anniversary International Convention & Exposition, “Safer, Smarter & Greener”
Technical Proceedings ISSN 1656-7757, September 28-30, 2011 Makati City, Philippines

probability at t = 0, 20 and 30 years. However if observation data is available at a specific time


then the pf can be updated to incorporate the new information. This is possible using the KF (for
linear models), EKF, UKF and SMCS (for non-linear models). The authors see future work in
using SMCS estimating the reliability of deteriorating structures and especially if the state or
observation models exhibit strong non-linearities and the probability density functions are non-
Gaussian.
REFERENCES

Anderson, B.D., and Moore, J.: The Kalman-Bucy filter as a true time-varying Wiener filter,
IEEE Trans. Syst., Man, Cybern., Vol. 1, pp. 119 – 128, 1971.

Chen, Z.: “Bayesian filtering: From Kalman filters to particle filters, and beyond,” Tech. Rep.
Adaptive Systems Lab, McMaster University, 2003. 68 pages.

Garciano,L. E. and Yoshida, I.: “Bayesian Updating and Recursive Monte Carlo”, 57th
National Congress of Theoretical and Applied Mechanics (NCTAM) Proc., 2C12, pp. 129 -
130, 2008

Gordon,N. J., Salmond, D. J. and Smith, A. F. M.: “Novel approach to nonlinear / non-
Gaussian Bayesian state estimation”, IEEE Proc. Radar Signal Processing, Vol. 140, No. 2,
pp. 107 - 113, 1993

Julier, S. and Uhlmann J.: A new extension of the Kalman filter to nonlinear systems, In Proc. of
AeroSense: The 11th Int. Symposium on Aerospace/Defense Sensing, Simulation and
Controls, USA, 1997.

Kalman, R. E.: A new approach to linear filtering and prediction problems, Trans. ASME, Ser.
D, J. Basic Eng., Vol. 83, pp. 95 - 107, 1961.

Kitagawa, G. Monte Carlo Filter and Smoother for non-Gaussian nonlinear state space models,
Journal of Computational and Graphical Statistics, Vol. 5, No. 1, pp. 1 - 5, 1996.

Yoshida, I., Akiyama, M., and Suzuki, S.: “Health monitoring algorithm by the Monte Carlo
Filter based on non-Gaussian noise”, Journal of Natural Disaster Science, Vol. 24 No. 2, pp.
101 - 107, 2002.

Yoshida, I., Akiyama, M., and Suzuki, S.: “Reliability Analysis of an Existing RC Structure
Updated by Inspection Data”, 10th International Conference on Structural Safety and
Reliability, pp. 2482 - 2489, 2009.
ABOUT THE AUTHORS
Lessandro Estelito Garciano received his Doctor of Engineering degree (Civil Engineering) from Tokyo
City University (formerly Musashi Institute of Technology). He can be contacted by e-mail at
logarciano@hotmail.com.

Ikumasa Yoshida is a Professor of the Information System and Management in Civil Engineering
Laboratory. He can be contacted at the Department of Urban and Civil Engineering of Tokyo City
University 1-28-1 Tamazutsumi, Setagaya-ku, Tokyo, 158-8557, Japan or by e-mail: iyoshida@tcu.ac.jp.

147

Вам также может понравиться