Professional Documents
Culture Documents
Reliability Estimation in Vibrating Structures by Girsanov'S Transformation With Markov Control Forces
Reliability Estimation in Vibrating Structures by Girsanov'S Transformation With Markov Control Forces
Reliability Estimation in Vibrating Structures by Girsanov'S Transformation With Markov Control Forces
This paper considers the problem of Monte Carlo simulation based time variant reliability analysis of
nonlinear dynamical systems driven by non-stationary Gaussian excitation. Attention is focused on an
importance sampling method based on Girsanov’s probability transformation technique. In this approach,
suitably designed artificial control forces are added to the dynamical system to drive the system response
to the failure region. A strategy to formulate closed loop control forces, which are dependent on the system
states, has been developed. These closed loop controls lead to a greater reduction in the sampling variance
of the failure probability estimator compared to the open loop controls which are being used in the existing
literature. Merits of proposed controls have been established through numerical studies on a randomly
driven duffing oscillator.
2 Problem statement ∧
The above estimator is unbiased, i.e, E P P F = PF ,
The governing equation of motion of a dynamical
and the variance of the estimator is given by
system driven by non-stationary Gaussian excitation
( ) R (T ) P
2
{ }
∧ 1
2
The sampling variance associated with the direct Monte controls developed in this study are applicable to linear/
Carlo estimator for PF depends only on the sample size nonlinear oscillators driven by non-stationary Gaussian
N, and is inversely proportional to N. Hence, a direct excitation. The advantage of the proposed control forces
Monte Carlo simulation study becomes computationally over the existing open loop controls, in terms of the
prohibitive, especially when PF is small. In this work sampling variance of the importance sampling estimator,
we adopt an importance sampling strategy based on have been illustrated through numerical simulations on
Girsanov’s transformation[9] to tackle this problem. This a randomly driven duffing oscillator.
is achieved by considering the modified SDE, The remaining paper is organized as follows. In
section 3 we describe the strategy for determining the
( )
dX ( t ) = A ( X ( t ) , t ) + Ã ( t ) u ( X ( t ) , t ) dt + Ã ( t ) dB ( t ) ;
state-dependent control forces. Here, we first layout
dR ( t ) = −u ( X ( t ) , t ) R ( t ) S dB ( t ) ; the formulation for obtaining the proposed closed loop
-1
(2)
X ( 0 ) = X 0 , R ( 0 ) = R0 , 0 ≤ t ≤ T controls and subsequently give the expressions for the
Radon-Nikodym derivative and the failure probability
obtained by introducing an artificial control force, estimator. Following this, the details of the numerical
u ( X ( t ) , t ) to the original dynamical system in Eq. (1).
study is provided in section 4. Finally, we conclude in
In Eq. (2), B ( t ) is an Ito’s process obtained according section 5.
to the transformation
dB ( t ) = −u ( X ( t ) , t ) dt + dB ( t ) ; B ( 0 ) = 0. (3) 3 Selection of control forces
By Girsanov’s theorem, the above transformation To determine the Girsanov’s control forces and the
changes the original probability measure P to a new associated Radon-Nikodym derivative, it is useful to
measure Q, such that B ( t ) is a standard Brownian view the problem of time variant reliability as a problem
motion with respect to Q. In Eq. (2), R(t) is a scalar in series system reliability estimation. Accordingly, the
process, called the Radon-Nikodym derivative, which time interval is expressed as a union of sub-intervals
accounts for the correction needed for modification of as 0 = T0 < T1 < T2 < Tm = T . If m is large, i.e., the
the original dynamical system. The estimator for failure length of sub-intervals is small, the failure event can be
probability then takes the form approximated as
(T )
{ } { }
∧ 1 R
k
F = h* − max h ( X ( t ) ) ≤ 0
N
P = F ∑ I h − max h X ( t ) ≤ 0
∗ k
( ) (4)
N k =1 R 0
k
0 ≤ t ≤T 0 ≤ t ≤T
m m
where, the sample realizations X ( t ) and Rk (t) are {h − h ( X (Ti ) ) ≤ 0 = } F.
k
*
≈ i (6)
obtained from Eq. (2). i =1 i =1
m
nonlinear system in Eq. (1) by a hypothetical time
the form Q (C ) = ∑ w Q (C )
i i
for any event C ∈ , varying linear system given by
i =1
∫( )
T
σ T2i ( t ) = S È i (Ti , r ) Ã ( r ) dr (10)
3.1 Formulation of state-dependent controls
t
For the nonlinear system in Eqs. (1) and (2), using
Substituting Eq. (10) into Eq. (7), the sub-optimal
stochastic control theory it can be shown[9] than an ideal
state-dependent Markov control for the nonlinear system
state-dependent control u i* ( X ( t ) , t ) exists which leads can be obtained as
to zero sampling variance of the importance sampling
estimator for the instantaneous failure probability u ( X ( t ) , t ) =
i
{ }
PFi = P h* − h ( X (Ti ) ) ≤ 0 . This ideal control has the
S φ ( z ( X ( t ) , t ) )
σ T ( t ) (1 − Φ ( z ( X ( t ) , t ) ) )
( Ã (t )) T
È ( Ti , t )
i
(11)
functional form given by i
1 ∂PFi ( t , X ( t ) ) ( X ( t ) , t ) *
h − µT
u
i*
( X ( t ) , t ) = S ( Ã(t ))
T
(7) where, z ( X ( t ) , t ) =
. In the above i
PF ( t , X ( t ) ) ∂X ( t ) σ T (t )
equations ψ (T , r ) = ( H¸ (T , r ) ) θ (T , r ) = [θ (T , r )
i T
with,
i i i i i
θ
for t ∈ [ 0, Ti ) and u i * ( X ( t ) , t ) = 0 for t θ≥ T(Ti . , r ) = [θ (T , r ) θ (T , r ) θ (T , r )], where each θ (T , r ) ; l = 1, , p
i i i 1 i
i i i i i
i 1 i 2 i p i l i
Here X ( t ) is the state at time t of the modified is the response at time Ti of the deterministic time
SDE in Eq. (2). In Eq. (7) the function varying system dZ (τ ) = G (τ ) Z (τ ) dτ ; τ ≥ r to initial i
*
{
}
PFi ( t , X ( t ) ) = P h − h ( X (Ti ) ) ≤ 0 X ( t ) = X ( t ) denotes
condition Z(r) set to zero everywhere except at the l-th
component which is set to unity. For the special case
the probability of exceedance of the response h(X(t))
of a linear system, the sub-optimal control u ( X ( t ) , t )
i
In order to enable determination of a sub- 3.2 Radon-Nikodym derivative and the failure
optimal control u ( X ( t ) , t ) analytically for nonlinear
i
probability estimator
systems as well, we consider a linear performance Based on the above formulation, the Radon-Nikodym
measure h ( X ( t ) ) = HX ( t ) and propose to obtain an derivative is given by
approximation of PFi ( t , X ( t ) ) by approximating the
Oindrila Kanjilal, et al
dRi ( t ) = − Ri ( t ) u i ( X ( t ) , t ) S -1dB i ( t )A; = 12.2, α = 0.2 s and α = 0.25 s . The system parameters are taken as w
o 1 2
x ( t ) + 2η ω x g ( t ) + ω x ( t ) =
2
from the two methods, is shown in Fig. 1. Failure
g g g g g
(15) probability estimates from Methods 1 and 2 have been
2η ω xd ( t ) + ω x ( t )
2
d d d
obtained using N = 500 samples. The accuracy of the
d
x ( t ) + 2η ω xd ( t ) + ω x ( t ) = e ( t ) W ( t )
2
d d d d d methods have been established by comparison with
where, W(t) is the Gaussian excitation with the estimates obtained from Method 3 using samples
E [W ( t ) W ( t + τ )] = Sδ (τ ) and e ( t ) = A ( e − e ) and
and has been shown in Fig. 2. In addition, to illustrate
− α 1t − α 2t
o
the relative merits of Methods 1 and 2, the mean and
P
0.0550 4.87 × 10−6 (0.064 ) 4.47 × 10−6 (0.126 ) [1] Langley, R.S., A first passage approximation for
normal stationary random processes, Journal of
0.0600 1.19 × 10−7 (0.065 ) 1.25 × 10−7 (0.161)
Sound and Vibration, Vol. 122 (2), pp. 261-275,
1988.
Table 3: Comparison of Methods 1 and 2 for
[2] Naess, A., Crossing rate statistics of quadratic
α = 25000 N m3
transformations of Gaussian processes,
( )
∧
PF , δ ∧ Probabilistic Engineering Mechanics, Vol. 16,
h*(m) PF
pp. 209-217, 2001.
Method 1 Method 2 [3] Lin, Y.K. and Cai, G.Q., Probabilistic Structural
0.0425 2.03 × 10−4 (0.102 ) 2.21× 10−4 (0.131)
Dynamics: Advanced Theory and Applications,
0.0450 3.92 × 10−5 (0.109 ) 3.94 × 10−5 (0.135 ) McGraw-Hill, Singapore, 1995.
4.29 × 10−6 (0.115 ) 4.45 × 10−6 (0.168 )
[4] Spencer, B.F. and Bergman, L.A., On the
0.0475
numerical solution of the Fokker-Planck equation
0.0500 3.90 × 10−7 (0.143) 4.11× 10−7 (0.193) for nonlinear stochastic systems, Nonlinear
Dynamics, Vol. 4, pp. 357-372, 1993.
Structural Engineering Convention (SEC-2016)
[5] Au, S.K. and Beck, J.L., Estimation of small [10] Macke, M. and Bucher, C., Importance sampling
failure probabilities in high dimensions by for randomly excited dynamical systems, Journal
subset simulation, Probabilistic Engineering of Sound and Vibrations, Vol. 268, pp. 269-290,
Mechanics, Vol. 16, pp. 263-277, 2001. 2003.
[6] Pradlwarter, H.J., Schueller, G.I., Koutsourelakis, [11] Ogawa, J. and Tanaka, H., Importance sampling
P.S. and Charmpis, D.C., Application of line for stochastic systems under stationary noise
sampling simulation method to reliability having a specified power spectrum, Probabilistic
benchmark problems, Structural Safety, Vol. Engineering Mechanics, Vol. 24, pp. 537-544,
29(3), pp. 208-221, 2007. 2009.
[7] Katafygiotis, L.S., Cheung, S.H. and Yuen K.V., [12] Sundar, V.S. and Manohar, C.S., Estimation of
Spherical subset simulation (S3) for solving time variant reliability of randomly parametered
nonlinear dynamical reliability problems, non-linear vibrating systems, Structural Safety,
International Journal of Reliability and Safety, Vol. 47, pp. 59-66, 2014.
Vol. 4, pp. 122-138, 2010. [13] Sundar, V.S., Ammanagi, S. and Manohar,
[8] Engelund, S. and Rackwitz, R., A benchmark C.S., System reliability of randomly vibrating
study on importance sampling techniques in structures: Computational modelling and
structural reliability, Structural Safety, Vol. laboratory testing, Journal of Sound and
12(4), pp. 299-302, 1993. Vibration, Vol. 351, pp. 189-205, 2015.
[9] Oksendal, B., Stochastic Differential Equations:
An Introduction and Applications, Springer,
New York, 2003.