Professional Documents
Culture Documents
Huynh Khac Tuan
Huynh Khac Tuan
Huynh Khac Tuan
Abstract
This work deals with a dynamic maintenance policy based on prognostic for a single-
unit deteriorating system. In order to take a step forward to practice, the system is
assumed operating under indirect condition monitoring. The degradation of system is
described by a crack growth evolution modelled by the Paris-Ergodan law, and a
non-destructive ultrasonic technique is used to monitor the system condition. A
procedure of state estimation and prediction of values associated with system health
from the noisy measured data are performed thanks to the well-known particle filter
algorithm. Then based on these estimated/predicted value, a dynamic maintenance
policy is developed and is compared with two more classical policies (i.e. block
replacement or periodic inspection/replacement [1]). The comparison results on the
optimal expected maintenance cost rate will show the performance of the proposed
dynamic maintenance.
1. Introduction
With the development of engineering structures, the role of the maintenance operation
becomes more and more important in keeping and improving the system availability
and safety but also the product quality. A widely used inspection scheme in CBM
framework is to inspect periodically due to its simplicity to implement. However this
scheme is sometimes inefficient to guarantee the system availability since at the same
time it might “over-inspect” or “under-inspect” failure times [2]. A non-periodic
inspection scheme [3, 4] is therefore preferred in order to avoid this weakness since it
can adapt the inter-inspection interval to the system state. But it seems difficult to
deal with for some practical applications. In keeping the strength of periodic scheme
but still saving in maintenance cost by guaranteeing the system availability and
avoiding inopportune maintenance spending, a dynamic maintenance model based on
prognostic is developed in this work. The dynamic maintenance model still bases on
periodic inspection scheme but it can intervene between the inspections thanks to a
predictive replacement time. The predictive replacement time adapted to system state
1
Proceedings of the 38th ESReDA Seminar, Pecs, May 04-05, 2010
allows bring out a timely maintenance decision; hence guarantee the system
availability, as well as saving in maintenance cost.
The Paris-Erdogan model adopted for describing the crack depth growth rate is
expressed by the following equation
dx / dt = C (∆K )n (1)
In order to keep the intrinsic randomness of the model, a normal random disturbance
ω ~ N (0, σ ω2 ) can be added and eq. (1) becomes
dx / dt = eω C ( β x ) n (2)
For time step ∆t sufficiently small, the state space model eq. (2) can be discretized as
xi = xi −1 + eωi C ( β xi −1 ) n ∆t (3)
where ωi, i =1,2, … are independent and identically distributed Gaussian random
variables. This equation represents a non-linear Markov process with independent,
but non-stationary increment whose degradation rate ∆xi = ( xi − xi −1 ) / ∆t follows a
lognormal distribution ∆xi ~ log N (ln(C ( β xi −1 ) n ), σ ω2i ) . As such, the expected value
2
Advanced Maintenance Modelling
and variance of crack depth increment per unit time are respectively
σ ω2i
σ2
E ( ∆xi ) = C ( β xi −1 ) e 2 and Var ( ∆xi ) = E 2 ( ∆xi )·(e ωi − 1) . From eq. (3), the
n
evolution of the crack depth can be explicitly represented by the following condition
probability density function (pdf)
−1/ σ ω2i
p( xi ∣xi −1 ) = e (σ ωi 2π C ( β xi −1 ) n ∆t ) (4)
which is usually taken as importance function in particle filter framework (see alg. 1).
The crack depth xi at time step ti is not directly observable, but can be assessed via
noisy measured condition information data. Thus a non-destructive ultrasonic
inspection/measurement technique is assumed, whose associated observation zi is
described by the following logit model [5, 10]
(6)
This will be used to update the importance weights in particle filtering (see alg. 1).
Assume that the initial pdf of crack depth p ( x0 ) is known. The crack growth model
described in sec. 2 (i.e. eq. (4) and (6)) expresses a hidden Markov model. We denote
by x 0:i {x0 , …, xi } and z 0:i {z0 , …, zi } , respectively, the system states and the
measurements up to time ti . The sequence of hidden states x 0:i can be approximated
by its expected values
3
Proceedings of the 38th ESReDA Seminar, Pecs, May 04-05, 2010
∑
Ns
Estimated posterior pdf: p (x 0:i ∣Z 0:i ) i =1
δ xˆ ( ) (x0:i ) N s
k
0:i
∑
Ns
Estimated states: xˆ 0:i i =1
xˆ (0:ki) N s
Figure 1 illustrates a case of crack depth estimation by applying the alg. 1 for the data
set C = 0.015, n = 0.35, b = 3.9, σ ω2 = 2.53, ∆t = 1, β 0 = 0.06, β1 = 1.25, σ v2 = 0.25, d = 6 ,
N s = 5000 and inter-inspection interval T = 10 . The figure shows that the system
states (crack depth) are recovered quite accurately from noisy measurements at each
inspection times.
4
Advanced Maintenance Modelling
Under the considered model, the system is assumed to be in failure state when the
crack depth xi exceeds a critical material thickness d . The sequence of observations
up to the current inspection time Ti , Z 0:i = {zT1 , zT2 , …, zTi } is assumed available. Our
goal is to predict at actual inspection Ti the conditional probability of the first failure
in a future unit time interval. Let p (i + j ) = P ( xi + j ≥ d ∣x0:i + j −1 < d , Z 0:i ) denotes the
conditional posterior probability of the crack depth first exceeding the critical
material thickness d in the interval (ti + j − 1, ti + j ) , given Z 0:i and knowing the
component had not failed up to time ti + j − 1 > Ti . In the same way as in [6], we obtain
p ( i + j ) ∑1{ xˆ( m) ≥ d } − ∑1{ xˆ( n) ≥ d } / ∑1{ xˆ (l ) < d } (8)
m i+ j n
i + j −1
l i+ j −1
(resp. ti + j ), which are extended from xˆ0:( i ) using the system crack depth evolution eq.
k
• For q = 1: j ,
k = 1: N s , sample xˆi(+ )j ~ p ( xi + q ∣xi(+ q) −1 ) and set x% (0:i)+ q {x% (0:i)+ q −1 , xˆi(+ q) }
k k k k k
For
The numerator of eq. (8) represents the number of sampled particles which fail for the
first time in the interval (ti + j − 1, ti + j ) and the denominator is those of the particles
which still survive up to time ti + j − 1 .
5
Proceedings of the 38th ESReDA Seminar, Pecs, May 04-05, 2010
4. Maintenance models
4.1 Assumptions and performance assessment
Two maintenance operations of the system are available: the preventive replacement
with cost Cp > Ci and the corrective replacement with cost Cc > Cp. We assume that a
replacement, whether preventive or corrective, restores the system to an as-good-as-
new condition and it takes negligible time.
C ∞ = lim C (t ) / t or C ∞ = lim E (C (t )) / t
t →∞ t →∞
where C(t) is the cumulated maintenance cost at time t. Due to the complexity of
considered model in this case study an analytical solution of expected cost rate is
almost impossible. We must therefore resort to stochastic Monte Carlo simulation:
C ∞ = lim C (t ) t = lim ∑ n =1 C ( S ( ) ) ∑ S(
N n N n)
n =1
(9)
t →∞ N →∞
where N is number of simulation histories and S(n) is n-th renewal cycle. All
numerical examples in the remainder of this section will be performed from the data
set Cp=50, Cc=100, Cd=250, Ci=5, C=0.0047, n=0.35, β=3.9, N=4000, ∆t=1, d=6,
β0=0.06, σ ω2 = 4.8326 , β1=1.25, σ v2 = 0.0625 , Ns=3000.
6
Advanced Maintenance Modelling
can be predicted by extending the trajectories xˆ (0:i) and using the following on-line
k
maintenance decision: “At the future time Ti + l , l ≥ 1 , if the system is still running, it is
preventively replaced with cost Cp; but if it is already failed, a corrective replacement
is performed with cost Cc. In this case, because of the system inactivity after failure,
an additional cost is incurred from the failure time until the replacement time at a
cost rate Cd”. And the expression of cost criterion to predict the replacement time
inferred from the on-line maintenance decision above is computed as
where:
p (i + 1) l =1
P (Ti + l ) = (11)
p (i + 1) + ∑ k = 2
l
(∏ k −1
j =1
(1 − p(i + j ) ) p(i + k ) ) l≥2
Consequently the suitable predictive replacement time Tpred is suggested by the time,
among all future time steps, which minimizes the cost criterion eq. (10)
( )
= Ci N i( ) + C p , if preventive replacement at the end of the cycle
n n
• CProg
( )
= Ci N i( ) + Cc + Cd Td( ) , if corrective replacement at the end of the cycle
n n n
• CProg
where N i(
n)
is the inspections number effectuated before replacement time, Td(
n)
is the
( ) n
downtime interval in the n − th renewal cycle S Prog . Using eq. (9), the expected cost
rate is given by
7
Proceedings of the 38th ESReDA Seminar, Pecs, May 04-05, 2010
= lim ∑ n =1 CProg
( )
∑ S P( rog
)
∞ N n N n
CProg n =1
(13)
N →∞
∞
The optimal inter-inspection time interval Topt satisfies CProg (Topt ) = inf CProg
∞
T
(
:T > 0 )
Figure 2. Long run expected cost rate of dynamic maintenance policy versus inter-inspection interval.
Maintenance policy: Under the considered block replacement policy, the system is
replaced at regular time interval T, either preventively if it still running at the end of
the replacement interval or correctively if a failure occurred since the last
replacement. The replacement period T is the only decision variable for this policy.
8
Advanced Maintenance Modelling
Figure 3. Long run expected cost rate of block replacement policy versus inter-inspection interval.
Maintenance policy: For this maintenance policy, the system is periodically inspected
with period T. At each inspection if estimated crack depth constructed from measured
data exceeds a threshold M, a preventive replacement is performed, but if the system
has already failed, it is correctively replaced. The inter-inspection interval T and the
preventive degradation threshold M are the two decision variables in this policy.
Numerical example: Figure 4 illustrates the optimal value CPI∞ opt = 2.8714 at Topt = 5
and M opt = 2 of the periodic inspection/replacement policy.
(a) Small variance in degradation increment. (b) High variance in degradation increment.
Figure 4. Long run expected cost rate of inspection/replacement policy.
9
Proceedings of the 38th ESReDA Seminar, Pecs, May 04-05, 2010
with the benchmark policies as a function of inspection cost Ci. The comparison is
performed under different variances of the crack growth process to keep the
generality of problem.
The numerical examples are based on the following data set: Cp=50, Cc=100,
Cd=250, ∆ t = 1, β 0 = 0.06, β1 = 1.25, σ v2 = 0.0625, d = 6, N = 4000, N s = 3000 . Figure 5
illustrates the comparison results for a case of small variance in deterioration of crack
depth ( C = 0.015, n = 0.35, b = 3.9, σ ω2 = 2.53 - Figure 5a) and a case of 10 times
greater in variance ( C = 0.0047 , n = 0.35, b = 3.9, σ ω2 = 4.8326 - Figure 5b).
(a) Small variance in degradation increment. (b) High variance in degradation increment.
Figure 5. Comparison of optimal long run expected cost rate.
In the figures, the circle-lines represents the optimal expected cost curves of the
dynamic maintenance policy; the diamond-lines and triangle-lines represent for two
more classical maintenance policies: block replacement and periodic inspection
/replacement respectively.
10
Advanced Maintenance Modelling
higher than one of block replacement policy. This can be explained by the fact that
for this inspection polices, it incurs not only cost of replacement (as in block
replacement policy) but also an inspection cost to know the system state when it is
still running. Regarding dynamic maintenance policy, owing to the preventive
replacement can be performed without inspection, this maintenance policy just
emulates a block replacement policy when inspection cost is high.
Acknowledgements
This work is part of the PhD research work of Khac Tuan Huynh supported in part by
11
Proceedings of the 38th ESReDA Seminar, Pecs, May 04-05, 2010
References
[1] Huynh, K.T., Barros, A., Bérenguer, C. and Castro, I.T. (2010) Value of
condition monitoring information for maintenance decision-making, In: Proc.
Ann. Reliability & Maintainability Symp, pp. 543-548.
[2] Yang, Y. and Klutke, G.A. (2000) Lifetime characteristics and inspection-
schemes for Levy degradation processes, IEEE Transactions on Reliability, vol.
49, pp. 377 –382.
[3] Grall, A., Dieulle, L., Bérenguer, C. and Roussignol, M. (2002) Continuous-
time predictive-maintenance scheduling for a deteriorating system, IEEE
Transactions on Reliability, vol. 51, pp. 144-150.
[4] Cui, L., Xie, M. and Loh, H.T. (2004) Inspection schemes for general systems,
IIE Transactions, vol. 36, pp. 817-825.
[5] Myötyri, E., Pulkkinen, U. and Simola, K. (2006) Application of stochastic
filtering for lifetime prediction, Engineering and System Safety, vol. 91, pp.
200-208.
[6] Cadini, F., Zio, E. and Avram, D. (2009) Model-based Monte Carlo state
estimation for condition-based component replacement, Engineering and
System Safety, vol. 94, pp. 752-758.
[7] Kozin, F. and Bogdanoff, J. (1989) Probabilistic models of fatigue crack
growth: results and speculations, Nuclear Engineering & Design, vol. 115, pp.
143-171.
[8] Bigerelle, M. and Iost, A. (1999) Bootstrap analysis of FCGR, application to
the Paris relationship and to life prediction, International Journal of Fatigue,
vol. 21, pp. 299-307.
[9] Whittaker, I. and Saunders, S. (1973) Application of reliability analysis to
aircraft structures subject to fatigue crack growth and periodic structural
inspection. Air Force Material Laboratories Technical Report, AFML-TR-73-
92.
[10] Simola, K. and Pulkkinen, U. (1998) Models for non-destructive inspection
data, Reliability Engineering and System Safety, vol. 60, pp. 1-12.
[11] Doucet, A., Godsill, S. and Andrieu, C. (2000) On sequential Monte Carlo
sampling methods for Bayesian filtering, Statistics and Computing, vol. 10, pp.
197-208.
[12] Arulampalam, M., Maskell, S., Gordon, N. and Clapp, T. (2002) A tutorial on
particle filters for online nonlinear/non-Gaussian Bayesian tracking, IEEE
Transaction on signal processing, vol. 520, pp. 174-188.
[13] Bérenguer, C. (2008) On the mathematical condition-based maintained
modeling for continuously deteriorating systems. International Journal of
Materials and Structural Reliability, vol. 6, pp. 133–151.
[14] Christer, A., Wang, W. and Sharp, J. (1997) A state space condition monitoring
model for furnace erosion prediction and replacement. European Journal of
Operational Research, vol. 101, pp. 1-14.
12