Professional Documents
Culture Documents
Abstract
The aim of this article is the identification of reliability parameters on the basis of small samples. First, we have
evaluated the errors affecting reliability models and parameters (e.g. mean up time) of non repairable systems.
Second, we have studied repairable systems and browsed in the broad line reliability theory characterizing these
systems. The different models detailed in this article are the homogeneous Poisson process, non-homogeneous
Poisson process and the generalized renewal process (GRP), this one is a function of a parameter q called ‘degree
of repair’ that characterizes efficiency of maintenance operations. To conclude the usefulness of these models in
an industrial context, we present a statistical analysis of data generated by Monte Carlo simulations. This study
points out the difficulty in determining an accurate model on the basis of a small sample.
Keywords: Reliability, Repairable systems, HPP, NHPP,GRP
N β η β η 0.9
5 1,07 46930 1,74 42006 ∆R
0.8
10 1,14 44845 1,50 41904 ∆t
50 failures
β=1,11 : η=33200
20 1,15 41433 1,39 39473 0.7
β
1000 1,29 40217 1,31 40075
β=1,3 : η=40000
0.4
Table 1. Mean values of the parameters β and η cal- β=1,11 : η=46000
0.3
culated by graphical method and MLE
0.2
N ∆MUTmax
∆tmax /MUTtheo ∆Rmax β=1,51 : η=33200
MUTtheo 0.1
R = 98% R = 99% R à t [h]
0
5 / 0,24 0,17 62% 18900 0 5 10 15
t x 1e4 [hours]
10 / 0,20 0,14 36% 32800
20 1,25 0,10 0,07 23% 38900 Figure 1. Variation field of an estimated reliability
50 0,47 0,05 0,04 14% 79000
distribution : ∆t and ∆R errors
100 0,32 0,04 0,03 11% 73800
1000 0,07 0,01 0,008 3% 61600
regression method to an accurate estimation of the
Table 2. Regression method - Maximal errors on MUT, parameters (50 failures against 1000),
time and reliability
2. estimations are not very precise with a poor col-
N ∆MUTmax
MUTtheo
∆tmax /MUTtheo ∆Rmax lection of data (5,10,20 failures).
R = 98% R = 99% R à t [h]
What is the incidence of this uncertainty on relia-
5 4,93 0,51 0,42 75% 14000
bility parameters as MUT, estimation time to failure
10 1,98 0,29 0,23 41% 31300
20 0,61 0,14 0,10 28% 38300 and reliability R? First, we have to note that β and η
50 0,26 0,06 0,04 14% 46100 follow normal distributions its means β and η. More-
100 0,20 0,04 0,03 10% 47800 over we know that for a normal distribution 95% of
1000 0,05 0,01 0,008 3% 54300 the values are included in interval m ± 2s (s is the
standard deviation). It has been verified that the re-
Table 3. MLE method - Maximal errors on MUT, time liability law is included in a variation field delimited
and reliability by the four curves corresponding to extreme values of
β and η. This envelope curves allow us to determine
errors on MUT, time to failure and reliability (fig. 1).
Results are presented in tables 2 and 3.
This method is a recursive one so we need an initial From these results, we note that the regression
estimation of β that may be obtained from a regression method and MLE method generate quite the same er-
analysis (see 2.1). rors with an advantage for MLE method concerning
From equation (10) we deduce : determination of MUT.
s
Pn β
β
i=1 ti 3 Repairable system : basic concepts
η = (14)
n
3.1 Definitions
Repairable system
2.5 Comparison graphical method / MLE A repairable system is a collection of items which, after
Reference [2] highlights errors committed in the esti- failing to perform at least one of its required functions,
mation of a distribution when we have a small set of can be restored to performing all of its required func-
data (i.e. less than 50 failure times) using the regres- tions by any method, other than replacement of the
sion method estimation. Now, we will compare MLE entire system.
and regression methods and see which one is the most
accurate. Let’s consider a component (eg. a bear- Stochastic point process
ing) which reliability law is a Weibull with β = 1, 3; A stochastic point process is a mathematical model
η = 40000; γ = 0 and MUTtheo = 39943 hours. for a physical phenomenon characterized by highly lo-
We intend to redetermining these parameters dis- calized events distributed randomly in a continuum.
tribution. For that purpose, we will generate thanks Applied to repairable systems, the continuum is time
Simulatrix 100 sets of 5, 10, 20, 50, 100 and 1000 fail- and the highly localized events are failures which are
ures. We get the results presented in table 1. We note: assumed to occur at instants within the time contin-
uum. To treat such a problem, we pose two assump-
1. the MLE method converges more rapidly than the tions: (1) the system is operated whenever possible,
0.7
0.6
0.5
0.4
ρ(t)
0.3
v(t)
where a > 0, b > 0. 0.3
0.2
Exponential Law
- ROCOF for the exponential law:
0.1
0.9 0.9
Repairable system
System failure Non−repairable system
0.8 0.8
0.7 0.7
0.6 0.6
R(t)
R(t)
0.5 0.5
0.4 0.4
System failure
0.3 0.3
Repairable system
Non−repairable system
0.2 0.2
0.1 0.1
0 0
0 2 4 6 8 10 12 14 16 18 20 0 2 4 6 8 10 12 14 16 18 20
t t
f (t|An ) = (33)
0.7
β−1 β β !
β t + An An t + An
exp − 0.6
η η η η R(t)
0.5
β−1 0.3
f (t|An ) β t + An
λ(t|An ) = = (34)
R(t|An ) η η 0.2
System failure
0.1
a=0,125 ; b=1.5
3 Simulated failures
V(t)=E(N(t)) 3
a=0,11 ; b=1.64
Best fitting line
2.5
2 a=0,73 ; b=1.18
2
1
ln(V(t))
ln(V(t))
1.5
a=0,15 ; b=1.40
0
1
−1
0.5
0 −2
−0.5
−2 −1 0 1 2 3 4 −3
0 0.5 1 1.5 2 2.5 3
ln(t) ln(t)
Figure 8. Estimation of NHPP parameters with 100 × Figure 9. Estimation of NHPP parameters with 1 × 5
20 failures failures
N. of repairs Instances N a b ρ2
1 1 5 0,16 1,40 0,99
2 5 10 0,20 1,33 0,99
3 11 20 0,17 1,47 0,99
4 20
5 17 Table 5. Estimation of a NHPP parameters with 100×
6 26 N failures
7 14
8 4
The main difficulty to apply this methodology in an
9 2
industrial context lies in fact we need many times for
Table 4. Number of minimal repairs each value of N (t) to determine V (t). As an example,
we have estimated NHPP parameters with 1 times 5
failures, the results are presented in figure 9 and ta-
ble 6. We note that if b parameter is rather closed
The component has undergone a minimum of 1 min-
to theoretical value 1, 5; it is quite different for a pa-
imal repair and a maximum of 9 minimal repairs. The
rameter that depends essentially on time of the first
mean is equal to 5,13 repairs. This result points out
failure, which is random. That is why, generally, it is
a general problem encountered in industry that the
preferred to focus the study of repairable systems on
amount of data collected is too small for a valid statis-
optimization of the availability of the system and its
tical analysis.
components.
4.3 Determination of NHPP parameters
(Power law) N a b ρ2
The aim of this section consists in determining the pre- 5 0,15 1,40 0,98
cision reached in practice when NHPP parameters are 5 0,11 1,64 0,97
evaluated from a sample of repair times. Let’s consider 5 0,73 1,18 0,99
the one-component system described in section 4.2. We
have simulated 100 sets of 5, 10, 20 repair times (we Table 6. Estimation of a NHPP parameters with 1 × 5
have reduced the discarded criterion to R = 1e − 8 to failures
get more than 9 failures). From equation 19 and 34,
we calculate the theoretical power law model parame-
ters: a = 0, 125 and b = 1, 5. Now, we will determine
5 Conclusion
a and b parameters on basis of failure data. For that, In this paper, we wonder about the applicability of
we simulate 100 times 5 (1st case), 10 (2nd case), and reliability theory in an industrial context and more
20 (3rd case) failures, reparations are supposed to be specifically on uncertainty affecting reliability param-
instantaneous. Next, in each case and for each value eters evaluated on the basis of a small set of data. For
of N (t) we calculate the mean time. After, we deter- this purpose the methodology used and implemented
mine the best line that fits these mean times (fig. 8). in a MATLAB library Simulatrix lies on Monte Carlo
The results obtained are presented in table 5, ρ2 is the simulations.
coefficient of correlation. We note that a and b values First, we have tried to determine model imprecisions
converge to the theoretical results. for non repairable systems. To this purpose, we have
considered a system of one component of which the ceptions, and their causes. New York: Marcel
reliability distribution is a Weibull. We have simulated Dekker, 1984.
several times of failures and after have identified the
distribution parameters. After, we have determined [2] O. Basile, P. Dehombreux, and E. Filippi.
the influence of this uncertainty on parameters such Computer-aided reliability analysis of mechanical
as mean up time and reliability prediction. The main systems ; some comments on estimation errors.
conclusion is we need almost 50 failures to get an error 6th National Congress on Theoretical and Applied
smaller than 10% on mean up time prediction. Mechanics, pages NCTAM–2003–102, 2003.
Second, we have payed a special attention to re- [3] Olivier Basile, Pierre Dehombreux, and Enrico
pairable systems and have summarized the broad lines Filippi. Estimations a priori et a posteriori
of the theory. Repairable systems approach is rather des lois de fiabilité : quelques réflexions sur
different than non repairable systems by the fact l’incertitude des grandeurs fiabilistes. Actes de
that for the first type it is preferred counting the Petom’03 Colloque francophone sur les Perfor-
events rather than studying their times distribution. mances et Nouvelles Technologies en Mainte-
For that, repairable systems theory defines counting nance, pages 329–346, 2003.
processes such as homogeneous Poisson process and
non-homogeneous Poisson process and to characterize [4] Christiane Cocozza-Thivent. Processus stochas-
maintenance operations renewal processes. In particu- tiques et fiabilité des systèmes. Springer, 1997.
lar, we have focused this article on non-homogeneous [5] Charles E. Ebeling. An Introduction to Reliability
Poisson process and generalized renewal process that and Maintainability Engineering. McGraw-Hill,
allow to take into consideration maintenance operation 1997.
efficiency.
Moreover, we have demonstrated that when inter- [6] I. Gertsbakh. Reliability Theory with Applications
arrival times between two failures (or repairs) are to Preventive Maintenance. Springer, 2000.
Weibull distributed and the system, after failing, is
[7] E. E. Lewis. Introduction to Reliability Engineer-
restored to the state just before failing, the counting
ing. McGraw-Hill, 2nd edition, 1994.
process of these events is the non-homogeneous Pois-
son process. Then, the rate of occurrence of failures [8] Patrick Lyonnet. La Maintenance, mathématiques
is continuous and its expression a power law. Before et méthodes. Tec & Doc, 4ème edition, 2000.
identifying NHPP parameters on the basis of simula-
tions, we wonder about the number of minimal repairs [9] Claude Marcovici and Jean-Claude Ligeron. Util-
allowed for a component restored to a minimal state isation des techniques de fiabilité en mécanique.
(as bad as old). For a single system of which the dis- Technique et documentation Paris, 1974.
tribution is Weibull with β = 1, 5 and η = 4 [ut], we
[10] Henri Procaccia and Patrick Morilhat. Fiabilité
have calculated a mean of 5,1 repairs for a discarded
des structures des installations industrielles. Ey-
reliability of 1%. This result points out the difficulty
rolles, 1996.
to get a large sample.
After, we have evaluated NHPP parameters on the [11] Medardo Yañez, Francisco Joglar, and Mahom-
basis of simulated life cycles. More precisely, we have mad Modarres. Generalized renewal process for
tried to identify these parameters on the basis of 5, 10, analysis of repairable systems with limited fail-
15 and 20 failures (we have reduced the discarded cri- ure experience. Reliability Engineering and Sys-
terion to get such number of failures). The results ob- tem Safety, 77:167–180, April 2002.
tained corroborates theoretical parameters values. But
because this identification needs a large number of fail-
ures, it is very difficult to apply this methodology in
an industrial context. This conclusion explains why
generally the optimization of availability is emphasized
rather than reliability for repairable systems.
Acknowledgments
The research reported here has been supported by Min-
istère de la Région Wallonne, DGTRE, under contracts
215206 (FUCaM) and 215362 (FPMs).
References
[1] Harold Ascher and Harry Feingold. Repairable
Systems Reliability - modeling, inference, miscon-