See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/357157823

Recent development on reliability assessment for a

multi-component system with masked failure data

Presentation · December 2021

CITATIONS READS

0 50

1 author:

Shaomin Wu
University of Kent
182 PUBLICATIONS 4,019 CITATIONS

SEE PROFILE

All content following this page was uploaded by Shaomin Wu on 18 December 2021.

The user has requested enhancement of the downloaded file.

Recent development on reliability assessment for a
multi-component system with masked failure data

Prof Shaomin Wu
Kent Business School

1
Repairable systems: failures, repair,
 Analysis of recurrent event data:

𝑡1 𝑡1 + 𝑧1 𝑡2 𝑡2 + 𝑧2 𝑡3 𝑡3 + 𝑧3

: failed
: repaired
𝑡𝑘 : time when the 𝑘th failure occurred
𝑡𝑘 + 𝑧𝑘 : time when the 𝑘th failure is fixed

𝑋1 𝑋2 𝑋3 𝑋4
𝑌1 𝑌2 𝑌3
𝑋𝑘 : the 𝑘th working time, or gap time
𝑌𝑘 : time duration of the 𝑘th repair,
2
Questions & Applications
 Recurrent event analysis! Below are some applications
– modelling times between failures of technical systems (software or hardware)
– modelling the number of insurance/warranty claims
– modelling the numbers of patient visits to their doctors

 We are interested in the following questions:

A. How can we estimate the length of working times/gap times, i.e., 𝑋𝑘 (𝑘 = 1,2, … )?
B. How many occurrences are there within a given time period?

3
Questions A & B
 Recurrent events:

We are interested in
 Notations:
𝑛
1. What are the distributions of the gap times? (Question A)
2. How many events occurred within a given time?
𝑆𝑛 = ෍ 𝑋𝑖
𝑖=1
(Question B)
∞
𝑍𝑡 = ෍ 𝜒{𝑆𝑛 ≤ 𝑡} = sup{𝑛: 𝑆𝑛 ≤ 𝑡}
Equivalently,
𝑛=1
1. What are the distributions of 𝑿𝒊 ?
𝑚 𝑡 = 𝐸[𝑍𝑡 ] 2. H𝐨𝐰 𝐜𝐚𝐧 𝐰𝐞 𝐞𝐬𝐭𝐢𝐦𝐚𝐭𝐞 𝒎(𝒕)?

The estimation becomes more complicated for

a multi-component system! 4
The challenge
 Consider a system composed of 𝑛 sockets in series, there is a component in each socket

 Superposition of renewal processes:

– If a component fails, it is replaced with a new, identical one
– The failure process of the system is a superimposed renewal process (SRP)

5
The challenge—cont’d
 In the real world
– The total number of failures of the SRP can be obtained if the reliability of each component is given;
– Nevertheless, if the times between failures of the system are available, but the components that cause the
system to fail are unknown, then the failure data are masked failure data
 However, the SRP requires a model of the failure process of each individual component. In practice,
– We may not have historical data of which component causes the system to fail
– We do not have many failure data

 Question: how can we build a model merely based on times-between-failures of the system,
without knowing the failure process model of each individual component?
6
Literature review

 Cox and Smith (1954) prove that the observed failure times of an SRP tend to be distributed with a
homogeneous Poisson process when the number of components goes to infinity and the time is far
from the origin;

 Khinchin (1956) further clarifies that the SRP tends to be a non-homogeneous

Poisson process (NHPP);

 Drenick (1960) proves the same property as that of Cox and Smith (1954) even if the failure
processes of the multi-sources follow heterogeneous renewal processes;

 Recent publications considers methods to approximate the SRP, considering different scenarios
(see Wu (2019 b) for a review)

 Conclusion: when the number of components goes to infinity and the time is far from the origin,
the SRP can be approximated by NHPP
7
Effectiveness of a repair

Scenario 1: Perfect repair, i.e., as good as new (AGAN)

new failed

Scenario 2: Minimal repair, i.e., as bad as old (ABAO)

new failed

Scenario 3: Imperfect repair, between AGAN and ABAO

new failed

Scenario 4: Better than new repair;

Scenario 5: Worse than old repair

To estimate the distribution of 𝑋𝑘 , one can use different methods of modelling the gap times,
including renewal processes, NHPP, etc
8
Existing modelling methods

 Differs from life data analysis, where events are assumed to be iid

 Non-parametric methods A ---- As good as new

 Parametric methods, for example

– Renewal process (RP)
– Nonhomogeneous Poisson process (NHPP)
B ---- As bad as old
– Geometric process
– Cox process (DPP)

C ---- between A & B

9
Renewal process: The model for perfect repair

 Given a sequence of random variables {𝑋𝑘 , 𝑘 = 1,2, … }, if they are independent and the
cdf of 𝑋𝑘 is given by 𝐹(𝑥) for 𝑘 = 1,2, …, then {𝑋𝑘 , 𝑘 = 1,2, ⋯ } is called a renewal
process (RP)

 The expected number of jumps observed up to some time 𝑡, i.e., renewal function, is
𝑡
𝑚 𝑡 = 𝐹 𝑡 + න 𝑚 𝑡 − 𝑢 𝑓 𝑢 𝑑𝑢
0

10
NHPP: The model for minimal repair

 Nonhomogeneous Poisson process (NHPP). Below are the assumptions

– 𝑁(0) = 0 (where 𝑁(𝑡) is the number of failures and is a random variable)
– The number of events in disjoint intervals are independent
– No events happen simultaneously

 Probabilities of a given number of failures for the NHPP model are calculated by
Λ(𝑡)𝑘 𝑒 −Λ(𝑡)
𝑃 𝑁(𝑡) = 𝑘 =
𝑘!
𝑡
where Λ(𝑡) is the cumulative failure intensity, i.e., Λ 𝑡 = ‫׬‬0 𝜆 𝑢 𝑑𝑢
– If Λ 𝑡 = 𝜆0 𝑡, then the NHPP reduces to the HPP
– The most widely used failure intensity function is the power law

𝜆 𝑡 = 𝛼𝑡 𝛽

11
Models for imperfect repair

 Age-modification methods. Let 𝑉𝑘 be the virtual age after the 𝑘th maintenance
– 𝑉𝑘 = 𝑉𝑘−1 + 𝐴𝑘 𝑋𝑘 , or 𝑉𝑘 = 𝐴𝑘 (𝑉𝑘−1 + 𝑋𝑘 )

 Intensity-modification methods. Let 𝜆𝑘 (𝑡) be the failure intensity of the 𝑘th

maintenance (𝐴𝑘−1 is the effectiveness of repair)
– 𝜆𝑘 𝑡 = 𝐴𝑘−1 𝜆𝑘−1 (𝑡)

 A hybrid method*
– Given a sequence of non-negative random variables {𝑋𝑘 , 𝑘 = 1,2, …}, if they are independent
and the cdf of 𝑋𝑘 is given by 𝐹(𝑎𝑘−1 𝑥 ℎ 𝑘 ) for 𝑘 = 1,2, …, where 𝑎 is a positive constant, ℎ(𝑘)
is a function of 𝑘 and the likelihood of the parameters in ℎ(𝑘) has a known closed form, and
ℎ(𝑘) > 0 for 𝑘 ∈ 𝑁 + , then {𝑋𝑘 , 𝑘 = 1,2, … } is called a doubly geometric process (DGP)
– 𝜆𝑘 𝑡 = 𝑎𝑘−1 𝜆𝑛−1 𝑎𝑘−1 𝑡 if ℎ 𝑘 = 1
12
Model I
 Consider a series system with three components with hazard functions 0.4𝑡,
0.45𝑡, and 0.5𝑡, respectively. If a component fails and then is replaced, at least a
failure intensity of inf{0.4𝑡, 0.45𝑡, 0.5𝑡} ≡ 𝜙1 (𝑡) is retained. As such, we may
assume the intensity function of the system is the sum of two sub-functions

– Sub-function 1: this part, or𝜙1 𝑡 , does not change if a failed component is repaired
 We may use the non-homogeneous Poisson process to model this part

– Sub-function 2: this part, or 𝜙2 𝑡 , models the changing intensity function upon

repair

Wu, S. and Scarf, P., 2017. Two new stochastic models of the failure process of a series system. European Journal of
Operational Research, 257(3), pp.763-772. 13
Sub-function 2: Modelling the changing intensity function
 Assume
– the number of components in a system is 𝑚, which is known
– the failure rate function of each component is 𝑚
1
𝜆2 (𝑡)

 Modelling the changing intensity function 𝜙2 𝑡 . If the components are

identical at time 𝑡 = 0, a convenient method is to assume that the components
fail in turn

14
15
16
17
18
The model

 Mathematically, the failure intensity of the two parts can be expressed by

𝜙1 𝑡 𝐻𝑡 − = 𝜆1 (𝑡)
and
𝜆2 𝑡 if 𝑁𝑡 = 0
1 𝑁 −1
𝜙2 𝑡 𝐻𝑡 − = 𝑚
𝑡
σ𝑘=0 𝜆2 𝑡−𝑇𝑁𝑡 −𝑘 + 𝑚−𝑁𝑡 𝜆2 (𝑡) if 1 ≤ 𝑁𝑡 < 𝑚
1 𝑁𝑡 −1
σ 𝜆 𝑡 − 𝑇𝑁𝑡 −𝑘 if 𝑁𝑡 ≥ 𝑚
𝑚 𝑘=0 2

respectively

19
Simple moving average method

 Given a time series 𝑥1 , 𝑥2 , … , 𝑥𝑡 , you are asked to provide a forecast value at time
𝑡 + 1, i.e., 𝑥ො𝑡+1 , where 𝑥ො𝑡+1 denotes a forecast of 𝑥𝑡+1

1 𝑝−1
 Let 𝑥ොt+1 = σ𝑖=0 𝑥𝑡−𝑖 , then 𝑥ොt+1 is a forecast with the simple moving average
𝑝
method

20
Numerical examples: settings

 Settings: Assume
– There are 𝑀 identical systems
– Each system may experience 𝑀𝑗 failures (times-between-failures of different systems may be
different)
– Each system is composed of 𝑚 components, the lifetime distribution of each component
𝑡 𝛽
follows 𝐹 𝑡 = 1 − exp − , where 𝛼 is randomly chosen from (12,60) (i.e., 𝛼 ∈ (12,60))
𝛼
and 𝛽 is randomly chosen from (0.5, 4), (i.e., 𝛽 ∈ (0.5,4))
 RP: renewal process;
 GP: geometric process;
 NHPP-PL: NHPP with the power law;
 GRP: Kijima model I

21
Numerical examples

22
Simple exponential method

 Given a time series 𝑥1 , 𝑥2 , … , 𝑥𝑡 , you are asked to provide a forecast value at time
𝑡 + 1, i.e., 𝑥ො𝑡+1 , where 𝑥ො𝑡+1 denotes a forecast of 𝑥𝑡+1

𝑝−1
 Let 𝑥ොt+1 = σ𝑖=0 𝛼 𝑡−𝑖 𝑥𝑡−𝑖 , then 𝑥ොt+1 is a forecast with the simple exponential
smoothing method

Wu, S., 2019 (a). A failure process model with the exponential smoothing of intensity functions. European
Journal of Operational Research, 275(2), pp.502-513.
23
Assumptions
 Suppose the failure process of a real series system of multiple components. Once
the system fails, the failed component can be immediately identified and replaced
with a new identical one;
1
 Assume that there are 𝑚 components, which have failure rate functions 𝜆 (𝑡),
𝑚 0
1 1 1 1
𝜌𝜆0 𝑡 , 𝑚 𝜌2 𝜆0 𝑡 , … , 𝑚 𝜌𝑚−2 𝜆0 (𝑡), and 𝑚 𝜌𝑚−1 𝜆0 (𝑡), respectively;
𝑚
 The effectiveness of repair is: once a repair is conducted, the component with
1
intensity function 𝜆 (𝑡) is replaced, and the intensity function of other
𝑚 0
1 1
component changes from 𝑚 𝜌𝑚−𝑘 𝜆0 (𝑡) to 𝑚 𝜌𝑚−𝑘−1 𝜆0 (𝑡)

24
Model II: failure process model with the exponential smoothing of intensity functions

 The model

 ESI: Exponential Smoothing of Intensity

 MAI: Moving Average of Intensity for the model with 𝜌 = 1

25
Experimental results on simulated data

26
Experimental results on real datasets

 The model outperforms many other existing models on real-world datasets and does
not need to assume “a failure component is replaced/renewed”

27
Comparison on 15 real-world failure datasets

28
Performance on the real world datasets

 If comparing the values of the AICc and BIC of the 15 models, then one can find that the
MAI has the smallest values of the AICc and BIC in 11 out of the 15 cases. That is, the
MAI outperforms the other models in terms of both AICc and BIC
29
Performance on a real-world dataset

 Cumulative air-conditioning failures and estimated mean cumulative functions

30
Conclusion and further development

 Conclusions
– Development of methods to estimate the failure process of a series system with masked
failure data can find applications in practice
– Sample size is usually small; there is a need to develop models with a small number of
parameters
– Real datasets are needed to validate any proposed methods

 Further development
– The above developments were on approaches to approximating the SRP (superposition of
renew processes)
– One may extend the above methods to the situation in which a failed component is
maintained with imperfect repair

31
References

 Cox, D.R. and Smith, W.L., 1954. On the superposition of renewal processes. Biometrika, 41(1-2), pp.91-99.
 Drenick, R.F., 1960. The failure law of complex equipment. Journal of the Society for Industrial and Applied
Mathematics, 8(4), pp.680-690.
 Khinchin, A.I., 1956. On Poisson streams of random events. Theory of Probability and Its Applications, 1, pp.248-
255.
 Wu, S., 2017. Doubly geometric processes and applications. Journal of the Operational Research Society, pp.1-13.
 Wu, S. and Scarf, P., 2017. Two new stochastic models of the failure process of a series system. European Journal
of Operational Research, 257(3), pp.763-772.
 Wu, S., 2019 (a). A failure process model with the exponential smoothing of intensity functions. European
Journal of Operational Research, 275(2), pp.502-513.
 Wu, S., 2019 (b). Superimposed Renewal Processes in Reliability. Wiley StatsRef: Statistics Reference Online, pp.1-
12.
 Wu, S., 2021. Two methods to approximate the superposition of imperfect failure processes. Reliability
Engineering & System Safety, 207, p.107332.

32
 This is a keynote talk given at Symposium on Recent Trends
in Engineering Research, 23th May 2021, College of
Engineering, national University of Science and Technology,
Sultanate of Oman.

Thanks the organiser for their invitation

33
View publication stats