A novel approach to modelling time-between failures of a repairable system. Dr Shaomin Wu Kent Business School

Size: px

Start display at page:

Download "A novel approach to modelling time-between failures of a repairable system. Dr Shaomin Wu Kent Business School"

Chastity Bond
5 years ago
Views:

1 A novel approach to modelling time-between failures of a repairable system Dr Shaomin Wu Kent Business School 1

2 Wu, S., A failure process model with the exponential smoothing of intensity functions. European Journal of Operational Research. 2

3 The failure process of a multiple component system Consider a series system composed of three components Assumptions Once a component fails, it is immediately replaced with an identical one, The failure mode is not recorded/unknown The failure process of the system is a super-imposed renewal process (SRP) RP RP RP SRP 3

4 Literature review The GP has been extensively studied since its introduction in 1988, mainly due to its elegance and convenience in deriving mathematical properties in applications Failure process models Simplest models Extended models Hybrid intensity models Other models Renewal process Poisson process Age modification models Intensity modification models (pp,11 pp) models Other models 4

5 Limitations in the existing models Limitations of the practical application of SRP (super-imposed renewal process): In reality, the number of failures of a system is limited Too many parameters in the SRP uncertainty of the parameters in a model becomes large A limitation in the existing models: Only considering the same effectiveness of maintenance upon failures in the lifecycle of a system Replacing different components causes different effectiveness of maintenance A possible solution Constraint: unknown failure modes + limited number of failures Solution: use the Exponential Smoothing methods in failure intensity function 5

6 Two forecasting models Assume we are interested in a time series xx 1, xx 2,, xx nn 6

7 Two forecasting models Assume we are interested in a time series xx 1, xx 2,, xx nn Model 1: Simple exponential smoothing method xx nn+1 = ααxx nn + (1 αα) xx nn where: xx nn is the last period s forecast; xx nn is the last periods actual value; and 0 < αα < 1 We modify the simple exponential smoothing method and think of the following method pp 1 xx nn+1 = 1 pp ii=0 αα ii xx nn ii 7

8 Two forecasting models Assume we are interested in a time series xx 1, xx 2,, xx nn Model 1: Simple exponential smoothing method xx nn+1 = ααxx nn + (1 αα) xx nn where: xx nn is the last period s forecast; xx nn is the last periods actual value; and 0 < αα < 1 We modify the simple exponential smoothing method and think of the following method pp 1 xx nn+1 = 1 pp ii=0 αα ii xx nn ii Understanding the time series from a failure process perspective xx nn is the time to the nnth failure, i.e., it is the time to the latest failure, i.e., youngest failure, xx nn pp+1 is the time to the nn pp + 1 th failure, i.e., older failure αα ii xx nn ii implies that more weight on the oldest failure 8

9 Two forecasting models Model 1: Simple exponential smoothing method We modify the simple exponential smoothing method and think of the following method pp 1 xx nn+1 = 1 pp ii=0 αα ii xx nn ii Understanding the time series from a failure process perspective xx nn is the time to the nnth failure, i.e., it is the time to the latest failure, i.e., youngest failure, xx nn pp+1 is the time to the nn pp + 1 th failure, i.e., older failure αα ii xx nn ii implies that more weight on the oldest failure 9

10 Two forecasting models Model 1: Simple exponential smoothing method pp 1 xx nn+1 = 1 pp ii=0 αα ii xx nn ii Model 2: Moving average method. Let αα = 1, then we obtain the moving average model pp 1 xx nn+1 = 1 pp ii=0 xx nn ii 10

11 Two forecasting models Model 1: Simple exponential smoothing method pp 1 xx nn+1 = 1 pp ii=0 αα ii xx nn ii Model 2: Moving average method. Let αα = 1, then we obtain the moving average model pp 1 xx nn+1 = 1 pp ii=0 xx nn ii We consider two models The exponential smoothing of intensity model (ESI) The simply moving average of intensity model (SMI) 11

12 The failure process of a multiple component system Consider a series system composed of three components

13 The failure process of a multiple component system Consider a series system composed of three components At the 1st failure, we assume one of the components, component 1 say, fails; 13

14 The failure process of a multiple component system Consider a series system composed of three components At the 1st failure, we assume one of the components, component 1 say, fails; At the 2 nd failure, we assume another component, component 2 say, fails 14

15 The failure process of a multiple component system Consider a series system composed of three components At the 1st failure, we assume one of the components, component 1 say, fails; At the 2 nd failure, we assume another component, component 2 say, fails At the 3 rd failure, we assume component 3 fails 15

16 The failure process of a multiple component system Consider a series system composed of three components At the 1st failure, we assume one of the components, component 1 say, fails; At the 2 nd failure, we assume another component, component 2 say, fails At the 3 rd failure, we assume component 3 fails At the 4 th failure, we assume component 1 fails, VC1 VC2 VC3 VC1 VC2 VC3 16

17 The exponential smoothing of intensity model Assume a real-world system is composed of mm components. The reliability of each component is unknown. We approximate the failure process of the above system with that of the following virtual system. Assume a system composed of mm components with failure rate functions 1 mm ρρmm 1 λλ 0 tt, 1 mm ρρmm 2 λλ 0 tt,, 1 mm λλ 0 tt, respectively. That is, before the 1 st failure of the system, the failure intensity function of the system is assumed to be mm 1 λλ(tt H tt ) = 1 mm kk=0 ρρ mm kk 1 λλ 0 tt 17

18 The exponential smoothing of intensity model Then for 1 NN tt < mm, we have λλ(tt H tt ) = 1 mm NN tt 1 kk=0 ρρ mm kk 1 λλ 0 tt TT NNkk kk mm 1 + ρρ mm kk 1 λλ 0 (tt) kk=nn tt 18

19 The exponential smoothing of intensity model Then for 1 NN tt < mm, we have NN tt 1 mm 1 λλ(tt H tt ) = 1 mm kk=0 ρρ mm kk 1 λλ 0 tt TT NNkk kk + ρρ mm kk 1 λλ 0 (tt) kk=nn tt Then for NN tt mm, we have mm 1 λλ(tt H tt ) = 1 mm kk=0 ρρ mm kk 1 λλ 0 tt TT NNkk kk 19

20 The exponential smoothing of intensity model In this model, we have parameter ρρ and parameters in λλ 0 tt. λλ 0 tt may be the power law, for example, λλ 0 tt = ααααtt ββ 1 on which the number of parameters in total is 3 20

21 The number of parameters Denote GG kk (tt) as the probability distribution of time to the kk-th failure, TT kk, of the component with intensity function λλ 0 (tt). 21

22 The moving average of intensity model If ρρ = 1, then 22

23 Comparison RP (renewal process) NHPP-PL (non-homogeneous Poisson process with the power law) GP (geometric process) Kijima I (virtual age process I) Kijima II (virtual age process II) ARI mm (Arithmetic Reduction of Intensity model) ARA mm (Arithmetic Reduction of Age model) BBIP(Bounded Bathtub Intensity Process) Model II (Wu & Scarf, 2017) ESI (Exponential Smoothing of Intensity model) MAI (Moving Average of Intensity model) 23

24 Number of parameters and performance metrics The numbers of parameters Performance metrics 24

25 n: the number of systems; m: the number of components in each system 25

26 Real world cases 26

27 odel performance n the real-world atasets 27

28 Results 28

29 Thank you. Questions? 29

A novel repair model for imperfect maintenance

IMA Journal of Management Mathematics (6) 7, 35 43 doi:.93/imaman/dpi36 Advance Access publication on July 4, 5 A novel repair model for imperfect maintenance SHAOMIN WU AND DEREK CLEMENTS-CROOME School