Introduction to repairable systems STK4400 Spring 2011

Size: px

Start display at page:

Download "Introduction to repairable systems STK4400 Spring 2011"

Osborne Douglas
5 years ago
Views:

1 Introduction to repairable systems STK4400 Spring 2011 Bo Lindqvist bo/ Bo Lindqvist Introduction to repairable systems

Definition of repairable system Ascher and Feingold (1984): A repairable system is a system which, after failing to perform one or more of its functions

2 Definition of repairable system Ascher and Feingold (1984): A repairable system is a system which, after failing to perform one or more of its functions satisfactorily, can be restored to fully satisfactory performance by any method, other than replacement of the entire system. Bo Lindqvist The trend-renewal process

3 Ascher and Feingold s mission in 1984 Ascher and Feingold presented the following example of a happy and sad system: Their claim: Reliability engineers do not recognize the difference between these cases since they always treat times between failures as i.i.d. and fit probability models like Weibull. Use nonstationary stochastic point process models to analyze repairable systems data! Bo Lindqvist The trend-renewal process

4 Today: Recurrent events extensively studied 0 T 1 T 2 T N τ Observe events occurring in time Applications: engineering and reliability studies, public health, clinical trials, politics, finance, insurance, sociology, etc. Reliability applications: breakdown or failure of a mechanical or electronic system discovery of a bug in an operating system software the occurrence of a crack in concrete structures the breakdown of a fiber in fibrous composites Warranty claims of manufactured products Bo Lindqvist The trend-renewal process

5 Important aspects for modelling and analysis 0 T 1 T 2 T N τ Trend in times between events? Renewals at events? Randomness of events? Dependence on covariates? Unobserved heterogeneity ( frailty, random effects ) among individual processes? Dependence between event process and the censoring at τ? Bo Lindqvist The trend-renewal process

6 Typical data format 0 T 11 T 21 T N11 τ 1 0 T 1j T 2j T Nj j τ j. 0 T 1m T 2m T Nmm τ m Covariates if available (fixed or time-varying):. X j (t); j = 1, 2,...,m Bo Lindqvist The trend-renewal process

7 Proschan (1963): The classical aircondition data Times of failures of aircondition system in a fleet of Boeing 720 airplanes Bo Lindqvist The trend-renewal process

8 Nelson (1995): Valve seat data Times of valve-seat replacements in a fleet of 41 diesel engines Event Plot for Valve Seat Replacements Diesel Engine Time Bo Lindqvist The trend-renewal process

9 Barlow and Davis (1977): Tractor data Bo Lindqvist The trend-renewal process

10 Basic models for repairable systems 0 T 1 T 2 T N τ RP(F): Renewal process with interarrival distribution F. Defining property: Times between events are i.i.d. with distribution F NHPP(λ( )): Nonhomogeneous Poisson process with intensity λ(t). Defining property: 1 Number of events in (0,t] is Poisson-distributed with expectation t 0 λ(u)du 2 Number of events in disjoint time intervals are stochastically independent Bo Lindqvist The trend-renewal process

11 Point process modelling of recurrent event processes t 0 T 1 T 2 T N τ φ(t F t ) F t = history of events until time t. Conditional intensity at t given history until time t, φ(t F t ) = lim t 0 Pr(failure in [t, t + t) F t ) t Bo Lindqvist The trend-renewal process

12 Special cases: The basic models NHPP(λ( )): φ(t F t ) = λ(t) so conditional intensity is independent of history. Interpreted as minimal repair at failures RP(F) (where F has hazard rate z( )): φ(t F t ) = z(t T N(t ) ) so conditional intensity depends (only) on time since last event. Interpreted as perfect repair at failures Between minimal and perfect repair? Imperfect repair models. Bo Lindqvist The trend-renewal process

13 Trend Renewal Process TRP (BL, Elvebakk and Heggland 2003) Trend function: λ(t) (cumulative Λ(t) = t 0 λ(u)du) Renewal distribution: F with expected value 1 (for uniqueness) 0 T 1 T 2 T 3 t TRP(F, λ( )) RP(F) 0 Λ(T 1 ) Λ(T 2 ) Λ(T 3 ) SPECIAL CASES: NHPP: F is standard exponential distribution RP: λ(t) is constant in t Bo Lindqvist The trend-renewal process

14 Motivation for TRP: Well known property of NHPP 0 T 1 T 2 T 3 t NHPP(λ( )) HPP(1) 0 Λ(T 1 ) Λ(T 2 ) Λ(T 3 ) Bo Lindqvist The trend-renewal process

15 Point process formulation of TRP: Conditional intensity of TRP(F, λ( )): φ(t F t ) = z(λ(t) Λ(T N(t ) ))λ(t) where z( ) is hazard rate of F Recall special cases: NHPP: z( ) 1, implies φ(t F t ) = λ(t) RP: λ( ) 1, implies φ(t F t ) = z(t T N(t ) ) Bo Lindqvist The trend-renewal process

16 Comparison of NHPP, TRP and RP: Conditional intensities Conditional intensity Time Example: Failures observed at times 1.0 and Conditional intensities for NHPP (solid); TRP (dashed); RP (dotted) Bo Lindqvist The trend-renewal process

17 Conditional intensities of TRP with DFR renewal distribution and increasing trend function Failures observed at times 1.0 and Renewal: F Weibull, shape 0.5 Trend: λ(t) = t 2 Bo Lindqvist The trend-renewal process

18 Interpretation of Λ(t) 0 T 1 T 2 T 3 t TRP(F, λ( )) RP(F) 0 Λ(T 1 ) Λ(T 2 ) Λ(T 3 ) By using results from renewal theory: N(t) lim t Λ(t) = 1 (a.s.) E(N(t)) lim = 1 t Λ(t) since F is assumed to have expected value 1. Bo Lindqvist The trend-renewal process

19 Alternative models: Cox modulated renewal model Cox (1972): φ(t F t ) = z(t T N(t ) )λ(t) Compare to TRP: φ(t F t ) = z(λ(t) Λ(T N(t ) ))λ(t) Bo Lindqvist The trend-renewal process

20 Alternative models: The model by Lawless and Thiagarajah Lawless and Thiagarajah (1996): φ(t F t ) = e θ h(t) where h(t) is a vector of observable functions, depending on t and possibly also on the history F t, and θ is a vector of parameters. This model allows the use of various dynamic covariates, such as number of previous failures of the system, N(t ) time since last event, t T N(t ) Bo Lindqvist The trend-renewal process

21 Alternative models: The model by Pena and Hollander Pena and Hollander (2004): φ(t F t, a) = a λ 0 (E(t)) ρ(n(t ); α) g(x(t), β) where a is a positive unobserved random variable (frailty) λ 0 ( ) is a baseline hazard rate function E(t) is effective age process, a predictable process modeling impact of performed interventions after each event. Possible choices Minimal intervention on repair: E(t) = t Perfect intervention on repair: E(t) = t T N(t ) Imperfect repair: E(t) = time since last perfect repair Age reduction models: E(t) is reduced by a certain factor at each repair ρ( ; ) can for example be on geometric form, ρ(k; α) = α k g(x(t), β) models impact of covariates, for example on Cox-form Bo Lindqvist The trend-renewal process

22 CONDITIONAL ROCOF BY MINIMAL REPAIR (NHPP) AND PERFECT REPAIR (RENEWAL PROCESS) 167

23 SIMPLE EXAMPLE WITH THREE SYSTEMS Sys. 1: 0 S 11 =5 S 21 =12 S 31 =17 τ 1 =20 Sys. 2: 0 S 12 =9 S 22 =23 τ 2 =30 Sys. 3: 0 S 13 =4 τ 3 =10 Proj: t Y(t): Y (t) =3 Y (t) =2 Y (t) =1 168

24 COMPUTATIONS FOR THE NELSON-AALEN ESTIMATOR t 1/Y (t) 1/Y (t) 2 Ŵ (t) VarŴ (t) SDŴ (t) 4 1/3 1/9 1/3 1/ /3 1/9 2/3 2/ /3 1/9 1 1/ /2 1/4 3/2 7/ /2 1/4 2 5/ /

25 ESTIMATED W (t) with 95% confidence limits (Nelson-Aalen) W(t) t 170

26 Simple Example With 3 Systems 174

27 Nelson-Aalen estimator for Cumulative ROCOF W (t) 1. Order all failure times as t 1 <t 2 <...t n. 2. Let d j (t i ) = # events in system j at t i. 3. Let d(t i )= m j=1 d j(t i ) = # events in all systems at t i. 4. Let Y j (t) = { 1 if system j is under observation at time t 0 otherwise 5. Let Y (t) = m j=1 Y j(t) = # systems under observation at time t. Then Under general assumptions: Ŵ (t) = Assuming NHPP: Var Ŵ (t) = t i t t i t d(t i ) {Y (t i )} 2 d(t i ) Y (t i ). Under general assumptions (MINITAB): Var Ŵ (t) = { m j=1 t i t Y j (t i ) Y (t i ) [ d j (t i ) d(t ] } 2 i) Y (t i ) 175

28 Illustration of last formula for Simple NHPP Example (Compare with MINITAB Output): Var Ŵ (4) = { 1 3 [ { 1 + 3]} 3 [ { 1 + 3]} 3 [ ]} 2 = 6 81 = Var Ŵ (5) = + + { 1 3 { 1 3 { 1 3 [ ] [ ] [ ] [ 1 1 ]} 2 3 [ 0 1 ]} 2 3 [ 0 1 ]} 2 3 = 6 81 =

29 Var Ŵ (9) = { 1 3 { { = 0 [ ] [ ] [ ] [ ] [ ] [ ] [ 0 1 ]} 2 3 [ 1 1 ]} 2 3 [ 0 1 ]} 2 3 Var Ŵ (12) = + + { 1 3 { 1 3 { 1 3 [ ] [ ] [ ] [ ] [ ] [ ] [ ] [ ] [ ]} [ 1 1 ]} 2 2 [ 0 1 ]} 2 2 = 1 8 =

30 Simple Example With 3 Systems Power Law NHPP Model: W (t; α, θ) =(t/θ) α 178

31 Simple Example With 3 Systems 179

32 RESIDUAL PROCESS: SIMPLE EXAMPLE. Data points (and endpoints on axes) are transformed with the estimated cumulative ROCOF, Ŵ (t) = t 1.20 Sys. 1: Sys. 2: Sys. 3: Times between events, plus censored times at the end of each axis, are on the next slide anlysed by MINITAB as a set of censored exponential variables. 180

33 181

34 182

35 VALVESEAT DATA 183

36 VALVESEAT DATA 185

37 VALVESEAT DATA 186

38 188

39 189

40 Grampus- data: Plot of (T i,t i+1 ) to investigate whether times between failures can be assumed independent. The figure does not indicate a correlation between successive times. 190

41 191

42 192

43 PROFILE LIKELIHOOD FOR BETA ( SIMPLE EXAMPLE ) ˆβ =1.20, ˆλ = profile beta 193

44 CONNECTION BETWEEN LAMBDA OG BETA ( SIMPLE EXAMPLE ) ˆβ =1.20, ˆλ = lambda beta 194

45 195

46 196

47 197

48 198

49 Valveseat Data 199

50 Valveseat Data 200

51 TTT-analysis Simple Example Row STTT ID Scaled , , , , , , ,00000 Parameter Estimates Standard 95% Normal CI Parameter Estimate Error Lower Upper Shape 1, ,511 0, ,25186 Scale 0, ,160-0, , Trend Tests MIL-Hdbk-189 Laplace s Anderson-Darling Test Statistic 9,59 0,12 0,24 P-Value 0,697 0,906 0,977 DF

52 Total Time on Test Plot for Simple Example System Column in ID 1,0 P arameter, M LE Shape Scale 1, , Scaled Total Time on Test 0,8 0,6 0,4 0,2 0,0 0,0 0,2 0,4 0,6 Scaled Failure Number 0,8 1,0 202

53 TTT-analysis of Valve Seat Data Parametric Growth Curve: C1 Model: Power-Law Process Estimation Method: Maximum Likelihood Parameter Estimates Standard 95% Normal CI Parameter Estimate Error Lower Upper Shape 1, ,202 1, ,79229 Scale 0, ,026 0, , Trend Tests MIL-Hdbk-189 Laplace s Anderson-Darling Test Statistic 68,72 2,03 3,17 P-Value 0,032 0,043 0,022 DF

54 Total Time on Test Plot for Valve Seat Data 1,0 Parameter, MLE Shape Scale 1, , Scaled Total Time on Test 0,8 0,6 0,4 0,2 0,0 0,0 0,2 0,4 0,6 Scaled Failure Number 0,8 1,0 204

e 4β e 4β + e β ˆβ =0.765

e 4β e 4β + e β ˆβ =0.765 SIMPLE EXAMPLE COX-REGRESSION i Y i x i δ i 1 5 12 0 2 10 10 1 3 40 3 0 4 80 5 0 5 120 3 1 6 400 4 1 7 600 1 0 Model: z(t x) =z 0 (t) exp{βx} Partial likelihood: L(β) = e 10β e 10β + e 3β + e 5β + e 3β