06_competing risk model

Upload: ayariseifallah

Post on 02-Jun-2018

220 views

Category:

Documents


0 download

TRANSCRIPT

  • 8/10/2019 06_Competing Risk Model

    1/23

    Competing Risks

    Competing risks occur when a piece of equipmentcan be taken out of service for several reasons. In

    principle, one could imagine there being a time atwhich the equipment would be taken out of service

    for each reason. However, one can observe only thesmallest such time, and usually also the associatedreason; hence the dierent out of service! reasonscompete to be the one that is actually observed.Competing risk theory seeks to provide models of the

    "statistical# relationships between these times so thatone can determine quantities, such as the failuretime distribution for the equipment, which may notbe directly observable.

  • 8/10/2019 06_Competing Risk Model

    2/23

    Competing Risks Model asa Renewal Process

    $odel data as renewal process% a sequence of iid variables&', &(). Competing risk models essentially describe asituation in which the minimum of two or more lifetimevariables is observed together with the indicator of whichvariable is the smallest. *he unobserved variable is said to be

    censoredby the observed variable.

    In the reliability area the variables involved are typicallyfrom the following classes%

    *ime to critical failure from a particular failure mode*ime to preventive maintenance*ime to equipment being taken o line for some e+traneous

    reason and being replacedrepaired

  • 8/10/2019 06_Competing Risk Model

    3/23

    Classical Competing Risksframework

    -pplications of the classical competing risk modelassume renewals of the equipment; that is, that theequipment is restored to as good as new eachtime.

    *he bestknown technique within competing risks isthat of the KaplanMeier (KM) estimator, whichprovides a nonparametric estimator assuming

    independent censoring. *he total time on test"***# statistic used in reliability is the estimatorused under the assumption of e+ponentiallydistributed lifetimes and independent censoring

  • 8/10/2019 06_Competing Risk Model

    4/23

    Competing risksframeworkIf the variable / is the time of failure due to a failure

    mode and 0 and is the time of 1$ actions. *hevariable / may be masked by 0. 2e only observethe smallest of / and 0, i.e the observed variable is

    In other words we observe the least of /and 0 and observe which one it is.

    -pplications of the competing risk modelassume renewal of equipment.

    ( )[ ])(,, ZXIZXMinY

  • 8/10/2019 06_Competing Risk Model

    5/23

    Competing riskframework2e are interested in the survival function of +

    and 3to asses the failure rate corresponding todierent competing risks

    *he competing risk theory deal with the subsurvival function and its related conditionalprobabilities

    2here the su4+ 5 represent the subfunctions

    )()(),()( tZPtStXPtSZX

    >=>=

    ),()(),,()( *

    XZtZPtSZXtXPtSZ

    *

    X ==

  • 8/10/2019 06_Competing Risk Model

    6/23

    Competing Risk Problem

    *he competing risk problem is that of identifying theunderlying "marginal# distributions from the operationaldata. In general, one is unable to identify the marginalsfrom the data without making untestable assumptions.*he motivation for studying problems is to know how

    things would change if one was able to abolish a failurecause, stop doing maintenance, or perform some otheraction that changes the underlying relationship of thevariables. 6rom an operational research point of view, thecompeting risk problem is one e+ample of the more

    general modelling problem in which there is a need toconstruct models that support decision making, but thesemodels cannot always be easily identi7ed from e+istingdata.

  • 8/10/2019 06_Competing Risk Model

    7/23

    Coal Mill Failure modes

    Internal of Mill !ternal of Mill

    Mill rollers failure8oot cause% wear and tear,overload, 7re

    "rinding table failure8oot cause% segmentdisplacement, fatigue, trampmetal and 7re

    #ir port ring failure8oot cause% rotating vanesstrikes on static ring

    Mill motor bearing failure8oot cause% stress from millstart and stop, mis

    alignment, temperature andlife e+pired

    "earbo! failure8oot cause% vibration, misalignment, foreign ob9ects

    and load

  • 8/10/2019 06_Competing Risk Model

    8/23

    Competing risks inreliability

    Competing risk models do not have to involvepreventive maintenance as a censoringmechanism, it is clear that preventivemaintenance provides an important category ofsuch models in the reliability conte+t. Hence, ine+tending the available classes of competing riskmodels, a better understanding of the impact ofpreventive maintenance is important.

    *he construction of a competing risk modelsinvolving maintenance clearly relate to theconstruction of maintenance optimisation models.

  • 8/10/2019 06_Competing Risk Model

    9/23

    $ptimisation modelsIt is acknowledged that one of the reasons that many

    models have found limited application is that theassumptions made are typically strong.

    :ome maintenance optimisation models aredeveloped by ekker, "'

  • 8/10/2019 06_Competing Risk Model

    10/23

    $pportunisticmaintenance

    $pportunistic maintenance can be de7nes as a strategyby which preventive maintenance actions are carried out ona component during the unavailability of other components"for e+ample arising through failure#, rather than accordingto a preset schedule.

    *he concept of opportunistic maintenance modeling is ascarce commodity within industrial power plant.ngineering %udgments and risk analysis for good

    reasons, held wave in the area of maintenance modeling.

    pportunistic maintenance is one of many routines for upkeeping or, if necessary, improving the level of reliability of

    components and systems.

  • 8/10/2019 06_Competing Risk Model

    11/23

    Competing riskincorporatingopportunisticmaintenance*he competing risk is to identify the

    marginal distribution from competing risk

    data, here we consider a simple

    opportunistic maintenance policies.

    *he focus of this work is to provideidentifythe needs and criteria for taking an

    opportunity to do maintenance, the risk ofusing an opportunity, and the approacheswhich can best satisfy the need of

    opportunistic maintenance.

  • 8/10/2019 06_Competing Risk Model

    12/23

    $pportunity agereplacement model(&ekker and &i%kstra)/D time of failure. pportunities for 1$ arise following a

    1oisson process, 1@"#. *he maintenance policy is to usethe 7rst opportunity occurring after the equipment hasbeen operating for time t0. 2e denote the time of 7rst

    opportunity byZ

    .

    In the competing risk problem we try to show that thedistribution 6/"fXor SX# of / can be identi7ed in terms ofthe subsurvivor functions :5"t#.

    n the interval E@, t@# / is not censored and on the intervalE@, F# the variable / is independently censored by ane+ponentially distributed variable. *he overall model isidenti7able. Crowder $.G., "(@@'#, indqvist et al., "(@@=#.

  • 8/10/2019 06_Competing Risk Model

    13/23

    $pportunity agereplacement modelsimulation e!ample

    *he lifetime Xmay be censored byZ, the 7rst opportunity after a giventime t0. 6or illustration we have initially taken t0D@.= which is half the

    mean value ofX. *he conditional subsurvivor functions generated by

    '@@@ samples.

    *he simulated empirical conditional subsurvivor function tells us the

    probability of an event at time greater than t, given that the event iseventually observed.

  • 8/10/2019 06_Competing Risk Model

    14/23

    Repair'alert model

    *he model assumes that with a 7+ed probability p, andindependently of the lifetime, a signal is given o bythe equipment, which is sub9ected to repair instantly.

    *he time at which this signal is given o is dependent

    on the underlying h"t# of the equipment, that is, thedensity function for the 1$ time given failure time is

    the conditional survivor function have to be ordered inthis case, with ,0,)(

    )(Xt

    XH

    th

    ).(~

    )(~ **

    tStSZX

  • 8/10/2019 06_Competing Risk Model

    15/23

    $pportunity'alert model

    2e now assume that the model only gives the time at whicha signal would be given and that actual repair can onlyoccur opportunistically after the signal has been given.

    *he dierence between this model and the repairalert

    model is that repair can only occur at an opportunityfollowing the alert.

    6or failures taking place at short lifetimes, there is a lowchance that an opportunity will take place prior to failure,and there is an increased probability that the failure willactually be observed "increased with respect to the repairalert model#.

  • 8/10/2019 06_Competing Risk Model

    16/23

    $pportunity'alertimulation e!ample

    2e assume / to be 2eibull "shape D'.J and scaleD '#, and probabilityof alert to be @.?. - simulation with '@@@ sample gives the empiricalconditional subsurvivor function

    Conditional subsurvivor functions "$ signi7es opportunisticmaintenance, that is opportunity taken after a signal, while 6 signi7esfailure#.

    time

    0

    0.2

    0.4

    0.6

    0.8

    1

    1.2

    0 1 2 3 4 5

    t

    probability KM

    X

    -0.2

    0

    0.2

    0.4

    0.6

    0.8

    1

    1.2

    0 1 2 3 4 5

    OM

    F

  • 8/10/2019 06_Competing Risk Model

    17/23

    ock'opportunity model

    *he idea in this model is that discrete eventoccur that change the underlying failure rate"events such as internale+ternal shocks#

    *he model requires us to de7ne the failure timedistribution through a process rather thanthrough the more conventional survival functionapproach

    2e de7ne an increasing sequence of shock timesand for each period between shocks

    , a failure rate

    ,...,,0 210 SSS =

    ],( 1 ii SS

    ],0(),( 1 iii SStt

  • 8/10/2019 06_Competing Risk Model

    18/23

    *he distribution for the failure time is de7nedconditionally% Kiven that the equipmentsurvives the i-1th shock, , the probability

    that it survives the ith shock, , is equal to

    H"t# has an e+ponential distribution with mean

    '. 2e 7rst simulate an e+ponential variable V

    with mean ', and inductively simulate shocktimes until the integrated conditional ha3ard

    e+ceed V.

    1

    0 )(exp

    ii SS

    i

    dtt

    iSX>

    ock'opportunity model

    1> iSX

    ++=101

    00

    1 )(...)()(ii SS

    i

    SS

    i dttdttSH

  • 8/10/2019 06_Competing Risk Model

    19/23

    ock'opportunity modelIn order to model the censoring process%

    how opportunities arise, andwhen opportunities are taken.

    If denotes the constant ha3ard between shocks. *his

    implies that the times between successive shocks follow ane+ponential distribution. *he e+pression for the conditionaldistribution function in terms of the shocks is given as

    2here is the ha3ard between shocks and is thelargest event time less than t.

    .1),......,,|( )()(

    21

    1

    1 1 jji

    j

    i ii StSSt

    n eSSStF + =

    = +

    i

    jS

    i

  • 8/10/2019 06_Competing Risk Model

    20/23

    ock'opportunity model

    2e assume that there are two shocks . *he7rst shock occurs with an e+ponential distributedtime after the equipment has been taken intoservice "as new#, and has mean @.(. *he secondshock occurs with an e+ponential distributed time

    after the 7rst shock, and has mean @.?.

    In addition we assume that the equipment hasfailure rates , for

    the time intervals and equal to@.@', @.' and ' respectively. *his speci7es thedistribution of the failure time.

    321 and, ),,[),,0[ 211 SSS ),,[ 3 S

    21 and SS

  • 8/10/2019 06_Competing Risk Model

    21/23

  • 8/10/2019 06_Competing Risk Model

    22/23

    *hree competing risk models that could be used to analysedata arising in a power plant. ne is an e+isting opportunitymaintenance model that has been described from acompeting risk point of view.

    *he 7rst model presented here has independent censoring,and hence from a dataanalysis point of view is identi7able onthe basis of competing risk data M the times of failure or 1$events together with the label of the event type.

    *he second and third models illustrate dependent censoring.

    *he Aaplan $eier estimator M often recommended in reliabilityte+ts M would be inappropriate in both cases, and indeed isoveroptimistic in its assessment of equipment lifetime.

  • 8/10/2019 06_Competing Risk Model

    23/23

    *ank +ou

    ,uestions-