appliedstatistics - university of...

Weibull Reliability Analysis

=⇒ http://www.rt.cs.boeing.com/MEA/stat/scholz/http://www.rt.cs.boeing.com/MEA/stat/reliability.html

http://www.rt.cs.boeing.com/MEA/stat/scholz/weibull.html

Fritz Scholz (425-865-3623, 7L-22)Boeing Phantom WorksMathematics & Computing TechnologyApplied Statistics

Weibull Reliability Analysis—FWS-5/21-23/2002—1

Wallodi Weibull


Seminal Paper


The Weibull Distribution

•Weibull distribution, useful uncertainty model for

– wearout failure time Twhen governed by wearout of weakest subpart

– material strength Twhen governed by embedded flaws or weaknesses,

• It has often been found useful based on empirical data (e.g. Y2K)

• It is also theoretically founded on the weakest link principle

T = min (X1, . . . , Xn) ,

with X1, . . . , Xn statistically independent random strengths orfailure times of the n “links” comprising the whole.The Xi must have a natural finite lower endpoint,e.g., link strength ≥ 0 or subpart time to failure ≥ 0. ↪→


Theoretical Basis

• Under weak conditions Extreme Value Theory shows1 that forlarge n

P (T ≤ t) ≈ 1− exp−

t− τα

β for t ≥ τ, α > 0, β > 0

• The above approximation has very much the same spirit as theCentral Limit Theorem which under some weak conditions onthe Xi asserts that the distribution of T = X1 + . . .+Xn isapproximately bell-shaped normal or Gaussian for large n

• Assuming a Weibull model for T , material strength or cycle time tofailure, amounts to treating the above approximation as an equality

F (t) = P (T ≤ t) = 1− exp−

t− τα

β for t ≥ τ, α > 0, β > 0

1see: E. Castillo, Extreme Value Theory in Engineering, Academic Press, 1988S. Coles, An Introduction to Statistical Modeling of Extreme Values, Springer Verlag, 2001


Weibull Reproductive Property

If X1, . . . ,Xn are statistically independent

with Xi ∼ Weibull(αi, β) then

P (min(X1, . . . ,Xn) > t) = P (X1 > t)× · · · × P (Xn > t)

=n∏i=1

exp−

tαi

β = exp

− tα

β

with α = n∑i=1α−βi

−1/β

Hence T = min(X1, . . . ,Xn) ∼ Weibull(α,β)

This is similar to the normal reproductive property

X1, . . . ,Xn independent normal =⇒ n∑i=1Xi is normal


Weibull Parameters

The Weibull distribution may be controlled by 2 or 3 parameters:

• the threshold parameter τ

T ≥ τ with probability 1

τ = 0 =⇒ 2-parameter Weibull model.

• the characteristic life or scale parameter α > 0

P (T ≤ τ + α) = 1− exp−

αα

β = 1− exp(−1) = .632

regardless of the value β > 0

• the shape parameter β > 0, usually β ≥ 1


2-Parameter Weibull Model

•We focus on analysis using the 2-parameter Weibull model

•Methods and software tools much better developed

• Estimation of τ in the 3-parameter Weibull model

leads to complications

•When a 3-parameter Weibull model is assumed,

it will be stated explicitly


Relation of α & β to Statistical Parameters (τ = 0)

• The expectation or mean value of T

µ = E(T ) =∫ ∞0 t f(t) dt = αΓ(1 + 1/β)

with Γ(t) =∫ ∞0 exp(−x)xt−1 dx

• The variance of T

σ2 = E(T − µ)2 = ∫ ∞0 (t− µ)2f(t) dt = α2 [Γ(1 + 2/β)− Γ2(1 + 1/β)

]

• p-quantile tp of T , i.e., by definition P (T ≤ tp) = p

tp = α [− log(1− p)]1/β , for p = 1− exp(−1) = .632 =⇒ tp = α


Weibull Density

• The cumulative distribution function F (t) = P (T ≤ t) is justone way to describe the distribution of the random quantity T

• The density function f(t) is another representation (τ = 0)

f(t) = F ′(t) =dF (t)dt

=β

α

tα

β−1

exp−

tα

β t ≥ 0

P (t ≤ T ≤ t+ dt) ≈ f(t) dt

F (t) =∫ t0 f(x) dx


Weibull Density & Distribution Function

0 5000 10000 15000 20000

cycles

Weibull density α = 10000, β = 2.5total area under density = 1

cumulative distribution functionp

p

0

1


Weibull Densities: Effect of τ

cycles

prob

abili

ty d

ensi

ty

0 2000 4000

τ = 0 τ = 1000 τ = 2000

α = 1000, β = 2.5


Weibull Densities: Effect of α

cycles

prob

abili

ty d

ensi

ty

0 2000 4000 6000 8000

α = 1000

α = 2000

α = 3000

τ = 0, β = 2.5


Weibull Densities: Effect of β

cycles

prob

abili

ty d

ensi

ty

0 1000 2000 3000 4000

β = .5

β = 1β = 2

β = 4

β = 7

τ = 0, α = 1000


Failure Rate or Hazard Function

• A third representation of the Weibull distribution is through thehazard or failure rate function

λ(t) =f(t)

1− F (t) =β

α

tα

β−1

=d

dt[− log(1− F (t))]

• λ(t) is increasing t for β > 1 (wearout)• λ(t) is decreasing t for β < 1• λ(t) is constant for β = 1 (exponential distribution)

P (t ≤ T ≤ t+ dt|T ≥ t) = P (t ≤ T ≤ t+ dt)P (T ≥ t) ≈ f(t) dt

1− F (t) = λ(t) dt

F (t) = 1−exp(− ∫ t

0 λ(x) dx)

and f(t) = λ(t) exp(− ∫ t

0 λ(x) dx)


Exponential Distribution

• The exponential distribution is a special case: β = 1 & τ = 0

F (t) = P (T ≤ t) = 1− exp− tα

for t ≥ 0

• This distribution is useful when parts fail due torandom external influences and not due to wear out

• Characterized by the memoryless property,a part that has not failed by time t is as good as new,past stresses without failure are water under the bridge

• Good for describing lifetimes of electronic components,failures due to external voltage spikes or overloads


Unknown Parameters

• Typically will not know the Weibull distribution: α, β unknown

•Will only have sample data =⇒ estimates α̂, β̂get estimated Weibull model for failure time distribution=⇒ double uncertaintyuncertainty of failure time & uncertainty of estimated model

• Samples of failure times are sometimes very small,only 7 fuse pins or 8 ball bearings tested until failure,long lifetimes make destructive testing difficult

• Variability issues are often not sufficiently appreciatedhow do small sample sizes affect our confidence inestimates and predictions concerning future failure experiences?


Estimation Uncertainty

0 20000 40000 60000 80000

0.0

0.00

002

0.00

004

cycles

A Weibull Population: Histogram for N = 10,000 & Density

• ••• •••• •

63.2% 36.8%

characteristic life = 30,000shape = 2.5

true modelestimated model from 9 data points


Weibull Parameters & Sample Estimates

t = t p p-quantile

p=P(T < t )

αcharacteristic life = 30,000

63.2% 36.8%

shape parameter β = 4

•• ••• •• •• • •• ••• ••• ••• •• • •••• • •

25390 3.02 33860 4.27 29410 5.01

parameter estimates from three samples of size n = 10


Generation of Weibull Samples

• Using the quantile relationship tp = α [− log(1− p)]1/β

one can generate a Weibull random sample of size n by

– generating a random sample U1, . . . , Un

from a uniform [0, 1] distribution

– and computing Ti = α [− log(1− Ui)]1/β, i = 1, . . . , n.

– Then T1, . . . , Tn can be viewed as a random sample of size n

from a Weibull population or Weibull distribution

with parameters α & β.

• Simulations are useful in gaining insight on estimation procedures


Graphical Methods

• Suppose we have a complete Weibull sample of size n: T1, . . . , Tn

• Sort these values from lowest to highest: T(1) ≤ T(2) ≤ . . . ≤ T(n)• Recall that the p-quantile is tp = α [− log(1− p)]1/β

• Compute tp1 < . . . < tpn for pi = (i− .5)/n, i = 1, . . . , n

• Plot the points (T(i), tpi), i = 1, . . . , n

We expect these points to cluster around main diagonal


Weibull Quantile-Quantile Plot: Known Parameters

•••••

•• • ••••••••••• •••• • • • ••• ••••••

••••••••••• ••••• •••••••

•••••••••••••

••••• ••••••••••• ••••

••••• ••

••

T

tp

0 5000 10000 15000

050

0010

000

1500

0


Weibull QQ-Plot: Unknown Parameters

• Previous plot requires knowledge of the unknown parameters α & β

• Note thatlog (tp) = log(α) + wp/β , where wp = log [− log(1− p)]

• Expect points (log[T(i)], wpi

), i = 1, . . . , n, to cluster around line

with slope 1/β and intercept log(α)

• This suggests estimating α & β from a fitted line, by eye or leastsquares


Maximum Likelihood Estimation

• If t1, . . . , tn are the observed sample values one can contemplate

the probability of obtaining such a sample or of values nearby, i.e.,

P (T1 ∈ [t1 − dt/2, t1 + dt/2], . . . , Tn ∈ [tn − dt/2, tn + dt/2])= P (T1 ∈ [t1 − dt/2, t1 + dt/2])× · · · × P (Tn ∈ [tn − dt/2, tn + dt/2])

≈ fα,β(t1)dt× · · · × fα,β(tn)dt

where f(t) = fα,β(t) is the Weibull density with parameters (α, β)

•Maximum likelihood estimation maximizes this probability

over α & β =⇒ maximum likelihood estimates (m.l.e.s) α̂ and β̂

=⇒ the most likely Weibull model “explaining” the data


General Remarks on Estimation

•MLEs tend to be optimal in large samples (lots of theory)

•Method is very versatile in extending to may other data scenarios

censoring and covariates

• Least squares method applied to QQ-plot is not entirely appropriate

tends to be unduly affected by stray observations

not as versatile to extend to other situations


Weibull Plot: n = 20

cycles/hours

prob

abili

ty

•

••

•• ••

••••••••

• •••

•

10 20 50 100 200 500 1000

.001

.010

.100

.200

.500

.900

.990

.632

true model, Weibull( 100 , 3 )m.l.e. model, Weibull( 108 , 3.338 )least squares model, Weibull( 107.5 , 3.684 )


Weibull Plot: n = 100

cycles/hours

prob

abili

ty

•

••

••• • • •••

••••••••••••••••••••••

•••••••••••••••••••

•••••••••••••••••••••

•••••••••••••••••••••••

•• • •

10 20 50 100 200 500 1000

.001

.010

.100

.200

.500

.900

.990

.632

true model, Weibull( 100 , 3 )m.l.e. model, Weibull( 102.8 , 2.981 )least squares model, Weibull( 102.2 , 3.144 )


Tests of Fit (Graphical)

• The Weibull plots provide an informal diagnostic

for checking the Weibull model assumption

• The anticipated linearity is based on the Weibull model properties

• Strong nonlinearity indicates that the model is not Weibull

• Sorting out nonlinearity from normal statistical point scatter

takes a lot of practice and a good sense for the effect

of sample size on the variation in point scatter (simulation helps)

• Formal tests of fit are available for complete samples2

and also for some other censored data scenarios2R.B. D’Agostino and M.A. Stephens, Goodness-of-Fit Techniques, Marcel Dekker 1986


Formal Goodness-of-Fit Tests

• Let Fα̂,β̂(t) be the fitted Weibull distribution function

• Let F̂n(t) = #{Ti ≤ t; i=1,...,n}n be the empirical distribution function

• Compute a discrepancy metric D between Fα̂,β̂ and F̂n,

DKS(Fα̂,β̂, F̂n) = supt

∣∣∣∣∣Fα̂,β̂(t)− F̂n(t)∣∣∣∣∣ Kolmogorov-Smirnov

DCvM(Fα̂,β̂, F̂n) =∫ ∞0

(Fα̂,β̂(t)− F̂n(t)

)2fα̂,β̂(t) dt Cramer-von Mises

DAD(Fα̂,β̂, F̂n) =∫ ∞0

(Fα̂,β̂(t)− F̂n(t)

)2Fα̂,β̂(t)(1− Fα̂,β̂(t))

fα̂,β̂(t) dt Anderson-Darling

• The distributions of D, when sampling from a Weibull population,

are known and p-values of observed values d of D can be calculated

p = P (D ≥ d) =⇒ BCSLIB: HSPFIT


Kolmogorov-Smirnov Distance

0 50 100 150 200 250

0.0

0.2

0.4

0.6

0.8

1.0

K-S distance

••• •• • •• • •

n = 10


Weibull Plots: n = 10 (solid line is true model)

•

•

•

••

•••

•

•

3.8 4.2 4.6 5.0

-3-2

-10

1 p(KS) = 0.62

p(CvM) = 0.54

p(AD) = 0.62

n = 10 •

•

•

••

•••

•

•

4.0 4.4 4.8-3

-2-1

01 p(KS) = 0.28

p(CvM) = 0.27

p(AD) = 0.35

•

•

•

••

•••

•

•

3.8 4.4 5.0

-3-2

-10

1 p(KS) = 0.88

p(CvM) = 0.62

p(AD) = 0.66

•

•

•

••••

••

•

3.8 4.2 4.6 5.0

-3-2

-10

1 p(KS) = 0.88

p(CvM) = 0.64

p(AD) = 0.69

•

•

•

•••

•••

•

4.0 4.4 4.8

-3-2

-10

1 p(KS) = 0.72

p(CvM) = 0.53

p(AD) = 0.6

•

•

•

••

••••

•

3.8 4.2 4.6

-3-2

-10

1 p(KS) = 0.16

p(CvM) = 0.21

p(AD) = 0.32

•

•

•

••

••••

•

4.0 4.4 4.8

-3-2

-10

1 p(KS) = 0.34

p(CvM) = 0.23

p(AD) = 0.22

•

•

•

•••

••

•

•

3.8 4.2 4.6

-3-2

-10

1 p(KS) = 0.86

p(CvM) = 0.56

p(AD) = 0.57


Weibull Plots: n = 20

•

•

••

•••••••

•••••••

••

3.5 4.0 4.5 5.0

-3-2

-10

1

p(KS) = 0.55

p(CvM) = 0.49

p(AD) = 0.49

n = 20 •

•

••

••

••••••••••••••

2.5 3.5 4.5-3

-2-1

01

p(KS) = 0.13

p(CvM) = 0.027

p(AD) = 0.028

•

•

••

••••••

••••

• •••

••

4.2 4.6

-3-2

-10

1

p(KS) = 0.12

p(CvM) = 0.053

p(AD) = 0.048

•

•

••

••••••

• • ••••••

••

3.8 4.2 4.6 5.0

-3-2

-10

1

p(KS) = 0.46

p(CvM) = 0.39

p(AD) = 0.38

•

•

••

••••

••••••••••••

3.6 4.0 4.4 4.8

-3-2

-10

1

p(KS) = 0.46

p(CvM) = 0.41

p(AD) = 0.41

•

•

••••••• ••

••••••

•••

3.0 4.0 5.0

-3-2

-10

1

p(KS) = 0.47

p(CvM) = 0.63

p(AD) = 0.69

•

•

••••••

•••••••

•••••

4.0 4.4 4.8

-3-2

-10

1

p(KS) = 0.85

p(CvM) = 0.6

p(AD) = 0.64

•

•

•••

•••••••••••• •

••

3.6 4.0 4.4 4.8

-3-2

-10

1

p(KS) = 0.04

p(CvM) = 0.017

p(AD) = 0.023



•

••••••••••

•••••••••••••••••••••••••••••••••••••••

2.0 3.0 4.0 5.0

-4-3

-2-1

01

p(KS) = 0.7

p(CvM) = 0.41

p(AD) = 0.43

n = 50 •

•••• • •••

•••••••• ••••

••••••••••••••

•••••••••••••••

3.0 4.0 5.0-4

-3-2

-10

1

p(KS) = 0.22

p(CvM) = 0.35

p(AD) = 0.32

•

•••• ••

•••••

• •••••••

••••••••••••• •••••

••••••••••••

3.5 4.0 4.5

-4-3

-2-1

01

p(KS) = 0.15

p(CvM) = 0.23

p(AD) = 0.12

•

••

•••••••••••••••

••••••••••••••••••••••

•••••••••

•

3.5 4.5

-4-3

-2-1

01

p(KS) = 0.76

p(CvM) = 0.49

p(AD) = 0.52

•

•••• ••

• ••••••

••••••••

••••••••••••••

••••••••

•••• •

•

3.5 4.0 4.5

-4-3

-2-1

01

p(KS) = 0.88

p(CvM) = 0.66

p(AD) = 0.72

•

••

••••••••••

••••••••••••••••••••••••••

•••••••

••••

3.5 4.0 4.5 5.0

-4-3

-2-1

01

p(KS) = 0.27

p(CvM) = 0.22

p(AD) = 0.23

•

••

••••••••

•••••••

••••••••••••••••••••••••••

•••••

•

3.5 4.0 4.5 5.0

-4-3

-2-1

01

p(KS) = 0.85

p(CvM) = 0.56

p(AD) = 0.62

•

•••

• ••••••••••••••••••••••••••••

•••••• ••••

•••••

••

3.5 4.0 4.5 5.0

-4-3

-2-1

01

p(KS) = 0.18

p(CvM) = 0.035

p(AD) = 0.039



•

••• • ••

•••••••••••••••••••••••••••••••••••••••••••••••••••••••••••••••••••••••••••••••••••

••••••••• •

3.0 4.0 5.0

-4-2

0

p(KS) = 0.58

p(CvM) = 0.26

p(AD) = 0.3

n = 100 •

••• ••

••••••••••••••••••••••••••••••••••••••••••••••••••••••••••••••••

••••••••••••••••••••••••

••••• •

3.5 4.5-4

-20

p(KS) = 0.86

p(CvM) = 0.68

p(AD) = 0.72

•

••• ••

••••••••••••••••••••

••••••••••••••••••••••••••••••••••••••••••••••••••••••

••••••••••••••••••

••

3.0 4.0 5.0

-4-2

0

p(KS) = 0.61

p(CvM) = 0.48

p(AD) = 0.55

•

••••

••••••••••••••••••••••••••••••••••••••••••••••••••••••••••••••••••••

•••••••••••••••••••••••••••

3.5 4.5

-4-2

0

p(KS) = 0.21

p(CvM) = 0.26

p(AD) = 0.43

•

•••••••••

••••••••••••••••••

•••••••••••••••••••••••••••••••••••••••••••••••••••••••••••••••••••

•••• •

3.0 4.0 5.0

-4-2

0

p(KS) = 0.47

p(CvM) = 0.3

p(AD) = 0.37

•

••

••••••••

••••••••••••••••••••••••••••••

•••••••••••••••••••••••••••••••••••••••••••••••••••••••••••

3.0 4.0 5.0

-4-2

0

p(KS) = 0.053

p(CvM) = 0.025

p(AD) = 0.032

•

•••••••

•••••••

••••••••••••••••••

•••••••••••••••••

••••••••••••••••••••••••••••••••••••••

••••••••••

••

3.5 4.5

-4-2

0

p(KS) = 0.87

p(CvM) = 0.58

p(AD) = 0.6

•

••• ••

• •••••••••••••••••••••••••••••••••••••••••••••••••••••••••••••••••••••••••••••••••••••••••••••

2.5 3.5 4.5

-4-2

0

p(KS) = 0.72

p(CvM) = 0.45

p(AD) = 0.43


Estimates and Confidence Bounds/Intervals

• For a target θ

θ = α, θ = β, θ = tp = α[− log(1− p)]1/β, or θ = Pα,β(T ≤ t)one gets corresponding m.l.e θ̂ by replacing (α, β) by (α̂, β̂)

• Such estimates θ̂ vary around the target θ due to sampling variation

• Capture the estimation uncertainty via confidence bounds

• For 0 < γ < 1 get 100γ% lower/upper confidence bounds θ̂L,γ & θ̂U,γ

P(θ̂L,γ ≤ θ

)= γ or P

(θ ≤ θ̂U,γ

)= γ

• For γ > .5 get a 100(2γ − 1)% confidence interval by [θ̂L,γ, θ̂U,γ]

P(θ̂L,γ ≤ θ ≤ θ̂U,γ

)= 2γ − 1 = γ%


Confidence Bounds & Sampling Variation, n = 10

90%

con

fiden

ce in

terv

als

1 2 3 4 5

2500

035

000

samples

L: Lawless Exact Method (RAP); B: Bain Exact Method (Tables); M: MLE Approximate Method

L B M

true α

90%

con

fiden

ce in

terv

als

1 2 3 4 5

02

46

8

L B M

true β


Confidence Bounds & Sampling Variation, n = 10

95%

low

er c

onfid

ence

bou

nds

1 2 3 4 5

050

0015

000

2500

0

samples

L: Lawless Exact Method (RAP); B: Bain Exact Method (Tables); M: MLE Approximate Method

L B Mtrue 10-percentile

95%

upp

er c

onfid

ence

bou

nds

1 2 3 4 5

0.0

0.10

0.20

L B M

true P(T < 10,000)


Confidence Bounds & Effect of n

• ••

• • • • •

95%

con

fiden

ce in

terv

als

1500

025

000

3500

0

•

•

•

• • • • •

•

• • •• • • •

5 10 20 50 100 200 500 1000sample size

true α

••

• • • • • •

95%

con

fiden

ce in

terv

als

12

34

56

7

•

••

• • • • •

•

•

••

• • • •

true β


Confidence Bounds & Effect of n

• • • • • • • •

95%

upp

er c

onfid

ence

bou

nds

0.0

0.05

0.15

sample size

•

•

•

• • • • •

5 10 20 50 100 200 500 1000

true P(T < 10000)

• • •• • • • •

95%

low

er c

onfid

ence

bou

nds

5000

1000

020

000

•

•

•

• •• • •

true t .10


Incomplete Data or Censored Failure Times

• Type I censoring or time censoring:units are tested until failure or until a prespecified time has elapsed

• Type II censoring or failure censoring:only the r lowest values of the total sample of size n become knownthis shows up when we put n units on test at the same timeand terminate the test after the first r units have failed

• Interval censoring, inspection data, grouped data:ith unit is only known to have failed between two knowninspection time points, i.e., failure time Ti falls in (si, ei]only bracketing intervals (si, ei), i = 1, . . . , n, become known


Incomplete Data or Censored Failure Times (continued)

•Random right censoring:units are observed until failed or removed from observationdue to other causes (different failure modes, competing risks)

•Multiple right censoring:units are put into service at different timesand times to failure or censoring are observed

• Data can also combine several of the above censoring phenomena

• It is important that the censoring mechanism should not correlatewith the (potential) failure times,i.e., no censoring of anticipated failures

•All data (censored & uncensored) should enter analysisotherwise bias results


Type I or Time Censored Data: n = 10

cycles

0 2000 4000 6000 8000 10000

1

2

3

4

5

6

7

8

9

10unit

?

?

X

X

?

X

X

?

?

X


Type II or Failure Censored Data: n = 10

cycles

0 2000 4000 6000 8000 10000

1

2

3

4

5

6

7

8

9

10unit

X

X

X

X

X

X

?

?

?

?


Interval Censored Data: n = 10

cycles

0 2000 4000 6000 8000 10000

1

2

3

4

5

6

7

8

9

10unit

•( ]

?

•( ?

•( ]

?

•( ]

•( ?

?

•( ]

?


Multiply Right Censored Data: n = 10

cycles

0 2000 4000 6000 8000 10000

1

2

3

4

5

6

7

8

9

10unit

X

?

X

X

X

X

?

?

?

?


Nonparametric MLEs of CDF

• For complete, type I & II censored data the nonparametric MLE is

F̂ (t) =#{Ti ≤ t; i = 1, . . . , n}

n,

−∞ < t <∞ for complete sample,−∞ < t ≤ t0 for type I censored (at t0) sample,−∞ < t ≤ T(r) for type II censored (at T(r), rth smallest failure time)

• For multiply right censored data with failures at t%1 < . . . < t%k the

nonparametric MLE (Kaplan-Meier or Product Limit Estimator) is

F̂ (t) = 1−[(1− p̂1)δ1(t) · · · (1− p̂k)δk(t)

], δi(t) =

1 for t%i ≤ t0 for t%i > t

where p̂i = di/ni, ni = # units known to be at risk just prior to t%iand di = # units that failed at t%i

• The nonparametric MLE for interval censored data is complicated


Empirical CDF (complete samples)

0 5000 15000 25000

0.0

0.4

0.8

n = 10

empirical and trueand Weibull mledistribution functions

0 5000 15000 25000

0.0

0.4

0.8

n = 30

0 5000 15000 25000

0.0

0.4

0.8

n = 50

0 5000 15000 25000

0.0

0.4

0.8

n = 100


Nonparametric MLE for Type I Censored Data

0 5000 15000 25000

0.0

0.4

0.8

...n = 10

nonparametric mle of CDFWeibull mle of CDFand true CDF

0 5000 15000 25000

0.0

0.4

0.8

... ... .n = 30

0 5000 15000 25000

0.0

0.4

0.8

.. . .... ..n = 50

0 5000 15000 25000

0.0

0.4

0.8

. ... . ... . ... .. .. . .n = 100


Nonparametric MLE for Type II Censored Data

0 5000 15000 25000

0.0

0.4

0.8

. . .n = 10

nonparametric mle of CDFWeibull mle of CDFand true CDF

0 5000 15000 25000

0.0

0.4

0.8

... . ... .n = 30

0 5000 15000 25000

0.0

0.4

0.8

..... ... .. . .. ..n = 50

0 5000 15000 25000

0.0

0.4

0.8

................. ....... . .n = 100


Kaplan-Meier Estimates (multiply right censored samples)

0 5000 15000 25000

0.0

0.4

0.8

.. .. ....n = 10

Kaplan-Meier CDFWeibull mle for CDFtrue CDF

0 5000 15000 25000

0.0

0.4

0.8

... . .. .. .. .. . .. .. .n = 30

0 5000 15000 25000

0.0

0.4

0.8

... ... .. .. .. .. ... .. . .. .. ... .n = 50

0 5000 15000 25000

0.0

0.4

0.8

.. ... .. .. ... .. .. . ... ..... .. ...... .. .. .. ... ... .. .... ... .. .. .n = 100


Nonparametric MLE for Interval Censored Data

cycles

CD

F

0 5000 10000 15000

0.0

0.2

0.4

0.6

0.8

1.0

superimposed is the true Weibull(10000,2.5)

that generated the 1,000 interval censored data cases

inspection points roughly 3000 cycles apart


Nonparametric MLE for Interval Censored Data

cycles

CD

F

0 5000 10000 15000

0.0

0.2

0.4

0.6

0.8

1.0

superimposed is the true Weibull(10000,2.5)

that generated the 10,000 interval censored data cases

inspection intervals randomly generated from same Weibull


Application to Y2K Questionnaire Return Data

days

prob

abili

ty o

f que

stio

nnai

re r

etur

n

0 50 100 150 200 250 300

0.0

0.2

0.4

0.6

0.8

1.0

estimated .99-quantile

8124 cases


Plotting Positions for Weibull Plots (Censored Case)

• For complete samples we plotted (log[T(i)], log [− log(1− pi)]) with

pi =i− .5n

=12[F̂ (T(i)) + F̂ (T(i−1))

]=

12

in+i− 1n

• For type I & II & multiply right censored samples

we plot in analogy (log[T %(i)], log [− log(1− pi)]), whereT %(1) < . . . < T

%(k) are the k distinct observed failure times and

pi =12

[F̂ (T %(i)) + F̂ (T

%(i−1))

]

and F̂ (t) is the nonparametric MLE of F (t) for the censored sample


Weibull Plot, Multiply Right Censored Sample n = 50

cycles/hours

prob

abili

ty

•

••

• •••••••• •••

•••••••• ••

•••

10 20 50 100 200 500 1000

.001

.010

.100

.200

.500

.900

.990

.632

true model, Weibull( 100 , 3 )m.l.e. model, Weibull( 96.27 , 2.641 ) n = 50least squares model, Weibull( 96.3 , 2.541 ) n = 50


Software: MLE’s & Confidence Bounds for Censored Data

• For complete & type II censored data RAP computes exact coverageconfidence bounds for α, β, tp, and P (T ≤ t).

•WEIBREG and commercial software (Weibull++, WeibullSMITH,common statistical packages) compute approximate confidencebounds for above targets, based on large sample m.l.e. theory

• RAP & WEIBREG are Boeing code, runs within DOS modeof Windows95 (interface clumsy, but job gets done), contact me

• Boeing has a site license for Weibull++ Version 4in the Puget Sound area (with upgrade pressure)contact David Twigg (425) 266-7919, [email protected]

• There is a Weibull++ Version 6 out


Other Software

• S-Plus Statistical data analysis environment (≈ $2000 for PC licence)• R is a GNU version of S-plus, freely availableeither from: http://cran.r-project.org/ for lots of operatingsystems or inside Boeing for Windows

http://starfly.ca.boeing.com/KIRTSweb/oss_toolkit/toolkit/optional.html#L048

• Both S-Plus and R have Terry Therneau’s survival analysis packagebuilt in or as add-on. Using it takes some sophistication.

•My web-based Weibull analysis tool is based on R

http://www.rt.cs.boeing.com/MEA/stat/scholz/weibull.html

• Also freely available is the very user friendly package SPLIDA by BillMeeker which can be added to S-Plus 4.5 or S-Plus 2000

http://www.public.iastate.edu/~splida/


Software in Progress (Needs Pull!)

• The current version of web-based Weibull analysis tool and alsomost other versions based on large sample approximations for theMLE’s have some defect for extreme sample with very few failures:

Confidence bounds for quantiles tp are not always monotoneincreasing in p, which does not make sense. Also quantile andprobability bounds are not inverses to each other.

• I have an updated version of WEIBREG based on a new methodthat avoids this problem.

It also includes a bootstrap version of the new monotone method.However, it is not yet released.

Quantile and probability bounds are true inverses of each other.

• The plan is to update the web based tool and to also create webbased versions of RAP and ABVAL to bypass the legacy issue.


Weibull Plot With Confidence Bound

1 2 3 5 10 20 50 100 200 500 1000

0.001

0.010

0.100

0.200

0.500

0.900

0.990

0.999

•

••

••••••••••

alpha = 102.4 [ 82.52 , 143.8 ] 95 % beta = 2.524 [ 1.352 , 3.696 ] 95 %n = 30 , r = 13


Weibull Plot With Confidence Bound (Variation)

1 2 3 5 10 20 50 100 200 500 1000

0.001

0.010

0.100

0.200

0.500

0.900

0.990

0.999

•

•••

•••••••••

alpha = 89.79 [ 77.4 , 109.4 ] 95 % beta = 3.674 [ 2.15 , 5.198 ] 95 %n = 30 , r = 13


cycles x 1000

prob

abili

ty

1 2 5 10 20 50 100

.001

.010

.100

.500

.900

.999

.632

m.l.e.95 % q−confidence bounds95 % p−confidence bounds

Problem Weibull Plot

α̂ = 38600 , 95 % conf. interval ( 30690 , 48560 )

β̂ = 6.088 , 95 % conf. interval ( 1.987 , 18.65 )

n = 5 , r = 2

Data: 15900+, 20000+, 25500, 36000+, 40010


Weibull Regression Model & Analysis

• Recall tp = α[− log(1− p)]1/β orlog(tp) = log(α) + wp/β with wp = log[− log(1− p)]

• Often we deal with failure data collected under different conditions

– different part types– different environmental conditions– different part users

• Linear regression model for log(α) (multiplicative on α)

log(αi) = b1Zi1 + . . .+ bpZip typically Zi1 ≡ 1

where Zi1, . . . , Zip are known covariates for ith unit

• Now we have p+ 1 unknown parameters β, b1, . . . , bp

• parameter estimates & confidence bounds using MLEs


Accelerated Life Testing: Inverse Power Law

• Units last too long under normal usage conditions

• Increase “stress” to accelerate failure time

• Increase voltage and accelerate life via inverse power law

T (Volt) = T (VoltU) VoltVoltU

c

, where usually c < 0

this meansα(Volt) = α (VoltU)

VoltVoltU

c

or log [α(Volt)] = log [α (VoltU)] + c log (Volt)− c log (VoltU)

= b1 + b2Z

with

b1 = log [α (VoltU)]− c log (VoltU) , b2 = c, and Z = log (Volt)


Accelerated Life Testing: Arrhenius Model

• Another way of accelerating failure in processes involvingchemical reaction rates is to increase the temperature

• Arrhenius proposed the following acceleration model withtemperature measured in Kelvin (tempK = temp◦C + 273.15)

T (temp) = T (tempU) exp κtemp

− κ

tempU

or

log [α(temp)] = log [α(tempU)] +κ

temp− κ

tempU

= b1 + b2Z

with

b1 = log [α(tempU)]−κ

tempU, b2 = κ, and Z = temp−1


Pooling Data With Different α’s

• Suppose we have three groups of failure dataT1, . . . , Tn1 ∼ W(α̃1, β), Tn1+1, . . . , Tn1+n2 ∼ W(α̃2, β),

Tn1+n2+1, . . . , Tn1+n2+n3 ∼ W(α̃3, β)

•We can analyze the whole data set of N = n1 + n2 + n3 values jointly

using the following model for αj and dummy covariates Z1,j, Z2,j & Z3,j

log(αj) = b1Z1,j + b2Z2,j + b3Z3,j , j = 1, 2, . . . , N

where b1 = log(α̃1), b2 = log(α̃2)−log(α̃1) and b3 = log(α̃3)−log(α̃1)Z1,j = 1 for j = 1, . . . , N , Z2,j = 1 for j = n1 + 1, . . . , n1 + n2, Z2,j = 0 else

and Z3,j = 1 for j = n1 + n2 + 1, . . . , N & Z3,j = 0 else.

• Advantage: Smaller estimation errorall data are used in estimating the (assumed) common β


Pooling Data (continued)

log(α1) = b1 · 1 + b2 · 0 + b3 · 0 = log(α̃1)... ...

log(αn1) = b1 · 1 + b2 · 0 + b3 · 0 = log(α̃1)... ...

log(αn1+1) = b1 · 1 + b2 · 1 + b3 · 0 = log(α̃1) + [log(α̃2)− log(α̃1)] = log(α̃2)... ...

log(αn1+n2) = b1 · 1 + b2 · 1 + b3 · 0 = log(α̃1) + [log(α̃2)− log(α̃1)] = log(α̃2)... ...

log(αn1+n2+1) = b1 · 1 + b2 · 0 + b3 · 1 = log(α̃1) + [log(α̃3)− log(α̃1)] = log(α̃3)... ...

log(αN) = b1 · 1 + b2 · 0 + b3 · 1 = log(α̃1) + [log(α̃3)− log(α̃1)] = log(α̃3)


Accelerated Life Testing: Model & Data

Voltage

thou

sand

cyc

les

•

•••••

• •••••

••

50 100 150 200 250 300 350 400 450 500

1

2

5

10

20

50

100

200


Accelerated Life Testing: Model, Data & MLEs

Voltage

thou

sand

cyc

les

•

•••••

• •••••

••

mle line

true line

mle for .01-quantile

95% lower bound for .01-quantile

50 100 150 200 250 300 350 400 450 500

1

2

5

10

20

50

100

200


Analysis for Data from Exponential Distribution

• T ∼ E(θ) Exponential with mean θ, i.e., E(θ) = W(α = θ, β = 1)

• It suffices to get confidence bounds for θ since all other quantities

of interest are explicit and monotone functions of θ

Fθ(t0) = Pθ(T ≤ t0) = 1− exp(−t0/θ)

Rθ(t0) = Pθ(T > t0) = exp(−t0/θ)and

tp(θ) = θ [− log(1− p)]


Complete & Type II Censored Exponential Samples

• T ∼ E(θ) Exponential with mean θ, E(θ) = W(α = θ, β = 1)

• For complete exponential samples T1, . . . , Tn ∼ E(θ)or for a type II censored sample of this type,

i.e., with observed r lowest values T(1) ≤ . . . ≤ T(r), 1 ≤ r ≤ n,one gets 100γ% lower confidence bounds for θ by computing

θ̂L,γ =2× TTTχ22r,γ

,

where TTT = T(1) + · · ·+ T(r) + (n− r)T(r) = Total Time on Test,

and χ22r,γ = γ-quantile of the χ22r distribution

get this in Excel via = GAMMAINV(γ, r, 1) or = CHIINV(1− γ, 2r)/2• These bounds have exact confidence coverage properties


Multiply Right Censored Exponential Data

• Here we define TTT = T1 + . . .+ Tn as the total time on test

but now Ti is either the observed failure time of the ith part

or the observed right censoring time of the ith part

The number r of observed failures is random here, 0 ≤ r ≤ n

• Approximate 100γ% lower confidence bounds for θ are computed as

θ̂L,γ =2× TTTχ22r+2,γ

• This also holds for r = 0, i.e., no failures at all over exposure period


Weibull Shortcut Methods for Known β

• T ∼ W(α, β) (Weibull) ⇒ Tβ ∼ E(θ) Exponential with mean θ = αβ

• As 100γ% lower confidence bound for α compute

α̂L,γ =

2× TTT (β)χ2f,γ

1/β

where for complete or type II censored samples we take f = 2r and

TTT = Tβ(1) + · · ·+ Tβ(r) + (n− r)Tβ(r)and for multiply right censored data we take f = 2r + 2 and

TTT = Tβ1 + · · ·+ Tβn .

• Sensitivity should be studied by repeating analyses for several β’s


A- and B-Allowables, ABVAL Program

•My first Weibull work led to the program ABVAL (1983)supporting Cecil Parsons and Ron Zabora w.r.t. MIL-HDBK-5

• ABVAL computes A- and B-Allowables based on samplesfrom the 3-parameter Weibull distribution.

– The A-Allowable is a 95% lower confidence boundfor 99% of the population, i.e., for the .01-quantile.

– The B-Allowable is a 95% lower confidence boundfor 90% of the population, i.e., for the .10-quantile.

• Hybrid approach of using Mann-Fertig threshold estimatecombined with maximum likelihood estimates for α and β,took lot of simulation and tuning and is very specialized in itscapabilities

• It is now a method in MIL-HDBK-5, I still maintain software


Selected References

• Abernethy, R.B. (1996), The New Weibull Handbook, 2nd edition,Reliability Analysis Center 800-526-4802

• Bain, L.J. (1978), Statistical Analysis of Reliability and Life-TestingModels, Marcel Dekker, New York

• Jeng, S.-L. and Meeker, W.Q. (2000), “Comparisons of approximateconfidence procedures for type I censored data,” J. Amer. Statist.Assoc. 42, 135-148.

• Kececioglu, D. (1993/4), Reliability & Life Testing Handbook Vol. 1& 2, Prentice Hall PTR, Upper Saddle River, NJ.

• Lawless, J.F. (1982), Statistical Models and Methods for LifetimeData, John Wiley & Sons, New York

•Mann, N.R., Schafer, R.E., and Singpurwalla, N.D. (1974), Methodsfor Statistical Analysis of Reliability and Life Data, John Wiley &Sons, New York


References (Continued)

•Meeker, W.Q. and Escobar, L.A. (1995), “Teaching aboutapproximate confidence regions based on maximum likelihoodestimation,” The American Statistician 49 48-53.

=⇒Meeker, W.Q. and Escobar, L.A. (1998), Statistical Methods forReliability Data, John Wiley & Sons, New York

•Meeker, W.Q. and Hahn, G.J. (1985), Volume 10: How to Plan anAccelerated Life Test—Some Practical Guidelines, American Societyfor Quality Control

• Robinson, J.A. (1983), “Bootstrap confidence intervals inlocation-scale models with progressive censoring,” Technometrics 25179-187.

• Nelson, W. (1982), Applied Life Data Analysis, John Wiley & Sons,New York


References (Continued)

• Nelson, W. (1985), “Weibull analysis of reliability data with few orno failures,” Journal of Quality Technology 17, 140-146

• Nelson, W. (1990), Accelerated Testing, Statistical Models, TestPlans and Data Analyses, John Wiley & Sons, New York

• Therneau, T.M. and Grambsch, P.M. (2000), Modeling SurvivalData, Extending the Cox Model, Springer Verlag, New York.

• Scholz, F.W. (1994), “Weibull and Gumbel distribution exactconfidence bounds.” BCSTECH-94-019 (RAP)

• Scholz, F.W. (1996) “Maximum likelihood estimation for type Icensored Weibull data including covariates.” ISSTECH-96-022

• Scholz, F.W. (1997), “Confidence bounds for type I censored Weibulldata including covariates.” SSGTECH-97-025 (WEIBREG)The Boeing Company, P.O. Box 3707, MS-7L-22, Seattle WA 98124-2207


appliedstatistics - university of...

Documents