Weibull Reliability Analysis: Boeing Phantom Works

Weibull Reliability Analysis
=⇒ http://www.rt.cs.boeing.com/MEA/stat/reliability.html
Fritz Scholz (425-865-3623, 7L-22)

Boeing Phantom Works
Mathematics & Computing Technology
Weibull Reliability Analysis—FWS-5/1999—1

Wallodi Weibull

Seminal Paper

The Weibull Distribution
• Weibull distribution, useful uncertainty model for

– wearout failure time T
when governed by wearout of weakest subpart
– material strength T
when 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 or
failure 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 for

large n
  β 
t−τ   
P (T ≤ t) ≈ 1 − exp − for t ≥ τ, α > 0, β > 0



  
α
• The above approximation has very much the same spirit as the
Central Limit Theorem which under some weak conditions on
the Xi asserts that the distribution of T = X1 + . . . + Xn is
approximately bell-shaped normal or Gaussian
• Assuming a Weibull model for T , material strength or cycle time to
failure, amounts to treating the above approximation as an equality
  β 
 t−τ  
F (t) = P (T ≤ t) = 1 − exp −    for t ≥ τ, α > 0, β > 0
α
1
see: E. Castillo, Extreme Value Theory in Engineering, Academic Press, 1988

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 t β n  t β 
Y    X
= exp −    = exp −   
 
i=1 αi i=1 αi
  β 
 t  
= exp −  
 

α
Hence T = min(X1, . . . , Xn) ∼ Weibull(α, β) with

 
n −1/β
α=  X
 αi−β 
i=1
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

   
α β
P (T ≤ τ + α) = 1 − exp −
    = 1 − exp(−1) = .632

α
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
• The expectation or mean value of T

Z
∞
µ = E(T ) = 0 t f (t) dt = αΓ(1 + 1/β)
Z
∞
with Γ(t) = 0 exp(−x) xt−1 dx
• The variance of T
Z
∞
σ = E(T − µ) =
2 2
0 (t − µ) f (t) dt = α Γ(1 + 2/β) − Γ (1 + 1/β)
2 2 2
• 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 just

one way to describe the distribution of the random quantity T
• The density function f (t) is another representation (τ = 0)

 β−1   β 
dF (t) β  t  t
f (t) = F 0(t) =
  
=   exp −    t≥0
dt α α α
P (t ≤ T ≤ t + dt) ≈ f (t) dt
Z
t
F (t) = 0 f (x) dx

Weibull Density & Distribution Function
Weibull density α = 10000, β = 2.5

total area under density = 1
p
cumulative distribution function
0
0 5000 10000 15000 20000

cycles

Weibull Densities: Effect of τ
τ=0 τ = 1000 τ = 2000
α = 1000, β = 2.5
probability density
0 2000 4000
cycles

Weibull Densities: Effect of α
α = 1000
τ = 0, β = 2.5
probability density
α = 2000
α = 3000
0 2000 4000 6000 8000
cycles

Weibull Densities: Effect of β
β=7
β = .5
probability density
β=4
β=1
β=2
τ = 0, α = 1000
0 1000 2000 3000 4000
cycles

Failure Rate or Hazard Function
• A third representation of the Weibull distribution is through the

hazard or failure rate function
 
f (t) β  t β−1
λ(t) = =  
1 − F (t) α α
• 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) f (t) dt
P (t ≤ T ≤ t + dt|T ≥ t) = ≈ = λ(t) dt
P (T ≥ t) 1 − F (t)
Z ! Z !
t t
F (t) = 1−exp − 0 λ(x) dx and f (t) = λ(t) exp − 0 λ(x) dx

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

 
t
F (t) = P (T ≤ t) = 1 − exp − 

 for t ≥ 0
α
• This distribution is useful when parts fail due to

random 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 α, c βc
get estimated Weibull model for failure time distribution
=⇒ double uncertainty
uncertainty 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 appreciated
how do small sample sizes affect our confidence in
estimates and predictions concerning future failure experiences?

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

0.00004
63.2% 36.8%
characteristic life = 30,000
shape = 2.5
0.00002
true model
estimated model from 9 data points
0.0
• • • •• • • • •
0 20000 40000 60000 80000
cycles

Weibull Parameters & Sample Estimates
shape parameter β = 4
63.2% 36.8%
characteristic life α = 30,000
p=P(T < t )
t =t p p-quantile
• •• • •• • • • •
• • • • • •• • ••
•• • ••••• • •
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 and expect
these points to cluster around main diagonal

Weibull Quantile-Quantile Plot: Known Parameters
•
•
15000 •
•
• •
•• ••
••
• ••
•
••
•• •••
10000
••
•••••
•
•
• ••
tp
• ••
••• • ••••
•• ••••••
•• •
•• •••
5000
•••
• • ••
•••
••
•
0
0 5000 10000 15000

Weibull QQ-Plot: Unknown Parameters
• Previous plot requires knowledge of the unknown parameters α & β
• Note that
log (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 least squares line

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

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

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
.990
•
.900 ••
• ••
.632
• ••
•
.500
• ••
•
.200 ••
probability
•
•
.100
•
•
.010 true model, Weibull( 100 , 3 )
m.l.e. model, Weibull( 108 , 3.338 )
least squares model, Weibull( 107.5 , 3.684 )
.001
10 20 50 100 200 500 1000
cycles/hours

Weibull Plot: n = 100
.990
• • •
•••
.900 •••
•
•
•••••
.632
•••
.500 • ••••••
•
••••••
••••
.200 •• •
probability
•
.100 ••••
• ••
•
•
•
•
.010 true model, Weibull( 100 , 3 )
m.l.e. model, Weibull( 102.8 , 2.981 )
•
least squares model, Weibull( 102.2 , 3.144 )
.001
10 20 50 100 200 500 1000
cycles/hours

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
• Formal tests of fit are available for complete samples2

and also for some other censored data scenarios
2
R.B. D’Agostino and M.A. Stephens, Goodness-of-Fit Techniques, Marcel Dekker 1986

Formal Goodness-of-Fit Tests
• Let Fα,
b βb (t) be the fitted Weibull distribution function
#{Ti ≤ t; i=1,...,n}
• Let F
d
n (t) = n be the empirical distribution function
• Compute a discrepancy metric D between Fα,
b β
d
b and Fn ,

b , Fn ) = sup F b b (t) − Fn (t)
d d
DKS (Fα,
b β α,β

Kolmogorov-Smirnov
t
Z !
∞ 2
b (t) − Fn (t)
d d
DCvM (Fα,
b βb , Fn ) =
0 Fα,
b β fα,
b βb (t) dt Cramer-von Mises
!
2
b (t) − Fn (t)
d
d
Z
∞ Fα,b β
DAD (Fα,
b βb , Fn ) = fα,
b βb (t) dt Anderson-Darling
0
Fα,
b βb (t)(1 − F b
b β
α, (t))
• 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
1.0
n = 10
0.8
K-S distance
0.6
0.4
0.2
0.0
••• •• • •• • •
0 50 100 150 200 250

Weibull Plots: n = 10
p(KS) = 0.62 • p(KS) = 0.28 • p(KS) = 0.88 • p(KS) = 0.88 •
1
p(CvM) = 0.54 p(CvM) = 0.27 p(CvM) = 0.62 p(CvM) = 0.64
• • • •
p(AD) = 0.62 p(AD) = 0.35 p(AD) = 0.66 p(AD) = 0.69
• • • •
0 • • • •
0
• • • •
• • • •
• • • •
-1
-1
-1
-1
• • • •
• • • •
-2
-2
-2
-2
• • • •
-3
-3
-3
-3
n = 10
3.8 4.2 4.6 5.0 4.0 4.4 4.8 3.8 4.4 5.0 3.8 4.2 4.6 5.0
p(KS) = 0.72 • p(KS) = 0.16 • p(KS) = 0.34 • p(KS) = 0.86 •

1
1
• • • •
p(AD) = 0.6 p(AD) = 0.32 p(AD) = 0.22 p(AD) = 0.57
• • • •
• • • •
0
0
• • • •
• • • •
• • • •
-1
-1
-1
-1
• • • •
• • • •
-2
-2
-2
-2
• • • •
-3
-3
-3
-3
4.0 4.4 4.8 3.8 4.2 4.6 4.0 4.4 4.8 3.8 4.2 4.6

p(KS) = 0.55 • p(KS) = 0.13 • p(KS) = 0.12 • p(KS) = 0.46 •
• • • •
1
• • • •
p(AD) = 0.49 • p(AD) = 0.028
•• p(AD) = 0.048
•• p(AD) = 0.38 •
••• ••• • • • ••
0 • •• •
0
•• • ••• •• ••
•
•• •• •• ••
• • • •
-1
-1
-1
-1
• • • •
• • • •
• • • •
-2
-2
-2
-2
• • • •
• • • •
-3
-3
-3
-3
• n = 20 • • •
3.5 4.0 4.5 5.0 2.5 3.5 4.5 4.2 4.6 3.8 4.2 4.6 5.0
p(KS) = 0.46 • p(KS) = 0.47 • p(KS) = 0.85 • p(KS) = 0.04 •

• • • •
1
1
• • • •
p(AD) = 0.41 •• p(AD) = 0.69
•• p(AD) = 0.64
•• p(AD) = 0.023
••
•• ••
•
••
•
••
•
•••
0
0
• • •• ••• •••
•• •• • • • •
• • • •
-1
-1
-1
-1
• • • •
• • • •
• • • •
-2
-2
-2
-2
• • • •
• • • •
-3
-3
-3
-3
• • • •
3.6 4.0 4.4 4.8 3.0 4.0 5.0 4.0 4.4 4.8 3.6 4.0 4.4 4.8

p(KS) = 0.7 • p(KS) = 0.22 • p(KS) = 0.15 • p(KS) = 0.76 •
•
• ••• ••• ••
p(CvM) = 0.41
•• p(CvM) = 0.35 p(CvM) = 0.23 p(CvM) = 0.49
••
1
•• ••• •
••• •
p(AD) = 0.43
••
• p(AD) = 0.32
•• • p(AD) = 0.12
•••• p(AD) = 0.52
••
•
0 •• •• •• • ••••
0
•••• ••••
• ••• ••
•• •• •• •••
• •• •••• •• •
• •••• ••
-1
-1
-1
-1
• •
••• •• •• • ••••
••• •• ••• ••
• • •••
•• • ••
-2
-2
-2
-2
•
• • • •
• • • •
• • • •
-3
-3
-3
-3
• • • •
-4
-4
-4
-4
• n = 50 • • •
2.0 3.0 4.0 5.0 3.0 4.0 5.0 3.5 4.0 4.5 3.5 4.5
p(KS) = 0.88 • p(KS) = 0.27 • p(KS) = 0.85 • p(KS) = 0.18 •

p(CvM) = 0.66 ••• p(CvM) = 0.22 ••
• p(CvM) = 0.56 ••
• p(CvM) = 0.035 ••
•
1
1
••• •• ••••
•••
••
p(AD) = 0.72 ••
•
p(AD) = 0.23
• •••• p(AD) = 0.62
•• p(AD) = 0.039
•• •
• •• •
•• •• •• •
•••
0
0
•••• ••
• •• •••
••• •• •
•• •
• • •• ••• •••
••••
-1
-1
-1
-1
•• • •
•• ••• •
••
•• •• • •• •
••
•• • • ••• •••
••
-2
-2
-2
-2
• •
• • • •
• • • •
• • • •
-3
-3
-3
-3
• • • •
-4
-4
-4
-4
• • • •
3.5 4.0 4.5 3.5 4.0 4.5 5.0 3.5 4.0 4.5 5.0 3.5 4.0 4.5 5.0

p(KS) = 0.58 • p(KS) = 0.86 • p(KS) = 0.61 • p(KS) = 0.21 •
•• ••• •• •••
p(CvM) = 0.26 •
••• p(CvM) = 0.68
•••• p(CvM) = 0.48 ••• p(CvM) = 0.26
•••
•
p(AD) = 0.3 ••• p(AD) = 0.72 ••
••• p(AD) = 0.55 ••• p(AD) = 0.43 ••
•••••
•••
• • •••
• ••
0 • •••• ••• ••
0
••• •••• •••••• ••••
••• ••• • ••
••
• • •• •• ••
••• •• •••• ••••
•••• •• ••• ••••
-2
-2
-2
-2
•
•• ••• ••• •••
•• ••• •• ••
• • •• •• ••
• • • •
• • • •
-4
-4
-4
-4
• • • •
• n = 100 • • •
3.0 4.0 5.0 3.5 4.5 3.0 4.0 5.0 3.5 4.5
p(KS) = 0.47 • p(KS) = 0.053 • p(KS) = 0.87 • p(KS) = 0.72 •

•• • •• •
p(CvM) = 0.3
••• p(CvM) = 0.025
•••• p(CvM) = 0.58
••• p(CvM) = 0.45
•••
•
• •• •
•• •• ••• •••
p(AD) = 0.37
•••
p(AD) = 0.032
•
p(AD) = 0.6
••••
•
p(AD) = 0.43
•••
•••
•
• ••• ••••
0
0
••• ••••• ••
•
•
•••
• ••• •••• •••
• ••• ••
••••• •••• •
••
•• ••••
•• • •• •••
••• •••• • •
••
•
-2
-2
-2
-2
•
• •• ••
•• • ••• ••• ••••
•• • ••• •• •
• • •• ••
• • • •
• • • •
-4
-4
-4
-4
• • • •
• • • •
3.0 4.0 5.0 3.0 4.0 5.0 3.5 4.5 2.5 3.5 4.5

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 θc by replacing (α, β) by (α, c
c β)
• Such estimates vary around the target due to sampling variation

• Capture the estimation uncertainty via confidence bounds
• For 0 < γ < 1 get 100γ% lower/upper confidence bounds θcL,γ & θcU,γ

P θL,γ ≤ θ = γ
c
or P θ ≤ θU,γ = γ
c
• For γ > .5 get a 100(2γ − 1)% confidence interval by [θcL,γ , θcU,γ ]

P θL,γ ≤ θ ≤ θU,γ = 2γ − 1 = γ ?
c c

Confidence Bounds & Sampling Variation, n = 10
L: Lawless Exact Method (RAP); B: Bain Exact Method (Tables); M: MLE Approximate Method
90% confidence intervals
35000
true α
25000
L BM
samples 1 2 3 4 5
8
true β
4
2
L BM
0
1 2 3 4 5

Confidence Bounds & Sampling Variation, n = 10
25000
95% lower confidence bounds
15000
true 10-percentile
5000
L BM
L: Lawless Exact Method (RAP); B: Bain Exact Method (Tables); M: MLE Approximate Method
0
samples 1 2 3 4 5
95% upper confidence bounds
L BM
0.20
true P(T < 10,000)

0.10
0.0
1 2 3 4 5

Confidence Bounds & Effect of n
35000
• •
• • •
• • •• ••
• • • •
• • •
•
25000
• • true α
•
•
15000
5 10 20 50 100 200 500 1000

sample size
•
7
•
6
•
5
• •
• • • • ••
• • •
4
• • • •
•
• •
3
• true β
•
2
•
1

Confidence Bounds & Effect of n
•
95% upper confidence bounds
0.15
•
true P(T < 10000)
•
0.05
• •
• • • • • •• •• •
0.0
5 10 20 50 100 200 500 1000

sample size
20000
95% lower confidence bounds
• • • • ••
• • • •
•
• •
true t .10
•
10000
•
5000

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 known
this shows up when we put n units on test simultaneously and
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 known
inspection 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 observation
due to other causes (different failure modes, competing risks)
• Multiple right censoring:

units are put into service at different times
and 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 correlate

with the (potential) failure times,
i.e., no censoring of anticipated failures
• All data (censored & uncensored) should enter analysis

Type I or Time Censored Data: n = 10
unit
10 X
9 ?
8 ?
7 X
6 X
5 ?
4 X
3 X
2 ?
1 ?
0 2000 4000 6000 8000 10000
cycles

Type II or Failure Censored Data: n = 10
unit
10 ?
9 ?
8 ?
7 ?
6 X
5 X
4 X
3 X
2 X
1 X
0 2000 4000 6000 8000 10000
cycles

Interval Censored Data: n = 10
unit
10 ?
9 ( •]
8 ?
7 ( • ?
6 ( • ]
5 ?
4 (• ]
3 ( • ?
2 ?
1 ( • ]
0 2000 4000 6000 8000 10000
cycles

Multiply Right Censored Data: n = 10
unit
10 ?
9 ?
8 ?
7 ?
6 X
5 X
4 X
3 X
2 ?
1 X
0 2000 4000 6000 8000 10000
cycles

Nonparametric MLEs of CDF
• For complete, type I & II censored data the nonparametric MLE is

d #{Ti ≤ t; i = 1, . . . , n}
F (t) = ,
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

 1 for t ≤ t
" # 
 ?
F (t) = 1 − (1 − p̂1)
d δ1 (t)
· · · (1 − p̂k )δk (t)
, δi(t) =  i
0 for t?i > t
where p̂i = di/ni, ni = # units known to be at risk just prior to t?i
and di = # units that failed at t?i
• The nonparametric MLE for interval censored data is complicated

Empirical CDF (complete samples)
empirical and true

0.8 and Weibull mle
0.8
distribution functions
0.4
0.4
n = 10 n = 30
0.0
0.0
0 5000 15000 25000 0 5000 15000 25000
0.8
0.8
0.4
0.4
n = 50 n = 100
0.0
0.0
0 5000 15000 25000 0 5000 15000 25000

Nonparametric MLE for Type I Censored Data
nonparametric mle of CDF

0.8 Weibull mle of CDF
0.8
and true CDF
0.4
0.4
n = 10 n = 30
0.0
0.0
... ... . .
0 5000 15000 25000 0 5000 15000 25000

0.8
0.8
0.4
0.4
n = 50 n = 100
0.0
0.0
. . .. .. ... ....... . .. .
0 5000 15000 25000 0 5000 15000 25000

Nonparametric MLE for Type II Censored Data
nonparametric mle of CDF

0.8 Weibull mle of CDF
0.8
and true CDF
0.4
0.4
n = 10 n = 30
0.0
0.0
.. . ... . ... .
0 5000 15000 25000 0 5000 15000 25000

0.8
0.8
0.4
0.4
n = 50 n = 100
0.0
0.0
... . .. .. . . .. .. . ........ .. ... . .
0 5000 15000 25000 0 5000 15000 25000

Kaplan-Meier Estimates (multiply right censored samples)
Kaplan-Meier CDF
0.8 Weibull mle for CDF
0.8
true CDF
0.4
0.4
n = 10 n = 30
0.0
0.0
. . . .. .. . .. . .. . . .. . . . .
0 5000 15000 25000 0 5000 15000 25000

0.8
0.8
0.4
0.4
n = 50 n = 100
0.0
0.0
.. ........ .... . .. . . . . .. . ..... ................. .......... . ... .
0 5000 15000 25000 0 5000 15000 25000

Nonparametric MLE for Interval Censored Data
1.0
superimposed is the true Weibull(10000,2.5)
that generated the 1,000 interval censored data cases
0.8
inspection points roughly 3000 cycles apart

0.6
CDF
0.4
0.2
0.0
0 5000 10000 15000
cycles

Nonparametric MLE for Interval Censored Data
1.0
superimposed is the true Weibull(10000,2.5)
that generated the 10,000 interval censored data cases
0.8
inspection intervals randomly generated from same Weibull

0.6
CDF
0.4
0.2
0.0
0 5000 10000 15000
cycles

Application to Y2K Questionnaire Return Data
1.0
probability of questionnaire return
0.8
8124 cases
0.6
0.4
0.2
estimated .99-quantile
0.0
0 50 100 150 200 250 300
days

Plotting Positions for Weibull Plots (Censored Case)
• For complete samples we plotted (log[T(i)], log [− log(1 − pi)]) with

 
i − .5 1 d d
1 i i − 1
pi = = F (T(i)) + F (T(i−1)) =  + 
n 2 2 n n
• For type I & II & multiply right censored samples

?
we plot in analogy (log[T(i) ], log [− log(1 − pi)]), where
? ?
T(1) < . . . < T(k) are the k distinct observed failure times and
1 "d ? d ?
#
pi = F (T(i)) + F (T(i−1))
2
d
and F (t) is the nonparametric MLE of F (t) for the censored sample

Weibull Plot, Multiply Right Censored Sample n = 50
.990
.900
•
•
•
.632
•
• ••
.500 •
••••
••
.200 ••••
probability
• ••
.100 •
•
•
•
.010 • true model, Weibull( 100 , 3 )

m.l.e. model, Weibull( 96.27 , 2.641 ) n = 50
least squares model, Weibull( 96.3 , 2.541 ) n = 50
.001
10 20 50 100 200 500 1000
cycles/hours

Confidence Bounds for Censored Data
• For complete & type II censored data RAP computes exact coverage
confidence bounds for α, β, tp, and P (T ≤ t).
• WEIBREG and commercial software (Weibull++, WeibullSMITH,
common statistical packages) compute approximate confidence
bounds for above targets, based on large sample m.l.e. theory
• RAP & WEIBREG are Boeing code, runs within DOS mode
of Windows95 (interface clumsy, but job gets done), contact me
• Boeing has a site license for Weibull++ Version 4
in the Puget Sound area
contact David Twigg (425) 717-1221, david.twigg@pss.boeing.com
• There is a Weibull++ Version 5 out

Weibull Plot With Confidence Bound
0.999
0.990
alpha = 102.4 [ 82.52 , 143.8 ] 95 %
0.900
beta = 2.524 [ 1.352 , 3.696 ] 95 %
n = 30 , r = 13
•
•
0.500 •
•
•
••
•
0.200
•
•
0.100 •
•
•
0.010
0.001
1 2 3 5 10 20 50 100 200 500 1000

Weibull Plot With Confidence Bound (Variation)
0.999
0.990
alpha = 89.79 [ 77.4 , 109.4 ] 95 %
0.900
beta = 3.674 [ 2.15 , 5.198 ] 95 %
n = 30 , r = 13
•
•
0.500
•
•
•
•
•
0.200 •
•
•
0.100 •
•
•
0.010
0.001
1 2 3 5 10 20 50 100 200 500 1000

Weibull Regression Model & Analysis
• Recall tp = α[− log(1 − p)]1/β or

log(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
• Regression model on log(α) (multiplicative on α)
log(αi) = b1Zi1 + . . . + bpZip
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
 c
Volt
T (Volt) = T (VoltU )  
 , where usually c < 0
VoltU
this means  c
Volt
α(Volt) = α (VoltU )  

VoltU
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 involving

chemical reaction rates is to increase the temperature
• Arrhenius proposed the following acceleration model with
temperature measured in Kelvin (tempK = temp◦C + 273.15)
 
 κ κ 
T (temp) = T (tempU ) exp  − 
temp tempU
or
κ κ
log [α(temp)] = log [α(tempU )] + −
temp tempU
= b1 + b2 Z
with
κ
b1 = log [α(tempU )] − , b2 = κ, and Z = temp−1
tempU

Pooling Data With Different α’s
• Suppose we have three groups of failure data

T1, . . . , 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 error

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
200
100
thousand cycles
50
20
10 • • •
• ••
• ••
5
•
50 100 150 200 250 300 350 400 450 500
Voltage

Accelerated Life Testing: Model, Data & MLEs
200
100
thousand cycles
50 true li
ne
mle lin
e
20
10 • • •
mle for .01-quantile • ••
• ••
5
•
95% lower bound for .01-quantile
50 100 150 200 250 300 350 400 450 500
Voltage

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
2 × TTT
θ̂L,γ = ,
χ22r,γ
where T T T = 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 T T T = 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
2 × TTT
θ̂L,γ = 2
χ2r+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) ⇒ T β ∼ E(θ) Exponential with mean θ = αβ
• As 100γ% lower confidence bound for α compute

 
1/β
2 × T T T (β) 


α̂L,γ = 
 2


χf,γ
where for complete or type II censored samples we take f = 2r and
β β β
T T T = T(1) + · · · + T(r) + (n − r)T(r)
and for multiply right censored data we take f = 2r + 2 and
T T T = T1β + · · · + Tnβ .
• 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 samples
from the 3-parameter Weibull distribution.
– The A-Allowable is a 95% lower confidence bound
for 99% of the population, i.e., for the .01-quantile.
– The B-Allowable is a 95% lower confidence bound
for 90% of the population, i.e., for the .10-quantile.
• Hybrid approach of using Mann-Fertig threshold estimate
combined with maximum likelihood estimates for α and β,
took lot of simulation and tuning and is very specialized in its
capabilities
• 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-Testing
Models, Marcel Dekker, New York
• 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 Lifetime
Data, John Wiley & Sons, New York
• Mann, N.R., Schafer, R.E., and Singpurwalla, N.D. (1974), Methods
for Statistical Analysis of Reliability and Life Data, John Wiley &
Sons, New York
⇒ Meeker, W.Q. and Escobar, L.A. (1998), Statistical Methods for
Reliability Data, John Wiley & Sons, New York

• Meeker, W.Q. and Hahn, G.J. (1985), Volume 10: How to Plan an
Accelerated Life Test—Some Practical Guidelines, American Society
for Quality Control
• Nelson, W. (1982), Applied Life Data Analysis, John Wiley & Sons,
New York
— (1985), “Weibull analysis of reliability data with few or no
failures,” Journal of Quality Technology 17, 140-146
— (1990), Accelerated Testing, Statistical Models, Test Plans and
Data Analyses, John Wiley & Sons, New York
• Scholz, F.W. (1994), “Weibull and Gumbel distribution exact
confidence bounds.” BCSTECH-94-019 (RAP)
— (1996) “Maximum likelihood estimation for type I censored
Weibull data including covariates.” ISSTECH-96-022
— (1997), “Confidence bounds for type I censored Weibull data
including covariates.” SSGTECH-97-025 (WEIBREG)
The Boeing Company, P.O. Box 3707, MS-7L-22, Seattle WA 98124-2207

Weibull Reliability Analysis: Boeing Phantom Works

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Weibull Reliability Analysis: Boeing Phantom Works

Uploaded by

Copyright:

Available Formats

Weibull Reliability Analysis

Fritz Scholz (425-865-3623, 7L-22)

Weibull Reliability Analysis—FWS-5/1999—1

Weibull Reliability Analysis—FWS-5/1999—2

Weibull Reliability Analysis—FWS-5/1999—3

• Weibull distribution, useful uncertainty model for

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

Weibull Reliability Analysis—FWS-5/1999—4

• Under weak conditions Extreme Value Theory shows1 that for

Weibull Reliability Analysis—FWS-5/1999—5

If X1, . . . , Xn are statistically independent

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

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

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

• the characteristic life or scale parameter α>0

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

Weibull Reliability Analysis—FWS-5/1999—7

• 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

• When a 3-parameter Weibull model is assumed,

Weibull Reliability Analysis—FWS-5/1999—8

• The expectation or mean value of T

• 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 Reliability Analysis—FWS-5/1999—9

• The cumulative distribution function F (t) = P (T ≤ t) is just

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

Weibull Reliability Analysis—FWS-5/1999—10

Weibull density α = 10000, β = 2.5

0 5000 10000 15000 20000

Weibull Reliability Analysis—FWS-5/1999—11

τ=0 τ = 1000 τ = 2000

Weibull Reliability Analysis—FWS-5/1999—12

0 2000 4000 6000 8000

Weibull Reliability Analysis—FWS-5/1999—13

0 1000 2000 3000 4000

Weibull Reliability Analysis—FWS-5/1999—14

• A third representation of the Weibull distribution is through the

Weibull Reliability Analysis—FWS-5/1999—15

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

• This distribution is useful when parts fail due to

Weibull Reliability Analysis—FWS-5/1999—16

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

Weibull Reliability Analysis—FWS-5/1999—17

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

0 20000 40000 60000 80000

Weibull Reliability Analysis—FWS-5/1999—18

characteristic life α = 30,000

parameter estimates from three samples of size n = 10

Weibull Reliability Analysis—FWS-5/1999—19

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

– 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

• Simulations are useful in gaining insight on estimation procedures

Weibull Reliability Analysis—FWS-5/1999—20

• 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)

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

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

these points to cluster around main diagonal

Weibull Reliability Analysis—FWS-5/1999—21

0 5000 10000 15000