Professional Documents
Culture Documents
Preliminaries
Tentative Content • The course grade will be based on three group report
projects plus an individual closed book final exam.
• Part 1: Introduction to reliability data and non-parametric We need to divide the class in eight groups of five stu-
estimation of reliability. Chapters 1, 2, and 3. dents each. I need to have the list of the eight groups by
Wednesday June 13. You are responsible for making up
the groups.
• Part 2: Log-location-scale models and probability plotting.
Chapters 4 and 6.
• The final exam will be administered on July 10 in class.
June 7, 2018
8h 14min 1-1 1-2
Chapter 1
• Rapid advances in technology.
Segment 1
• Development of highly sophisticated products.
Basic Reliability Concepts
1-3 1-4
Failure avoidance.
• Check the veracity of an advertising claim.
Life data.
Chapter 1 Survival data.
1-7 1-8
0 50 100 150 200 at three years for a product in the field only six months).
Millions of Cycles
1 - 11 1 - 12
Issues in Failure-Time Data Analysis
IC Data
1 - 13 1 - 14
n = 4,128 IC’s tested for 1,370 hours at 80◦C and 80% Integrated Circuit Failure Data After 1370 Hours
relative humidity, there were 28 failures (Meeker 1987). Count
Questions: Row
1 2
2
3
4 2
5
• Proportion of units that will fail in 105 hours (proportion of 6
7
defective units)? 8 2
9
10 2
11
12 2
13 2
• Hazard function 14 2
15
16
17
For the product population? 18
19
20
21
For the defective subpopulation? 22 4128
A shock absorber with both Mode 1 and Mode 2 active. 0 5000 10000 15000 20000 25000 30000
Kilometers
1 - 17 1 - 18
Strategy for
Data Analysis, Modeling, and Inference Computer Software
• Model-free graphical data analysis. Capabilities for reliability data analysis can be found in:
• Sensitivity analysis. With JMP one can accomplish most of the data analyses
needed for this course.
• Conclusions.
1 - 19 1 - 20
Segment 3 • Data from 3 plants with same design. Each plant entered
into service at different dates.
Failure-time Data Based on Inspections
Cracked tubes at the end of each year,
1 - 21 1 - 22
Heat Exchanger Tube Crack Inspection Data Heat Exchanger Tube Crack Inspection Data
in Calendar Time in Operating Time
100 new 1 failure 2 failures 2 failures 100 new 1 failure 2 failures 2 failures
tubes tubes
Plant 1 95 Plant 1 95
.. .. .. .. .. ..
. . . . . .
100 new 100 new
tubes 2 failures 3 failures 2 failures 3 failures
tubes
Plant 2 95 Plant 2 95
.. .. .. ..
. . . .
100 new
1 failure 100 new
tubes tubes 1 failure
Plant 3 99 Plant 3
99
.. ..
. .
1981 1982 1983 Year 1 Year 2 Year 3
1 - 23 1 - 24
Turbine Wheel Data (Nelson 1982)
Current-Status Data
Excessive wear found in an aircraft jackscrew triggered • Need to schedule regular inspections.
inspection of similar aircraft.
A crack detected in an aircraft engine rotating compo- • Estimate the distribution of time to crack.
nent triggered inspection of all such components in ser-
vice.
• Is the reliability of the wheels getting worse as the wheels age?
Event Plot
Turbine Wheel Inspection Data (Nelson 1982)
Turbine Wheel Data
Summary at Time of Study
Turbine Wheel Crack Initiation Data
Count
100-hours of # Cracked # Not Cracked Row
Exposure Left Censored Right Censored 1 39
2 ? | 4
3 49
4 0 39 4 ? | 2
5 31
10 4 49 6 ? | 7
14 2 31 7
8 ? |
66
5
18 7 66 9 25
10 ? | 9
22 5 25 11 30
12 ? | 9
26 9 30 13 33
14 ? | 6
30 9 33 15 7
34 6 7 16
17
? | 22
12
38 22 12 18
19
? | 21
19
42 21 19 20 ? | 21
21 15
46 21 15
0 10 20 30 40 50
Hundreds of Hours
1 - 27 1 - 28
250
Chapter 1
200
Segment 4
50
1 - 29 1 - 30
Scatter Plot of Printed Circuit Board Accelerated Life
Test Data (Meeker & LuValle 1995)
Degradation Data
11/68 censored
x
x x
x • Becoming more common in certain areas of component re-
xxx
x x liability where few or no failures expected in life tests.
x
x xx
Hours to Failure
x
x
x
x xxx 0/70 censored
500
x xx
x xxx xx
xxx xxx 0/70 censored • Important connections with physical mechanisms of failure
xx xx
x
xx xxx
x xx xx and failure-time reliability models.
xx xxx
xx
xx x
xxx
50
xxx
xxx x
Destructive degradation data (Chapter 19).
xx
x x
x
Repeated measures degradation data (Chapter 20).
10
40 50 60 70 80 90
Relative Humidity
1 - 31 1 - 32
Fatigue Crack Size as a Function of Number of Cycles Change in Resistance Over Time of Carbon-Film
(Bogdanoff & Kozin 1985) Resistors (Shiomi & Yanagisawa 1979)
1.8 3
10.0
9
7
Critical crack size 1 12
1.6 5.0 173 Degrees C
Crack Length (inches)
Percent Increase
14
1.4
18
133 Degrees C
21
1.0
1.2
83 Degrees C
0.5
0.0 0.02 0.04 0.06 0.08 0.10 0.12 0 2000 4000 6000 8000 10000
1 - 33 1 - 34
Recurrent-Event Data
Valve Seat Replacements
Repairable Systems and Nonrepairable Units
408
tem tests or monitoring. 406
404
402
Time to first failure of a system. 400
398
396
394
392
390
• Repairable Systems result in recurrent-event data that 331
329
327
describe the failure trends and patterns of a system (Chap- 251
ter 21).
0 200 400 600
Engine Age in Days
1 - 35 1 - 36
Chapter 2
Chapter 2
Models, Censoring, and Likelihood for
Failure-Time Data Models, Censoring, and Likelihood for
Failure-Time Data
1 0.8
0.6
0.2
Segment 1 0
0.0 0.5 1.0 1.5 2.0 2.5 0.0 0.5 1.0 1.5 2.0 2.5
t t
Failure-time Models and Metrics
Survival Function Hazard Function
1
2.00
1.00
0.50
S (t ) 0.5 h (t )
0.20
0.10
0.05
0
0.0 0.5 1.0 1.5 2.0 2.5 0.0 0.5 1.0 1.5 2.0 2.5
t t
2-3 2-4
2-7 2-8
Probability Density Function Cumulative Distribution Function Pr(T ≤ tp) = F (tp) ≥ p, where 0 < p < 1.
0.8 1.0 t0.20 is the time by which 20% of the population will fail.
For, F (t) = 1 − exp(−t1.7), p = F (tp) gives tp = [− log(1 −
0.8 p)]1/1.7 and t0.2 = [− log(1 − 0.2)]1/1.7 = 0.414.
0.6
• When F (t) is strictly increasing there is a unique value tp
0.6 that satisfies F (tp) = p, and we write
f (t ) 0.4 F (t )
tp = F −1(p).
0.4
2-9 2 - 10
2 - 11 2 - 12
Distribution of Remaining Life
Distribution of Remaining Life
• Certain applications require consideration of the distribution
of remaining life:
... ... F (t 2)
0.6
F (t 1)
Times need not be equally spaced. 0.2
π1
0.0
0.0 0.5 1.0 1.5 2.0
t0 t1 t2 t3 t4
2 - 17 2 - 18
Models for Discrete Data
from a Continuous Time Processes
(t0, t1], (t1, t2], . . . , (tm−1, tm], (tm, tm+1) Following from the previous result,
πi = piS(ti−1)
Define,
i
πi = Pr(ti−1 < T ≤ ti) = F (ti) − F (ti−1) S(ti) = 1 − pj , i = 1, . . . , m + 1
j=1
F (ti) − F (ti−1)
pi = Pr(ti−1 < T ≤ ti | T > ti−1) = Either π = (π1, . . . , πm+1) or p = (p1, . . . , pm) can be used as
1 − F (ti−1)
“nonparametric parameters.”
Because the πi values are multinomial probabilities, πi ≥ 0
2 - 19 2 - 20
2 - 21 2 - 22
0.6
• Assumed model.
f (t ) 0.4
• Identifiability/parameterization.
0.0
0.0 0.5 1.0 1.5 2.0 2.5
t
2 - 25 2 - 26
• For a life table the data are: the number of failures (di),
• Interval-censored observations: right censored (ri), and left censored (i) units on each of
t
i the nonoverlapping interval (ti−1, ti], i = 1, . . . , m+1, t0 = 0.
Li = f (t) dt = F (ti) − F (ti−1).
ti−1
• The likelihood (probability of the data) for a single obser-
If a unit is still operating at t = 1.0 but has failed at t = 1.5
vation, datai, in (ti−1, ti] is
inspection, Li = F (1.5) − F (1.0) = 0.231.
Li(π ; datai) = F (ti; π ) − F (ti−1; π ).
• Left-censored observations:
t
Li =
i
f (t) dt = F (ti) − F (0) = F (ti). • Assuming that the censoring is at ti (note that F (t) depends
0 on either π or p):
If a failure is found at the first inspection time t = 0.5,
Type of Characteristic Number Likelihood of
Li = F (0.5) = 0.265. Censoring of Cases Responses Li(π ; datai)
• Right-censored observations: Left at ti T ≤ ti i [F (ti)]i
∞
Li = f (t) dt = F (∞) − F (ti) = 1 − F (ti). Interval ti−1 < T ≤ ti di F (ti) − F (ti−1) di
ti
If a unit has not failed by the last inspection at t = 2, Right at ti T > ti ri [1 − F (ti)]ri
Li = 1 − F (2) = 0.0388.
2 - 27 2 - 28
Likelihood: Probability of the Failure-time Data Likelihood for Arbitrary Censored Failure-Time Data
• The total likelihood, or joint probability of the DATA, for • In general, observation i consists of an interval (tL i , ti ], i =
n independent observations is (note that F (t) depends on 1, . . . , n (tL
i < ti ) that contains the time event T for individ-
either π or p): ual i.
n
The intervals (tL
i , ti ] may overlap and their union may not
L(π ; DATA) = C Li(π ; datai)
cover the entire time line (0, ∞). In general tLi = ti−1 .
i=1
m+1
• Assuming that the censoring is at ti
=C [F (ti)]i F (ti) − F (ti−1) di [1 − F (ti)]ri
i=1 Type of Characteristic Likelihood of a single
where n = m+1 Censoring Response Li(π ; datai)
j=1 dj + rj + j and C is a constant depend-
ing on the sampling inspection scheme but not on π (so we Left at ti T ≤ ti F (ti)
can take C = 1). Interval i < T ≤ ti
tL F (ti) − F (tL
i)
• Want to find π so that L(π ; DATA) is large. Right at ti T > ti 1 − F (ti)
2 - 29 2 - 30
Likelihood for General Failure-Time Data
k
w
j
L(π ; DATA) = Lj (π ; dataj ) .
j=1
2 - 31
Chapter 3 Chapter 3
Nonparametric Estimation
Nonparametric Estimation
of a Failure-Time Distribution
Topics discussed in this chapter are:
• Use of the binomial distribution to estimate F (t) from in-
terval and singly right censored failure-time data, without
assumptions about the form of F (t). This is called non-
parametric estimation.
Cracked tubes
Chapter 3
100 tubes at start Year 1 Year 2 Year 3 Uncracked tubes
Segment 1 Plant 1 1 2 2 95
3-3 3-4
3-7 3-8
3-9 3 - 10
• For a specified value of ti, an approximate 100(1 − α)% • A 100(1 − α)% approximate confidence interval for F (ti) is
confidence interval for F (ti) is
F (ti) = qbeta(α/2; nF + 0.5, n − nF + 0.5)
[F (ti), F̃ (ti)] = F (ti) ∓ z(1−α/2)seF . ⎧ ⎫−1
where z(1−α/2) is the 1−α/2 quantile of the standard normal
⎪
⎨ (n − nF + 0.5)F(1−α/2;2n−2nF+1,2nF+1) ⎪
⎬
= 1+
⎪
⎩ nF + 0.5 ⎪
⎭
distribution and seF = F (ti) 1 − F (ti) /n is an estimate of
the standard error of F (ti). F̃ (ti) = qbeta(1 − α/2; nF + 0.5, n − nF + 0.5)
⎧ ⎫−1
⎪
⎨ ⎪
⎬
n − nF + 0.5
= 1+
• This confidence interval is based on ⎪
⎩ (nF + 0.5)F(1−α/2;2nF+1,2n−2nF+1) ⎪
⎭
F (ti) − F (ti)
ZF = ∼
˙ NORM(0, 1). where F = F (ti) and F(1−α/2;ν1,ν2) is the 100(1−α/2) quan-
seF
tile of the F distribution with (ν1, ν2) degrees of freedom.
• This method is computationally simple but is not recom- • This method is based on a Bayesian interval using a Jef-
mended because of its poor coverage probability properties freys prior distribution and has coverage probability prop-
(see Section 6.2.6 of Meeker, Hahn, and Escobar, 2017). erties that are much better than the Wald method (see
Instead, use the Jeffreys or the Conservative method. Section 6.2.6 of Meeker, Hahn, and Escobar, 2017).
3 - 11 3 - 12
Pointwise Conservative Plant 1 Heat Exchanger Tube Crack Nonparametric
Confidence Interval for F (ti) Estimate with Conservative Pointwise 95% Confidence
Intervals Based on Binomial Theory
Proportion Failing
F̃ (ti) = qbeta(1 − α/2; nF + 1, n − nF )
⎧ ⎫−1 0.10
⎪
⎨ ⎪
⎬
n − nF
= 1+
⎪
⎩ (nF + 1)F(1−α/2;2nF+2,2n−2nF) ⎪
⎭
0.05
where F = F (ti) and F(1−α/2;ν1,ν2) is the 100(1−α/2) quan-
tile of the F distribution with (ν1, ν2) degrees of freedom.
0.0
• This confidence interval is conservative in that the actual 0.0 0.5 1.0 1.5 2.0 2.5 3.0
coverage probability is greater than or equal to 1 − α. Years
3 - 13 3 - 14
Year ti di F (ti )
se F
Pointwise Confidence Interval Integrated Circuit (IC) Failure Times in Hours
F (ti ) F̃ (ti ) Data from Meeker (1987)
(0 − 1] 1 1 0.01 0.00995
95% Confidence Intervals for F (1)
Wald [−0.0095, 0.0295]
Jeffreys [0.0011, 0.0458]
0.10 0.10 0.15 0.60 0.80 0.80
Conservative [0.0003, 0.0545] 1.20 2.50 3.00 4.00 4.00 6.00
10.00 10.00 12.50 20.00 20.00 43.00
(1 − 2] 2 2 0.03 0.01706 43.00 48.00 48.00 54.00 74.00 84.00
95% Confidence Intervals for F (2) 94.00 168.00 263.00 593.00
Wald [−0.0034, 0.0634]
Jeffreys [0.0085, 0.0779]
Conservative [0.0062, 0.0852] When the test ended at 1370 hours, there were 28
observed failures and 4128 unfailed units.
(2 − 3] 3 2 0.05 0.02179
95% Confidence Intervals for F (3) Note: Ties in the data. Reason?
Wald [0.0073, 0.0927]
Jeffreys [0.0193, 0.1061]
Conservative [0.0164, 0.1128]
3 - 15 3 - 16
Event Plot
Nonparametric Estimator of F (t) Integrated Circuit Life Test Data
Based on Binomial Methods for Exact Failures
and Singly Right-Censored Data
10 2
ods give 11
12 2
# of failures up to time te 13 2
F (te) = 14 2
n 15
16
F (te ) 1 − F (te ) 17
18
seF = . 19
20
n 21
22 4128
• F (t) is defined for all t in the interval (0, tc] where tc is the
0.012 single censoring time.
0.008
1/n at each exact failure time (unless there are ties).
0.006
Sometimes the step size is an integer multiple of 1/n be-
0.004 cause there are ties in the failure times.
0.002
• When there is no censoring, F (t) is the well known empiri-
cal cdf.
0.0
Plant 1
100 99 97
1 2 2 95
Plant 2
100 98
2 3 95
Chapter 3 100
Plant 3 1 99
Segment 3
All Plants
300 197 97
4 5 2 95
Failure Probability π1 π2 π3 π4
4 5 2 95 95 99
Likelihood: L( π
__) = C [π 1] [π 2] [ π 3] [π4] [π 3+ π 4] [π 2+ π 3+ π 4]
3 - 21 3 - 22
The pooled data heat exchanger tube crack data are multiply
censored and the simple binomial method to estimate F (ti)
cannot be used. Summary of the Nonparametric Estimate of F (ti) for
the Pooled Heat Exchanger Tube Crack Data
Consider the more general nonparametric estimator of F (ti)
based on the probability model introduced in Chapter 2: Year ti ni di ri pi 1 − pi )
S(t F (ti)
i
F (ti) = 1 − S(t
)
i (0 − 1] 1 300 4 99 4/300 296/300 0.9867 0.0133
i
dj (1 − 2] 2 197 5 95 5/197 192/197 0.9616 0.0384
where )=
S(t 1 − pj with pj =
i
j=1 nj (2 − 3] 3 97 2 95 2/97 95/97 0.9418 0.0582
i
• Recall, F (ti) = 1 − S(t
) and S(t
i
)=
i j=1 1 − p
j .
0.20
• Then Var F (ti) = Var S(t
) .
i
) is
• A Taylor series first-order approximation of S(t
0.15 i
Proportion Failing
i
∂S
) ≈ S(t ) +
S(t qj − qj
i i
0.10 j=1 ∂qj qj
where qj = 1 − pj .
3 - 27 3 - 28
• The endpoints F (ti) and F̃ (ti) will always lie between 0 and 1.
3 - 29 3 - 30
Pointwise Wald-Approximate Confidence Intervals for Summary of the Nonparametric Pointwise Confidence
the Heat Exchanger Tube Crack Data Intervals for F (ti) Based on the Heat Exchanger Tube
Crack Data
Computation of approximate 95% confidence intervals:
Year ti F (ti )
se F
Pointwise Confidence Intervals
√
• For F (1) with F (t1) = 0.0133, se
=
F (t1 )
0.0000438 = 0.00662 (0 − 1] 1 0.0133 0.00662
Based on: ZF = [F (t1) − F (t1)]/se
∼ ˙ NORM(0, 1). 95% Confidence Intervals for F (1)
F
Based on Zlogit(F) ∼
˙ NORM(0, 1) [0.0050, 0.0350]
[F (t1), F̃ (t1)] = 0.0133 ∓ 1.96(0.00662) = [0.0003, 0.0263]. Based on ZF ∼˙ NORM(0, 1) [0.0003, 0.0263]
Based on: Zlogit(F) = [logit[F (t1)]−logit[F (t1)]/se
) ∼˙ NORM(0, 1).
logit(F (1 − 2] 2 .0384 .0128
" #
0.0133 0.0133 95% Confidence Intervals for F (2)
[F (t1), F̃ (t1)] = ,
0.0133 + (1 − 0.0133) × w 0.0133 + (1 − 0.0133)/w Based on Zlogit(F) ∼
˙ NORM(0, 1) [0.0198, 0.0730]
= [0.0050, 0.0350]. Based on ZF ∼˙ NORM(0, 1) [0.0133, 0.0635]
where w = exp{1.96(0.00662)/[0.0133(1 − 0.0133)]} = 2.687816.
√ (2 − 3] 3 .0582 .0187
• For F (2) with F (t2) = 0.0384, se
=
F (t )
0.0001639 = 0.0128
2
95% Confidence Intervals for F (3)
Based on: ZF, [F (t2), F̃ (t2)] = [0.0133, 0.0635]. Based on Zlogit(F) ∼
˙ NORM(0, 1) [0.0307, 0.1076]
Based on ZF ∼˙ NORM(0, 1) [0.0216, 0.0949]
Based on: Zlogit(F) , [F (t2), F̃ (t2)] = [0.0198, 0.0730].
3 - 31 3 - 32
0.20
Chapter 3
Segment 4
0.15
Proportion Failing
0.05
0.0
Event Plot
Shock Absorber Field-Failure Data
Shock Absorber Failure Data Failure Mode Information Ignored
Row
1
2
• Failure times, in number of kilometers of use, of vehicle 3
4
5
6
shock absorbers. 7
8
9
10
11
• Two failure modes, denoted by M1 and M2.
12
13
14
15
16
17
• One might be interested in the distribution of time to fail- 18
19
20
21
22
ure for mode M1, mode M2, or in the overall failure-time 23
24
25
distribution of the part. 26
27
28
29
30
31
Here we do not differentiate between modes M1 and M2. 32
33
34
35
We will estimate the distribution of time to failure by either 36
37
38
mode M1 or M2.
0 5000 10000 15000 20000 25000 30000
Kilometers
3 - 35 3 - 36
Nonparametric Estimation of F (t) with Exact Failures Kaplan-Meier Estimates for the Shock Absorber Data
Using the Kaplan-Meier Estimator up to 12,220 km
3 - 37 3 - 38
1.0
Chapter 3
0.8
Segment 5
Proportion Failing
0.6
Nonparametric Simultaneous Confidence Bands
for F (t)
0.4
0.2
0.0
ticular value of t.
0.6
0.0
Comments:
0.6
• These particular simultaneous confidence bands are known
as “equal precision” or “EP” bands.
0.4
• The approximate factors e(a,b,1−α/2) can be computed from
0.2 a large-sample approximation given in Nair (1984).
3 - 45 3 - 46
Nonparametric Estimate with Simultaneous 95% Nonparametric Estimate with Pointwise Wald 95%
Confidence Bands Based on Zmax logit(F) Confidence Intervals Based on Zlogit(F)
for the Heat Exchanger Tube Crack Data for the Heat Exchanger Tube Crack Data
0.20 0.20
0.15 0.15
Proportion Failing
Proportion Failing
0.10 0.10
0.05 0.05
0.0 0.0
0.0 0.5 1.0 1.5 2.0 2.5 3.0 0.0 0.5 1.0 1.5 2.0 2.5 3.0
Years Years
3 - 47 3 - 48
Nonparametric Estimation of F (ti)
with Arbitrary Censoring
Truncated data.
3 - 49 3 - 50
0 39 4 49 2 31 21 15
0.6
Proportion Failing
0 4 10 18 26 34 42 46
0.2
Hundreds of Hours
0.0
0 10 20 30 40 50
Hundreds of Hours
3 - 53 3 - 54
Nonparametric ML Estimate with Pointwise
Nonparametric Estimation of F (t)
Approximate 95% Confidence Intervals for F (ti) Based
with Current-Status Data
on Zlogit(F) for the Turbine Wheel Data
Proportion Failing
L(π ) = L(π ; DATA) = C × [π1]0 × [π2 + · · · + π12]39×
×[π1 + π2]4 × [π3 + · · · + π12]49 0.4
×[π1 + · · · + π3]2 × [π4 + · · · + π12]31
...
0.2
×[π1 + · · · + π11]21 × [π12]15
0.8
0.6
Proportion Failing
0.4
0.2
0.0
0 10 20 30 40 50
Hundreds of Hours
3 - 57
Chapter 4 Chapter 4
Location-Scale and Log-Location-Scale Parametric
Location-Scale and Log-Location-Scale Parametric Distributions
Distributions
Topics discussed in this chapter are:
• Motivation, definition, metrics, and importance of location-
scale and log-location-scale parametric distributions.
Based on the authors’ book Statistical Methods for Relia- • Description and properties of the generalized gamma distri-
bility Data, Second Edition, John Wiley & Sons Inc., to be bution.
published in 2019.
• Description and properties of threshold log-location-scale
June 7, 2018 distributions.
8h 15min 4-1 4-2
4-3 4-4
The distribution of the random variable Y belongs to the The distribution of the random variable T belongs to the log-
location-scale family of distributions if the cdf of Y can be location-scale family of distributions if the cdf of T can be
expressed as expressed as
( )
y−μ log(t) − μ
F (y; μ, σ) = Pr(Y ≤ y) = Φ , −∞ < y < ∞ F (t; μ, σ) = Pr(T ≤ t) = Φ , 0<t<∞
σ σ
where −∞ < μ < ∞ is a location parameter and σ > 0 is a where 0 < exp(μ) < ∞ is a scale parameter and σ > 0 is a
scale parameter. shape parameter.
Φ is the cdf of Y when μ = 0 and σ = 1 and Φ does not Φ is the cdf of log(T ) when μ = 0 and σ = 1 and Φ does not
depend on any unknown parameters. depend on any unknown parameters.
Note: Z = (Y − μ)/σ is a standardized random variable and Note: Z = (log(T ) − μ)/σ is a standardized random variable
Z ∼ Φ(z). Thus the distribution of Z does not depend on and Z ∼ Φ(z). Thus the distribution of Z does not depend
any unknown parameters. on any unknown parameters.
4-5 4-6
Metrics/Functions of the Parameters Metrics/Functions of the Parameters-Continued
• Cumulative distribution function (cdf) of T • The mean time to failure, MTTF, of T (also known as the
expectation of T )
F (t; θ ) = Pr(T ≤ t), t > 0. ∞ ∞
E(T ) = tf (t; θ ) dt = [1 − F (t; θ )] dt.
0 0
• The p quantile of T is the smallest value tp such that If 0∞ tf (t; θ ) dt = ∞, we say that the mean of T does not
exist.
F (tp; θ ) ≥ p.
• The variance (or the second central moment) of T and the
standard deviation
• Hazard function of T ∞
Var(T ) = [t − E(T )]2f (t; θ ) dt
f (t; θ ) *0
h(t; θ ) = , t > 0.
1 − F (t; θ ) SD(T ) = Var(T ).
4-7 4-8
• Methods of inference, statistical theory, and computer soft- Exponential, Normal, Lognormal, Smallest Extreme
ware generated for the general family can be applied to this Value, and Weibull Distributions
large, important class of models.
4-9 4 - 10
0.0
E(T ) = γ + θ, Var(T ) = θ2. 0.0 0.5 1.0 1.5 2.0 2.5 3.0
t
4 - 11 4 - 12
Motivation for the Exponential Distribution
Normal (Gaussian) Distribution
Cumulative Distribution Function Probability Density Function If T ∼ LNORM(μ, σ) then log(T ) ∼ NORM(μ, σ) with
1
1.2 log(t) − μ
1.0 F (t; μ, σ) = Φnorm
σ
F (t ) 0.5 f (t )
0.8
0.6 1 log(t) − μ
0.4 f (t; μ, σ) = φnorm , t > 0.
0.2 σt σ
0 0.0
φnorm and Φnorm are the pdf and cdf for the standard normal
3 4 5 6 7 3 4 5 6 7
t t distribution. exp(μ) is a scale parameter; σ > 0 is a shape
Hazard Function
parameter.
20 σ μ
0.3 5 Quantiles: tp = exp μ + σΦ−1 −1
norm (p) , where Φnorm (p) is the
15
0.5 5 p quantile for the standard normal distribution.
h (t ) 10 0.8 5
Moments: For integer m > 0, E(T m) = exp mμ + m2σ 2/2 .
5
0 E(T ) = exp μ + σ 2/2 , Var(T ) = exp 2μ + σ 2 exp(σ 2) − 1 .
3 4 5 6 7
t
4 - 15 4 - 16
Cumulative Distribution Function Probability Density Function • The lognormal distribution is a common model for fail-
1 5 ure times.
4
0
components with a decreasing hazard function (due to a
0.0 0.5 1.0 1.5 2.0 2.5 small proportion of defects in the population).
t
4 - 17 4 - 18
Examples of Smallest Extreme Value Distributions
Smallest Extreme Value Distribution
4 - 19 4 - 20
Weibull Distribution
Examples of Weibull Distributions
T ∼ WEIB(η, β)
⎡ ( ) ⎤
β
t ⎦
F (t; η, β) = Pr(T ≤ t) = 1 − exp⎣− Cumulative Distribution Function Probability Density Function
η 1 2.0
( )β−1 ⎡ ( ) ⎤
β
β t t ⎦ 1.5
f (t; η, β) = exp⎣−
η η η F (t ) 0.5 f (t ) 1.0
( )β−1
β t 0.5
h(t; η, β) =, t>0
η η 0 0.0
0.0 0.5 1.0 1.5 2.0 0.0 0.5 1.0 1.5 2.0
where η > 0 is a scale parameter and β > 0 is a shape pa- t t
rameter.
Hazard Function
∞ 0.0
where Γ(κ) = wκ−1 exp(−w)dw is the gamma function. 0.0 0.5 1.0
t
1.5 2.0
0
Note: When β = 1 then T ∼ EXP(η).
4 - 21 4 - 22
Note: If T ∼ WEIB(η, β) then Y = log(T ) ∼ SEV(μ, σ). • The theory of extreme values shows that the Weibull dis-
Then the Weibull cdf can be written as tribution can be used to model the minimum of a large
log(t) − μ number of independent positive random variables from cer-
F (t; μ, σ) = Pr(T ≤ t) = Φsev
σ tain classes of distributions.
1 log(t) − μ
f (t; μ, σ) = φsev Failure of the weakest link in a chain with many links
σt σ
with failure mechanisms (e.g., creep or fatigue) in each
where σ = 1/β is a scale parameter for log(T ) (shape param-
link acting approximately independent.
eter for T ), μ = log(η) is a location parameter for log(T ),
and Failure of a system with a large number of components in
series and with approximately independent failure mech-
φsev(z) = exp[z − exp(z)] anisms in each component.
Φsev(z) = 1 − exp[− exp(z)].
• A common justification for its use is empirical: the Weibull
Quantiles:
distribution can be used to model failure-time data with a
tp = exp μ + σΦ−1
sev (p) , t>0 decreasing or an increasing hazard function.
where Φ−1
sev (p) is the p quantile for the standard SEV (i.e., • When compared with the lognormal distribution, Weibull in-
μ = 0, σ = 1) distribution. ferences are conservative when extrapolating into either tail.
4 - 23 4 - 24
Largest Extreme Value Distribution
4 - 25 4 - 26
0 0.00
0 5 10 15 20 25 0 5 10 15 20 25
Notes:
t t
0.5
σ μ limy→∞ h(y; μ, σ) = 1/σ.
0.4
2 9
0.3 3 9
h (t ) 5 9
• If Y ∼ LEV(μ, σ) then −Y ∼ SEV(−μ, σ).
0.2
0.1
0.0
0 5 10 15 20 25
t
4 - 27 4 - 28
T ∼ FREC(η, β)
Cumulative Distribution Function Probability Density Function
ηβ
F (t; η, β) = Pr(T ≤ t) = exp − , 1 1.2
t 1.0
β
β η β η 0.8
f (t; η, β) = exp − F (t ) 0.5 f (t ) 0.6
t t t
0.4
β η β 0.2
η β 0 1 2 3 4 0 1 2 3 4
exp −1 t t
t
Hazard Function
where η > 0 is a scale parameter and β > 0 is a shape pa-
rameter. Quantiles: tp = η[− log(p)]−1/β . 1.5
β η
1 1
2 1
h (t ) 1.0 3 1
• If T ∼ FREC(η, β), then 1/T ∼ WEIB(1/η, β) 0.5
4 - 29 4 - 30
Fréchet Distribution- Continued
Moments: For integer m > 0, the mth Fréchet moment Alternative Fréchet Parameterization
E(T m) exists if and only if β > m. In particular,
⎧ ( ) If T ∼ FREC(η, β) then Y = log(T ) ∼ LEV(μ, σ). Then
⎪
⎪
⎨η m Γ 1 − m for β > m
E(T m) = β log(t) − μ
⎪
⎪ F (t; μ, σ) = Φlev
⎩∞ otherwise σ
where Γ(κ) is the gamma function. Then 1 log(t) − μ
f (t; μ, σ) = φlev , t>0
σt σ
E(T ) = ηΓ(1 − 1/β), β>1
where σ = 1/β, μ = log(η), and
Var(T ) = η 2 Γ(1 − 2/β) − Γ2(1 − 1/β) , β > 2.
Φlev(z) = exp[− exp(−z)]
φlev(z) = exp[−z − exp(−z)].
• The Fréchet distribution can be used to model the max-
imum of a large number of independent identically dis- Quantiles:
tributed (iid) positive random variables.
tp = exp μ + σΦ−1
lev (p)
• When extrapolating into either tail of a failure-time distri-
where Φ−1lev (p) is the p quantile for the standard LEV (i.e.,
bution, the Fréchet distribution provides more optimistic
μ = 0, σ = 1).
(anti-conservative) estimates relative to the lognormal or
the Weibull distribution.
4 - 31 4 - 32
1.0
φlogis and Φlogis are pdf and cdf for the standard logistic σ μ
0.8
distribution defined by 1 10
0.6 2 10
exp(z) h (t ) 5 10
0.4
φlogis(z) =
[1 + exp(z)]2 0.2
exp(z) 0.0
Φlogis(z) = . 0 5 10 15 20
1 + exp(z) t
4 - 33 4 - 34
Logistic Distribution-Continued
Loglogistic Distribution
Quantiles: yp = μ + σΦ−1
logis (p) = μ + σ log[p/(1 − p)], where If Y ∼ LOGIS(μ, σ) then T = exp(Y ) ∼ LOGLOGIS(μ, σ)
Φ−1
logis (p) = log[p/(1 − p)] is the p quantile for the standard with
logistic distribution.
log(t) − μ
F (t; μ, σ) = Φlogis
σ
E[(Y − μ) ]
4 - 35 4 - 36
Examples of Loglogistic Distributions
Loglogistic Distribution-Continued
Cumulative Distribution Function Probability Density Function
1 3.0
2.5 Quantiles: tp = exp μ + σΦ−1 σ
logis(p) = exp(μ)[p/(1 − p)] .
2.0
Moments: For integer m > 0,
F (t ) 0.5 f (t ) 1.5
1.0
E(T m) = exp(mμ) Γ(1 + mσ) Γ(1 − mσ).
0.5
0 0.0 The m moment is not finite when mσ ≥ 1.
0.0 0.5 1.0 1.5 0.0 0.5 1.0 1.5
t t
For σ < 1,
Hazard Function
4
σ μ E(T ) = exp(μ) Γ(1 + σ) Γ(1 − σ),
3 0.1 0.1
0.3 −1 and for σ < 1/2,
h (t ) 2
1.7 −2
1
Var(T ) = exp(2μ) Γ(1 + 2σ) Γ(1 − 2σ) − Γ2(1 + σ) Γ2(1 − σ) .
0
0.0 0.5 1.0 1.5
t
4 - 37 4 - 38
T ∼ GENG(μ, σ, λ)
⎧
⎪
⎪
⎪ Φlg λω + log(λ−2); λ−2 if λ>0
⎪
⎨
F (t; μ, σ, λ) = Φnorm(ω) if λ=0
⎪
⎪
⎪
⎪
⎩ 1 − Φlg λω + log(λ−2); λ−2 if λ<0
⎧
|λ|
⎨ σt φlg λω + log(λ−2); λ−2 if λ
⎪ = 0
Chapter 4
f (t; μ, σ, λ) =
⎪
⎩ 1
Segment 4 σt φnorm (ω) if λ = 0
where t > 0, ω = [log(t) − μ]/σ and
The Generalized Gamma Distribution Φlg(z; κ) = ΓI[exp(z); κ],
1
φlg(z; κ) = exp[κz − exp(z)].
Γ(κ)
−∞ < μ < ∞, exp(μ) is a scale parameter and −∞ < λ < ∞
and σ > 0 are shape parameters and ΓI(v; κ) is the incomplete
gamma function defined by
v xκ−1 exp(−x) dx
ΓI(v; κ) = 0 , v > 0.
Γ(κ)
4 - 39 4 - 40
• When λ = 0, E(T ) and Var(T ) are the same as that for the
lognormal distribution.
4 - 41 4 - 42
Generalized Gamma Distribution Special Cases
• If λ = 1, T ∼ WEIB(μ, σ).
Segment 4
• If λ = −1, T ∼ FREC(μ, σ).
Threshold Log-Location-Scale Distributions
• When λ = σ, T is has a Gamma distribution with a scale
parameter λ2 exp(μ) and a shape parameter λ−2.
4 - 43 4 - 44
4 - 47 4 - 48
Chapter 5
System Reliability Concepts and Methods
System Reliability Concepts and Methods
Chapter 5 Objectives
Definitions
5-3 5-4
System Structures and System Failure Probability For the time to first failure of a new system with s compo-
nents (all components starting a time 0)
System failure probability FT (t; θ ) is the probability that • The cdf for component i is Fi = Fi(t; θ i). The correspond-
a system fails before t (time to first failure for a repairable ing survival probability is Si = Si(t; θ i) = 1 − Fi(t; θ i). The
system). θ is may have some elements in common. Here θ denotes
the unique elements in (θ 1, . . . , θ s).
The failure probability of the system depends on:
• The cdf for the system is denoted by FT = FT (t; θ ). This cdf
• Time in operation (or other measure of use) denoted by t is determined by the Fi’s and the system structure. Then
• System structure. FT (t; θ ) = g[F1(t; θ 1), . . . , Fs(t; θ s)]
• Reliability of system components, including interconnec- or one of the simpler forms
tions, interfaces, and human operators.
FT (θ ) = g[F1(θ 1), . . . , Fs(θ s)]
• Environmental conditions. FT = g(F1, . . . , Fs).
To simplify the presentation, time-(and parameter)-dependency
will usually be suppressed in this chapter.
5-5 5-6
Examples of Series Systems
• Multi-cell battery
• Cell phone
5-7 5-8
5-9 5 - 10
• For a series system with two components and dependent • The distributions of log failure times for the individual com-
failure times ponents is bivariate normal with the same (arbitrary) mean
FT (t) = Pr(T ≤ t) = 1 − Pr(T > t) = 1 − Pr(T1 > t ∩ T2 > t). and standard deviation for both components and correla-
tion ρ.
In this case, the evaluation has to be done with respect to
the bivariate distribution of T1 and T2. • The reliability 1 − FT (t) of the system can be expressed as
a function of the individual reliability components 1 − F (t)
• If the correlation between the two components is positive, and ρ.
then the assumption of independence is conservative in the
sense that the actual FT (t) is smaller than that predicted • When ρ = 1 (so the two components are perfectly depen-
by the independent-component model. dent and will fail at exactly the same time), the system
reliability 1 − FT (t) is the same as the reliability for a single
• These results extend to the s components in series, the component.
system FT (t) would have to be computed with respect to
the underlying s-variate distribution. Such computations • When ρ = 0 (so the two components are independent),
are, in general, difficult. 1 − FT (t) corresponds to the system reliability for an s = 2
series system with independent components.
5 - 11 5 - 12
Reliability of a Series System
with 2 Identical Dependent Components
Chapter 5
0.95
Segment 2
0.85
ρ=1
0.9
0.7
0.80 0.4 0
5 - 13 5 - 14
A Parallel System with Two Components Examples of Systems with Components in Parallel
• Automobile headlights
1
• RAID array computer disk systems
5 - 15 5 - 16
= F1 F2
0.90
5 - 17 5 - 18
Effect of Positive Dependency in a Parallel System
Effect of Positive Dependency in a with Two Identical Components
Two-Component Parallel System Having Lognormal Failure Times
1.00
Two−Component Parallel−System Reliability
Chapter 5
0.95
Segment 3
0.85
0.80
0.5 0.6 0.7 0.8 0.9 1.0
Individual Component Reliability
5 - 21 5 - 22
Series and parallel structures are the basis for building mod- A Series-Parallel System Structure
els for more complicated systems which use redundancy to with Component-Level Redundancy
increase reliability.
5 - 23 5 - 24
Systems with Component-Level Redundancy
FT (t) = 1 − (1 − F r )k
5 - 25 5 - 26
5 - 27 5 - 28
Series-Parallel Structure
with System-Level Redundancy
where Fij , j = 1, 2 are the cdfs for the series system i. More Complicated System Structures
• For a r × k structure with independent components
⎡ ⎤
r
k
FT (t) = ⎣1 − 1 − Fij ⎦
i=1 j=1
A Bridge System Structure • Let Ai be the event that the i unit is working.
5 - 31 5 - 32
5 - 33 5 - 34
For a 2 out of 3 independent components where δ = (δ1, . . . , δs) with δi = 1 indicating failure of unit
i by time t and δi = 0 otherwise and Aj is the set of all δ
FT (t) = Pr(T ≤ t) such that δ δ = j.
= Pr(exactly two fail) + Pr(exactly three fail)
• FT (t) can also be viewed as the distribution of the sum of
= F1F2(1 − F3) + F1F3(1 − F2) + F2F3(1 − F1) + F1F2F3
s independent non-identically distributed Bernoulli random
= F1F2 + F1F3 + F2F3 − 2F1F2F3
variables, and is known as the Poisson-binomial distribution.
5 - 35 5 - 36
Other System Structures
5 - 37 5 - 38
5 - 39 5 - 40
Repairable Systems
5 - 41 5 - 42
Markov and Other More General Models
• The Markov models are also limited by the life and repair
distributions that can be employed.
5 - 43
Chapter 6
June 7, 2018
8h 15min 6-1 6-2
• The transformed axes are relabeled in terms of the original We plot tp on the horizontal axis and p on the vertical axis.
probability and time variables. γ is the intercept on the time axis and 1/θ is equal to the
slope of the cdf line.
The resulting probability axis is generally nonlinear and is
called the probability scale. The data axis is usually a linear
Note:
axis or a log axis.
Changing θ changes the slope of the line and changing γ
changes the position of the line.
6-5 6-6
Plot with Exponential Distribution Probability Scales
Showing Exponential cdfs as Straight Lines for Linearizing the Normal Distribution cdf
Combinations of Parameters θ = 50, 200 and γ = 0, 200
tp = γ − θ log(1 − p)
cdf: p = F (y; μ, σ) = Φnorm y−μ
σ , −∞ < y < ∞.
Quantiles : yp = μ + σΦ−1
norm (p).
Φ−1
.99
norm (p) is the p quantile of the standard normal distribution.
4
.98
θ, γ = 50, 0
Conclusion:
50, 200
.9 μ is the point at the time axis where the cdf intersects the
2
Φ−1(p) = 0 line (i.e., p = 0.5). The slope of the cdf line on
200, 200
.8
the graph is 1/σ.
.7
1
.6
Note:
.4
.2 Any normal distribution cdf plots as a straight line with a
.01 0
positive slope. Also, any straight line with positive slope
0 200 400 600 800 1000
corresponds to a normal cdf.
Time
6-7 6-8
Plot with Normal Distribution Probability Scales Linearizing the Lognormal Distribution cdf
Showing Normal distribution cdfs as Straight Lines for
Combinations of Parameters μ = 40, 80 and σ = 5, 10
yp = μ + σΦ−1
norm (p)
cdf: p = F (t; μ, σ) = Φnorm log(t)−μ
σ , t > 0.
Quantiles : tp = exp μ + σΦ−1
norm (p) .
.99
.98 2 Then log(tp) = μ + σΦ−1 norm (p)
.95 Conclusion:
.9 { log(tp) versus Φ−1
norm (p) } will plot as a straight line.
1
.8
Standard quantile
Proportion Failing
.7 The median exp(μ) can be read from the time axis at the
.6
.5 0
point where the cdf intersects the horizontal line Φ−1
norm (p) =
.4 0 (which corresponds to p = 0.50. The slope of the cdf
.3
line on the graph is 1/σ (but in the computations use base e
.2
80, 10
-1 logarithms for the times rather than the base 10 logarithms
.1
shown on the figures).
.05
μ, σ = 40, 10
40, 5 -2
.02 80, 5
Note:
.01
Any lognormal distribution cdf plots as a straight line with
0 20 40 60 80 100
a positive slope. Also, any straight line with positive slope
Time
corresponds to a lognormal distribution.
6-9 6 - 10
.99
.98 2 cdf: p = F (t; μ, σ) = Φsev log(t)−μ
σ , t > 0.
.95
.9 Quantiles : tp = exp μ + σΦ−1
sev (p) = η[− log(1 − p)]
1/β ,
1
where Φ−1
sev (p) = log[− log(1 − p)], η = exp(μ), β = 1/σ.
.8
Standard quantile
Proportion Failing
.7
.6
.5 0
.4
This leads to
.3
500, 2 1
.2 log(tp) = μ + σ log[− log(1 − p)] = log(η) + log[− log(1 − p)]
-1 β
exp(μ), σ = 50, 2
.1
Conclusion:
.05
50, 1 500, 1
-2
{ log(tp) versus log[− log(1 − p)] } will plot as a straight line.
.02
.01
Time
6 - 11 6 - 12
Plot with Weibull Distribution Probability Scales
Showing Weibull cdfs as Straight Lines for
Linearizing the Weibull Distribution cdf-Continued Combinations of η = 50, 500 and β = 0.5, 1
log(tp) = log(η) + β1 log[− log(1 − p)]
Comments:
0.0 0.5 1.0 1.5 2.0 2.5 3.0
Standard quantile
500, 0.5
Proportion Failing
the computations use base e logarithms for the times rather η, β = 50, 0.5 -1
.3
than the base 10 logarithms used for the figures). .2
equal to 1. .01
Time
6 - 13 6 - 14
6 - 15 6 - 16
0.4
• Display of maximum likelihood results with data.
0.2
• Estimation of parameters.
0.0
1+ , .9
pi = F t(i) + Δ + F t(i) − Δ . .7
0
2 .5
.3
.2
Standard quantile
Proportion Failing
−2
.1
.03
.02 −4
.01
• When the model fits well, the ML line approximately goes
.005
through the points. .003
−6
.001
• Need to adjust these plotting positions when there are ties. 6000 8000 10000 12000 14000 16000 20000 24000 28000
Kilometers
6 - 19 6 - 20
.9
1 + ,
.8
Plotting Positions: ti versus i−0.5
n .
.7
.6
.5 0 • Justification:
Standard quantile
1+ ,
Proportion Failing
i − 0.5
.4
.3
= F t + Δ + F t(i) − Δ
.2
n 2 (i)
−1
i − 0.5
.1
E t(i) ≈ F −1 .
.05 n
−2
.02 where Δ is positive and small.
.01
.005
Kilometers
6 - 21 6 - 22
1
Chapter 6
Segment 3 0.8
0.6
0.2
Thousand Cycles
6 - 23 6 - 24
Weibull Probability Plot for the Alloy T7987 Fatigue Lognormal Probability Plot for the Alloy T7987
Life and Simultaneous Approximate 95% Confidence Fatigue Life and Simultaneous Approximate 95%
Bands for F (t) Confidence Bands for F (t)
1.9 2.0 2.1 2.2 2.3 2.4 2.5 1.9 2.0 2.1 2.2 2.3 2.4 2.5
.99 .99
.98
.98 2
.9
.95
.7
0 .9
.5
1
.8
.3
.7
.2
Standard quantile
Standard quantile
.6
Proportion Failing
Proportion Failing
−2 .5 0
.1
.4
.05 .3
.03 .2
−1
.02 −4 .1
.01 .05
.005 .02 −2
.003 .01
−6 .005
.002
−3
.001 .001
80 100 120 140 160 180 200 220 240 280 320 360 80 100 120 140 160 180 200 220 240 280 320 360
6 - 25 6 - 26
Plotting Positions: Interval Censored Inspection Data Weibull Distribution Probability Plot of the
Heat-Exchanger Tube Crack Data and Simultaneous
Approximate 95% Confidence Bands for F (t)
• Let (t0, t1], . . . , (tm−1, tm] be the inspection times.
−
pi = F (ti) .05
Proportion Failing
−
• When there are no censored observations beyond tm,
−
F (tm) = 1 and this point cannot be plotted on probability .01
paper.
Years
6 - 27 6 - 28
.2
− Chapter 6
.15
−
.1
Segment 4
−
.05
Using Simulation to Calibrate
Proportion Failing
−
.01
.001
1.0 1.2 1.4 1.6 1.8 2.0 2.2 2.4 2.6 2.8 3.0
Years
6 - 29 6 - 30
Random Normal Variates Plotted on Normal Random Exponential Variates Plotted on Normal
Probability Plots with Sample Sizes of n=10, 20, and Probability Plots with Sample Sizes of n=10, 20, and
40. Five Replications of Each Probability Plot 40. Five Replications of Each Probability Plot
-0.5 0.5 1.5 -2 -1 0 1 2 -1 0 1 2 -1.5 -0.5 0.5 -0.5 0.5 0 1 2 3 4 5 0.0 1.0 2.0 0 1 2 3 0.0 1.0 2.0 3.0 0.0 0.4 0.8 1.2
-2 -1 0 1 -2 -1 0 1 -2 -1 0 1 2 -2 -1 0 1 2 -1.0 0.0 1.0 0.0 1.0 0.0 1.0 2.0 0.0 1.0 2.0 0 1 2 3 4 0.5 1.5 2.5
6 - 31 6 - 32
6 - 33 6 - 34
Row
1
2 39
52
3
4 46
• Field data from 2256 systems in the field; staggered entry– 5
6
7
8
31
48
102
9
10
11 158
multiple censoring. 12
13
14
2
312
101
15
16 2
101
17
18 2
19
20 100
9
21
22
23
24 100
23
25
• Unexpected failures. 26
27
28
29
30
2
56
27
55
31
32 20
33
34 56
35
36 22
37
38 55
39
40 22
• What is going on??
41
42 56
11
43
44
45
46 53
11
47
48 55
20
49
50 55
51
52 8
55
53
54 4
55
55
56 2
152
• The Weibull probability plot suggest changes in the failure 57
58
59
60
3
152
3
6 - 35 6 - 36
Bleed System
Bleed System (All Bases)
Separate Weibull Probability Plots
Weibull Probability Plot
for Base D and Other Bases
.05
.03 .1 DBase
.02 OtherBase
.05
.01
.03
.005 .02
.003
Fraction Failing
.01
Fraction Failing
.001
.005
.0005 .003
.0002
.0001 .001
.00005 .0005
.0003
.00002
.0002
.00001
.0001
20 50 100 200 500 1000 2000 5000
20 50 100 200 500 1000 2000 5000
Hours
Hours
6 - 37 6 - 38
Bleed System
Failure Data Analysis-Conclusions Transmitter Vacuum Tube Data
(Davis 1952)
• Separate analyses of the Base D data and the data from Days
the other bases indicated different failure distributions.
Interval Endpoint Number
• The large slope (β ≈ 5) for Base D indicated strong wearout. Lower Upper Failing
• The relatively small slope for the other bases (β ≈ 0.85) 0 25 109
suggested infant mortality or accidental failures. 25 50 42
50 75 17
• The problem at base D was caused by salt air. A change in 75 100 7
maintenance procedures there solved the main part of the 100 ∞ 13
reliability problem with the bleed systems.
6 - 39 6 - 40
.98
0.5 −
1 ? | 109
−
1.0
1.0
.9
2 | ? | 42 −
−
1.5 0.5
Standard quantile
−
3 | ? | 17
.7 − −
2.0
4 | ? | 7 0.0
2.5 .5
5 13
− −0.5
3.0
.3 −1.0
0 20 40 60 80 100 3.5
20 30 40 50 60 70 80 90 100
Days
Days
6 - 41 6 - 42
Chapter 7 Chapter 7
Parametric Likelihood Fitting Concepts:
Parametric Likelihood Fitting Concepts:
Exponential Distribution Exponential Distribution
Objectives
W. Q. Meeker, L. A. Escobar, and F. G. Pascual • The use of likelihood and Wald methods of computing con-
Iowa State University, Louisiana State University, and Wash- fidence intervals for model parameters and other quantities
ington State University. of interest.
Copyright 2018 W. Q. Meeker, L. A. Escobar, and F. G. Pas- • The appropriate use of the density approximation for obser-
cual. vations reported as exact failures.
Based on the authors’ book Statistical Methods for Relia- • The effect that sample size has on confidence interval width
bility Data, Second Edition, John Wiley & Sons Inc., to be and the likelihood shape.
published in 2019.
• How to make exponential distribution inferences with zero-
June 7, 2018 failures.
8h 15min 7-1 7-2
• n =10,220 observations.
Likelihood for Interval-Censored Data
Times Between α-Particle Arrivals • Data binned into intervals from 0 to 4000 time
units. Interval sizes ranging from 25 to 100 units. Ad-
ditional interval for observed times exceeding 4,000 time
units.
7-3 7-4
Interarrival Times
Time Frequency of Occurrence
0 100 1609 41 3
Frequency
7-5 7-6
Exponential Probability Plot of the n = 200 Sample of Parametric Likelihood
α-Particle Interarrival Time Data. The Plot also Probability of the Data
Shows Approximate 95% Simultaneous Nonparametric
Confidence Bands.
• Using the model Pr(T ≤ t) = F (t; θ ) for continuous T ,
the likelihood (probability) for a single observation in the
interval (ti−1, ti] is
.99
Li(θ ; datai) = Pr(ti−1 < T ≤ ti) = F (ti; θ ) − F (ti−1; θ ).
−
−
.9 suming n independent observations
−
n
.8 − L(θ ) = L(θ ; DATA) = C Li(θ ; datai).
−
.7 i=1
−
.6 −
−
.5
.4
.3 − −
−
• As explained in Chapter 2, we take C = 1.
.2
.1 −
1/5000 Seconds
7-7 7-8
1.2
• The exponential distribution cdf is Estimate of the mean using all n=10220 times →
t
F (t; θ) = 1 − exp −
, t > 0. 1.0
θ
Relative Likelihood R (θ)
Confidence Level
0.60
• The interval-data likelihood for the α-particle data is
0.6 0.70
200
200
L(θ) = Li(θ) = F (ti; θ) − F (ti−1; θ) 0.80
0.4 n=200 →
i=1 i=1
8
d 8 "
tj−1
tj
#d
j 0.90
= F (tj ; θ) − F (tj−1; θ) = exp − − exp − 0.2
j=1 j=1 θ θ 0.95
(i.e., the number of times between tj−1 and tj ). 200 300 400 500 600 700 900
Mean Time Between Arrivals θ
7-9 7 - 10
.99
.98
Chapter 7
Segment 2
.95
.9
and Methods for Confidence Intervals
.8
.7
.6
.5
.4
.3
.2 θ^ = 572.3
.1
1/5000 Seconds
7 - 11 7 - 12
Likelihood as a Tool for Modeling and Inference Likelihood as a Tool for Modeling and Inference
(Continued)
What can we do with the (log) likelihood?
n
• Regions of high likelihood are credible; regions of low likeli-
L(θ ) = log[L(θ )] = Li(θ ).
hood are not credible (suggests confidence regions for pa-
i=1
rameters).
7 - 13 7 - 14
Large-Sample Approximate Theory for Likelihood for the n = 200 α-Particle Interarrival
R(θ) = L(θ)/L(θ)
Ratios for a Scalar Parameter Time Data. Vertical Lines Give an Approximate 95%
Likelihood-Based Confidence Interval for θ
• Relative likelihood for θ is
L(θ)
R(θ) =
.
L(θ)
1.2
Estimate of the mean using all n=10220 times →
• If evaluated at the true θ, then, asymptotically, −2 log[R(θ)]
follows, a chi-square distribution with 1 degree of freedom. 1.0
0.8 0.50
region for θ is the set of all values of θ such that
Confidence Level
0.60
−2 log[R(θ)] < χ2
(1−α;1) 0.6 0.70
or, equivalently, the set defined by
0.4 n=200 →
0.80
R(θ) > exp −χ2
(1−α;1)/2 .
0.90
• A one-sided approximate 100(1 − α)% lower or upper confi- 0.2
0.95
dence bound is obtained by replacing 1 − α with 1 − 2α and
0.99
using the appropriate endpoint. 0.0
200 300 400 500 600 700 900
Mean Time Between Arrivals θ
• General theory in the Appendix.
7 - 15 7 - 16
7 - 17 7 - 18
Comparison of Confidence Intervals for the α-Particle
Confidence Intervals for Functions of θ
Data
All Times Sample of Times
• For one-parameter distributions, confidence intervals for n =10,220 n = 200 n=20
θ can be translated directly into confidence intervals for
monotone functions of θ. Mean Time Between Arrivals θ
ML Estimate θ 596 572 440
Standard Error se 6.1 41.7 101
θ
• The arrival rate λ = 1/θ is a decreasing function of θ. 95% Confidence Intervals
for θ Based on
[λ, λ̃] = [1/θ̃, 1/θ ] Likelihood [585, 608] [498, 662] [289, 713]
Zlog(
θ)
∼
˙ NORM(0, 1) [585, 608] [496, 660] [281, 690]
= [1/662, 1/498] = [.00151, .00201]. Z ∼˙ NORM(0, 1) [585, 608] [491, 654] [242, 638]
θ
7 - 19 7 - 20
7 - 21 7 - 22
ML Estimates for the Exponential Distribution Mean Confidence Interval for the Mean Life of a New
Based on the Density Approximation Insulating Material
• Using the observed curvature in the likelihood: 1.15, 3.16, 10.38, 10.75, 12.53, 16.74, 22.54, 25.01, 33.02,
3 33.93, 36.17, 39.06, 44.56, 46.65, 55.93
−1
d2L(θ) θ2 θ
seθ = − 2
= =√ . Then TTT = 1.15 + · · · + 55.93 + 10 × 55.93 = 950.88 hours.
dθ r r
θ
• The ML estimate of θ and a 95% confidence interval are:
• If the data are complete or failure censored, 2TTT /θ ∼ χ2
2r . θ = 950.88/15 = 63.392 hours
Then an exact 100(1 − α)% confidence interval for θ is ⎡ ⎤
" #
⎡ ⎤ 2(950.88) 2(950.88) 1901.76 1901.76
2(TTT ) 2(TTT ) ⎦ θ , θ̃ = ⎣ 2 , ⎦= ,
[θ , θ̃] = ⎣ 2 , . χ(0.975;30) χ2
(0.025;30)
46.98 16.79
χ(1−α/2;2r) χ2
(α/2;2r) = [40.48, 113.26].
7 - 23 7 - 24
Example. α-Particle Pseudo Data Constructed
with Constant Proportion within Each Bin
Interarrival Times
Time Frequency of Occurrence
Interval Endpoint Samples of Times
Chapter 7
lower upper n=20000 n=2000 n=200 n=20
Segment 4 tj−1 tj dj
0 100 3000 300 30 3
Effect of Sample Size on Confidence Interval Width 100 300 5000 500 50 5
and the Likelihood Shape 300 500 3000 300 30 3
500 700 3000 300 30 3
700 1000 2000 200 20 2
1000 2000 3000 300 30 3
2000 4000 1000 100 10 1
4000 ∞ 0000 000 0 0
20000 2000 200 20
7 - 25 7 - 26
1.2 1.0
← Estimate using all
n = 20,000 times
1.0
0.8 0.50
Relative Likelihood R (θ)
0.60
Relative Likelihood R (θ)
Confidence Level
0.8 0.50
Confidence Level
0.6 0.70
0.60
0.6 0.70
0.80
n=20 → 0.4
0.80
0.4
← n=2,000 0.90
0.90 0.2
n=200 →
0.2 0.95
0.95
0.99
0.99 0.0
0.0
0 10 20 30 40 50
300 400 500 600 700 900 1100 1300 1600 θ
Mean Time Between Arrivals θ
7 - 27 7 - 28
0.8 0.50
• In large samples, the log-likelihood can be approximated
Relative Likelihood R (θ)
0.60
well by a quadratic function.
Confidence Level
0.6 0.70
7 - 29 7 - 30
Exponential Distribution Inferences with Zero Failures
Analysis of the Diesel Generator Fan Data Relative Likelihood for the Diesel Generator Fan Data
After 200 Hours of Service After 200 Hours of Service and Zero Failures
• Suppose that all fans were removed from service after 200
hours of operation at which time all of the 70 fans were still
running (zero failures). 1.0
0.60
• Again, using the entire data set, the upper confidence bound
on F (10000; θ) is F (10000) = F (10000; θ ) = 1−exp(−10000/18485
0.4178. This shows how little information is available from
a short test with few or zero failures.
7 - 35
Chapter 8 Chapter 8
Maximum Likelihood
Maximum Likelihood for Log-Location-Scale Distributions
for Log-Location-Scale Distributions Objectives
W. Q. Meeker, L. A. Escobar, and F. G. Pascual • How to construct and interpret likelihood confidence in-
Iowa State University, Louisiana State University, and Wash- tervals for parameters and for functions of parameters of
ington State University. log-location-scale distributions.
Copyright 2018 W. Q. Meeker, L. A. Escobar, and F. G. Pas-
cual. • The construction of Wald (normal-approximation) confi-
dence intervals for parameters and functions of parameters
Based on the authors’ book Statistical Methods for Relia- of log-location-scale distributions.
bility Data, Second Edition, John Wiley & Sons Inc., to be
published in 2019.
• Inference for log-location-scale distribution parameters and
June 7, 2018 functions of parameters with a given shape parameter.
8h 15min 8-1 8-2
.98
.9
.7
0
Chapter 8 .5
.3
Segment 1 .2
Standard quantile
Proportion Failing
−2
.1
.03
.02 −4
.01
.005
.003
−6
.001
Kilometers
8-3 8-4
• The likelihood has the form • The likelihood has the form
n
n
L(μ, σ) = Li (μ, σ; datai ) L(μ, σ) = Li (μ, σ; datai )
i=1 i=1
n
δ 1−δi
n
δ 1−δi
= [f (ti ; μ, σ)] i [1 − F (ti ; μ, σ)] = [f (ti ; μ, σ)] i [1 − F (ti ; μ, σ)]
i=1 i=1
n " #δi " #1−δi n " #δi " #1−δi
1 log(ti ) − μ log(ti ) − μ 1 log(ti ) − μ log(ti ) − μ
= φsev × 1 − Φsev = φnorm × 1 − Φnorm
i=1
σti σ σ i=1
σti σ σ
4 4
1 if ti is an exact observation 1 if ti is an exact observation
δi = δi =
0 if ti is a right censored observation 0 if ti is a right censored observation
φsev(z) is the standardized smallest extreme value density. φnorm(z) is the standardized normal density.
8-5 8-6
Weibull Relative Likelihood Weibull Relative Likelihood
for the Shock Absorber Data for the Shock Absorber Data
ML Estimates: μ = 10.23 and σ = .3164 ML Estimates: μ = 10.23 and σ = .3164
log(σ
R(μ, log(σ)) = L(μ, log(σ))/L(μ, )) σ
R(μ, σ) = L(μ, σ)/L(μ, )
0.7
0.6
0 0.2 0.4 0.6 0.8 1
Relative Likelihood
0.5 0.1
0.01
0.2
0.4
0.4
0.7
0.9
σ
0.3
10
.8
10
.6
10
.4
10 -1.6
μ .2
-1.2
-1.4
10 -1 0.2
-0.8
-0.6 log(σ)
0.001 0.001
μ
8-7 8-8
Weibull Probability Plot of Shock Absorber Failure Lognormal Probability Plots of Shock Absorber Data
Times (Both Failure Modes) with Maximum with ML Estimates and Wald 95% Pointwise
Likelihood Estimates and Wald 95% Pointwise Confidence Intervals for F (t). The Curved Line is the
Confidence Intervals for F (t) Weibull ML Estimate.
.9 0.95
0.9
.7
0.8
.5
0.7
.3 0.6
0.5
.2
0.4
0.3
.1
Proportion Failing
Proportion Failing
0.2
.05 0.1
.03 0.05
.02 0.02
0.01
.01 0.005
0.002
.005 0.001
0.0005
.003 ^ = 27719
η Lognormal Distribution ML Fit
Weibull Distribution ML Fit
0.0001
0.00005 Lognormal Distribution 95% Pointwise Confidence Intervals
^
β = 3.16
.001 0.00001
6000 8000 10000 12000 14000 16000 20000 24000 28000 5000 10000 15000 20000 25000 30000 35000
Kilometers Kilometers
8-9 8 - 10
8 - 11 8 - 12
Weibull Relative Likelihood Weibull Likelihood-Based Joint Confidence Regions for
for the Shock Absorber Data μ and σ for the Shock Absorber Data
ML Estimates: μ = 10.23 and σ = .3164 ML Estimates: μ = 10.23 and σ
= .3164
σ
R(μ, σ) = L(μ, σ)/L(μ, ) R(μ, σ) > exp −χ2
(1−α;2)
/2 = α gives a 100(1 − α)% region
0.7 0.7
0.6 0.6
95
0.5 0.1 0.5 90
0.01 99
0.2 70 80
0.4 60
0.4 0.4 50
0.7
25
0.9
σ σ
0.3 0.3
0.2 0.2
0.001 0.001
9.8 10.0 10.2 10.4 10.6 10.8 9.8 10.0 10.2 10.4 10.6 10.8
μ μ
8 - 13 8 - 14
Weibull Profile Likelihood R(μ) (exp(μ) = η =≈ t0.63) Weibull Profile Likelihood R(σ) (σ = 1/β)
for the Shock Absorber
Data for the Shock Absorber
Data
R(μ) = max
σ
L(μ,σ)
,σ
L(μ )
R(σ) = max
μ
L(μ,σ)
,σ
L(μ )
1.0 1.0
Confidence Level
Profile Likelihood
Profile Likelihood
0.60 0.60
0.80 0.80
0.4 0.4
0.90 0.90
0.2 0.2
0.95 0.95
0.99 0.99
0.0 0.0
9.8 10.0 10.2 10.4 10.6 10.8 0.2 0.3 0.4 0.5 0.6
μ σ
8 - 15 8 - 16
Need: Inferences on subset θ 1, from the partition θ = (θ 1, θ 2). • An approximate 100(1 − α)% likelihood-based confidence
region for θ 1 is the set of all values of θ 1 such that
• General theory in the Appendix. • Can improve the asymptotic approximation with simulation
(only small effect except in very small samples).
8 - 17 8 - 18
Confidence Regions and Intervals for
Functions of μ and σ
8 - 19 8 - 20
0.6
L(tp, σ).
σ
0.4
t0.1
8 - 21 8 - 22
1.0
1.0
0.8 0.50
0.8 0.50
Confidence Level
Profile Likelihood
0.60
Confidence Level
Profile Likelihood
0.60
0.6 0.70 0.6 0.70
0.80 0.80
0.4 0.4
0.90 0.90
0.2 0.2
0.95 0.95
0.99
0.99 0.0
0.0
0.00 0.02 0.04 0.06 0.08 0.10 0.12 0.14
6000 8000 10000 12000 14000 16000 18000 20000
F(10000)
t0.1
8 - 23 8 - 24
Large-Sample Approximation Theory of ML Estimation
8 - 25 8 - 26
• Alternative asymptotic theory is based on the large-sample • An Wald approximate 100(1 − α)% confidence region for θ
distribution of quadratic forms (Wald’s statistic). is the set of all values of θ in the ellipsoid
−1
− θ ) Σ̂
(θ − θ ) ≤ χ2
(θ
θ (1−α;k).
• Let Σ̂ be a consistent estimator of Σ , the asymptotic
θ θ
. For example,
covariance matrix of θ
• This is sometimes known as the normal-approximation con-
−1
∂ 2L(θ ) fidence region.
Σ̂ = −
θ ∂ θ∂ θ
.
where the derivatives are evaluated at θ • Can specialize to functions or subsets of θ .
• Asymptotically, the Wald statistic • Wald confidence intervals are not transformation invariant.
−1
− θ ) Σ̂ − θ) Thus there are multiple ways to compute a Wald interval.
w(θ ) = (θ (θ
θ
when evaluated at the true θ , follows a chi-square distri-
bution with k degrees of freedom, where k is the length • Can try to find a transformation that results in a log-
of θ . likelihood with approximate quadratic shape.
8 - 27 8 - 28
g =
se ! g1) =
Var( ! μ
Var( ) + ! σ
Var( ) + 2 , σ
Cov(μ )
1
∂μ ∂σ ∂μ ∂σ
)−log(σ)]/s
• Assuming that Zlog(σ) = [log(σ elog(σ) ∼
˙ NORM(0, 1)
an approximate 100(1 − α)% confidence interval for σ is • Partial derivatives evaluated at μ,
σ.
[σ , /w,
σ̃] = [σ × w]
σ • General theory in the appendix.
*
where w = exp z(1−α/2)seσ /σ
and s
eσ = ! σ
Var( ).
8 - 29 8 - 30
Wald Confidence Interval for F (te; μ, σ)–Continued
F = F (te; μ,
σ ) = Φ(ze ) • The asymptotic normality of ze = [log(te) − μ]/
σ
• Expressions for seF and seze are obtained by using the delta
method.
8 - 31 8 - 32
Comments:
• Continuous-run test for a newly designed bearing.
8 - 33 8 - 34
Bearing-A Weibull Probability Plot and ML Estimate Bearing-A Weibull Probability Plot and ML Estimate
.9 .9
.7 .7
.5 .5
.3 .3
Proportion Failing
Proportion Failing
.2 .2
.1 .1
.05 .05
.03 .03
.02 .02
^ = 3417
η ^ = 3417
η
.01 .01
^ ^
β = 0.984 β = 0.984
.005 .005
200 300 400 500 600 700 900 1100 200 300 400 500 600 700 900 1100
Thousands of Cycles Thousands of Cycles
8 - 35 8 - 36
Bearing-A Life Test Example Conclusions
Segment 5
• When the ML estimate does not go through the points, it
is an indication that the Weibull distribution does not agree Weibull Distribution Inference
with the data. with Few Failures
• It is important to find out the reason that the line does not
fit.
8 - 37 8 - 38
• Suppose that β is given. Knowledge of the failure mecha- • n = 1703 units had been introduced into service over time;
nism will often provide information about β. oldest unit had 2220 hours of operation.
• Simplifies problem. Only one parameter with r failures and • 6 units had failed observed.
t1, . . . , tn failures and censor times
⎛
⎞1/β 3
β
nt η̂ 1 • Design life specification was B10=t0.1 = 8000 hours.
η̂ = ⎝ i=1 i ⎠ , seη̂ = .
r β r
• Requires sensitivity analysis because β is unknown. • How many replacement parts will be needed?
8 - 39 8 - 40
Row
1 288 .9
2 148
3 .7
4 124 .5
5
6 111 .3
7 .2
8 106
9 99 .1
10 110
.05
Proportion Failing
11 114
12 119 .03
13 127 .02
14
15 .01
16 123 .005
17 93
18 47 .003
19 41
20 27
21
.001
22 11 .0005
23 6
24 .0003
25 2 .0002
^ = 11792
η
.0001
.00005 ^
β = 2.035
0 500 1000 1500 2000 .00003
Hours 100 200 500 1000 2000 5000 10000
Hours
8 - 41 8 - 42
Weibull Probability Plots Bearing-Cage Fracture Data Weibull Probability Plots Bearing-Cage Fracture Data
with Weibull ML Estimates and Sets of 95% Pointwise with Weibull ML Estimates and Sets of 95% Pointwise
Confidence Intervals for F (t) with Given β = 1.5. Confidence Intervals for F (t) with Given β = 2.
.9 .9
.7 .7
.5 .5
.3 .3
.2 .2
.1 .1
.05 .05
Proportion Failing
Proportion Failing
.03 .03
.02 .02
.01 .01
.005 .005
.003 .003
.001 .001
.0005 .0005
.0003 .0003
.0002 .0002
^ = 28987
η ^ = 12320
η
.0001 .0001
.00005 β = 1.5 .00005 β=2
.00003 .00003
100 200 500 1000 2000 5000 10000 100 200 500 1000 2000 5000 10000
Hours Hours
8 - 43 8 - 44
Weibull Probability Plots Bearing-Cage Fracture Data Lognormal and Weibull Comparison
with Weibull ML Estimates and Sets of 95% Pointwise Bearing-Cage Fracture Field Data
Confidence Intervals for F (t) with Given β = 3. Lognormal Probability Paper
.9 .8
.7 .7
.5 .6
.3 .5
.2 .4
.1 .3
.05 .2
Proportion Failing
.03
Proportion Failing
.02 .1
.01 .05
.005 .02
.003
.01
.005
.001
.002
.0005 .001
.0003 .0005
.0002
^ = 5408
η Lognormal Distribution ML Fit
.0001 .0001 Weibull Distribution ML Fit
.00005 Lognormal Distribution 95% Pointwise Confidence Intervals
.00005 β=3
.00003 .00001
100 200 500 1000 2000 5000 10000 100 200 500 1000 2000 5000 10000
Hours Hours
8 - 45 8 - 46
1.0
14
12 0.8
0.1
10
0.1
0.6
sigma
sigma
6 0.4
0.2
0.2
4
0.3
0.3 0.2
2 0.5 0.5
0.9 0.9
0 0.0
mu mu
8 - 47 8 - 48
Weibull Distribution with Given β
and Zero Failures
8 - 49 8 - 50
.2
eter is near β = 2, and almost certainly between 1.5 and 2
2.5. .1
3
• Many copies of a newly designed component were put into
.05
.03
service during the past year and no failures have been re- .02
4
ported. β = 1.5
.01
5 .005 β = 2.0
Hours: 500 1000 1500 2000 2500 3000 3500 4000 .003
β = 2.5
Number of Units: 10 12 8 9 7 9 6 3 6
7 .0005
• Can the replacement time be safely increased from 2000 500 1000 1500 2000 2500 3000 3500 4500
8 - 51 8 - 52
Regularity Conditions
• Can exchange the order of differentiation of log likelihood • Multiple failure modes (Chapter 15).
w.r.t. θ and integration w.r.t. data.
Analysis of Data with More than One Failure Mode Analysis of Data with More than One Failure Mode
Chapter 15 Objectives
June 7, 2018
8h 16min 15 - 1 15 - 2
Basic Ideas Behind the Analysis of Data with Multiple Multiple Failure Modes Data and
Failure Modes the Series-System Model
• There is a joint distribution of failure times for all failure • Let Tj be the lifetime from failure mode j, j = 1, . . . J.
modes.
• The failure-time of the system is
• When failure times for the different failure modes are in-
dependent, it is possible to analyze the individual failure
modes separately.
Chapter 15
• Create a separate data set, with n cases, for each failure
mode. Segment 2
15 - 7 15 - 8
Device-G Data
Device-G Background
15 - 9 15 - 10
Row
.98
1 W
2 S
3 W
.9
4 S
5 W
6 S
7 S .7
8 S
9
10 S .5
11 S
12
13
Proportion Failing
.3
14 W
15
16 .2
17
18 S
19 S
20 W .1
21 S
22 S
23 S
24 S .05
25
26 S
.03
27
28 S ^ = 242.6
η
29 W .02
30 W
^
β = 0.9268
.01
0 50 100 150 200 250 300
1 2 5 10 20 50 100 200 500
Kilocycles
Kilocycles
15 - 11 15 - 12
Weibull Analyses of Device-G Data
Individual Failure Modes
Series-System Model for Device G
• Let TS be the lifetime from the electrical surge failure mode .98 Surge
Wearout
(S) and TW be the lifetime from the wearout failure mode (W). .9
.5
T = min(TS, TW)
Proportion Failing
.3
.2
• When the failure modes are independent, the cdf for the
failure-time is .1
.02
= 1 − [1 − FS(t)][1 − FW(t)].
.01
Kilocycles
15 - 13 15 - 14
Weibull Analyses of Device-G Data Weibull Analyses of Device-G Data Estimating Time
Individual Failure Modes to Failure Using Series System Model
2 8
.7 .7
.5 .5
Proportion Failing
Proportion Failing
.3 .3
.2 .2
.1 .1
.05 .05
.03 .03
.02 .02
.01 .01
Kilocycles Kilocycles
15 - 15 15 - 16
• The model ignoring the cause of failure information is S & W μSW 5.49 0.23 5.04 5.94
( )
log(t) − μSW σSW 1.08 .21 0.74 1.57
F (t) = Φsev .
σSW For Mode S alone, LS = −101.36 for Mode W alone, LW =
−47.16, and for both modes together, LSW = −142.62.
15 - 17 15 - 18
Some Comments on the
Weibull Analyses of Device-G Data
19 1
20 1
• Need to estimate the failure-time distribution for the shock 21
22
23
24
absorbers. 25
26
27 2
28
29
30
31 2
32 1
33
• One might be interested in the failure-time distribution for 34 1
35
36 1
37
38
Weibull Analyses of Shock Absorber Data Estimating Weibull Analyses of Shock Absorber Data
Time to Failure Ignoring the Cause of Failure Individual Failure Modes
.9 .9 Mode1
Mode2
.7 .7
.5 .5
.3 .3
.2 .2
.1 .1
Proportion Failing
Proportion Failing
.05 .05
.03 .03
.02 .02
.01 .01
.005 .005
.003 ^ = 27719
η .003
^
β = 3.16
.001 .001
6000 8000 10000 12000 14000 16000 20000 24000 28000 6000 8000 10000 12000 14000 16000 20000 24000 28000
Kilometers Kilometers
15 - 23 15 - 24
Weibull Analyses of Shock Absorber Data Weibull Analyses of Shock Absorber Data Estimating
Individual Failure Modes Time to Failure Using Series System Model
.9 Mode1 .9 Mode1
Mode2 Mode2
.7 .7 Combined
.5 .5
.3 .3
.2 .2
.1 .1
Proportion Failing
Proportion Failing
.05 .05
.03 .03
.02 .02
.01 .01
.005 .005
.003 .003
.001 .001
6000 8000 10000 12000 14000 16000 20000 24000 28000 6000 8000 10000 12000 14000 16000 20000 24000 28000
Kilometers Kilometers
15 - 25 15 - 26
Weibull Analyses of Shock Absorber Data Estimating Weibull Analyses of Shock Absorber Data Estimating
Time to Failure Ignoring the Cause of Failure Time to Failure Using Series System Model
.9 .9 Mode1
Mode2
.7 .7 Combined
.5 .5
.3 .3
.2 .2
.1 .1
Proportion Failing
Proportion Failing
.05 .05
.03 .03
.02 .02
.01 .01
.005 .005
.003 ^ = 27719
η .003
^
β = 3.16
.001 .001
6000 8000 10000 12000 14000 16000 20000 24000 28000 6000 8000 10000 12000 14000 16000 20000 24000 28000
Kilometers Kilometers
15 - 27 15 - 28
There is a need to predict causes of future failures to as- Other Topics Related to Multiple Failure Modes
sure that a sufficient number of appropriate replacement
parts are available.
When failure modes are not identified or are only partially • The common assumption of independent failure modes is
identified for some units, it is still possible to estimate the sometimes unrealistic.
individual Fi(t) distributions by using maximum likelihood.
When there is dependence, still one can use the relationship
! F̂ ) = ∂FT ∂FT
• A combination of external information and sensitivity anal- Var( T Σ̂
∂θ θ ∂θ
ysis can be used to help specify a copula. .
where the derivatives are evaluated at θ
W. Q. Meeker, L. A. Escobar, and F. G. Pascual • Describe simple regression models to relate life to explana-
Iowa State University, Louisiana State University, and Wash- tory variables.
ington State University.
Copyright 2018 W. Q. Meeker, L. A. Escobar, and F. G. Pas- • Illustrate the use of likelihood methods for censored regres-
cual. sion data.
Segment 1
• Information from the Unix uptime command.
Introduction to Failure-Time Regression
16 - 3 16 - 4
Scatter Plot of Computer Program Execution Time Explanatory Variables for Failure Times
Versus System Load
Useful explanatory variables explain/predict why some units
fail quickly and some units survive a long time.
600
• Discrete variables like number of hardening treatments or
500
Seconds
300 •
• • Categorical variables like manufacturer, design, and loca-
• ••
200 tion.
•
100 ••• ••• •• • •
16 - 7 16 - 8
• 4
200 • 1 exact observation
• δi =
0 right censored observation
•
• •
• • • φnorm(z) is the standardized normal pdf and Φnorm(z) is the
100 • •
corresponding normal cdf.
••
50 The parameters are θ = (β0, β1, σ).
0 1 2 3 4 5 6 7
System Load
16 - 9 16 - 10
16 - 11 16 - 12
Standard Errors and Confidence Intervals for
Quantities at Specific Explanatory Variable Conditions
16 - 13 16 - 14
stress in ksi.
100
• 26 units tested; 4 units did not fail.
50
16 - 15 16 - 16
100
Pseudo-stress (ksi)
16 - 17 16 - 18
Weibull Distribution Quadratic Regression Model
Likelihood for Weibull Distribution Quadratic
with Nonconstant β = 1/σ
Regression Model with Right Censored Data
The parameters are θ = (β0, β1, β2, σ). which is not quadratic in x.
16 - 19 16 - 20
500
The likelihood for n independent observations has the form
• [μ] [μ] [μ] [σ]
L(β0 , β1 , β2 , β0 , β1 )
[σ]
•••
• • n
Thousands of Cycles
100
[μ] [μ] [μ] [σ] [σ]
= Li(β0 , β1 , β2 , β0 , β1 ; datai)
50 •
• i=1
n 4
:δ 4 :1−δ
1 log(ti) − μi i
log(ti) − μi i
••• = φsev 1 − Φsev .
• ••
• • 90% i=1 σi ti σi σi
10
••
•• 50% [μ] [μ] [μ] [σ] [σ]
5 10%
where μi = β0 + β1 xi + β2 x2 i and σi = exp β0 + β1 xi .
[μ] [μ] [μ] [σ] [σ]
Parameters are θ = (β0 , β1 , β2 , β0 , β1 ).
1
Pseudo-stress (ksi)
16 - 21 16 - 22
16 - 23 16 - 24
Definition of Residuals
Other residual plots. • With models for positive random variables like Weibull, log-
normal, and loglogistic, standardized residuals are defined
as
• Most analytical tests can be suitably generalized, at least ( )1/σ
log(ti) − log(t̂i) ti
approximately, for censored data (especially using likelihood exp(i) = exp =
σ t̂i
ratio tests).
i) and when ti is a censored observation,
where t̂i = exp(μ
the corresponding residual is also censored.
16 - 25 16 - 26
Plot of Standardized Residuals Versus Fitted Values Probability Plot of the Standardized Residuals from
for the Log-Quadratic Weibull Regression Model Fit the Log-Quadratic Weibull Regression Model
to the Super Alloy Data on Log-Log Axes Fit to the Super Alloy Data
.98
•
5.000 .9 •
• •
2.000 • • ••
• • .7
••
1.000
••
• • • • ••
• • .5
••
0.500 • • •• ••
Standardized Residuals
• .3 ••
••
Probability
0.200
0.100 • .2
•
• •
0.050 • .1 •
0.020 •
.05
0.010
.03
0.005
.02 •
0.002
•
0.001 .01
5 10 20 50 100 200 500 0.001 0.005 0.020 0.050 0.200 0.500 2.000 5.000
16 - 29 16 - 30
Transformations of a
Positive Explanatory Variable
Chapter 16
• For example, curvature in a plot of the data may suggest
Segment 4 that a model quadratic in x will provide a better fit than one
that is linear in x. In this case, the response ti is modeled
Transformations of a Positive Explanatory Variable using x∗i = x2i.
x∗i = xi
1/3
0.333 • The Box–Cox transformation has the following properties:
√
0.5 x∗i = xi
The transformed value x∗i is an increasing function of xi.
1 x∗i = xi
2 x∗i = x2 For fixed, xi, x∗i is a continuous function of λ through 0.
i
16 - 33 16 - 34
• Units tested at 270, 280, 300, 340, and 380 MPa. 15000
10000
• Some “runouts” at 270 and 280 MPa (8 and 2, respec-
tively).
5000
16 - 35 16 - 36
Laminate Panel Data
Laminate Panel Data Scatter Plot
Multiple Weibull Probability Plot
Log–Log Axes
Different Shape Parameters
.99 270MPa
.98
280MPa
50000 .9 300MPa
340MPa
380MPa
.7
20000
.5
10000
.3
5000
.2
Thousands of Cycles
Proportion Failing
2000 .1
1000 .05
500 .03
.02
200
.01
100
.005
50 .003
20
.001
260 280 300 320 340 360 380
101 102 103 104 105
MPa
Thousands of Cycles
16 - 37 16 - 38
.8 .8
.7 .7
.6 .6
Proportion Failing
Proportion Failing
.5 .5
.4 .4
.3 .3
.2 .2
.1 .1
.05 .05
.02 .02
.01 .01
.005 .005
.002 .002
.001 .001
1 2 3 4 5
10 10 10 10 10 101 102 103 104 105
16 - 39 16 - 40
.8 .8
.7 .7
.6 .6
Proportion Failing
Proportion Failing
.5 .5
.4 .4
.3 .3
.2 .2
.1 .1
.05 .05
.02 .02
.01 .01
.005 .005
.002 .002
.001 .001
1 2 3 4 5
10 10 10 10 10 101 102 103 104 105
16 - 41 16 - 42
Laminate Panel Data Laminate Panel Data
Model Plot Model Plot
Box-Cox Power Law Model λ = −2 Log Transformation
105 105
104 104
Thousands of Cycles
Thousands of Cycles
103 103
90%
102 102 90%
50%
50%
10%
10%
101 101
260 270 280 290 300 310 330 350 370 390 260 270 280 290 300 310 330 350 370 390
MPa on Box−Cox(−2) Scale MPa on Log Scale
16 - 43 16 - 44
16 - 45 16 - 46
1.0
Approximate Profile Likelihood
Chapter 16
0.8 0.50
Confidence Level
0.60
0.6
Segment 5
0.70
0.4
0.80
Failure-Time Regression Analysis
0.90
with Two Explanatory Variables
0.2
0.95
0.99
0.0
-4 -3 -2 -1 0
Stress Box-Cox Transformation Power
16 - 47 16 - 48
Scatter Plot the Effect of Voltage and Temperature
on Glass Capacitor Life
1200
• Experiment designed to determine the effect of voltage and 170 Degrees C
180 Degrees C
temperature on capacitor life.
1000
800
• 2 × 4 factorial, 8 units at each combination.
Hours
600
.6 170DegreesC;200Volts .7 170DegreesC;200Volts
170DegreesC;250Volts 170DegreesC;250Volts
170DegreesC;300Volts 170DegreesC;300Volts
170DegreesC;350Volts 170DegreesC;350Volts
.5 180DegreesC;200Volts 180DegreesC;200Volts
180DegreesC;250Volts .5 180DegreesC;250Volts
180DegreesC;300Volts 180DegreesC;300Volts
180DegreesC;350Volts 180DegreesC;350Volts
.4
.3
Proportion Failing
Proportion Failing
.3
.2
.2
.1
.1
.05 .05
100 200 500 1000 2000 100 200 500 1000 2000
Hours Hours
16 - 51 16 - 52
.7 170DegreesC;200Volts
170DegreesC;250Volts
170DegreesC;300Volts
170DegreesC;350Volts
• Additive model
180DegreesC;200Volts
.3
• The interaction model
Proportion Failing
Hours
16 - 53 16 - 54
Weibull Probability Plot Weibull Probability Plot
Glass Capacitor Life Test Results Glass Capacitor Life Test Results
Arrhenius/Power Law Model Common Shape Parameter
.7 170DegreesC;200Volts .7 170DegreesC;200Volts
170DegreesC;250Volts 170DegreesC;250Volts
170DegreesC;300Volts 170DegreesC;300Volts
170DegreesC;350Volts 170DegreesC;350Volts
180DegreesC;200Volts 180DegreesC;200Volts
.5 180DegreesC;250Volts .5 180DegreesC;250Volts
180DegreesC;300Volts 180DegreesC;300Volts
180DegreesC;350Volts 180DegreesC;350Volts
.3 .3
Proportion Failing
Proportion Failing
.2 .2
.1 .1
.05 .05
100 200 500 1000 2000 100 200 500 1000 2000
Hours Hours
16 - 55 16 - 56
600
Hours