Chapter 23 Accelerated Destructive Degradation Test Data, Models, and Analysis

advertisement
Chapter 23
Accelerated Destructive Degradation Test Data,
Models, and Analysis
William Q. Meeker and Luis A. Escobar
Iowa State University and Louisiana State University
Copyright 2002-2003 W. Q. Meeker and L. A. Escobar.
Complements to the authors’ text Statistical Methods for
Reliability Data, John Wiley & Sons Inc., 1998.
Work done jointly with Danny L. Kugler and Laura L. Kramer (Imaging
& Printing Group, Hewlett-Packard)
January 13, 2014
3h 42min
23 - 1
Accelerated Destructive Degradation Tests
Data, Models, and Data Analysis
Chapter 23 Objectives
• Describe useful accelerated destructive degradation test (ADDT)
reliability models.
• Show the connection between degradation reliability models
and failure-time reliability models.
• Present methods of data analysis and reliability inference
for ADDT.
• Discuss the use of ADDT data to estimate acceleration
factors.
• Illustrate methods with examples.
23 - 2
AdhesiveBondB ADDT Data
• Objective: Assess the strength of an AdhesiveBond over
time; estimate the proportion of devices with a strength
below 40 Newtons after 5 years of operation (approximately
260 weeks) at 25◦C.
• The test is destructive; each unit can be measured only
once. There were 6 right censored observations.
• Test plan: 8 units with no ageing were measured at the
start of the experiment. A total of 80 additional units were
aged and measured according to the following temperatures
and time schedule.
Temp
◦C
—
50
60
70
Totals
0
Weeks Aged
2 4
6 12
Totals
16
8
8
8
6
6
20
0
0
6
6
8
6
4
18
8
6
9
23
7
6
0
13
8
31
24
25
88
23 - 3
DegreesC
AdhesiveBondB ADDT Plan
Number of Units Measured at Each
(Time, Temperature) Combination
70
6
60
50
25
6
4
9
6
6
6
6
8
8
8
7
6
12
16
8
0
2
4
Weeks
23 - 4
AdhesiveBondB ADDT Data
Scatter Plot
Linear–Linear Axes
AdhesiveBondB Data
Destructive Degradation Scatter Plot
Resp:Linear,Time:Linear
100
Newtons
80
60
40
50DegreesC
60DegreesC
70DegreesC
20
0
0
5
10
15
20
Weeks
23 - 5
AdhesiveBondB ADDT Data
Scatter Plot at Individual Conditions of Temperature
Linear–Linear Axes
AdhesiveBondB Data
Resp:Linear,Time:Linear, Dist:Normal
50DegreesC
60DegreesC
100
100
80
80
60
60
40
40
20
20
0
0
5
10
15
20
15
20
0
5
10
15
20
Newtons
0
70DegreesC
100
80
60
40
20
0
0
5
10
Weeks
23 - 6
AdhesiveBondB ADDT Data
Scatter Plot at Individual Conditions of Temperature
Square Root–Log Axes
AdhesiveBondB Data
Resp:Log,Time:Square root, Dist:Normal
50DegreesC
60DegreesC
80
40
20
10
80
40
20
10
5
10 15
0
5
10 15
Newtons
0
70DegreesC
80
40
20
10
0
5
10 15
Weeks
23 - 7
Model for Accelerated Degradation Path
• Actual degradation path model: Actual path of a unit at
ti and accelerating variable(s) level (combinations) AccVarj
(e.g., temperature or temperature and relative humidity) is
Dij = D(τi, xj , β )
where τi = ht(ti) is a known monotone increasing transformations of ti and xj = ha(AccVarj ) are known transformations of the accelerating variables at the jth level (or
combination of levels) AccVarj .
When there is no possibility of confusion, τi and xj are called
the time and the AccVar, respectively.
• Rates in the model are with respect to τ = ht(t).
• Path parameters: the elements of β are fixed but unknown.
• Though no commonly used in accelerated testing it is possible to have models with interactions among the accelerating
variables.
23 - 8
Statistical Model for ADDT Data
• Sample path model: For observation k at time τi and
accelerating variable(s) level xj the model is
yijk = hd(Dij ) + ǫijk
= µij + ǫijk , k = 1, . . . , nij
where µij = hd(Dij ) and yijk are, respectively, monotone
increasing transformations of Dij and the observed degradation.
nij is the number of observations at time ti and AccVarj
vector level xj .
ǫijk is a residual deviation which describes unit-to-unit variability with (ǫijk /σ) ∼ Φ(z).
• The transformations for the observed degradation, Dij , ti,,
and the AccVarj might be suggested by physical/chemical
theory, past experience, or the data.
23 - 9
Degradation Path Model: Single Temperature Level
D(τ, x, β ) = exp[β0 + β1 exp(β2x)τ ]
√
τ = Weeks and x is Arrhenius-Transformed Temp
Linear–Linear Axes
100
70 DegreesC
Newtons
80
60
40
20
0
0
10
20
30
40
50
60
Weeks
23 - 10
Degradation Path Model: Multiple Temperature Levels
D(τ, x, β ) = exp[β0 + β1 exp(β2x)τ ]
√
τ = Weeks, x is Arrhenius-Transformed Temp
Linear–Linear Axes
100
70 DegreesC
60 DegreesC
50 DegreesC
Newtons
80
60
40
20
0
0
10
20
30
40
50
60
Weeks
23 - 11
Using Transformations
to Linearize a Degradation Model
log[D(τ, x, β )] = β0 + β1 exp(β2x)τ
√
τ = Weeks, x is Arrhenius-Transformed Temp
Square Root–Log Axes
100
Newtons
50
20
70 DegreesC
60 DegreesC
50 DegreesC
10
0
10
20
30
40
50
60
Weeks
23 - 12
A Class of Linear Degradation Models
• These models are of the form
yijk = µij + ǫijk
= β0 + β1 exp(β ′2xj )τi + ǫijk ,
k = 1, . . . , nij
where yijk , τi, and xj may be transformations of the measured degradation, ti, and the accelerating variable(s) AccVarj ,
respectively.
• The model is linear in the sense that for specified AccVar
vector xj , the degradation is linear in τi.
◮ For a single (scalar) AccVar xj , β ′2xj = β2xj .
◮ For multiple AccVar, xj , xj = (x2, . . . , xp)′, β ′2 = (β2, β3, . . . ),
and β ′2xj is a linear combination of the AccVar, i.e.,
β ′2xj =
p
X
i=2
βixji = β2xj2 + β3xj3 + · · · + βpxjp.
23 - 13
Linear Degradation Model
Interpretation of the Parameters
For the linear degradation model
yijk = β0 + β1 exp(β ′2xj )τi + ǫijk
• β0 is path intercept parameter when τ = 0. For example,
◮ If τi =
√
ti then β0 is the intercept at time t = 0.
◮ If τi = log(ti) then β0 is intercept at time t = 1.
• The degradation rate at xj is υ(xj )(AccVarj ) = β1 exp(β ′2xj ).
• The sign (±) of β1 determines if the degradation is increasing or decreasing in time.
• For a power transformation of time τ = tκ, the components
of the vector parameter β 2 are related to the amount of acceleration obtained by increasing the accelerating variables
AccVar.
23 - 14
Linear Degradation Model for
the AdhesiveBondB Data
For the AdhesiveBondB data, the strength of the adhesive
as a function of time and temperature is modeled by
yijk = β0 + β1 exp(β2xj )τi + ǫijk
where
yijk = log(Newtonsijk )
p
√
ti = Weeksi
τi =
xj = −11605/(◦Cj + 273.15)
(ǫijk /σ) ∼ Φnor (z).
The ǫijk term contains model and measurement errors.
23 - 15
ADDT Individual Analysis
Likelihood for Fixed AccVar level xj
• For the data at fixed xj of the AccVar with exact failure
times and right-censored observations, the likelihood is
nij Y Y
1
yijk − µij 1−δijk
yijk − µij δijk × 1−Φ
φ
Lj (θ |DATA) =
σ
σ
i k=1 σ
where µij ≡ µ(τi, xj , β ) = β0 + β1 exp(β ′2xj )τi, δijk indicates
whether observation yijk is a failure (δijk = 1) or a right
censored observation (δijk = 0), θ = (β0, β1, β ′2, σ), and nij
is the number of observations at (τi, xj ).
• For fixed xj , the model parameters are: the spread σ, the
intercept β0, and the slope of the line υj = β1 exp(β ′2xj ).
The parameter υj can be interpreted as the degradation
rate of µ(τ, xj , β ) with respect to the transformed time τ.
23 - 16
AdhesiveBondB ADDT Data
Normal Model ML Individual Line Fits
• For each temperature level three individual ML estimates
[j]
b [j].
are obtained: βb0 , υb [j], and σ
• A summary of the Normal model individual ML fits for the
AdhesiveBondB data is
ML Estimates
AccVarj
[j]
βb0
50◦C
4.490
60◦C
4.489
70◦C
4.400
υb [j]
−0.1088
−0.2089
−0.3626
sceυb[j]
0.01494
0.02214
0.01944
95% Approximate
Confidence Interval
for υ [j]
Lower
Upper
−0.1424
−0.08309
−0.2571
−0.4028
−0.16969
−0.32643
23 - 17
AdhesiveBondB ADDT Data
Individual Normal Model ML Fits
Square Root–Log Axes
[j]
b [j]τ
b [j] = βb0 + υ
µ
AdhesiveBondB Data
Resp:Log,Time:Square root, Dist:Normal
50DegreesC
60DegreesC
80
40
20
10
80
40
20
10
5
10 15
0
5
10 15
Newtons
0
70DegreesC
80
40
20
10
0
5
10 15
Weeks
23 - 18
AdhesiveBondB ADDT Data
Overlay of Individual Normal Model Fits
Square Root–Log Axes
[j]
b [j]τ
b [j] = βb0 + υ
µ
AdhesiveBondB Data
Destructive Degradation Individual Regression Analyses
Resp:Log,Time:Square root, Dist:Normal
100
80
Newtons
60
50
40
30
20
50DegreesC
60DegreesC
70DegreesC
10
0
5
10
15
20
Weeks
23 - 19
Individual Degradation Rate Estimates
• The ML estimates υb [j] (slopes of the individual lines) can
be used to identify the relationship between the degradation
rate and the AccVar.
• When the degradation is decreasing, use absolute values of
the degradation rate.
• Because
log(| υj |) = log(| β1 |) + β ′2xj
the surface log(| υb [j] |) versus the AccVar xj should be approximately linear in the xj if the model relating degradation
rate and the AccVar is adequate. Then
◮ For a single AccVar xj , the plot of log(| υb [j] |) versus xj
should be approximately linear.
◮ For an AccVar vector xj , the plot of
log(| υb [j] |) versus any of the accelerating variables, conditional on fixed values of the remaining accelerating
variables, should be approximately linear.
23 - 20
Arrhenius Plot of |υb [j]|
Individual Degradation Rates
Normal Model ML Estimates
Degradation rate versus DegreesC on Arrhenius Scale for
AdhesiveBondB Data
Resp:Log,Time:Square root,x:arrhenius, Dist:Normal
0.40
0.30
0.25
Slope
0.20
0.15
0.10
0.05
45
50
55
60
65
70
75
DegreesC on Arrhenius Scale
23 - 21
Likelihood for All Data With Right Censored Data
• For a sample of n units consisting of exact failure times and
right-censored observations, the likelihood can be expressed
as
L(θ |DATA) =
L(θ |DATA) =
Y
j
Lj (θ |DATA)
Y 1
ijk
σ
φ
yijk − µij δijk
σ
× 1−Φ
yijk − µij 1−δijk
σ
where θ = (β0, β1, β ′2, σ)′, and δijk indicates whether observation ijk is a failure (δijk = 1) or a right censored observation (δijk = 0).
• Note that µij is a nonlinear function of the accelerating
variable xj .
• For the AdhesiveBondB data where xj is a scalar,
µij = β0 + β1 exp(β2xj )τi
11605
xj = − ◦
.
Cj + 273.15
23 - 22
AdhesiveBondB Data ML Estimates
for the Acceleration Model Fit
• Parameter estimates
Parameter
β0
β1
β2
σ
95% Approximate
Confidence Interval
Lower
Upper
4.396
4.547
ML
Estimate
4.471
Standard
Error
0.03864
−8.641 × 108
1.595 × 109
−3.989 × 109
2.261 × 109
0.1580
0.01233
0.1356
0.1841
0.6364
0.05488
0.5375
0.7536
• Normal model ML for the slopes (degradation rates) at
b ◦Cj ) = βb1 exp(βb2xj ),
each temperature are obtained from υ(
where xj = −11605/(◦Cj + 273.15). In this case for the four
temperatures of interest the ML estimates are
b
υ(25)
= −0.0151,
b
υ(60)
= −0.2035,
b
υ(50)
= −0.1025
b
υ(70)
= −0.3883
23 - 23
AdhesiveBondB ADDT Data
Normal Model Arrhenius ML Fit
b ij = βb0 + βb1 exp(βb2xj )τi
µ
AdhesiveBondB data
Destructive Degradation Regression Analyses
Resp:Log,Time:Square root,DegreesC:Arrhenius, Dist:Normal
100
80
Newtons
60
50
40
30
20
50DegreesC
60DegreesC
70DegreesC
25DegreesC
10
0
5
10
15
20
Weeks
23 - 24
Model Checking
Residual Plots
• Residuals versus fitted values.
• Residuals versus accelerating variables.
• Residuals versus time of exposure.
• Residuals versus observation order.
observations are taken sequentially.
This is useful when
• Residual probability plot.
23 - 25
AdhesiveBondB ADDT Data
Residuals Versus Fitted Values
AdhesiveBondB Data
Destructive Degradation Residuals versus Fitted Values
Resp:Log,Time:Square root,DegreesC:Arrhenius, Dist:Normal
4
Standardized Residuals
2
.95
0
.50
.05
-2
-4
-6
3.0
3.2
3.4
3.6
3.8
4.0
4.2
4.4
4.6
Fitted Values
23 - 26
AdhesiveBondB ADDT Data
Residuals Versus Temperature Conditions
AdhesiveBondB Data
Destructive Degradation Residuals versus DegreesC
Resp:Log,Time:Square root,DegreesC:Arrhenius, Dist:Normal
4
Standardized Residuals
2
.95
0
.50
.05
-2
-4
-6
50
55
60
65
70
DegreesC
23 - 27
AdhesiveBondB ADDT Data
Sev Residual Probability Plot
AdhesiveBondB Data
Destructive Degradation Residual Probability Plot with 95% Simultaneous Confidence Bands
Resp:Log,Time:Square root,DegreesC:Arrhenius, Dist:Normal
Normal Probability Plot
.9999
.999
Probability
.995
.98
.95
.9
.8
.6
.4
.2
.1
.05
.01
.002
.0005
.0001
-3
-2
-1
0
1
2
3
Standardized Residuals
23 - 28
Some Comments on the AdhesiveBondB Residuals
• The standardized residuals look approximately like a random
sample from a NOR(0, 1).
• The vertical line at 0 is the median life of the standardized distribution. Then approximately 50% of the residuals
should be below that line.
• There appears to be some evidence of nonconstant variance, but it is not systematic with temperature or times.
23 - 29
Distribution of Degradation at (Possibly Transformed)
Time and AccVar (t, AccVar)
• The degradation distribution, for given time t and AccVar
vector x is
FY (y; τ, x) = P (Y ≤ y; τ, x) = Φ
"
y − µ(τ, x, β )
σ
#
where y = hd(degradation), µ(τ, x, β ) = β0 + β1 exp(β ′2x)τ.
• The p quantiles of the distribution is yp = µ(t, x, β )+σΦ−1(p).
• The ML estimate of the degradation distribution for given
(t, x) is
b
y−µ
b
FY (y; τ, x) = Φ
b
σ
b x)τ , τ = h (t), x = h (AccVar),
b = βb0 + βb1 exp(β
where µ
a
t
2
b ,σ
and βb0, βb1, β
2 b are ML estimates of the corresponding parameters. When AccVar is a vector, ha is generally a different function for each of its elements.
23 - 30
AdhesiveBondB ADDT Data
Distribution of Degradation at
Given Time and Temperature (t, Temp)
• For the AdhesiveBondB data, the Normal model ML estimate of FY (y; τ, x) at time t and temperature ◦C is
b
y−µ
FbY (y; τ, x) = Φnor
b
σ
√
b
b
b
b = β0+β1 exp(β2x)τ , τ = Weeks, x = −11605/(◦C+
where µ
273.15), βb0 = 4.471, βb1 = −864064160, βb2 = 0.6364,
b = 0.1580
σ
• The ML estimate of the p quantile (log Newtons) is
b+σ
b Φ−1
ybp = µ
nor (p).
23 - 31
Induced Failure Time Distribution at
Fixed Values of (AccVar, Df )
for Decreasing Linear Degradation
• Observe that T ≤ t [i.e., ht(T ) ≤ τ ] is equivalent to observed
degradation less than Df (i.e., Y ≤ µf ,) where µf = hd(Df ).
Then
FT (t; x, β ) = FY (µf ; x, β ) = Φ
τ −ν
= Φ
ς
"
µf − µ(τ, x, β )
σ
#
for t ≥ 0
where τ = ht(t),
(β0 − µf ) exp(−β ′2x)
ν=
| β1 |
and
σ exp(−β ′2x)
ς=
.
| β1 |
• The failure time distribution is a mixed distribution with a
spike of Pr(T = 0) = Φ [(β0 − µf )/σ ] at t = 0.
For t > 0 the cdf is continuous and it agrees with the cdf
of a log-location-scale variable with standardized cdf Φ(z),
location ν and scale ς.
23 - 32
50
Newtons
100
ML Estimate Showing Proportion Failing as
a Function of Time at 25◦C
b Φ−1
ybp = βb0 + βb1 exp(βb2x)τ + σ
nor (p)
1%
0.1%
0
20
100
200
500
1000
Weeks
23 - 33
Quantiles for the Induced Failure Time Distribution at
AccVar vector x and Critical Condition Df
Decreasing Linear Degradation
• For p ≥ Φ [(β0 − µf )/σ ]
ht(tp) = τp = ν + ςΦ−1(p)
where
(β0 − µf ) exp(−β ′2x)
ν=
| β1 |
Thus
σ exp(−β ′2x)
ς=
.
| β1 |
and
h
i
−1
−1
t p = ht
ν + ςΦ (p) .
• Substituting the expressions for ν and ς and after simplifications, shows that
log[ht(tp)] = log(τp) = −β ′2x + log
"
(β0 − µf ) + σΦ−1(p)
| β1 |
#
.
Thus, the log of the transformed failure-time distribution
quantiles are linear in the transformed AccVar vector x.
23 - 34
ML Estimates for Quantiles of the Induced Failure
Time Distribution at AccVar x and Critical
Degradation Df
Decreasing Linear Degradation
• For p ≥ Φ [(β0 − µf )/σ ]
h
i
−1
−1
tbp = ht
νb + ςbΦ (p)
where h−1
t (·) is the inverse of the ht (·) function.
• An approximate 100(1−α)% confidence interval for tp based
on the large-sample approximate NOR(0, 1) distribution of
Zlog(bt ) = [log(tbp) − log(tp)]/scelog(bt ) is
p
p
[tp,
e
t̃p] = [tbp/w,
where w = exp[z(1−α/2)scebt /tbp].
tbp × w]
p
23 - 35
AdhesiveBondB ADDT Data
Failure Time Distribution Quantile Estimates
• For the AdhesiveBondB accelerated test data, ht(t) =
Then
log[
q
√
t.
#
−1
b
b Φ (p)
(β 0 − µ f ) + σ
.
tbp] = −βb2x + log
b
| β1 |
"
This implies that
#
−1
b
b
(β0 − µf ) + σ Φ (p)
.
log(tbp) = −2βb2x + 2 log
b
| β1 |
"
Thus the logs of the quantiles of the failure-time distribution
are linear in the AccVar vector x.
23 - 36
AdhesiveBondB Data
Model Plot ML Estimate of Failure Time Distribution
as Functionn of Temperature
o
−1
b
b
b
b
b Φ (p)]/ | β1 |
log(tp) = −2β2x + 2 log [(β0 − µf ) + σ
Model plot for AdhesiveBondB data
Resp:log,Time:Square root,DegreesC:arrhenius, Dist:normal
Failure−time distribution for degradation failure level of 40 Newtons
10000
Weeks
1000
100
10
90%
50%
10%
1
Probability spike at time zero = 3.6473e−007
20
30
40
50
60
70
80
DegreesC on Arrhenius Scale
23 - 37
Acceleration Factors
• Here we consider time power transformations, i.e., τ =
ht(t) = tκ, where κ 6= 0.
• To obtain the accelerating effect of the AccVar x, let τ (x)
and τ (xU ) be the (transformed) times to reach the critical
degradation Df when the (transformed) accelerating variable take values x and xU , respectively.
• Solving for τ (x) and τxU the equation
Df = D[τ (x), x, β ] = D[τ (xU ), xU , β ]
gives
τ (x U )
ht[t(xU )]
=
= exp[β ′2(x − xU )].
τ (x )
ht[t(x)]
Using τ (x) = ht[t(x)] = [t(x)]κ and solving for t(xU )/t(x)
yields
1
t(xU )
= exp β ′2(x − xU ) .
AF (x) =
t(x)
κ
23 - 38
AdhesiveBondB ADDT Data
ML Estimates of Acceleration Factors
√
• For the AdhesiveBondB, ht(t) = t, then κ = 1/2 and the
acceleration factor for temperature is
"
AF (temp) = exp 2 β2
11605
11605
−
tempU K temp K
!#
where temp K = temp ◦C + 273.15 is temperature in the absolute Kelvin scale.
• The Normal model ML estimate are
"
d (temp) = exp 2 βb
AF
2
11605
11605
−
tempU K temp K
!#
where βb2 = 0.6364.
• The ML estimate of the Arrhenius activation energy
is 2 × βb2 = 2 × 0.6364 = 1.2728
d (temp) is done easily using SPLIDA.
• The computation of AF
d (60) = 182.12
For example, AF
23 - 39
AdhesiveBondB Data
Normal Model Acceleration Factors
as a Function of Temperature
Acceleration Factor relative to 25 Degrees C
Arrhenius relationship with activation energy in units of eV,
200
1.2728
100
50
20
10
5
2
1
25
30
35
40
45
50
55
60
Degrees C
23 - 40
A Two-Variables Linear Degradation Model
With Temperature and Relative Humidity as AccVar
• AdhesiveBondD Model
yijk = β0 + β1 exp(β ′2xj )τi + ǫijk
= β0 + β1 exp(β2xj2 + β3xj3)τi + ǫijk
where
yijk = log(Newtonsijk )
p
√
τi =
ti = Weeksi
(ǫijk /σ) ∼ Φnor (z).
• The AccVar transformations are Arrhenius for Temp (Temperature) and logit for RH (Relative Humidity), i.e.,
11605
xj2 = − ◦
Cj + 273.15
RH
xj3 = log
.
1 − RH
23 - 41
AdhesiveBondD ADDT Data
• Objective: Assess the strength of an AdhesiveBond over
time.
There is interest in estimating the proportion of devices with
a strength below 25 Newtons after 5 years of operation at
room temperature of 25◦C and relative humidity of 50%.
• The test is destructive; each unit can be measured only
once.
• Experimental factors
◮ AccVar:
Temp (Temperature):
RH (Relative Humidity):
50◦C, 60◦C, 70◦C
20%, 80%
◮ Weeks : Measurements after 0, 1, 2, 4, 6, 8, 12, 16 Weeks
• Sample size: 6 units allocated at each (Weeks, RH, Temp)
combination. This gives a sample size equal
to: 6 × 8 × 2 × 3 = 288 units.
23 - 42
AdhesiveBondD ADDT Data
Structure of the Data
The test plan is a completely balanced 8 × 3 × 2 full factorial arrangement in Weeks, Temp, and RH, with 6 units
allocated at each combination of the experimental factors.
RH
%
20
Temp
◦C
50
60
70
80
50
60
70
Totals
0
1
6
6
6
6
6
6
36
6
6
6
6
6
6
36
Weeks Aged
2
4
6
8
6
6
6
6
6
6
36
6
6
6
6
6
6
36
6
6
6
6
6
6
36
6
6
6
6
6
6
36
Totals
12
16
6
6
6
6
6
6
36
6
6
6
6
6
6
36
48
48
48
48
48
48
288
23 - 43
AdhesiveBondD ADDT Data
Acceleration Factors
√
• For the AdhesiveBondD, ht(t) = t, then acceleration factor for temperature and relative humidity is
t(xU )
1
AF (x) =
= exp β ′2(x − xU )
t(x)
κ
where κ = 1/2 because the square root transformation on
time. Then
(
!
11605
11605
AF (temp, RH) = exp 2 β2
−
+
tempU K temp K
!#)
"
RH
RHU
− log
2 β3 log
1 − RH
1 − RHU
where tempU K = 25◦C + 273.15 and RHU = 50% are the
AccVar at use conditions.
• The estimation of AF (temp, RH) will be programmed in SPLIDA.
23 - 44
Another Linear Degradation Model for
an Insulation Breakdown
Nelson 1990, Chapter 11, discusses the following linear
degradation model for time to an insulation Breakdown.
yijk = β0 + β1 exp(β2xj )τi + ǫijk
where
yijk = log[(Breakdown Voltage)ijk ]
τi = ti = Weeksi
xj = −11605/(◦Cj + 273.15)
(ǫijk /σ) ∼ Φnor (z).
Notice that in this instance the variable time is used in the
linear scale.
23 - 45
ADT Models with Interactions
• Most accelerated test models used in practice do not contain interaction terms.
• Interaction terms imply curvature in the degradation rate
versus the AccVar surface, i.e.,
µij = β0 + β1 exp(β2xj2 + β3xj3 + β4xj2xj3) τi.
• Extrapolation is hazardous, especially with surfaces involving curvature.
• A physical model could suggest interactions.
• Current version of the SPLIDA software does not allow for
interactions.
23 - 46
Future Research in ADDT Data, Models, and Analysis
• Nonlinear degradation models.
• Coarse data.
• Stochastic variability in the degradation response.
• Prediction in non-constant environments.
• Use of prior information.
• Random initiation times.
23 - 47
Technical Details
The following slides give technical details used in SPLIDA
to implement the methodology.
23 - 48
A Reparameterization of the
Linear ADDT Model for Numerical Stability
• The model is as before
yij = β0 + β1 exp(β ′2xj )τi + ǫij .
• Suppose that x̄ is the centroid of the stress variables
[i.e., x̄ = (x̄1, . . . , x̄k )′] and τ̄ is an average transformed
time. Then the model can be reparameterized as
yij = γ0 + γ1
exp
′
xj − x̄ γ 2 τi − τ̄ + ǫij .
where γ0 is the intercept for the average stress line (i.e.,
degradation line for x̄) at τ̄ ; γ1 is the slope of the average
stress line; and γ 2 = β 2 are the regression coefficients
corresponding to the x variables.
• It can be shown that γ1 is the geometric mean of the slopes
β1 exp(β ′2xj ), for the values of xj stress variables in the data
set.
23 - 49
Relationship Between Stable and Original Parameters
• Then
γ0 = β0 + β1 exp(β ′2x̄)τ̄
γ1 = β1 exp(β ′2x̄)
γ 2 = β 2.
• Solving for the βs
β2 = γ 2
β1 = γ1 exp(−x̄′γ 2)
β0 = γ0 − γ1τ̄ .
23 - 50
Induced Failure Time Distribution at
Use Conditions x and Critical Level Df
Decreasing Linear Degradation
• In this case
τ −ν
FT (t, x, β ) = Φ
ς
for t ≥ 0
where
(β0 − µf ) exp(−β ′2x)
ν=
| β1 |
and
σ exp(−β ′2x)
ς=
.
| β1 |
• The failure time distribution is a mixed distribution with a
spike of Φ [(µf − β0)/σ ] = Φ (− ν/ς ) at t = 0.
For t > 0 the cdf is continuous and it agrees with the cdf of a
location-scale variable with standardized cdf Φ(z), location
ν and scale ς.
23 - 51
Induced Failure Time Distribution at
Use Conditions x and Critical Level Df
Decreasing Linear Degradation (Continued)
• In particular, the following follows:
Y = hd(degradation) ∼ Φ
and
τ −ν
ht(T ) ∼ Φ
,
ς
"
y − µ(τ, x, β )
σ
#
t > 0.
• This implies that ht(T ) is location-scale distributed with
parameters (ν, ς) and standardized distribution Φ(·).
23 - 52
Density to Plot Induced Failure Time Distribution at
Use Conditions x and Critical Level Df
Decreasing Linear Degradation
Let W = log(T ), one needs to plot fW (w) which is giving
by
exp(w)
τ −ν
fW (w, β ) =
×φ
ς
ς
where
dτ × dt t=exp(w)
for − ∞ < w < ∞
τ = 
ht[exp(w)]

1
if not transformation on time
dτ =
1

dt t=exp(w)
if square-root transformation on time

2τ
and
(β0 − µf ) exp(−β ′2x)
ν=
| β1 |
and
σ exp(−β ′2x)
ς=
.
| β1 |
23 - 53
Induced Failure Time Distribution at
Use Conditions x and Critical Level Df
Increasing Linear Degradation
• In this case
−τ − ν
FT (t, x, β ) = 1 − Φ
ς
where as before
(β0 − µf ) exp(−β ′2x)
ν=
| β1 |
and
for t ≥ 0
σ exp(−β ′2x)
ς=
.
| β1 |
• The failure time distribution is a mixed distribution with a
spike of 1 − Φ [(µf − β0)/σ ] = 1 − Φ (− ν/ς ) at t = 0.
For t > 0 the cdf is continuous and it agrees with the cdf
of continuous random variable.
23 - 54
Induced Failure Time Distribution at
Use Conditions x and Critical Level Df
Increasing Linear Degradation
• In particular the following follows:
"
y − µ(τ, x, β )
Y = hd(degradation) ∼ Φ
σ
and
−τ − ν
−ht(T ) ∼ Φ
,
ς
#
t > 0.
• This implies that that −ht(T ) is location-scale distributed
with parameters (ν, ς) and standardized distribution Φ(·).
23 - 55
Density to Plot Induced Failure Time Distribution at
Use Conditions x and Critical Level Df
Increasing Linear Degradation
Let W = log(T ), you need to plot fW (w) which is giving by
exp(w)
−τ − ν
fW (w, β ) =
×φ
ς
ς
where
dτ × dt t=exp(w)
for − ∞ < w < ∞
τ = (
ht[exp(w)]
dτ 1
if not transformation on time
=
1
if square-root transformation on time
dt t=exp(w)
2τ
and
(β0 − µf ) exp(−β ′2x)
ν=
| β1 |
and
σ exp(−β ′2x)
ς=
.
| β1 |
23 - 56
Quantiles of the Failure Time Distribution
• For decreasing degradation, the quantiles of the failure time
distribution are

ν

 0
if p ≤ Φ −
ς
tp =
h
i

 h−1 ν + ςΦ−1(p)
otherwise
t
• For increasing degradation, the quantiles of the failure time
distribution are

ν

 0
if p ≤ 1 − Φ −
ς
tp =
n h
io

 h−1 − ν + ςΦ−1(1 − p)
otherwise
t
23 - 57
Download