Degradation Testing and Analysis

8
DEGRADATION TESTING AND

ANALYSIS
8.1 INTRODUCTION
As explained in Chapter 7, the accelerated life test is an important task in nearly

all effective reliability programs. In today’s competitive business environment,
the time allowed for testing is continuously reduced. On the other hand, products
become more reliable thanks to advancements in technology and manufacturing
capability. So it is not uncommon that accelerated life tests yield no or few
failures at low stress levels. In these situations it is difficult or impossible to ana-
lyze the life data and make meaningful inferences about product reliability. For
some products whose performance characteristics degrade over time, a failure
is said to have occurred when a performance characteristic crosses a specified
threshold. Such a failure, called a soft failure, is discussed in Chapters 2 and 5.
Degradation of the characteristic indicates a deterioration in reliability. The mea-
surements of the characteristic contain much useful and credible information
about product reliability. Therefore, it is possible to infer reliability by analyzing
the degradation data.
A degradation test has several advantages over a life test. Reliability analy-
sis using degradation data directly relates reliability to physical characteristics.
This distinction enables reliability to be estimated even before a test unit fails
and thus greatly shortens the test time. Degradation analysis often yields more
accurate estimates than those from life data analysis, especially when a test is
highly censored. As shown in Chapter 7, an unfailed unit affects an estimate only
through its censoring time. The time from censoring to failure, or the remaining
Life Cycle Reliability Engineering. Guangbin Yang

Copyright  2007 John Wiley & Sons, Inc. ISBN: 978-0-471-71529-0
332
DETERMINATION OF THE CRITICAL PERFORMANCE CHARACTERISTIC 333
life, is unknown and not taken into account in life data analysis. In contrast,
a degradation test measures the performance characteristics of an unfailed unit
at different times, including the censoring time. The degradation process and
the distance between the last measurement and a specified threshold are known.
In degradation analysis, such information is also utilized to estimate reliability.
Certainly, degradation analysis has drawbacks and limitations. For example, it
usually requires intensive computations.
In this chapter we describe different techniques for reliability estimation from
degradation data, which may be generated from nondestructive or destructive
inspections. The principle and method for accelerated degradation test with tight-
ened thresholds are also presented. We also give a brief survey of the optimal
design of degradation test plans.
8.2 DETERMINATION OF THE CRITICAL PERFORMANCE

CHARACTERISTIC
The performance of a product is usually measured by multiple characteristics.

Typically, a small component such as a capacitor or resistor has three or more
characteristics, and a complicated system such as the automobile can have dozens,
if not hundreds. Each characteristic reflects to some degree the level of product
reliability. In degradation analysis, quantitatively relating reliability to all char-
acteristics is difficult, if not impossible. In fact, most characteristics are neither
independent of each other nor equally important. In many applications, there
is one critical characteristic, which describes the dominant degradation process.
This one can be used to characterize product reliability. The robust reliability
design discussed in Chapter 5 is based on such a characteristic.
In many applications, the critical characteristic is fairly obvious and can be
identified using physical knowledge, customer requirement, or experience. For a
component to be installed in a system, the critical characteristic of the compo-
nent is often the one that has the greatest impact on system performance. The
drift and variation in this characteristic cause the system performance to deterio-
rate remarkably. The relationship between component characteristics and system
performance may be explored using a design of experiment. For a commercial
product, the critical characteristic is the one that most concerns customers. The
degradation in this characteristic closely reflects customer dissatisfaction on the
product. Determination of the critical characteristic may be aided by using the
quality deployment function, which translates customer expectations to product
performance characteristics. This technique was described in detail in Chapter 3.
The critical characteristic chosen to characterize product reliability must be
increasing or decreasing monotonically over time. That is, the degradation in the
characteristic is irreversible. This requirement is satisfied in most applications.
For example, attrition of automobile tires always increases with mileage, and the
bond strength of solder joints always decreases with age. For some electronic
products, however, the characteristic may not be monotone in early life, due to
burn-in effects. The characteristic fluctuates in a short period of time and then
334 DEGRADATION TESTING AND ANALYSIS
becomes monotone afterward. An example is light-emitting diodes, the luminous

power of which may increase in the first few hundred hours of operation and
then decreases over time. For these products it is important that the degradation
analysis excludes observations collected in the burn-in period.
8.3 RELIABILITY ESTIMATION FROM PSEUDOLIFE

Suppose that a degradation test uses a sample of n units. During testing, each
unit is inspected periodically to measure the critical performance characteristic
y. The inspection is nondestructive, meaning that the unit is not damaged by the
inspection and resumes its function after inspection. Let yij denote the measure-
ment of y on unit i at time tij , where i = 1, 2, . . . , n, j = 1, 2, . . . , mi , and mi is
the number of measurements on unit i. The degradation path can be modeled by
yij = g(tij ; β1i , β2i , . . . , βpi ) + eij , (8.1)
where g(tij ; β1i , β2i , . . . , βpi ) is the true degradation of y of unit i at time tij
and eij is the error term. Often, the error term is independent over i and j and
is modeled with the normal distribution with mean zero and standard deviation
σe , where σe is constant. Although measurements are taken on the same unit, the
potential autocorrelation among eij may be ignored if the readings are widely
spaced. In (8.1), β1i , β2i , . . . , βpi are unknown degradation model parameters
for unit i and should be estimated from test data, and p is the number of such
parameters.
During testing, the inspections on unit i yield the data points (ti1 , yi1 ), (ti2 , yi2 ),
. . . , (timi , yimi ). Since eij ∼ N (0, σe2 ), the log likelihood Li for the measurement
data of unit i is
1
mi
mi
Li = − ln(2π) − mi ln(σe ) − [yij − g(tij ; β1i , β2i , . . . , βpi )]2 .
2 2σe2 j =1
(8.2)
The estimates β̂1i , β̂2i , . . . , β̂pi and σ̂e are obtained by maximizing Li directly.
The parameters may also be estimated by the least squares method. This is
done by minimizing the sum of squares of the deviations of the measurements
from the true degradation path, which is given by

mi
mi
SSDi = eij2 = [yij − g(tij ; β1i , β2i , . . . , βpi )]2 , (8.3)
j =1 j =1
where SSDi is the sum of squares of deviations for unit i. Note that the maximum
likelihood estimates are the same as the least squares estimates.
Once the estimates β̂1i , β̂2i , . . . , β̂pi are obtained, we can calculate the pseu-
dolife. If a failure occurs when y crosses a specified threshold, denoted G, the
life of unit i is given by
tî = g −1 (G; β̂1i , β̂2i , . . . , β̂pi ), (8.4)

RELIABILITY ESTIMATION FROM PSEUDOLIFE 335
0 t1 t2 tn
t
FIGURE 8.1 Relation of degradation path, pseudolife, and life distribution
where g −1 is the inverse of g. Applying (8.4) to each test unit yields the lifetime
estimates tˆ1 , tˆ2 , . . . , tˆn . Apparently, pseudolifetimes are complete exact data. In
Chapter 7 we described probability plotting for this type of data. By using the
graphical method, we can determine a life distribution that fits these life data
adequately and estimate the distribution parameters. As explained in Chapter 7,
the maximum likelihood method should be used for estimation of distribution
parameters and other quantities of interest when commercial software is avail-
able. Figure 8.1 depicts the relation of degradation path, pseudolife, and life
distribution.
For some products, the true degradation path is simple and can be written in
a linear form:
g(t) = β1i + β2i t, (8.5)
where g(t), t, or both may represent a log transformation. Some examples fol-
low. The wear of an automobile tire is directly proportional to mileage and β1i =
0. Tseng et al. (1995) model the log luminous flux of the fluorescent lamp as a lin-
ear function of time. K. Yang and Yang (1998) establish a log-log linear relation-
ship between the variation ratio of luminous power and the aging time for a type
of infrared light-emitting diodes. The MOS field-effect transistors have a linear
relationship between the log current and log time, according to J. Lu et al. (1997).
For (8.5), the least squares estimates of β1i and β2i are
β̂1i = y i − β̂2i t i ,
i mi mi
mi m j =1 yij tij − j =1 yij j =1 tij
β̂2i = ,
mi 2 mi 2
mi j =1 tij − j =1 tij
1 1
mi mi
yi = yij , ti = tij .
mi j =1 mi j =1
Then the pseudolife of unit i is

G − β̂1i
tî = .
β̂2i
100
15
10
y
0.1
10 100 1000 10,000 100,000
t (s)
FIGURE 8.2 Percent transconductance degradation over time
Example 8.1 J. Lu et al. (1997) give the percent transconductance degradation

data taken at different times for five units of an MOS field-effect transistor.
The testing was censored at 40,000 seconds. The failure criterion is the percent
transconductance degradation greater than or equal to 15%. The data are shown in
Table 8.11 of Problem 8.10 and plotted on the log-log scale in Figure 8.2, where
the vertical axis is the percent transconductance degradation and the horizontal
axis is the time in seconds. The plot suggests a log-log linear degradation model:
ln(yij ) = β1i + β2i ln(t) + eij .
The degradation model above is fitted to each degradation path. Simple linear
regression analysis suggests that the degradation model is adequate. The least
squares estimates for the five paths are shown in Table 8.1. After obtaining the
estimates, we calculate the pseudolifetimes. For example, for unit 5 we have
ln(15) + 2.217
ln(tˆ5 ) = = 12.859 or tˆ5 = 384,285 seconds.
0.383
TABLE 8.1 Least Squares Estimates of Model

Parameters
Unit
Estimate 1 2 3 4 5
β̂1i −2.413 −2.735 −2.056 −2.796 −2.217

β̂2i 0.524 0.525 0.424 0.465 0.383
DEGRADATION ANALYSIS WITH RANDOM-EFFECT MODELS 337
99
95
90
80
70
60
Percent
50
40
30
20
10
5
1000 10,000 100,000 1,000,000 10,000,000

Time to failure (s)
FIGURE 8.3 Lognormal plot, ML fits, and percentile confidence intervals for the pseu-
dolife data of Example 8.1
Similarly, the pseudolifetimes for the other four units are tˆ1 = 17,553, tˆ2 =
31,816, tˆ3 = 75,809, and tˆ4 = 138,229 seconds. Among the commonly used
life distributions, the lognormal provides the best fit to these data. Figure 8.3
shows the lognormal plot, ML fit, and two-sided 90% confidence interval for
percentiles. The ML estimates of the scale and shape parameters are µ̂ = 11.214
and σ̂ = 1.085.
8.4 DEGRADATION ANALYSIS WITH RANDOM-EFFECT MODELS
8.4.1 Random-Effect Models

From Example 8.1 it can be seen that the values of the degradation model param-
eters β1 and β2 are different for each unit. In general, the values of some or all of
the model parameters β1 , β2 , . . . , βp in (8.1) vary from unit to unit. They can be
considered to be drawn randomly from the respective populations. The random
effects exist because of material property change, manufacturing process varia-
tion, stress variation, and many other uncontrollable factors. The parameters are
not independent; they usually form a joint distribution. In practice, a multivariate
normal distribution with mean vector µβ and variance–covariance matrix β
is often used for simplicity. For example, Ahmad and Sheikh (1984) employ a
bivariate normal model to study the tool-wear problem. Using the multivariate
normal distribution and (8.1), the likelihood function l for all measurement data
of the n units can be written as
n ∞ ∞
1
l= ···
i=1 −∞ −∞ (2π)
(p+mi )/2 σemi |β |1/2
  
 1 mi 
· exp −  zij2 + (βi − µβ ) β−1 (βi − µβ ) dβ1i · · · dβpi , (8.6)
 2 
j =1
where zij = [yij − g(tij ; β1i , β2i , . . . , βpi )]/σe , βi = [β1i , β2i , . . . , βpi ], and |β |
is the determinant of β . Conceptually, the model parameters, including the
mean vector µβ , the variance–covariance matrix β , and the standard deviation
of error σe , can be estimated by directly maximizing the likelihood. In practice,
however, the calculation is extremely difficult unless the true degradation path
takes a simple linear form such as in (8.5).
Here we provide a multivariate approach to estimating the model parameters
µβ and β . The approach is approximately accurate, yet very simple. First, we
fit the degradation model (8.1) to each individual degradation path and calculate
the parameter estimates β̂1i , β̂2i , . . . , β̂pi (i = 1, 2, . . . , n) by maximizing the log
likelihood (8.2) or by minimizing the sum of squares of the deviations (8.3). The
estimates of each parameter are considered as a sample of n observations. The
sample mean vector is
β = [β 1 , β 2 , . . . , β p ], (8.7)
where
1
n
βj = β̂j i , j = 1, 2, . . . , p. (8.8)
n i=1
The sample variance–covariance matrix is given by

 
s11 s12 ··· s1p
 s21 s22 ··· s2p 
S=
···
, (8.9)
··· ··· ··· 
sp1 sp2 ··· spp
where
1
n
skj = (β̂ki − β k )(β̂j i − β j ), (8.10)
n i=1
for k = 1, 2, . . . , p and j = 1, 2, . . . , p. skj (k = j ) is the covariance of βk and

βj . When k = j, skk is the sample variance of βk . When the sample size is small,
say n ≤ 15, the variance–covariance component is corrected by replacing n with
n − 1 to obtain the unbiased estimate. The computations can be done easily using
a statistical software package such as Minitab.
The association between βk and βj can be measured by the correlation coef-
ficient ρkj , which is defined as
skj
ρkj = √ √ . (8.11)
skk sjj
99
95
90 b2
b1
80
70
Percent
60
50
40
30
20
10
5
−3 −2 −1 0
Model parameter
FIGURE 8.4 Normal fits to estimates of β1 and β2
β and S are estimates of µβ and β , respectively. The estimates are useful in

calculating degradation reliability, as we will see in Section 8.4.2.
Example 8.2 In Example 8.1 we have fitted the log-log linear model to each
degradation path and obtained the estimates β̂1 and β̂2 for each path. Now β1
and β2 are considered as random variables, and we want to estimate µβ and β
by using the multivariate approach.
SOLUTION First we plot the five estimates of β1 and β2 on normal probability

paper, as shown in Figure 8.4. Loosely, the plot indicates that β1 and β2 have
joint normal distribution. The mean of β1 is β 1 = (−2.413 − 2.735 − 2.056 −
2.796 − 2.217)/5 = −2.443. Similarly, β 2 = 0.464. Then the estimate of the
mean vector is β = [β 1 , β 2 ] = [−2.443, 0.464]. The variance estimate of β1 is
(−2.413 + 2.443)2 + · · · + (−2.217 + 2.443)2

s11 = = 0.1029.
5−1
Similarly, s22 = 0.00387. The estimate of the covariance of β1 and β2 is
1
s12 = s21 = [(−2.413 + 2.443)(0.524 − 0.464) + · · ·
5−1
+(−2.217 + 2.443)(0.383 − 0.464)] = −0.01254.
The estimate of the variance–covariance matrix is

s11 s12 0.1029 −0.01254
S= = .
s21 s22 −0.01254 0.00387
distribution of y
G
y
life distribution
0 ti
t
FIGURE 8.5 Relation of degradation to failure
The correlation coefficient between β1 and β2 is
s12 −0.01254
ρ12 = √ √ =√ √ = −0.628.
s11 s22 0.1029 0.00387
The large absolute value suggests that the correlation between the two parameters
cannot be ignored, while the negative sign indicates that β1 increases as β2
decreases, and vice versa. In other words, in this particular case, a unit with a
smaller degradation percentage early in the test time (t = 1 second) will have a
greater degradation rate.
8.4.2 Relation of Degradation to Failure

As we saw earlier, some or all of the parameters in the true degradation path
g(t; β1 , β2 , . . . , βp ) are random variables. As such, the degradation amount at a
given time varies from unit to unit, and some units may have crossed a specified
threshold by this time. The probability of a critical characteristic crossing a
threshold at a given time equals the probability of failure at that time. As the
time proceeds, these probabilities increase. Figure 8.5 depicts the relation of the
distribution of characteristic y to the life distribution of the product for the case
where failure is defined in terms of y ≥ G, where G is a specified threshold. In
Figure 8.5, the shaded fraction of y distribution at time ti is equal to the shaded
fraction of the life distribution at that time, representing that the probability of
y(ti ) ≥ G equals the probability of T ≤ ti , where T denotes the time to failure.
In general, the probability of failure at time t can be expressed as
F (t) = Pr(T ≤ t) = Pr[y(t) ≥ G] = Pr[g(t; β1 , β2 , . . . , βp ) ≥ G]. (8.12)
For a monotonically decreasing performance characteristic, the probability of

failure is calculated by replacing y ≥ G in (8.12) with y ≤ G. In some simple
cases that will be presented later, (8.12) may lead to a closed form for F (t). In
most applications, however, (8.12) has to be evaluated numerically for the given
distributions of the model parameters.
As (8.12) indicates, the probability of failure depends on distribution of the
model parameters, which, in turn, is a function of stress level. As such, an acceler-
ated test is often conducted at an elevated stress level to generate more failures or
a larger amount of degradation before the test is censored. Sufficient degradation
reduces the statistical uncertainty of the estimate of the probability of failure. On
the other hand, (8.12) also indicates that the probability of failure is influenced
by the threshold. Essentially, a threshold is subjective and may be changed in
specific applications. For a monotonically increasing performance characteristic,
the smaller the threshold, the shorter the life and the larger the probability of fail-
ure. In this sense, a threshold can be considered as a stress; tightening a threshold
accelerates the test. This acceleration method was mentioned in Chapter 7 and is
discussed in detail in this chapter.
8.4.3 Reliability Estimation by Monte Carlo Simulation

Once the estimates of µβ and β are obtained for β1 , β2 , . . . , βp , we can use
a Monte Carlo simulation to generate a large number of degradation paths. The
probability of failure F (t) is approximated by the percentage of simulated degra-
dation paths crossing a specified threshold at the time of interest. The steps for
evaluating F (t) are as follows:
1. Generate n (a large number, say 100,000) sets of realizations of
β1 , β2 , . . . , βp from a multivariate normal distribution with mean vector β
and variance–covariance matrix S. The n sets of realizations are denoted
by β1i , β2i , . . . , βpi

, where i = 1, 2, . . . , n.
2. For each i, calculate the true degradation yi at the given time t by substi-
tuting β1i , β2i , . . . , βpi

into g(t; β1 , β2 , . . . , βp ).
3. Count the number of yi (i = 1, 2, . . . , n) crossing the threshold. Let r
denote this number.
4. The probability of failure at time t is approximated by F (t) ≈ r/n. Then
the reliability is 1 − F (t).
Example 8.3 In Example 8.2 we computed the mean vector β and the vari-
ance–covariance matrix S for an MOS field-effect transistor. Now we want to
evaluate the probability of failure at 1000, 2000, 3000, . . ., 900,000 seconds by
Monte Carlo simulation.
SOLUTION Using Minitab we generated 65,000 sets of β1 and β2 from the
bivariate normal distribution with mean vector β and the variance–covariance
matrix S calculated in Example 8.2. At a given time (e.g., t = 40,000), we
computed the percent transconductance degradation for each set (β1 , β2 ). Then
count the number of degradation percentages greater than 15%. For t =
40,000, the number is r = 21,418. The probability of failure is F (40,000) ≈
0.8
0.6
F(t)
0.4
0.2
0
0 200 400 600 800 1000
t (1000 s)
FIGURE 8.6 Failure probabilities computed from the Monte Carlo simulation data
21,418/65,000 = 0.3295. Repeat the calculation for other times, and plot the
probabilities of failure (Figure 8.6). In the next subsection this plot is compared
with others obtained using different approaches.
8.4.4 Reliability Estimation with a Bivariate Normal Distribution

We use a bivariate normal distribution to model the joint distribution of β1 and
β2 , where the means are µβ1 and µβ2 , the variances are σβ21 and σβ22 , and the
correlation coefficient is ρ12 . For the linear model (8.5), if a failure is said to
have occurred when y ≥ G, the probability of failure can be written as
F (t) = Pr[g(t) ≥ G] = Pr(β1 + β2 t ≥ G)

µβ1 + tµβ2 − G
= , (8.13)
(σβ21 + t 2 σβ22 + 2tρ12 σβ1 σβ2 )1/2
where (·) is the cumulative distribution function (cdf) of the standard normal
distribution. To evaluate F (t), the means, variances, and correlation coefficient
in (8.13) are substituted with their estimates.
Example 8.4 In Example 8.2 we computed the mean vector β and the vari-
ance–covariance matrix S for the MOS field-effect transistor. Now we use (8.13)
to evaluate the probabilities of failure at 1000, 2000, 3000, . . ., 900,000 seconds.
SOLUTION Because the degradation path is linear on the log-log scale as

shown in Example 8.1, both G and t in (8.13) must take the log transformation.
For t = 40,000, for example, the probability of failure is
F (40,000) =

−2.443+ 10.5966× 0.464− 2.7081
√ √
(0.1029+ 10.59662 × 0.00387− 2× 10.5966× 0.6284× 0.1029 0.00387)1/2
= (−0.4493) = 0.3266.
0.8
pseudolife simulation bivariate normal
0.6
F(t)
0.4
0.2
0
0 200 400 600 800 1000
t (1000 s)
FIGURE 8.7 Probabilities of failure calculated using different methods
The probabilities of failure at other times are computed similarly. Figure 8.7
shows probabilities at various times calculated from (8.13). For comparison,
probabilities using Monte Carlo simulation and pseudolife calculation are also
shown in Figure 8.7. The probability plots generated from (8.13) and the Monte
Carlo simulation cannot be differentiated visually, indicating that estimates from
the two approaches are considerably close. In contrast, the pseudolife calcula-
tion gives significantly different results, especially when the time is greater than
150,000 seconds. Let’s look at the numerical differences at the censoring time t =
40,000 seconds. Using (8.13), the probability of failure is F (40,000) = 0.3266.
The Monte Carlo simulation gave the probability as F (40,000) = 0.3295, as
shown in Example 8.3. The percentage difference is only 0.9%. Using the pseu-
dolife approach, the probability is

ln(40,000) − 11.214
F (40,000) = = 0.2847.
1.085
It deviates from F (40,000) = 0.3295 (the Monte Carlo simulation result) by
13.6%. In general, compared with the other two methods, the pseudolife method
provides less accurate results. But its simplicity is an obvious appeal.
8.4.5 Reliability Estimation with a Univariate Normal Distribution

As in (8.5), let’s consider the simple linear model g(t) = β1 + β2 t, where both
g(t) and t are on the original scale (no log transformation). Suppose that the
parameter β1 is fixed and β2 varies from unit to unit. That is, β1 representing
the initial degradation amount at time zero is common to all units, whereas β2
representing the degradation rate is a random variable. An important special
case is β1 equal to zero. For example, automobile tires do not wear before use.
Suppose that β2 has a normal distribution with mean µβ2 and standard deviation
σβ2 . For a monotonically increasing characteristic, the probability of failure can
be written as

G − β1
F (t) = Pr[g(t) ≥ G] = Pr(β1 + β2 t ≥ G) = Pr β2 ≥
t

µβ2 /(G − β1 ) − 1/t
= . (8.14)
σβ2 /(G − β1 )
Now let’s consider the case where ln(β2 ) can be modeled with a normal dis-
tribution with mean µβ2 and standard deviation σβ2 ; that is, β2 has a lognormal
distribution with scale parameter µβ2 and shape parameter σβ2 . For a monotoni-
cally increasing characteristic, the probability of failure can be expressed as

ln(t) − [ln(G − β1 ) − µβ2 ]
F (t) = Pr[g(t) ≥ G] = . (8.15)
σβ2
This indicates that the time to failure also has a lognormal distribution; the
scale parameter is ln(G − β1 ) − µβ2 and the shape parameter is σβ2 . Substituting
the estimates of the degradation model parameters and G into (8.15) gives an
estimate of the probability of failure.
Example 8.5 A solenoid valve is used to control the airflow at a desired rate.
As the valve ages, the actual airflow rate deviates from the rate desired. The
deviation represents the performance degradation of the valve. A sample of 11
valves was tested, and the percent deviation of the airflow rate from that desired
was measured at different numbers of cycles. Figure 8.8 plots the degradation
paths of the 11 units. Assuming that the valve fails when the percent deviation is
greater than or equal to 20%, estimate the reliability of the valve at 50,000 cycles.
SOLUTION As shown in Figure 8.8, the degradation data suggest a linear

relationship between the percent deviation and the number of cycles. Since mea-
surements of percent deviation at the first inspection time (1055 cycles) are very
12
10
Percent deviation
0
0 5000 10,000 15,000 20,000 25,000
Cycles
FIGURE 8.8 Degradation paths of the solenoid valves
DEGRADATION ANALYSIS FOR DESTRUCTIVE INSPECTIONS 345
99
95
90
80
70
Percent
60
50
40
30
20
10
5
0.1 1.0
b^2
FIGURE 8.9 Lognormal plot, ML fits, and percentile confidence intervals for the esti-
mates of β2
small, deviations at time zero are negligible. Thus, the true degradation path can
be modeled by g(t) = β2 t, where t is in thousands of cycles. The simple linear
regression analyses yielded estimates of β2 for the 11 test units. The estimates are
0.2892, 0.2809, 0.1994, 0.2303, 0.3755, 0.3441, 0.3043, 0.4726, 0.3467, 0.2624,
and 0.3134. Figure 8.9 shows the lognormal plot of these estimates with ML fit
and two-sided 90% confidence intervals for percentiles. It is seen that β2 can be
approximated adequately by a lognormal distribution with µ̂β2 = −1.19406 and
σ̂β2 = 0.22526.
From (8.15), the cycles to failure of the valve also have a lognormal distribu-
tion. The probability of failure at 50,000 cycles is

ln(50) − [ln(20) + 1.19406]
F (50) = = 0.1088.
0.22526
The reliability at this time is 0.8912, indicating that about 89% of the valves
will survive 50,000 cycles of operation.
8.5 DEGRADATION ANALYSIS FOR DESTRUCTIVE INSPECTIONS
For some products, inspection for measurement of performance characteristics

(often, monotonically decreasing strengths) must damage the function of the
units either completely or partially. Such units cannot restart the same func-
tion as before the inspection and are discontinued from testing. Thus, each unit
of the products can be inspected only once during testing and generates one
measurement. For example, a solder joint has to be sheared or pulled off to
get its joint strength, and measuring the dielectric strength of an insulator must
break down the insulator. For such products, the degradation analysis methods
described earlier in the chapter are not applicable. In this section we present
the random-process method for degradation testing and data analysis. It is worth
noting that this method is equally applicable to nondestructive products and is
especially suitable for cases when degradation models are complicated. Examples
of such application include K. Yang and Xue (1996), K. Yang and Yang (1998),
and W. Wang and Dragomir-Daescu (2002).
8.5.1 Test Methods

The test uses n samples and m destructive inspections. The inspection times are
t1 , t2 , . . . , tm . At t1 (which may be time zero), n1 units are destructed and yield the
measurements yi1 (i = 1, 2, . . . , n1 ). Then the degradation test continues (starts)
on the remaining n − n1 units until t2 , at which point n2 units are inspected. The
inspection gives the measurements yi2 (i = 1, 2, . . . , n2 ). The process is repeated
until (n − n1 − n2 − · · · − nm−1 ) units are measured at tm . The last inspection
produces the measurements yim (i = 1, 2, . . . , n − n1 − n2 − · · · − nm−1 ). Obvi-
ously, m j =1 nj = n. The n test units may be allocated equally to each inspection
time; however, this allocation may not be statistically efficient. A better alloca-
tion should consider the shape of the degradation path as well as the change in
performance dispersion over time. A degradation path may be linear, convex,
or concave. On the other hand, the measurement dispersion may be constant
or depend on time. Figure 8.10 shows examples of the three shapes of the
degradation paths and the decreasing performance dispersion for the linear path.
Decreasing dispersion for other shapes is possible but is not shown in Figure 8.10.
Principles for better sample allocation are given below. Optimal allocation as a
part of test planning deserves further study.
ž If a degradation path is linear and performance dispersion is constant over
time, the n test units should be apportioned equally to each inspection.
ž If a degradation path is linear and performance dispersion decreases with
time, more units should be allocated to low time inspections.
concave
y
convex linear
0 t1 t2 tm
t
FIGURE 8.10 Three shapes of degradation paths and decreasing performance dispersion
ž For the convex decreasing degradation path in Figure 8.10, the degradation
rate becomes smaller as time increases. For the degradation amount between
two consecutive inspection times to be noticeable, more units should be
allocated to high time inspections regardless of the performance dispersion.
The effect of unit-to-unit variability is usually less important than the aging
effect.
ž For the concave decreasing degradation path in Figure 8.10, the degradation
rate is flat at a low time. More units should be assigned to low time inspec-
tions. This principle applies to both constant and decreasing performance
dispersion.
8.5.2 Data Analysis

The following assumptions are needed for analyzing destructive degradation data:
ž The (log) performance characteristic y at any given time has a (transformed)
location-scale distribution (Weibull, normal, or lognormal). Other distribu-
tions are less common, but may be used.
ž The location parameter µy is a function of time t and is denoted by µy (t; β),
where β = [β1 , β2 , . . . , βp ] is the vector of p unknown parameters. µy (t; β)
represents a (transformed) “typical” degradation path. Its specific form is
known from experience or test data. The simplest case is µy as a linear
function of (log) time t:
µy (t; β1 , β2 ) = β1 + β2 t. (8.16)
ž If the scale parameter σy is a function of time t, the specific form of
the function σy (t; θ ) is known from experience or test data, where θ =
[θ1 , θ2 , . . . , θk ] is the vector of k unknown parameters. Often, σy is assumed
to be independent of time. For example, Nelson (1990, 2004) uses a con-
stant σy for insulation breakdown data, and W. Wang and Dragomir-Daescu
(2002) model degradation of the relative orbit area of an induction motor
with an invariable σy .
As described in the test method, the inspection at time tj yields nj measure-
ments yij , where j = 1, 2, . . . , m and i = 1, 2, . . . , nj . Now we want to use the
measurements yij to estimate the reliability at a given time. As described earlier,
the performance characteristic at any time has a (transformed) location-scale dis-
tribution. Let fy (y; t) denote the conditional probability density function (pdf)
of the distribution at time t. fy (y; t) contains the unknown parameters µy (t; β)
and σy (t; θ ), which, in turn, depend on the unknown parameters β and θ . These
parameters can be estimated using the maximum likelihood method. Because the
observations yij come from different test units, they are independent of each
other. The total sample log likelihood can be written as

m
nj
L(β, θ ) = ln[fy (yij ; tj )]. (8.17)
j =1 i=1
Maximizing (8.17) directly gives the estimates of the model parameters β and θ .
If σy is constant and µy (t; β) is a linear function of (log) time as given in (8.16),
then (8.17) will be greatly simplified. In this special case, commercial software
packages such as Minitab and Reliasoft ALTA for accelerated life test analysis
can apply to estimate the model parameters β1 , β2 , and σy . This is done by
treating (8.16) as an acceleration relationship, where tj is considered a stress
level. Once the estimates are computed, the conditional cdf Fy (y; t) is readily
available. Now let’s consider the following cases.
Case 1: Weibull Performance Suppose that the performance characteristic y has
a Weibull distribution with shape parameter βy and characteristic life αy , where
βy is constant and ln(αy ) = β1 + β2 t. Since the measurements are complete exact
data, from (7.59) and (8.17) the total sample log likelihood is

m
nj

L(β1 , β2 , βy ) = ln(βy ) − βy (β1 + β2 tj )
j =1 i=1
yij βy
+ (βy − 1) ln(yij ) − . (8.18)
eβ1 +β2 tj
The estimates β̂1 , β̂2 , and β̂y can be calculated by maximizing (8.18) directly.
As in accelerated life test data analysis, commercial software can be employed
to obtain these estimates. In computation, we treat the performance characteristic
y as life, the linear relationship ln(αy ) = β1 + β2 t as an acceleration model, the
inspection time t as a stress, m as the number of stress levels, and nj as the
number of units tested at “stress level” tj . If y is a monotonically decreasing
characteristic such as strength, the effect of time on y is analogous to that of
stress on life. If y is a monotonically increasing characteristic (mostly in non-
destructive cases), the effects are exactly opposite. Such a difference does not
impair the applicability of the software to this type of characteristic. In this case,
the parameter β2 is positive.
The conditional cdf for y can be written as
β̂y
y
Fy (y; t) = 1 − exp − . (8.19)
eβ̂1 +β̂2 t
Case 2: Lognormal Performance If the performance characteristic y has a log-
normal distribution with scale parameter µy and shape parameter σy , ln(y) has
the normal distribution with mean µy and standard deviation σy . If σy is constant
and µy = β1 + β2 t is used, the total sample log likelihood can be obtained easily
from (7.73) and (8.17). As in the Weibull case, the model parameters may be cal-
culated by maximizing the likelihood directly or by using the existing commercial
software.
The conditional cdf for y can be expressed as

ln(y) − β̂1 − β̂2 t
Fy (y; t) = . (8.20)
σ̂y
Case 3: Normal Performance with Nonconstant σy Sometimes the perfor-

mance characteristic y can be modeled with a normal distribution with mean µy
and standard deviation σy , where
ln(µy ) = β1 + β2 t and ln(σy ) = θ1 + θ2 t. (8.21)
From (7.73), (8.17), and (8.21), the total sample log likelihood is
nj
m
1
L(β1 , β2 , θ1 , θ2 ) = − ln(2π) − θ1 − θ2 tj
j =1 i=1
2

1 yij − exp(β1 + β2 tj ) 2
− . (8.22)
2 exp(θ1 + θ2 tj )
The existing commercial software do not handle nonconstant σy cases; the esti-
mates β̂1 , β̂2 , θ̂1 , and θ̂2 are calculated by maximizing (8.22) directly. This will
be illustrated in Example 8.6.
The conditional cdf for y is

y − exp(β̂1 + β̂2 t)
Fy (y; t) = . (8.23)
exp(θ̂1 + θ̂2 t)
The three cases above illustrate how to determine the conditional cdf for the
performance characteristic. Now we want to relate the performance distribution
to a life distribution. Similar to the nondestructive inspection case described in
Section 8.4.2, the probability of failure at a given time is equal to the probability
of the performance characteristic crossing a threshold at that time. In particular,
if a failure is defined in terms of y ≤ G, the probability of failure at time t equals
the probability of y(t) ≤ G: namely,
F (t) = Pr(T ≤ t) = Pr[y(t) ≤ G] = Fy (G; t). (8.24)
In some simple cases, it is possible to express F (t) in a closed form. For example,
for case 2, the probability of failure is given by

ln(G) − β̂1 − β̂2 t t − [ln(G) − β̂1 ]/β̂2
F (t) = = . (8.25)
σ̂y −σ̂y /β̂2
This indicates that the time to failure has a normal distribution with mean
[ln(G) − β̂1 ]/β̂2 and standard deviation −σ̂y /β̂2 . Note that β2 is negative for
a monotonically decreasing characteristic, and thus −σ̂y /β̂2 is positive.
Example 8.6 In Section 5.13 we presented a case study on the robust reliabil-
ity design of IC wire bonds. The purpose of the study was to select a setting
of bonding parameters that maximizes robustness and reliability. In the experi-
ment, wire bonds were generated with different settings of bonding parameters
according to the experimental design. The bonds generated with the same setting
were divided into two groups each with 140 bonds. One group underwent level
1 thermal cycling and the other group was subjected to level 2. For each group
a sample of 20 bonds were sheared at 0, 50, 100, 200, 300, 500, and 800 cycles,
respectively, for the measurement of bonding strength. In this example, we want
to estimate the reliability of the wire bonds after 1000 cycles of level 2 thermal
cycling, where the bonds were generated at the optimal setting of the bonding
parameters. The optimal setting is a stage temperature of 150◦ C, ultrasonic power
of 7 units, bonding force of 60 gf, and bonding time of 40 ms. As described in
Section 5.13, the minimum acceptable bonding strength is 18 grams.
SOLUTION The strength measurements [in grams (g)] at each inspection time
can be modeled with a normal distribution. Figure 8.11 shows normal fits to the
strength data at, for example, 0, 300, and 800 cycles. In Section 5.13 we show
that the normal mean and standard deviation decrease with the number of ther-
mal cycles. Their relationships can be modeled using (8.21). Figure 8.12 plots the
relationships for the wire bonds that underwent level 2 thermal cycling. Simple
linear regression analysis gives β̂1 = 4.3743, β̂2 = −0.000716, θ̂1 = 2.7638, and
θ̂2 = −0.000501. Substituting these estimates into (8.23) and (8.24) can yield an
estimate of the probability of failure at a given time. To improve the accuracy of
the estimate, we use the maximum likelihood method. The parameters are esti-
mated by maximizing (8.22), where yij are the strength measurements, m = 7,
and nj = 20 for all j . The estimates obtained from linear regression analysis serve
as the initial values. The maximum likelihood estimates are β̂1 = 4.3744, β̂2 =
−0.000712, θ̂1 = 2.7623, and θ̂2 = −0.000495. From (8.23) and (8.24), the esti-
mate of reliability at 1000 cycles is

18 − exp(4.3744 − 0.000712 × 1000)
R(1000) = 1 − = 0.985.
exp(2.7623 − 0.000495 × 1000)
99 300 cycles
95
90
800 cycles 0 cycles
80
70
Percent
60
50
40
30
20
10
5
1
20 30 40 50 60 70 80 90 100 110 120
Bonding strength (g)
FIGURE 8.11 Normal fits to the strength data at different cycles
STRESS-ACCELERATED DEGRADATION TESTS 351
4.5
4
ln(my)
3.5
3
ln(sy)
2.5
2
0 200 400 600 800 1000
t (cycles)
FIGURE 8.12 ln(µy ) and ln(σy ) versus the number of cycles
8.6 STRESS-ACCELERATED DEGRADATION TESTS
In the preceding sections we described techniques for degradation analysis at

a single stress level, which may be a use stress level or an elevated level. In
many situations, testing a sample at a use stress level yields a small amount of
degradation in a reasonable length of time. Insufficient degradation inevitably
provides biased reliability estimates. To overcome this problem, as in accelerated
life testing, two or more groups of test units are subjected to higher-than-use
stress levels. The degradation data so obtained are extrapolated to estimate the
reliability at a use stress level. In this section we present methods for accelerated
degradation analysis.
8.6.1 Pseudo Accelerated Life Test Method

Suppose that nk units are tested at stress level Sk , where k = 1, 2, . . . , q, and q
is the number of stress levels. Tested at Sk , unit i (i = 1, 2, . . . , nk ) is inspected
at time tij k and produces the measurement yij k , where j = 1, 2, . . . , mik and
mik is the number of measurements on unit i tested at Sk . Having obtained the
measurements, we calculate the approximate life tîk of unit i tested at Sk using the
pseudolife method described in Section 8.3. Treat the life data tˆ1k , tˆ2k , . . . , tˆnk k as
if they came from an accelerated life test at Sk . Then the life distribution at the use
stress level can be estimated from q sets of such life data by using the graphical
or maximum likelihood method presented in Section 7.7. The calculation is illus-
trated in Example 8.7.
Example 8.7 Stress relaxation is the loss of stress in a component subjected to

a constant strain over time. The contacts of electrical connectors often fail due to
excessive stress relaxation. The relationship between stress relaxation and aging
time at temperature T can be expressed as

s Ea
= At exp −
B
, (8.26)
s0 kT
30
25
100 °C
20 85 °C
65 °C
∆s
s0
15
10
0
0 500 1000 1500 2000 2500 3000
t (hr)
FIGURE 8.13 Stress relaxation at different times and temperatures
where s is the stress loss by time t, s0 the initial stress, the ratio s/s0 the stress
relaxation (in percent), A and B are unknowns, and other notation is as in (7.4).
Here A usually varies from unit to unit, and B is a fixed-effect parameter. At a
given temperature, (8.26) can be written as

s
ln = β1 + β2 ln(t), (8.27)
s0
where β1 = ln(A) − Ea /kT and β2 = B. To estimate the reliability of a type of

connectors, a sample of 18 units was randomly selected and divided into three
equal groups, which were tested at 65, 85, and 100◦ C. The stress relaxation data
measured during testing are plotted in Figure 8.13. The electrical connector is
said to have failed if the stress relaxation exceeds 30%. Estimate the probability
of failure of the connectors operating at 40◦ C (use temperature) for 15 years
(design life).
SOLUTION First we fit (8.27) to each degradation path, and estimate the para-
meters β1 and β2 for each unit using the least squares method. Figure 8.14 plots
the fits of the degradation model to the data and indicates that the model is
adequate. Then we calculate the approximate lifetime of each test unit using
tˆ = exp{[ln(30) − β̂1 ]/β̂2 }. The resulting lifetimes are 15,710, 20,247, 21,416,
29,690, 41,167, and 42,666 hours at 65◦ C; 3676, 5524, 7077, 7142, 10,846, and
10,871 hours at 85◦ C; and 1702, 1985, 2434, 2893, 3343, and 3800 hours at
100◦ C. The life data are plotted on the lognormal probability paper, as shown in
Figure 8.15. It is seen that the lifetimes at the three temperatures are reasonably
lognormal with a common shape parameter σ .
30
25
100 °c
85 °c
20
65 °c
15
∆s
s0
10
0
0 500 1000 1500 2000 2500 3000
t (h)
FIGURE 8.14 Degradation model fitted to the measurement data of stress relaxation
99
95
90
100 °C 85 °C 65 °C
80
70
Percent
60
50
40
30
20
10
5
1
1000 10,000 100,000
Time to failure (h)
FIGURE 8.15 Lognormal fits to the pseudolifetimes at different temperatures
Since a failure occurs when s/s0 ≥ 30, from (8.26) the nominal life can be
written as 1/B
30 Ea
t= exp .
A kBT
Thus, it is legitimate to assume that the lognormal scale parameter µ is a lin-

ear function of 1/T : namely, µ = γ0 + γ1 /T , where γ0 and γ1 are constants.
By using the maximum likelihood method for accelerated life data analysis as
described in Section 7.7, we obtain the ML estimates as γ̂0 = −14.56, γ̂1 =
8373.35, and σ̂ = 0.347. The estimate of the scale parameter at 40◦ C is µ̂ =
−14.56 + 8373.35/313.15 = 12.179. Then the probability of failure of the con-
nectors operating at 40◦ C for 15 years (131,400 hours) is

ln(131,400) − 12.179
F (131,400) = = 0.129 or 12.9%.
0.347
That is, an estimated 12.9% of the connectors will fail by 15 years when used
at 40◦ C.
8.6.2 Random-Effect Method

In Section 8.4 we delineated use of the random-effect model to analyze degrada-
tion data obtained at a single stress level. Now we extend the analysis to multiple
stress levels. The analysis still consists of two steps: estimating degradation model
parameters and evaluating the probability of failure. As in Section 8.4, the model
parameters can be estimated using the maximum likelihood method, but the cal-
culation requires much effort. For computational simplicity, readers may use the
data analysis method for destructive inspection described in Section 8.6.3 or the
pseudo accelerated life test method presented in Section 8.6.1.
Estimating the Model Parameters The degradation model g(t; β1 , β2 , . . . , βp )

describes the true degradation path at a given stress level. The stress effect
is implicitly embedded into one or more parameters among β1 , β2 , . . . , βp . In
other words, at least one of these parameters depends on the stress level. For
example, β1 in Example 8.7 depends on temperature, whereas β2 does not.
The temperature effect on stress relaxation can be described by ln(s/s0 ) =
ln(A) − Ea /kT + β2 ln(t), where the parameters A and β2 vary from unit to
unit, and the activation energy Ea is usually of fixed-effect type. In general,
because of the additional stress term, the total number of model parameters
increases, but is still denoted by p here for notional convenience. Thus, the
true degradation path can be written as g(t, S; β1 , β2 , . . . , βp ), where S denotes
the (transformed) stress, and the model parameters are free of stress effects. The
mean vector µβ and the variance–covariance matrix β may be estimated directly
by maximizing the likelihood (8.6), where the true degradation path is replaced
with g(tij , Si ; β1i , β2i , . . . , βpi ) and Si is the stress level of unit i. Note that the
stress parameters are constant and the corresponding variances and covariance
elements are zero. As pointed out earlier, estimating µβ and β from (8.6) is
computationally intensive.
Evaluating the Probability of Failure Similar to (8.12), for a monotonically

increasing characteristic, the probability of failure at the use stress level S0 can
be expressed as
F (t) = Pr[g(t, S0 ; β1 , β2 , . . . , βp ) ≥ G]. (8.28)

In some simple cases, F (t) can be expressed in closed form. Let’s consider
the true degradation path
g(t, S) = β1 + γ S + β2 t,
where β1 and β2 have bivariate normal distribution and γ is a constant. The

probability of failure at the use stress level S0 is
F (t) = Pr[g(t, S0 ) ≥ G] = Pr(β1 + γ S0 + β2 t ≥ G)

µβ1 + γ S0 + tµβ2 − G
= , (8.29)
(σβ21 + t 2 σβ22 + 2tρ12 σβ1 σβ2 )1/2
which is a closed-form expression for F (t). In general, if F (t) can be expressed

in closed form, the evaluation is done simply by substituting estimates of µβ
and β into the F (t) expression. Otherwise, we use the Monte Carlo simulation
method described in Section 8.4.
8.6.3 Random-Process Method

In Section 8.5 we described the random-process method for degradation analysis
of destructive and nondestructive measurements from a single stress level. As
shown below, the method can be extended to multiple stress levels. The extension
uses the following assumptions:
ž The (log) performance characteristic y at a given time t and stress S has

a (transformed) location-scale distribution (Weibull, normal, or lognormal).
Other distributions are less common, but can be used.
ž The location parameter µy is a monotone function of t and S. The function
is denoted by µy (t, S; β), where β = [β1 , β2 , . . . , βp ]. The specific form of
µy (t, S; β) is known from experience or test data. The simplest case is µy
as a linear function of (log) time and (transformed) stress:
µy (t, S; β) = β1 + β2 t + β3 S. (8.30)
ž If the scale parameter σy is a function of t and S, the specific form of

the function σy (t, S; θ ) is known from experience or test data, where θ =
[θ1 , θ2 , . . . , θk ]. Often, σy is assumed to be independent of t and S.
Estimating the Model Parameters For an accelerated degradation test with
destructive inspections, the test method described in Section 8.5 applies to each
of k stress levels. Let yij k denote measurement at time tj k on unit i of stress
level Sk , where i = 1, 2, . . . , nj k , j = 1, 2, . . . , mk , and k = 1, 2, . . . , q; nj k is
the number of units inspected at tj k ; mk is the qnumber k of inspections at Sk ; and
q is the number of stress levels. Obviously, k=1 m j =1 nj k = n, where n is the
total sample size. Let fy (y; t, S) denote the pdf of y distribution conditional on
t and S. Similar to (8.17), the total sample log likelihood can be expressed as

q mk nj k
L(β, θ ) = ln[fy (yij k ; tj k , Sk )]. (8.31)
k=1 j =1 i=1
For nondestructive inspections, the notations above are slightly different. yij k
denotes the measurement at time tij k on unit i of stress level Sk , where i =
1, 2, . . . , nk , j = 1, 2, . . . , mik , and k = 1, 2, . . . , q; nk is the number of units at
S
k ; and mik is the number of inspections on unit i of stress level Sk . Clearly,
q
k=1 nk = n. The total sample log likelihood is similar to (8.31), but the nota-
tion changes accordingly. Note that the likelihood may be approximately cor-
rect for nondestructive inspections because of potential autocorrelation among
the measurements. To reduce the autocorrelation, inspections should be widely
spaced.
Maximizing the log likelihood directly yields estimates of the model parame-
ters β and θ . If σy is constant and µy (t, S; β) is a linear function of (log) time
and (transformed) stress given by (8.30), then (8.31) will be greatly simplified. In
this special case, commercial software packages for accelerated life test analysis
can be used to estimate the model parameters β1 , β2 , β3 , and σy . In calculation,
we treat (8.30) as a two-variable acceleration relationship, where the time t is
considered as a stress.
Evaluating the Probability of Failure After obtaining the estimates β̂ and θ̂,
we can calculate the conditional cdf for y, denoted by Fy (y; t, S). If a failure
is defined in terms of y ≤ G, the probability of failure at time t and use stress
level S0 is given by
F (t, S0 ) = Fy (G; t, S0 ). (8.32)
For example, if y has the lognormal distribution with µy modeled by (8.30) and
constant σy , the estimate of the probability of failure at t and S0 is

ln(G) − β̂1 − β̂2 t − β̂3 S0
F (t, S0 ) =
σ̂y

t − [ln(G) − β̂1 − β̂3 S0 ]/β̂2
= . (8.33)
−σ̂y /β̂2
Note that (8.33) is similar to (8.25).

The following example illustrates application of the random-process method
to nondestructive inspections.
Example 8.8 Refer to Example 8.7. Using the random-process method, estimate
the probability of failure of connectors operating at 40◦ C (use temperature) for
15 years (design life).
99 99
95 95
90 90
80 80
70 70
Percent
Percent
60 60
50 50
40 40
30 30
20 20
235 h 1355 h 2848 h 105 h 1025 h 1842 h
10 10
5 5
1 1
1 10 1 10
Stress relaxation Stress relaxation
(a) (b)
99
95
90
80
70
Percent
60
50
40
30
20
50 h 633 h 1238 h
10
5
1
10 100
Stress relaxation
(c)
FIGURE 8.16 Lognormal fits to stress relaxation measurements at (a) 65◦ C, (b) 85◦ C,
and (c) 100◦ C
SOLUTION Figures 8.16 shows examples of lognormal fits to measurements

of stress relaxation at different times and temperatures. It is seen that the stress
relaxation can be approximated adequately using lognormal distribution and the
shape parameter is reasonably independent of time and temperature. From (8.26),
the scale parameter can be written as (8.30), where t is the log time and S =
1/T . In this example we have q = 3, n1 = n2 = n3 = 6, and mi1 = 11, mi2 =
10, mi3 = 10 for i = 1, 2, . . . , 6. The total sample log likelihood is

3 6 mik
1
L(β, σy ) = − ln(2π) − ln(σy )
k=1 i=1 j =1
2

1 β3 2
− ln(yij k ) − β1 − β2 ln(tij k ) − .
2σy2 Tk
Directly maximizing the likelihood yields the estimates β̂1 = 9.5744, β̂2 = 0.4519,
β̂3 = −3637.75, and σ̂y = 0.1532. The calculation was performed using the
Solver feature of Microsoft Excel. Alternatively, as described earlier, we may
treat the measurement data as if they came from an accelerated life test. The pseu-
dotest involves two stresses (temperature and time), and the acceleration model
0.8
random-process
method
0.6
F(t)
pseudo accelerated
0.4
life test method
0.2
0
0 10 20 30 40 50
t (yr)
FIGURE 8.17 Probabilities of failure calculated from two methods
combines the Arrhenius relationship and the inverse power relationship. Minitab
gave the estimates β̂1 = 9.5810, β̂2 = 0.4520, β̂3 = −3640.39, and σ̂y = 0.1532,
which are close to these from the Excel calculation and are used for subsequent
analysis.
Because stress relaxation is a monotonically increasing characteristic, the prob-
ability of failure at t and S0 is the complement of the probability given in (8.33)
and can be written as

ln(t) − [ln(G) − β̂1 − β̂3 S0 ]/β̂2 ln(t) − 12.047
F (t) = = . (8.34)
σ̂y /β̂2 0.3389
This shows that the time to failure has a lognormal distribution with scale param-
eter 12.047 and shape parameter 0.3389. Note that, in Example 8.7, the pseudo
accelerated life test method resulted in a lognormal distribution with scale and
shape parameters equal to 12.179 and 0.347. At the design life of 15 years
(131,400 hours), from (8.34) the probability of failure is 0.2208. This estimate
should be more accurate than that in Example 8.7. For comparison, the probabili-
ties at different times calculated from the two methods are plotted in Figure 8.17.
It is seen that the random-process method always gives a higher probability of
failure than the other method in this case. In general, the random-process method
results in more accurate estimates.
8.7 ACCELERATED DEGRADATION TESTS WITH TIGHTENED

THRESHOLDS
As shown earlier, degradation analysis often involves complicated modeling and
intensive computation. In contrast, as presented in Chapter 7, methods for the
analysis of accelerated life test data are fairly easy to implement. The methods,
however, become inefficient when there are no or few failures at low stress
levels. Such difficulties may be overcome by measuring performance degradation
in accelerated life testing and tightening a failure threshold. Because life is
a function of threshold, tightening a threshold produces more failures. Here,
ACCELERATED DEGRADATION TESTS WITH TIGHTENED THRESHOLDS 359
tightening a threshold refers to reducing the threshold for a monotonically increas-

ing characteristic and increasing the threshold for a monotonically decreasing one.
In this section we discuss the relationship between threshold and life and describe
a test method and life data analysis.
8.7.1 Relationship Between Life and Threshold

A degrading product is said to have failed if a performance characteristic crosses
its failure threshold. The life is the time at which the characteristic reaches the
threshold. Apparently, the life of a unit depends on the value of the threshold.
For a monotonically increasing characteristic, the smaller the value, the shorter
the life, and vice versa. A unit that does not fail at a usual threshold may have
failed at a tightened one. Thus, a tightened threshold yields more failures in a
censored test. In this sense, life can be accelerated by tightening a threshold. Such
an acceleration method does not induce failure modes different from those at the
usual threshold because failures at different values of a threshold are caused by
the same degradation mechanisms. The relationship between life and threshold
for a monotonically increasing characteristic is depicted in Figure 8.18, where
G0 is the usual threshold, G1 and G2 are reduced thresholds, and f (t|Gi ) is the
pdf of life conditional on Gi .
A quantitative relationship between life and threshold is useful and impor-
tant. The threshold for a component is usually specified by the manufacturer of
the component to meet functional requirements in most applications. The spec-
ification is somewhat subjective and arbitrary. Certainly, it can be changed in
specific system designs. In other words, the value of a threshold for a compo-
nent depends on the system specifications. For example, a component installed
in a critical military system is often required to have a tighter threshold than
when installed in a commercial product. Provided with the relationship, system
designers can estimate the reliability of the component at a desirable value of
the threshold without repeating reliability testing. The life–threshold relation-
ship finds applications in the design of automobile catalytic converters, which
convert the engine exhausts (hydrocarbon and NOx , a group of highly reactive
gases containing nitrogen and oxygen in varying amounts) to less harmful gases.
f (t| G0)
f (t| G1)
G0
f (t| G2)
G1
y
G2
0
t
FIGURE 8.18 Relationship between life and threshold for an increasing characteristic
The performance of the catalytic converter is characterized by the index ratio,

which increases with mileage. A smaller threshold for the index ratio is needed
to reduce the actual emission level, but that increases warranty repairs. In the
early design phase, design engineers often rely on the life–threshold relationship
to evaluate the impact on warranty costs of lowering the index ratio threshold.
Example 8.9 In Example 8.7 the usual threshold for stress relaxation is 30%.
Reducing the threshold shortens the time to failure. Determine the relationship
between the time to failure and the threshold, and evaluate the effect of threshold
on the probability of failure.
SOLUTION From (8.34), the estimate of the mean log life (location parameter
µt ) is
ln(G) − β̂1 − β̂3 S0
µ̂t = = 4.5223 + 2.2124 ln(G),
β̂2
where β̂1 , β̂2 , and β̂3 are as obtained in Example 8.8. It is seen that the mean log
life is a linear function of the log threshold. The influence of the threshold on
life is significant because of the large slope. Figure 8.19 plots the probabilities of
failure with thresholds 30%, 25%, and 20%. It is seen that reducing the threshold
greatly increases the probability of failure. For instance, at a design life of 15
years, the probabilities of failure at the three thresholds are 0.2208, 0.6627, and
0.9697, respectively.
8.7.2 Test Method

The test uses a sample of size n, q stress levels and m tightened thresholds.
q of nk units is tested at stress level Sk (k = 1, 2, . . . , q) until time ηk .
A group
Here k=1 nk = n. The times are recorded when the performance measurement
of a unit reaches the tightened thresholds G1 , G2 , . . . , Gm , where G1 is closest
to and Gm is farthest from G0 (the usual threshold). As a result, a test unit
0.8
0.6 20% 25% 30%

F(t)
0.4
0.2
0
0 10 20 30 40 50
t (yr)
FIGURE 8.19 Probabilities of failure with different thresholds
can have m “life” observations. Let tij k denote the “failure” time of unit i at Sk
and at Gj , where i = 1, 2, . . . , nk , j = 1, 2, . . . , m, and k = 1, 2, . . . , q. The life
distribution at the use stress level S0 and the usual threshold G0 is estimated by
utilizing these “life” data.
The most severe threshold Gm should be as tight as possible to maximize the
threshold range and the number of failures at Gm , but it must not fall in the
fluctuating degradation stage caused by the burn-in effects at the beginning of a
test. The space between two thresholds should be as wide as possible to reduce
the potential autocorrelation among the failure times. For this purpose, we usually
use m ≤ 4. In Section 8.8 we describe optimal design of the test plans.
8.7.3 Life Data Analysis

The life data analysis uses the following assumptions:
ž The (log) time to failure at a given threshold has a (transformed) location-

scale distribution (Weibull, normal, or lognormal). Other distributions are
less common, but may be used.
ž The scale parameter σ does not depend on stress and threshold.
ž The location parameter µ is a linear function of (transformed) stress and
threshold: namely,
µ(S, G) = β1 + β2 S + β3 G. (8.35)
where β1 , β2 , and β3 are parameters to be estimated from test data. Nonlinear

relationships may be used, but the analysis will be greatly complicated.
The life data analysis is to estimate the life distribution at the use stress
level and the usual threshold. This can be done by using the graphical or max-
imum likelihood method described in Section 7.7. The analysis is illustrated in
Example 8.10.
Example 8.10 Infrared light-emitting diodes (IRLEDs) are high-reliability opto-

electronic devices used widely in communication systems. The devices studied
here are GaAs/GaAs IRLEDs, the wavelength of which is 880 nm and the design
operating current, 50 mA. The performance of the devices is measured mainly
by the variation ratio of luminous power. A failure is said to have occurred if
the ratio is greater than 30%. To estimate the reliability at the operating current
of 50 mA, 40 units were sampled and divided into two groups. A group of 25
units was tested at 170 mA; the group of 15 units was tested at 320 mA. Two
reduced thresholds (10% and 19%) were used in testing. Table 8.2 summarizes
the test plan.
The test units were inspected for luminous power before and during testing.
The inspections were scheduled at 0, 24, 48, 96, 155, 368, 768, 1130, 1536,
1905, 2263, and 2550 hours for the 170 mA group, and at 0, 6, 12, 24, 48, 96,
156, 230, 324, 479, and 635 hours for the 320 mA group. Figure 8.20 shows
TABLE 8.2 Accelerated Degradation Test Plan for the IRLEDs

Current Threshold Censoring Time
Group Sample Size (mA) (%) (h)
1 15 320 19 635
2 15 320 10 635
3 25 170 19 2550
4 25 170 10 2550
40
Power degradation (%)
20
10
500 1500 2500

Aging time (h)
(a)
80
70
60
Power degradation(%)
50
40
30
20
10
0
0 100 200 300 400 500 600 700

Aging time (h)
(b)
FIGURE 8.20 Degradation of luminous power at (a) 170 mA and (b) 320 mA
the values of the variation ratio at each inspection time. The data are shown in
Tables 8.9 and 8.10 of Problem 8.8.
The variation ratio of luminous power is a function of time and current, and
can be written as
A
y = B tC, (8.36)
I
TABLE 8.3 Failure Times (Hours) at Different

Levels of Current and Threshold
Test 320 mA 170 mA
Unit 19% 10% 19% 10%
1 176 39 1751 1004

2 544 243 2550a 735
3 489 117 2550a 2550a
4 320 183 2550a 2550a
5 108 35 1545 867
6 136 47 2550a 1460
7 241 113 2550a 1565
8 73 27 1480 710
9 126 59 2550a 1389
10 92 35 2550a 2550a
11 366 74 2550a 1680
12 158 58 2550a 2550a
13 322 159 2550a 873
14 220 64 2550a 2550a
15 325 195 2550a 2138
16 2550a 840
17 2550a 2550a
18 2550a 2550a
19 2550a 1185
20 2550a 898
21 2550a 636
22 2550a 2329
23 2550a 1965
24 2550a 2550a
25 1611 618
a
Censored.
where y is the variation ratio, I the current, A and C random-effect parameters,

and B a fixed-effect constant.
To calculate the approximate failure times at a threshold of 10%, we fitted the
log transformed (8.36) to each of the degradation paths whose y values are greater
than or equal to 10% at their censoring times. Interpolating y to 10% gives the
failure times at this threshold. The units with y value less than 10% are considered
censored. To calculate the failure times at the threshold of 19%, we interpolated y
to 19% for the units with y values greater than or equal to 19% at their censoring
times. The remaining units are considered censored. The resulting exact failure
times are shown in Table 8.3. Alternatively, the failure times can be expressed
as interval data by using the measurements and corresponding inspection times.
This approach does not require fitting degradation model and interpolation. For
example, the measurement on unit 1 of 170 mA was 16% at the inspection time
of 1536 hours, and 22.5% at 1905 hours. If the reduced threshold is 19%, the
unit failed between the two consecutive inspections. Therefore, the failure time
is within the interval [1536,1905]. In contrast, the exact time from interpolation
is 1751.
Figure 8.21 plots the lognormal fits to the failure times listed in Table 8.3. It
is shown that the shape parameter σ is approximately independent of current and
threshold. In addition, from (8.36), the scale parameter µ is a linear function of
the log current and the log threshold: namely,
µ(I, G) = β1 + β2 ln(I ) + β3 ln(G). (8.37)
The estimates of the model parameters were computed using Minitab as β̂1 =
28.913, β̂2 = −4.902, β̂3 = 1.601, and σ̂ = 0.668. The mean log life at the
operating current of 50 mA and the usual threshold of 30% is µ̂ = 15.182. The
probability of failure in 10 years (87,600 hours) is negligible, so the device has
ultrahigh reliability.
8.8 ACCELERATED DEGRADATION TEST PLANNING

Like accelerated life tests, degradation tests merit careful planning. In this section
we present compromise test plans for accelerated degradation tests with tightened
thresholds and survey briefly other degradation test plans.
8.8.1 Compromise Plans for Tightened Threshold Tests

The test plans here use the 22 full factorial design in Figure 7.21, where S1
is a two-level stress and S2 is a two-level threshold. The sample is divided
into two groups. One group is tested at the low level of S1 , and the other is
subjected to the high level. The two groups employ two levels of threshold. Since
99 170 mA, 10%

320 mA, 10%
95
90 320 mA, 19%
80
70
Percent
60
50
40
30
20
170 mA, 19%
10
5
1
10 100 1000 10,000
Time to failure (h)
FIGURE 8.21 Lognormal fits to the failure times at different levels of current and
threshold
ACCELERATED DEGRADATION TEST PLANNING 365
the threshold is thought of as a stress, the accelerated degradation tests can be

perceived of as accelerated life tests with two accelerating variables. As such, with
the assumptions stated in Section 8.7.3, the compromise test plans for accelerated
degradation tests can be developed using the methods for accelerated life tests
described in Sections 7.8.4 and 7.8.5. We use the notation in Sections 7.8.4, and
choose ξ11 (low stress level), ξ21 (low threshold level), and π11 (proportion of
sample size allocated to ξ11 ), to minimize the asymptotic variance of the MLE of
the mean log life at the use stress and the usual threshold. Note the uniqueness of
the test plans here: π11 = π12 , π21 = π22 , and π11 + π22 = 1. G. Yang and Yang
(2002) describe in detail development of the test plans.
Tables 8.4 and 8.5 list the test plans for Weibull and lognormal distributions,
respectively. In the tables, a1 is the standardized censoring time at S1 , and a2
at S2 . It is seen that a few plans have ξ11 , ξ21 , or both at the use (usual) level.
In these cases, the tests become partially accelerated degradation tests. To find a
plan from the tables, one looks up the value of c first, then b, a2 , and a1 in order.
Linear interpolation may be needed for a combination (a1 , a2 , b, c) not given in
the tables. Extrapolation outside the tables is not valid. Instead, use a numerical
algorithm for solution. After obtaining the standardized values, we convert them
to the actual values using the formulas given in Section 7.8.4.
The test plans depend on the values of µ00 , µ22 , µ02 , µ20 , and σ . They
are unknown at the test planning stage. Generally, µ00 can be approximated by
using experience, similar data, or a reliability prediction handbook such as MIL-
HDBK-217F (U.S. DoD, 1995), as described in Chapter 7. µ22 , µ20 , and σ can
be preestimated using a preliminary degradation test conducted at a high level of
stress. Since G2 is the tightest threshold, most test units will fail at this threshold
when the test is terminated. The failure times are used to preestimate µ22 and σ .
On the other hand, the preestimate of µ20 can be calculated by employing the
approximate pseudolife method described earlier in the chapter. The unknown
µ02 is estimated by µ00 + µ22 − µ20 , provided that (8.35) holds.
Example 8.11 A type of new tire has a nominal tread depth of 10 32

inch and
2
must be retreaded when the tread is worn to 32 inch. To evaluate the wear life
of the tires under a load of 1350 pounds, a sample of 35 units is to be placed
in test. The sample is divided into two groups, one loaded with the maximum
allowable load of 2835 pounds, and the other with a lighter load. Each group
8
will be run for 13,000 miles and subject to two thresholds, with one being 32
inch (the tightest threshold). The wear life (in miles) can be modeled with the
Weibull distribution with µ(P , G) = β1 + β2 ln(P ) + β3 ln(G), where µ is the
log Weibull characteristic life, P the load, and G the threshold. A preliminary
test was conducted on six tires loaded with 2835 pounds of pressure and run for
6500 miles. The tread loss was measured every 500 miles until test termination.
The degradation data yielded the preestimates µ22 = 8.56, µ20 = 9.86, and σ =
0.45. The fleet test data of the prior-generation tires resulted in the preestimate
µ00 = 11.32. Develop a compromise test plan for the experiment.
TABLE 8.4 Compromise Test Plans for a Weibull Distribution

No. a1 a2 b c π11 ξ11 ξ21 V No. a1 a2 b c π11 ξ11 ξ21 V
1 2 2 2 2 0.542 1.000 1.000 5.8 46 4 3 3 4 0.399 1.000 0.775 7.8
2 3 2 2 2 0.842 0.893 1.000 3.1 47 5 3 3 4 0.792 0.756 0.746 4.5
3 2 2 3 2 0.559 0.807 1.000 10.1 48 2 2 4 4 0.429 0.617 0.587 31.1
4 3 2 3 2 0.669 0.775 1.000 5.4 49 3 2 4 4 0.456 0.599 0.604 19.6
5 4 2 3 2 0.850 0.843 1.000 2.9 50 4 2 4 4 0.675 0.541 0.599 11.5
6 2 2 4 2 0.575 0.625 0.930 16.3 51 5 3 4 4 0.714 0.645 0.687 6.6
7 3 2 4 2 0.622 0.664 1.000 8.5 52 4 4 4 4 0.450 0.992 0.995 7.9
8 4 2 4 2 0.745 0.710 1.000 4.6 53 5 4 4 4 0.432 1.000 0.987 6.1
9 2 2 5 2 0.577 0.506 0.812 24.2 54 6 4 4 4 0.782 0.749 0.790 4.0
10 3 2 5 2 0.612 0.555 0.918 12.4 55 2 2 5 4 0.458 0.502 0.557 41.7
11 4 2 5 2 0.689 0.614 0.960 6.8 56 3 2 5 4 0.468 0.516 0.588 25.2
12 2 2 2 3 0.416 1.000 0.821 10.0 57 4 2 5 4 0.633 0.479 0.568 15.2
13 3 2 2 3 0.726 0.770 0.787 6.2 58 3 3 5 4 0.468 0.636 0.757 20.1
14 4 2 2 3 0.908 0.891 0.894 3.1 59 4 3 5 4 0.438 0.732 0.779 13.5
15 2 2 3 3 0.456 0.806 0.781 15.6 60 5 3 5 4 0.666 0.566 0.644 8.9
16 3 2 3 3 0.556 0.714 0.777 9.6 61 5 4 5 4 0.635 0.589 0.665 8.8
17 4 2 3 3 0.771 0.716 0.808 5.2 62 6 4 5 4 0.723 0.652 0.732 5.5
18 3 3 3 3 0.473 1.000 1.000 7.2 63 2 2 6 4 0.478 0.423 0.524 54.0
19 4 3 3 3 0.501 1.000 0.984 5.0 64 3 2 6 4 0.482 0.447 0.566 31.7
20 5 3 3 3 0.858 0.852 0.932 2.8 65 4 2 6 4 0.603 0.432 0.545 19.3
21 2 2 4 3 0.492 0.621 0.731 23.1 66 3 3 6 4 0.488 0.534 0.735 26.4
22 3 2 4 3 0.522 0.632 0.765 13.5 67 4 3 6 4 0.463 0.617 0.765 17.2
23 4 2 4 3 0.695 0.611 0.752 7.8 68 5 3 6 4 0.634 0.507 0.610 11.6
24 3 3 4 3 0.509 0.791 1.000 10.9 69 5 4 6 4 0.602 0.533 0.639 11.5
25 4 3 4 3 0.487 0.901 1.000 7.1 70 6 4 6 4 0.686 0.578 0.686 7.3
26 5 3 4 3 0.757 0.721 0.852 4.4 71 2 2 7 4 0.492 0.365 0.490 68.0
27 2 2 5 3 0.512 0.505 0.672 32.4 72 3 2 7 4 0.495 0.392 0.540 39.1
28 3 2 5 3 0.529 0.535 0.728 18.3 73 4 2 7 4 0.582 0.394 0.525 23.8
29 4 2 5 3 0.648 0.538 0.712 10.7 74 3 3 7 4 0.502 0.460 0.707 33.8
30 3 3 5 3 0.530 0.640 0.960 15.5 75 4 3 7 4 0.480 0.534 0.746 21.5
31 4 3 5 3 0.513 0.738 0.996 9.8 76 5 3 7 4 0.612 0.460 0.582 14.5
32 5 3 5 3 0.698 0.628 0.794 6.2 77 4 4 7 4 0.514 0.575 0.977 18.5
33 2 2 6 3 0.524 0.425 0.614 43.4 78 5 4 7 4 0.489 0.676 0.998 12.7
34 3 2 6 3 0.538 0.459 0.682 23.9 79 6 4 7 4 0.467 0.783 1.000 9.4
35 4 2 6 3 0.618 0.480 0.680 14.0 80 2 2 2 5 0.280 1.000 0.516 22.1
36 3 3 6 3 0.541 0.536 0.905 21.2 81 3 2 2 5 0.672 0.600 0.501 15.3
37 4 3 6 3 0.528 0.621 0.956 13.0 82 4 2 2 5 0.811 0.698 0.581 8.1
38 5 3 6 3 0.661 0.558 0.749 8.3 83 2 2 3 5 0.331 0.800 0.502 30.2
39 2 2 2 4 0.335 1.000 0.635 15.5 84 3 2 3 5 0.546 0.538 0.483 21.0
40 3 2 2 4 0.388 0.875 0.622 10.9 85 4 2 3 5 0.721 0.568 0.534 11.8
41 4 2 2 4 0.844 0.778 0.703 5.3 86 3 3 3 5 0.337 1.000 0.638 14.3
42 2 2 3 4 0.384 0.803 0.613 22.3 87 4 3 3 5 0.332 1.000 0.633 11.2
43 3 2 3 4 0.535 0.620 0.597 14.8 88 5 3 3 5 0.757 0.685 0.625 6.6
44 4 2 3 4 0.737 0.630 0.641 8.2 89 2 2 4 5 0.379 0.613 0.487 40.2
45 3 3 3 4 0.394 1.000 0.790 10.5 90 3 2 4 5 0.417 0.563 0.494 26.9
TABLE 8.4 (continued )

91 4 2 4 5 0.665 0.489 0.501 15.9 136 6 5 8 5 0.470 0.710 0.992 12.1
92 3 3 4 5 0.384 0.782 0.629 19.3 137 2 2 3 6 0.291 0.798 0.425 39.2
93 4 3 4 5 0.346 0.894 0.640 14.1 138 3 2 3 6 0.558 0.480 0.408 28.2
94 5 3 4 5 0.690 0.587 0.580 9.2 139 4 2 3 6 0.712 0.520 0.460 15.9
95 4 4 4 5 0.392 0.988 0.799 10.4 140 3 3 3 6 0.294 1.000 0.535 18.8
96 5 4 4 5 0.370 1.000 0.800 8.3 141 4 3 3 6 0.285 1.000 0.534 15.1
97 6 4 4 5 0.740 0.686 0.670 5.6 142 5 3 3 6 0.737 0.629 0.539 9.1
98 2 2 5 5 0.413 0.498 0.470 52.0 143 2 2 4 6 0.339 0.610 0.415 50.5
99 3 2 5 5 0.421 0.498 0.488 33.4 144 3 2 4 6 0.416 0.504 0.413 35.3
100 4 2 5 5 0.627 0.434 0.476 20.4 145 4 2 4 6 0.661 0.449 0.432 20.8
101 3 3 5 5 0.419 0.632 0.620 25.3 146 4 3 4 6 0.303 0.890 0.537 18.5
102 4 3 5 5 0.383 0.726 0.635 17.8 147 5 3 4 6 0.675 0.542 0.503 12.1
103 5 3 5 5 0.647 0.519 0.545 12.0 148 4 4 4 6 0.347 0.985 0.667 13.2
104 5 4 5 5 0.598 0.551 0.567 11.8 149 5 4 4 6 0.323 1.000 0.670 10.9
105 6 4 5 5 0.691 0.601 0.625 7.5 150 6 4 4 6 0.712 0.637 0.583 7.5
106 5 5 5 5 0.439 0.971 0.980 8.1 151 2 2 5 6 0.376 0.495 0.404 63.5
107 6 5 5 5 0.408 1.000 0.991 6.6 152 3 2 5 6 0.387 0.480 0.415 42.6
108 2 2 6 5 0.438 0.420 0.450 65.6 153 4 2 5 6 0.625 0.399 0.411 26.1
109 3 2 6 5 0.438 0.435 0.477 40.7 154 3 3 5 6 0.378 0.629 0.523 31.0
110 4 2 6 5 0.599 0.393 0.456 25.3 155 4 3 5 6 0.340 0.722 0.534 22.7
111 3 3 6 5 0.444 0.531 0.608 32.2 156 5 3 5 6 0.636 0.481 0.474 15.4
112 4 3 6 5 0.411 0.612 0.628 22.0 157 4 4 5 6 0.384 0.792 0.666 16.9
113 5 3 6 5 0.618 0.466 0.518 15.1 158 5 4 5 6 0.345 0.929 0.672 13.1
114 5 4 6 5 0.418 0.781 0.805 12.8 159 6 4 5 6 0.323 1.000 0.670 10.9
115 6 4 6 5 0.388 0.904 0.807 10.1 160 5 5 5 6 0.393 0.967 0.816 10.1
116 5 5 6 5 0.463 0.810 0.982 10.5 161 6 5 5 6 0.362 1.000 0.825 8.4
117 6 5 6 5 0.433 0.944 0.987 8.0 162 3 3 6 6 0.406 0.528 0.516 38.6
118 2 2 7 5 0.456 0.364 0.429 81.0 163 4 3 6 6 0.370 0.608 0.530 27.3
119 3 2 7 5 0.454 0.384 0.462 48.9 164 5 3 6 6 0.609 0.434 0.452 19.1
120 4 2 7 5 0.577 0.361 0.441 30.5 165 4 4 6 6 0.411 0.662 0.664 21.2
121 3 3 7 5 0.462 0.458 0.594 40.2 166 5 4 6 6 0.375 0.778 0.672 16.0
122 4 3 7 5 0.433 0.530 0.619 26.8 167 6 4 6 6 0.344 0.900 0.674 12.8
123 5 3 7 5 0.598 0.425 0.495 18.6 168 5 5 6 6 0.420 0.807 0.817 12.7
124 5 4 7 5 0.439 0.672 0.803 15.8 169 6 5 6 6 0.388 0.940 0.821 10.0
125 6 4 7 5 0.412 0.778 0.806 12.1 170 6 6 6 6 0.432 0.958 0.972 8.1
126 5 5 7 5 0.481 0.695 0.983 13.2 171 3 3 7 6 0.428 0.455 0.508 47.3
127 6 5 7 5 0.454 0.810 0.990 9.9 172 4 3 7 6 0.393 0.526 0.525 32.6
128 3 2 8 5 0.467 0.342 0.445 58.0 173 5 3 7 6 0.589 0.396 0.433 23.0
129 4 2 8 5 0.561 0.335 0.428 36.1 174 4 4 7 6 0.433 0.569 0.662 26.2
130 3 3 8 5 0.476 0.402 0.578 49.3 175 5 4 7 6 0.398 0.669 0.671 19.2
131 4 3 8 5 0.449 0.467 0.607 32.1 176 6 4 7 6 0.369 0.775 0.674 15.2
132 4 4 8 5 0.484 0.502 0.784 27.4 177 5 5 7 6 0.441 0.693 0.818 15.7
133 5 4 8 5 0.456 0.590 0.801 19.2 178 6 5 7 6 0.411 0.807 0.823 12.1
134 6 4 8 5 0.431 0.684 0.806 14.5 179 6 6 7 6 0.452 0.822 0.974 10.1
135 5 5 8 5 0.494 0.609 0.984 16.4 180 3 3 8 6 0.444 0.400 0.498 56.9

181 4 3 8 6 0.413 0.463 0.519 38.4 220 5 5 7 7 0.407 0.691 0.700 18.4
182 5 3 8 6 0.574 0.366 0.417 27.2 221 6 5 7 7 0.375 0.805 0.704 14.5
183 4 4 8 6 0.449 0.499 0.659 31.7 222 6 6 7 7 0.416 0.819 0.833 11.9
184 5 4 8 6 0.417 0.587 0.670 22.9 223 4 4 8 7 0.419 0.498 0.567 36.4
185 6 4 8 6 0.389 0.680 0.673 17.7 224 5 4 8 7 0.384 0.585 0.575 27.0
186 5 5 8 6 0.457 0.607 0.819 19.1 225 6 4 8 7 0.355 0.678 0.578 21.4
187 6 5 8 6 0.429 0.707 0.824 14.5 226 5 5 8 7 0.425 0.605 0.701 22.1
188 6 6 8 6 0.468 0.719 0.977 12.3 227 6 5 8 7 0.395 0.705 0.705 17.1
189 2 2 4 7 0.307 0.607 0.361 62.0 228 6 6 8 7 0.434 0.717 0.834 14.3
190 3 2 4 7 0.466 0.423 0.350 44.7 229 2 2 5 8 0.317 0.491 0.313 90.1
191 3 3 4 7 0.308 0.777 0.456 30.1 230 3 2 5 8 0.352 0.434 0.317 64.6
192 4 3 4 7 0.269 0.887 0.463 23.5 231 3 3 5 8 0.317 0.625 0.398 44.4
193 5 3 4 7 0.667 0.505 0.445 15.4 232 4 3 5 8 0.278 0.715 0.404 34.3
194 4 4 4 7 0.312 0.983 0.573 16.4 233 5 3 5 8 0.626 0.423 0.379 23.3
195 5 4 4 7 0.287 1.000 0.576 13.7 234 4 4 5 8 0.319 0.787 0.501 24.4
196 6 4 4 7 0.693 0.596 0.517 9.6 235 5 4 5 8 0.281 0.923 0.504 19.7
197 2 2 5 7 0.344 0.493 0.353 76.2 236 6 4 5 8 0.258 1.000 0.504 17.0
198 3 2 5 7 0.363 0.460 0.360 53.0 237 5 5 5 8 0.325 0.962 0.610 14.7
199 4 2 5 7 0.626 0.371 0.363 32.4 238 6 5 5 8 0.295 1.000 0.616 12.6
200 3 3 5 7 0.345 0.627 0.452 37.4 239 3 2 7 8 0.368 0.358 0.312 85.0
201 4 3 5 7 0.306 0.718 0.460 28.2 240 3 3 7 8 0.371 0.452 0.391 63.2
202 5 3 5 7 0.629 0.449 0.421 19.2 241 4 3 7 8 0.333 0.520 0.400 46.1
203 4 4 5 7 0.349 0.789 0.572 20.5 242 5 3 7 8 0.581 0.352 0.348 32.9
204 5 4 5 7 0.310 0.926 0.576 16.2 243 4 4 7 8 0.373 0.566 0.499 35.2
205 6 4 5 7 0.287 1.000 0.576 13.7 244 5 4 7 8 0.335 0.664 0.504 27.1
206 5 5 5 7 0.356 0.964 0.699 12.3 245 6 4 7 8 0.305 0.769 0.506 22.3
207 6 5 5 7 0.325 1.000 0.706 10.4 246 5 5 7 8 0.378 0.689 0.612 21.3
208 3 3 6 7 0.374 0.526 0.448 45.7 247 6 5 7 8 0.345 0.802 0.615 17.1
209 4 3 6 7 0.336 0.604 0.458 33.3 248 6 6 7 8 0.385 0.818 0.727 13.8
210 5 3 6 7 0.603 0.407 0.402 23.4 249 3 3 8 8 0.391 0.397 0.387 74.1
211 4 4 6 7 0.378 0.660 0.570 25.2 250 4 3 8 8 0.354 0.458 0.398 52.9
212 5 4 6 7 0.340 0.775 0.576 19.4 251 5 3 8 8 0.566 0.327 0.336 38.1
213 6 4 6 7 0.308 0.897 0.578 16.0 252 4 4 8 8 0.393 0.496 0.497 41.5
214 5 5 6 7 0.384 0.805 0.699 15.2 253 5 4 8 8 0.356 0.583 0.504 31.4
215 6 5 6 7 0.351 0.937 0.702 12.1 254 6 4 8 8 0.326 0.675 0.506 25.4
216 6 6 6 7 0.394 0.956 0.831 9.7 255 5 5 8 8 0.397 0.604 0.612 25.3
217 4 4 7 7 0.401 0.567 0.569 30.5 256 6 5 8 8 0.366 0.703 0.616 19.9
218 5 4 7 7 0.364 0.666 0.576 23.0 257 6 6 8 8 0.405 0.716 0.728 16.4
219 6 4 7 7 0.334 0.772 0.578 18.5
SOLUTION From the preestimates we obtain µ02 = 11.32 + 8.56 − 9.86 =

10.02, a1 = a2 = 2.03, b = 3.24, and c = 2.89. Make the approximations a1 =
a2 ≈ 2 and c ≈ 3. Since the desired value of b is not included in Table 8.4, lin-
ear interpolation is needed. First find the plans for (a1 , a2 , b, c) = (2, 2, 3, 3) and
(2, 2, 4, 3). Then interpolation to (2, 2, 3.24, 3) gives π11 = 0.465, ξ11 = 0.762,
ξ21 = 0.769, and V = 17.4. For comparison, we calculated the optimization model
TABLE 8.5 Compromise Test Plans for a Lognormal Distribution

1 2 2 2 2 0.478 1.000 1.000 5.6 46 4 3 3 4 0.677 0.685 0.659 6.5
2 3 2 2 2 0.846 0.879 1.000 2.6 47 5 3 3 4 0.805 0.781 0.765 3.7
3 2 2 3 2 0.472 0.837 1.000 9.8 48 2 2 4 4 0.379 0.696 0.678 26.7
4 3 2 3 2 0.704 0.711 0.897 4.8 49 3 2 4 4 0.572 0.489 0.514 16.7
5 4 2 3 2 0.850 0.844 1.000 2.4 50 4 2 4 4 0.681 0.559 0.603 9.4
6 2 2 4 2 0.495 0.654 1.000 15.4 51 5 3 4 4 0.722 0.667 0.703 5.4
7 3 2 4 2 0.646 0.593 0.814 7.5 52 4 4 4 4 0.417 1.000 1.000 7.2
8 4 2 4 2 0.747 0.699 0.952 3.9 53 5 4 4 4 0.697 0.678 0.709 5.4
9 2 2 5 2 0.537 0.442 0.645 21.5 54 6 4 4 4 0.786 0.771 0.810 3.3
10 3 2 5 2 0.616 0.507 0.747 10.8 55 2 2 5 4 0.409 0.561 0.676 35.9
11 4 2 5 2 0.691 0.600 0.874 5.8 56 3 2 5 4 0.552 0.427 0.482 21.8
12 2 2 2 3 0.374 1.000 0.823 9.4 57 4 2 5 4 0.641 0.489 0.563 12.5
13 3 2 2 3 0.751 0.767 0.753 5.1 58 3 3 5 4 0.501 0.467 0.510 20.9
14 4 2 2 3 0.923 0.910 0.896 2.5 59 4 3 5 4 0.600 0.510 0.570 12.2
15 2 2 3 3 0.395 0.912 0.883 14.0 60 5 3 5 4 0.674 0.584 0.655 7.4
16 3 2 3 3 0.646 0.630 0.681 8.1 61 5 4 5 4 0.657 0.592 0.661 7.3
17 4 2 3 3 0.779 0.732 0.802 4.3 62 6 4 5 4 0.725 0.675 0.752 4.6
18 3 3 3 3 0.423 1.000 1.000 7.0 63 2 2 6 4 0.431 0.470 0.671 46.6
19 4 3 3 3 0.731 0.753 0.809 4.2 64 3 2 6 4 0.398 0.563 0.675 29.7
20 5 3 3 3 0.867 0.868 0.936 2.3 65 4 2 6 4 0.615 0.436 0.531 15.9
21 2 2 4 3 0.432 0.693 0.878 20.7 66 3 3 6 4 0.507 0.404 0.476 26.6
22 3 2 4 3 0.600 0.534 0.627 11.7 67 4 3 6 4 0.584 0.452 0.537 15.6
23 4 2 4 3 0.703 0.618 0.734 6.4 68 5 3 6 4 0.644 0.520 0.616 9.6
24 3 3 4 3 0.554 0.573 0.656 11.3 69 5 4 6 4 0.633 0.525 0.621 9.5
25 4 3 4 3 0.667 0.638 0.741 6.3 70 6 4 6 4 0.687 0.600 0.706 6.1
26 5 3 4 3 0.763 0.735 0.853 3.6 71 2 2 7 4 0.447 0.405 0.665 58.8
27 2 2 5 3 0.457 0.560 0.868 28.9 72 3 2 7 4 0.417 0.487 0.671 36.6
28 3 2 5 3 0.578 0.462 0.583 15.9 73 4 2 7 4 0.597 0.394 0.503 19.7
29 4 2 5 3 0.659 0.537 0.681 8.9 74 3 3 7 4 0.511 0.358 0.448 32.8
30 3 3 5 3 0.549 0.487 0.605 15.5 75 4 3 7 4 0.573 0.406 0.509 19.3
31 4 3 5 3 0.632 0.553 0.688 8.7 76 5 3 7 4 0.623 0.468 0.584 12.0
32 5 3 5 3 0.705 0.638 0.790 5.2 77 4 4 7 4 0.567 0.410 0.514 19.2
33 2 2 6 3 0.473 0.470 0.852 38.6 78 5 4 7 4 0.615 0.473 0.587 11.9
34 3 2 6 3 0.565 0.407 0.547 20.7 79 6 4 7 4 0.660 0.541 0.667 7.7
35 4 2 6 3 0.631 0.475 0.638 11.7 80 2 2 2 5 0.263 1.000 0.541 19.0
36 3 3 6 3 0.546 0.423 0.564 20.2 81 3 2 2 5 0.684 0.622 0.505 12.0
37 4 3 6 3 0.611 0.487 0.644 11.5 82 4 2 2 5 0.826 0.732 0.605 6.5
38 5 3 6 3 0.669 0.564 0.739 6.9 83 2 2 3 5 0.293 0.925 0.551 25.1
39 2 2 2 4 0.308 1.000 0.658 13.8 84 3 2 3 5 0.596 0.522 0.466 16.9
40 3 2 2 4 0.706 0.685 0.604 8.2 85 4 2 3 5 0.728 0.597 0.554 9.5
41 4 2 2 4 0.859 0.808 0.721 4.3 86 3 3 3 5 0.319 1.000 0.710 12.7
42 2 2 3 4 0.337 0.920 0.679 19.1 87 4 3 3 5 0.642 0.632 0.558 9.2
43 3 2 3 4 0.614 0.570 0.552 12.2 88 5 3 3 5 0.769 0.714 0.649 5.5
44 4 2 3 4 0.746 0.655 0.653 6.7 89 2 2 4 5 0.338 0.699 0.551 33.7
45 3 3 3 4 0.368 1.000 0.882 9.5 90 3 2 4 5 0.295 0.828 0.551 24.2

91 4 2 4 5 0.669 0.512 0.514 12.8 136 6 5 8 5 0.621 0.464 0.554 12.1
92 3 3 4 5 0.351 0.883 0.698 16.5 137 2 2 3 6 0.260 0.929 0.463 31.9
93 4 3 4 5 0.319 1.000 0.710 12.7 138 3 2 3 6 0.229 1.000 0.457 25.1
94 5 3 4 5 0.697 0.614 0.601 7.6 139 4 2 3 6 0.719 0.551 0.482 12.7
95 4 4 4 5 0.563 0.564 0.533 12.1 140 3 3 3 6 0.281 1.000 0.594 16.3
96 5 4 4 5 0.659 0.630 0.607 7.4 141 4 3 3 6 0.285 1.000 0.546 13.9
97 6 4 4 5 0.746 0.712 0.693 4.7 142 5 3 3 6 0.748 0.661 0.565 7.5
98 2 2 5 5 0.370 0.562 0.550 43.8 143 2 2 4 6 0.304 0.701 0.463 41.5
99 3 2 5 5 0.329 0.671 0.552 30.2 144 3 2 4 6 0.261 0.833 0.464 30.9
100 4 2 5 5 0.631 0.452 0.483 16.5 145 4 2 4 6 0.663 0.475 0.450 16.7
101 3 3 5 5 0.384 0.710 0.700 21.6 146 4 3 4 6 0.577 0.509 0.454 16.1
102 4 3 5 5 0.349 0.854 0.708 15.9 147 5 3 4 6 0.682 0.571 0.526 10.0
103 5 3 5 5 0.319 1.000 0.710 12.7 148 4 4 4 6 0.528 0.537 0.469 15.6
104 5 4 5 5 0.372 1.000 0.870 9.2 149 5 4 4 6 0.631 0.591 0.532 9.7
105 6 4 5 5 0.364 1.000 0.884 8.1 150 6 4 4 6 0.718 0.665 0.608 6.3
106 5 5 5 5 0.417 1.000 1.000 7.2 151 2 2 5 6 0.338 0.563 0.463 52.5
107 6 5 5 5 0.404 1.000 1.000 6.4 152 3 2 5 6 0.295 0.673 0.465 37.6
108 2 2 6 5 0.395 0.471 0.549 55.4 153 4 2 5 6 0.626 0.421 0.424 21.0
109 3 2 6 5 0.356 0.565 0.552 37.0 154 3 3 5 6 0.349 0.712 0.587 26.2
110 4 2 6 5 0.606 0.405 0.457 20.6 155 4 3 5 6 0.311 0.856 0.592 19.9
111 3 3 6 5 0.467 0.397 0.420 33.4 156 5 3 5 6 0.281 1.000 0.594 16.2
112 4 3 6 5 0.563 0.425 0.462 20.0 157 4 4 5 6 0.361 0.869 0.718 14.6
113 5 3 6 5 0.627 0.484 0.532 12.5 158 5 4 5 6 0.331 1.000 0.727 11.6
114 5 4 6 5 0.609 0.492 0.537 12.4 159 6 4 5 6 0.323 1.000 0.738 10.4
115 6 4 6 5 0.662 0.560 0.611 8.0 160 5 5 5 6 0.377 1.000 0.864 8.9
116 5 5 6 5 0.605 0.494 0.539 12.3 161 6 5 5 6 0.367 1.000 0.888 8.0
117 6 5 6 5 0.656 0.562 0.614 8.0 162 3 3 6 6 0.375 0.595 0.587 32.6
118 2 2 7 5 0.415 0.405 0.546 68.5 163 4 3 6 6 0.339 0.716 0.595 24.1
119 3 2 7 5 0.378 0.488 0.552 44.6 164 5 3 6 6 0.309 0.843 0.597 19.3
120 4 2 7 5 0.588 0.368 0.435 25.0 165 4 4 6 6 0.388 0.727 0.720 18.3
121 3 3 7 5 0.480 0.347 0.394 40.6 166 5 4 6 6 0.358 0.859 0.726 14.1
122 4 3 7 5 0.554 0.383 0.440 24.4 167 6 4 6 6 0.332 0.996 0.728 11.6
123 5 3 7 5 0.608 0.438 0.505 15.4 168 5 5 6 6 0.402 0.865 0.859 11.2
124 5 4 7 5 0.595 0.444 0.509 15.2 169 6 5 6 6 0.377 1.000 0.864 8.9
125 6 4 7 5 0.640 0.506 0.580 10.0 170 6 6 6 6 0.417 1.000 1.000 7.2
126 5 5 7 5 0.593 0.445 0.511 15.2 171 3 3 7 6 0.396 0.512 0.588 39.9
127 6 5 7 5 0.636 0.508 0.582 10.0 172 4 3 7 6 0.362 0.616 0.597 28.8
128 3 2 8 5 0.395 0.430 0.551 53.0 173 5 3 7 6 0.332 0.725 0.599 22.6
129 4 2 8 5 0.576 0.337 0.416 29.7 174 4 4 7 6 0.408 0.625 0.721 22.5
130 3 3 8 5 0.487 0.311 0.373 48.4 175 5 4 7 6 0.380 0.738 0.728 17.0
131 4 3 8 5 0.548 0.349 0.421 29.1 176 6 4 7 6 0.355 0.856 0.731 13.7
132 4 4 8 5 0.541 0.353 0.425 28.9 177 5 5 7 6 0.422 0.744 0.861 13.8
133 5 4 8 5 0.585 0.405 0.486 18.3 178 6 5 7 6 0.399 0.862 0.866 10.8
134 6 4 8 5 0.624 0.463 0.553 12.1 179 6 6 7 6 0.437 0.868 1.000 9.0
135 5 5 8 5 0.584 0.406 0.488 18.3 180 3 3 8 6 0.412 0.449 0.588 48.0

181 4 3 8 6 0.381 0.540 0.598 34.1 220 5 5 7 7 0.390 0.745 0.740 16.1
182 5 3 8 6 0.352 0.637 0.602 26.2 221 6 5 7 7 0.365 0.863 0.744 12.9
183 4 4 8 6 0.424 0.548 0.722 27.2 222 6 6 7 7 0.403 0.868 0.864 10.5
184 5 4 8 6 0.398 0.647 0.730 20.3 223 4 4 8 7 0.396 0.549 0.621 31.1
185 6 4 8 6 0.375 0.751 0.734 16.1 224 5 4 8 7 0.368 0.648 0.627 23.7
186 5 5 8 6 0.437 0.652 0.863 16.8 225 6 4 8 7 0.343 0.751 0.630 19.2
187 6 5 8 6 0.416 0.757 0.869 13.0 226 5 5 8 7 0.408 0.653 0.741 19.3
188 6 6 8 6 0.452 0.762 1.000 10.9 227 6 5 8 7 0.384 0.757 0.746 15.2
189 2 2 4 7 0.277 0.702 0.400 50.1 228 6 6 8 7 0.420 0.761 0.866 12.7
190 3 2 4 7 0.233 0.836 0.401 38.5 229 2 2 5 8 0.285 0.552 0.356 72.5
191 3 3 4 7 0.285 0.888 0.504 25.1 230 3 2 5 8 0.243 0.676 0.353 55.0
192 4 3 4 7 0.251 1.000 0.511 20.3 231 3 3 5 8 0.294 0.714 0.443 36.7
193 5 3 4 7 0.673 0.535 0.469 12.7 232 4 3 5 8 0.256 0.859 0.447 29.2
194 4 4 4 7 0.498 0.516 0.420 19.4 233 5 3 5 8 0.227 1.000 0.448 24.8
195 5 4 4 7 0.609 0.559 0.475 12.3 234 4 4 5 8 0.303 0.873 0.542 20.8
196 6 4 4 7 0.699 0.625 0.542 8.1 235 5 4 5 8 0.272 1.000 0.548 17.1
197 2 2 5 7 0.311 0.564 0.400 62.1 236 6 4 5 8 0.263 1.000 0.554 15.7
198 3 2 5 7 0.266 0.675 0.401 45.9 237 5 5 5 8 0.313 1.000 0.651 12.8
199 4 2 5 7 0.625 0.395 0.379 25.9 238 6 5 5 8 0.302 1.000 0.665 11.7
200 3 3 5 7 0.319 0.713 0.505 31.2 239 3 2 7 8 0.294 0.489 0.353 73.5
201 4 3 5 7 0.281 0.857 0.509 24.3 240 3 3 7 8 0.345 0.513 0.444 52.4
202 5 3 5 7 0.251 1.000 0.511 20.3 241 4 3 7 8 0.308 0.617 0.449 39.8
203 4 4 5 7 0.329 0.871 0.618 17.6 242 5 3 7 8 0.277 0.726 0.451 32.4
204 5 4 5 7 0.299 1.000 0.625 14.2 243 4 4 7 8 0.353 0.626 0.544 29.9
205 6 4 5 7 0.290 1.000 0.633 12.9 244 5 4 7 8 0.323 0.739 0.548 23.6
206 5 5 5 7 0.342 1.000 0.743 10.8 245 6 4 7 8 0.296 0.857 0.550 19.7
207 6 5 5 7 0.331 1.000 0.760 9.7 246 5 5 7 8 0.363 0.745 0.649 18.6
208 3 3 6 7 0.346 0.596 0.506 38.2 247 6 5 7 8 0.337 0.864 0.652 15.1
209 4 3 6 7 0.309 0.717 0.511 28.9 248 6 6 7 8 0.374 0.869 0.758 12.2
210 5 3 6 7 0.279 0.844 0.512 23.6 249 3 3 8 8 0.364 0.449 0.445 61.5
211 4 4 6 7 0.357 0.728 0.619 21.6 250 4 3 8 8 0.328 0.541 0.450 45.8
212 5 4 6 7 0.326 0.860 0.623 17.0 251 5 3 8 8 0.297 0.637 0.452 36.7
213 6 4 6 7 0.299 0.997 0.625 14.2 252 4 4 8 8 0.372 0.549 0.545 35.3
214 5 5 6 7 0.369 0.866 0.739 13.2 253 5 4 8 8 0.342 0.648 0.549 27.4
215 6 5 6 7 0.342 1.000 0.743 10.8 254 6 4 8 8 0.316 0.752 0.551 22.5
216 6 6 6 7 0.381 1.000 0.864 8.6 255 5 5 8 8 0.382 0.653 0.650 21.9
217 4 4 7 7 0.379 0.626 0.620 26.1 256 6 5 8 8 0.357 0.758 0.653 17.6
218 5 4 7 7 0.349 0.739 0.625 20.2 257 6 6 8 8 0.392 0.761 0.759 14.5
219 6 4 7 7 0.323 0.857 0.627 16.6
for directly (2.03, 2.03, 3.24, 2.89) and got π11 = 0.477, ξ11 = 0.758, ξ21 =
0.800, and V = 16.1. In this case, the approximation and interpolation yield
fairly accurate results. The standardized plan is then converted to the actual
plan, as shown in Table 8.6. In implementing this plan, the six tires used for the
preliminary test should continue being tested until 13,000 miles as part of the
group at 2,835 pounds. Thus, this group requires only an additional 12 units.
TABLE 8.6 Actual Test Plan for the Tires

Sample Load Threshold Censoring Time
Group Size (pounds) (inches) (miles)
1 17 1611 3/32 13,000

2 17 1611 8/32 13,000
3 18 2835 3/32 13,000
4 18 2835 8/32 13,000
8.8.2 Survey of Degradation Test Plans

As stated earlier, a degradation test may be conducted at a use stress level or
higher than a use stress level. For a test at a use stress level, we need to choose the
sample size, inspection times (inspection frequency), and test termination time.
In addition to these variables, we need to determine the stress levels and sample
size allocated to each stress level for a constant-stress accelerated degradation
test. The common design of the test plans selects the optimum values of these
variables to minimize the statistical error, total test cost, or both. In particular,
the test plans can be formulated as one of the following optimization problems:
ž Minimize the asymptotic variance (or mean square error) of the estimate of a
life percentile or other quantity at a use stress level, subject to a prespecified
cost budget.
ž Minimize the total test cost, subject to the allowable statistical error.
ž Minimize the asymptotic variance (or mean square error) and the total test
cost simultaneously.
Nelson (2005) provides a nearly complete bibliography of test plans. Most

plans are formulated as a first optimization problem. Yu (2003) designs optimum
accelerated degradation test plans using the nonlinear integer programming tech-
nique. The plans are designed to choose the optimal combination of sample size,
inspection frequency, and test termination time at each stress level by minimiz-
ing the mean square error of the estimated 100pth percentile of the product life
distribution at a use stress level. The optimization is subject to the constraint that
the total test cost must not exceed a given budget. In an earlier work, Yu (1999)
presents similar test plans but minimizes the variance of the estimate. Q. Li and
Kececioglu (2003) present a four-step approach to the calculation of sample size,
test stress levels, respective sample allocation assigned to each stress level, mea-
surement times, and test termination time. The decision variables are optimized
to minimize the mean square error of the estimate of the mean life subject to a
cost budget. Using a simple constant rate relationship between the applied stress
and product performance, Park and Yum (1997) devise optimum test plans for
destructive inspections. The plans also determine the stress levels, proportion of
test units allocated to each stress level, and measurement times, which minimize
the asymptotic variance of the MLE of the mean life at a use stress level. Park
PROBLEMS 373
and Yum (1999) compare numerically plans they developed for accelerated life
and degradation tests. Unsurprisingly, they conclude that accelerated degradation
test plans provide more accurate estimates of life percentiles, especially when the
probabilities of failure are small. Boulanger and Escobar (1994) design optimum
accelerated degradation test plans for a particular degradation model that may be
suitable to describe a degradation process in which the amount of degradation
over time levels off toward a stress-dependent plateau (maximum degradation).
The design consists of three steps, the first of which is to determine the stress
levels and corresponding proportions of test units by minimizing the variance
of the weighted least squares estimate of the mean of the log plateaus at the
use condition. The second step is to optimize the times at which to measure the
units at a selected stress level, then the results of the two steps are combined to
determine the total number of test units.
For degradation tests at a single constant-stress level, S. Wu and Chang (2002)
propose an approach to the determination of sample size, inspection frequency,
and test termination time. The optimization criterion is to minimize the variance
of a life percentile estimate subject to total test cost. Marseguerra et al. (2003)
develop test plans similar to those of S. Wu and Chang (2002), and consider
additionally simultaneous minimization of the variance and total test cost. The
latter part deals with the third optimization problem described above.
Few publications deal with the second optimization problem. Yu and Chiao
(2002) design optimal plans for fractional factorial degradation experiments with
the aim of improving product reliability. The plans select the inspection fre-
quency, sample size, and test termination time at each run by minimizing the
total test cost subject to a prespecified correct decision probability. Tang et al.
(2004) conduct the optimal design of step-stress accelerated degradation tests.
The objective of the design is to minimize the total test cost subject to a vari-
ance constraint. The minimization yields the optimal sample size, number of
inspections at each intermediate stress level, and number of total inspections.
Generally speaking, the optimal design of accelerated degradation tests is
considerably more difficult than that of accelerated life tests, mainly because the
former involves complicated degradation models and more decision variables. It
is not surprising that there is scant literature on this subject. As we may have
observed, degradation testing and analysis is a promising and rewarding tech-
nique. Increasingly wide applications of this technique require more practically
useful test plans.
PROBLEMS
8.1 A product usually has more than one performance characteristic. Describe
the general approaches to determination of the critical characteristic. Explain
why the critical characteristic selected must be monotone.
8.2 Discuss the advantages and disadvantages of pseudolife analysis. Compared
with the traditional life data analysis, do you expect a pseudolife analysis
to give a more accurate estimate? Why?
TABLE 8.7 Valve Recession Data (Inches)

Valve
Time (h) 1 2 3 4 5 6 7
0 0 0 0 0 0 0 0
15 0.001472 0.001839 0.001472 0.001839 0.001839 0.001839 0.002575
45 0.002943 0.004047 0.003311 0.002943 0.003311 0.002943 0.003679
120 0.005886 0.00699 0.005886 0.004415 0.005518 0.005886 0.005886
150 0.006254 0.008093 0.006622 0.005150 0.006254 0.006990 0.007726
180 0.008461 0.009933 0.008093 0.006622 0.007726 0.008461 0.010301
8.3 K.Yang and Xue (1996) performed degradation analysis of the exhaust
valves installed in a certain internal combustion engine. The degradation
is the valve recession, representing the amount of wear in a valve over
time; the valve recession data at different inspection times are shown in
Table 8.7. Develop a model to describe the degradation paths. If the valve
is said to have failed when the recession reaches 0.025 inch, estimate the
probability of failure at 500 hours through pseudolife analysis.
8.4 Refer to Problem 8.3. Suppose that the degradation model parameters for
the valve recession have random effects and have a bivariate normal dis-
tribution. Calculate the mean vector and variance–covariance matrix using
the multivariate approach. Estimate the probability of failure of the valve
at 500 hours through Monte Carlo simulation.
8.5 Refer to Problem 8.4. Calculate the probability of failure of the valve at
500 hours by using (8.13). Compare the result with those in Problems 8.3
and 8.4.
8.6 In Section 8.5.1 we describe methods for sample allocation to destructive
inspections. Explain how the methods improve the statistical accuracy of a
reliability estimate.
8.7 A type of new polymer was exposed to the alkaline environment at ele-
vated temperatures to evaluate the long-term reliability of the material. The
experiment tested standard bars of the material at 50, 65, and 80◦ C, each
with 25 units. Five units were inspected destructively for tensile strength
at each inspection time during testing. The degradation performance is the
ratio of the tensile strength to the original standard strength. The material
is said to have failed when the ratio is less than 60%. Table 8.8 shows the
values of the ratio at different inspection times (in days) and temperatures.
(a) For each combination of temperatures and inspection times, plot the
data and ML fits on lognormal paper. Comment on the adequacy of the
lognormal distribution.
(b) Does the shape parameter change with time?
PROBLEMS 375
TABLE 8.8 Polymer Degradation Data

Temperature (◦ C)
Days 50 65 80
8 98.3 94.2 96.5 98.1 96.0 87.5 85.2 93.3 90.0 88.4 80.8 82.3 83.7 86.6 81.1
25 92.4 88.1 90.5 93.4 90.2 83.2 80.5 85.7 86.3 84.2 73.3 72.3 71.9 74.5 76.8
75 86.2 82.7 84.2 86.1 85.5 77.0 73.2 79.8 75.4 76.2 67.4 65.4 64.3 65.3 64.5
130 82.3 78.5 79.4 81.8 82.3 73.9 70.1 75.8 72.3 71.7 64.3 60.4 58.6 58.9 59.7
180 77.7 74.6 76.1 77.9 79.2 68.7 65.3 69.8 67.4 66.6 60.4 55.3 56.7 57.3 55.7
TABLE 8.9 Degradation Data for the IRLEDs at 170 mA

Time (h)
Unit 24 48 96 155 368 768 1130 1536 1905 2263 2550
1 0.1 0.3 0.7 1.2 3.0 6.6 12.1 16.0 22.5 25.3 30.0
2 2.0 2.3 4.7 5.9 8.2 9.3 12.6 12.9 17.5 16.4 16.3
3 0.3 0.5 0.9 1.3 2.2 3.8 5.5 5.7 8.5 9.8 10.7
4 0.3 0.5 0.8 1.1 1.5 2.4 3.2 5.1 4.7 6.5 6.0
5 0.2 0.4 0.9 1.6 3.9 8.2 11.8 19.5 26.1 29.5 32.0
6 0.6 1.0 1.6 2.2 4.6 6.2 10.5 10.2 11.2 11.6 14.6
7 0.2 0.4 0.7 1.1 2.4 4.9 7.1 10.4 10.8 13.7 18.0
8 0.5 0.9 1.8 2.7 6.5 10.2 13.4 22.4 23.0 32.2 25.0
9 1.4 1.9 2.6 3.4 6.1 7.9 9.9 10.2 11.1 12.2 13.1
10 0.7 0.8 1.4 1.8 2.6 5.2 5.7 7.1 7.6 9.0 9.6
11 0.2 0.5 0.8 1.1 2.5 5.6 7.0 9.8 11.5 12.2 14.2
12 0.2 0.3 0.6 0.9 1.6 2.9 3.5 5.3 6.4 6.6 9.2
13 2.1 3.4 4.1 4.9 7.2 8.6 10.8 13.7 13.2 17.0 13.9
14 0.1 0.2 0.5 0.7 1.2 2.3 3.0 4.3 5.4 5.5 6.1
15 0.7 0.9 1.5 1.9 4.0 4.7 7.1 7.4 10.1 11.0 10.5
16 1.8 2.3 3.7 4.7 6.1 9.4 11.4 14.4 16.2 15.6 16.6
17 0.1 0.2 0.5 0.8 1.6 3.2 3.7 5.9 7.2 6.1 8.8
18 0.1 0.1 0.2 0.3 0.7 1.7 2.2 3.0 3.5 4.2 4.6
19 0.5 0.7 1.3 1.9 4.8 7.7 9.1 12.8 12.9 15.5 19.3
20 1.9 2.3 3.3 4.1 5.2 8.9 11.8 13.8 14.1 16.2 17.1
21 3.7 4.8 7.3 8.3 9.0 10.9 11.5 12.2 13.5 12.4 13.8
22 1.5 2.2 3.0 3.7 5.1 5.9 8.1 7.8 9.2 8.8 11.1
23 1.2 1.7 2.0 2.5 4.5 6.9 7.5 9.2 8.5 12.7 11.6
24 3.2 4.2 5.1 6.2 8.3 10.6 14.9 17.5 16.6 18.4 15.8
25 1.0 1.6 3.4 4.7 7.4 10.7 15.9 16.7 17.4 28.7 25.9
(c) Develop a degradation model for a given temperature.

(d) For each temperature, estimate the degradation model parameters using
the least squares method. Which model parameter(s) depend on temper-
ature?
(e) For each temperature, estimate the probability of failure at the test ter-
mination time.
(f) Repeat part (e) for a design life of 10 years.
8.8 Example 8.10 describes the degradation analysis for the IRLEDs. The de-
gradation data of the device at different measurement times (in hours) and
current levels (in mA) are shown in Tables 8.9 and 8.10.
(a) Fit the degradation model (8.36) to each degradation path, and estimate
the model parameters using the least squares method.
(b) Compute the pseudolife of each unit.
(c) For each current, plot the pseudolife data and the ML fits on lognor-
mal probability paper. Comment on the adequacy of the lognormal
distribution.
(d) Is it evident that the shape parameter depends on current?
(e) Use the inverse power relationship for current, and estimate the log-
normal scale parameter and the probability of failure at a use cur-
rent of 50 mA. Comment on the difference in results from those in
Example 8.10.
8.9 Refer to Problem 8.7.

(a) Is it evident that the shape parameter depends on temperature?
(b) Develop a degradation model to describe the relationship between the
mean log ratio and time and temperature.
TABLE 8.10 Degradation Data for the IRLEDs at 320 mA

Time (h)
Unit 6 12 24 48 96 156 230 324 479 635
1 4.3 5.8 9.5 10.2 13.8 20.6 19.7 25.3 33.4 27.9
2 0.5 0.9 1.4 3.3 5.0 6.1 9.9 13.2 17.0 20.7
3 2.6 3.6 4.6 6.9 9.5 13.0 15.3 13.5 19.0 19.5
4 0.2 0.4 0.9 2.4 4.5 7.1 13.4 21.2 30.7 41.7
5 3.7 5.6 8.0 12.8 16.0 23.7 26.7 38.4 49.2 47.2
6 3.2 4.3 5.8 9.9 15.2 20.3 26.2 33.6 39.5 53.2
7 0.8 1.7 2.8 4.6 7.9 12.4 20.2 24.8 32.5 45.4
8 4.3 6.5 7.8 13.0 21.7 33.0 42.1 49.9 59.9 78.6
9 1.4 2.7 5.0 7.8 14.5 23.3 29.0 43.3 59.8 77.4
10 3.4 4.6 7.8 13.0 16.8 26.8 34.1 41.5 67.0 65.5
11 3.6 4.7 6.2 9.1 11.7 13.8 14.5 15.5 23.1 24.0
12 2.3 3.7 5.6 8.8 13.7 17.2 24.8 29.1 42.9 45.3
13 0.5 0.9 1.9 3.5 5.9 10.0 14.4 22.0 26.0 31.8
14 2.6 4.4 6.0 8.7 14.6 16.8 17.9 23.2 27.0 31.3
15 0.1 0.4 0.7 2.0 3.5 6.6 12.2 18.8 32.3 47.0
PROBLEMS 377
TABLE 8.11 Degradation Data for an MOS

Field-Effect Transistor
Unit
Time (s) 1 2 3 4 5
100 1.05 0.58 0.86 0.6 0.62

200 1.4 0.9 1.25 0.6 0.64
300 1.75 1.2 1.45 0.6 1.25
400 2.1 1.75 1.75 0.9 1.3
500 2.1 2.01 1.75 0.9 0.95
600 2.8 2 2 1.2 1.25
700 2.8 2 2 1.5 1.55
800 2.8 2 2 1.5 1.9
900 3.2 2.3 2.3 1.5 1.25
1,000 3.4 2.6 2.3 1.7 1.55
1,200 3.8 2.9 2.6 2.1 1.5
1,400 4.2 2.9 2.8 2.1 1.55
1,600 4.2 3.2 3.15 1.8 1.9
1,800 4.5 3.6 3.2 2.1 1.85
2,000 4.9 3.8 3.2 2.1 2.2
2,500 5.6 4.2 3.8 2.4 2.2
3,000 5.9 4.4 3.8 2.7 2.5
3,500 6.3 4.8 4 2.7 2.2
4,000 6.6 5 4.2 3 2.8
4,500 7 5.6 4.4 3 2.8
5,000 7.8 5.9 4.6 3 2.8
6,000 8.6 6.2 4.9 3.6 3.1
7,000 9.1 6.8 5.2 3.6 3.1
8,000 9.5 7.4 5.8 4.2 3.1
9,000 10.5 7.7 6.1 4.6 3.7
10,000 11.1 8.4 6.3 4.2 4.4
12,000 12.2 8.9 7 4.8 3.7
14,000 13 9.5 7.2 5.1 4.4
16,000 14 10 7.6 4.8 4.4
18,000 15 10.4 7.7 5.3 4.1
20,000 16 10.9 8.1 5.8 4.1
25,000 18.5 12.6 8.9 5.7 4.7
30,000 20.3 13.2 9.5 6.2 4.7
35,000 22.1 15.4 11.2 8 6.4
40,000 24.2 18.1 14 10.9 9.4
Source: J. Lu et al. (1997).
(c) Write down the total sample log likelihood.

(d) Estimate the model parameters.
(e) Calculate the probability of failure of the polymer after 10 years of
continuous use at 40◦ C.
8.10 Refer to Example 8.1. The percent transconductance degradation data taken
at different times (in seconds) for five units of an MOS field-effect transistor
are shown in Table 8.11.
(a) Without fitting a degradation model, determine the failure time intervals
for each unit at thresholds 3%, 8%, and 13%, respectively.
(b) Estimate the life distributions at each threshold.
(c) Does the scale parameter depend on threshold?
(d) Develop a relationship between the distribution location parameter and
the threshold.
(e) Estimate the distribution location parameter at the usual threshold of
15%. Compare the result with that in Example 8.1.
8.11 To estimate the reliability of a resistor at a use temperature of 50◦ C, the
manufacturer plans to sample 45 units and divide them into two groups,
each tested at an elevated temperature. The failure of the resistor is defined
in terms of a resistance drift greater than 500 ppm (parts per million). The
tightest failure criterion is 100 ppm, and the maximum allowable tempera-
ture is 175◦ C. The time to failure of the resistor has a Weibull distribution
with shape parameter 1.63. Preestimates of the log characteristic life are
µ00 = 12.1, µ20 = 8.3, and µ22 = 5.9. Each group is tested for 2350 hours
or until all units fail, whichever is sooner. Develop a compromise test plan
for the experiment.

Degradation Testing and Analysis

Uploaded by

Copyright:

Available Formats

Degradation Testing and Analysis

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Degradation Testing and Analysis

Uploaded by

Copyright:

Available Formats

8

DEGRADATION TESTING AND

As explained in Chapter 7, the accelerated life test is an important task in nearly

Life Cycle Reliability Engineering. Guangbin Yang

8.2 DETERMINATION OF THE CRITICAL PERFORMANCE

The performance of a product is usually measured by multiple characteristics.

becomes monotone afterward. An example is light-emitting diodes, the luminous

8.3 RELIABILITY ESTIMATION FROM PSEUDOLIFE

tˆi = g −1 (G; β̂1i , β̂2i , . . . , β̂pi ), (8.4)

Then the pseudolife of unit i is

Example 8.1 J. Lu et al. (1997) give the percent transconductance degradation

ln(yij ) = β1i + β2i ln(t) + eij .

TABLE 8.1 Least Squares Estimates of Model

β̂1i −2.413 −2.735 −2.056 −2.796 −2.217

1000 10,000 100,000 1,000,000 10,000,000

8.4 DEGRADATION ANALYSIS WITH RANDOM-EFFECT MODELS

8.4.1 Random-Effect Models

The sample variance–covariance matrix is given by

for k = 1, 2, . . . , p and j = 1, 2, . . . , p. skj (k = j ) is the covariance of βk and

FIGURE 8.4 Normal ﬁts to estimates of β1 and β2

β and S are estimates of µβ and β , respectively. The estimates are useful in

SOLUTION First we plot the ﬁve estimates of β1 and β2 on normal probability

(−2.413 + 2.443)2 + · · · + (−2.217 + 2.443)2

The estimate of the variance–covariance matrix is

FIGURE 8.5 Relation of degradation to failure

The correlation coefﬁcient between β1 and β2 is

8.4.2 Relation of Degradation to Failure

F (t) = Pr(T ≤ t) = Pr[y(t) ≥ G] = Pr[g(t; β1 , β2 , . . . , βp ) ≥ G]. (8.12)

For a monotonically decreasing performance characteristic, the probability of

8.4.3 Reliability Estimation by Monte Carlo Simulation

8.4.4 Reliability Estimation with a Bivariate Normal Distribution

SOLUTION Because the degradation path is linear on the log-log scale as

8.4.5 Reliability Estimation with a Univariate Normal Distribution

SOLUTION As shown in Figure 8.8, the degradation data suggest a linear

8.5 DEGRADATION ANALYSIS FOR DESTRUCTIVE INSPECTIONS

For some products, inspection for measurement of performance characteristics

8.5.1 Test Methods

8.5.2 Data Analysis

Case 3: Normal Performance with Nonconstant σy Sometimes the perfor-

8.6 STRESS-ACCELERATED DEGRADATION TESTS

In the preceding sections we described techniques for degradation analysis at

8.6.1 Pseudo Accelerated Life Test Method

Example 8.7 Stress relaxation is the loss of stress in a component subjected to

FIGURE 8.13 Stress relaxation at different times and temperatures

where β1 = ln(A) − Ea /kT and β2 = B. To estimate the reliability of a type of

FIGURE 8.15 Lognormal ﬁts to the pseudolifetimes at different temperatures

Thus, it is legitimate to assume that the lognormal scale parameter µ is a lin-

8.6.2 Random-Effect Method

Estimating the Model Parameters The degradation model g(t; β1 , β2 , . . . , βp )

Evaluating the Probability of Failure Similar to (8.12), for a monotonically

F (t) = Pr[g(t, S0 ; β1 , β2 , . . . , βp ) ≥ G]. (8.28)

where β1 and β2 have bivariate normal distribution and γ is a constant. The

F (t) = Pr[g(t, S0 ) ≥ G] = Pr(β1 + γ S0 + β2 t ≥ G)

which is a closed-form expression for F (t). In general, if F (t) can be expressed

8.6.3 Random-Process Method

ž The (log) performance characteristic y at a given time t and stress S has

ž If the scale parameter σy is a function of t and S, the speciﬁc form of

Note that (8.33) is similar to (8.25).

SOLUTION Figures 8.16 shows examples of lognormal ﬁts to measurements