Meeker 2014

Degradation Models and Analyses
By William Meeker,1 Yili Hong,2 and Luis Escobar3

Keywords: Accelerated degradation tests, Accelerated destructive degradation tests, Condition based main-
tenance, General path model, Reliability assessment, Wiener process
Abstract: For products/systems requiring high reliability or availability, reliability assess-

ment and maintenance actions are often needed in a timely manner. For these systems/
products, information on system state and performance can be collected over the system’s
operating time, providing what is known as degradation data. With appropriate models
and data analysis techniques, degradation data are converted to system reliability informa-
tion and this information is useful for short and long-term reliability assessment and for
planning maintenance actions. This article provides a review of degradation data models
and methods that have been used successfully in this area. Commonly used methods for
analyzing degradation data, general path models and stochastic models, are described
in detail. Models and methods for analyzing data from accelerated degradation tests and
accelerated destructive degradation tests are also introduced. Dynamic condition moni-
toring based on degradation data is covered. Examples are used to illustrate the models
and methods covered in this article. The conclusion provides a discussion of test planning,
software implementation and the future of degradation data analysis.
1 Introduction
1.1 Background
For products or systems requiring high reliability or availability, reliability assessment and maintenance
actions should be done in timely manner. Obtaining such assessments based on system failure-time
information is challenging because products/systems nowadays can operate for long periods of time with
few or no failures. Information on system state and performance can, however, in some applications,
1 Iowa State University

2 Virginia Tech
3 Louisiana State University
This article was originally published online in 2011 in Encyclopedia of Statistical Sciences, c John Wiley & Sons, Inc. and
republished in Wiley StatsRef: Statistics Reference Online, 2014.
Copyright c 2011 by John Wiley & Sons, Inc. All rights reserved. 1
be collected over a system’s operating time. With appropriate models and data analysis techniques, this
information is converted to system reliability information, which can be used to provide assessments
of short- and long-term reliability and for planning maintenance actions like repairs and replacements.
Information on system state and performance collected over time is referred to as degradation data.
Depending on how the measurement is taken, there are two types of degradation: repeated measurement
degradation data and destructive degradation data. In repeated measurement degradation data, measure-
ments are taken on the same unit at different points in time. In destructive degradation data, only one
measurement can be taken on a unit because the measurement is destructive.
For many products/systems used in the field, there will be deterioration of physical conditions, such
as the wear of automobile tires, change in chemical concentration of a certain material, or the growth of a
fatigue crack in a metal component. Deteriorations can happen in system performance such as the power
output of an amplifier or the intensity of an LED. Such deteriorations will eventually lead to complete
or partial loss of system functionality. Most system failures are associated with a underlying degradation
process. For example, when the tire-tread depth has decreased to a certain level, the tire should, because
of safety concerns, no longer be used and a failure is said to have occurred.
Such failures are often referred as soft failures because the product/system does not lose complete
functionality at the time of failure. In contrast, failures that result in a complete loss of functionality are
called hard failures. Hard failures, however, could also be caused by some random events that do not
correspond exactly with a particular level of degradation. In this article, we describe statistical models and
data analysis techniques for failures that are caused by system degradation.
1.2 Data and Examples
Degradation measurements on a unit can be taken at specific time points. For example, the crack size on
the surface of a metal component in a system might be monitored every month. Measurements can also
be taken continuously in time. Continuous measurements are often done with sensors or other automatic
data collecting technologies.
Depending on the application, various types of system-condition information can be used as a system
degradation indicator. The degradation indicator could be the vibration level of a system, the number of
particles or amount of chemical elements that are released into the environment by a system, the size
of a fatigue crack, the temperature of a system, or the resistance or conductivity of certain components.
In material sciences, for example, researchers are interested in the performance of solar cells 16 and the
power output of solar cells is a natural degradation indicator. In the pharmaceutical area, the researchers
are interested in the synergistic degradation of ibuprofen 21 and the drug concentration can be used as
the indicator.
The following two degradation datasets will be used in subsequent sections for illustrations of statistical
methods.
Example 1. Laser Life Data Meeker and Escobar [22, Ex. 13.10], Example 13.10, describes a dataset from a
laboratory degradation test of a laser unit, a device that was used in telecommunication applications. The
device has a feedback circuit to maintain constant light output. To compensate for the degradation in the
lasing material, current is increased over time. Thus the operating current was used as the degradation
indicator. The current was measured at certain time points. Figure 1 displays the degradation paths for
the 15 units in the test. The x-axis is hours in operation and the y-axis is the percentage increase in
operating current. The degradation paths shown in Figure 1 are approximately linear. A 10% increase in
current was the specified failure level for this application.
2 Copyright c 2011 by John Wiley & Sons, Inc. All rights reserved.
Figure 1. Observed degradation paths for laser device units.
Example 2. Outdoor Weathering Data Vaca-Trigo and Meeker 33 describe photodegradation of an epoxy
coating exposed to outdoor weathering. The damage is mainly caused by exposure to outdoor ultraviolet
(UV) radiation. The damage to the compound was measured after every few days of exposure using
Fourier Transform Infrared Spectroscopy (FTIR), a technology that provides measurements that are
proportional to the concentration of chemical compounds in the coating. In this application, the effective
dosage is, scientifically, the appropriate time scale for the degradation process. The effective dosage is a
number that is proportional to the number of photons absorbed into the experimental specimens. Figure
2 shows a plot of the damage to an epoxy compound due to outdoor weathering for a subset of specimens.
For purpose of illustration, the definition of failure was set at a damage level of −0.45. This damage level
is shown as a horizontal line in Figure 2. As a result, for this level of damage, there were 13 failures in
a sample of 55 specimens. Note that Figure 2 shows only a representative subset of specimens with 4
failures.
1.3 Statistical Models
The main task in developing a statistical model for degradation data is to identify a model for the
degradation paths. The shape of the degradation path as a function of the variable measuring the lifetime
of a unit (i.e., time in service, cycles, or other amount of use) depends on the nature of the degradation
process and it can take numerous different shapes. Figure 3 illustrates linear, convex, and concave
Figure 2. Plot of the damage to an epoxy compound due to outdoor weathering for a subset of specimens.
A unit with damage below the critical level of −0.45 is considered a failed unit with a failure time equal to the
crossing time (the time scale here is the effective dosage).
degradation paths. The laser data in Figure 1 give an example of linear degradation paths. The shape of
outdoor degradation paths is usually more complex with more variability due to uncontrolled factors (e.g.,
the degradation paths in Figure 2). Although examples given in this article all have monotone degradation
paths, non-monotone degradation paths do exist. For example, Bae and Kvam 1 gave an application of the
degradation path of vacuum fluorescent displays (VFD) where the degradation path is non-monotone.
In their application, the brightness of the VFD for some units is increasing at the burn-in stage and is
decreasing for all units in the long run.
For each unit in the population, the degradation model implies a degradation path with level, say, D(t)
at time t. The first time, tD , at which D(t) reaches the failure-definition level D f a soft failure occurs and
we say that the unit has failed. Thus
tD = min{t : D(t) reachesD f }. (1)
In particular, when D(t) is monotone increasing,
tD = min{t : D(t) ≥ D f }.
Figure 3 illustrates the relationship between the degradation levels and the soft failure times. Define T
as the collection of the failure times tD for all the units in the population. The cumulative distribution
function (cdf) of T, which gives the fraction of the population failing before time t, is defined by
F (t) = Pr(T ≤ t).

Figure 3. Illustration of possible shapes of degradation paths and the relationship between the degradation
level and the failure time.
Reliability assessments and maintenance scheduling can be made based on the estimates of F(t). The
estimate of F(t) is obtained using the observed degradation paths of the units.
Statistical models that specify the degradation paths are useful in degradation analysis. Two large
classes of degradation models are general path models and stochastic models. When the interest is in
dynamic condition monitoring, Bayesian updating models are often used. The following list outlines the
general ideas and summarizes the related literature.
r General path modeling: The basic idea is to find a parametric model, with vector parameter h, that
fits all paths well. Some elements of h may be random to account for unit-to-unit variabilities. This
parametric model determines the cdf of the failure-time distribution. Important literature in this
class includes Lu and Meeker 20, Meeker, Escobar, and Lu 24, and Meeker and Escobar [22, Chaps.
13 and 21].
r Stochastic modeling: This class of models assumes that the data are generated from a stochastic
process, such as the Wiener process (See Brownian motion) (e.g., Whitmore 36), gamma process
(e.g., Lawless and Crowder 19), inverse Gaussian process (e.g., Wang and Xu 35), etc. The cdf of
the failure-time random variable is determined by using the properties of the assumed underlying
stochastic process. Some important references for these models are Whitmore 36, Whitemore,
Crowder and Lawless 37, Bagdonavičius and Nikulin [2, Chaps. 3 and 13], Lawless and Crowder 19,
Park and Padgett 27, Wang 34, and Wang and Xu 35.
r Dynamic condition monitoring: Degradation data can also be used to do dynamic condition monitor-
ing. Based on the degradation data available at time ti , the conditional (conditional on the degradation
level at ti ) failure-time distribution is obtained. Risk assessment and maintenance scheduling deci-
sions are based on the estimated conditional failure-time distribution. When more data are available
at the next time point ti+1 , one updates the conditional failure-time distribution. Gebraeel and Pan
10 considered dynamic condition monitoring using Bayesian updating models.
1.4 Overview
The rest of this article is organized as follows. Section titled General Path Modeling introduces general
degradation path modeling. Various examples will also be given to illustrate the use of general degradation
path modeling. Section titled Stochastic Modeling introduces several stochastic models for modeling
degradation data. Section titled Dynamic Condition Monitoring describes several applications and gives
examples where dynamic condition monitoring is used or needed. Section titled Accelerated Destructive
Degradation Test introduces a model for accelerated destructive degradation data and illustrates the use of
the model with an example dataset. Section titled Concluding Remarks gives some concluding remarks.
2 General Path Modeling
2.1 Physically Motivated Path Models
General path modeling uses parametric forms for D(t). Because most degradation processes are connected
to the underlying physical/chemical mechanism, physically motivated parametric forms of D(t) can, for
many applications, provide a good fit to the data. Here we introduce some examples of physically motivated
path models for degradation data.
Example 3. Linear Degradation Path In some simple wear process, a linear degradation assumption
is adequate. Automobile tire wear provides a good example. For the linear degradation path, the degra-
dation rate is dD(t)/dt = C where C is constant over time. The degradation path can be expressed as
D(t) = D(0) + Ct. The parameters D(0) and C can be random from unit to unit to account for unit-to-unit
variability.
Example 4. Convex Degradation Path Convex degradation paths occur when the degradation rate
increases with the degradation level. The growth of fatigue cracks in a metal plate provides a good
example of a convex degradation path. Let D(t) denote the size of a crack at time t. For cracks within a
certain size range, the crack size growth rate can be modeled by a simple deterministic version of the
Paris-rule model (e.g., Dowling 7). That is,
dD(t) m
= C × 1K [D(t)] . (2)
dt
Here C and m are constants that depend on material properties. 1K[D(t)] is known as the stress intensity
range function which is a function of crack size, the range of applied
p stress, part dimensions, and geometry.
In some application, 1K[D(t)] can be modeled as 1K [D(t)] = S BD(t) where S is the stress on the metal.
For this case, the solution to the differential equation (2) is

D(t)

 1− m m

 [D(0)] 2 + 1 − C

 2


= √ m 2

 ×(S B) t } 2 − m , m 6= 2



 h i
 D(0)exp C(S√B)2 t , m = 2.
Example 5. Concave Degradation Path Concave degradation paths occur when the degradation rate
decreases with the degradation level. As an example, Meeker and LuValle 25 describe models for growth
of failure-causing conducting filaments of chlorine-copper compounds in printed-circuit boards. These
filaments cause failure when they reach from one copper-plated through-hole to another. Let A(t) be
the amount of chlorine available for reaction and D(t) be proportional to the amount of failure-causing
chlorine-copper compounds. At time t, copper combines with chlorine A(t) to produce the chlorine-copper
compound D(t) with rate constant k, that is
dA(t) dD(t)
= −k A(t) and = k A(t).
dt dt
The solution of this differential equation system is
A(t) = A(0)exp(−kt)
D(t) = D(0) + A(0)[1 − exp(−kt)]
where A(0) and D(0) are initial amounts. If D(0) = 0, the solution for the degradation path D(t) is
D(t) = A(0)[1 − exp(−kt)],
which is a concave function of t.
2.2 Approximate Degradation Analysis
The approximate degradation data analysis approach is simple and commonly used by engineers. It
consists of the following steps:
1. Find a transformation for the observed degradation path and/or the time such that the transformed
degradation paths are approximately linear functions of the transformed time. If the observed paths
are already approximately linear, no transformation is needed.
2. Fit separate simple linear regression lines to the observed degradation path for each unit.
3. Use these fitted lines to predict the time at which the unit will reach the specified failure-definition
level D f . These times are called “pseudo-failure times.”
4. Use the failure-time data methods approach to analyze the pseudo-failure times to provide estimates
of F(t).
The method is called the approximate method because the analysis is based on the approximate pseudo
failure times. For simple problems, the approximate degradation analysis method is attractive because
the computations are relatively simple. The approximate method is less appealing, however, when the
degradation paths cannot be transformed to be approximately linear.
Figure 4. Weibull probability plot for the laser pseudo failure times showing the ML estimate for F(t) and the
pointwise approximate 95% confidence intervals. The dots track the Kaplan-Meier estimate of the cdf F(t).
The following gives an example of applying the approximate methods to the laser data. For the laser
data, no transformation is needed because the original paths are already approximately linear.
Example 6. Laser Life Data Analysis The laser data are shown in Figure 1. The failure times (for paths
exceeding a 10% increase in current before 4000 hours) and the pseudo-failure times were obtained by
fitting straight lines through the data for each path. In particular, these pseudo-failure-time values are
3702, 4194, 5847, 6172, 5301, 3592, 6051, 6538, 5110, 3306, 5326, 4995 4721, 5689, and 6102. A Weibull
distribution was fitted to these pseudo-failures (e.g., Meeker and Escobar [22, Chap. 8].
Figure 4 displays the Weibull-fitted distribution in a Weibull probability plot. The plot shows the ML
estimate for F(t) and pointwise-approximate 95% confidence intervals for F(t). Figure 4 also shows the
Kaplan-Meier 17 estimate of F(t), which is a step (staircase) function of time that jumps at each distinct
failure time. The dots at each failure time are plotted halfway up each step in the Kaplan-Meier estimate,
as explained in Lawless [18, Sec. 3.3].
2.3 Random-Effects Path Model
The random-effects path model is a more general approach for analyzing degradation data. Individual
random-effects are used to account for unit-to-unit variability in the degradation paths. This variability
may be caused, for example, by initial conditions, material properties, or component dimensions. That
is, individual units will vary with respect to the amount of material available to wear, the initial level of
degradation, amount of harmful degradation-causing material, operating and environmental conditions,
etc. Conditional on a unit’s parameters and operating/environmental conditions, however, the degrada-
tion path for the unit is a deterministic function of time. A member of this class of models is sometimes
referred to as a deterministic model.
In practice, the values of the degradation level, D(t), are observed or measured at discrete points in
time. The observed degradation for unit i at time tij is
yij = Di (tij ; ", $i ) + ,ij , (3)
where D i (t) is the actual path and ,ij is the residual deviation for unit i at tij . Here " = ("1 , . . . , "l )′
denotes the fixed parameters in the degradation path; these parameters are common across all units.
$i = ($1i , . . . , $ki )′ is random from unit to unit to account for the unit-to-unit variability and k is the
number of random parameters. The random parameter vector is commonly assumed to be multivariate
normal distributed with mean : $ and variance-covariance matrix 6$ . The residual deviations ,ij are
assumed to be independent with a normal distribution [indicated by NOR(0, F , )] where F , is the standard
deviation of ,. When measurement times are not too close together, this independence assumption is a
plausible assumption.
2.3.1 Parameter Estimation
The parameters for the degradation path D(t) in model (3), denoted by h consist of the parameters : $
and 6$ that determine the distribution of the random effects, plus the fixed parameters h. The likelihood
function based on model (3) is
L (2, F, |DATA)
 
n Z
Y ∞ Z ∞ Ymi
 1
= ... Nnor (zij ) (4)
i=1 −∞ −∞ j =1
F,
× f $ ($i ; : $ , 6$ )d$1i . . . d$ki ,
where zij = [yij − D(tij ; ", $i )]/F, (see Section 13.3 of Meeker and Escobar 22 for examples), Nnor is the
probability density function (pdf) of the standard normal distribution, and f $ ($i ; : $ , 6$ ) is the pdf of the
multivariate normal distribution. The maximum likelihood (ML) estimates of h and F , , denoted by 2̂ and
Fb, are obtained by finding the values of h and F , that maximize (4).
2.3.2 Evaluation of F(t) and Confidence Intervals
For the random effects path model, the cdf of the failure-time distribution is
F (t) = F (t; 2) = Pr[D(t) ≥ D f ]

if the degradation is increasing. Note that F(t) only depends on h. There is a similar definition for
decreasing degradation. In some simple cases, it is possible to write down an explicit form for F(t). The
following give a simple example where the F(t) can be obtained explicitly.
Example 7. Linear Degradation with a Lognormal Rate Suppose the actual degradation path is
D(t) = " + $t
where " is fixed and $ has a lognormal distribution with parameters : and F . The explicit form for the
cdf of T is
F (t; "1 , : , F ) = Pr[D(t) ≥ D f ]
(5)
log(t) − [log(D f − "1 ) − :]
= 8nor
F
where 8nor (·) is the standard normal cdf.
For most practical path models, it is not possible to obtain an explicit form for the cdf, especially when
D(t) is a nonlinear function of time and the number of random parameters is more than one. It is then
necessary to compute F(t) by using numerical methods such as numerical integration or Monte Carlo
simulation. As described in Meeker and Escobar [22, Sec. 13.5], the failure-time distribution F(t) can be
evaluated using simulation as outlined in the following algorithm:
Algorithm 1. Evaluate F(t) by simulation
1. Generate N simulated realizations $∗ of h from a multivariate normal distribution with mean : c$

and variance-covariance matrix 6 c$ , where N is a large number (e.g., N = 100,000).
2. Compute the N simulated failure times corresponding to the N realizations of $∗ . This can be done
ˆ $∗ ) and finding the crossing time for each unit.
by substituting the realizations of $∗ into D(t; ",
3. For any t use
F (t) ≈
Simulated First Crossing Times ≤ t
N
as an evaluation of F(t).
To account for the uncertainties of the estimates in F(t), as described in Meeker and Escobar [22, Sec.
13.7], the following algorithm can be used to obtain the bootstrap confidence intervals for F(t).
Algorithm 2. Compute confidence intervals
∧
1. Use the Algorithm 1 to compute the estimate of F (t) at t.
2. Generate a large number B (e.g., B = 4000) of bootstrap samples that mimic the original sample
∧∗
and compute the corresponding bootstrap estimates of F (t) according to the following steps:
∧
a. Generate n simulated realizations of the random path parameters $i∗ , i = 1, . . . , n from $.
b. Using the same sampling scheme as in the original experiment, compute n simulated
observed paths from
∧
yij∗ = D(tij ; ", $∗ ) + ,∗ij
up to the planned stopping time tc where the ,∗ij values are independent simulated residual
values generated from NOR(0, F,2 ).
c. Use the n simulated paths to estimate parameters of the path model, giving the bootstrap
∧∗ ∧∗
estimates 2 and F , .
∧∗ ∧∗
d. Use the above algorithm with 2 as input to compute the bootstrap estimates F (t) at t.
∧
3. For each t, the bootstrap confidence interval for F (t) is computed using the following steps:
∧∗ ∧∗ ∧∗
a. Sort the B bootstrap estimates F (t)1 , . . . , F (t) B in increasing order giving F (t)(b) , b =
1, . . . , B.
b. The bootstrap bias-corrected percentile confidence intervals (Efron and Tibshirani 8) are
used here. The lower and upper bounds of pointwise-approximate 100(1 − ")% confidence
∧
intervals for the F (t) are
∧∗ ∧∗
[ũF (t), F̃ (t)] = [ F (t)(l ) , F (t)(u) ]
where
!
l = B8nor 2zq + z "
2
!
u = B8nor 2zq + z " ,
1−
2
∧∗ ∧
z" = 8−1
nor ("), and q is the proportion of the B values of F (t) that are less than F (t).
The following two examples illustrate the analysis of degradation data using random-effects models.
Example 8. Laser Data Analysis with a Lognormal Rate Here we use the degradation path model in
Example 7 to analyze the laser data. The parameters that need to be estimated are ", :, F , and F , . The
∧ ∧ ∧ ∧
ML estimates obtained by maximizing (4) are "= 0.00804, : = −6.2146, F = 0.20419, and F , = 0.1988.
An estimate of the cdf F(t) is obtained by substituting the ML estimates into (5). A confidence interval
for F(t) is obtained by using Algorithm 2 . Figure 5 shows the degradation model estimate of F(t) with
pointwise-approximate 90% confidence bootstrap intervals for the laser data, based on the random effects
degradation model. The dots track the Kaplan-Meier estimate of the cdf F(t). The estimate from the
degradation model is rising more slowly because the degradation paths of the unfailed units suggest this
future behavior.
Example 9. Analysis of Weathering Data using a Random Effects Model This example uses the weathering
data shown in Figure 2. The degradation model used here is the same as in Vaca-Trigo and Meeker 33 ;
that is,

log(t) − $1 −1
D(t) = "1 1 + exp − ,
exp($2 )
where t is the effective dosage, "1 is a fixed effect, and $ = ($1 , $2 )′ is random from unit to unit. This
degradation path is a nonlinear function of t. h is assumed to have a bivariate normal distribution
with mean : $ and covariance matrix 6$ . The estimates for the unknown parameters "1 , : $ , and 6$
were obtained by maximizing the likelihood function in (4). The estimated failure-time cdf F(t) and the
bootstrap confidence intervals are shown in Figure 6.
Figure 5. Degradation model estimate of F(t) with pointwise-approximate 90% confidence bootstrap intervals
(CIs) for the laser data, based on the random effects degradation model. The dots track the Kaplan-Meier estimate
of the cdf F(t) for those units that had reached the failure level.
2.4 Accelerated Degradation Data
Sections titled Physically Motivated Path Models through Random-Effects Path Model focus on modeling
degradation as a function of time. It is common in degradation data to have explanatory variable infor-
mation such as temperature, loading, voltage, environmental humidity, etc. In accelerated degradation
tests, variables like temperature and voltage are often used as accelerating factors. Thus it is desirable and
useful to have models that can incorporate the explanatory variable information. Below is an example of
degradation data with temperature as an explanatory variable from an accelerated degradation test.
Example 10. Device-B Power Output Degradation Meeker and Escobar [22, Ex. 21.1] describe the Device-
B (an integrated circuit RF power amplifier device) degradation data. Design engineers of Device-B
needed to assess the proportion of these devices that would “fail” before 15 years of operation at 80◦ C.
This information would be used to determine the amount of redundancy required in the full system.
At standard operating temperatures (i.e., 80◦ C), the devices will degrade too slowly to provide useful
information in 6 months. Because units at low temperature degrade more slowly, they had to be run
for longer periods of time to accumulate appreciable degradation. This assessment was based on an
accelerated repeated measures degradation test that used three levels of temperature (150◦ C, 195◦ C, and
237◦ C) and lasted about 6 months. Figure 7 shows the accelerated degradation test results. The failure
for an individual device was defined as the time at which the output power of a device dropped to a level
0.5 decibels (dB) below the initial output.
Figure 6. Degradation model estimate of F(t) with pointwise-approximate 90% confidence bootstrap intervals
(CIs) for the outdoor weathering data, based on the random effects degradation model. The dots track the
Kaplan-Meier estimate of the cdf F(t) for those units that had reached the failure level.
For temperature the following model is often used,
D(t; temp)
(6)
= D∞ {1 − exp[−RU AF (temp)t]},
where R U is the reaction rate constant at use temperature tempU , RU AF (temp) is the reaction rate constant
at temperature temp, and D ∞ is the asymptote. The Arrhenius acceleration factor is
AF (temp) = AF (temp, tempU , E a )

11605
= exp E a
tempU + 273.15

11605
−
temp + 273.15
which depends on the two temperature levels and the effective activation energy Ea . If temp > tempU ,
then AF (temp, tempU , E a ) > 1. In general, rate-acceleration parameters are unknown fixed-effect
parameters. That is, this form of the Arrhenius model suggests no unit-to-unit variability in activation
energy.
The following example illustrates the analysis of degradation data with explanatory variables.
Figure 7. Accelerated degradation test results giving power drop in Device-B output for a sample of units tested
at three levels of junction temperature.
Example 11. Device-B Data Analysis We assume R U and D ∞ are random from unit to unit. Then a
possible parametrization would be $ = ($1 , $2 )′ = [log(RU ), log(−D∞ )]′ , which are random effects
distributed with a bivariate normal distribution. The activation energy " = Ea is a fixed effect. That is, Ea
is regarded as a material property that does not depend on temperature and that is constant from unit to
unit. To improve the stability and robustness of the approximate ML algorithm, we use $1 = log[R (195)],
$2 = log(−D ∞ ) where R (195) = RU AF (195) is the reaction rate constant at 195◦ C. The ML estimates
∧ ∧
obtained by maximizing (4) are "= 0.6670, : $ = (−7.572, 0.3510)′ ,
!
0.15021 −0.02918
b$ =
6 ,
−0.02918 0.01809
∧
and F , = 0.0233. Figure 8 shows the fitted degradation paths for the degradation data.
∧ ∧ ∧ ∧
Using 2, one obtains the failure-time distribution estimate from F (t) = F (t; 2). Numerical integration
was used to compute the cdf and its estimate. There is, in general, no simple procedure to compute the
∧
standard errors for F (t). A simulation approach is useful for computing the confidence intervals for
∧
F (t).
∧
Example 12. Estimates of and Confidence Intervals for F(t) for the Device-B Data Figure 9 shows F (t)
for the Device-B based on the power-drop data with failure defined as power drop of D f = −0.5 dB at
80◦ C. These estimates were computed with Algorithm 1 . Figure 8 also shows a set of pointwise two-
Figure 8. Device-B power-drop observations and fitted degradation model for Device-B.
sided approximate 90% and 80% bootstrap bias-corrected percentile confidence intervals for F(t) using
Algorithm 2 .
3 Stochastic Modeling
Stochastic process models are also useful in the analysis of degradation data. In this section, we introduce
the Wiener process (e.g., Whitmore 36), the gamma process (e.g., Lawless and Crowder 19), and the inverse
Gaussian process (e.g., Wang and Xu 35) models that have appeared in the literature on degradation data
analysis. Without loss of generality, in this section, we assume that the degradation path D(t) is monotone
increasing. For a monotone decreasing degradation path, the model can be transformed into increasing
degradation using −D(t).
3.1 Wiener Process
The degradation path using the Wiener process model is

D(t) = 0(t) + > W [0(t)],
where 0(t) is a continuous monotone increasing function reflecting trend, W(t) is the standard Wiener
process, and > > 0 is a scale parameter. Some basic properties of the Wiener process can be found in, for
Figure 9. Estimates of the Device-B failure-time distribution at 80◦ C (solid line) with approximate 80% and
90% pointwise two-sided bootstrap confidence intervals (dotted lines) based on the power-drop data with failure
defined as a power drop of D f = −0.5 dB.
example, Cox and Miller 5. It can be shown that the model has the important independent increments
property. That is, the increments 1D(t) = D(t + 1t) − D(t) are independent. Also, 1D(t) has a normal
distribution with mean 10(t) = 0(t + 1t) − 0(t) and variance > 2 10(t). In particular, the pdf of 1D(t) is
f [1D(t); 10(t), > ]

" #
1 1D(t) − 10(t) (7)
= p Nnor p .
> 10(t) > 10(t)
The failure time T is defined as the first time that D(t) crosses the failure threshold D f . The induced
failure-time distribution is identical with 0−1 (TIG ) where TIG is a random variable with an IG distribution
IG(D f , D 2f /> 2 ). In particular, the pdf of an IG(:, 8) is
f (t; : , 8)
1 " #
8 2 −8(t − :)2 (8)
= exp , t > 0,
2Bt 3 2: 2 t
where : > 0 is the mean of the IG distribution and 8 > 0 is a shape parameter. The cdf of an IG
distribution is
"r #
8 t
G(t; : , 8) = 8nor −1
t :
" r # (9)
28 8 t
+exp 8nor − +1 .
: t :
More details about the IG distribution can be found in, for example, Chikkara and Folks 4.
3.2 Gamma Process
The gamma process D(t) can be used to model monotone degradation paths. A monotone increasing
gamma process is characterized by a monotone increasing function 0(t) and a scale parameter > . The
gamma process has the independent increment property. The increments 1D(t) have a gamma distribu-
tion with scale parameter > and shape parameter equal to 10(t) = 0(t + 1t) − 0(t). The pdf of a gamma
distribution with shape parameter k > 0 and scale parameter > > 0 is given by

1 t
f (t; k, > ) = k t k−1 exp − (10)
> Ŵ(k) >
where
Rt t > 0 and Ŵ(·) is the gamma function. The cdf of a gamma distribution is G(t; k, > ) =
0 f (u; k, > ) du. The induced failure-time cdf is given by F(t) = 1 − G[D f ;0(t), > ].
3.3 Inverse Gaussian Process
Wang and Xu 35 describe the inverse Gaussian (IG) process as an alternative to the Wiener and gamma
processes. The inverse Gaussian process D(t) has a mean function 0(t) and a scale parameter > . Like
the Wiener and the gamma processes, the IG process has the independent increment property. The
increments 1D(t) have an IG distribution IG{10(t); > [10(t)]2 }. The failure-time distribution is F (t) =
1 − G{D f ; 0(t), > [0(t)]2 } where G is the cdf of the IG as given in (9).
3.4 Parameter Estimation
Here we briefly describe the parameter estimation procedures for stochastic degradation models. The
likelihood is
L (", > |DATA)
n m
Y Yi −1
(11)
= f [1Di (tij ); 10(tij ), > ],
i=1 j =0
where α denotes the parameter(s) in the trend function 0(t), 1D i (tij ) = D i (ti, j+1 ) − D i (ti, j ), 10(tij ) =
0(ti, j +1 ; ") − 0(tij ; "), ti0 = 0, and mi is the number of observations for unit i. The expression for
f[1D i (tij );10(tij ), > ] is obtained from the pdf in (7), (10), and (8) for the Wiener, gamma, and IG processes,
respectively.
Figure 10. Induced cdf estimate from the three stochastic models and the Kaplan-Meier estimate of the cdf
with approximate 95% simultaneous nonparametric confidence bands.
A parametric form of 0(t) is needed for model fitting. A common choice for the trend function 0(t) is
the power law relationship
"1
t
0(t) = 0(t; ") = , (12)
"2
where " = ("1 , "2 )′ . The ML estimate of (", > )′ is obtained by finding the value of (", > )′ that maximizes
the likelihood function in (11).
Example 13. Stochastic Model Fitting for the Outdoor Weathering Data Here we apply stochastic models
to the weathering data for the purpose of illustration. We fit the power law relationship in (12). Table 1
gives parameter estimates for different stochastic models and the negative log likelihoods. Note that the
parameters have different interpretations for the three different stochastic models. Thus the estimates
are not comparable across different models. Figure 10 shows the estimated induced cdf from the three
stochastic models and the Kaplan-Meier estimates of the cdf with the 95% simultaneous confidence
bands. The simultaneous nonparametric confidence bands were computed using the method in Nair 26.
Based on the plots, there is not a clear indication of which stochastic model fits the data best. Overall the
random effects model in Example 9 fits better than any of the three stochastic models considered here.
Table 1. Parameter estimates and the negative log likelihood for the different stochastic models
∧ ∧ ∧ ∧ ∧
Model "1 (se) "2 (se) > (b
se) −log likelihood
Wiener 0.7848 (0.0165) 4.3482 (0.1729) 0.1122 (0.0028) −3136.039

Gamma 0.7846 (0.0120) 0.0071 (0.0006) 0.0065 (0.0003) −3434.878
IG 0.7777 (0.0099) 4.3963 (0.1225) 142.50 (7.9907) −1487.961
4 Dynamic Condition Monitoring
4.1 Applications
Another application area of degradation data is condition monitoring, which can provide more timely
information for decision and maintenance. Modern sensor and communications technology along with
smart chips can provide system status or health information in real time.
The monitoring of modern locomotive engines provides one example. Hahn and Doganaksoy [11,
Sec. 9.9] describe the monitoring of modern locomotive engines with sensors. These sensors monitor
the operating parameters such as oil pressure, oil temperature, and water temperature. This real-time
information is used to detect and avoid catastrophic failures by automatically shutting down a locomotive
in a timely manner. More details will be given in Example 14 .
The monitoring of high-voltage power transformer oil cooling tanks provides another example. These
transformers can be monitored by an automatic dissolved gas analyzer (DGA) system (e.g., Spurgeon et
al. 32). The DGA system performs periodic (typically hourly) analyses to indicate the presence of different
kinds of dissolved gases in the transformer insulating oil. Certain combinations of dissolved gases can
signal a system health problem. Such information would enable operators to take appropriate actions
such as reducing the load on the transformer, or in an extreme case, shutting it down to avoid a costly
catastrophic failure.
4.2 Examples
Here we give two examples to illustrate the ideas of conditional monitoring.

Example 14. Locomotive Engine Oil Pressure Monitoring As described in Doganaksoy, Hahn, and Meeker
6, modern locomotive engines are equipped with numerous sensors that read operating parameters such
as oil pressure, oil temperature, and water (coolant) temperature. They considered an apparent large drop
in engine oil pressure that could be caused by any one of the three reasons. Appropriate actions, ranging
from activating a backup sensor to shutting down the engine, are taken according to the associated cause
and problem severity.
The causes for the indication of a drop in oil pressure are identified from the sensor data on oil
pressure, oil temperature, and water temperature changes over time. In particular,
1. If the oil-pressure sensor fails, the drop in the oil pressure reading would not be accompanied by
any discernible change in either oil or water temperature.
2. If the cooling system fails, a sharp rise in water temperature would precede the drop in measured
oil pressure, which was also accompanied by a sharp rise in oil temperature.
3. If the oil pump fails, a sharp rise in oil temperature would be followed by a drop in measured oil
pressure. Water temperature would remain unchanged.
By analysis of the patterns in the dynamic sensor data, appropriate proactive actions can be taken in
response to the observed oil pressure drop.
Example 15. Bayesian Model Updating Gebraeel and Pan 10 described the monitoring of the vibration
signal of a bearing and use the information to update the remaining life distribution. The vibration signal,
which has an increasing trend, gives the degree of degradation. The bearing fails when the signal reaches
the failure threshold D f . The model for the observed degradation data is
yij = Dij (tij ; ", $i ) + ,ij .
In a Bayesian framework, it is assumed that there is prior information on $i , denoted by B($i ). When
the degradation data are available at time ti , one obtains the posterior distribution p($i ). The distribution
of the remaining life is obtained from p($i ). Risk assessment and scheduled maintenance decisions can
be made based on the remaining life distribution. When more data become available at the next time
point ti+1 , one updates the remaining life distribution and the corresponding decision is correspondingly
updated.
5 Accelerated Destructive Degradation Test
5.1 ADDT data
Different from repeated measurement degradation data, destructive degradation data occur when the
degradation measurements destroy or change the physical/mechanical characteristics of test units. In
such cases, only one meaningful measurement can be taken on each test unit. Acceleration is usually
used in such degradation tests. The data obtained from this kind of test are referred to as accelerated
destructive degradation test (ADDT) data. The following gives an example of ADDT data.
Example 16. Adhesive Bond B ADDT Data Escobar et al. 9 describe data from an ADDT test of an adhesive
Bond, which is called Adhesive Bond B. The objective of the experiment was to assess the strength of
an adhesive bond over time. Engineers were interested in estimating the proportion of devices with a
strength below 40 Newtons after 5 years of operation at the use temperature of 25◦ C. Acceleration was
required because the test needed to be completed in 16 weeks. The test was destructive because strength
of the bond could be measured only once on each unit.
Table 2 shows the test plan. There were 8 units measured at the start of the experiment, with no aging,
and there were an additional 80 units aged and measured at the times and temperatures shown in Table
2. Figure 11 shows the strength data. There were six observations, at three different levels of temperature,
marked with a N. These observations are considered as right-censored values, indicating that the actual
level of strength is unknown, but certainly larger than the recorded value. These observations were right
censored because the low response values indicated that the test specimens were most likely defective
due to fabrication errors.
5.2 Models for ADDT Data
The actual degradation path D ij of a unit at time ti and at a particular accelerating variable condition
AccVarj is modeled as Dij = D(Ji , x j , $). Here J i = ht (ti ) and xj = ha (AccVarj ) are known monotone
increasing transformations of the original time scale ti and the accelerating variable condition AccVarj ,
respectively. The path parameters β are fixed but unknown.
Figure 11. Adhesive Bond B ADDT data scatter plot.
Table 2. Adhesive Bond B test plan
Temp ◦ C Weeks Aged

0 2 4 6 12 16
— 8
50 8 0 8 8 7
60 6 0 6 6 6
70 6 6 4 9 0
The form of the function D ij and the appropriate transformations need to be chosen for particular
applications. Physical-chemical theory, past experience, or the data can be used to suggest appropriate
transformations. Examples of such transformations can be found in Meeker and LuValle 25 and Meeker,
Escobar, and Lu 24. Using the Arrhenius relationship for the temperature acceleration, the form of D ij is
Dij = exp[$0 + $1 exp($2 x j )Ji ] (13)
where xj = −11605/(◦ Cj + 273.15) is Arrhenius-transformed temperature.
The sample degradation path for unit k at time J i and accelerating variable condition xj is modeled as
yijk = : ij + ,ijk , (14)
where yijk is the observed degradation, : ij = hd (D ij ) is a monotone increasing transformation of the actual
degradation path D ij , and ,ijk is a residual deviation that describes unit-to-unit variability. The cdf of
(,ijk /F ) is 8(z), which needs to be specified. For example, 8(z) = 8nor (z) provides a normal model and
8(z) = 8sev (z) provides a smallest extreme value model.
Using a log transformation of D ij [i.e., hd (u) = log(u)] and the Arrhenius relationship in (13), the model
in (14) becomes
yijk = $0 + $1 exp($2 x j )Ji + ,ijk . (15)
For multiple AccVar situations (e.g., temperature and humidity), the term $2 xj can be replaced by a linear
combination of accelerating variables $′2 x j . The interpretation of the degradation model parameters is as
follows. $0 is the (transformed) degradation level when J i = 0 and L (xj ) = $1 exp($2 xj ) is the degradation
rate with respect to transformed time scale J at AccVar level xj . The sign (±) of $1 determines whether
degradation is increasing or decreasing in time.
5.3 Parameter Estimation
For a sample of n units consisting of exact observations and right-censored observations, the likelihood
based on model (15) is
L (2|DATA)
Y
= L j (2|DATA)
j
Y1
yijk − : ij
*ijk
(16)
= N
F F
ijk

yijk − : ij 1−*ijk
× 1−8
F
where 2 = ($0 , $1 , $2 , F ) and *ijk indicates whether observation ijk is an exact observation (*ijk = 1) or
a right censored observation (*ijk = 0).
Example 17. Adhesive Bond B data Model Fit This example uses model (15) and ML estimation to fit the
Adhesive Bond B data. Figure 12 shows the ML estimates for the data at the three levels of temperature
used in the experiment and the estimate at use conditions of 25◦ C.
5.4 Induced Failure-Time Distribution
For the special case of a linearly decreasing degradation, the cdf of the induced failure-time distribution
at a fixed value of (AccVar, D f ) is
FT (t; x, $) = Pr(T ≤ t)

: f − :(J, x, $)
=8
F (17)

J −<
=8
l
for t > 0 where : f = h d (D f ), J = ht (t), < = ($0 − : f )exp(−$2 x)/|$1 |, and l = F exp(−$2 x)/|$1 |.
Figure 12. Normal distribution Arrhenius model fit to the Adhesive Bond B ADDT data.
The failure-time distribution in (17) is a mixed distribution with a spike of probability, Pr(T = 0) =
8[($0 − : f )/F ] at t = 0. For t > 0 the cdf is continuous and it agrees with the cdf of a log-location-scale
variable with standardized cdf 8(z), with location parameter < and scale parameter F
Example 18. Adhesive Bond B cdf For the Adhesive Bond B application with D f = 40, Figure 13 provides
a visualization of the failure-time distribution induced by the degradation model at 25◦ C, based on the ML
estimates of the Adhesive Bond B example. The figure shows clearly the reason for the spike of probability
at time zero.
6 Concluding Remarks
In this section, we provide some discussion on aspects of degradation models and data analysis that have
not been covered in the main text.
6.1 Test Planning
This section briefly discusses some issues in accelerated degradation test planning. Because degradation
data can provide more timely information than the common failure-time data, accelerated degradation
tests are desirable for obtaining product reliability information. There is some literature on this topic.
Meeker and Escobar [22, Chap. 22.5] describe how to use simulation methods to help planning accelerated
Figure 13. ML estimate showing the proportion failing as a function of time at condition 25◦ C.
degradation tests. Boulanger and Escobar 3 propose methods for planning repeated measures degradation
tests. Shi et al. 30 describe methods for finding good ADDT plans for a class of destructive degradation
models.
In test planning, one needs to specify the factor-level combinations of an accelerating variable such
as temperature and voltage. One also needs to determine the length of the test and how to allocate units
to those combinations. The goal of test planning is to find a plan that minimizes a criterion such as the
large sample approximate variance of the ML estimator of a specified quantile of the distribution that is of
interest and meets practical constraints (e.g., robustness to misspecification of inputs). Both large-sample
approximations and simulation methods can be used to assess the performance of the test plan.
6.2 Software Implementations
Most of the degradation model and analysis examples in Meeker and Escobar 22 are implemented in
SPLIDA 23, a free S-Plus 31 add-on package that can be downloaded from the Internet. An R 28 version
of SPLIDA, which is named RSPLIDA, is also available. The R 28 statistical package is a free package and
provides a flexible statistical computing environment for researchers. Some degradation models can be
implemented by the built-in functions, such as lme or nlme functions R.
In terms of commercial software, JMP 9 15 has tools for both repeated measures degradation (the
simple method) and destructive degradation. Weibull++ 29 implements the approximate degradation
analysis approach, which uses models such as linear, exponential, and power-rule to extrapolate the
failure times of a product based on the observed degradation data.
6.3 The Future of Degradation Data Analysis
We expect to see a continuing increase in the use of degradation data in reliability applications. For
systems and products that are used outdoors and in other dynamic environments, there is often more
variability in the system reliability. Modern data collection technology, degradation data, and environ-
mental information can be recorded and transmitted over the network. Hong and Meeker [12, 13] used
userate information and dynamic environmental information in predicting product reliability. Degrada-
tion models also need to be developed to analyze data from the performance tests of solar cells 16 and
other new areas.
Acknowledgments. Figures 1, 4, 5, 7, 8, and 9 were taken from Meeker and Escobar 22, and Figures 2,
3, and 6 were taken from Hong, Meeker, and Escobar 14, with permission from John Wiley & Sons, Inc.
7 Related Articles
Destructive Testing ; Physical Degradation Models; Optimal Sample Size Allocation for Accelerated
Degradation Test Based on Wiener Process; Degradation Processes; Degradation and Shock Models;
Degradation and Failure.
References
[1] Bae, S. J. and Kvam, P. H. (2004). A nonlinear random-coefficients model for degradation testing. Technometrics, 46,
460–469.
[2] Bagdonavičius, V. and Nikulin, M. (2001). Accelerated Life Models: Modeling and Statistical Analysis. Chapman & Hall/
CRC.
[3] Boulanger, M. and Escobar, L. A. (1994). Experimental design for a class of accelerated degradation tests. Technometrics,
36, 260–272.
[4] Chikkara, R. S. and Folks, J. L. (1989). The Inverse Gaussian Distribution. Marcell Dekker.
[5] Cox, D. R. and Miller, H. D. (1965). The Theory of Stochastic Processes. Chapman & Hall.
[6] Doganaksoy, N., Hahn, G. J., and Meeker, W. Q. (2008). Statistics roundtable: The pros of proactive product servicing.
Quality Progress, 40, 60–62.
[7] Dowling, N. E. (1993). Mechanical Behavior of Materials. Prentice Hall.
[8] Efron, B. and Tibshirani, R. (1993). An Introduction to the Bootstrap. Chapman and Hall/CRC.
[9] Escobar, L. A., Meeker, W. Q., Kugler, D. L., and Kramer, L. L. (2003). Accelerated destructive degradation tests: Data,
models, and analysis. In Mathematical and Statistical Methods in Reliability. B. H. Lindqvist and K. A. Doksum, (eds.),
World Scientific Publishing Company.
[10] Gebraeel, N. and Pan, J. (2008). Prognostic degradation models for computing and updating residual life distributions
in a time-varying environment. IEEE Transactions on Reliability, 57, 539–550.
[11] Hahn, G. J. and Doganaksoy, N. (2008). The Role of Statistics in Business and Industry. Wiley, New York.
[12] Hong, Y. and Meeker, W. Q. (2010). Field-failure and warranty prediction based on auxiliary use-rate information.
Technometrics, 52, 148–159.
[13] Hong, Y. and Meeker, W. Q. (2010). Field Failure Prediction Using Dynamic Environmental Data. Mathematical and
Statistical Methods in Reliability. Applications to Medicine, Finance and Quality Control. Birkhauser, Boston, in press.
[14] Hong, Y., Meeker, W. Q., and Escobar, L. A. (2010). The condition-based paradigm. The Wiley Encyclopedia of Operations
Research and Management Science. Wiley, New York.
[15] JMP (2010). JMP 9. http://www.jmp.com/.
[16] Jøgensen, M., Norrman, K., and Krebs, F. C. (2008). Stability/degradation of polymer solar cells. Solar Energy Materials
& Solar Cells, 92, 686–714.
[17] Kaplan, E. L. and Meier, P. (1958). Nonparametric estimation from incomplete observations. Journal of the American
Statistical Association, 53, 457–481.
[18] Lawless, J. F. (2003). Statistical Models and Methods for Lifetime Data, 2nd ed. Wiley, New York.
[19] Lawless, J. F. and Crowder, M. (2004). Covariates and random effects in a gamma process model with application to
degradation and failure. Lifetime Data Analysis, 10, 213–227.
[20] Lu, C. J. and Meeker, W. Q. (1993). Using degradation measures to estimate a time-to-failure distribution. Technometrics,
34, 161–174.
[21] Madhavan, J., Grieser, F., and Ashokkumar, M. (2010). Combined advanced oxidation processes for the synergistic
degradation of ibuprofen in aqueous environments. Journal of Hazardous Materials, 178, 202–208.
[22] Meeker, W. Q. and Escobar, L. A. (1998). Statistical Methods for Reliability Data. John Wiley & Sons, Inc.
[23] Meeker, W. Q. and Escobar, L. A. (2008). SPLIDA (S-PLUS Life Data Analysis). www.public.iastate.edu/∼splida.
[24] Meeker, W. Q., Escobar, L. A., and Lu, C. J. (1998). Accelerated degradation tests: modeling and analysis. Technometrics,
40, 89–99.
[25] Meeker, W. Q. and LuValle, M. J. (1995). An accelerated life test model based on reliability kinetics. Technometrics, 37,
133–146.
[26] Nair, V. N. (1984). Confidence bands for survival functions with censored data: a comparative study. Technometrics, 26,
265–275.
[27] Park, C. and Padgett, W. J. (2005). Accelerated degradation models for failure based on geometric Brownian motion and
gamma processes. Lifetime Data Analysis, 11, 511–527.
[28] R (2010). The R Project for Statistical Computing. http://www.r-project.org/.
[29] ReliaSoft (2010). Weibull++. http://weibull.reliasoft.com/.
[30] Shi, Y., Escobar, L. A., and Meeker, W. Q. (2009) Accelerated destructive degradation test planning. Technometrics, 51,
1–13.
[31] S-Plus (2010). S-Plus. http://spotfire.tibco.com/products/s-plus/statistical-analysis-software.aspx.
[32] Spurgeon, K., Tang, W. H., Wu, Q. H., Richardson, Z. J., and Moss, G. (2005). Dissolved gas analysis using evidential
reasoning. IEE Proceedings - Science, Measurement & Technology, 152, 110–117.
[33] Vaca-Trigo, I. and Meeker, W. Q. (2009). A statistical model for linking field and laboratory exposure results for a model
coating. In Service Life Prediction of Polymeric Materials. J. Martin, R. A. Ryntz, J. Chin, and R. A. Dickie (eds.). Springer.
[34] Wang, X. (2007). Physical degradation models. In Encyclopedia of Statistics in Quality and Reliability F. Ruggeri, R. Kenett,
and F. W. Faltin, (eds.), pp. 1356–1361. Wiley, New York.
[35] Wang, X. and Xu, D. (2010). An inverse Gaussian process model for degradation data. Technometrics, 52, 188–197.
[36] Whitmore, G. A. (1995). Estimation degradation by a Wiener diffusion process subject to measurement error. Lifetime
Data Analysis, 1, 307–319.
[37] Whitmore, G. A., Crowder, M. J., and Lawless, J. F. (1998). Failure inference from a marker process based on a bivariate
Wiener model. Lifetime Data Analysis, 4, 229–251.

Meeker 2014

Uploaded by

Copyright:

Available Formats

Meeker 2014

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

Meeker 2014

Uploaded by

Copyright:

Available Formats

Degradation Models and Analyses

By William Meeker,1 Yili Hong,2 and Luis Escobar3

Abstract: For products/systems requiring high reliability or availability, reliability assess-

1 Iowa State University

1.2 Data and Examples

Figure 1. Observed degradation paths for laser device units.

1.3 Statistical Models

tD = min{t : D(t) reachesD f }. (1)

In particular, when D(t) is monotone increasing,

F (t) = Pr(T ≤ t).

2 General Path Modeling

2.1 Physically Motivated Path Models

For this case, the solution to the differential equation (2) is

2.2 Approximate Degradation Analysis

2.3 Random-Effects Path Model

yij = Di (tij ; ", $i ) + ,ij , (3)

2.3.1 Parameter Estimation

× f $ ($i ; : $ , 6$ )d$1i . . . d$ki ,

2.3.2 Evaluation of F(t) and Confidence Intervals

F (t) = F (t; 2) = Pr[D(t) ≥ D f ]

1. Generate N simulated realizations $∗ of h from a multivariate normal distribution with mean : c$

2.4 Accelerated Degradation Data

For temperature the following model is often used,

AF (temp) = AF (temp, tempU , E a )

3.1 Wiener Process

The degradation path using the Wiener process model is

f [1D(t); 10(t), > ]

3.2 Gamma Process

3.3 Inverse Gaussian Process

3.4 Parameter Estimation

Wiener 0.7848 (0.0165) 4.3482 (0.1729) 0.1122 (0.0028) −3136.039

4 Dynamic Condition Monitoring

Here we give two examples to illustrate the ideas of conditional monitoring.

yij = Dij (tij ; ", $i ) + ,ij .

5 Accelerated Destructive Degradation Test

5.1 ADDT data

5.2 Models for ADDT Data

Figure 11. Adhesive Bond B ADDT data scatter plot.

Table 2. Adhesive Bond B test plan

Temp ◦ C Weeks Aged

5.3 Parameter Estimation

5.4 Induced Failure-Time Distribution

6.1 Test Planning

6.2 Software Implementations

6.3 The Future of Degradation Data Analysis

You might also like