See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/305514796

Planning Fatigue Tests for Polymer Composites

Article  in  Journal of Quality Technology · July 2016

DOI: 10.1080/00224065.2016.11918163

CITATIONS READS

10 976

5 authors, including:

Caleb Bridges King Yili Hong

SAS Institute Virginia Polytechnic Institute and State University
19 PUBLICATIONS   80 CITATIONS    111 PUBLICATIONS   1,180 CITATIONS   

SEE PROFILE SEE PROFILE

Stephanie Dehart Rong Pan

Eastman Arizona State University
5 PUBLICATIONS   66 CITATIONS    143 PUBLICATIONS   1,246 CITATIONS   

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

quality engineering View project

Generative Models for Functional Data View project

All content following this page was uploaded by Yili Hong on 01 August 2016.

The user has requested enhancement of the downloaded file.

 Planning Fatigue Tests for Polymer Composites
Caleb B. King1∗, Yili Hong2†, Stephanie P. DeHart3‡,
Patrick A. DeFeo4§, and Rong Pan5¶
1 Statistics & Surveillance Assessment Group, Sandia National Labs., Albuquerque, NM 87123
2 Department of Statistics, Virginia Tech, Blacksburg, VA 24061
3 Applied Statistics Group, Eastman Chemical Company, Kingsport, TN 37662
4 Applied Statistics Group, DuPont, Wilmington, DE 19803
5 School of CIDSE, Arizona State University, Tempe, AZ 85287

Abstract

Polymer composite materials have become key components in the transportation

and alternative energy industries as they are more lightweight than homogeneous met-
als and alloys yet still retain comparable levels of strength and endurance. To un-
derstand how these polymer composites perform after long periods of use, material
manufacturers commonly use cyclic fatigue testing. The current industrial standards
include test plans with balanced designs and equal spacing of stress levels in many
cases are not the most statistically efficient designs. In this paper, we present op-
timal designs with the goal of minimizing the weighted sum of asymptotic variances
of an estimated lifetime percentile at selected design stress levels. These designs are
based on a physical model adapted from the fatigue literature which is more suitable
for modeling cyclic fatigue of polymer composites than the model used in the current
industrial standards. We provide a comparison between our optimal designs and the
traditional designs currently in use and ultimately propose a compromise design for use
∗
Dr. King is a Research Statistician at Sandia National Labs. His email address is calking@sandia.gov.
†
Dr. Hong is an Associate Professor at Virginia Tech. He is the corresponding author. His email address
is yilihong@vt.edu.
‡
Dr. DeHart is a Principal Statistician at Eastman Chemical Company. She is a full member of ASQ.
Her email address is stephaniedehart@eastman.com.
§
Dr. DeFeo is a Statistician Principal Consultant at DuPont. His email address is
patrick.a.defeo@dupont.com.
¶
Dr. Pan is an Associate Professor at Arizona State University. He is a Senior Member of ASQ. His email
address is rong.pan@asu.edu.

1
 by practitioners in order to ensure robustness against deviations from the underlying
assumptions.
Key Words: Fatigue lifetime model, large-sample approximate variance, lognor-
mal, optimal design, tolerance limits, Weibull.

1 Introduction
1.1 Motivation
All materials, regardless of their physical makeup, exhibit some form of weakening after long
periods of field use. This weakening is known as “fatigue” in the material sciences and is
of great importance to material manufacturers. For example, manufacturers may use this
fatigue behavior knowledge when determining their marketing promotions, product pricing
and differentiation as well as warranties. It may also help manufacturers improve upon
existing products or even allow for the development of new offerings. Accelerated tests are
often used to collect data necessary to assess material fatigue. The most common form of
fatigue testing is known as constant amplitude fatigue testing. In this form of testing, a
sample of material is either stretched or compressed to a maximum stress and then cycled
between that stress and a lower level of stress until failure is achieved. Figure 1 gives a
graphical illustration of this procedure.
An important class of materials is polymer composites, which are made from polymers
or combinations of polymers with other materials. Polymer composites have become more
commonplace as they are more lightweight than homogeneous metals and alloys yet still
retain comparable levels of strength and endurance, a property that makes them more en-
ergy efficient. As a result, polymer composites have become key components in several large
industries related to energy consumption, including transportation manufacturing and al-
ternative energy production. Acquiring knowledge of polymer composite fatigue requires
proper experimentation, which provides opportunities and challenges for test planning. The
majority of testing performed in this field is in accordance with the standards provided in
ASTM E739 (2010) for testing and analysis of polymer composite fatigue. All of the test
plans discussed within this standard are of a balanced nature with equal replication and
spacing of the test samples.
A key characteristic often used in setting standards for and assessment of fatigue behavior
is a specified quantile of the material’s lifetime distribution. For example, a manufacturer
may want to be confident that 95% of its materials will last for five million cycles at a
certain stress level. Thus, it is important for manufacturers to be able to estimate this

2
 Table 1: Models for material fatigue from the literature.
Model Formula
Wöhler/ASTM log (N ) = A − Bσ
M 
Weibull log (N + D) = A − B log σσM −σe

       u −σe   −2C 
Spindel and Haibach log Ne = A log σe − B log σe + B C1 log 1 + σσMe
N σM σM

   
+N1
Kohout and Vechet log σσMe = A log N N +N2
 
1 B

B  σu   γ(α)−1
σu −γ(α)
Epaarachchi and Clausen log (N ) = B log A fr σM − 1 σM [1 − ψ(R)] +1

quantity with great precision. While balanced plans are very appealing, research in the
statistical literature on optimal test planning has shown that these plans may not result in
the most precise information. Optimal test plans tend to be more unbalanced yet yield more
precise estimates of a selected quantity than could be achieved under a balanced design. As
such, they may be better suited for the task of gathering information on lifetime quantiles
and consequently on polymer composite fatigue behavior. Thus, it is our goal to introduce
statistically optimal test planning procedures to polymer composite fatigue research.
Much of the optimal design literature on accelerated testing consider a single use stress
level as part of the design criterion. However, it is more likely that a material will experience
several varying levels of stress over its lifetime. For example, the polymer composite materials
in motor vehicles are subject to a variety of stresses during the lifetime of the vehicle, such
as changes in the terrain (potholes, train tracks, etc.) or weather conditions to name a
few. Thus, we propose an optimality criterion consisting of a weighted sum of asymptotic
variances over a distribution of design stress levels, which would be more relevant to fatigue
researchers. The weights are selected so as to correspond to the frequency with which each
stress level is expected to be encountered.
In addition to specifying more relevant criteria, we also seek to incorporate existing fatigue
knowledge through the specification of an appropriate model. Many of the optimal designs
proposed for fatigue testing are based on linear models, which is reflected in the industrial
standards (e.g., ASTM E739 2010). Instead, we incorporate subject-matter expertise and
propose using a physical model taken from the fatigue literature as the basis for our optimal
design. We also propose compromise designs that utilize more design points so as to further
increase robustness to any assumption violations. Finally, we provide tools for assessment
of test plans and guidelines for practitioners.

3
 Stress σ

σmax

cycle
σmin

Time

Figure 1: Example of constant-amplitude fatigue testing.

1.2 Related Literature

There is an abundance of literature on models for the lifetime of polymer composite materials.
Here, the focus will be on accelerated failure time (AFT) models for the stress-life (S-N) curve
under cyclic constant amplitude stress testing. Table 1 summarizes some of the important
models in the fatigue regression literature. Here, N is the lifetime in loading cycles, σM is
the maximum stress applied during testing, σe is the endurance limit below which lifetime is
theoretically infinite, and σu is the ultimate strength of the material, which represents the
maximum stress the material can withstand before instantaneous failure.
In the engineering literature, the earliest models (e.g., Wöhler 1870) were simple linear
regression models on either semi-log or log-log data. Industry standards, such as ASTM
E739 (2010), still present these models for fatigue data analysis. Weibull (1949) introduced a
model with an offset term D, an endurance limit σe , and an ultimate tensile strength σu based
on previous work. Spindel and Haibach (1981) incorporated the endurance limit through
a smooth splicing of two lines. Kohout and Ve?chet (2001) incorporated the transitions
from low-cycle (N1 ) or high-cycle (N2 ) lifetimes to intermediate lifetimes simultaneously.
Distributional models, which attempt to describe the conditional distribution of lifetime
given stress, are available in Castillo and Hadi (1995), Castillo and Fernández-Canteli (2009),
and Pascual and Meeker (1999).
All of the aforementioned models are essentially based on linear models. However, as

4
Castillo et al. (1985) noted, fatigue data is nonlinear by nature. Epaarachchi and Clausen
(2003) introduced a physically motivated nonlinear model, which was derived from assump-
tions on the accumulation of fatigue damage in polymer composite materials. It also considers
the effect of variables involved in the testing procedure based on knowledge acquired from
experimental data. This, combined with its parsimonious nature (only parameters A and
B are unknown), provided us with sufficient justification to consider it as the basis for our
optimal designs. Further details on its parameterization are discussed in Section 2.
Regarding test planning, the industry standards for creating balanced test plans are
given in ASTM D3479 (2012) and ASTM E122 (2009). With regards to optimal accelerated
test planning, Nelson and Kielpinski (1976) derived plans based on normal and lognormal
distributions with the goal of minimizing the variance of the maximum likelihood (ML)
estimator of the mean/median lifetime. Nelson and Meeker (1978) considered the Weibull
and extreme value distributions and the goal of minimizing the variance of the ML estimator
of a specified quantile. Meeker (1984) compared plans based on Weibull and lognormal
distributions and concluded that compromise plans may be more suitable as they are more
robust to departures of model assumptions. Meeter and Meeker (1994) derive plans in which
the scale parameter is non-constant and Pascual (2006) discusses plans that are robust to
model misspecifications. Pascual (2003) derives several plans based on his random fatigue
limit model. Most of the existing literature considers test planning for a single use condition.
However, as noted in Pan and Yang (2014), multiple use conditions are more common in
practice. A useful reference for accelerated test planning is Nelson (1990).

1.3 Overview
This paper will present optimal designs and assessments for cyclic fatigue testing with the
goal of minimizing the weighted sum of asymptotic variances of an estimated lifetime per-
centile at selected design stress levels. Section 2 discusses details on the model used to build
the test plans. The construction of the optimality criterion is also presented. Section 3
presents the settings and notation for derivation of the test plans along with a description
and example of each type of test plan. In Section 4 we discuss the effects of the design
parameters and modeling assumptions on the test plans. A simulation study is performed
using selected test planning values to compare the asymptotic variance with the actual vari-
ance observed under the simulated test plans. Finally, Section 5 presents conclusions and
recommendations for fatigue test planning.

5
2 The Statistical Model
2.1 Model
For our test planning purposes, the cycles to failure (T ) in the fatigue data are assumed to
come from a log-location-scale family of distributions with the probability density function
(pdf) and cumulative density function (cdf) given as
   
1 log (t) − µ log (t) − µ
f (t) = φ , and F (t) = Φ ,
νt ν ν
respectively. Here, φ(·) and Φ(·) are the standard forms of the pdf and cdf, respectively.
Common examples of the log-location-scale family of distributions include the lognormal
and Weibull distribution, which are very popular in fatigue modeling. The scale parameter
ν is assumed to be constant while the location parameter µ = µ(σM ) is specified to be a
function of the stress. As mentioned in Section 1.2, we specify the relationship according to
the Epaarachchi and Clausen (2003) model as
(    γ(α)−1 )
1 B σu σu
µ(σM ; A, B) = log frB −1 [1 − ψ(R)] −γ(α)
+1 . (1)
B A σM σM

The unknown parameters are θ = (A, B, ν)′ , where A represents environmental effects on
the material fatigue and B is a material-specific parameter which represents effects from
the material itself. The remaining parameters are known from the testing procedure. The
ultimate strength σu is the same as described previously and fr is the frequency of the
cyclic testing. The parameter ψ is a function of the ratio R = σm /σM , where ψ(R) = R if
−∞ < R < 1 and ψ(R) = 1/R if 1 < R < ∞, and γ(α) = 1.6 −ψ| sin (α)| is a function of the
smallest angle α between the testing direction and the fiber direction. These relationships
are based on previous experimental results.
There are several reasons for using this model as the basis for our optimal test plans:

• The model is physically motivated as opposed to a curve-fitting technique or an exten-

sion thereof. It is derived on the assumption of a power law relationship for damage
within the material structure.

• It incorporates current knowledge regarding the effects of the testing parameters on

the relationship between stress and lifetime in the material. None of the other models
discussed in Section 1.2 even considered these effects.

• It is parsimonious in terms of the number of unknown parameters to be considered.

Linear models have at least two parameters to be estimated and the same goes for this

6
 model. It is also a flexible model because it can fit many kinds of polymer materials
well.

• Each of the unknown parameters has a physical meaning. In a linear or linearized

model, the slope and intercept may not have a physical meaning or may yield values
that are physically impossible.

One primary weakness of this model in comparison to a linear model is that it is highly
nonlinear in its structure, resulting in a more complex estimation procedure. However,
current computing power and technology can easily be used to overcome this complexity.

2.2 Data and Maximum Likelihood Estimation

Let the data from a fatigue test be denoted by {Tij , dij }, j = 1, . . . , ki , where Tij is the
lifetime of the jth sample at stress level i and dij is a censoring indicator. In particular,
dij = 1 if Tij ≤ TM and dij = 0 if Tij > TM , where TM is a pre-specified maximum test
time. It is assumed for now that TM is the same for each stress level. The total time under
P
test can be computed as Ttot. = si=1 ki TM . The likelihood function based on this data and
model (1) is
Y  1  log(Tij ) − µi (A, B) dij  
log(TM ) − µi (A, B)
1−dij
L(θ|DATA) = φ 1−Φ ,
i,j
νTij ν ν

where µi (A, B) = µ(σM i ; A, B) is the Epaarachchi and Clausen (2003) model as defined
in (1). The log-likelihood function is
X
l(θ|DATA) = log[L(θ|DATA)] = lij (θ), (2)
ij

where lij (θ) is the log-likelihood contribution from observation j at stress level i and θ =
(A, B, ν)′ is the vector of model parameters. In particular,

lij (θ) = dij {− log(ν) − log(Tij ) + log[φ(zij )]} + (1 − dij ) log[1 − Φ(zM i )],

where zij = [log(Tij ) − µi (A, B)]/ν and zM i = [log(TM ) − µi (A, B)]/ν. The maximum
likelihood (ML) estimates are those values of θ that maximize (2). Our main objective is to
estimate a specific
h quantile
i of the lifetime distribution at the use condition. Thus, the ML
estimator log Tbp (quse,i ) is then derived as
h i
log Tbp (quse,i ) = µuse,i (A,
b B)
b + zp νb, (3)

7
 600
500
400
Max Stress

300
200

Wohler/ASTM S and H
K and V E and C
100

Weibull

0 5 10 15

Cycles to Failure (Log)

Figure 2: Examples of fatigue models.

bB
where A, b and νb are the ML estimators of A, B and ν, respectively, and zp = Φ−1 (p).
As an illustration, we apply the ML estimation procedure to a motivating dataset from
industry, which we will refer to as Material B. Due to the sensitive nature of the experiment
and the resulting dataset, the data used here have been rescaled and relabeled. The ML
b = 0.00967, B
estimate for the Material B data are A b = 0.391, and νb = 0.478. Figure 2 shows
the estimated S-N curve based on model (1). As a comparison, we also visualize the fitted
curves from the other models listed in Table 1, revealing the inadequacy of a linear model
for explaining the relationship between stress and lifetime for polymer composites.

2.3 Computation of Optimality Criterion

The large-sample approximate variance of the estimator in (3) is derived from the Fisher
information matrix  2 
∂ l(θ)
Iθ = E − .
∂θ∂θ ′
The details of the calculation of Iθ are given in Appendix A. Specifically, under the standard
regularity conditions, the large-sample approximate variance-covariance matrix of the ML
estimators of A, B, and ν is Σθ = Iθ−1 . In industrial reliability, the p quantile of the product
lifetime distribution at a specified use level is commonly used for marketing purposes or as
part of industrial standards to ensure products are performing to proper specifications. As

8
such, C-optimality is a commonly used criterion because n it hminimizesiothe variance of the p
quantile. In particular, the optimality criterion is AVar log Tbp (quse,i ) , which is computed
as
n h io  ∂µ   ′
b use,i ∂µuse,i ∂µuse,i ∂µuse,i
AVar log Tp (quse,i ) = , , zp Σθ , , zp , (4)
∂A ∂B ∂A ∂B
where µuse,i = µ(σuse,i ; A, B). The exact variances and covariances of the ML estimates of
the parameters or functions of parameters from nonlinear models can be difficult to obtain
in closed form. Thus the asymptotic variance is often used. The design of optimal test plans
involves minimizing the asymptotic variance directly or a function of this variance.

3 Test Plan Procedures

3.1 Test Setting and Traditional Test Plan
Consider a fatigue test with s stress levels and ki samples tested at stress levels i = 1, . . . , s.
P
The total number of test units is k = si=1 ki . Let σM i be the maximum stress for level
i. We define the design points for the test plan as qi = σM i /σu ; the ratio of maximum
stress to the ultimate tensile strength. Note that 0 < qi ≤ 1. In practice, a planning range
of qi is specified prior to experimentation. We denote the planning range by qi ∈ [qL , qU ],
where qL and qU are the lower and upper bounds, respectively. An example of a planning
range could be qi ∈ [0.35, 0.75]. We denote the proportion of samples allocated to level i
P
by πi = ki /k. Note that 0 ≤ πi ≤ 1 and si=1 πi = 1. Here, we enforce πi to be a valid
proportion of the total sample size. That is, πi = r1 /r2 for some positive integers r1 and
r2 . This is done to ensure that the resulting test plans yield sensible proportions of the
total sample making for easier implementation. Finally, our test plan consists of the vector
η = (q1 , · · · , qs , π1 , · · · , πs )′ containing the design points where samples are to be allocated
and the proportion of the sample allocated to each point. These are the parameters that
can be adjusted to create the desired test plan.
The test plan we will use to represent the current methodology, which we will denote as
the “traditional” test plan, is represented by η trad = (0.35, 0.50, 0.75, 1/3, 1/3, 1/3)′. This
plan is representative of what is commonly used in fatigue testing and will serve as the basis
for comparison.

3.2 Optimal Test Plan

We will denote the “optimal” test plan as the vector η opt = (q1∗ , q2∗ , . . . , qs∗ , π1∗ , π2∗ , . . . , πs∗ )′ ,
where qi∗ and πi∗ are the design points and sample allocation values that optimize a desired

9
criterion. In practice, the product use level typically consists of a range of values rather
than one specific value. Here, we extend the criterion in (4) to minimizing the weighted sum
of the large-sample approximate variance of an estimator of the p quantile of the lifetime
distribution at a vector of specified use levels. Specifically, let quse = (quse,1 , quse,2 , . . . , quse,n )′
denote a vector of use levels and w = (w1 , w2 , . . . , wn )′ denote a vector of weights, where
Pn
i=1 wi = 1. The vectors quse and w are chosen to best represent the stresses and associated
frequencies the material under consideration would experience in the field. Thus, our optimal
design can be represented as
n
X n h io
η optim = arg min wi AVar log Tbp (quse,i ) (5)
η
i=1

where Tbp is the estimator of the p quantile of the lifetime distribution. This criterion is
subject to the following constraints:

(C1) The optimal design points must fall within the pre-specified planning range (qi∗ ∈
[qL , qU ]).

(C2) The optimal sample allocations must be valid reasonable proportions (πi∗ ∈ [0, 1], kπi∗ =
Xs
ki and πi∗ = 1).
i=1

(C3) The expected total time of the fatigue test must not exceed a pre-specified total test
s
X
time Ttot. (k πi∗ E[Ti ] ≤ Ttot. ).
i=1

The expected lifetime E[Ti ] at stress level i is computed according to the assumed distri-
bution. For example, in the case of the lognormal distribution E[Ti ] = exp (µi + ν 2 /2) and
in the case of the Weibull distribution E[Ti ] = exp (µi )Γ(1 + ν), where µi is the location
parameter for stress level i.

3.3 Compromise Test Plans

While optimal test plans are attractive due to their ability to extract the greatest amount of
information from data, their reliance on an assumed model structure is a weakness. Optimal
test plans are economical in terms of the number of design points required to implement
them; however, they often do not leave much room for experimenters to determine if their
modeling assumptions were correct. In order to alleviate this risk, many researchers pro-
pose a “compromise” plan in which additional design points and/or more allocation of the
sample to these points are included. These additions make the test plan more robust to

10
possible misspecification of the model. In a similar manner, we consider a compromise plan
η comp = (q1∗∗ , q2∗∗ , . . . , qs∗∗ , π1∗∗ , π2∗∗ , . . . , πs∗∗ )′ in which additional constraints are imposed on
the optimality criterion given in (5). Specifically, the optimality criterion is now subject to
the following constraints:

(C’1) The optimal design points must fall within the pre-specified planning range (qi∗∗ ∈
[qL , qU ]).

(C’2) The optimal sample allocations must be valid reasonable proportions and allocate no
less than a specified minimum πmin
∗∗
to each design point. (πi∗∗ ∈ [πmin
∗∗
, 1], kπi∗∗ = ki and
X s
πi∗∗ = 1).
i=1

(C’3) The expected total time of the fatigue test must not exceed a pre-specified total test
s
X
time Ttot. (k πi∗∗ E[Ti ] ≤ Ttot. ).
i=1

(C’4) The number of design points must be no less than a pre-specified number s∗∗ (s ≥ s∗∗ ).

The compromise plans we consider here have s∗∗ = 3 for constraint (C’4) as the inclusion
of more design points would result in test plans that are larger than the traditional plan
rendering them less efficient and so less desirable.

3.4 Optimization and Results

Optimization over the test plan parameters was performed using the optimization procedure
given in Appendix B. Under the constraints given in Section 3.2, optimization consistently
resulted in test plans with only two design points at each end of the planning range. This is
a common occurrence in ALT plans as the optimal number of design points is usually equal
to the number of parameters to be estimated in the model, of which there are only two here.
In light of these results, a minimum distance of 0.1 was imposed between any two design
points for optimization of the compromise test plans to ensure they did not revert to the
optimal two design points.
Because the quantities in Σθ usually depend on the true value of the model parameters,
planning values for the parameters are assumed so as to yield numeric values for the opti-
mality criterion. We suggest that planning values for any experimental design project should
be chosen based on expert knowledge of the fatigue process or on the results of experiments
performed on similar materials. For the purpose of exposition, our planning values corre-
spond to the ML estimates on Material B. Examples of each test plan can be seen in Table 2
in Section 4.

11
4 Assessment of Test Plans
The test plans will be assessed in two areas: the effects of the design parameters (total sam-
ple size, total time on test and use stress level distribution) and the effects of the modeling
assumptions (planning values and lifetime distribution). These assessments will allow us to
investigate the behavior of our proposed test plans under a variety of possible circumstances
and make appropriate suggestions accordingly. Unless otherwise specified, the default set-
tings for the design parameters are p = 0.05, k = 12, and TM = 1.0 × 107 cycles. The
planning values are based on those estimated from the motivating dataset and a lognormal
distribution is used as the default lifetime distribution.

4.1 Design Parameters

The range of sample sizes consists of k ∈ {9, 12, 15, 18, 21, 24, 27, 30}. As each of these sample
sizes is a multiple of three, this makes for a more direct comparison with the traditional test
plans. The effect of the total time on test is expressed through the individual censoring time
TM . The range of censoring times considered is TM ∈ {(0.5, 0.75, 1.0, 1.5, 2.0, 3.0, 5.0) × 106 }
cycles and was chosen to ensure fair comparison among the test plan types. More restrictive
censoring times would lead to poor test plans and less restrictive censoring times would have
no effect. The range of use stress levels consisted of 20 stress levels ranging from 5% to 25%
of the ultimate strength of the material. Four distributions representing different possible
weighting schemes or use profiles were selected and are presented in Figure 3. The effect of
the design parameters will be assessed under each use profile.

4.1.1 Sample Size

The configuration of each type of test plan under varying sample sizes and use profiles is
presented in Table 2. Figure 4 summarizes the effect of the sample size on each type of
test plan under each use distribution. As expected, all of the test plans yield more precise
estimates as the sample size increases. However, it is clear that the optimal and compromise
test plans yield more precise estimators across the board. Furthermore, they are both much
more efficient than the traditional test plan. For example, in order to obtain a more precise
estimator than that resulting from an optimal test plan with k = 9 samples under the Skewed
Right profile, the traditional test plan would require k = 15 samples, a roughly 67% increase
in the required number of samples.
When looking at the test plans in Table 2, one can see a marked difference in the al-
location of samples between the traditional and alternative plans. For both the optimal

12
 (a) Profile 1: Skewed right (b) Profile 2: Skewed left

(c) Profile 3: Unimodal symmetric (d) Profile 4: Bimodal asymmetric

Figure 3: Distribution profiles of use stress levels for the test plan assessment. Stress levels
increase towards the right.

13
and compromise test plans, the majority of the samples are allocated to the lower end of
the planning range. For the optimal test plan, this allocation is roughly 2:1 while for the
compromise plan it is roughly 5:1:3. This is common in optimal and compromise ALT plans.

14
 Table 2: Test plans by sample size k and use profile“P”.
Traditional Plan Optimum Plan Compromise Plan
P k Stress Level Allocation Stress Level Allocation Stress Level Allocation
AVar AVar AVar
q1 q2 q3 k1 k2 k3 q1 q2 k1 k2 q1 q2 q3 k1 k2 k3
9 0.45 0.35 0.50 0.75 3 3 3 0.31 0.35 0.75 6 3 0.33 0.35 0.45 0.75 5 1 3
12 0.34 0.35 0.50 0.75 4 4 4 0.23 0.35 0.75 8 4 0.26 0.35 0.65 0.75 7 2 3
15 0.27 0.35 0.50 0.75 5 5 5 0.19 0.35 0.75 10 5 0.20 0.35 0.65 0.75 9 2 4
18 0.22 0.35 0.50 0.75 6 6 6 0.15 0.35 0.75 12 6 0.17 0.35 0.65 0.75 11 2 5
1
21 0.19 0.35 0.50 0.75 7 7 7 0.13 0.35 0.75 14 7 0.15 0.35 0.65 0.75 13 3 5
24 0.17 0.35 0.50 0.75 8 8 8 0.12 0.35 0.75 15 9 0.13 0.35 0.65 0.75 14 3 7
27 0.15 0.35 0.50 0.75 9 9 9 0.10 0.35 0.75 17 10 0.11 0.35 0.65 0.75 16 3 8
30 0.13 0.35 0.50 0.75 10 10 10 0.09 0.35 0.75 19 11 0.10 0.35 0.65 0.75 18 3 9
9 0.23 0.35 0.50 0.75 3 3 3 0.16 0.35 0.75 6 3 0.17 0.35 0.65 0.75 6 1 2
12 0.17 0.35 0.50 0.75 4 4 4 0.12 0.35 0.75 9 3 0.13 0.35 0.65 0.75 8 2 2
15 0.14 0.35 0.50 0.75 5 5 5 0.09 0.35 0.75 11 4 0.10 0.35 0.65 0.75 10 2 3
18 0.12 0.35 0.50 0.75 6 6 6 0.08 0.35 0.75 13 5 0.08 0.35 0.65 0.75 12 2 4
2
21 0.10 0.35 0.50 0.75 7 7 7 0.07 0.35 0.75 15 6 0.07 0.35 0.65 0.75 14 3 4
15

24 0.09 0.35 0.50 0.75 8 8 8 0.06 0.35 0.75 17 7 0.06 0.35 0.65 0.75 16 3 5
27 0.08 0.35 0.50 0.75 9 9 9 0.05 0.35 0.75 19 8 0.06 0.35 0.65 0.75 18 3 6
30 0.07 0.35 0.50 0.75 10 10 10 0.05 0.35 0.75 22 8 0.05 0.35 0.65 0.75 20 3 7
9 0.32 0.35 0.50 0.75 3 3 3 0.22 0.35 0.75 6 3 0.24 0.35 0.65 0.75 6 1 2
12 0.24 0.35 0.50 0.75 4 4 4 0.16 0.35 0.75 8 4 0.18 0.35 0.65 0.75 7 2 3
15 0.19 0.35 0.50 0.75 5 5 5 0.13 0.35 0.75 10 5 0.14 0.35 0.65 0.75 9 2 4
18 0.16 0.35 0.50 0.75 6 6 6 0.11 0.35 0.75 12 6 0.12 0.35 0.65 0.75 11 2 5
3
21 0.14 0.35 0.50 0.75 7 7 7 0.09 0.35 0.75 14 7 0.10 0.35 0.65 0.75 13 3 5
24 0.12 0.35 0.50 0.75 8 8 8 0.08 0.35 0.75 16 8 0.09 0.35 0.65 0.75 15 3 6
27 0.11 0.35 0.50 0.75 9 9 9 0.07 0.35 0.75 18 9 0.08 0.35 0.65 0.75 17 3 7
30 0.10 0.35 0.50 0.75 10 10 10 0.07 0.35 0.75 20 10 0.07 0.35 0.65 0.75 19 3 8
9 0.34 0.35 0.50 0.75 3 3 3 0.24 0.35 0.75 6 3 0.26 0.35 0.65 0.75 6 1 2
12 0.26 0.35 0.50 0.75 4 4 4 0.18 0.35 0.75 8 4 0.20 0.35 0.65 0.75 7 2 3
15 0.21 0.35 0.50 0.75 5 5 5 0.14 0.35 0.75 10 5 0.16 0.35 0.65 0.75 9 2 4
18 0.17 0.35 0.50 0.75 6 6 6 0.12 0.35 0.75 12 6 0.13 0.35 0.65 0.75 11 2 5
4
21 0.15 0.35 0.50 0.75 7 7 7 0.10 0.35 0.75 14 7 0.11 0.35 0.65 0.75 13 3 5
24 0.13 0.35 0.50 0.75 8 8 8 0.09 0.35 0.75 16 8 0.10 0.35 0.65 0.75 15 3 6
27 0.11 0.35 0.50 0.75 9 9 9 0.08 0.35 0.75 18 9 0.09 0.35 0.65 0.75 17 3 7
30 0.10 0.35 0.50 0.75 10 10 10 0.07 0.35 0.75 20 10 0.08 0.35 0.65 0.75 19 3 8
 0.8

0.8
Traditional plan Traditional plan
Optimal plan Optimal plan
Comprimise plan Comprimise plan
0.6

0.6
Avar

Avar
0.4

0.4
0.2

0.2
0.0

0.0
10 15 20 25 30 10 15 20 25 30

Sample Size Sample Size

(a) Profile 1: Skewed Right (b) Profile 2: Skewed Left

0.8

0.8

Traditional plan Traditional plan

Optimal plan Optimal plan
Comprimise plan Comprimise plan
0.6

0.6
Avar

Avar
0.4

0.4
0.2

0.2
0.0

0.0

10 15 20 25 30 10 15 20 25 30

Sample Size Sample Size

(c) Profile 3: Unimodal Symmetric (d) Profile 4: Bimodal Asymmetric

Figure 4: Sample size effect for each plan under each use profile. The lines connecting the
observation points only serve as a visual reference and have no physical meaning.

Because the data collected in the planning range must then be used to extrapolate to the
range of interest, optimal test plans tend to assign more samples to the lower end of the
planning range as these design points are closest to the range of interest and so will yield
more information regarding the lifetime behavior there. One other notable result is the
relative similarity of the asymptotic variance between the optimal and compromise plans.
The difference is likely to be inconsequential in practice for even small samples and only
continues to decrease as the sample size increases. This is certainly comforting as compromise
test plans tend to be more suitable in practice than strictly optimal test plans.

16
 With regards to the effect of the use profile, for the Skewed Right profile, which places
most of the weight on use stress levels further away from the lower bound of the planning
range, the test plans yield less precise estimators. This is to be expected as the farther away
the use levels are from the planning range, the more uncertainty there is in the extrapolation.
Furthermore, when the majority of the weight is on use stress levels closer to the planning
range, as is the case for the Skewed Left profile, more of the sample tends to be allocated
to the lower stress levels in the planning range. Once again, because these regions of the
planning range are closer to the use levels (in these situations, much closer), it is intuitive to
assign even more samples to these regions as they are sufficiently close to yield the maximum
amount of information. For the remaining use profiles, there is no difference in the test plans
overall.

4.1.2 Censoring Time

The effect of the censoring time is summarized in Figure 5. The configuration of each type
of test plan under varying censoring times and use distributions is presented in Table 3. The
efficiency of the optimal and compromise plans remains resistant to severe time constraints.
This is most evident when TM = 500,000 cycles. The traditional plan is untenable as the
location of the design points and allocation of samples yields a test plan that will require a
longer average time of completion than is possible under this constraint. Both the optimal
and compromise plans are able to be run within this time constraint as they allow for more
flexible locations of the design points while still maintaining more samples at the lower design
points. The degree of change during transitions to less restrictive time constraints is fairly
consistent for all but the Skewed Left profile, the change in the test plans occurs in sooner
and remains unchanged over the remaining time constraints. As with the previous results,
more weight on use stresses closer to the lowest end of the planning range yield test plans
that attempt to place more of the data at the closest stress level as soon as is feasible.

4.2 Modeling Assumptions

For the assessment of sensitivity to the planning values, the range of values was chosen to
roughly cover a span of one standard deviation away from the ML estimates. With regards
to sensitivity to the lifetime distribution, in addition to the default lognormal distribution,
the Weibull distribution will be considered as an alternative due to its popularity in modeling
fatigue data. For the purposes of this portion of the assessment, the use profile was fixed at
Skewed Right from Figure 3, which is a common profile encountered in practice.

17
 0.8

0.8
Traditional plan Traditional plan
Optimal plan Optimal plan
Comprimise plan Comprimise plan
0.6

0.6
Avar

Avar
0.4

0.4
0.2

0.2
0.0

0.0

1e+06 2e+06 3e+06 4e+06 5e+06 1e+06 2e+06 3e+06 4e+06 5e+06

Censoring Time Censoring Time

(a) Profile 1: Skewed right (b) Profile 2: Skewed left

0.8

0.8

Traditional plan Traditional plan

Optimal plan Optimal plan
Comprimise plan Comprimise plan
0.6

0.6
Avar

Avar
0.4

0.4
0.2

0.2
0.0

0.0

1e+06 2e+06 3e+06 4e+06 5e+06 1e+06 2e+06 3e+06 4e+06 5e+06

Censoring Time Censoring Time

(c) Profile 3: Unimodal symmetric (d) Profile 4: Bimodal asymmetric

Figure 5: Censoring time effect for each plan under each use profile.

18
 Table 3: Test plans by censoring time TM (in millions of cycles) and use profile “P”. For TM = 500,000 cycles, the traditional
plan was untenable, requiring more than the allotted total time to implement.
Traditional Plan Optimum Plan Compromise Plan
P TM Stress Level Allocation Stress Level Allocation Stress Level Allocation
AVar AVar AVar
q1 q2 q3 k1 k2 k3 q1 q2 k1 k2 q1 q2 q3 k1 k2 k3
0.50 - 0.35 0.50 0.75 4 4 4 0.41 0.41 0.75 8 4 0.47 0.41 0.51 0.75 6 2 4
0.75 0.57 0.35 0.50 0.75 4 4 4 0.33 0.39 0.75 8 4 0.37 0.38 0.48 0.75 6 2 4
1.00 0.44 0.35 0.50 0.75 4 4 4 0.28 0.37 0.75 8 4 0.33 0.37 0.65 0.75 6 2 4
1 1.50 0.35 0.35 0.50 0.75 4 4 4 0.24 0.35 0.75 8 4 0.27 0.35 0.65 0.75 8 2 2
2.00 0.34 0.35 0.50 0.75 4 4 4 0.23 0.35 0.75 8 4 0.26 0.35 0.65 0.75 7 2 3
3.00 0.34 0.35 0.50 0.75 4 4 4 0.23 0.35 0.75 8 4 0.26 0.35 0.65 0.75 7 2 3
5.00 0.34 0.35 0.50 0.75 4 4 4 0.23 0.35 0.75 8 4 0.26 0.35 0.65 0.75 7 2 3
0.50 - 0.35 0.50 0.75 4 4 4 0.20 0.41 0.75 8 4 0.23 0.41 0.51 0.75 7 2 3
0.75 0.29 0.35 0.50 0.75 4 4 4 0.16 0.38 0.75 9 3 0.18 0.38 0.65 0.75 8 2 2
1.00 0.22 0.35 0.50 0.75 4 4 4 0.14 0.36 0.75 9 3 0.16 0.36 0.65 0.75 8 2 2
2 1.50 0.18 0.35 0.50 0.75 4 4 4 0.12 0.35 0.75 9 3 0.13 0.35 0.65 0.75 8 2 2
19

2.00 0.17 0.35 0.50 0.75 4 4 4 0.12 0.35 0.75 9 3 0.13 0.35 0.65 0.75 8 2 2
3.00 0.17 0.35 0.50 0.75 4 4 4 0.12 0.35 0.75 9 3 0.13 0.35 0.65 0.75 8 2 2
5.00 0.17 0.35 0.50 0.75 4 4 4 0.12 0.35 0.75 9 3 0.13 0.35 0.65 0.75 8 2 2
0.50 - 0.35 0.50 0.75 4 4 4 0.28 0.41 0.75 8 4 0.33 0.41 0.51 0.75 6 2 4
0.75 0.41 0.35 0.50 0.75 4 4 4 0.23 0.39 0.75 8 4 0.26 0.38 0.48 0.75 6 2 4
1.00 0.31 0.35 0.50 0.75 4 4 4 0.20 0.37 0.75 8 4 0.23 0.36 0.65 0.75 8 2 2
3 1.50 0.25 0.35 0.50 0.75 4 4 4 0.17 0.35 0.75 8 4 0.19 0.35 0.65 0.75 8 2 2
2.00 0.24 0.35 0.50 0.75 4 4 4 0.17 0.35 0.75 8 4 0.19 0.35 0.65 0.75 7 2 3
3.00 0.24 0.35 0.50 0.75 4 4 4 0.16 0.35 0.75 8 4 0.18 0.35 0.65 0.75 7 2 3
5.00 0.24 0.35 0.50 0.75 4 4 4 0.16 0.35 0.75 8 4 0.18 0.35 0.65 0.75 7 2 3
0.50 - 0.35 0.50 0.75 4 4 4 0.31 0.41 0.75 8 4 0.35 0.41 0.51 0.75 6 2 4
0.75 0.44 0.35 0.50 0.75 4 4 4 0.25 0.39 0.75 8 4 0.28 0.38 0.48 0.75 6 2 4
1.00 0.33 0.35 0.50 0.75 4 4 4 0.22 0.37 0.75 8 4 0.25 0.36 0.46 0.75 8 2 2
4 1.50 0.27 0.35 0.50 0.75 4 4 4 0.18 0.35 0.75 8 4 0.21 0.35 0.45 0.75 7 2 3
2.00 0.26 0.35 0.50 0.75 4 4 4 0.18 0.35 0.75 8 4 0.20 0.35 0.65 0.75 7 2 3
3.00 0.26 0.35 0.50 0.75 4 4 4 0.18 0.35 0.75 8 4 0.20 0.35 0.65 0.75 7 2 3
5.00 0.26 0.35 0.50 0.75 4 4 4 0.18 0.35 0.75 8 4 0.20 0.35 0.65 0.75 7 2 3
 350

350
A=0.0015,B=0.350 A=0.0556,B=0.350 A=0.0385,B=0.350 A=0.0385,B=0.423
A=0.0169,B=0.350 A=0.0750,B=0.350 A=0.0385,B=0.371 A=0.0385,B=0.450
A=0.0363,B=0.350 A=0.0385,B=0.397
300

300
250

250
200

200
Stress

Stress
150

150
100

100
50

50
0

0
0 5 10 15 20 25 30 0 5 10 15 20 25

Log−Cycles Log−Cycles

(a) Effect of environmental parameter A (b) Effect of material parameter B

Figure 6: Parameter effects on Epaarachchi and Clausen (2003) model.

4.2.1 Planning Values

Before discussing the assessment of the effect of the planning values, it is important to explain
the effect of the unknown model parameters A and B on the model in Epaarachchi and
Clausen (2003). In light of their descriptions given in Section 2, we will for the remainder of
this section refer to A as the “environmental” parameter and B as the “material” parameter.
The effects of each of these parameters on the model curve are summarized in Figure 6.
From this, we can clearly see that perturbing the environmental parameter causes a much
greater change in the overall curve than perturbations in the material parameter. This is in
agreement with what one would expect with small changes in the environmental factors of
the testing procedure as opposed to small changes in the material makeup. Note that the
effect is inversely related to the direction of increase for both parameters. This knowledge
will prove useful in evaluating the effects of changes in the planning values for the test plans.
Figure 7 gives contour plots of the weighted sum of asymptotic variances for each com-
bination of parameter values for the three test plan designs. Figures 8 and 9 give shade
plots for the design points and sample allocation for the optimal and compromise plans,
respectively. For both plans, the location of the upper design point was fixed at 0.75 for
all combinations of the planning values. Both the optimal and compromise test plans are
more robust to small changes in the planning values than the traditional plan. Only for very
small values of the environmental parameter do all of the test plans become more sensitive

20
 0.44

4
0.44

0.2
0.32
0.42

0.42
0.40

0.40
B

0.38
0.38

5
5
0.3

0.2
0.33

0.2
0.3
6

4
0.4

0.
35
2

0.
23

0.2
0.4
0.5

0.36
0.
0.36

34
1

8
2

0.3

0.2
8

0.2
0.34
0.4

0.3

5
6

9
0.3
7
6

0.005 0.010 0.015 0.020 0.005 0.010 0.015 0.020

A A

(a) Traditional Plan (b) Optimal Plan

0.44

7
0.2
0.25
0.42
0.40
B

0.38

8
0.2

0.2
7
0.3
0.36

0.2
6
0.2
0.4

0.005 0.010 0.015 0.020

(c) Compromise Plan

Figure 7: Planning value effect on optimality criterion for each design type. The shaded
regions correspond to values that yield untenable plans due to the time restrictions.

21
 0.41

0.44
0.40

0.42
0.39

0.40
B

0.38
0.38

0.37

0.36
0.36

0.35
0.005 0.010 0.015 0.020

Figure 8: Changes in lower design point qL for the optimal design. The sample allocation
at this point was always 8 samples. The upper design point qU was always fixed at 0.75 and
the sample allocation was always 4 samples.

to changes in the planning values. The reason for this is due to the inverse effect discussed
previously. As the planning value of the environmental parameter deviates to smaller values,
the lifetime increases much more dramatically resulting in more time required to evaluate
the test plans. Thus, the effect is equivalent to imposing a more severe time constraint on
the runs and both the optimal and compromise plans have the advantage of adjusting the
test plans to accommodate this change.

4.2.2 Distribution Assumption

The effect of lifetime distribution choice is summarized in Figure 10 and the specific layout of
each test plan is presented in Table 4. It was decided to compare lifetime distributions across
a variety of sample sizes both for ease of comparison and to detect any changes in sample
size effect. In both the asymptotic variance and the layout of the test plans, there is almost
no practical difference in using the Weibull lifetime distribution. However, it is important to
note that our criterion is based on asymptotic values. The accuracy of the criterion may be
dependent on the choice of lifetime distribution, especially when that lifetime distribution is
far from symmetric.
In order to assess the validity of the asymptotic variances under each choice of lifetime
distribution, we performed a simulation study to ascertain the accuracy of our criterion

22
 0.65
0.41
0.44

0.44
0.40
0.60
0.42

0.42
0.39
0.40

0.40
0.55
B

B
0.38
0.38

0.38
0.37

0.50

0.36
0.36

0.36
0.35 0.45
0.005 0.010 0.015 0.020 0.005 0.010 0.015 0.020

A A

(a) Lower design point qL (b) Middle design point qM

0.44

0.44

8 4
0.42

0.42
0.40

0.40

7 3
B

B
0.38

0.38

6 2
0.36

0.36

0.005 0.010 0.015 0.020 0.005 0.010 0.015 0.020

A A

(c) Sample allocation kL to qL (d) Sample allocation kU to qU

Figure 9: Changes in selected design parameters for the compromise design. The upper
design point qU was always fixed at 0.75 and the middle design point was always allocated
2 samples.

23
 0.6
Optim−Lognorm
Compr−Lognorm
Optim−SEV

0.5
Compr−SEV

0.4
Avar

0.3
0.2
0.1
0.0

10 15 20 25 30

Sample Size

Figure 10: Effect of distributions on optimum and compromise plans. Lines are for visual
representation only.

under each of the selected lifetime distributions and sample sizes. For each combination of
distribution and sample size, an optimal and compromise test plan as given in Table 4 was
implemented and the resulting data were used to estimate the unknown parameters A, B and
ν. The estimated quantile at each of the default use stress levels was then computed. This
simulation and estimation was performed 10,000 times and the final “exact” criterion was
computed using the empirical variance of the estimates with the weights selected according
to the Skewed Right use distribution. The results are summarized in Table 5. Note that the
asymptotic criterion severely underestimates the true value of the criterion when the Weibull
lifetime distribution is assumed. This is because the Weibull distribution is not symmetric
on the logarithmic scale and so it takes larger sample size for the asymptotic normality of the
ML estimator to perform well. In practice, however, a simulation can be used to determine
the number of samples required for a particular lifetime distribution and planning values.

5 Concluding Remarks
In this paper, we sought to derive optimal test plans for cyclic constant amplitude fatigue
testing of polymer composites based on the model proposed by Epaarachchi and Clausen
(2003). The optimality criterion consisted of the weighted sum of asymptotic variances of

24
Table 4: Test plans by sample size k and lifetime distribution “D”. The abbreviations are
“L” for lognormal and “W” for Weibull.

Optimum Plan Compromise Plan

D k Stress Level Allocation Stress Level Allocation
AVar AVar
q1 q2 k1 k2 q1 q2 q3 k1 k2 k3
9 0.31 0.35 0.75 6 3 0.33 0.35 0.45 0.75 5 1 3
12 0.23 0.35 0.75 8 4 0.26 0.35 0.65 0.75 7 2 3
15 0.19 0.35 0.75 10 5 0.20 0.35 0.65 0.75 9 2 4
18 0.15 0.35 0.75 12 6 0.17 0.35 0.65 0.75 11 2 5
L
21 0.13 0.35 0.75 14 7 0.15 0.35 0.65 0.75 13 3 5
24 0.12 0.35 0.75 15 9 0.13 0.35 0.65 0.75 14 3 7
27 0.10 0.35 0.75 17 10 0.11 0.35 0.65 0.75 16 3 8
30 0.09 0.35 0.75 19 11 0.10 0.35 0.65 0.75 18 3 9
9 0.29 0.35 0.75 6 3 0.32 0.35 0.45 0.75 5 1 3
12 0.22 0.35 0.75 8 4 0.25 0.35 0.65 0.75 7 2 3
15 0.18 0.35 0.75 10 5 0.19 0.35 0.65 0.75 9 2 4
18 0.15 0.35 0.75 12 6 0.16 0.35 0.65 0.75 11 2 5
W
21 0.13 0.35 0.75 14 7 0.14 0.35 0.65 0.75 13 3 5
24 0.11 0.35 0.75 15 9 0.12 0.35 0.65 0.75 14 3 7
27 0.10 0.35 0.75 17 10 0.11 0.35 0.65 0.75 16 3 8
30 0.09 0.35 0.75 19 11 0.09 0.35 0.65 0.75 18 3 9

Table 5: Asymptotic and simulated variances for the lognormal and Weibull lifetime distri-
butions for selected sample sizes.

Lognormal Weibull
k Optimum Compromise Optimum Compromise
True Asymp True Asymp True Asymp True Asymp
9 0.31 0.31 0.33 0.33 0.55 0.29 0.58 0.32
12 0.24 0.23 0.26 0.26 0.42 0.22 0.41 0.25
15 0.19 0.19 0.20 0.20 0.33 0.18 0.32 0.19
18 0.16 0.15 0.17 0.17 0.27 0.15 0.27 0.16
21 0.13 0.13 0.15 0.15 0.24 0.13 0.23 0.14
24 0.11 0.12 0.12 0.13 0.21 0.11 0.21 0.12
27 0.10 0.10 0.11 0.11 0.19 0.10 0.18 0.11
30 0.09 0.09 0.10 0.10 0.16 0.09 0.16 0.09

25
a quantile estimator for a vector of desired use stress levels. The model and criterion were
both chosen to better incorporate subject-matter knowledge into the test planning procedure.
Optimal and compromise plans were derived and assessed in terms of the effects of the design
parameters and modeling assumptions. The results of this assessment lead us to make the
following observations and suggestions:

• Assigning more samples to the design point closest to the use levels yields more precise
estimates of quantiles at those use levels. Furthermore, if the use levels closest to the
planning range are more likely to be encountered, than even more of the sample should
be allocated to the lower end of the planning range.

• Both the optimal and compromise plans are robust to slight miscalculations in the
planning values for all but the smallest values of the environmental parameter A.

• For symmetric distributions, the optimal and compromise designs are more effective
and efficient than the traditional plans for achieving the goal of quantile estimation.
For asymmetric distributions, although the asymptotic variance is underestimated, a
simulation can be used to determine the number of samples required for a particular
lifetime distribution and planning values.

Regarding the decision to use one type of test plan over another, we propose using the
compromise test plan over the optimal test plan in general unless knowledge of the underlying
model is very strong. In addition to its ability to detect more curvature than the optimal test
plan, the difference in the criterion value between the two plans is rather small, especially
for large samples. Even when the asymptotic criterion results in underestimation, the bias
is roughly the same for both test plans for the distributions considered. This, along with the
robustness and consistency in the test plans with respect to deviations from the planning
values, compels for us to advocate the use of compromise test plans for general test planning.
The plans presented here are specific to cyclic constant amplitude fatigue testing, which
is the most popular form of fatigue testing used in the industry. However, it is well known
that such testing procedures do not adequately represent the stress that materials experience
in the field. As such, procedures have been developed that allow the stress to vary over the
testing of the material. Epaarachchi (2006) presents an extension of their model for the
case of block testing, a simple extension of constant amplitude testing in which a sample is
subjected to blocks of cyclic testing with the amplitude varying among blocks. The number
of blocks and the amplitude for each block can be chosen to suit the experimenter’s need. A
logical next step would be to derive optimal test plans for fatigue testing under this scheme.

26
Future work could then include consideration of spectral testing, which is a more complex
extension of block testing.
We can also consider a Bayesian framework in modeling and test planning (e.g., Li and
Meeker 2014, Li, Meeker, and Thompson 2014, and Hong et al. 2015), which can allow for
computing the exact variance estimates and the incorporation of uncertainty in the choice of
planning values. Computational challenges in variance calculation and optimization have to
be addressed in the use of the Bayesian framework. In some test settings, random effects may
be involved (e.g., Li and Doganaksoy 2014, and Kensler, Freeman, and Vining 2015) and so
it would also be interesting to consider random effects in future research. In addition, newly
developed technology can allow for constant monitoring of fatigue development in critical
components. Incorporating this new technology into test planning would certainly make for
an interesting research topic.

Acknowledgments
We would like to thank the editor, an associate editor, and two referees, for their valuable
comments that lead to improvement of this paper. The Advanced Research Computing at
Virginia Tech is acknowledged for providing computational resources. The work by Hong
and King was supported by the National Science Foundation under Grant CMMI-1068933
to Virginia Tech and the 2011 DuPont Young Professor Grant.

A Formulae for the Fisher Information Matrix

This appendix provides the derivatives of the log likelihood given in Section 2.2, and the
details of the approach that we used to calculate needed expectations, and the large-sample
approximate variance-covariance matrix of the ML estimators of θ = (A, B, ν)′ .
Let c = (σu /σ − 1) (σu /σ)γ(α)−1 [1 − ψ(R)]−γ(α) . Then
  
1 B B log(D) − log(A)
µ(A, B) = log fr c + 1 = ,
B A B

27
where D = A + BfrB c. Let
∂µ(A, B) 1 1
µ1 = =− +
∂A AB BD
∂µ(A, B) log(A) log(D) E
µ2 = = 2
− 2
+
∂B B B BD
2
∂ µ(A, B) 1 1
µ11 = 2
= 2 −
∂A A B BD 2
2
∂ µ(A, B) 1 1 E
µ12 = = 2
− 2 −
∂A∂B AB B D BD 2
2
∂ µ(A, B) 2 log(A) 2 log(D) 2E E2 F
µ22 = = − + − − + .
∂B 2 B3 B3 B 2 D BD 2 BD
Here, E = BfrB log(f )c + frB c, and F = BfrB [log(fr )]2 c + 2fr B log(fr )c.
Considering a log-likelihood for a single observation, the first partial derivatives with
respect to the unknown parameters can be expressed as
∂l d φ′ (z) (1 − d) φ(zm )
=− µ1 + µ1
∂A ν φ(z) ν 1 − Φ(zm )
∂l d φ′ (z) (1 − d) φ(zm )
=− µ2 + µ2
∂B ν φ(z) ν 1 − Φ(zm )
 
∂l d φ′ (z)z (1 − d) φ(zm )zm
=− 1+ + .
∂ν ν φ(z) ν 1 − Φ(zm )
The second partial derivatives are
 ′  ′
∂2l d φ′ (z) 2 (1 − d) φ(zm ) 2 d φ′ (z) (1 − d) φ(zm )
2
= 2
µ 1 − 2
µ 1 − µ11 + µ11
∂A ν φ(z) ν 1 − Φ(zm ) ν φ(z) ν 1 − Φ(zm )
 ′  ′
∂2l d φ′ (z) 2 (1 − d) φ(zm ) 2 d φ′ (z) (1 − d) φ(zm )
2
= 2
µ 2 − 2
µ 2 − µ22 + µ22
∂B ν φ(z) ν 1 − Φ(zm ) ν φ(z) ν 1 − Φ(zm )
  ′ ′    ′ 
∂2l d 2φ′ (z)z φ (z) 2 (1 − d) 2φ(zm )zm φ(zm )
= 1+ + z − + z2
∂ν 2 ν 2 φ(z) φ(z) ν2 1 − Φ(zm ) 1 − Φ(zm ) m
 ′  ′
∂2l d φ′ (z) (1 − d) φ(zm )
= µ1 µ2 − µ1 µ2
∂A∂B ν 2 φ(z) ν2 1 − Φ(zm )
d φ′ (z) (1 − d) φ(zm )
− µ12 + µ12
ν φ(z) ν 1 − Φ(zm )
  ′ ′    ′ 
∂2l d φ′ (z) φ (z) (1 − d) φ(zm ) φ(zm )
= + z µ1 − + zm µ1
∂A∂ν ν 2 φ(z) φ(z) ν2 1 − Φ(zm ) 1 − Φ(zm )
  ′ ′    ′ 
∂2l d φ′ (z) φ (z) (1 − d) φ(zm ) φ(zm )
= + z µ2 − + zm µ2 .
∂B∂ν ν 2 φ(z) φ(z) ν2 1 − Φ(zm ) 1 − Φ(zm )
Here,  ′    ′  
φ′ (z) ∂ φ′ (z) φ(z) ∂ φ(z)
= , and = .
φ(z) ∂z φ(z) 1 − Φ(z) ∂z 1 − Φ(z)

28
Detailed formulae of Φ, φ, φ/(1−Φ), [φ/(1−Φ)]′ , φ′ /φ, and [φ′ /φ]′ for commonly-used location-
scale distributions (normal, smallest extreme value, largest extreme value, and logistic, corre-
sponding to the lognormal, Weibull, Fréchet, and loglogistic log-location-scale distributions)
can be found in Table I of Hong, Ma, and Meeker (2010). The scaled expectations of the
second derivatives are
  Z zm   ′ ′ 
2 ∂2l 2 φ (z) φ′ (z)
ν E − 2 =− µ1 φ(z) − νµ11 φ(z) dz
∂A −∞ φ(z) φ(z)
 ′
2 φ(zm )
+ µ1 [1 − Φ(zm )] − νµ11 φ(zm )
1 − Φ(zm )
  Z zm   ′ ′ 
2 ∂2l 2 φ (z) φ′ (z)
ν E − 2 =− µ2 φ(z) − νµ22 φ(z) dz
∂B −∞ φ(z) φ(z)
 ′
2 φ(zm )
+ µ2 [1 − Φ(zm )] − νµ22 φ(zm )
1 − Φ(zm )
  Z zm   ′ ′ 
2 ∂2l 2φ′ (z)z φ (z)
ν E − 2 =− 1+ + z 2 φ(z)dz
∂ν −∞ φ(z) φ(z)
  ′ 
2φ(zm )zm φ(zm )
+ + z 2 [1 − Φ(zm )]
1 − Φ(zm ) 1 − Φ(zm ) m

  Z zm   ′ ′ 
2 ∂2l φ (z) φ′ (z)
ν E − =− µ1 µ2 φ(z) − νµ12 φ(z) dz
∂A∂B −∞ φ(z) φ(z)
 ′
φ(zm )
+ µ1 µ2 [1 − Φ(zm )] − νµ12 φ(zm )
1 − Φ(zm )
  Z zm  ′  ′ ′ 
2 ∂2l φ (z) φ (z)
ν E − =− µ1 + z φ(z)dz
∂A∂ν −∞ φ(z) φ(z)
  ′ 
φ(zm ) φ(zm )
+ µ1 + zm [1 − Φ(zm )]
1 − Φ(zm ) 1 − Φ(zm )
  Z zm  ′  ′ ′ 
2 ∂2l φ (z) φ (z)
ν E − =− µ2 + z φ(z)dz
∂B∂ν −∞ φ(z) φ(z)
  ′ 
φ(zm ) φ(zm )
+ µ2 + zm [1 − Φ(zm )] .
1 − Φ(zm ) 1 − Φ(zm )

B Optimization Procedure
The following algorithm represents the procedure used during optimization:

1. Given k and s, determine all P possible sample allocation schemes π p = (π1 , . . . , πs )′ , p =

1, . . . , P . If πmin is specified, all sample allocations that include values of πi < πmin

29
 are excluded. This first step takes advantage of the constraint on ki = kπi being an
integer value.

2. For each π p :

(a) Divide the planning range into a grid of values separated by a specified grid size
“q−step”. Select all D possible combinations qpd = (q1 , . . . , qs )′ , d = 1, . . . , D of
design points using these grid points. If a minimum distance between any two
design points is specified, set q−step equal to this minimum distance. Evaluate
each plan η pd = ((qpd )′ , (π p )′ )′ and select the qpd
min that yields the minimum value
of the optimality criterion. This step was required due to the sensitivity of the
optimization to the initial values.
(b) Using qpd
min as the initial values, optimize over the design point space to find the
overall minimum value of the optimality criterion for the given π p . This was
performed using the free statistical software package R using the BFGS algorithm
option in the optim() function. To ensure stability in the optimization, the design
points were first transformed by the following function, which maps values from
the range [a, b] to the real line. Specifically, given a value x the transformed value
x
e is given by, x
e = − log [(b − x)/(x − a)].
To enforce the time constraint, a constant penalty term M = 10,000 was added
to the result if the expected total time exceeded the maximum total time.

3. Once a test plan η p is obtained for each π p , select η min that yields the minimum value
of the optimality criterion across all π p , giving the optimal/compromise plan.

References
ASTM D3479 (2012). Test method for tension-tension fatigue of polymer matrix composite
materials. Technical report, ASTM International.
ASTM E122 (2009). Practice for calculating sample size to estimate, with specified pre-
cision, the average for a characteristic of a lot or process. Technical report, ASTM
International.
ASTM E739 (2010). Practice for statistical analysis of linear or linearized stress-life (S-N)
and strain-life (σ-N) fatigue data. Technical report, ASTM International.
Castillo, E., A. F. Canteli, V. Esslinger, and B. Thurlimann (1985). Statistical model for
fatigue analysis of wires, strands and cables. In International Association for Bridge

30
 and Structural Engineering Proceedings P-82/85, Zurich, pp. 1–40.
Castillo, E. and A. Fernández-Canteli (2009). A unified statistical methodology for modeling
fatigue damage. Netherlands: Springer.
Castillo, E. and A. S. Hadi (1995). Modeling lifetime data with application to fatigue
models. Journal of the American Statistical Association 90 (431), 1041–1054.
Epaarachchi, J. A. (2006). A study on estimation of damage accumulation of glass fi-
bre reinforce plastic (GFRP) composites under a block loading situation. Composite
Structures 75 (1-4), 88–92.
Epaarachchi, J. A. and P. D. Clausen (2003). An empirical model for fatigue behavior
prediction of glass fibre-reinforced plastic composites for various stress ratios and test
frequencies. Composites Part A: Applied Science and Manufacturing 34 (4), 313–326.
Hong, Y., C. King, Y. Zhang, and W. Q. Meeker (2015). Bayesian life test planning for the
log-location scale family of distributions. Journal of Quality Technology 47, 336–350.
Hong, Y., H. Ma, and W. Q. Meeker (2010). A tool for evaluating time-varying-stress
accelerated life test plans with log-location-scale distributions. IEEE Transactions on
Reliability 59, 620–627.
Kensler, J. L., L. J. Freeman, and G. G. Vining (2015). Analysis of reliability experiments
with random blocks and subsampling. Journal of Quality Technology 47, 235–251.
Kohout, J. and S. Ve?chet (2001). A new function for fatigue curves characterization and
its multiple merits. International Journal of Fatigue 23 (2), 175–183.
Li, M. and N. Doganaksoy (2014). Batch variability in accelerated-degradation testing.
Journal of Quality Technology 46, 171–180.
Li, M. and W. Q. Meeker (2014). Application of Bayesian methods in reliability data
analyses. Journal of Quality Technology 46, 1–23.
Li, M., W. Q. Meeker, and R. B. Thompson (2014). Physical model-assisted probability
of detection of flaws in titanium forgings using ultrasonic nondestructive evaluation.
Technometrics 56, 78–91.
Meeker, W. Q. (1984). A comparison of accelerated life test plans for Weibull and lognor-
mal distributions and type I censoring. Technometrics 26 (2), 157–171.
Meeter, C. A. and W. Q. Meeker (1994). Optimum accelerated life tests with a nonconstant
scale parameter. Technometrics 36 (1), 71–83.
Nelson, W. (1990). Accelerated Testing: Statistical Models, Test Plans, and Data Analysis.
New York: Wiley-Interscience.

31
 Nelson, W. and T. J. Kielpinski (1976). Theory for optimum censored accelerated life tests
for normal and lognormal life distributions. Technometrics 18 (1), 105–114.
Nelson, W. and W. Q. Meeker (1978). Theory for optimum accelerated censored life tests
for weibull and extreme value distributions. Technometrics 20 (2), 171–177.
Pan, R. and T. Yang (2014). Design and evaluation of accelerated life testing plans with
dual objectives. Journal of Quality Technology 46, 114–126.
Pascual, F. G. (2003). Theory for optimal test plans for the random fatigue-limit model.
Technometrics 45 (2), 130–141.
Pascual, F. G. (2006). Accelerated life test plans robust to misspecification of the stress-life
relationship. Technometrics 48 (1), 11–25.
Pascual, F. G. and W. Q. Meeker (1999). Estimating fatigue curves with the random
fatigue-limit model. Technometrics 41 (4), 277–290.
Spindel, J. E. and E. Haibach (1981). Some considerations in the statistical determination
of the shape of S-N curves. In R. E. Little and J. C. Ekvall (Eds.), Statistical Analysis of
Fatigue Data, ASTM STP 744, pp. 89–113. American Society for Testing and Materials.
Weibull, W. (1949). A statistical report of fatigue failure in solids. In Transactions, Vol-
ume 27, Stockholm. Royal Institute of Technology of Sweden.
Wöhler, A. (1870). Über die Festigkeitsversuche mit Eisen und Stahl. Zeitschrift für
Bauwesen 20, 73–106.

32

View publication stats