You are on page 1of 7

ARTICLE IN PRESS

Reliability Engineering and System Safety 92 (2007) 286292


www.elsevier.com/locate/ress

Design of PH-based accelerated life testing plans


under multiple-stress-type
E.A. Elsayed, Hao Zhang
Department of Industrial and System Engineering, Rutgers University, 96 Felinghuysen Road, Piscataway, NJ 08854-8018, USA
Available online 9 June 2006

Abstract
Accelerated life testing (ALT) is used to obtain failure time data quickly under high stress levels in order to predict product life
performance under design stress conditions. Most of the previous work on designing ALT plans is focused on the application of a single
stress. However, as components or products become more reliable due to technological advances, it becomes more difcult to obtain
signicant amount of failure data within reasonable amount of time using single stress only. Multiple-stress-type ALTs have been
employed as a means of overcoming such difculties. In this paper, we design optimum multiple-stress-type ALT plans based on the
proportional hazards model. The optimum combinations of stresses and their levels are determined such that the variance of the
reliability estimate of the product over a specied period of time is minimized. The use of the model is illustrated using numerical
example, and sensitivity analysis shows that the resultant optimum ALT plan is robust to the deviation in model parameters.
r 2006 Elsevier Ltd. All rights reserved.
Keywords: Accelerated life testing (ALT); Reliability predication; ALT plans; Proportional hazards model; Sensitivity analysis

1. Introduction
Globalization of the world economy has intensied the
competition for decreasing both cost and time to market
while improving the quality and reliability of products.
Moreover, such competition has lead to new strategies for
product marketing such as unusually long or extended
warranties at the time of product sale. This has indeed
deemed the traditional reliability testing to be obsolete due
to the long time needed for testing and its inability to
provide reliability prediction at conditions different from
the test conditions. Accelerated life testing is an alternative
that overcomes the shortcomings of the traditional
reliability testing.
Accelerated life testing (ALT) is used to obtain failure
time data quickly under high stress levels in order to
predict product reliability performance under the design
stress. In order to predict reliability at design stresses, an
accurate reliability model is used to relate failure times
Corresponding author. Tel.: +1 732 445 3859; fax: +1 732 445 5467.

E-mail addresses: elsayed@rci.rutgers.edu (E.A. Elsayed),


haoz@eden.rutgers.edu (H. Zhang).
0951-8320/$ - see front matter r 2006 Elsevier Ltd. All rights reserved.
doi:10.1016/j.ress.2006.04.016

(or failure rates) at stress conditions to those at normal or


design stress conditions. The accuracy of the model is
important as it usually predicts reliability at tens or
hundreds of years. The most common models include: (1)
accelerated failure time models (AFT) where the failure
times are inversely proportional to the applied stresses, i.e.,
failure times of the products at higher stresses are shorter
than those tested at lower stresses [1,2]; (2) proportional
hazards (PH) models where the failure rates are proportional to the applied stresses, i.e., the failure rates at higher
stresses are higher than those at lower stresses [3,4];
(3) extended linear hazards regression models (ELHR)
where both AFT and PH models are special cases of ELHR
when some conditions are satised [5]; (4) proportional
mean residual life models where the mean residual lives are
inversely proportional to the stresses, i.e., the mean
residual life at higher stresses is shorter than that at lower
stresses [6]; and (5) proportional odds (PO) models where
the odds functions under different stress levels are
proportional to each other [79].
ALT plans are used to design the ALT in order to
increase the reliability prediction accuracy by determining
the optimum stress levels assigned to testing units, the

ARTICLE IN PRESS
E.A. Elsayed, H. Zhang / Reliability Engineering and System Safety 92 (2007) 286292

number of units at every stress level, and the duration of


the test. Most of the previous work on designing ALT plan
is focused on the application of a single stress [1015],
which includes multi-stress-level, constant stress, and stepstress ALTs. However, as components or products become
more reliable due to technological advances, it becomes
more difcult to obtain signicant amount of failure data
within a reasonable amount of time using single stress only,
even with multiple stress levels. Testing time and cost
cannot be reduced by just increasing stress levels for a
single stress accelerated life testing. This is due to the fact
that increasing stress to some level could induce different
failure modes such that the assumed statistical regression
model does not hold any more. Therefore multiple-stresstype ALTs have been employed as a means of overcoming
such difculties. For instance, Kobayashi et al. [16],
Minford [17], Mogilevsky and Shirn [18], and Munikoti
and Dhar [19] use two stresses to test capacitors, and Weis
et al. [20] employ two stresses to estimate the lifetime of
silicon photodetectors.
Unlike the case of the single stress ALT, little work has
been focused on designing multiple-stress-type accelerated
life testing plans. Escobar and Meeker [21] develop
statistically optimal and practical plans with two stresses
and no interaction between them. However, if prior
information does not support the nonexistence of interaction or if the so-called sliding level technique cannot be
employed to avoid the potential interaction, then the
analysis based on the main effects only could lead to
serious bias in estimation. Park and Yum [1] develop ALT
plans in which two stresses are employed with possible
interaction between them. The lifetime distribution of test
units is assumed to be exponential.
In this paper, we design two-stress ALT plans based on
the proportional hazards (PH) model. We do not make
assumptions about the failure time distribution of the
test units. The plans are optimized such that the variance of the reliability prediction at normal operating
condition is minimized. This work is based on Elsayed and
Zhang [8].
2. ALT plans based on PH model with two stresses
2.1. Assumptions
Fig. 1 shows the candidates of test stress levels for the
two-stress ALT plans. The stresses z1D and z2D are the
design stress levels corresponding to stresses z1 and z2 ,
respectively. Stress level zlm (l 1; 2 and m 0; 1) represents the type l stress at the m level.
We assume the following model for the accelerated life
test with two stresses.
1. Two stresses z1 and z2 are used in the test, let z z1 ; z2
be the vector of stress levels.
2. The proportional hazards model is employed to relate
the reliability performance under different stress levels,

287

(z1H, z2H)

z2

High
(z21)

Low
(z20)

(z1D, z2D)

p01

p11

p00

p10

Low
(z10)

High
(z11)

z1

Fig. 1. Candidate test combinations in two-stress ALT plans.

it is expressed as
lt; z l0 t expb1 z1 b2 z2 ,
where
b1 and b2 are unknown model parameters.
3. The baseline hazard function l0 t is quadratic
l0 t g0 g1 t g2 t2 ,
where
g0 , g1 and g2 are unknown parameters.
4. Four candidates of test stress level combinations are
considered, as shown in Fig. 1. The upper bounds of
stresses are pre-specied, whereas two levels for each
stress are to be optimally determined. The upper bounds
are dened as the highest possible stress levels beyond
which different failure modes will be induced to violate
the proportional hazard assumption.
5. The total number of test units n is given. The proportion
of units allocated to the ith level of the rst stress z1 and
jth level of the second stress z2 is denoted by pij
(i 0; 1; j 0; 1) and is optimally determined.
6. The lifetimes of test units are s-independent.
7. The test is terminated at a pre-specied censoring time t.
The proposed factorial arrangement of test units shown
in Fig. 1 is not statistically optimal; however, this
arrangement is motivated by the actual practice of
reliability engineers. Since any other arrangement with
the same number of testing points will result in more stress
levels for each stress type, it enables one to conduct the
entire test simultaneously by utilizing the available equipment in an efcient manner, which can lead to signicant
savings in completion time and cost. In addition, it allows
for testing of interactions after the data are collected.
The hazard function lt; z is obtained by substituting
l0 t into the PH model as
lt; z g0 g1 t g2 t2 expb1 z1 b2 z2 .

(1)

ARTICLE IN PRESS
E.A. Elsayed, H. Zhang / Reliability Engineering and System Safety 92 (2007) 286292

288

We obtain the corresponding cumulative hazard function Lt; z, reliability function Rt; z and density function
f t; z, respectively, as follows:
Z t
Lt; z
lu du
0
Z t
g0 g1 u g2 u2 expb1 z1 b2 z2 du

0


g t2 g t3 b1 z1 b2 z2
g0 t 1 2
,
2
e
2
3
Rt; z expLt; z
 


g1 t2 g2 t3 b1 z1 b2 z2
exp  g0 t

,
e
2
3

f t; z lt; zRt; z
g0 g1 t g2 t2 eb1 z1 b2 z2
 


g t2 g t3 b1 z1 b2 z2
 exp  g0 t 1 2
.
e
2
3

In order to obtain the values of the model parameters, we


utilize the maximum likelihood estimation (MLE) procedure as described below.
2.2. The log likelihood function
The log likelihood of an observation t (time to failure)
under two stresses z1 and z2 with Type I censoring is given
below. We dene the indicator function I Itpt in
terms of the censoring time t as
(
1 if tpt; failure observed before time t;
I Itpt
0 if t4t; censored at time t:
The log likelihood of a Type I censored observation at
stresses z1 and z2 is
ln Lt; z I lnlt; z  Lt; z
I lng0 g1 t g2 t2 Ib1 z1 Ib2 z2


g t2 g t3 b1 z1 b2 z2
 g0 t 1 2
.
e
2
3

Supposing that the ith observation ti corresponds to a


vector of stress levels zi and the corresponding log
likelihood is l i , then the sample log likelihood l for n
independent observations is
l l1 l2    ln.
2.3. Maximum likelihood estimation
For a single observation, the ve rst partial derivatives
with respect to the model parameters are
q ln Lt; z
I

 teb1 z1 b2 z2 ,
qg0
g0 g1 t g2 t 2

(6)

q ln Lt; z
It
t2 b1 z1 b2 z2
e


,
qg1
g0 g1 t g2 t 2 2

(7)

q ln Lt; z
It2
t3

 eb1 z1 b2 z2 ,
2
qg2
g0 g1 t g2 t
3

(8)


q ln Lt; z
g
g 
Iz1  z1 g0 t 1 t2 1 t3 eb1 z1 b2 z2 ,
qb1
2
3

(9)


q ln Lt; z
g
g 
Iz2  z2 g0 t 1 t2 1 t3 eb1 z1 b2 z2 .
qb2
2
3

(10)

Summing Eqs. (6)(10) over all test units, setting them


equal to zero, and solving the resultant equations
simultaneously will provide the maximum likelihood
estimates for the model parameters.
Suppose that we only consider the correlation among g0 ,
g1 and g2 , the second-partial derivatives are as shown
below:
q2 ln Lt; z
I

,
qg20
g0 g1 t g2 t2 2

(11)

q2 ln Lt; z
It2

,
2
qg1
g0 g1 t g2 t2 2

(12)

q2 ln Lt; z
It4

,
2
qg2
g0 g1 t g2 t2 2

(13)

q2 ln Lt; z
It

,
qg0 qg1
g0 g1 t g2 t2 2

(14)

q2 ln Lt; z
It2

,
qg0 qg2
g0 g1 t g2 t2 2

(15)

q2 ln Lt; z
It3

,
qg1 qg2
g0 g1 t g2 t2 2

(16)


q2 ln Lt; z
g
g 
z21 g0 t 1 t2 2 t3 eb1 z1 b2 z2 ,
2
2
3
qb1

(17)


q2 ln Lt; z
g
g 
z22 g0 t 1 t2 2 t3 eb1 z1 b2 z2 .
2
2
3
qb2

(18)

Eqs. (11)(18) are given in terms of the random


quantities I and z and the model parameters. The elements
of the Fisher information matrix for an observation are the
negative expectations of the above equations:
 2

q ln Lt; z
E 
qg2
Z t 0
1
eb1 z1 b2 z2

2
g

g
t

g
t
0
0
2

1
g1 2 g2 3  b1 z1 b2 z2 
 exp  g0 t t t e
dt,
19
2
3

ARTICLE IN PRESS
E.A. Elsayed, H. Zhang / Reliability Engineering and System Safety 92 (2007) 286292


q2 ln Lt; z
E 
qg21
Z t
t2

eb1 z1 b2 z2
2
g

g
t

g
t
0
0
2
 1

g
g 
 exp  g0 t 1 t2 2 t3 eb1 z1 b2 z2 dt,
2
3
 2

q ln Lt; z
E 
qg22
Z t
t4

eb1 z1 b2 z2
2
g

g
t

g
t
0
0
2
 1

g
g 
 exp  g0 t 1 t2 2 t3 eb1 z1 b2 z2 dt,
2
3
 2

q ln Lt; z
E 
qg g
Z t 0 1
t

eb1 z1 b2 z2
2
0 g0 g1 t g2 t
 

g
g 
 exp  g0 t 1 t2 2 t3 eb1 z1 b2 z2 dt,
2
3

q2 ln Lt; z
E 
qg0 g2
Z t
t2

eb1 z1 b2 z2
2
0 g0 g1 t g2 t
 

g
g 
 exp  g0 t 1 t2 2 t3 eb1 z1 b2 z2 dt,
2
3

20

21


q2 ln Lt; z
E 
qg1 g2
Z t
t3

eb1 z1 b2 z2 exp
2
0 g0 g1 t g2 t
 

g
g 
  g0 t 1 t2 2 t3 eb1 z1 b2 z2 dt,
2
3

Var^g0
6
6 Cov^g0 ; g^ 1
6
6
S 6 Cov^g0 ; g^ 2

6
6 Cov^g ; b^
0 1
4
Cov^g0 ; b^ 2

Cov^g0 ; g^ 1

Cov^g0 ; g^ 2

Var^g1

Cov^g1 ; g^ 2

Cov^g1 ; g^ 2
Var^g2
Cov^g1 ; b^ 1 Cov^g2 ; b^ 1
Cov^g ; b^ Cov^g ; b^
1

Cov^g0 ; b^ 1
Cov^g ; b^
1

Covb^ 1 ; b^ 2

23

24

Each ALT plan is characterized by the levels of all


stresses, the proportions of test units allocated to each
test combination, the total number of test units available,
and the censoring time. We assume that both the design
and the highest levels of each stress, the total number
of test units, and the censoring time are pre-specied.
The objective is to determine the low and high levels of
each stress and the proportion of test units allocated to
each test condition such that the asymptotic variance of the
MLE of the reliability estimate at the normal operating
conditions is minimized. This is a nonlinear optimization
problem which is formulated as described in the next
section.

2.4. Problem formulation


25

Objective function
Min

q ln Lt; z
qb21
Z 1
g
g 
z22
g0 t 1 t2 2 t3 eb1 z1 b2 z2 f t; z1 ; z2 dt.
2
3
0

Cov^g2 ; b^ 1
Varb^

3
Cov^g0 ; b^ 2
7
Cov^g1 ; b^ 2 7
7
7
Cov^g2 ; b^ 2 7 F 1 .
7
Covb^ 1 ; b^ 2 7
5
Varb^ 2

(27)

)
q2 ln Lt; z
E 
qb21
Z 1
g
g 
z21
g0 t 1 t2 2 t3 eb1 z1 b2 z2 f t; z1 ; z2 dt,
2
3
0

combination z10 , z20 , p01 the proportion of test units


allocated to test stress level combination z10 ; z21 , p10 the
proportion of test units allocated to test stress level
combination z11 ; z20 , p11 the proportion of test units
allocated to test stress level combination z11 ; z21 , F 00 the
Fishers information matrix for a unit at test combination
z10 ; z20 , F 01 the Fishers information matrix for a unit at
test combination z10 ; z21 , F 10 the Fishers information
matrix for a unit at test combination z11 ; z20 , and F 11 the
Fishers information matrix for a unit at test combination
z11 ; z21 .
The Fisher information matrix is a function of
g0 ; g1 ; g2 ; b1 ; b2 ; zlm l 1; 2; m 0; 1 and pij i 0; 1;
j 0; 1. The asymptotic variancecovariance matrix S of

the ML estimates g^ 0 ; g^ 1 ; g^ 2 ; b^ 1 and b^ 2 is the inverse of the
Fisher information matrix F
2

22

289

RT
0

Var^g0 g^ 1 t g^ 2 t2
^

eb1 z1D b2 z2D  dt

E 

s:t:
26

0opij o1; i 0; 1; j 0; 1;
P
pij 1;
i;j

The Fisher information matrix F for all units under test


is expressed as

S F 1 ;

F np00 F 00 np01 F 01 np10 F 10 np11 F 11 ,

z2D pz20 oz21 pz2H ;


npij Prtptjz1i ; z2j XMNF;

where n is the total number of test units placed on test, p00


the proportion of test units allocated to test stress level

z1D pz10 oz11 pz1H ;

i 0; 1; j 0; 1;

ARTICLE IN PRESS
E.A. Elsayed, H. Zhang / Reliability Engineering and System Safety 92 (2007) 286292

290

where T is the time period over which the reliability


estimate is interested, and MNF is the minimum number of
failures. The stresses z1H and z2H are the highest stress
levels corresponding to stresses z1 and z2 , respectively, that
can be applied without causing different failure modes
from those at normal conditions. The last constraint
ensures a desired minimum number of failures at any of
the test stress combinations.
^
^
The term Var^g0 g^ 1 t g^ 2 t2 eb1 z1D b2 z2D  is calculated
by Delta method as follows:

4. The optimization problem is formulated as objective


function
RT
Min f x 0 Var^g0 g^ 1 t g^ 2 t2
^

eb1 z1D b2 z2D  dt


s:t:
0opij o1;
P
pij 1;

i 0; 1; j 0; 1;

i;j

S F 1 ;


Var^g0 g^ 1 t g^ 2 t2 eb1 z1D b2 z2D 


^ g ; g^ ; g^ ; b^ ; b^ 
Varl^
0 1 2 1 2
"
# "
#0
ql^ ql^ ql^ ql^ ql^
ql^ ql^ ql^ ql^ ql^

;
;
;
;
;
;
;
;
.
S
q^g0 q^g1 q^g2 qb^ 1 qb^ 2  q^g0 q^g1 q^g2 qb^ 1 qb^ 2
The nonlinear optimization problem can be only solved
by numerical methods [2225]. This is a typical constrained nonlinear optimization problem. Since the derivatives of the unknown parameter have complicated forms,
we adopt a direct search algorithm by Powell [25] to
avoid the calculation of derivatives. The globally optimum solution may be reached by trying different initial
values.

3. Numerical example
Suppose we develop an accelerated life test plan for a
certain type of capacitor using two stresses: temperature
and electric voltage. The reliability estimate at the design
condition over a 10-year period of time is of interest. The
design condition is characterized by 50  C and 5 V. By
engineering judgment, the highest levels (upper bounds) of
temperature and voltage are pre-specied as 250  C and
10 V, respectively. The allowed test duration is 200 h, and
the total number of test units placed under test is 200. The
minimum number of failures at any test combination is
specied as 10. The test plan is determined through the
following steps:
1. According to the Arrehenius model, we use 1/(absolute
temperature) as the rst covariate z1 and 1/(voltage) as
the second covariate z2 in the ALT model.
2. The PH model is used in conducting reliability data
analysis and designing the optimal ALT plan. The
model is given by
lt; z l0 t expb1 z1 b2 z2 ,
where l0 t g0 g1 t g2 t2 .
3. A baseline experiment [26] is conducted to obtain
initial estimates for the model parameters. These
values are: g^ 0 0:0001, g^ 1 0:5, g^ 2 0, b^ 1 3800,
and b^ 2 10.

50  Cp1=z10  273:15  Co1=z11


273:15  Cp250  C;
5p1=z20 o1=z21 p10;
npij Prtptjz1i ; z2j XMNF;
i 0; 1; j 0; 1;
where
MNF 10,
T 87 600,
z1D 0:00309,
z2D 0:2, n 200, t 200, g^ 0 ; g^ 1 ; g^ 2 ; b^ 1 ; b^ 2 are given
in step 3, and x z10 ; z11 ; z20 ; z21 ; p00 ; p01 ; p10 .
5. We use numerical method as mentioned in Section 2.5
to solve the optimization problem: input the initial
baseline values for the model parameters g0 , g1 , g2 , b1 ,
and b2 ; the design stress levels z1D , z2D and highest
stress levels z1H , z2H as well as total test units n, test
duration t and minimum number of failures for each
stress level MNF.
6. The optimum plan that optimizes the objective
function and meets the constraints is shown as follows:
T L 1=z10  273:15  C 140  C,
T H 1=z11  273:15  C 209  C,
V L 1=z20 5:04 V; V H 1=z21 7:48 V,
p00 0:37;

p01 0:17;

p10 0:32;

p11 0:14.

4. Sensitivity analysis
To solve the nonlinear optimization problem given in
Sections 2 and 3, we rst obtain estimates of the values of
the model parameters g0 , g1 , g2 , b1 , and b2 . Since these are
point estimates it is important to investigate the sensitivity
of the reliability estimates to variations of the parameters
estimates changes. Therefore, we investigate and analyze
the sensitivity of the solution of the proposed optimum
ALT plan to changes in the model parameters. If a small
change in a parameter results in relatively large changes in
the solution of the optimum ALT plan, the ALT plan is
said to be sensitive to that parameter. This means that this
specic parameter needs to be investigated further before
we design the optimum ALT plan. Meanwhile some
parameters in the nonlinear optimization problem are
given arbitrarily values or are based on engineering
judgment, e.g. the censoring time t, minimum required
failure units MNF, and the total period of time, T, in which

ARTICLE IN PRESS
E.A. Elsayed, H. Zhang / Reliability Engineering and System Safety 92 (2007) 286292
Table 1
Sensitivity analysis: effect of model parameter uncertainty on stress levels

T H  C

V L

210
208
208
211
210
212
202
214
201
214
209
209

5.16
5.08
5.13
5.25
5.06
5.01
5.84
5.63
5.31
5.84
5.04
5.04

g^ 0

20
20
20
20
20
20
20
20
20
20
20
20

140
140
140
141
139
141
133
151
146
149
140
140

g^ 1
g^ 2
b^ 1
b^ 2
T

7.47
7.49
7.40
7.35
7.44
7.45
6.87
7.85
7.03
7.53
7.48
7.48

Table 2
Sensitivity analysis: effect of model parameter uncertainty on unit
allocations
Parameter

Deviation (%)

p00

p01

p10

p11

g^ 0

20
20
20
20
20
20
20
20
20
20
20
20

0.35
0.34
0.37
0.37
0.37
0.35
0.42
0.34
0.38
0.35
0.37
0.37

0.21
0.18
0.20
0.15
0.19
0.19
0.16
0.16
0.20
0.16
0.17
0.17

0.33
0.31
0.30
0.32
0.30
0.31
0.29
0.33
0.29
0.33
0.32
0.32

0.11
0.17
0.13
0.16
0.14
0.15
0.13
0.17
0.13
0.16
0.14
0.14

g^ 1
g^ 2
b^ 1
b^ 2
T

8
Stress Level (Voltage)

Deviation (%)

V H

6
5
4
3
Lower level
High level

2
1
0
10

12

14
16
18
Minimum Number of Failure

20

Fig. 3. Stress (voltage) levels vs. minimum number of failures.

250
Stress Level (Temperature)

Parameter

T L  C

291

200
150
100
Lower level
High level

50
0
200

220

260
240
Censoring Time

280

300

Fig. 4. Stress (temperature) levels vs. censoring time.

8
250
Stress Level (Voltage)

Stress Level

200
150
100
Lower level
High level

50

6
5
4
3
2

0
200

0
10

12

14
16
18
Minimum Number of Failures

Lower level
High level

20

220

240
260
Censoring Time

280

300

Fig. 5. Stress (voltage) levels vs. censoring time.

Fig. 2. Stress (temperature) levels vs. minimum number of failures.

we are supposed to estimate the reliability performance of


product at the design stress levels. If the solution of the
optimum ALT plan is sensitive to any of the above
mentioned parameters, then changes in these parameters
are needed in order to accurately estimate the reliability at
design conditions.
To conduct the sensitivity analysis, we keep changing
the value of one of the model parameter estimates

g^ 0 ; g^ 1 ; g^ 2 ; b^ 1 , and b2 , as well as t, MNF, and T, and keep


the other values unchanged, then we solve the nonlinear
optimization problem to obtain the corresponding optimum ALT plan. If a small change in any parameter results
in a relatively large change in the optimum solution, then
the ALT plan is sensitive to that parameter.
The result of sensitivity analysis is summarized in
Tables 1 and 2. As shown the proposed optimum ALT
plan derived in this example is robust to the deviations of

ARTICLE IN PRESS
292

E.A. Elsayed, H. Zhang / Reliability Engineering and System Safety 92 (2007) 286292

the model parameters. Specially the ALT plan is robust to


changes in the model parameters g^ 0 ; g^ 1 ; g^ 2 , and strongly
robust to the changes in the parameter T. But it is relatively
more sensitive to the deviations of the stress coefcients b^ 1
and b^ 2 than other parameters. So the estimates of baseline
values of model parameters b^ 1 and b^ 2 should be accurate to
ensure accurate optimum ALT plans.
Figs. 2 and 3 describe the relationship between the
optimum stress (temperature and voltage, respectively)
levels and the required minimum number of failures
(MNF). The gures show minor increase of optimum
stress levels corresponding to increasing required minimum
number of failures. Figs. 4 and 5 show the relationship
between the optimum stress (temperature and voltage,
respectively) levels and the censoring time (t). A decreasing
trend of optimum stress levels is displayed corresponding
to the increases in censoring time. Although both trends
exist, the increasing (decreasing) slopes are minor, say
0.005 as shown in Fig. 5. Therefore the optimum ALT plan
given in this example is also robust to the parameters MNF
and t.
5. Conclusions
We develop the rst analytical accelerated life testing
plans that utilize the proportional hazards model with
multiple stresses for reliability prediction. The proposed
ALT plans determine the levels for all stress types and the
number of test units allocated to each level in order to
minimize the variance of reliability estimate at normal
operating conditions over some time period (life of the
product). The proposed factorial arrangement of test units
shown in Fig. 1 enables one to conduct the entire test
simultaneously by utilizing the available equipment in an
efcient manner, which can lead to signicant savings in
test completion time and cost. Multiple-stress-type ALTs
and ALT plans are employed to overcome the difculties in
obtaining signicant amount of failure time data within the
limited amount of testing time using single stress only. A
numerical example is also given to demonstrate the
proposed accelerated life testing plans. The result of
sensitivity analysis of the given example shows that the
solution of the proposed optimum ALT plan is robust to
the deviations in model parameters.
References
[1] Park JW, Yum BJ. Optimum design of accelerated life tests with two
stresses. Nav Res Logistic 1996;43:86384.
[2] Tang LC, Goh TN, Sun YS, Ong HL. Planning accelerated life tests
for censored two-parameter exponential distributions. Nav Res
Logistic 1999;46(2):16986.

[3] Cox DR. Regression models and life tables (with discussion). R Stat
Soc B 1972;34:187208.
[4] Elsayed EA. Reliability engineering. MA: Addison-Wesley Longman;
1996.
[5] Elsayed EA, Wang X. An extended hazard regression model for
ALT. Eighth industrial engineering research conference, Phoenix,
Arizona, May 2324, 1999.
[6] Zhao W, Elsayed EA. General model for variable-stress accelerated
life testing. Twelfth industrial engineering research conference,
Houston, Texas, May 1618, 2004.
[7] Brass W. On the scale of mortality. In: Brass W, editor., Biological
aspects of mortality, Symposia of the society for the study of human
biology, vol. X. London: Taylor & Francis Ltd; 1971. p. 69110.
[8] Elsayed EA, Zhang H. Design of optimum reliability test plans under
multiple stresses. QUALITA 2005, quality and dependability,
Bordeaux, France, March 1618, 2005.
[9] Hannerz H. An extension of relational methods in mortality
estimation. Demographic Res 2001;4:33768.
[10] Elsayed EA, Jiao L. Optimal design of proportional hazards based
accelerated life testing plans. Int J Mater Product Technol 2002;
17:41124.
[11] Elsayed EA, Zhang H. Design of optimum simple step-stress
accelerated life testing plans. Proceedings of 2005 international
workshop on recent advances in stochastic operations research,
Canmore, Canada; 2005.
[12] Nelson W, Meeker W. Theory for optimum censored accelerated life
tests for Weibull and extreme value distributions. Technometrics
1978;20:1717.
[13] Nelson W. Accelerated testing: statistical models, test plans, and data
analyses. New York: Wiley; 1990.
[14] Tang LC. Planning for accelerated life tests. Int J Reliab Quality Saf
Eng 1999;6:26575.
[15] Yang GB. Optimum constant-stress accelerated life test plans. IEEE
Trans Reliab 1994;43:57581.
[16] Kobayashi T, Ariyoshi H, Masua A. Reliability evaluation and
failure analysis for multilayer ceramic chip capacitors. IEEE Trans
Components Hybrids Manuf Technol 1978;CHMT-1:31624.
[17] Minford WJ. Accelerated life testing and reliability of high K
multilayer ceramic capacitors. IEEE Trans Components Hybrids
Manuf Technol 1982;CHMT-5:297300.
[18] Mogilevsky BM, Shirn GA. Accelerated life tests of ceramic
capacitors. IEEE Trans Components Hybrids Manuf Technol 1988;
CHMT-11:3517.
[19] Munikoti R, Dhar P. Highly accelerated life testing (HALT) for
multilayer ceramic capacitor qualication. IEEE Trans Components
Hybrids Manuf Technol 1988;CHMT-11:3425.
[20] Weis EA, Caldararu D, Snyder MM, Croitoru N. Reliability
evaluation and prediction for silicon photodetectors. IEEE Trans
Reliab 1988;R-37:1423.
[21] Escobar LA, Meeker WQ. Planning accelerated life tests with two or
more experimental factors. Technometrics 1995;37(4):41127.
[22] Fletcher R. Practical methods of optimization. New York: Wiley;
1987.
[23] Gill PE, Murray W, Wright MH. Practical optimization. London:
Academic Press; 1981.
[24] Hock W, Schittkowski K. A comparative performance evaluation of
27 nonlinear programming codes. Computing 1983;30:335.
[25] Powell, MJD. A direct search optimization method that models the
objective and constraint functions by linear interpolation. DAMTP/
NA5. Cambridge, England; 1992.
[26] Chernoff H. Optimal accelerated life design for estimation. Technometrics 1962;4:381408.

You might also like