Sie sind auf Seite 1von 7

IEEE TRANSACTIONS ON RELIABILITY, VOL. R-32, NO.

1, APRIL 1983

59

Optimum Simple Step-Stress Plans for Accelerated Life Testing


Robert Miller, Member ASQC General Electric Co., Schenectady Wayne Nelson, Senior Member IEEE & ASQC General Electric Co., Schenectady
Key Words-Accelerated test, Step-stress, Constant-stress test, Exponential life distribution, Maximum likelihood estimation.

section 2. The test data are used to estimate the parameters of the relationship. The relationship is then extrapolated to estimate life at a constant low design stress. Estimation is usually done by maximum likelihood (ML) rather than least squares since the assumed life distribution is often not s-normal but exponential or Weibull.
Step Stress

Special math needed for explanations: Statistics, Maximum likelihood theory Special math needed to use results: None Results useful to: Engineers and accelerated test planners stresses) step-stress tests where all units are run to failure. Such plans minimize the asymptotic variance of the maximum likelihood estimator (MLE) of the mean life at a design stress. The life-test model consists of: 1) an exponential life distribution with 2) a mean that is a log-linear function of stress, and 3) a cumulative exposure model for the effect of changing stress. Two types of simple step-stress tests are considered: 1) a timestep test and 2) a failure-step test. A time-step test runs a specified time at the first stress, whereas, a failure-step test runs until a specified proportion of units fail at the first stress. New results include: 1) the optimum time at the first stress for time-step test and 2) the optimum proportion failing at the low stress for a failure-step test, and 3) the asymptotic variance of these optimum tests. Both the optimum time-step and failurestep tests have the same asymptotic variance as the corresponding optimum constant-stress test. Thus step-stress tests yield the same amount of

Reader AidsPurpose: Widen state of the art

In a time-step-stress accelerated life test, stress on each unit is not constant but is increased by planned steps at

the uni es

planned times. A test unit starts at a specified low stress. the unit does not fail in a specified time, stress on it d f a specified time , stress raised and held a specified time. Specimen stress

If

Abstract-This paper presents optimum plans for simple (two

is is is repeatedly increased and held, until the specimen fails. The step-stress pattern is chosen to assure failures quickly.
As with the constant-stress test, one estimates the

parameters of a model for life under step stressing. The parameter estimates are used to estimate life at a constant design stress. However, one now needs a model that relates the life distribution under step stressing to that under constant stress. Cumulative exposure models for life under step-stressing have been presented by [1, 8, 10, 17].
Overview

information as constant-stress tests.

1. ACCELERATED LIFE TESTING

Accelerated life testing quickly yields information on product life. Test units are run at high stress and fail sooner than at design stress. The short lives are then extrapolated to estimate life at design conditions. Such overstress testing reduces time and cost. Overstress involves high temperature, voltage, pressure, vibration, cycling rate, load, etc., or some combination of them. Engineering experience usually suggests such accelerating variables for a specific product or material.
Constant Stress

to x2; and the test is continued until all units run to failure. In a failure-step test, stress is changed to x2 when a fraction P1 of the units fail at stress x, and the remaining fraction are run to failure. Such tests are called simple step-stress tests since they use only two stresses. This paper presents optimum simple step-stress tests employing an exponential life distribution at constant stress and the cumulative exposure model of [8, 10]. While simple, these results provide much needed insight into the design of step-stress tests. Further work is needed to extend results to other distributions and censored tests. Section 2 describes the accelerated test and cumulative exposure models. Section 3 presents the optimum simple time-step test. The optimum simple failure-step test is

at stress x, and run until time T,, when the stress is changed

This paper considers step-stress tests using only two stresses. In a time-step test, units are initially placed on test

In a constant-stress accelerated test, units are each run at a (possibly different) constant high stress. Life at a design stress is estimated by regression methods as follows. Notation A relationship between life and constant stress is assumed. For example, for voltage stress, the inverse n total sample size power law is often used; it is a linear relationship between XD, XH, XL, X1, X2 transformed stresses (Design, High, log mean life and log voltage. It is discussed in [6] and in Low, 1-st, 2-nd) 0018-9529/83/0400-0059$01 .00 )1983 IEEE

presented in section 4. Section 5 discusses the optimum constant-stress test. Section 6 compares the properties of the optimum test plans.

60

IEEE TRANSACTIONS ON RELIABILITY, VOL. R-32, NO. 1, APRIL 1983

0D9 OH9'L9 01 C02 The F-th fractile of the life distribution at a stress value mean life at the subscripted stress x is: F(t, x) Cdf for stress x Fi(t) Cdf at stress i (2.3) yo, y, parameters of the log linear function between tF(X) = - e(x) ln[I -1]. stress x and the mean life e; i.e., In{e(x)} = Yo + Thus

(XL - XD) from XL to XD In{tF(x)} = In[O(x)] + In[-In(l - F)] as a multiple of the test range (XH - XL); 4 (XL - XD)/(XH - XL) > 0 = yO + y, x + In[-In(I -F)]. (2.4) T, length of time at stress 1 of a simple time-step test TL, H optimum test time at XL (xH) in a simple time-step So, on semi-log paper, percentiles versus stress follow test with XL (XH) first parallel straight lines as shown in figure 1. PH, PL optimum fraction of test units that fail at the low stress for a simple failure-step test with XH (XL) if (LIFE) first Pc optimum fraction of test units allocated to the low stress of a simple constant-stress test EP* mean proportion of test units failing at stress 1, x', L for an optimum simple time-step test co desired relative error in 0D; co > 1 (1 + f)/2 fractile of the standard s-normal \99%FAILED Kp H distribution tF(X) F-th fractile of the life distribution at stress x 630
4

y'X amount of extrapolation

2. MODELS, ASSUMPTIONS, AND PROBLEM STATEMENT


This section presents the assumptions and models for: 1) the distribution of life as a function of constant stress, and 2) the cumulative effect of exposure in a step-stress
XD

1/

XL

XH

STRESS X

test.

Fig. 1. Constant-Stress Model.

2.1 Constant-Stress Test Models and Relationships For omeprouct undr cnstnt tres, te flloing model adequately describes life as a function of the stress. Its assumptions are:

1. For any constant value of the stress, the life distribution is exponential. 2. The mean G (the characteristic life, and roughly 63-rd percentile) is a log-linear function of a (possibly transformed) stress x; that is, ln[e(x)] = yo + yix, x
(x) = exp(yo + y x).
(2.1)

The yo and yl are parameters of the product and the test method. If x is the log of voltage stress, then (2.1) is the inverse power law [7, 9, 13]. If x is the reciprocal of absolute temperature, then (2.1) is the Arrhenius relationship [11, 2.2 Cumulative Exposure Model 12]. Other relationships are presented by [2, 3, 6, 15, 16]. Under the assumptions, the fraction F(t; x) of units To analyze data from a step-stress test, one needs a failing by time t under a constant stress x is: model that relates the life distribution under step-stressing to the distribution under a constant stress. Such a model F(t; x) = 1 - exp[-t/exp(yO + y,x)], t > 0. (2.2) follows. The model assumes that the remaining life of a unit This model is depicted in figure 1, which has a log scale for depends only on the 'exposure' it has seen, and the unit time and a linear scale for (transformed) stress x. does not remember how the exposure was accumulated.

Example. Ref. [9] reports an accelerated life test with 76 times (in minutes) to breakdown of an insulating fluid at constant voltage stresses (kV). The inverse power law model was used, and the extreme (transformed) test stresses are XL = In(26.0) = 3.2581, and XH = In(38.0) = 3.6376. The (transformed) design stress is XD = In(20.0) = 2.9957. This test is used to illustrate the optimum tests. The ML estimates of the model parameters for those data are y0 = 64.912, and Yi = - 17.704. These estimates were obtained with the STATPAC program [14]. The ML estimates of the means at the high and low stresses are EH = 1.67 min. and 0L = 1380 min. The estimate of the mean life at the design stress is OD = exp[64.912 - 17.704(2.9957)] 144,000 min., and the estimate of the asymptotic variance of the In estimate is Var{In(6D)} = 0.427.
'

MILLER/NELSON: OPTIMUM SIMPLE STEP-STRESS PLANS FOR ACCELERATED LIFE TESTING

61

More precisely, if different 'large' groups of units each have a different exposure history but the same fraction failed, then survivors of all such groups have the same remaining life distribution. Moreover, if then held at a constant stress, the survivors will continue failing according to the life distribution for that stress but starting at the age corresponding to the previous fraction failed. Figure 2 depicts this model for a simple test with two stress levels.
STRESS x
(A)

When step 2 starts, units have equivalent age T' which would have produced the same fraction failed seen at the end of step 1; that is, F2(T') = FI(T,) (2.6)

T' = F2 [Fl (T1)];

(2.7)

here FT'[ ] inverse of F2[ ]. SO T' is the Fi(T1) fractile of the distribution F2. The population Cdf of units failing
by time t> Ti is;

X2 '
xl

F0(t) = F2(t - T1 + T'), t > Tr.


x

(2.8)

o *(*x x x x 0

TIME t t

FI tj(t)
X2

(B) 0

i /<fit

So Fo(t) consists of segments FI(t) and F2(t) as shown in part C of figure 2. A different step-stress pattern would have different Fo(t). Ref. [10] presents this cumulative exposure model in more detail and for any number of steps. It describes an application employing the Weibull distribution. /The preceding general theory is specialized to the model of section 2.1 as follows. By (2.2),

O i |TIME | O T T

Fi(t)
t

exp(-t/0j);
=

(2.9)
exp(y0 +

i,Fo (t)
(C)
0 z
0

this exponential distribution has a mean Oi y xI). For step 1, (2.5) becomes:

Fo(t)

I - exp(-t/01),

0 ( t ( T.

(2.10)

i TIME t

The equivalent time T' at x2 that produces a fraction failing Fi(T1) at xi is given by (2.7) as T- = T-(02/01). For step 2, (2.8) becomes

FF(t) =

1 - exp{

[t - Ti + Ti (02/0)1/021,} t>T,. (2.11)

For the exponential distribution only, this model corresponds to Cox's proportional hazard model with a time varying covariate. However, Cox's model assumes no parametric form of the distribution, and fitting methods Part A of the figure shows a step-stress pattern with two for it (e.g., [16]) do not apply here. stresses and shows sample failure times. Part B depicts the two Cdf's for stresses xl and x2. The arrows in part B show 2.3 Problem Statement that the units first follow the distribution for the first stress. When stress 2 is applied, the survivors continue Further assumptions. along the life distribution for stress 2, starting with the fraction failed at stress 1. The Cdf under the step-stress 3. A test uses only two stresses, xi and x2. Such a test is pattern appears in part C; it consists of the segments of the called simple. Units unfailed at first stress xl are run to failure at stress x2. distributions for the two constant stresses. This model can be mathematically expressed as follows 4. Any two test stresses in the allowable test range to obtain the Cdf Fo(t) of time t to failure under this step- from the lowest XL to the highest XH can be used. The XL stress pattern. and XH are given. Suppose step 1 at stress x, runs to time T,. Let Fi(t) -Cdf 5. The specified design stress XD is below XL. Of time t to failure for units run at constant stress xi, i = 1, 2. 6. The random variations of the lives of the test units The population Cdf of units failed by time tin step 1 is: are all s-independent.

Fig. 2. Relationship Between Constant- and Step-stress Distributions.

Fo(t) = F1(t),

0 4 t S T,.

(2.5)

Usually XL is chosen as low as practical subject to the

62

IEEE TRANSACTIONS ON RELIABILITY, VOL. R-32, NO. 1, APRIL 1983

constraint that the test end by a desired time. Usually XH is chosen as high as possible subject to the relationship (2.1) holding over the range XH to XD. Estimates. Several estimation methods are reviewed and compared in [4]. ML estimation is used here rather than least squares, linear estimation, or the other analytic methods, because 1. Optimum plans for ML estimation are easier to calculate than those for linear estimation. 2. ML estimates have a minimum statidard deviation for large samples. For small samples, their standard deviation is generally comparable to that for other estimates. 3. In a time-step test, the lives of test units at the first stress are time censored at time T1. Linear methods assume the data have failure censoring and are not strictly correct for the time censoring considered here, whereas ML methods apply to both time and failure censoring. 4. The optimum design for ML estimation is undoubtedly close to optimum for any other methods of estimation, even a graphical method. 5. ML fitting is easily done by means of STATPAC [14] and SURVREG [16]. SURVREG [14] and [16].. 6. Asymptotic theory gives approximate s-confidence limits for parameters and percentiles [5, 10]. Optimization criterion. An optimum plan provides a "best" estimate of a quantity. Here, it minimizes the asymptotic (large-sample) variance of the ML estimator of the mean at a specified design stress. For a simple time-step test, this variance is minimized by the optimum choice of: 1) the first test stress x, 2) the second test stress x2, and 3) T1 the time at stress x,. For a simple failure-step test, the variance is minimized by the optimum choice of: 1) xl, 2) x2, and 3) Pi, the specified sample fraction failing at x when stress is switched to x2. For the constant-stress test, it is minimized by the optimum choice of: 1) x,, 2) x2, and 3) Pc' the proportion of test units run at the lower stress. These choices also minimize the asymptotic coefficient of variation of any percentile estimator, that is, minimize the relative error.
3. OPTIMUM SIMPLE TIME-STEP TEST

and the cumulative exposure model above, the optimum variance is the same whether high or low stress is first. For a Weibull life distribution, which stress is better first undoubtedly depends on whether the Weibull shape parameter is greater or less than 1. Proofs of the following results appear in [8]. 3.2 Low Stress First The optimum time at 'low stress first' is:
TL = 0L 2n[(24 + 1)/4];

(3.1)

Example. For the insulating fluid example, the optimum test has 4 = (3.2581-2.9957)/(3.6376-3.2581) = 0.6914. So TL = OL In[(2 * (0.6914) + 1)/0.6914] = 1.237 eL = 1.237(1380) = 1707 min. The ML estimate of )L was given as 1380 min in section 2.1.

4 - (XL - XD)/(XH - XL) is the amount of extrapolation as a multiple of the test range (XH - XL).

3.3 High Stress First . .,. ~~~~The optimum time at 'high stress first' iS:
TH= 0H ln[(24+ 1)/( +

1)].

(3.2)

test runs at the high stress for *r = eH In[(2(0.6914) + 1)/(0.6914 + 1)] = 0.3427 OH = 0.3427(1.67) = 0.57 min. The ML estimate of 0H was given as 1.67 min in section 2.1.

Example. For the insulating fluid example, the optimum

3.4 Properties Figure 3 shows .r*/eL and T

*/G

as functions

of4

ticipated. In particular, when the low

1. As 4

0 (no extrapolation),

T-r/eL

-->

design stress (4 = 0), units are run only at the low stress.

stress

equals the

o, as an-

3.1 Optimum Stresses The optimum simple time-step test uses the extreme stresses (XL and XH) of the allowed test range. They minimize the asymptotic variance of the ML estimate of the mean life at the design stress. In practice, XH should not cause failure modes different from those at the design stress, so the model will be valid over the range of the test| and design stresses. XH(XL) iS chosen as high (low) as pOSSible to minimize the variance of the estimate of mean life at the design stress. That the optimum test uses the extreme tion in ordinary linear regression. Traditionally, low stress is used first. 'High stress first' is also considered below. For an exponential life distribution

TL*/L |
A
.0

In(2)

TH /@H

__
Vlllllllll

ll _

Fig. 3. Optimum Specified Times vs 4.

MILLER/NELSON: OPTIMUM SIMPLE STEP-STRESS PLANS FOR ACCELERATED LIFE TESTING


-

63

-* In(2) 2. As 4 oo, P* l/2. That is, half of the units are 2. As 4 -- 0, T 0.69. Large 4 corresponds to plans with great extrapolation compared to the failed at each stress when the design stress is far from the test range. Then half fail at each stress. test range. 3. TH and TL are related by exp(- */e)H) + exp(--r//eL) 3. Always 1/2< P* < 1. That is, always the majority of = 1. units fail at the low stress.

mean

Optimum TL and TH are in terms of a true unknown life, EL or 0H. This is typical of optimum plans. Such a plan is called locally optimum in [2]. In practice, one must approximate the unknown mean with a guess or with similar data. The effect of using a wrong value for 0L or 0H is discussed in section 6.4.
4. OPTIMUM SIMPLE FAILURE-STEP TEST

.00C.
l
0. 80
0.70
0

---_

__

The optimum failure-step test specifies the fraction of units P* that must fail at stress x1 before the switch to stress x2. In practice, such a test can require constant monitoring and might be inconvenient. Proofs of the following results appear in [8]. 4.1 Optimum Stresses The optimum simple failure-step test uses the extreme stresses (XL and xH) of the allowed test range.

_
060

0.60

0.50 0

Fig. 4. Optimum Proportion P* of Lowest Test Stress Versus 4.


5. OPTIMUM CONSTANT-STRESS TEST

4.2 Low Stress First

A simple constant-stress test uses two test stresses and The optimum simple failure-step test switches from XL low stress) allocates the n sample units to them. Chernoff [2] shows to XH when the sample fraction failed at XL (the that the optimum test uses the extreme stresses XL and XH reaches: and that the optimum proportion PC of test units at XL is (4.1) equal to the optimum specified fraction failing in a simple PL = (1 + 4)/(24 + 1). step-stress test, i.e., PC = PL = P*, which are all denoted Example. For the insulating fluid example, the op- by P* in figure 4. timum simple failure-step test runs until a proportion PL = (1 + 0.6914)/(2(0.6914) + 1) = 0.710 fail. If there are 6. PROPERTIES OF THE OPTIMUM PLANS 76 test units, 0.710 x 76 54 units must fail at the low 1 Overview stress. stress. The previous sections discussed five optimum simple 4.3 High Stress First accelerated life tests: 1. Constant-stress test (section 5). The optimum simple failure-step test switches from XH to XL when the sample fraction remaining (and failing at 2. Time-step test starting at the low stress (section 3.2). the low stress) is: 3. Time-step test starting at the high stress (section 3.3). 4. Failure-step test starting at the low stress (section (4.2) PH = (1 + 4)/(24 + 1). 4.2). Eqs. (4.1) and (4.2) show PL = PH. That is, the optimum 5. Failure-step test starting at the high stress (section fraction failing at the low stress is the same whether the 4.3). first stress is low or high. Figure 4 shows PL = PH, denoted Eahotestssispimmntesnetatt by~ there, aafucino4. P. minimizes the (large-sample) variance of the ML estimate Proertes.Theoptmumspcifed racionfaiingatof the mean at a specified design stress. This section the lw stesshasthe olloing poperies:discusses properties of the optimum tests, including the L 1. As 4 - 0 (no extrapolation), P* -~1, as anticipated. variance and the fraction failing at each stress. Ref. [8] That is, when the low stress equals the design stress (4 = 0), presents proofs and the mean test durations and further all units are run to failure at the low stress. properties.
-

~~~~~~~~~~~6.

64

IEEE TRANSACTIONS ON RELIABILITY, VOL. R-32, NO. 1, APRIL 1983

6.2 Optimum Variance and Sample Size 11/in(1.20)}2 = 462, and OD will be within about (1.20 - 1) = 20% of eD with probability 0.90. The optimum asymptotic variance is the same for all five optimum simple plans,fiv optimum simple plans, namely, 6.3 Fraction Failing at Each Stress . namely,.

Var*{in(AD)} = (24 + 1)2/n

(6.1)

Figure 5 shows this variance as a function of 4. For a as a function of 4 Section 4.3 comments on the proportion Weibull life distribution with shape parameter different as a function of . Example. For the insulating fluid example, the opfrom 1, we anticipate that the five variances are not all equal. timum simple time-step test that starts with 26.0 kV has a mean proportion failing at 26.0 kV of EPL = (0.6914 + _ l00 1)/[2(0.6914) + 1] - 0.710. So we anticipate about 76(0.710) = 54 of the 76 test units to fail at the low stress of t____ 7f___ 26.0 kV. Similarly, an optimum simple failure-step test that starts at 26.0 kV runs there until 54 of the 76 units fail___ __I_l ____ ed. An optimum constant-stress plan runs 54 of the 76 Z n Var / units at 26.0 kV.
106.4

Comparison. For given xL and XH, the (mean or specified) proportion of test units failing at XL 1S the same function P*(4) for the 5 optimum plans. Figure 4 shows P*

Wrong eLoreH

_____|___d_ 0(

I__________

35

'i

To use an optimum simple time-step test, one must know Oe(GH) in order to calculate the optimum time Tr (Tr) at stress XL (XH). xLFirst. Suppose one incorrectly uses Oj for Oe and -r = 0L In[(24 + 1)/1] for the true optimum TL = 0L In[(24 + l)/4]. Then the actual plan is not optimum and has a higher asymptotic variance

Fig. 5. Optimum n

Var*{in(e6D)} Versus 4.

Var[{ln(kD)}

(4 +

1)1{1

[4/(24 + 1)]IL/eL}-1
(6.3)

+ 42[(24 +

Example. For the insulating fluid example, suppose an optimum simple step-stress test begins with low stress and Figure 6 shows Var'{ln(eD)}/Var{ln(OD)} as a function uses n = 76 units. Then, Var*{in(0D)} = [2(0.6914) + of OJGL for selected values of 4. Of course, the variance 112/76 = 0.075. In contrast, for the actual non-optimum ratio is minimum for 3L/'L = 1. plan, Var{in(OD)} = 0.427, larger by a factor of 5.7. That 100 o -is, the actual plan requires 5.7 times as many specimens as the optimum plan for the same accuracy. 50 50 Sample size. For an optimum simple step-stress plan, I Van. the accuracy of the estimator G, is determined by the sam_i V ple size n. Suppose one wishes OD to be within a factor )> ar>L 1 of the trueeD with (high) probability ,that is:

I)/4]EL/eL.

Pr{OD/cJ) < OD < (AJOD}


n

lo f
5
I

t0

10

The approximate sample size that achieves this is:

WoMKg3(2S, + W)/lN(0)]2;

(6.2)

s fractile. For co near 1, (co C I Iv K,, =tandard s-normal (1 + I/2 - 1) is approximately the relative error in G3D. Eq. (6.2) isI based on the assumption that n is large enough that the . 5 01 0 distribution of in(GD) is close to s-normal. Example. For the insulating fluid, suppose that the 0D iSeLL to be within a factor co = 1.20 of GD with 0 =0.90 probability. Then, for the optimum plan, n = { 1.645(2(0.6914) + Fig. 6. Effect of Wrong OL on VarL.

|ll|
0

MILLER/NELSON: OPTIMUM SIMPLE STEP-STRESS PLANS FOR ACCELERATED LIFE TESTING

65

Figure 6 shows that underestimating E)L by some factor increases Var2 L less than does overestimating E)L by the saefactor. o i samefcor So it iS safer .to underestimate ,eL. L XH First. Suppose one incorrectly uses O'H for O. The H asymptotic variance is:

Var'{ln(QD)}

= 42{1 [( + 1)/(24 +
)2[(24+

1)] OH/OH} -1
ixiO'..8/O@,
+ (1+ 4)[24 1)14+ )](64)

Figure 7 shows Var'{ln(eD)}/VarH{ln(0D)} as a function of 3/(0H and 4. The same conclusions are true for figure 7 for figure 6. However, effect of Incorrectly as for figure 6. However, the effect of incorrectly estimating E0H tends to be smaller than that for incorrectly estimating 0L; i.e., for E3/E)H = O'L/OL, (Var'/VarH) < (Var2/Var*). This suggests considering using high stress
first, contrary to usual practice.

[91 W.B. Nelson, "Statistical methods for accelerated life test data-The inverse power law model," General Electric Research & Development TIS Report 71-C-011, 1970.* Graphical methods published in IEEE Trans. Reliability, vol R-21, 1972 Feb, pp 2-11; correction, p 195.* Least squares methods published in IEEE Trans. Reliability, vol R-24, Jun 1975, pp 103-106. [10] W.B. Nelson, "Faster accelerated life testing by step-stress-models and data analyses," General Electric Research & Development TIS Report 78CRD051, 1978.* Also published in IEEE Trans. Reliability, vol R-29, 1980 June, pp 103-108, in a simpler presentation.*

[8] R. Miller, W.B. Nelson, "Optimum simple step-stress tests for accelerated life testing," General Electric Research & Development TIS Report 79CRD262, 1979.* TI eot7CD6,17.

[11] W.B. Nelson, T.J. Kielpinski, "Optimum accelerated life tests for normal and lognormal life distributions," IEEE Trans. Reliability,
vol R-24, 1975 Dec, pp

sored life tests for Weibull and extreme value distributions," Technometrics, vol 20, 1978, pp 171-177.* 10 T ici ---=--F =--t_ X-_Z ~ -7 [14] W.B. Nelson, C.B. Morgan, P. Caporal, "1979 STATPAC ___ _ _ + __ ! j }i HISimplified-A short introduction to how to run STATPAC, a Var L--' H -_ ____ ls general statistical package for data analysis," General Electric __ ____ __ Va* H r __ 10 __ __ _ -4 _ Research & Development TIS Report 78CRD276, 1978.* I_ 1 _ _ 4 1L) _ [15] R.L. Prentice, "Exponential survivals with censoring and ex1\\ / / planatory variables," Biometrika, vol 60, 1973, pp 279-288. - 2- L ___ __ 9 _D.L. Preston, D.B. Clarkson, "SURVREG: An interactive program [16] for regression analysis with censored survival data," ASA Pro0.1 1 z G D0.1 ceedings on Statistical Computing from American Statistical Association annual meeting, Houston, TX, 1980. 1 05 [17] W. Yurkowski, R.E. Schafer, J.M. Finkelstein, "Accelerated OH/0H testing technology," Rome Air Development Center Technical Report RADC-TR-67-420, Griffiss AFB, NY, 1967. Fig. 7. Effect of Incorrect EH on VarA.

celerated tests for normal and lognormal life distributions," Technometrics, vol 18, 1976, pp 105-114.* [13] W.B. Nelson, W.Q. Meeker, "Theory for optimum accelerated cen-

[12] W.B.

Nelson,

T.J. Kielpinski,

310-320.* "Theory for

optimum censored ac-

0__

REFERENCES

[1] W.R. Allen, "Notes oIn some statistical aspects of design and analysis of accelerated life tests," private communication, 68 Magnolia Lane, Princeton, NJ 08540 USA, 1965. [2] H. Chernoff, "Optical accelerated life designs for estimation," Technometrics, vol 4, 1962,pp 381-408. [3] P. Feigl, M. Zelen, "Estimation of exponential survival probabilities with concomitant information," Biometrics, vol 21, 1965, pp 826-838. [4] G.J. Hahn, W.B. Nelson, "A review and comparison of methods for regression analysis of censored data," General Electric Research & Development TIS Report 71-C-196, 1971.* Also published in IEEE Trans. Reliability, vol R-23, 1974 Apr, pp 2-10. [5] B. Hoadley, "Asymptotic properties of maximum likelihood estimators for the independent not identically distributed case," Ann. Mathematical Statistics, vol 42, 1971, pp 1977-1991. [61 N.R. Mann, R.E. Schafer, N.D. Singpurwalla, Methods of Statistical Analysis of Reliability and Life Data, Wiley, 1974. [7] W.Q. Meeker, W.B. Nelson, "Optimum accelerated life tests for Weibull and extreme value distributions and censored data," IEEE Trans. Reliability, vol R-24, 1975, pp 321-332.

*These reprints and General Electric TIS Reports are available from the Technical Information Exchange; 5-317; GE Corp. Research & Development; SchncayNY135UA henectady, NY 12345 USA.
AUTHORS
Dr. Robert W. Miller; MD: K60; GE Aircraft Engine Bus.; Evendale, OH 45215 USA. Robert W. Miller was born in Albany, New York, on 1950 March 20. He received the BA degree in Biology/Mathematics from Hartwick College in 1972, an MS degree in Applied Statistics and a PhD degree in Administrative and Engineering Systems from Union University in 1978. He worked as a virologist at Sterling-Winthrop Research Institute and a consultant statistician at Allied Chemical Corporation. He is a statistician at General Electric's Aircraft Engine Group, Cincinnati. He is a member of ASA and ASQC and is an ASQC-Certified Quality Engineer and Reliability Engineer. Dr. Wayne Nelson; K-1, 3A32; GE CRD; Schenectady, NY 12345 USA. Wayne Nelson: For biography see vol R-30, 1981 Jun, p 155.

Manuscript TR81-85 received 1981 July 31; revised 1982 September 29.

s TII1
*

st

1983 SEPTEMBER 21-23 *j6 Raleigh *North Carolina *USA

Das könnte Ihnen auch gefallen