You are on page 1of 6

S T A T I S T I C S

How To Estimate the Parameters


Of a Weibull Distribution
Data from an incomplete life test in overstress conditions
can be used to determine product reliability
by
Mitchell O. Locks

Q
UALITY PROFESSIONALS OFTEN the characteristic lifetime, the location parameter µ
need to determine time to failure is the minimum lifetime and there is a shape para-
and reliability of a product from a meter β. Since the minimum lifetime is not neces-
life test. sarily zero (0, the beginning of the life test), the
Sometimes this analysis must be value of µ must frequently either be assumed,
done when the life test is incomplete. approximated or obtained by trial and error.
The technique described in this article Estimation, therefore, is done with just the scale
shows how to estimate the Weibull and shape parameters, and the Weibull simplifies to
parameters when the incomplete life test is per- a two-parameter distribution.
formed under overstress conditions with measurable Let t denote the time to failure. The Weibull prob-
degradation of the product during the test. ability density function is
The article also shows how to extend this tech-
nique to find confidence bounds on the reliable
lifetime of the product. For this purpose we use .
(1a)
a standardized ex- The distribution function, the cumulative proba-
treme-value statistic bility, is
that is obtained by a
logarithmic trans- (1b) .
formation of the The reliability, the probability of successfully
Weibull parameters. operating until t, is
A Weibull distrib-
ution is a member of
Weibull distribution a family of three- .
(1c)
is a member of a family of parameter time-to- If a nonzero location parameter is assumed, sub-
three-parameter time-to-failure failure probability stitute (t-µ) for t in the three previous equations so
probability distributions. distributions. The the computations and analysis proceed in a two-
scale parameter δ is parameter format.

QU A L I T Y P R O G R E S S I A U G U S T 2 0 0 2 I 59
H O W T O E S T I M AT E T H E PA R A M E T E R S O F A W E I B U L L D I S T R I B U T I O N

The extreme-value distribution This BLI technique applies both to complete sam-
Like the better known normal (Gaussian) distribu- ples, where all n items are tested to failure, and to cen-
tion, the extreme-value distribution is a two-parame- sored samples, which are terminated before all items
ter distribution with both location and scale. Unlike have failed. It takes into account n (the sample size),
the Weibull, there is no shape parameter. By a change m (the censoring number) and i (the rank order num-
of variables, making the random variable the natural ber of an observation).
(base-e) logarithm n(t) of the time to failure t, For each (n,m) pair, two sets of weights are provid-
instead of t, the Weibull is transformed to an extreme- ed. The weights for the location parameter u are
value distribution. As a result of the change, the loca- called a(n,m,i), where i is the rank of the observation
tion parameter is and the sum of these weights is 1. The weights for
the scale parameter b are called c(n,m,i), and their
(2a) sum is 0.
and the scale parameter is This is similar to the case with the normal distri-
bution, because the sum of the weights for the
. (2b) location parameter is zero and the sum of the
As a result of this change, the random variable is weights for the scale parameter is one. The formu-
expressed as a linear combination of the location las are
and scale parameters, n(t) = u + bx, where x can be and
any real number. There is an analogy to the normal
distribution: Let Y:N(µ, σ) represent a normal distri-
bution with random variable y, location parameter µ .
(the mean) and scale parameter σ (the standard
deviation). Also let X:N(0,1) represent the “stan- The weights a(n,m,i) and c(n,m,i) are published for
dard” normal distribution. Then any normal ran- sample sizes n, 2 ≤ n ≤ 25 and censoring numbers
dom variable can be expressed as a linear function m, 2 ≤ m ≤ n.4 Because two unique sets of weights are
of x, µ and σ or needed for every (n,m) pair, the tables are volumi-
y = µ + σx nous, but they are easy to use because the resulting
As another consequence of the change of variables, computations are simple weighted averages of the
you are able to extract good estimates of the parame- logarithms of the times to failure.
ters that are linear combinations (weighted averages) The Weibull parameters are obtained by reversing
of the sample data. Good estimates are those that sat- the changes of parameters in equations 2a and 2b:
isfy some statistical criterion such lack of bias, suffi-
ciency, efficiency or invariance.
If we refer again to the normal distribution as a
familiar example, the unbiased estimate of the mean Confidence bounds
µ for a random sample of n observations is the sam-
Reliability specifications are usually stated in terms
ple mean. This is the sum of the values divided by n.
of fixed probabilities such as 0.9, 0.95 or 0.99. Given a
In other words, the estimate of the location parameter
set of life test data, it is useful to determine the confi-
µ is a weighted average of the observations, where
dence level at which the specified reliability is
each observation has an equal weight of 1/n and the
attained. To facilitate confidence assessment, Mann,
sum of the weights is 1. Wilfrid Dixon has shown an
Kenneth W. Fertig and Ernest M. Scheuer (whom I
efficient linearized estimate of the scale parameter σ
will refer to as MFS) generated a set of tables of a stan-
can also be obtained as a weighted average of the
dardized statistic
ranked observations, where the weights depend both
on n and the ranks, and the sum of the weights is
zero (0).1, 2
where tR is a required lifetime and the reliability R is
The Mann BLI technique either 0.9, 0.95 or 0.99.5
Nancy R. Mann developed a linearized technique, These tables provide values of VR at 11 different
called best linear invariant (BLI), for estimating the confidence levels ranging from 0.02 to 0.98 for the
location and scale parameters of an extreme value dis- same (n,m) pairs as in the BLI tables, providing m ≤ 3.
tribution and the scale and shape parameters of a Like the Mann tables, each table in the set R is unique
Weibull distribution, based on life test data.3 for a specified sample size n and censoring number m.

60 I A U G U S T 2 0 0 2 I W W W . A S Q . O R G
are tested for 4,000 hours. A measurement is taken of
FIGURE 1 GaAs Laser 4,000-Hour Degradation the current input for each unit about every 250 hours.
Data Three units failed: one at 3,374, one at 3,621 and one
14
at 3,781 hours. Since an electrical current input history
was maintained for each of the other 12 units that
12
didn’t fail, it would be possible to predict by projec-
Percentageincreasein operating current

10
tion the time at which failure would occur.
Figure 1 is a graph of the current input histories for
8
all 15 units for all 4,000 hours of testing.
6
Case 1: the original data
4
Of 15 units tested, three failures were recorded at
2
3,374, 3,521 and 3,781 hours. The test was truncated at
4,000 hours. Because a long time passed between the
0
last of the three failures and termination of the test,
0 1,000 2,000 3,000 4,000 these results are treated as if they are a censored sam-
Hours
ple with four failures, with the fourth failure occur-
Note: Identical to Figure 2 in William Q. Meeker, Necip Doganaksoy
and Gerald J. Hahn, “Using Degradation Data for Product Reliability ring at 4,000 hours.
Analysis,” Quality Progress, June 2001, p. 60. The weights for (n = 15, m = 4) are given in Table 1
(p. 62). The computations of the estimates of the para-
meters u and b are illustrated in Table 2 (p. 62). Table 2
After specifying t and R and calculating u and b, the indicates the estimates of the extreme value parame-
user calculates VR for the sample data and compares it ters are u = 8.403 and b = 0.08327. From equations 2a
to the values in the table to determine a range of confi- and 2b, the Weibull parameters are δ = 4,462 and β =
dence levels for the specified R. The tables of confi- 12.0. From equation 1c the calculated reliability for
dence bounds are published in references 4 and 5. 5,000 hours of operation (the probability of operating
without failure for 5,000 hours) is 0.0198.
The example Figure 2 is the Weibull probability graph for case 1.
The example to illustrate the Mann and MFS The straight line is the plot of the reliability function
methods was taken from a Quality Progress article by
Meeker, Doganaksoy and Hahn6 (whom I will refer
with β = 12.0 and δ = 4,462. The four dots in the lower
to as MDH) that was based upon an example by
left-hand corner are the four failures with the abscissa
William Q. Meeker and L.A. Escobar, 7 in which a
at n(t) and the ordinate based on the widely used
similar analysis was performed on data generated
median rank plotting points
by an overstress truncated life test, using the lognor-
mal distribution.
The MFS article included three different cases, each .
one representing a different way of projecting the
results of the test. To help better illustrate the use of
the technique, I will perform Weibull analysis via the
Mann-MFS methodology for the same three cases and FIGURE 2 Weibull Reliability Graph for Case 1
add two additional cases.
The device, a gallium arsenide laser, was required 2
to operate at accelerated test conditions of a high tem- Ln[-Ln(R(t))]
1
perature of 80° C for 5,000 hours. The actual required
lifetime is 200,000 hours at nominal 20° C. 0
Based on the fact that a higher temperature is used, -1
an engineering judgment makes the acceleration fac-
-2
tor 40, so the required lifetime under accelerated test
conditions is reduced to 5,000 hours. -3
As the device ages, more electrical current is needed -4
to obtain the rated light output. Failure is defined as that 8.0 8.1 8.2 8.3 8.4 8.5
point of time at which a 10% increase in current input is Ln(t)
needed to sustain the rated light intensity. Fifteen units

QU A L I T Y P R O G R E S S I A U G U S T 2 0 0 2 I 61
H O W T O E S T I M AT E T H E PA R A M E T E R S O F A W E I B U L L D I S T R I B U T I O N

If we use the same technique as in case 1, the


TABLE 1 Weights for Case 1 weights taken from Table A-7 of reference 4 are
a(15,6,i) and c(15,6,i) respectively. Table 3 is a work-
i a(15,4,i) c(15,4,i) sheet with the calculations of the parameters.
1 -0.329324 -0.241651
The extreme value parameters are u = 8.658 and b =
0.187. The corresponding Weibull parameters are δ =
2 -0.301829 -0.234806
5,754 and β = 5.358. The calculated reliability, the
3 -0.252948 -0.213548 probability of running 5,000 hours with no failures, is
4 1.884101 0.690005 0.624. For confidence assessment, the calculated value
of VR is
From:
1. Nancy R. Mann, “Tables for Obtaining the Best Linear Invariant Estimates
of the Weibull Distribution,” Technometrics, Vol. 9, 1967, pp. 629-646. .
2. Mitchell Locks, Reliability, Maintainability and Availability Assessment,
second edition (Milwaukee: ASQ Quality Press, 1995), Table A-7. For n = 15, m = 6 and R = 0.9 from Table A-8a in ref-
erence 4, the 2% confidence bound is 1.29. Since 0.752
is less than 1.29, there is less than 2% confidence in a
For confidence assessment of case 1, from the equa- reliability of 0.9, and hence also for 0.95 and 0.99.
tion in the first paragraph found under the subhead The Weibull reliability graph is Figure 3. As in
“confidence bounds,” we have Figure 2, the straight line represents the Weibull prob-
ability function, and the median rank plotting points
. are for the six failures.
From Table A-8b in reference 4, the 2% confidence
bound for V0.9 for 0.9 reliability for a sample of 15 cen-
FIGURE 3 Weibull Reliability Graph for Case 2
sored at the fourth failure is 0.77. Because the calculat-
ed value of -1.37 is below this level, there is less than
2
2% confidence in a reliability of 0.9. Because the confi- Ln[-Ln(R(t))]
dence is that low for a reliability of 0.9, it is even lower 1
for 0.95 and for 0.99. 0

Case 2: degradation data—the pseudofailures -1


Since electrical current usage data were maintained -2
for all 15 units on test, MDH were able to project esti-
-3
mated failure times for each of the 12 units that did
not fail before 4,000 hours. -4
8.0 8.1 8.2 8.3 8.4 8.5
MDH projected three additional units would fail
Ln(t)
before the required lifetime of 5,000 hours—at
4,194, 4,721 and 4,995 hours. By appending those
three pseudofailures to the original three, we now Case 3: 2,000-hour degradation data
have the equivalent of a 15-unit test censored at the
As another example of a way to use the degrada-
sixth failure.
tion data, MDH projected failure times starting from
2,000 hours as if the testing had been-
truncated at 2,000 hours before an
TABLE 2 Spreadsheet for Calculating the Extreme-Value failures had occurred.
Parameters for Case 1 The seven projected pseudofailure
times up to 5,000 hours were 3,229,
I t a(15,4,i) c(15,4,i) n(t) a(15,4,I)* n(t) c(15,4,I)* n(t) 3,514, 3,742, 4,047, 4,282, 4,781 and
1 3,374 -0.32932 -0.24165 8.1238543 -2.675380181 -1.963137507 4,969 hours. The spreadsheet for cal-
2 3,521 -0.30183 -0.23481 8.1665003 -2.464886625 -1.917543274 culating the extreme-value parame-
3 3,781 -0.25295 -0.21355 8.2377438 -2.08372082 -1.759153714 ters is shown in Table 4.
The extreme-value parameters are
4 4,001 1.884101 0.69001 8.2942996 15.62729819 5.723108202
u = 8.6085 and b = 0.181, and the cor-
sums 1 0 8.403310561 0.083273707 responding Weibull parameters are δ
= 5,478 and β = 5.525. The estimated

62 I A U G U S T 2 0 0 2 I W W W . A S Q . O R G
original nominal requirement was
TABLE 3 Spreadsheet for Calculating the Extreme-Value 200,000 hours at 20° C. However, the
Parameters for Case 2 test was actually performed at 80° C.
The 5,000-hour requirement was
I t a(15,6,i) c(15,6,i) n(t) a(15,6,I)* n(t) c(15,6,I)* n(t) based on a judgment call that the
1 3,374 -0.1365 -0.1566 8.1238543 -1.108889859 -1.272171206 overstress factor was 40. By a different
2 3,521 -0.12452 -0.15656 8.1665003 -1.016876287 -1.278571789 judgment call, the overstress factor
3 3,781 -0.1034 -0.14752 8.2377438 -0.851790947 -1.215207253 could just as easily have been 50, so
4 4,194 -0.07561 -0.13218 8.3414102 -0.630727392 -1.102584285
the requirement could be changed to
4,000 hours.
5 4,721 -0.04168 -0.11122 8.4597759 -0.35260346 -0.940853979
In this case, it is not necessary to
6 4,995 1.481712 0.70407 8.5161927 12.6185449 5.996029853 use projected pseudofailures, as in
sums 1.000001 0 8.65765696 0.186641341 cases 2 and 3. The relevant calculations
are in the spreadsheet for case 1. From
Table 2, the extreme value parameters
reliability, the probability of not failing before 5,000 are u = 8.403 and b = 0.08327, and the Weibull parame-
hours, is 0.5467. For confidence assessment ters are δ = 4,462 and β = 12.0. The estimated reliability,
however, is 0.764 rather than 0.0198 as it was in case 1,
. where the reliable life requirement was 5,000 hours.
Since this is less than the V0.9 of 1.33 for n = 15, m =
7 at the 2% level, as in cases 1 and 2, there is less than FIGURE 4 Weibull Reliability Graph for Case 3
2% confidence in a reliability of 0.9 or higher. Figure 4
is the Weibull reliability graph of case 3. 2
Ln[-Ln(R(t))]
Cases 4 and 5: changing the reliable lifetime requirement 1
Cases 4 and 5 did not originate with MDH but 0
are based on their data. They illustrate what could
-1
happen if the reliable lifetime requirement is
changed. Because a lower reliable life is required, -2
the reliability estimate is higher. This also helps -3
better illustrate reliability-confidence assessment
-4
by the MFS tables.
8.0 8.1 8.2 8.3 8.4 8.5
Suppose the specification is changed so a lifetime of Ln(t)
4,000 hours instead of 5,000 hours is required. This
might be justified because the choice of a reliable life-
time is arbitrary. For example, in this problem the Another consequence of lowering the reliable life-
time requirement is that more mean-
ingful confidence assessment could
TABLE 4 Spreadsheet for Calculating the Extreme-Value be performed for higher levels of reli-
Parameters for Case 3 ability. In cases 1, 2 and 3, VR was
always less than the 2% confidence
I t a(15,7,i) c(15,7,i) n(t) a(15,7,I)* n(t) c(15,7,I)* n(t) bound. In case 4
1 3,229 -0.090036 -0.131891 8.0799278 -0.727484377 -1.065669754
.
2 3,514 -0.08085 -0.133342 8.1645103 -0.660100655 -1.088672128
3 3,742 -0.065446 -0.127335 8.2273755 -0.538448817 -1.04763286 For V0.9 with n = 15, m = 4, the 5%
4 4,047 -0.045441 -0.116138 8.3057311 -0.377420729 -0.964611004 confidence bound is 1.26, and the
5 4,282 -0.021137 -0.100291 8.3621755 -0.176751303 -0.83865094
10% bound is 1.59. Therefore, a relia-
bility of 0.9 is in the confidence range
6 4,781 0.007597 -0.079774 8.472405 0.064364861 -0.675877637
between 5% and 10%. For V0.95, the
7 4,969 1.295312 0.688771 8.5109739 11.02436661 5.862111998 2% bound is 1.77. Therefore, there is
sums 0.999999 0 8.608525593 0.180997675 less than 2% confidence in a reliabili-
ty of 0.95.

QU A L I T Y P R O G R E S S I A U G U S T 2 0 0 2 I 63
H O W T O E S T I M AT E T H E PA R A M E T E R S O F A W E I B U L L D I S T R I B U T I O N

In case 5, by further reducing the reliable lifetime 5. Nancy R. Mann, Kenneth W. Fertig and Ernest M.
requirement, we not only improve the estimate of the Scheuer, “Confidence and Tolerance Bounds and a New
reliability but also get to exercise the confidence assess- Goodness of Fit Test for Two-Parameter Weibull or Extreme-
ment tables more deeply. As with case 4, the relevant Value Distributions,” Dayton, OH, Aerospace Research
calculations are all in the spreadsheet for case 1. Let the Laboratories, U. S. Air Force, ARL 71-0077, 1971.
reliable lifetime requirement be reduced to 3,000 hours. 6. William Q. Meeker, Necip Doganaksoy and Gerald Hahn,
Then the calculated reliability is 0.9915 and “Using Degradation Data for Product Reliability Analysis,”
Quality Progress, June 2001, pp. 60-65.
. 7. William Q. Meeker and L.A. Escobar, Statistical Methods for
For R = 0.9, V9 is between the 90% bound of 4.78 Reliability Data (New York: John Wiley & Sons, 1998).
and the 95% bound of 5.95; for R = 0.95, it is between
the 60% bound of 4.07 and the 75% bound of 4.99; for MITCHELL O. LOCKS is principal of LG Associates, Beverly
R = 0.99, it is between the 25% bound of 4.55 and the Hills, CA. He earned a doctorate in economics from the
40% bound of 5.40. University of Chicago. Locks is a Fellow of ASQ and an ASQ cer-
Results summary tified reliability engineer. QP

This article explains how to apply the Mann-MFS


methodology to life test data for small censored sam-
ples for Weibull and extreme value estimation and I F YOU WOULD LIKE to comment on this article,
reliability confidence assessment. please post your remarks on the Quality Progress
The MDH data, which are also censored, provided Discussion Board at www.asqnet.org, or e-mail
an opportunity to demonstrate these techniques could them to editor@asq.org.
be used effectively for the purposes intended when
the data are obtained under accelerated conditions
with measurable degradation.
Cases 1, 2 and 3 all exhibited very low reliabilities;
hence the confidence levels were extremely small for
high reliability values. It is doubtful a product that
unreliable could be on the market very long. Cases 4
Emerging Sectors
and 5 show results that could be more typical for
Education, healthcare, government and
higher reliability products.
There is a difference in the interpretation of the term service industries are increasingly adopting the
“confidence” between the Mann-MFS methodology and tools of quality so common to manufacturing.
regression based or maximum likelihood techniques such Quality Progress magazine, therefore,
as those used by MDH. For MDH, a confidence interval
features a monthly column called “Emerging
is a range of values for the reliability. For Mann-MFS, the
confidence bound is a benchmark, a point on the proba- Sectors.”
bility distribution for a fixed high reliability value. We invite readers to submit articles on how
quality tools have been successfully implemented
NOTE
in these sectors. Submissions should be typewrit-
Calculations on the tables in this article were done with
Excel spreadsheets. ten, double-spaced and limited to 1,500 words.

REFERENCES
Submit your questions or article to:
1. Wilfrid J. Dixon, “Estimates of the Mean and Standard
Deviation of a Normal Population,” Annals of Mathematical Susan Daniels, Associate Editor
Statistics, Vol. 28, 1957, pp. 806-809. ASQ/Quality Progress
2. Wilfrid Dixon and F.J. Massey, Introduction to Statistical PO Box 3005
Analysis, second edition (New York: McGraw Hill, 1957). Milwaukee, WI 53201-3005
3. Nancy R. Mann, “Tables for Obtaining the Best Linear 800-248-1946 OR 414-272-8575
Invariant Estimates of the Weibull Distribution,” Technometrics, E-mail: sdaniels@asq.org
Vol. 9, 1967, pp. 629-646.
4. Mitchell Locks, Reliability, Maintainability and Availability
Assessment, second edition (Milwaukee: ASQ Quality Press, 1995).

64 I A U G U S T 2 0 0 2 I W W W . A S Q . O R G

You might also like