ANSI/IEEE Std 930-1987

An American National Standard

IEEE Guide for the Statistical Analysis of

Electrical Insulation Voltage Endurance
Data

Sponsor
Statistical Technical Committee
of the
IEEE Dielectrics and Electrical Insulation Society

Approved June 19, 1986

IEEE Standards Board

Approved November 17, 1986

American National Standards Institute

© Copyright 1987
The Institute of Electrical and Electronics Engineers, Inc
345 East 47th Street, New York, NY 10017, USA
No part of this publication may be reproduced in any form, in an electronic retrieval system or otherwise, without the
prior written permission of the publisher.
IEEE Standards documents are developed within the Technical Committees of the IEEE Societies and the Standards
Coordinating Committees of the IEEE Standards Board. Members of the committees serve voluntarily and without
compensation. They are not necessarily members of the Institute. The standards developed within IEEE represent a
consensus of the broad expertise on the subject within the Institute as well as those activities outside of IEEE which
have expressed an interest in participating in the development of the standard.

Use of an IEEE Standard is wholly voluntary. The existence of an IEEE Standard does not imply that there are no other
ways to produce, test, measure, purchase, market, or provide other goods and services related to the scope of the IEEE
Standard. Furthermore, the viewpoint expressed at the time a standard is approved and issued is subject to change
brought about through developments in the state of the art and comments received from users of the standard. Every
IEEE Standard is subjected to review at least once every Þve years for revision or reafÞrmation. When a document is
more than Þve years old, and has not been reafÞrmed, it is reasonable to conclude that its contents, although still of
some value, do not wholly reßect the present state of the art. Users are cautioned to check to determine that they have
the latest edition of any IEEE Standard.

Comments for revision of IEEE Standards are welcome from any interested party, regardless of membership afÞliation
with IEEE. Suggestions for changes in documents should be in the form of a proposed change of text, together with
appropriate supporting comments.

Interpretations: Occasionally questions may arise regarding the meaning of portions of standards as they relate to
speciÞc applications. When the need for interpretations is brought to the attention of IEEE, the Institute will initiate
action to prepare appropriate responses. Since IEEE Standards represent a consensus of all concerned interests, it is
important to ensure that any interpretation has also received the concurrence of a balance of interests. For this reason
IEEE and the members of its technical committees are not able to provide an instant response to interpretation requests
except in those cases where the matter has previously received formal consideration.

Comments on standards and requests for interpretations should be addressed to:

Secretary, IEEE Standards Board

345 East 47th Street
New York, NY 10017
USA
Foreword

(This Foreword is not a part of ANSI/IEEE Std 930-1987, IEEE Guide for the Statistical Analysis of Electrical Insulation Voltage
Endurance Data.)

This document is the Þrst version of a guide to promote standardized methods for the statistical analysis of data from
voltage endurance tests on electrical insulation. Simple methods to calculate conÞdence intervals have received special
attention. We strongly encourage users of this guide to calculate conÞdence intervals for their test data, as this will help
interested parties in evaluating the precision of the tests. Much work remains to be done in extrapolating the results of
tests from accelerated conditions to normal operating stress (Section 6).

The following members of the Statistics Technical Committee of the IEEE Dielectrics and Electrical Insulation
Society have contributed to the development of this guide

H. Rosen, Chair

P. E. Alexander J. C. Fothergill G. A. Vincent

A. Bulinski H. N. Galpern B. E. Ward
T. L. Claussen J. F. Lawless W. D. Wilkens
T. W. Dakin W. Nelson
E. M. Fort G. C. Stone

The following persons were on the balloting committee that approved this document for submission to the IEEE
Standards Board:

G. C. Stone, Chair

P. E. Alexander E. M. Fort W. Nelson

A. Bulinski J. C. Fothergill H. Rosen
T. L. Claussen H. N. Galpern G. A. Vincent
T. W. Dakin J. F. Lawless B. E. Ward

When the IEEE Standards Board approved this standard on June 19, 1986, it had the following membership:

John E. May, Chair

Irving Kolodny, Vice Chair
Sava I. Sherr, Secretary

James H. Beall Joseph L. Koepfinger* Martha Sloan

Fletcher J. Buckley Edward Lohse Oley Wanaselja
Paul G. Cummings Lawrence V. McCall J. Richard Weger
Donald C. Fleckenstein Donald T. Michael* William B. Wilkens
Jay Forster Marco W. Migliaro Helen M. Wood
Daniel L. Goldberg Stanley Owens Charles J. Wylie
Kenneth D. Hendrix John P. Riganati Donald W. Zipse
Irvin N. Howell Frank L. Rose
Jack Kinn Robert E. Rountree

* Member emeritus

iii
CLAUSE PAGE
1. Scope and References .........................................................................................................................................1

1.1 Scope .......................................................................................................................................................... 1

1.2 References .................................................................................................................................................. 1

2. Introduction .........................................................................................................................................................2

2.1 Steps Required to Analyze Voltage-Aging Data ....................................................................................... 2

3. Data Analysis ......................................................................................................................................................3

3.1 Extreme-Value Distributions ..................................................................................................................... 3

3.2 The Lognormal Distribution ...................................................................................................................... 4
3.3 Censored Data ............................................................................................................................................ 5
3.4 Probability Graph Papers ........................................................................................................................... 5

4. Estimation of Distribution Parameters ................................................................................................................9

4.1 Graphical Parameter Estimates .................................................................................................................. 9

4.2 Accurate Parameter Estimates.................................................................................................................. 12
4.3 Confidence Intervals for Parameters........................................................................................................ 14
4.4 Confidence Intervals for Percentage Failed ............................................................................................. 21

5. Comparison Tests..............................................................................................................................................23

5.1 Rigorous Hypothesis Tests....................................................................................................................... 23

5.2 Simplified Method to Compare Percentiles of Extreme Value Distributions.......................................... 23

6. Failure Models ..................................................................................................................................................25

6.1 Inverse Power Model ............................................................................................................................... 26

6.2 Exponential Model ................................................................................................................................... 26
6.3 Fitting Data to Failure Models ................................................................................................................. 26

Annex A (Informative) Program to Estimate Weibull and Gumbel Parameters ..........................................................29

Annex B (Informative) Examples .................................................................................................................................31

iv
An American National Standard

IEEE Guide for the Statistical Analysis of

Electrical Insulation Voltage Endurance
Data

1. Scope and References

1.1 Scope

This guide describes, with examples, statistical methods to analyze the data for time-to-failure from constant-stress
voltage endurance tests or breakdown voltage from progressive-stress tests on specimens or systems of electrical
insulation. Methods to compare test data are also given. The methods are principally applied to data from tests on solid
insulation, but may also apply to the analyses of data from tests on gas, liquid, and composite systems. The statistical
methods discussed here do not take into consideration the physical mechanism of voltage aging. The methods assume
that the only aging stress is alternating voltage of constant frequency. The methods may not apply if there is more than
one aging stress. Methods to ascertain the short time withstand voltage or operating voltage of an insulation system are
not presented in this guide. The mathematical techniques contained in this guide may not directly apply to the
estimation of equipment life.

1.2 References

This guide shall be used in conjunction with the following publications:

[1] IEEE Std 101-1972 (R 1980) , IEEE Guide for the Statistical Analysis of Thermal Life Test Data.1

[2] BROOKES, A. S. The Weibull Distribution: Effect of Length and Conductor Size of Test Cables. Electra no 33, p
49.

[3] COCHRAN, W. G. and SNEDECOR, G. W. Statistical Methods. Iowa State, 1976, 6th ed.

[4] DAKIN, T. W. and STUDNIARZ, S. A. The Voltage Endurance of Cast Epoxy Resins. 1978 IEEE International
Symposium on Electrical Insulation. Philadelphia, PA: p 216.

[5] DISSADO, L. A., FOTHERGILL, J. C., HILL, R. M., and WOLFE, S. V. Weibull Statistics in Dielectric
Breakdown, Theoretical Basis, Applications, and Implications. IEEE Transactions, EI-19, June 1984, p 227.

[6] ENDICOTT, H. S. and WEBER, K. H. Area Effect and its Extremal Basis for the Electric Breakdown of
Transformer Oil. AIEE Transactions pt III, vol 75, June 1956, p 371.

1 IEEE publications are available from IEEE Service Center, 445 Hoes Lane, PO Box 1331, Piscataway, NJ 08855-1331.

Copyright © 1998 IEEE All Rights Reserved 1

ANSI/IEEE Std 930-1987 IEEE GUIDE FOR THE STATISTICAL ANALYSIS OF

[7] FURTIG, K. W., MANN, N. R., and SCHEURER, E. W. ConÞdence Bounds and a New Goodness-of-Fit Test for
the Two-Parameter Weibull or Extreme-Value Distributions with Tables for Censored Samples of Size 3(1)25.
Aerospace Research Laboratories Report, ARL 71-0077, Wright-Patterson AFB, 1971.

[8] GLADSTEIN, D., NELSON, W., and SCHMEE, J. ConÞdence Limits for Parameters of a Normal Distribution
from Singly Censored Samples Using Maximum Likelihood. Technometrics, May 1985.

[9] HAHN, G. J. and NELSON, W. A Comparison of Methods for Analyzing Censored Life Data to Estimate
Relationships Between Stress and Product Life. IEEE Transactions on Reliability, R-23, April 1974, p 2.

[10] KIERSZTYN, S. E. Formal Theoretical Foundation of Electrical Aging of Dielectrics. IEEE Transactions (Power
Apparatus and Systems), PAS-100, Nov 1981, p 4333.

[11] LAWLESS, J. F. Statistical Models and Methods for Lifetime Data. New York: John Wiley and Sons, 1982.

[12] LAWLESS, J. F. and STONE, G. C. The Application of Weibull Statistics to Insulation Aging Tests. IEEE
Transactions, EI-14, Oct 1979, p 233.

[13] NATRELLA, M. G. Experimental Statistics. NBS Handbook, 91, 1966, pp 2Ð14.

[14] NELSON, W. Applied Life Data Analysis. New York: John Wiley and Sons, 1982.

[15] ROSEN, H., STONE, G. C. Some Graphical Techniques for Estimating Weibull ConÞdence Intervals IEEE
Transactions on Reliability, R-33, Dec 1984, p 362.

[16] SIMONI, L. Voltage Endurance of Electrical Insulation. Bologna: Technoprint Publ, Dec 1974.

2. Introduction

2.1 Steps Required to Analyze Voltage-Aging Data

2.1.1

Select a probability distribution that adequately represents the test data. Data on solid insulation tests are most often
represented by the (two-parameter) Weibull distribution [5]2 whereas the Gumbel (smallest extreme-value)
distribution is often employed for liquids [6]. The lognormal distribution may also be relevant. The adequacy of the
assumed distribution should always be tested, since another distribution could possibly Þt better, unless experience or
theory suggests a particular distribution.

2.1.2

Estimate the distribution parameters and their conÞdence intervals. The conÞdence interval objectively quantiÞes the
uncertainty in a parameter estimate. Generally the more data, the narrower are these conÞdence intervals.

2.1.3

Perform a hypothesis test if data are to be compared. A hypothesis test is an objective method of ascertaining if the
probability distributions of two data sets differ by a convincing amount, or whether such a conclusion is possible due
to the inherent variability of the test data.

2 The numbers in brackets correspond to those of the references in 1.2.

2 Copyright © 1998 IEEE All Rights Reserved

ELECTRICAL INSULATION VOLTAGE ENDURANCE DATA ANSI/IEEE Std 930-1987

2.1.4

Estimate the life distribution at normal operating stress by extrapolating the accelerated aging test data. This is
particularly prone to error and should be done with care, as the mathematical model, test conditions, etc, may be in
error. Any conclusions derived from statistical analysis of test data are rigorously valid only for the test conditions and
particular specimens used. In general, voltage endurance data are best used to make comparisons between different
groups of specimens at the accelerated test levels.

3. Data Analysis

This section presents the Weibull and Gumbel probability distributions. These are the distributions most often
employed to represent breakdown times, or breakdown voltages of insulation, or both. The lognormal probability
distribution is also described since it is sometimes used in data analysis, especially in connection with failure models
(Section 6). Also described are some of the graphical methods for statistical analysis.

3.1 Extreme-Value Distributions

The Weibull and Gumbel distributions are two types of extreme-value distributions. The extreme-value distributions
have been found to represent a wide variety of failure data [2], [11], and [14].

Furthermore, the effect of the size of the test specimens (that is, thickness, area, or volume, or a combination of these)
on life or breakdown voltage can easily be modeled with these distributions [2] and [14].

Statistical techniques commonly used for estimating normal distribution quantities do not apply to the extreme-value
distributions. In particular, the estimates of the mean standard deviation, and the conÞdence intervals and hypothesis
tests for the normal distribution do not apply to extreme-value distributions.

3.1.1 The Weibull Distribution

The Weibull distribution is most often used to represent both the times-to-breakdown (obtained from a constant-stress
test) and the breakdown voltages (obtained from a progressive-stress test) in tests on solid insulation. The Weibull
cumulative distribution function for the population fraction below x is

b
Ð( x ¤ a )
F( x) = 1 Ð e ,x³0 (1)

where

a = scale parameter and is positive

b = shape parameter and is positive
x = random variable, usually the time to breakdown or the breakdown voltage
F(x) = probability of failure at voltage or time (less than or equal to) x

The probability of failure F(x) is zero at x = 0. The probability of failure rises continuously as x increases. As the time
or voltage increases, the probability of failure approaches certainty, that is, F(¥) = 1.

The scale parameter a represents the time (or voltage) for which the failure probability is 0.632. It is analogous to the
mean of the normal distribution.3 The units of a are the same as x, that is, voltage, electric stress, or time.

3 Note that a mean and a standard deviation can be determined for the Weibull distribution. These quantities, however, are seldom used.

Copyright © 1998 IEEE All Rights Reserved 3

ANSI/IEEE Std 930-1987 IEEE GUIDE FOR THE STATISTICAL ANALYSIS OF

The shape parameter b is a measure of the spread of the failure times or voltages. The larger b is, the smaller is the
range of breakdown voltages or times.

The two-parameter Weibull distribution of Eq 1 is a special case of the three-parameter Weibull distribution that has
the cumulative distribution function
b
Ð[ ( ( x g ) ¤ a ) ]
F( x) = 1 Ð e ,x³g
= 0, x < g (2)

The additional term g is called the location parameter. F(x) = 0 for x < g, that is, the probability of failure for x < g is
zero. The value of g is usually assumed to be zero in practical applications.

3.1.2 The Gumbel Distribution

The Gumbel distribution is most often used to represent the breakdown voltages of liquid and solid insulations [6]. The
cumulative Gumbel (smallest extreme-value) distribution function for the population fraction below y is

( y Ð u) ¤ b
Ðe
G( y) = 1 Ð e ,Ð¥£y£¥ (3)

where

u = location parameter and may have any value

b = scale parameter and is positive
y = random variable, usually the breakdown voltage
G(y) = probability of failure at voltage or time £ y

The Gumbel distribution is asymmetrical and can have a physically impossible Þnite probability of breakdown for y <
0. This distribution is also called the smallest extreme-value (that is, weakest-link) distribution. If y is voltage, then the
units of u and b are also voltage.

The Gumbel distribution is closely related to the Weibull distribution. That is, if x has a Weibull distribution then y =
ln (x) has a Gumbel distribution

where
u = ln ( a )
b = 1¤b (4)

Estimation techniques (see Section 4) for one distribution (Gumbel or Weibull) apply to the other if the transformation
equation, Eq 4, is utilized.

3.2 The Lognormal Distribution

The lognormal distribution has sometimes been used to represent failure data from insulation systems, but it has not
been used nearly as often as the extreme-value distributions in 3.1. However, since this probability distribution is a
simple logarithmic transformation of the well-known normal or Gaussian distribution, methods for data analysis are
available in all standard statistical references. The probability density function of the lognormal distribution is

2 2
[ Ð ( z Ð u ) ¤ 2s ]
f ( z ) = ( 1 ¤ 2ps )e (5)

4 Copyright © 1998 IEEE All Rights Reserved

ELECTRICAL INSULATION VOLTAGE ENDURANCE DATA ANSI/IEEE Std 930-1987

where

z = log x
x = breakdown voltage or time-to-failure
m = logarithmic mean
s = logarithmic standard deviation

The cumulative density function is the integral of the above. There is no closed-form equation for the integral. Values
of the distribution are found in [3], [13], or can be obtained from statistical calculators or computer programs.

3.3 Censored Data

Censoring of data occurs when n specimens are started on the test together and the failures of only r (< n) are observed.

Censoring is encountered mainly with constant stress tests, where the data are analyzed or the test is terminated before
all the specimens fail. Censoring can also occur with progressive stress tests where ßashovers or spurious breakdowns
occur on specimens that would otherwise survive to high stresses. Censoring can occur by plan or by accident in many
insulation tests, and this factor must be taken into account in the data analysis.

There are also other kinds of censoring, such as progressive and multiple censoring. These types of censoring result
when failed specimens are disqualiÞed because of specimen failure by spurious mechanisms. Analytical treatment of
such data is difÞcult. See [11] and [14].

3.4 Probability Graph Papers

Special graph papers for the Weibull and Gumbel distributions are shown in Figs 1, 2, 4, and 5 respectively. One axis
of the graph paper is a nonlinear cumulative probability scale (This scale is the same for the Weibull and the Gumbel
papers). The other axis is the data axis; logarithmic for the Weibull distribution and linear for the Gumbel distribution.
This axis is for the failure times or voltages. The axes are scaled so that plotted data from the two-parameter Weibull
and Gumbel distributions tend to follow a straight line. Such graph paper is commercially available.4

Similar graph paper is also available for the lognormal distribution. The probability scale is the same as for normal
distribution graph paper. The other axis however, is logarithmic. If lognormal paper is not available, normal paper can
be employed by taking logarithms of the data and plotting them on the linear scale.

The primary uses of the probability plotting paper are to assess the adequacy of the assumed distribution to represent
the test data and to obtain rough parameter estimates. Linearity on probability paper is the only easy method to assess
which of the Weibull, Gumbel, or lognormal distributions best represents the data. All test data should Þrst be plotted
to check that the assumed distribution is adequate.

3.4.1 Weibull Probability Paper

Samples of Weibull probability paper are shown in Figs 1 and 2. To use this paper, order the failure times or voltages
from smallest to largest. The most commonly used approximation of the cumulative probability Fi is

F i = æ ------------ö × 100%
i
è n + 1ø
(6)

The Weibull example data in Table 1 are plotted in Figs 1 and 2. Data from unfailed specimens should not be plotted,
although they should be included in n.

4 TEAM, Box 25, Tamworth, NH 03886.

Copyright © 1998 IEEE All Rights Reserved 5

ANSI/IEEE Std 930-1987 IEEE GUIDE FOR THE STATISTICAL ANALYSIS OF

Table 1ÑBreakdown Data on Epoxy Specimens

Specimen Breakdown
No Fi Time ti
i % (h)

1 10 15.3

2 20 30.3

3 30 48.5

4 40 89.4

5 50 90.4

6 60 105.7

7 70 144.9

8 80 Ñ

9 90 Ñ

If a reasonably straight line is produced by the plotted data, then it may be reasonable to assume that the data are
adequately represented by the Weibull distribution. Some random deviations from a straight line are normally
expected. If, however, there is a consistent departure from a straight line (for example, curvature), then another
distribution may Þt the data better. Figure 3 shows plots on Weibull probability paper of other distribution functions.

3.4.2 Gumbel Probability Paper

A sample of Gumbel probability paper is shown in Figs 4 and 5. The n failure voltages (or times) are ordered from
smallest to largest, and the ith smallest voltage yi can be assigned the plotting position

G i = æ ------------ö × 100%
i
è n + 1ø
(7)

The Gumbel data in Table 2 are plotted in Figs 4 and 5. If a reasonably straight line is produced by the plotted data on
the Gumbel paper, then the Gumbel distribution adequately represents the failure data. Examples of data from other
distributions are plotted on Gumbel paper in Fig 6.

3.4.3 Lognormal Probability Paper

A sample of lognormal probability paper is shown in Fig 7, and contains the data presented in Table 3. The failure data
are ordered from smallest to largest, and the ith smallest voltage xi can be assigned the plotting position

F i = æ ------------ö × 100%
i
è n + 1ø
(8)

If a reasonably straight line is produced by the plotted data on the lognormal paper, then this distribution adequately
represents the data.

6 Copyright © 1998 IEEE All Rights Reserved

ELECTRICAL INSULATION VOLTAGE ENDURANCE DATA ANSI/IEEE Std 930-1987

Figure 1ÑWeibull Plot of Failure Data With Line Fit By Eye

Copyright © 1998 IEEE All Rights Reserved 7

ANSI/IEEE Std 930-1987 IEEE GUIDE FOR THE STATISTICAL ANALYSIS OF

Figure 2ÑWeibull Plot of 90% Confidence Bounds for Percentiles and Maximum Likelihood Fitted
Line

Figure 3ÑPlots of Other Distributions on Two-Parameter Weibull Paper

8 Copyright © 1998 IEEE All Rights Reserved

ELECTRICAL INSULATION VOLTAGE ENDURANCE DATA ANSI/IEEE Std 930-1987

Table 2ÑBreakdown Voltage of Oil

Specimen Breakdown
No Gi Voltage yi
i % (kV)

1 9 5.0

2 18 5.0

3 27 5.2

4 36 5.6

5 45 5.7

6 54 5.7

7 64 5.8

8 73 5.8

9 82 Ñ

10 91 Ñ

An example of the Weibull distribution function plotted on lognormal paper is shown in Fig 8.

4. Estimation of Distribution Parameters

4.1 Graphical Parameter Estimates

The probability graph papers described in 3.4 can be used to obtain approximate estimates of the parameters. The slope
and intercept of a straight line Þtted by eye to the data yield parameter estimates. More objective estimates can be
calculated by Þtting a straight line to the plotted data using least-squares regression. Parameter estimates based on a
least-squares analysis are not as accurate as other estimates, and are therefore not preferred. There are no graphical
conÞdence intervals for these estimated parameters.

4.1.1 Weibull Graphical Estimates

Plot the test data on Weibull probability paper, as described in 3.4.1. Fit a straight line by eye to the data points. The
estimate for the scale parameter a, denoted by aö is the time (or voltage) corresponding to F(x) = 63.2%. An estimate
for the shape parameter, denoted as bö , is obtained as follows. The line labelled A-A through a point labeled origin in
Fig 1, is drawn parallel to the Þtted line. The point where this line interesects the shape scale is an estimate for b. In the
example, the estimate is bö = 1.15.

4.1.2 Gumbel Graphical Estimates

Plot the test data on Gumbel probability paper, as described in 3.4.2. Fit a straight line by eye to the data points. The
estimate of the location parameter u, denoted by ž, is the voltage corresponding to G(y) = 63.2%. An estimate of the
scale parameter b, denoted by bö , is the difference in y between the 63% and 31% points.

For the data in Table 2, plotted in Fig 4, ž = 5.84 kV. Using the voltages corresponding to the 63rd and 31st percentiles,
bö = 5.84 kV - 5.43 kV = 0.41 kV.

Copyright © 1998 IEEE All Rights Reserved 9

ANSI/IEEE Std 930-1987 IEEE GUIDE FOR THE STATISTICAL ANALYSIS OF

4.1.3 Lognormal Graphical Estimates

Plot the test data on lognormal probability paper, as described in 3.4.3. Fit a straight line by eye to the data points. An
estimate of the log mean, denoted by zö , is the voltage (or time) corresponding to F(z) = 50%. An estimate of the log
standard deviation, denoted by s, is the difference in z (= log x) between the 16% and 50% points. Note that z is
dimensionless, though it can be converted back to time or voltage by taking the antilog.

For the data in Table 3, plotted in Fig 7, z = 1.113 (x = 13 h) and s = 0.16.

Figure 4ÑGumbel Probability Plot of Data In Table 2 With The Line Fitted By Eye

10 Copyright © 1998 IEEE All Rights Reserved

ELECTRICAL INSULATION VOLTAGE ENDURANCE DATA ANSI/IEEE Std 930-1987

Table 3ÑBreakdown Data

Specimen Breakdown
No Fi Time xi Log (time)
i % (h) zi

1 10 7.0 0.84

2 20 8.5 0.93

3 30 11 1.04

4 40 12 1.08

5 50 12 1.08

6 60 17 1.23

7 70 18 1.25

8 80 18 1.25

9 90 21 1.32

Figure 5ÑGumbel Plot of 90% Confidence Bounds on Maximum Likelihood Fitted Line
Copyright © 1998 IEEE All Rights Reserved 11
ANSI/IEEE Std 930-1987 IEEE GUIDE FOR THE STATISTICAL ANALYSIS OF

4.2 Accurate Parameter Estimates

Graphical estimates of the distribution parameters are not the most accurate. There are, however, more accurate
methods, which also yield conÞdence intervals. For the extreme-value distributions, the most widely employed
parameter estimates are the maximum likelihood estimates [11] and [14]. The parameter estimates for the lognormal
distribution are based on the well-known formulae used for the normal distribution.

4.2.1 Weibull Distribution

For the two-parameter Weibull distribution, the maximum likelihood estimate of b from singly censored data is the
iterative solution of the maximum likelihood equations (see Appendix A). A typical short BASIC program for solving
these equations on a personal computer is shown in Appendix A. This program requires an initial approximate value
for b. This is most easily obtained from a Weibull probability plot. Otherwise, an initial b of 10 may be used for
progressive-stress tests. For constant-stress tests, an initial b of 2 may apply.

The maximum likelihood estimates for the data in Table 1 are aö = 115 h and bö = 1.5 h. The line on Weibull probability
paper that Þts this data is shown in Fig 2. The line is plotted from two points obtained by substituting two values of x,
along with aö and bö into 1 and calculating the corresponding values of F(x).

4.2.2 Gumbel Distribution

There are equations for the maximum likelihood estimates for the Gumbel distribution. However, it is simpler to
transform the failure data and then use a computer program that Þts the Weibull distribution. In the Gumbel data yi are
transformed into Weibull data with

yi
xi = e
(9)

12 Copyright © 1998 IEEE All Rights Reserved

ELECTRICAL INSULATION VOLTAGE ENDURANCE DATA ANSI/IEEE Std 930-1987

Figure 6ÑPlots of Other Distributions on Gumbel Probability Paper

An initial estimate for b = 1/b can be obtained from a Gumbel probability plot. Alternatively, assume an initial b of 0.1
(that is, b = 10) or 0.5(b = 2). The program output of b and a are then converted to Gumbel parameter estimates by
using Eq 4.

For the Gumbel data in Table 2, ž = 5.73 kV and bö = 0.26 kV. With bö and ž and two sets of y values corresponding
values of G(y) can be calculated. These two points deÞne the straight line on Gumbel paper (Fig 5).

4.2.3 Lognormal Distribution

Exact estimates for the lognormal parameters are available if there is no censoring, that is, r = n. These estimates are
obtained by taking the logarithms of the failure voltages (or times) and using the transformed data with well-known
formulae for the mean and standard deviation of the normal distribution. Using the notation in Eq 5.

Sz
z = ------
n (10)
S(z Ð z )2
s = ----------------------
(n Ð 1)

Copyright © 1998 IEEE All Rights Reserved 13

ANSI/IEEE Std 930-1987 IEEE GUIDE FOR THE STATISTICAL ANALYSIS OF

Figure 7ÑLognormal Probability Plot of Data In Table 3 With The Line Fitted By Eye

The quantities are easily calculated with statistical calculators or computer software, which is widely available. If the
data are censored, then the special methods described in [13] are required.

4.3 Confidence Intervals for Parameters

If the same experiment involving the testing of many specimens is performed a number of times, the values of the
parameter estimates aö , bö ; ž, bö or x , s from each experiment differ. This variation in estimates results from the
statistical nature of insulation breakdown. Therefore, any parameter estimate differs from the true parameter value that
is obtained from an experiment involving an inÞnitely large number of specimens. Hence, it is common to give with
each parameter estimate a conÞdence interval that encloses the true parameter value with high probability. In general,
the more specimens tested, the narrower the conÞdence interval. Enough specimens should be tested so as to obtain
sufÞciently narrow conÞdence intervals for practical purposes. If the conÞdence intervals are calculated to be adequate
before all the samples have failed, the test can be aborted.

If an experiment is poorly performed, for example, if the applied voltage is not held constant in a constant stress test,
the conÞdence intervals are inaccurate. ConÞdence intervals are valid only for identically tested specimens.

14 Copyright © 1998 IEEE All Rights Reserved

ELECTRICAL INSULATION VOLTAGE ENDURANCE DATA ANSI/IEEE Std 930-1987

4.3.1 Approximate Weibull Parameter Intervals

There are various methods of estimating conÞdence intervals for Weibull parameters [11] and [14]. One method,
which is valid for censored data and all sample sizes, requires a computer program that is not generally available [12].
Many other computer programs give parameter estimates and approximate conÞdence intervals. Some of these are
outlined in references [11] and [14]. Some of the methods used in these programs may not be accurate, especially with
small sample sizes. If a particular program yields results within approximately 10% of those shown in Appendix B,
then the computation method is probably valid for the given sample size. In addition to computer programs,
conÞdence-limit tables have been calculated for the extreme-value distributions [7], [11], and [14].

Figure 8ÑPlot of a Weibull Function on Lognormal Probability Paper

The conÞdence interval tables for a and b, which cover all possible censoring fractions are very extensive [7]. For
simplicity these tables are represented as curves in Figs 9, 10, and 11. These curves are approximate and are for
experiments when up to 25 specimens are tested. Figures 9, 10, and 11 give the 90% conÞdence intervals only [15]. For
more accurate intervals, other sample sizes, conÞdence limits other than 90%, etc, a computer-based method or
reference to Tables is necessary [7], [11], [12], and [14].

Copyright © 1998 IEEE All Rights Reserved 15

ANSI/IEEE Std 930-1987 IEEE GUIDE FOR THE STATISTICAL ANALYSIS OF

Figure 9 is used to obtain factors Wl and Wu for the two-sided 90% conÞdence limits for the shape parameter b

b 1 = W 1 bö

b u = W u bö (11)

where

bl and bu= lower and upper limits, respectively, for the interval

Figure 9ÑW Factors for Calculating 90% Confidence Intervals for b and b

As seen in Fig 9, Wl and Wu are mainly functions of r, the number of failures. For the Weibull data in Table 1, Wl = 0.42
and Wu = 1.46, and since bö = 1.5 (4.2.1), the 90% conÞdence limits for b are bl = 0.42 X 1.5 = 0.63 and bu = 1.46 X
1.5 = 2.19.

16 Copyright © 1998 IEEE All Rights Reserved

ELECTRICAL INSULATION VOLTAGE ENDURANCE DATA ANSI/IEEE Std 930-1987

Figures 10 and 11 are used to calculate the 90% conÞdence interval for a
Ð Z 1 ¤ bö
a 1 = aö
Ð Z u ¤ bö (12)
a u = aö
where

al and au= lower and upper conÞdence limits, respectively

The factor Zl (Fig 10), is primarily a function of n, the number of specimens put on test. However, Zu (Fig 11), is a
function of both r, the number of specimens that actually failed, and n. For the Weibull data in Table 1, Zl = 0.75 and
Zu = 1.0. Therefore, the 90% conÞdence limits for a are al = 69 and au = 220 h.

4.3.2 Approximate Gumbel Parameter Intervals

The curves for the Weibull parameter intervals apply to the Gumbel parameters. The curves in Figs 9, 10, and 11 yield
approximate 90% conÞdence intervals.

The approximate 90% conÞdence interval for b has lower and upper limits

b 1 = bö ¤ W u
b u = bö ¤ W 1 (13)

Copyright © 1998 IEEE All Rights Reserved 17

ANSI/IEEE Std 930-1987 IEEE GUIDE FOR THE STATISTICAL ANALYSIS OF

Figure 10ÑZ1 Factor to Calculate 90% Confidence Interval for a1 and u1

Note that Wu factor from Fig 9 is for the lower bound and Wl is for the upper bound. For the Gumbel data in Table 2

Wu = 1.44
Wl = 0.45

thus

bl = 0.18 kV
bu = 0.58 kV

are the 90% conÞdence limits for b.

18 Copyright © 1998 IEEE All Rights Reserved

ELECTRICAL INSULATION VOLTAGE ENDURANCE DATA ANSI/IEEE Std 930-1987

The approximate 90% conÞdence interval for u has lower and upper limits

u 1 = uö Ð bö Z 1
u u = uö + bö Z u (14)

Figure 11ÑZu Factor to Calculate 90% Confidence Bounds for au and uu

The Z factors are obtained from Figs 10 and 11. For the Gumbel data in Table 2, Zl = 0.70 and Zu = 1.12. Thus the 90%
conÞdence interval for u has limits ul = 5.57 kV and uu = 6.04 kV.

4.3.3 Lognormal Parameter Intervals

The conÞdence intervals for the log mean and log standard deviation are easily found using the studentÕs t and X2
distributions, if there is no censoring. The 90% conÞdence interval for z has the lower and upper limits

Copyright © 1998 IEEE All Rights Reserved 19

ANSI/IEEE Std 930-1987 IEEE GUIDE FOR THE STATISTICAL ANALYSIS OF

t 0.05 *s
z 1 = z ----------------
n
t 0.05 *s
z 1 = z + ----------------
n (15)

Figure 12ÑV1 Factor for Calculating the 90% Tolerance Bounds for Gumbel and Weibull 100 pth
Percentiles

where t0.05 is obtained from a studentÕs t table, with n-1 degrees of freedom. Such a table is found in any standard
statistics textbook [3]. The lower and upper limits for the 90% conÞdence interval for the log standard deviation are
2
( n Ð 1 )s
s1 = ----------------------
2
c 0.05
(16)
2
( n Ð 1 )s
su = ----------------------
2
c 0.95
(17)

where X2 is from the X2 tables, with n-1 degrees of freedom. Such tables are in any standard statistics textbook [3].
ConÞdence intervals for the parameters when only censored data are available, can be estimated from the methods
described in [8].

20 Copyright © 1998 IEEE All Rights Reserved

ELECTRICAL INSULATION VOLTAGE ENDURANCE DATA ANSI/IEEE Std 930-1987

4.4 Confidence Intervals for Percentage Failed

Often it is of interest to estimate the time to failure or the failure voltage at a low probability of failure. The values of
x or y at a particular probability of failure are called percentiles. ConÞdence intervals for percentiles should also be
calculated. Although not often used, percentiles and their conÞdence intervals for the lognormal distribution are
estimated by the methods described in [8] and [13].

Many Weibull computer programs give estimates and conÞdence intervals (sometimes known as tolerance bounds) for
percentiles. Since some programs use conÞdence limits or approximations that are not valid for data with few failures,
programs should be checked for accuracy with the example data in Appendix B.

Approximate methods are available to calculate conÞdence intervals for percentiles, with the same theoretical basis as
the intervals in 4.3, for those without access to a valid computer program. Tables upon which the curves in Figs 12 and
13 are based are only available for the 1st, 5th, and 10th percentiles, and 90% conÞdence intervals.

Figure 13ÑVu Factor for Calculating the 90% Tolerance Bounds for Gumbel and Weibull 100 pth
Percentiles

4.4.1 Weibull Percentiles

The maximum likelihood estimate of the pth quantile, or the 100 pth percentile, for the Weibull distribution is

1 ¤ bö
xö p = aö [ Ð ln ( 1 Ð p ) ]
(18)

Copyright © 1998 IEEE All Rights Reserved 21

ANSI/IEEE Std 930-1987 IEEE GUIDE FOR THE STATISTICAL ANALYSIS OF

where

aö and bö = maximum likelihood parameter estimates

The 90% conÞdence interval for xp is

[ Ð V 1 ( p ) ¤ bö ]
x 1 ( p ) = aö
[ Ð V u ( p ) ¤ bö ] (19)
x u ( p ) = aö
where

xl (p) and xu (p)= lower and upper limits for x(p), the 100 pth percentile, respectively

The V factors are obtained from Figs 12 and 13 for the 1st, 5th, and 10th percentiles (p = 0.01, 0.05, and 0.10
respectively). From Fig 12 Vl (p) is primarily only a function of r whereas Vu (p) from Fig 13 is primarily a function
of n.

For the Weibull data in Table 1, for the Þrst percentile, Vl (0.01) = 10.4 and Vu (0.01) = 3.1. Thus the 90% conÞdence
limits are xl (0.01) = 0.11 and xu (0.01) = 14 h. The maximum likelihood estimate of the Þrst percentile is xö p = 5.3 h.

The conÞdence limits for the percentiles, together with the conÞdence interval for a (the 63rd percentile) can be
usefully displayed on Weibull probability paper. For the upper limit, plot the calculated limits (x values) corresponding
to the 1st, 5th, 10th, and 63.2 (a) percentiles on the graph paper. Join these four points with a smooth line. Similarly,
draw a line through the plotted lower conÞdence limits. Such conÞdence curves enclose any particular percentile of the
true population with 90% probability. The greater the number of specimens tested, the closer the upper and lower
curves.

4.4.2 Gumbel Percentiles

The maximum likelihood estimate of the pth quantile, or the 100 pth percentile, for the Gumbel distribution is

yö p = uö + bö ln ln æ ------------ö
1
è 1 Ð pø
(20)

where

ž and bö = maximum likelihood parameter estimates

The 90% conÞdence interval for yp is

y 1 ( p ) = uö Ð bö V 1 ( p )
y u ( p ) = uö Ð bö V u ( p ) (21)

where

yl (p) and yu (p)=the lower and upper limits for the 100 pth percentile. The Vl (p) and Vu (p) factors for the relevant
percentile are obtained from Figs 12 and 13, respectively.

For the Gumbel data in Table 2, for the 5th percentile, Vl = 6.4 and Vu = 2.0, thus the 90% conÞdence limits are yl (0.05)
= 4.1 and yu (0.05) = 5.2 kV. The maximum likelihood estimate of the 5th percentile is yp = 4.95 kV. The conÞdence
bounds for the percentiles, together with the conÞdence bounds for b, can be plotted on Gumbel paper (Fig 5).

22 Copyright © 1998 IEEE All Rights Reserved

ELECTRICAL INSULATION VOLTAGE ENDURANCE DATA ANSI/IEEE Std 930-1987

5. Comparison Tests

A common situation involves testing two or more insulation types or groups of specimens to determine which of the
two is superior. To analyze results from such a test involves testing the hypothesis to verify that there is no difference
between the probability distributions of the data for the two types of insulation. The hypothesis test usually compares
the parameters or a given percentile, for example, 63rd, 50th, or 10th. For power system insulation, it is usually of more
interest to compare low percentiles in ranges of one percent or lower.

It is easiest to compare test data from two types of insulation by plotting data sets on probability paper. However,
visual comparison of two plots is subjective.

5.1 Rigorous Hypothesis Tests

Various methods are available to test the hypothesis to verify that the theoretical distribution of one set of test data
differs signiÞcantly (convincingly) from another. Methods for the lognormal distribution are easily adapted from the
normal distribution, using the log transformation. See [3], ch 4 for an example. Rigorous hypothesis tests for the
Weibull and Gumbel distributions, described in [11] and [14], are usually difÞcult to perform and often require
sophisticated computer programs.

5.2 Simplified Method to Compare Percentiles of Extreme Value Distributions

If two data sets differ convincingly the following method is used to give an approximate assessment. Determine if the
conÞdence intervals for a chosen percentile of the two distributions overlap. If there is no overlap, for example, at the
10th percentile, then the two 10th percentiles differ signiÞcantly at the selected conÞdence level. This comparison does
not assume that the two shape parameters are equal. The conÞdence intervals for the percentiles are calculated as
described in Section 4.

It is always useful to compare sets of test data on Weibull probability paper. Plot the data from the two (or more) tests
on the same graph paper.

Copyright © 1998 IEEE All Rights Reserved 23

ANSI/IEEE Std 930-1987 IEEE GUIDE FOR THE STATISTICAL ANALYSIS OF

Table 4ÑBreakdown Stresses (kV/mm) Polyethylene for Two Manufacturing Processes

Failure Process 1 Process 2
No (Unscreened) (Screened)

1 35 39

2 35 45

3 36 49

4 40 49

5 43 53

6 43 53

7 43 53

8 46 53

9 46 55
10 48 55

11 48 57

12 48 57

13 48 57

14 48 57

15 48 61

16 51 64

17 51 64

18 51 65

19 51 67

20 57 68

24 Copyright © 1998 IEEE All Rights Reserved

ELECTRICAL INSULATION VOLTAGE ENDURANCE DATA ANSI/IEEE Std 930-1987

Figure 14ÑComparison of Data Sets

As described in 4.4, plot the 90% conÞdence bounds for percentiles for each data set. The test data in [12] and Table
4 are plotted in Fig 14 with the 90% conÞdence intervals. For percentiles above approximately 10%, the two intervals
do not overlap. Therefore, high percentiles of the two insulations differ signiÞcantly. Note that for low percentiles, the
intervals overlap, and those percentiles probably do not differ signiÞcantly. In principle, more specimens need to be
tested to show if low percentiles differ signiÞcantly.

6. Failure Models

A failure model may result in an equation that relates the test stress to the time-to-failure. In general as the test voltage
decreases, the failure times increase. The general experimental procedure is to test several specimens at two or more
voltages. Since the times-to-failure at a constant stress are variable, Þnding the best failure model is often difÞcult.

Copyright © 1998 IEEE All Rights Reserved 25

ANSI/IEEE Std 930-1987 IEEE GUIDE FOR THE STATISTICAL ANALYSIS OF

Two major classes of failure models have been suggested; the inverse power model and the exponential model.
Examples of their use can be found in [4], [9], [10], and [16]. No model has been conclusively proven to be valid, thus
caution is required in employing the results of an analysis.

6.1 Inverse Power Model

The inverse power model is described by

ÐN
L = kV (22)

where

L = time-to-failure at voltage V
V = voltage
k = constant
N = constant

Usually L is the Weibull scale parameter a, the mean or some other failure percentile. As the voltage increases, the life
decreases. The inverse power model is a good description if the data (the life at each test voltage) are plotted on log-
log graph paper, and a straight line results (Fig 15).

6.2 Exponential Model

One form of the exponential model is described by

Ð kV
L = ce (23)

where

L = time-to-failure at voltage V
V = voltage
c = constant
k = constant

The exponential model is a good description if the failure data plot as a straight line on semilog paper, with time-to-
failure the logarithmic scale, and voltage the linear scale.

Other variations on Eq 23 have been suggested [4] and [8]. Equations 22 and 23 can result in very large differences in
predicted times-to-failure when extrapolated to low-voltage stresses.

6.3 Fitting Data to Failure Models

Objective methods for estimating the parameters of a failure model are based upon transforming the equation into
linear form, and employing linear regression (curve-Þtting). Since simple linear regression requires the times-to-
failure at any particular voltage to be a symmetric distribution, such a regression on failure data that follow an extreme-
value distribution is not strictly valid. Various approximate procedures, generally valid for tests with many samples,
are discussed in [9] and [14], ch 12.

26 Copyright © 1998 IEEE All Rights Reserved

ELECTRICAL INSULATION VOLTAGE ENDURANCE DATA ANSI/IEEE Std 930-1987

An alternative approach is to assume the failure data are lognormally distributed. The logarithms of the failure times
are then normally distributed, and thus simple linear regression may be applied. This approach is the basis of IEEE Std
101-1972 (R 1980) [1], where the Arrhenius equation is the failure model for thermal aging.

Figure 15ÑInverse Power Model

The parameters k and log c are then estimated in exactly the same way as for Eq 24.

6.3.1 Inverse Power Model

The linearized form of Eq 22 is

log L = log k Ð N log V (24)

This equation has the linear form

where

y = mx + b
y = log L
x = log V
m =ÑN
b = log k

The parameters are to be estimated by linear regression. The assumption is that at any particular test voltage V, log L
is normally distributed.

Computer software for estimating log k and N are widely available. Such software often yields conÞdence intervals for
the equation parameters. Estimates of the life at any particular voltage with conÞdence intervals, can also be obtained.
Such estimates assume that the failure mechanism is unchanged at the selected voltage.

Copyright © 1998 IEEE All Rights Reserved 27

ANSI/IEEE Std 930-1987 IEEE GUIDE FOR THE STATISTICAL ANALYSIS OF

6.3.2 Exponential Model

The linearized form of the simple exponential model of Eq 23 is

ln L = ln c Ð kV . (25)

28 Copyright © 1998 IEEE All Rights Reserved

ELECTRICAL INSULATION VOLTAGE ENDURANCE DATA ANSI/IEEE Std 930-1987

Annex A

(Informative)

Program to Estimate Weibull and Gumbel Parameters

(These Appendixes are not a part of ANSI/IEEE Std 930-1987, IEEE Guide for the Statistical Analysis of Electrical Insulation
Voltage Endurance Data, but are included for information only.)

The maximum likelihood parameter estimates for the two-parameter Weibull distribution require the iterative solution
of

A 1
-----2- Ð --- Ð C = 0
A 1 bö
and

A 1 1 ¤ bö
aö = æè ------öø
r

where

r
bö bö
å xi
kÐ1 kÐ1
Ak = ( ln x i ) + ( n Ð r )x s ( ln x s )
i=1

k = 1, 2, 3
r
1
C = ---
r å ln xi
i=1

xs = the highest common unfailed running time or test voltage

n = total number of specimens tested
r = number of specimens failed

A BASIC program for solving this equation, based on a Newton-Raphson routine, is shown in Fig A-1. A3 is required
by the program. Note that the logarithms in the program are to the base e.

Copyright © 1998 IEEE All Rights Reserved 29

ANSI/IEEE Std 930-1987 IEEE GUIDE FOR THE STATISTICAL ANALYSIS OF

Figure A-1ÑBasic Program to Calculate aö and bö

30 Copyright © 1998 IEEE All Rights Reserved

Annex B

(Informative)

Examples

The following examples are the exact 90% conÞdence bounds for the given data [11]. Valid computation methods
should yield similar results. The Weibull distribution is assumed.

Example 1:

Failure Times (minutes) 11, 22, 33, 60, 100, 133, 151, 159, 205, 319, 340, 378, 605, 629, 1060, 1132, 1279

n = 20
r = 17
aö = 575 min
bö = 0.77
bl = 0.50
bu = 1.00
al = 355
au = 1089

1st percentile:

tl = 0.05
tu = 6.7

5th percentile:

tl = 1.3
tu = 35

10th percentile:

tl = 5.4
tu = 75
ANSI/IEEE Std 930-1987 IEEE GUIDE FOR THE STATISTICAL ANALYSIS OF

Example 2:

Failure Times (minutes) 140, 160, 185, 280, 290, 295, 320, 320

n = 14
r =8
aö = 342 min
bö = 3.8
bl = 1.7
bu = 5.5
al = 299
au = 473

1st percentile:

tl = 26.4
tu = 149

5th percentile:

tl = 67.8
tu = 203

10th percentile:

tl = 102
tu = 234

32 Copyright © 1998 IEEE All Rights Reserved