IEEE Guide For The Statistical Analysis of Thermal Life Test Data
IEEE Guide For The Statistical Analysis of Thermal Life Test Data
IEEE Guide For The Statistical Analysis of Thermal Life Test Data
Sponsor
Statistics Technical Committee
of the
IEEE Dielectrics and Electrical Insulation Society
© Copyright 1988 by
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 charge
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.
IEEE Standards documents are adopted by the Institute of Electrical and Electronics Engineers without regard to
whether their adoption may involve patents on articles, materials, or processes. Such adoption does not assume
any liability to any patent owner, nor does it assume any obligation whatever to parties adopting the standards
documents.
Foreword
(This Foreword is not a part of ANSI/IEEE Std 101-1987, IEEE Guide for the Statistical Analysis of Thermal Life Test Data.)
ANSI/IEEE Std 101-1987 is substantially revised from IEEE Std 101-1972. Extensive revision was deemed necessary
to reßect the widespread access of users of this document to advanced calculators and computers. Thus the tedious
workbook approach used in previous versions was no longer necessary. However, Annex 3 of this document does
contain a detailed worked example to provide guidance to those individuals without advanced calculators or
computers.
This document was prepared by the Statistics Technical Committee of the IEEE Dielectrics and Electrical Insulation
Society. The members of the working group were as follows:
G. C. Stone, Chair
T. W. Dakin, Past Chair
The members of the balloting committee that approved this document for submission to the IEEE Standards Board
were as follows:
When the IEEE Standards Board approved this standard on September 10, 1987, it had the following membership:
* Member emeritus
iii
CLAUSE PAGE
1. Introduction .........................................................................................................................................................1
5. Bibliography......................................................................................................................................................19
iv
An American National Standard
1. Introduction
This revision of IEEE Std 101-1972 describes statistical analyses for data from thermally accelerated aging tests. It
explains the basis and use of statistical calculations for an engineer or scientist. Statistical methods discussed in [2] 1
through [6] provide more information. Life test data analysis is dealt with more speciÞcally in [7] through [13]. The
subject of this guide is treated extensively in [11] and [12].
ANSI/IEEE Std 1-1986 (see Annex 1) discusses the principles for temperature rating of electrical insulation. These
principles are carried forward in ANSI/IEEE Std 98-1972 and ANSI/IEEE Std 99-1980 (see Annex 1) which
respectively outline test procedures for the experimental estimation of the life of insulating materials and systems. Life
test procedures for speciÞc insulating materials and systems are outlined in other IEEE publications (see Annex 1).
Insulation life test procedures are also described in ASTM (American Society for Testing and Materials), NEMA
(National Electrical Manufacturers Association), and IEC (International Electrotechnical Commission) standards (see
Annex 1). Also, proposed life test procedures continue to appear in the literature. All such publications assume that the
life of insulation with organic materials is a decreasing function of temperature.
Accelerated test procedures usually call for a number of specimens to be aged at each of several temperatures
appreciably above normal operating temperatures. High temperatures are chosen to produce specimen failures
(according to speciÞed failure criteria) in typically one week to one year. The test objective is to determine the
dependence of median life on temperature from the data and to estimate, by extrapolation, the median life to be
expected at service temperature. This guide presents methods for analyzing such data and for comparing test data on
different materials.
The Arrhenius equation gives the rate of a chemical reaction as a function of temperature. It has been adapted [1] to
approximate the relationship between insulation life and temperature as follows.
1The numbers in brackets refer to those of the bibliographic entries listed in Section 5..
k = D exp ( Ð E ¤ RT ) (1)
where
The median life L of insulation specimens is assumed to be inversely proportional to the chemical reaction rate k. This
yields
M ( X ) = A + BX (3)
in which
The coefÞcients A and B are estimated by Þtting the above equation to experimental data. This Þtting can be done
graphically or, more precisely, by the method of least squares. These methods are presented in sections 2 and 3.
Section 3 gives conÞdence limits that indicate the uncertainty in estimates from data. Throughout, population values
are denoted by capital letters (A, B, M(X), etc), and their sample estimates based on experimental data by lower-case
letters (a, b, m(X), etc) to distinguish them.
Theoretically, Eq 3 is valid only if a single chemical reaction and failure mode control the insulation failure
mechanism. Other reactions may occur, but if they are not dominant, application of the Arrhenius equation may still be
valid. The Arrhenius equation application is often valid in practice. Sometimes one reaction and failure mode
dominates over a temperature range, but another reaction with a different temperature coefÞcient and/or failure mode
dominates at lower or higher temperatures. Deviations from the simple Arrhenius relation may be caused by different
failure modes dominant at different temperatures or by variations in mechanical stress, with temperature, that affect
the life. Therefore, while the Arrhenius relation will often Þt insulation life-temperature data, it will not always apply.
Presented in [8] are valid analyses of such data with two or more failure modes identiÞed by examination of failed
specimens.
With these reservations, this guide assumes the Arrhenius equation applies and outlines statistical calculations to Þt
this equation to test data.
Statistical analysis will not compensate for invalid test data. The following recommendations on test procedures ensure
correct test data. Some test conditions needed for validity of the Arrhenius relation were mentioned in section 1.2.
Periodic Òproof testsÓ are often used to measure an insulation property that deteriorates. Failure to withstand the
speciÞed proof test level is then only detected after a ÒcycleÓ of aging. It is more rigorous to treat the failure as if it
occurred at the midpoint of the cycle. If there are different modes of failure, the data require special data analysis
methods [8].
In planning and carrying out life tests, one should aim toward valid failure time data. One obtains more accurate life
estimates at design temperatures from a test with a larger number of specimens at the low end of the test temperature
range and fewer at the high end and the middle of this range [11]. It is also best to select the lowest accelerated test
temperature as close to the anticipated service use temperature as is possibleÑconsistent with a reasonable estimated
test time and enough failures in the allotted test time. Further comments on planning the aging test are given in ANSI/
IEEE Std 1-1986 (see Annex 1).
2. Data Analysis
At any particular test temperature, the log times to failure will be distributed about a mean. A quick, easy, and
informative analysis of data at each test temperature is to plot the life times on log-normal probability paper, since the
distribution is assumed to be log-normal [11].
First, arrange the n failure times at a temperature in order from lowest to highest (see Table 1). Then calculate the
corresponding cumulative probability (cumulative fraction failed), Pj, for the jth failure time:
Pj = 100j/(n+1)
Then, on log-normal probability paper, plot each failure time versus its Pj, as in the example of Fig 1. Then a line is
drawn by eye to best Þt the points.
The same procedure is followed for the data at each test temperature. Use log-normal probability paper with enough
decades so that the data for all test temperatures Þt on the same graph, as shown in Fig 1.
200 1 84 0.09
200 2 84 0.18
The plotted points for each temperature should be approximately linear. The plots for the different temperatures should
be nearly parallel, indicating equal standard deviations of log life. The analysis of such data according to the Arrhenius
model is based on the assumptions that the distribution lines for the true populations are straight and parallel. Some
departure from strict linearity or parallelism may be acceptable, due to expected randomness in small samples. Gross
departure from linearity or parallelism, however, may indicate signiÞcant uncontrolled differences in test conditions,
specimens, or failure modes at the different temperatures. This would lead to incorrect estimates of life at lower
temperatures and incorrect conÞdence limits.
If the above described plotting is Þtted well by straight parallel lines, as illustrated in Fig 1, then the 50% points on the
lines give an estimate of the median life at each test temperature. An estimate s of the standard deviation s at
temperature i is the difference between the log of median life and the log of the 16% life on the line for temperature i.
It is sometimes desirable to analyze the data at individual temperatures to compare two sets of data on different
insulations or duplicate tests on the same insulation.
The sample mean and standard deviation of the log life times are calculated for each temperature. Such estimates will
generally be close to the graphical estimates. Graphs may, however, reveal important information that the calculations
do not. These calculations should be made after converting the life times Lij, of specimen j at temperature i, to their
base 10 logs:
Yij = log(Lij)
n
å Y ij (4A)
j=1
Y i = -----------------
ni
where ni is the number of specimens at that temperature i, and the sum runs over all n values at the temperature. The
subscript j denotes a specimen value within one temperature, i.
It is important to note that if the mean and standard deviation are calculated with the log values (assuming a normal
distribution for the log values) that the antilogs of the upper and lower standard deviation interval from the mean will
not be symmetric about the mean. The asymmetry will not be great, however, unless the distribution is quite wide.
The sample standard deviation of the log failure times at one temperature is
n
2
å ( Y ij Ð Y i )
0.5
si = ¤ ( ni Ð 1 )
j=1 (4B)
n n 2
2 æ ö
å Y ij Ð çè å Y ij÷ø
0.5
= ni ¤ [ ni ( ni Ð 1 ) ] (4C)
j=1 j=1
Equation 4B gives explicitly the sum of the squares of the individual deviation Yij from the mean Y i, but Eq 4C is
easier to use in calculators and computers, where sums are easily accumulated as data are entered. Some calculators
have programs for doing this. Equation 4C is prone to roundoff errors.
The calculated s should be close to the graphical estimate. The calculated values for s from the data of Table 1 are
plotted with the failure times in Fig 1. The antilogs of the log mean are plotted at the 50% line, the antilogs of the mean
plus the standard deviation at the 84% line, and the antilogs of the mean minus the standard deviation at the 16% line.
Calculation of the standard deviation at individual temperatures permits a statistical comparison of those standard
deviations. The statistical signiÞcance of their differences can be assessed with Bartlett's test, which is outlined in
[11]2.
It is often desirable to analyze data from a partially completed life test at one or more temperatures [7]. For example,
only half of a set of test specimens at one temperature may have failed. Thus, failure times are established for part of
the set and a running time for the remainder. Such data with unfailed specimens are called Òcensored.Ó By analyzing
data before all specimens fail, one can save considerable time and cost and still get reliable estimates. Many
applications require early analyses. Censored life data at a temperature are called Òsingly censoredÓ if the unfailed
units have a common running time and all the failure times are below that. Data for a temperature are called Òmultiply
censoredÓ if failure and running times differ and are intermixed. The latter type of data result when specimens are put
on test at different times.
It is very useful to graph incomplete (censored) test data on log normal probability paper, as in Fig 2. The cumulative
probability, P, is calculated for each failure time, as in section 2.1, using the total number of failed and unfailed
specimens for n. This yields a plot of the failure times in the lower tail of the sample. A line is drawn through these
points parallel to the line(s) for other temperature(s); this line estimates the distribution.
The example in Fig 2 uses the complete data at 200 °C (from Fig 1), the lowest 4 of 6 points at 175 °C, and the lowest
4 of 10 points at 150 °C. The lines were drawn parallel by eye through the data points. The 50% points on these lines
give estimates of medians that are close to those obtained in Fig 1 using the complete data.
More extensive graphical and numerical analyses of such incomplete test data are discussed in [7], [9], [10], which in
turn give further references.
As the Þrst step, plot the lives (and the antilogs of the averages of log lives, calculated in section 2.2) against
temperature on Arrhenius paper, as shown in Fig 3. Arrhenius paper3 has a log scale for time and a reciprocal Kelvin
scale for temperature. Figure 3 was plotted using the data of Table 1.
The plotted points should scatter around a straight line whose algebraic form was given in Eq 3. It is better to Þt the line
by eye to the median life times obtained from Fig 1 or to the antilogs of the log averages calculated from Eq 4A. This
line is used to obtain an estimate of the median life at a temperature, particularly, a design temperature. If a good
graphical Þt to a straight line is obtained, only slightly greater accuracy in Þtting the line is gained by the more
objective least squares regression, described in the next section. However, the least squares Þt yields much more
information, particularly conÞdence limits.
The graphical method does not provide conÞdence limits or objective statistical comparisons, such as comparing data
from two insulation systems.
This section discusses, without details, the calculation of the regression line and conÞdence limits.
1) The relation of log life to reciprocal Kelvin temperature is linear over the temperature range of interest (test
and design temperatures).
2) The specimen life times are statistically independent (ie, the time to failure of any one specimen does not
inßuence others).
3) Temperature measurement errors are negligible.
4) The specimens are randomly selected from the population of interest.
5) The random variations in log life have a normal distribution with the same standard deviation at all
temperatures in the range of interest.
Section 3.3 presents the equations for the calculations. Quantities calculated in early steps are used in the following
steps. Ordinarily one would use a standard least squares computer program to do such calculations.
The following describes both the calculations and the interpretation of results. Programs for the linear regression
calculation are available. Such programs can be adapted by converting the input data to log time and reciprocal Kelvin
temperature, respectively. Done by hand, or even with a calculator, the least squares arithmetic is tedious and error
prone. Yet one can calculate the quantities by hand following the steps in Annex 2. Also the steps can be programmed
into a computer without too much difÞculty. It is best to use an existing program, if available, for these calculations
[12].
The least squares Þtted line minimizes the sum of the squares of the deviations of the observed log lives from the line
on Arrhenius paper.
The calculation of the estimates a and b of the intercept, A, and the slope, B, of Eq 3, is outlined step-wise in sections
3.3.1 through 3.3.4. The calculation is in a form convenient for a programmable calculator or computer or hand
calculator. Section 3.3 also includes calculations for the conÞdence limits for the population line and other population
quantities.
There are statistical methods for Þtting regression lines to censored data [7], [9], [10]. Graphical methods for this are
in section 2.3.
The conÞdence limits indicate the accuracy of estimate of the Þtted regression line for mean log life, which is
discussed in 3.2.2.1 and in the calculation outline, 3.3. This estimate of the Arrhenius line differs from the population
line due to the typically small number of test data points used to calculate it.
The calculation of the conÞdence intervals uses values from the student's ÒtÓ distribution. Several t percentiles are
given in Table 2.
Two types of probability limits are described here because of their frequent application and the relative simplicity of
their calculation. The Þrst conÞdence limits applies to the mean life and the second to future individual life values.
Which is more appropriate depends on the speciÞc use of the regression line.
Equations for calculating the conÞdence limits appear in sections 3.3.5 and 3.3.6. The result of a simple calculation
appears in Annex 2 and is illustrated in Fig 4. For a speciÞed temperature, these limits will contain the true population
mean log life with the selected percent probability (conÞdence). However, individual life times do fall outside this
interval for the mean. The interval indicates the uncertainty between the Þtted line and the true Arrhenius line. The
percent conÞdence levels, CL, suggested are 90%, 95%, or 99%. Intervals with other conÞdence levels may be
calculated with the student's t value for the desired percentage from a larger t table. The higher the conÞdence, the
wider the conÞdence interval.
ConÞdence limits can be calculated for both sides, or only one side of the population median line. If one is interested
in an interval that brackets the population line (above and below the median life), select the factor t for Eqs 8a and 8b
from Table 2. However, for electrical insulation, if one is interested in reliability, one may only be interested in a lower
conÞdence limit for population median life. This is calculated by using the student's t¢ factor (from Table 2) in Eq 8b.
Note that the one-sided conÞdence limit is closer to the median estimate than the two-sided limitsÑfor the same
percent conÞdence. For example, the student's t¢ values for 97.5% conÞdence for the one-sided interval equal the t
values for 95% conÞdence for the two-sided interval.
tN-2, 90
N-2* t¢N-2, 90 t¢N-2, 95 tN-2,95 t¢N-2,99 tN-2,99
tN-2, 90
N-2* t¢N-2, 90 t¢N-2, 95 tN-2,95 t¢N-2,99 tN-2,99
Figure 4ÑGraph of Arrhenius Regression Line and Confidence Limits (Table 1 Data)
It is, of course, customary and convenient to convert the limits for mean log life back to their corresponding times in
hours or other original time units.
3.2.2.2 Prediction Limits for the Log Life of Individual Future Units
The second type of probability limits enclose the log life of a single future specimen at a speciÞed temperature, Tc. As
might be expected, these limits are signiÞcantly wider than those for the mean log life. Equations for these limits
appear in section 3.3.6.
It is probably prudent for persons using these statistical calculations to consider these conÞdence limits, since they
indicate the breadth of variation that can be expected for an individual specimen life from the mean. Since the lower
limit is usually more important, the student's t¢ value might preferably be selected for this calculation, using Eq 9B. If
desired, the two-sided limits can also be considered and calculated using the student's t value in Eqs 9A and 9B.
Unlike the conÞdence interval for the mean (see Eqs 8a and 8b), the prediction interval for an individual future life
does not become arbitrarily small as the number of test points is made increasingly large.
X = 1/(T + 273)
These conversions can be done with a calculator or computer as the data are entered.
where these sums are calculated over the total number of specimens, N, from all test temperatures. The number of
specimens at each temperature need not be the same.
Store these for future calculations. They must be accurate to at least seven signiÞcant Þgures to avoid roundoff error in
the subsequent calculations. Some calculators calculate these sums as the (X, Y) pairs are entered.
a = ( SY Ð bSX ) ¤ N (5B)
For a selected temperature Tc, the sample estimate of the population mean log life, M(Tc), is calculated from
m ( T c ) = a + b [ 1 ¤ ( T c + 273 ) ] (5C)
The antilog of m(Tc) is the estimate of median life in hours at Tc. Although only two mean estimates (preferably far
apart) are algebraically necessary to deÞne the line on an Arrhenius plot, it is usually recommended or required that
tests be run to establish the mean at three or four temperatures, since this permits checking the linearity of the Þt of the
data to the Arrhenius relation.
2 0.5
s = { S [ Y Ð ( a + bX ) ] ¤ ( N Ð 2 ) } } (6)
For a selected temperature, Tc, where Xc = 1/(Tc + 273), calculate the factor:
2
{ X c Ð ( SX ¤ N ) }
V ( T c ) = ------------------------------------------
2
SX Ð ( SX ) ¤ N
2 (7)
The sums were calculated in section 3.3.2. s is the same for all temperatures, but V(Tc) must be calculated for each
temperature, Tc, for which conÞdence limits are desired. Store these factors for future use.
The two-sided upper, mU (Tc), and lower mL (Tc) conÞdence limits for the mean log life M(Tc) are
0.5
m U ( T c ) = ( a + bX c ) + t N Ð 2 s { ( 1 ¤ N ) + V ( T c ) }
(8A)
0.5
m L ( T c ) = ( a + bX c ) + t N Ð 2 s { ( 1 ¤ N ) + V ( T c ) }
(8B)
where tN-2 is the value of the t statistic for the selected conÞdence level (for example: 90% or 95%) from Table 2.
For just a lower (one-sided) conÞdence limit, use factor t' and Eq 8b. Refer to the explanation in section 3.2.2.1.
The conÞdence limits would normally be calculated for selected service and test temperatures, since the limits are not
linear functions of X. (See Fig 4 for a plot of such limits and the calculation example in Annex 2.)
It is usually desirable to convert these log life limits back to their corresponding life times, by taking their antilogs.
These limits are wider than those for the regression line for mean life. See the discussion of this quantity in section
3.2.2.2.
Calculate:
0.5
Y U ( T c ) = ( a + bX c ) + t N Ð 2 s { 1 + ( 1 ¤ N ) + V ( T c ) } (9A)
0.5
Y L ( T c ) = ( a + bX c ) Ð t N Ð 2 s { 1 + ( 1 ¤ N ) + V ( T c ) } (9B)
where tN-2 is the same as in section 3.3.5. If only the lower prediction limit, YL (Tc) for an individual future value is
desired, t¢ is used in the equation.
It is usually desirable to convert these values of the log life limits back to life times, by taking their antilogs.
The prediction limits for an individual future value of log life, analogous to the conÞdence limits for mean log life,
may be plotted on the same graph with the Þtted regression line. The lower prediction curve is illustrated in Fig 4 for
the data from Table 1.
3.3.7 Linearity
The statistical calculations which Þt Eq 2 to the life versus temperature data assume that the straight line (Arrhenius
Equation) correctly Þts such data. Changes in the slope of the line for log life versus reciprocal absolute temperature
may indicate changes in aging mechanisms. Since extrapolation based on such data without additional considerations
may be of questionable validity, it is necessary to determine that acceptable linearity exists. See Annex 3 for a
discussion on linearity test procedures.
The following method compares two sample means (averages) of log life at the same temperature and calculated as in
section 2.2. A ÒtÓ test is used to assess whether the higher log (mean) is statistically signiÞcantly higher than the lower
log (mean).
These quantities are calculated from Eqs 4A, 4B, and 4C. The comparison assumes that the two population standard
deviations are not too different. With a difference of less than about 10%, it is reasonable to consider the two standard
deviations the same.
From the two standard deviations a ÒpooledÓ standard deviation is calculated with the following equation:
2 2 0.5
( n 1 Ð 1 )S 1 + ( n 2 Ð 1 )S 2
s p = ---------------------------------------------------------
( n1 + n2 Ð 2 ) (10)
To Þnd the signiÞcance level of this statistic, enter the t table (Table 2) at the row for the above degrees of freedom. Go
across to a value of t that most closely matches that calculated above for the difference. One-hundred percent minus the
percentage heading on this column for t¢ gives the signiÞcance level of the statistic. If this signiÞcance level is small
(<5%), then there is convincing evidence of a difference in the means.
Preferably, a conÞdence interval for the difference between the two population means may be calculated for a speciÞed
conÞdence level (for example, 95%) using the corresponding t value:
1 0.5
[ m 1 ( T c ) Ð m 2 ( T c ) ] ± t s p æ ----- + -----ö
1
è n 1 n 2ø
(12)
If this interval encloses zero, the two sample means do not differ signiÞcantly. If the interval does not enclose zero,
then the sample means differ signiÞcantly; that is, they differ convincingly. Thus one acts as if the mean log lives for
the two insulations are different. m1 (Tc) - m2 (Tc) is an estimate of how much they differ.
Thermal life tests are often used to compare two different systems or materials. Arrhenius lines are Þtted to data for
both materials. There is a question whether one insulation system is signiÞcantly better than the other, ie, are the two
lines signiÞcantly different. The observed difference might be due to chance variation. To assess whether or not there
is a convincing difference, use the following comparison between the two sets of data. The calculation is similar to that
for comparing two averages.
The comparison between the Þtted lines is made at one or more temperatures, usually at a service temperature. It may
also be useful to compare at a temperature in the test range.
The calculation utilizes some quantities obtained in section 3.3.5, namely the standard deviations s1 and s2 (Eq 6) and
V1 (Tc) and V2(Tc) (Eq 7). These are subscripted 1 and 2 to identify the insulation. The s values are independent of
temperature, but the V(Tc) values vary with temperature and must be calculated for each selected temperature, Tc,
where the comparison is made. (Annex 2 illustrates the variation in V(Tc) for Þve temperatures.)
where N1 and N2 are the numbers of specimens used to estimate the respective Þtted lines. Then calculate the mean log
life M(Tc) for each line at the selected temperature, Tc. These estimates are calculated using Eq 5c, using their
respective coefÞcients, a and b
Enter a table of t percentiles (Table 2) at the row for N1 +N2-4 degrees of freedom. Go across this row to the t
percentile most nearly equal to the t statistic. One hundred percent minus the percentage at the head of this column
gives the ÒsigniÞcance levelÓ of the t statistic.
Alternatively, conÞdence limits for the difference between the log means at the selected temperature (Tc) may be
calculated for a selected conÞdence level and corresponding t value:
1 1 0.5
m 1 ( T c ) Ð m 2 ( T c ) ± ts p ------- + V 1 ( T c ) + ------- + V 2 ( T c )
N1 N2
(15)
If these limits enclose zero, the two log means do not differ signiÞcantly. If the limits do not enclose zero, the means
are signiÞcantly different.
5. Bibliography
[1] DAKIN, T. W. Electrical Insulation Deterioration Treated as a Chemical Rate Phenomenon. AIEE Transactions,
vol 67, part 1, 1948, pp 113-122.
[2] BOX, G. E. P., HUNTER, W. G., and HUNTER. J. S. Statistics for Experimenters. New York: Wiley, 1978.
[3] SNEDECOR, G. W. and COCHRAN, W. G. Statistical Methods. Iowa State, 6th Ed, 1967.
[4] HAHN, G. J. and SHAPIRO, S.S. Statistical Models in Engineering. New York: Wiley, 1967.
[5] NETER, J., WASSERMAN, W., and KUTNER, M. H. Applied Linear Regression Models. Homewood, IL: Irwin,
1983.
[6] DRAPER, N. R. and SMITH, H. Applied Regression Analysis. New York: Wiley, 2nd Ed, 1981.
[7] NELSON, W. B. Hazard Plotting for Incomplete Failure Data. J. Quality Technol., vol 1, Jan 1969, pp 27-52.
[8] NELSON, W. B. Hazard Plotting Methods for Analysis of Life Data with Different Failure Modes. J. Quality
Technol., vol 2, July 1970.
[9] HAHN, G. J. and MILLER, J. M. Methods and Computer Programs for Estimating Parameters in a Regression
Model from Censored Data. General Electric Co., TIS Report 68-C-277, July 1968.4
[10] HAHN, G. J. and MILLER, J. M. Methods and Computer Programs for Estimating Parameters of Several
Normal Populations from Censored Data. General Electric Co., TIS Report 67-C472, December 1967.5
[11] NELSON, W. B. Analysis of Accelerated Life Test Data: I Ñ The Arrhenius Model and Graphical Methods.
IEEE Trans. Elec. Insulation, vol EI-6, Dec 1971; II Ñ Numerical Methods and Test Planning, vol EI-7, Mar 1972;
III Ñ Product Comparisons and Checks on the Validity of the Model and Data, vol EI-7, June 1972.
[12] DAKIN, T. W. Computer Programs for Calculating the Arrhenius Linear Regression Line and Related Statistics.
IEEE Trans. Elec. Insulation, vol EI-18, Apr 1983, p 172; Erratum, vol EI-18, Aug 1983, p 466.
[13] NELSON, W. B. Accelerated Testing: Statistical Models, Test Plans, and Data Analysis. New York: Wiley,
1988.
4Available from Distribution Unit, GE Research and Development Center, PO Box 43, Bldg 5, Schenectady, NY 12305.
5 See footnote 4.
Annex 1
(Informative)
List of Thermal Life Test and Related Standards for Electrical Insulation
(These Appendixes are not part of ANSI/IEEE Std 101-1987, IEEE Guide for the Statistical Analysis of Thermal Life Test Data, but
are included for information only.)
The following are lists of IEEE and IEC standards or guides to which apply the statistical techniques described in this
guide.
[A1] ANSI/IEEE Std 1-1986, IEEE Standard General Principles for Temperature Limits in the Rating of Electrical
Equipment and for the Evaluation of Electrical Insulation.6
[A2] ANSI/IEEE Std 98-1984, IEEE Standard for the Preparation of Test Procedures for the Thermal Evaluation of
Solid Electrical Insulating Materials.
[A3] ANSI/IEEE Std 99-1980, IEEE Recommended Practice for the Preparation of Test Procedures for the Thermal
Evaluation of Insulation Systems for Electrical Equipment.
[A4] ANSI/IEEE Std 117-1974 (R1985), IEEE Standard Test Procedure for the Evaluation of Systems of Insulating
Materials for Random-Wound AC Electric Machinery.
[A5] ANSI/IEEE Std 275-1981, IEEE Recommended Practice for Thermal Evaluation of Insulation Systems for AC
Electric Machinery Employing Form-Wound Pre-Insulated Coils, Machines Rated 6900 V and Below.
[A6] ANSI/IEEE Std 304-1977 (R1982), IEEE Standard Test Procedures for Evaluation and ClassiÞcation of
Insulation Systems for DC Machines.
[A7] ANSI/IEEE Std 429-1972, IEEE Standard Test Procedure for the Evaluation of Sealed Insulation Systems for AC
Electric Machinery Employing Form-Wound Stator Coils.
[A8] ANSI/IEEE Std 434-1973 (R1985), IEEE Guide for Functional Evaluation of Insulation Systems for Large High-
Voltage Machines.
[A9] ANSI/IEEE Std 620-1987, IEEE Guide for Construction and Interpretation of Thermal Limit Curves for
Squirrel-Cage Motors Over 500 hp.
[A10] ANSI/IEEE Std 930-1987, IEEE Guide for the Statistical Analysis of Voltage Endurance Data for Electrical
Insulation.
[A11] ANSI/IEEE C57.91-1981, IEEE Guide for Loading Mineral-Oil-Immersed Overhead and Pad-Mounted
Distribution Transformers Rated 500 kVA and Less with 65 °C or 55 °C Average Winding Rise.
[A12] ANSI/IEEE C57.92-1981, IEEE Guide for Loading Mineral-Oil-Immersed Power Transformers up to and
Including 100 MVA with 55 °C or 65 °C Winding Rise.
6ANSI/IEEE publications can be obtained from the Sales Department, American National Standards Institute, 1430 Broadway, New York, NY
10018, or from the Institute of Electrical and Electronics Engineers, Service Center, 445 Hoes Lane, Piscataway, NJ 08854-4150.
[A13] IEEE Std 266-1969, (R1981), IEEE Test Procedures for Evaluation of Insulation Systems for Electronics Power
Transformers.
[A1] IEC 172 (1987), Test Procedure for the Determination of the Temperature Index of Enamelled Winding Wires.7
[A2] IEC 216, Guide for the Determination of Thermal Endurance Properties of Electrical Insulating Materials.
[A3] IEC 216-1 (1987), Part 1: General Guidelines for Aging and Evaluation of Test Results.
[A4] IEC 216-2 (1984), Part 2: List of Materials and Available Tests.
[A6] IEC 216-4 (1980), Part 4: Instructions for Calculating the Thermal Endurance ProÞle.
[A7] IEC 290 (1969), Evaluation of the Thermal Endurance of Electrical Insulating Varnishes by the Helical Coil
Bond Test.
[A8] IEC 370 (1971), Test Procedure for Thermal Endurance of Insulating Varnishes--Electrical Strength Method.
7IEC publications can be obtained from Bureau Central de la Commission Electrotechnique Internationale, 3 rue de Varemb???, 1211 Geneve 20,
Suisse (Switzerland). They may also be obtained from the American National Standards Institute, 1430 Broadway, New York NY 10018.
Annex 2
(Informative)
Example Calculations
This appendix presents an example of the calculations of statistical quantities discussed above. The data from Table 1,
discussed in section 2.2, are used here to calculate the Þtted line and its conÞdence limits.
The calculations employ Table A2-1. This table organizes conversions of the input temperature and failure times to X
and Y values for calculations. The table is convenient for a hand calculator. Fill in the columns in the table. Start with
the conversion to X and Y in columns 5 and 6. Then calculate the squares and products in columns 7, 8, and 9. The sums
of these columns are then inserted into the equations for the desired quantities.
Computer calculations are usually similarly organized, but the intermediate values are usually not printed out unless
they might be useful for further calculations. The program listed in [12]8 is, however, speciÞcally designed for this
application.
If only the Þtted line is calculated, the sums for individual temperatures, as shown in Table A2-1, are not needed. These
sums may have already been calculated in analyzing data at individual temperatures as in Table 1. Also, some are used
in further statistical calculations. So it is suggested that they be calculated.
126
1 4 150 2.36406619E-03 3.10037055 5.58880897E-06 9.61229752 7.32948119E-03
0
126
1 5 150 2.36406619E-03 3.10037055 5.58880897E-06 9.61229752 7.32948119E-03
0
147
1 6 150 2.36406619E-03 3.16908636 5.58880897E-06 10.0431083 7.49192993E-03
6
147
1 7 150 2.36406619E-03 3.16908636 5.58880897E-06 10.0431083 7.49192993E-03
6
147
1 8 150 2.36406619E-03 3.16908636 5.58880897E-06 10.0431083 7.49192993E-03
6
198
1 9 150 2.36406619E-03 3.29666519 5.58880897E-06 10.8680014 7.79353473E-03
0
219
1 10 150 2.36406619E-03 3.34163234 5.58880897E-06 11.1665067 7.89984004E-03
6
NS ( XY ) Ð ( SX ) ( SY )
b = --------------------------------------------------------
2 2 (5a)
NS ( X ) Ð ( SX )
All of these sums are taken over all the N specimens (26 in this case) and appear in Table A2-1 at the bottom. Inserting
the sums into the equation gives
26 ( 0.153712413 ) Ð ( 0.0581751681 ) ( 68.1500115 )
b = ----------------------------------------------------------------------------------------------------------------------------
2
-
26 ( 0.000130479793 ) Ð ( 0.0581751681 )
3.996522738 Ð 3.964638375
= --------------------------------------------------------------------------- = 3924.5
0.0033924746 Ð 0.0033843502
Note that the differences in the numerator and denominator are in the third and fourth signiÞcant Þgure. Thus it is
essential to work with at least seven signiÞcant Þgures in intermediate calculations.
a = [ SY Ð b ( SX ) ] ¤ N (5b)
m ( T c ) = a + bX c
(5c)
By inserting values of Xc = 1/(273 + Tc) for selected temperatures, Tc, the corresponding estimates of mean log life are
calculated. Take the antilog to get the median life estimate in hours. For example, for 150 °C,
Although only two points are needed to determine the line, values of m(Tc) at the other temperatures are useful.
The standard deviation s of the data points about the Þtted line is calculated with12
2 0.5
s = { S [ Y Ð a Ð bX ] ¤ ( N Ð 2 ) } (6)
The sums are taken over all test points at all temperatures. This calculation is shown in Table A2-2. The temperatures
are tabulated in column 1, the measured Y values in column 2, and the calculated mi values are in column 3. In column
4 are the calculated values of differences between columns 2 and 3. Finally, in column 5 are the squares of the
differences in column 4. Insertion of the sum of the Þfth column into Eq 6 yields the value
0.5
s = [ 0.438733372 ¤ ( 26-2 ) ]
= 0.135206
The quanitity V(Tc) must be calculated for each temperature for which conÞdence limits are desired:13
2 2 2
V ( T c ) = [ X c Ð ( SX ) ¤ N ] ¤ [ SX Ð ( SX ) ¤ N ]
(7)
An example calculation for 150 °C is given here. Use values of the appropriate sums from Table A2-1 to calculate
2
[ ( 0.0023641 ) Ð ( 0.0581752 ¤ 26 ) ]
V ( T c ) = -----------------------------------------------------------------------------------------
2
-
( 0.000130479 ) Ð ( ( 0.0581752 ) ¤ 26 )
2
( 0.0023641 Ð 0.00223751 )
= ------------------------------------------------------------------------- = 0.0513
( 0.000130479 Ð 0.000130167 )
Note that here the difference in the denominator is in the 4th and 5th signiÞcant Þgure or the 7th and 8th Þgure after
the decimal point. This emphasizes the need to carry 8 or 9 signiÞcant Þgures in the sums of X and Y, etc.
Tc V(Tc)
100 0.6273
130 0.1903
150 0.0513
175 0
200 0.0486
Now all of the appropriate quantities are available for the calculation of the conÞdence limit values in log form. The
appropriate t and t' values for 95% conÞdence are obtained for N-2 (= 24 here) degrees of freedom from the t table in
Table 2. The upper and lower (two-sided) conÞdence limits for the regression line are14
0.5
m U ( T c ) = ( a + bX c ) + t N Ð 2 s ( ( 1 ¤ N ) + V ( T c ) )
(8a)
0.5
m L ( T c ) = ( a + bX c ) Ð t N Ð 2 s ( ( 1 ¤ N ) + V ( T c ) ) (8b)
For 150 °C, necessary values have already been calculated in this appendix. Insertion of these values yields
m U ( 150 ) = ( Ð 6.15994 + ( 3924.5 ) ¤ ( 273.16 + 150 ) )
0.5
+ 2.064 ( 0.135206 ) [ ( 1 ¤ 26 ) + 0.0513 ]
= 3.11784 + 0.08361 = 3.20145
and Eq 8b yields
The antilogs of mU(150) and mL(150) are then taken to get the limit in hours. The corresponding conÞdence limits for
median life at 150 °C are
For a comparison, the estimate of median life calculated above is 1311 hours. Such conÞdence limits have been
calculated for other temperatures for the data of Table 1. They are plotted in Fig 4. Note that the lower 95% (two-sided)
conÞdence limit calculated above is the one-sided 97.5% conÞdence limit.
3.05154543E-
150 °C 3.10037055 3.11783922 -0.174686732
04
3.05154543E-
150 °C 3.10037055 3.11783922 -0.174686732
04
1.11416814E-
175 °C 2.61066016 2.60010475 0.0105554163
04
2.73080561E-
200 °C 2.12057393 2.13709908 -0.0165251493
04
3.13885868E-
200 °C 2.12057393 2.13709908 0.0560255181
03
3.13885868E-
200 °C 2.1931246 2.13709908 0.0560255181
03
3.13885868E-
200 °C 2.1931246 2.13709908 0.0560255181
03
3.13885868E-
200 °C 2.1931246 2.13709908 0.0560255181
03
Sum Ñ Ñ Ñ 0.438733372
Limits for other percent conÞdence levels can be easily calculated by inserting the corresponding t or t' values into the
above equations. All other values in the equations remain the same; only the t value is changed.
Annex 3
(Informative)
A3.1 F-Test
The F-Test (for the statistician Fisher) [3] is based on the premise that in a truly linear relationship the variance of the
sample life times at each temperature about the empirical regression line (designated as sL2) should be the same as the
pooled variance of the sample life times at each temperature (designated as si2).
The ratio, F = sL2/si2, is compared to the tabulated values of Fcrit in Table A3-1 at the appropriate degrees of freedom
and if F > Fcrit, then the null hypothesis of linearity is rejected.
SpeciÞcally, the variance of the Y values at each temperature, i, is calculated with Eq A3-1, or obtained by squaring
each side of Eq 4C (section 2.2):
2 2
2 n i SY i Ð ( SY i )
si = -------------------------------------
ni ( ni Ð 1 )
(A3-1)
The sums are to be taken over all the Yi values at each test temperature, and the weighted mean of these variances
calculated:
2
2 S ( n i Ð 1 )s i
si = --------------------------
-
S ( ni Ð 1 )
(A3-2)
where ni are the number of specimens at each temperature, i. The variance, sL2 of the Yi values about corresponding
points, mi, on the regression line is calculated with Eq A3-3:
2 2
2 Sn i ( Y i Ð m i ) Sn i ( Y i Ð ( a + bX i ) )
s L = --------------------------------
- = -------------------------------------------------
-
IÐ2 IÐ2 (A3-3)
where a and b are the intercept and slope of the regression line and I is the number of test temperatures.
E(ni-1) 1 2 3 4 5 6 8 12 24
120 3.92 3.07 2.68 2.45 229 2.18 2.02 1.83 1.61
2 2
F = sL ¤ s1
(A3-4)
This should be compared to the values in Table A3-1 for critical values of F at the 5% signiÞcance level. Enter the table
at the column for I-2 and the row for the degrees of freedom for all individual tests values: S (ni - 1), which is the
denominator in Eq A3-2.
If the calculated F value is less than the value in Table A3-1 at I-2 and S(ni- 1), there is less than 0.05 probability that
the data do not correctly Þt a linear (ie, Arrhenius) relation. The calculated F value and the corresponding F value from
the table should be reported together with a statement of signiÞcance.
Although the F-Test procedure is reasonably simple and statistically valid, it may reject data that were acquired at
considerable expense and time, and from which useful information can nevertheless be gleaned. This is particularly the
case for tightly clustered data at each temperature (s1 is small), suggesting good experimental results but which yield
higher F values than would more widely scatteredÑand presumably poorerÑexperimental data.
In recognition of this, at least one forthcoming revision of IEC Publication 216-3 (see Annex 1, A1.2) permits a data
ÒrescueÓ operation that is effected by an artiÞcial increase of s1. The limits to this increase are determined by further
constraints, primarily by the maximum permissible difference between the empirical regression line and its lower 95%
conÞdence bound.